tools/perf/builtin-stat.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * builtin-stat.c
0004  *
0005  * Builtin stat command: Give a precise performance counters summary
0006  * overview about any workload, CPU or specific PID.
0007  *
0008  * Sample output:
0009
0010    $ perf stat ./hackbench 10
0011
0012   Time: 0.118
0013
0014   Performance counter stats for './hackbench 10':
0015
0016        1708.761321 task-clock                #   11.037 CPUs utilized
0017             41,190 context-switches          #    0.024 M/sec
0018              6,735 CPU-migrations            #    0.004 M/sec
0019             17,318 page-faults               #    0.010 M/sec
0020      5,205,202,243 cycles                    #    3.046 GHz
0021      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
0022      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
0023      2,603,501,247 instructions              #    0.50  insns per cycle
0024                                              #    1.48  stalled cycles per insn
0025        484,357,498 branches                  #  283.455 M/sec
0026          6,388,934 branch-misses             #    1.32% of all branches
0027
0028         0.154822978  seconds time elapsed
0029
0030  *
0031  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
0032  *
0033  * Improvements and fixes by:
0034  *
0035  *   Arjan van de Ven <arjan@linux.intel.com>
0036  *   Yanmin Zhang <yanmin.zhang@intel.com>
0037  *   Wu Fengguang <fengguang.wu@intel.com>
0038  *   Mike Galbraith <efault@gmx.de>
0039  *   Paul Mackerras <paulus@samba.org>
0040  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
0041  */
0042
0043 #include "builtin.h"
0044 #include "perf.h"
0045 #include "util/cgroup.h"
0046 #include <subcmd/parse-options.h>
0047 #include "util/parse-events.h"
0048 #include "util/pmu.h"
0049 #include "util/event.h"
0050 #include "util/evlist.h"
0051 #include "util/evlist-hybrid.h"
0052 #include "util/evsel.h"
0053 #include "util/debug.h"
0054 #include "util/color.h"
0055 #include "util/stat.h"
0056 #include "util/header.h"
0057 #include "util/cpumap.h"
0058 #include "util/thread_map.h"
0059 #include "util/counts.h"
0060 #include "util/topdown.h"
0061 #include "util/session.h"
0062 #include "util/tool.h"
0063 #include "util/string2.h"
0064 #include "util/metricgroup.h"
0065 #include "util/synthetic-events.h"
0066 #include "util/target.h"
0067 #include "util/time-utils.h"
0068 #include "util/top.h"
0069 #include "util/affinity.h"
0070 #include "util/pfm.h"
0071 #include "util/bpf_counter.h"
0072 #include "util/iostat.h"
0073 #include "util/pmu-hybrid.h"
0074 #include "asm/bug.h"
0075
0076 #include <linux/time64.h>
0077 #include <linux/zalloc.h>
0078 #include <api/fs/fs.h>
0079 #include <errno.h>
0080 #include <signal.h>
0081 #include <stdlib.h>
0082 #include <sys/prctl.h>
0083 #include <inttypes.h>
0084 #include <locale.h>
0085 #include <math.h>
0086 #include <sys/types.h>
0087 #include <sys/stat.h>
0088 #include <sys/wait.h>
0089 #include <unistd.h>
0090 #include <sys/time.h>
0091 #include <sys/resource.h>
0092 #include <linux/err.h>
0093
0094 #include <linux/ctype.h>
0095 #include <perf/evlist.h>
0096
0097 #define DEFAULT_SEPARATOR   " "
0098 #define FREEZE_ON_SMI_PATH  "devices/cpu/freeze_on_smi"
0099
0100 static void print_counters(struct timespec *ts, int argc, const char **argv);
0101
0102 /* Default events used for perf stat -T */
0103 static const char *transaction_attrs = {
0104     "task-clock,"
0105     "{"
0106     "instructions,"
0107     "cycles,"
0108     "cpu/cycles-t/,"
0109     "cpu/tx-start/,"
0110     "cpu/el-start/,"
0111     "cpu/cycles-ct/"
0112     "}"
0113 };
0114
0115 /* More limited version when the CPU does not have all events. */
0116 static const char * transaction_limited_attrs = {
0117     "task-clock,"
0118     "{"
0119     "instructions,"
0120     "cycles,"
0121     "cpu/cycles-t/,"
0122     "cpu/tx-start/"
0123     "}"
0124 };
0125
0126 static const char * topdown_attrs[] = {
0127     "topdown-total-slots",
0128     "topdown-slots-retired",
0129     "topdown-recovery-bubbles",
0130     "topdown-fetch-bubbles",
0131     "topdown-slots-issued",
0132     NULL,
0133 };
0134
0135 static const char *topdown_metric_attrs[] = {
0136     "slots",
0137     "topdown-retiring",
0138     "topdown-bad-spec",
0139     "topdown-fe-bound",
0140     "topdown-be-bound",
0141     NULL,
0142 };
0143
0144 static const char *topdown_metric_L2_attrs[] = {
0145     "slots",
0146     "topdown-retiring",
0147     "topdown-bad-spec",
0148     "topdown-fe-bound",
0149     "topdown-be-bound",
0150     "topdown-heavy-ops",
0151     "topdown-br-mispredict",
0152     "topdown-fetch-lat",
0153     "topdown-mem-bound",
0154     NULL,
0155 };
0156
0157 #define TOPDOWN_MAX_LEVEL           2
0158
0159 static const char *smi_cost_attrs = {
0160     "{"
0161     "msr/aperf/,"
0162     "msr/smi/,"
0163     "cycles"
0164     "}"
0165 };
0166
0167 static struct evlist    *evsel_list;
0168 static bool all_counters_use_bpf = true;
0169
0170 static struct target target = {
0171     .uid    = UINT_MAX,
0172 };
0173
0174 #define METRIC_ONLY_LEN 20
0175
0176 static volatile pid_t       child_pid           = -1;
0177 static int          detailed_run            =  0;
0178 static bool         transaction_run;
0179 static bool         topdown_run         = false;
0180 static bool         smi_cost            = false;
0181 static bool         smi_reset           = false;
0182 static int          big_num_opt         =  -1;
0183 static bool         group               = false;
0184 static const char       *pre_cmd            = NULL;
0185 static const char       *post_cmd           = NULL;
0186 static bool         sync_run            = false;
0187 static bool         forever             = false;
0188 static bool         force_metric_only       = false;
0189 static struct timespec      ref_time;
0190 static bool         append_file;
0191 static bool         interval_count;
0192 static const char       *output_name;
0193 static int          output_fd;
0194
0195 struct perf_stat {
0196     bool             record;
0197     struct perf_data     data;
0198     struct perf_session *session;
0199     u64          bytes_written;
0200     struct perf_tool     tool;
0201     bool             maps_allocated;
0202     struct perf_cpu_map *cpus;
0203     struct perf_thread_map *threads;
0204     enum aggr_mode       aggr_mode;
0205 };
0206
0207 static struct perf_stat     perf_stat;
0208 #define STAT_RECORD     perf_stat.record
0209
0210 static volatile int done = 0;
0211
0212 static struct perf_stat_config stat_config = {
0213     .aggr_mode      = AGGR_GLOBAL,
0214     .scale          = true,
0215     .unit_width     = 4, /* strlen("unit") */
0216     .run_count      = 1,
0217     .metric_only_len    = METRIC_ONLY_LEN,
0218     .walltime_nsecs_stats   = &walltime_nsecs_stats,
0219     .ru_stats       = &ru_stats,
0220     .big_num        = true,
0221     .ctl_fd         = -1,
0222     .ctl_fd_ack     = -1,
0223     .iostat_run     = false,
0224 };
0225
0226 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
0227 {
0228     if (!a->core.cpus && !b->core.cpus)
0229         return true;
0230
0231     if (!a->core.cpus || !b->core.cpus)
0232         return false;
0233
0234     if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
0235         return false;
0236
0237     for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
0238         if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
0239             perf_cpu_map__cpu(b->core.cpus, i).cpu)
0240             return false;
0241     }
0242
0243     return true;
0244 }
0245
0246 static void evlist__check_cpu_maps(struct evlist *evlist)
0247 {
0248     struct evsel *evsel, *pos, *leader;
0249     char buf[1024];
0250
0251     if (evlist__has_hybrid(evlist))
0252         evlist__warn_hybrid_group(evlist);
0253
0254     evlist__for_each_entry(evlist, evsel) {
0255         leader = evsel__leader(evsel);
0256
0257         /* Check that leader matches cpus with each member. */
0258         if (leader == evsel)
0259             continue;
0260         if (cpus_map_matched(leader, evsel))
0261             continue;
0262
0263         /* If there's mismatch disable the group and warn user. */
0264         WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
0265         evsel__group_desc(leader, buf, sizeof(buf));
0266         pr_warning("  %s\n", buf);
0267
0268         if (verbose) {
0269             cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
0270             pr_warning("     %s: %s\n", leader->name, buf);
0271             cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
0272             pr_warning("     %s: %s\n", evsel->name, buf);
0273         }
0274
0275         for_each_group_evsel(pos, leader)
0276             evsel__remove_from_group(pos, leader);
0277     }
0278 }
0279
0280 static inline void diff_timespec(struct timespec *r, struct timespec *a,
0281                  struct timespec *b)
0282 {
0283     r->tv_sec = a->tv_sec - b->tv_sec;
0284     if (a->tv_nsec < b->tv_nsec) {
0285         r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
0286         r->tv_sec--;
0287     } else {
0288         r->tv_nsec = a->tv_nsec - b->tv_nsec ;
0289     }
0290 }
0291
0292 static void perf_stat__reset_stats(void)
0293 {
0294     int i;
0295
0296     evlist__reset_stats(evsel_list);
0297     perf_stat__reset_shadow_stats();
0298
0299     for (i = 0; i < stat_config.stats_num; i++)
0300         perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
0301 }
0302
0303 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
0304                      union perf_event *event,
0305                      struct perf_sample *sample __maybe_unused,
0306                      struct machine *machine __maybe_unused)
0307 {
0308     if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
0309         pr_err("failed to write perf data, error: %m\n");
0310         return -1;
0311     }
0312
0313     perf_stat.bytes_written += event->header.size;
0314     return 0;
0315 }
0316
0317 static int write_stat_round_event(u64 tm, u64 type)
0318 {
0319     return perf_event__synthesize_stat_round(NULL, tm, type,
0320                          process_synthesized_event,
0321                          NULL);
0322 }
0323
0324 #define WRITE_STAT_ROUND_EVENT(time, interval) \
0325     write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
0326
0327 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
0328
0329 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
0330                    struct perf_counts_values *count)
0331 {
0332     struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
0333     struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
0334
0335     return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
0336                        process_synthesized_event, NULL);
0337 }
0338
0339 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
0340                    int thread, struct timespec *rs)
0341 {
0342     switch(counter->tool_event) {
0343         case PERF_TOOL_DURATION_TIME: {
0344             u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
0345             struct perf_counts_values *count =
0346                 perf_counts(counter->counts, cpu_map_idx, thread);
0347             count->ena = count->run = val;
0348             count->val = val;
0349             return 0;
0350         }
0351         case PERF_TOOL_USER_TIME:
0352         case PERF_TOOL_SYSTEM_TIME: {
0353             u64 val;
0354             struct perf_counts_values *count =
0355                 perf_counts(counter->counts, cpu_map_idx, thread);
0356             if (counter->tool_event == PERF_TOOL_USER_TIME)
0357                 val = ru_stats.ru_utime_usec_stat.mean;
0358             else
0359                 val = ru_stats.ru_stime_usec_stat.mean;
0360             count->ena = count->run = val;
0361             count->val = val;
0362             return 0;
0363         }
0364         default:
0365         case PERF_TOOL_NONE:
0366             return evsel__read_counter(counter, cpu_map_idx, thread);
0367         case PERF_TOOL_MAX:
0368             /* This should never be reached */
0369             return 0;
0370     }
0371 }
0372
0373 /*
0374  * Read out the results of a single counter:
0375  * do not aggregate counts across CPUs in system-wide mode
0376  */
0377 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
0378 {
0379     int nthreads = perf_thread_map__nr(evsel_list->core.threads);
0380     int thread;
0381
0382     if (!counter->supported)
0383         return -ENOENT;
0384
0385     for (thread = 0; thread < nthreads; thread++) {
0386         struct perf_counts_values *count;
0387
0388         count = perf_counts(counter->counts, cpu_map_idx, thread);
0389
0390         /*
0391          * The leader's group read loads data into its group members
0392          * (via evsel__read_counter()) and sets their count->loaded.
