tools/perf/builtin-sched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf.h"
   4 #include "perf-sys.h"
   5
   6 #include "util/cpumap.h"
   7 #include "util/evlist.h"
   8 #include "util/evsel.h"
   9 #include "util/evsel_fprintf.h"
  10 #include "util/symbol.h"
  11 #include "util/thread.h"
  12 #include "util/header.h"
  13 #include "util/session.h"
  14 #include "util/tool.h"
  15 #include "util/cloexec.h"
  16 #include "util/thread_map.h"
  17 #include "util/color.h"
  18 #include "util/stat.h"
  19 #include "util/string2.h"
  20 #include "util/callchain.h"
  21 #include "util/time-utils.h"
  22
  23 #include <subcmd/pager.h>
  24 #include <subcmd/parse-options.h>
  25 #include "util/trace-event.h"
  26
  27 #include "util/debug.h"
  28 #include "util/event.h"
  29
  30 #include <linux/kernel.h>
  31 #include <linux/log2.h>
  32 #include <linux/zalloc.h>
  33 #include <sys/prctl.h>
  34 #include <sys/resource.h>
  35 #include <inttypes.h>
  36
  37 #include <errno.h>
  38 #include <semaphore.h>
  39 #include <pthread.h>
  40 #include <math.h>
  41 #include <api/fs/fs.h>
  42 #include <perf/cpumap.h>
  43 #include <linux/time64.h>
  44 #include <linux/err.h>
  45
  46 #include <linux/ctype.h>
  47
  48 #define PR_SET_NAME             15               /* Set process name */
  49 #define MAX_CPUS                4096
  50 #define COMM_LEN                20
  51 #define SYM_LEN                 129
  52 #define MAX_PID                 1024000
  53
  54 static const char *cpu_list;
  55 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  56
  57 struct sched_atom;
  58
  59 struct task_desc {
  60         unsigned long           nr;
  61         unsigned long           pid;
  62         char                    comm[COMM_LEN];
  63
  64         unsigned long           nr_events;
  65         unsigned long           curr_event;
  66         struct sched_atom       **atoms;
  67
  68         pthread_t               thread;
  69         sem_t                   sleep_sem;
  70
  71         sem_t                   ready_for_work;
  72         sem_t                   work_done_sem;
  73
  74         u64                     cpu_usage;
  75 };
  76
  77 enum sched_event_type {
  78         SCHED_EVENT_RUN,
  79         SCHED_EVENT_SLEEP,
  80         SCHED_EVENT_WAKEUP,
  81         SCHED_EVENT_MIGRATION,
  82 };
  83
  84 struct sched_atom {
  85         enum sched_event_type   type;
  86         int                     specific_wait;
  87         u64                     timestamp;
  88         u64                     duration;
  89         unsigned long           nr;
  90         sem_t                   *wait_sem;
  91         struct task_desc        *wakee;
  92 };
  93
  94 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  95
  96 /* task state bitmask, copied from include/linux/sched.h */
  97 #define TASK_RUNNING            0
  98 #define TASK_INTERRUPTIBLE      1
  99 #define TASK_UNINTERRUPTIBLE    2
 100 #define __TASK_STOPPED          4
 101 #define __TASK_TRACED           8
 102 /* in tsk->exit_state */
 103 #define EXIT_DEAD               16
 104 #define EXIT_ZOMBIE             32
 105 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
 106 /* in tsk->state again */
 107 #define TASK_DEAD               64
 108 #define TASK_WAKEKILL           128
 109 #define TASK_WAKING             256
 110 #define TASK_PARKED             512
 111
 112 enum thread_state {
 113         THREAD_SLEEPING = 0,
 114         THREAD_WAIT_CPU,
 115         THREAD_SCHED_IN,
 116         THREAD_IGNORE
 117 };
 118
 119 struct work_atom {
 120         struct list_head        list;
 121         enum thread_state       state;
 122         u64                     sched_out_time;
 123         u64                     wake_up_time;
 124         u64                     sched_in_time;
 125         u64                     runtime;
 126 };
 127
 128 struct work_atoms {
 129         struct list_head        work_list;
 130         struct thread           *thread;
 131         struct rb_node          node;
 132         u64                     max_lat;
 133         u64                     max_lat_at;
 134         u64                     total_lat;
 135         u64                     nb_atoms;
 136         u64                     total_runtime;
 137         int                     num_merged;
 138 };
 139
 140 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 141
 142 struct perf_sched;
 143
 144 struct trace_sched_handler {
 145         int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 146                             struct perf_sample *sample, struct machine *machine);
 147
 148         int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 149                              struct perf_sample *sample, struct machine *machine);
 150
 151         int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 152                             struct perf_sample *sample, struct machine *machine);
 153
 154         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 155         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 156                           struct machine *machine);
 157
 158         int (*migrate_task_event)(struct perf_sched *sched,
 159                                   struct evsel *evsel,
 160                                   struct perf_sample *sample,
 161                                   struct machine *machine);
 162 };
 163
 164 #define COLOR_PIDS PERF_COLOR_BLUE
 165 #define COLOR_CPUS PERF_COLOR_BG_RED
 166
 167 struct perf_sched_map {
 168         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 169         int                     *comp_cpus;
 170         bool                     comp;
 171         struct perf_thread_map *color_pids;
 172         const char              *color_pids_str;
 173         struct perf_cpu_map     *color_cpus;
 174         const char              *color_cpus_str;
 175         struct perf_cpu_map     *cpus;
 176         const char              *cpus_str;
 177 };
 178
 179 struct perf_sched {
 180         struct perf_tool tool;
 181         const char       *sort_order;
 182         unsigned long    nr_tasks;
 183         struct task_desc **pid_to_task;
 184         struct task_desc **tasks;
 185         const struct trace_sched_handler *tp_handler;
 186         pthread_mutex_t  start_work_mutex;
 187         pthread_mutex_t  work_done_wait_mutex;
 188         int              profile_cpu;
 189 /*
 190  * Track the current task - that way we can know whether there's any
 191  * weird events, such as a task being switched away that is not current.
