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/kernel.h>
  26 #include <linux/log2.h>
  27 #include <sys/prctl.h>
  28 #include <sys/resource.h>
  29 #include <inttypes.h>
  30
  31 #include <errno.h>
  32 #include <semaphore.h>
  33 #include <pthread.h>
  34 #include <math.h>
  35 #include <api/fs/fs.h>
  36 #include <linux/time64.h>
  37
  38 #include "sane_ctype.h"
  39
  40 #define PR_SET_NAME             15               /* Set process name */
  41 #define MAX_CPUS                4096
  42 #define COMM_LEN                20
  43 #define SYM_LEN                 129
  44 #define MAX_PID                 1024000
  45
  46 struct sched_atom;
  47
  48 struct task_desc {
  49         unsigned long           nr;
  50         unsigned long           pid;
  51         char                    comm[COMM_LEN];
  52
  53         unsigned long           nr_events;
  54         unsigned long           curr_event;
  55         struct sched_atom       **atoms;
  56
  57         pthread_t               thread;
  58         sem_t                   sleep_sem;
  59
  60         sem_t                   ready_for_work;
  61         sem_t                   work_done_sem;
  62
  63         u64                     cpu_usage;
  64 };
  65
  66 enum sched_event_type {
  67         SCHED_EVENT_RUN,
  68         SCHED_EVENT_SLEEP,
  69         SCHED_EVENT_WAKEUP,
  70         SCHED_EVENT_MIGRATION,
  71 };
  72
  73 struct sched_atom {
  74         enum sched_event_type   type;
  75         int                     specific_wait;
  76         u64                     timestamp;
  77         u64                     duration;
  78         unsigned long           nr;
  79         sem_t                   *wait_sem;
  80         struct task_desc        *wakee;
  81 };
  82
  83 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  84
  85 /* task state bitmask, copied from include/linux/sched.h */
  86 #define TASK_RUNNING            0
  87 #define TASK_INTERRUPTIBLE      1
  88 #define TASK_UNINTERRUPTIBLE    2
  89 #define __TASK_STOPPED          4
  90 #define __TASK_TRACED           8
  91 /* in tsk->exit_state */
  92 #define EXIT_DEAD               16
  93 #define EXIT_ZOMBIE             32
  94 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
  95 /* in tsk->state again */
  96 #define TASK_DEAD               64
  97 #define TASK_WAKEKILL           128
  98 #define TASK_WAKING             256
  99 #define TASK_PARKED             512
 100
 101 enum thread_state {
 102         THREAD_SLEEPING = 0,
 103         THREAD_WAIT_CPU,
 104         THREAD_SCHED_IN,
 105         THREAD_IGNORE
 106 };
 107
 108 struct work_atom {
 109         struct list_head        list;
 110         enum thread_state       state;
 111         u64                     sched_out_time;
 112         u64                     wake_up_time;
 113         u64                     sched_in_time;
 114         u64                     runtime;
 115 };
 116
 117 struct work_atoms {
 118         struct list_head        work_list;
 119         struct thread           *thread;
 120         struct rb_node          node;
 121         u64                     max_lat;
 122         u64                     max_lat_at;
 123         u64                     total_lat;
 124         u64                     nb_atoms;
 125         u64                     total_runtime;
 126         int                     num_merged;
 127 };
 128
 129 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 130
 131 struct perf_sched;
 132
 133 struct trace_sched_handler {
 134         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 135                             struct perf_sample *sample, struct machine *machine);
 136
 137         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 138                              struct perf_sample *sample, struct machine *machine);
 139
 140         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 141                             struct perf_sample *sample, struct machine *machine);
 142
 143         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 144         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 145                           struct machine *machine);
 146
 147         int (*migrate_task_event)(struct perf_sched *sched,
 148                                   struct perf_evsel *evsel,
 149                                   struct perf_sample *sample,
 150                                   struct machine *machine);
 151 };
 152
 153 #define COLOR_PIDS PERF_COLOR_BLUE
 154 #define COLOR_CPUS PERF_COLOR_BG_RED
 155
 156 struct perf_sched_map {
 157         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 158         int                     *comp_cpus;
 159         bool                     comp;
 160         struct thread_map       *color_pids;
 161         const char              *color_pids_str;
 162         struct cpu_map          *color_cpus;
 163         const char              *color_cpus_str;
 164         struct cpu_map          *cpus;
 165         const char              *cpus_str;
 166 };
 167
 168 struct perf_sched {
 169         struct perf_tool tool;
 170         const char       *sort_order;
 171         unsigned long    nr_tasks;
 172         struct task_desc **pid_to_task;
 173         struct task_desc **tasks;
 174         const struct trace_sched_handler *tp_handler;
 175         pthread_mutex_t  start_work_mutex;
 176         pthread_mutex_t  work_done_wait_mutex;
 177         int              profile_cpu;
 178 /*
 179  * Track the current task - that way we can know whether there's any
 180  * weird events, such as a task being switched away that is not current.
 181  */
 182         int              max_cpu;
 183         u32              curr_pid[MAX_CPUS];
 184         struct thread    *curr_thread[MAX_CPUS];
 185         char             next_shortname1;
 186         char             next_shortname2;
 187         unsigned int     replay_repeat;
 188         unsigned long    nr_run_events;
 189         unsigned long    nr_sleep_events;
 190         unsigned long    nr_wakeup_events;
 191         unsigned long    nr_sleep_corrections;
 192         unsigned long    nr_run_events_optimized;
 193         unsigned long    targetless_wakeups;
 194         unsigned long    multitarget_wakeups;
 195         unsigned long    nr_runs;
 196         unsigned long    nr_timestamps;
 197         unsigned long    nr_unordered_timestamps;
 198         unsigned long    nr_context_switch_bugs;
 199         unsigned long    nr_events;
 200         unsigned long    nr_lost_chunks;
 201         unsigned long    nr_lost_events;
 202         u64              run_measurement_overhead;
 203         u64              sleep_measurement_overhead;
 204         u64              start_time;
 205         u64              cpu_usage;
 206         u64              runavg_cpu_usage;
 207         u64              parent_cpu_usage;
 208         u64              runavg_parent_cpu_usage;
 209         u64              sum_runtime;
 210         u64              sum_fluct;
 211         u64              run_avg;
 212         u64              all_runtime;
 213         u64              all_count;
 214         u64              cpu_last_switched[MAX_CPUS];
 215         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 216         struct list_head sort_list, cmp_pid;
 217         bool force;
 218         bool skip_merge;
 219         struct perf_sched_map map;
 220
 221         /* options for timehist command */
 222         bool            summary;
 223         bool            summary_only;
 224         bool            idle_hist;
 225         bool            show_callchain;
 226         unsigned int    max_stack;
 227         bool            show_cpu_visual;
 228         bool            show_wakeups;
 229         bool            show_next;
 230         bool            show_migrations;
 231         bool            show_state;
 232         u64             skipped_samples;
 233         const char      *time_str;
 234         struct perf_time_interval ptime;
 235         struct perf_time_interval hist_time;
 236 };
 237
 238 /* per thread run time data */
 239 struct thread_runtime {
 240         u64 last_time;      /* time of previous sched in/out event */
 241         u64 dt_run;         /* run time */
 242         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 243         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 244         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 245         u64 dt_delay;       /* time between wakeup and sched-in */
 246         u64 ready_to_run;   /* time of wakeup */
 247
 248         struct stats run_stats;
 249         u64 total_run_time;
 250         u64 total_sleep_time;
 251         u64 total_iowait_time;
 252         u64 total_preempt_time;
 253         u64 total_delay_time;
 254
 255         int last_state;
 256         u64 migrations;
 257 };
 258
 259 /* per event run time data */
 260 struct evsel_runtime {
 261         u64 *last_time; /* time this event was last seen per cpu */
 262         u32 ncpu;       /* highest cpu slot allocated */
 263 };
 264
 265 /* per cpu idle time data */
 266 struct idle_thread_runtime {
 267         struct thread_runtime   tr;
 268         struct thread           *last_thread;
 269         struct rb_root          sorted_root;
 270         struct callchain_root   callchain;
 271         struct callchain_cursor cursor;
 272 };
 273
 274 /* track idle times per cpu */
 275 static struct thread **idle_threads;
 276 static int idle_max_cpu;
 277 static char idle_comm[] = "<idle>";
 278
 279 static u64 get_nsecs(void)
 280 {
 281         struct timespec ts;
 282
 283         clock_gettime(CLOCK_MONOTONIC, &ts);
 284
 285         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 286 }
 287
 288 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 289 {
 290         u64 T0 = get_nsecs(), T1;
 291
 292         do {
 293                 T1 = get_nsecs();
 294         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 295 }
 296
 297 static void sleep_nsecs(u64 nsecs)
 298 {
 299         struct timespec ts;
 300
 301         ts.tv_nsec = nsecs % 999999999;
 302         ts.tv_sec = nsecs / 999999999;
 303
 304         nanosleep(&ts, NULL);
 305 }
 306
 307 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 308 {
 309         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 310         int i;
 311
 312         for (i = 0; i < 10; i++) {
 313                 T0 = get_nsecs();
 314                 burn_nsecs(sched, 0);
 315                 T1 = get_nsecs();
 316                 delta = T1-T0;
 317                 min_delta = min(min_delta, delta);
 318         }
 319         sched->run_measurement_overhead = min_delta;
 320
 321         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 322 }
 323
 324 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 325 {
 326         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 327         int i;
 328
 329         for (i = 0; i < 10; i++) {
 330                 T0 = get_nsecs();
 331                 sleep_nsecs(10000);
 332                 T1 = get_nsecs();
 333                 delta = T1-T0;
 334                 min_delta = min(min_delta, delta);
 335         }
 336         min_delta -= 10000;
 337         sched->sleep_measurement_overhead = min_delta;
 338
 339         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 340 }
 341
 342 static struct sched_atom *
 343 get_new_event(struct task_desc *task, u64 timestamp)
 344 {
 345         struct sched_atom *event = zalloc(sizeof(*event));
 346         unsigned long idx = task->nr_events;
 347         size_t size;
 348
 349         event->timestamp = timestamp;
 350         event->nr = idx;
 351
 352         task->nr_events++;
 353         size = sizeof(struct sched_atom *) * task->nr_events;
 354         task->atoms = realloc(task->atoms, size);
 355         BUG_ON(!task->atoms);
 356
 357         task->atoms[idx] = event;
 358
 359         return event;
 360 }
 361
 362 static struct sched_atom *last_event(struct task_desc *task)
 363 {
 364         if (!task->nr_events)
 365                 return NULL;
 366
 367         return task->atoms[task->nr_events - 1];
 368 }
 369
 370 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 371                                 u64 timestamp, u64 duration)
 372 {
 373         struct sched_atom *event, *curr_event = last_event(task);
 374
 375         /*
 376          * optimize an existing RUN event by merging this one
 377          * to it:
 378          */
 379         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 380                 sched->nr_run_events_optimized++;
 381                 curr_event->duration += duration;
 382                 return;
 383         }
 384
 385         event = get_new_event(task, timestamp);
 386
 387         event->type = SCHED_EVENT_RUN;
 388         event->duration = duration;
 389
 390         sched->nr_run_events++;
 391 }
 392
 393 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 394                                    u64 timestamp, struct task_desc *wakee)
 395 {
 396         struct sched_atom *event, *wakee_event;
 397
 398         event = get_new_event(task, timestamp);
 399         event->type = SCHED_EVENT_WAKEUP;
 400         event->wakee = wakee;
 401
 402         wakee_event = last_event(wakee);
 403         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 404                 sched->targetless_wakeups++;
 405                 return;
 406         }
 407         if (wakee_event->wait_sem) {
 408                 sched->multitarget_wakeups++;
 409                 return;
 410         }
 411
 412         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 413         sem_init(wakee_event->wait_sem, 0, 0);
 414         wakee_event->specific_wait = 1;
 415         event->wait_sem = wakee_event->wait_sem;
 416
 417         sched->nr_wakeup_events++;
 418 }
 419
 420 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 421                                   u64 timestamp, u64 task_state __maybe_unused)
 422 {
 423         struct sched_atom *event = get_new_event(task, timestamp);
 424
 425         event->type = SCHED_EVENT_SLEEP;
 426
 427         sched->nr_sleep_events++;
 428 }
 429
 430 static struct task_desc *register_pid(struct perf_sched *sched,
 431                                       unsigned long pid, const char *comm)
 432 {
 433         struct task_desc *task;
 434         static int pid_max;
 435
 436         if (sched->pid_to_task == NULL) {
