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 #include "util/stat.h"
  17 #include "util/callchain.h"
  18 #include "util/time-utils.h"
  19
  20 #include <subcmd/parse-options.h>
  21 #include "util/trace-event.h"
  22
  23 #include "util/debug.h"
  24
  25 #include <linux/log2.h>
  26 #include <sys/prctl.h>
  27 #include <sys/resource.h>
  28
  29 #include <semaphore.h>
  30 #include <pthread.h>
  31 #include <math.h>
  32 #include <api/fs/fs.h>
  33 #include <linux/time64.h>
  34
  35 #define PR_SET_NAME             15               /* Set process name */
  36 #define MAX_CPUS                4096
  37 #define COMM_LEN                20
  38 #define SYM_LEN                 129
  39 #define MAX_PID                 1024000
  40
  41 struct sched_atom;
  42
  43 struct task_desc {
  44         unsigned long           nr;
  45         unsigned long           pid;
  46         char                    comm[COMM_LEN];
  47
  48         unsigned long           nr_events;
  49         unsigned long           curr_event;
  50         struct sched_atom       **atoms;
  51
  52         pthread_t               thread;
  53         sem_t                   sleep_sem;
  54
  55         sem_t                   ready_for_work;
  56         sem_t                   work_done_sem;
  57
  58         u64                     cpu_usage;
  59 };
  60
  61 enum sched_event_type {
  62         SCHED_EVENT_RUN,
  63         SCHED_EVENT_SLEEP,
  64         SCHED_EVENT_WAKEUP,
  65         SCHED_EVENT_MIGRATION,
  66 };
  67
  68 struct sched_atom {
  69         enum sched_event_type   type;
  70         int                     specific_wait;
  71         u64                     timestamp;
  72         u64                     duration;
  73         unsigned long           nr;
  74         sem_t                   *wait_sem;
  75         struct task_desc        *wakee;
  76 };
  77
  78 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  79
  80 /* task state bitmask, copied from include/linux/sched.h */
  81 #define TASK_RUNNING            0
  82 #define TASK_INTERRUPTIBLE      1
  83 #define TASK_UNINTERRUPTIBLE    2
  84 #define __TASK_STOPPED          4
  85 #define __TASK_TRACED           8
  86 /* in tsk->exit_state */
  87 #define EXIT_DEAD               16
  88 #define EXIT_ZOMBIE             32
  89 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
  90 /* in tsk->state again */
  91 #define TASK_DEAD               64
  92 #define TASK_WAKEKILL           128
  93 #define TASK_WAKING             256
  94 #define TASK_PARKED             512
  95
  96 enum thread_state {
  97         THREAD_SLEEPING = 0,
  98         THREAD_WAIT_CPU,
  99         THREAD_SCHED_IN,
 100         THREAD_IGNORE
 101 };
 102
 103 struct work_atom {
 104         struct list_head        list;
 105         enum thread_state       state;
 106         u64                     sched_out_time;
 107         u64                     wake_up_time;
 108         u64                     sched_in_time;
 109         u64                     runtime;
 110 };
 111
 112 struct work_atoms {
 113         struct list_head        work_list;
 114         struct thread           *thread;
 115         struct rb_node          node;
 116         u64                     max_lat;
 117         u64                     max_lat_at;
 118         u64                     total_lat;
 119         u64                     nb_atoms;
 120         u64                     total_runtime;
 121         int                     num_merged;
 122 };
 123
 124 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 125
 126 struct perf_sched;
 127
 128 struct trace_sched_handler {
 129         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 130                             struct perf_sample *sample, struct machine *machine);
 131
 132         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 133                              struct perf_sample *sample, struct machine *machine);
 134
 135         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 136                             struct perf_sample *sample, struct machine *machine);
 137
 138         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 139         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 140                           struct machine *machine);
 141
 142         int (*migrate_task_event)(struct perf_sched *sched,
 143                                   struct perf_evsel *evsel,
 144                                   struct perf_sample *sample,
 145                                   struct machine *machine);
 146 };
 147
 148 #define COLOR_PIDS PERF_COLOR_BLUE
 149 #define COLOR_CPUS PERF_COLOR_BG_RED
 150
 151 struct perf_sched_map {
 152         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 153         int                     *comp_cpus;
 154         bool                     comp;
 155         struct thread_map       *color_pids;
 156         const char              *color_pids_str;
 157         struct cpu_map          *color_cpus;
 158         const char              *color_cpus_str;
 159         struct cpu_map          *cpus;
 160         const char              *cpus_str;
 161 };
 162
 163 struct perf_sched {
 164         struct perf_tool tool;
 165         const char       *sort_order;
 166         unsigned long    nr_tasks;
 167         struct task_desc **pid_to_task;
 168         struct task_desc **tasks;
 169         const struct trace_sched_handler *tp_handler;
 170         pthread_mutex_t  start_work_mutex;
 171         pthread_mutex_t  work_done_wait_mutex;
 172         int              profile_cpu;
 173 /*
 174  * Track the current task - that way we can know whether there's any
 175  * weird events, such as a task being switched away that is not current.
 176  */
 177         int              max_cpu;
 178         u32              curr_pid[MAX_CPUS];
 179         struct thread    *curr_thread[MAX_CPUS];
 180         char             next_shortname1;
 181         char             next_shortname2;
 182         unsigned int     replay_repeat;
 183         unsigned long    nr_run_events;
 184         unsigned long    nr_sleep_events;
 185         unsigned long    nr_wakeup_events;
 186         unsigned long    nr_sleep_corrections;
 187         unsigned long    nr_run_events_optimized;
 188         unsigned long    targetless_wakeups;
 189         unsigned long    multitarget_wakeups;
 190         unsigned long    nr_runs;
 191         unsigned long    nr_timestamps;
 192         unsigned long    nr_unordered_timestamps;
 193         unsigned long    nr_context_switch_bugs;
 194         unsigned long    nr_events;
 195         unsigned long    nr_lost_chunks;
 196         unsigned long    nr_lost_events;
 197         u64              run_measurement_overhead;
 198         u64              sleep_measurement_overhead;
 199         u64              start_time;
 200         u64              cpu_usage;
 201         u64              runavg_cpu_usage;
 202         u64              parent_cpu_usage;
 203         u64              runavg_parent_cpu_usage;
 204         u64              sum_runtime;
 205         u64              sum_fluct;
 206         u64              run_avg;
 207         u64              all_runtime;
 208         u64              all_count;
 209         u64              cpu_last_switched[MAX_CPUS];
 210         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 211         struct list_head sort_list, cmp_pid;
 212         bool force;
 213         bool skip_merge;
 214         struct perf_sched_map map;
 215
 216         /* options for timehist command */
 217         bool            summary;
 218         bool            summary_only;
 219         bool            idle_hist;
 220         bool            show_callchain;
 221         unsigned int    max_stack;
 222         bool            show_cpu_visual;
 223         bool            show_wakeups;
 224         bool            show_migrations;
 225         bool            show_state;
 226         u64             skipped_samples;
 227         const char      *time_str;
 228         struct perf_time_interval ptime;
 229         struct perf_time_interval hist_time;
 230 };
 231
 232 /* per thread run time data */
 233 struct thread_runtime {
 234         u64 last_time;      /* time of previous sched in/out event */
 235         u64 dt_run;         /* run time */
 236         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 237         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 238         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 239         u64 dt_delay;       /* time between wakeup and sched-in */
 240         u64 ready_to_run;   /* time of wakeup */
 241
 242         struct stats run_stats;
 243         u64 total_run_time;
 244         u64 total_sleep_time;
 245         u64 total_iowait_time;
 246         u64 total_preempt_time;
 247         u64 total_delay_time;
 248
 249         int last_state;
 250         u64 migrations;
 251 };
 252
 253 /* per event run time data */
 254 struct evsel_runtime {
 255         u64 *last_time; /* time this event was last seen per cpu */
 256         u32 ncpu;       /* highest cpu slot allocated */
 257 };
 258
 259 /* per cpu idle time data */
 260 struct idle_thread_runtime {
 261         struct thread_runtime   tr;
 262         struct thread           *last_thread;
 263         struct rb_root          sorted_root;
 264         struct callchain_root   callchain;
 265         struct callchain_cursor cursor;
 266 };
 267
 268 /* track idle times per cpu */
 269 static struct thread **idle_threads;
 270 static int idle_max_cpu;
 271 static char idle_comm[] = "<idle>";
 272
 273 static u64 get_nsecs(void)
 274 {
 275         struct timespec ts;
 276
 277         clock_gettime(CLOCK_MONOTONIC, &ts);
 278
 279         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 280 }
 281
 282 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 283 {
 284         u64 T0 = get_nsecs(), T1;
 285
 286         do {
 287                 T1 = get_nsecs();
 288         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 289 }
 290
 291 static void sleep_nsecs(u64 nsecs)
 292 {
 293         struct timespec ts;
 294
 295         ts.tv_nsec = nsecs % 999999999;
 296         ts.tv_sec = nsecs / 999999999;
 297
 298         nanosleep(&ts, NULL);
 299 }
 300
 301 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 302 {
 303         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 304         int i;
 305
 306         for (i = 0; i < 10; i++) {
 307                 T0 = get_nsecs();
 308                 burn_nsecs(sched, 0);
 309                 T1 = get_nsecs();
 310                 delta = T1-T0;
 311                 min_delta = min(min_delta, delta);
 312         }
 313         sched->run_measurement_overhead = min_delta;
 314
 315         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 316 }
 317
 318 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 319 {
 320         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 321         int i;
 322
 323         for (i = 0; i < 10; i++) {
 324                 T0 = get_nsecs();
 325                 sleep_nsecs(10000);
 326                 T1 = get_nsecs();
 327                 delta = T1-T0;
 328                 min_delta = min(min_delta, delta);
 329         }
 330         min_delta -= 10000;
 331         sched->sleep_measurement_overhead = min_delta;
 332
 333         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 334 }
 335
 336 static struct sched_atom *
 337 get_new_event(struct task_desc *task, u64 timestamp)
 338 {
 339         struct sched_atom *event = zalloc(sizeof(*event));
 340         unsigned long idx = task->nr_events;
 341         size_t size;
 342
 343         event->timestamp = timestamp;
 344         event->nr = idx;
 345
 346         task->nr_events++;
 347         size = sizeof(struct sched_atom *) * task->nr_events;
 348         task->atoms = realloc(task->atoms, size);
 349         BUG_ON(!task->atoms);
 350
 351         task->atoms[idx] = event;
 352
 353         return event;
 354 }
 355
 356 static struct sched_atom *last_event(struct task_desc *task)
 357 {
 358         if (!task->nr_events)
 359                 return NULL;
 360
 361         return task->atoms[task->nr_events - 1];
 362 }
 363
 364 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 365                                 u64 timestamp, u64 duration)
 366 {
 367         struct sched_atom *event, *curr_event = last_event(task);
 368
 369         /*
 370          * optimize an existing RUN event by merging this one
 371          * to it:
 372          */
 373         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 374                 sched->nr_run_events_optimized++;
 375                 curr_event->duration += duration;
 376                 return;
 377         }
 378
 379         event = get_new_event(task, timestamp);
 380
 381         event->type = SCHED_EVENT_RUN;
 382         event->duration = duration;
 383
 384         sched->nr_run_events++;
 385 }
 386
 387 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 388                                    u64 timestamp, struct task_desc *wakee)
 389 {
 390         struct sched_atom *event, *wakee_event;
 391
 392         event = get_new_event(task, timestamp);
 393         event->type = SCHED_EVENT_WAKEUP;
 394         event->wakee = wakee;
 395
 396         wakee_event = last_event(wakee);
 397         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 398                 sched->targetless_wakeups++;
 399                 return;
 400         }
 401         if (wakee_event->wait_sem) {
 402                 sched->multitarget_wakeups++;
 403                 return;
 404         }
 405
 406         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 407         sem_init(wakee_event->wait_sem, 0, 0);
 408         wakee_event->specific_wait = 1;
 409         event->wait_sem = wakee_event->wait_sem;
 410
 411         sched->nr_wakeup_events++;
 412 }
 413
 414 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 415                                   u64 timestamp, u64 task_state __maybe_unused)
 416 {
 417         struct sched_atom *event = get_new_event(task, timestamp);
 418
 419         event->type = SCHED_EVENT_SLEEP;
 420
 421         sched->nr_sleep_events++;
 422 }
 423
 424 static struct task_desc *register_pid(struct perf_sched *sched,
 425                                       unsigned long pid, const char *comm)
 426 {
 427         struct task_desc *task;
 428         static int pid_max;
 429
 430         if (sched->pid_to_task == NULL) {
