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                                         perf_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 (perf_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         perf_evlist__print_counters(evsel_list, &stat_config, &target,
 978                                     ts, argc, argv);
 979 }
 980
 981 static volatile int signr = -1;
 982
 983 static void skip_signal(int signo)
 984 {
 985         if ((child_pid == -1) || stat_config.interval)
 986                 done = 1;
 987
 988         signr = signo;
 989         /*
 990          * render child_pid harmless
 991          * won't send SIGTERM to a random
 992          * process in case of race condition
 993          * and fast PID recycling
 994          */
 995         child_pid = -1;
 996 }
 997
 998 static void sig_atexit(void)
 999 {
1000         sigset_t set, oset;
1001
1002         /*
1003          * avoid race condition with SIGCHLD handler
1004          * in skip_signal() which is modifying child_pid
1005          * goal is to avoid send SIGTERM to a random
1006          * process
1007          */
1008         sigemptyset(&set);
1009         sigaddset(&set, SIGCHLD);
1010         sigprocmask(SIG_BLOCK, &set, &oset);
1011
1012         if (child_pid != -1)
1013                 kill(child_pid, SIGTERM);
1014
1015         sigprocmask(SIG_SETMASK, &oset, NULL);
1016
1017         if (signr == -1)
1018                 return;
1019
1020         signal(signr, SIG_DFL);
1021         kill(getpid(), signr);
1022 }
1023
1024 void perf_stat__set_big_num(int set)
1025 {
1026         stat_config.big_num = (set != 0);
1027 }
1028
1029 static int stat__set_big_num(const struct option *opt __maybe_unused,
1030                              const char *s __maybe_unused, int unset)
1031 {
1032         big_num_opt = unset ? 0 : 1;
1033         perf_stat__set_big_num(!unset);
1034         return 0;
1035 }
1036
1037 static int enable_metric_only(const struct option *opt __maybe_unused,
1038                               const char *s __maybe_unused, int unset)
1039 {
1040         force_metric_only = true;
1041         stat_config.metric_only = !unset;
1042         return 0;
1043 }
1044
1045 static int parse_metric_groups(const struct option *opt,
1046                                const char *str,
1047                                int unset __maybe_unused)
1048 {
1049         return metricgroup__parse_groups(opt, str,
1050                                          stat_config.metric_no_group,
1051                                          stat_config.metric_no_merge,
1052                                          &stat_config.metric_events);
1053 }
1054
1055 static int parse_control_option(const struct option *opt,
1056                                 const char *str,
1057                                 int unset __maybe_unused)
1058 {
1059         struct perf_stat_config *config = opt->value;
1060
1061         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1062 }
1063
1064 static int parse_stat_cgroups(const struct option *opt,
1065                               const char *str, int unset)
1066 {
1067         if (stat_config.cgroup_list) {
1068                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1069                 return -1;
1070         }
1071
1072         return parse_cgroups(opt, str, unset);
1073 }
1074
1075 static struct option stat_options[] = {
1076         OPT_BOOLEAN('T', "transaction", &transaction_run,
1077                     "hardware transaction statistics"),
1078         OPT_CALLBACK('e', "event", &evsel_list, "event",
1079                      "event selector. use 'perf list' to list available events",
1080                      parse_events_option),
1081         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1082                      "event filter", parse_filter),
1083         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1084                     "child tasks do not inherit counters"),
1085         OPT_STRING('p', "pid", &target.pid, "pid",
1086                    "stat events on existing process id"),
1087         OPT_STRING('t', "tid", &target.tid, "tid",
1088                    "stat events on existing thread id"),
1089         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1090                     "system-wide collection from all CPUs"),
1091         OPT_BOOLEAN('g', "group", &group,
1092                     "put the counters into a counter group"),
1093         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1094                     "Use --no-scale to disable counter scaling for multiplexing"),
1095         OPT_INCR('v', "verbose", &verbose,
1096                     "be more verbose (show counter open errors, etc)"),
1097         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1098                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1099         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1100                     "display details about each run (only with -r option)"),
1101         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1102                     "null run - dont start any counters"),
1103         OPT_INCR('d', "detailed", &detailed_run,
1104                     "detailed run - start a lot of events"),
1105         OPT_BOOLEAN('S', "sync", &sync_run,
1106                     "call sync() before starting a run"),
1107         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1108                            "print large numbers with thousands\' separators",
1109                            stat__set_big_num),
1110         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1111                     "list of cpus to monitor in system-wide"),
1112         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1113                     "disable CPU count aggregation", AGGR_NONE),
1114         OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1115         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1116                    "print counts with custom separator"),
1117         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1118                      "monitor event in cgroup name only", parse_stat_cgroups),
1119         OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1120                     "expand events for each cgroup"),
1121         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1122         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1123         OPT_INTEGER(0, "log-fd", &output_fd,
1124                     "log output to fd, instead of stderr"),
1125         OPT_STRING(0, "pre", &pre_cmd, "command",
1126                         "command to run prior to the measured command"),
1127         OPT_STRING(0, "post", &post_cmd, "command",
1128                         "command to run after to the measured command"),
1129         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1130                     "print counts at regular interval in ms "
1131                     "(overhead is possible for values <= 100ms)"),
1132         OPT_INTEGER(0, "interval-count", &stat_config.times,
1133                     "print counts for fixed number of times"),
1134         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1135                     "clear screen in between new interval"),
1136         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1137                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1138         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1139                      "aggregate counts per processor socket", AGGR_SOCKET),
1140         OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1141                      "aggregate counts per processor die", AGGR_DIE),
1142         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1143                      "aggregate counts per physical processor core", AGGR_CORE),
