tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/cgroup.h"
  47 #include "util/util.h"
  48 #include <subcmd/parse-options.h>
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/drv_configs.h"
  56 #include "util/color.h"
  57 #include "util/stat.h"
  58 #include "util/header.h"
  59 #include "util/cpumap.h"
  60 #include "util/thread.h"
  61 #include "util/thread_map.h"
  62 #include "util/counts.h"
  63 #include "util/group.h"
  64 #include "util/session.h"
  65 #include "util/tool.h"
  66 #include "util/string2.h"
  67 #include "util/metricgroup.h"
  68 #include "util/top.h"
  69 #include "asm/bug.h"
  70
  71 #include <linux/time64.h>
  72 #include <api/fs/fs.h>
  73 #include <errno.h>
  74 #include <signal.h>
  75 #include <stdlib.h>
  76 #include <sys/prctl.h>
  77 #include <inttypes.h>
  78 #include <locale.h>
  79 #include <math.h>
  80 #include <sys/types.h>
  81 #include <sys/stat.h>
  82 #include <sys/wait.h>
  83 #include <unistd.h>
  84 #include <sys/time.h>
  85 #include <sys/resource.h>
  86 #include <sys/wait.h>
  87
  88 #include "sane_ctype.h"
  89
  90 #define DEFAULT_SEPARATOR       " "
  91 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  92
  93 static void print_counters(struct timespec *ts, int argc, const char **argv);
  94
  95 /* Default events used for perf stat -T */
  96 static const char *transaction_attrs = {
  97         "task-clock,"
  98         "{"
  99         "instructions,"
 100         "cycles,"
 101         "cpu/cycles-t/,"
 102         "cpu/tx-start/,"
 103         "cpu/el-start/,"
 104         "cpu/cycles-ct/"
 105         "}"
 106 };
 107
 108 /* More limited version when the CPU does not have all events. */
 109 static const char * transaction_limited_attrs = {
 110         "task-clock,"
 111         "{"
 112         "instructions,"
 113         "cycles,"
 114         "cpu/cycles-t/,"
 115         "cpu/tx-start/"
 116         "}"
 117 };
 118
 119 static const char * topdown_attrs[] = {
 120         "topdown-total-slots",
 121         "topdown-slots-retired",
 122         "topdown-recovery-bubbles",
 123         "topdown-fetch-bubbles",
 124         "topdown-slots-issued",
 125         NULL,
 126 };
 127
 128 static const char *smi_cost_attrs = {
 129         "{"
 130         "msr/aperf/,"
 131         "msr/smi/,"
 132         "cycles"
 133         "}"
 134 };
 135
 136 static struct perf_evlist       *evsel_list;
 137
 138 static struct target target = {
 139         .uid    = UINT_MAX,
 140 };
 141
 142 #define METRIC_ONLY_LEN 20
 143
 144 static volatile pid_t           child_pid                       = -1;
 145 static int                      detailed_run                    =  0;
 146 static bool                     transaction_run;
 147 static bool                     topdown_run                     = false;
 148 static bool                     smi_cost                        = false;
 149 static bool                     smi_reset                       = false;
 150 static int                      big_num_opt                     =  -1;
 151 static bool                     group                           = false;
 152 static const char               *pre_cmd                        = NULL;
 153 static const char               *post_cmd                       = NULL;
 154 static bool                     sync_run                        = false;
 155 static bool                     forever                         = false;
 156 static bool                     force_metric_only               = false;
 157 static struct timespec          ref_time;
 158 static bool                     append_file;
 159 static bool                     interval_count;
 160 static const char               *output_name;
 161 static int                      output_fd;
 162
 163 struct perf_stat {
 164         bool                     record;
 165         struct perf_data         data;
 166         struct perf_session     *session;
 167         u64                      bytes_written;
 168         struct perf_tool         tool;
 169         bool                     maps_allocated;
 170         struct cpu_map          *cpus;
 171         struct thread_map       *threads;
 172         enum aggr_mode           aggr_mode;
 173 };
 174
 175 static struct perf_stat         perf_stat;
 176 #define STAT_RECORD             perf_stat.record
 177
 178 static volatile int done = 0;
 179
 180 static struct perf_stat_config stat_config = {
 181         .aggr_mode              = AGGR_GLOBAL,
 182         .scale                  = true,
 183         .unit_width             = 4, /* strlen("unit") */
 184         .run_count              = 1,
 185         .metric_only_len        = METRIC_ONLY_LEN,
 186         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 187         .big_num                = true,
 188 };
 189
 190 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 191                                  struct timespec *b)
 192 {
 193         r->tv_sec = a->tv_sec - b->tv_sec;
 194         if (a->tv_nsec < b->tv_nsec) {
 195                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 196                 r->tv_sec--;
 197         } else {
 198                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 199         }
 200 }
 201
 202 static void perf_stat__reset_stats(void)
 203 {
 204         int i;
 205
 206         perf_evlist__reset_stats(evsel_list);
 207         perf_stat__reset_shadow_stats();
 208
 209         for (i = 0; i < stat_config.stats_num; i++)
 210                 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
 211 }
 212
 213 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 214                                      union perf_event *event,
 215                                      struct perf_sample *sample __maybe_unused,
 216                                      struct machine *machine __maybe_unused)
 217 {
 218         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 219                 pr_err("failed to write perf data, error: %m\n");
 220                 return -1;
 221         }
 222
 223         perf_stat.bytes_written += event->header.size;
 224         return 0;
 225 }
 226
 227 static int write_stat_round_event(u64 tm, u64 type)
 228 {
 229         return perf_event__synthesize_stat_round(NULL, tm, type,
 230                                                  process_synthesized_event,
 231                                                  NULL);
 232 }
 233
 234 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 235         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 236
 237 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 238
 239 static int
 240 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 241                              struct perf_counts_values *count)
 242 {
 243         struct perf_sample_id *sid = SID(counter, cpu, thread);
 244
 245         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 246                                            process_synthesized_event, NULL);
 247 }
 248
 249 /*
 250  * Read out the results of a single counter:
 251  * do not aggregate counts across CPUs in system-wide mode
 252  */
 253 static int read_counter(struct perf_evsel *counter)
 254 {
 255         int nthreads = thread_map__nr(evsel_list->threads);
 256         int ncpus, cpu, thread;
 257
 258         if (target__has_cpu(&target) && !target__has_per_thread(&target))
 259                 ncpus = perf_evsel__nr_cpus(counter);
 260         else
 261                 ncpus = 1;
 262
 263         if (!counter->supported)
 264                 return -ENOENT;
 265
 266         if (counter->system_wide)
 267                 nthreads = 1;
 268
 269         for (thread = 0; thread < nthreads; thread++) {
 270                 for (cpu = 0; cpu < ncpus; cpu++) {
 271                         struct perf_counts_values *count;
 272
 273                         count = perf_counts(counter->counts, cpu, thread);
 274
 275                         /*
 276                          * The leader's group read loads data into its group members
 277                          * (via perf_evsel__read_counter) and sets threir count->loaded.