 437                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 438                         pid_max = MAX_PID;
 439                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 440         }
 441         if (pid >= (unsigned long)pid_max) {
 442                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 443                         sizeof(struct task_desc *))) == NULL);
 444                 while (pid >= (unsigned long)pid_max)
 445                         sched->pid_to_task[pid_max++] = NULL;
 446         }
 447
 448         task = sched->pid_to_task[pid];
 449
 450         if (task)
 451                 return task;
 452
 453         task = zalloc(sizeof(*task));
 454         task->pid = pid;
 455         task->nr = sched->nr_tasks;
 456         strcpy(task->comm, comm);
 457         /*
 458          * every task starts in sleeping state - this gets ignored
 459          * if there's no wakeup pointing to this sleep state:
 460          */
 461         add_sched_event_sleep(sched, task, 0, 0);
 462
 463         sched->pid_to_task[pid] = task;
 464         sched->nr_tasks++;
 465         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 466         BUG_ON(!sched->tasks);
 467         sched->tasks[task->nr] = task;
 468
 469         if (verbose > 0)
 470                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 471
 472         return task;
 473 }
 474
 475
 476 static void print_task_traces(struct perf_sched *sched)
 477 {
 478         struct task_desc *task;
 479         unsigned long i;
 480
 481         for (i = 0; i < sched->nr_tasks; i++) {
 482                 task = sched->tasks[i];
 483                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 484                         task->nr, task->comm, task->pid, task->nr_events);
 485         }
 486 }
 487
 488 static void add_cross_task_wakeups(struct perf_sched *sched)
 489 {
 490         struct task_desc *task1, *task2;
 491         unsigned long i, j;
 492
 493         for (i = 0; i < sched->nr_tasks; i++) {
 494                 task1 = sched->tasks[i];
 495                 j = i + 1;
 496                 if (j == sched->nr_tasks)
 497                         j = 0;
 498                 task2 = sched->tasks[j];
 499                 add_sched_event_wakeup(sched, task1, 0, task2);
 500         }
 501 }
 502
 503 static void perf_sched__process_event(struct perf_sched *sched,
 504                                       struct sched_atom *atom)
 505 {
 506         int ret = 0;
 507
 508         switch (atom->type) {
 509                 case SCHED_EVENT_RUN:
 510                         burn_nsecs(sched, atom->duration);
 511                         break;
 512                 case SCHED_EVENT_SLEEP:
 513                         if (atom->wait_sem)
 514                                 ret = sem_wait(atom->wait_sem);
 515                         BUG_ON(ret);
 516                         break;
 517                 case SCHED_EVENT_WAKEUP:
 518                         if (atom->wait_sem)
 519                                 ret = sem_post(atom->wait_sem);
 520                         BUG_ON(ret);
 521                         break;
 522                 case SCHED_EVENT_MIGRATION:
 523                         break;
 524                 default:
 525                         BUG_ON(1);
 526         }
 527 }
 528
 529 static u64 get_cpu_usage_nsec_parent(void)
 530 {
 531         struct rusage ru;
 532         u64 sum;
 533         int err;
 534
 535         err = getrusage(RUSAGE_SELF, &ru);
 536         BUG_ON(err);
 537
 538         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 539         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 540
 541         return sum;
 542 }
 543
 544 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 545 {
 546         struct perf_event_attr attr;
 547         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 548         int fd;
 549         struct rlimit limit;
 550         bool need_privilege = false;
 551
 552         memset(&attr, 0, sizeof(attr));
 553
 554         attr.type = PERF_TYPE_SOFTWARE;
 555         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 556
 557 force_again:
 558         fd = sys_perf_event_open(&attr, 0, -1, -1,
 559                                  perf_event_open_cloexec_flag());
 560
 561         if (fd < 0) {
 562                 if (errno == EMFILE) {
 563                         if (sched->force) {
 564                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 565                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 566                                 if (limit.rlim_cur > limit.rlim_max) {
 567                                         limit.rlim_max = limit.rlim_cur;
 568                                         need_privilege = true;
 569                                 }
 570                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 571                                         if (need_privilege && errno == EPERM)
 572                                                 strcpy(info, "Need privilege\n");
 573                                 } else
 574                                         goto force_again;
 575                         } else
 576                                 strcpy(info, "Have a try with -f option\n");
 577                 }
 578                 pr_err("Error: sys_perf_event_open() syscall returned "
 579                        "with %d (%s)\n%s", fd,
 580                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 581                 exit(EXIT_FAILURE);
 582         }
 583         return fd;
 584 }
 585
 586 static u64 get_cpu_usage_nsec_self(int fd)
 587 {
 588         u64 runtime;
 589         int ret;
 590
 591         ret = read(fd, &runtime, sizeof(runtime));
 592         BUG_ON(ret != sizeof(runtime));
 593
 594         return runtime;
 595 }
 596
 597 struct sched_thread_parms {
 598         struct task_desc  *task;
 599         struct perf_sched *sched;
 600         int fd;
 601 };
 602
 603 static void *thread_func(void *ctx)
 604 {
 605         struct sched_thread_parms *parms = ctx;
 606         struct task_desc *this_task = parms->task;
 607         struct perf_sched *sched = parms->sched;
 608         u64 cpu_usage_0, cpu_usage_1;
 609         unsigned long i, ret;
 610         char comm2[22];
 611         int fd = parms->fd;
 612
 613         zfree(&parms);
 614
 615         sprintf(comm2, ":%s", this_task->comm);
 616         prctl(PR_SET_NAME, comm2);
 617         if (fd < 0)
 618                 return NULL;
 619 again:
 620         ret = sem_post(&this_task->ready_for_work);
 621         BUG_ON(ret);
 622         ret = pthread_mutex_lock(&sched->start_work_mutex);
 623         BUG_ON(ret);
 624         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 625         BUG_ON(ret);
 626
 627         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 628
 629         for (i = 0; i < this_task->nr_events; i++) {
 630                 this_task->curr_event = i;
 631                 perf_sched__process_event(sched, this_task->atoms[i]);
 632         }
 633
 634         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 635         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 636         ret = sem_post(&this_task->work_done_sem);
 637         BUG_ON(ret);
 638
 639         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 640         BUG_ON(ret);
 641         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 642         BUG_ON(ret);
 643
 644         goto again;
 645 }
 646
 647 static void create_tasks(struct perf_sched *sched)
 648 {
 649         struct task_desc *task;
 650         pthread_attr_t attr;
 651         unsigned long i;
 652         int err;
 653
 654         err = pthread_attr_init(&attr);
 655         BUG_ON(err);
 656         err = pthread_attr_setstacksize(&attr,
 657                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 658         BUG_ON(err);
 659         err = pthread_mutex_lock(&sched->start_work_mutex);
 660         BUG_ON(err);
 661         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 662         BUG_ON(err);
 663         for (i = 0; i < sched->nr_tasks; i++) {
 664                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 665                 BUG_ON(parms == NULL);
 666                 parms->task = task = sched->tasks[i];
 667                 parms->sched = sched;
 668                 parms->fd = self_open_counters(sched, i);
 669                 sem_init(&task->sleep_sem, 0, 0);
 670                 sem_init(&task->ready_for_work, 0, 0);
 671                 sem_init(&task->work_done_sem, 0, 0);
 672                 task->curr_event = 0;
 673                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 674                 BUG_ON(err);
 675         }
 676 }
 677
 678 static void wait_for_tasks(struct perf_sched *sched)
 679 {
 680         u64 cpu_usage_0, cpu_usage_1;
 681         struct task_desc *task;
 682         unsigned long i, ret;
 683
 684         sched->start_time = get_nsecs();
 685         sched->cpu_usage = 0;
 686         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 687
 688         for (i = 0; i < sched->nr_tasks; i++) {
 689                 task = sched->tasks[i];
 690                 ret = sem_wait(&task->ready_for_work);
 691                 BUG_ON(ret);
 692                 sem_init(&task->ready_for_work, 0, 0);
 693         }
 694         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 695         BUG_ON(ret);
 696
 697         cpu_usage_0 = get_cpu_usage_nsec_parent();
 698
 699         pthread_mutex_unlock(&sched->start_work_mutex);
 700
 701         for (i = 0; i < sched->nr_tasks; i++) {
 702                 task = sched->tasks[i];
 703                 ret = sem_wait(&task->work_done_sem);
 704                 BUG_ON(ret);
 705                 sem_init(&task->work_done_sem, 0, 0);
 706                 sched->cpu_usage += task->cpu_usage;
 707                 task->cpu_usage = 0;
 708         }
 709
 710         cpu_usage_1 = get_cpu_usage_nsec_parent();
 711         if (!sched->runavg_cpu_usage)
 712                 sched->runavg_cpu_usage = sched->cpu_usage;
 713         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 714
 715         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 716         if (!sched->runavg_parent_cpu_usage)
 717                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 718         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 719                                          sched->parent_cpu_usage)/sched->replay_repeat;
 720
 721         ret = pthread_mutex_lock(&sched->start_work_mutex);
 722         BUG_ON(ret);
 723
 724         for (i = 0; i < sched->nr_tasks; i++) {
 725                 task = sched->tasks[i];
 726                 sem_init(&task->sleep_sem, 0, 0);
 727                 task->curr_event = 0;
 728         }
 729 }
 730
 731 static void run_one_test(struct perf_sched *sched)
 732 {
 733         u64 T0, T1, delta, avg_delta, fluct;
 734
 735         T0 = get_nsecs();
 736         wait_for_tasks(sched);
 737         T1 = get_nsecs();
 738
 739         delta = T1 - T0;
 740         sched->sum_runtime += delta;
 741         sched->nr_runs++;
 742
 743         avg_delta = sched->sum_runtime / sched->nr_runs;
 744         if (delta < avg_delta)
 745                 fluct = avg_delta - delta;
 746         else
 747                 fluct = delta - avg_delta;
 748         sched->sum_fluct += fluct;
 749         if (!sched->run_avg)
 750                 sched->run_avg = delta;
 751         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 752
 753         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 754
 755         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 756
 757         printf("cpu: %0.2f / %0.2f",
 758                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 759
 760 #if 0
 761         /*
 762          * rusage statistics done by the parent, these are less
 763          * accurate than the sched->sum_exec_runtime based statistics:
 764          */
 765         printf(" [%0.2f / %0.2f]",
 766                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 767                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 768 #endif
 769
 770         printf("\n");
 771
 772         if (sched->nr_sleep_corrections)
 773                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 774         sched->nr_sleep_corrections = 0;
 775 }
 776
 777 static void test_calibrations(struct perf_sched *sched)
 778 {
 779         u64 T0, T1;
 780
 781         T0 = get_nsecs();
 782         burn_nsecs(sched, NSEC_PER_MSEC);
 783         T1 = get_nsecs();
 784
 785         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 786
 787         T0 = get_nsecs();
 788         sleep_nsecs(NSEC_PER_MSEC);
 789         T1 = get_nsecs();
 790
 791         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 792 }
 793
 794 static int
 795 replay_wakeup_event(struct perf_sched *sched,
 796                     struct perf_evsel *evsel, struct perf_sample *sample,
 797                     struct machine *machine __maybe_unused)
 798 {
 799         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 800         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 801         struct task_desc *waker, *wakee;
 802
 803         if (verbose > 0) {
 804                 printf("sched_wakeup event %p\n", evsel);
 805
 806                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 807         }
 808
 809         waker = register_pid(sched, sample->tid, "<unknown>");
 810         wakee = register_pid(sched, pid, comm);
 811
 812         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 813         return 0;
 814 }
 815
 816 static int replay_switch_event(struct perf_sched *sched,
 817                                struct perf_evsel *evsel,
 818                                struct perf_sample *sample,
 819                                struct machine *machine __maybe_unused)
 820 {
 821         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 822                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 823         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 824                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 825         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 826         struct task_desc *prev, __maybe_unused *next;
 827         u64 timestamp0, timestamp = sample->time;
 828         int cpu = sample->cpu;
 829         s64 delta;
 830
 831         if (verbose > 0)