0393          */
0394         if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
0395             read_single_counter(counter, cpu_map_idx, thread, rs)) {
0396             counter->counts->scaled = -1;
0397             perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
0398             perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
0399             return -1;
0400         }
0401
0402         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
0403
0404         if (STAT_RECORD) {
0405             if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
0406                 pr_err("failed to write stat event\n");
0407                 return -1;
0408             }
0409         }
0410
0411         if (verbose > 1) {
0412             fprintf(stat_config.output,
0413                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
0414                     evsel__name(counter),
0415                     perf_cpu_map__cpu(evsel__cpus(counter),
0416                               cpu_map_idx).cpu,
0417                     count->val, count->ena, count->run);
0418         }
0419     }
0420
0421     return 0;
0422 }
0423
0424 static int read_affinity_counters(struct timespec *rs)
0425 {
0426     struct evlist_cpu_iterator evlist_cpu_itr;
0427     struct affinity saved_affinity, *affinity;
0428
0429     if (all_counters_use_bpf)
0430         return 0;
0431
0432     if (!target__has_cpu(&target) || target__has_per_thread(&target))
0433         affinity = NULL;
0434     else if (affinity__setup(&saved_affinity) < 0)
0435         return -1;
0436     else
0437         affinity = &saved_affinity;
0438
0439     evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
0440         struct evsel *counter = evlist_cpu_itr.evsel;
0441
0442         if (evsel__is_bpf(counter))
0443             continue;
0444
0445         if (!counter->err) {
0446             counter->err = read_counter_cpu(counter, rs,
0447                             evlist_cpu_itr.cpu_map_idx);
0448         }
0449     }
0450     if (affinity)
0451         affinity__cleanup(&saved_affinity);
0452
0453     return 0;
0454 }
0455
0456 static int read_bpf_map_counters(void)
0457 {
0458     struct evsel *counter;
0459     int err;
0460
0461     evlist__for_each_entry(evsel_list, counter) {
0462         if (!evsel__is_bpf(counter))
0463             continue;
0464
0465         err = bpf_counter__read(counter);
0466         if (err)
0467             return err;
0468     }
0469     return 0;
0470 }
0471
0472 static void read_counters(struct timespec *rs)
0473 {
0474     struct evsel *counter;
0475
0476     if (!stat_config.stop_read_counter) {
0477         if (read_bpf_map_counters() ||
0478             read_affinity_counters(rs))
0479             return;
0480     }
0481
0482     evlist__for_each_entry(evsel_list, counter) {
0483         if (counter->err)
0484             pr_debug("failed to read counter %s\n", counter->name);
0485         if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
0486             pr_warning("failed to process counter %s\n", counter->name);
0487         counter->err = 0;
0488     }
0489 }
0490
0491 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
0492 {
0493     int i;
0494
0495     config->stats = calloc(nthreads, sizeof(struct runtime_stat));
0496     if (!config->stats)
0497         return -1;
0498
0499     config->stats_num = nthreads;
0500
0501     for (i = 0; i < nthreads; i++)
0502         runtime_stat__init(&config->stats[i]);
0503
0504     return 0;
0505 }
0506
0507 static void runtime_stat_delete(struct perf_stat_config *config)
0508 {
0509     int i;
0510
0511     if (!config->stats)
0512         return;
0513
0514     for (i = 0; i < config->stats_num; i++)
0515         runtime_stat__exit(&config->stats[i]);
0516
0517     zfree(&config->stats);
0518 }
0519
0520 static void runtime_stat_reset(struct perf_stat_config *config)
0521 {
0522     int i;
0523
0524     if (!config->stats)
0525         return;
0526
0527     for (i = 0; i < config->stats_num; i++)
0528         perf_stat__reset_shadow_per_stat(&config->stats[i]);
0529 }
0530
0531 static void process_interval(void)
0532 {
0533     struct timespec ts, rs;
0534
0535     clock_gettime(CLOCK_MONOTONIC, &ts);
0536     diff_timespec(&rs, &ts, &ref_time);
0537
0538     perf_stat__reset_shadow_per_stat(&rt_stat);
0539     runtime_stat_reset(&stat_config);
0540     read_counters(&rs);
0541
0542     if (STAT_RECORD) {
0543         if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
0544             pr_err("failed to write stat round event\n");
0545     }
0546
0547     init_stats(&walltime_nsecs_stats);
0548     update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
0549     print_counters(&rs, 0, NULL);
0550 }
0551
0552 static bool handle_interval(unsigned int interval, int *times)
0553 {
0554     if (interval) {
0555         process_interval();
0556         if (interval_count && !(--(*times)))
0557             return true;
0558     }
0559     return false;
0560 }
0561
0562 static int enable_counters(void)
0563 {
0564     struct evsel *evsel;
0565     int err;
0566
0567     evlist__for_each_entry(evsel_list, evsel) {
0568         if (!evsel__is_bpf(evsel))
0569             continue;
0570
0571         err = bpf_counter__enable(evsel);
0572         if (err)
0573             return err;
0574     }
0575
0576     if (stat_config.initial_delay < 0) {
0577         pr_info(EVLIST_DISABLED_MSG);
0578         return 0;
0579     }
0580
0581     if (stat_config.initial_delay > 0) {
0582         pr_info(EVLIST_DISABLED_MSG);
0583         usleep(stat_config.initial_delay * USEC_PER_MSEC);
0584     }
0585
0586     /*
0587      * We need to enable counters only if:
0588      * - we don't have tracee (attaching to task or cpu)
0589      * - we have initial delay configured
0590      */
0591     if (!target__none(&target) || stat_config.initial_delay) {
0592         if (!all_counters_use_bpf)
0593             evlist__enable(evsel_list);
0594         if (stat_config.initial_delay > 0)
0595             pr_info(EVLIST_ENABLED_MSG);
0596     }
0597     return 0;
0598 }
0599
0600 static void disable_counters(void)
0601 {
0602     struct evsel *counter;
0603
0604     /*
0605      * If we don't have tracee (attaching to task or cpu), counters may
0606      * still be running. To get accurate group ratios, we must stop groups
0607      * from counting before reading their constituent counters.
0608      */
0609     if (!target__none(&target)) {
0610         evlist__for_each_entry(evsel_list, counter)
0611             bpf_counter__disable(counter);
0612         if (!all_counters_use_bpf)
0613             evlist__disable(evsel_list);
0614     }
0615 }
0616
0617 static volatile int workload_exec_errno;
0618
0619 /*
0620  * evlist__prepare_workload will send a SIGUSR1
0621  * if the fork fails, since we asked by setting its
0622  * want_signal to true.
