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