 192  */
 193         int              max_cpu;
 194         u32              curr_pid[MAX_CPUS];
 195         struct thread    *curr_thread[MAX_CPUS];
 196         char             next_shortname1;
 197         char             next_shortname2;
 198         unsigned int     replay_repeat;
 199         unsigned long    nr_run_events;
 200         unsigned long    nr_sleep_events;
 201         unsigned long    nr_wakeup_events;
 202         unsigned long    nr_sleep_corrections;
 203         unsigned long    nr_run_events_optimized;
 204         unsigned long    targetless_wakeups;
 205         unsigned long    multitarget_wakeups;
 206         unsigned long    nr_runs;
 207         unsigned long    nr_timestamps;
 208         unsigned long    nr_unordered_timestamps;
 209         unsigned long    nr_context_switch_bugs;
 210         unsigned long    nr_events;
 211         unsigned long    nr_lost_chunks;
 212         unsigned long    nr_lost_events;
 213         u64              run_measurement_overhead;
 214         u64              sleep_measurement_overhead;
 215         u64              start_time;
 216         u64              cpu_usage;
 217         u64              runavg_cpu_usage;
 218         u64              parent_cpu_usage;
 219         u64              runavg_parent_cpu_usage;
 220         u64              sum_runtime;
 221         u64              sum_fluct;
 222         u64              run_avg;
 223         u64              all_runtime;
 224         u64              all_count;
 225         u64              cpu_last_switched[MAX_CPUS];
 226         struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 227         struct list_head sort_list, cmp_pid;
 228         bool force;
 229         bool skip_merge;
 230         struct perf_sched_map map;
 231
 232         /* options for timehist command */
 233         bool            summary;
 234         bool            summary_only;
 235         bool            idle_hist;
 236         bool            show_callchain;
 237         unsigned int    max_stack;
 238         bool            show_cpu_visual;
 239         bool            show_wakeups;
 240         bool            show_next;
 241         bool            show_migrations;
 242         bool            show_state;
 243         u64             skipped_samples;
 244         const char      *time_str;
 245         struct perf_time_interval ptime;
 246         struct perf_time_interval hist_time;
 247 };
 248
 249 /* per thread run time data */
 250 struct thread_runtime {
 251         u64 last_time;      /* time of previous sched in/out event */
 252         u64 dt_run;         /* run time */
 253         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 254         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 255         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 256         u64 dt_delay;       /* time between wakeup and sched-in */
 257         u64 ready_to_run;   /* time of wakeup */
 258
 259         struct stats run_stats;
 260         u64 total_run_time;
 261         u64 total_sleep_time;
 262         u64 total_iowait_time;
 263         u64 total_preempt_time;
 264         u64 total_delay_time;
 265
 266         int last_state;
 267
 268         char shortname[3];
 269         bool comm_changed;
 270
 271         u64 migrations;
 272 };
 273
 274 /* per event run time data */
 275 struct evsel_runtime {
 276         u64 *last_time; /* time this event was last seen per cpu */
 277         u32 ncpu;       /* highest cpu slot allocated */
 278 };
 279
 280 /* per cpu idle time data */
 281 struct idle_thread_runtime {
 282         struct thread_runtime   tr;
 283         struct thread           *last_thread;
 284         struct rb_root_cached   sorted_root;
 285         struct callchain_root   callchain;
 286         struct callchain_cursor cursor;
 287 };
 288
 289 /* track idle times per cpu */
 290 static struct thread **idle_threads;
 291 static int idle_max_cpu;
 292 static char idle_comm[] = "<idle>";
 293
 294 static u64 get_nsecs(void)
 295 {
 296         struct timespec ts;
 297
 298         clock_gettime(CLOCK_MONOTONIC, &ts);
 299
 300         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 301 }
 302
 303 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 304 {
 305         u64 T0 = get_nsecs(), T1;
 306
 307         do {
 308                 T1 = get_nsecs();
 309         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 310 }
 311
 312 static void sleep_nsecs(u64 nsecs)
 313 {
 314         struct timespec ts;
 315
 316         ts.tv_nsec = nsecs % 999999999;
 317         ts.tv_sec = nsecs / 999999999;
 318
 319         nanosleep(&ts, NULL);
 320 }
 321
 322 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 323 {
 324         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 325         int i;
 326
 327         for (i = 0; i < 10; i++) {
 328                 T0 = get_nsecs();
 329                 burn_nsecs(sched, 0);
 330                 T1 = get_nsecs();
 331                 delta = T1-T0;
 332                 min_delta = min(min_delta, delta);
 333         }
 334         sched->run_measurement_overhead = min_delta;
 335
 336         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 337 }
 338
 339 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 340 {
 341         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 342         int i;
 343
 344         for (i = 0; i < 10; i++) {
 345                 T0 = get_nsecs();
 346                 sleep_nsecs(10000);
 347                 T1 = get_nsecs();
 348                 delta = T1-T0;
 349                 min_delta = min(min_delta, delta);
 350         }
 351         min_delta -= 10000;
 352         sched->sleep_measurement_overhead = min_delta;
 353
 354         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 355 }
 356
 357 static struct sched_atom *
 358 get_new_event(struct task_desc *task, u64 timestamp)
 359 {
 360         struct sched_atom *event = zalloc(sizeof(*event));
 361         unsigned long idx = task->nr_events;
 362         size_t size;
 363
 364         event->timestamp = timestamp;
 365         event->nr = idx;
 366
 367         task->nr_events++;
 368         size = sizeof(struct sched_atom *) * task->nr_events;
 369         task->atoms = realloc(task->atoms, size);
 370         BUG_ON(!task->atoms);
 371
 372         task->atoms[idx] = event;
 373
 374         return event;
 375 }
 376
 377 static struct sched_atom *last_event(struct task_desc *task)
 378 {
 379         if (!task->nr_events)
 380                 return NULL;
 381
 382         return task->atoms[task->nr_events - 1];
 383 }
 384
 385 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 386                                 u64 timestamp, u64 duration)
 387 {
 388         struct sched_atom *event, *curr_event = last_event(task);
 389
 390         /*
 391          * optimize an existing RUN event by merging this one
 392          * to it:
 393          */
 394         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 395                 sched->nr_run_events_optimized++;
 396                 curr_event->duration += duration;
 397                 return;
 398         }
 399
 400         event = get_new_event(task, timestamp);
 401
 402         event->type = SCHED_EVENT_RUN;
 403         event->duration = duration;
 404
 405         sched->nr_run_events++;
 406 }
 407
 408 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 409                                    u64 timestamp, struct task_desc *wakee)
 410 {
 411         struct sched_atom *event, *wakee_event;
 412
 413         event = get_new_event(task, timestamp);
 414         event->type = SCHED_EVENT_WAKEUP;
 415         event->wakee = wakee;
 416
 417         wakee_event = last_event(wakee);
 418         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 419                 sched->targetless_wakeups++;
 420                 return;
 421         }
 422         if (wakee_event->wait_sem) {
 423                 sched->multitarget_wakeups++;
 424                 return;
 425         }
 426
 427         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 428         sem_init(wakee_event->wait_sem, 0, 0);
 429         wakee_event->specific_wait = 1;
 430         event->wait_sem = wakee_event->wait_sem;
 431
 432         sched->nr_wakeup_events++;
 433 }
 434
 435 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 436                                   u64 timestamp, u64 task_state __maybe_unused)
 437 {
 438         struct sched_atom *event = get_new_event(task, timestamp);
 439
 440         event->type = SCHED_EVENT_SLEEP;
 441
 442         sched->nr_sleep_events++;
 443 }
 444
 445 static struct task_desc *register_pid(struct perf_sched *sched,
 446                                       unsigned long pid, const char *comm)
 447 {
 448         struct task_desc *task;
 449         static int pid_max;
 450
 451         if (sched->pid_to_task == NULL) {
 452                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 453                         pid_max = MAX_PID;
 454                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 455         }
 456         if (pid >= (unsigned long)pid_max) {
 457                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 458                         sizeof(struct task_desc *))) == NULL);
 459                 while (pid >= (unsigned long)pid_max)
 460                         sched->pid_to_task[pid_max++] = NULL;
 461         }
 462
 463         task = sched->pid_to_task[pid];
 464
 465         if (task)
 466                 return task;
 467
 468         task = zalloc(sizeof(*task));
 469         task->pid = pid;
 470         task->nr = sched->nr_tasks;
 471         strcpy(task->comm, comm);
 472         /*