0623  */
0624 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
0625                     void *ucontext __maybe_unused)
0626 {
0627     workload_exec_errno = info->si_value.sival_int;
0628 }
0629
0630 static bool evsel__should_store_id(struct evsel *counter)
0631 {
0632     return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
0633 }
0634
0635 static bool is_target_alive(struct target *_target,
0636                 struct perf_thread_map *threads)
0637 {
0638     struct stat st;
0639     int i;
0640
0641     if (!target__has_task(_target))
0642         return true;
0643
0644     for (i = 0; i < threads->nr; i++) {
0645         char path[PATH_MAX];
0646
0647         scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
0648               threads->map[i].pid);
0649
0650         if (!stat(path, &st))
0651             return true;
0652     }
0653
0654     return false;
0655 }
0656
0657 static void process_evlist(struct evlist *evlist, unsigned int interval)
0658 {
0659     enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
0660
0661     if (evlist__ctlfd_process(evlist, &cmd) > 0) {
0662         switch (cmd) {
0663         case EVLIST_CTL_CMD_ENABLE:
0664             if (interval)
0665                 process_interval();
0666             break;
0667         case EVLIST_CTL_CMD_DISABLE:
0668             if (interval)
0669                 process_interval();
0670             break;
0671         case EVLIST_CTL_CMD_SNAPSHOT:
0672         case EVLIST_CTL_CMD_ACK:
0673         case EVLIST_CTL_CMD_UNSUPPORTED:
0674         case EVLIST_CTL_CMD_EVLIST:
0675         case EVLIST_CTL_CMD_STOP:
0676         case EVLIST_CTL_CMD_PING:
0677         default:
0678             break;
0679         }
0680     }
0681 }
0682
0683 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
0684             int *time_to_sleep)
0685 {
0686     int tts = *time_to_sleep;
0687     struct timespec time_diff;
0688
0689     diff_timespec(&time_diff, time_stop, time_start);
0690
0691     tts -= time_diff.tv_sec * MSEC_PER_SEC +
0692            time_diff.tv_nsec / NSEC_PER_MSEC;
0693
0694     if (tts < 0)
0695         tts = 0;
0696
0697     *time_to_sleep = tts;
0698 }
0699
0700 static int dispatch_events(bool forks, int timeout, int interval, int *times)
0701 {
0702     int child_exited = 0, status = 0;
0703     int time_to_sleep, sleep_time;
0704     struct timespec time_start, time_stop;
0705
0706     if (interval)
0707         sleep_time = interval;
0708     else if (timeout)
0709         sleep_time = timeout;
0710     else
0711         sleep_time = 1000;
0712
0713     time_to_sleep = sleep_time;
0714
0715     while (!done) {
0716         if (forks)
0717             child_exited = waitpid(child_pid, &status, WNOHANG);
0718         else
0719             child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
0720
0721         if (child_exited)
0722             break;
0723
0724         clock_gettime(CLOCK_MONOTONIC, &time_start);
0725         if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
0726             if (timeout || handle_interval(interval, times))
0727                 break;
0728             time_to_sleep = sleep_time;
0729         } else { /* fd revent */
0730             process_evlist(evsel_list, interval);
0731             clock_gettime(CLOCK_MONOTONIC, &time_stop);
0732             compute_tts(&time_start, &time_stop, &time_to_sleep);
0733         }
0734     }
0735
0736     return status;
0737 }
0738
0739 enum counter_recovery {
0740     COUNTER_SKIP,
0741     COUNTER_RETRY,
0742     COUNTER_FATAL,
0743 };
0744
0745 static enum counter_recovery stat_handle_error(struct evsel *counter)
0746 {
0747     char msg[BUFSIZ];
0748     /*
0749      * PPC returns ENXIO for HW counters until 2.6.37
0750      * (behavior changed with commit b0a873e).
0751      */
0752     if (errno == EINVAL || errno == ENOSYS ||
0753         errno == ENOENT || errno == EOPNOTSUPP ||
0754         errno == ENXIO) {
0755         if (verbose > 0)
0756             ui__warning("%s event is not supported by the kernel.\n",
0757                     evsel__name(counter));
0758         counter->supported = false;
0759         /*
0760          * errored is a sticky flag that means one of the counter's
0761          * cpu event had a problem and needs to be reexamined.
0762          */
0763         counter->errored = true;
0764
0765         if ((evsel__leader(counter) != counter) ||
0766             !(counter->core.leader->nr_members > 1))
0767             return COUNTER_SKIP;
0768     } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
0769         if (verbose > 0)
0770             ui__warning("%s\n", msg);
0771         return COUNTER_RETRY;
0772     } else if (target__has_per_thread(&target) &&
0773            evsel_list->core.threads &&
0774            evsel_list->core.threads->err_thread != -1) {
0775         /*
0776          * For global --per-thread case, skip current
0777          * error thread.
0778          */
0779         if (!thread_map__remove(evsel_list->core.threads,
0780                     evsel_list->core.threads->err_thread)) {
0781             evsel_list->core.threads->err_thread = -1;
0782             return COUNTER_RETRY;
0783         }
0784     }
0785
0786     evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
0787     ui__error("%s\n", msg);
0788
0789     if (child_pid != -1)
0790         kill(child_pid, SIGTERM);
0791     return COUNTER_FATAL;
0792 }
0793
0794 static int __run_perf_stat(int argc, const char **argv, int run_idx)
0795 {
0796     int interval = stat_config.interval;
0797     int times = stat_config.times;
0798     int timeout = stat_config.timeout;
0799     char msg[BUFSIZ];
0800     unsigned long long t0, t1;
0801     struct evsel *counter;
0802     size_t l;
0803     int status = 0;
0804     const bool forks = (argc > 0);
0805     bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
0806     struct evlist_cpu_iterator evlist_cpu_itr;
0807     struct affinity saved_affinity, *affinity = NULL;
0808     int err;
0809     bool second_pass = false;
0810
0811     if (forks) {
0812         if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
0813             perror("failed to prepare workload");
0814             return -1;
0815         }
0816         child_pid = evsel_list->workload.pid;
0817     }
0818
0819     if (group)
0820         evlist__set_leader(evsel_list);
0821
0822     if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
0823         if (affinity__setup(&saved_affinity) < 0)
0824             return -1;
0825         affinity = &saved_affinity;
0826     }
0827
0828     evlist__for_each_entry(evsel_list, counter) {
0829         counter->reset_group = false;
0830         if (bpf_counter__load(counter, &target))
0831             return -1;
0832         if (!evsel__is_bpf(counter))
0833             all_counters_use_bpf = false;
0834     }
0835
0836     evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
0837         counter = evlist_cpu_itr.evsel;
0838
0839         /*
0840          * bperf calls evsel__open_per_cpu() in bperf__load(), so
0841          * no need to call it again here.
0842          */
0843         if (target.use_bpf)
0844             break;
0845
0846         if (counter->reset_group || counter->errored)
0847             continue;
0848         if (evsel__is_bpf(counter))
0849             continue;
0850 try_again:
0851         if (create_perf_stat_counter(counter, &stat_config, &target,
0852                          evlist_cpu_itr.cpu_map_idx) < 0) {
0853
0854             /*
0855              * Weak group failed. We cannot just undo this here
0856              * because earlier CPUs might be in group mode, and the kernel
0857              * doesn't support mixing group and non group reads. Defer
0858              * it to later.
0859              * Don't close here because we're in the wrong affinity.
0860              */
0861             if ((errno == EINVAL || errno == EBADF) &&
0862                 evsel__leader(counter) != counter &&
0863                 counter->weak_group) {
0864                 evlist__reset_weak_group(evsel_list, counter, false);
0865                 assert(counter->reset_group);
0866                 second_pass = true;
0867                 continue;
0868             }
0869
0870             switch (stat_handle_error(counter)) {
0871             case COUNTER_FATAL:
0872                 return -1;
0873             case COUNTER_RETRY:
0874                 goto try_again;
0875             case COUNTER_SKIP:
0876                 continue;
0877             default:
0878                 break;
0879             }
0880
0881         }
0882         counter->supported = true;
0883     }
0884
0885     if (second_pass) {
0886         /*
0887          * Now redo all the weak group after closing them,
0888          * and also close errored counters.
0889          */
0890
0891         /* First close errored or weak retry */
0892         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
0893             counter = evlist_cpu_itr.evsel;
0894
0895             if (!counter->reset_group && !counter->errored)
0896                 continue;
0897
0898             perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
0899         }
0900         /* Now reopen weak */
0901         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
0902             counter = evlist_cpu_itr.evsel;
0903
0904             if (!counter->reset_group && !counter->errored)
0905                 continue;
0906             if (!counter->reset_group)
0907                 continue;
0908 try_again_reset:
0909             pr_debug2("reopening weak %s\n", evsel__name(counter));
0910             if (create_perf_stat_counter(counter, &stat_config, &target,
0911                              evlist_cpu_itr.cpu_map_idx) < 0) {
0912
0913                 switch (stat_handle_error(counter)) {
0914                 case COUNTER_FATAL:
0915                     return -1;
0916                 case COUNTER_RETRY:
0917                     goto try_again_reset;
0918                 case COUNTER_SKIP:
0919                     continue;
0920                 default:
0921                     break;
0922                 }
0923             }
0924             counter->supported = true;
0925         }
0926     }
0927     affinity__cleanup(affinity);
0928
0929     evlist__for_each_entry(evsel_list, counter) {
0930         if (!counter->supported) {
0931             perf_evsel__free_fd(&counter->core);
0932             continue;
0933         }
0934
0935         l = strlen(counter->unit);
0936         if (l > stat_config.unit_width)
0937             stat_config.unit_width = l;
0938
0939         if (evsel__should_store_id(counter) &&
0940             evsel__store_ids(counter, evsel_list))
0941             return -1;
0942     }
0943
0944     if (evlist__apply_filters(evsel_list, &counter)) {
0945         pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
0946             counter->filter, evsel__name(counter), errno,
0947             str_error_r(errno, msg, sizeof(msg)));
0948         return -1;
0949     }
0950
0951     if (STAT_RECORD) {
0952         int fd = perf_data__fd(&perf_stat.data);
0953
0954         if (is_pipe) {
0955             err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
0956         } else {
0957             err = perf_session__write_header(perf_stat.session, evsel_list,
0958                              fd, false);
0959         }
0960
0961         if (err < 0)
0962             return err;
0963
0964         err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
0965                              process_synthesized_event, is_pipe);
0966         if (err < 0)
0967             return err;
0968     }
0969
0970     err = enable_counters();
0971     if (err)
0972         return -1;
0973
0974     /* Exec the command, if any */
0975     if (forks)
0976         evlist__start_workload(evsel_list);
0977
0978     t0 = rdclock();
0979     clock_gettime(CLOCK_MONOTONIC, &ref_time);
0980
0981     if (forks) {
0982         if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
0983             status = dispatch_events(forks, timeout, interval, &times);
0984         if (child_pid != -1) {
0985             if (timeout)
0986                 kill(child_pid, SIGTERM);
0987             wait4(child_pid, &status, 0, &stat_config.ru_data);
0988         }
0989
0990         if (workload_exec_errno) {
0991             const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
0992             pr_err("Workload failed: %s\n", emsg);
0993             return -1;
0994         }
0995
0996         if (WIFSIGNALED(status))
0997             psignal(WTERMSIG(status), argv[0]);
0998     } else {
0999         status = dispatch_events(forks, timeout, interval, &times);
1000     }
1001
1002     disable_counters();
1003
1004     t1 = rdclock();
1005
1006     if (stat_config.walltime_run_table)
1007         stat_config.walltime_run[run_idx] = t1 - t0;
1008
1009     if (interval && stat_config.summary) {
1010         stat_config.interval = 0;
1011         stat_config.stop_read_counter = true;
1012         init_stats(&walltime_nsecs_stats);
1013         update_stats(&walltime_nsecs_stats, t1 - t0);
1014
1015         if (stat_config.aggr_mode == AGGR_GLOBAL)
1016             evlist__save_aggr_prev_raw_counts(evsel_list);
1017
1018         evlist__copy_prev_raw_counts(evsel_list);
1019         evlist__reset_prev_raw_counts(evsel_list);
1020         runtime_stat_reset(&stat_config);
1021         perf_stat__reset_shadow_per_stat(&rt_stat);
1022     } else {
1023         update_stats(&walltime_nsecs_stats, t1 - t0);
1024         update_rusage_stats(&ru_stats, &stat_config.ru_data);
1025     }
1026
1027     /*
1028      * Closing a group leader splits the group, and as we only disable
1029      * group leaders, results in remaining events becoming enabled. To
1030      * avoid arbitrary skew, we must read all counters before closing any
1031      * group leaders.
