tools/perf/builtin-sched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf.h"
   4
   5 #include "util/util.h"
   6 #include "util/evlist.h"
   7 #include "util/cache.h"
   8 #include "util/evsel.h"
   9 #include "util/symbol.h"
  10 #include "util/thread.h"
  11 #include "util/header.h"
  12 #include "util/session.h"
  13 #include "util/tool.h"
  14 #include "util/cloexec.h"
  15 #include "util/thread_map.h"
  16 #include "util/color.h"
  17 #include "util/stat.h"
  18 #include "util/callchain.h"
  19 #include "util/time-utils.h"
  20
  21 #include <subcmd/parse-options.h>
  22 #include "util/trace-event.h"
  23
  24 #include "util/debug.h"
  25
  26 #include <linux/kernel.h>
  27 #include <linux/log2.h>
  28 #include <sys/prctl.h>
  29 #include <sys/resource.h>
  30 #include <inttypes.h>
  31
  32 #include <errno.h>
  33 #include <semaphore.h>
  34 #include <pthread.h>
  35 #include <math.h>
  36 #include <api/fs/fs.h>
  37 #include <linux/time64.h>
  38
  39 #include "sane_ctype.h"
  40
  41 #define PR_SET_NAME             15               /* Set process name */
  42 #define MAX_CPUS                4096
  43 #define COMM_LEN                20
  44 #define SYM_LEN                 129
  45 #define MAX_PID                 1024000
  46
  47 struct sched_atom;
  48
  49 struct task_desc {
  50         unsigned long           nr;
  51         unsigned long           pid;
  52         char                    comm[COMM_LEN];
  53
  54         unsigned long           nr_events;
  55         unsigned long           curr_event;
  56         struct sched_atom       **atoms;
  57
  58         pthread_t               thread;
  59         sem_t                   sleep_sem;
  60
  61         sem_t                   ready_for_work;
  62         sem_t                   work_done_sem;
  63
  64         u64                     cpu_usage;
  65 };
  66
  67 enum sched_event_type {
  68         SCHED_EVENT_RUN,
  69         SCHED_EVENT_SLEEP,
  70         SCHED_EVENT_WAKEUP,
  71         SCHED_EVENT_MIGRATION,
  72 };
  73
  74 struct sched_atom {
  75         enum sched_event_type   type;
  76         int                     specific_wait;
  77         u64                     timestamp;
  78         u64                     duration;
  79         unsigned long           nr;
  80         sem_t                   *wait_sem;
  81         struct task_desc        *wakee;
  82 };
  83
  84 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  85
  86 /* task state bitmask, copied from include/linux/sched.h */
  87 #define TASK_RUNNING            0
  88 #define TASK_INTERRUPTIBLE      1
  89 #define TASK_UNINTERRUPTIBLE    2
  90 #define __TASK_STOPPED          4
  91 #define __TASK_TRACED           8
  92 /* in tsk->exit_state */
  93 #define EXIT_DEAD               16
  94 #define EXIT_ZOMBIE             32
  95 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
  96 /* in tsk->state again */
  97 #define TASK_DEAD               64
  98 #define TASK_WAKEKILL           128
  99 #define TASK_WAKING             256
 100 #define TASK_PARKED             512
 101
 102 enum thread_state {
 103         THREAD_SLEEPING = 0,
 104         THREAD_WAIT_CPU,
 105         THREAD_SCHED_IN,
 106         THREAD_IGNORE
 107 };
 108
 109 struct work_atom {
 110         struct list_head        list;
 111         enum thread_state       state;
 112         u64                     sched_out_time;
 113         u64                     wake_up_time;
 114         u64                     sched_in_time;
 115         u64                     runtime;
 116 };
 117
 118 struct work_atoms {
 119         struct list_head        work_list;
 120         struct thread           *thread;
 121         struct rb_node          node;
 122         u64                     max_lat;
 123         u64                     max_lat_at;
 124         u64                     total_lat;
 125         u64                     nb_atoms;
 126         u64                     total_runtime;
 127         int                     num_merged;
 128 };
 129
 130 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 131
 132 struct perf_sched;
 133
 134 struct trace_sched_handler {
 135         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 136                             struct perf_sample *sample, struct machine *machine);
 137
 138         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 139                              struct perf_sample *sample, struct machine *machine);
 140
 141         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 142                             struct perf_sample *sample, struct machine *machine);
 143
 144         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 145         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 146                           struct machine *machine);
 147
 148         int (*migrate_task_event)(struct perf_sched *sched,
 149                                   struct perf_evsel *evsel,
 150                                   struct perf_sample *sample,
 151                                   struct machine *machine);
 152 };
 153
 154 #define COLOR_PIDS PERF_COLOR_BLUE
 155 #define COLOR_CPUS PERF_COLOR_BG_RED
 156
 157 struct perf_sched_map {
 158         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 159         int                     *comp_cpus;
 160         bool                     comp;
 161         struct thread_map       *color_pids;
 162         const char              *color_pids_str;
 163         struct cpu_map          *color_cpus;
 164         const char              *color_cpus_str;
 165         struct cpu_map          *cpus;
 166         const char              *cpus_str;
 167 };
 168
 169 struct perf_sched {
 170         struct perf_tool tool;
 171         const char       *sort_order;
 172         unsigned long    nr_tasks;
 173         struct task_desc **pid_to_task;
 174         struct task_desc **tasks;
 175         const struct trace_sched_handler *tp_handler;
 176         pthread_mutex_t  start_work_mutex;
 177         pthread_mutex_t  work_done_wait_mutex;
 178         int              profile_cpu;
 179 /*
 180  * Track the current task - that way we can know whether there's any
 181  * weird events, such as a task being switched away that is not current.
 182  */
 183         int              max_cpu;
 184         u32              curr_pid[MAX_CPUS];
 185         struct thread    *curr_thread[MAX_CPUS];
 186         char             next_shortname1;
 187         char             next_shortname2;
 188         unsigned int     replay_repeat;
 189         unsigned long    nr_run_events;
 190         unsigned long    nr_sleep_events;
 191         unsigned long    nr_wakeup_events;
 192         unsigned long    nr_sleep_corrections;
 193         unsigned long    nr_run_events_optimized;
 194         unsigned long    targetless_wakeups;
 195         unsigned long    multitarget_wakeups;
 196         unsigned long    nr_runs;
 197         unsigned long    nr_timestamps;
 198         unsigned long    nr_unordered_timestamps;
 199         unsigned long    nr_context_switch_bugs;
 200         unsigned long    nr_events;
 201         unsigned long    nr_lost_chunks;
 202         unsigned long    nr_lost_events;
 203         u64              run_measurement_overhead;
 204         u64              sleep_measurement_overhead;
 205         u64              start_time;
 206         u64              cpu_usage;
 207         u64              runavg_cpu_usage;
 208         u64              parent_cpu_usage;
 209         u64              runavg_parent_cpu_usage;
 210         u64              sum_runtime;
 211         u64              sum_fluct;
 212         u64              run_avg;
 213         u64              all_runtime;
 214         u64              all_count;
 215         u64              cpu_last_switched[MAX_CPUS];
 216         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 217         struct list_head sort_list, cmp_pid;
 218         bool force;
 219         bool skip_merge;
 220         struct perf_sched_map map;
 221
 222         /* options for timehist command */
 223         bool            summary;
 224         bool            summary_only;
 225         bool            idle_hist;
 226         bool            show_callchain;
 227         unsigned int    max_stack;
 228         bool            show_cpu_visual;
 229         bool            show_wakeups;
 230         bool            show_next;
 231         bool            show_migrations;
 232         bool            show_state;
 233         u64             skipped_samples;
 234         const char      *time_str;
 235         struct perf_time_interval ptime;
 236         struct perf_time_interval hist_time;
 237 };
 238
 239 /* per thread run time data */
 240 struct thread_runtime {
 241         u64 last_time;      /* time of previous sched in/out event */
 242         u64 dt_run;         /* run time */
 243         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 244         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 245         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 246         u64 dt_delay;       /* time between wakeup and sched-in */
 247         u64 ready_to_run;   /* time of wakeup */
 248
 249         struct stats run_stats;
 250         u64 total_run_time;
 251         u64 total_sleep_time;
 252         u64 total_iowait_time;
 253         u64 total_preempt_time;
 254         u64 total_delay_time;
 255
 256         int last_state;
 257         u64 migrations;
 258 };
 259
 260 /* per event run time data */
 261 struct evsel_runtime {
 262         u64 *last_time; /* time this event was last seen per cpu */
 263         u32 ncpu;       /* highest cpu slot allocated */
 264 };
 265
 266 /* per cpu idle time data */
 267 struct idle_thread_runtime {
 268         struct thread_runtime   tr;
 269         struct thread           *last_thread;
 270         struct rb_root          sorted_root;
 271         struct callchain_root   callchain;
 272         struct callchain_cursor cursor;
 273 };
 274
 275 /* track idle times per cpu */
 276 static struct thread **idle_threads;
 277 static int idle_max_cpu;
 278 static char idle_comm[] = "<idle>";
 279
 280 static u64 get_nsecs(void)
 281 {
 282         struct timespec ts;
 283
 284         clock_gettime(CLOCK_MONOTONIC, &ts);
 285
 286         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 287 }
 288
 289 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 290 {
 291         u64 T0 = get_nsecs(), T1;
 292
 293         do {
 294                 T1 = get_nsecs();
 295         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 296 }
 297
 298 static void sleep_nsecs(u64 nsecs)
 299 {
 300         struct timespec ts;
 301
 302         ts.tv_nsec = nsecs % 999999999;
 303         ts.tv_sec = nsecs / 999999999;
 304
 305         nanosleep(&ts, NULL);
 306 }
 307
 308 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 309 {
 310         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 311         int i;
 312
 313         for (i = 0; i < 10; i++) {
 314                 T0 = get_nsecs();
 315                 burn_nsecs(sched, 0);
 316                 T1 = get_nsecs();
 317                 delta = T1-T0;
 318                 min_delta = min(min_delta, delta);
 319         }
 320         sched->run_measurement_overhead = min_delta;
 321
 322         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 323 }
 324
 325 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 326 {
 327         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 328         int i;
 329
 330         for (i = 0; i < 10; i++) {
 331                 T0 = get_nsecs();
 332                 sleep_nsecs(10000);
 333                 T1 = get_nsecs();
 334                 delta = T1-T0;
 335                 min_delta = min(min_delta, delta);
 336         }
 337         min_delta -= 10000;
 338         sched->sleep_measurement_overhead = min_delta;
 339
 340         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 341 }
 342
 343 static struct sched_atom *
 344 get_new_event(struct task_desc *task, u64 timestamp)
 345 {
 346         struct sched_atom *event = zalloc(sizeof(*event));
 347         unsigned long idx = task->nr_events;
 348         size_t size;
 349
 350         event->timestamp = timestamp;
 351         event->nr = idx;
 352
 353         task->nr_events++;
 354         size = sizeof(struct sched_atom *) * task->nr_events;
 355         task->atoms = realloc(task->atoms, size);
 356         BUG_ON(!task->atoms);
 357
 358         task->atoms[idx] = event;
 359
 360         return event;
 361 }
 362
 363 static struct sched_atom *last_event(struct task_desc *task)
 364 {
 365         if (!task->nr_events)
 366                 return NULL;
 367
 368         return task->atoms[task->nr_events - 1];
 369 }
 370
 371 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 372                                 u64 timestamp, u64 duration)
 373 {
 374         struct sched_atom *event, *curr_event = last_event(task);
 375
 376         /*
 377          * optimize an existing RUN event by merging this one
 378          * to it:
 379          */
 380         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 381                 sched->nr_run_events_optimized++;
 382                 curr_event->duration += duration;
 383                 return;
 384         }
 385
 386         event = get_new_event(task, timestamp);
 387
 388         event->type = SCHED_EVENT_RUN;
 389         event->duration = duration;
 390
 391         sched->nr_run_events++;
 392 }
 393
 394 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 395                                    u64 timestamp, struct task_desc *wakee)
 396 {
 397         struct sched_atom *event, *wakee_event;
 398
 399         event = get_new_event(task, timestamp);
 400         event->type = SCHED_EVENT_WAKEUP;
 401         event->wakee = wakee;
 402
 403         wakee_event = last_event(wakee);
 404         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 405                 sched->targetless_wakeups++;
 406                 return;
 407         }
 408         if (wakee_event->wait_sem) {
 409                 sched->multitarget_wakeups++;
 410                 return;
 411         }
 412
 413         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 414         sem_init(wakee_event->wait_sem, 0, 0);
 415         wakee_event->specific_wait = 1;
 416         event->wait_sem = wakee_event->wait_sem;
 417
 418         sched->nr_wakeup_events++;
 419 }
 420
 421 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 422                                   u64 timestamp, u64 task_state __maybe_unused)
 423 {
 424         struct sched_atom *event = get_new_event(task, timestamp);
 425
