tools/perf/builtin-stat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-stat.c
   4  *
   5  * Builtin stat command: Give a precise performance counters summary
   6  * overview about any workload, CPU or specific PID.
   7  *
   8  * Sample output:
   9
  10    $ perf stat ./hackbench 10
  11
  12   Time: 0.118
  13
  14   Performance counter stats for './hackbench 10':
  15
  16        1708.761321 task-clock                #   11.037 CPUs utilized
  17             41,190 context-switches          #    0.024 M/sec
  18              6,735 CPU-migrations            #    0.004 M/sec
  19             17,318 page-faults               #    0.010 M/sec
  20      5,205,202,243 cycles                    #    3.046 GHz
  21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  23      2,603,501,247 instructions              #    0.50  insns per cycle
  24                                              #    1.48  stalled cycles per insn
  25        484,357,498 branches                  #  283.455 M/sec
  26          6,388,934 branch-misses             #    1.32% of all branches
  27
  28         0.154822978  seconds time elapsed
  29
  30  *
  31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  32  *
  33  * Improvements and fixes by:
  34  *
  35  *   Arjan van de Ven <arjan@linux.intel.com>
  36  *   Yanmin Zhang <yanmin.zhang@intel.com>
  37  *   Wu Fengguang <fengguang.wu@intel.com>
  38  *   Mike Galbraith <efault@gmx.de>
  39  *   Paul Mackerras <paulus@samba.org>
  40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  41  */
  42
  43 #include "builtin.h"
  44 #include "perf.h"
  45 #include "util/cgroup.h"
  46 #include <subcmd/parse-options.h>
  47 #include "util/parse-events.h"
  48 #include "util/pmu.h"
  49 #include "util/event.h"
  50 #include "util/evlist.h"
  51 #include "util/evsel.h"
  52 #include "util/debug.h"
  53 #include "util/color.h"
  54 #include "util/stat.h"
  55 #include "util/header.h"
  56 #include "util/cpumap.h"
  57 #include "util/thread_map.h"
  58 #include "util/counts.h"
  59 #include "util/topdown.h"
  60 #include "util/session.h"
  61 #include "util/tool.h"
  62 #include "util/string2.h"
  63 #include "util/metricgroup.h"
  64 #include "util/synthetic-events.h"
  65 #include "util/target.h"
  66 #include "util/time-utils.h"
  67 #include "util/top.h"
  68 #include "util/affinity.h"
  69 #include "util/pfm.h"
  70 #include "util/bpf_counter.h"
  71 #include "util/iostat.h"
  72 #include "asm/bug.h"
  73
  74 #include <linux/time64.h>
  75 #include <linux/zalloc.h>
  76 #include <api/fs/fs.h>
  77 #include <errno.h>
  78 #include <signal.h>
  79 #include <stdlib.h>
  80 #include <sys/prctl.h>
  81 #include <inttypes.h>
  82 #include <locale.h>
  83 #include <math.h>
  84 #include <sys/types.h>
  85 #include <sys/stat.h>
  86 #include <sys/wait.h>
  87 #include <unistd.h>
  88 #include <sys/time.h>
  89 #include <sys/resource.h>
  90 #include <linux/err.h>
  91
  92 #include <linux/ctype.h>
  93 #include <perf/evlist.h>
  94
  95 #define DEFAULT_SEPARATOR       " "
  96 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  97
  98 static void print_counters(struct timespec *ts, int argc, const char **argv);
  99
 100 /* Default events used for perf stat -T */
 101 static const char *transaction_attrs = {
 102         "task-clock,"
 103         "{"
 104         "instructions,"
 105         "cycles,"
 106         "cpu/cycles-t/,"
 107         "cpu/tx-start/,"
 108         "cpu/el-start/,"
 109         "cpu/cycles-ct/"
 110         "}"
 111 };
 112
 113 /* More limited version when the CPU does not have all events. */
 114 static const char * transaction_limited_attrs = {
 115         "task-clock,"
 116         "{"
 117         "instructions,"
 118         "cycles,"
 119         "cpu/cycles-t/,"
 120         "cpu/tx-start/"
 121         "}"
 122 };
 123
 124 static const char * topdown_attrs[] = {
 125         "topdown-total-slots",
 126         "topdown-slots-retired",
 127         "topdown-recovery-bubbles",
 128         "topdown-fetch-bubbles",
 129         "topdown-slots-issued",
 130         NULL,
 131 };
 132
 133 static const char *topdown_metric_attrs[] = {
 134         "slots",
 135         "topdown-retiring",
 136         "topdown-bad-spec",
 137         "topdown-fe-bound",
 138         "topdown-be-bound",
 139         NULL,
 140 };
 141
 142 static const char *topdown_metric_L2_attrs[] = {
 143         "slots",
 144         "topdown-retiring",
 145         "topdown-bad-spec",
 146         "topdown-fe-bound",
 147         "topdown-be-bound",
 148         "topdown-heavy-ops",
 149         "topdown-br-mispredict",
 150         "topdown-fetch-lat",
 151         "topdown-mem-bound",
 152         NULL,
 153 };
 154
 155 static const char *smi_cost_attrs = {
 156         "{"
 157         "msr/aperf/,"
 158         "msr/smi/,"
 159         "cycles"
 160         "}"
 161 };
 162
 163 static struct evlist    *evsel_list;
 164 static bool all_counters_use_bpf = true;
 165
 166 static struct target target = {
 167         .uid    = UINT_MAX,
 168 };
 169
 170 #define METRIC_ONLY_LEN 20
 171
 172 static volatile pid_t           child_pid                       = -1;
 173 static int                      detailed_run                    =  0;
 174 static bool                     transaction_run;
 175 static bool                     topdown_run                     = false;
 176 static bool                     smi_cost                        = false;
 177 static bool                     smi_reset                       = false;
 178 static int                      big_num_opt                     =  -1;
 179 static bool                     group                           = false;
 180 static const char               *pre_cmd                        = NULL;
 181 static const char               *post_cmd                       = NULL;
 182 static bool                     sync_run                        = false;
 183 static bool                     forever                         = false;
 184 static bool                     force_metric_only               = false;
 185 static struct timespec          ref_time;
 186 static bool                     append_file;
 187 static bool                     interval_count;
 188 static const char               *output_name;
 189 static int                      output_fd;
 190
 191 struct perf_stat {
 192         bool                     record;
 193         struct perf_data         data;
 194         struct perf_session     *session;
 195         u64                      bytes_written;
 196         struct perf_tool         tool;
 197         bool                     maps_allocated;
 198         struct perf_cpu_map     *cpus;
 199         struct perf_thread_map *threads;
 200         enum aggr_mode           aggr_mode;
 201 };
 202
 203 static struct perf_stat         perf_stat;
 204 #define STAT_RECORD             perf_stat.record
 205
 206 static volatile int done = 0;
 207
 208 static struct perf_stat_config stat_config = {
 209         .aggr_mode              = AGGR_GLOBAL,
 210         .scale                  = true,
 211         .unit_width             = 4, /* strlen("unit") */
 212         .run_count              = 1,
 213         .metric_only_len        = METRIC_ONLY_LEN,
 214         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 215         .big_num                = true,
 216         .ctl_fd                 = -1,
 217         .ctl_fd_ack             = -1,
 218         .iostat_run             = false,
 219 };
 220
 221 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
 222 {
 223         if (!a->core.cpus && !b->core.cpus)
 224                 return true;
 225
 226         if (!a->core.cpus || !b->core.cpus)
 227                 return false;
 228
 229         if (a->core.cpus->nr != b->core.cpus->nr)
 230                 return false;
 231
 232         for (int i = 0; i < a->core.cpus->nr; i++) {
 233                 if (a->core.cpus->map[i] != b->core.cpus->map[i])
 234                         return false;
 235         }
 236
 237         return true;
 238 }
 239
 240 static void evlist__check_cpu_maps(struct evlist *evlist)
 241 {
 242         struct evsel *evsel, *pos, *leader;
 243         char buf[1024];
 244
 245         evlist__for_each_entry(evlist, evsel) {
 246                 leader = evsel->leader;
 247
 248                 /* Check that leader matches cpus with each member. */
 249                 if (leader == evsel)
 250                         continue;
 251                 if (cpus_map_matched(leader, evsel))
 252                         continue;
 253
 254                 /* If there's mismatch disable the group and warn user. */
 255                 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
 256                 evsel__group_desc(leader, buf, sizeof(buf));
 257                 pr_warning("  %s\n", buf);
 258
 259                 if (verbose) {
 260                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 261                         pr_warning("     %s: %s\n", leader->name, buf);
 262                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 263                         pr_warning("     %s: %s\n", evsel->name, buf);
 264                 }
 265
 266                 for_each_group_evsel(pos, leader) {
 267                         pos->leader = pos;
 268                         pos->core.nr_members = 0;
 269                 }
 270                 evsel->leader->core.nr_members = 0;
 271         }
 272 }
 273
 274 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 275                                  struct timespec *b)
 276 {
 277         r->tv_sec = a->tv_sec - b->tv_sec;
 278         if (a->tv_nsec < b->tv_nsec) {
 279                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 280                 r->tv_sec--;
 281         } else {
 282                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 283         }
 284 }
 285
 286 static void perf_stat__reset_stats(void)
 287 {
 288         int i;
 289
 290         evlist__reset_stats(evsel_list);
 291         perf_stat__reset_shadow_stats();
 292
 293         for (i = 0; i < stat_config.stats_num; i++)
 294                 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
 295 }
 296
 297 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 298                                      union perf_event *event,
 299                                      struct perf_sample *sample __maybe_unused,
 300                                      struct machine *machine __maybe_unused)
 301 {
 302         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 303                 pr_err("failed to write perf data, error: %m\n");
 304                 return -1;
 305         }
 306
 307         perf_stat.bytes_written += event->header.size;
 308         return 0;
 309 }
 310
 311 static int write_stat_round_event(u64 tm, u64 type)
 312 {
 313         return perf_event__synthesize_stat_round(NULL, tm, type,
 314                                                  process_synthesized_event,
 315                                                  NULL);
 316 }
 317
 318 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 319         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 320
 321 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 322
 323 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
 324                                    struct perf_counts_values *count)
 325 {
 326         struct perf_sample_id *sid = SID(counter, cpu, thread);
 327
 328         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 329                                            process_synthesized_event, NULL);
 330 }
 331
 332 static int read_single_counter(struct evsel *counter, int cpu,
 333                                int thread, struct timespec *rs)
 334 {
 335         if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
 336                 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 337                 struct perf_counts_values *count =
 338                         perf_counts(counter->counts, cpu, thread);
 339                 count->ena = count->run = val;
 340                 count->val = val;
 341                 return 0;
 342         }
 343         return evsel__read_counter(counter, cpu, thread);
 344 }
 345
 346 /*
 347  * Read out the results of a single counter:
 348  * do not aggregate counts across CPUs in system-wide mode
 349  */
 350 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
 351 {
 352         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 353         int thread;
 354
 355         if (!counter->supported)
 356                 return -ENOENT;
 357
 358         if (counter->core.system_wide)
 359                 nthreads = 1;
 360
 361         for (thread = 0; thread < nthreads; thread++) {
 362                 struct perf_counts_values *count;
 363
 364                 count = perf_counts(counter->counts, cpu, thread);
 365
 366                 /*
 367                  * The leader's group read loads data into its group members
 368                  * (via evsel__read_counter()) and sets their count->loaded.