 278                          */
 279                         if (!count->loaded &&
 280                             perf_evsel__read_counter(counter, cpu, thread)) {
 281                                 counter->counts->scaled = -1;
 282                                 perf_counts(counter->counts, cpu, thread)->ena = 0;
 283                                 perf_counts(counter->counts, cpu, thread)->run = 0;
 284                                 return -1;
 285                         }
 286
 287                         count->loaded = false;
 288
 289                         if (STAT_RECORD) {
 290                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 291                                         pr_err("failed to write stat event\n");
 292                                         return -1;
 293                                 }
 294                         }
 295
 296                         if (verbose > 1) {
 297                                 fprintf(stat_config.output,
 298                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 299                                                 perf_evsel__name(counter),
 300                                                 cpu,
 301                                                 count->val, count->ena, count->run);
 302                         }
 303                 }
 304         }
 305
 306         return 0;
 307 }
 308
 309 static void read_counters(void)
 310 {
 311         struct perf_evsel *counter;
 312         int ret;
 313
 314         evlist__for_each_entry(evsel_list, counter) {
 315                 ret = read_counter(counter);
 316                 if (ret)
 317                         pr_debug("failed to read counter %s\n", counter->name);
 318
 319                 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
 320                         pr_warning("failed to process counter %s\n", counter->name);
 321         }
 322 }
 323
 324 static void process_interval(void)
 325 {
 326         struct timespec ts, rs;
 327
 328         read_counters();
 329
 330         clock_gettime(CLOCK_MONOTONIC, &ts);
 331         diff_timespec(&rs, &ts, &ref_time);
 332
 333         if (STAT_RECORD) {
 334                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 335                         pr_err("failed to write stat round event\n");
 336         }
 337
 338         init_stats(&walltime_nsecs_stats);
 339         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
 340         print_counters(&rs, 0, NULL);
 341 }
 342
 343 static void enable_counters(void)
 344 {
 345         if (stat_config.initial_delay)
 346                 usleep(stat_config.initial_delay * USEC_PER_MSEC);
 347
 348         /*
 349          * We need to enable counters only if:
 350          * - we don't have tracee (attaching to task or cpu)
 351          * - we have initial delay configured
 352          */
 353         if (!target__none(&target) || stat_config.initial_delay)
 354                 perf_evlist__enable(evsel_list);
 355 }
 356
 357 static void disable_counters(void)
 358 {
 359         /*
 360          * If we don't have tracee (attaching to task or cpu), counters may
 361          * still be running. To get accurate group ratios, we must stop groups
 362          * from counting before reading their constituent counters.
 363          */
 364         if (!target__none(&target))
 365                 perf_evlist__disable(evsel_list);
 366 }
 367
 368 static volatile int workload_exec_errno;
 369
 370 /*
 371  * perf_evlist__prepare_workload will send a SIGUSR1
 372  * if the fork fails, since we asked by setting its
 373  * want_signal to true.
 374  */
 375 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 376                                         void *ucontext __maybe_unused)
 377 {
 378         workload_exec_errno = info->si_value.sival_int;
 379 }
 380
 381 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
 382 {
 383         return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
 384 }
 385
 386 static bool is_target_alive(struct target *_target,
 387                             struct thread_map *threads)
 388 {
 389         struct stat st;
 390         int i;
 391
 392         if (!target__has_task(_target))
 393                 return true;
 394
 395         for (i = 0; i < threads->nr; i++) {
 396                 char path[PATH_MAX];
 397
 398                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 399                           threads->map[i].pid);
 400
 401                 if (!stat(path, &st))
 402                         return true;
 403         }
 404
 405         return false;
 406 }
 407
 408 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 409 {
 410         int interval = stat_config.interval;
 411         int times = stat_config.times;
 412         int timeout = stat_config.timeout;
 413         char msg[BUFSIZ];
 414         unsigned long long t0, t1;
 415         struct perf_evsel *counter;
 416         struct timespec ts;
 417         size_t l;
 418         int status = 0;
 419         const bool forks = (argc > 0);
 420         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 421         struct perf_evsel_config_term *err_term;
 422
 423         if (interval) {
 424                 ts.tv_sec  = interval / USEC_PER_MSEC;
 425                 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
 426         } else if (timeout) {
 427                 ts.tv_sec  = timeout / USEC_PER_MSEC;
 428                 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
 429         } else {
 430                 ts.tv_sec  = 1;
 431                 ts.tv_nsec = 0;
 432         }
 433
 434         if (forks) {
 435                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 436                                                   workload_exec_failed_signal) < 0) {
 437                         perror("failed to prepare workload");
 438                         return -1;
 439                 }
 440                 child_pid = evsel_list->workload.pid;
 441         }
 442
 443         if (group)
 444                 perf_evlist__set_leader(evsel_list);
 445
 446         evlist__for_each_entry(evsel_list, counter) {
 447 try_again:
 448                 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
 449
 450                         /* Weak group failed. Reset the group. */
 451                         if ((errno == EINVAL || errno == EBADF) &&
 452                             counter->leader != counter &&
 453                             counter->weak_group) {
 454                                 counter = perf_evlist__reset_weak_group(evsel_list, counter);
 455                                 goto try_again;
 456                         }
 457
 458                         /*
 459                          * PPC returns ENXIO for HW counters until 2.6.37
 460                          * (behavior changed with commit b0a873e).
 461                          */
 462                         if (errno == EINVAL || errno == ENOSYS ||
 463                             errno == ENOENT || errno == EOPNOTSUPP ||
 464                             errno == ENXIO) {
 465                                 if (verbose > 0)
 466                                         ui__warning("%s event is not supported by the kernel.\n",
 467                                                     perf_evsel__name(counter));
 468                                 counter->supported = false;
 469
 470                                 if ((counter->leader != counter) ||
 471                                     !(counter->leader->nr_members > 1))
 472                                         continue;
 473                         } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 474                                 if (verbose > 0)
 475                                         ui__warning("%s\n", msg);
 476                                 goto try_again;
 477                         } else if (target__has_per_thread(&target) &&
 478                                    evsel_list->threads &&
 479                                    evsel_list->threads->err_thread != -1) {
 480                                 /*
 481                                  * For global --per-thread case, skip current
 482                                  * error thread.