 473          * every task starts in sleeping state - this gets ignored
 474          * if there's no wakeup pointing to this sleep state:
 475          */
 476         add_sched_event_sleep(sched, task, 0, 0);
 477
 478         sched->pid_to_task[pid] = task;
 479         sched->nr_tasks++;
 480         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 481         BUG_ON(!sched->tasks);
 482         sched->tasks[task->nr] = task;
 483
 484         if (verbose > 0)
 485                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 486
 487         return task;
 488 }
 489
 490
 491 static void print_task_traces(struct perf_sched *sched)
 492 {
 493         struct task_desc *task;
 494         unsigned long i;
 495
 496         for (i = 0; i < sched->nr_tasks; i++) {
 497                 task = sched->tasks[i];
 498                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 499                         task->nr, task->comm, task->pid, task->nr_events);
 500         }
 501 }
 502
 503 static void add_cross_task_wakeups(struct perf_sched *sched)
 504 {
 505         struct task_desc *task1, *task2;
 506         unsigned long i, j;
 507
 508         for (i = 0; i < sched->nr_tasks; i++) {
 509                 task1 = sched->tasks[i];
 510                 j = i + 1;
 511                 if (j == sched->nr_tasks)
 512                         j = 0;
 513                 task2 = sched->tasks[j];
 514                 add_sched_event_wakeup(sched, task1, 0, task2);
 515         }
 516 }
 517
 518 static void perf_sched__process_event(struct perf_sched *sched,
 519                                       struct sched_atom *atom)
 520 {
 521         int ret = 0;
 522
 523         switch (atom->type) {
 524                 case SCHED_EVENT_RUN:
 525                         burn_nsecs(sched, atom->duration);
 526                         break;
 527                 case SCHED_EVENT_SLEEP:
 528                         if (atom->wait_sem)
 529                                 ret = sem_wait(atom->wait_sem);
 530                         BUG_ON(ret);
 531                         break;
 532                 case SCHED_EVENT_WAKEUP:
 533                         if (atom->wait_sem)
 534                                 ret = sem_post(atom->wait_sem);
 535                         BUG_ON(ret);
 536                         break;
 537                 case SCHED_EVENT_MIGRATION:
 538                         break;
 539                 default:
 540                         BUG_ON(1);
 541         }
 542 }
 543
 544 static u64 get_cpu_usage_nsec_parent(void)
 545 {
 546         struct rusage ru;
 547         u64 sum;
 548         int err;
 549
 550         err = getrusage(RUSAGE_SELF, &ru);
 551         BUG_ON(err);
 552
 553         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 554         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 555
 556         return sum;
 557 }
 558
 559 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 560 {
 561         struct perf_event_attr attr;
 562         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 563         int fd;
 564         struct rlimit limit;
 565         bool need_privilege = false;
 566
 567         memset(&attr, 0, sizeof(attr));
 568
 569         attr.type = PERF_TYPE_SOFTWARE;
 570         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 571
 572 force_again:
 573         fd = sys_perf_event_open(&attr, 0, -1, -1,
 574                                  perf_event_open_cloexec_flag());
 575
 576         if (fd < 0) {
 577                 if (errno == EMFILE) {
 578                         if (sched->force) {
 579                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 580                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 581                                 if (limit.rlim_cur > limit.rlim_max) {
 582                                         limit.rlim_max = limit.rlim_cur;
 583                                         need_privilege = true;
 584                                 }
 585                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 586                                         if (need_privilege && errno == EPERM)
 587                                                 strcpy(info, "Need privilege\n");
 588                                 } else
 589                                         goto force_again;
 590                         } else
 591                                 strcpy(info, "Have a try with -f option\n");
 592                 }
 593                 pr_err("Error: sys_perf_event_open() syscall returned "
 594                        "with %d (%s)\n%s", fd,
 595                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 596                 exit(EXIT_FAILURE);
 597         }
 598         return fd;
 599 }
 600
 601 static u64 get_cpu_usage_nsec_self(int fd)
 602 {
 603         u64 runtime;
 604         int ret;
 605
 606         ret = read(fd, &runtime, sizeof(runtime));
 607         BUG_ON(ret != sizeof(runtime));
 608
 609         return runtime;
 610 }
 611
 612 struct sched_thread_parms {
 613         struct task_desc  *task;
 614         struct perf_sched *sched;
 615         int fd;
 616 };
 617
 618 static void *thread_func(void *ctx)
 619 {
 620         struct sched_thread_parms *parms = ctx;
 621         struct task_desc *this_task = parms->task;
 622         struct perf_sched *sched = parms->sched;
 623         u64 cpu_usage_0, cpu_usage_1;
 624         unsigned long i, ret;
 625         char comm2[22];
 626         int fd = parms->fd;
 627
 628         zfree(&parms);
 629
 630         sprintf(comm2, ":%s", this_task->comm);
 631         prctl(PR_SET_NAME, comm2);
 632         if (fd < 0)
 633                 return NULL;
 634 again:
 635         ret = sem_post(&this_task->ready_for_work);
 636         BUG_ON(ret);
 637         ret = pthread_mutex_lock(&sched->start_work_mutex);
 638         BUG_ON(ret);
 639         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 640         BUG_ON(ret);
 641
 642         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 643
 644         for (i = 0; i < this_task->nr_events; i++) {
 645                 this_task->curr_event = i;
 646                 perf_sched__process_event(sched, this_task->atoms[i]);
 647         }
 648
 649         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 650         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 651         ret = sem_post(&this_task->work_done_sem);
 652         BUG_ON(ret);
 653
 654         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 655         BUG_ON(ret);
 656         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 657         BUG_ON(ret);
 658
 659         goto again;
 660 }
 661
 662 static void create_tasks(struct perf_sched *sched)
 663 {
 664         struct task_desc *task;
 665         pthread_attr_t attr;
 666         unsigned long i;
 667         int err;
 668
 669         err = pthread_attr_init(&attr);
 670         BUG_ON(err);
 671         err = pthread_attr_setstacksize(&attr,
 672                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 673         BUG_ON(err);
 674         err = pthread_mutex_lock(&sched->start_work_mutex);
 675         BUG_ON(err);
 676         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 677         BUG_ON(err);
 678         for (i = 0; i < sched->nr_tasks; i++) {
 679                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 680                 BUG_ON(parms == NULL);
 681                 parms->task = task = sched->tasks[i];
 682                 parms->sched = sched;
 683                 parms->fd = self_open_counters(sched, i);
 684                 sem_init(&task->sleep_sem, 0, 0);
 685                 sem_init(&task->ready_for_work, 0, 0);
 686                 sem_init(&task->work_done_sem, 0, 0);
 687                 task->curr_event = 0;
 688                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 689                 BUG_ON(err);
 690         }
 691 }
 692
 693 static void wait_for_tasks(struct perf_sched *sched)
 694 {
 695         u64 cpu_usage_0, cpu_usage_1;
 696         struct task_desc *task;
 697         unsigned long i, ret;
 698
 699         sched->start_time = get_nsecs();
 700         sched->cpu_usage = 0;
 701         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 702
 703         for (i = 0; i < sched->nr_tasks; i++) {
 704                 task = sched->tasks[i];
 705                 ret = sem_wait(&task->ready_for_work);
 706                 BUG_ON(ret);
 707                 sem_init(&task->ready_for_work, 0, 0);
 708         }
 709         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 710         BUG_ON(ret);
 711
 712         cpu_usage_0 = get_cpu_usage_nsec_parent();
 713
 714         pthread_mutex_unlock(&sched->start_work_mutex);
 715
 716         for (i = 0; i < sched->nr_tasks; i++) {
 717                 task = sched->tasks[i];
 718                 ret = sem_wait(&task->work_done_sem);
 719                 BUG_ON(ret);
 720                 sem_init(&task->work_done_sem, 0, 0);
 721                 sched->cpu_usage += task->cpu_usage;
 722                 task->cpu_usage = 0;
 723         }
 724
 725         cpu_usage_1 = get_cpu_usage_nsec_parent();
 726         if (!sched->runavg_cpu_usage)
 727                 sched->runavg_cpu_usage = sched->cpu_usage;
 728         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 729
 730         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 731         if (!sched->runavg_parent_cpu_usage)
 732                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 733         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 734                                          sched->parent_cpu_usage)/sched->replay_repeat;
 735