1144         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1145                      "aggregate counts per thread", AGGR_THREAD),
1146         OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1147                      "aggregate counts per numa node", AGGR_NODE),
1148         OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1149                     "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1150         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1151                         "Only print computed metrics. No raw values", enable_metric_only),
1152         OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1153                        "don't group metric events, impacts multiplexing"),
1154         OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1155                        "don't try to share events between metrics in a group"),
1156         OPT_BOOLEAN(0, "topdown", &topdown_run,
1157                         "measure topdown level 1 statistics"),
1158         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1159                         "measure SMI cost"),
1160         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1161                      "monitor specified metrics or metric groups (separated by ,)",
1162                      parse_metric_groups),
1163         OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1164                          "Configure all used events to run in kernel space.",
1165                          PARSE_OPT_EXCLUSIVE),
1166         OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1167                          "Configure all used events to run in user space.",
1168                          PARSE_OPT_EXCLUSIVE),
1169         OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1170                     "Use with 'percore' event qualifier to show the event "
1171                     "counts of one hardware thread by sum up total hardware "
1172                     "threads of same physical core"),
1173         OPT_BOOLEAN(0, "summary", &stat_config.summary,
1174                        "print summary for interval mode"),
1175         OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1176                         "don't print output (useful with record)"),
1177 #ifdef HAVE_LIBPFM
1178         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1179                 "libpfm4 event selector. use 'perf list' to list available events",
1180                 parse_libpfm_events_option),
1181 #endif
1182         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1183                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1184                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1185                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1186                       parse_control_option),
1187         OPT_END()
1188 };
1189
1190 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1191                                  struct perf_cpu_map *map, int cpu)
1192 {
1193         return cpu_map__get_socket(map, cpu, NULL);
1194 }
1195
1196 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1197                               struct perf_cpu_map *map, int cpu)
1198 {
1199         return cpu_map__get_die(map, cpu, NULL);
1200 }
1201
1202 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1203                                struct perf_cpu_map *map, int cpu)
1204 {
1205         return cpu_map__get_core(map, cpu, NULL);
1206 }
1207
1208 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1209                                struct perf_cpu_map *map, int cpu)
1210 {
1211         return cpu_map__get_node(map, cpu, NULL);
1212 }
1213
1214 static int perf_stat__get_aggr(struct perf_stat_config *config,
1215                                aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1216 {
1217         int cpu;
1218
1219         if (idx >= map->nr)
1220                 return -1;
1221
1222         cpu = map->map[idx];
1223
1224         if (config->cpus_aggr_map->map[cpu] == -1)
1225                 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1226
1227         return config->cpus_aggr_map->map[cpu];
1228 }
1229
1230 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
1231                                         struct perf_cpu_map *map, int idx)
1232 {
1233         return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1234 }
1235
1236 static int perf_stat__get_die_cached(struct perf_stat_config *config,
1237                                         struct perf_cpu_map *map, int idx)
1238 {
1239         return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1240 }
1241
1242 static int perf_stat__get_core_cached(struct perf_stat_config *config,
1243                                       struct perf_cpu_map *map, int idx)
1244 {
1245         return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1246 }
1247
1248 static int perf_stat__get_node_cached(struct perf_stat_config *config,
1249                                       struct perf_cpu_map *map, int idx)
1250 {
1251         return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1252 }
1253
1254 static bool term_percore_set(void)
1255 {
1256         struct evsel *counter;
1257
1258         evlist__for_each_entry(evsel_list, counter) {
1259                 if (counter->percore)
1260                         return true;
1261         }
1262
1263         return false;
1264 }
1265
1266 static int perf_stat_init_aggr_mode(void)
1267 {
1268         int nr;
1269
1270         switch (stat_config.aggr_mode) {
1271         case AGGR_SOCKET:
1272                 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1273                         perror("cannot build socket map");
1274                         return -1;
1275                 }
1276                 stat_config.aggr_get_id = perf_stat__get_socket_cached;
1277                 break;
1278         case AGGR_DIE:
1279                 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1280                         perror("cannot build die map");
1281                         return -1;
1282                 }
1283                 stat_config.aggr_get_id = perf_stat__get_die_cached;
1284                 break;
1285         case AGGR_CORE:
1286                 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1287                         perror("cannot build core map");
1288                         return -1;
1289                 }
1290                 stat_config.aggr_get_id = perf_stat__get_core_cached;
1291                 break;
1292         case AGGR_NODE:
1293                 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1294                         perror("cannot build core map");
1295                         return -1;
1296                 }
1297                 stat_config.aggr_get_id = perf_stat__get_node_cached;
1298                 break;
1299         case AGGR_NONE:
1300                 if (term_percore_set()) {
1301                         if (cpu_map__build_core_map(evsel_list->core.cpus,
1302                                                     &stat_config.aggr_map)) {
1303                                 perror("cannot build core map");
1304                                 return -1;
1305                         }
1306                         stat_config.aggr_get_id = perf_stat__get_core_cached;
1307                 }
1308                 break;
1309         case AGGR_GLOBAL:
1310         case AGGR_THREAD:
1311         case AGGR_UNSET:
1312         default:
1313                 break;
1314         }
1315
1316         /*
1317          * The evsel_list->cpus is the base we operate on,
1318          * taking the highest cpu number to be the size of
1319          * the aggregation translate cpumap.