 483                                  */
 484                                 if (!thread_map__remove(evsel_list->threads,
 485                                                         evsel_list->threads->err_thread)) {
 486                                         evsel_list->threads->err_thread = -1;
 487                                         goto try_again;
 488                                 }
 489                         }
 490
 491                         perf_evsel__open_strerror(counter, &target,
 492                                                   errno, msg, sizeof(msg));
 493                         ui__error("%s\n", msg);
 494
 495                         if (child_pid != -1)
 496                                 kill(child_pid, SIGTERM);
 497
 498                         return -1;
 499                 }
 500                 counter->supported = true;
 501
 502                 l = strlen(counter->unit);
 503                 if (l > stat_config.unit_width)
 504                         stat_config.unit_width = l;
 505
 506                 if (perf_evsel__should_store_id(counter) &&
 507                     perf_evsel__store_ids(counter, evsel_list))
 508                         return -1;
 509         }
 510
 511         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 512                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 513                         counter->filter, perf_evsel__name(counter), errno,
 514                         str_error_r(errno, msg, sizeof(msg)));
 515                 return -1;
 516         }
 517
 518         if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
 519                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
 520                       err_term->val.drv_cfg, perf_evsel__name(counter), errno,
 521                       str_error_r(errno, msg, sizeof(msg)));
 522                 return -1;
 523         }
 524
 525         if (STAT_RECORD) {
 526                 int err, fd = perf_data__fd(&perf_stat.data);
 527
 528                 if (is_pipe) {
 529                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 530                 } else {
 531                         err = perf_session__write_header(perf_stat.session, evsel_list,
 532                                                          fd, false);
 533                 }
 534
 535                 if (err < 0)
 536                         return err;
 537
 538                 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
 539                                                   process_synthesized_event, is_pipe);
 540                 if (err < 0)
 541                         return err;
 542         }
 543
 544         /*
 545          * Enable counters and exec the command:
 546          */
 547         t0 = rdclock();
 548         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 549
 550         if (forks) {
 551                 perf_evlist__start_workload(evsel_list);
 552                 enable_counters();
 553
 554                 if (interval || timeout) {
 555                         while (!waitpid(child_pid, &status, WNOHANG)) {
 556                                 nanosleep(&ts, NULL);
 557                                 if (timeout)
 558                                         break;
 559                                 process_interval();
 560                                 if (interval_count && !(--times))
 561                                         break;
 562                         }
 563                 }
 564                 wait4(child_pid, &status, 0, &stat_config.ru_data);
 565
 566                 if (workload_exec_errno) {
 567                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 568                         pr_err("Workload failed: %s\n", emsg);
 569                         return -1;
 570                 }
 571
 572                 if (WIFSIGNALED(status))
 573                         psignal(WTERMSIG(status), argv[0]);
 574         } else {
 575                 enable_counters();
 576                 while (!done) {
 577                         nanosleep(&ts, NULL);
 578                         if (!is_target_alive(&target, evsel_list->threads))
 579                                 break;
 580                         if (timeout)
 581                                 break;
 582                         if (interval) {
 583                                 process_interval();
 584                                 if (interval_count && !(--times))
 585                                         break;
 586                         }
 587                 }
 588         }
 589
 590         disable_counters();
 591
 592         t1 = rdclock();
 593
 594         if (stat_config.walltime_run_table)
 595                 stat_config.walltime_run[run_idx] = t1 - t0;
 596
 597         update_stats(&walltime_nsecs_stats, t1 - t0);
 598
 599         /*
 600          * Closing a group leader splits the group, and as we only disable
 601          * group leaders, results in remaining events becoming enabled. To
 602          * avoid arbitrary skew, we must read all counters before closing any
 603          * group leaders.
 604          */
 605         read_counters();
 606         perf_evlist__close(evsel_list);
 607
 608         return WEXITSTATUS(status);
 609 }
 610
 611 static int run_perf_stat(int argc, const char **argv, int run_idx)
 612 {
 613         int ret;
 614
 615         if (pre_cmd) {
 616                 ret = system(pre_cmd);
 617                 if (ret)
 618                         return ret;
 619         }
 620
 621         if (sync_run)
 622                 sync();
 623
 624         ret = __run_perf_stat(argc, argv, run_idx);
 625         if (ret)
 626                 return ret;
 627
 628         if (post_cmd) {
 629                 ret = system(post_cmd);
 630                 if (ret)
 631                         return ret;
 632         }
 633
 634         return ret;
 635 }
 636
 637 static void print_counters(struct timespec *ts, int argc, const char **argv)
 638 {
 639         /* Do not print anything if we record to the pipe. */
 640         if (STAT_RECORD && perf_stat.data.is_pipe)
 641                 return;
 642
 643         perf_evlist__print_counters(evsel_list, &stat_config, &target,
 644                                     ts, argc, argv);
 645 }
 646
 647 static volatile int signr = -1;
 648
 649 static void skip_signal(int signo)
 650 {
 651         if ((child_pid == -1) || stat_config.interval)
 652                 done = 1;
 653
 654         signr = signo;
 655         /*
 656          * render child_pid harmless
 657          * won't send SIGTERM to a random
 658          * process in case of race condition
 659          * and fast PID recycling
 660          */
 661         child_pid = -1;
 662 }
 663
 664 static void sig_atexit(void)
 665 {
 666         sigset_t set, oset;
 667
 668         /*
 669          * avoid race condition with SIGCHLD handler
 670          * in skip_signal() which is modifying child_pid
 671          * goal is to avoid send SIGTERM to a random
 672          * process
 673          */
 674         sigemptyset(&set);
 675         sigaddset(&set, SIGCHLD);
 676         sigprocmask(SIG_BLOCK, &set, &oset);
 677
 678         if (child_pid != -1)
 679                 kill(child_pid, SIGTERM);
 680
 681         sigprocmask(SIG_SETMASK, &oset, NULL);
 682
 683         if (signr == -1)
 684                 return;
 685
 686         signal(signr, SIG_DFL);
 687         kill(getpid(), signr);
 688 }
 689
 690 static int stat__set_big_num(const struct option *opt __maybe_unused,
 691                              const char *s __maybe_unused, int unset)
 692 {
 693         big_num_opt = unset ? 0 : 1;
 694         return 0;
 695 }
 696
 697 static int enable_metric_only(const struct option *opt __maybe_unused,
 698                               const char *s __maybe_unused, int unset)
 699 {
 700         force_metric_only = true;
 701         stat_config.metric_only = !unset;
 702         return 0;
 703 }
 704
 705 static int parse_metric_groups(const struct option *opt,
 706                                const char *str,
 707                                int unset __maybe_unused)
 708 {
 709         return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
 710 }
 711
 712 static const struct option stat_options[] = {
 713         OPT_BOOLEAN('T', "transaction", &transaction_run,
 714                     "hardware transaction statistics"),
 715         OPT_CALLBACK('e', "event", &evsel_list, "event",
 716                      "event selector. use 'perf list' to list available events",
 717                      parse_events_option),
 718         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 719                      "event filter", parse_filter),
 720         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
 721                     "child tasks do not inherit counters"),