 426         event->type = SCHED_EVENT_SLEEP;
 427
 428         sched->nr_sleep_events++;
 429 }
 430
 431 static struct task_desc *register_pid(struct perf_sched *sched,
 432                                       unsigned long pid, const char *comm)
 433 {
 434         struct task_desc *task;
 435         static int pid_max;
 436
 437         if (sched->pid_to_task == NULL) {
 438                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 439                         pid_max = MAX_PID;
 440                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 441         }
 442         if (pid >= (unsigned long)pid_max) {
 443                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 444                         sizeof(struct task_desc *))) == NULL);
 445                 while (pid >= (unsigned long)pid_max)
 446                         sched->pid_to_task[pid_max++] = NULL;
 447         }
 448
 449         task = sched->pid_to_task[pid];
 450
 451         if (task)
 452                 return task;
 453
 454         task = zalloc(sizeof(*task));
 455         task->pid = pid;
 456         task->nr = sched->nr_tasks;
 457         strcpy(task->comm, comm);
 458         /*
 459          * every task starts in sleeping state - this gets ignored
 460          * if there's no wakeup pointing to this sleep state:
 461          */
 462         add_sched_event_sleep(sched, task, 0, 0);
 463
 464         sched->pid_to_task[pid] = task;
 465         sched->nr_tasks++;
 466         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 467         BUG_ON(!sched->tasks);
 468         sched->tasks[task->nr] = task;
 469
 470         if (verbose > 0)
 471                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 472
 473         return task;
 474 }
 475
 476
 477 static void print_task_traces(struct perf_sched *sched)
 478 {
 479         struct task_desc *task;
 480         unsigned long i;
 481
 482         for (i = 0; i < sched->nr_tasks; i++) {
 483                 task = sched->tasks[i];
 484                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 485                         task->nr, task->comm, task->pid, task->nr_events);
 486         }
 487 }
 488
 489 static void add_cross_task_wakeups(struct perf_sched *sched)
 490 {
 491         struct task_desc *task1, *task2;
 492         unsigned long i, j;
 493
 494         for (i = 0; i < sched->nr_tasks; i++) {
 495                 task1 = sched->tasks[i];
 496                 j = i + 1;
 497                 if (j == sched->nr_tasks)
 498                         j = 0;
 499                 task2 = sched->tasks[j];
 500                 add_sched_event_wakeup(sched, task1, 0, task2);
 501         }
 502 }
 503
 504 static void perf_sched__process_event(struct perf_sched *sched,
 505                                       struct sched_atom *atom)
 506 {
 507         int ret = 0;
 508
 509         switch (atom->type) {
 510                 case SCHED_EVENT_RUN:
 511                         burn_nsecs(sched, atom->duration);
 512                         break;
 513                 case SCHED_EVENT_SLEEP:
 514                         if (atom->wait_sem)
 515                                 ret = sem_wait(atom->wait_sem);
 516                         BUG_ON(ret);
 517                         break;
 518                 case SCHED_EVENT_WAKEUP:
 519                         if (atom->wait_sem)
 520                                 ret = sem_post(atom->wait_sem);
 521                         BUG_ON(ret);
 522                         break;
 523                 case SCHED_EVENT_MIGRATION:
 524                         break;
 525                 default:
 526                         BUG_ON(1);
 527         }
 528 }
 529
 530 static u64 get_cpu_usage_nsec_parent(void)
 531 {
 532         struct rusage ru;
 533         u64 sum;
 534         int err;
 535
 536         err = getrusage(RUSAGE_SELF, &ru);
 537         BUG_ON(err);
 538
 539         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 540         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 541
 542         return sum;
 543 }
 544
 545 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 546 {
 547         struct perf_event_attr attr;
 548         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 549         int fd;
 550         struct rlimit limit;
 551         bool need_privilege = false;
 552
 553         memset(&attr, 0, sizeof(attr));
 554
 555         attr.type = PERF_TYPE_SOFTWARE;
 556         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 557
 558 force_again:
 559         fd = sys_perf_event_open(&attr, 0, -1, -1,
 560                                  perf_event_open_cloexec_flag());
 561
 562         if (fd < 0) {
 563                 if (errno == EMFILE) {
 564                         if (sched->force) {
 565                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 566                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 567                                 if (limit.rlim_cur > limit.rlim_max) {
 568                                         limit.rlim_max = limit.rlim_cur;
 569                                         need_privilege = true;
 570                                 }
 571                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 572                                         if (need_privilege && errno == EPERM)
 573                                                 strcpy(info, "Need privilege\n");
 574                                 } else
 575                                         goto force_again;
 576                         } else
 577                                 strcpy(info, "Have a try with -f option\n");
 578                 }
 579                 pr_err("Error: sys_perf_event_open() syscall returned "
 580                        "with %d (%s)\n%s", fd,
 581                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 582                 exit(EXIT_FAILURE);
 583         }
 584         return fd;
 585 }
 586
 587 static u64 get_cpu_usage_nsec_self(int fd)
 588 {
 589         u64 runtime;
 590         int ret;
 591
 592         ret = read(fd, &runtime, sizeof(runtime));
 593         BUG_ON(ret != sizeof(runtime));
 594
 595         return runtime;
 596 }
 597
 598 struct sched_thread_parms {
 599         struct task_desc  *task;
 600         struct perf_sched *sched;
 601         int fd;
 602 };
 603
 604 static void *thread_func(void *ctx)
 605 {
 606         struct sched_thread_parms *parms = ctx;
 607         struct task_desc *this_task = parms->task;
 608         struct perf_sched *sched = parms->sched;
 609         u64 cpu_usage_0, cpu_usage_1;
 610         unsigned long i, ret;
 611         char comm2[22];
 612         int fd = parms->fd;
 613
 614         zfree(&parms);
 615
 616         sprintf(comm2, ":%s", this_task->comm);
 617         prctl(PR_SET_NAME, comm2);
 618         if (fd < 0)
 619                 return NULL;
 620 again:
 621         ret = sem_post(&this_task->ready_for_work);
 622         BUG_ON(ret);
 623         ret = pthread_mutex_lock(&sched->start_work_mutex);
 624         BUG_ON(ret);
 625         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 626         BUG_ON(ret);
 627
 628         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 629
 630         for (i = 0; i < this_task->nr_events; i++) {
 631                 this_task->curr_event = i;
 632                 perf_sched__process_event(sched, this_task->atoms[i]);
 633         }
 634
 635         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 636         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 637         ret = sem_post(&this_task->work_done_sem);
 638         BUG_ON(ret);
 639
 640         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 641         BUG_ON(ret);
 642         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 643         BUG_ON(ret);
 644
 645         goto again;
 646 }
 647
 648 static void create_tasks(struct perf_sched *sched)
 649 {
 650         struct task_desc *task;
 651         pthread_attr_t attr;
 652         unsigned long i;
 653         int err;
 654
 655         err = pthread_attr_init(&attr);
 656         BUG_ON(err);
 657         err = pthread_attr_setstacksize(&attr,
 658                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 659         BUG_ON(err);
 660         err = pthread_mutex_lock(&sched->start_work_mutex);
 661         BUG_ON(err);
 662         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 663         BUG_ON(err);
 664         for (i = 0; i < sched->nr_tasks; i++) {
 665                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 666                 BUG_ON(parms == NULL);
 667                 parms->task = task = sched->tasks[i];
 668                 parms->sched = sched;
 669                 parms->fd = self_open_counters(sched, i);
 670                 sem_init(&task->sleep_sem, 0, 0);
 671                 sem_init(&task->ready_for_work, 0, 0);
 672                 sem_init(&task->work_done_sem, 0, 0);
 673                 task->curr_event = 0;
 674                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 675                 BUG_ON(err);
 676         }
 677 }
 678
 679 static void wait_for_tasks(struct perf_sched *sched)
 680 {
 681         u64 cpu_usage_0, cpu_usage_1;
 682         struct task_desc *task;
 683         unsigned long i, ret;
 684
 685         sched->start_time = get_nsecs();
 686         sched->cpu_usage = 0;
 687         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 688
 689         for (i = 0; i < sched->nr_tasks; i++) {
 690                 task = sched->tasks[i];
 691                 ret = sem_wait(&task->ready_for_work);
 692                 BUG_ON(ret);
 693                 sem_init(&task->ready_for_work, 0, 0);
 694         }
 695         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 696         BUG_ON(ret);
 697
 698         cpu_usage_0 = get_cpu_usage_nsec_parent();
 699
 700         pthread_mutex_unlock(&sched->start_work_mutex);
 701
 702         for (i = 0; i < sched->nr_tasks; i++) {
 703                 task = sched->tasks[i];
 704                 ret = sem_wait(&task->work_done_sem);
 705                 BUG_ON(ret);
 706                 sem_init(&task->work_done_sem, 0, 0);
 707                 sched->cpu_usage += task->cpu_usage;
 708                 task->cpu_usage = 0;
 709         }
 710
 711         cpu_usage_1 = get_cpu_usage_nsec_parent();
 712         if (!sched->runavg_cpu_usage)
 713                 sched->runavg_cpu_usage = sched->cpu_usage;
 714         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 715
 716         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 717         if (!sched->runavg_parent_cpu_usage)
 718                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 719         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 720                                          sched->parent_cpu_usage)/sched->replay_repeat;
 721
 722         ret = pthread_mutex_lock(&sched->start_work_mutex);
 723         BUG_ON(ret);
 724
 725         for (i = 0; i < sched->nr_tasks; i++) {
 726                 task = sched->tasks[i];
 727                 sem_init(&task->sleep_sem, 0, 0);
 728                 task->curr_event = 0;
 729         }
 730 }
 731
 732 static void run_one_test(struct perf_sched *sched)
 733 {
 734         u64 T0, T1, delta, avg_delta, fluct;
 735
 736         T0 = get_nsecs();
 737         wait_for_tasks(sched);
 738         T1 = get_nsecs();
 739
 740         delta = T1 - T0;
 741         sched->sum_runtime += delta;
 742         sched->nr_runs++;
 743
 744         avg_delta = sched->sum_runtime / sched->nr_runs;
 745         if (delta < avg_delta)
 746                 fluct = avg_delta - delta;
 747         else
 748                 fluct = delta - avg_delta;
 749         sched->sum_fluct += fluct;
 750         if (!sched->run_avg)
 751                 sched->run_avg = delta;
 752         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 753
 754         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 755
 756         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 757
 758         printf("cpu: %0.2f / %0.2f",
 759                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 760
 761 #if 0
 762         /*
 763          * rusage statistics done by the parent, these are less
 764          * accurate than the sched->sum_exec_runtime based statistics:
 765          */
 766         printf(" [%0.2f / %0.2f]",
 767                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 768                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 769 #endif
 770
 771         printf("\n");
 772
 773         if (sched->nr_sleep_corrections)
 774                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 775         sched->nr_sleep_corrections = 0;
 776 }
 777
 778 static void test_calibrations(struct perf_sched *sched)
 779 {
 780         u64 T0, T1;
 781
 782         T0 = get_nsecs();
 783         burn_nsecs(sched, NSEC_PER_MSEC);
 784         T1 = get_nsecs();
 785
 786         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 787
 788         T0 = get_nsecs();
 789         sleep_nsecs(NSEC_PER_MSEC);
 790         T1 = get_nsecs();
 791
 792         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 793 }
 794
 795 static int
 796 replay_wakeup_event(struct perf_sched *sched,
 797                     struct perf_evsel *evsel, struct perf_sample *sample,
 798                     struct machine *machine __maybe_unused)
 799 {
 800         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 801         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 802         struct task_desc *waker, *wakee;
 803
 804         if (verbose > 0) {
 805                 printf("sched_wakeup event %p\n", evsel);
 806
 807                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 808         }
 809
 810         waker = register_pid(sched, sample->tid, "<unknown>");
 811         wakee = register_pid(sched, pid, comm);
 812
 813         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 814         return 0;
 815 }
 816
 817 static int replay_switch_event(struct perf_sched *sched,
 818                                struct perf_evsel *evsel,
 819                                struct perf_sample *sample,
 820                                struct machine *machine __maybe_unused)
 821 {
 822         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 823                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 824         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 825                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 826         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 827         struct task_desc *prev, __maybe_unused *next;