1320          */
1321         nr = perf_cpu_map__max(evsel_list->core.cpus);
1322         stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
1323         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1324 }
1325
1326 static void perf_stat__exit_aggr_mode(void)
1327 {
1328         perf_cpu_map__put(stat_config.aggr_map);
1329         perf_cpu_map__put(stat_config.cpus_aggr_map);
1330         stat_config.aggr_map = NULL;
1331         stat_config.cpus_aggr_map = NULL;
1332 }
1333
1334 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1335 {
1336         int cpu;
1337
1338         if (idx > map->nr)
1339                 return -1;
1340
1341         cpu = map->map[idx];
1342
1343         if (cpu >= env->nr_cpus_avail)
1344                 return -1;
1345
1346         return cpu;
1347 }
1348
1349 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1350 {
1351         struct perf_env *env = data;
1352         int cpu = perf_env__get_cpu(env, map, idx);
1353
1354         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1355 }
1356
1357 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1358 {
1359         struct perf_env *env = data;
1360         int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1361
1362         if (cpu != -1) {
1363                 /*
1364                  * Encode socket in bit range 15:8
1365                  * die_id is relative to socket,
1366                  * we need a global id. So we combine
1367                  * socket + die id
1368                  */
1369                 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1370                         return -1;
1371
1372                 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1373                         return -1;
1374
1375                 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1376         }
1377
1378         return die_id;
1379 }
1380
1381 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1382 {
1383         struct perf_env *env = data;
1384         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1385
1386         if (cpu != -1) {
1387                 /*
1388                  * Encode socket in bit range 31:24
1389                  * encode die id in bit range 23:16
1390                  * core_id is relative to socket and die,
1391                  * we need a global id. So we combine
1392                  * socket + die id + core id
1393                  */
1394                 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1395                         return -1;
1396
1397                 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1398                         return -1;
1399
1400                 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1401                         return -1;
1402
1403                 core = (env->cpu[cpu].socket_id << 24) |
1404                        (env->cpu[cpu].die_id << 16) |
1405                        (env->cpu[cpu].core_id & 0xffff);
1406         }
1407
1408         return core;
1409 }
1410
1411 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1412 {
1413         int cpu = perf_env__get_cpu(data, map, idx);
1414
1415         return perf_env__numa_node(data, cpu);
1416 }
1417
1418 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1419                                       struct perf_cpu_map **sockp)
1420 {
1421         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1422 }
1423
1424 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1425                                    struct perf_cpu_map **diep)
1426 {
1427         return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1428 }
1429
1430 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1431                                     struct perf_cpu_map **corep)
1432 {
1433         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1434 }
1435
1436 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1437                                     struct perf_cpu_map **nodep)
1438 {
1439         return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1440 }
1441
1442 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1443                                       struct perf_cpu_map *map, int idx)
1444 {
1445         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1446 }
1447 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1448                                    struct perf_cpu_map *map, int idx)
1449 {
1450         return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1451 }
1452
1453 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1454                                     struct perf_cpu_map *map, int idx)
1455 {
1456         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1457 }
1458
1459 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1460                                     struct perf_cpu_map *map, int idx)
1461 {
1462         return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1463 }
1464
1465 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1466 {
1467         struct perf_env *env = &st->session->header.env;
1468
1469         switch (stat_config.aggr_mode) {
1470         case AGGR_SOCKET:
1471                 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1472                         perror("cannot build socket map");
1473                         return -1;
1474                 }
1475                 stat_config.aggr_get_id = perf_stat__get_socket_file;
1476                 break;
1477         case AGGR_DIE:
1478                 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1479                         perror("cannot build die map");
1480                         return -1;
1481                 }
1482                 stat_config.aggr_get_id = perf_stat__get_die_file;
1483                 break;
1484         case AGGR_CORE:
1485                 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1486                         perror("cannot build core map");
1487                         return -1;
1488                 }
1489                 stat_config.aggr_get_id = perf_stat__get_core_file;
1490                 break;
1491         case AGGR_NODE:
1492                 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1493                         perror("cannot build core map");
1494                         return -1;
1495                 }
1496                 stat_config.aggr_get_id = perf_stat__get_node_file;
1497                 break;
1498         case AGGR_NONE:
1499         case AGGR_GLOBAL:
1500         case AGGR_THREAD:
1501         case AGGR_UNSET:
1502         default:
1503                 break;
1504         }
1505
1506         return 0;
1507 }
1508
1509 /*
1510  * Add default attributes, if there were no attributes specified or
1511  * if -d/--detailed, -d -d or -d -d -d is used:
1512  */
1513 static int add_default_attributes(void)
1514 {
1515         int err;
1516         struct perf_event_attr default_attrs0[] = {
1517
1518   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1519   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1520   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1521   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1522
1523   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1524 };
1525         struct perf_event_attr frontend_attrs[] = {
1526   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1527 };
1528         struct perf_event_attr backend_attrs[] = {
1529   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1530 };
1531         struct perf_event_attr default_attrs1[] = {