 736         ret = pthread_mutex_lock(&sched->start_work_mutex);
 737         BUG_ON(ret);
 738
 739         for (i = 0; i < sched->nr_tasks; i++) {
 740                 task = sched->tasks[i];
 741                 sem_init(&task->sleep_sem, 0, 0);
 742                 task->curr_event = 0;
 743         }
 744 }
 745
 746 static void run_one_test(struct perf_sched *sched)
 747 {
 748         u64 T0, T1, delta, avg_delta, fluct;
 749
 750         T0 = get_nsecs();
 751         wait_for_tasks(sched);
 752         T1 = get_nsecs();
 753
 754         delta = T1 - T0;
 755         sched->sum_runtime += delta;
 756         sched->nr_runs++;
 757
 758         avg_delta = sched->sum_runtime / sched->nr_runs;
 759         if (delta < avg_delta)
 760                 fluct = avg_delta - delta;
 761         else
 762                 fluct = delta - avg_delta;
 763         sched->sum_fluct += fluct;
 764         if (!sched->run_avg)
 765                 sched->run_avg = delta;
 766         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 767
 768         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 769
 770         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 771
 772         printf("cpu: %0.2f / %0.2f",
 773                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 774
 775 #if 0
 776         /*
 777          * rusage statistics done by the parent, these are less
 778          * accurate than the sched->sum_exec_runtime based statistics:
 779          */
 780         printf(" [%0.2f / %0.2f]",
 781                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 782                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 783 #endif
 784
 785         printf("\n");
 786
 787         if (sched->nr_sleep_corrections)
 788                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 789         sched->nr_sleep_corrections = 0;
 790 }
 791
 792 static void test_calibrations(struct perf_sched *sched)
 793 {
 794         u64 T0, T1;
 795
 796         T0 = get_nsecs();
 797         burn_nsecs(sched, NSEC_PER_MSEC);
 798         T1 = get_nsecs();
 799
 800         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 801
 802         T0 = get_nsecs();
 803         sleep_nsecs(NSEC_PER_MSEC);
 804         T1 = get_nsecs();
 805
 806         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 807 }
 808
 809 static int
 810 replay_wakeup_event(struct perf_sched *sched,
 811                     struct evsel *evsel, struct perf_sample *sample,
 812                     struct machine *machine __maybe_unused)
 813 {
 814         const char *comm = evsel__strval(evsel, sample, "comm");
 815         const u32 pid    = evsel__intval(evsel, sample, "pid");
 816         struct task_desc *waker, *wakee;
 817
 818         if (verbose > 0) {
 819                 printf("sched_wakeup event %p\n", evsel);
 820
 821                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 822         }
 823
 824         waker = register_pid(sched, sample->tid, "<unknown>");
 825         wakee = register_pid(sched, pid, comm);
 826
 827         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 828         return 0;
 829 }
 830
 831 static int replay_switch_event(struct perf_sched *sched,
 832                                struct evsel *evsel,
 833                                struct perf_sample *sample,
 834                                struct machine *machine __maybe_unused)
 835 {
 836         const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 837                    *next_comm  = evsel__strval(evsel, sample, "next_comm");
 838         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 839                   next_pid = evsel__intval(evsel, sample, "next_pid");
 840         const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
 841         struct task_desc *prev, __maybe_unused *next;
 842         u64 timestamp0, timestamp = sample->time;
 843         int cpu = sample->cpu;
 844         s64 delta;
 845
 846         if (verbose > 0)
 847                 printf("sched_switch event %p\n", evsel);
 848
 849         if (cpu >= MAX_CPUS || cpu < 0)
 850                 return 0;
 851
 852         timestamp0 = sched->cpu_last_switched[cpu];
 853         if (timestamp0)
 854                 delta = timestamp - timestamp0;
 855         else
 856                 delta = 0;
 857
 858         if (delta < 0) {
 859                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 860                 return -1;
 861         }
 862
 863         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 864                  prev_comm, prev_pid, next_comm, next_pid, delta);
 865
 866         prev = register_pid(sched, prev_pid, prev_comm);
 867         next = register_pid(sched, next_pid, next_comm);
 868
 869         sched->cpu_last_switched[cpu] = timestamp;
 870
 871         add_sched_event_run(sched, prev, timestamp, delta);
 872         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 873
 874         return 0;
 875 }
 876
 877 static int replay_fork_event(struct perf_sched *sched,
 878                              union perf_event *event,
 879                              struct machine *machine)
 880 {
 881         struct thread *child, *parent;
 882
 883         child = machine__findnew_thread(machine, event->fork.pid,
 884                                         event->fork.tid);
 885         parent = machine__findnew_thread(machine, event->fork.ppid,
 886                                          event->fork.ptid);
 887
 888         if (child == NULL || parent == NULL) {
 889                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 890                                  child, parent);
 891                 goto out_put;
 892         }
 893
 894         if (verbose > 0) {
 895                 printf("fork event\n");
 896                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 897                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 898         }
 899
 900         register_pid(sched, parent->tid, thread__comm_str(parent));
 901         register_pid(sched, child->tid, thread__comm_str(child));
 902 out_put:
 903         thread__put(child);
 904         thread__put(parent);
 905         return 0;
 906 }
 907
 908 struct sort_dimension {
 909         const char              *name;
 910         sort_fn_t               cmp;
 911         struct list_head        list;
 912 };
 913
 914 /*
 915  * handle runtime stats saved per thread
 916  */
 917 static struct thread_runtime *thread__init_runtime(struct thread *thread)
 918 {
 919         struct thread_runtime *r;
 920
 921         r = zalloc(sizeof(struct thread_runtime));
 922         if (!r)
 923                 return NULL;
 924
 925         init_stats(&r->run_stats);
 926         thread__set_priv(thread, r);
 927
 928         return r;
 929 }
 930
 931 static struct thread_runtime *thread__get_runtime(struct thread *thread)
 932 {
 933         struct thread_runtime *tr;
 934
 935         tr = thread__priv(thread);
 936         if (tr == NULL) {
 937                 tr = thread__init_runtime(thread);
 938                 if (tr == NULL)
 939                         pr_debug("Failed to malloc memory for runtime data.\n");
 940         }
 941
 942         return tr;
 943 }
 944
 945 static int
 946 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 947 {
 948         struct sort_dimension *sort;
 949         int ret = 0;
 950
 951         BUG_ON(list_empty(list));
 952
 953         list_for_each_entry(sort, list, list) {
 954                 ret = sort->cmp(l, r);
 955                 if (ret)
 956                         return ret;
 957         }
 958
 959         return ret;
 960 }
 961
 962 static struct work_atoms *
 963 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 964                          struct list_head *sort_list)
 965 {
 966         struct rb_node *node = root->rb_root.rb_node;
 967         struct work_atoms key = { .thread = thread };
 968
 969         while (node) {
 970                 struct work_atoms *atoms;
 971                 int cmp;
 972
 973                 atoms = container_of(node, struct work_atoms, node);
 974
 975                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 976                 if (cmp > 0)
 977                         node = node->rb_left;
 978                 else if (cmp < 0)
 979                         node = node->rb_right;
 980                 else {
 981                         BUG_ON(thread != atoms->thread);
 982                         return atoms;
 983                 }
 984         }
 985         return NULL;
 986 }
 987
 988 static void
 989 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 990                          struct list_head *sort_list)
 991 {
 992         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 993         bool leftmost = true;
 994
 995         while (*new) {
 996                 struct work_atoms *this;
 997                 int cmp;
 998
 999                 this = container_of(*new, struct work_atoms, node);
1000                 parent = *new;
1001
1002                 cmp = thread_lat_cmp(sort_list, data, this);
1003
1004                 if (cmp > 0)
1005                         new = &((*new)->rb_left);
1006                 else {
1007                         new = &((*new)->rb_right);
1008                         leftmost = false;
1009                 }
1010         }
1011
1012         rb_link_node(&data->node, parent, new);
1013         rb_insert_color_cached(&data->node, root, leftmost);
1014 }
1015
1016 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1017 {
1018         struct work_atoms *atoms = zalloc(sizeof(*atoms));
1019         if (!atoms) {
1020                 pr_err("No memory at %s\n", __func__);