 431                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 432                         pid_max = MAX_PID;
 433                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 434         }
 435         if (pid >= (unsigned long)pid_max) {
 436                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 437                         sizeof(struct task_desc *))) == NULL);
 438                 while (pid >= (unsigned long)pid_max)
 439                         sched->pid_to_task[pid_max++] = NULL;
 440         }
 441
 442         task = sched->pid_to_task[pid];
 443
 444         if (task)
 445                 return task;
 446
 447         task = zalloc(sizeof(*task));
 448         task->pid = pid;
 449         task->nr = sched->nr_tasks;
 450         strcpy(task->comm, comm);
 451         /*
 452          * every task starts in sleeping state - this gets ignored
 453          * if there's no wakeup pointing to this sleep state:
 454          */
 455         add_sched_event_sleep(sched, task, 0, 0);
 456
 457         sched->pid_to_task[pid] = task;
 458         sched->nr_tasks++;
 459         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 460         BUG_ON(!sched->tasks);
 461         sched->tasks[task->nr] = task;
 462
 463         if (verbose > 0)
 464                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 465
 466         return task;
 467 }
 468
 469
 470 static void print_task_traces(struct perf_sched *sched)
 471 {
 472         struct task_desc *task;
 473         unsigned long i;
 474
 475         for (i = 0; i < sched->nr_tasks; i++) {
 476                 task = sched->tasks[i];
 477                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 478                         task->nr, task->comm, task->pid, task->nr_events);
 479         }
 480 }
 481
 482 static void add_cross_task_wakeups(struct perf_sched *sched)
 483 {
 484         struct task_desc *task1, *task2;
 485         unsigned long i, j;
 486
 487         for (i = 0; i < sched->nr_tasks; i++) {
 488                 task1 = sched->tasks[i];
 489                 j = i + 1;
 490                 if (j == sched->nr_tasks)
 491                         j = 0;
 492                 task2 = sched->tasks[j];
 493                 add_sched_event_wakeup(sched, task1, 0, task2);
 494         }
 495 }
 496
 497 static void perf_sched__process_event(struct perf_sched *sched,
 498                                       struct sched_atom *atom)
 499 {
 500         int ret = 0;
 501
 502         switch (atom->type) {
 503                 case SCHED_EVENT_RUN:
 504                         burn_nsecs(sched, atom->duration);
 505                         break;
 506                 case SCHED_EVENT_SLEEP:
 507                         if (atom->wait_sem)
 508                                 ret = sem_wait(atom->wait_sem);
 509                         BUG_ON(ret);
 510                         break;
 511                 case SCHED_EVENT_WAKEUP:
 512                         if (atom->wait_sem)
 513                                 ret = sem_post(atom->wait_sem);
 514                         BUG_ON(ret);
 515                         break;
 516                 case SCHED_EVENT_MIGRATION:
 517                         break;
 518                 default:
 519                         BUG_ON(1);
 520         }
 521 }
 522
 523 static u64 get_cpu_usage_nsec_parent(void)
 524 {
 525         struct rusage ru;
 526         u64 sum;
 527         int err;
 528
 529         err = getrusage(RUSAGE_SELF, &ru);
 530         BUG_ON(err);
 531
 532         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 533         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 534
 535         return sum;
 536 }
 537
 538 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 539 {
 540         struct perf_event_attr attr;
 541         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 542         int fd;
 543         struct rlimit limit;
 544         bool need_privilege = false;
 545
 546         memset(&attr, 0, sizeof(attr));
 547
 548         attr.type = PERF_TYPE_SOFTWARE;
 549         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 550
 551 force_again:
 552         fd = sys_perf_event_open(&attr, 0, -1, -1,
 553                                  perf_event_open_cloexec_flag());
 554
 555         if (fd < 0) {
 556                 if (errno == EMFILE) {
 557                         if (sched->force) {
 558                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 559                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 560                                 if (limit.rlim_cur > limit.rlim_max) {
 561                                         limit.rlim_max = limit.rlim_cur;
 562                                         need_privilege = true;
 563                                 }
 564                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 565                                         if (need_privilege && errno == EPERM)
 566                                                 strcpy(info, "Need privilege\n");
 567                                 } else
 568                                         goto force_again;
 569                         } else
 570                                 strcpy(info, "Have a try with -f option\n");
 571                 }
 572                 pr_err("Error: sys_perf_event_open() syscall returned "
 573                        "with %d (%s)\n%s", fd,
 574                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 575                 exit(EXIT_FAILURE);
 576         }
 577         return fd;
 578 }
 579
 580 static u64 get_cpu_usage_nsec_self(int fd)
 581 {
 582         u64 runtime;
 583         int ret;
 584
 585         ret = read(fd, &runtime, sizeof(runtime));
 586         BUG_ON(ret != sizeof(runtime));
 587
 588         return runtime;
 589 }
 590
 591 struct sched_thread_parms {
 592         struct task_desc  *task;
 593         struct perf_sched *sched;
 594         int fd;
 595 };
 596
 597 static void *thread_func(void *ctx)
 598 {
 599         struct sched_thread_parms *parms = ctx;
 600         struct task_desc *this_task = parms->task;
 601         struct perf_sched *sched = parms->sched;
 602         u64 cpu_usage_0, cpu_usage_1;
 603         unsigned long i, ret;
 604         char comm2[22];
 605         int fd = parms->fd;
 606
 607         zfree(&parms);
 608
 609         sprintf(comm2, ":%s", this_task->comm);
 610         prctl(PR_SET_NAME, comm2);
 611         if (fd < 0)
 612                 return NULL;
 613 again:
 614         ret = sem_post(&this_task->ready_for_work);
 615         BUG_ON(ret);
 616         ret = pthread_mutex_lock(&sched->start_work_mutex);
 617         BUG_ON(ret);
 618         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 619         BUG_ON(ret);
 620
 621         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 622
 623         for (i = 0; i < this_task->nr_events; i++) {
 624                 this_task->curr_event = i;
 625                 perf_sched__process_event(sched, this_task->atoms[i]);
 626         }
 627
 628         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 629         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 630         ret = sem_post(&this_task->work_done_sem);
 631         BUG_ON(ret);
 632
 633         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 634         BUG_ON(ret);
 635         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 636         BUG_ON(ret);
 637
 638         goto again;
 639 }
 640
 641 static void create_tasks(struct perf_sched *sched)
 642 {
 643         struct task_desc *task;
 644         pthread_attr_t attr;
 645         unsigned long i;
 646         int err;
 647
 648         err = pthread_attr_init(&attr);
 649         BUG_ON(err);
 650         err = pthread_attr_setstacksize(&attr,
 651                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 652         BUG_ON(err);
 653         err = pthread_mutex_lock(&sched->start_work_mutex);
 654         BUG_ON(err);
 655         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 656         BUG_ON(err);
 657         for (i = 0; i < sched->nr_tasks; i++) {
 658                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 659                 BUG_ON(parms == NULL);
 660                 parms->task = task = sched->tasks[i];
 661                 parms->sched = sched;
 662                 parms->fd = self_open_counters(sched, i);
 663                 sem_init(&task->sleep_sem, 0, 0);
 664                 sem_init(&task->ready_for_work, 0, 0);
 665                 sem_init(&task->work_done_sem, 0, 0);
 666                 task->curr_event = 0;
 667                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 668                 BUG_ON(err);
 669         }
 670 }
 671
 672 static void wait_for_tasks(struct perf_sched *sched)
 673 {
 674         u64 cpu_usage_0, cpu_usage_1;
 675         struct task_desc *task;
 676         unsigned long i, ret;
 677
 678         sched->start_time = get_nsecs();
 679         sched->cpu_usage = 0;
 680         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 681
 682         for (i = 0; i < sched->nr_tasks; i++) {
 683                 task = sched->tasks[i];
 684                 ret = sem_wait(&task->ready_for_work);
 685                 BUG_ON(ret);
 686                 sem_init(&task->ready_for_work, 0, 0);
 687         }
 688         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 689         BUG_ON(ret);
 690
 691         cpu_usage_0 = get_cpu_usage_nsec_parent();
 692
 693         pthread_mutex_unlock(&sched->start_work_mutex);
 694
 695         for (i = 0; i < sched->nr_tasks; i++) {
 696                 task = sched->tasks[i];
 697                 ret = sem_wait(&task->work_done_sem);
 698                 BUG_ON(ret);
 699                 sem_init(&task->work_done_sem, 0, 0);
 700                 sched->cpu_usage += task->cpu_usage;
 701                 task->cpu_usage = 0;
 702         }
 703
 704         cpu_usage_1 = get_cpu_usage_nsec_parent();
 705         if (!sched->runavg_cpu_usage)
 706                 sched->runavg_cpu_usage = sched->cpu_usage;
 707         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 708
 709         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 710         if (!sched->runavg_parent_cpu_usage)
 711                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 712         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 713                                          sched->parent_cpu_usage)/sched->replay_repeat;
 714
 715         ret = pthread_mutex_lock(&sched->start_work_mutex);
 716         BUG_ON(ret);
 717
 718         for (i = 0; i < sched->nr_tasks; i++) {
 719                 task = sched->tasks[i];
 720                 sem_init(&task->sleep_sem, 0, 0);
 721                 task->curr_event = 0;
 722         }
 723 }
 724
 725 static void run_one_test(struct perf_sched *sched)
 726 {
 727         u64 T0, T1, delta, avg_delta, fluct;
 728
 729         T0 = get_nsecs();
 730         wait_for_tasks(sched);
 731         T1 = get_nsecs();
 732
 733         delta = T1 - T0;
 734         sched->sum_runtime += delta;
 735         sched->nr_runs++;
 736
 737         avg_delta = sched->sum_runtime / sched->nr_runs;
 738         if (delta < avg_delta)
 739                 fluct = avg_delta - delta;
 740         else
 741                 fluct = delta - avg_delta;
 742         sched->sum_fluct += fluct;
 743         if (!sched->run_avg)
 744                 sched->run_avg = delta;
 745         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 746
 747         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 748
 749         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 750
 751         printf("cpu: %0.2f / %0.2f",
 752                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 753
 754 #if 0
 755         /*
 756          * rusage statistics done by the parent, these are less
 757          * accurate than the sched->sum_exec_runtime based statistics:
 758          */
 759         printf(" [%0.2f / %0.2f]",
 760                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 761                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 762 #endif
 763
 764         printf("\n");
 765
 766         if (sched->nr_sleep_corrections)
 767                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 768         sched->nr_sleep_corrections = 0;
 769 }
 770
 771 static void test_calibrations(struct perf_sched *sched)
 772 {
 773         u64 T0, T1;
 774
 775         T0 = get_nsecs();
 776         burn_nsecs(sched, NSEC_PER_MSEC);
 777         T1 = get_nsecs();
 778
 779         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 780
 781         T0 = get_nsecs();
 782         sleep_nsecs(NSEC_PER_MSEC);
 783         T1 = get_nsecs();
 784
 785         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 786 }
 787
 788 static int
 789 replay_wakeup_event(struct perf_sched *sched,
 790                     struct perf_evsel *evsel, struct perf_sample *sample,
 791                     struct machine *machine __maybe_unused)
 792 {
 793         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 794         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 795         struct task_desc *waker, *wakee;
 796
 797         if (verbose > 0) {
 798                 printf("sched_wakeup event %p\n", evsel);
 799
 800                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 801         }
 802
 803         waker = register_pid(sched, sample->tid, "<unknown>");
 804         wakee = register_pid(sched, pid, comm);
 805
 806         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 807         return 0;
 808 }
 809
 810 static int replay_switch_event(struct perf_sched *sched,
 811                                struct perf_evsel *evsel,
 812                                struct perf_sample *sample,
 813                                struct machine *machine __maybe_unused)
 814 {
 815         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 816                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 817         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 818                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 819         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 820         struct task_desc *prev, __maybe_unused *next;
 821         u64 timestamp0, timestamp = sample->time;
 822         int cpu = sample->cpu;
 823         s64 delta;
 824