 832                 printf("sched_switch event %p\n", evsel);
 833
 834         if (cpu >= MAX_CPUS || cpu < 0)
 835                 return 0;
 836
 837         timestamp0 = sched->cpu_last_switched[cpu];
 838         if (timestamp0)
 839                 delta = timestamp - timestamp0;
 840         else
 841                 delta = 0;
 842
 843         if (delta < 0) {
 844                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 845                 return -1;
 846         }
 847
 848         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 849                  prev_comm, prev_pid, next_comm, next_pid, delta);
 850
 851         prev = register_pid(sched, prev_pid, prev_comm);
 852         next = register_pid(sched, next_pid, next_comm);
 853
 854         sched->cpu_last_switched[cpu] = timestamp;
 855
 856         add_sched_event_run(sched, prev, timestamp, delta);
 857         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 858
 859         return 0;
 860 }
 861
 862 static int replay_fork_event(struct perf_sched *sched,
 863                              union perf_event *event,
 864                              struct machine *machine)
 865 {
 866         struct thread *child, *parent;
 867
 868         child = machine__findnew_thread(machine, event->fork.pid,
 869                                         event->fork.tid);
 870         parent = machine__findnew_thread(machine, event->fork.ppid,
 871                                          event->fork.ptid);
 872
 873         if (child == NULL || parent == NULL) {
 874                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 875                                  child, parent);
 876                 goto out_put;
 877         }
 878
 879         if (verbose > 0) {
 880                 printf("fork event\n");
 881                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 882                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 883         }
 884
 885         register_pid(sched, parent->tid, thread__comm_str(parent));
 886         register_pid(sched, child->tid, thread__comm_str(child));
 887 out_put:
 888         thread__put(child);
 889         thread__put(parent);
 890         return 0;
 891 }
 892
 893 struct sort_dimension {
 894         const char              *name;
 895         sort_fn_t               cmp;
 896         struct list_head        list;
 897 };
 898
 899 static int
 900 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 901 {
 902         struct sort_dimension *sort;
 903         int ret = 0;
 904
 905         BUG_ON(list_empty(list));
 906
 907         list_for_each_entry(sort, list, list) {
 908                 ret = sort->cmp(l, r);
 909                 if (ret)
 910                         return ret;
 911         }
 912
 913         return ret;
 914 }
 915
 916 static struct work_atoms *
 917 thread_atoms_search(struct rb_root *root, struct thread *thread,
 918                          struct list_head *sort_list)
 919 {
 920         struct rb_node *node = root->rb_node;
 921         struct work_atoms key = { .thread = thread };
 922
 923         while (node) {
 924                 struct work_atoms *atoms;
 925                 int cmp;
 926
 927                 atoms = container_of(node, struct work_atoms, node);
 928
 929                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 930                 if (cmp > 0)
 931                         node = node->rb_left;
 932                 else if (cmp < 0)
 933                         node = node->rb_right;
 934                 else {
 935                         BUG_ON(thread != atoms->thread);
 936                         return atoms;
 937                 }
 938         }
 939         return NULL;
 940 }
 941
 942 static void
 943 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 944                          struct list_head *sort_list)
 945 {
 946         struct rb_node **new = &(root->rb_node), *parent = NULL;
 947
 948         while (*new) {
 949                 struct work_atoms *this;
 950                 int cmp;
 951
 952                 this = container_of(*new, struct work_atoms, node);
 953                 parent = *new;
 954
 955                 cmp = thread_lat_cmp(sort_list, data, this);
 956
 957                 if (cmp > 0)
 958                         new = &((*new)->rb_left);
 959                 else
 960                         new = &((*new)->rb_right);
 961         }
 962
 963         rb_link_node(&data->node, parent, new);
 964         rb_insert_color(&data->node, root);
 965 }
 966
 967 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 968 {
 969         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 970         if (!atoms) {
 971                 pr_err("No memory at %s\n", __func__);
 972                 return -1;
 973         }
 974
 975         atoms->thread = thread__get(thread);
 976         INIT_LIST_HEAD(&atoms->work_list);
 977         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 978         return 0;
 979 }
 980
 981 static char sched_out_state(u64 prev_state)
 982 {
 983         const char *str = TASK_STATE_TO_CHAR_STR;
 984
 985         return str[prev_state];
 986 }
 987
 988 static int
 989 add_sched_out_event(struct work_atoms *atoms,
 990                     char run_state,
 991                     u64 timestamp)
 992 {
 993         struct work_atom *atom = zalloc(sizeof(*atom));
 994         if (!atom) {
 995                 pr_err("Non memory at %s", __func__);
 996                 return -1;
 997         }
 998
 999         atom->sched_out_time = timestamp;
1000
1001         if (run_state == 'R') {
1002                 atom->state = THREAD_WAIT_CPU;
1003                 atom->wake_up_time = atom->sched_out_time;
1004         }
1005
1006         list_add_tail(&atom->list, &atoms->work_list);
1007         return 0;
1008 }
1009
1010 static void
1011 add_runtime_event(struct work_atoms *atoms, u64 delta,
1012                   u64 timestamp __maybe_unused)
1013 {
1014         struct work_atom *atom;
1015
1016         BUG_ON(list_empty(&atoms->work_list));
1017
1018         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1019
1020         atom->runtime += delta;
1021         atoms->total_runtime += delta;
1022 }
1023
1024 static void
1025 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1026 {
1027         struct work_atom *atom;
1028         u64 delta;
1029
1030         if (list_empty(&atoms->work_list))
1031                 return;
1032
1033         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1034
1035         if (atom->state != THREAD_WAIT_CPU)
1036                 return;
1037
1038         if (timestamp < atom->wake_up_time) {
1039                 atom->state = THREAD_IGNORE;
1040                 return;
1041         }
1042
1043         atom->state = THREAD_SCHED_IN;
1044         atom->sched_in_time = timestamp;
1045
1046         delta = atom->sched_in_time - atom->wake_up_time;
1047         atoms->total_lat += delta;
1048         if (delta > atoms->max_lat) {
1049                 atoms->max_lat = delta;
1050                 atoms->max_lat_at = timestamp;
1051         }
1052         atoms->nb_atoms++;
1053 }
1054
1055 static int latency_switch_event(struct perf_sched *sched,
1056                                 struct perf_evsel *evsel,
1057                                 struct perf_sample *sample,
1058                                 struct machine *machine)
1059 {
1060         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1061                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1062         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1063         struct work_atoms *out_events, *in_events;
1064         struct thread *sched_out, *sched_in;
1065         u64 timestamp0, timestamp = sample->time;
1066         int cpu = sample->cpu, err = -1;
1067         s64 delta;
1068
1069         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1070
1071         timestamp0 = sched->cpu_last_switched[cpu];
1072         sched->cpu_last_switched[cpu] = timestamp;
1073         if (timestamp0)
1074                 delta = timestamp - timestamp0;
1075         else
1076                 delta = 0;
1077
1078         if (delta < 0) {
1079                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1080                 return -1;
1081         }
1082
1083         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1084         sched_in = machine__findnew_thread(machine, -1, next_pid);
1085         if (sched_out == NULL || sched_in == NULL)
1086                 goto out_put;
1087
1088         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1089         if (!out_events) {
1090                 if (thread_atoms_insert(sched, sched_out))
1091                         goto out_put;
1092                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1093                 if (!out_events) {
1094                         pr_err("out-event: Internal tree error");
1095                         goto out_put;
1096                 }
1097         }
1098         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1099                 return -1;
1100
1101         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1102         if (!in_events) {
1103                 if (thread_atoms_insert(sched, sched_in))
1104                         goto out_put;
1105                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1106                 if (!in_events) {
1107                         pr_err("in-event: Internal tree error");
1108                         goto out_put;
1109                 }
1110                 /*
1111                  * Take came in we have not heard about yet,
1112                  * add in an initial atom in runnable state:
1113                  */
1114                 if (add_sched_out_event(in_events, 'R', timestamp))
1115                         goto out_put;
1116         }
1117         add_sched_in_event(in_events, timestamp);
1118         err = 0;
1119 out_put:
1120         thread__put(sched_out);
1121         thread__put(sched_in);
1122         return err;
1123 }
1124
1125 static int latency_runtime_event(struct perf_sched *sched,
1126                                  struct perf_evsel *evsel,
1127                                  struct perf_sample *sample,
1128                                  struct machine *machine)
1129 {
1130         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1131         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1132         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1133         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1134         u64 timestamp = sample->time;
1135         int cpu = sample->cpu, err = -1;
1136
1137         if (thread == NULL)
1138                 return -1;
1139
1140         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1141         if (!atoms) {
1142                 if (thread_atoms_insert(sched, thread))
1143                         goto out_put;
1144                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1145                 if (!atoms) {
1146                         pr_err("in-event: Internal tree error");
1147                         goto out_put;
1148                 }
1149                 if (add_sched_out_event(atoms, 'R', timestamp))
1150                         goto out_put;
1151         }
1152
1153         add_runtime_event(atoms, runtime, timestamp);
1154         err = 0;
1155 out_put:
1156         thread__put(thread);
1157         return err;
1158 }
1159
1160 static int latency_wakeup_event(struct perf_sched *sched,
1161                                 struct perf_evsel *evsel,
1162                                 struct perf_sample *sample,
1163                                 struct machine *machine)
1164 {
1165         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1166         struct work_atoms *atoms;
1167         struct work_atom *atom;
1168         struct thread *wakee;
1169         u64 timestamp = sample->time;
1170         int err = -1;
1171
1172         wakee = machine__findnew_thread(machine, -1, pid);
1173         if (wakee == NULL)
1174                 return -1;
1175         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1176         if (!atoms) {
1177                 if (thread_atoms_insert(sched, wakee))
1178                         goto out_put;
1179                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1180                 if (!atoms) {
1181                         pr_err("wakeup-event: Internal tree error");
1182                         goto out_put;
1183                 }
1184                 if (add_sched_out_event(atoms, 'S', timestamp))
1185                         goto out_put;
1186         }
1187
1188         BUG_ON(list_empty(&atoms->work_list));
1189
1190         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1191
1192         /*
1193          * As we do not guarantee the wakeup event happens when
1194          * task is out of run queue, also may happen when task is
1195          * on run queue and wakeup only change ->state to TASK_RUNNING,
1196          * then we should not set the ->wake_up_time when wake up a
1197          * task which is on run queue.
1198          *
1199          * You WILL be missing events if you've recorded only
1200          * one CPU, or are only looking at only one, so don't
1201          * skip in this case.
1202          */
1203         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1204                 goto out_ok;
1205
1206         sched->nr_timestamps++;
1207         if (atom->sched_out_time > timestamp) {
1208                 sched->nr_unordered_timestamps++;
1209                 goto out_ok;
1210         }
1211
1212         atom->state = THREAD_WAIT_CPU;
1213         atom->wake_up_time = timestamp;
1214 out_ok:
1215         err = 0;
1216 out_put:
1217         thread__put(wakee);
1218         return err;
1219 }
1220
1221 static int latency_migrate_task_event(struct perf_sched *sched,
1222                                       struct perf_evsel *evsel,
1223                                       struct perf_sample *sample,
1224                                       struct machine *machine)
1225 {
1226         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1227         u64 timestamp = sample->time;
1228         struct work_atoms *atoms;
1229         struct work_atom *atom;
1230         struct thread *migrant;
1231         int err = -1;
1232
1233         /*
1234          * Only need to worry about migration when profiling one CPU.