1021                 return -1;
1022         }
1023
1024         atoms->thread = thread__get(thread);
1025         INIT_LIST_HEAD(&atoms->work_list);
1026         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1027         return 0;
1028 }
1029
1030 static char sched_out_state(u64 prev_state)
1031 {
1032         const char *str = TASK_STATE_TO_CHAR_STR;
1033
1034         return str[prev_state];
1035 }
1036
1037 static int
1038 add_sched_out_event(struct work_atoms *atoms,
1039                     char run_state,
1040                     u64 timestamp)
1041 {
1042         struct work_atom *atom = zalloc(sizeof(*atom));
1043         if (!atom) {
1044                 pr_err("Non memory at %s", __func__);
1045                 return -1;
1046         }
1047
1048         atom->sched_out_time = timestamp;
1049
1050         if (run_state == 'R') {
1051                 atom->state = THREAD_WAIT_CPU;
1052                 atom->wake_up_time = atom->sched_out_time;
1053         }
1054
1055         list_add_tail(&atom->list, &atoms->work_list);
1056         return 0;
1057 }
1058
1059 static void
1060 add_runtime_event(struct work_atoms *atoms, u64 delta,
1061                   u64 timestamp __maybe_unused)
1062 {
1063         struct work_atom *atom;
1064
1065         BUG_ON(list_empty(&atoms->work_list));
1066
1067         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1068
1069         atom->runtime += delta;
1070         atoms->total_runtime += delta;
1071 }
1072
1073 static void
1074 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1075 {
1076         struct work_atom *atom;
1077         u64 delta;
1078
1079         if (list_empty(&atoms->work_list))
1080                 return;
1081
1082         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1083
1084         if (atom->state != THREAD_WAIT_CPU)
1085                 return;
1086
1087         if (timestamp < atom->wake_up_time) {
1088                 atom->state = THREAD_IGNORE;
1089                 return;
1090         }
1091
1092         atom->state = THREAD_SCHED_IN;
1093         atom->sched_in_time = timestamp;
1094
1095         delta = atom->sched_in_time - atom->wake_up_time;
1096         atoms->total_lat += delta;
1097         if (delta > atoms->max_lat) {
1098                 atoms->max_lat = delta;
1099                 atoms->max_lat_at = timestamp;
1100         }
1101         atoms->nb_atoms++;
1102 }
1103
1104 static int latency_switch_event(struct perf_sched *sched,
1105                                 struct evsel *evsel,
1106                                 struct perf_sample *sample,
1107                                 struct machine *machine)
1108 {
1109         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1110                   next_pid = evsel__intval(evsel, sample, "next_pid");
1111         const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1112         struct work_atoms *out_events, *in_events;
1113         struct thread *sched_out, *sched_in;
1114         u64 timestamp0, timestamp = sample->time;
1115         int cpu = sample->cpu, err = -1;
1116         s64 delta;
1117
1118         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1119
1120         timestamp0 = sched->cpu_last_switched[cpu];
1121         sched->cpu_last_switched[cpu] = timestamp;
1122         if (timestamp0)
1123                 delta = timestamp - timestamp0;
1124         else
1125                 delta = 0;
1126
1127         if (delta < 0) {
1128                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1129                 return -1;
1130         }
1131
1132         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1133         sched_in = machine__findnew_thread(machine, -1, next_pid);
1134         if (sched_out == NULL || sched_in == NULL)
1135                 goto out_put;
1136
1137         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1138         if (!out_events) {
1139                 if (thread_atoms_insert(sched, sched_out))
1140                         goto out_put;
1141                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1142                 if (!out_events) {
1143                         pr_err("out-event: Internal tree error");
1144                         goto out_put;
1145                 }
1146         }
1147         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1148                 return -1;
1149
1150         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1151         if (!in_events) {
1152                 if (thread_atoms_insert(sched, sched_in))
1153                         goto out_put;
1154                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1155                 if (!in_events) {
1156                         pr_err("in-event: Internal tree error");
1157                         goto out_put;
1158                 }
1159                 /*
1160                  * Take came in we have not heard about yet,
1161                  * add in an initial atom in runnable state:
1162                  */
1163                 if (add_sched_out_event(in_events, 'R', timestamp))
1164                         goto out_put;
1165         }
1166         add_sched_in_event(in_events, timestamp);
1167         err = 0;
1168 out_put:
1169         thread__put(sched_out);
1170         thread__put(sched_in);
1171         return err;
1172 }
1173
1174 static int latency_runtime_event(struct perf_sched *sched,
1175                                  struct evsel *evsel,
1176                                  struct perf_sample *sample,
1177                                  struct machine *machine)
1178 {
1179         const u32 pid      = evsel__intval(evsel, sample, "pid");
1180         const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1181         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1182         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1183         u64 timestamp = sample->time;
1184         int cpu = sample->cpu, err = -1;
1185
1186         if (thread == NULL)
1187                 return -1;
1188
1189         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1190         if (!atoms) {
1191                 if (thread_atoms_insert(sched, thread))
1192                         goto out_put;
1193                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1194                 if (!atoms) {
1195                         pr_err("in-event: Internal tree error");
1196                         goto out_put;
1197                 }
1198                 if (add_sched_out_event(atoms, 'R', timestamp))
1199                         goto out_put;
1200         }
1201
1202         add_runtime_event(atoms, runtime, timestamp);
1203         err = 0;
1204 out_put:
1205         thread__put(thread);
1206         return err;
1207 }
1208
1209 static int latency_wakeup_event(struct perf_sched *sched,
1210                                 struct evsel *evsel,
1211                                 struct perf_sample *sample,
1212                                 struct machine *machine)
1213 {
1214         const u32 pid     = evsel__intval(evsel, sample, "pid");
1215         struct work_atoms *atoms;
1216         struct work_atom *atom;
1217         struct thread *wakee;
1218         u64 timestamp = sample->time;
1219         int err = -1;
1220
1221         wakee = machine__findnew_thread(machine, -1, pid);
1222         if (wakee == NULL)
1223                 return -1;
1224         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1225         if (!atoms) {
1226                 if (thread_atoms_insert(sched, wakee))
1227                         goto out_put;
1228                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1229                 if (!atoms) {
1230                         pr_err("wakeup-event: Internal tree error");
1231                         goto out_put;
1232                 }
1233                 if (add_sched_out_event(atoms, 'S', timestamp))
1234                         goto out_put;
1235         }
1236
1237         BUG_ON(list_empty(&atoms->work_list));
1238
1239         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1240
1241         /*
1242          * As we do not guarantee the wakeup event happens when
1243          * task is out of run queue, also may happen when task is
1244          * on run queue and wakeup only change ->state to TASK_RUNNING,
1245          * then we should not set the ->wake_up_time when wake up a
1246          * task which is on run queue.
1247          *
1248          * You WILL be missing events if you've recorded only
1249          * one CPU, or are only looking at only one, so don't
1250          * skip in this case.
1251          */
1252         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1253                 goto out_ok;
1254
1255         sched->nr_timestamps++;
1256         if (atom->sched_out_time > timestamp) {
1257                 sched->nr_unordered_timestamps++;
1258                 goto out_ok;
1259         }
1260
1261         atom->state = THREAD_WAIT_CPU;
1262         atom->wake_up_time = timestamp;
1263 out_ok:
1264         err = 0;
1265 out_put:
1266         thread__put(wakee);
1267         return err;
1268 }
1269
1270 static int latency_migrate_task_event(struct perf_sched *sched,
1271                                       struct evsel *evsel,
1272                                       struct perf_sample *sample,
1273                                       struct machine *machine)
1274 {
1275         const u32 pid = evsel__intval(evsel, sample, "pid");
1276         u64 timestamp = sample->time;
1277         struct work_atoms *atoms;
1278         struct work_atom *atom;
1279         struct thread *migrant;
1280         int err = -1;
1281
1282         /*
1283          * Only need to worry about migration when profiling one CPU.