1532   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1533   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1534   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1535
1536 };
1537
1538 /*
1539  * Detailed stats (-d), covering the L1 and last level data caches:
1540  */
1541         struct perf_event_attr detailed_attrs[] = {
1542
1543   { .type = PERF_TYPE_HW_CACHE,
1544     .config =
1545          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1546         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1547         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1548
1549   { .type = PERF_TYPE_HW_CACHE,
1550     .config =
1551          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1552         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1553         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1554
1555   { .type = PERF_TYPE_HW_CACHE,
1556     .config =
1557          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1558         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1559         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1560
1561   { .type = PERF_TYPE_HW_CACHE,
1562     .config =
1563          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1564         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1565         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1566 };
1567
1568 /*
1569  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1570  */
1571         struct perf_event_attr very_detailed_attrs[] = {
1572
1573   { .type = PERF_TYPE_HW_CACHE,
1574     .config =
1575          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1576         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1577         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1578
1579   { .type = PERF_TYPE_HW_CACHE,
1580     .config =
1581          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1582         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1583         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1584
1585   { .type = PERF_TYPE_HW_CACHE,
1586     .config =
1587          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1588         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1589         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1590
1591   { .type = PERF_TYPE_HW_CACHE,
1592     .config =
1593          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1594         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1595         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1596
1597   { .type = PERF_TYPE_HW_CACHE,
1598     .config =
1599          PERF_COUNT_HW_CACHE_ITLB               <<  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_ITLB               <<  0  |
1606         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1607         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1608
1609 };
1610
1611 /*
1612  * Very, very detailed stats (-d -d -d), adding prefetch events:
1613  */
1614         struct perf_event_attr very_very_detailed_attrs[] = {
1615
1616   { .type = PERF_TYPE_HW_CACHE,
1617     .config =
1618          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1619         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1620         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1621
1622   { .type = PERF_TYPE_HW_CACHE,
1623     .config =
1624          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1625         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1626         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1627 };
1628         struct parse_events_error errinfo;
1629
1630         /* Set attrs if no event is selected and !null_run: */
1631         if (stat_config.null_run)
1632                 return 0;
1633
1634         bzero(&errinfo, sizeof(errinfo));
1635         if (transaction_run) {
1636                 /* Handle -T as -M transaction. Once platform specific metrics
1637                  * support has been added to the json files, all archictures
1638                  * will use this approach. To determine transaction support
1639                  * on an architecture test for such a metric name.
1640                  */
1641                 if (metricgroup__has_metric("transaction")) {
1642                         struct option opt = { .value = &evsel_list };
1643
1644                         return metricgroup__parse_groups(&opt, "transaction",
1645                                                          stat_config.metric_no_group,
1646                                                         stat_config.metric_no_merge,
1647                                                          &stat_config.metric_events);
1648                 }
1649
1650                 if (pmu_have_event("cpu", "cycles-ct") &&
1651                     pmu_have_event("cpu", "el-start"))
1652                         err = parse_events(evsel_list, transaction_attrs,
1653                                            &errinfo);
1654                 else
1655                         err = parse_events(evsel_list,
1656                                            transaction_limited_attrs,
1657                                            &errinfo);
1658                 if (err) {
1659                         fprintf(stderr, "Cannot set up transaction events\n");
1660                         parse_events_print_error(&errinfo, transaction_attrs);
1661                         return -1;
1662                 }
1663                 return 0;
1664         }
1665
1666         if (smi_cost) {
1667                 int smi;
1668
1669                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1670                         fprintf(stderr, "freeze_on_smi is not supported.\n");
1671                         return -1;
1672                 }
1673
1674                 if (!smi) {
1675                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1676                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1677                                 return -1;
1678                         }
1679                         smi_reset = true;
1680                 }
1681
1682                 if (pmu_have_event("msr", "aperf") &&
1683                     pmu_have_event("msr", "smi")) {
1684                         if (!force_metric_only)
1685                                 stat_config.metric_only = true;
1686                         err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1687                 } else {
1688                         fprintf(stderr, "To measure SMI cost, it needs "
1689                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1690                         parse_events_print_error(&errinfo, smi_cost_attrs);
1691                         return -1;
1692                 }
1693                 if (err) {
1694                         parse_events_print_error(&errinfo, smi_cost_attrs);
1695                         fprintf(stderr, "Cannot set up SMI cost events\n");
1696                         return -1;
1697                 }
1698                 return 0;
1699         }
1700
1701         if (topdown_run) {
1702                 char *str = NULL;
1703                 bool warn = false;
1704
1705                 if (!force_metric_only)
1706                         stat_config.metric_only = true;
1707
1708                 if (topdown_filter_events(topdown_metric_attrs, &str, 1) < 0) {
1709                         pr_err("Out of memory\n");
1710                         return -1;
1711                 }
1712                 if (topdown_metric_attrs[0] && str) {