 369                  */
 370                 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
 371                     read_single_counter(counter, cpu, thread, rs)) {
 372                         counter->counts->scaled = -1;
 373                         perf_counts(counter->counts, cpu, thread)->ena = 0;
 374                         perf_counts(counter->counts, cpu, thread)->run = 0;
 375                         return -1;
 376                 }
 377
 378                 perf_counts__set_loaded(counter->counts, cpu, thread, false);
 379
 380                 if (STAT_RECORD) {
 381                         if (evsel__write_stat_event(counter, cpu, thread, count)) {
 382                                 pr_err("failed to write stat event\n");
 383                                 return -1;
 384                         }
 385                 }
 386
 387                 if (verbose > 1) {
 388                         fprintf(stat_config.output,
 389                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 390                                         evsel__name(counter),
 391                                         cpu,
 392                                         count->val, count->ena, count->run);
 393                 }
 394         }
 395
 396         return 0;
 397 }
 398
 399 static int read_affinity_counters(struct timespec *rs)
 400 {
 401         struct evsel *counter;
 402         struct affinity affinity;
 403         int i, ncpus, cpu;
 404
 405         if (all_counters_use_bpf)
 406                 return 0;
 407
 408         if (affinity__setup(&affinity) < 0)
 409                 return -1;
 410
 411         ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
 412         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 413                 ncpus = 1;
 414         evlist__for_each_cpu(evsel_list, i, cpu) {
 415                 if (i >= ncpus)
 416                         break;
 417                 affinity__set(&affinity, cpu);
 418
 419                 evlist__for_each_entry(evsel_list, counter) {
 420                         if (evsel__cpu_iter_skip(counter, cpu))
 421                                 continue;
 422                         if (evsel__is_bpf(counter))
 423                                 continue;
 424                         if (!counter->err) {
 425                                 counter->err = read_counter_cpu(counter, rs,
 426                                                                 counter->cpu_iter - 1);
 427                         }
 428                 }
 429         }
 430         affinity__cleanup(&affinity);
 431         return 0;
 432 }
 433
 434 static int read_bpf_map_counters(void)
 435 {
 436         struct evsel *counter;
 437         int err;
 438
 439         evlist__for_each_entry(evsel_list, counter) {
 440                 if (!evsel__is_bpf(counter))
 441                         continue;
 442
 443                 err = bpf_counter__read(counter);
 444                 if (err)
 445                         return err;
 446         }
 447         return 0;
 448 }
 449
 450 static void read_counters(struct timespec *rs)
 451 {
 452         struct evsel *counter;
 453
 454         if (!stat_config.stop_read_counter) {
 455                 if (read_bpf_map_counters() ||
 456                     read_affinity_counters(rs))
 457                         return;
 458         }
 459
 460         evlist__for_each_entry(evsel_list, counter) {
 461                 if (counter->err)
 462                         pr_debug("failed to read counter %s\n", counter->name);
 463                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 464                         pr_warning("failed to process counter %s\n", counter->name);
 465                 counter->err = 0;
 466         }
 467 }
 468
 469 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
 470 {
 471         int i;
 472
 473         config->stats = calloc(nthreads, sizeof(struct runtime_stat));
 474         if (!config->stats)
 475                 return -1;
 476
 477         config->stats_num = nthreads;
 478
 479         for (i = 0; i < nthreads; i++)
 480                 runtime_stat__init(&config->stats[i]);
 481
 482         return 0;
 483 }
 484
 485 static void runtime_stat_delete(struct perf_stat_config *config)
 486 {
 487         int i;
 488
 489         if (!config->stats)
 490                 return;
 491
 492         for (i = 0; i < config->stats_num; i++)
 493                 runtime_stat__exit(&config->stats[i]);
 494
 495         zfree(&config->stats);
 496 }
 497
 498 static void runtime_stat_reset(struct perf_stat_config *config)
 499 {
 500         int i;
 501
 502         if (!config->stats)
 503                 return;
 504
 505         for (i = 0; i < config->stats_num; i++)
 506                 perf_stat__reset_shadow_per_stat(&config->stats[i]);
 507 }
 508
 509 static void process_interval(void)
 510 {
 511         struct timespec ts, rs;
 512
 513         clock_gettime(CLOCK_MONOTONIC, &ts);
 514         diff_timespec(&rs, &ts, &ref_time);
 515
 516         perf_stat__reset_shadow_per_stat(&rt_stat);
 517         runtime_stat_reset(&stat_config);
 518         read_counters(&rs);
 519
 520         if (STAT_RECORD) {
 521                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 522                         pr_err("failed to write stat round event\n");
 523         }
 524
 525         init_stats(&walltime_nsecs_stats);
 526         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 527         print_counters(&rs, 0, NULL);
 528 }
 529
 530 static bool handle_interval(unsigned int interval, int *times)
 531 {
 532         if (interval) {
 533                 process_interval();
 534                 if (interval_count && !(--(*times)))
 535                         return true;
 536         }
 537         return false;
 538 }
 539
 540 static int enable_counters(void)
 541 {
 542         struct evsel *evsel;
 543         int err;
 544
 545         evlist__for_each_entry(evsel_list, evsel) {
 546                 if (!evsel__is_bpf(evsel))
 547                         continue;
 548
 549                 err = bpf_counter__enable(evsel);
 550                 if (err)
 551                         return err;
 552         }
 553
 554         if (stat_config.initial_delay < 0) {
 555                 pr_info(EVLIST_DISABLED_MSG);
 556                 return 0;
 557         }
 558
 559         if (stat_config.initial_delay > 0) {
 560                 pr_info(EVLIST_DISABLED_MSG);
 561                 usleep(stat_config.initial_delay * USEC_PER_MSEC);
 562         }
 563
 564         /*
 565          * We need to enable counters only if:
 566          * - we don't have tracee (attaching to task or cpu)
 567          * - we have initial delay configured
 568          */
 569         if (!target__none(&target) || stat_config.initial_delay) {
 570                 evlist__enable(evsel_list);
 571                 if (stat_config.initial_delay > 0)
 572                         pr_info(EVLIST_ENABLED_MSG);
 573         }
 574         return 0;
 575 }
 576
 577 static void disable_counters(void)
 578 {
 579         /*
 580          * If we don't have tracee (attaching to task or cpu), counters may
 581          * still be running. To get accurate group ratios, we must stop groups
 582          * from counting before reading their constituent counters.
 583          */
 584         if (!target__none(&target))
 585                 evlist__disable(evsel_list);
 586 }
 587
 588 static volatile int workload_exec_errno;
 589
 590 /*
 591  * evlist__prepare_workload will send a SIGUSR1
 592  * if the fork fails, since we asked by setting its
 593  * want_signal to true.