1235          */
1236         if (sched->profile_cpu == -1)
1237                 return 0;
1238
1239         migrant = machine__findnew_thread(machine, -1, pid);
1240         if (migrant == NULL)
1241                 return -1;
1242         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1243         if (!atoms) {
1244                 if (thread_atoms_insert(sched, migrant))
1245                         goto out_put;
1246                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1247                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1248                 if (!atoms) {
1249                         pr_err("migration-event: Internal tree error");
1250                         goto out_put;
1251                 }
1252                 if (add_sched_out_event(atoms, 'R', timestamp))
1253                         goto out_put;
1254         }
1255
1256         BUG_ON(list_empty(&atoms->work_list));
1257
1258         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1259         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1260
1261         sched->nr_timestamps++;
1262
1263         if (atom->sched_out_time > timestamp)
1264                 sched->nr_unordered_timestamps++;
1265         err = 0;
1266 out_put:
1267         thread__put(migrant);
1268         return err;
1269 }
1270
1271 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1272 {
1273         int i;
1274         int ret;
1275         u64 avg;
1276         char max_lat_at[32];
1277
1278         if (!work_list->nb_atoms)
1279                 return;
1280         /*
1281          * Ignore idle threads:
1282          */
1283         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1284                 return;
1285
1286         sched->all_runtime += work_list->total_runtime;
1287         sched->all_count   += work_list->nb_atoms;
1288
1289         if (work_list->num_merged > 1)
1290                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1291         else
1292                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1293
1294         for (i = 0; i < 24 - ret; i++)
1295                 printf(" ");
1296
1297         avg = work_list->total_lat / work_list->nb_atoms;
1298         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1299
1300         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1301               (double)work_list->total_runtime / NSEC_PER_MSEC,
1302                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1303                  (double)work_list->max_lat / NSEC_PER_MSEC,
1304                  max_lat_at);
1305 }
1306
1307 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1308 {
1309         if (l->thread == r->thread)
1310                 return 0;
1311         if (l->thread->tid < r->thread->tid)
1312                 return -1;
1313         if (l->thread->tid > r->thread->tid)
1314                 return 1;
1315         return (int)(l->thread - r->thread);
1316 }
1317
1318 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1319 {
1320         u64 avgl, avgr;
1321
1322         if (!l->nb_atoms)
1323                 return -1;
1324
1325         if (!r->nb_atoms)
1326                 return 1;
1327
1328         avgl = l->total_lat / l->nb_atoms;
1329         avgr = r->total_lat / r->nb_atoms;
1330
1331         if (avgl < avgr)
1332                 return -1;
1333         if (avgl > avgr)
1334                 return 1;
1335
1336         return 0;
1337 }
1338
1339 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1340 {
1341         if (l->max_lat < r->max_lat)
1342                 return -1;
1343         if (l->max_lat > r->max_lat)
1344                 return 1;
1345
1346         return 0;
1347 }
1348
1349 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1350 {
1351         if (l->nb_atoms < r->nb_atoms)
1352                 return -1;
1353         if (l->nb_atoms > r->nb_atoms)
1354                 return 1;
1355
1356         return 0;
1357 }
1358
1359 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1360 {
1361         if (l->total_runtime < r->total_runtime)
1362                 return -1;
1363         if (l->total_runtime > r->total_runtime)
1364                 return 1;
1365
1366         return 0;
1367 }
1368
1369 static int sort_dimension__add(const char *tok, struct list_head *list)
1370 {
1371         size_t i;
1372         static struct sort_dimension avg_sort_dimension = {
1373                 .name = "avg",
1374                 .cmp  = avg_cmp,
1375         };
1376         static struct sort_dimension max_sort_dimension = {
1377                 .name = "max",
1378                 .cmp  = max_cmp,
1379         };
1380         static struct sort_dimension pid_sort_dimension = {
1381                 .name = "pid",
1382                 .cmp  = pid_cmp,
1383         };
1384         static struct sort_dimension runtime_sort_dimension = {
1385                 .name = "runtime",
1386                 .cmp  = runtime_cmp,
1387         };
1388         static struct sort_dimension switch_sort_dimension = {
1389                 .name = "switch",
1390                 .cmp  = switch_cmp,
1391         };
1392         struct sort_dimension *available_sorts[] = {
1393                 &pid_sort_dimension,
1394                 &avg_sort_dimension,
1395                 &max_sort_dimension,
1396                 &switch_sort_dimension,
1397                 &runtime_sort_dimension,
1398         };
1399
1400         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1401                 if (!strcmp(available_sorts[i]->name, tok)) {
1402                         list_add_tail(&available_sorts[i]->list, list);
1403
1404                         return 0;
1405                 }
1406         }
1407
1408         return -1;
1409 }
1410
1411 static void perf_sched__sort_lat(struct perf_sched *sched)
1412 {
1413         struct rb_node *node;
1414         struct rb_root *root = &sched->atom_root;
1415 again:
1416         for (;;) {
1417                 struct work_atoms *data;
1418                 node = rb_first(root);
1419                 if (!node)
1420                         break;
1421
1422                 rb_erase(node, root);
1423                 data = rb_entry(node, struct work_atoms, node);
1424                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1425         }
1426         if (root == &sched->atom_root) {
1427                 root = &sched->merged_atom_root;
1428                 goto again;
1429         }
1430 }
1431
1432 static int process_sched_wakeup_event(struct perf_tool *tool,
1433                                       struct perf_evsel *evsel,
1434                                       struct perf_sample *sample,
1435                                       struct machine *machine)
1436 {
1437         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1438
1439         if (sched->tp_handler->wakeup_event)
1440                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1441
1442         return 0;
1443 }
1444
1445 union map_priv {
1446         void    *ptr;
1447         bool     color;
1448 };
1449
1450 static bool thread__has_color(struct thread *thread)
1451 {
1452         union map_priv priv = {
1453                 .ptr = thread__priv(thread),
1454         };
1455
1456         return priv.color;
1457 }
1458
1459 static struct thread*
1460 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1461 {
1462         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1463         union map_priv priv = {
1464                 .color = false,
1465         };
1466
1467         if (!sched->map.color_pids || !thread || thread__priv(thread))
1468                 return thread;
1469
1470         if (thread_map__has(sched->map.color_pids, tid))
1471                 priv.color = true;
1472
1473         thread__set_priv(thread, priv.ptr);
1474         return thread;
1475 }
1476
1477 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1478                             struct perf_sample *sample, struct machine *machine)
1479 {
1480         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1481         struct thread *sched_in;
1482         int new_shortname;
1483         u64 timestamp0, timestamp = sample->time;
1484         s64 delta;
1485         int i, this_cpu = sample->cpu;
1486         int cpus_nr;
1487         bool new_cpu = false;
1488         const char *color = PERF_COLOR_NORMAL;
1489         char stimestamp[32];
1490
1491         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1492
1493         if (this_cpu > sched->max_cpu)
1494                 sched->max_cpu = this_cpu;
1495
1496         if (sched->map.comp) {
1497                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1498                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1499                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1500                         new_cpu = true;
1501                 }
1502         } else
1503                 cpus_nr = sched->max_cpu;
1504
1505         timestamp0 = sched->cpu_last_switched[this_cpu];
1506         sched->cpu_last_switched[this_cpu] = timestamp;
1507         if (timestamp0)
1508                 delta = timestamp - timestamp0;
1509         else
1510                 delta = 0;
1511
1512         if (delta < 0) {
1513                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1514                 return -1;
1515         }
1516
1517         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1518         if (sched_in == NULL)
1519                 return -1;
1520
1521         sched->curr_thread[this_cpu] = thread__get(sched_in);
1522
1523         printf("  ");
1524
1525         new_shortname = 0;
1526         if (!sched_in->shortname[0]) {
1527                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1528                         /*
1529                          * Don't allocate a letter-number for swapper:0
1530                          * as a shortname. Instead, we use '.' for it.
1531                          */
1532                         sched_in->shortname[0] = '.';
1533                         sched_in->shortname[1] = ' ';
1534                 } else {
1535                         sched_in->shortname[0] = sched->next_shortname1;
1536                         sched_in->shortname[1] = sched->next_shortname2;
1537
1538                         if (sched->next_shortname1 < 'Z') {
1539                                 sched->next_shortname1++;
1540                         } else {
1541                                 sched->next_shortname1 = 'A';
1542                                 if (sched->next_shortname2 < '9')
1543                                         sched->next_shortname2++;
1544                                 else
1545                                         sched->next_shortname2 = '0';
1546                         }
1547                 }
1548                 new_shortname = 1;
1549         }
1550
1551         for (i = 0; i < cpus_nr; i++) {
1552                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1553                 struct thread *curr_thread = sched->curr_thread[cpu];
1554                 const char *pid_color = color;
1555                 const char *cpu_color = color;
1556
1557                 if (curr_thread && thread__has_color(curr_thread))
1558                         pid_color = COLOR_PIDS;
1559
1560                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1561                         continue;
1562
1563                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1564                         cpu_color = COLOR_CPUS;
1565
1566                 if (cpu != this_cpu)
1567                         color_fprintf(stdout, color, " ");
1568                 else
1569                         color_fprintf(stdout, cpu_color, "*");
1570
1571                 if (sched->curr_thread[cpu])
1572                         color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1573                 else
1574                         color_fprintf(stdout, color, "   ");
1575         }
1576
1577         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1578                 goto out;
1579
1580         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1581         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1582         if (new_shortname || (verbose > 0 && sched_in->tid)) {
1583                 const char *pid_color = color;
1584
1585                 if (thread__has_color(sched_in))
1586                         pid_color = COLOR_PIDS;
1587
1588                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1589                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1590         }
1591
1592         if (sched->map.comp && new_cpu)
1593                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1594
1595 out:
1596         color_fprintf(stdout, color, "\n");
1597
1598         thread__put(sched_in);
1599
1600         return 0;
1601 }
1602
1603 static int process_sched_switch_event(struct perf_tool *tool,
1604                                       struct perf_evsel *evsel,
1605                                       struct perf_sample *sample,
1606                                       struct machine *machine)
1607 {
1608         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1609         int this_cpu = sample->cpu, err = 0;
1610         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1611             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1612
1613         if (sched->curr_pid[this_cpu] != (u32)-1) {
1614                 /*
1615                  * Are we trying to switch away a PID that is
1616                  * not current?
1617                  */
1618                 if (sched->curr_pid[this_cpu] != prev_pid)
1619                         sched->nr_context_switch_bugs++;
1620         }
1621
1622         if (sched->tp_handler->switch_event)
1623                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1624
1625         sched->curr_pid[this_cpu] = next_pid;
1626         return err;
1627 }
1628
1629 static int process_sched_runtime_event(struct perf_tool *tool,
1630                                        struct perf_evsel *evsel,
1631                                        struct perf_sample *sample,
1632                                        struct machine *machine)
1633 {
1634         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1635
1636         if (sched->tp_handler->runtime_event)
1637                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1638
1639         return 0;
1640 }
1641
1642 static int perf_sched__process_fork_event(struct perf_tool *tool,
1643                                           union perf_event *event,
1644                                           struct perf_sample *sample,
1645                                           struct machine *machine)
1646 {
1647         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1648
1649         /* run the fork event through the perf machineruy */
1650         perf_event__process_fork(tool, event, sample, machine);
1651
1652         /* and then run additional processing needed for this command */
1653         if (sched->tp_handler->fork_event)
1654                 return sched->tp_handler->fork_event(sched, event, machine);
1655
1656         return 0;
1657 }
1658
1659 static int process_sched_migrate_task_event(struct perf_tool *tool,
1660                                             struct perf_evsel *evsel,
1661                                             struct perf_sample *sample,
1662                                             struct machine *machine)
1663 {
1664         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1665
1666         if (sched->tp_handler->migrate_task_event)
1667                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1668
1669         return 0;
1670 }
1671
1672 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1673                                   struct perf_evsel *evsel,
1674                                   struct perf_sample *sample,
1675                                   struct machine *machine);
1676
1677 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1678                                                  union perf_event *event __maybe_unused,
1679                                                  struct perf_sample *sample,
1680                                                  struct perf_evsel *evsel,
1681                                                  struct machine *machine)
1682 {
1683         int err = 0;
1684
1685         if (evsel->handler != NULL) {
1686                 tracepoint_handler f = evsel->handler;
1687                 err = f(tool, evsel, sample, machine);
1688         }
1689
1690         return err;
1691 }
1692
1693 static int perf_sched__read_events(struct perf_sched *sched)
1694 {
1695         const struct perf_evsel_str_handler handlers[] = {
1696                 { "sched:sched_switch",       process_sched_switch_event, },
1697                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1698                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1699                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1700                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1701         };
1702         struct perf_session *session;
1703         struct perf_data data = {
1704                 .file      = {
1705                         .path = input_name,
1706                 },
1707                 .mode      = PERF_DATA_MODE_READ,
1708                 .force     = sched->force,
1709         };
1710         int rc = -1;
1711
1712         session = perf_session__new(&data, false, &sched->tool);
1713         if (session == NULL) {
1714                 pr_debug("No Memory for session\n");
1715                 return -1;
1716         }
1717
1718         symbol__init(&session->header.env);
1719
1720         if (perf_session__set_tracepoints_handlers(session, handlers))
1721                 goto out_delete;
1722
1723         if (perf_session__has_traces(session, "record -R")) {
1724                 int err = perf_session__process_events(session);
1725                 if (err) {
1726                         pr_err("Failed to process events, error %d", err);
1727                         goto out_delete;
1728                 }
1729
1730                 sched->nr_events      = session->evlist->stats.nr_events[0];
1731                 sched->nr_lost_events = session->evlist->stats.total_lost;
1732                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1733         }
1734
1735         rc = 0;
1736 out_delete:
1737         perf_session__delete(session);
1738         return rc;
1739 }
1740
1741 /*
1742  * scheduling times are printed as msec.usec
1743  */
1744 static inline void print_sched_time(unsigned long long nsecs, int width)
1745 {
1746         unsigned long msecs;
1747         unsigned long usecs;
1748
1749         msecs  = nsecs / NSEC_PER_MSEC;
1750         nsecs -= msecs * NSEC_PER_MSEC;
1751         usecs  = nsecs / NSEC_PER_USEC;
1752         printf("%*lu.%03lu ", width, msecs, usecs);
1753 }
1754
1755 /*
1756  * returns runtime data for event, allocating memory for it the
1757  * first time it is used.