1713                         if (!stat_config.interval && !stat_config.metric_only) {
1714                                 fprintf(stat_config.output,
1715                                         "Topdown accuracy may decrease when measuring long periods.\n"
1716                                         "Please print the result regularly, e.g. -I1000\n");
1717                         }
1718                         goto setup_metrics;
1719                 }
1720
1721                 zfree(&str);
1722
1723                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1724                     stat_config.aggr_mode != AGGR_CORE) {
1725                         pr_err("top down event configuration requires --per-core mode\n");
1726                         return -1;
1727                 }
1728                 stat_config.aggr_mode = AGGR_CORE;
1729                 if (nr_cgroups || !target__has_cpu(&target)) {
1730                         pr_err("top down event configuration requires system-wide mode (-a)\n");
1731                         return -1;
1732                 }
1733
1734                 if (topdown_filter_events(topdown_attrs, &str,
1735                                 arch_topdown_check_group(&warn)) < 0) {
1736                         pr_err("Out of memory\n");
1737                         return -1;
1738                 }
1739                 if (topdown_attrs[0] && str) {
1740                         if (warn)
1741                                 arch_topdown_group_warn();
1742 setup_metrics:
1743                         err = parse_events(evsel_list, str, &errinfo);
1744                         if (err) {
1745                                 fprintf(stderr,
1746                                         "Cannot set up top down events %s: %d\n",
1747                                         str, err);
1748                                 parse_events_print_error(&errinfo, str);
1749                                 free(str);
1750                                 return -1;
1751                         }
1752                 } else {
1753                         fprintf(stderr, "System does not support topdown\n");
1754                         return -1;
1755                 }
1756                 free(str);
1757         }
1758
1759         if (!evsel_list->core.nr_entries) {
1760                 if (target__has_cpu(&target))
1761                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1762
1763                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1764                         return -1;
1765                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1766                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1767                                 return -1;
1768                 }
1769                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1770                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1771                                 return -1;
1772                 }
1773                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1774                         return -1;
1775         }
1776
1777         /* Detailed events get appended to the event list: */
1778
1779         if (detailed_run <  1)
1780                 return 0;
1781
1782         /* Append detailed run extra attributes: */
1783         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1784                 return -1;
1785
1786         if (detailed_run < 2)
1787                 return 0;
1788
1789         /* Append very detailed run extra attributes: */
1790         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1791                 return -1;
1792
1793         if (detailed_run < 3)
1794                 return 0;
1795
1796         /* Append very, very detailed run extra attributes: */
1797         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1798 }
1799
1800 static const char * const stat_record_usage[] = {
1801         "perf stat record [<options>]",
1802         NULL,
1803 };
1804
1805 static void init_features(struct perf_session *session)
1806 {
1807         int feat;
1808
1809         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1810                 perf_header__set_feat(&session->header, feat);
1811
1812         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1813         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1814         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1815         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1816         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1817 }
1818
1819 static int __cmd_record(int argc, const char **argv)
1820 {
1821         struct perf_session *session;
1822         struct perf_data *data = &perf_stat.data;
1823
1824         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1825                              PARSE_OPT_STOP_AT_NON_OPTION);
1826
1827         if (output_name)
1828                 data->path = output_name;
1829
1830         if (stat_config.run_count != 1 || forever) {
1831                 pr_err("Cannot use -r option with perf stat record.\n");
1832                 return -1;
1833         }
1834
1835         session = perf_session__new(data, false, NULL);
1836         if (IS_ERR(session)) {
1837                 pr_err("Perf session creation failed\n");
1838                 return PTR_ERR(session);
1839         }
1840
1841         init_features(session);
1842
1843         session->evlist   = evsel_list;
1844         perf_stat.session = session;
1845         perf_stat.record  = true;
1846         return argc;
1847 }
1848
1849 static int process_stat_round_event(struct perf_session *session,
1850                                     union perf_event *event)
1851 {
1852         struct perf_record_stat_round *stat_round = &event->stat_round;
1853         struct evsel *counter;
1854         struct timespec tsh, *ts = NULL;
1855         const char **argv = session->header.env.cmdline_argv;
1856         int argc = session->header.env.nr_cmdline;
1857
1858         evlist__for_each_entry(evsel_list, counter)
1859                 perf_stat_process_counter(&stat_config, counter);
1860
1861         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1862                 update_stats(&walltime_nsecs_stats, stat_round->time);
1863
1864         if (stat_config.interval && stat_round->time) {
1865                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1866                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1867                 ts = &tsh;
1868         }
1869
1870         print_counters(ts, argc, argv);
1871         return 0;
1872 }
1873
1874 static
1875 int process_stat_config_event(struct perf_session *session,
1876                               union perf_event *event)
1877 {
1878         struct perf_tool *tool = session->tool;
1879         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1880
1881         perf_event__read_stat_config(&stat_config, &event->stat_config);
1882
1883         if (perf_cpu_map__empty(st->cpus)) {
1884                 if (st->aggr_mode != AGGR_UNSET)
1885                         pr_warning("warning: processing task data, aggregation mode not set\n");
1886                 return 0;
1887         }
1888
1889         if (st->aggr_mode != AGGR_UNSET)
1890                 stat_config.aggr_mode = st->aggr_mode;
1891
1892         if (perf_stat.data.is_pipe)
1893                 perf_stat_init_aggr_mode();
1894         else
1895                 perf_stat_init_aggr_mode_file(st);
1896
1897         return 0;
1898 }
1899