 825         if (verbose > 0)
 826                 printf("sched_switch event %p\n", evsel);
 827
 828         if (cpu >= MAX_CPUS || cpu < 0)
 829                 return 0;
 830
 831         timestamp0 = sched->cpu_last_switched[cpu];
 832         if (timestamp0)
 833                 delta = timestamp - timestamp0;
 834         else
 835                 delta = 0;
 836
 837         if (delta < 0) {
 838                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 839                 return -1;
 840         }
 841
 842         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 843                  prev_comm, prev_pid, next_comm, next_pid, delta);
 844
 845         prev = register_pid(sched, prev_pid, prev_comm);
 846         next = register_pid(sched, next_pid, next_comm);
 847
 848         sched->cpu_last_switched[cpu] = timestamp;
 849
 850         add_sched_event_run(sched, prev, timestamp, delta);
 851         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 852
 853         return 0;
 854 }
 855
 856 static int replay_fork_event(struct perf_sched *sched,
 857                              union perf_event *event,
 858                              struct machine *machine)
 859 {
 860         struct thread *child, *parent;
 861
 862         child = machine__findnew_thread(machine, event->fork.pid,
 863                                         event->fork.tid);
 864         parent = machine__findnew_thread(machine, event->fork.ppid,
 865                                          event->fork.ptid);
 866
 867         if (child == NULL || parent == NULL) {
 868                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 869                                  child, parent);
 870                 goto out_put;
 871         }
 872
 873         if (verbose > 0) {
 874                 printf("fork event\n");
 875                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 876                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 877         }
 878
 879         register_pid(sched, parent->tid, thread__comm_str(parent));
 880         register_pid(sched, child->tid, thread__comm_str(child));
 881 out_put:
 882         thread__put(child);
 883         thread__put(parent);
 884         return 0;
 885 }
 886
 887 struct sort_dimension {
 888         const char              *name;
 889         sort_fn_t               cmp;
 890         struct list_head        list;
 891 };
 892
 893 static int
 894 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 895 {
 896         struct sort_dimension *sort;
 897         int ret = 0;
 898
 899         BUG_ON(list_empty(list));
 900
 901         list_for_each_entry(sort, list, list) {
 902                 ret = sort->cmp(l, r);
 903                 if (ret)
 904                         return ret;
 905         }
 906
 907         return ret;
 908 }
 909
 910 static struct work_atoms *
 911 thread_atoms_search(struct rb_root *root, struct thread *thread,
 912                          struct list_head *sort_list)
 913 {
 914         struct rb_node *node = root->rb_node;
 915         struct work_atoms key = { .thread = thread };
 916
 917         while (node) {
 918                 struct work_atoms *atoms;
 919                 int cmp;
 920
 921                 atoms = container_of(node, struct work_atoms, node);
 922
 923                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 924                 if (cmp > 0)
 925                         node = node->rb_left;
 926                 else if (cmp < 0)
 927                         node = node->rb_right;
 928                 else {
 929                         BUG_ON(thread != atoms->thread);
 930                         return atoms;
 931                 }
 932         }
 933         return NULL;
 934 }
 935
 936 static void
 937 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 938                          struct list_head *sort_list)
 939 {
 940         struct rb_node **new = &(root->rb_node), *parent = NULL;
 941
 942         while (*new) {
 943                 struct work_atoms *this;
 944                 int cmp;
 945
 946                 this = container_of(*new, struct work_atoms, node);
 947                 parent = *new;
 948
 949                 cmp = thread_lat_cmp(sort_list, data, this);
 950
 951                 if (cmp > 0)
 952                         new = &((*new)->rb_left);
 953                 else
 954                         new = &((*new)->rb_right);
 955         }
 956
 957         rb_link_node(&data->node, parent, new);
 958         rb_insert_color(&data->node, root);
 959 }
 960
 961 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 962 {
 963         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 964         if (!atoms) {
 965                 pr_err("No memory at %s\n", __func__);
 966                 return -1;
 967         }
 968
 969         atoms->thread = thread__get(thread);
 970         INIT_LIST_HEAD(&atoms->work_list);
 971         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 972         return 0;
 973 }
 974
 975 static char sched_out_state(u64 prev_state)
 976 {
 977         const char *str = TASK_STATE_TO_CHAR_STR;
 978
 979         return str[prev_state];
 980 }
 981
 982 static int
 983 add_sched_out_event(struct work_atoms *atoms,
 984                     char run_state,
 985                     u64 timestamp)
 986 {
 987         struct work_atom *atom = zalloc(sizeof(*atom));
 988         if (!atom) {
 989                 pr_err("Non memory at %s", __func__);
 990                 return -1;
 991         }
 992
 993         atom->sched_out_time = timestamp;
 994
 995         if (run_state == 'R') {
 996                 atom->state = THREAD_WAIT_CPU;
 997                 atom->wake_up_time = atom->sched_out_time;
 998         }
 999
1000         list_add_tail(&atom->list, &atoms->work_list);
1001         return 0;
1002 }
1003
1004 static void
1005 add_runtime_event(struct work_atoms *atoms, u64 delta,
1006                   u64 timestamp __maybe_unused)
1007 {
1008         struct work_atom *atom;
1009
1010         BUG_ON(list_empty(&atoms->work_list));
1011
1012         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1013
1014         atom->runtime += delta;
1015         atoms->total_runtime += delta;
1016 }
1017
1018 static void
1019 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1020 {
1021         struct work_atom *atom;
1022         u64 delta;
1023
1024         if (list_empty(&atoms->work_list))
1025                 return;
1026
1027         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1028
1029         if (atom->state != THREAD_WAIT_CPU)
1030                 return;
1031
1032         if (timestamp < atom->wake_up_time) {
1033                 atom->state = THREAD_IGNORE;
1034                 return;
1035         }
1036
1037         atom->state = THREAD_SCHED_IN;
1038         atom->sched_in_time = timestamp;
1039
1040         delta = atom->sched_in_time - atom->wake_up_time;
1041         atoms->total_lat += delta;
1042         if (delta > atoms->max_lat) {
1043                 atoms->max_lat = delta;
1044                 atoms->max_lat_at = timestamp;
1045         }
1046         atoms->nb_atoms++;
1047 }
1048
1049 static int latency_switch_event(struct perf_sched *sched,
1050                                 struct perf_evsel *evsel,
1051                                 struct perf_sample *sample,
1052                                 struct machine *machine)
1053 {
1054         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1055                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1056         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1057         struct work_atoms *out_events, *in_events;
1058         struct thread *sched_out, *sched_in;
1059         u64 timestamp0, timestamp = sample->time;
1060         int cpu = sample->cpu, err = -1;
1061         s64 delta;
1062
1063         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1064
1065         timestamp0 = sched->cpu_last_switched[cpu];
1066         sched->cpu_last_switched[cpu] = timestamp;
1067         if (timestamp0)
1068                 delta = timestamp - timestamp0;
1069         else
1070                 delta = 0;
1071
1072         if (delta < 0) {
1073                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1074                 return -1;
1075         }
1076
1077         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1078         sched_in = machine__findnew_thread(machine, -1, next_pid);
1079         if (sched_out == NULL || sched_in == NULL)
1080                 goto out_put;
1081
1082         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1083         if (!out_events) {
1084                 if (thread_atoms_insert(sched, sched_out))
1085                         goto out_put;
1086                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1087                 if (!out_events) {
1088                         pr_err("out-event: Internal tree error");
1089                         goto out_put;
1090                 }
1091         }
1092         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1093                 return -1;
1094
1095         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1096         if (!in_events) {
1097                 if (thread_atoms_insert(sched, sched_in))
1098                         goto out_put;
1099                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1100                 if (!in_events) {
1101                         pr_err("in-event: Internal tree error");
1102                         goto out_put;
1103                 }
1104                 /*
1105                  * Take came in we have not heard about yet,
1106                  * add in an initial atom in runnable state:
1107                  */
1108                 if (add_sched_out_event(in_events, 'R', timestamp))
1109                         goto out_put;
1110         }
1111         add_sched_in_event(in_events, timestamp);
1112         err = 0;
1113 out_put:
1114         thread__put(sched_out);
1115         thread__put(sched_in);
1116         return err;
1117 }
1118
1119 static int latency_runtime_event(struct perf_sched *sched,
1120                                  struct perf_evsel *evsel,
1121                                  struct perf_sample *sample,
1122                                  struct machine *machine)
1123 {
1124         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1125         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1126         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1127         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1128         u64 timestamp = sample->time;
1129         int cpu = sample->cpu, err = -1;
1130
1131         if (thread == NULL)
1132                 return -1;
1133
1134         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1135         if (!atoms) {
1136                 if (thread_atoms_insert(sched, thread))
1137                         goto out_put;
1138                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1139                 if (!atoms) {
1140                         pr_err("in-event: Internal tree error");
1141                         goto out_put;
1142                 }
1143                 if (add_sched_out_event(atoms, 'R', timestamp))
1144                         goto out_put;
1145         }
1146
1147         add_runtime_event(atoms, runtime, timestamp);
1148         err = 0;
1149 out_put:
1150         thread__put(thread);
1151         return err;
1152 }
1153
1154 static int latency_wakeup_event(struct perf_sched *sched,
1155                                 struct perf_evsel *evsel,
1156                                 struct perf_sample *sample,
1157                                 struct machine *machine)
1158 {
1159         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1160         struct work_atoms *atoms;
1161         struct work_atom *atom;
1162         struct thread *wakee;
1163         u64 timestamp = sample->time;
1164         int err = -1;
1165
1166         wakee = machine__findnew_thread(machine, -1, pid);
1167         if (wakee == NULL)
1168                 return -1;
1169         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1170         if (!atoms) {
1171                 if (thread_atoms_insert(sched, wakee))
1172                         goto out_put;
1173                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1174                 if (!atoms) {
1175                         pr_err("wakeup-event: Internal tree error");
1176                         goto out_put;
1177                 }
1178                 if (add_sched_out_event(atoms, 'S', timestamp))
1179                         goto out_put;
1180         }
1181
1182         BUG_ON(list_empty(&atoms->work_list));
1183
1184         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1185
1186         /*
1187          * As we do not guarantee the wakeup event happens when
1188          * task is out of run queue, also may happen when task is
1189          * on run queue and wakeup only change ->state to TASK_RUNNING,
1190          * then we should not set the ->wake_up_time when wake up a
1191          * task which is on run queue.
1192          *
1193          * You WILL be missing events if you've recorded only
1194          * one CPU, or are only looking at only one, so don't
1195          * skip in this case.
1196          */
1197         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1198                 goto out_ok;
1199
1200         sched->nr_timestamps++;
1201         if (atom->sched_out_time > timestamp) {
1202                 sched->nr_unordered_timestamps++;
1203                 goto out_ok;
1204         }
1205
1206         atom->state = THREAD_WAIT_CPU;
1207         atom->wake_up_time = timestamp;
1208 out_ok:
1209         err = 0;
1210 out_put:
1211         thread__put(wakee);
1212         return err;
1213 }
1214
1215 static int latency_migrate_task_event(struct perf_sched *sched,
1216                                       struct perf_evsel *evsel,
1217                                       struct perf_sample *sample,
1218                                       struct machine *machine)
1219 {
1220         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1221         u64 timestamp = sample->time;
1222         struct work_atoms *atoms;
1223         struct work_atom *atom;
1224         struct thread *migrant;
1225         int err = -1;
1226
1227         /*
1228          * Only need to worry about migration when profiling one CPU.