1284          */
1285         if (sched->profile_cpu == -1)
1286                 return 0;
1287
1288         migrant = machine__findnew_thread(machine, -1, pid);
1289         if (migrant == NULL)
1290                 return -1;
1291         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1292         if (!atoms) {
1293                 if (thread_atoms_insert(sched, migrant))
1294                         goto out_put;
1295                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1296                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1297                 if (!atoms) {
1298                         pr_err("migration-event: Internal tree error");
1299                         goto out_put;
1300                 }
1301                 if (add_sched_out_event(atoms, 'R', timestamp))
1302                         goto out_put;
1303         }
1304
1305         BUG_ON(list_empty(&atoms->work_list));
1306
1307         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1308         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1309
1310         sched->nr_timestamps++;
1311
1312         if (atom->sched_out_time > timestamp)
1313                 sched->nr_unordered_timestamps++;
1314         err = 0;
1315 out_put:
1316         thread__put(migrant);
1317         return err;
1318 }
1319
1320 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1321 {
1322         int i;
1323         int ret;
1324         u64 avg;
1325         char max_lat_at[32];
1326
1327         if (!work_list->nb_atoms)
1328                 return;
1329         /*
1330          * Ignore idle threads:
1331          */
1332         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1333                 return;
1334
1335         sched->all_runtime += work_list->total_runtime;
1336         sched->all_count   += work_list->nb_atoms;
1337
1338         if (work_list->num_merged > 1)
1339                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1340         else
1341                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1342
1343         for (i = 0; i < 24 - ret; i++)
1344                 printf(" ");
1345
1346         avg = work_list->total_lat / work_list->nb_atoms;
1347         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1348
1349         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1350               (double)work_list->total_runtime / NSEC_PER_MSEC,
1351                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1352                  (double)work_list->max_lat / NSEC_PER_MSEC,
1353                  max_lat_at);
1354 }
1355
1356 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1357 {
1358         if (l->thread == r->thread)
1359                 return 0;
1360         if (l->thread->tid < r->thread->tid)
1361                 return -1;
1362         if (l->thread->tid > r->thread->tid)
1363                 return 1;
1364         return (int)(l->thread - r->thread);
1365 }
1366
1367 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1368 {
1369         u64 avgl, avgr;
1370
1371         if (!l->nb_atoms)
1372                 return -1;
1373
1374         if (!r->nb_atoms)
1375                 return 1;
1376
1377         avgl = l->total_lat / l->nb_atoms;
1378         avgr = r->total_lat / r->nb_atoms;
1379
1380         if (avgl < avgr)
1381                 return -1;
1382         if (avgl > avgr)
1383                 return 1;
1384
1385         return 0;
1386 }
1387
1388 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1389 {
1390         if (l->max_lat < r->max_lat)
1391                 return -1;
1392         if (l->max_lat > r->max_lat)
1393                 return 1;
1394
1395         return 0;
1396 }
1397
1398 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1399 {
1400         if (l->nb_atoms < r->nb_atoms)
1401                 return -1;
1402         if (l->nb_atoms > r->nb_atoms)
1403                 return 1;
1404
1405         return 0;
1406 }
1407
1408 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1409 {
1410         if (l->total_runtime < r->total_runtime)
1411                 return -1;
1412         if (l->total_runtime > r->total_runtime)
1413                 return 1;
1414
1415         return 0;
1416 }
1417
1418 static int sort_dimension__add(const char *tok, struct list_head *list)
1419 {
1420         size_t i;
1421         static struct sort_dimension avg_sort_dimension = {
1422                 .name = "avg",
1423                 .cmp  = avg_cmp,
1424         };
1425         static struct sort_dimension max_sort_dimension = {
1426                 .name = "max",
1427                 .cmp  = max_cmp,
1428         };
1429         static struct sort_dimension pid_sort_dimension = {
1430                 .name = "pid",
1431                 .cmp  = pid_cmp,
1432         };
1433         static struct sort_dimension runtime_sort_dimension = {
1434                 .name = "runtime",
1435                 .cmp  = runtime_cmp,
1436         };
1437         static struct sort_dimension switch_sort_dimension = {
1438                 .name = "switch",
1439                 .cmp  = switch_cmp,
1440         };
1441         struct sort_dimension *available_sorts[] = {
1442                 &pid_sort_dimension,
1443                 &avg_sort_dimension,
1444                 &max_sort_dimension,
1445                 &switch_sort_dimension,
1446                 &runtime_sort_dimension,
1447         };
1448
1449         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1450                 if (!strcmp(available_sorts[i]->name, tok)) {
1451                         list_add_tail(&available_sorts[i]->list, list);
1452
1453                         return 0;
1454                 }
1455         }
1456
1457         return -1;
1458 }
1459
1460 static void perf_sched__sort_lat(struct perf_sched *sched)
1461 {
1462         struct rb_node *node;
1463         struct rb_root_cached *root = &sched->atom_root;
1464 again:
1465         for (;;) {
1466                 struct work_atoms *data;
1467                 node = rb_first_cached(root);
1468                 if (!node)
1469                         break;
1470
1471                 rb_erase_cached(node, root);
1472                 data = rb_entry(node, struct work_atoms, node);
1473                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1474         }
1475         if (root == &sched->atom_root) {
1476                 root = &sched->merged_atom_root;
1477                 goto again;
1478         }
1479 }
1480
1481 static int process_sched_wakeup_event(struct perf_tool *tool,
1482                                       struct evsel *evsel,
1483                                       struct perf_sample *sample,
1484                                       struct machine *machine)
1485 {
1486         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1487
1488         if (sched->tp_handler->wakeup_event)
1489                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1490
1491         return 0;
1492 }
1493
1494 union map_priv {
1495         void    *ptr;
1496         bool     color;
1497 };
1498
1499 static bool thread__has_color(struct thread *thread)
1500 {
1501         union map_priv priv = {
1502                 .ptr = thread__priv(thread),
1503         };
1504
1505         return priv.color;
1506 }
1507
1508 static struct thread*
1509 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1510 {
1511         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1512         union map_priv priv = {
1513                 .color = false,
1514         };
1515
1516         if (!sched->map.color_pids || !thread || thread__priv(thread))
1517                 return thread;
1518
1519         if (thread_map__has(sched->map.color_pids, tid))
1520                 priv.color = true;
1521
1522         thread__set_priv(thread, priv.ptr);
1523         return thread;
1524 }
1525
1526 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1527                             struct perf_sample *sample, struct machine *machine)
1528 {
1529         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1530         struct thread *sched_in;
1531         struct thread_runtime *tr;
1532         int new_shortname;
1533         u64 timestamp0, timestamp = sample->time;
1534         s64 delta;
1535         int i, this_cpu = sample->cpu;
1536         int cpus_nr;
1537         bool new_cpu = false;
1538         const char *color = PERF_COLOR_NORMAL;
1539         char stimestamp[32];
1540
1541         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1542
1543         if (this_cpu > sched->max_cpu)
1544                 sched->max_cpu = this_cpu;
1545
1546         if (sched->map.comp) {
1547                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1548                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1549                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1550                         new_cpu = true;
1551                 }
1552         } else
1553                 cpus_nr = sched->max_cpu;
1554
1555         timestamp0 = sched->cpu_last_switched[this_cpu];
1556         sched->cpu_last_switched[this_cpu] = timestamp;
1557         if (timestamp0)
1558                 delta = timestamp - timestamp0;
1559         else
1560                 delta = 0;
1561
1562         if (delta < 0) {
1563                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1564                 return -1;
1565         }
1566
1567         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1568         if (sched_in == NULL)
1569                 return -1;
1570
1571         tr = thread__get_runtime(sched_in);
1572         if (tr == NULL) {
1573                 thread__put(sched_in);
1574                 return -1;
1575         }
1576
1577         sched->curr_thread[this_cpu] = thread__get(sched_in);
1578
1579         printf("  ");
1580
1581         new_shortname = 0;
1582         if (!tr->shortname[0]) {
1583                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1584                         /*
1585                          * Don't allocate a letter-number for swapper:0
1586                          * as a shortname. Instead, we use '.' for it.