1032      */
1033     read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1034
1035     /*
1036      * We need to keep evsel_list alive, because it's processed
1037      * later the evsel_list will be closed after.
1038      */
1039     if (!STAT_RECORD)
1040         evlist__close(evsel_list);
1041
1042     return WEXITSTATUS(status);
1043 }
1044
1045 static int run_perf_stat(int argc, const char **argv, int run_idx)
1046 {
1047     int ret;
1048
1049     if (pre_cmd) {
1050         ret = system(pre_cmd);
1051         if (ret)
1052             return ret;
1053     }
1054
1055     if (sync_run)
1056         sync();
1057
1058     ret = __run_perf_stat(argc, argv, run_idx);
1059     if (ret)
1060         return ret;
1061
1062     if (post_cmd) {
1063         ret = system(post_cmd);
1064         if (ret)
1065             return ret;
1066     }
1067
1068     return ret;
1069 }
1070
1071 static void print_counters(struct timespec *ts, int argc, const char **argv)
1072 {
1073     /* Do not print anything if we record to the pipe. */
1074     if (STAT_RECORD && perf_stat.data.is_pipe)
1075         return;
1076     if (stat_config.quiet)
1077         return;
1078
1079     evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1080 }
1081
1082 static volatile int signr = -1;
1083
1084 static void skip_signal(int signo)
1085 {
1086     if ((child_pid == -1) || stat_config.interval)
1087         done = 1;
1088
1089     signr = signo;
1090     /*
1091      * render child_pid harmless
1092      * won't send SIGTERM to a random
1093      * process in case of race condition
1094      * and fast PID recycling
1095      */
1096     child_pid = -1;
1097 }
1098
1099 static void sig_atexit(void)
1100 {
1101     sigset_t set, oset;
1102
1103     /*
1104      * avoid race condition with SIGCHLD handler
1105      * in skip_signal() which is modifying child_pid
1106      * goal is to avoid send SIGTERM to a random
1107      * process
1108      */
1109     sigemptyset(&set);
1110     sigaddset(&set, SIGCHLD);
1111     sigprocmask(SIG_BLOCK, &set, &oset);
1112
1113     if (child_pid != -1)
1114         kill(child_pid, SIGTERM);
1115
1116     sigprocmask(SIG_SETMASK, &oset, NULL);
1117
1118     if (signr == -1)
1119         return;
1120
1121     signal(signr, SIG_DFL);
1122     kill(getpid(), signr);
1123 }
1124
1125 void perf_stat__set_big_num(int set)
1126 {
1127     stat_config.big_num = (set != 0);
1128 }
1129
1130 void perf_stat__set_no_csv_summary(int set)
1131 {
1132     stat_config.no_csv_summary = (set != 0);
1133 }
1134
1135 static int stat__set_big_num(const struct option *opt __maybe_unused,
1136                  const char *s __maybe_unused, int unset)
1137 {
1138     big_num_opt = unset ? 0 : 1;
1139     perf_stat__set_big_num(!unset);
1140     return 0;
1141 }
1142
1143 static int enable_metric_only(const struct option *opt __maybe_unused,
1144                   const char *s __maybe_unused, int unset)
1145 {
1146     force_metric_only = true;
1147     stat_config.metric_only = !unset;
1148     return 0;
1149 }
1150
1151 static int parse_metric_groups(const struct option *opt,
1152                    const char *str,
1153                    int unset __maybe_unused)
1154 {
1155     return metricgroup__parse_groups(opt, str,
1156                      stat_config.metric_no_group,
1157                      stat_config.metric_no_merge,
1158                      &stat_config.metric_events);
1159 }
1160
1161 static int parse_control_option(const struct option *opt,
1162                 const char *str,
1163                 int unset __maybe_unused)
1164 {
1165     struct perf_stat_config *config = opt->value;
1166
1167     return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1168 }
1169
1170 static int parse_stat_cgroups(const struct option *opt,
1171                   const char *str, int unset)
1172 {
1173     if (stat_config.cgroup_list) {
1174         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1175         return -1;
1176     }
1177
1178     return parse_cgroups(opt, str, unset);
1179 }
1180
1181 static int parse_hybrid_type(const struct option *opt,
1182                  const char *str,
1183                  int unset __maybe_unused)
1184 {
1185     struct evlist *evlist = *(struct evlist **)opt->value;
1186
1187     if (!list_empty(&evlist->core.entries)) {
1188         fprintf(stderr, "Must define cputype before events/metrics\n");
1189         return -1;
1190     }
1191
1192     evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1193     if (!evlist->hybrid_pmu_name) {
1194         fprintf(stderr, "--cputype %s is not supported!\n", str);
1195         return -1;
1196     }
1197
1198     return 0;
1199 }
1200
1201 static struct option stat_options[] = {
1202     OPT_BOOLEAN('T', "transaction", &transaction_run,
1203             "hardware transaction statistics"),
1204     OPT_CALLBACK('e', "event", &evsel_list, "event",
1205              "event selector. use 'perf list' to list available events",
1206              parse_events_option),
1207     OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1208              "event filter", parse_filter),
1209     OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1210             "child tasks do not inherit counters"),
1211     OPT_STRING('p', "pid", &target.pid, "pid",
1212            "stat events on existing process id"),
1213     OPT_STRING('t', "tid", &target.tid, "tid",
1214            "stat events on existing thread id"),
1215 #ifdef HAVE_BPF_SKEL
1216     OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1217            "stat events on existing bpf program id"),
1218     OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1219             "use bpf program to count events"),
1220     OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1221            "path to perf_event_attr map"),
1222 #endif
1223     OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1224             "system-wide collection from all CPUs"),
1225     OPT_BOOLEAN('g', "group", &group,
1226             "put the counters into a counter group"),
1227     OPT_BOOLEAN(0, "scale", &stat_config.scale,
1228             "Use --no-scale to disable counter scaling for multiplexing"),
1229     OPT_INCR('v', "verbose", &verbose,
1230             "be more verbose (show counter open errors, etc)"),
1231     OPT_INTEGER('r', "repeat", &stat_config.run_count,
1232             "repeat command and print average + stddev (max: 100, forever: 0)"),
1233     OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1234             "display details about each run (only with -r option)"),
1235     OPT_BOOLEAN('n', "null", &stat_config.null_run,
1236             "null run - dont start any counters"),
1237     OPT_INCR('d', "detailed", &detailed_run,
1238             "detailed run - start a lot of events"),
1239     OPT_BOOLEAN('S', "sync", &sync_run,
1240             "call sync() before starting a run"),
1241     OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1242                "print large numbers with thousands\' separators",
1243                stat__set_big_num),
1244     OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1245             "list of cpus to monitor in system-wide"),
1246     OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1247             "disable CPU count aggregation", AGGR_NONE),
1248     OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1249     OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1250             "Merge identical named hybrid events"),
1251     OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1252            "print counts with custom separator"),
1253     OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1254            "print counts in JSON format"),
1255     OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1256              "monitor event in cgroup name only", parse_stat_cgroups),
1257     OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1258             "expand events for each cgroup"),
1259     OPT_STRING('o', "output", &output_name, "file", "output file name"),
1260     OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1261     OPT_INTEGER(0, "log-fd", &output_fd,
1262             "log output to fd, instead of stderr"),
1263     OPT_STRING(0, "pre", &pre_cmd, "command",
1264             "command to run prior to the measured command"),
1265     OPT_STRING(0, "post", &post_cmd, "command",
1266             "command to run after to the measured command"),
1267     OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1268             "print counts at regular interval in ms "
1269             "(overhead is possible for values <= 100ms)"),
1270     OPT_INTEGER(0, "interval-count", &stat_config.times,
1271             "print counts for fixed number of times"),
1272     OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1273             "clear screen in between new interval"),
1274     OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1275             "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1276     OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1277              "aggregate counts per processor socket", AGGR_SOCKET),
1278     OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1279              "aggregate counts per processor die", AGGR_DIE),
1280     OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1281              "aggregate counts per physical processor core", AGGR_CORE),
1282     OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1283              "aggregate counts per thread", AGGR_THREAD),
1284     OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1285              "aggregate counts per numa node", AGGR_NODE),
1286     OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1287             "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1288     OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1289             "Only print computed metrics. No raw values", enable_metric_only),
1290     OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1291                "don't group metric events, impacts multiplexing"),
1292     OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1293                "don't try to share events between metrics in a group"),
1294     OPT_BOOLEAN(0, "topdown", &topdown_run,
1295             "measure top-down statistics"),
1296     OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1297             "Set the metrics level for the top-down statistics (0: max level)"),
1298     OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1299             "measure SMI cost"),
1300     OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1301              "monitor specified metrics or metric groups (separated by ,)",
1302              parse_metric_groups),
1303     OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1304              "Configure all used events to run in kernel space.",
1305              PARSE_OPT_EXCLUSIVE),
1306     OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1307              "Configure all used events to run in user space.",
1308              PARSE_OPT_EXCLUSIVE),
1309     OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1310             "Use with 'percore' event qualifier to show the event "
1311             "counts of one hardware thread by sum up total hardware "
1312             "threads of same physical core"),
1313     OPT_BOOLEAN(0, "summary", &stat_config.summary,
1314                "print summary for interval mode"),
1315     OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1316                "don't print 'summary' for CSV summary output"),
1317     OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1318             "don't print output (useful with record)"),
1319     OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1320              "Only enable events on applying cpu with this type "