 722         OPT_STRING('p', "pid", &target.pid, "pid",
 723                    "stat events on existing process id"),
 724         OPT_STRING('t', "tid", &target.tid, "tid",
 725                    "stat events on existing thread id"),
 726         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
 727                     "system-wide collection from all CPUs"),
 728         OPT_BOOLEAN('g', "group", &group,
 729                     "put the counters into a counter group"),
 730         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
 731         OPT_INCR('v', "verbose", &verbose,
 732                     "be more verbose (show counter open errors, etc)"),
 733         OPT_INTEGER('r', "repeat", &stat_config.run_count,
 734                     "repeat command and print average + stddev (max: 100, forever: 0)"),
 735         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
 736                     "display details about each run (only with -r option)"),
 737         OPT_BOOLEAN('n', "null", &stat_config.null_run,
 738                     "null run - dont start any counters"),
 739         OPT_INCR('d', "detailed", &detailed_run,
 740                     "detailed run - start a lot of events"),
 741         OPT_BOOLEAN('S', "sync", &sync_run,
 742                     "call sync() before starting a run"),
 743         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 744                            "print large numbers with thousands\' separators",
 745                            stat__set_big_num),
 746         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
 747                     "list of cpus to monitor in system-wide"),
 748         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
 749                     "disable CPU count aggregation", AGGR_NONE),
 750         OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
 751         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
 752                    "print counts with custom separator"),
 753         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 754                      "monitor event in cgroup name only", parse_cgroups),
 755         OPT_STRING('o', "output", &output_name, "file", "output file name"),
 756         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
 757         OPT_INTEGER(0, "log-fd", &output_fd,
 758                     "log output to fd, instead of stderr"),
 759         OPT_STRING(0, "pre", &pre_cmd, "command",
 760                         "command to run prior to the measured command"),
 761         OPT_STRING(0, "post", &post_cmd, "command",
 762                         "command to run after to the measured command"),
 763         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
 764                     "print counts at regular interval in ms "
 765                     "(overhead is possible for values <= 100ms)"),
 766         OPT_INTEGER(0, "interval-count", &stat_config.times,
 767                     "print counts for fixed number of times"),
 768         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
 769                     "clear screen in between new interval"),
 770         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
 771                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
 772         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
 773                      "aggregate counts per processor socket", AGGR_SOCKET),
 774         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
 775                      "aggregate counts per physical processor core", AGGR_CORE),
 776         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
 777                      "aggregate counts per thread", AGGR_THREAD),
 778         OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
 779                      "ms to wait before starting measurement after program start"),
 780         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
 781                         "Only print computed metrics. No raw values", enable_metric_only),
 782         OPT_BOOLEAN(0, "topdown", &topdown_run,
 783                         "measure topdown level 1 statistics"),
 784         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
 785                         "measure SMI cost"),
 786         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
 787                      "monitor specified metrics or metric groups (separated by ,)",
 788                      parse_metric_groups),
 789         OPT_END()
 790 };
 791
 792 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
 793                                  struct cpu_map *map, int cpu)
 794 {
 795         return cpu_map__get_socket(map, cpu, NULL);
 796 }
 797
 798 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
 799                                struct cpu_map *map, int cpu)
 800 {
 801         return cpu_map__get_core(map, cpu, NULL);
 802 }
 803
 804 static int cpu_map__get_max(struct cpu_map *map)
 805 {
 806         int i, max = -1;
 807
 808         for (i = 0; i < map->nr; i++) {
 809                 if (map->map[i] > max)
 810                         max = map->map[i];
 811         }
 812
 813         return max;
 814 }
 815
 816 static int perf_stat__get_aggr(struct perf_stat_config *config,
 817                                aggr_get_id_t get_id, struct cpu_map *map, int idx)
 818 {
 819         int cpu;
 820
 821         if (idx >= map->nr)
 822                 return -1;
 823
 824         cpu = map->map[idx];
 825
 826         if (config->cpus_aggr_map->map[cpu] == -1)
 827                 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
 828
 829         return config->cpus_aggr_map->map[cpu];
 830 }
 831
 832 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
 833                                         struct cpu_map *map, int idx)
 834 {
 835         return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
 836 }
 837
 838 static int perf_stat__get_core_cached(struct perf_stat_config *config,
 839                                       struct cpu_map *map, int idx)
 840 {
 841         return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
 842 }
 843
 844 static int perf_stat_init_aggr_mode(void)
 845 {
 846         int nr;
 847
 848         switch (stat_config.aggr_mode) {
 849         case AGGR_SOCKET:
 850                 if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
 851                         perror("cannot build socket map");
 852                         return -1;
 853                 }
 854                 stat_config.aggr_get_id = perf_stat__get_socket_cached;
 855                 break;
 856         case AGGR_CORE:
 857                 if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
 858                         perror("cannot build core map");
 859                         return -1;
 860                 }
 861                 stat_config.aggr_get_id = perf_stat__get_core_cached;
 862                 break;
 863         case AGGR_NONE:
 864         case AGGR_GLOBAL:
 865         case AGGR_THREAD:
 866         case AGGR_UNSET:
 867         default:
 868                 break;
 869         }
 870
 871         /*
 872          * The evsel_list->cpus is the base we operate on,
 873          * taking the highest cpu number to be the size of
 874          * the aggregation translate cpumap.
 875          */
 876         nr = cpu_map__get_max(evsel_list->cpus);
 877         stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
 878         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
 879 }
 880
 881 static void perf_stat__exit_aggr_mode(void)
 882 {
 883         cpu_map__put(stat_config.aggr_map);
 884         cpu_map__put(stat_config.cpus_aggr_map);
 885         stat_config.aggr_map = NULL;
 886         stat_config.cpus_aggr_map = NULL;
 887 }
 888
 889 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
 890 {
 891         int cpu;
 892
 893         if (idx > map->nr)
 894                 return -1;
 895
 896         cpu = map->map[idx];
 897
 898         if (cpu >= env->nr_cpus_avail)
 899                 return -1;
 900
 901         return cpu;
 902 }
 903
 904 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
 905 {
 906         struct perf_env *env = data;
 907         int cpu = perf_env__get_cpu(env, map, idx);
 908
 909         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
 910 }
 911
 912 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
 913 {
 914         struct perf_env *env = data;
 915         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
 916
 917         if (cpu != -1) {
 918                 int socket_id = env->cpu[cpu].socket_id;
 919
 920                 /*
 921                  * Encode socket in upper 16 bits
 922                  * core_id is relative to socket, and
 923                  * we need a global id. So we combine
 924                  * socket + core id.