1587                          */
1588                         tr->shortname[0] = '.';
1589                         tr->shortname[1] = ' ';
1590                 } else {
1591                         tr->shortname[0] = sched->next_shortname1;
1592                         tr->shortname[1] = sched->next_shortname2;
1593
1594                         if (sched->next_shortname1 < 'Z') {
1595                                 sched->next_shortname1++;
1596                         } else {
1597                                 sched->next_shortname1 = 'A';
1598                                 if (sched->next_shortname2 < '9')
1599                                         sched->next_shortname2++;
1600                                 else
1601                                         sched->next_shortname2 = '0';
1602                         }
1603                 }
1604                 new_shortname = 1;
1605         }
1606
1607         for (i = 0; i < cpus_nr; i++) {
1608                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1609                 struct thread *curr_thread = sched->curr_thread[cpu];
1610                 struct thread_runtime *curr_tr;
1611                 const char *pid_color = color;
1612                 const char *cpu_color = color;
1613
1614                 if (curr_thread && thread__has_color(curr_thread))
1615                         pid_color = COLOR_PIDS;
1616
1617                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1618                         continue;
1619
1620                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1621                         cpu_color = COLOR_CPUS;
1622
1623                 if (cpu != this_cpu)
1624                         color_fprintf(stdout, color, " ");
1625                 else
1626                         color_fprintf(stdout, cpu_color, "*");
1627
1628                 if (sched->curr_thread[cpu]) {
1629                         curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1630                         if (curr_tr == NULL) {
1631                                 thread__put(sched_in);
1632                                 return -1;
1633                         }
1634                         color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1635                 } else
1636                         color_fprintf(stdout, color, "   ");
1637         }
1638
1639         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1640                 goto out;
1641
1642         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1643         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1644         if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1645                 const char *pid_color = color;
1646
1647                 if (thread__has_color(sched_in))
1648                         pid_color = COLOR_PIDS;
1649
1650                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1651                        tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1652                 tr->comm_changed = false;
1653         }
1654
1655         if (sched->map.comp && new_cpu)
1656                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1657
1658 out:
1659         color_fprintf(stdout, color, "\n");
1660
1661         thread__put(sched_in);
1662
1663         return 0;
1664 }
1665
1666 static int process_sched_switch_event(struct perf_tool *tool,
1667                                       struct evsel *evsel,
1668                                       struct perf_sample *sample,
1669                                       struct machine *machine)
1670 {
1671         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1672         int this_cpu = sample->cpu, err = 0;
1673         u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1674             next_pid = evsel__intval(evsel, sample, "next_pid");
1675
1676         if (sched->curr_pid[this_cpu] != (u32)-1) {
1677                 /*
1678                  * Are we trying to switch away a PID that is
1679                  * not current?
1680                  */
1681                 if (sched->curr_pid[this_cpu] != prev_pid)
1682                         sched->nr_context_switch_bugs++;
1683         }
1684
1685         if (sched->tp_handler->switch_event)
1686                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1687
1688         sched->curr_pid[this_cpu] = next_pid;
1689         return err;
1690 }
1691
1692 static int process_sched_runtime_event(struct perf_tool *tool,
1693                                        struct evsel *evsel,
1694                                        struct perf_sample *sample,
1695                                        struct machine *machine)
1696 {
1697         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1698
1699         if (sched->tp_handler->runtime_event)
1700                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1701
1702         return 0;
1703 }
1704
1705 static int perf_sched__process_fork_event(struct perf_tool *tool,
1706                                           union perf_event *event,
1707                                           struct perf_sample *sample,
1708                                           struct machine *machine)
1709 {
1710         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1711
1712         /* run the fork event through the perf machineruy */
1713         perf_event__process_fork(tool, event, sample, machine);
1714
1715         /* and then run additional processing needed for this command */
1716         if (sched->tp_handler->fork_event)
1717                 return sched->tp_handler->fork_event(sched, event, machine);
1718
1719         return 0;
1720 }
1721
1722 static int process_sched_migrate_task_event(struct perf_tool *tool,
1723                                             struct evsel *evsel,
1724                                             struct perf_sample *sample,
1725                                             struct machine *machine)
1726 {
1727         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1728
1729         if (sched->tp_handler->migrate_task_event)
1730                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1731
1732         return 0;
1733 }
1734
1735 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1736                                   struct evsel *evsel,
1737                                   struct perf_sample *sample,
1738                                   struct machine *machine);
1739
1740 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1741                                                  union perf_event *event __maybe_unused,
1742                                                  struct perf_sample *sample,
1743                                                  struct evsel *evsel,
1744                                                  struct machine *machine)
1745 {
1746         int err = 0;
1747
1748         if (evsel->handler != NULL) {
1749                 tracepoint_handler f = evsel->handler;
1750                 err = f(tool, evsel, sample, machine);
1751         }
1752
1753         return err;
1754 }
1755
1756 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1757                                     union perf_event *event,
1758                                     struct perf_sample *sample,
1759                                     struct machine *machine)
1760 {
1761         struct thread *thread;
1762         struct thread_runtime *tr;
1763         int err;
1764
1765         err = perf_event__process_comm(tool, event, sample, machine);
1766         if (err)
1767                 return err;
1768
1769         thread = machine__find_thread(machine, sample->pid, sample->tid);
1770         if (!thread) {
1771                 pr_err("Internal error: can't find thread\n");
1772                 return -1;
1773         }
1774
1775         tr = thread__get_runtime(thread);
1776         if (tr == NULL) {
1777                 thread__put(thread);
1778                 return -1;
1779         }
1780
1781         tr->comm_changed = true;
1782         thread__put(thread);
1783
1784         return 0;
1785 }
1786
1787 static int perf_sched__read_events(struct perf_sched *sched)
1788 {
1789         const struct evsel_str_handler handlers[] = {
1790                 { "sched:sched_switch",       process_sched_switch_event, },
1791                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1792                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1793                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1794                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1795         };
1796         struct perf_session *session;
1797         struct perf_data data = {
1798                 .path  = input_name,
1799                 .mode  = PERF_DATA_MODE_READ,
1800                 .force = sched->force,
1801         };
1802         int rc = -1;
1803
1804         session = perf_session__new(&data, false, &sched->tool);
1805         if (IS_ERR(session)) {
1806                 pr_debug("Error creating perf session");
1807                 return PTR_ERR(session);
1808         }
1809
1810         symbol__init(&session->header.env);
1811
1812         if (perf_session__set_tracepoints_handlers(session, handlers))
1813                 goto out_delete;
1814
1815         if (perf_session__has_traces(session, "record -R")) {
1816                 int err = perf_session__process_events(session);
1817                 if (err) {
1818                         pr_err("Failed to process events, error %d", err);
1819                         goto out_delete;
1820                 }
1821
1822                 sched->nr_events      = session->evlist->stats.nr_events[0];
1823                 sched->nr_lost_events = session->evlist->stats.total_lost;
1824                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1825         }
1826
1827         rc = 0;
1828 out_delete:
1829         perf_session__delete(session);
1830         return rc;
1831 }
1832
1833 /*
1834  * scheduling times are printed as msec.usec
1835  */
1836 static inline void print_sched_time(unsigned long long nsecs, int width)
1837 {
1838         unsigned long msecs;
1839         unsigned long usecs;
1840
1841         msecs  = nsecs / NSEC_PER_MSEC;
1842         nsecs -= msecs * NSEC_PER_MSEC;
1843         usecs  = nsecs / NSEC_PER_USEC;
1844         printf("%*lu.%03lu ", width, msecs, usecs);
1845 }
1846
1847 /*
1848  * returns runtime data for event, allocating memory for it the
1849  * first time it is used.