1321              "for hybrid platform (e.g. core or atom)",
1322              parse_hybrid_type),
1323 #ifdef HAVE_LIBPFM
1324     OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1325         "libpfm4 event selector. use 'perf list' to list available events",
1326         parse_libpfm_events_option),
1327 #endif
1328     OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1329              "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1330              "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1331              "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1332               parse_control_option),
1333     OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1334                 "measure I/O performance metrics provided by arch/platform",
1335                 iostat_parse),
1336     OPT_END()
1337 };
1338
1339 static const char *const aggr_mode__string[] = {
1340     [AGGR_CORE] = "core",
1341     [AGGR_DIE] = "die",
1342     [AGGR_GLOBAL] = "global",
1343     [AGGR_NODE] = "node",
1344     [AGGR_NONE] = "none",
1345     [AGGR_SOCKET] = "socket",
1346     [AGGR_THREAD] = "thread",
1347     [AGGR_UNSET] = "unset",
1348 };
1349
1350 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1351                         struct perf_cpu cpu)
1352 {
1353     return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1354 }
1355
1356 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1357                          struct perf_cpu cpu)
1358 {
1359     return aggr_cpu_id__die(cpu, /*data=*/NULL);
1360 }
1361
1362 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1363                           struct perf_cpu cpu)
1364 {
1365     return aggr_cpu_id__core(cpu, /*data=*/NULL);
1366 }
1367
1368 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1369                           struct perf_cpu cpu)
1370 {
1371     return aggr_cpu_id__node(cpu, /*data=*/NULL);
1372 }
1373
1374 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1375                           aggr_get_id_t get_id, struct perf_cpu cpu)
1376 {
1377     struct aggr_cpu_id id = aggr_cpu_id__empty();
1378
1379     if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1380         config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1381
1382     id = config->cpus_aggr_map->map[cpu.cpu];
1383     return id;
1384 }
1385
1386 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1387                                struct perf_cpu cpu)
1388 {
1389     return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1390 }
1391
1392 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1393                             struct perf_cpu cpu)
1394 {
1395     return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1396 }
1397
1398 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1399                              struct perf_cpu cpu)
1400 {
1401     return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1402 }
1403
1404 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1405                              struct perf_cpu cpu)
1406 {
1407     return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1408 }
1409
1410 static bool term_percore_set(void)
1411 {
1412     struct evsel *counter;
1413
1414     evlist__for_each_entry(evsel_list, counter) {
1415         if (counter->percore)
1416             return true;
1417     }
1418
1419     return false;
1420 }
1421
1422 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1423 {
1424     switch (aggr_mode) {
1425     case AGGR_SOCKET:
1426         return aggr_cpu_id__socket;
1427     case AGGR_DIE:
1428         return aggr_cpu_id__die;
1429     case AGGR_CORE:
1430         return aggr_cpu_id__core;
1431     case AGGR_NODE:
1432         return aggr_cpu_id__node;
1433     case AGGR_NONE:
1434         if (term_percore_set())
1435             return aggr_cpu_id__core;
1436
1437         return NULL;
1438     case AGGR_GLOBAL:
1439     case AGGR_THREAD:
1440     case AGGR_UNSET:
1441     case AGGR_MAX:
1442     default:
1443         return NULL;
1444     }
1445 }
1446
1447 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1448 {
1449     switch (aggr_mode) {
1450     case AGGR_SOCKET:
1451         return perf_stat__get_socket_cached;
1452     case AGGR_DIE:
1453         return perf_stat__get_die_cached;
1454     case AGGR_CORE:
1455         return perf_stat__get_core_cached;
1456     case AGGR_NODE:
1457         return perf_stat__get_node_cached;
1458     case AGGR_NONE:
1459         if (term_percore_set()) {
1460             return perf_stat__get_core_cached;
1461         }
1462         return NULL;
1463     case AGGR_GLOBAL:
1464     case AGGR_THREAD:
1465     case AGGR_UNSET:
1466     case AGGR_MAX:
1467     default:
1468         return NULL;
1469     }
1470 }
1471
1472 static int perf_stat_init_aggr_mode(void)
1473 {
1474     int nr;
1475     aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1476
1477     if (get_id) {
1478         stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1479                              get_id, /*data=*/NULL);
1480         if (!stat_config.aggr_map) {
1481             pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1482             return -1;
1483         }
1484         stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1485     }
1486
1487     /*
1488      * The evsel_list->cpus is the base we operate on,
1489      * taking the highest cpu number to be the size of
1490      * the aggregation translate cpumap.
1491      */
1492     if (evsel_list->core.user_requested_cpus)
1493         nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1494     else
1495         nr = 0;
1496     stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1497     return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1498 }
1499
1500 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1501 {
1502     if (map) {
1503         WARN_ONCE(refcount_read(&map->refcnt) != 0,
1504               "cpu_aggr_map refcnt unbalanced\n");
1505         free(map);
1506     }
1507 }
1508
1509 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1510 {
1511     if (map && refcount_dec_and_test(&map->refcnt))
1512         cpu_aggr_map__delete(map);
1513 }
1514
1515 static void perf_stat__exit_aggr_mode(void)
1516 {
1517     cpu_aggr_map__put(stat_config.aggr_map);
1518     cpu_aggr_map__put(stat_config.cpus_aggr_map);
1519     stat_config.aggr_map = NULL;
1520     stat_config.cpus_aggr_map = NULL;
1521 }
1522
1523 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1524 {
1525     struct perf_env *env = data;
1526     struct aggr_cpu_id id = aggr_cpu_id__empty();
1527
1528     if (cpu.cpu != -1)
1529         id.socket = env->cpu[cpu.cpu].socket_id;
1530
1531     return id;
1532 }
1533
1534 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1535 {
1536     struct perf_env *env = data;
1537     struct aggr_cpu_id id = aggr_cpu_id__empty();
1538
1539     if (cpu.cpu != -1) {
1540         /*
1541          * die_id is relative to socket, so start
1542          * with the socket ID and then add die to
1543          * make a unique ID.
1544          */
1545         id.socket = env->cpu[cpu.cpu].socket_id;
1546         id.die = env->cpu[cpu.cpu].die_id;
1547     }
1548
1549     return id;
1550 }
1551
1552 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1553 {
1554     struct perf_env *env = data;
1555     struct aggr_cpu_id id = aggr_cpu_id__empty();
1556
1557     if (cpu.cpu != -1) {
1558         /*
1559          * core_id is relative to socket and die,
1560          * we need a global id. So we set
1561          * socket, die id and core id
1562          */
1563         id.socket = env->cpu[cpu.cpu].socket_id;
1564         id.die = env->cpu[cpu.cpu].die_id;
1565         id.core = env->cpu[cpu.cpu].core_id;
1566     }
1567
1568     return id;
1569 }
1570
1571 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1572 {
1573     struct aggr_cpu_id id = aggr_cpu_id__empty();
1574
1575     id.node = perf_env__numa_node(data, cpu);
1576     return id;
1577 }
1578
1579 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1580                              struct perf_cpu cpu)
1581 {
1582     return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1583 }
1584 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1585                           struct perf_cpu cpu)
1586 {
1587     return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1588 }
1589
1590 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1591                            struct perf_cpu cpu)
1592 {
1593     return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1594 }
1595
1596 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1597                            struct perf_cpu cpu)
1598 {
1599     return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1600 }
1601
1602 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1603 {
1604     switch (aggr_mode) {
1605     case AGGR_SOCKET:
1606         return perf_env__get_socket_aggr_by_cpu;
1607     case AGGR_DIE:
1608         return perf_env__get_die_aggr_by_cpu;
1609     case AGGR_CORE:
1610         return perf_env__get_core_aggr_by_cpu;
1611     case AGGR_NODE:
1612         return perf_env__get_node_aggr_by_cpu;
1613     case AGGR_NONE:
1614     case AGGR_GLOBAL:
1615     case AGGR_THREAD:
1616     case AGGR_UNSET:
1617     case AGGR_MAX:
1618     default:
1619         return NULL;
1620     }
1621 }
1622
1623 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1624 {
1625     switch (aggr_mode) {
1626     case AGGR_SOCKET:
1627         return perf_stat__get_socket_file;
1628     case AGGR_DIE:
1629         return perf_stat__get_die_file;
1630     case AGGR_CORE:
1631         return perf_stat__get_core_file;
1632     case AGGR_NODE:
1633         return perf_stat__get_node_file;
1634     case AGGR_NONE:
1635     case AGGR_GLOBAL:
1636     case AGGR_THREAD:
1637     case AGGR_UNSET:
1638     case AGGR_MAX:
1639     default:
1640         return NULL;
1641     }
1642 }
1643
1644 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1645 {
1646     struct perf_env *env = &st->session->header.env;
1647     aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1648
1649     if (!get_id)
1650         return 0;
1651
1652     stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1653     if (!stat_config.aggr_map) {
1654         pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1655         return -1;
1656     }
1657     stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1658     return 0;
1659 }
1660
1661 /*
1662  * Add default attributes, if there were no attributes specified or
1663  * if -d/--detailed, -d -d or -d -d -d is used:
1664  */
1665 static int add_default_attributes(void)
1666 {
1667     int err;
1668     struct perf_event_attr default_attrs0[] = {
1669
1670   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
1671   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES    },
1672   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS      },
1673   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
1674
1675   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
1676 };
1677     struct perf_event_attr frontend_attrs[] = {
1678   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1679 };
1680     struct perf_event_attr backend_attrs[] = {
1681   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1682 };
1683     struct perf_event_attr default_attrs1[] = {
1684   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS        },
1685   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1686   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES       },
1687
1688 };
1689
1690 /*
1691  * Detailed stats (-d), covering the L1 and last level data caches:
1692  */
1693     struct perf_event_attr detailed_attrs[] = {
1694
1695   { .type = PERF_TYPE_HW_CACHE,
1696     .config =
1697      PERF_COUNT_HW_CACHE_L1D        <<  0  |