1758  */
1759 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1760 {
1761         struct evsel_runtime *r = evsel->priv;
1762
1763         if (r == NULL) {
1764                 r = zalloc(sizeof(struct evsel_runtime));
1765                 evsel->priv = r;
1766         }
1767
1768         return r;
1769 }
1770
1771 /*
1772  * save last time event was seen per cpu
1773  */
1774 static void perf_evsel__save_time(struct perf_evsel *evsel,
1775                                   u64 timestamp, u32 cpu)
1776 {
1777         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1778
1779         if (r == NULL)
1780                 return;
1781
1782         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1783                 int i, n = __roundup_pow_of_two(cpu+1);
1784                 void *p = r->last_time;
1785
1786                 p = realloc(r->last_time, n * sizeof(u64));
1787                 if (!p)
1788                         return;
1789
1790                 r->last_time = p;
1791                 for (i = r->ncpu; i < n; ++i)
1792                         r->last_time[i] = (u64) 0;
1793
1794                 r->ncpu = n;
1795         }
1796
1797         r->last_time[cpu] = timestamp;
1798 }
1799
1800 /* returns last time this event was seen on the given cpu */
1801 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1802 {
1803         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1804
1805         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1806                 return 0;
1807
1808         return r->last_time[cpu];
1809 }
1810
1811 static int comm_width = 30;
1812
1813 static char *timehist_get_commstr(struct thread *thread)
1814 {
1815         static char str[32];
1816         const char *comm = thread__comm_str(thread);
1817         pid_t tid = thread->tid;
1818         pid_t pid = thread->pid_;
1819         int n;
1820
1821         if (pid == 0)
1822                 n = scnprintf(str, sizeof(str), "%s", comm);
1823
1824         else if (tid != pid)
1825                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1826
1827         else
1828                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1829
1830         if (n > comm_width)
1831                 comm_width = n;
1832
1833         return str;
1834 }
1835
1836 static void timehist_header(struct perf_sched *sched)
1837 {
1838         u32 ncpus = sched->max_cpu + 1;
1839         u32 i, j;
1840
1841         printf("%15s %6s ", "time", "cpu");
1842
1843         if (sched->show_cpu_visual) {
1844                 printf(" ");
1845                 for (i = 0, j = 0; i < ncpus; ++i) {
1846                         printf("%x", j++);
1847                         if (j > 15)
1848                                 j = 0;
1849                 }
1850                 printf(" ");
1851         }
1852
1853         printf(" %-*s  %9s  %9s  %9s", comm_width,
1854                 "task name", "wait time", "sch delay", "run time");
1855
1856         if (sched->show_state)
1857                 printf("  %s", "state");
1858
1859         printf("\n");
1860
1861         /*
1862          * units row
1863          */
1864         printf("%15s %-6s ", "", "");
1865
1866         if (sched->show_cpu_visual)
1867                 printf(" %*s ", ncpus, "");
1868
1869         printf(" %-*s  %9s  %9s  %9s", comm_width,
1870                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1871
1872         if (sched->show_state)
1873                 printf("  %5s", "");
1874
1875         printf("\n");
1876
1877         /*
1878          * separator
1879          */
1880         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1881
1882         if (sched->show_cpu_visual)
1883                 printf(" %.*s ", ncpus, graph_dotted_line);
1884
1885         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1886                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1887                 graph_dotted_line);
1888
1889         if (sched->show_state)
1890                 printf("  %.5s", graph_dotted_line);
1891
1892         printf("\n");
1893 }
1894
1895 static char task_state_char(struct thread *thread, int state)
1896 {
1897         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1898         unsigned bit = state ? ffs(state) : 0;
1899
1900         /* 'I' for idle */
1901         if (thread->tid == 0)
1902                 return 'I';
1903
1904         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1905 }
1906
1907 static void timehist_print_sample(struct perf_sched *sched,
1908                                   struct perf_evsel *evsel,
1909                                   struct perf_sample *sample,
1910                                   struct addr_location *al,
1911                                   struct thread *thread,
1912                                   u64 t, int state)
1913 {
1914         struct thread_runtime *tr = thread__priv(thread);
1915         const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1916         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1917         u32 max_cpus = sched->max_cpu + 1;
1918         char tstr[64];
1919         char nstr[30];
1920         u64 wait_time;
1921
1922         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1923         printf("%15s [%04d] ", tstr, sample->cpu);
1924
1925         if (sched->show_cpu_visual) {
1926                 u32 i;
1927                 char c;
1928
1929                 printf(" ");
1930                 for (i = 0; i < max_cpus; ++i) {
1931                         /* flag idle times with 'i'; others are sched events */
1932                         if (i == sample->cpu)
1933                                 c = (thread->tid == 0) ? 'i' : 's';
1934                         else
1935                                 c = ' ';
1936                         printf("%c", c);
1937                 }
1938                 printf(" ");
1939         }
1940
1941         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1942
1943         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
1944         print_sched_time(wait_time, 6);
1945
1946         print_sched_time(tr->dt_delay, 6);
1947         print_sched_time(tr->dt_run, 6);
1948
1949         if (sched->show_state)
1950                 printf(" %5c ", task_state_char(thread, state));
1951
1952         if (sched->show_next) {
1953                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
1954                 printf(" %-*s", comm_width, nstr);
1955         }
1956
1957         if (sched->show_wakeups && !sched->show_next)
1958                 printf("  %-*s", comm_width, "");
1959
1960         if (thread->tid == 0)
1961                 goto out;
1962
1963         if (sched->show_callchain)
1964                 printf("  ");
1965
1966         sample__fprintf_sym(sample, al, 0,
1967                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1968                             EVSEL__PRINT_CALLCHAIN_ARROW |
1969                             EVSEL__PRINT_SKIP_IGNORED,
1970                             &callchain_cursor, stdout);
1971
1972 out:
1973         printf("\n");
1974 }
1975
1976 /*
1977  * Explanation of delta-time stats:
1978  *
1979  *            t = time of current schedule out event
1980  *        tprev = time of previous sched out event
1981  *                also time of schedule-in event for current task
1982  *    last_time = time of last sched change event for current task
1983  *                (i.e, time process was last scheduled out)
1984  * ready_to_run = time of wakeup for current task
1985  *
1986  * -----|------------|------------|------------|------
1987  *    last         ready        tprev          t
1988  *    time         to run
1989  *
1990  *      |-------- dt_wait --------|
1991  *                   |- dt_delay -|-- dt_run --|
1992  *
1993  *   dt_run = run time of current task
1994  *  dt_wait = time between last schedule out event for task and tprev
1995  *            represents time spent off the cpu
1996  * dt_delay = time between wakeup and schedule-in of task
1997  */
1998
1999 static void timehist_update_runtime_stats(struct thread_runtime *r,
2000                                          u64 t, u64 tprev)
2001 {
2002         r->dt_delay   = 0;
2003         r->dt_sleep   = 0;
2004         r->dt_iowait  = 0;
2005         r->dt_preempt = 0;
2006         r->dt_run     = 0;
2007
2008         if (tprev) {
2009                 r->dt_run = t - tprev;
2010                 if (r->ready_to_run) {
2011                         if (r->ready_to_run > tprev)
2012                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2013                         else
2014                                 r->dt_delay = tprev - r->ready_to_run;
2015                 }
2016
2017                 if (r->last_time > tprev)
2018                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2019                 else if (r->last_time) {
2020                         u64 dt_wait = tprev - r->last_time;
2021
2022                         if (r->last_state == TASK_RUNNING)
2023                                 r->dt_preempt = dt_wait;
2024                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2025                                 r->dt_iowait = dt_wait;
2026                         else
2027                                 r->dt_sleep = dt_wait;
2028                 }
2029         }
2030
2031         update_stats(&r->run_stats, r->dt_run);
2032
2033         r->total_run_time     += r->dt_run;
2034         r->total_delay_time   += r->dt_delay;
2035         r->total_sleep_time   += r->dt_sleep;
2036         r->total_iowait_time  += r->dt_iowait;
2037         r->total_preempt_time += r->dt_preempt;
2038 }
2039
2040 static bool is_idle_sample(struct perf_sample *sample,
2041                            struct perf_evsel *evsel)
2042 {
2043         /* pid 0 == swapper == idle task */
2044         if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2045                 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2046
2047         return sample->pid == 0;
2048 }
2049
2050 static void save_task_callchain(struct perf_sched *sched,
2051                                 struct perf_sample *sample,
2052                                 struct perf_evsel *evsel,
2053                                 struct machine *machine)
2054 {
2055         struct callchain_cursor *cursor = &callchain_cursor;
2056         struct thread *thread;
2057
2058         /* want main thread for process - has maps */
2059         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2060         if (thread == NULL) {
2061                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2062                 return;
2063         }
2064
2065         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2066                 return;
2067
2068         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2069                                       NULL, NULL, sched->max_stack + 2) != 0) {
2070                 if (verbose > 0)
2071                         pr_err("Failed to resolve callchain. Skipping\n");
2072
2073                 return;
2074         }
2075
2076         callchain_cursor_commit(cursor);
2077
2078         while (true) {
2079                 struct callchain_cursor_node *node;
2080                 struct symbol *sym;
2081
2082                 node = callchain_cursor_current(cursor);
2083                 if (node == NULL)
2084                         break;
2085
2086                 sym = node->sym;
2087                 if (sym) {
2088                         if (!strcmp(sym->name, "schedule") ||
2089                             !strcmp(sym->name, "__schedule") ||
2090                             !strcmp(sym->name, "preempt_schedule"))
2091                                 sym->ignore = 1;
2092                 }
2093
2094                 callchain_cursor_advance(cursor);
2095         }
2096 }
2097
2098 static int init_idle_thread(struct thread *thread)
2099 {
2100         struct idle_thread_runtime *itr;
2101
2102         thread__set_comm(thread, idle_comm, 0);
2103
2104         itr = zalloc(sizeof(*itr));
2105         if (itr == NULL)
2106                 return -ENOMEM;
2107
2108         init_stats(&itr->tr.run_stats);
2109         callchain_init(&itr->callchain);
2110         callchain_cursor_reset(&itr->cursor);
2111         thread__set_priv(thread, itr);
2112
2113         return 0;
2114 }
2115
2116 /*
2117  * Track idle stats per cpu by maintaining a local thread
2118  * struct for the idle task on each cpu.