1900 static int set_maps(struct perf_stat *st)
1901 {
1902         if (!st->cpus || !st->threads)
1903                 return 0;
1904
1905         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1906                 return -EINVAL;
1907
1908         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1909
1910         if (evlist__alloc_stats(evsel_list, true))
1911                 return -ENOMEM;
1912
1913         st->maps_allocated = true;
1914         return 0;
1915 }
1916
1917 static
1918 int process_thread_map_event(struct perf_session *session,
1919                              union perf_event *event)
1920 {
1921         struct perf_tool *tool = session->tool;
1922         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1923
1924         if (st->threads) {
1925                 pr_warning("Extra thread map event, ignoring.\n");
1926                 return 0;
1927         }
1928
1929         st->threads = thread_map__new_event(&event->thread_map);
1930         if (!st->threads)
1931                 return -ENOMEM;
1932
1933         return set_maps(st);
1934 }
1935
1936 static
1937 int process_cpu_map_event(struct perf_session *session,
1938                           union perf_event *event)
1939 {
1940         struct perf_tool *tool = session->tool;
1941         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1942         struct perf_cpu_map *cpus;
1943
1944         if (st->cpus) {
1945                 pr_warning("Extra cpu map event, ignoring.\n");
1946                 return 0;
1947         }
1948
1949         cpus = cpu_map__new_data(&event->cpu_map.data);
1950         if (!cpus)
1951                 return -ENOMEM;
1952
1953         st->cpus = cpus;
1954         return set_maps(st);
1955 }
1956
1957 static const char * const stat_report_usage[] = {
1958         "perf stat report [<options>]",
1959         NULL,
1960 };
1961
1962 static struct perf_stat perf_stat = {
1963         .tool = {
1964                 .attr           = perf_event__process_attr,
1965                 .event_update   = perf_event__process_event_update,
1966                 .thread_map     = process_thread_map_event,
1967                 .cpu_map        = process_cpu_map_event,
1968                 .stat_config    = process_stat_config_event,
1969                 .stat           = perf_event__process_stat_event,
1970                 .stat_round     = process_stat_round_event,
1971         },
1972         .aggr_mode = AGGR_UNSET,
1973 };
1974
1975 static int __cmd_report(int argc, const char **argv)
1976 {
1977         struct perf_session *session;
1978         const struct option options[] = {
1979         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1980         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1981                      "aggregate counts per processor socket", AGGR_SOCKET),
1982         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1983                      "aggregate counts per processor die", AGGR_DIE),
1984         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1985                      "aggregate counts per physical processor core", AGGR_CORE),
1986         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
1987                      "aggregate counts per numa node", AGGR_NODE),
1988         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1989                      "disable CPU count aggregation", AGGR_NONE),
1990         OPT_END()
1991         };
1992         struct stat st;
1993         int ret;
1994
1995         argc = parse_options(argc, argv, options, stat_report_usage, 0);
1996
1997         if (!input_name || !strlen(input_name)) {
1998                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1999                         input_name = "-";
2000                 else
2001                         input_name = "perf.data";
2002         }
2003
2004         perf_stat.data.path = input_name;
2005         perf_stat.data.mode = PERF_DATA_MODE_READ;
2006
2007         session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2008         if (IS_ERR(session))
2009                 return PTR_ERR(session);
2010
2011         perf_stat.session  = session;
2012         stat_config.output = stderr;
2013         evsel_list         = session->evlist;
2014
2015         ret = perf_session__process_events(session);
2016         if (ret)
2017                 return ret;
2018
2019         perf_session__delete(session);
2020         return 0;
2021 }
2022
2023 static void setup_system_wide(int forks)
2024 {
2025         /*
2026          * Make system wide (-a) the default target if
2027          * no target was specified and one of following
2028          * conditions is met:
2029          *
2030          *   - there's no workload specified
2031          *   - there is workload specified but all requested
2032          *     events are system wide events
2033          */
2034         if (!target__none(&target))
2035                 return;
2036
2037         if (!forks)
2038                 target.system_wide = true;
2039         else {
2040                 struct evsel *counter;
2041
2042                 evlist__for_each_entry(evsel_list, counter) {
2043                         if (!counter->core.system_wide &&
2044                             strcmp(counter->name, "duration_time")) {
2045                                 return;
2046                         }
2047                 }
2048
2049                 if (evsel_list->core.nr_entries)
2050                         target.system_wide = true;
2051         }
2052 }
2053
2054 int cmd_stat(int argc, const char **argv)
2055 {
2056         const char * const stat_usage[] = {
2057                 "perf stat [<options>] [<command>]",
2058                 NULL
2059         };
2060         int status = -EINVAL, run_idx;
2061         const char *mode;
2062         FILE *output = stderr;
2063         unsigned int interval, timeout;
2064         const char * const stat_subcommands[] = { "record", "report" };
2065
2066         setlocale(LC_ALL, "");
2067
2068         evsel_list = evlist__new();
2069         if (evsel_list == NULL)
2070                 return -ENOMEM;
2071
2072         parse_events__shrink_config_terms();
2073
2074         /* String-parsing callback-based options would segfault when negated */
2075         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2076         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2077         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2078
2079         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2080                                         (const char **) stat_usage,
2081                                         PARSE_OPT_STOP_AT_NON_OPTION);
2082         perf_stat__collect_metric_expr(evsel_list);
2083         perf_stat__init_shadow_stats();
2084
2085         if (stat_config.csv_sep) {
2086                 stat_config.csv_output = true;
2087                 if (!strcmp(stat_config.csv_sep, "\\t"))
2088                         stat_config.csv_sep = "\t";
2089         } else
2090                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2091
2092         if (argc && !strncmp(argv[0], "rec", 3)) {
2093                 argc = __cmd_record(argc, argv);
2094                 if (argc < 0)
2095                         return -1;
2096         } else if (argc && !strncmp(argv[0], "rep", 3))
2097                 return __cmd_report(argc, argv);
2098
2099         interval = stat_config.interval;
2100         timeout = stat_config.timeout;
2101
2102         /*
2103          * For record command the -o is already taken care of.