 925                  */
 926                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
 927         }
 928
 929         return core;
 930 }
 931
 932 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
 933                                       struct cpu_map **sockp)
 934 {
 935         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
 936 }
 937
 938 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
 939                                     struct cpu_map **corep)
 940 {
 941         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
 942 }
 943
 944 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
 945                                       struct cpu_map *map, int idx)
 946 {
 947         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
 948 }
 949
 950 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
 951                                     struct cpu_map *map, int idx)
 952 {
 953         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
 954 }
 955
 956 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
 957 {
 958         struct perf_env *env = &st->session->header.env;
 959
 960         switch (stat_config.aggr_mode) {
 961         case AGGR_SOCKET:
 962                 if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
 963                         perror("cannot build socket map");
 964                         return -1;
 965                 }
 966                 stat_config.aggr_get_id = perf_stat__get_socket_file;
 967                 break;
 968         case AGGR_CORE:
 969                 if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
 970                         perror("cannot build core map");
 971                         return -1;
 972                 }
 973                 stat_config.aggr_get_id = perf_stat__get_core_file;
 974                 break;
 975         case AGGR_NONE:
 976         case AGGR_GLOBAL:
 977         case AGGR_THREAD:
 978         case AGGR_UNSET:
 979         default:
 980                 break;
 981         }
 982
 983         return 0;
 984 }
 985
 986 static int topdown_filter_events(const char **attr, char **str, bool use_group)
 987 {
 988         int off = 0;
 989         int i;
 990         int len = 0;
 991         char *s;
 992
 993         for (i = 0; attr[i]; i++) {
 994                 if (pmu_have_event("cpu", attr[i])) {
 995                         len += strlen(attr[i]) + 1;
 996                         attr[i - off] = attr[i];
 997                 } else
 998                         off++;
 999         }
1000         attr[i - off] = NULL;
1001
1002         *str = malloc(len + 1 + 2);
1003         if (!*str)
1004                 return -1;
1005         s = *str;
1006         if (i - off == 0) {
1007                 *s = 0;
1008                 return 0;
1009         }
1010         if (use_group)
1011                 *s++ = '{';
1012         for (i = 0; attr[i]; i++) {
1013                 strcpy(s, attr[i]);
1014                 s += strlen(s);
1015                 *s++ = ',';
1016         }
1017         if (use_group) {
1018                 s[-1] = '}';
1019                 *s = 0;
1020         } else
1021                 s[-1] = 0;
1022         return 0;
1023 }
1024
1025 __weak bool arch_topdown_check_group(bool *warn)
1026 {
1027         *warn = false;
1028         return false;
1029 }
1030
1031 __weak void arch_topdown_group_warn(void)
1032 {
1033 }
1034
1035 /*
1036  * Add default attributes, if there were no attributes specified or
1037  * if -d/--detailed, -d -d or -d -d -d is used:
1038  */
1039 static int add_default_attributes(void)
1040 {
1041         int err;
1042         struct perf_event_attr default_attrs0[] = {
1043
1044   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1045   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1046   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1047   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1048
1049   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1050 };
1051         struct perf_event_attr frontend_attrs[] = {
1052   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1053 };
1054         struct perf_event_attr backend_attrs[] = {
1055   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1056 };
1057         struct perf_event_attr default_attrs1[] = {
1058   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1059   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1060   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1061
1062 };
1063
1064 /*
1065  * Detailed stats (-d), covering the L1 and last level data caches:
1066  */
1067         struct perf_event_attr detailed_attrs[] = {
1068
1069   { .type = PERF_TYPE_HW_CACHE,
1070     .config =
1071          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1072         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1073         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1074
1075   { .type = PERF_TYPE_HW_CACHE,
1076     .config =
1077          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1078         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1079         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1080
1081   { .type = PERF_TYPE_HW_CACHE,
1082     .config =
1083          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1084         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1085         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1086
1087   { .type = PERF_TYPE_HW_CACHE,
1088     .config =
1089          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1090         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1091         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1092 };
1093
1094 /*
1095  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1096  */
1097         struct perf_event_attr very_detailed_attrs[] = {
1098
1099   { .type = PERF_TYPE_HW_CACHE,
1100     .config =
1101          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1102         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1103         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1104
1105   { .type = PERF_TYPE_HW_CACHE,
1106     .config =
1107          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1108         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1109         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1110
1111   { .type = PERF_TYPE_HW_CACHE,
1112     .config =
1113          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1114         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1115         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1116
1117   { .type = PERF_TYPE_HW_CACHE,
1118     .config =
1119          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1120         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1121         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1122
1123   { .type = PERF_TYPE_HW_CACHE,
1124     .config =
1125          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1126         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1127         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1128
1129   { .type = PERF_TYPE_HW_CACHE,
1130     .config =
1131          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1132         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1133         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1134
1135 };
1136
1137 /*
1138  * Very, very detailed stats (-d -d -d), adding prefetch events:
1139  */
1140         struct perf_event_attr very_very_detailed_attrs[] = {
1141
1142   { .type = PERF_TYPE_HW_CACHE,
1143     .config =
1144          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1145         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1146         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1147
1148   { .type = PERF_TYPE_HW_CACHE,
1149     .config =
1150          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1151         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1152         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1153 };
1154         struct parse_events_error errinfo;
1155
1156         /* Set attrs if no event is selected and !null_run: */
1157         if (stat_config.null_run)
1158                 return 0;
1159
1160         if (transaction_run) {
1161                 /* Handle -T as -M transaction. Once platform specific metrics
1162                  * support has been added to the json files, all archictures
1163                  * will use this approach. To determine transaction support
1164                  * on an architecture test for such a metric name.
1165                  */
1166                 if (metricgroup__has_metric("transaction")) {
1167                         struct option opt = { .value = &evsel_list };
1168
1169                         return metricgroup__parse_groups(&opt, "transaction",
1170                                                          &stat_config.metric_events);
1171                 }
1172
1173                 if (pmu_have_event("cpu", "cycles-ct") &&
1174                     pmu_have_event("cpu", "el-start"))
1175                         err = parse_events(evsel_list, transaction_attrs,
1176                                            &errinfo);
1177                 else
1178                         err = parse_events(evsel_list,
1179                                            transaction_limited_attrs,
1180                                            &errinfo);
1181                 if (err) {
1182                         fprintf(stderr, "Cannot set up transaction events\n");
1183                         parse_events_print_error(&errinfo, transaction_attrs);
1184                         return -1;
1185                 }
1186                 return 0;
1187         }
1188
1189         if (smi_cost) {
