tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/evlist.h"
   6 #include "util/cache.h"
   7 #include "util/evsel.h"
   8 #include "util/symbol.h"
   9 #include "util/thread.h"
  10 #include "util/header.h"
  11 #include "util/session.h"
  12 #include "util/tool.h"
  13 #include "util/cloexec.h"
  14 #include "util/thread_map.h"
  15 #include "util/color.h"
  16
  17 #include <subcmd/parse-options.h>
  18 #include "util/trace-event.h"
  19
  20 #include "util/debug.h"
  21
  22 #include <sys/prctl.h>
  23 #include <sys/resource.h>
  24
  25 #include <semaphore.h>
  26 #include <pthread.h>
  27 #include <math.h>
  28 #include <api/fs/fs.h>
  29 #include <linux/time64.h>
  30
  31 #define PR_SET_NAME             15               /* Set process name */
  32 #define MAX_CPUS                4096
  33 #define COMM_LEN                20
  34 #define SYM_LEN                 129
  35 #define MAX_PID                 1024000
  36
  37 struct sched_atom;
  38
  39 struct task_desc {
  40         unsigned long           nr;
  41         unsigned long           pid;
  42         char                    comm[COMM_LEN];
  43
  44         unsigned long           nr_events;
  45         unsigned long           curr_event;
  46         struct sched_atom       **atoms;
  47
  48         pthread_t               thread;
  49         sem_t                   sleep_sem;
  50
  51         sem_t                   ready_for_work;
  52         sem_t                   work_done_sem;
  53
  54         u64                     cpu_usage;
  55 };
  56
  57 enum sched_event_type {
  58         SCHED_EVENT_RUN,
  59         SCHED_EVENT_SLEEP,
  60         SCHED_EVENT_WAKEUP,
  61         SCHED_EVENT_MIGRATION,
  62 };
  63
  64 struct sched_atom {
  65         enum sched_event_type   type;
  66         int                     specific_wait;
  67         u64                     timestamp;
  68         u64                     duration;
  69         unsigned long           nr;
  70         sem_t                   *wait_sem;
  71         struct task_desc        *wakee;
  72 };
  73
  74 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  75
  76 enum thread_state {
  77         THREAD_SLEEPING = 0,
  78         THREAD_WAIT_CPU,
  79         THREAD_SCHED_IN,
  80         THREAD_IGNORE
  81 };
  82
  83 struct work_atom {
  84         struct list_head        list;
  85         enum thread_state       state;
  86         u64                     sched_out_time;
  87         u64                     wake_up_time;
  88         u64                     sched_in_time;
  89         u64                     runtime;
  90 };
  91
  92 struct work_atoms {
  93         struct list_head        work_list;
  94         struct thread           *thread;
  95         struct rb_node          node;
  96         u64                     max_lat;
  97         u64                     max_lat_at;
  98         u64                     total_lat;
  99         u64                     nb_atoms;
 100         u64                     total_runtime;
 101         int                     num_merged;
 102 };
 103
 104 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 105
 106 struct perf_sched;
 107
 108 struct trace_sched_handler {
 109         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 110                             struct perf_sample *sample, struct machine *machine);
 111
 112         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 113                              struct perf_sample *sample, struct machine *machine);
 114
 115         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 116                             struct perf_sample *sample, struct machine *machine);
 117
 118         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 119         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 120                           struct machine *machine);
 121
 122         int (*migrate_task_event)(struct perf_sched *sched,
 123                                   struct perf_evsel *evsel,
 124                                   struct perf_sample *sample,
 125                                   struct machine *machine);
 126 };
 127
 128 #define COLOR_PIDS PERF_COLOR_BLUE
 129 #define COLOR_CPUS PERF_COLOR_BG_RED
 130
 131 struct perf_sched_map {
 132         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 133         int                     *comp_cpus;
 134         bool                     comp;
 135         struct thread_map       *color_pids;
 136         const char              *color_pids_str;
 137         struct cpu_map          *color_cpus;
 138         const char              *color_cpus_str;
 139         struct cpu_map          *cpus;
 140         const char              *cpus_str;
 141 };
 142
 143 struct perf_sched {
 144         struct perf_tool tool;
 145         const char       *sort_order;
 146         unsigned long    nr_tasks;
 147         struct task_desc **pid_to_task;
 148         struct task_desc **tasks;
 149         const struct trace_sched_handler *tp_handler;
 150         pthread_mutex_t  start_work_mutex;
 151         pthread_mutex_t  work_done_wait_mutex;
 152         int              profile_cpu;
 153 /*
 154  * Track the current task - that way we can know whether there's any
 155  * weird events, such as a task being switched away that is not current.
 156  */
 157         int              max_cpu;
 158         u32              curr_pid[MAX_CPUS];
 159         struct thread    *curr_thread[MAX_CPUS];
 160         char             next_shortname1;
 161         char             next_shortname2;
 162         unsigned int     replay_repeat;
 163         unsigned long    nr_run_events;
 164         unsigned long    nr_sleep_events;
 165         unsigned long    nr_wakeup_events;
 166         unsigned long    nr_sleep_corrections;
 167         unsigned long    nr_run_events_optimized;
 168         unsigned long    targetless_wakeups;
 169         unsigned long    multitarget_wakeups;
 170         unsigned long    nr_runs;
 171         unsigned long    nr_timestamps;
 172         unsigned long    nr_unordered_timestamps;
 173         unsigned long    nr_context_switch_bugs;
 174         unsigned long    nr_events;
 175         unsigned long    nr_lost_chunks;
 176         unsigned long    nr_lost_events;
 177         u64              run_measurement_overhead;
 178         u64              sleep_measurement_overhead;
 179         u64              start_time;
 180         u64              cpu_usage;
 181         u64              runavg_cpu_usage;
 182         u64              parent_cpu_usage;
 183         u64              runavg_parent_cpu_usage;
 184         u64              sum_runtime;
 185         u64              sum_fluct;
 186         u64              run_avg;
 187         u64              all_runtime;
 188         u64              all_count;
 189         u64              cpu_last_switched[MAX_CPUS];
 190         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 191         struct list_head sort_list, cmp_pid;
 192         bool force;
 193         bool skip_merge;
 194         struct perf_sched_map map;
 195 };
 196
 197 static u64 get_nsecs(void)
 198 {
 199         struct timespec ts;
 200
 201         clock_gettime(CLOCK_MONOTONIC, &ts);
 202
 203         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 204 }
 205
 206 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 207 {
 208         u64 T0 = get_nsecs(), T1;
 209
 210         do {
 211                 T1 = get_nsecs();
 212         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 213 }
 214
 215 static void sleep_nsecs(u64 nsecs)
 216 {
 217         struct timespec ts;
 218
 219         ts.tv_nsec = nsecs % 999999999;
 220         ts.tv_sec = nsecs / 999999999;
 221
 222         nanosleep(&ts, NULL);
 223 }
 224
 225 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 226 {
 227         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 228         int i;
 229
 230         for (i = 0; i < 10; i++) {
 231                 T0 = get_nsecs();
 232                 burn_nsecs(sched, 0);
 233                 T1 = get_nsecs();
 234                 delta = T1-T0;
 235                 min_delta = min(min_delta, delta);
 236         }
 237         sched->run_measurement_overhead = min_delta;
 238
 239         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 240 }
 241
 242 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 243 {
 244         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 245         int i;
 246
 247         for (i = 0; i < 10; i++) {
 248                 T0 = get_nsecs();
 249                 sleep_nsecs(10000);
 250                 T1 = get_nsecs();
 251                 delta = T1-T0;
 252                 min_delta = min(min_delta, delta);
 253         }
 254         min_delta -= 10000;
 255         sched->sleep_measurement_overhead = min_delta;
 256
 257         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 258 }
 259
 260 static struct sched_atom *
 261 get_new_event(struct task_desc *task, u64 timestamp)
 262 {
 263         struct sched_atom *event = zalloc(sizeof(*event));