 594  */
 595 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 596                                         void *ucontext __maybe_unused)
 597 {
 598         workload_exec_errno = info->si_value.sival_int;
 599 }
 600
 601 static bool evsel__should_store_id(struct evsel *counter)
 602 {
 603         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 604 }
 605
 606 static bool is_target_alive(struct target *_target,
 607                             struct perf_thread_map *threads)
 608 {
 609         struct stat st;
 610         int i;
 611
 612         if (!target__has_task(_target))
 613                 return true;
 614
 615         for (i = 0; i < threads->nr; i++) {
 616                 char path[PATH_MAX];
 617
 618                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 619                           threads->map[i].pid);
 620
 621                 if (!stat(path, &st))
 622                         return true;
 623         }
 624
 625         return false;
 626 }
 627
 628 static void process_evlist(struct evlist *evlist, unsigned int interval)
 629 {
 630         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 631
 632         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 633                 switch (cmd) {
 634                 case EVLIST_CTL_CMD_ENABLE:
 635                         if (interval)
 636                                 process_interval();
 637                         break;
 638                 case EVLIST_CTL_CMD_DISABLE:
 639                         if (interval)
 640                                 process_interval();
 641                         break;
 642                 case EVLIST_CTL_CMD_SNAPSHOT:
 643                 case EVLIST_CTL_CMD_ACK:
 644                 case EVLIST_CTL_CMD_UNSUPPORTED:
 645                 case EVLIST_CTL_CMD_EVLIST:
 646                 case EVLIST_CTL_CMD_STOP:
 647                 case EVLIST_CTL_CMD_PING:
 648                 default:
 649                         break;
 650                 }
 651         }
 652 }
 653
 654 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 655                         int *time_to_sleep)
 656 {
 657         int tts = *time_to_sleep;
 658         struct timespec time_diff;
 659
 660         diff_timespec(&time_diff, time_stop, time_start);
 661
 662         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 663                time_diff.tv_nsec / NSEC_PER_MSEC;
 664
 665         if (tts < 0)
 666                 tts = 0;
 667
 668         *time_to_sleep = tts;
 669 }
 670
 671 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 672 {
 673         int child_exited = 0, status = 0;
 674         int time_to_sleep, sleep_time;
 675         struct timespec time_start, time_stop;
 676
 677         if (interval)
 678                 sleep_time = interval;
 679         else if (timeout)
 680                 sleep_time = timeout;
 681         else
 682                 sleep_time = 1000;
 683
 684         time_to_sleep = sleep_time;
 685
 686         while (!done) {
 687                 if (forks)
 688                         child_exited = waitpid(child_pid, &status, WNOHANG);
 689                 else
 690                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 691
 692                 if (child_exited)
 693                         break;
 694
 695                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 696                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 697                         if (timeout || handle_interval(interval, times))
 698                                 break;
 699                         time_to_sleep = sleep_time;
 700                 } else { /* fd revent */
 701                         process_evlist(evsel_list, interval);
 702                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 703                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 704                 }
 705         }
 706
 707         return status;
 708 }
 709
 710 enum counter_recovery {
 711         COUNTER_SKIP,
 712         COUNTER_RETRY,
 713         COUNTER_FATAL,
 714 };
 715
 716 static enum counter_recovery stat_handle_error(struct evsel *counter)
 717 {
 718         char msg[BUFSIZ];
 719         /*
 720          * PPC returns ENXIO for HW counters until 2.6.37
 721          * (behavior changed with commit b0a873e).
 722          */
 723         if (errno == EINVAL || errno == ENOSYS ||
 724             errno == ENOENT || errno == EOPNOTSUPP ||
 725             errno == ENXIO) {
 726                 if (verbose > 0)
 727                         ui__warning("%s event is not supported by the kernel.\n",
 728                                     evsel__name(counter));
 729                 counter->supported = false;
 730                 /*
 731                  * errored is a sticky flag that means one of the counter's
 732                  * cpu event had a problem and needs to be reexamined.
 733                  */
 734                 counter->errored = true;
 735
 736                 if ((counter->leader != counter) ||
 737                     !(counter->leader->core.nr_members > 1))
 738                         return COUNTER_SKIP;
 739         } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
 740                 if (verbose > 0)
 741                         ui__warning("%s\n", msg);
 742                 return COUNTER_RETRY;
 743         } else if (target__has_per_thread(&target) &&
 744                    evsel_list->core.threads &&
 745                    evsel_list->core.threads->err_thread != -1) {
 746                 /*
 747                  * For global --per-thread case, skip current
 748                  * error thread.
 749                  */
 750                 if (!thread_map__remove(evsel_list->core.threads,
 751                                         evsel_list->core.threads->err_thread)) {
 752                         evsel_list->core.threads->err_thread = -1;
 753                         return COUNTER_RETRY;
 754                 }
 755         }
 756
 757         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 758         ui__error("%s\n", msg);
 759
 760         if (child_pid != -1)
 761                 kill(child_pid, SIGTERM);
 762         return COUNTER_FATAL;
 763 }
 764
 765 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 766 {
 767         int interval = stat_config.interval;
 768         int times = stat_config.times;
 769         int timeout = stat_config.timeout;
 770         char msg[BUFSIZ];
 771         unsigned long long t0, t1;
 772         struct evsel *counter;
 773         size_t l;
 774         int status = 0;
 775         const bool forks = (argc > 0);
 776         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 777         struct affinity affinity;
 778         int i, cpu, err;
 779         bool second_pass = false;
 780
 781         if (forks) {
 782                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 783                         perror("failed to prepare workload");
 784                         return -1;
 785                 }
 786                 child_pid = evsel_list->workload.pid;
 787         }
 788
 789         if (group)
 790                 evlist__set_leader(evsel_list);
 791
 792         if (affinity__setup(&affinity) < 0)
 793                 return -1;
 794
 795         evlist__for_each_entry(evsel_list, counter) {
 796                 if (bpf_counter__load(counter, &target))
 797                         return -1;
 798                 if (!evsel__is_bpf(counter))
 799                         all_counters_use_bpf = false;
 800         }
 801
 802         evlist__for_each_cpu (evsel_list, i, cpu) {
 803                 /*
 804                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 805                  * no need to call it again here.
 806                  */
 807                 if (target.use_bpf)
 808                         break;
 809                 affinity__set(&affinity, cpu);
 810
 811                 evlist__for_each_entry(evsel_list, counter) {
 812                         if (evsel__cpu_iter_skip(counter, cpu))
 813                                 continue;
 814                         if (counter->reset_group || counter->errored)
 815                                 continue;
 816                         if (evsel__is_bpf(counter))
 817                                 continue;
 818 try_again:
 819                         if (create_perf_stat_counter(counter, &stat_config, &target,
 820                                                      counter->cpu_iter - 1) < 0) {
 821
 822                                 /*
 823                                  * Weak group failed. We cannot just undo this here
 824                                  * because earlier CPUs might be in group mode, and the kernel
 825                                  * doesn't support mixing group and non group reads. Defer
 826                                  * it to later.
 827                                  * Don't close here because we're in the wrong affinity.
 828                                  */
 829                                 if ((errno == EINVAL || errno == EBADF) &&
 830                                     counter->leader != counter &&
 831                                     counter->weak_group) {
 832                                         evlist__reset_weak_group(evsel_list, counter, false);
 833                                         assert(counter->reset_group);
 834                                         second_pass = true;
 835                                         continue;
 836                                 }
 837
 838                                 switch (stat_handle_error(counter)) {
 839                                 case COUNTER_FATAL:
 840                                         return -1;
 841                                 case COUNTER_RETRY:
 842                                         goto try_again;
 843                                 case COUNTER_SKIP:
 844                                         continue;
 845                                 default:
 846                                         break;
 847                                 }
 848
 849                         }
 850                         counter->supported = true;
 851                 }
 852         }
 853
 854         if (second_pass) {
 855                 /*
 856                  * Now redo all the weak group after closing them,
 857                  * and also close errored counters.