1190                 int smi;
1191
1192                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1193                         fprintf(stderr, "freeze_on_smi is not supported.\n");
1194                         return -1;
1195                 }
1196
1197                 if (!smi) {
1198                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1199                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1200                                 return -1;
1201                         }
1202                         smi_reset = true;
1203                 }
1204
1205                 if (pmu_have_event("msr", "aperf") &&
1206                     pmu_have_event("msr", "smi")) {
1207                         if (!force_metric_only)
1208                                 stat_config.metric_only = true;
1209                         err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1210                 } else {
1211                         fprintf(stderr, "To measure SMI cost, it needs "
1212                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1213                         parse_events_print_error(&errinfo, smi_cost_attrs);
1214                         return -1;
1215                 }
1216                 if (err) {
1217                         fprintf(stderr, "Cannot set up SMI cost events\n");
1218                         return -1;
1219                 }
1220                 return 0;
1221         }
1222
1223         if (topdown_run) {
1224                 char *str = NULL;
1225                 bool warn = false;
1226
1227                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1228                     stat_config.aggr_mode != AGGR_CORE) {
1229                         pr_err("top down event configuration requires --per-core mode\n");
1230                         return -1;
1231                 }
1232                 stat_config.aggr_mode = AGGR_CORE;
1233                 if (nr_cgroups || !target__has_cpu(&target)) {
1234                         pr_err("top down event configuration requires system-wide mode (-a)\n");
1235                         return -1;
1236                 }
1237
1238                 if (!force_metric_only)
1239                         stat_config.metric_only = true;
1240                 if (topdown_filter_events(topdown_attrs, &str,
1241                                 arch_topdown_check_group(&warn)) < 0) {
1242                         pr_err("Out of memory\n");
1243                         return -1;
1244                 }
1245                 if (topdown_attrs[0] && str) {
1246                         if (warn)
1247                                 arch_topdown_group_warn();
1248                         err = parse_events(evsel_list, str, &errinfo);
1249                         if (err) {
1250                                 fprintf(stderr,
1251                                         "Cannot set up top down events %s: %d\n",
1252                                         str, err);
1253                                 free(str);
1254                                 parse_events_print_error(&errinfo, str);
1255                                 return -1;
1256                         }
1257                 } else {
1258                         fprintf(stderr, "System does not support topdown\n");
1259                         return -1;
1260                 }
1261                 free(str);
1262         }
1263
1264         if (!evsel_list->nr_entries) {
1265                 if (target__has_cpu(&target))
1266                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1267
1268                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1269                         return -1;
1270                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1271                         if (perf_evlist__add_default_attrs(evsel_list,
1272                                                 frontend_attrs) < 0)
1273                                 return -1;
1274                 }
1275                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1276                         if (perf_evlist__add_default_attrs(evsel_list,
1277                                                 backend_attrs) < 0)
1278                                 return -1;
1279                 }
1280                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1281                         return -1;
1282         }
1283
1284         /* Detailed events get appended to the event list: */
1285
1286         if (detailed_run <  1)
1287                 return 0;
1288
1289         /* Append detailed run extra attributes: */
1290         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1291                 return -1;
1292
1293         if (detailed_run < 2)
1294                 return 0;
1295
1296         /* Append very detailed run extra attributes: */
1297         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1298                 return -1;
1299
1300         if (detailed_run < 3)
1301                 return 0;
1302
1303         /* Append very, very detailed run extra attributes: */
1304         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1305 }
1306
1307 static const char * const stat_record_usage[] = {
1308         "perf stat record [<options>]",
1309         NULL,
1310 };
1311
1312 static void init_features(struct perf_session *session)
1313 {
1314         int feat;
1315
1316         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1317                 perf_header__set_feat(&session->header, feat);
1318
1319         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1320         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1321         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1322         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1323 }
1324
1325 static int __cmd_record(int argc, const char **argv)
1326 {
1327         struct perf_session *session;
1328         struct perf_data *data = &perf_stat.data;
1329
1330         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1331                              PARSE_OPT_STOP_AT_NON_OPTION);
1332
1333         if (output_name)
1334                 data->file.path = output_name;
1335
1336         if (stat_config.run_count != 1 || forever) {
1337                 pr_err("Cannot use -r option with perf stat record.\n");
1338                 return -1;
1339         }
1340
1341         session = perf_session__new(data, false, NULL);
1342         if (session == NULL) {
1343                 pr_err("Perf session creation failed.\n");
1344                 return -1;
1345         }
1346
1347         init_features(session);
1348
1349         session->evlist   = evsel_list;
1350         perf_stat.session = session;
1351         perf_stat.record  = true;
1352         return argc;
1353 }
1354
1355 static int process_stat_round_event(struct perf_session *session,
1356                                     union perf_event *event)
1357 {
1358         struct stat_round_event *stat_round = &event->stat_round;
1359         struct perf_evsel *counter;
1360         struct timespec tsh, *ts = NULL;
1361         const char **argv = session->header.env.cmdline_argv;
1362         int argc = session->header.env.nr_cmdline;
1363
1364         evlist__for_each_entry(evsel_list, counter)
1365                 perf_stat_process_counter(&stat_config, counter);
1366
1367         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1368                 update_stats(&walltime_nsecs_stats, stat_round->time);
1369
1370         if (stat_config.interval && stat_round->time) {
1371                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1372                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1373                 ts = &tsh;
1374         }
1375
1376         print_counters(ts, argc, argv);
1377         return 0;
1378 }
1379
1380 static
1381 int process_stat_config_event(struct perf_session *session,
1382                               union perf_event *event)
1383 {
1384         struct perf_tool *tool = session->tool;
1385         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1386
1387         perf_event__read_stat_config(&stat_config, &event->stat_config);
1388
1389         if (cpu_map__empty(st->cpus)) {
1390                 if (st->aggr_mode != AGGR_UNSET)
1391                         pr_warning("warning: processing task data, aggregation mode not set\n");
1392                 return 0;
1393         }
1394
1395         if (st->aggr_mode != AGGR_UNSET)
1396                 stat_config.aggr_mode = st->aggr_mode;
1397
1398         if (perf_stat.data.is_pipe)
1399                 perf_stat_init_aggr_mode();
1400         else
1401                 perf_stat_init_aggr_mode_file(st);
1402
1403         return 0;
1404 }
1405
1406 static int set_maps(struct perf_stat *st)
1407 {
1408         if (!st->cpus || !st->threads)
1409                 return 0;
1410
1411         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1412                 return -EINVAL;
1413
1414         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1415
1416         if (perf_evlist__alloc_stats(evsel_list, true))
1417                 return -ENOMEM;
1418
1419         st->maps_allocated = true;
1420         return 0;
1421 }
1422
1423 static
1424 int process_thread_map_event(struct perf_session *session,
1425                              union perf_event *event)
1426 {
1427         struct perf_tool *tool = session->tool;
1428         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1429
1430         if (st->threads) {
1431                 pr_warning("Extra thread map event, ignoring.\n");
1432                 return 0;
1433         }
1434
1435         st->threads = thread_map__new_event(&event->thread_map);
1436         if (!st->threads)
1437                 return -ENOMEM;