2104          */
2105         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2106                 output = NULL;
2107
2108         if (output_name && output_fd) {
2109                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2110                 parse_options_usage(stat_usage, stat_options, "o", 1);
2111                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2112                 goto out;
2113         }
2114
2115         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2116                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2117                 goto out;
2118         }
2119
2120         if (stat_config.metric_only && stat_config.run_count > 1) {
2121                 fprintf(stderr, "--metric-only is not supported with -r\n");
2122                 goto out;
2123         }
2124
2125         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2126                 fprintf(stderr, "--table is only supported with -r\n");
2127                 parse_options_usage(stat_usage, stat_options, "r", 1);
2128                 parse_options_usage(NULL, stat_options, "table", 0);
2129                 goto out;
2130         }
2131
2132         if (output_fd < 0) {
2133                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2134                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2135                 goto out;
2136         }
2137
2138         if (!output && !stat_config.quiet) {
2139                 struct timespec tm;
2140                 mode = append_file ? "a" : "w";
2141
2142                 output = fopen(output_name, mode);
2143                 if (!output) {
2144                         perror("failed to create output file");
2145                         return -1;
2146                 }
2147                 clock_gettime(CLOCK_REALTIME, &tm);
2148                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2149         } else if (output_fd > 0) {
2150                 mode = append_file ? "a" : "w";
2151                 output = fdopen(output_fd, mode);
2152                 if (!output) {
2153                         perror("Failed opening logfd");
2154                         return -errno;
2155                 }
2156         }
2157
2158         stat_config.output = output;
2159
2160         /*
2161          * let the spreadsheet do the pretty-printing
2162          */
2163         if (stat_config.csv_output) {
2164                 /* User explicitly passed -B? */
2165                 if (big_num_opt == 1) {
2166                         fprintf(stderr, "-B option not supported with -x\n");
2167                         parse_options_usage(stat_usage, stat_options, "B", 1);
2168                         parse_options_usage(NULL, stat_options, "x", 1);
2169                         goto out;
2170                 } else /* Nope, so disable big number formatting */
2171                         stat_config.big_num = false;
2172         } else if (big_num_opt == 0) /* User passed --no-big-num */
2173                 stat_config.big_num = false;
2174
2175         setup_system_wide(argc);
2176
2177         /*
2178          * Display user/system times only for single
2179          * run and when there's specified tracee.
2180          */
2181         if ((stat_config.run_count == 1) && target__none(&target))
2182                 stat_config.ru_display = true;
2183
2184         if (stat_config.run_count < 0) {
2185                 pr_err("Run count must be a positive number\n");
2186                 parse_options_usage(stat_usage, stat_options, "r", 1);
2187                 goto out;
2188         } else if (stat_config.run_count == 0) {
2189                 forever = true;
2190                 stat_config.run_count = 1;
2191         }
2192
2193         if (stat_config.walltime_run_table) {
2194                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2195                 if (!stat_config.walltime_run) {
2196                         pr_err("failed to setup -r option");
2197                         goto out;
2198                 }
2199         }
2200
2201         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2202                 !target__has_task(&target)) {
2203                 if (!target.system_wide || target.cpu_list) {
2204                         fprintf(stderr, "The --per-thread option is only "
2205                                 "available when monitoring via -p -t -a "
2206                                 "options or only --per-thread.\n");
2207                         parse_options_usage(NULL, stat_options, "p", 1);
2208                         parse_options_usage(NULL, stat_options, "t", 1);
2209                         goto out;
2210                 }
2211         }
2212
2213         /*
2214          * no_aggr, cgroup are for system-wide only
2215          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2216          */
2217         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2218               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2219             !target__has_cpu(&target)) {
2220                 fprintf(stderr, "both cgroup and no-aggregation "
2221                         "modes only available in system-wide mode\n");
2222
2223                 parse_options_usage(stat_usage, stat_options, "G", 1);
2224                 parse_options_usage(NULL, stat_options, "A", 1);
2225                 parse_options_usage(NULL, stat_options, "a", 1);
2226                 goto out;
2227         }
2228
2229         if (add_default_attributes())
2230                 goto out;
2231
2232         if (stat_config.cgroup_list) {
2233                 if (nr_cgroups > 0) {
2234                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2235                         parse_options_usage(stat_usage, stat_options, "G", 1);
2236                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2237                         goto out;
2238                 }
2239
2240                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2241                                           &stat_config.metric_events, true) < 0) {
2242                         parse_options_usage(stat_usage, stat_options,
2243                                             "for-each-cgroup", 0);
2244                         goto out;
2245                 }
2246         }
2247
2248         target__validate(&target);
2249
2250         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2251                 target.per_thread = true;
2252
2253         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2254                 if (target__has_task(&target)) {
2255                         pr_err("Problems finding threads of monitor\n");
2256                         parse_options_usage(stat_usage, stat_options, "p", 1);
2257                         parse_options_usage(NULL, stat_options, "t", 1);
2258                 } else if (target__has_cpu(&target)) {
2259                         perror("failed to parse CPUs map");
2260                         parse_options_usage(stat_usage, stat_options, "C", 1);
2261                         parse_options_usage(NULL, stat_options, "a", 1);
2262                 }
2263                 goto out;
2264         }
2265
2266         evlist__check_cpu_maps(evsel_list);
2267
2268         /*
2269          * Initialize thread_map with comm names,
2270          * so we could print it out on output.