 264         unsigned long idx = task->nr_events;
 265         size_t size;
 266
 267         event->timestamp = timestamp;
 268         event->nr = idx;
 269
 270         task->nr_events++;
 271         size = sizeof(struct sched_atom *) * task->nr_events;
 272         task->atoms = realloc(task->atoms, size);
 273         BUG_ON(!task->atoms);
 274
 275         task->atoms[idx] = event;
 276
 277         return event;
 278 }
 279
 280 static struct sched_atom *last_event(struct task_desc *task)
 281 {
 282         if (!task->nr_events)
 283                 return NULL;
 284
 285         return task->atoms[task->nr_events - 1];
 286 }
 287
 288 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 289                                 u64 timestamp, u64 duration)
 290 {
 291         struct sched_atom *event, *curr_event = last_event(task);
 292
 293         /*
 294          * optimize an existing RUN event by merging this one
 295          * to it:
 296          */
 297         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 298                 sched->nr_run_events_optimized++;
 299                 curr_event->duration += duration;
 300                 return;
 301         }
 302
 303         event = get_new_event(task, timestamp);
 304
 305         event->type = SCHED_EVENT_RUN;
 306         event->duration = duration;
 307
 308         sched->nr_run_events++;
 309 }
 310
 311 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 312                                    u64 timestamp, struct task_desc *wakee)
 313 {
 314         struct sched_atom *event, *wakee_event;
 315
 316         event = get_new_event(task, timestamp);
 317         event->type = SCHED_EVENT_WAKEUP;
 318         event->wakee = wakee;
 319
 320         wakee_event = last_event(wakee);
 321         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 322                 sched->targetless_wakeups++;
 323                 return;
 324         }
 325         if (wakee_event->wait_sem) {
 326                 sched->multitarget_wakeups++;
 327                 return;
 328         }
 329
 330         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 331         sem_init(wakee_event->wait_sem, 0, 0);
 332         wakee_event->specific_wait = 1;
 333         event->wait_sem = wakee_event->wait_sem;
 334
 335         sched->nr_wakeup_events++;
 336 }
 337
 338 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 339                                   u64 timestamp, u64 task_state __maybe_unused)
 340 {
 341         struct sched_atom *event = get_new_event(task, timestamp);
 342
 343         event->type = SCHED_EVENT_SLEEP;
 344
 345         sched->nr_sleep_events++;
 346 }
 347
 348 static struct task_desc *register_pid(struct perf_sched *sched,
 349                                       unsigned long pid, const char *comm)
 350 {
 351         struct task_desc *task;
 352         static int pid_max;
 353
 354         if (sched->pid_to_task == NULL) {
 355                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 356                         pid_max = MAX_PID;
 357                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 358         }
 359         if (pid >= (unsigned long)pid_max) {
 360                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 361                         sizeof(struct task_desc *))) == NULL);
 362                 while (pid >= (unsigned long)pid_max)
 363                         sched->pid_to_task[pid_max++] = NULL;
 364         }
 365
 366         task = sched->pid_to_task[pid];
 367
 368         if (task)
 369                 return task;
 370
 371         task = zalloc(sizeof(*task));
 372         task->pid = pid;
 373         task->nr = sched->nr_tasks;
 374         strcpy(task->comm, comm);
 375         /*
 376          * every task starts in sleeping state - this gets ignored
 377          * if there's no wakeup pointing to this sleep state:
 378          */
 379         add_sched_event_sleep(sched, task, 0, 0);
 380
 381         sched->pid_to_task[pid] = task;
 382         sched->nr_tasks++;
 383         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 384         BUG_ON(!sched->tasks);
 385         sched->tasks[task->nr] = task;
 386
 387         if (verbose)
 388                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 389
 390         return task;
 391 }
 392
 393
 394 static void print_task_traces(struct perf_sched *sched)
 395 {
 396         struct task_desc *task;
 397         unsigned long i;
 398
 399         for (i = 0; i < sched->nr_tasks; i++) {
 400                 task = sched->tasks[i];
 401                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 402                         task->nr, task->comm, task->pid, task->nr_events);
 403         }
 404 }
 405
 406 static void add_cross_task_wakeups(struct perf_sched *sched)
 407 {
 408         struct task_desc *task1, *task2;
 409         unsigned long i, j;
 410
 411         for (i = 0; i < sched->nr_tasks; i++) {
 412                 task1 = sched->tasks[i];
 413                 j = i + 1;
 414                 if (j == sched->nr_tasks)
 415                         j = 0;
 416                 task2 = sched->tasks[j];
 417                 add_sched_event_wakeup(sched, task1, 0, task2);
 418         }
 419 }
 420
 421 static void perf_sched__process_event(struct perf_sched *sched,
 422                                       struct sched_atom *atom)
 423 {
 424         int ret = 0;
 425
 426         switch (atom->type) {
 427                 case SCHED_EVENT_RUN:
 428                         burn_nsecs(sched, atom->duration);
 429                         break;
 430                 case SCHED_EVENT_SLEEP:
 431                         if (atom->wait_sem)
 432                                 ret = sem_wait(atom->wait_sem);
 433                         BUG_ON(ret);
 434                         break;
 435                 case SCHED_EVENT_WAKEUP:
 436                         if (atom->wait_sem)
 437                                 ret = sem_post(atom->wait_sem);
 438                         BUG_ON(ret);
 439                         break;
 440                 case SCHED_EVENT_MIGRATION:
 441                         break;
 442                 default:
 443                         BUG_ON(1);
 444         }
 445 }
 446
 447 static u64 get_cpu_usage_nsec_parent(void)
 448 {
 449         struct rusage ru;
 450         u64 sum;
 451         int err;
 452
 453         err = getrusage(RUSAGE_SELF, &ru);
 454         BUG_ON(err);
 455
 456         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 457         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 458
 459         return sum;
 460 }
 461
 462 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 463 {
 464         struct perf_event_attr attr;
 465         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 466         int fd;
 467         struct rlimit limit;
 468         bool need_privilege = false;
 469
 470         memset(&attr, 0, sizeof(attr));
 471
 472         attr.type = PERF_TYPE_SOFTWARE;
 473         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 474
 475 force_again:
 476         fd = sys_perf_event_open(&attr, 0, -1, -1,
 477                                  perf_event_open_cloexec_flag());
 478
 479         if (fd < 0) {
 480                 if (errno == EMFILE) {
 481                         if (sched->force) {
 482                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 483                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 484                                 if (limit.rlim_cur > limit.rlim_max) {
 485                                         limit.rlim_max = limit.rlim_cur;
 486                                         need_privilege = true;
 487                                 }
 488                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 489                                         if (need_privilege && errno == EPERM)
 490                                                 strcpy(info, "Need privilege\n");
 491                                 } else
 492                                         goto force_again;
 493                         } else
 494                                 strcpy(info, "Have a try with -f option\n");
 495                 }
 496                 pr_err("Error: sys_perf_event_open() syscall returned "
 497                        "with %d (%s)\n%s", fd,
 498                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 499                 exit(EXIT_FAILURE);
 500         }
 501         return fd;
 502 }
 503
 504 static u64 get_cpu_usage_nsec_self(int fd)
 505 {
 506         u64 runtime;
 507         int ret;
 508
 509         ret = read(fd, &runtime, sizeof(runtime));
 510         BUG_ON(ret != sizeof(runtime));
 511
 512         return runtime;
 513 }
 514
 515 struct sched_thread_parms {
 516         struct task_desc  *task;
 517         struct perf_sched *sched;
 518         int fd;
 519 };
 520
 521 static void *thread_func(void *ctx)
 522 {
 523         struct sched_thread_parms *parms = ctx;
 524         struct task_desc *this_task = parms->task;