 828         u64 timestamp0, timestamp = sample->time;
 829         int cpu = sample->cpu;
 830         s64 delta;
 831
 832         if (verbose > 0)
 833                 printf("sched_switch event %p\n", evsel);
 834
 835         if (cpu >= MAX_CPUS || cpu < 0)
 836                 return 0;
 837
 838         timestamp0 = sched->cpu_last_switched[cpu];
 839         if (timestamp0)
 840                 delta = timestamp - timestamp0;
 841         else
 842                 delta = 0;
 843
 844         if (delta < 0) {
 845                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 846                 return -1;
 847         }
 848
 849         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 850                  prev_comm, prev_pid, next_comm, next_pid, delta);
 851
 852         prev = register_pid(sched, prev_pid, prev_comm);
 853         next = register_pid(sched, next_pid, next_comm);
 854
 855         sched->cpu_last_switched[cpu] = timestamp;
 856
 857         add_sched_event_run(sched, prev, timestamp, delta);
 858         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 859
 860         return 0;
 861 }
 862
 863 static int replay_fork_event(struct perf_sched *sched,
 864                              union perf_event *event,
 865                              struct machine *machine)
 866 {
 867         struct thread *child, *parent;
 868
 869         child = machine__findnew_thread(machine, event->fork.pid,
 870                                         event->fork.tid);
 871         parent = machine__findnew_thread(machine, event->fork.ppid,
 872                                          event->fork.ptid);
 873
 874         if (child == NULL || parent == NULL) {
 875                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 876                                  child, parent);
 877                 goto out_put;
 878         }
 879
 880         if (verbose > 0) {
 881                 printf("fork event\n");
 882                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 883                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 884         }
 885
 886         register_pid(sched, parent->tid, thread__comm_str(parent));
 887         register_pid(sched, child->tid, thread__comm_str(child));
 888 out_put:
 889         thread__put(child);
 890         thread__put(parent);
 891         return 0;
 892 }
 893
 894 struct sort_dimension {
 895         const char              *name;
 896         sort_fn_t               cmp;
 897         struct list_head        list;
 898 };
 899
 900 static int
 901 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 902 {
 903         struct sort_dimension *sort;
 904         int ret = 0;
 905
 906         BUG_ON(list_empty(list));
 907
 908         list_for_each_entry(sort, list, list) {
 909                 ret = sort->cmp(l, r);
 910                 if (ret)
 911                         return ret;
 912         }
 913
 914         return ret;
 915 }
 916
 917 static struct work_atoms *
 918 thread_atoms_search(struct rb_root *root, struct thread *thread,
 919                          struct list_head *sort_list)
 920 {
 921         struct rb_node *node = root->rb_node;
 922         struct work_atoms key = { .thread = thread };
 923
 924         while (node) {
 925                 struct work_atoms *atoms;
 926                 int cmp;
 927
 928                 atoms = container_of(node, struct work_atoms, node);
 929
 930                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 931                 if (cmp > 0)
 932                         node = node->rb_left;
 933                 else if (cmp < 0)
 934                         node = node->rb_right;
 935                 else {
 936                         BUG_ON(thread != atoms->thread);
 937                         return atoms;
 938                 }
 939         }
 940         return NULL;
 941 }
 942
 943 static void
 944 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 945                          struct list_head *sort_list)
 946 {
 947         struct rb_node **new = &(root->rb_node), *parent = NULL;
 948
 949         while (*new) {
 950                 struct work_atoms *this;
 951                 int cmp;
 952
 953                 this = container_of(*new, struct work_atoms, node);
 954                 parent = *new;
 955
 956                 cmp = thread_lat_cmp(sort_list, data, this);
 957
 958                 if (cmp > 0)
 959                         new = &((*new)->rb_left);
 960                 else
 961                         new = &((*new)->rb_right);
 962         }
 963
 964         rb_link_node(&data->node, parent, new);
 965         rb_insert_color(&data->node, root);
 966 }
 967
 968 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 969 {
 970         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 971         if (!atoms) {
 972                 pr_err("No memory at %s\n", __func__);
 973                 return -1;
 974         }
 975
 976         atoms->thread = thread__get(thread);
 977         INIT_LIST_HEAD(&atoms->work_list);
 978         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 979         return 0;
 980 }
 981
 982 static char sched_out_state(u64 prev_state)
 983 {
 984         const char *str = TASK_STATE_TO_CHAR_STR;
 985
 986         return str[prev_state];
 987 }
 988
 989 static int
 990 add_sched_out_event(struct work_atoms *atoms,
 991                     char run_state,
 992                     u64 timestamp)
 993 {
 994         struct work_atom *atom = zalloc(sizeof(*atom));
 995         if (!atom) {
 996                 pr_err("Non memory at %s", __func__);
 997                 return -1;
 998         }
 999
1000         atom->sched_out_time = timestamp;
1001
1002         if (run_state == 'R') {
1003                 atom->state = THREAD_WAIT_CPU;
1004                 atom->wake_up_time = atom->sched_out_time;
1005         }
1006
1007         list_add_tail(&atom->list, &atoms->work_list);
1008         return 0;
1009 }
1010
1011 static void
1012 add_runtime_event(struct work_atoms *atoms, u64 delta,
1013                   u64 timestamp __maybe_unused)
1014 {
1015         struct work_atom *atom;
1016
1017         BUG_ON(list_empty(&atoms->work_list));
1018
1019         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1020
1021         atom->runtime += delta;
1022         atoms->total_runtime += delta;
1023 }
1024
1025 static void
1026 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1027 {
1028         struct work_atom *atom;
1029         u64 delta;
1030
1031         if (list_empty(&atoms->work_list))
1032                 return;
1033
1034         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1035
1036         if (atom->state != THREAD_WAIT_CPU)
1037                 return;
1038
1039         if (timestamp < atom->wake_up_time) {
1040                 atom->state = THREAD_IGNORE;
1041                 return;
1042         }
1043
1044         atom->state = THREAD_SCHED_IN;
1045         atom->sched_in_time = timestamp;
1046
1047         delta = atom->sched_in_time - atom->wake_up_time;
1048         atoms->total_lat += delta;
1049         if (delta > atoms->max_lat) {
1050                 atoms->max_lat = delta;
1051                 atoms->max_lat_at = timestamp;
1052         }
1053         atoms->nb_atoms++;
1054 }
1055
1056 static int latency_switch_event(struct perf_sched *sched,
1057                                 struct perf_evsel *evsel,
1058                                 struct perf_sample *sample,
1059                                 struct machine *machine)
1060 {
1061         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1062                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1063         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1064         struct work_atoms *out_events, *in_events;
1065         struct thread *sched_out, *sched_in;
1066         u64 timestamp0, timestamp = sample->time;
1067         int cpu = sample->cpu, err = -1;
1068         s64 delta;
1069
1070         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1071
1072         timestamp0 = sched->cpu_last_switched[cpu];
1073         sched->cpu_last_switched[cpu] = timestamp;
1074         if (timestamp0)
1075                 delta = timestamp - timestamp0;
1076         else
1077                 delta = 0;
1078
1079         if (delta < 0) {
1080                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1081                 return -1;
1082         }
1083
1084         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1085         sched_in = machine__findnew_thread(machine, -1, next_pid);
1086         if (sched_out == NULL || sched_in == NULL)
1087                 goto out_put;
1088
1089         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1090         if (!out_events) {
1091                 if (thread_atoms_insert(sched, sched_out))
1092                         goto out_put;
1093                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1094                 if (!out_events) {
1095                         pr_err("out-event: Internal tree error");
1096                         goto out_put;
1097                 }
1098         }
1099         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1100                 return -1;
1101
1102         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1103         if (!in_events) {
1104                 if (thread_atoms_insert(sched, sched_in))
1105                         goto out_put;
1106                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1107                 if (!in_events) {
1108                         pr_err("in-event: Internal tree error");
1109                         goto out_put;
1110                 }
1111                 /*
1112                  * Take came in we have not heard about yet,
1113                  * add in an initial atom in runnable state:
1114                  */
1115                 if (add_sched_out_event(in_events, 'R', timestamp))
1116                         goto out_put;
1117         }
1118         add_sched_in_event(in_events, timestamp);
1119         err = 0;
1120 out_put:
1121         thread__put(sched_out);
1122         thread__put(sched_in);
1123         return err;
1124 }
1125
1126 static int latency_runtime_event(struct perf_sched *sched,
1127                                  struct perf_evsel *evsel,
1128                                  struct perf_sample *sample,
1129                                  struct machine *machine)
1130 {
1131         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1132         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1133         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1134         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1135         u64 timestamp = sample->time;
1136         int cpu = sample->cpu, err = -1;
1137
1138         if (thread == NULL)
1139                 return -1;
1140
1141         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1142         if (!atoms) {
1143                 if (thread_atoms_insert(sched, thread))
1144                         goto out_put;
1145                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1146                 if (!atoms) {
1147                         pr_err("in-event: Internal tree error");
1148                         goto out_put;
1149                 }
1150                 if (add_sched_out_event(atoms, 'R', timestamp))
1151                         goto out_put;
1152         }
1153
1154         add_runtime_event(atoms, runtime, timestamp);
1155         err = 0;
1156 out_put:
1157         thread__put(thread);
1158         return err;
1159 }
1160
1161 static int latency_wakeup_event(struct perf_sched *sched,
1162                                 struct perf_evsel *evsel,
1163                                 struct perf_sample *sample,
1164                                 struct machine *machine)
1165 {
1166         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1167         struct work_atoms *atoms;
1168         struct work_atom *atom;
1169         struct thread *wakee;
1170         u64 timestamp = sample->time;
1171         int err = -1;
1172
1173         wakee = machine__findnew_thread(machine, -1, pid);
1174         if (wakee == NULL)
1175                 return -1;
1176         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1177         if (!atoms) {
1178                 if (thread_atoms_insert(sched, wakee))
1179                         goto out_put;
1180                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1181                 if (!atoms) {
1182                         pr_err("wakeup-event: Internal tree error");
1183                         goto out_put;
1184                 }
1185                 if (add_sched_out_event(atoms, 'S', timestamp))
1186                         goto out_put;
1187         }
1188
1189         BUG_ON(list_empty(&atoms->work_list));
1190
1191         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1192
1193         /*
1194          * As we do not guarantee the wakeup event happens when
1195          * task is out of run queue, also may happen when task is
1196          * on run queue and wakeup only change ->state to TASK_RUNNING,
1197          * then we should not set the ->wake_up_time when wake up a
1198          * task which is on run queue.
1199          *
1200          * You WILL be missing events if you've recorded only
1201          * one CPU, or are only looking at only one, so don't
1202          * skip in this case.
1203          */
1204         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1205                 goto out_ok;
1206
1207         sched->nr_timestamps++;
1208         if (atom->sched_out_time > timestamp) {
1209                 sched->nr_unordered_timestamps++;
1210                 goto out_ok;
1211         }
1212
1213         atom->state = THREAD_WAIT_CPU;
1214         atom->wake_up_time = timestamp;
1215 out_ok:
1216         err = 0;
1217 out_put:
1218         thread__put(wakee);
1219         return err;
1220 }
1221
1222 static int latency_migrate_task_event(struct perf_sched *sched,
1223                                       struct perf_evsel *evsel,
1224                                       struct perf_sample *sample,
1225                                       struct machine *machine)
1226 {
1227         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1228         u64 timestamp = sample->time;
1229         struct work_atoms *atoms;
1230         struct work_atom *atom;
1231         struct thread *migrant;
1232         int err = -1;
1233
1234         /*
1235          * Only need to worry about migration when profiling one CPU.