2119  */
2120 static int init_idle_threads(int ncpu)
2121 {
2122         int i, ret;
2123
2124         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2125         if (!idle_threads)
2126                 return -ENOMEM;
2127
2128         idle_max_cpu = ncpu;
2129
2130         /* allocate the actual thread struct if needed */
2131         for (i = 0; i < ncpu; ++i) {
2132                 idle_threads[i] = thread__new(0, 0);
2133                 if (idle_threads[i] == NULL)
2134                         return -ENOMEM;
2135
2136                 ret = init_idle_thread(idle_threads[i]);
2137                 if (ret < 0)
2138                         return ret;
2139         }
2140
2141         return 0;
2142 }
2143
2144 static void free_idle_threads(void)
2145 {
2146         int i;
2147
2148         if (idle_threads == NULL)
2149                 return;
2150
2151         for (i = 0; i < idle_max_cpu; ++i) {
2152                 if ((idle_threads[i]))
2153                         thread__delete(idle_threads[i]);
2154         }
2155
2156         free(idle_threads);
2157 }
2158
2159 static struct thread *get_idle_thread(int cpu)
2160 {
2161         /*
2162          * expand/allocate array of pointers to local thread
2163          * structs if needed
2164          */
2165         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2166                 int i, j = __roundup_pow_of_two(cpu+1);
2167                 void *p;
2168
2169                 p = realloc(idle_threads, j * sizeof(struct thread *));
2170                 if (!p)
2171                         return NULL;
2172
2173                 idle_threads = (struct thread **) p;
2174                 for (i = idle_max_cpu; i < j; ++i)
2175                         idle_threads[i] = NULL;
2176
2177                 idle_max_cpu = j;
2178         }
2179
2180         /* allocate a new thread struct if needed */
2181         if (idle_threads[cpu] == NULL) {
2182                 idle_threads[cpu] = thread__new(0, 0);
2183
2184                 if (idle_threads[cpu]) {
2185                         if (init_idle_thread(idle_threads[cpu]) < 0)
2186                                 return NULL;
2187                 }
2188         }
2189
2190         return idle_threads[cpu];
2191 }
2192
2193 static void save_idle_callchain(struct idle_thread_runtime *itr,
2194                                 struct perf_sample *sample)
2195 {
2196         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2197                 return;
2198
2199         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2200 }
2201
2202 /*
2203  * handle runtime stats saved per thread
2204  */
2205 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2206 {
2207         struct thread_runtime *r;
2208
2209         r = zalloc(sizeof(struct thread_runtime));
2210         if (!r)
2211                 return NULL;
2212
2213         init_stats(&r->run_stats);
2214         thread__set_priv(thread, r);
2215
2216         return r;
2217 }
2218
2219 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2220 {
2221         struct thread_runtime *tr;
2222
2223         tr = thread__priv(thread);
2224         if (tr == NULL) {
2225                 tr = thread__init_runtime(thread);
2226                 if (tr == NULL)
2227                         pr_debug("Failed to malloc memory for runtime data.\n");
2228         }
2229
2230         return tr;
2231 }
2232
2233 static struct thread *timehist_get_thread(struct perf_sched *sched,
2234                                           struct perf_sample *sample,
2235                                           struct machine *machine,
2236                                           struct perf_evsel *evsel)
2237 {
2238         struct thread *thread;
2239
2240         if (is_idle_sample(sample, evsel)) {
2241                 thread = get_idle_thread(sample->cpu);
2242                 if (thread == NULL)
2243                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2244
2245         } else {
2246                 /* there were samples with tid 0 but non-zero pid */
2247                 thread = machine__findnew_thread(machine, sample->pid,
2248                                                  sample->tid ?: sample->pid);
2249                 if (thread == NULL) {
2250                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2251                                  sample->tid);
2252                 }
2253
2254                 save_task_callchain(sched, sample, evsel, machine);
2255                 if (sched->idle_hist) {
2256                         struct thread *idle;
2257                         struct idle_thread_runtime *itr;
2258
2259                         idle = get_idle_thread(sample->cpu);
2260                         if (idle == NULL) {
2261                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2262                                 return NULL;
2263                         }
2264
2265                         itr = thread__priv(idle);
2266                         if (itr == NULL)
2267                                 return NULL;
2268
2269                         itr->last_thread = thread;
2270
2271                         /* copy task callchain when entering to idle */
2272                         if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2273                                 save_idle_callchain(itr, sample);
2274                 }
2275         }
2276
2277         return thread;
2278 }
2279
2280 static bool timehist_skip_sample(struct perf_sched *sched,
2281                                  struct thread *thread,
2282                                  struct perf_evsel *evsel,
2283                                  struct perf_sample *sample)
2284 {
2285         bool rc = false;
2286
2287         if (thread__is_filtered(thread)) {
2288                 rc = true;
2289                 sched->skipped_samples++;
2290         }
2291
2292         if (sched->idle_hist) {
2293                 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2294                         rc = true;
2295                 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2296                          perf_evsel__intval(evsel, sample, "next_pid") != 0)
2297                         rc = true;
2298         }
2299
2300         return rc;
2301 }
2302
2303 static void timehist_print_wakeup_event(struct perf_sched *sched,
2304                                         struct perf_evsel *evsel,
2305                                         struct perf_sample *sample,
2306                                         struct machine *machine,
2307                                         struct thread *awakened)
2308 {
2309         struct thread *thread;
2310         char tstr[64];
2311
2312         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2313         if (thread == NULL)
2314                 return;
2315
2316         /* show wakeup unless both awakee and awaker are filtered */
2317         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2318             timehist_skip_sample(sched, awakened, evsel, sample)) {
2319                 return;
2320         }
2321
2322         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2323         printf("%15s [%04d] ", tstr, sample->cpu);
2324         if (sched->show_cpu_visual)
2325                 printf(" %*s ", sched->max_cpu + 1, "");
2326
2327         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2328
2329         /* dt spacer */
2330         printf("  %9s  %9s  %9s ", "", "", "");
2331
2332         printf("awakened: %s", timehist_get_commstr(awakened));
2333
2334         printf("\n");
2335 }
2336
2337 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2338                                        union perf_event *event __maybe_unused,
2339                                        struct perf_evsel *evsel,
2340                                        struct perf_sample *sample,
2341                                        struct machine *machine)
2342 {
2343         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2344         struct thread *thread;
2345         struct thread_runtime *tr = NULL;
2346         /* want pid of awakened task not pid in sample */
2347         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2348
2349         thread = machine__findnew_thread(machine, 0, pid);
2350         if (thread == NULL)
2351                 return -1;
2352
2353         tr = thread__get_runtime(thread);
2354         if (tr == NULL)
2355                 return -1;
2356
2357         if (tr->ready_to_run == 0)
2358                 tr->ready_to_run = sample->time;
2359
2360         /* show wakeups if requested */
2361         if (sched->show_wakeups &&
2362             !perf_time__skip_sample(&sched->ptime, sample->time))
2363                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2364
2365         return 0;
2366 }
2367
2368 static void timehist_print_migration_event(struct perf_sched *sched,
2369                                         struct perf_evsel *evsel,
2370                                         struct perf_sample *sample,
2371                                         struct machine *machine,
2372                                         struct thread *migrated)
2373 {
2374         struct thread *thread;
2375         char tstr[64];
2376         u32 max_cpus = sched->max_cpu + 1;
2377         u32 ocpu, dcpu;
2378
2379         if (sched->summary_only)
2380                 return;
2381
2382         max_cpus = sched->max_cpu + 1;
2383         ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2384         dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2385
2386         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2387         if (thread == NULL)
2388                 return;
2389
2390         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2391             timehist_skip_sample(sched, migrated, evsel, sample)) {
2392                 return;
2393         }
2394
2395         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2396         printf("%15s [%04d] ", tstr, sample->cpu);
2397
2398         if (sched->show_cpu_visual) {
2399                 u32 i;
2400                 char c;
2401
2402                 printf("  ");
2403                 for (i = 0; i < max_cpus; ++i) {
2404                         c = (i == sample->cpu) ? 'm' : ' ';
2405                         printf("%c", c);
2406                 }
2407                 printf("  ");
2408         }
2409
2410         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2411
2412         /* dt spacer */
2413         printf("  %9s  %9s  %9s ", "", "", "");
2414
2415         printf("migrated: %s", timehist_get_commstr(migrated));
2416         printf(" cpu %d => %d", ocpu, dcpu);
2417
2418         printf("\n");
2419 }
2420
2421 static int timehist_migrate_task_event(struct perf_tool *tool,
2422                                        union perf_event *event __maybe_unused,
2423                                        struct perf_evsel *evsel,
2424                                        struct perf_sample *sample,
2425                                        struct machine *machine)
2426 {
2427         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2428         struct thread *thread;
2429         struct thread_runtime *tr = NULL;
2430         /* want pid of migrated task not pid in sample */
2431         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2432
2433         thread = machine__findnew_thread(machine, 0, pid);
2434         if (thread == NULL)
2435                 return -1;
2436
2437         tr = thread__get_runtime(thread);
2438         if (tr == NULL)
2439                 return -1;
2440
2441         tr->migrations++;
2442
2443         /* show migrations if requested */
2444         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2445
2446         return 0;
2447 }
2448
2449 static int timehist_sched_change_event(struct perf_tool *tool,
2450                                        union perf_event *event,
2451                                        struct perf_evsel *evsel,
2452                                        struct perf_sample *sample,
2453                                        struct machine *machine)
2454 {
2455         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2456         struct perf_time_interval *ptime = &sched->ptime;
2457         struct addr_location al;
2458         struct thread *thread;
2459         struct thread_runtime *tr = NULL;
2460         u64 tprev, t = sample->time;
2461         int rc = 0;
2462         int state = perf_evsel__intval(evsel, sample, "prev_state");
2463
2464
2465         if (machine__resolve(machine, &al, sample) < 0) {
2466                 pr_err("problem processing %d event. skipping it\n",
2467                        event->header.type);
2468                 rc = -1;
2469                 goto out;
2470         }
2471
2472         thread = timehist_get_thread(sched, sample, machine, evsel);
2473         if (thread == NULL) {
2474                 rc = -1;
2475                 goto out;
2476         }
2477
2478         if (timehist_skip_sample(sched, thread, evsel, sample))
2479                 goto out;
2480
2481         tr = thread__get_runtime(thread);
2482         if (tr == NULL) {
2483                 rc = -1;
2484                 goto out;
2485         }
2486
2487         tprev = perf_evsel__get_time(evsel, sample->cpu);
2488
2489         /*
2490          * If start time given:
2491          * - sample time is under window user cares about - skip sample
2492          * - tprev is under window user cares about  - reset to start of window
2493          */
2494         if (ptime->start && ptime->start > t)
2495                 goto out;
2496
2497         if (tprev && ptime->start > tprev)
2498                 tprev = ptime->start;
2499
2500         /*
2501          * If end time given:
2502          * - previous sched event is out of window - we are done
2503          * - sample time is beyond window user cares about - reset it
2504          *   to close out stats for time window interest
2505          */
2506         if (ptime->end) {
2507                 if (tprev > ptime->end)
2508                         goto out;
2509
2510                 if (t > ptime->end)
2511                         t = ptime->end;
2512         }
2513
2514         if (!sched->idle_hist || thread->tid == 0) {
2515                 timehist_update_runtime_stats(tr, t, tprev);
2516
2517                 if (sched->idle_hist) {
2518                         struct idle_thread_runtime *itr = (void *)tr;
2519                         struct thread_runtime *last_tr;
2520
2521                         BUG_ON(thread->tid != 0);
2522
2523                         if (itr->last_thread == NULL)
2524                                 goto out;
2525
2526                         /* add current idle time as last thread's runtime */
2527                         last_tr = thread__get_runtime(itr->last_thread);
2528                         if (last_tr == NULL)
2529                                 goto out;
2530
2531                         timehist_update_runtime_stats(last_tr, t, tprev);
2532                         /*
2533                          * remove delta time of last thread as it's not updated
2534                          * and otherwise it will show an invalid value next
2535                          * time.  we only care total run time and run stat.