 525         struct perf_sched *sched = parms->sched;
 526         u64 cpu_usage_0, cpu_usage_1;
 527         unsigned long i, ret;
 528         char comm2[22];
 529         int fd = parms->fd;
 530
 531         zfree(&parms);
 532
 533         sprintf(comm2, ":%s", this_task->comm);
 534         prctl(PR_SET_NAME, comm2);
 535         if (fd < 0)
 536                 return NULL;
 537 again:
 538         ret = sem_post(&this_task->ready_for_work);
 539         BUG_ON(ret);
 540         ret = pthread_mutex_lock(&sched->start_work_mutex);
 541         BUG_ON(ret);
 542         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 543         BUG_ON(ret);
 544
 545         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 546
 547         for (i = 0; i < this_task->nr_events; i++) {
 548                 this_task->curr_event = i;
 549                 perf_sched__process_event(sched, this_task->atoms[i]);
 550         }
 551
 552         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 553         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 554         ret = sem_post(&this_task->work_done_sem);
 555         BUG_ON(ret);
 556
 557         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 558         BUG_ON(ret);
 559         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 560         BUG_ON(ret);
 561
 562         goto again;
 563 }
 564
 565 static void create_tasks(struct perf_sched *sched)
 566 {
 567         struct task_desc *task;
 568         pthread_attr_t attr;
 569         unsigned long i;
 570         int err;
 571
 572         err = pthread_attr_init(&attr);
 573         BUG_ON(err);
 574         err = pthread_attr_setstacksize(&attr,
 575                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 576         BUG_ON(err);
 577         err = pthread_mutex_lock(&sched->start_work_mutex);
 578         BUG_ON(err);
 579         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 580         BUG_ON(err);
 581         for (i = 0; i < sched->nr_tasks; i++) {
 582                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 583                 BUG_ON(parms == NULL);
 584                 parms->task = task = sched->tasks[i];
 585                 parms->sched = sched;
 586                 parms->fd = self_open_counters(sched, i);
 587                 sem_init(&task->sleep_sem, 0, 0);
 588                 sem_init(&task->ready_for_work, 0, 0);
 589                 sem_init(&task->work_done_sem, 0, 0);
 590                 task->curr_event = 0;
 591                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 592                 BUG_ON(err);
 593         }
 594 }
 595
 596 static void wait_for_tasks(struct perf_sched *sched)
 597 {
 598         u64 cpu_usage_0, cpu_usage_1;
 599         struct task_desc *task;
 600         unsigned long i, ret;
 601
 602         sched->start_time = get_nsecs();
 603         sched->cpu_usage = 0;
 604         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 605
 606         for (i = 0; i < sched->nr_tasks; i++) {
 607                 task = sched->tasks[i];
 608                 ret = sem_wait(&task->ready_for_work);
 609                 BUG_ON(ret);
 610                 sem_init(&task->ready_for_work, 0, 0);
 611         }
 612         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 613         BUG_ON(ret);
 614
 615         cpu_usage_0 = get_cpu_usage_nsec_parent();
 616
 617         pthread_mutex_unlock(&sched->start_work_mutex);
 618
 619         for (i = 0; i < sched->nr_tasks; i++) {
 620                 task = sched->tasks[i];
 621                 ret = sem_wait(&task->work_done_sem);
 622                 BUG_ON(ret);
 623                 sem_init(&task->work_done_sem, 0, 0);
 624                 sched->cpu_usage += task->cpu_usage;
 625                 task->cpu_usage = 0;
 626         }
 627
 628         cpu_usage_1 = get_cpu_usage_nsec_parent();
 629         if (!sched->runavg_cpu_usage)
 630                 sched->runavg_cpu_usage = sched->cpu_usage;
 631         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 632
 633         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 634         if (!sched->runavg_parent_cpu_usage)
 635                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 636         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 637                                          sched->parent_cpu_usage)/sched->replay_repeat;
 638
 639         ret = pthread_mutex_lock(&sched->start_work_mutex);
 640         BUG_ON(ret);
 641
 642         for (i = 0; i < sched->nr_tasks; i++) {
 643                 task = sched->tasks[i];
 644                 sem_init(&task->sleep_sem, 0, 0);
 645                 task->curr_event = 0;
 646         }
 647 }
 648
 649 static void run_one_test(struct perf_sched *sched)
 650 {
 651         u64 T0, T1, delta, avg_delta, fluct;
 652
 653         T0 = get_nsecs();
 654         wait_for_tasks(sched);
 655         T1 = get_nsecs();
 656
 657         delta = T1 - T0;
 658         sched->sum_runtime += delta;
 659         sched->nr_runs++;
 660
 661         avg_delta = sched->sum_runtime / sched->nr_runs;
 662         if (delta < avg_delta)
 663                 fluct = avg_delta - delta;
 664         else
 665                 fluct = delta - avg_delta;
 666         sched->sum_fluct += fluct;
 667         if (!sched->run_avg)
 668                 sched->run_avg = delta;
 669         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 670
 671         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 672
 673         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 674
 675         printf("cpu: %0.2f / %0.2f",
 676                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 677
 678 #if 0
 679         /*
 680          * rusage statistics done by the parent, these are less
 681          * accurate than the sched->sum_exec_runtime based statistics:
 682          */
 683         printf(" [%0.2f / %0.2f]",
 684                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 685                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 686 #endif
 687
 688         printf("\n");
 689
 690         if (sched->nr_sleep_corrections)
 691                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 692         sched->nr_sleep_corrections = 0;
 693 }
 694
 695 static void test_calibrations(struct perf_sched *sched)
 696 {
 697         u64 T0, T1;
 698
 699         T0 = get_nsecs();
 700         burn_nsecs(sched, NSEC_PER_MSEC);
 701         T1 = get_nsecs();
 702
 703         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 704
 705         T0 = get_nsecs();
 706         sleep_nsecs(NSEC_PER_MSEC);
 707         T1 = get_nsecs();
 708
 709         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 710 }
 711
 712 static int
 713 replay_wakeup_event(struct perf_sched *sched,
 714                     struct perf_evsel *evsel, struct perf_sample *sample,
 715                     struct machine *machine __maybe_unused)
 716 {
 717         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 718         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 719         struct task_desc *waker, *wakee;
 720
 721         if (verbose) {
 722                 printf("sched_wakeup event %p\n", evsel);
 723
 724                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 725         }
 726
 727         waker = register_pid(sched, sample->tid, "<unknown>");
 728         wakee = register_pid(sched, pid, comm);
 729
 730         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 731         return 0;
 732 }
 733
 734 static int replay_switch_event(struct perf_sched *sched,
 735                                struct perf_evsel *evsel,
 736                                struct perf_sample *sample,
 737                                struct machine *machine __maybe_unused)
 738 {
 739         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 740                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 741         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 742                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 743         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 744         struct task_desc *prev, __maybe_unused *next;
 745         u64 timestamp0, timestamp = sample->time;
 746         int cpu = sample->cpu;
 747         s64 delta;
 748
 749         if (verbose)
 750                 printf("sched_switch event %p\n", evsel);
 751
 752         if (cpu >= MAX_CPUS || cpu < 0)
 753                 return 0;
 754
 755         timestamp0 = sched->cpu_last_switched[cpu];
 756         if (timestamp0)
 757                 delta = timestamp - timestamp0;
 758         else
 759                 delta = 0;
 760
 761         if (delta < 0) {
 762                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 763                 return -1;
 764         }
 765
 766         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 767                  prev_comm, prev_pid, next_comm, next_pid, delta);
 768
 769         prev = register_pid(sched, prev_pid, prev_comm);