1236          */
1237         if (sched->profile_cpu == -1)
1238                 return 0;
1239
1240         migrant = machine__findnew_thread(machine, -1, pid);
1241         if (migrant == NULL)
1242                 return -1;
1243         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1244         if (!atoms) {
1245                 if (thread_atoms_insert(sched, migrant))
1246                         goto out_put;
1247                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1248                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1249                 if (!atoms) {
1250                         pr_err("migration-event: Internal tree error");
1251                         goto out_put;
1252                 }
1253                 if (add_sched_out_event(atoms, 'R', timestamp))
1254                         goto out_put;
1255         }
1256
1257         BUG_ON(list_empty(&atoms->work_list));
1258
1259         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1260         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1261
1262         sched->nr_timestamps++;
1263
1264         if (atom->sched_out_time > timestamp)
1265                 sched->nr_unordered_timestamps++;
1266         err = 0;
1267 out_put:
1268         thread__put(migrant);
1269         return err;
1270 }
1271
1272 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1273 {
1274         int i;
1275         int ret;
1276         u64 avg;
1277         char max_lat_at[32];
1278
1279         if (!work_list->nb_atoms)
1280                 return;
1281         /*
1282          * Ignore idle threads:
1283          */
1284         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1285                 return;
1286
1287         sched->all_runtime += work_list->total_runtime;
1288         sched->all_count   += work_list->nb_atoms;
1289
1290         if (work_list->num_merged > 1)
1291                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1292         else
1293                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1294
1295         for (i = 0; i < 24 - ret; i++)
1296                 printf(" ");
1297
1298         avg = work_list->total_lat / work_list->nb_atoms;
1299         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1300
1301         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1302               (double)work_list->total_runtime / NSEC_PER_MSEC,
1303                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1304                  (double)work_list->max_lat / NSEC_PER_MSEC,
1305                  max_lat_at);
1306 }
1307
1308 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1309 {
1310         if (l->thread == r->thread)
1311                 return 0;
1312         if (l->thread->tid < r->thread->tid)
1313                 return -1;
1314         if (l->thread->tid > r->thread->tid)
1315                 return 1;
1316         return (int)(l->thread - r->thread);
1317 }
1318
1319 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1320 {
1321         u64 avgl, avgr;
1322
1323         if (!l->nb_atoms)
1324                 return -1;
1325
1326         if (!r->nb_atoms)
1327                 return 1;
1328
1329         avgl = l->total_lat / l->nb_atoms;
1330         avgr = r->total_lat / r->nb_atoms;
1331
1332         if (avgl < avgr)
1333                 return -1;
1334         if (avgl > avgr)
1335                 return 1;
1336
1337         return 0;
1338 }
1339
1340 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1341 {
1342         if (l->max_lat < r->max_lat)
1343                 return -1;
1344         if (l->max_lat > r->max_lat)
1345                 return 1;
1346
1347         return 0;
1348 }
1349
1350 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1351 {
1352         if (l->nb_atoms < r->nb_atoms)
1353                 return -1;
1354         if (l->nb_atoms > r->nb_atoms)
1355                 return 1;
1356
1357         return 0;
1358 }
1359
1360 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1361 {
1362         if (l->total_runtime < r->total_runtime)
1363                 return -1;
1364         if (l->total_runtime > r->total_runtime)
1365                 return 1;
1366
1367         return 0;
1368 }
1369
1370 static int sort_dimension__add(const char *tok, struct list_head *list)
1371 {
1372         size_t i;
1373         static struct sort_dimension avg_sort_dimension = {
1374                 .name = "avg",
1375                 .cmp  = avg_cmp,
1376         };
1377         static struct sort_dimension max_sort_dimension = {
1378                 .name = "max",
1379                 .cmp  = max_cmp,
1380         };
1381         static struct sort_dimension pid_sort_dimension = {
1382                 .name = "pid",
1383                 .cmp  = pid_cmp,
1384         };
1385         static struct sort_dimension runtime_sort_dimension = {
1386                 .name = "runtime",
1387                 .cmp  = runtime_cmp,
1388         };
1389         static struct sort_dimension switch_sort_dimension = {
1390                 .name = "switch",
1391                 .cmp  = switch_cmp,
1392         };
1393         struct sort_dimension *available_sorts[] = {
1394                 &pid_sort_dimension,
1395                 &avg_sort_dimension,
1396                 &max_sort_dimension,
1397                 &switch_sort_dimension,
1398                 &runtime_sort_dimension,
1399         };
1400
1401         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1402                 if (!strcmp(available_sorts[i]->name, tok)) {
1403                         list_add_tail(&available_sorts[i]->list, list);
1404
1405                         return 0;
1406                 }
1407         }
1408
1409         return -1;
1410 }
1411
1412 static void perf_sched__sort_lat(struct perf_sched *sched)
1413 {
1414         struct rb_node *node;
1415         struct rb_root *root = &sched->atom_root;
1416 again:
1417         for (;;) {
1418                 struct work_atoms *data;
1419                 node = rb_first(root);
1420                 if (!node)
1421                         break;
1422
1423                 rb_erase(node, root);
1424                 data = rb_entry(node, struct work_atoms, node);
1425                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1426         }
1427         if (root == &sched->atom_root) {
1428                 root = &sched->merged_atom_root;
1429                 goto again;
1430         }
1431 }
1432
1433 static int process_sched_wakeup_event(struct perf_tool *tool,
1434                                       struct perf_evsel *evsel,
1435                                       struct perf_sample *sample,
1436                                       struct machine *machine)
1437 {
1438         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1439
1440         if (sched->tp_handler->wakeup_event)
1441                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1442
1443         return 0;
1444 }
1445
1446 union map_priv {
1447         void    *ptr;
1448         bool     color;
1449 };
1450
1451 static bool thread__has_color(struct thread *thread)
1452 {
1453         union map_priv priv = {
1454                 .ptr = thread__priv(thread),
1455         };
1456
1457         return priv.color;
1458 }
1459
1460 static struct thread*
1461 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1462 {
1463         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1464         union map_priv priv = {
1465                 .color = false,
1466         };
1467
1468         if (!sched->map.color_pids || !thread || thread__priv(thread))
1469                 return thread;
1470
1471         if (thread_map__has(sched->map.color_pids, tid))
1472                 priv.color = true;
1473
1474         thread__set_priv(thread, priv.ptr);
1475         return thread;
1476 }
1477
1478 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1479                             struct perf_sample *sample, struct machine *machine)
1480 {
1481         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1482         struct thread *sched_in;
1483         int new_shortname;
1484         u64 timestamp0, timestamp = sample->time;
1485         s64 delta;
1486         int i, this_cpu = sample->cpu;
1487         int cpus_nr;
1488         bool new_cpu = false;
1489         const char *color = PERF_COLOR_NORMAL;
1490         char stimestamp[32];
1491
1492         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1493
1494         if (this_cpu > sched->max_cpu)
1495                 sched->max_cpu = this_cpu;
1496
1497         if (sched->map.comp) {
1498                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1499                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1500                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1501                         new_cpu = true;
1502                 }
1503         } else
1504                 cpus_nr = sched->max_cpu;
1505
1506         timestamp0 = sched->cpu_last_switched[this_cpu];
1507         sched->cpu_last_switched[this_cpu] = timestamp;
1508         if (timestamp0)
1509                 delta = timestamp - timestamp0;
1510         else
1511                 delta = 0;
1512
1513         if (delta < 0) {
1514                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1515                 return -1;
1516         }
1517
1518         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1519         if (sched_in == NULL)
1520                 return -1;
1521
1522         sched->curr_thread[this_cpu] = thread__get(sched_in);
1523
1524         printf("  ");
1525
1526         new_shortname = 0;
1527         if (!sched_in->shortname[0]) {
1528                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1529                         /*
1530                          * Don't allocate a letter-number for swapper:0
1531                          * as a shortname. Instead, we use '.' for it.
1532                          */
1533                         sched_in->shortname[0] = '.';
1534                         sched_in->shortname[1] = ' ';
1535                 } else {
1536                         sched_in->shortname[0] = sched->next_shortname1;
1537                         sched_in->shortname[1] = sched->next_shortname2;
1538
1539                         if (sched->next_shortname1 < 'Z') {
1540                                 sched->next_shortname1++;
1541                         } else {
1542                                 sched->next_shortname1 = 'A';
1543                                 if (sched->next_shortname2 < '9')
1544                                         sched->next_shortname2++;
1545                                 else
1546                                         sched->next_shortname2 = '0';
1547                         }
1548                 }
1549                 new_shortname = 1;
1550         }
1551
1552         for (i = 0; i < cpus_nr; i++) {
1553                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1554                 struct thread *curr_thread = sched->curr_thread[cpu];
1555                 const char *pid_color = color;
1556                 const char *cpu_color = color;
1557
1558                 if (curr_thread && thread__has_color(curr_thread))
1559                         pid_color = COLOR_PIDS;
1560
1561                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1562                         continue;
1563
1564                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1565                         cpu_color = COLOR_CPUS;
1566
1567                 if (cpu != this_cpu)
1568                         color_fprintf(stdout, color, " ");
1569                 else
1570                         color_fprintf(stdout, cpu_color, "*");
1571
1572                 if (sched->curr_thread[cpu])
1573                         color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1574                 else
1575                         color_fprintf(stdout, color, "   ");
1576         }
1577
1578         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1579                 goto out;
1580
1581         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1582         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1583         if (new_shortname || (verbose > 0 && sched_in->tid)) {
1584                 const char *pid_color = color;
1585
1586                 if (thread__has_color(sched_in))
1587                         pid_color = COLOR_PIDS;
1588
1589                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1590                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1591         }
1592
1593         if (sched->map.comp && new_cpu)
1594                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1595
1596 out:
1597         color_fprintf(stdout, color, "\n");
1598
1599         thread__put(sched_in);
1600
1601         return 0;
1602 }
1603
1604 static int process_sched_switch_event(struct perf_tool *tool,
1605                                       struct perf_evsel *evsel,
1606                                       struct perf_sample *sample,
1607                                       struct machine *machine)
1608 {
1609         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1610         int this_cpu = sample->cpu, err = 0;
1611         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1612             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1613
1614         if (sched->curr_pid[this_cpu] != (u32)-1) {
1615                 /*
1616                  * Are we trying to switch away a PID that is
1617                  * not current?
1618                  */
1619                 if (sched->curr_pid[this_cpu] != prev_pid)
1620                         sched->nr_context_switch_bugs++;
1621         }
1622
1623         if (sched->tp_handler->switch_event)
1624                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1625
1626         sched->curr_pid[this_cpu] = next_pid;
1627         return err;
1628 }
1629
1630 static int process_sched_runtime_event(struct perf_tool *tool,
1631                                        struct perf_evsel *evsel,
1632                                        struct perf_sample *sample,
1633                                        struct machine *machine)
1634 {
1635         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1636
1637         if (sched->tp_handler->runtime_event)
1638                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1639
1640         return 0;
1641 }
1642
1643 static int perf_sched__process_fork_event(struct perf_tool *tool,
1644                                           union perf_event *event,
1645                                           struct perf_sample *sample,
1646                                           struct machine *machine)
1647 {
1648         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1649
1650         /* run the fork event through the perf machineruy */
1651         perf_event__process_fork(tool, event, sample, machine);
1652
1653         /* and then run additional processing needed for this command */
1654         if (sched->tp_handler->fork_event)
1655                 return sched->tp_handler->fork_event(sched, event, machine);
1656
1657         return 0;
1658 }
1659
1660 static int process_sched_migrate_task_event(struct perf_tool *tool,
1661                                             struct perf_evsel *evsel,
1662                                             struct perf_sample *sample,
1663                                             struct machine *machine)
1664 {
1665         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1666
1667         if (sched->tp_handler->migrate_task_event)
1668                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1669
1670         return 0;
1671 }
1672
1673 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1674                                   struct perf_evsel *evsel,
1675                                   struct perf_sample *sample,
1676                                   struct machine *machine);
1677
1678 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1679                                                  union perf_event *event __maybe_unused,
1680                                                  struct perf_sample *sample,
1681                                                  struct perf_evsel *evsel,
1682                                                  struct machine *machine)
1683 {
1684         int err = 0;
1685
1686         if (evsel->handler != NULL) {
1687                 tracepoint_handler f = evsel->handler;
1688                 err = f(tool, evsel, sample, machine);
1689         }
1690
1691         return err;
1692 }
1693
1694 static int perf_sched__read_events(struct perf_sched *sched)
1695 {
1696         const struct perf_evsel_str_handler handlers[] = {
1697                 { "sched:sched_switch",       process_sched_switch_event, },
1698                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1699                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1700                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1701                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1702         };
1703         struct perf_session *session;
1704         struct perf_data data = {
1705                 .file      = {
1706                         .path = input_name,
1707                 },
1708                 .mode      = PERF_DATA_MODE_READ,
1709                 .force     = sched->force,
1710         };
1711         int rc = -1;
1712
1713         session = perf_session__new(&data, false, &sched->tool);
1714         if (session == NULL) {
1715                 pr_debug("No Memory for session\n");
1716                 return -1;
1717         }
1718
1719         symbol__init(&session->header.env);
1720
1721         if (perf_session__set_tracepoints_handlers(session, handlers))
1722                 goto out_delete;
1723
1724         if (perf_session__has_traces(session, "record -R")) {
1725                 int err = perf_session__process_events(session);
1726                 if (err) {
1727                         pr_err("Failed to process events, error %d", err);
1728                         goto out_delete;
1729                 }
1730
1731                 sched->nr_events      = session->evlist->stats.nr_events[0];
1732                 sched->nr_lost_events = session->evlist->stats.total_lost;
1733                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1734         }
1735
1736         rc = 0;
1737 out_delete:
1738         perf_session__delete(session);
1739         return rc;
1740 }
1741
1742 /*
1743  * scheduling times are printed as msec.usec
1744  */
1745 static inline void print_sched_time(unsigned long long nsecs, int width)
1746 {
1747         unsigned long msecs;
1748         unsigned long usecs;
1749
1750         msecs  = nsecs / NSEC_PER_MSEC;
1751         nsecs -= msecs * NSEC_PER_MSEC;
1752         usecs  = nsecs / NSEC_PER_USEC;
1753         printf("%*lu.%03lu ", width, msecs, usecs);
1754 }
1755
1756 /*
1757  * returns runtime data for event, allocating memory for it the
1758  * first time it is used.