2536                          */
2537                         last_tr->dt_run = 0;
2538                         last_tr->dt_delay = 0;
2539                         last_tr->dt_sleep = 0;
2540                         last_tr->dt_iowait = 0;
2541                         last_tr->dt_preempt = 0;
2542
2543                         if (itr->cursor.nr)
2544                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2545
2546                         itr->last_thread = NULL;
2547                 }
2548         }
2549
2550         if (!sched->summary_only)
2551                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2552
2553 out:
2554         if (sched->hist_time.start == 0 && t >= ptime->start)
2555                 sched->hist_time.start = t;
2556         if (ptime->end == 0 || t <= ptime->end)
2557                 sched->hist_time.end = t;
2558
2559         if (tr) {
2560                 /* time of this sched_switch event becomes last time task seen */
2561                 tr->last_time = sample->time;
2562
2563                 /* last state is used to determine where to account wait time */
2564                 tr->last_state = state;
2565
2566                 /* sched out event for task so reset ready to run time */
2567                 tr->ready_to_run = 0;
2568         }
2569
2570         perf_evsel__save_time(evsel, sample->time, sample->cpu);
2571
2572         return rc;
2573 }
2574
2575 static int timehist_sched_switch_event(struct perf_tool *tool,
2576                              union perf_event *event,
2577                              struct perf_evsel *evsel,
2578                              struct perf_sample *sample,
2579                              struct machine *machine __maybe_unused)
2580 {
2581         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2582 }
2583
2584 static int process_lost(struct perf_tool *tool __maybe_unused,
2585                         union perf_event *event,
2586                         struct perf_sample *sample,
2587                         struct machine *machine __maybe_unused)
2588 {
2589         char tstr[64];
2590
2591         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2592         printf("%15s ", tstr);
2593         printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2594
2595         return 0;
2596 }
2597
2598
2599 static void print_thread_runtime(struct thread *t,
2600                                  struct thread_runtime *r)
2601 {
2602         double mean = avg_stats(&r->run_stats);
2603         float stddev;
2604
2605         printf("%*s   %5d  %9" PRIu64 " ",
2606                comm_width, timehist_get_commstr(t), t->ppid,
2607                (u64) r->run_stats.n);
2608
2609         print_sched_time(r->total_run_time, 8);
2610         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2611         print_sched_time(r->run_stats.min, 6);
2612         printf(" ");
2613         print_sched_time((u64) mean, 6);
2614         printf(" ");
2615         print_sched_time(r->run_stats.max, 6);
2616         printf("  ");
2617         printf("%5.2f", stddev);
2618         printf("   %5" PRIu64, r->migrations);
2619         printf("\n");
2620 }
2621
2622 static void print_thread_waittime(struct thread *t,
2623                                   struct thread_runtime *r)
2624 {
2625         printf("%*s   %5d  %9" PRIu64 " ",
2626                comm_width, timehist_get_commstr(t), t->ppid,
2627                (u64) r->run_stats.n);
2628
2629         print_sched_time(r->total_run_time, 8);
2630         print_sched_time(r->total_sleep_time, 6);
2631         printf(" ");
2632         print_sched_time(r->total_iowait_time, 6);
2633         printf(" ");
2634         print_sched_time(r->total_preempt_time, 6);
2635         printf(" ");
2636         print_sched_time(r->total_delay_time, 6);
2637         printf("\n");
2638 }
2639
2640 struct total_run_stats {
2641         struct perf_sched *sched;
2642         u64  sched_count;
2643         u64  task_count;
2644         u64  total_run_time;
2645 };
2646
2647 static int __show_thread_runtime(struct thread *t, void *priv)
2648 {
2649         struct total_run_stats *stats = priv;
2650         struct thread_runtime *r;
2651
2652         if (thread__is_filtered(t))
2653                 return 0;
2654
2655         r = thread__priv(t);
2656         if (r && r->run_stats.n) {
2657                 stats->task_count++;
2658                 stats->sched_count += r->run_stats.n;
2659                 stats->total_run_time += r->total_run_time;
2660
2661                 if (stats->sched->show_state)
2662                         print_thread_waittime(t, r);
2663                 else
2664                         print_thread_runtime(t, r);
2665         }
2666
2667         return 0;
2668 }
2669
2670 static int show_thread_runtime(struct thread *t, void *priv)
2671 {
2672         if (t->dead)
2673                 return 0;
2674
2675         return __show_thread_runtime(t, priv);
2676 }
2677
2678 static int show_deadthread_runtime(struct thread *t, void *priv)
2679 {
2680         if (!t->dead)
2681                 return 0;
2682
2683         return __show_thread_runtime(t, priv);
2684 }
2685
2686 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2687 {
2688         const char *sep = " <- ";
2689         struct callchain_list *chain;
2690         size_t ret = 0;
2691         char bf[1024];
2692         bool first;
2693
2694         if (node == NULL)
2695                 return 0;
2696
2697         ret = callchain__fprintf_folded(fp, node->parent);
2698         first = (ret == 0);
2699
2700         list_for_each_entry(chain, &node->val, list) {
2701                 if (chain->ip >= PERF_CONTEXT_MAX)
2702                         continue;
2703                 if (chain->ms.sym && chain->ms.sym->ignore)
2704                         continue;
2705                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2706                                callchain_list__sym_name(chain, bf, sizeof(bf),
2707                                                         false));
2708                 first = false;
2709         }
2710
2711         return ret;
2712 }
2713
2714 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2715 {
2716         size_t ret = 0;
2717         FILE *fp = stdout;
2718         struct callchain_node *chain;
2719         struct rb_node *rb_node = rb_first(root);
2720
2721         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2722         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2723                graph_dotted_line);
2724
2725         while (rb_node) {
2726                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2727                 rb_node = rb_next(rb_node);
2728
2729                 ret += fprintf(fp, "  ");
2730                 print_sched_time(chain->hit, 12);
2731                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2732                 ret += fprintf(fp, " %8d  ", chain->count);
2733                 ret += callchain__fprintf_folded(fp, chain);
2734                 ret += fprintf(fp, "\n");
2735         }
2736
2737         return ret;
2738 }
2739
2740 static void timehist_print_summary(struct perf_sched *sched,
2741                                    struct perf_session *session)
2742 {
2743         struct machine *m = &session->machines.host;
2744         struct total_run_stats totals;
2745         u64 task_count;
2746         struct thread *t;
2747         struct thread_runtime *r;
2748         int i;
2749         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2750
2751         memset(&totals, 0, sizeof(totals));
2752         totals.sched = sched;
2753
2754         if (sched->idle_hist) {
2755                 printf("\nIdle-time summary\n");
2756                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2757                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2758         } else if (sched->show_state) {
2759                 printf("\nWait-time summary\n");
2760                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2761                 printf("   run-time      sleep      iowait     preempt       delay\n");
2762         } else {
2763                 printf("\nRuntime summary\n");
2764                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2765                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2766         }
2767         printf("%*s            (count)  ", comm_width, "");
2768         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2769                sched->show_state ? "(msec)" : "%");
2770         printf("%.117s\n", graph_dotted_line);
2771
2772         machine__for_each_thread(m, show_thread_runtime, &totals);
2773         task_count = totals.task_count;
2774         if (!task_count)
2775                 printf("<no still running tasks>\n");
2776
2777         printf("\nTerminated tasks:\n");
2778         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2779         if (task_count == totals.task_count)
2780                 printf("<no terminated tasks>\n");
2781
2782         /* CPU idle stats not tracked when samples were skipped */
2783         if (sched->skipped_samples && !sched->idle_hist)
2784                 return;
2785
2786         printf("\nIdle stats:\n");
2787         for (i = 0; i < idle_max_cpu; ++i) {
2788                 t = idle_threads[i];
2789                 if (!t)
2790                         continue;
2791
2792                 r = thread__priv(t);
2793                 if (r && r->run_stats.n) {
2794                         totals.sched_count += r->run_stats.n;
2795                         printf("    CPU %2d idle for ", i);
2796                         print_sched_time(r->total_run_time, 6);
2797                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2798                 } else
2799                         printf("    CPU %2d idle entire time window\n", i);
2800         }
2801
2802         if (sched->idle_hist && symbol_conf.use_callchain) {
2803                 callchain_param.mode  = CHAIN_FOLDED;
2804                 callchain_param.value = CCVAL_PERIOD;
2805
2806                 callchain_register_param(&callchain_param);
2807
2808                 printf("\nIdle stats by callchain:\n");
2809                 for (i = 0; i < idle_max_cpu; ++i) {
2810                         struct idle_thread_runtime *itr;
2811
2812                         t = idle_threads[i];
2813                         if (!t)
2814                                 continue;
2815
2816                         itr = thread__priv(t);
2817                         if (itr == NULL)
2818                                 continue;
2819
2820                         callchain_param.sort(&itr->sorted_root, &itr->callchain,
2821                                              0, &callchain_param);
2822
2823                         printf("  CPU %2d:", i);
2824                         print_sched_time(itr->tr.total_run_time, 6);
2825                         printf(" msec\n");
2826                         timehist_print_idlehist_callchain(&itr->sorted_root);
2827                         printf("\n");
2828                 }
2829         }
2830
2831         printf("\n"
2832                "    Total number of unique tasks: %" PRIu64 "\n"
2833                "Total number of context switches: %" PRIu64 "\n",
2834                totals.task_count, totals.sched_count);
2835
2836         printf("           Total run time (msec): ");
2837         print_sched_time(totals.total_run_time, 2);
2838         printf("\n");
2839
2840         printf("    Total scheduling time (msec): ");
2841         print_sched_time(hist_time, 2);
2842         printf(" (x %d)\n", sched->max_cpu);
2843 }
2844
2845 typedef int (*sched_handler)(struct perf_tool *tool,
2846                           union perf_event *event,
2847                           struct perf_evsel *evsel,
2848                           struct perf_sample *sample,
2849                           struct machine *machine);
2850
2851 static int perf_timehist__process_sample(struct perf_tool *tool,
2852                                          union perf_event *event,
2853                                          struct perf_sample *sample,
2854                                          struct perf_evsel *evsel,
2855                                          struct machine *machine)
2856 {
2857         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2858         int err = 0;
2859         int this_cpu = sample->cpu;
2860
2861         if (this_cpu > sched->max_cpu)
2862                 sched->max_cpu = this_cpu;
2863
2864         if (evsel->handler != NULL) {
2865                 sched_handler f = evsel->handler;
2866
2867                 err = f(tool, event, evsel, sample, machine);
2868         }
2869
2870         return err;
2871 }
2872
2873 static int timehist_check_attr(struct perf_sched *sched,
2874                                struct perf_evlist *evlist)
2875 {
2876         struct perf_evsel *evsel;
2877         struct evsel_runtime *er;
2878
2879         list_for_each_entry(evsel, &evlist->entries, node) {
2880                 er = perf_evsel__get_runtime(evsel);
2881                 if (er == NULL) {
2882                         pr_err("Failed to allocate memory for evsel runtime data\n");
2883                         return -1;
2884                 }
2885
2886                 if (sched->show_callchain &&
2887                     !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2888                         pr_info("Samples do not have callchains.\n");
2889                         sched->show_callchain = 0;
2890                         symbol_conf.use_callchain = 0;
2891                 }
2892         }
2893
2894         return 0;
2895 }
2896
2897 static int perf_sched__timehist(struct perf_sched *sched)
2898 {
2899         const struct perf_evsel_str_handler handlers[] = {
2900                 { "sched:sched_switch",       timehist_sched_switch_event, },
2901                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2902                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2903         };
2904         const struct perf_evsel_str_handler migrate_handlers[] = {
2905                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2906         };
2907         struct perf_data data = {
2908                 .file      = {
2909                         .path = input_name,
2910                 },
2911                 .mode      = PERF_DATA_MODE_READ,
2912                 .force     = sched->force,
2913         };
2914
2915         struct perf_session *session;
2916         struct perf_evlist *evlist;
2917         int err = -1;
2918
2919         /*
2920          * event handlers for timehist option
2921          */
2922         sched->tool.sample       = perf_timehist__process_sample;
2923         sched->tool.mmap         = perf_event__process_mmap;
2924         sched->tool.comm         = perf_event__process_comm;
2925         sched->tool.exit         = perf_event__process_exit;
2926         sched->tool.fork         = perf_event__process_fork;
2927         sched->tool.lost         = process_lost;
2928         sched->tool.attr         = perf_event__process_attr;
2929         sched->tool.tracing_data = perf_event__process_tracing_data;
2930         sched->tool.build_id     = perf_event__process_build_id;
2931
2932         sched->tool.ordered_events = true;
2933         sched->tool.ordering_requires_timestamps = true;
2934
2935         symbol_conf.use_callchain = sched->show_callchain;
2936
2937         session = perf_session__new(&data, false, &sched->tool);
2938         if (session == NULL)
2939                 return -ENOMEM;
2940
2941         evlist = session->evlist;
2942
2943         symbol__init(&session->header.env);
2944
2945         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2946                 pr_err("Invalid time string\n");
2947                 return -EINVAL;
2948         }
2949
2950         if (timehist_check_attr(sched, evlist) != 0)
2951                 goto out;
2952
2953         setup_pager();
2954
2955         /* setup per-evsel handlers */
2956         if (perf_session__set_tracepoints_handlers(session, handlers))
2957                 goto out;
2958
2959         /* sched_switch event at a minimum needs to exist */
2960         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2961                                                   "sched:sched_switch")) {
2962                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2963                 goto out;
2964         }
2965
2966         if (sched->show_migrations &&
2967             perf_session__set_tracepoints_handlers(session, migrate_handlers))
2968                 goto out;
2969
2970         /* pre-allocate struct for per-CPU idle stats */
2971         sched->max_cpu = session->header.env.nr_cpus_online;
2972         if (sched->max_cpu == 0)
2973                 sched->max_cpu = 4;
2974         if (init_idle_threads(sched->max_cpu))
2975                 goto out;
2976
2977         /* summary_only implies summary option, but don't overwrite summary if set */
2978         if (sched->summary_only)
2979                 sched->summary = sched->summary_only;
2980
2981         if (!sched->summary_only)
2982                 timehist_header(sched);
2983
2984         err = perf_session__process_events(session);
2985         if (err) {
2986                 pr_err("Failed to process events, error %d", err);
2987                 goto out;
2988         }
2989
2990         sched->nr_events      = evlist->stats.nr_events[0];
2991         sched->nr_lost_events = evlist->stats.total_lost;
2992         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2993
2994         if (sched->summary)
2995                 timehist_print_summary(sched, session);
2996
2997 out:
2998         free_idle_threads();
2999         perf_session__delete(session);
3000
3001         return err;
3002 }
3003
3004
3005 static void print_bad_events(struct perf_sched *sched)
3006 {
3007         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3008                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3009                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3010                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3011         }
3012         if (sched->nr_lost_events && sched->nr_events) {
3013                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3014                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3015                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3016         }
3017         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3018                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3019                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3020                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3021                 if (sched->nr_lost_events)
3022                         printf(" (due to lost events?)");