 770         next = register_pid(sched, next_pid, next_comm);
 771
 772         sched->cpu_last_switched[cpu] = timestamp;
 773
 774         add_sched_event_run(sched, prev, timestamp, delta);
 775         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 776
 777         return 0;
 778 }
 779
 780 static int replay_fork_event(struct perf_sched *sched,
 781                              union perf_event *event,
 782                              struct machine *machine)
 783 {
 784         struct thread *child, *parent;
 785
 786         child = machine__findnew_thread(machine, event->fork.pid,
 787                                         event->fork.tid);
 788         parent = machine__findnew_thread(machine, event->fork.ppid,
 789                                          event->fork.ptid);
 790
 791         if (child == NULL || parent == NULL) {
 792                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 793                                  child, parent);
 794                 goto out_put;
 795         }
 796
 797         if (verbose) {
 798                 printf("fork event\n");
 799                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 800                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 801         }
 802
 803         register_pid(sched, parent->tid, thread__comm_str(parent));
 804         register_pid(sched, child->tid, thread__comm_str(child));
 805 out_put:
 806         thread__put(child);
 807         thread__put(parent);
 808         return 0;
 809 }
 810
 811 struct sort_dimension {
 812         const char              *name;
 813         sort_fn_t               cmp;
 814         struct list_head        list;
 815 };
 816
 817 static int
 818 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 819 {
 820         struct sort_dimension *sort;
 821         int ret = 0;
 822
 823         BUG_ON(list_empty(list));
 824
 825         list_for_each_entry(sort, list, list) {
 826                 ret = sort->cmp(l, r);
 827                 if (ret)
 828                         return ret;
 829         }
 830
 831         return ret;
 832 }
 833
 834 static struct work_atoms *
 835 thread_atoms_search(struct rb_root *root, struct thread *thread,
 836                          struct list_head *sort_list)
 837 {
 838         struct rb_node *node = root->rb_node;
 839         struct work_atoms key = { .thread = thread };
 840
 841         while (node) {
 842                 struct work_atoms *atoms;
 843                 int cmp;
 844
 845                 atoms = container_of(node, struct work_atoms, node);
 846
 847                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 848                 if (cmp > 0)
 849                         node = node->rb_left;
 850                 else if (cmp < 0)
 851                         node = node->rb_right;
 852                 else {
 853                         BUG_ON(thread != atoms->thread);
 854                         return atoms;
 855                 }
 856         }
 857         return NULL;
 858 }
 859
 860 static void
 861 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 862                          struct list_head *sort_list)
 863 {
 864         struct rb_node **new = &(root->rb_node), *parent = NULL;
 865
 866         while (*new) {
 867                 struct work_atoms *this;
 868                 int cmp;
 869
 870                 this = container_of(*new, struct work_atoms, node);
 871                 parent = *new;
 872
 873                 cmp = thread_lat_cmp(sort_list, data, this);
 874
 875                 if (cmp > 0)
 876                         new = &((*new)->rb_left);
 877                 else
 878                         new = &((*new)->rb_right);
 879         }
 880
 881         rb_link_node(&data->node, parent, new);
 882         rb_insert_color(&data->node, root);
 883 }
 884
 885 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 886 {
 887         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 888         if (!atoms) {
 889                 pr_err("No memory at %s\n", __func__);
 890                 return -1;
 891         }
 892
 893         atoms->thread = thread__get(thread);
 894         INIT_LIST_HEAD(&atoms->work_list);
 895         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 896         return 0;
 897 }
 898
 899 static char sched_out_state(u64 prev_state)
 900 {
 901         const char *str = TASK_STATE_TO_CHAR_STR;
 902
 903         return str[prev_state];
 904 }
 905
 906 static int
 907 add_sched_out_event(struct work_atoms *atoms,
 908                     char run_state,
 909                     u64 timestamp)
 910 {
 911         struct work_atom *atom = zalloc(sizeof(*atom));
 912         if (!atom) {
 913                 pr_err("Non memory at %s", __func__);
 914                 return -1;
 915         }
 916
 917         atom->sched_out_time = timestamp;
 918
 919         if (run_state == 'R') {
 920                 atom->state = THREAD_WAIT_CPU;
 921                 atom->wake_up_time = atom->sched_out_time;
 922         }
 923
 924         list_add_tail(&atom->list, &atoms->work_list);
 925         return 0;
 926 }
 927
 928 static void
 929 add_runtime_event(struct work_atoms *atoms, u64 delta,
 930                   u64 timestamp __maybe_unused)
 931 {
 932         struct work_atom *atom;
 933
 934         BUG_ON(list_empty(&atoms->work_list));
 935
 936         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 937
 938         atom->runtime += delta;
 939         atoms->total_runtime += delta;
 940 }
 941
 942 static void
 943 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 944 {
 945         struct work_atom *atom;
 946         u64 delta;
 947
 948         if (list_empty(&atoms->work_list))
 949                 return;
 950
 951         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 952
 953         if (atom->state != THREAD_WAIT_CPU)
 954                 return;
 955
 956         if (timestamp < atom->wake_up_time) {
 957                 atom->state = THREAD_IGNORE;
 958                 return;
 959         }
 960
 961         atom->state = THREAD_SCHED_IN;
 962         atom->sched_in_time = timestamp;
 963
 964         delta = atom->sched_in_time - atom->wake_up_time;
 965         atoms->total_lat += delta;
 966         if (delta > atoms->max_lat) {
 967                 atoms->max_lat = delta;
 968                 atoms->max_lat_at = timestamp;
 969         }
 970         atoms->nb_atoms++;
 971 }
 972
 973 static int latency_switch_event(struct perf_sched *sched,
 974                                 struct perf_evsel *evsel,
 975                                 struct perf_sample *sample,
 976                                 struct machine *machine)
 977 {
 978         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 979                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 980         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 981         struct work_atoms *out_events, *in_events;
 982         struct thread *sched_out, *sched_in;
 983         u64 timestamp0, timestamp = sample->time;
 984         int cpu = sample->cpu, err = -1;
 985         s64 delta;
 986
 987         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 988
 989         timestamp0 = sched->cpu_last_switched[cpu];
 990         sched->cpu_last_switched[cpu] = timestamp;
 991         if (timestamp0)
 992                 delta = timestamp - timestamp0;
 993         else
 994                 delta = 0;
 995
 996         if (delta < 0) {
 997                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 998                 return -1;
 999         }
1000
1001         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1002         sched_in = machine__findnew_thread(machine, -1, next_pid);
1003         if (sched_out == NULL || sched_in == NULL)
1004                 goto out_put;
1005
1006         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1007         if (!out_events) {
1008                 if (thread_atoms_insert(sched, sched_out))
1009                         goto out_put;
1010                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1011                 if (!out_events) {
1012                         pr_err("out-event: Internal tree error");
1013                         goto out_put;
1014                 }
1015         }
1016         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1017                 return -1;
1018
1019         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1020         if (!in_events) {
1021                 if (thread_atoms_insert(sched, sched_in))
1022                         goto out_put;
1023                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1024                 if (!in_events) {
1025                         pr_err("in-event: Internal tree error");
1026                         goto out_put;
1027                 }
1028                 /*
1029                  * Take came in we have not heard about yet,
1030                  * add in an initial atom in runnable state:
1031                  */