 858                  */
 859
 860                 evlist__for_each_cpu(evsel_list, i, cpu) {
 861                         affinity__set(&affinity, cpu);
 862                         /* First close errored or weak retry */
 863                         evlist__for_each_entry(evsel_list, counter) {
 864                                 if (!counter->reset_group && !counter->errored)
 865                                         continue;
 866                                 if (evsel__cpu_iter_skip_no_inc(counter, cpu))
 867                                         continue;
 868                                 perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
 869                         }
 870                         /* Now reopen weak */
 871                         evlist__for_each_entry(evsel_list, counter) {
 872                                 if (!counter->reset_group && !counter->errored)
 873                                         continue;
 874                                 if (evsel__cpu_iter_skip(counter, cpu))
 875                                         continue;
 876                                 if (!counter->reset_group)
 877                                         continue;
 878 try_again_reset:
 879                                 pr_debug2("reopening weak %s\n", evsel__name(counter));
 880                                 if (create_perf_stat_counter(counter, &stat_config, &target,
 881                                                              counter->cpu_iter - 1) < 0) {
 882
 883                                         switch (stat_handle_error(counter)) {
 884                                         case COUNTER_FATAL:
 885                                                 return -1;
 886                                         case COUNTER_RETRY:
 887                                                 goto try_again_reset;
 888                                         case COUNTER_SKIP:
 889                                                 continue;
 890                                         default:
 891                                                 break;
 892                                         }
 893                                 }
 894                                 counter->supported = true;
 895                         }
 896                 }
 897         }
 898         affinity__cleanup(&affinity);
 899
 900         evlist__for_each_entry(evsel_list, counter) {
 901                 if (!counter->supported) {
 902                         perf_evsel__free_fd(&counter->core);
 903                         continue;
 904                 }
 905
 906                 l = strlen(counter->unit);
 907                 if (l > stat_config.unit_width)
 908                         stat_config.unit_width = l;
 909
 910                 if (evsel__should_store_id(counter) &&
 911                     evsel__store_ids(counter, evsel_list))
 912                         return -1;
 913         }
 914
 915         if (evlist__apply_filters(evsel_list, &counter)) {
 916                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 917                         counter->filter, evsel__name(counter), errno,
 918                         str_error_r(errno, msg, sizeof(msg)));
 919                 return -1;
 920         }
 921
 922         if (STAT_RECORD) {
 923                 int fd = perf_data__fd(&perf_stat.data);
 924
 925                 if (is_pipe) {
 926                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 927                 } else {
 928                         err = perf_session__write_header(perf_stat.session, evsel_list,
 929                                                          fd, false);
 930                 }
 931
 932                 if (err < 0)
 933                         return err;
 934
 935                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 936                                                          process_synthesized_event, is_pipe);
 937                 if (err < 0)
 938                         return err;
 939         }
 940
 941         /*
 942          * Enable counters and exec the command:
 943          */
 944         if (forks) {
 945                 evlist__start_workload(evsel_list);
 946                 err = enable_counters();
 947                 if (err)
 948                         return -1;
 949
 950                 t0 = rdclock();
 951                 clock_gettime(CLOCK_MONOTONIC, &ref_time);
 952
 953                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 954                         status = dispatch_events(forks, timeout, interval, &times);
 955                 if (child_pid != -1) {
 956                         if (timeout)
 957                                 kill(child_pid, SIGTERM);
 958                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 959                 }
 960
 961                 if (workload_exec_errno) {
 962                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 963                         pr_err("Workload failed: %s\n", emsg);
 964                         return -1;
 965                 }
 966
 967                 if (WIFSIGNALED(status))
 968                         psignal(WTERMSIG(status), argv[0]);
 969         } else {
 970                 err = enable_counters();
 971                 if (err)
 972                         return -1;
 973
 974                 t0 = rdclock();
 975                 clock_gettime(CLOCK_MONOTONIC, &ref_time);
 976
 977                 status = dispatch_events(forks, timeout, interval, &times);
 978         }
 979
 980         disable_counters();
 981
 982         t1 = rdclock();
 983
 984         if (stat_config.walltime_run_table)
 985                 stat_config.walltime_run[run_idx] = t1 - t0;
 986
 987         if (interval && stat_config.summary) {
 988                 stat_config.interval = 0;
 989                 stat_config.stop_read_counter = true;
 990                 init_stats(&walltime_nsecs_stats);
 991                 update_stats(&walltime_nsecs_stats, t1 - t0);
 992
 993                 if (stat_config.aggr_mode == AGGR_GLOBAL)
 994                         evlist__save_aggr_prev_raw_counts(evsel_list);
 995
 996                 evlist__copy_prev_raw_counts(evsel_list);
 997                 evlist__reset_prev_raw_counts(evsel_list);
 998                 runtime_stat_reset(&stat_config);
 999                 perf_stat__reset_shadow_per_stat(&rt_stat);
1000         } else
1001                 update_stats(&walltime_nsecs_stats, t1 - t0);
1002
1003         /*
1004          * Closing a group leader splits the group, and as we only disable
1005          * group leaders, results in remaining events becoming enabled. To
1006          * avoid arbitrary skew, we must read all counters before closing any
1007          * group leaders.
1008          */
1009         read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1010
1011         /*
1012          * We need to keep evsel_list alive, because it's processed
1013          * later the evsel_list will be closed after.
1014          */
1015         if (!STAT_RECORD)
1016                 evlist__close(evsel_list);
1017
1018         return WEXITSTATUS(status);
1019 }
1020
1021 static int run_perf_stat(int argc, const char **argv, int run_idx)
1022 {
1023         int ret;
1024
1025         if (pre_cmd) {
1026                 ret = system(pre_cmd);
1027                 if (ret)
1028                         return ret;
1029         }
1030
1031         if (sync_run)
1032                 sync();
1033
1034         ret = __run_perf_stat(argc, argv, run_idx);
1035         if (ret)
1036                 return ret;
1037
1038         if (post_cmd) {
1039                 ret = system(post_cmd);
1040                 if (ret)
1041                         return ret;
1042         }
1043
1044         return ret;
1045 }
1046
1047 static void print_counters(struct timespec *ts, int argc, const char **argv)
1048 {
1049         /* Do not print anything if we record to the pipe. */
1050         if (STAT_RECORD && perf_stat.data.is_pipe)
1051                 return;
1052         if (stat_config.quiet)
1053                 return;
1054
1055         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1056 }
1057
1058 static volatile int signr = -1;
1059
1060 static void skip_signal(int signo)
1061 {
1062         if ((child_pid == -1) || stat_config.interval)
1063                 done = 1;
1064
1065         signr = signo;
1066         /*
1067          * render child_pid harmless
1068          * won't send SIGTERM to a random
1069          * process in case of race condition
1070          * and fast PID recycling
1071          */
1072         child_pid = -1;
1073 }
1074
1075 static void sig_atexit(void)
1076 {
1077         sigset_t set, oset;
1078
1079         /*
1080          * avoid race condition with SIGCHLD handler
1081          * in skip_signal() which is modifying child_pid
1082          * goal is to avoid send SIGTERM to a random
1083          * process
1084          */
1085         sigemptyset(&set);
1086         sigaddset(&set, SIGCHLD);
1087         sigprocmask(SIG_BLOCK, &set, &oset);
1088
1089         if (child_pid != -1)
1090                 kill(child_pid, SIGTERM);
1091
1092         sigprocmask(SIG_SETMASK, &oset, NULL);
1093
1094         if (signr == -1)
1095                 return;
1096
1097         signal(signr, SIG_DFL);
1098         kill(getpid(), signr);
1099 }
1100
1101 void perf_stat__set_big_num(int set)
1102 {
1103         stat_config.big_num = (set != 0);
1104 }
1105
1106 void perf_stat__set_no_csv_summary(int set)
1107 {
1108         stat_config.no_csv_summary = (set != 0);
1109 }
1110
1111 static int stat__set_big_num(const struct option *opt __maybe_unused,
1112                              const char *s __maybe_unused, int unset)
1113 {
1114         big_num_opt = unset ? 0 : 1;
1115         perf_stat__set_big_num(!unset);
1116         return 0;
1117 }
1118
1119 static int enable_metric_only(const struct option *opt __maybe_unused,
1120                               const char *s __maybe_unused, int unset)
1121 {
1122         force_metric_only = true;
1123         stat_config.metric_only = !unset;
1124         return 0;
1125 }
1126
1127 static int parse_metric_groups(const struct option *opt,
1128                                const char *str,
1129                                int unset __maybe_unused)
1130 {
1131         return metricgroup__parse_groups(opt, str,
1132                                          stat_config.metric_no_group,
1133                                          stat_config.metric_no_merge,
1134                                          &stat_config.metric_events);
1135 }
1136
1137 static int parse_control_option(const struct option *opt,
1138                                 const char *str,
1139                                 int unset __maybe_unused)
1140 {
1141         struct perf_stat_config *config = opt->value;
1142
1143         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1144 }
1145
1146 static int parse_stat_cgroups(const struct option *opt,
1147                               const char *str, int unset)
1148 {
1149         if (stat_config.cgroup_list) {
1150                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1151                 return -1;
1152         }
1153
1154         return parse_cgroups(opt, str, unset);
1155 }
1156
1157 static struct option stat_options[] = {
1158         OPT_BOOLEAN('T', "transaction", &transaction_run,
1159                     "hardware transaction statistics"),
1160         OPT_CALLBACK('e', "event", &evsel_list, "event",
1161                      "event selector. use 'perf list' to list available events",
1162                      parse_events_option),
1163         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1164                      "event filter", parse_filter),
1165         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1166                     "child tasks do not inherit counters"),
1167         OPT_STRING('p', "pid", &target.pid, "pid",
1168                    "stat events on existing process id"),
1169         OPT_STRING('t', "tid", &target.tid, "tid",
1170                    "stat events on existing thread id"),
1171 #ifdef HAVE_BPF_SKEL
1172         OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1173                    "stat events on existing bpf program id"),
1174         OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1175                     "use bpf program to count events"),
1176         OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1177                    "path to perf_event_attr map"),
1178 #endif
1179         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1180                     "system-wide collection from all CPUs"),
1181         OPT_BOOLEAN('g', "group", &group,
1182                     "put the counters into a counter group"),
1183         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1184                     "Use --no-scale to disable counter scaling for multiplexing"),
1185         OPT_INCR('v', "verbose", &verbose,
1186                     "be more verbose (show counter open errors, etc)"),
1187         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1188                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1189         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1190                     "display details about each run (only with -r option)"),