3023                 printf("\n");
3024         }
3025 }
3026
3027 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3028 {
3029         struct rb_node **new = &(root->rb_node), *parent = NULL;
3030         struct work_atoms *this;
3031         const char *comm = thread__comm_str(data->thread), *this_comm;
3032
3033         while (*new) {
3034                 int cmp;
3035
3036                 this = container_of(*new, struct work_atoms, node);
3037                 parent = *new;
3038
3039                 this_comm = thread__comm_str(this->thread);
3040                 cmp = strcmp(comm, this_comm);
3041                 if (cmp > 0) {
3042                         new = &((*new)->rb_left);
3043                 } else if (cmp < 0) {
3044                         new = &((*new)->rb_right);
3045                 } else {
3046                         this->num_merged++;
3047                         this->total_runtime += data->total_runtime;
3048                         this->nb_atoms += data->nb_atoms;
3049                         this->total_lat += data->total_lat;
3050                         list_splice(&data->work_list, &this->work_list);
3051                         if (this->max_lat < data->max_lat) {
3052                                 this->max_lat = data->max_lat;
3053                                 this->max_lat_at = data->max_lat_at;
3054                         }
3055                         zfree(&data);
3056                         return;
3057                 }
3058         }
3059
3060         data->num_merged++;
3061         rb_link_node(&data->node, parent, new);
3062         rb_insert_color(&data->node, root);
3063 }
3064
3065 static void perf_sched__merge_lat(struct perf_sched *sched)
3066 {
3067         struct work_atoms *data;
3068         struct rb_node *node;
3069
3070         if (sched->skip_merge)
3071                 return;
3072
3073         while ((node = rb_first(&sched->atom_root))) {
3074                 rb_erase(node, &sched->atom_root);
3075                 data = rb_entry(node, struct work_atoms, node);
3076                 __merge_work_atoms(&sched->merged_atom_root, data);
3077         }
3078 }
3079
3080 static int perf_sched__lat(struct perf_sched *sched)
3081 {
3082         struct rb_node *next;
3083
3084         setup_pager();
3085
3086         if (perf_sched__read_events(sched))
3087                 return -1;
3088
3089         perf_sched__merge_lat(sched);
3090         perf_sched__sort_lat(sched);
3091
3092         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3093         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3094         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3095
3096         next = rb_first(&sched->sorted_atom_root);
3097
3098         while (next) {
3099                 struct work_atoms *work_list;
3100
3101                 work_list = rb_entry(next, struct work_atoms, node);
3102                 output_lat_thread(sched, work_list);
3103                 next = rb_next(next);
3104                 thread__zput(work_list->thread);
3105         }
3106
3107         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3108         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3109                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3110
3111         printf(" ---------------------------------------------------\n");
3112
3113         print_bad_events(sched);
3114         printf("\n");
3115
3116         return 0;
3117 }
3118
3119 static int setup_map_cpus(struct perf_sched *sched)
3120 {
3121         struct cpu_map *map;
3122
3123         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3124
3125         if (sched->map.comp) {
3126                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3127                 if (!sched->map.comp_cpus)
3128                         return -1;
3129         }
3130
3131         if (!sched->map.cpus_str)
3132                 return 0;
3133
3134         map = cpu_map__new(sched->map.cpus_str);
3135         if (!map) {
3136                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3137                 return -1;
3138         }
3139
3140         sched->map.cpus = map;
3141         return 0;
3142 }
3143
3144 static int setup_color_pids(struct perf_sched *sched)
3145 {
3146         struct thread_map *map;
3147
3148         if (!sched->map.color_pids_str)
3149                 return 0;
3150
3151         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3152         if (!map) {
3153                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3154                 return -1;
3155         }
3156
3157         sched->map.color_pids = map;
3158         return 0;
3159 }
3160
3161 static int setup_color_cpus(struct perf_sched *sched)
3162 {
3163         struct cpu_map *map;
3164
3165         if (!sched->map.color_cpus_str)
3166                 return 0;
3167
3168         map = cpu_map__new(sched->map.color_cpus_str);
3169         if (!map) {
3170                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3171                 return -1;
3172         }
3173
3174         sched->map.color_cpus = map;
3175         return 0;
3176 }
3177
3178 static int perf_sched__map(struct perf_sched *sched)
3179 {
3180         if (setup_map_cpus(sched))
3181                 return -1;
3182
3183         if (setup_color_pids(sched))
3184                 return -1;
3185
3186         if (setup_color_cpus(sched))
3187                 return -1;
3188
3189         setup_pager();
3190         if (perf_sched__read_events(sched))
3191                 return -1;
3192         print_bad_events(sched);
3193         return 0;
3194 }
3195
3196 static int perf_sched__replay(struct perf_sched *sched)
3197 {
3198         unsigned long i;
3199
3200         calibrate_run_measurement_overhead(sched);
3201         calibrate_sleep_measurement_overhead(sched);
3202
3203         test_calibrations(sched);
3204
3205         if (perf_sched__read_events(sched))
3206                 return -1;
3207
3208         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3209         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3210         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3211
3212         if (sched->targetless_wakeups)
3213                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3214         if (sched->multitarget_wakeups)
3215                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3216         if (sched->nr_run_events_optimized)
3217                 printf("run atoms optimized: %ld\n",
3218                         sched->nr_run_events_optimized);
3219
3220         print_task_traces(sched);
3221         add_cross_task_wakeups(sched);
3222
3223         create_tasks(sched);
3224         printf("------------------------------------------------------------\n");
3225         for (i = 0; i < sched->replay_repeat; i++)
3226                 run_one_test(sched);
3227
3228         return 0;
3229 }
3230
3231 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3232                           const char * const usage_msg[])
3233 {
3234         char *tmp, *tok, *str = strdup(sched->sort_order);
3235
3236         for (tok = strtok_r(str, ", ", &tmp);
3237                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3238                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3239                         usage_with_options_msg(usage_msg, options,
3240                                         "Unknown --sort key: `%s'", tok);
3241                 }
3242         }
3243
3244         free(str);
3245
3246         sort_dimension__add("pid", &sched->cmp_pid);
3247 }
3248
3249 static int __cmd_record(int argc, const char **argv)
3250 {
3251         unsigned int rec_argc, i, j;
3252         const char **rec_argv;
3253         const char * const record_args[] = {
3254                 "record",
3255                 "-a",
3256                 "-R",
3257                 "-m", "1024",
3258                 "-c", "1",
3259                 "-e", "sched:sched_switch",
3260                 "-e", "sched:sched_stat_wait",
3261                 "-e", "sched:sched_stat_sleep",
3262                 "-e", "sched:sched_stat_iowait",
3263                 "-e", "sched:sched_stat_runtime",
3264                 "-e", "sched:sched_process_fork",
3265                 "-e", "sched:sched_wakeup",
3266                 "-e", "sched:sched_wakeup_new",
3267                 "-e", "sched:sched_migrate_task",
3268         };
3269
3270         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3271         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3272
3273         if (rec_argv == NULL)
3274                 return -ENOMEM;
3275
3276         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3277                 rec_argv[i] = strdup(record_args[i]);
3278
3279         for (j = 1; j < (unsigned int)argc; j++, i++)
3280                 rec_argv[i] = argv[j];
3281
3282         BUG_ON(i != rec_argc);
3283
3284         return cmd_record(i, rec_argv);
3285 }
3286
3287 int cmd_sched(int argc, const char **argv)
3288 {
3289         const char default_sort_order[] = "avg, max, switch, runtime";
3290         struct perf_sched sched = {
3291                 .tool = {
3292                         .sample          = perf_sched__process_tracepoint_sample,
3293                         .comm            = perf_event__process_comm,
3294                         .namespaces      = perf_event__process_namespaces,
3295                         .lost            = perf_event__process_lost,
3296                         .fork            = perf_sched__process_fork_event,
3297                         .ordered_events = true,
3298                 },
3299                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3300                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3301                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3302                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3303                 .sort_order           = default_sort_order,
3304                 .replay_repeat        = 10,
3305                 .profile_cpu          = -1,
3306                 .next_shortname1      = 'A',
3307                 .next_shortname2      = '0',
3308                 .skip_merge           = 0,
3309                 .show_callchain       = 1,
3310                 .max_stack            = 5,
3311         };
3312         const struct option sched_options[] = {
3313         OPT_STRING('i', "input", &input_name, "file",
3314                     "input file name"),
3315         OPT_INCR('v', "verbose", &verbose,
3316                     "be more verbose (show symbol address, etc)"),
3317         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3318                     "dump raw trace in ASCII"),
3319         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3320         OPT_END()
3321         };
3322         const struct option latency_options[] = {
3323         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3324                    "sort by key(s): runtime, switch, avg, max"),
3325         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3326                     "CPU to profile on"),
3327         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3328                     "latency stats per pid instead of per comm"),
3329         OPT_PARENT(sched_options)
3330         };
3331         const struct option replay_options[] = {
3332         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3333                      "repeat the workload replay N times (-1: infinite)"),
3334         OPT_PARENT(sched_options)
3335         };
3336         const struct option map_options[] = {
3337         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3338                     "map output in compact mode"),
3339         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3340                    "highlight given pids in map"),
3341         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3342                     "highlight given CPUs in map"),
3343         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3344                     "display given CPUs in map"),
3345         OPT_PARENT(sched_options)
3346         };
3347         const struct option timehist_options[] = {
3348         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3349                    "file", "vmlinux pathname"),
3350         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3351                    "file", "kallsyms pathname"),
3352         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3353                     "Display call chains if present (default on)"),
3354         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3355                    "Maximum number of functions to display backtrace."),
3356         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3357                     "Look for files with symbols relative to this directory"),
3358         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3359                     "Show only syscall summary with statistics"),
3360         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3361                     "Show all syscalls and summary with statistics"),
3362         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3363         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3364         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3365         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3366         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3367         OPT_STRING(0, "time", &sched.time_str, "str",
3368                    "Time span for analysis (start,stop)"),
3369         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3370         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3371                    "analyze events only for given process id(s)"),
3372         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3373                    "analyze events only for given thread id(s)"),
3374         OPT_PARENT(sched_options)
3375         };
3376
3377         const char * const latency_usage[] = {
3378                 "perf sched latency [<options>]",
3379                 NULL
3380         };
3381         const char * const replay_usage[] = {
3382                 "perf sched replay [<options>]",
3383                 NULL
3384         };
3385         const char * const map_usage[] = {
3386                 "perf sched map [<options>]",
3387                 NULL
3388         };
3389         const char * const timehist_usage[] = {
3390                 "perf sched timehist [<options>]",
3391                 NULL
3392         };
3393         const char *const sched_subcommands[] = { "record", "latency", "map",
3394                                                   "replay", "script",
3395                                                   "timehist", NULL };
3396         const char *sched_usage[] = {
3397                 NULL,
3398                 NULL
3399         };
3400         struct trace_sched_handler lat_ops  = {
3401                 .wakeup_event       = latency_wakeup_event,
3402                 .switch_event       = latency_switch_event,
3403                 .runtime_event      = latency_runtime_event,
3404                 .migrate_task_event = latency_migrate_task_event,
3405         };
3406         struct trace_sched_handler map_ops  = {
3407                 .switch_event       = map_switch_event,
3408         };
3409         struct trace_sched_handler replay_ops  = {
3410                 .wakeup_event       = replay_wakeup_event,
3411                 .switch_event       = replay_switch_event,
3412                 .fork_event         = replay_fork_event,
3413         };
3414         unsigned int i;
3415
3416         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3417                 sched.curr_pid[i] = -1;
3418
3419         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3420                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3421         if (!argc)
3422                 usage_with_options(sched_usage, sched_options);
3423
3424         /*
3425          * Aliased to 'perf script' for now:
3426          */
3427         if (!strcmp(argv[0], "script"))
3428                 return cmd_script(argc, argv);
3429
3430         if (!strncmp(argv[0], "rec", 3)) {
3431                 return __cmd_record(argc, argv);
3432         } else if (!strncmp(argv[0], "lat", 3)) {
3433                 sched.tp_handler = &lat_ops;
3434                 if (argc > 1) {
3435                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3436                         if (argc)
3437                                 usage_with_options(latency_usage, latency_options);
3438                 }
3439                 setup_sorting(&sched, latency_options, latency_usage);
3440                 return perf_sched__lat(&sched);
3441         } else if (!strcmp(argv[0], "map")) {
3442                 if (argc) {
3443                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3444                         if (argc)
3445                                 usage_with_options(map_usage, map_options);
3446                 }
3447                 sched.tp_handler = &map_ops;
3448                 setup_sorting(&sched, latency_options, latency_usage);
3449                 return perf_sched__map(&sched);
3450         } else if (!strncmp(argv[0], "rep", 3)) {
3451                 sched.tp_handler = &replay_ops;
3452                 if (argc) {
3453                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3454                         if (argc)
3455                                 usage_with_options(replay_usage, replay_options);
3456                 }
3457                 return perf_sched__replay(&sched);
3458         } else if (!strcmp(argv[0], "timehist")) {
3459                 if (argc) {
3460                         argc = parse_options(argc, argv, timehist_options,
3461                                              timehist_usage, 0);
3462                         if (argc)
3463                                 usage_with_options(timehist_usage, timehist_options);
3464                 }
3465                 if ((sched.show_wakeups || sched.show_next) &&
3466                     sched.summary_only) {
3467                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3468                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3469                         if (sched.show_wakeups)
3470                                 parse_options_usage(NULL, timehist_options, "w", true);
3471                         if (sched.show_next)
3472                                 parse_options_usage(NULL, timehist_options, "n", true);
3473                         return -EINVAL;
3474                 }
3475
3476                 return perf_sched__timehist(&sched);
3477         } else {
3478                 usage_with_options(sched_usage, sched_options);
3479         }
3480
3481         return 0;
3482 }