1032                 if (add_sched_out_event(in_events, 'R', timestamp))
1033                         goto out_put;
1034         }
1035         add_sched_in_event(in_events, timestamp);
1036         err = 0;
1037 out_put:
1038         thread__put(sched_out);
1039         thread__put(sched_in);
1040         return err;
1041 }
1042
1043 static int latency_runtime_event(struct perf_sched *sched,
1044                                  struct perf_evsel *evsel,
1045                                  struct perf_sample *sample,
1046                                  struct machine *machine)
1047 {
1048         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1049         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1050         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1051         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1052         u64 timestamp = sample->time;
1053         int cpu = sample->cpu, err = -1;
1054
1055         if (thread == NULL)
1056                 return -1;
1057
1058         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1059         if (!atoms) {
1060                 if (thread_atoms_insert(sched, thread))
1061                         goto out_put;
1062                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1063                 if (!atoms) {
1064                         pr_err("in-event: Internal tree error");
1065                         goto out_put;
1066                 }
1067                 if (add_sched_out_event(atoms, 'R', timestamp))
1068                         goto out_put;
1069         }
1070
1071         add_runtime_event(atoms, runtime, timestamp);
1072         err = 0;
1073 out_put:
1074         thread__put(thread);
1075         return err;
1076 }
1077
1078 static int latency_wakeup_event(struct perf_sched *sched,
1079                                 struct perf_evsel *evsel,
1080                                 struct perf_sample *sample,
1081                                 struct machine *machine)
1082 {
1083         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1084         struct work_atoms *atoms;
1085         struct work_atom *atom;
1086         struct thread *wakee;
1087         u64 timestamp = sample->time;
1088         int err = -1;
1089
1090         wakee = machine__findnew_thread(machine, -1, pid);
1091         if (wakee == NULL)
1092                 return -1;
1093         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1094         if (!atoms) {
1095                 if (thread_atoms_insert(sched, wakee))
1096                         goto out_put;
1097                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1098                 if (!atoms) {
1099                         pr_err("wakeup-event: Internal tree error");
1100                         goto out_put;
1101                 }
1102                 if (add_sched_out_event(atoms, 'S', timestamp))
1103                         goto out_put;
1104         }
1105
1106         BUG_ON(list_empty(&atoms->work_list));
1107
1108         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1109
1110         /*
1111          * As we do not guarantee the wakeup event happens when
1112          * task is out of run queue, also may happen when task is
1113          * on run queue and wakeup only change ->state to TASK_RUNNING,
1114          * then we should not set the ->wake_up_time when wake up a
1115          * task which is on run queue.
1116          *
1117          * You WILL be missing events if you've recorded only
1118          * one CPU, or are only looking at only one, so don't
1119          * skip in this case.
1120          */
1121         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1122                 goto out_ok;
1123
1124         sched->nr_timestamps++;
1125         if (atom->sched_out_time > timestamp) {
1126                 sched->nr_unordered_timestamps++;
1127                 goto out_ok;
1128         }
1129
1130         atom->state = THREAD_WAIT_CPU;
1131         atom->wake_up_time = timestamp;
1132 out_ok:
1133         err = 0;
1134 out_put:
1135         thread__put(wakee);
1136         return err;
1137 }
1138
1139 static int latency_migrate_task_event(struct perf_sched *sched,
1140                                       struct perf_evsel *evsel,
1141                                       struct perf_sample *sample,
1142                                       struct machine *machine)
1143 {
1144         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1145         u64 timestamp = sample->time;
1146         struct work_atoms *atoms;
1147         struct work_atom *atom;
1148         struct thread *migrant;
1149         int err = -1;
1150
1151         /*
1152          * Only need to worry about migration when profiling one CPU.
1153          */
1154         if (sched->profile_cpu == -1)
1155                 return 0;
1156
1157         migrant = machine__findnew_thread(machine, -1, pid);
1158         if (migrant == NULL)
1159                 return -1;
1160         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1161         if (!atoms) {
1162                 if (thread_atoms_insert(sched, migrant))
1163                         goto out_put;
1164                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1165                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1166                 if (!atoms) {
1167                         pr_err("migration-event: Internal tree error");
1168                         goto out_put;
1169                 }
1170                 if (add_sched_out_event(atoms, 'R', timestamp))
1171                         goto out_put;
1172         }
1173
1174         BUG_ON(list_empty(&atoms->work_list));
1175
1176         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1177         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1178
1179         sched->nr_timestamps++;
1180
1181         if (atom->sched_out_time > timestamp)
1182                 sched->nr_unordered_timestamps++;
1183         err = 0;
1184 out_put:
1185         thread__put(migrant);
1186         return err;
1187 }
1188
1189 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1190 {
1191         int i;
1192         int ret;
1193         u64 avg;
1194
1195         if (!work_list->nb_atoms)
1196                 return;
1197         /*
1198          * Ignore idle threads:
1199          */
1200         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1201                 return;
1202
1203         sched->all_runtime += work_list->total_runtime;
1204         sched->all_count   += work_list->nb_atoms;
1205
1206         if (work_list->num_merged > 1)
1207                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1208         else
1209                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1210
1211         for (i = 0; i < 24 - ret; i++)
1212                 printf(" ");
1213
1214         avg = work_list->total_lat / work_list->nb_atoms;
1215
1216         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1217               (double)work_list->total_runtime / NSEC_PER_MSEC,
1218                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1219                  (double)work_list->max_lat / NSEC_PER_MSEC,
1220                  (double)work_list->max_lat_at / NSEC_PER_SEC);
1221 }
1222
1223 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1224 {
1225         if (l->thread == r->thread)
1226                 return 0;
1227         if (l->thread->tid < r->thread->tid)
1228                 return -1;
1229         if (l->thread->tid > r->thread->tid)
1230                 return 1;
1231         return (int)(l->thread - r->thread);
1232 }
1233
1234 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1235 {
1236         u64 avgl, avgr;
1237
1238         if (!l->nb_atoms)
1239                 return -1;
1240
1241         if (!r->nb_atoms)
1242                 return 1;
1243
1244         avgl = l->total_lat / l->nb_atoms;
1245         avgr = r->total_lat / r->nb_atoms;
1246
1247         if (avgl < avgr)
1248                 return -1;
1249         if (avgl > avgr)
1250                 return 1;
1251
1252         return 0;
1253 }
1254
1255 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1256 {
1257         if (l->max_lat < r->max_lat)
1258                 return -1;
1259         if (l->max_lat > r->max_lat)
1260                 return 1;
1261
1262         return 0;
1263 }
1264
1265 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1266 {
1267         if (l->nb_atoms < r->nb_atoms)
1268                 return -1;
1269         if (l->nb_atoms > r->nb_atoms)
1270                 return 1;
1271
1272         return 0;
1273 }
1274
1275 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1276 {
1277         if (l->total_runtime < r->total_runtime)
1278                 return -1;
1279         if (l->total_runtime > r->total_runtime)
1280                 return 1;
1281
1282         return 0;
1283 }
1284
1285 static int sort_dimension__add(const char *tok, struct list_head *list)
1286 {
1287         size_t i;
1288         static struct sort_dimension avg_sort_dimension = {
1289                 .name = "avg",