1229          */
1230         if (sched->profile_cpu == -1)
1231                 return 0;
1232
1233         migrant = machine__findnew_thread(machine, -1, pid);
1234         if (migrant == NULL)
1235                 return -1;
1236         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1237         if (!atoms) {
1238                 if (thread_atoms_insert(sched, migrant))
1239                         goto out_put;
1240                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1241                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1242                 if (!atoms) {
1243                         pr_err("migration-event: Internal tree error");
1244                         goto out_put;
1245                 }
1246                 if (add_sched_out_event(atoms, 'R', timestamp))
1247                         goto out_put;
1248         }
1249
1250         BUG_ON(list_empty(&atoms->work_list));
1251
1252         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1253         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1254
1255         sched->nr_timestamps++;
1256
1257         if (atom->sched_out_time > timestamp)
1258                 sched->nr_unordered_timestamps++;
1259         err = 0;
1260 out_put:
1261         thread__put(migrant);
1262         return err;
1263 }
1264
1265 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1266 {
1267         int i;
1268         int ret;
1269         u64 avg;
1270         char max_lat_at[32];
1271
1272         if (!work_list->nb_atoms)
1273                 return;
1274         /*
1275          * Ignore idle threads:
1276          */
1277         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1278                 return;
1279
1280         sched->all_runtime += work_list->total_runtime;
1281         sched->all_count   += work_list->nb_atoms;
1282
1283         if (work_list->num_merged > 1)
1284                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1285         else
1286                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1287
1288         for (i = 0; i < 24 - ret; i++)
1289                 printf(" ");
1290
1291         avg = work_list->total_lat / work_list->nb_atoms;
1292         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1293
1294         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1295               (double)work_list->total_runtime / NSEC_PER_MSEC,
1296                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1297                  (double)work_list->max_lat / NSEC_PER_MSEC,
1298                  max_lat_at);
1299 }
1300
1301 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1302 {
1303         if (l->thread == r->thread)
1304                 return 0;
1305         if (l->thread->tid < r->thread->tid)
1306                 return -1;
1307         if (l->thread->tid > r->thread->tid)
1308                 return 1;
1309         return (int)(l->thread - r->thread);
1310 }
1311
1312 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1313 {
1314         u64 avgl, avgr;
1315
1316         if (!l->nb_atoms)
1317                 return -1;
1318
1319         if (!r->nb_atoms)
1320                 return 1;
1321
1322         avgl = l->total_lat / l->nb_atoms;
1323         avgr = r->total_lat / r->nb_atoms;
1324
1325         if (avgl < avgr)
1326                 return -1;
1327         if (avgl > avgr)
1328                 return 1;
1329
1330         return 0;
1331 }
1332
1333 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1334 {
1335         if (l->max_lat < r->max_lat)
1336                 return -1;
1337         if (l->max_lat > r->max_lat)
1338                 return 1;
1339
1340         return 0;
1341 }
1342
1343 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1344 {
1345         if (l->nb_atoms < r->nb_atoms)
1346                 return -1;
1347         if (l->nb_atoms > r->nb_atoms)
1348                 return 1;
1349
1350         return 0;
1351 }
1352
1353 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1354 {
1355         if (l->total_runtime < r->total_runtime)
1356                 return -1;
1357         if (l->total_runtime > r->total_runtime)
1358                 return 1;
1359
1360         return 0;
1361 }
1362
1363 static int sort_dimension__add(const char *tok, struct list_head *list)
1364 {
1365         size_t i;
1366         static struct sort_dimension avg_sort_dimension = {
1367                 .name = "avg",
1368                 .cmp  = avg_cmp,
1369         };
1370         static struct sort_dimension max_sort_dimension = {
1371                 .name = "max",
1372                 .cmp  = max_cmp,
1373         };
1374         static struct sort_dimension pid_sort_dimension = {
1375                 .name = "pid",
1376                 .cmp  = pid_cmp,
1377         };
1378         static struct sort_dimension runtime_sort_dimension = {
1379                 .name = "runtime",
1380                 .cmp  = runtime_cmp,
1381         };
1382         static struct sort_dimension switch_sort_dimension = {
1383                 .name = "switch",
1384                 .cmp  = switch_cmp,
1385         };
1386         struct sort_dimension *available_sorts[] = {
1387                 &pid_sort_dimension,
1388                 &avg_sort_dimension,
1389                 &max_sort_dimension,
1390                 &switch_sort_dimension,
1391                 &runtime_sort_dimension,
1392         };
1393
1394         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1395                 if (!strcmp(available_sorts[i]->name, tok)) {
1396                         list_add_tail(&available_sorts[i]->list, list);
1397
1398                         return 0;
1399                 }
1400         }
1401
1402         return -1;
1403 }
1404
1405 static void perf_sched__sort_lat(struct perf_sched *sched)
1406 {
1407         struct rb_node *node;
1408         struct rb_root *root = &sched->atom_root;
1409 again:
1410         for (;;) {
1411                 struct work_atoms *data;
1412                 node = rb_first(root);
1413                 if (!node)
1414                         break;
1415
1416                 rb_erase(node, root);
1417                 data = rb_entry(node, struct work_atoms, node);
1418                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1419         }
1420         if (root == &sched->atom_root) {
1421                 root = &sched->merged_atom_root;
1422                 goto again;
1423         }
1424 }
1425
1426 static int process_sched_wakeup_event(struct perf_tool *tool,
1427                                       struct perf_evsel *evsel,
1428                                       struct perf_sample *sample,
1429                                       struct machine *machine)
1430 {
1431         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1432
1433         if (sched->tp_handler->wakeup_event)
1434                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1435
1436         return 0;
1437 }
1438
1439 union map_priv {
1440         void    *ptr;
1441         bool     color;
1442 };
1443
1444 static bool thread__has_color(struct thread *thread)
1445 {
1446         union map_priv priv = {
1447                 .ptr = thread__priv(thread),
1448         };
1449
1450         return priv.color;
1451 }
1452
1453 static struct thread*
1454 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1455 {
1456         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1457         union map_priv priv = {
1458                 .color = false,
1459         };
1460
1461         if (!sched->map.color_pids || !thread || thread__priv(thread))
1462                 return thread;
1463
1464         if (thread_map__has(sched->map.color_pids, tid))
1465                 priv.color = true;
1466
1467         thread__set_priv(thread, priv.ptr);
1468         return thread;
1469 }
1470
1471 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1472                             struct perf_sample *sample, struct machine *machine)
1473 {
1474         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1475         struct thread *sched_in;
1476         int new_shortname;
1477         u64 timestamp0, timestamp = sample->time;
1478         s64 delta;
1479         int i, this_cpu = sample->cpu;
1480         int cpus_nr;
1481         bool new_cpu = false;
1482         const char *color = PERF_COLOR_NORMAL;
1483         char stimestamp[32];
1484
1485         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1486
1487         if (this_cpu > sched->max_cpu)
1488                 sched->max_cpu = this_cpu;
1489
1490         if (sched->map.comp) {
1491                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1492                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1493                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1494                         new_cpu = true;
1495                 }
1496         } else
1497                 cpus_nr = sched->max_cpu;
1498
1499         timestamp0 = sched->cpu_last_switched[this_cpu];
1500         sched->cpu_last_switched[this_cpu] = timestamp;
1501         if (timestamp0)
1502                 delta = timestamp - timestamp0;
1503         else
1504                 delta = 0;
1505
1506         if (delta < 0) {
1507                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1508                 return -1;
1509         }
1510
1511         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1512         if (sched_in == NULL)
1513                 return -1;
1514
1515         sched->curr_thread[this_cpu] = thread__get(sched_in);
1516
1517         printf("  ");
1518
1519         new_shortname = 0;
1520         if (!sched_in->shortname[0]) {
1521                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1522                         /*
1523                          * Don't allocate a letter-number for swapper:0
1524                          * as a shortname. Instead, we use '.' for it.
1525                          */
1526                         sched_in->shortname[0] = '.';
1527                         sched_in->shortname[1] = ' ';
1528                 } else {
1529                         sched_in->shortname[0] = sched->next_shortname1;
1530                         sched_in->shortname[1] = sched->next_shortname2;
1531
1532                         if (sched->next_shortname1 < 'Z') {
1533                                 sched->next_shortname1++;
1534                         } else {
1535                                 sched->next_shortname1 = 'A';
1536                                 if (sched->next_shortname2 < '9')
1537                                         sched->next_shortname2++;
1538                                 else
1539                                         sched->next_shortname2 = '0';
1540                         }
1541                 }
1542                 new_shortname = 1;
1543         }
1544
1545         for (i = 0; i < cpus_nr; i++) {
1546                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1547                 struct thread *curr_thread = sched->curr_thread[cpu];
1548                 const char *pid_color = color;
1549                 const char *cpu_color = color;
1550
1551                 if (curr_thread && thread__has_color(curr_thread))
1552                         pid_color = COLOR_PIDS;
1553
1554                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1555                         continue;
1556
1557                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1558                         cpu_color = COLOR_CPUS;
1559
1560                 if (cpu != this_cpu)
1561                         color_fprintf(stdout, color, " ");
1562                 else
1563                         color_fprintf(stdout, cpu_color, "*");
1564
1565                 if (sched->curr_thread[cpu])
1566                         color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1567                 else
1568                         color_fprintf(stdout, color, "   ");
1569         }
1570
1571         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1572                 goto out;
1573
1574         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1575         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1576         if (new_shortname || (verbose > 0 && sched_in->tid)) {
1577                 const char *pid_color = color;
1578
1579                 if (thread__has_color(sched_in))
1580                         pid_color = COLOR_PIDS;
1581
1582                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1583                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1584         }
1585
1586         if (sched->map.comp && new_cpu)
1587                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1588
1589 out:
1590         color_fprintf(stdout, color, "\n");
1591
1592         thread__put(sched_in);
1593
1594         return 0;
1595 }
1596
1597 static int process_sched_switch_event(struct perf_tool *tool,
1598                                       struct perf_evsel *evsel,
1599                                       struct perf_sample *sample,
1600                                       struct machine *machine)
1601 {
1602         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1603         int this_cpu = sample->cpu, err = 0;
1604         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1605             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1606
1607         if (sched->curr_pid[this_cpu] != (u32)-1) {
1608                 /*
1609                  * Are we trying to switch away a PID that is
1610                  * not current?
1611                  */
1612                 if (sched->curr_pid[this_cpu] != prev_pid)
1613                         sched->nr_context_switch_bugs++;
1614         }
1615
1616         if (sched->tp_handler->switch_event)
1617                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1618
1619         sched->curr_pid[this_cpu] = next_pid;
1620         return err;
1621 }
1622
1623 static int process_sched_runtime_event(struct perf_tool *tool,
1624                                        struct perf_evsel *evsel,
1625                                        struct perf_sample *sample,
1626                                        struct machine *machine)
1627 {
1628         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1629
1630         if (sched->tp_handler->runtime_event)
1631                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1632
1633         return 0;
1634 }
1635
1636 static int perf_sched__process_fork_event(struct perf_tool *tool,
1637                                           union perf_event *event,
1638                                           struct perf_sample *sample,
1639                                           struct machine *machine)
1640 {
1641         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1642
1643         /* run the fork event through the perf machineruy */
1644         perf_event__process_fork(tool, event, sample, machine);
1645
1646         /* and then run additional processing needed for this command */
1647         if (sched->tp_handler->fork_event)
1648                 return sched->tp_handler->fork_event(sched, event, machine);
1649
1650         return 0;
1651 }
1652
1653 static int process_sched_migrate_task_event(struct perf_tool *tool,
1654                                             struct perf_evsel *evsel,
1655                                             struct perf_sample *sample,
1656                                             struct machine *machine)
1657 {
1658         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1659
1660         if (sched->tp_handler->migrate_task_event)
1661                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1662
1663         return 0;
1664 }
1665
1666 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1667                                   struct perf_evsel *evsel,
1668                                   struct perf_sample *sample,
1669                                   struct machine *machine);
1670
1671 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1672                                                  union perf_event *event __maybe_unused,
1673                                                  struct perf_sample *sample,
1674                                                  struct perf_evsel *evsel,
1675                                                  struct machine *machine)
1676 {
1677         int err = 0;
1678
1679         if (evsel->handler != NULL) {
1680                 tracepoint_handler f = evsel->handler;
1681                 err = f(tool, evsel, sample, machine);
1682         }
1683
1684         return err;
1685 }
1686
1687 static int perf_sched__read_events(struct perf_sched *sched)
1688 {
1689         const struct perf_evsel_str_handler handlers[] = {
1690                 { "sched:sched_switch",       process_sched_switch_event, },
1691                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1692                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1693                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1694                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1695         };
1696         struct perf_session *session;
1697         struct perf_data_file file = {
1698                 .path = input_name,
1699                 .mode = PERF_DATA_MODE_READ,
1700                 .force = sched->force,
1701         };
1702         int rc = -1;
1703
1704         session = perf_session__new(&file, false, &sched->tool);
1705         if (session == NULL) {
1706                 pr_debug("No Memory for session\n");
1707                 return -1;
1708         }
1709
1710         symbol__init(&session->header.env);
1711
1712         if (perf_session__set_tracepoints_handlers(session, handlers))
1713                 goto out_delete;
1714
1715         if (perf_session__has_traces(session, "record -R")) {
1716                 int err = perf_session__process_events(session);
1717                 if (err) {
1718                         pr_err("Failed to process events, error %d", err);
1719                         goto out_delete;
1720                 }
1721
1722                 sched->nr_events      = session->evlist->stats.nr_events[0];
1723                 sched->nr_lost_events = session->evlist->stats.total_lost;
1724                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1725         }
1726
1727         rc = 0;
1728 out_delete:
1729         perf_session__delete(session);
1730         return rc;
1731 }
1732
1733 /*
1734  * scheduling times are printed as msec.usec
1735  */
1736 static inline void print_sched_time(unsigned long long nsecs, int width)
1737 {
1738         unsigned long msecs;
1739         unsigned long usecs;
1740
1741         msecs  = nsecs / NSEC_PER_MSEC;
1742         nsecs -= msecs * NSEC_PER_MSEC;
1743         usecs  = nsecs / NSEC_PER_USEC;
1744         printf("%*lu.%03lu ", width, msecs, usecs);
1745 }
1746
1747 /*
1748  * returns runtime data for event, allocating memory for it the
1749  * first time it is used.