1759  */
1760 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1761 {
1762         struct evsel_runtime *r = evsel->priv;
1763
1764         if (r == NULL) {
1765                 r = zalloc(sizeof(struct evsel_runtime));
1766                 evsel->priv = r;
1767         }
1768
1769         return r;
1770 }
1771
1772 /*
1773  * save last time event was seen per cpu
1774  */
1775 static void perf_evsel__save_time(struct perf_evsel *evsel,
1776                                   u64 timestamp, u32 cpu)
1777 {
1778         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1779
1780         if (r == NULL)
1781                 return;
1782
1783         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1784                 int i, n = __roundup_pow_of_two(cpu+1);
1785                 void *p = r->last_time;
1786
1787                 p = realloc(r->last_time, n * sizeof(u64));
1788                 if (!p)
1789                         return;
1790
1791                 r->last_time = p;
1792                 for (i = r->ncpu; i < n; ++i)
1793                         r->last_time[i] = (u64) 0;
1794
1795                 r->ncpu = n;
1796         }
1797
1798         r->last_time[cpu] = timestamp;
1799 }
1800
1801 /* returns last time this event was seen on the given cpu */
1802 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1803 {
1804         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1805
1806         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1807                 return 0;
1808
1809         return r->last_time[cpu];
1810 }
1811
1812 static int comm_width = 30;
1813
1814 static char *timehist_get_commstr(struct thread *thread)
1815 {
1816         static char str[32];
1817         const char *comm = thread__comm_str(thread);
1818         pid_t tid = thread->tid;
1819         pid_t pid = thread->pid_;
1820         int n;
1821
1822         if (pid == 0)
1823                 n = scnprintf(str, sizeof(str), "%s", comm);
1824
1825         else if (tid != pid)
1826                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1827
1828         else
1829                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1830
1831         if (n > comm_width)
1832                 comm_width = n;
1833
1834         return str;
1835 }
1836
1837 static void timehist_header(struct perf_sched *sched)
1838 {
1839         u32 ncpus = sched->max_cpu + 1;
1840         u32 i, j;
1841
1842         printf("%15s %6s ", "time", "cpu");
1843
1844         if (sched->show_cpu_visual) {
1845                 printf(" ");
1846                 for (i = 0, j = 0; i < ncpus; ++i) {
1847                         printf("%x", j++);
1848                         if (j > 15)
1849                                 j = 0;
1850                 }
1851                 printf(" ");
1852         }
1853
1854         printf(" %-*s  %9s  %9s  %9s", comm_width,
1855                 "task name", "wait time", "sch delay", "run time");
1856
1857         if (sched->show_state)
1858                 printf("  %s", "state");
1859
1860         printf("\n");
1861
1862         /*
1863          * units row
1864          */
1865         printf("%15s %-6s ", "", "");
1866
1867         if (sched->show_cpu_visual)
1868                 printf(" %*s ", ncpus, "");
1869
1870         printf(" %-*s  %9s  %9s  %9s", comm_width,
1871                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1872
1873         if (sched->show_state)
1874                 printf("  %5s", "");
1875
1876         printf("\n");
1877
1878         /*
1879          * separator
1880          */
1881         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1882
1883         if (sched->show_cpu_visual)
1884                 printf(" %.*s ", ncpus, graph_dotted_line);
1885
1886         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1887                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1888                 graph_dotted_line);
1889
1890         if (sched->show_state)
1891                 printf("  %.5s", graph_dotted_line);
1892
1893         printf("\n");
1894 }
1895
1896 static char task_state_char(struct thread *thread, int state)
1897 {
1898         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1899         unsigned bit = state ? ffs(state) : 0;
1900
1901         /* 'I' for idle */
1902         if (thread->tid == 0)
1903                 return 'I';
1904
1905         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1906 }
1907
1908 static void timehist_print_sample(struct perf_sched *sched,
1909                                   struct perf_evsel *evsel,
1910                                   struct perf_sample *sample,
1911                                   struct addr_location *al,
1912                                   struct thread *thread,
1913                                   u64 t, int state)
1914 {
1915         struct thread_runtime *tr = thread__priv(thread);
1916         const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1917         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1918         u32 max_cpus = sched->max_cpu + 1;
1919         char tstr[64];
1920         char nstr[30];
1921         u64 wait_time;
1922
1923         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1924         printf("%15s [%04d] ", tstr, sample->cpu);
1925
1926         if (sched->show_cpu_visual) {
1927                 u32 i;
1928                 char c;
1929
1930                 printf(" ");
1931                 for (i = 0; i < max_cpus; ++i) {
1932                         /* flag idle times with 'i'; others are sched events */
1933                         if (i == sample->cpu)
1934                                 c = (thread->tid == 0) ? 'i' : 's';
1935                         else
1936                                 c = ' ';
1937                         printf("%c", c);
1938                 }
1939                 printf(" ");
1940         }
1941
1942         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1943
1944         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
1945         print_sched_time(wait_time, 6);
1946
1947         print_sched_time(tr->dt_delay, 6);
1948         print_sched_time(tr->dt_run, 6);
1949
1950         if (sched->show_state)
1951                 printf(" %5c ", task_state_char(thread, state));
1952
1953         if (sched->show_next) {
1954                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
1955                 printf(" %-*s", comm_width, nstr);
1956         }
1957
1958         if (sched->show_wakeups && !sched->show_next)
1959                 printf("  %-*s", comm_width, "");
1960
1961         if (thread->tid == 0)
1962                 goto out;
1963
1964         if (sched->show_callchain)
1965                 printf("  ");
1966
1967         sample__fprintf_sym(sample, al, 0,
1968                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1969                             EVSEL__PRINT_CALLCHAIN_ARROW |
1970                             EVSEL__PRINT_SKIP_IGNORED,
1971                             &callchain_cursor, stdout);
1972
1973 out:
1974         printf("\n");
1975 }
1976
1977 /*
1978  * Explanation of delta-time stats:
1979  *
1980  *            t = time of current schedule out event
1981  *        tprev = time of previous sched out event
1982  *                also time of schedule-in event for current task
1983  *    last_time = time of last sched change event for current task
1984  *                (i.e, time process was last scheduled out)
1985  * ready_to_run = time of wakeup for current task
1986  *
1987  * -----|------------|------------|------------|------
1988  *    last         ready        tprev          t
1989  *    time         to run
1990  *
1991  *      |-------- dt_wait --------|
1992  *                   |- dt_delay -|-- dt_run --|
1993  *
1994  *   dt_run = run time of current task
1995  *  dt_wait = time between last schedule out event for task and tprev
1996  *            represents time spent off the cpu
1997  * dt_delay = time between wakeup and schedule-in of task
1998  */
1999
2000 static void timehist_update_runtime_stats(struct thread_runtime *r,
2001                                          u64 t, u64 tprev)
2002 {
2003         r->dt_delay   = 0;
2004         r->dt_sleep   = 0;
2005         r->dt_iowait  = 0;
2006         r->dt_preempt = 0;
2007         r->dt_run     = 0;
2008
2009         if (tprev) {
2010                 r->dt_run = t - tprev;
2011                 if (r->ready_to_run) {
2012                         if (r->ready_to_run > tprev)
2013                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2014                         else
2015                                 r->dt_delay = tprev - r->ready_to_run;
2016                 }
2017
2018                 if (r->last_time > tprev)
2019                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2020                 else if (r->last_time) {
2021                         u64 dt_wait = tprev - r->last_time;
2022
2023                         if (r->last_state == TASK_RUNNING)
2024                                 r->dt_preempt = dt_wait;
2025                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2026                                 r->dt_iowait = dt_wait;
2027                         else
2028                                 r->dt_sleep = dt_wait;
2029                 }
2030         }
2031
2032         update_stats(&r->run_stats, r->dt_run);
2033
2034         r->total_run_time     += r->dt_run;
2035         r->total_delay_time   += r->dt_delay;
2036         r->total_sleep_time   += r->dt_sleep;
2037         r->total_iowait_time  += r->dt_iowait;
2038         r->total_preempt_time += r->dt_preempt;
2039 }
2040
2041 static bool is_idle_sample(struct perf_sample *sample,
2042                            struct perf_evsel *evsel)
2043 {
2044         /* pid 0 == swapper == idle task */
2045         if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2046                 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2047
2048         return sample->pid == 0;
2049 }
2050
2051 static void save_task_callchain(struct perf_sched *sched,
2052                                 struct perf_sample *sample,
2053                                 struct perf_evsel *evsel,
2054                                 struct machine *machine)
2055 {
2056         struct callchain_cursor *cursor = &callchain_cursor;
2057         struct thread *thread;
2058
2059         /* want main thread for process - has maps */
2060         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2061         if (thread == NULL) {
2062                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2063                 return;
2064         }
2065
2066         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2067                 return;
2068
2069         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2070                                       NULL, NULL, sched->max_stack + 2) != 0) {
2071                 if (verbose > 0)
2072                         pr_err("Failed to resolve callchain. Skipping\n");
2073
2074                 return;
2075         }
2076
2077         callchain_cursor_commit(cursor);
2078
2079         while (true) {
2080                 struct callchain_cursor_node *node;
2081                 struct symbol *sym;
2082
2083                 node = callchain_cursor_current(cursor);
2084                 if (node == NULL)
2085                         break;
2086
2087                 sym = node->sym;
2088                 if (sym) {
2089                         if (!strcmp(sym->name, "schedule") ||
2090                             !strcmp(sym->name, "__schedule") ||
2091                             !strcmp(sym->name, "preempt_schedule"))
2092                                 sym->ignore = 1;
2093                 }
2094
2095                 callchain_cursor_advance(cursor);
2096         }
2097 }
2098
2099 static int init_idle_thread(struct thread *thread)
2100 {
2101         struct idle_thread_runtime *itr;
2102
2103         thread__set_comm(thread, idle_comm, 0);
2104
2105         itr = zalloc(sizeof(*itr));
2106         if (itr == NULL)
2107                 return -ENOMEM;
2108
2109         init_stats(&itr->tr.run_stats);
2110         callchain_init(&itr->callchain);
2111         callchain_cursor_reset(&itr->cursor);
2112         thread__set_priv(thread, itr);
2113
2114         return 0;
2115 }
2116
2117 /*
2118  * Track idle stats per cpu by maintaining a local thread
2119  * struct for the idle task on each cpu.