1290                 .cmp  = avg_cmp,
1291         };
1292         static struct sort_dimension max_sort_dimension = {
1293                 .name = "max",
1294                 .cmp  = max_cmp,
1295         };
1296         static struct sort_dimension pid_sort_dimension = {
1297                 .name = "pid",
1298                 .cmp  = pid_cmp,
1299         };
1300         static struct sort_dimension runtime_sort_dimension = {
1301                 .name = "runtime",
1302                 .cmp  = runtime_cmp,
1303         };
1304         static struct sort_dimension switch_sort_dimension = {
1305                 .name = "switch",
1306                 .cmp  = switch_cmp,
1307         };
1308         struct sort_dimension *available_sorts[] = {
1309                 &pid_sort_dimension,
1310                 &avg_sort_dimension,
1311                 &max_sort_dimension,
1312                 &switch_sort_dimension,
1313                 &runtime_sort_dimension,
1314         };
1315
1316         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1317                 if (!strcmp(available_sorts[i]->name, tok)) {
1318                         list_add_tail(&available_sorts[i]->list, list);
1319
1320                         return 0;
1321                 }
1322         }
1323
1324         return -1;
1325 }
1326
1327 static void perf_sched__sort_lat(struct perf_sched *sched)
1328 {
1329         struct rb_node *node;
1330         struct rb_root *root = &sched->atom_root;
1331 again:
1332         for (;;) {
1333                 struct work_atoms *data;
1334                 node = rb_first(root);
1335                 if (!node)
1336                         break;
1337
1338                 rb_erase(node, root);
1339                 data = rb_entry(node, struct work_atoms, node);
1340                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1341         }
1342         if (root == &sched->atom_root) {
1343                 root = &sched->merged_atom_root;
1344                 goto again;
1345         }
1346 }
1347
1348 static int process_sched_wakeup_event(struct perf_tool *tool,
1349                                       struct perf_evsel *evsel,
1350                                       struct perf_sample *sample,
1351                                       struct machine *machine)
1352 {
1353         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1354
1355         if (sched->tp_handler->wakeup_event)
1356                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1357
1358         return 0;
1359 }
1360
1361 union map_priv {
1362         void    *ptr;
1363         bool     color;
1364 };
1365
1366 static bool thread__has_color(struct thread *thread)
1367 {
1368         union map_priv priv = {
1369                 .ptr = thread__priv(thread),
1370         };
1371
1372         return priv.color;
1373 }
1374
1375 static struct thread*
1376 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1377 {
1378         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1379         union map_priv priv = {
1380                 .color = false,
1381         };
1382
1383         if (!sched->map.color_pids || !thread || thread__priv(thread))
1384                 return thread;
1385
1386         if (thread_map__has(sched->map.color_pids, tid))
1387                 priv.color = true;
1388
1389         thread__set_priv(thread, priv.ptr);
1390         return thread;
1391 }
1392
1393 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1394                             struct perf_sample *sample, struct machine *machine)
1395 {
1396         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1397         struct thread *sched_in;
1398         int new_shortname;
1399         u64 timestamp0, timestamp = sample->time;
1400         s64 delta;
1401         int i, this_cpu = sample->cpu;
1402         int cpus_nr;
1403         bool new_cpu = false;
1404         const char *color = PERF_COLOR_NORMAL;
1405
1406         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1407
1408         if (this_cpu > sched->max_cpu)
1409                 sched->max_cpu = this_cpu;
1410
1411         if (sched->map.comp) {
1412                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1413                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1414                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1415                         new_cpu = true;
1416                 }
1417         } else
1418                 cpus_nr = sched->max_cpu;
1419
1420         timestamp0 = sched->cpu_last_switched[this_cpu];
1421         sched->cpu_last_switched[this_cpu] = timestamp;
1422         if (timestamp0)
1423                 delta = timestamp - timestamp0;
1424         else
1425                 delta = 0;
1426
1427         if (delta < 0) {
1428                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1429                 return -1;
1430         }
1431
1432         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1433         if (sched_in == NULL)
1434                 return -1;
1435
1436         sched->curr_thread[this_cpu] = thread__get(sched_in);
1437
1438         printf("  ");
1439
1440         new_shortname = 0;
1441         if (!sched_in->shortname[0]) {
1442                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1443                         /*
1444                          * Don't allocate a letter-number for swapper:0
1445                          * as a shortname. Instead, we use '.' for it.
1446                          */
1447                         sched_in->shortname[0] = '.';
1448                         sched_in->shortname[1] = ' ';
1449                 } else {
1450                         sched_in->shortname[0] = sched->next_shortname1;
1451                         sched_in->shortname[1] = sched->next_shortname2;
1452
1453                         if (sched->next_shortname1 < 'Z') {
1454                                 sched->next_shortname1++;
1455                         } else {
1456                                 sched->next_shortname1 = 'A';
1457                                 if (sched->next_shortname2 < '9')
1458                                         sched->next_shortname2++;
1459                                 else
1460                                         sched->next_shortname2 = '0';
1461                         }
1462                 }
1463                 new_shortname = 1;
1464         }
1465
1466         for (i = 0; i < cpus_nr; i++) {
1467                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1468                 struct thread *curr_thread = sched->curr_thread[cpu];
1469                 const char *pid_color = color;
1470                 const char *cpu_color = color;
1471
1472                 if (curr_thread && thread__has_color(curr_thread))
1473                         pid_color = COLOR_PIDS;
1474
1475                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1476                         continue;
1477
1478                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1479                         cpu_color = COLOR_CPUS;
1480
1481                 if (cpu != this_cpu)
1482                         color_fprintf(stdout, cpu_color, " ");
1483                 else
1484                         color_fprintf(stdout, cpu_color, "*");
1485
1486                 if (sched->curr_thread[cpu])
1487                         color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1488                 else
1489                         color_fprintf(stdout, color, "   ");
1490         }
1491
1492         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1493                 goto out;
1494
1495         color_fprintf(stdout, color, "  %12.6f secs ", (double)timestamp / NSEC_PER_SEC);
1496         if (new_shortname) {
1497                 const char *pid_color = color;
1498
1499                 if (thread__has_color(sched_in))
1500                         pid_color = COLOR_PIDS;
1501
1502                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1503                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1504         }
1505
1506         if (sched->map.comp && new_cpu)
1507                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1508
1509 out:
1510         color_fprintf(stdout, color, "\n");
1511
1512         thread__put(sched_in);
1513
1514         return 0;
1515 }
1516
1517 static int process_sched_switch_event(struct perf_tool *tool,
1518                                       struct perf_evsel *evsel,
1519                                       struct perf_sample *sample,
1520                                       struct machine *machine)
1521 {
1522         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1523         int this_cpu = sample->cpu, err = 0;
1524         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1525             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1526
1527         if (sched->curr_pid[this_cpu] != (u32)-1) {
1528                 /*
1529                  * Are we trying to switch away a PID that is
1530                  * not current?