1750  */
1751 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1752 {
1753         struct evsel_runtime *r = evsel->priv;
1754
1755         if (r == NULL) {
1756                 r = zalloc(sizeof(struct evsel_runtime));
1757                 evsel->priv = r;
1758         }
1759
1760         return r;
1761 }
1762
1763 /*
1764  * save last time event was seen per cpu
1765  */
1766 static void perf_evsel__save_time(struct perf_evsel *evsel,
1767                                   u64 timestamp, u32 cpu)
1768 {
1769         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1770
1771         if (r == NULL)
1772                 return;
1773
1774         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1775                 int i, n = __roundup_pow_of_two(cpu+1);
1776                 void *p = r->last_time;
1777
1778                 p = realloc(r->last_time, n * sizeof(u64));
1779                 if (!p)
1780                         return;
1781
1782                 r->last_time = p;
1783                 for (i = r->ncpu; i < n; ++i)
1784                         r->last_time[i] = (u64) 0;
1785
1786                 r->ncpu = n;
1787         }
1788
1789         r->last_time[cpu] = timestamp;
1790 }
1791
1792 /* returns last time this event was seen on the given cpu */
1793 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1794 {
1795         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1796
1797         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1798                 return 0;
1799
1800         return r->last_time[cpu];
1801 }
1802
1803 static int comm_width = 30;
1804
1805 static char *timehist_get_commstr(struct thread *thread)
1806 {
1807         static char str[32];
1808         const char *comm = thread__comm_str(thread);
1809         pid_t tid = thread->tid;
1810         pid_t pid = thread->pid_;
1811         int n;
1812
1813         if (pid == 0)
1814                 n = scnprintf(str, sizeof(str), "%s", comm);
1815
1816         else if (tid != pid)
1817                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1818
1819         else
1820                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1821
1822         if (n > comm_width)
1823                 comm_width = n;
1824
1825         return str;
1826 }
1827
1828 static void timehist_header(struct perf_sched *sched)
1829 {
1830         u32 ncpus = sched->max_cpu + 1;
1831         u32 i, j;
1832
1833         printf("%15s %6s ", "time", "cpu");
1834
1835         if (sched->show_cpu_visual) {
1836                 printf(" ");
1837                 for (i = 0, j = 0; i < ncpus; ++i) {
1838                         printf("%x", j++);
1839                         if (j > 15)
1840                                 j = 0;
1841                 }
1842                 printf(" ");
1843         }
1844
1845         printf(" %-*s  %9s  %9s  %9s", comm_width,
1846                 "task name", "wait time", "sch delay", "run time");
1847
1848         if (sched->show_state)
1849                 printf("  %s", "state");
1850
1851         printf("\n");
1852
1853         /*
1854          * units row
1855          */
1856         printf("%15s %-6s ", "", "");
1857
1858         if (sched->show_cpu_visual)
1859                 printf(" %*s ", ncpus, "");
1860
1861         printf(" %-*s  %9s  %9s  %9s", comm_width,
1862                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1863
1864         if (sched->show_state)
1865                 printf("  %5s", "");
1866
1867         printf("\n");
1868
1869         /*
1870          * separator
1871          */
1872         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1873
1874         if (sched->show_cpu_visual)
1875                 printf(" %.*s ", ncpus, graph_dotted_line);
1876
1877         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1878                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1879                 graph_dotted_line);
1880
1881         if (sched->show_state)
1882                 printf("  %.5s", graph_dotted_line);
1883
1884         printf("\n");
1885 }
1886
1887 static char task_state_char(struct thread *thread, int state)
1888 {
1889         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1890         unsigned bit = state ? ffs(state) : 0;
1891
1892         /* 'I' for idle */
1893         if (thread->tid == 0)
1894                 return 'I';
1895
1896         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1897 }
1898
1899 static void timehist_print_sample(struct perf_sched *sched,
1900                                   struct perf_sample *sample,
1901                                   struct addr_location *al,
1902                                   struct thread *thread,
1903                                   u64 t, int state)
1904 {
1905         struct thread_runtime *tr = thread__priv(thread);
1906         u32 max_cpus = sched->max_cpu + 1;
1907         char tstr[64];
1908         u64 wait_time;
1909
1910         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1911         printf("%15s [%04d] ", tstr, sample->cpu);
1912
1913         if (sched->show_cpu_visual) {
1914                 u32 i;
1915                 char c;
1916
1917                 printf(" ");
1918                 for (i = 0; i < max_cpus; ++i) {
1919                         /* flag idle times with 'i'; others are sched events */
1920                         if (i == sample->cpu)
1921                                 c = (thread->tid == 0) ? 'i' : 's';
1922                         else
1923                                 c = ' ';
1924                         printf("%c", c);
1925                 }
1926                 printf(" ");
1927         }
1928
1929         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1930
1931         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
1932         print_sched_time(wait_time, 6);
1933
1934         print_sched_time(tr->dt_delay, 6);
1935         print_sched_time(tr->dt_run, 6);
1936
1937         if (sched->show_state)
1938                 printf(" %5c ", task_state_char(thread, state));
1939
1940         if (sched->show_wakeups)
1941                 printf("  %-*s", comm_width, "");
1942
1943         if (thread->tid == 0)
1944                 goto out;
1945
1946         if (sched->show_callchain)
1947                 printf("  ");
1948
1949         sample__fprintf_sym(sample, al, 0,
1950                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1951                             EVSEL__PRINT_CALLCHAIN_ARROW |
1952                             EVSEL__PRINT_SKIP_IGNORED,
1953                             &callchain_cursor, stdout);
1954
1955 out:
1956         printf("\n");
1957 }
1958
1959 /*
1960  * Explanation of delta-time stats:
1961  *
1962  *            t = time of current schedule out event
1963  *        tprev = time of previous sched out event
1964  *                also time of schedule-in event for current task
1965  *    last_time = time of last sched change event for current task
1966  *                (i.e, time process was last scheduled out)
1967  * ready_to_run = time of wakeup for current task
1968  *
1969  * -----|------------|------------|------------|------
1970  *    last         ready        tprev          t
1971  *    time         to run
1972  *
1973  *      |-------- dt_wait --------|
1974  *                   |- dt_delay -|-- dt_run --|
1975  *
1976  *   dt_run = run time of current task
1977  *  dt_wait = time between last schedule out event for task and tprev
1978  *            represents time spent off the cpu
1979  * dt_delay = time between wakeup and schedule-in of task
1980  */
1981
1982 static void timehist_update_runtime_stats(struct thread_runtime *r,
1983                                          u64 t, u64 tprev)
1984 {
1985         r->dt_delay   = 0;
1986         r->dt_sleep   = 0;
1987         r->dt_iowait  = 0;
1988         r->dt_preempt = 0;
1989         r->dt_run     = 0;
1990
1991         if (tprev) {
1992                 r->dt_run = t - tprev;
1993                 if (r->ready_to_run) {
1994                         if (r->ready_to_run > tprev)
1995                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
1996                         else
1997                                 r->dt_delay = tprev - r->ready_to_run;
1998                 }
1999
2000                 if (r->last_time > tprev)
2001                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2002                 else if (r->last_time) {
2003                         u64 dt_wait = tprev - r->last_time;
2004
2005                         if (r->last_state == TASK_RUNNING)
2006                                 r->dt_preempt = dt_wait;
2007                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2008                                 r->dt_iowait = dt_wait;
2009                         else
2010                                 r->dt_sleep = dt_wait;
2011                 }
2012         }
2013
2014         update_stats(&r->run_stats, r->dt_run);
2015
2016         r->total_run_time     += r->dt_run;
2017         r->total_delay_time   += r->dt_delay;
2018         r->total_sleep_time   += r->dt_sleep;
2019         r->total_iowait_time  += r->dt_iowait;
2020         r->total_preempt_time += r->dt_preempt;
2021 }
2022
2023 static bool is_idle_sample(struct perf_sample *sample,
2024                            struct perf_evsel *evsel)
2025 {
2026         /* pid 0 == swapper == idle task */
2027         if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2028                 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2029
2030         return sample->pid == 0;
2031 }
2032
2033 static void save_task_callchain(struct perf_sched *sched,
2034                                 struct perf_sample *sample,
2035                                 struct perf_evsel *evsel,
2036                                 struct machine *machine)
2037 {
2038         struct callchain_cursor *cursor = &callchain_cursor;
2039         struct thread *thread;
2040
2041         /* want main thread for process - has maps */
2042         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2043         if (thread == NULL) {
2044                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2045                 return;
2046         }
2047
2048         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2049                 return;
2050
2051         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2052                                       NULL, NULL, sched->max_stack + 2) != 0) {
2053                 if (verbose > 0)
2054                         error("Failed to resolve callchain. Skipping\n");
2055
2056                 return;
2057         }
2058
2059         callchain_cursor_commit(cursor);
2060
2061         while (true) {
2062                 struct callchain_cursor_node *node;
2063                 struct symbol *sym;
2064
2065                 node = callchain_cursor_current(cursor);
2066                 if (node == NULL)
2067                         break;
2068
2069                 sym = node->sym;
2070                 if (sym) {
2071                         if (!strcmp(sym->name, "schedule") ||
2072                             !strcmp(sym->name, "__schedule") ||
2073                             !strcmp(sym->name, "preempt_schedule"))
2074                                 sym->ignore = 1;
2075                 }
2076
2077                 callchain_cursor_advance(cursor);
2078         }
2079 }
2080
2081 static int init_idle_thread(struct thread *thread)
2082 {
2083         struct idle_thread_runtime *itr;
2084
2085         thread__set_comm(thread, idle_comm, 0);
2086
2087         itr = zalloc(sizeof(*itr));
2088         if (itr == NULL)
2089                 return -ENOMEM;
2090
2091         init_stats(&itr->tr.run_stats);
2092         callchain_init(&itr->callchain);
2093         callchain_cursor_reset(&itr->cursor);
2094         thread__set_priv(thread, itr);
2095
2096         return 0;
2097 }
2098
2099 /*
2100  * Track idle stats per cpu by maintaining a local thread
2101  * struct for the idle task on each cpu.