1698     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1699     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1700
1701   { .type = PERF_TYPE_HW_CACHE,
1702     .config =
1703      PERF_COUNT_HW_CACHE_L1D        <<  0  |
1704     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1705     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1706
1707   { .type = PERF_TYPE_HW_CACHE,
1708     .config =
1709      PERF_COUNT_HW_CACHE_LL         <<  0  |
1710     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1711     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1712
1713   { .type = PERF_TYPE_HW_CACHE,
1714     .config =
1715      PERF_COUNT_HW_CACHE_LL         <<  0  |
1716     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1717     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1718 };
1719
1720 /*
1721  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1722  */
1723     struct perf_event_attr very_detailed_attrs[] = {
1724
1725   { .type = PERF_TYPE_HW_CACHE,
1726     .config =
1727      PERF_COUNT_HW_CACHE_L1I        <<  0  |
1728     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1729     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1730
1731   { .type = PERF_TYPE_HW_CACHE,
1732     .config =
1733      PERF_COUNT_HW_CACHE_L1I        <<  0  |
1734     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1735     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1736
1737   { .type = PERF_TYPE_HW_CACHE,
1738     .config =
1739      PERF_COUNT_HW_CACHE_DTLB       <<  0  |
1740     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1741     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1742
1743   { .type = PERF_TYPE_HW_CACHE,
1744     .config =
1745      PERF_COUNT_HW_CACHE_DTLB       <<  0  |
1746     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1747     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1748
1749   { .type = PERF_TYPE_HW_CACHE,
1750     .config =
1751      PERF_COUNT_HW_CACHE_ITLB       <<  0  |
1752     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1753     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1754
1755   { .type = PERF_TYPE_HW_CACHE,
1756     .config =
1757      PERF_COUNT_HW_CACHE_ITLB       <<  0  |
1758     (PERF_COUNT_HW_CACHE_OP_READ        <<  8) |
1759     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1760
1761 };
1762
1763 /*
1764  * Very, very detailed stats (-d -d -d), adding prefetch events:
1765  */
1766     struct perf_event_attr very_very_detailed_attrs[] = {
1767
1768   { .type = PERF_TYPE_HW_CACHE,
1769     .config =
1770      PERF_COUNT_HW_CACHE_L1D        <<  0  |
1771     (PERF_COUNT_HW_CACHE_OP_PREFETCH    <<  8) |
1772     (PERF_COUNT_HW_CACHE_RESULT_ACCESS  << 16)              },
1773
1774   { .type = PERF_TYPE_HW_CACHE,
1775     .config =
1776      PERF_COUNT_HW_CACHE_L1D        <<  0  |
1777     (PERF_COUNT_HW_CACHE_OP_PREFETCH    <<  8) |
1778     (PERF_COUNT_HW_CACHE_RESULT_MISS    << 16)              },
1779 };
1780
1781     struct perf_event_attr default_null_attrs[] = {};
1782
1783     /* Set attrs if no event is selected and !null_run: */
1784     if (stat_config.null_run)
1785         return 0;
1786
1787     if (transaction_run) {
1788         struct parse_events_error errinfo;
1789         /* Handle -T as -M transaction. Once platform specific metrics
1790          * support has been added to the json files, all architectures
1791          * will use this approach. To determine transaction support
1792          * on an architecture test for such a metric name.
1793          */
1794         if (metricgroup__has_metric("transaction")) {
1795             struct option opt = { .value = &evsel_list };
1796
1797             return metricgroup__parse_groups(&opt, "transaction",
1798                              stat_config.metric_no_group,
1799                             stat_config.metric_no_merge,
1800                              &stat_config.metric_events);
1801         }
1802
1803         parse_events_error__init(&errinfo);
1804         if (pmu_have_event("cpu", "cycles-ct") &&
1805             pmu_have_event("cpu", "el-start"))
1806             err = parse_events(evsel_list, transaction_attrs,
1807                        &errinfo);
1808         else
1809             err = parse_events(evsel_list,
1810                        transaction_limited_attrs,
1811                        &errinfo);
1812         if (err) {
1813             fprintf(stderr, "Cannot set up transaction events\n");
1814             parse_events_error__print(&errinfo, transaction_attrs);
1815         }
1816         parse_events_error__exit(&errinfo);
1817         return err ? -1 : 0;
1818     }
1819
1820     if (smi_cost) {
1821         struct parse_events_error errinfo;
1822         int smi;
1823
1824         if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1825             fprintf(stderr, "freeze_on_smi is not supported.\n");
1826             return -1;
1827         }
1828
1829         if (!smi) {
1830             if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1831                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1832                 return -1;
1833             }
1834             smi_reset = true;
1835         }
1836
1837         if (!pmu_have_event("msr", "aperf") ||
1838             !pmu_have_event("msr", "smi")) {
1839             fprintf(stderr, "To measure SMI cost, it needs "
1840                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1841             return -1;
1842         }
1843         if (!force_metric_only)
1844             stat_config.metric_only = true;
1845
1846         parse_events_error__init(&errinfo);
1847         err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1848         if (err) {
1849             parse_events_error__print(&errinfo, smi_cost_attrs);
1850             fprintf(stderr, "Cannot set up SMI cost events\n");
1851         }
1852         parse_events_error__exit(&errinfo);
1853         return err ? -1 : 0;
1854     }
1855
1856     if (topdown_run) {
1857         const char **metric_attrs = topdown_metric_attrs;
1858         unsigned int max_level = 1;
1859         char *str = NULL;
1860         bool warn = false;
1861         const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1862
1863         if (!force_metric_only)
1864             stat_config.metric_only = true;
1865
1866         if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1867             metric_attrs = topdown_metric_L2_attrs;
1868             max_level = 2;
1869         }
1870
1871         if (stat_config.topdown_level > max_level) {
1872             pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1873             return -1;
1874         } else if (!stat_config.topdown_level)
1875             stat_config.topdown_level = max_level;
1876
1877         if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1878             pr_err("Out of memory\n");
1879             return -1;
1880         }
1881
1882         if (metric_attrs[0] && str) {
1883             if (!stat_config.interval && !stat_config.metric_only) {
1884                 fprintf(stat_config.output,
1885                     "Topdown accuracy may decrease when measuring long periods.\n"
1886                     "Please print the result regularly, e.g. -I1000\n");
1887             }
1888             goto setup_metrics;
1889         }
1890
1891         zfree(&str);
1892
1893         if (stat_config.aggr_mode != AGGR_GLOBAL &&
1894             stat_config.aggr_mode != AGGR_CORE) {
1895             pr_err("top down event configuration requires --per-core mode\n");
1896             return -1;
1897         }
1898         stat_config.aggr_mode = AGGR_CORE;
1899         if (nr_cgroups || !target__has_cpu(&target)) {
1900             pr_err("top down event configuration requires system-wide mode (-a)\n");
1901             return -1;
1902         }
1903
1904         if (topdown_filter_events(topdown_attrs, &str,
1905                 arch_topdown_check_group(&warn),
1906                 pmu_name) < 0) {
1907             pr_err("Out of memory\n");
1908             return -1;
1909         }
1910
1911         if (topdown_attrs[0] && str) {
1912             struct parse_events_error errinfo;
1913             if (warn)
1914                 arch_topdown_group_warn();
1915 setup_metrics:
1916             parse_events_error__init(&errinfo);
1917             err = parse_events(evsel_list, str, &errinfo);
1918             if (err) {
1919                 fprintf(stderr,
1920                     "Cannot set up top down events %s: %d\n",
1921                     str, err);
1922                 parse_events_error__print(&errinfo, str);
1923                 parse_events_error__exit(&errinfo);
1924                 free(str);
1925                 return -1;
1926             }
1927             parse_events_error__exit(&errinfo);
1928         } else {
1929             fprintf(stderr, "System does not support topdown\n");
1930             return -1;
1931         }
1932         free(str);
1933     }
1934
1935     if (!stat_config.topdown_level)
1936         stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1937
1938     if (!evsel_list->core.nr_entries) {
1939         if (target__has_cpu(&target))
1940             default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1941
1942         if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1943             return -1;
1944         if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1945             if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1946                 return -1;
1947         }
1948         if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1949             if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1950                 return -1;
1951         }
1952         if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1953             return -1;
1954         /* Platform specific attrs */
1955         if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
1956             return -1;
1957     }
1958
1959     /* Detailed events get appended to the event list: */
1960
1961     if (detailed_run <  1)
1962         return 0;
1963
1964     /* Append detailed run extra attributes: */
1965     if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1966         return -1;
1967
1968     if (detailed_run < 2)
1969         return 0;
1970
1971     /* Append very detailed run extra attributes: */
1972     if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1973         return -1;
1974
1975     if (detailed_run < 3)
1976         return 0;
1977
1978     /* Append very, very detailed run extra attributes: */
1979     return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1980 }
1981
1982 static const char * const stat_record_usage[] = {
1983     "perf stat record [<options>]",
1984     NULL,
1985 };
1986
1987 static void init_features(struct perf_session *session)
1988 {
1989     int feat;
1990
1991     for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1992         perf_header__set_feat(&session->header, feat);
1993
1994     perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1995     perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1996     perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1997     perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1998     perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1999 }
2000
2001 static int __cmd_record(int argc, const char **argv)
2002 {
2003     struct perf_session *session;
2004     struct perf_data *data = &perf_stat.data;
2005
2006     argc = parse_options(argc, argv, stat_options, stat_record_usage,
2007                  PARSE_OPT_STOP_AT_NON_OPTION);
2008
2009     if (output_name)
2010         data->path = output_name;
2011
2012     if (stat_config.run_count != 1 || forever) {
2013         pr_err("Cannot use -r option with perf stat record.\n");
2014         return -1;
2015     }
2016