1531                  */
1532                 if (sched->curr_pid[this_cpu] != prev_pid)
1533                         sched->nr_context_switch_bugs++;
1534         }
1535
1536         if (sched->tp_handler->switch_event)
1537                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1538
1539         sched->curr_pid[this_cpu] = next_pid;
1540         return err;
1541 }
1542
1543 static int process_sched_runtime_event(struct perf_tool *tool,
1544                                        struct perf_evsel *evsel,
1545                                        struct perf_sample *sample,
1546                                        struct machine *machine)
1547 {
1548         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1549
1550         if (sched->tp_handler->runtime_event)
1551                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1552
1553         return 0;
1554 }
1555
1556 static int perf_sched__process_fork_event(struct perf_tool *tool,
1557                                           union perf_event *event,
1558                                           struct perf_sample *sample,
1559                                           struct machine *machine)
1560 {
1561         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1562
1563         /* run the fork event through the perf machineruy */
1564         perf_event__process_fork(tool, event, sample, machine);
1565
1566         /* and then run additional processing needed for this command */
1567         if (sched->tp_handler->fork_event)
1568                 return sched->tp_handler->fork_event(sched, event, machine);
1569
1570         return 0;
1571 }
1572
1573 static int process_sched_migrate_task_event(struct perf_tool *tool,
1574                                             struct perf_evsel *evsel,
1575                                             struct perf_sample *sample,
1576                                             struct machine *machine)
1577 {
1578         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1579
1580         if (sched->tp_handler->migrate_task_event)
1581                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1582
1583         return 0;
1584 }
1585
1586 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1587                                   struct perf_evsel *evsel,
1588                                   struct perf_sample *sample,
1589                                   struct machine *machine);
1590
1591 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1592                                                  union perf_event *event __maybe_unused,
1593                                                  struct perf_sample *sample,
1594                                                  struct perf_evsel *evsel,
1595                                                  struct machine *machine)
1596 {
1597         int err = 0;
1598
1599         if (evsel->handler != NULL) {
1600                 tracepoint_handler f = evsel->handler;
1601                 err = f(tool, evsel, sample, machine);
1602         }
1603
1604         return err;
1605 }
1606
1607 static int perf_sched__read_events(struct perf_sched *sched)
1608 {
1609         const struct perf_evsel_str_handler handlers[] = {
1610                 { "sched:sched_switch",       process_sched_switch_event, },
1611                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1612                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1613                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1614                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1615         };
1616         struct perf_session *session;
1617         struct perf_data_file file = {
1618                 .path = input_name,
1619                 .mode = PERF_DATA_MODE_READ,
1620                 .force = sched->force,
1621         };
1622         int rc = -1;
1623
1624         session = perf_session__new(&file, false, &sched->tool);
1625         if (session == NULL) {
1626                 pr_debug("No Memory for session\n");
1627                 return -1;
1628         }
1629
1630         symbol__init(&session->header.env);
1631
1632         if (perf_session__set_tracepoints_handlers(session, handlers))
1633                 goto out_delete;
1634
1635         if (perf_session__has_traces(session, "record -R")) {
1636                 int err = perf_session__process_events(session);
1637                 if (err) {
1638                         pr_err("Failed to process events, error %d", err);
1639                         goto out_delete;
1640                 }
1641
1642                 sched->nr_events      = session->evlist->stats.nr_events[0];
1643                 sched->nr_lost_events = session->evlist->stats.total_lost;
1644                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1645         }
1646
1647         rc = 0;
1648 out_delete:
1649         perf_session__delete(session);
1650         return rc;
1651 }
1652
1653 static void print_bad_events(struct perf_sched *sched)
1654 {
1655         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1656                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1657                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1658                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1659         }
1660         if (sched->nr_lost_events && sched->nr_events) {
1661                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1662                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1663                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1664         }
1665         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1666                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1667                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1668                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1669                 if (sched->nr_lost_events)
1670                         printf(" (due to lost events?)");
1671                 printf("\n");
1672         }
1673 }
1674
1675 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
1676 {
1677         struct rb_node **new = &(root->rb_node), *parent = NULL;
1678         struct work_atoms *this;
1679         const char *comm = thread__comm_str(data->thread), *this_comm;
1680
1681         while (*new) {
1682                 int cmp;
1683
1684                 this = container_of(*new, struct work_atoms, node);
1685                 parent = *new;
1686
1687                 this_comm = thread__comm_str(this->thread);
1688                 cmp = strcmp(comm, this_comm);
1689                 if (cmp > 0) {
1690                         new = &((*new)->rb_left);
1691                 } else if (cmp < 0) {
1692                         new = &((*new)->rb_right);
1693                 } else {
1694                         this->num_merged++;
1695                         this->total_runtime += data->total_runtime;
1696                         this->nb_atoms += data->nb_atoms;
1697                         this->total_lat += data->total_lat;
1698                         list_splice(&data->work_list, &this->work_list);
1699                         if (this->max_lat < data->max_lat) {
1700                                 this->max_lat = data->max_lat;
1701                                 this->max_lat_at = data->max_lat_at;
1702                         }
1703                         zfree(&data);
1704                         return;
1705                 }
1706         }
1707
1708         data->num_merged++;
1709         rb_link_node(&data->node, parent, new);
1710         rb_insert_color(&data->node, root);
1711 }
1712
1713 static void perf_sched__merge_lat(struct perf_sched *sched)
1714 {
1715         struct work_atoms *data;
1716         struct rb_node *node;
1717
1718         if (sched->skip_merge)
1719                 return;
1720
1721         while ((node = rb_first(&sched->atom_root))) {
1722                 rb_erase(node, &sched->atom_root);
1723                 data = rb_entry(node, struct work_atoms, node);
1724                 __merge_work_atoms(&sched->merged_atom_root, data);
1725         }
1726 }
1727
1728 static int perf_sched__lat(struct perf_sched *sched)
1729 {
1730         struct rb_node *next;
1731
1732         setup_pager();
1733
1734         if (perf_sched__read_events(sched))
1735                 return -1;
1736
1737         perf_sched__merge_lat(sched);
1738         perf_sched__sort_lat(sched);
1739
1740         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1741         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1742         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1743
1744         next = rb_first(&sched->sorted_atom_root);
1745
1746         while (next) {
1747                 struct work_atoms *work_list;
1748
1749                 work_list = rb_entry(next, struct work_atoms, node);
1750                 output_lat_thread(sched, work_list);
1751                 next = rb_next(next);
1752                 thread__zput(work_list->thread);
1753         }
1754
1755         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1756         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1757                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
1758
1759         printf(" ---------------------------------------------------\n");
1760
1761         print_bad_events(sched);
1762         printf("\n");
1763
1764         return 0;
1765 }
1766
1767 static int setup_map_cpus(struct perf_sched *sched)
1768 {
1769         struct cpu_map *map;
1770
1771         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
1772
1773         if (sched->map.comp) {
1774                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
1775                 if (!sched->map.comp_cpus)
1776                         return -1;
1777         }
1778
1779         if (!sched->map.cpus_str)
1780                 return 0;
1781
1782         map = cpu_map__new(sched->map.cpus_str);
1783         if (!map) {
1784                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
1785                 return -1;
1786         }
1787
1788         sched->map.cpus = map;
1789         return 0;
1790 }
1791
1792 static int setup_color_pids(struct perf_sched *sched)
1793 {
1794         struct thread_map *map;
1795
1796         if (!sched->map.color_pids_str)
1797                 return 0;
1798
1799         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
1800         if (!map) {
1801                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
1802                 return -1;
1803         }
1804
1805         sched->map.color_pids = map;
1806         return 0;
1807 }
1808
1809 static int setup_color_cpus(struct perf_sched *sched)
1810 {
1811         struct cpu_map *map;
1812
1813         if (!sched->map.color_cpus_str)
1814                 return 0;
1815
1816         map = cpu_map__new(sched->map.color_cpus_str);
1817         if (!map) {
1818                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
1819                 return -1;