2102  */
2103 static int init_idle_threads(int ncpu)
2104 {
2105         int i, ret;
2106
2107         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2108         if (!idle_threads)
2109                 return -ENOMEM;
2110
2111         idle_max_cpu = ncpu;
2112
2113         /* allocate the actual thread struct if needed */
2114         for (i = 0; i < ncpu; ++i) {
2115                 idle_threads[i] = thread__new(0, 0);
2116                 if (idle_threads[i] == NULL)
2117                         return -ENOMEM;
2118
2119                 ret = init_idle_thread(idle_threads[i]);
2120                 if (ret < 0)
2121                         return ret;
2122         }
2123
2124         return 0;
2125 }
2126
2127 static void free_idle_threads(void)
2128 {
2129         int i;
2130
2131         if (idle_threads == NULL)
2132                 return;
2133
2134         for (i = 0; i < idle_max_cpu; ++i) {
2135                 if ((idle_threads[i]))
2136                         thread__delete(idle_threads[i]);
2137         }
2138
2139         free(idle_threads);
2140 }
2141
2142 static struct thread *get_idle_thread(int cpu)
2143 {
2144         /*
2145          * expand/allocate array of pointers to local thread
2146          * structs if needed
2147          */
2148         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2149                 int i, j = __roundup_pow_of_two(cpu+1);
2150                 void *p;
2151
2152                 p = realloc(idle_threads, j * sizeof(struct thread *));
2153                 if (!p)
2154                         return NULL;
2155
2156                 idle_threads = (struct thread **) p;
2157                 for (i = idle_max_cpu; i < j; ++i)
2158                         idle_threads[i] = NULL;
2159
2160                 idle_max_cpu = j;
2161         }
2162
2163         /* allocate a new thread struct if needed */
2164         if (idle_threads[cpu] == NULL) {
2165                 idle_threads[cpu] = thread__new(0, 0);
2166
2167                 if (idle_threads[cpu]) {
2168                         if (init_idle_thread(idle_threads[cpu]) < 0)
2169                                 return NULL;
2170                 }
2171         }
2172
2173         return idle_threads[cpu];
2174 }
2175
2176 static void save_idle_callchain(struct idle_thread_runtime *itr,
2177                                 struct perf_sample *sample)
2178 {
2179         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2180                 return;
2181
2182         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2183 }
2184
2185 /*
2186  * handle runtime stats saved per thread
2187  */
2188 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2189 {
2190         struct thread_runtime *r;
2191
2192         r = zalloc(sizeof(struct thread_runtime));
2193         if (!r)
2194                 return NULL;
2195
2196         init_stats(&r->run_stats);
2197         thread__set_priv(thread, r);
2198
2199         return r;
2200 }
2201
2202 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2203 {
2204         struct thread_runtime *tr;
2205
2206         tr = thread__priv(thread);
2207         if (tr == NULL) {
2208                 tr = thread__init_runtime(thread);
2209                 if (tr == NULL)
2210                         pr_debug("Failed to malloc memory for runtime data.\n");
2211         }
2212
2213         return tr;
2214 }
2215
2216 static struct thread *timehist_get_thread(struct perf_sched *sched,
2217                                           struct perf_sample *sample,
2218                                           struct machine *machine,
2219                                           struct perf_evsel *evsel)
2220 {
2221         struct thread *thread;
2222
2223         if (is_idle_sample(sample, evsel)) {
2224                 thread = get_idle_thread(sample->cpu);
2225                 if (thread == NULL)
2226                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2227
2228         } else {
2229                 /* there were samples with tid 0 but non-zero pid */
2230                 thread = machine__findnew_thread(machine, sample->pid,
2231                                                  sample->tid ?: sample->pid);
2232                 if (thread == NULL) {
2233                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2234                                  sample->tid);
2235                 }
2236
2237                 save_task_callchain(sched, sample, evsel, machine);
2238                 if (sched->idle_hist) {
2239                         struct thread *idle;
2240                         struct idle_thread_runtime *itr;
2241
2242                         idle = get_idle_thread(sample->cpu);
2243                         if (idle == NULL) {
2244                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2245                                 return NULL;
2246                         }
2247
2248                         itr = thread__priv(idle);
2249                         if (itr == NULL)
2250                                 return NULL;
2251
2252                         itr->last_thread = thread;
2253
2254                         /* copy task callchain when entering to idle */
2255                         if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2256                                 save_idle_callchain(itr, sample);
2257                 }
2258         }
2259
2260         return thread;
2261 }
2262
2263 static bool timehist_skip_sample(struct perf_sched *sched,
2264                                  struct thread *thread,
2265                                  struct perf_evsel *evsel,
2266                                  struct perf_sample *sample)
2267 {
2268         bool rc = false;
2269
2270         if (thread__is_filtered(thread)) {
2271                 rc = true;
2272                 sched->skipped_samples++;
2273         }
2274
2275         if (sched->idle_hist) {
2276                 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2277                         rc = true;
2278                 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2279                          perf_evsel__intval(evsel, sample, "next_pid") != 0)
2280                         rc = true;
2281         }
2282
2283         return rc;
2284 }
2285
2286 static void timehist_print_wakeup_event(struct perf_sched *sched,
2287                                         struct perf_evsel *evsel,
2288                                         struct perf_sample *sample,
2289                                         struct machine *machine,
2290                                         struct thread *awakened)
2291 {
2292         struct thread *thread;
2293         char tstr[64];
2294
2295         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2296         if (thread == NULL)
2297                 return;
2298
2299         /* show wakeup unless both awakee and awaker are filtered */
2300         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2301             timehist_skip_sample(sched, awakened, evsel, sample)) {
2302                 return;
2303         }
2304
2305         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2306         printf("%15s [%04d] ", tstr, sample->cpu);
2307         if (sched->show_cpu_visual)
2308                 printf(" %*s ", sched->max_cpu + 1, "");
2309
2310         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2311
2312         /* dt spacer */
2313         printf("  %9s  %9s  %9s ", "", "", "");
2314
2315         printf("awakened: %s", timehist_get_commstr(awakened));
2316
2317         printf("\n");
2318 }
2319
2320 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2321                                        union perf_event *event __maybe_unused,
2322                                        struct perf_evsel *evsel,
2323                                        struct perf_sample *sample,
2324                                        struct machine *machine)
2325 {
2326         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2327         struct thread *thread;
2328         struct thread_runtime *tr = NULL;
2329         /* want pid of awakened task not pid in sample */
2330         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2331
2332         thread = machine__findnew_thread(machine, 0, pid);
2333         if (thread == NULL)
2334                 return -1;
2335
2336         tr = thread__get_runtime(thread);
2337         if (tr == NULL)
2338                 return -1;
2339
2340         if (tr->ready_to_run == 0)
2341                 tr->ready_to_run = sample->time;
2342
2343         /* show wakeups if requested */
2344         if (sched->show_wakeups &&
2345             !perf_time__skip_sample(&sched->ptime, sample->time))
2346                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2347
2348         return 0;
2349 }
2350
2351 static void timehist_print_migration_event(struct perf_sched *sched,
2352                                         struct perf_evsel *evsel,
2353                                         struct perf_sample *sample,
2354                                         struct machine *machine,
2355                                         struct thread *migrated)
2356 {
2357         struct thread *thread;
2358         char tstr[64];
2359         u32 max_cpus = sched->max_cpu + 1;
2360         u32 ocpu, dcpu;
2361
2362         if (sched->summary_only)
2363                 return;
2364
2365         max_cpus = sched->max_cpu + 1;
2366         ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2367         dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2368
2369         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2370         if (thread == NULL)
2371                 return;
2372
2373         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2374             timehist_skip_sample(sched, migrated, evsel, sample)) {
2375                 return;
2376         }
2377
2378         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2379         printf("%15s [%04d] ", tstr, sample->cpu);
2380
2381         if (sched->show_cpu_visual) {
2382                 u32 i;
2383                 char c;
2384
2385                 printf("  ");
2386                 for (i = 0; i < max_cpus; ++i) {
2387                         c = (i == sample->cpu) ? 'm' : ' ';
2388                         printf("%c", c);
2389                 }
2390                 printf("  ");
2391         }
2392
2393         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2394
2395         /* dt spacer */
2396         printf("  %9s  %9s  %9s ", "", "", "");
2397
2398         printf("migrated: %s", timehist_get_commstr(migrated));
2399         printf(" cpu %d => %d", ocpu, dcpu);
2400
2401         printf("\n");
2402 }
2403
2404 static int timehist_migrate_task_event(struct perf_tool *tool,
2405                                        union perf_event *event __maybe_unused,
2406                                        struct perf_evsel *evsel,
2407                                        struct perf_sample *sample,
2408                                        struct machine *machine)
2409 {
2410         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2411         struct thread *thread;
2412         struct thread_runtime *tr = NULL;
2413         /* want pid of migrated task not pid in sample */
2414         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2415
2416         thread = machine__findnew_thread(machine, 0, pid);
2417         if (thread == NULL)
2418                 return -1;
2419
2420         tr = thread__get_runtime(thread);
2421         if (tr == NULL)
2422                 return -1;
2423
2424         tr->migrations++;
2425
2426         /* show migrations if requested */
2427         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2428
2429         return 0;
2430 }
2431
2432 static int timehist_sched_change_event(struct perf_tool *tool,
2433                                        union perf_event *event,
2434                                        struct perf_evsel *evsel,
2435                                        struct perf_sample *sample,
2436                                        struct machine *machine)
2437 {
2438         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2439         struct perf_time_interval *ptime = &sched->ptime;
2440         struct addr_location al;
2441         struct thread *thread;
2442         struct thread_runtime *tr = NULL;
2443         u64 tprev, t = sample->time;
2444         int rc = 0;
2445         int state = perf_evsel__intval(evsel, sample, "prev_state");
2446
2447
2448         if (machine__resolve(machine, &al, sample) < 0) {
2449                 pr_err("problem processing %d event. skipping it\n",
2450                        event->header.type);
2451                 rc = -1;
2452                 goto out;
2453         }
2454
2455         thread = timehist_get_thread(sched, sample, machine, evsel);
2456         if (thread == NULL) {
2457                 rc = -1;
2458                 goto out;
2459         }
2460
2461         if (timehist_skip_sample(sched, thread, evsel, sample))
2462                 goto out;
2463
2464         tr = thread__get_runtime(thread);
2465         if (tr == NULL) {
2466                 rc = -1;
2467                 goto out;
2468         }
2469
2470         tprev = perf_evsel__get_time(evsel, sample->cpu);
2471
2472         /*
2473          * If start time given:
2474          * - sample time is under window user cares about - skip sample
2475          * - tprev is under window user cares about  - reset to start of window
2476          */
2477         if (ptime->start && ptime->start > t)
2478                 goto out;
2479
2480         if (tprev && ptime->start > tprev)
2481                 tprev = ptime->start;
2482
2483         /*
2484          * If end time given:
2485          * - previous sched event is out of window - we are done
2486          * - sample time is beyond window user cares about - reset it
2487          *   to close out stats for time window interest
2488          */
2489         if (ptime->end) {
2490                 if (tprev > ptime->end)
2491                         goto out;
2492
2493                 if (t > ptime->end)
2494                         t = ptime->end;
2495         }
2496
2497         if (!sched->idle_hist || thread->tid == 0) {
2498                 timehist_update_runtime_stats(tr, t, tprev);
2499
2500                 if (sched->idle_hist) {
2501                         struct idle_thread_runtime *itr = (void *)tr;
2502                         struct thread_runtime *last_tr;
2503
2504                         BUG_ON(thread->tid != 0);
2505
2506                         if (itr->last_thread == NULL)
2507                                 goto out;
2508
2509                         /* add current idle time as last thread's runtime */
2510                         last_tr = thread__get_runtime(itr->last_thread);
2511                         if (last_tr == NULL)
2512                                 goto out;
2513
2514                         timehist_update_runtime_stats(last_tr, t, tprev);
2515                         /*
2516                          * remove delta time of last thread as it's not updated
2517                          * and otherwise it will show an invalid value next
2518                          * time.  we only care total run time and run stat.