1191         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1192                     "null run - dont start any counters"),
1193         OPT_INCR('d', "detailed", &detailed_run,
1194                     "detailed run - start a lot of events"),
1195         OPT_BOOLEAN('S', "sync", &sync_run,
1196                     "call sync() before starting a run"),
1197         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1198                            "print large numbers with thousands\' separators",
1199                            stat__set_big_num),
1200         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1201                     "list of cpus to monitor in system-wide"),
1202         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1203                     "disable CPU count aggregation", AGGR_NONE),
1204         OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1205         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1206                    "print counts with custom separator"),
1207         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1208                      "monitor event in cgroup name only", parse_stat_cgroups),
1209         OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1210                     "expand events for each cgroup"),
1211         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1212         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1213         OPT_INTEGER(0, "log-fd", &output_fd,
1214                     "log output to fd, instead of stderr"),
1215         OPT_STRING(0, "pre", &pre_cmd, "command",
1216                         "command to run prior to the measured command"),
1217         OPT_STRING(0, "post", &post_cmd, "command",
1218                         "command to run after to the measured command"),
1219         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1220                     "print counts at regular interval in ms "
1221                     "(overhead is possible for values <= 100ms)"),
1222         OPT_INTEGER(0, "interval-count", &stat_config.times,
1223                     "print counts for fixed number of times"),
1224         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1225                     "clear screen in between new interval"),
1226         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1227                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1228         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1229                      "aggregate counts per processor socket", AGGR_SOCKET),
1230         OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1231                      "aggregate counts per processor die", AGGR_DIE),
1232         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1233                      "aggregate counts per physical processor core", AGGR_CORE),
1234         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1235                      "aggregate counts per thread", AGGR_THREAD),
1236         OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1237                      "aggregate counts per numa node", AGGR_NODE),
1238         OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1239                     "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1240         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1241                         "Only print computed metrics. No raw values", enable_metric_only),
1242         OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1243                        "don't group metric events, impacts multiplexing"),
1244         OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1245                        "don't try to share events between metrics in a group"),
1246         OPT_BOOLEAN(0, "topdown", &topdown_run,
1247                         "measure top-down statistics"),
1248         OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1249                         "Set the metrics level for the top-down statistics (0: max level)"),
1250         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1251                         "measure SMI cost"),
1252         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1253                      "monitor specified metrics or metric groups (separated by ,)",
1254                      parse_metric_groups),
1255         OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1256                          "Configure all used events to run in kernel space.",
1257                          PARSE_OPT_EXCLUSIVE),
1258         OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1259                          "Configure all used events to run in user space.",
1260                          PARSE_OPT_EXCLUSIVE),
1261         OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1262                     "Use with 'percore' event qualifier to show the event "
1263                     "counts of one hardware thread by sum up total hardware "
1264                     "threads of same physical core"),
1265         OPT_BOOLEAN(0, "summary", &stat_config.summary,
1266                        "print summary for interval mode"),
1267         OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1268                        "don't print 'summary' for CSV summary output"),
1269         OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1270                         "don't print output (useful with record)"),
1271 #ifdef HAVE_LIBPFM
1272         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1273                 "libpfm4 event selector. use 'perf list' to list available events",
1274                 parse_libpfm_events_option),
1275 #endif
1276         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1277                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1278                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1279                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1280                       parse_control_option),
1281         OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1282                             "measure I/O performance metrics provided by arch/platform",
1283                             iostat_parse),
1284         OPT_END()
1285 };
1286
1287 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1288                                  struct perf_cpu_map *map, int cpu)
1289 {
1290         return cpu_map__get_socket(map, cpu, NULL);
1291 }
1292
1293 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1294                               struct perf_cpu_map *map, int cpu)
1295 {
1296         return cpu_map__get_die(map, cpu, NULL);
1297 }
1298
1299 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1300                                struct perf_cpu_map *map, int cpu)
1301 {
1302         return cpu_map__get_core(map, cpu, NULL);
1303 }
1304
1305 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1306                                struct perf_cpu_map *map, int cpu)
1307 {
1308         return cpu_map__get_node(map, cpu, NULL);
1309 }
1310
1311 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1312                                aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1313 {
1314         int cpu;
1315         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1316
1317         if (idx >= map->nr)
1318                 return id;
1319
1320         cpu = map->map[idx];
1321
1322         if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1323                 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1324
1325         id = config->cpus_aggr_map->map[cpu];
1326         return id;
1327 }
1328
1329 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1330                                         struct perf_cpu_map *map, int idx)
1331 {
1332         return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1333 }
1334
1335 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1336                                         struct perf_cpu_map *map, int idx)
1337 {
1338         return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1339 }
1340
1341 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1342                                       struct perf_cpu_map *map, int idx)
1343 {
1344         return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1345 }
1346
1347 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1348                                       struct perf_cpu_map *map, int idx)
1349 {
1350         return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1351 }
1352
1353 static bool term_percore_set(void)
1354 {
1355         struct evsel *counter;
1356
1357         evlist__for_each_entry(evsel_list, counter) {
1358                 if (counter->percore)
1359                         return true;
1360         }
1361
1362         return false;
1363 }
1364
1365 static int perf_stat_init_aggr_mode(void)
1366 {
1367         int nr;
1368
1369         switch (stat_config.aggr_mode) {
1370         case AGGR_SOCKET:
1371                 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1372                         perror("cannot build socket map");
1373                         return -1;
1374                 }
1375                 stat_config.aggr_get_id = perf_stat__get_socket_cached;
1376                 break;
1377         case AGGR_DIE:
1378                 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1379                         perror("cannot build die map");
1380                         return -1;
1381                 }
1382                 stat_config.aggr_get_id = perf_stat__get_die_cached;
1383                 break;
1384         case AGGR_CORE:
1385                 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1386                         perror("cannot build core map");
1387                         return -1;
1388                 }
1389                 stat_config.aggr_get_id = perf_stat__get_core_cached;
1390                 break;
1391         case AGGR_NODE:
1392                 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1393                         perror("cannot build core map");
1394                         return -1;
1395                 }
1396                 stat_config.aggr_get_id = perf_stat__get_node_cached;
1397                 break;
1398         case AGGR_NONE:
1399                 if (term_percore_set()) {
1400                         if (cpu_map__build_core_map(evsel_list->core.cpus,
1401                                                     &stat_config.aggr_map)) {
1402                                 perror("cannot build core map");
1403                                 return -1;
1404                         }
1405                         stat_config.aggr_get_id = perf_stat__get_core_cached;
1406                 }
1407                 break;
1408         case AGGR_GLOBAL:
1409         case AGGR_THREAD:
1410         case AGGR_UNSET:
1411         default:
1412                 break;
1413         }
1414
1415         /*
1416          * The evsel_list->cpus is the base we operate on,
1417          * taking the highest cpu number to be the size of
1418          * the aggregation translate cpumap.
1419          */
1420         nr = perf_cpu_map__max(evsel_list->core.cpus);
1421         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1422         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1423 }
1424
1425 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1426 {
1427         if (map) {
1428                 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1429                           "cpu_aggr_map refcnt unbalanced\n");
1430                 free(map);
1431         }
1432 }
1433
1434 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1435 {
1436         if (map && refcount_dec_and_test(&map->refcnt))
1437                 cpu_aggr_map__delete(map);
1438 }
1439
1440 static void perf_stat__exit_aggr_mode(void)
1441 {
1442         cpu_aggr_map__put(stat_config.aggr_map);
1443         cpu_aggr_map__put(stat_config.cpus_aggr_map);
1444         stat_config.aggr_map = NULL;
1445         stat_config.cpus_aggr_map = NULL;
1446 }
1447
1448 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1449 {
1450         int cpu;
1451
1452         if (idx > map->nr)
1453                 return -1;
1454
1455         cpu = map->map[idx];
1456
1457         if (cpu >= env->nr_cpus_avail)
1458                 return -1;
1459
1460         return cpu;
1461 }
1462
1463 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1464 {
1465         struct perf_env *env = data;
1466         int cpu = perf_env__get_cpu(env, map, idx);
1467         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1468
1469         if (cpu != -1)
1470                 id.socket = env->cpu[cpu].socket_id;
1471
1472         return id;
1473 }
1474
1475 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1476 {
1477         struct perf_env *env = data;
1478         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1479         int cpu = perf_env__get_cpu(env, map, idx);
1480
1481         if (cpu != -1) {
1482                 /*
1483                  * die_id is relative to socket, so start
1484                  * with the socket ID and then add die to
1485                  * make a unique ID.