2120  */
2121 static int init_idle_threads(int ncpu)
2122 {
2123         int i, ret;
2124
2125         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2126         if (!idle_threads)
2127                 return -ENOMEM;
2128
2129         idle_max_cpu = ncpu;
2130
2131         /* allocate the actual thread struct if needed */
2132         for (i = 0; i < ncpu; ++i) {
2133                 idle_threads[i] = thread__new(0, 0);
2134                 if (idle_threads[i] == NULL)
2135                         return -ENOMEM;
2136
2137                 ret = init_idle_thread(idle_threads[i]);
2138                 if (ret < 0)
2139                         return ret;
2140         }
2141
2142         return 0;
2143 }
2144
2145 static void free_idle_threads(void)
2146 {
2147         int i;
2148
2149         if (idle_threads == NULL)
2150                 return;
2151
2152         for (i = 0; i < idle_max_cpu; ++i) {
2153                 if ((idle_threads[i]))
2154                         thread__delete(idle_threads[i]);
2155         }
2156
2157         free(idle_threads);
2158 }
2159
2160 static struct thread *get_idle_thread(int cpu)
2161 {
2162         /*
2163          * expand/allocate array of pointers to local thread
2164          * structs if needed
2165          */
2166         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2167                 int i, j = __roundup_pow_of_two(cpu+1);
2168                 void *p;
2169
2170                 p = realloc(idle_threads, j * sizeof(struct thread *));
2171                 if (!p)
2172                         return NULL;
2173
2174                 idle_threads = (struct thread **) p;
2175                 for (i = idle_max_cpu; i < j; ++i)
2176                         idle_threads[i] = NULL;
2177
2178                 idle_max_cpu = j;
2179         }
2180
2181         /* allocate a new thread struct if needed */
2182         if (idle_threads[cpu] == NULL) {
2183                 idle_threads[cpu] = thread__new(0, 0);
2184
2185                 if (idle_threads[cpu]) {
2186                         if (init_idle_thread(idle_threads[cpu]) < 0)
2187                                 return NULL;
2188                 }
2189         }
2190
2191         return idle_threads[cpu];
2192 }
2193
2194 static void save_idle_callchain(struct idle_thread_runtime *itr,
2195                                 struct perf_sample *sample)
2196 {
2197         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2198                 return;
2199
2200         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2201 }
2202
2203 /*
2204  * handle runtime stats saved per thread
2205  */
2206 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2207 {
2208         struct thread_runtime *r;
2209
2210         r = zalloc(sizeof(struct thread_runtime));
2211         if (!r)
2212                 return NULL;
2213
2214         init_stats(&r->run_stats);
2215         thread__set_priv(thread, r);
2216
2217         return r;
2218 }
2219
2220 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2221 {
2222         struct thread_runtime *tr;
2223
2224         tr = thread__priv(thread);
2225         if (tr == NULL) {
2226                 tr = thread__init_runtime(thread);
2227                 if (tr == NULL)
2228                         pr_debug("Failed to malloc memory for runtime data.\n");
2229         }
2230
2231         return tr;
2232 }
2233
2234 static struct thread *timehist_get_thread(struct perf_sched *sched,
2235                                           struct perf_sample *sample,
2236                                           struct machine *machine,
2237                                           struct perf_evsel *evsel)
2238 {
2239         struct thread *thread;
2240
2241         if (is_idle_sample(sample, evsel)) {
2242                 thread = get_idle_thread(sample->cpu);
2243                 if (thread == NULL)
2244                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2245
2246         } else {
2247                 /* there were samples with tid 0 but non-zero pid */
2248                 thread = machine__findnew_thread(machine, sample->pid,
2249                                                  sample->tid ?: sample->pid);
2250                 if (thread == NULL) {
2251                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2252                                  sample->tid);
2253                 }
2254
2255                 save_task_callchain(sched, sample, evsel, machine);
2256                 if (sched->idle_hist) {
2257                         struct thread *idle;
2258                         struct idle_thread_runtime *itr;
2259
2260                         idle = get_idle_thread(sample->cpu);
2261                         if (idle == NULL) {
2262                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2263                                 return NULL;
2264                         }
2265
2266                         itr = thread__priv(idle);
2267                         if (itr == NULL)
2268                                 return NULL;
2269
2270                         itr->last_thread = thread;
2271
2272                         /* copy task callchain when entering to idle */
2273                         if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2274                                 save_idle_callchain(itr, sample);
2275                 }
2276         }
2277
2278         return thread;
2279 }
2280
2281 static bool timehist_skip_sample(struct perf_sched *sched,
2282                                  struct thread *thread,
2283                                  struct perf_evsel *evsel,
2284                                  struct perf_sample *sample)
2285 {
2286         bool rc = false;
2287
2288         if (thread__is_filtered(thread)) {
2289                 rc = true;
2290                 sched->skipped_samples++;
2291         }
2292
2293         if (sched->idle_hist) {
2294                 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2295                         rc = true;
2296                 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2297                          perf_evsel__intval(evsel, sample, "next_pid") != 0)
2298                         rc = true;
2299         }
2300
2301         return rc;
2302 }
2303
2304 static void timehist_print_wakeup_event(struct perf_sched *sched,
2305                                         struct perf_evsel *evsel,
2306                                         struct perf_sample *sample,
2307                                         struct machine *machine,
2308                                         struct thread *awakened)
2309 {
2310         struct thread *thread;
2311         char tstr[64];
2312
2313         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2314         if (thread == NULL)
2315                 return;
2316
2317         /* show wakeup unless both awakee and awaker are filtered */
2318         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2319             timehist_skip_sample(sched, awakened, evsel, sample)) {
2320                 return;
2321         }
2322
2323         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2324         printf("%15s [%04d] ", tstr, sample->cpu);
2325         if (sched->show_cpu_visual)
2326                 printf(" %*s ", sched->max_cpu + 1, "");
2327
2328         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2329
2330         /* dt spacer */
2331         printf("  %9s  %9s  %9s ", "", "", "");
2332
2333         printf("awakened: %s", timehist_get_commstr(awakened));
2334
2335         printf("\n");
2336 }
2337
2338 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2339                                        union perf_event *event __maybe_unused,
2340                                        struct perf_evsel *evsel,
2341                                        struct perf_sample *sample,
2342                                        struct machine *machine)
2343 {
2344         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2345         struct thread *thread;
2346         struct thread_runtime *tr = NULL;
2347         /* want pid of awakened task not pid in sample */
2348         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2349
2350         thread = machine__findnew_thread(machine, 0, pid);
2351         if (thread == NULL)
2352                 return -1;
2353
2354         tr = thread__get_runtime(thread);
2355         if (tr == NULL)
2356                 return -1;
2357
2358         if (tr->ready_to_run == 0)
2359                 tr->ready_to_run = sample->time;
2360
2361         /* show wakeups if requested */
2362         if (sched->show_wakeups &&
2363             !perf_time__skip_sample(&sched->ptime, sample->time))
2364                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2365
2366         return 0;
2367 }
2368
2369 static void timehist_print_migration_event(struct perf_sched *sched,
2370                                         struct perf_evsel *evsel,
2371                                         struct perf_sample *sample,
2372                                         struct machine *machine,
2373                                         struct thread *migrated)
2374 {
2375         struct thread *thread;
2376         char tstr[64];
2377         u32 max_cpus = sched->max_cpu + 1;
2378         u32 ocpu, dcpu;
2379
2380         if (sched->summary_only)
2381                 return;
2382
2383         max_cpus = sched->max_cpu + 1;
2384         ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2385         dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2386
2387         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2388         if (thread == NULL)
2389                 return;
2390
2391         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2392             timehist_skip_sample(sched, migrated, evsel, sample)) {
2393                 return;
2394         }
2395
2396         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2397         printf("%15s [%04d] ", tstr, sample->cpu);
2398
2399         if (sched->show_cpu_visual) {
2400                 u32 i;
2401                 char c;
2402
2403                 printf("  ");
2404                 for (i = 0; i < max_cpus; ++i) {
2405                         c = (i == sample->cpu) ? 'm' : ' ';
2406                         printf("%c", c);
2407                 }
2408                 printf("  ");
2409         }
2410
2411         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2412
2413         /* dt spacer */
2414         printf("  %9s  %9s  %9s ", "", "", "");
2415
2416         printf("migrated: %s", timehist_get_commstr(migrated));
2417         printf(" cpu %d => %d", ocpu, dcpu);
2418
2419         printf("\n");
2420 }
2421
2422 static int timehist_migrate_task_event(struct perf_tool *tool,
2423                                        union perf_event *event __maybe_unused,
2424                                        struct perf_evsel *evsel,
2425                                        struct perf_sample *sample,
2426                                        struct machine *machine)
2427 {
2428         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2429         struct thread *thread;
2430         struct thread_runtime *tr = NULL;
2431         /* want pid of migrated task not pid in sample */
2432         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2433
2434         thread = machine__findnew_thread(machine, 0, pid);
2435         if (thread == NULL)
2436                 return -1;
2437
2438         tr = thread__get_runtime(thread);
2439         if (tr == NULL)
2440                 return -1;
2441
2442         tr->migrations++;
2443
2444         /* show migrations if requested */
2445         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2446
2447         return 0;
2448 }
2449
2450 static int timehist_sched_change_event(struct perf_tool *tool,
2451                                        union perf_event *event,
2452                                        struct perf_evsel *evsel,
2453                                        struct perf_sample *sample,
2454                                        struct machine *machine)
2455 {
2456         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2457         struct perf_time_interval *ptime = &sched->ptime;
2458         struct addr_location al;
2459         struct thread *thread;
2460         struct thread_runtime *tr = NULL;
2461         u64 tprev, t = sample->time;
2462         int rc = 0;
2463         int state = perf_evsel__intval(evsel, sample, "prev_state");
2464
2465
2466         if (machine__resolve(machine, &al, sample) < 0) {
2467                 pr_err("problem processing %d event. skipping it\n",
2468                        event->header.type);
2469                 rc = -1;
2470                 goto out;
2471         }
2472
2473         thread = timehist_get_thread(sched, sample, machine, evsel);
2474         if (thread == NULL) {
2475                 rc = -1;
2476                 goto out;
2477         }
2478
2479         if (timehist_skip_sample(sched, thread, evsel, sample))
2480                 goto out;
2481
2482         tr = thread__get_runtime(thread);
2483         if (tr == NULL) {
2484                 rc = -1;
2485                 goto out;
2486         }
2487
2488         tprev = perf_evsel__get_time(evsel, sample->cpu);
2489
2490         /*
2491          * If start time given:
2492          * - sample time is under window user cares about - skip sample
2493          * - tprev is under window user cares about  - reset to start of window
2494          */
2495         if (ptime->start && ptime->start > t)
2496                 goto out;
2497
2498         if (tprev && ptime->start > tprev)
2499                 tprev = ptime->start;
2500
2501         /*
2502          * If end time given:
2503          * - previous sched event is out of window - we are done
2504          * - sample time is beyond window user cares about - reset it
2505          *   to close out stats for time window interest
2506          */
2507         if (ptime->end) {
2508                 if (tprev > ptime->end)
2509                         goto out;
2510
2511                 if (t > ptime->end)
2512                         t = ptime->end;
2513         }
2514
2515         if (!sched->idle_hist || thread->tid == 0) {
2516                 timehist_update_runtime_stats(tr, t, tprev);
2517
2518                 if (sched->idle_hist) {
2519                         struct idle_thread_runtime *itr = (void *)tr;
2520                         struct thread_runtime *last_tr;
2521
2522                         BUG_ON(thread->tid != 0);
2523
2524                         if (itr->last_thread == NULL)
2525                                 goto out;
2526
2527                         /* add current idle time as last thread's runtime */
2528                         last_tr = thread__get_runtime(itr->last_thread);
2529                         if (last_tr == NULL)
2530                                 goto out;
2531
2532                         timehist_update_runtime_stats(last_tr, t, tprev);
2533                         /*
2534                          * remove delta time of last thread as it's not updated
2535                          * and otherwise it will show an invalid value next
2536                          * time.  we only care total run time and run stat.