2519                          */
2520                         last_tr->dt_run = 0;
2521                         last_tr->dt_delay = 0;
2522                         last_tr->dt_sleep = 0;
2523                         last_tr->dt_iowait = 0;
2524                         last_tr->dt_preempt = 0;
2525
2526                         if (itr->cursor.nr)
2527                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2528
2529                         itr->last_thread = NULL;
2530                 }
2531         }
2532
2533         if (!sched->summary_only)
2534                 timehist_print_sample(sched, sample, &al, thread, t, state);
2535
2536 out:
2537         if (sched->hist_time.start == 0 && t >= ptime->start)
2538                 sched->hist_time.start = t;
2539         if (ptime->end == 0 || t <= ptime->end)
2540                 sched->hist_time.end = t;
2541
2542         if (tr) {
2543                 /* time of this sched_switch event becomes last time task seen */
2544                 tr->last_time = sample->time;
2545
2546                 /* last state is used to determine where to account wait time */
2547                 tr->last_state = state;
2548
2549                 /* sched out event for task so reset ready to run time */
2550                 tr->ready_to_run = 0;
2551         }
2552
2553         perf_evsel__save_time(evsel, sample->time, sample->cpu);
2554
2555         return rc;
2556 }
2557
2558 static int timehist_sched_switch_event(struct perf_tool *tool,
2559                              union perf_event *event,
2560                              struct perf_evsel *evsel,
2561                              struct perf_sample *sample,
2562                              struct machine *machine __maybe_unused)
2563 {
2564         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2565 }
2566
2567 static int process_lost(struct perf_tool *tool __maybe_unused,
2568                         union perf_event *event,
2569                         struct perf_sample *sample,
2570                         struct machine *machine __maybe_unused)
2571 {
2572         char tstr[64];
2573
2574         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2575         printf("%15s ", tstr);
2576         printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2577
2578         return 0;
2579 }
2580
2581
2582 static void print_thread_runtime(struct thread *t,
2583                                  struct thread_runtime *r)
2584 {
2585         double mean = avg_stats(&r->run_stats);
2586         float stddev;
2587
2588         printf("%*s   %5d  %9" PRIu64 " ",
2589                comm_width, timehist_get_commstr(t), t->ppid,
2590                (u64) r->run_stats.n);
2591
2592         print_sched_time(r->total_run_time, 8);
2593         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2594         print_sched_time(r->run_stats.min, 6);
2595         printf(" ");
2596         print_sched_time((u64) mean, 6);
2597         printf(" ");
2598         print_sched_time(r->run_stats.max, 6);
2599         printf("  ");
2600         printf("%5.2f", stddev);
2601         printf("   %5" PRIu64, r->migrations);
2602         printf("\n");
2603 }
2604
2605 static void print_thread_waittime(struct thread *t,
2606                                   struct thread_runtime *r)
2607 {
2608         printf("%*s   %5d  %9" PRIu64 " ",
2609                comm_width, timehist_get_commstr(t), t->ppid,
2610                (u64) r->run_stats.n);
2611
2612         print_sched_time(r->total_run_time, 8);
2613         print_sched_time(r->total_sleep_time, 6);
2614         printf(" ");
2615         print_sched_time(r->total_iowait_time, 6);
2616         printf(" ");
2617         print_sched_time(r->total_preempt_time, 6);
2618         printf(" ");
2619         print_sched_time(r->total_delay_time, 6);
2620         printf("\n");
2621 }
2622
2623 struct total_run_stats {
2624         struct perf_sched *sched;
2625         u64  sched_count;
2626         u64  task_count;
2627         u64  total_run_time;
2628 };
2629
2630 static int __show_thread_runtime(struct thread *t, void *priv)
2631 {
2632         struct total_run_stats *stats = priv;
2633         struct thread_runtime *r;
2634
2635         if (thread__is_filtered(t))
2636                 return 0;
2637
2638         r = thread__priv(t);
2639         if (r && r->run_stats.n) {
2640                 stats->task_count++;
2641                 stats->sched_count += r->run_stats.n;
2642                 stats->total_run_time += r->total_run_time;
2643
2644                 if (stats->sched->show_state)
2645                         print_thread_waittime(t, r);
2646                 else
2647                         print_thread_runtime(t, r);
2648         }
2649
2650         return 0;
2651 }
2652
2653 static int show_thread_runtime(struct thread *t, void *priv)
2654 {
2655         if (t->dead)
2656                 return 0;
2657
2658         return __show_thread_runtime(t, priv);
2659 }
2660
2661 static int show_deadthread_runtime(struct thread *t, void *priv)
2662 {
2663         if (!t->dead)
2664                 return 0;
2665
2666         return __show_thread_runtime(t, priv);
2667 }
2668
2669 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2670 {
2671         const char *sep = " <- ";
2672         struct callchain_list *chain;
2673         size_t ret = 0;
2674         char bf[1024];
2675         bool first;
2676
2677         if (node == NULL)
2678                 return 0;
2679
2680         ret = callchain__fprintf_folded(fp, node->parent);
2681         first = (ret == 0);
2682
2683         list_for_each_entry(chain, &node->val, list) {
2684                 if (chain->ip >= PERF_CONTEXT_MAX)
2685                         continue;
2686                 if (chain->ms.sym && chain->ms.sym->ignore)
2687                         continue;
2688                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2689                                callchain_list__sym_name(chain, bf, sizeof(bf),
2690                                                         false));
2691                 first = false;
2692         }
2693
2694         return ret;
2695 }
2696
2697 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2698 {
2699         size_t ret = 0;
2700         FILE *fp = stdout;
2701         struct callchain_node *chain;
2702         struct rb_node *rb_node = rb_first(root);
2703
2704         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2705         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2706                graph_dotted_line);
2707
2708         while (rb_node) {
2709                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2710                 rb_node = rb_next(rb_node);
2711
2712                 ret += fprintf(fp, "  ");
2713                 print_sched_time(chain->hit, 12);
2714                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2715                 ret += fprintf(fp, " %8d  ", chain->count);
2716                 ret += callchain__fprintf_folded(fp, chain);
2717                 ret += fprintf(fp, "\n");
2718         }
2719
2720         return ret;
2721 }
2722
2723 static void timehist_print_summary(struct perf_sched *sched,
2724                                    struct perf_session *session)
2725 {
2726         struct machine *m = &session->machines.host;
2727         struct total_run_stats totals;
2728         u64 task_count;
2729         struct thread *t;
2730         struct thread_runtime *r;
2731         int i;
2732         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2733
2734         memset(&totals, 0, sizeof(totals));
2735         totals.sched = sched;
2736
2737         if (sched->idle_hist) {
2738                 printf("\nIdle-time summary\n");
2739                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2740                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2741         } else if (sched->show_state) {
2742                 printf("\nWait-time summary\n");
2743                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2744                 printf("   run-time      sleep      iowait     preempt       delay\n");
2745         } else {
2746                 printf("\nRuntime summary\n");
2747                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2748                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2749         }
2750         printf("%*s            (count)  ", comm_width, "");
2751         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2752                sched->show_state ? "(msec)" : "%");
2753         printf("%.117s\n", graph_dotted_line);
2754
2755         machine__for_each_thread(m, show_thread_runtime, &totals);
2756         task_count = totals.task_count;
2757         if (!task_count)
2758                 printf("<no still running tasks>\n");
2759
2760         printf("\nTerminated tasks:\n");
2761         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2762         if (task_count == totals.task_count)
2763                 printf("<no terminated tasks>\n");
2764
2765         /* CPU idle stats not tracked when samples were skipped */
2766         if (sched->skipped_samples && !sched->idle_hist)
2767                 return;
2768
2769         printf("\nIdle stats:\n");
2770         for (i = 0; i < idle_max_cpu; ++i) {
2771                 t = idle_threads[i];
2772                 if (!t)
2773                         continue;
2774
2775                 r = thread__priv(t);
2776                 if (r && r->run_stats.n) {
2777                         totals.sched_count += r->run_stats.n;
2778                         printf("    CPU %2d idle for ", i);
2779                         print_sched_time(r->total_run_time, 6);
2780                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2781                 } else
2782                         printf("    CPU %2d idle entire time window\n", i);
2783         }
2784
2785         if (sched->idle_hist && symbol_conf.use_callchain) {
2786                 callchain_param.mode  = CHAIN_FOLDED;
2787                 callchain_param.value = CCVAL_PERIOD;
2788
2789                 callchain_register_param(&callchain_param);
2790
2791                 printf("\nIdle stats by callchain:\n");
2792                 for (i = 0; i < idle_max_cpu; ++i) {
2793                         struct idle_thread_runtime *itr;
2794
2795                         t = idle_threads[i];
2796                         if (!t)
2797                                 continue;
2798
2799                         itr = thread__priv(t);
2800                         if (itr == NULL)
2801                                 continue;
2802
2803                         callchain_param.sort(&itr->sorted_root, &itr->callchain,
2804                                              0, &callchain_param);
2805
2806                         printf("  CPU %2d:", i);
2807                         print_sched_time(itr->tr.total_run_time, 6);
2808                         printf(" msec\n");
2809                         timehist_print_idlehist_callchain(&itr->sorted_root);
2810                         printf("\n");
2811                 }
2812         }
2813
2814         printf("\n"
2815                "    Total number of unique tasks: %" PRIu64 "\n"
2816                "Total number of context switches: %" PRIu64 "\n",
2817                totals.task_count, totals.sched_count);
2818
2819         printf("           Total run time (msec): ");
2820         print_sched_time(totals.total_run_time, 2);
2821         printf("\n");
2822
2823         printf("    Total scheduling time (msec): ");
2824         print_sched_time(hist_time, 2);
2825         printf(" (x %d)\n", sched->max_cpu);
2826 }
2827
2828 typedef int (*sched_handler)(struct perf_tool *tool,
2829                           union perf_event *event,
2830                           struct perf_evsel *evsel,
2831                           struct perf_sample *sample,
2832                           struct machine *machine);
2833
2834 static int perf_timehist__process_sample(struct perf_tool *tool,
2835                                          union perf_event *event,
2836                                          struct perf_sample *sample,
2837                                          struct perf_evsel *evsel,
2838                                          struct machine *machine)
2839 {
2840         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2841         int err = 0;
2842         int this_cpu = sample->cpu;
2843
2844         if (this_cpu > sched->max_cpu)
2845                 sched->max_cpu = this_cpu;
2846
2847         if (evsel->handler != NULL) {
2848                 sched_handler f = evsel->handler;
2849
2850                 err = f(tool, event, evsel, sample, machine);
2851         }
2852
2853         return err;
2854 }
2855
2856 static int timehist_check_attr(struct perf_sched *sched,
2857                                struct perf_evlist *evlist)
2858 {
2859         struct perf_evsel *evsel;
2860         struct evsel_runtime *er;
2861
2862         list_for_each_entry(evsel, &evlist->entries, node) {
2863                 er = perf_evsel__get_runtime(evsel);
2864                 if (er == NULL) {
2865                         pr_err("Failed to allocate memory for evsel runtime data\n");
2866                         return -1;
2867                 }
2868
2869                 if (sched->show_callchain &&
2870                     !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2871                         pr_info("Samples do not have callchains.\n");
2872                         sched->show_callchain = 0;
2873                         symbol_conf.use_callchain = 0;
2874                 }
2875         }
2876
2877         return 0;
2878 }
2879
2880 static int perf_sched__timehist(struct perf_sched *sched)
2881 {
2882         const struct perf_evsel_str_handler handlers[] = {
2883                 { "sched:sched_switch",       timehist_sched_switch_event, },
2884                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2885                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2886         };
2887         const struct perf_evsel_str_handler migrate_handlers[] = {
2888                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2889         };
2890         struct perf_data_file file = {
2891                 .path = input_name,
2892                 .mode = PERF_DATA_MODE_READ,
2893                 .force = sched->force,
2894         };
2895
2896         struct perf_session *session;
2897         struct perf_evlist *evlist;
2898         int err = -1;
2899
2900         /*
2901          * event handlers for timehist option
2902          */
2903         sched->tool.sample       = perf_timehist__process_sample;
2904         sched->tool.mmap         = perf_event__process_mmap;
2905         sched->tool.comm         = perf_event__process_comm;
2906         sched->tool.exit         = perf_event__process_exit;
2907         sched->tool.fork         = perf_event__process_fork;
2908         sched->tool.lost         = process_lost;
2909         sched->tool.attr         = perf_event__process_attr;
2910         sched->tool.tracing_data = perf_event__process_tracing_data;
2911         sched->tool.build_id     = perf_event__process_build_id;
2912
2913         sched->tool.ordered_events = true;
2914         sched->tool.ordering_requires_timestamps = true;
2915
2916         symbol_conf.use_callchain = sched->show_callchain;
2917
2918         session = perf_session__new(&file, false, &sched->tool);
2919         if (session == NULL)
2920                 return -ENOMEM;
2921
2922         evlist = session->evlist;
2923
2924         symbol__init(&session->header.env);
2925
2926         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2927                 pr_err("Invalid time string\n");
2928                 return -EINVAL;
2929         }
2930
2931         if (timehist_check_attr(sched, evlist) != 0)
2932                 goto out;
2933
2934         setup_pager();
2935
2936         /* setup per-evsel handlers */
2937         if (perf_session__set_tracepoints_handlers(session, handlers))
2938                 goto out;
2939
2940         /* sched_switch event at a minimum needs to exist */
2941         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2942                                                   "sched:sched_switch")) {
2943                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2944                 goto out;
2945         }
2946
2947         if (sched->show_migrations &&
2948             perf_session__set_tracepoints_handlers(session, migrate_handlers))
2949                 goto out;
2950
2951         /* pre-allocate struct for per-CPU idle stats */
2952         sched->max_cpu = session->header.env.nr_cpus_online;
2953         if (sched->max_cpu == 0)
2954                 sched->max_cpu = 4;
2955         if (init_idle_threads(sched->max_cpu))
2956                 goto out;
2957
2958         /* summary_only implies summary option, but don't overwrite summary if set */
2959         if (sched->summary_only)
2960                 sched->summary = sched->summary_only;
2961
2962         if (!sched->summary_only)
2963                 timehist_header(sched);
2964
2965         err = perf_session__process_events(session);
2966         if (err) {
2967                 pr_err("Failed to process events, error %d", err);
2968                 goto out;
2969         }
2970
2971         sched->nr_events      = evlist->stats.nr_events[0];
2972         sched->nr_lost_events = evlist->stats.total_lost;
2973         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2974
2975         if (sched->summary)
2976                 timehist_print_summary(sched, session);
2977
2978 out:
2979         free_idle_threads();
2980         perf_session__delete(session);
2981
2982         return err;
2983 }
2984
2985
2986 static void print_bad_events(struct perf_sched *sched)
2987 {
2988         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
2989                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
2990                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
2991                         sched->nr_unordered_timestamps, sched->nr_timestamps);
2992         }
2993         if (sched->nr_lost_events && sched->nr_events) {
2994                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
2995                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
2996                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
2997         }
2998         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