1438
1439         return set_maps(st);
1440 }
1441
1442 static
1443 int process_cpu_map_event(struct perf_session *session,
1444                           union perf_event *event)
1445 {
1446         struct perf_tool *tool = session->tool;
1447         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1448         struct cpu_map *cpus;
1449
1450         if (st->cpus) {
1451                 pr_warning("Extra cpu map event, ignoring.\n");
1452                 return 0;
1453         }
1454
1455         cpus = cpu_map__new_data(&event->cpu_map.data);
1456         if (!cpus)
1457                 return -ENOMEM;
1458
1459         st->cpus = cpus;
1460         return set_maps(st);
1461 }
1462
1463 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1464 {
1465         int i;
1466
1467         config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1468         if (!config->stats)
1469                 return -1;
1470
1471         config->stats_num = nthreads;
1472
1473         for (i = 0; i < nthreads; i++)
1474                 runtime_stat__init(&config->stats[i]);
1475
1476         return 0;
1477 }
1478
1479 static void runtime_stat_delete(struct perf_stat_config *config)
1480 {
1481         int i;
1482
1483         if (!config->stats)
1484                 return;
1485
1486         for (i = 0; i < config->stats_num; i++)
1487                 runtime_stat__exit(&config->stats[i]);
1488
1489         free(config->stats);
1490 }
1491
1492 static const char * const stat_report_usage[] = {
1493         "perf stat report [<options>]",
1494         NULL,
1495 };
1496
1497 static struct perf_stat perf_stat = {
1498         .tool = {
1499                 .attr           = perf_event__process_attr,
1500                 .event_update   = perf_event__process_event_update,
1501                 .thread_map     = process_thread_map_event,
1502                 .cpu_map        = process_cpu_map_event,
1503                 .stat_config    = process_stat_config_event,
1504                 .stat           = perf_event__process_stat_event,
1505                 .stat_round     = process_stat_round_event,
1506         },
1507         .aggr_mode = AGGR_UNSET,
1508 };
1509
1510 static int __cmd_report(int argc, const char **argv)
1511 {
1512         struct perf_session *session;
1513         const struct option options[] = {
1514         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1515         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1516                      "aggregate counts per processor socket", AGGR_SOCKET),
1517         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1518                      "aggregate counts per physical processor core", AGGR_CORE),
1519         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1520                      "disable CPU count aggregation", AGGR_NONE),
1521         OPT_END()
1522         };
1523         struct stat st;
1524         int ret;
1525
1526         argc = parse_options(argc, argv, options, stat_report_usage, 0);
1527
1528         if (!input_name || !strlen(input_name)) {
1529                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1530                         input_name = "-";
1531                 else
1532                         input_name = "perf.data";
1533         }
1534
1535         perf_stat.data.file.path = input_name;
1536         perf_stat.data.mode      = PERF_DATA_MODE_READ;
1537
1538         session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1539         if (session == NULL)
1540                 return -1;
1541
1542         perf_stat.session  = session;
1543         stat_config.output = stderr;
1544         evsel_list         = session->evlist;
1545
1546         ret = perf_session__process_events(session);
1547         if (ret)
1548                 return ret;
1549
1550         perf_session__delete(session);
1551         return 0;
1552 }
1553
1554 static void setup_system_wide(int forks)
1555 {
1556         /*
1557          * Make system wide (-a) the default target if
1558          * no target was specified and one of following
1559          * conditions is met:
1560          *
1561          *   - there's no workload specified
1562          *   - there is workload specified but all requested
1563          *     events are system wide events
1564          */
1565         if (!target__none(&target))
1566                 return;
1567
1568         if (!forks)
1569                 target.system_wide = true;
1570         else {
1571                 struct perf_evsel *counter;
1572
1573                 evlist__for_each_entry(evsel_list, counter) {
1574                         if (!counter->system_wide)
1575                                 return;
1576                 }
1577
1578                 if (evsel_list->nr_entries)
1579                         target.system_wide = true;
1580         }
1581 }
1582
1583 int cmd_stat(int argc, const char **argv)
1584 {
1585         const char * const stat_usage[] = {
1586                 "perf stat [<options>] [<command>]",
1587                 NULL
1588         };
1589         int status = -EINVAL, run_idx;
1590         const char *mode;
1591         FILE *output = stderr;
1592         unsigned int interval, timeout;
1593         const char * const stat_subcommands[] = { "record", "report" };
1594
1595         setlocale(LC_ALL, "");
1596
1597         evsel_list = perf_evlist__new();
1598         if (evsel_list == NULL)
1599                 return -ENOMEM;
1600
1601         parse_events__shrink_config_terms();
1602         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1603                                         (const char **) stat_usage,
1604                                         PARSE_OPT_STOP_AT_NON_OPTION);
1605         perf_stat__collect_metric_expr(evsel_list);
1606         perf_stat__init_shadow_stats();
1607
1608         if (stat_config.csv_sep) {
1609                 stat_config.csv_output = true;
1610                 if (!strcmp(stat_config.csv_sep, "\\t"))
1611                         stat_config.csv_sep = "\t";
1612         } else
1613                 stat_config.csv_sep = DEFAULT_SEPARATOR;
1614
1615         if (argc && !strncmp(argv[0], "rec", 3)) {
1616                 argc = __cmd_record(argc, argv);
1617                 if (argc < 0)
1618                         return -1;
1619         } else if (argc && !strncmp(argv[0], "rep", 3))
1620                 return __cmd_report(argc, argv);
1621
1622         interval = stat_config.interval;
1623         timeout = stat_config.timeout;
1624
1625         /*
1626          * For record command the -o is already taken care of.
1627          */
1628         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1629                 output = NULL;
1630
1631         if (output_name && output_fd) {
1632                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1633                 parse_options_usage(stat_usage, stat_options, "o", 1);
1634                 parse_options_usage(NULL, stat_options, "log-fd", 0);
1635                 goto out;
1636         }
1637
1638         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1639                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1640                 goto out;
1641         }
1642
1643         if (stat_config.metric_only && stat_config.run_count > 1) {
1644                 fprintf(stderr, "--metric-only is not supported with -r\n");
1645                 goto out;
1646         }
1647
1648         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1649                 fprintf(stderr, "--table is only supported with -r\n");
1650                 parse_options_usage(stat_usage, stat_options, "r", 1);
1651                 parse_options_usage(NULL, stat_options, "table", 0);
1652                 goto out;
1653         }
1654
1655         if (output_fd < 0) {
1656                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1657                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1658                 goto out;
1659         }
1660
1661         if (!output) {
1662                 struct timespec tm;
1663                 mode = append_file ? "a" : "w";
1664
1665                 output = fopen(output_name, mode);
1666                 if (!output) {
1667                         perror("failed to create output file");
1668                         return -1;
1669                 }
1670                 clock_gettime(CLOCK_REALTIME, &tm);
1671                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1672         } else if (output_fd > 0) {
1673                 mode = append_file ? "a" : "w";
1674                 output = fdopen(output_fd, mode);
1675                 if (!output) {
1676                         perror("Failed opening logfd");
1677                         return -errno;
1678                 }
1679         }
1680
1681         stat_config.output = output;
1682
1683         /*
1684          * let the spreadsheet do the pretty-printing
1685          */
1686         if (stat_config.csv_output) {
1687                 /* User explicitly passed -B? */
1688                 if (big_num_opt == 1) {
1689                         fprintf(stderr, "-B option not supported with -x\n");
1690                         parse_options_usage(stat_usage, stat_options, "B", 1);
1691                         parse_options_usage(NULL, stat_options, "x", 1);
1692                         goto out;
1693                 } else /* Nope, so disable big number formatting */
1694                         stat_config.big_num = false;
1695         } else if (big_num_opt == 0) /* User passed --no-big-num */
1696                 stat_config.big_num = false;
1697
1698         setup_system_wide(argc);
1699
1700         /*
1701          * Display user/system times only for single
1702          * run and when there's specified tracee.