2537                          */
2538                         last_tr->dt_run = 0;
2539                         last_tr->dt_delay = 0;
2540                         last_tr->dt_sleep = 0;
2541                         last_tr->dt_iowait = 0;
2542                         last_tr->dt_preempt = 0;
2543
2544                         if (itr->cursor.nr)
2545                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2546
2547                         itr->last_thread = NULL;
2548                 }
2549         }
2550
2551         if (!sched->summary_only)
2552                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2553
2554 out:
2555         if (sched->hist_time.start == 0 && t >= ptime->start)
2556                 sched->hist_time.start = t;
2557         if (ptime->end == 0 || t <= ptime->end)
2558                 sched->hist_time.end = t;
2559
2560         if (tr) {
2561                 /* time of this sched_switch event becomes last time task seen */
2562                 tr->last_time = sample->time;
2563
2564                 /* last state is used to determine where to account wait time */
2565                 tr->last_state = state;
2566
2567                 /* sched out event for task so reset ready to run time */
2568                 tr->ready_to_run = 0;
2569         }
2570
2571         perf_evsel__save_time(evsel, sample->time, sample->cpu);
2572
2573         return rc;
2574 }
2575
2576 static int timehist_sched_switch_event(struct perf_tool *tool,
2577                              union perf_event *event,
2578                              struct perf_evsel *evsel,
2579                              struct perf_sample *sample,
2580                              struct machine *machine __maybe_unused)
2581 {
2582         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2583 }
2584
2585 static int process_lost(struct perf_tool *tool __maybe_unused,
2586                         union perf_event *event,
2587                         struct perf_sample *sample,
2588                         struct machine *machine __maybe_unused)
2589 {
2590         char tstr[64];
2591
2592         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2593         printf("%15s ", tstr);
2594         printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2595
2596         return 0;
2597 }
2598
2599
2600 static void print_thread_runtime(struct thread *t,
2601                                  struct thread_runtime *r)
2602 {
2603         double mean = avg_stats(&r->run_stats);
2604         float stddev;
2605
2606         printf("%*s   %5d  %9" PRIu64 " ",
2607                comm_width, timehist_get_commstr(t), t->ppid,
2608                (u64) r->run_stats.n);
2609
2610         print_sched_time(r->total_run_time, 8);
2611         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2612         print_sched_time(r->run_stats.min, 6);
2613         printf(" ");
2614         print_sched_time((u64) mean, 6);
2615         printf(" ");
2616         print_sched_time(r->run_stats.max, 6);
2617         printf("  ");
2618         printf("%5.2f", stddev);
2619         printf("   %5" PRIu64, r->migrations);
2620         printf("\n");
2621 }
2622
2623 static void print_thread_waittime(struct thread *t,
2624                                   struct thread_runtime *r)
2625 {
2626         printf("%*s   %5d  %9" PRIu64 " ",
2627                comm_width, timehist_get_commstr(t), t->ppid,
2628                (u64) r->run_stats.n);
2629
2630         print_sched_time(r->total_run_time, 8);
2631         print_sched_time(r->total_sleep_time, 6);
2632         printf(" ");
2633         print_sched_time(r->total_iowait_time, 6);
2634         printf(" ");
2635         print_sched_time(r->total_preempt_time, 6);
2636         printf(" ");
2637         print_sched_time(r->total_delay_time, 6);
2638         printf("\n");
2639 }
2640
2641 struct total_run_stats {
2642         struct perf_sched *sched;
2643         u64  sched_count;
2644         u64  task_count;
2645         u64  total_run_time;
2646 };
2647
2648 static int __show_thread_runtime(struct thread *t, void *priv)
2649 {
2650         struct total_run_stats *stats = priv;
2651         struct thread_runtime *r;
2652
2653         if (thread__is_filtered(t))
2654                 return 0;
2655
2656         r = thread__priv(t);
2657         if (r && r->run_stats.n) {
2658                 stats->task_count++;
2659                 stats->sched_count += r->run_stats.n;
2660                 stats->total_run_time += r->total_run_time;
2661
2662                 if (stats->sched->show_state)
2663                         print_thread_waittime(t, r);
2664                 else
2665                         print_thread_runtime(t, r);
2666         }
2667
2668         return 0;
2669 }
2670
2671 static int show_thread_runtime(struct thread *t, void *priv)
2672 {
2673         if (t->dead)
2674                 return 0;
2675
2676         return __show_thread_runtime(t, priv);
2677 }
2678
2679 static int show_deadthread_runtime(struct thread *t, void *priv)
2680 {
2681         if (!t->dead)
2682                 return 0;
2683
2684         return __show_thread_runtime(t, priv);
2685 }
2686
2687 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2688 {
2689         const char *sep = " <- ";
2690         struct callchain_list *chain;
2691         size_t ret = 0;
2692         char bf[1024];
2693         bool first;
2694
2695         if (node == NULL)
2696                 return 0;
2697
2698         ret = callchain__fprintf_folded(fp, node->parent);
2699         first = (ret == 0);
2700
2701         list_for_each_entry(chain, &node->val, list) {
2702                 if (chain->ip >= PERF_CONTEXT_MAX)
2703                         continue;
2704                 if (chain->ms.sym && chain->ms.sym->ignore)
2705                         continue;
2706                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2707                                callchain_list__sym_name(chain, bf, sizeof(bf),
2708                                                         false));
2709                 first = false;
2710         }
2711
2712         return ret;
2713 }
2714
2715 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2716 {
2717         size_t ret = 0;
2718         FILE *fp = stdout;
2719         struct callchain_node *chain;
2720         struct rb_node *rb_node = rb_first(root);
2721
2722         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2723         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2724                graph_dotted_line);
2725
2726         while (rb_node) {
2727                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2728                 rb_node = rb_next(rb_node);
2729
2730                 ret += fprintf(fp, "  ");
2731                 print_sched_time(chain->hit, 12);
2732                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2733                 ret += fprintf(fp, " %8d  ", chain->count);
2734                 ret += callchain__fprintf_folded(fp, chain);
2735                 ret += fprintf(fp, "\n");
2736         }
2737
2738         return ret;
2739 }
2740
2741 static void timehist_print_summary(struct perf_sched *sched,
2742                                    struct perf_session *session)
2743 {
2744         struct machine *m = &session->machines.host;
2745         struct total_run_stats totals;
2746         u64 task_count;
2747         struct thread *t;
2748         struct thread_runtime *r;
2749         int i;
2750         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2751
2752         memset(&totals, 0, sizeof(totals));
2753         totals.sched = sched;
2754
2755         if (sched->idle_hist) {
2756                 printf("\nIdle-time summary\n");
2757                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2758                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2759         } else if (sched->show_state) {
2760                 printf("\nWait-time summary\n");
2761                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2762                 printf("   run-time      sleep      iowait     preempt       delay\n");
2763         } else {
2764                 printf("\nRuntime summary\n");
2765                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2766                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2767         }
2768         printf("%*s            (count)  ", comm_width, "");
2769         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2770                sched->show_state ? "(msec)" : "%");
2771         printf("%.117s\n", graph_dotted_line);
2772
2773         machine__for_each_thread(m, show_thread_runtime, &totals);
2774         task_count = totals.task_count;
2775         if (!task_count)
2776                 printf("<no still running tasks>\n");
2777
2778         printf("\nTerminated tasks:\n");
2779         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2780         if (task_count == totals.task_count)
2781                 printf("<no terminated tasks>\n");
2782
2783         /* CPU idle stats not tracked when samples were skipped */
2784         if (sched->skipped_samples && !sched->idle_hist)
2785                 return;
2786
2787         printf("\nIdle stats:\n");
2788         for (i = 0; i < idle_max_cpu; ++i) {
2789                 t = idle_threads[i];
2790                 if (!t)
2791                         continue;
2792
2793                 r = thread__priv(t);
2794                 if (r && r->run_stats.n) {
2795                         totals.sched_count += r->run_stats.n;
2796                         printf("    CPU %2d idle for ", i);
2797                         print_sched_time(r->total_run_time, 6);
2798                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2799                 } else
2800                         printf("    CPU %2d idle entire time window\n", i);
2801         }
2802
2803         if (sched->idle_hist && symbol_conf.use_callchain) {
2804                 callchain_param.mode  = CHAIN_FOLDED;
2805                 callchain_param.value = CCVAL_PERIOD;
2806
2807                 callchain_register_param(&callchain_param);
2808
2809                 printf("\nIdle stats by callchain:\n");
2810                 for (i = 0; i < idle_max_cpu; ++i) {
2811                         struct idle_thread_runtime *itr;
2812
2813                         t = idle_threads[i];
2814                         if (!t)
2815                                 continue;
2816
2817                         itr = thread__priv(t);
2818                         if (itr == NULL)
2819                                 continue;
2820
2821                         callchain_param.sort(&itr->sorted_root, &itr->callchain,
2822                                              0, &callchain_param);
2823
2824                         printf("  CPU %2d:", i);
2825                         print_sched_time(itr->tr.total_run_time, 6);
2826                         printf(" msec\n");
2827                         timehist_print_idlehist_callchain(&itr->sorted_root);
2828                         printf("\n");
2829                 }
2830         }
2831
2832         printf("\n"
2833                "    Total number of unique tasks: %" PRIu64 "\n"
2834                "Total number of context switches: %" PRIu64 "\n",
2835                totals.task_count, totals.sched_count);
2836
2837         printf("           Total run time (msec): ");
2838         print_sched_time(totals.total_run_time, 2);
2839         printf("\n");
2840
2841         printf("    Total scheduling time (msec): ");
2842         print_sched_time(hist_time, 2);
2843         printf(" (x %d)\n", sched->max_cpu);
2844 }
2845
2846 typedef int (*sched_handler)(struct perf_tool *tool,
2847                           union perf_event *event,
2848                           struct perf_evsel *evsel,
2849                           struct perf_sample *sample,
2850                           struct machine *machine);
2851
2852 static int perf_timehist__process_sample(struct perf_tool *tool,
2853                                          union perf_event *event,
2854                                          struct perf_sample *sample,
2855                                          struct perf_evsel *evsel,
2856                                          struct machine *machine)
2857 {
2858         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2859         int err = 0;
2860         int this_cpu = sample->cpu;
2861
2862         if (this_cpu > sched->max_cpu)
2863                 sched->max_cpu = this_cpu;
2864
2865         if (evsel->handler != NULL) {
2866                 sched_handler f = evsel->handler;
2867
2868                 err = f(tool, event, evsel, sample, machine);
2869         }
2870
2871         return err;
2872 }
2873
2874 static int timehist_check_attr(struct perf_sched *sched,
2875                                struct perf_evlist *evlist)
2876 {
2877         struct perf_evsel *evsel;
2878         struct evsel_runtime *er;
2879
2880         list_for_each_entry(evsel, &evlist->entries, node) {
2881                 er = perf_evsel__get_runtime(evsel);
2882                 if (er == NULL) {
2883                         pr_err("Failed to allocate memory for evsel runtime data\n");
2884                         return -1;
2885                 }
2886
2887                 if (sched->show_callchain &&
2888                     !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2889                         pr_info("Samples do not have callchains.\n");
2890                         sched->show_callchain = 0;
2891                         symbol_conf.use_callchain = 0;
2892                 }
2893         }
2894
2895         return 0;
2896 }
2897
2898 static int perf_sched__timehist(struct perf_sched *sched)
2899 {
2900         const struct perf_evsel_str_handler handlers[] = {
2901                 { "sched:sched_switch",       timehist_sched_switch_event, },
2902                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2903                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2904         };
2905         const struct perf_evsel_str_handler migrate_handlers[] = {
2906                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2907         };
2908         struct perf_data data = {
2909                 .file      = {
2910                         .path = input_name,
2911                 },
2912                 .mode      = PERF_DATA_MODE_READ,
2913                 .force     = sched->force,
2914         };
2915
2916         struct perf_session *session;
2917         struct perf_evlist *evlist;
2918         int err = -1;
2919
2920         /*
2921          * event handlers for timehist option
2922          */
2923         sched->tool.sample       = perf_timehist__process_sample;
2924         sched->tool.mmap         = perf_event__process_mmap;
2925         sched->tool.comm         = perf_event__process_comm;
2926         sched->tool.exit         = perf_event__process_exit;
2927         sched->tool.fork         = perf_event__process_fork;
2928         sched->tool.lost         = process_lost;
2929         sched->tool.attr         = perf_event__process_attr;
2930         sched->tool.tracing_data = perf_event__process_tracing_data;
2931         sched->tool.build_id     = perf_event__process_build_id;
2932
2933         sched->tool.ordered_events = true;
2934         sched->tool.ordering_requires_timestamps = true;
2935
2936         symbol_conf.use_callchain = sched->show_callchain;
2937
2938         session = perf_session__new(&data, false, &sched->tool);
2939         if (session == NULL)
2940                 return -ENOMEM;
2941
2942         evlist = session->evlist;
2943
2944         symbol__init(&session->header.env);
2945
2946         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2947                 pr_err("Invalid time string\n");
2948                 return -EINVAL;
2949         }
2950
2951         if (timehist_check_attr(sched, evlist) != 0)
2952                 goto out;
2953
2954         setup_pager();
2955
2956         /* setup per-evsel handlers */
2957         if (perf_session__set_tracepoints_handlers(session, handlers))
2958                 goto out;
2959
2960         /* sched_switch event at a minimum needs to exist */
2961         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2962                                                   "sched:sched_switch")) {
2963                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2964                 goto out;
2965         }
2966
2967         if (sched->show_migrations &&
2968             perf_session__set_tracepoints_handlers(session, migrate_handlers))
2969                 goto out;
2970
2971         /* pre-allocate struct for per-CPU idle stats */
2972         sched->max_cpu = session->header.env.nr_cpus_online;
2973         if (sched->max_cpu == 0)
2974                 sched->max_cpu = 4;
2975         if (init_idle_threads(sched->max_cpu))
2976                 goto out;
2977
2978         /* summary_only implies summary option, but don't overwrite summary if set */
2979         if (sched->summary_only)
2980                 sched->summary = sched->summary_only;
2981
2982         if (!sched->summary_only)
2983                 timehist_header(sched);
2984
2985         err = perf_session__process_events(session);
2986         if (err) {
2987                 pr_err("Failed to process events, error %d", err);
2988                 goto out;
2989         }
2990
2991         sched->nr_events      = evlist->stats.nr_events[0];
2992         sched->nr_lost_events = evlist->stats.total_lost;
2993         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2994
2995         if (sched->summary)
2996                 timehist_print_summary(sched, session);
2997
2998 out:
2999         free_idle_threads();
3000         perf_session__delete(session);
3001
3002         return err;
3003 }
3004
3005
3006 static void print_bad_events(struct perf_sched *sched)
3007 {
3008         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3009                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3010                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3011                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3012         }
3013         if (sched->nr_lost_events && sched->nr_events) {
3014                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3015                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3016                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3017         }
3018         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3019                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3020                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3021                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3022                 if (sched->nr_lost_events)
3023                         printf(" (due to lost events?)");