2999                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3000                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3001                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3002                 if (sched->nr_lost_events)
3003                         printf(" (due to lost events?)");
3004                 printf("\n");
3005         }
3006 }
3007
3008 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3009 {
3010         struct rb_node **new = &(root->rb_node), *parent = NULL;
3011         struct work_atoms *this;
3012         const char *comm = thread__comm_str(data->thread), *this_comm;
3013
3014         while (*new) {
3015                 int cmp;
3016
3017                 this = container_of(*new, struct work_atoms, node);
3018                 parent = *new;
3019
3020                 this_comm = thread__comm_str(this->thread);
3021                 cmp = strcmp(comm, this_comm);
3022                 if (cmp > 0) {
3023                         new = &((*new)->rb_left);
3024                 } else if (cmp < 0) {
3025                         new = &((*new)->rb_right);
3026                 } else {
3027                         this->num_merged++;
3028                         this->total_runtime += data->total_runtime;
3029                         this->nb_atoms += data->nb_atoms;
3030                         this->total_lat += data->total_lat;
3031                         list_splice(&data->work_list, &this->work_list);
3032                         if (this->max_lat < data->max_lat) {
3033                                 this->max_lat = data->max_lat;
3034                                 this->max_lat_at = data->max_lat_at;
3035                         }
3036                         zfree(&data);
3037                         return;
3038                 }
3039         }
3040
3041         data->num_merged++;
3042         rb_link_node(&data->node, parent, new);
3043         rb_insert_color(&data->node, root);
3044 }
3045
3046 static void perf_sched__merge_lat(struct perf_sched *sched)
3047 {
3048         struct work_atoms *data;
3049         struct rb_node *node;
3050
3051         if (sched->skip_merge)
3052                 return;
3053
3054         while ((node = rb_first(&sched->atom_root))) {
3055                 rb_erase(node, &sched->atom_root);
3056                 data = rb_entry(node, struct work_atoms, node);
3057                 __merge_work_atoms(&sched->merged_atom_root, data);
3058         }
3059 }
3060
3061 static int perf_sched__lat(struct perf_sched *sched)
3062 {
3063         struct rb_node *next;
3064
3065         setup_pager();
3066
3067         if (perf_sched__read_events(sched))
3068                 return -1;
3069
3070         perf_sched__merge_lat(sched);
3071         perf_sched__sort_lat(sched);
3072
3073         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3074         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3075         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3076
3077         next = rb_first(&sched->sorted_atom_root);
3078
3079         while (next) {
3080                 struct work_atoms *work_list;
3081
3082                 work_list = rb_entry(next, struct work_atoms, node);
3083                 output_lat_thread(sched, work_list);
3084                 next = rb_next(next);
3085                 thread__zput(work_list->thread);
3086         }
3087
3088         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3089         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3090                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3091
3092         printf(" ---------------------------------------------------\n");
3093
3094         print_bad_events(sched);
3095         printf("\n");
3096
3097         return 0;
3098 }
3099
3100 static int setup_map_cpus(struct perf_sched *sched)
3101 {
3102         struct cpu_map *map;
3103
3104         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3105
3106         if (sched->map.comp) {
3107                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3108                 if (!sched->map.comp_cpus)
3109                         return -1;
3110         }
3111
3112         if (!sched->map.cpus_str)
3113                 return 0;
3114
3115         map = cpu_map__new(sched->map.cpus_str);
3116         if (!map) {
3117                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3118                 return -1;
3119         }
3120
3121         sched->map.cpus = map;
3122         return 0;
3123 }
3124
3125 static int setup_color_pids(struct perf_sched *sched)
3126 {
3127         struct thread_map *map;
3128
3129         if (!sched->map.color_pids_str)
3130                 return 0;
3131
3132         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3133         if (!map) {
3134                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3135                 return -1;
3136         }
3137
3138         sched->map.color_pids = map;
3139         return 0;
3140 }
3141
3142 static int setup_color_cpus(struct perf_sched *sched)
3143 {
3144         struct cpu_map *map;
3145
3146         if (!sched->map.color_cpus_str)
3147                 return 0;
3148
3149         map = cpu_map__new(sched->map.color_cpus_str);
3150         if (!map) {
3151                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3152                 return -1;
3153         }
3154
3155         sched->map.color_cpus = map;
3156         return 0;
3157 }
3158
3159 static int perf_sched__map(struct perf_sched *sched)
3160 {
3161         if (setup_map_cpus(sched))
3162                 return -1;
3163
3164         if (setup_color_pids(sched))
3165                 return -1;
3166
3167         if (setup_color_cpus(sched))
3168                 return -1;
3169
3170         setup_pager();
3171         if (perf_sched__read_events(sched))
3172                 return -1;
3173         print_bad_events(sched);
3174         return 0;
3175 }
3176
3177 static int perf_sched__replay(struct perf_sched *sched)
3178 {
3179         unsigned long i;
3180
3181         calibrate_run_measurement_overhead(sched);
3182         calibrate_sleep_measurement_overhead(sched);
3183
3184         test_calibrations(sched);
3185
3186         if (perf_sched__read_events(sched))
3187                 return -1;
3188
3189         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3190         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3191         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3192
3193         if (sched->targetless_wakeups)
3194                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3195         if (sched->multitarget_wakeups)
3196                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3197         if (sched->nr_run_events_optimized)
3198                 printf("run atoms optimized: %ld\n",
3199                         sched->nr_run_events_optimized);
3200
3201         print_task_traces(sched);
3202         add_cross_task_wakeups(sched);
3203
3204         create_tasks(sched);
3205         printf("------------------------------------------------------------\n");
3206         for (i = 0; i < sched->replay_repeat; i++)
3207                 run_one_test(sched);
3208
3209         return 0;
3210 }
3211
3212 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3213                           const char * const usage_msg[])
3214 {
3215         char *tmp, *tok, *str = strdup(sched->sort_order);
3216
3217         for (tok = strtok_r(str, ", ", &tmp);
3218                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3219                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3220                         usage_with_options_msg(usage_msg, options,
3221                                         "Unknown --sort key: `%s'", tok);
3222                 }
3223         }
3224
3225         free(str);
3226
3227         sort_dimension__add("pid", &sched->cmp_pid);
3228 }
3229
3230 static int __cmd_record(int argc, const char **argv)
3231 {
3232         unsigned int rec_argc, i, j;
3233         const char **rec_argv;
3234         const char * const record_args[] = {
3235                 "record",
3236                 "-a",
3237                 "-R",
3238                 "-m", "1024",
3239                 "-c", "1",
3240                 "-e", "sched:sched_switch",
3241                 "-e", "sched:sched_stat_wait",
3242                 "-e", "sched:sched_stat_sleep",
3243                 "-e", "sched:sched_stat_iowait",
3244                 "-e", "sched:sched_stat_runtime",
3245                 "-e", "sched:sched_process_fork",
3246                 "-e", "sched:sched_wakeup",
3247                 "-e", "sched:sched_wakeup_new",
3248                 "-e", "sched:sched_migrate_task",
3249         };
3250
3251         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3252         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3253
3254         if (rec_argv == NULL)
3255                 return -ENOMEM;
3256
3257         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3258                 rec_argv[i] = strdup(record_args[i]);
3259
3260         for (j = 1; j < (unsigned int)argc; j++, i++)
3261                 rec_argv[i] = argv[j];
3262
3263         BUG_ON(i != rec_argc);
3264
3265         return cmd_record(i, rec_argv, NULL);
3266 }
3267
3268 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
3269 {
3270         const char default_sort_order[] = "avg, max, switch, runtime";
3271         struct perf_sched sched = {
3272                 .tool = {
3273                         .sample          = perf_sched__process_tracepoint_sample,
3274                         .comm            = perf_event__process_comm,
3275                         .lost            = perf_event__process_lost,
3276                         .fork            = perf_sched__process_fork_event,
3277                         .ordered_events = true,
3278                 },
3279                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3280                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3281                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3282                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3283                 .sort_order           = default_sort_order,
3284                 .replay_repeat        = 10,
3285                 .profile_cpu          = -1,
3286                 .next_shortname1      = 'A',
3287                 .next_shortname2      = '0',
3288                 .skip_merge           = 0,
3289                 .show_callchain       = 1,
3290                 .max_stack            = 5,
3291         };
3292         const struct option sched_options[] = {
3293         OPT_STRING('i', "input", &input_name, "file",
3294                     "input file name"),
3295         OPT_INCR('v', "verbose", &verbose,
3296                     "be more verbose (show symbol address, etc)"),
3297         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3298                     "dump raw trace in ASCII"),
3299         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3300         OPT_END()
3301         };
3302         const struct option latency_options[] = {
3303         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3304                    "sort by key(s): runtime, switch, avg, max"),
3305         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3306                     "CPU to profile on"),
3307         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3308                     "latency stats per pid instead of per comm"),
3309         OPT_PARENT(sched_options)
3310         };
3311         const struct option replay_options[] = {
3312         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3313                      "repeat the workload replay N times (-1: infinite)"),
3314         OPT_PARENT(sched_options)
3315         };
3316         const struct option map_options[] = {
3317         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3318                     "map output in compact mode"),
3319         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3320                    "highlight given pids in map"),
3321         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3322                     "highlight given CPUs in map"),
3323         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3324                     "display given CPUs in map"),
3325         OPT_PARENT(sched_options)
3326         };
3327         const struct option timehist_options[] = {
3328         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3329                    "file", "vmlinux pathname"),
3330         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3331                    "file", "kallsyms pathname"),
3332         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3333                     "Display call chains if present (default on)"),
3334         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3335                    "Maximum number of functions to display backtrace."),
3336         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3337                     "Look for files with symbols relative to this directory"),
3338         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3339                     "Show only syscall summary with statistics"),
3340         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3341                     "Show all syscalls and summary with statistics"),
3342         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3343         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3344         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3345         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3346         OPT_STRING(0, "time", &sched.time_str, "str",
3347                    "Time span for analysis (start,stop)"),
3348         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3349         OPT_PARENT(sched_options)
3350         };
3351
3352         const char * const latency_usage[] = {
3353                 "perf sched latency [<options>]",
3354                 NULL
3355         };
3356         const char * const replay_usage[] = {
3357                 "perf sched replay [<options>]",
3358                 NULL
3359         };
3360         const char * const map_usage[] = {
3361                 "perf sched map [<options>]",
3362                 NULL
3363         };
3364         const char * const timehist_usage[] = {
3365                 "perf sched timehist [<options>]",
3366                 NULL
3367         };
3368         const char *const sched_subcommands[] = { "record", "latency", "map",
3369                                                   "replay", "script",
3370                                                   "timehist", NULL };
3371         const char *sched_usage[] = {
3372                 NULL,
3373                 NULL
3374         };
3375         struct trace_sched_handler lat_ops  = {
3376                 .wakeup_event       = latency_wakeup_event,
3377                 .switch_event       = latency_switch_event,
3378                 .runtime_event      = latency_runtime_event,
3379                 .migrate_task_event = latency_migrate_task_event,
3380         };
3381         struct trace_sched_handler map_ops  = {
3382                 .switch_event       = map_switch_event,
3383         };
3384         struct trace_sched_handler replay_ops  = {
3385                 .wakeup_event       = replay_wakeup_event,
3386                 .switch_event       = replay_switch_event,
3387                 .fork_event         = replay_fork_event,
3388         };
3389         unsigned int i;
3390
3391         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3392                 sched.curr_pid[i] = -1;
3393
3394         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3395                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3396         if (!argc)
3397                 usage_with_options(sched_usage, sched_options);
3398
3399         /*
3400          * Aliased to 'perf script' for now:
3401          */
3402         if (!strcmp(argv[0], "script"))
3403                 return cmd_script(argc, argv, prefix);
3404
3405         if (!strncmp(argv[0], "rec", 3)) {
3406                 return __cmd_record(argc, argv);
3407         } else if (!strncmp(argv[0], "lat", 3)) {
3408                 sched.tp_handler = &lat_ops;
3409                 if (argc > 1) {
3410                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3411                         if (argc)
3412                                 usage_with_options(latency_usage, latency_options);
3413                 }
3414                 setup_sorting(&sched, latency_options, latency_usage);
3415                 return perf_sched__lat(&sched);
3416         } else if (!strcmp(argv[0], "map")) {
3417                 if (argc) {
3418                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3419                         if (argc)
3420                                 usage_with_options(map_usage, map_options);
3421                 }
3422                 sched.tp_handler = &map_ops;
3423                 setup_sorting(&sched, latency_options, latency_usage);
3424                 return perf_sched__map(&sched);
3425         } else if (!strncmp(argv[0], "rep", 3)) {
3426                 sched.tp_handler = &replay_ops;
3427                 if (argc) {
3428                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3429                         if (argc)
3430                                 usage_with_options(replay_usage, replay_options);
3431                 }
3432                 return perf_sched__replay(&sched);
3433         } else if (!strcmp(argv[0], "timehist")) {
3434                 if (argc) {
3435                         argc = parse_options(argc, argv, timehist_options,
3436                                              timehist_usage, 0);
3437                         if (argc)
3438                                 usage_with_options(timehist_usage, timehist_options);
3439                 }
3440                 if (sched.show_wakeups && sched.summary_only) {
3441                         pr_err(" Error: -s and -w are mutually exclusive.\n");
3442                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3443                         parse_options_usage(NULL, timehist_options, "w", true);
3444                         return -EINVAL;
3445                 }
3446
3447                 return perf_sched__timehist(&sched);
3448         } else {
3449                 usage_with_options(sched_usage, sched_options);
3450         }
3451
3452         return 0;
3453 }