1486                  */
1487                 id.socket = env->cpu[cpu].socket_id;
1488                 id.die = env->cpu[cpu].die_id;
1489         }
1490
1491         return id;
1492 }
1493
1494 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1495 {
1496         struct perf_env *env = data;
1497         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1498         int cpu = perf_env__get_cpu(env, map, idx);
1499
1500         if (cpu != -1) {
1501                 /*
1502                  * core_id is relative to socket and die,
1503                  * we need a global id. So we set
1504                  * socket, die id and core id
1505                  */
1506                 id.socket = env->cpu[cpu].socket_id;
1507                 id.die = env->cpu[cpu].die_id;
1508                 id.core = env->cpu[cpu].core_id;
1509         }
1510
1511         return id;
1512 }
1513
1514 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1515 {
1516         int cpu = perf_env__get_cpu(data, map, idx);
1517         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1518
1519         id.node = perf_env__numa_node(data, cpu);
1520         return id;
1521 }
1522
1523 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1524                                       struct cpu_aggr_map **sockp)
1525 {
1526         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1527 }
1528
1529 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1530                                    struct cpu_aggr_map **diep)
1531 {
1532         return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1533 }
1534
1535 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1536                                     struct cpu_aggr_map **corep)
1537 {
1538         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1539 }
1540
1541 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1542                                     struct cpu_aggr_map **nodep)
1543 {
1544         return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1545 }
1546
1547 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1548                                       struct perf_cpu_map *map, int idx)
1549 {
1550         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1551 }
1552 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1553                                    struct perf_cpu_map *map, int idx)
1554 {
1555         return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1556 }
1557
1558 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1559                                     struct perf_cpu_map *map, int idx)
1560 {
1561         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1562 }
1563
1564 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1565                                     struct perf_cpu_map *map, int idx)
1566 {
1567         return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1568 }
1569
1570 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1571 {
1572         struct perf_env *env = &st->session->header.env;
1573
1574         switch (stat_config.aggr_mode) {
1575         case AGGR_SOCKET:
1576                 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1577                         perror("cannot build socket map");
1578                         return -1;
1579                 }
1580                 stat_config.aggr_get_id = perf_stat__get_socket_file;
1581                 break;
1582         case AGGR_DIE:
1583                 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1584                         perror("cannot build die map");
1585                         return -1;
1586                 }
1587                 stat_config.aggr_get_id = perf_stat__get_die_file;
1588                 break;
1589         case AGGR_CORE:
1590                 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1591                         perror("cannot build core map");
1592                         return -1;
1593                 }
1594                 stat_config.aggr_get_id = perf_stat__get_core_file;
1595                 break;
1596         case AGGR_NODE:
1597                 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1598                         perror("cannot build core map");
1599                         return -1;
1600                 }
1601                 stat_config.aggr_get_id = perf_stat__get_node_file;
1602                 break;
1603         case AGGR_NONE:
1604         case AGGR_GLOBAL:
1605         case AGGR_THREAD:
1606         case AGGR_UNSET:
1607         default:
1608                 break;
1609         }
1610
1611         return 0;
1612 }
1613
1614 /*
1615  * Add default attributes, if there were no attributes specified or
1616  * if -d/--detailed, -d -d or -d -d -d is used:
1617  */
1618 static int add_default_attributes(void)
1619 {
1620         int err;
1621         struct perf_event_attr default_attrs0[] = {
1622
1623   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1624   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1625   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1626   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1627
1628   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1629 };
1630         struct perf_event_attr frontend_attrs[] = {
1631   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1632 };
1633         struct perf_event_attr backend_attrs[] = {
1634   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1635 };
1636         struct perf_event_attr default_attrs1[] = {
1637   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1638   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1639   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1640
1641 };
1642
1643 /*
1644  * Detailed stats (-d), covering the L1 and last level data caches:
1645  */
1646         struct perf_event_attr detailed_attrs[] = {
1647
1648   { .type = PERF_TYPE_HW_CACHE,
1649     .config =
1650          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1651         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1652         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1653
1654   { .type = PERF_TYPE_HW_CACHE,
1655     .config =
1656          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1657         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1658         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1659
1660   { .type = PERF_TYPE_HW_CACHE,
1661     .config =
1662          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1663         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1664         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1665
1666   { .type = PERF_TYPE_HW_CACHE,
1667     .config =
1668          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1669         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1670         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1671 };
1672
1673 /*
1674  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1675  */
1676         struct perf_event_attr very_detailed_attrs[] = {
1677
1678   { .type = PERF_TYPE_HW_CACHE,
1679     .config =
1680          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1681         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1682         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1683
1684   { .type = PERF_TYPE_HW_CACHE,
1685     .config =
1686          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1687         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1688         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1689
1690   { .type = PERF_TYPE_HW_CACHE,
1691     .config =
1692          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1693         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1694         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1695
1696   { .type = PERF_TYPE_HW_CACHE,
1697     .config =
1698          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1699         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1700         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1701
1702   { .type = PERF_TYPE_HW_CACHE,
1703     .config =
1704          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1705         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1706         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1707
1708   { .type = PERF_TYPE_HW_CACHE,
1709     .config =
1710          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1711         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1712         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1713
1714 };
1715
1716 /*
1717  * Very, very detailed stats (-d -d -d), adding prefetch events:
1718  */
1719         struct perf_event_attr very_very_detailed_attrs[] = {
1720
1721   { .type = PERF_TYPE_HW_CACHE,
1722     .config =
1723          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1724         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1725         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1726
1727   { .type = PERF_TYPE_HW_CACHE,
1728     .config =
1729          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1730         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1731         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1732 };
1733         struct parse_events_error errinfo;
1734
1735         /* Set attrs if no event is selected and !null_run: */
1736         if (stat_config.null_run)
1737                 return 0;
1738
1739         bzero(&errinfo, sizeof(errinfo));
1740         if (transaction_run) {
1741                 /* Handle -T as -M transaction. Once platform specific metrics
1742                  * support has been added to the json files, all architectures
1743                  * will use this approach. To determine transaction support
1744                  * on an architecture test for such a metric name.
1745                  */
1746                 if (metricgroup__has_metric("transaction")) {
1747                         struct option opt = { .value = &evsel_list };
1748
1749                         return metricgroup__parse_groups(&opt, "transaction",
1750                                                          stat_config.metric_no_group,
1751                                                         stat_config.metric_no_merge,
1752                                                          &stat_config.metric_events);
1753                 }
1754
1755                 if (pmu_have_event("cpu", "cycles-ct") &&
1756                     pmu_have_event("cpu", "el-start"))
1757                         err = parse_events(evsel_list, transaction_attrs,
1758                                            &errinfo);
1759                 else
1760                         err = parse_events(evsel_list,
1761                                            transaction_limited_attrs,
1762                                            &errinfo);
1763                 if (err) {
1764                         fprintf(stderr, "Cannot set up transaction events\n");
1765                         parse_events_print_error(&errinfo, transaction_attrs);
1766                         return -1;
1767                 }
1768                 return 0;
1769         }
1770
1771         if (smi_cost) {
1772                 int smi;
1773
1774                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1775                         fprintf(stderr, "freeze_on_smi is not supported.\n");
1776                         return -1;
1777                 }
1778
1779                 if (!smi) {
1780                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1781                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1782                                 return -1;
1783                         }
1784                         smi_reset = true;
1785                 }
1786
1787                 if (pmu_have_event("msr", "aperf") &&
1788                     pmu_have_event("msr", "smi")) {
1789                         if (!force_metric_only)
1790                                 stat_config.metric_only = true;
1791                         err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1792                 } else {
1793                         fprintf(stderr, "To measure SMI cost, it needs "
1794                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1795                         parse_events_print_error(&errinfo, smi_cost_attrs);
1796                         return -1;
1797                 }
1798                 if (err) {
1799                         parse_events_print_error(&errinfo, smi_cost_attrs);
1800                         fprintf(stderr, "Cannot set up SMI cost events\n");
1801                         return -1;
1802                 }
1803                 return 0;
1804         }
1805
1806         if (topdown_run) {
1807                 const char **metric_attrs = topdown_metric_attrs;
1808                 unsigned int max_level = 1;
1809                 char *str = NULL;
1810                 bool warn = false;
1811
1812                 if (!force_metric_only)
1813                         stat_config.metric_only = true;
1814
1815                 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1816                         metric_attrs = topdown_metric_L2_attrs;
1817                         max_level = 2;
1818                 }
1819
1820                 if (stat_config.topdown_level > max_level) {
1821                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1822                         return -1;
1823                 } else if (!stat_config.topdown_level)
1824                         stat_config.topdown_level = max_level;
1825
1826                 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1827                         pr_err("Out of memory\n");
1828                         return -1;
1829                 }
1830                 if (metric_attrs[0] && str) {
1831                         if (!stat_config.interval && !stat_config.metric_only) {
1832                                 fprintf(stat_config.output,
1833                                         "Topdown accuracy may decrease when measuring long periods.\n"
1834                                         "Please print the result regularly, e.g. -I1000\n");
1835                         }
1836                         goto setup_metrics;
1837                 }
1838
1839                 zfree(&str);
1840
1841                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1842                     stat_config.aggr_mode != AGGR_CORE) {
1843                         pr_err("top down event configuration requires --per-core mode\n");
1844                         return -1;
1845                 }
1846                 stat_config.aggr_mode = AGGR_CORE;
1847                 if (nr_cgroups || !target__has_cpu(&target)) {
1848                         pr_err("top down event configuration requires system-wide mode (-a)\n");
1849                         return -1;
1850                 }
1851
1852                 if (topdown_filter_events(topdown_attrs, &str,
1853                                 arch_topdown_check_group(&warn)) < 0) {
1854                         pr_err("Out of memory\n");
1855                         return -1;
1856                 }
1857                 if (topdown_attrs[0] && str) {
1858                         if (warn)
1859                                 arch_topdown_group_warn();
1860 setup_metrics:
1861                         err = parse_events(evsel_list, str, &errinfo);
1862                         if (err) {
1863                                 fprintf(stderr,
1864                                         "Cannot set up top down events %s: %d\n",
1865                                         str, err);
1866                                 parse_events_print_error(&errinfo, str);
1867                                 free(str);
1868                                 return -1;
1869                         }
1870                 } else {
1871                         fprintf(stderr, "System does not support topdown\n");
1872                         return -1;
1873                 }
1874                 free(str);
1875         }
1876
1877         if (!evsel_list->core.nr_entries) {
1878                 if (target__has_cpu(&target))