1703          */
1704         if ((stat_config.run_count == 1) && target__none(&target))
1705                 stat_config.ru_display = true;
1706
1707         if (stat_config.run_count < 0) {
1708                 pr_err("Run count must be a positive number\n");
1709                 parse_options_usage(stat_usage, stat_options, "r", 1);
1710                 goto out;
1711         } else if (stat_config.run_count == 0) {
1712                 forever = true;
1713                 stat_config.run_count = 1;
1714         }
1715
1716         if (stat_config.walltime_run_table) {
1717                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1718                 if (!stat_config.walltime_run) {
1719                         pr_err("failed to setup -r option");
1720                         goto out;
1721                 }
1722         }
1723
1724         if ((stat_config.aggr_mode == AGGR_THREAD) &&
1725                 !target__has_task(&target)) {
1726                 if (!target.system_wide || target.cpu_list) {
1727                         fprintf(stderr, "The --per-thread option is only "
1728                                 "available when monitoring via -p -t -a "
1729                                 "options or only --per-thread.\n");
1730                         parse_options_usage(NULL, stat_options, "p", 1);
1731                         parse_options_usage(NULL, stat_options, "t", 1);
1732                         goto out;
1733                 }
1734         }
1735
1736         /*
1737          * no_aggr, cgroup are for system-wide only
1738          * --per-thread is aggregated per thread, we dont mix it with cpu mode
1739          */
1740         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1741               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1742             !target__has_cpu(&target)) {
1743                 fprintf(stderr, "both cgroup and no-aggregation "
1744                         "modes only available in system-wide mode\n");
1745
1746                 parse_options_usage(stat_usage, stat_options, "G", 1);
1747                 parse_options_usage(NULL, stat_options, "A", 1);
1748                 parse_options_usage(NULL, stat_options, "a", 1);
1749                 goto out;
1750         }
1751
1752         if (add_default_attributes())
1753                 goto out;
1754
1755         target__validate(&target);
1756
1757         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1758                 target.per_thread = true;
1759
1760         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1761                 if (target__has_task(&target)) {
1762                         pr_err("Problems finding threads of monitor\n");
1763                         parse_options_usage(stat_usage, stat_options, "p", 1);
1764                         parse_options_usage(NULL, stat_options, "t", 1);
1765                 } else if (target__has_cpu(&target)) {
1766                         perror("failed to parse CPUs map");
1767                         parse_options_usage(stat_usage, stat_options, "C", 1);
1768                         parse_options_usage(NULL, stat_options, "a", 1);
1769                 }
1770                 goto out;
1771         }
1772
1773         /*
1774          * Initialize thread_map with comm names,
1775          * so we could print it out on output.
1776          */
1777         if (stat_config.aggr_mode == AGGR_THREAD) {
1778                 thread_map__read_comms(evsel_list->threads);
1779                 if (target.system_wide) {
1780                         if (runtime_stat_new(&stat_config,
1781                                 thread_map__nr(evsel_list->threads))) {
1782                                 goto out;
1783                         }
1784                 }
1785         }
1786
1787         if (stat_config.times && interval)
1788                 interval_count = true;
1789         else if (stat_config.times && !interval) {
1790                 pr_err("interval-count option should be used together with "
1791                                 "interval-print.\n");
1792                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1793                 parse_options_usage(stat_usage, stat_options, "I", 1);
1794                 goto out;
1795         }
1796
1797         if (timeout && timeout < 100) {
1798                 if (timeout < 10) {
1799                         pr_err("timeout must be >= 10ms.\n");
1800                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
1801                         goto out;
1802                 } else
1803                         pr_warning("timeout < 100ms. "
1804                                    "The overhead percentage could be high in some cases. "
1805                                    "Please proceed with caution.\n");
1806         }
1807         if (timeout && interval) {
1808                 pr_err("timeout option is not supported with interval-print.\n");
1809                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1810                 parse_options_usage(stat_usage, stat_options, "I", 1);
1811                 goto out;
1812         }
1813
1814         if (perf_evlist__alloc_stats(evsel_list, interval))
1815                 goto out;
1816
1817         if (perf_stat_init_aggr_mode())
1818                 goto out;
1819
1820         /*
1821          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1822          * while avoiding that older tools show confusing messages.
1823          *
1824          * However for pipe sessions we need to keep it zero,
1825          * because script's perf_evsel__check_attr is triggered
1826          * by attr->sample_type != 0, and we can't run it on
1827          * stat sessions.
1828          */
1829         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1830
1831         /*
1832          * We dont want to block the signals - that would cause
1833          * child tasks to inherit that and Ctrl-C would not work.
1834          * What we want is for Ctrl-C to work in the exec()-ed
1835          * task, but being ignored by perf stat itself:
1836          */
1837         atexit(sig_atexit);
1838         if (!forever)
1839                 signal(SIGINT,  skip_signal);
1840         signal(SIGCHLD, skip_signal);
1841         signal(SIGALRM, skip_signal);
1842         signal(SIGABRT, skip_signal);
1843
1844         status = 0;
1845         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1846                 if (stat_config.run_count != 1 && verbose > 0)
1847                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1848                                 run_idx + 1);
1849
1850                 status = run_perf_stat(argc, argv, run_idx);
1851                 if (forever && status != -1) {
1852                         print_counters(NULL, argc, argv);
1853                         perf_stat__reset_stats();
1854                 }
1855         }
1856
1857         if (!forever && status != -1 && !interval)
1858                 print_counters(NULL, argc, argv);
1859
1860         if (STAT_RECORD) {
1861                 /*
1862                  * We synthesize the kernel mmap record just so that older tools
1863                  * don't emit warnings about not being able to resolve symbols
1864                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1865                  * a saner message about no samples being in the perf.data file.
1866                  *
1867                  * This also serves to suppress a warning about f_header.data.size == 0
1868                  * in header.c at the moment 'perf stat record' gets introduced, which
1869                  * is not really needed once we start adding the stat specific PERF_RECORD_
1870                  * records, but the need to suppress the kptr_restrict messages in older
1871                  * tools remain  -acme
1872                  */
1873                 int fd = perf_data__fd(&perf_stat.data);
1874                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1875                                                              process_synthesized_event,
1876                                                              &perf_stat.session->machines.host);
1877                 if (err) {
1878                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1879                                    "older tools may produce warnings about this file\n.");
1880                 }
1881
1882                 if (!interval) {
1883                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1884                                 pr_err("failed to write stat round event\n");
1885                 }
1886
1887                 if (!perf_stat.data.is_pipe) {
1888                         perf_stat.session->header.data_size += perf_stat.bytes_written;
1889                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1890                 }
1891
1892                 perf_session__delete(perf_stat.session);
1893         }
1894
1895         perf_stat__exit_aggr_mode();
1896         perf_evlist__free_stats(evsel_list);
1897 out:
1898         free(stat_config.walltime_run);
1899
1900         if (smi_cost && smi_reset)
1901                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1902
1903         perf_evlist__delete(evsel_list);
1904
1905         runtime_stat_delete(&stat_config);
1906
1907         return status;
1908 }