1820         }
1821
1822         sched->map.color_cpus = map;
1823         return 0;
1824 }
1825
1826 static int perf_sched__map(struct perf_sched *sched)
1827 {
1828         if (setup_map_cpus(sched))
1829                 return -1;
1830
1831         if (setup_color_pids(sched))
1832                 return -1;
1833
1834         if (setup_color_cpus(sched))
1835                 return -1;
1836
1837         setup_pager();
1838         if (perf_sched__read_events(sched))
1839                 return -1;
1840         print_bad_events(sched);
1841         return 0;
1842 }
1843
1844 static int perf_sched__replay(struct perf_sched *sched)
1845 {
1846         unsigned long i;
1847
1848         calibrate_run_measurement_overhead(sched);
1849         calibrate_sleep_measurement_overhead(sched);
1850
1851         test_calibrations(sched);
1852
1853         if (perf_sched__read_events(sched))
1854                 return -1;
1855
1856         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1857         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1858         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1859
1860         if (sched->targetless_wakeups)
1861                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1862         if (sched->multitarget_wakeups)
1863                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1864         if (sched->nr_run_events_optimized)
1865                 printf("run atoms optimized: %ld\n",
1866                         sched->nr_run_events_optimized);
1867
1868         print_task_traces(sched);
1869         add_cross_task_wakeups(sched);
1870
1871         create_tasks(sched);
1872         printf("------------------------------------------------------------\n");
1873         for (i = 0; i < sched->replay_repeat; i++)
1874                 run_one_test(sched);
1875
1876         return 0;
1877 }
1878
1879 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1880                           const char * const usage_msg[])
1881 {
1882         char *tmp, *tok, *str = strdup(sched->sort_order);
1883
1884         for (tok = strtok_r(str, ", ", &tmp);
1885                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1886                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1887                         usage_with_options_msg(usage_msg, options,
1888                                         "Unknown --sort key: `%s'", tok);
1889                 }
1890         }
1891
1892         free(str);
1893
1894         sort_dimension__add("pid", &sched->cmp_pid);
1895 }
1896
1897 static int __cmd_record(int argc, const char **argv)
1898 {
1899         unsigned int rec_argc, i, j;
1900         const char **rec_argv;
1901         const char * const record_args[] = {
1902                 "record",
1903                 "-a",
1904                 "-R",
1905                 "-m", "1024",
1906                 "-c", "1",
1907                 "-e", "sched:sched_switch",
1908                 "-e", "sched:sched_stat_wait",
1909                 "-e", "sched:sched_stat_sleep",
1910                 "-e", "sched:sched_stat_iowait",
1911                 "-e", "sched:sched_stat_runtime",
1912                 "-e", "sched:sched_process_fork",
1913                 "-e", "sched:sched_wakeup",
1914                 "-e", "sched:sched_wakeup_new",
1915                 "-e", "sched:sched_migrate_task",
1916         };
1917
1918         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1919         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1920
1921         if (rec_argv == NULL)
1922                 return -ENOMEM;
1923
1924         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1925                 rec_argv[i] = strdup(record_args[i]);
1926
1927         for (j = 1; j < (unsigned int)argc; j++, i++)
1928                 rec_argv[i] = argv[j];
1929
1930         BUG_ON(i != rec_argc);
1931
1932         return cmd_record(i, rec_argv, NULL);
1933 }
1934
1935 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1936 {
1937         const char default_sort_order[] = "avg, max, switch, runtime";
1938         struct perf_sched sched = {
1939                 .tool = {
1940                         .sample          = perf_sched__process_tracepoint_sample,
1941                         .comm            = perf_event__process_comm,
1942                         .lost            = perf_event__process_lost,
1943                         .fork            = perf_sched__process_fork_event,
1944                         .ordered_events = true,
1945                 },
1946                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1947                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1948                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1949                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1950                 .sort_order           = default_sort_order,
1951                 .replay_repeat        = 10,
1952                 .profile_cpu          = -1,
1953                 .next_shortname1      = 'A',
1954                 .next_shortname2      = '0',
1955                 .skip_merge           = 0,
1956         };
1957         const struct option latency_options[] = {
1958         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1959                    "sort by key(s): runtime, switch, avg, max"),
1960         OPT_INCR('v', "verbose", &verbose,
1961                     "be more verbose (show symbol address, etc)"),
1962         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1963                     "CPU to profile on"),
1964         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1965                     "dump raw trace in ASCII"),
1966         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
1967                     "latency stats per pid instead of per comm"),
1968         OPT_END()
1969         };
1970         const struct option replay_options[] = {
1971         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1972                      "repeat the workload replay N times (-1: infinite)"),
1973         OPT_INCR('v', "verbose", &verbose,
1974                     "be more verbose (show symbol address, etc)"),
1975         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1976                     "dump raw trace in ASCII"),
1977         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1978         OPT_END()
1979         };
1980         const struct option sched_options[] = {
1981         OPT_STRING('i', "input", &input_name, "file",
1982                     "input file name"),
1983         OPT_INCR('v', "verbose", &verbose,
1984                     "be more verbose (show symbol address, etc)"),
1985         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1986                     "dump raw trace in ASCII"),
1987         OPT_END()
1988         };
1989         const struct option map_options[] = {
1990         OPT_BOOLEAN(0, "compact", &sched.map.comp,
1991                     "map output in compact mode"),
1992         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
1993                    "highlight given pids in map"),
1994         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
1995                     "highlight given CPUs in map"),
1996         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
1997                     "display given CPUs in map"),
1998         OPT_END()
1999         };
2000         const char * const latency_usage[] = {
2001                 "perf sched latency [<options>]",
2002                 NULL
2003         };
2004         const char * const replay_usage[] = {
2005                 "perf sched replay [<options>]",
2006                 NULL
2007         };
2008         const char * const map_usage[] = {
2009                 "perf sched map [<options>]",
2010                 NULL
2011         };
2012         const char *const sched_subcommands[] = { "record", "latency", "map",
2013                                                   "replay", "script", NULL };
2014         const char *sched_usage[] = {
2015                 NULL,
2016                 NULL
2017         };
2018         struct trace_sched_handler lat_ops  = {
2019                 .wakeup_event       = latency_wakeup_event,
2020                 .switch_event       = latency_switch_event,
2021                 .runtime_event      = latency_runtime_event,
2022                 .migrate_task_event = latency_migrate_task_event,
2023         };
2024         struct trace_sched_handler map_ops  = {
2025                 .switch_event       = map_switch_event,
2026         };
2027         struct trace_sched_handler replay_ops  = {
2028                 .wakeup_event       = replay_wakeup_event,
2029                 .switch_event       = replay_switch_event,
2030                 .fork_event         = replay_fork_event,
2031         };
2032         unsigned int i;
2033
2034         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
2035                 sched.curr_pid[i] = -1;
2036
2037         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
2038                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2039         if (!argc)
2040                 usage_with_options(sched_usage, sched_options);
2041
2042         /*
2043          * Aliased to 'perf script' for now:
2044          */
2045         if (!strcmp(argv[0], "script"))
2046                 return cmd_script(argc, argv, prefix);
2047
2048         if (!strncmp(argv[0], "rec", 3)) {
2049                 return __cmd_record(argc, argv);
2050         } else if (!strncmp(argv[0], "lat", 3)) {
2051                 sched.tp_handler = &lat_ops;
2052                 if (argc > 1) {
2053                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
2054                         if (argc)
2055                                 usage_with_options(latency_usage, latency_options);
2056                 }
2057                 setup_sorting(&sched, latency_options, latency_usage);
2058                 return perf_sched__lat(&sched);
2059         } else if (!strcmp(argv[0], "map")) {
2060                 if (argc) {
2061                         argc = parse_options(argc, argv, map_options, map_usage, 0);
2062                         if (argc)
2063                                 usage_with_options(map_usage, map_options);
2064                 }
2065                 sched.tp_handler = &map_ops;
2066                 setup_sorting(&sched, latency_options, latency_usage);
2067                 return perf_sched__map(&sched);
2068         } else if (!strncmp(argv[0], "rep", 3)) {
2069                 sched.tp_handler = &replay_ops;
2070                 if (argc) {
2071                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
2072                         if (argc)
2073                                 usage_with_options(replay_usage, replay_options);
2074                 }
2075                 return perf_sched__replay(&sched);
2076         } else {
2077                 usage_with_options(sched_usage, sched_options);
2078         }
2079
2080         return 0;
2081 }