2017     session = perf_session__new(data, NULL);
2018     if (IS_ERR(session)) {
2019         pr_err("Perf session creation failed\n");
2020         return PTR_ERR(session);
2021     }
2022
2023     init_features(session);
2024
2025     session->evlist   = evsel_list;
2026     perf_stat.session = session;
2027     perf_stat.record  = true;
2028     return argc;
2029 }
2030
2031 static int process_stat_round_event(struct perf_session *session,
2032                     union perf_event *event)
2033 {
2034     struct perf_record_stat_round *stat_round = &event->stat_round;
2035     struct evsel *counter;
2036     struct timespec tsh, *ts = NULL;
2037     const char **argv = session->header.env.cmdline_argv;
2038     int argc = session->header.env.nr_cmdline;
2039
2040     evlist__for_each_entry(evsel_list, counter)
2041         perf_stat_process_counter(&stat_config, counter);
2042
2043     if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2044         update_stats(&walltime_nsecs_stats, stat_round->time);
2045
2046     if (stat_config.interval && stat_round->time) {
2047         tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2048         tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2049         ts = &tsh;
2050     }
2051
2052     print_counters(ts, argc, argv);
2053     return 0;
2054 }
2055
2056 static
2057 int process_stat_config_event(struct perf_session *session,
2058                   union perf_event *event)
2059 {
2060     struct perf_tool *tool = session->tool;
2061     struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2062
2063     perf_event__read_stat_config(&stat_config, &event->stat_config);
2064
2065     if (perf_cpu_map__empty(st->cpus)) {
2066         if (st->aggr_mode != AGGR_UNSET)
2067             pr_warning("warning: processing task data, aggregation mode not set\n");
2068         return 0;
2069     }
2070
2071     if (st->aggr_mode != AGGR_UNSET)
2072         stat_config.aggr_mode = st->aggr_mode;
2073
2074     if (perf_stat.data.is_pipe)
2075         perf_stat_init_aggr_mode();
2076     else
2077         perf_stat_init_aggr_mode_file(st);
2078
2079     return 0;
2080 }
2081
2082 static int set_maps(struct perf_stat *st)
2083 {
2084     if (!st->cpus || !st->threads)
2085         return 0;
2086
2087     if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2088         return -EINVAL;
2089
2090     perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2091
2092     if (evlist__alloc_stats(evsel_list, true))
2093         return -ENOMEM;
2094
2095     st->maps_allocated = true;
2096     return 0;
2097 }
2098
2099 static
2100 int process_thread_map_event(struct perf_session *session,
2101                  union perf_event *event)
2102 {
2103     struct perf_tool *tool = session->tool;
2104     struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2105
2106     if (st->threads) {
2107         pr_warning("Extra thread map event, ignoring.\n");
2108         return 0;
2109     }
2110
2111     st->threads = thread_map__new_event(&event->thread_map);
2112     if (!st->threads)
2113         return -ENOMEM;
2114
2115     return set_maps(st);
2116 }
2117
2118 static
2119 int process_cpu_map_event(struct perf_session *session,
2120               union perf_event *event)
2121 {
2122     struct perf_tool *tool = session->tool;
2123     struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2124     struct perf_cpu_map *cpus;
2125
2126     if (st->cpus) {
2127         pr_warning("Extra cpu map event, ignoring.\n");
2128         return 0;
2129     }
2130
2131     cpus = cpu_map__new_data(&event->cpu_map.data);
2132     if (!cpus)
2133         return -ENOMEM;
2134
2135     st->cpus = cpus;
2136     return set_maps(st);
2137 }
2138
2139 static const char * const stat_report_usage[] = {
2140     "perf stat report [<options>]",
2141     NULL,
2142 };
2143
2144 static struct perf_stat perf_stat = {
2145     .tool = {
2146         .attr       = perf_event__process_attr,
2147         .event_update   = perf_event__process_event_update,
2148         .thread_map = process_thread_map_event,
2149         .cpu_map    = process_cpu_map_event,
2150         .stat_config    = process_stat_config_event,
2151         .stat       = perf_event__process_stat_event,
2152         .stat_round = process_stat_round_event,
2153     },
2154     .aggr_mode = AGGR_UNSET,
2155 };
2156
2157 static int __cmd_report(int argc, const char **argv)
2158 {
2159     struct perf_session *session;
2160     const struct option options[] = {
2161     OPT_STRING('i', "input", &input_name, "file", "input file name"),
2162     OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2163              "aggregate counts per processor socket", AGGR_SOCKET),
2164     OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2165              "aggregate counts per processor die", AGGR_DIE),
2166     OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2167              "aggregate counts per physical processor core", AGGR_CORE),
2168     OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2169              "aggregate counts per numa node", AGGR_NODE),
2170     OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2171              "disable CPU count aggregation", AGGR_NONE),
2172     OPT_END()
2173     };
2174     struct stat st;
2175     int ret;
2176
2177     argc = parse_options(argc, argv, options, stat_report_usage, 0);
2178
2179     if (!input_name || !strlen(input_name)) {
2180         if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2181             input_name = "-";
2182         else
2183             input_name = "perf.data";
2184     }
2185
2186     perf_stat.data.path = input_name;
2187     perf_stat.data.mode = PERF_DATA_MODE_READ;
2188
2189     session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2190     if (IS_ERR(session))
2191         return PTR_ERR(session);
2192
2193     perf_stat.session  = session;
2194     stat_config.output = stderr;
2195     evsel_list         = session->evlist;
2196
2197     ret = perf_session__process_events(session);
2198     if (ret)
2199         return ret;
2200
2201     perf_session__delete(session);
2202     return 0;
2203 }
2204
2205 static void setup_system_wide(int forks)
2206 {
2207     /*
2208      * Make system wide (-a) the default target if
2209      * no target was specified and one of following
2210      * conditions is met:
2211      *
2212      *   - there's no workload specified
2213      *   - there is workload specified but all requested
2214      *     events are system wide events
2215      */
2216     if (!target__none(&target))
2217         return;
2218
2219     if (!forks)
2220         target.system_wide = true;
2221     else {
2222         struct evsel *counter;
2223
2224         evlist__for_each_entry(evsel_list, counter) {
2225             if (!counter->core.requires_cpu &&
2226                 strcmp(counter->name, "duration_time")) {
2227                 return;
2228             }
2229         }
2230
2231         if (evsel_list->core.nr_entries)
2232             target.system_wide = true;
2233     }
2234 }
2235
2236 int cmd_stat(int argc, const char **argv)
2237 {
2238     const char * const stat_usage[] = {
2239         "perf stat [<options>] [<command>]",
2240         NULL
2241     };
2242     int status = -EINVAL, run_idx, err;
2243     const char *mode;
2244     FILE *output = stderr;
2245     unsigned int interval, timeout;
2246     const char * const stat_subcommands[] = { "record", "report" };
2247     char errbuf[BUFSIZ];
2248
2249     setlocale(LC_ALL, "");
2250
2251     evsel_list = evlist__new();
2252     if (evsel_list == NULL)
2253         return -ENOMEM;
2254
2255     parse_events__shrink_config_terms();
2256
2257     /* String-parsing callback-based options would segfault when negated */
2258     set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2259     set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2260     set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2261
2262     argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2263                     (const char **) stat_usage,
2264                     PARSE_OPT_STOP_AT_NON_OPTION);
2265     perf_stat__collect_metric_expr(evsel_list);
2266     perf_stat__init_shadow_stats();
2267
2268     if (stat_config.csv_sep) {
2269         stat_config.csv_output = true;
2270         if (!strcmp(stat_config.csv_sep, "\\t"))
2271             stat_config.csv_sep = "\t";
2272     } else
2273         stat_config.csv_sep = DEFAULT_SEPARATOR;
2274
2275     if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2276         argc = __cmd_record(argc, argv);
2277         if (argc < 0)
2278             return -1;
2279     } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2280         return __cmd_report(argc, argv);
2281
2282     interval = stat_config.interval;
2283     timeout = stat_config.timeout;
2284
2285     /*
2286      * For record command the -o is already taken care of.
2287      */
2288     if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2289         output = NULL;
2290
2291     if (output_name && output_fd) {
2292         fprintf(stderr, "cannot use both --output and --log-fd\n");
2293         parse_options_usage(stat_usage, stat_options, "o", 1);
2294         parse_options_usage(NULL, stat_options, "log-fd", 0);
2295         goto out;
2296     }
2297
2298     if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2299         fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2300         goto out;
2301     }
2302
2303     if (stat_config.metric_only && stat_config.run_count > 1) {
2304         fprintf(stderr, "--metric-only is not supported with -r\n");
2305         goto out;
2306     }
2307
2308     if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2309         fprintf(stderr, "--table is only supported with -r\n");
2310         parse_options_usage(stat_usage, stat_options, "r", 1);
2311         parse_options_usage(NULL, stat_options, "table", 0);
2312         goto out;
2313     }
2314
2315     if (output_fd < 0) {
2316         fprintf(stderr, "argument to --log-fd must be a > 0\n");
2317         parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2318         goto out;
2319     }
2320
2321     if (!output && !stat_config.quiet) {
2322         struct timespec tm;
2323         mode = append_file ? "a" : "w";
2324
2325         output = fopen(output_name, mode);
2326         if (!output) {
2327             perror("failed to create output file");
2328             return -1;
2329         }
2330         clock_gettime(CLOCK_REALTIME, &tm);
2331         fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2332     } else if (output_fd > 0) {
2333         mode = append_file ? "a" : "w";
2334         output = fdopen(output_fd, mode);
2335         if (!output) {
2336             perror("Failed opening logfd");
2337             return -errno;
2338         }
2339     }
2340
2341     stat_config.output = output;
2342
2343     /*
2344      * let the spreadsheet do the pretty-printing
2345      */
2346     if (stat_config.csv_output) {
2347         /* User explicitly passed -B? */
2348         if (big_num_opt == 1) {
2349             fprintf(stderr, "-B option not supported with -x\n");
2350             parse_options_usage(stat_usage, stat_options, "B", 1);
2351             parse_options_usage(NULL, stat_options, "x", 1);
2352             goto out;
2353         } else /* Nope, so disable big number formatting */
2354             stat_config.big_num = false;
2355     } else if (big_num_opt == 0) /* User passed --no-big-num */
2356         stat_config.big_num = false;
2357
2358     err = target__validate(&target);
2359     if (err) {
2360         target__strerror(&target, err, errbuf, BUFSIZ);
2361         pr_warning("%s\n", errbuf);
2362     }
2363
2364     setup_system_wide(argc);
2365
2366     /*
2367      * Display user/system times only for single
2368      * run and when there's specified tracee.