3024                 printf("\n");
3025         }
3026 }
3027
3028 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3029 {
3030         struct rb_node **new = &(root->rb_node), *parent = NULL;
3031         struct work_atoms *this;
3032         const char *comm = thread__comm_str(data->thread), *this_comm;
3033
3034         while (*new) {
3035                 int cmp;
3036
3037                 this = container_of(*new, struct work_atoms, node);
3038                 parent = *new;
3039
3040                 this_comm = thread__comm_str(this->thread);
3041                 cmp = strcmp(comm, this_comm);
3042                 if (cmp > 0) {
3043                         new = &((*new)->rb_left);
3044                 } else if (cmp < 0) {
3045                         new = &((*new)->rb_right);
3046                 } else {
3047                         this->num_merged++;
3048                         this->total_runtime += data->total_runtime;
3049                         this->nb_atoms += data->nb_atoms;
3050                         this->total_lat += data->total_lat;
3051                         list_splice(&data->work_list, &this->work_list);
3052                         if (this->max_lat < data->max_lat) {
3053                                 this->max_lat = data->max_lat;
3054                                 this->max_lat_at = data->max_lat_at;
3055                         }
3056                         zfree(&data);
3057                         return;
3058                 }
3059         }
3060
3061         data->num_merged++;
3062         rb_link_node(&data->node, parent, new);
3063         rb_insert_color(&data->node, root);
3064 }
3065
3066 static void perf_sched__merge_lat(struct perf_sched *sched)
3067 {
3068         struct work_atoms *data;
3069         struct rb_node *node;
3070
3071         if (sched->skip_merge)
3072                 return;
3073
3074         while ((node = rb_first(&sched->atom_root))) {
3075                 rb_erase(node, &sched->atom_root);
3076                 data = rb_entry(node, struct work_atoms, node);
3077                 __merge_work_atoms(&sched->merged_atom_root, data);
3078         }
3079 }
3080
3081 static int perf_sched__lat(struct perf_sched *sched)
3082 {
3083         struct rb_node *next;
3084
3085         setup_pager();
3086
3087         if (perf_sched__read_events(sched))
3088                 return -1;
3089
3090         perf_sched__merge_lat(sched);
3091         perf_sched__sort_lat(sched);
3092
3093         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3094         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3095         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3096
3097         next = rb_first(&sched->sorted_atom_root);
3098
3099         while (next) {
3100                 struct work_atoms *work_list;
3101
3102                 work_list = rb_entry(next, struct work_atoms, node);
3103                 output_lat_thread(sched, work_list);
3104                 next = rb_next(next);
3105                 thread__zput(work_list->thread);
3106         }
3107
3108         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3109         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3110                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3111
3112         printf(" ---------------------------------------------------\n");
3113
3114         print_bad_events(sched);
3115         printf("\n");
3116
3117         return 0;
3118 }
3119
3120 static int setup_map_cpus(struct perf_sched *sched)
3121 {
3122         struct cpu_map *map;
3123
3124         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3125
3126         if (sched->map.comp) {
3127                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3128                 if (!sched->map.comp_cpus)
3129                         return -1;
3130         }
3131
3132         if (!sched->map.cpus_str)
3133                 return 0;
3134
3135         map = cpu_map__new(sched->map.cpus_str);
3136         if (!map) {
3137                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3138                 return -1;
3139         }
3140
3141         sched->map.cpus = map;
3142         return 0;
3143 }
3144
3145 static int setup_color_pids(struct perf_sched *sched)
3146 {
3147         struct thread_map *map;
3148
3149         if (!sched->map.color_pids_str)
3150                 return 0;
3151
3152         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3153         if (!map) {
3154                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3155                 return -1;
3156         }
3157
3158         sched->map.color_pids = map;
3159         return 0;
3160 }
3161
3162 static int setup_color_cpus(struct perf_sched *sched)
3163 {
3164         struct cpu_map *map;
3165
3166         if (!sched->map.color_cpus_str)
3167                 return 0;
3168
3169         map = cpu_map__new(sched->map.color_cpus_str);
3170         if (!map) {
3171                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3172                 return -1;
3173         }
3174
3175         sched->map.color_cpus = map;
3176         return 0;
3177 }
3178
3179 static int perf_sched__map(struct perf_sched *sched)
3180 {
3181         if (setup_map_cpus(sched))
3182                 return -1;
3183
3184         if (setup_color_pids(sched))
3185                 return -1;
3186
3187         if (setup_color_cpus(sched))
3188                 return -1;
3189
3190         setup_pager();
3191         if (perf_sched__read_events(sched))
3192                 return -1;
3193         print_bad_events(sched);
3194         return 0;
3195 }
3196
3197 static int perf_sched__replay(struct perf_sched *sched)
3198 {
3199         unsigned long i;
3200
3201         calibrate_run_measurement_overhead(sched);
3202         calibrate_sleep_measurement_overhead(sched);
3203
3204         test_calibrations(sched);
3205
3206         if (perf_sched__read_events(sched))
3207                 return -1;
3208
3209         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3210         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3211         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3212
3213         if (sched->targetless_wakeups)
3214                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3215         if (sched->multitarget_wakeups)
3216                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3217         if (sched->nr_run_events_optimized)
3218                 printf("run atoms optimized: %ld\n",
3219                         sched->nr_run_events_optimized);
3220
3221         print_task_traces(sched);
3222         add_cross_task_wakeups(sched);
3223
3224         create_tasks(sched);
3225         printf("------------------------------------------------------------\n");
3226         for (i = 0; i < sched->replay_repeat; i++)
3227                 run_one_test(sched);
3228
3229         return 0;
3230 }
3231
3232 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3233                           const char * const usage_msg[])
3234 {
3235         char *tmp, *tok, *str = strdup(sched->sort_order);
3236
3237         for (tok = strtok_r(str, ", ", &tmp);
3238                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3239                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3240                         usage_with_options_msg(usage_msg, options,
3241                                         "Unknown --sort key: `%s'", tok);
3242                 }
3243         }
3244
3245         free(str);
3246
3247         sort_dimension__add("pid", &sched->cmp_pid);
3248 }
3249
3250 static int __cmd_record(int argc, const char **argv)
3251 {
3252         unsigned int rec_argc, i, j;
3253         const char **rec_argv;
3254         const char * const record_args[] = {
3255                 "record",
3256                 "-a",
3257                 "-R",
3258                 "-m", "1024",
3259                 "-c", "1",
3260                 "-e", "sched:sched_switch",
3261                 "-e", "sched:sched_stat_wait",
3262                 "-e", "sched:sched_stat_sleep",
3263                 "-e", "sched:sched_stat_iowait",
3264                 "-e", "sched:sched_stat_runtime",
3265                 "-e", "sched:sched_process_fork",
3266                 "-e", "sched:sched_wakeup",
3267                 "-e", "sched:sched_wakeup_new",
3268                 "-e", "sched:sched_migrate_task",
3269         };
3270
3271         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3272         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3273
3274         if (rec_argv == NULL)
3275                 return -ENOMEM;
3276
3277         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3278                 rec_argv[i] = strdup(record_args[i]);
3279
3280         for (j = 1; j < (unsigned int)argc; j++, i++)
3281                 rec_argv[i] = argv[j];
3282
3283         BUG_ON(i != rec_argc);
3284
3285         return cmd_record(i, rec_argv);
3286 }
3287
3288 int cmd_sched(int argc, const char **argv)
3289 {
3290         const char default_sort_order[] = "avg, max, switch, runtime";
3291         struct perf_sched sched = {
3292                 .tool = {
3293                         .sample          = perf_sched__process_tracepoint_sample,
3294                         .comm            = perf_event__process_comm,
3295                         .namespaces      = perf_event__process_namespaces,
3296                         .lost            = perf_event__process_lost,
3297                         .fork            = perf_sched__process_fork_event,
3298                         .ordered_events = true,
3299                 },
3300                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3301                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3302                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3303                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3304                 .sort_order           = default_sort_order,
3305                 .replay_repeat        = 10,
3306                 .profile_cpu          = -1,
3307                 .next_shortname1      = 'A',
3308                 .next_shortname2      = '0',
3309                 .skip_merge           = 0,
3310                 .show_callchain       = 1,
3311                 .max_stack            = 5,
3312         };
3313         const struct option sched_options[] = {
3314         OPT_STRING('i', "input", &input_name, "file",
3315                     "input file name"),
3316         OPT_INCR('v', "verbose", &verbose,
3317                     "be more verbose (show symbol address, etc)"),
3318         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3319                     "dump raw trace in ASCII"),
3320         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3321         OPT_END()
3322         };
3323         const struct option latency_options[] = {
3324         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3325                    "sort by key(s): runtime, switch, avg, max"),
3326         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3327                     "CPU to profile on"),
3328         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3329                     "latency stats per pid instead of per comm"),
3330         OPT_PARENT(sched_options)
3331         };
3332         const struct option replay_options[] = {
3333         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3334                      "repeat the workload replay N times (-1: infinite)"),
3335         OPT_PARENT(sched_options)
3336         };
3337         const struct option map_options[] = {
3338         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3339                     "map output in compact mode"),
3340         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3341                    "highlight given pids in map"),
3342         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3343                     "highlight given CPUs in map"),
3344         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3345                     "display given CPUs in map"),
3346         OPT_PARENT(sched_options)
3347         };
3348         const struct option timehist_options[] = {
3349         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3350                    "file", "vmlinux pathname"),
3351         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3352                    "file", "kallsyms pathname"),
3353         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3354                     "Display call chains if present (default on)"),
3355         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3356                    "Maximum number of functions to display backtrace."),
3357         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3358                     "Look for files with symbols relative to this directory"),
3359         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3360                     "Show only syscall summary with statistics"),
3361         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3362                     "Show all syscalls and summary with statistics"),
3363         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3364         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3365         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3366         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3367         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3368         OPT_STRING(0, "time", &sched.time_str, "str",
3369                    "Time span for analysis (start,stop)"),
3370         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3371         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3372                    "analyze events only for given process id(s)"),
3373         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3374                    "analyze events only for given thread id(s)"),
3375         OPT_PARENT(sched_options)
3376         };
3377
3378         const char * const latency_usage[] = {
3379                 "perf sched latency [<options>]",
3380                 NULL
3381         };
3382         const char * const replay_usage[] = {
3383                 "perf sched replay [<options>]",
3384                 NULL
3385         };
3386         const char * const map_usage[] = {
3387                 "perf sched map [<options>]",
3388                 NULL
3389         };
3390         const char * const timehist_usage[] = {
3391                 "perf sched timehist [<options>]",
3392                 NULL
3393         };
3394         const char *const sched_subcommands[] = { "record", "latency", "map",
3395                                                   "replay", "script",
3396                                                   "timehist", NULL };
3397         const char *sched_usage[] = {
3398                 NULL,
3399                 NULL
3400         };
3401         struct trace_sched_handler lat_ops  = {
3402                 .wakeup_event       = latency_wakeup_event,
3403                 .switch_event       = latency_switch_event,
3404                 .runtime_event      = latency_runtime_event,
3405                 .migrate_task_event = latency_migrate_task_event,
3406         };
3407         struct trace_sched_handler map_ops  = {
3408                 .switch_event       = map_switch_event,
3409         };
3410         struct trace_sched_handler replay_ops  = {
3411                 .wakeup_event       = replay_wakeup_event,
3412                 .switch_event       = replay_switch_event,
3413                 .fork_event         = replay_fork_event,
3414         };
3415         unsigned int i;
3416
3417         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3418                 sched.curr_pid[i] = -1;
3419
3420         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3421                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3422         if (!argc)
3423                 usage_with_options(sched_usage, sched_options);
3424
3425         /*
3426          * Aliased to 'perf script' for now:
3427          */
3428         if (!strcmp(argv[0], "script"))
3429                 return cmd_script(argc, argv);
3430
3431         if (!strncmp(argv[0], "rec", 3)) {
3432                 return __cmd_record(argc, argv);
3433         } else if (!strncmp(argv[0], "lat", 3)) {
3434                 sched.tp_handler = &lat_ops;
3435                 if (argc > 1) {
3436                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3437                         if (argc)
3438                                 usage_with_options(latency_usage, latency_options);
3439                 }
3440                 setup_sorting(&sched, latency_options, latency_usage);
3441                 return perf_sched__lat(&sched);
3442         } else if (!strcmp(argv[0], "map")) {
3443                 if (argc) {
3444                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3445                         if (argc)
3446                                 usage_with_options(map_usage, map_options);
3447                 }
3448                 sched.tp_handler = &map_ops;
3449                 setup_sorting(&sched, latency_options, latency_usage);
3450                 return perf_sched__map(&sched);
3451         } else if (!strncmp(argv[0], "rep", 3)) {
3452                 sched.tp_handler = &replay_ops;
3453                 if (argc) {
3454                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3455                         if (argc)
3456                                 usage_with_options(replay_usage, replay_options);
3457                 }
3458                 return perf_sched__replay(&sched);
3459         } else if (!strcmp(argv[0], "timehist")) {
3460                 if (argc) {
3461                         argc = parse_options(argc, argv, timehist_options,
3462                                              timehist_usage, 0);
3463                         if (argc)
3464                                 usage_with_options(timehist_usage, timehist_options);
3465                 }
3466                 if ((sched.show_wakeups || sched.show_next) &&
3467                     sched.summary_only) {
3468                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3469                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3470                         if (sched.show_wakeups)
3471                                 parse_options_usage(NULL, timehist_options, "w", true);
3472                         if (sched.show_next)
3473                                 parse_options_usage(NULL, timehist_options, "n", true);
3474                         return -EINVAL;
3475                 }
3476
3477                 return perf_sched__timehist(&sched);
3478         } else {
3479                 usage_with_options(sched_usage, sched_options);
3480         }
3481
3482         return 0;
3483 }