2271          */
2272         if (stat_config.aggr_mode == AGGR_THREAD) {
2273                 thread_map__read_comms(evsel_list->core.threads);
2274                 if (target.system_wide) {
2275                         if (runtime_stat_new(&stat_config,
2276                                 perf_thread_map__nr(evsel_list->core.threads))) {
2277                                 goto out;
2278                         }
2279                 }
2280         }
2281
2282         if (stat_config.aggr_mode == AGGR_NODE)
2283                 cpu__setup_cpunode_map();
2284
2285         if (stat_config.times && interval)
2286                 interval_count = true;
2287         else if (stat_config.times && !interval) {
2288                 pr_err("interval-count option should be used together with "
2289                                 "interval-print.\n");
2290                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2291                 parse_options_usage(stat_usage, stat_options, "I", 1);
2292                 goto out;
2293         }
2294
2295         if (timeout && timeout < 100) {
2296                 if (timeout < 10) {
2297                         pr_err("timeout must be >= 10ms.\n");
2298                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2299                         goto out;
2300                 } else
2301                         pr_warning("timeout < 100ms. "
2302                                    "The overhead percentage could be high in some cases. "
2303                                    "Please proceed with caution.\n");
2304         }
2305         if (timeout && interval) {
2306                 pr_err("timeout option is not supported with interval-print.\n");
2307                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2308                 parse_options_usage(stat_usage, stat_options, "I", 1);
2309                 goto out;
2310         }
2311
2312         if (evlist__alloc_stats(evsel_list, interval))
2313                 goto out;
2314
2315         if (perf_stat_init_aggr_mode())
2316                 goto out;
2317
2318         /*
2319          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2320          * while avoiding that older tools show confusing messages.
2321          *
2322          * However for pipe sessions we need to keep it zero,
2323          * because script's perf_evsel__check_attr is triggered
2324          * by attr->sample_type != 0, and we can't run it on
2325          * stat sessions.
2326          */
2327         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2328
2329         /*
2330          * We dont want to block the signals - that would cause
2331          * child tasks to inherit that and Ctrl-C would not work.
2332          * What we want is for Ctrl-C to work in the exec()-ed
2333          * task, but being ignored by perf stat itself:
2334          */
2335         atexit(sig_atexit);
2336         if (!forever)
2337                 signal(SIGINT,  skip_signal);
2338         signal(SIGCHLD, skip_signal);
2339         signal(SIGALRM, skip_signal);
2340         signal(SIGABRT, skip_signal);
2341
2342         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2343                 goto out;
2344
2345         status = 0;
2346         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2347                 if (stat_config.run_count != 1 && verbose > 0)
2348                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2349                                 run_idx + 1);
2350
2351                 if (run_idx != 0)
2352                         evlist__reset_prev_raw_counts(evsel_list);
2353
2354                 status = run_perf_stat(argc, argv, run_idx);
2355                 if (forever && status != -1 && !interval) {
2356                         print_counters(NULL, argc, argv);
2357                         perf_stat__reset_stats();
2358                 }
2359         }
2360
2361         if (!forever && status != -1 && (!interval || stat_config.summary))
2362                 print_counters(NULL, argc, argv);
2363
2364         evlist__finalize_ctlfd(evsel_list);
2365
2366         if (STAT_RECORD) {
2367                 /*
2368                  * We synthesize the kernel mmap record just so that older tools
2369                  * don't emit warnings about not being able to resolve symbols
2370                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2371                  * a saner message about no samples being in the perf.data file.
2372                  *
2373                  * This also serves to suppress a warning about f_header.data.size == 0
2374                  * in header.c at the moment 'perf stat record' gets introduced, which
2375                  * is not really needed once we start adding the stat specific PERF_RECORD_
2376                  * records, but the need to suppress the kptr_restrict messages in older
2377                  * tools remain  -acme
2378                  */
2379                 int fd = perf_data__fd(&perf_stat.data);
2380                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2381                                                              process_synthesized_event,
2382                                                              &perf_stat.session->machines.host);
2383                 if (err) {
2384                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2385                                    "older tools may produce warnings about this file\n.");
2386                 }
2387
2388                 if (!interval) {
2389                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2390                                 pr_err("failed to write stat round event\n");
2391                 }
2392
2393                 if (!perf_stat.data.is_pipe) {
2394                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2395                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2396                 }
2397
2398                 evlist__close(evsel_list);
2399                 perf_session__delete(perf_stat.session);
2400         }
2401
2402         perf_stat__exit_aggr_mode();
2403         evlist__free_stats(evsel_list);
2404 out:
2405         zfree(&stat_config.walltime_run);
2406
2407         if (smi_cost && smi_reset)
2408                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2409
2410         evlist__delete(evsel_list);
2411
2412         metricgroup__rblist_exit(&stat_config.metric_events);
2413         runtime_stat_delete(&stat_config);
2414         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2415
2416         return status;
2417 }