1879                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1880
1881                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1882                         return -1;
1883                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1884                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1885                                 return -1;
1886                 }
1887                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1888                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1889                                 return -1;
1890                 }
1891                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1892                         return -1;
1893
1894                 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1895                         return -1;
1896         }
1897
1898         /* Detailed events get appended to the event list: */
1899
1900         if (detailed_run <  1)
1901                 return 0;
1902
1903         /* Append detailed run extra attributes: */
1904         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1905                 return -1;
1906
1907         if (detailed_run < 2)
1908                 return 0;
1909
1910         /* Append very detailed run extra attributes: */
1911         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1912                 return -1;
1913
1914         if (detailed_run < 3)
1915                 return 0;
1916
1917         /* Append very, very detailed run extra attributes: */
1918         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1919 }
1920
1921 static const char * const stat_record_usage[] = {
1922         "perf stat record [<options>]",
1923         NULL,
1924 };
1925
1926 static void init_features(struct perf_session *session)
1927 {
1928         int feat;
1929
1930         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1931                 perf_header__set_feat(&session->header, feat);
1932
1933         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1934         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1935         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1936         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1937         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1938 }
1939
1940 static int __cmd_record(int argc, const char **argv)
1941 {
1942         struct perf_session *session;
1943         struct perf_data *data = &perf_stat.data;
1944
1945         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1946                              PARSE_OPT_STOP_AT_NON_OPTION);
1947
1948         if (output_name)
1949                 data->path = output_name;
1950
1951         if (stat_config.run_count != 1 || forever) {
1952                 pr_err("Cannot use -r option with perf stat record.\n");
1953                 return -1;
1954         }
1955
1956         session = perf_session__new(data, false, NULL);
1957         if (IS_ERR(session)) {
1958                 pr_err("Perf session creation failed\n");
1959                 return PTR_ERR(session);
1960         }
1961
1962         init_features(session);
1963
1964         session->evlist   = evsel_list;
1965         perf_stat.session = session;
1966         perf_stat.record  = true;
1967         return argc;
1968 }
1969
1970 static int process_stat_round_event(struct perf_session *session,
1971                                     union perf_event *event)
1972 {
1973         struct perf_record_stat_round *stat_round = &event->stat_round;
1974         struct evsel *counter;
1975         struct timespec tsh, *ts = NULL;
1976         const char **argv = session->header.env.cmdline_argv;
1977         int argc = session->header.env.nr_cmdline;
1978
1979         evlist__for_each_entry(evsel_list, counter)
1980                 perf_stat_process_counter(&stat_config, counter);
1981
1982         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1983                 update_stats(&walltime_nsecs_stats, stat_round->time);
1984
1985         if (stat_config.interval && stat_round->time) {
1986                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1987                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1988                 ts = &tsh;
1989         }
1990
1991         print_counters(ts, argc, argv);
1992         return 0;
1993 }
1994
1995 static
1996 int process_stat_config_event(struct perf_session *session,
1997                               union perf_event *event)
1998 {
1999         struct perf_tool *tool = session->tool;
2000         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2001
2002         perf_event__read_stat_config(&stat_config, &event->stat_config);
2003
2004         if (perf_cpu_map__empty(st->cpus)) {
2005                 if (st->aggr_mode != AGGR_UNSET)
2006                         pr_warning("warning: processing task data, aggregation mode not set\n");
2007                 return 0;
2008         }
2009
2010         if (st->aggr_mode != AGGR_UNSET)
2011                 stat_config.aggr_mode = st->aggr_mode;
2012
2013         if (perf_stat.data.is_pipe)
2014                 perf_stat_init_aggr_mode();
2015         else
2016                 perf_stat_init_aggr_mode_file(st);
2017
2018         return 0;
2019 }
2020
2021 static int set_maps(struct perf_stat *st)
2022 {
2023         if (!st->cpus || !st->threads)
2024                 return 0;
2025
2026         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2027                 return -EINVAL;
2028
2029         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2030
2031         if (evlist__alloc_stats(evsel_list, true))
2032                 return -ENOMEM;
2033
2034         st->maps_allocated = true;
2035         return 0;
2036 }
2037
2038 static
2039 int process_thread_map_event(struct perf_session *session,
2040                              union perf_event *event)
2041 {
2042         struct perf_tool *tool = session->tool;
2043         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2044
2045         if (st->threads) {
2046                 pr_warning("Extra thread map event, ignoring.\n");
2047                 return 0;
2048         }
2049
2050         st->threads = thread_map__new_event(&event->thread_map);
2051         if (!st->threads)
2052                 return -ENOMEM;
2053
2054         return set_maps(st);
2055 }
2056
2057 static
2058 int process_cpu_map_event(struct perf_session *session,
2059                           union perf_event *event)
2060 {
2061         struct perf_tool *tool = session->tool;
2062         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2063         struct perf_cpu_map *cpus;
2064
2065         if (st->cpus) {
2066                 pr_warning("Extra cpu map event, ignoring.\n");
2067                 return 0;
2068         }
2069
2070         cpus = cpu_map__new_data(&event->cpu_map.data);
2071         if (!cpus)
2072                 return -ENOMEM;
2073
2074         st->cpus = cpus;
2075         return set_maps(st);
2076 }
2077
2078 static const char * const stat_report_usage[] = {
2079         "perf stat report [<options>]",
2080         NULL,
2081 };
2082
2083 static struct perf_stat perf_stat = {
2084         .tool = {
2085                 .attr           = perf_event__process_attr,
2086                 .event_update   = perf_event__process_event_update,
2087                 .thread_map     = process_thread_map_event,
2088                 .cpu_map        = process_cpu_map_event,
2089                 .stat_config    = process_stat_config_event,
2090                 .stat           = perf_event__process_stat_event,
2091                 .stat_round     = process_stat_round_event,
2092         },
2093         .aggr_mode = AGGR_UNSET,
2094 };
2095
2096 static int __cmd_report(int argc, const char **argv)
2097 {
2098         struct perf_session *session;
2099         const struct option options[] = {
2100         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2101         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2102                      "aggregate counts per processor socket", AGGR_SOCKET),
2103         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2104                      "aggregate counts per processor die", AGGR_DIE),
2105         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2106                      "aggregate counts per physical processor core", AGGR_CORE),
2107         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2108                      "aggregate counts per numa node", AGGR_NODE),
2109         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2110                      "disable CPU count aggregation", AGGR_NONE),
2111         OPT_END()
2112         };
2113         struct stat st;
2114         int ret;
2115
2116         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2117
2118         if (!input_name || !strlen(input_name)) {
2119                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2120                         input_name = "-";
2121                 else
2122                         input_name = "perf.data";
2123         }
2124
2125         perf_stat.data.path = input_name;
2126         perf_stat.data.mode = PERF_DATA_MODE_READ;
2127
2128         session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2129         if (IS_ERR(session))
2130                 return PTR_ERR(session);
2131
2132         perf_stat.session  = session;
2133         stat_config.output = stderr;
2134         evsel_list         = session->evlist;
2135
2136         ret = perf_session__process_events(session);
2137         if (ret)
2138                 return ret;
2139
2140         perf_session__delete(session);
2141         return 0;
2142 }
2143
2144 static void setup_system_wide(int forks)
2145 {
2146         /*
2147          * Make system wide (-a) the default target if
2148          * no target was specified and one of following
2149          * conditions is met:
2150          *
2151          *   - there's no workload specified
2152          *   - there is workload specified but all requested
2153          *     events are system wide events
2154          */
2155         if (!target__none(&target))
2156                 return;
2157
2158         if (!forks)
2159                 target.system_wide = true;
2160         else {
2161                 struct evsel *counter;
2162
2163                 evlist__for_each_entry(evsel_list, counter) {
2164                         if (!counter->core.system_wide &&
2165                             strcmp(counter->name, "duration_time")) {
2166                                 return;
2167                         }
2168                 }
2169
2170                 if (evsel_list->core.nr_entries)
2171                         target.system_wide = true;
2172         }
2173 }
2174
2175 int cmd_stat(int argc, const char **argv)
2176 {
2177         const char * const stat_usage[] = {
2178                 "perf stat [<options>] [<command>]",
2179                 NULL
2180         };
2181         int status = -EINVAL, run_idx, err;
2182         const char *mode;
2183         FILE *output = stderr;
2184         unsigned int interval, timeout;
2185         const char * const stat_subcommands[] = { "record", "report" };
2186         char errbuf[BUFSIZ];
2187
2188         setlocale(LC_ALL, "");
2189
2190         evsel_list = evlist__new();
2191         if (evsel_list == NULL)
2192                 return -ENOMEM;
2193
2194         parse_events__shrink_config_terms();
2195
2196         /* String-parsing callback-based options would segfault when negated */
2197         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2198         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2199         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2200
2201         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2202                                         (const char **) stat_usage,
2203                                         PARSE_OPT_STOP_AT_NON_OPTION);
2204         perf_stat__collect_metric_expr(evsel_list);
2205         perf_stat__init_shadow_stats();
2206
2207         if (stat_config.csv_sep) {
2208                 stat_config.csv_output = true;
2209                 if (!strcmp(stat_config.csv_sep, "\\t"))
2210                         stat_config.csv_sep = "\t";
2211         } else
2212                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2213
2214         if (argc && !strncmp(argv[0], "rec", 3)) {
2215                 argc = __cmd_record(argc, argv);
2216                 if (argc < 0)
2217                         return -1;
2218         } else if (argc && !strncmp(argv[0], "rep", 3))
2219                 return __cmd_report(argc, argv);
2220
2221         interval = stat_config.interval;
2222         timeout = stat_config.timeout;
2223
2224         /*
2225          * For record command the -o is already taken care of.
2226          */
2227         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2228                 output = NULL;
2229
2230         if (output_name && output_fd) {
2231                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2232                 parse_options_usage(stat_usage, stat_options, "o", 1);
2233                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2234                 goto out;
2235         }
2236
2237         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2238                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2239                 goto out;
2240         }
2241
2242         if (stat_config.metric_only && stat_config.run_count > 1) {
2243                 fprintf(stderr, "--metric-only is not supported with -r\n");
2244                 goto out;
2245         }
2246
2247         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2248                 fprintf(stderr, "--table is only supported with -r\n");
2249                 parse_options_usage(stat_usage, stat_options, "r", 1);
2250                 parse_options_usage(NULL, stat_options, "table", 0);
2251                 goto out;
2252         }
2253
2254         if (output_fd < 0) {
2255                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2256                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2257                 goto out;
2258         }
2259
2260         if (!output && !stat_config.quiet) {
2261                 struct timespec tm;
2262                 mode = append_file ? "a" : "w";
2263
2264                 output = fopen(output_name, mode);
2265                 if (!output) {
2266                         perror("failed to create output file");
2267                         return -1;
2268                 }
2269                 clock_gettime(CLOCK_REALTIME, &tm);
2270                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2271         } else if (output_fd > 0) {
2272                 mode = append_file ? "a" : "w";
2273                 output = fdopen(output_fd, mode);
2274                 if (!output) {
2275                         perror("Failed opening logfd");
2276                         return -errno;
2277                 }
2278         }
2279
2280         stat_config.output = output;
2281
2282         /*
2283          * let the spreadsheet do the pretty-printing
2284          */
2285         if (stat_config.csv_output) {
2286                 /* User explicitly passed -B? */
2287                 if (big_num_opt == 1) {
2288                         fprintf(stderr, "-B option not supported with -x\n");
2289                         parse_options_usage(stat_usage, stat_options, "B", 1);
2290                         parse_options_usage(NULL, stat_options, "x", 1);
2291                         goto out;
2292                 } else /* Nope, so disable big number formatting */
2293                         stat_config.big_num = false;
2294         } else if (big_num_opt == 0) /* User passed --no-big-num */
2295                 stat_config.big_num = false;
2296
2297         err = target__validate(&target);
2298         if (err) {
2299                 target__strerror(&target, err, errbuf, BUFSIZ);
2300                 pr_warning("%s\n", errbuf);
2301         }
2302
2303         setup_system_wide(argc);
2304
2305         /*
2306          * Display user/system times only for single
2307          * run and when there's specified tracee.