2369      */
2370     if ((stat_config.run_count == 1) && target__none(&target))
2371         stat_config.ru_display = true;
2372
2373     if (stat_config.run_count < 0) {
2374         pr_err("Run count must be a positive number\n");
2375         parse_options_usage(stat_usage, stat_options, "r", 1);
2376         goto out;
2377     } else if (stat_config.run_count == 0) {
2378         forever = true;
2379         stat_config.run_count = 1;
2380     }
2381
2382     if (stat_config.walltime_run_table) {
2383         stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2384         if (!stat_config.walltime_run) {
2385             pr_err("failed to setup -r option");
2386             goto out;
2387         }
2388     }
2389
2390     if ((stat_config.aggr_mode == AGGR_THREAD) &&
2391         !target__has_task(&target)) {
2392         if (!target.system_wide || target.cpu_list) {
2393             fprintf(stderr, "The --per-thread option is only "
2394                 "available when monitoring via -p -t -a "
2395                 "options or only --per-thread.\n");
2396             parse_options_usage(NULL, stat_options, "p", 1);
2397             parse_options_usage(NULL, stat_options, "t", 1);
2398             goto out;
2399         }
2400     }
2401
2402     /*
2403      * no_aggr, cgroup are for system-wide only
2404      * --per-thread is aggregated per thread, we dont mix it with cpu mode
2405      */
2406     if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2407           stat_config.aggr_mode != AGGR_THREAD) ||
2408          (nr_cgroups || stat_config.cgroup_list)) &&
2409         !target__has_cpu(&target)) {
2410         fprintf(stderr, "both cgroup and no-aggregation "
2411             "modes only available in system-wide mode\n");
2412
2413         parse_options_usage(stat_usage, stat_options, "G", 1);
2414         parse_options_usage(NULL, stat_options, "A", 1);
2415         parse_options_usage(NULL, stat_options, "a", 1);
2416         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2417         goto out;
2418     }
2419
2420     if (stat_config.iostat_run) {
2421         status = iostat_prepare(evsel_list, &stat_config);
2422         if (status)
2423             goto out;
2424         if (iostat_mode == IOSTAT_LIST) {
2425             iostat_list(evsel_list, &stat_config);
2426             goto out;
2427         } else if (verbose)
2428             iostat_list(evsel_list, &stat_config);
2429         if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2430             target.system_wide = true;
2431     }
2432
2433     if (add_default_attributes())
2434         goto out;
2435
2436     if (stat_config.cgroup_list) {
2437         if (nr_cgroups > 0) {
2438             pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2439             parse_options_usage(stat_usage, stat_options, "G", 1);
2440             parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2441             goto out;
2442         }
2443
2444         if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2445                       &stat_config.metric_events, true) < 0) {
2446             parse_options_usage(stat_usage, stat_options,
2447                         "for-each-cgroup", 0);
2448             goto out;
2449         }
2450     }
2451
2452     if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2453         target.per_thread = true;
2454
2455     if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2456         pr_err("failed to use cpu list %s\n", target.cpu_list);
2457         goto out;
2458     }
2459
2460     target.hybrid = perf_pmu__has_hybrid();
2461     if (evlist__create_maps(evsel_list, &target) < 0) {
2462         if (target__has_task(&target)) {
2463             pr_err("Problems finding threads of monitor\n");
2464             parse_options_usage(stat_usage, stat_options, "p", 1);
2465             parse_options_usage(NULL, stat_options, "t", 1);
2466         } else if (target__has_cpu(&target)) {
2467             perror("failed to parse CPUs map");
2468             parse_options_usage(stat_usage, stat_options, "C", 1);
2469             parse_options_usage(NULL, stat_options, "a", 1);
2470         }
2471         goto out;
2472     }
2473
2474     evlist__check_cpu_maps(evsel_list);
2475
2476     /*
2477      * Initialize thread_map with comm names,
2478      * so we could print it out on output.
2479      */
2480     if (stat_config.aggr_mode == AGGR_THREAD) {
2481         thread_map__read_comms(evsel_list->core.threads);
2482         if (target.system_wide) {
2483             if (runtime_stat_new(&stat_config,
2484                 perf_thread_map__nr(evsel_list->core.threads))) {
2485                 goto out;
2486             }
2487         }
2488     }
2489
2490     if (stat_config.aggr_mode == AGGR_NODE)
2491         cpu__setup_cpunode_map();
2492
2493     if (stat_config.times && interval)
2494         interval_count = true;
2495     else if (stat_config.times && !interval) {
2496         pr_err("interval-count option should be used together with "
2497                 "interval-print.\n");
2498         parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2499         parse_options_usage(stat_usage, stat_options, "I", 1);
2500         goto out;
2501     }
2502
2503     if (timeout && timeout < 100) {
2504         if (timeout < 10) {
2505             pr_err("timeout must be >= 10ms.\n");
2506             parse_options_usage(stat_usage, stat_options, "timeout", 0);
2507             goto out;
2508         } else
2509             pr_warning("timeout < 100ms. "
2510                    "The overhead percentage could be high in some cases. "
2511                    "Please proceed with caution.\n");
2512     }
2513     if (timeout && interval) {
2514         pr_err("timeout option is not supported with interval-print.\n");
2515         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2516         parse_options_usage(stat_usage, stat_options, "I", 1);
2517         goto out;
2518     }
2519
2520     if (evlist__alloc_stats(evsel_list, interval))
2521         goto out;
2522
2523     if (perf_stat_init_aggr_mode())
2524         goto out;
2525
2526     /*
2527      * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2528      * while avoiding that older tools show confusing messages.
2529      *
2530      * However for pipe sessions we need to keep it zero,
2531      * because script's perf_evsel__check_attr is triggered
2532      * by attr->sample_type != 0, and we can't run it on
2533      * stat sessions.
2534      */
2535     stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2536
2537     /*
2538      * We dont want to block the signals - that would cause
2539      * child tasks to inherit that and Ctrl-C would not work.
2540      * What we want is for Ctrl-C to work in the exec()-ed
2541      * task, but being ignored by perf stat itself:
2542      */
2543     atexit(sig_atexit);
2544     if (!forever)
2545         signal(SIGINT,  skip_signal);
2546     signal(SIGCHLD, skip_signal);
2547     signal(SIGALRM, skip_signal);
2548     signal(SIGABRT, skip_signal);
2549
2550     if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2551         goto out;
2552
2553     /* Enable ignoring missing threads when -p option is defined. */
2554     evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2555     status = 0;
2556     for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2557         if (stat_config.run_count != 1 && verbose > 0)
2558             fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2559                 run_idx + 1);
2560
2561         if (run_idx != 0)
2562             evlist__reset_prev_raw_counts(evsel_list);
2563
2564         status = run_perf_stat(argc, argv, run_idx);
2565         if (forever && status != -1 && !interval) {
2566             print_counters(NULL, argc, argv);
2567             perf_stat__reset_stats();
2568         }
2569     }
2570
2571     if (!forever && status != -1 && (!interval || stat_config.summary))
2572         print_counters(NULL, argc, argv);
2573
2574     evlist__finalize_ctlfd(evsel_list);
2575
2576     if (STAT_RECORD) {
2577         /*
2578          * We synthesize the kernel mmap record just so that older tools
2579          * don't emit warnings about not being able to resolve symbols
2580          * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2581          * a saner message about no samples being in the perf.data file.
2582          *
2583          * This also serves to suppress a warning about f_header.data.size == 0
2584          * in header.c at the moment 'perf stat record' gets introduced, which
2585          * is not really needed once we start adding the stat specific PERF_RECORD_
2586          * records, but the need to suppress the kptr_restrict messages in older
2587          * tools remain  -acme
2588          */
2589         int fd = perf_data__fd(&perf_stat.data);
2590
2591         err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2592                              process_synthesized_event,
2593                              &perf_stat.session->machines.host);
2594         if (err) {
2595             pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2596                    "older tools may produce warnings about this file\n.");
2597         }
2598
2599         if (!interval) {
2600             if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2601                 pr_err("failed to write stat round event\n");
2602         }
2603
2604         if (!perf_stat.data.is_pipe) {
2605             perf_stat.session->header.data_size += perf_stat.bytes_written;
2606             perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2607         }
2608
2609         evlist__close(evsel_list);
2610         perf_session__delete(perf_stat.session);
2611     }
2612
2613     perf_stat__exit_aggr_mode();
2614     evlist__free_stats(evsel_list);
2615 out:
2616     if (stat_config.iostat_run)
2617         iostat_release(evsel_list);
2618
2619     zfree(&stat_config.walltime_run);
2620
2621     if (smi_cost && smi_reset)
2622         sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2623
2624     evlist__delete(evsel_list);
2625
2626     metricgroup__rblist_exit(&stat_config.metric_events);
2627     runtime_stat_delete(&stat_config);
2628     evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2629
2630     return status;
2631 }