2308          */
2309         if ((stat_config.run_count == 1) && target__none(&target))
2310                 stat_config.ru_display = true;
2311
2312         if (stat_config.run_count < 0) {
2313                 pr_err("Run count must be a positive number\n");
2314                 parse_options_usage(stat_usage, stat_options, "r", 1);
2315                 goto out;
2316         } else if (stat_config.run_count == 0) {
2317                 forever = true;
2318                 stat_config.run_count = 1;
2319         }
2320
2321         if (stat_config.walltime_run_table) {
2322                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2323                 if (!stat_config.walltime_run) {
2324                         pr_err("failed to setup -r option");
2325                         goto out;
2326                 }
2327         }
2328
2329         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2330                 !target__has_task(&target)) {
2331                 if (!target.system_wide || target.cpu_list) {
2332                         fprintf(stderr, "The --per-thread option is only "
2333                                 "available when monitoring via -p -t -a "
2334                                 "options or only --per-thread.\n");
2335                         parse_options_usage(NULL, stat_options, "p", 1);
2336                         parse_options_usage(NULL, stat_options, "t", 1);
2337                         goto out;
2338                 }
2339         }
2340
2341         /*
2342          * no_aggr, cgroup are for system-wide only
2343          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2344          */
2345         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2346               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2347             !target__has_cpu(&target)) {
2348                 fprintf(stderr, "both cgroup and no-aggregation "
2349                         "modes only available in system-wide mode\n");
2350
2351                 parse_options_usage(stat_usage, stat_options, "G", 1);
2352                 parse_options_usage(NULL, stat_options, "A", 1);
2353                 parse_options_usage(NULL, stat_options, "a", 1);
2354                 goto out;
2355         }
2356
2357         if (stat_config.iostat_run) {
2358                 status = iostat_prepare(evsel_list, &stat_config);
2359                 if (status)
2360                         goto out;
2361                 if (iostat_mode == IOSTAT_LIST) {
2362                         iostat_list(evsel_list, &stat_config);
2363                         goto out;
2364                 } else if (verbose)
2365                         iostat_list(evsel_list, &stat_config);
2366         }
2367
2368         if (add_default_attributes())
2369                 goto out;
2370
2371         if (stat_config.cgroup_list) {
2372                 if (nr_cgroups > 0) {
2373                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2374                         parse_options_usage(stat_usage, stat_options, "G", 1);
2375                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2376                         goto out;
2377                 }
2378
2379                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2380                                           &stat_config.metric_events, true) < 0) {
2381                         parse_options_usage(stat_usage, stat_options,
2382                                             "for-each-cgroup", 0);
2383                         goto out;
2384                 }
2385         }
2386
2387         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2388                 target.per_thread = true;
2389
2390         if (evlist__create_maps(evsel_list, &target) < 0) {
2391                 if (target__has_task(&target)) {
2392                         pr_err("Problems finding threads of monitor\n");
2393                         parse_options_usage(stat_usage, stat_options, "p", 1);
2394                         parse_options_usage(NULL, stat_options, "t", 1);
2395                 } else if (target__has_cpu(&target)) {
2396                         perror("failed to parse CPUs map");
2397                         parse_options_usage(stat_usage, stat_options, "C", 1);
2398                         parse_options_usage(NULL, stat_options, "a", 1);
2399                 }
2400                 goto out;
2401         }
2402
2403         evlist__check_cpu_maps(evsel_list);
2404
2405         /*
2406          * Initialize thread_map with comm names,
2407          * so we could print it out on output.
2408          */
2409         if (stat_config.aggr_mode == AGGR_THREAD) {
2410                 thread_map__read_comms(evsel_list->core.threads);
2411                 if (target.system_wide) {
2412                         if (runtime_stat_new(&stat_config,
2413                                 perf_thread_map__nr(evsel_list->core.threads))) {
2414                                 goto out;
2415                         }
2416                 }
2417         }
2418
2419         if (stat_config.aggr_mode == AGGR_NODE)
2420                 cpu__setup_cpunode_map();
2421
2422         if (stat_config.times && interval)
2423                 interval_count = true;
2424         else if (stat_config.times && !interval) {
2425                 pr_err("interval-count option should be used together with "
2426                                 "interval-print.\n");
2427                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2428                 parse_options_usage(stat_usage, stat_options, "I", 1);
2429                 goto out;
2430         }
2431
2432         if (timeout && timeout < 100) {
2433                 if (timeout < 10) {
2434                         pr_err("timeout must be >= 10ms.\n");
2435                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2436                         goto out;
2437                 } else
2438                         pr_warning("timeout < 100ms. "
2439                                    "The overhead percentage could be high in some cases. "
2440                                    "Please proceed with caution.\n");
2441         }
2442         if (timeout && interval) {
2443                 pr_err("timeout option is not supported with interval-print.\n");
2444                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2445                 parse_options_usage(stat_usage, stat_options, "I", 1);
2446                 goto out;
2447         }
2448
2449         if (evlist__alloc_stats(evsel_list, interval))
2450                 goto out;
2451
2452         if (perf_stat_init_aggr_mode())
2453                 goto out;
2454
2455         /*
2456          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2457          * while avoiding that older tools show confusing messages.
2458          *
2459          * However for pipe sessions we need to keep it zero,
2460          * because script's perf_evsel__check_attr is triggered
2461          * by attr->sample_type != 0, and we can't run it on
2462          * stat sessions.
2463          */
2464         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2465
2466         /*
2467          * We dont want to block the signals - that would cause
2468          * child tasks to inherit that and Ctrl-C would not work.
2469          * What we want is for Ctrl-C to work in the exec()-ed
2470          * task, but being ignored by perf stat itself:
2471          */
2472         atexit(sig_atexit);
2473         if (!forever)
2474                 signal(SIGINT,  skip_signal);
2475         signal(SIGCHLD, skip_signal);
2476         signal(SIGALRM, skip_signal);
2477         signal(SIGABRT, skip_signal);
2478
2479         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2480                 goto out;
2481
2482         status = 0;
2483         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2484                 if (stat_config.run_count != 1 && verbose > 0)
2485                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2486                                 run_idx + 1);
2487
2488                 if (run_idx != 0)
2489                         evlist__reset_prev_raw_counts(evsel_list);
2490
2491                 status = run_perf_stat(argc, argv, run_idx);
2492                 if (forever && status != -1 && !interval) {
2493                         print_counters(NULL, argc, argv);
2494                         perf_stat__reset_stats();
2495                 }
2496         }
2497
2498         if (!forever && status != -1 && (!interval || stat_config.summary))
2499                 print_counters(NULL, argc, argv);
2500
2501         evlist__finalize_ctlfd(evsel_list);
2502
2503         if (STAT_RECORD) {
2504                 /*
2505                  * We synthesize the kernel mmap record just so that older tools
2506                  * don't emit warnings about not being able to resolve symbols
2507                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2508                  * a saner message about no samples being in the perf.data file.
2509                  *
2510                  * This also serves to suppress a warning about f_header.data.size == 0
2511                  * in header.c at the moment 'perf stat record' gets introduced, which
2512                  * is not really needed once we start adding the stat specific PERF_RECORD_
2513                  * records, but the need to suppress the kptr_restrict messages in older
2514                  * tools remain  -acme
2515                  */
2516                 int fd = perf_data__fd(&perf_stat.data);
2517
2518                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2519                                                          process_synthesized_event,
2520                                                          &perf_stat.session->machines.host);
2521                 if (err) {
2522                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2523                                    "older tools may produce warnings about this file\n.");
2524                 }
2525
2526                 if (!interval) {
2527                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2528                                 pr_err("failed to write stat round event\n");
2529                 }
2530
2531                 if (!perf_stat.data.is_pipe) {
2532                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2533                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2534                 }
2535
2536                 evlist__close(evsel_list);
2537                 perf_session__delete(perf_stat.session);
2538         }
2539
2540         perf_stat__exit_aggr_mode();
2541         evlist__free_stats(evsel_list);
2542 out:
2543         if (stat_config.iostat_run)
2544                 iostat_release(evsel_list);
2545
2546         zfree(&stat_config.walltime_run);
2547
2548         if (smi_cost && smi_reset)
2549                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2550
2551         evlist__delete(evsel_list);
2552
2553         metricgroup__rblist_exit(&stat_config.metric_events);
2554         runtime_stat_delete(&stat_config);
2555         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2556
2557         return status;
2558 }