1850  */
1851 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1852 {
1853         struct evsel_runtime *r = evsel->priv;
1854
1855         if (r == NULL) {
1856                 r = zalloc(sizeof(struct evsel_runtime));
1857                 evsel->priv = r;
1858         }
1859
1860         return r;
1861 }
1862
1863 /*
1864  * save last time event was seen per cpu
1865  */
1866 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1867 {
1868         struct evsel_runtime *r = evsel__get_runtime(evsel);
1869
1870         if (r == NULL)
1871                 return;
1872
1873         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1874                 int i, n = __roundup_pow_of_two(cpu+1);
1875                 void *p = r->last_time;
1876
1877                 p = realloc(r->last_time, n * sizeof(u64));
1878                 if (!p)
1879                         return;
1880
1881                 r->last_time = p;
1882                 for (i = r->ncpu; i < n; ++i)
1883                         r->last_time[i] = (u64) 0;
1884
1885                 r->ncpu = n;
1886         }
1887
1888         r->last_time[cpu] = timestamp;
1889 }
1890
1891 /* returns last time this event was seen on the given cpu */
1892 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1893 {
1894         struct evsel_runtime *r = evsel__get_runtime(evsel);
1895
1896         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1897                 return 0;
1898
1899         return r->last_time[cpu];
1900 }
1901
1902 static int comm_width = 30;
1903
1904 static char *timehist_get_commstr(struct thread *thread)
1905 {
1906         static char str[32];
1907         const char *comm = thread__comm_str(thread);
1908         pid_t tid = thread->tid;
1909         pid_t pid = thread->pid_;
1910         int n;
1911
1912         if (pid == 0)
1913                 n = scnprintf(str, sizeof(str), "%s", comm);
1914
1915         else if (tid != pid)
1916                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1917
1918         else
1919                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1920
1921         if (n > comm_width)
1922                 comm_width = n;
1923
1924         return str;
1925 }
1926
1927 static void timehist_header(struct perf_sched *sched)
1928 {
1929         u32 ncpus = sched->max_cpu + 1;
1930         u32 i, j;
1931
1932         printf("%15s %6s ", "time", "cpu");
1933
1934         if (sched->show_cpu_visual) {
1935                 printf(" ");
1936                 for (i = 0, j = 0; i < ncpus; ++i) {
1937                         printf("%x", j++);
1938                         if (j > 15)
1939                                 j = 0;
1940                 }
1941                 printf(" ");
1942         }
1943
1944         printf(" %-*s  %9s  %9s  %9s", comm_width,
1945                 "task name", "wait time", "sch delay", "run time");
1946
1947         if (sched->show_state)
1948                 printf("  %s", "state");
1949
1950         printf("\n");
1951
1952         /*
1953          * units row
1954          */
1955         printf("%15s %-6s ", "", "");
1956
1957         if (sched->show_cpu_visual)
1958                 printf(" %*s ", ncpus, "");
1959
1960         printf(" %-*s  %9s  %9s  %9s", comm_width,
1961                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1962
1963         if (sched->show_state)
1964                 printf("  %5s", "");
1965
1966         printf("\n");
1967
1968         /*
1969          * separator
1970          */
1971         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1972
1973         if (sched->show_cpu_visual)
1974                 printf(" %.*s ", ncpus, graph_dotted_line);
1975
1976         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1977                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1978                 graph_dotted_line);
1979
1980         if (sched->show_state)
1981                 printf("  %.5s", graph_dotted_line);
1982
1983         printf("\n");
1984 }
1985
1986 static char task_state_char(struct thread *thread, int state)
1987 {
1988         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1989         unsigned bit = state ? ffs(state) : 0;
1990
1991         /* 'I' for idle */
1992         if (thread->tid == 0)
1993                 return 'I';
1994
1995         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1996 }
1997
1998 static void timehist_print_sample(struct perf_sched *sched,
1999                                   struct evsel *evsel,
2000                                   struct perf_sample *sample,
2001                                   struct addr_location *al,
2002                                   struct thread *thread,
2003                                   u64 t, int state)
2004 {
2005         struct thread_runtime *tr = thread__priv(thread);
2006         const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2007         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2008         u32 max_cpus = sched->max_cpu + 1;
2009         char tstr[64];
2010         char nstr[30];
2011         u64 wait_time;
2012
2013         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2014                 return;
2015
2016         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2017         printf("%15s [%04d] ", tstr, sample->cpu);
2018
2019         if (sched->show_cpu_visual) {
2020                 u32 i;
2021                 char c;
2022
2023                 printf(" ");
2024                 for (i = 0; i < max_cpus; ++i) {
2025                         /* flag idle times with 'i'; others are sched events */
2026                         if (i == sample->cpu)
2027                                 c = (thread->tid == 0) ? 'i' : 's';
2028                         else
2029                                 c = ' ';
2030                         printf("%c", c);
2031                 }
2032                 printf(" ");
2033         }
2034
2035         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2036
2037         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2038         print_sched_time(wait_time, 6);
2039
2040         print_sched_time(tr->dt_delay, 6);
2041         print_sched_time(tr->dt_run, 6);
2042
2043         if (sched->show_state)
2044                 printf(" %5c ", task_state_char(thread, state));
2045
2046         if (sched->show_next) {
2047                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2048                 printf(" %-*s", comm_width, nstr);
2049         }
2050
2051         if (sched->show_wakeups && !sched->show_next)
2052                 printf("  %-*s", comm_width, "");
2053
2054         if (thread->tid == 0)
2055                 goto out;
2056
2057         if (sched->show_callchain)
2058                 printf("  ");
2059
2060         sample__fprintf_sym(sample, al, 0,
2061                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2062                             EVSEL__PRINT_CALLCHAIN_ARROW |
2063                             EVSEL__PRINT_SKIP_IGNORED,
2064                             &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2065
2066 out:
2067         printf("\n");
2068 }
2069
2070 /*
2071  * Explanation of delta-time stats:
2072  *
2073  *            t = time of current schedule out event
2074  *        tprev = time of previous sched out event
2075  *                also time of schedule-in event for current task
2076  *    last_time = time of last sched change event for current task
2077  *                (i.e, time process was last scheduled out)
2078  * ready_to_run = time of wakeup for current task
2079  *
2080  * -----|------------|------------|------------|------
2081  *    last         ready        tprev          t
2082  *    time         to run
2083  *
2084  *      |-------- dt_wait --------|
2085  *                   |- dt_delay -|-- dt_run --|
2086  *
2087  *   dt_run = run time of current task
2088  *  dt_wait = time between last schedule out event for task and tprev
2089  *            represents time spent off the cpu
2090  * dt_delay = time between wakeup and schedule-in of task
2091  */
2092
2093 static void timehist_update_runtime_stats(struct thread_runtime *r,
2094                                          u64 t, u64 tprev)
2095 {
2096         r->dt_delay   = 0;
2097         r->dt_sleep   = 0;
2098         r->dt_iowait  = 0;
2099         r->dt_preempt = 0;
2100         r->dt_run     = 0;
2101
2102         if (tprev) {
2103                 r->dt_run = t - tprev;
2104                 if (r->ready_to_run) {
2105                         if (r->ready_to_run > tprev)
2106                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2107                         else
2108                                 r->dt_delay = tprev - r->ready_to_run;
2109                 }
2110
2111                 if (r->last_time > tprev)
2112                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2113                 else if (r->last_time) {
2114                         u64 dt_wait = tprev - r->last_time;
2115
2116                         if (r->last_state == TASK_RUNNING)
2117                                 r->dt_preempt = dt_wait;
2118                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2119                                 r->dt_iowait = dt_wait;
2120                         else
2121                                 r->dt_sleep = dt_wait;
2122                 }
2123         }
2124
2125         update_stats(&r->run_stats, r->dt_run);
2126
2127         r->total_run_time     += r->dt_run;
2128         r->total_delay_time   += r->dt_delay;
2129         r->total_sleep_time   += r->dt_sleep;
2130         r->total_iowait_time  += r->dt_iowait;
2131         r->total_preempt_time += r->dt_preempt;
2132 }
2133
2134 static bool is_idle_sample(struct perf_sample *sample,
2135                            struct evsel *evsel)
2136 {
2137         /* pid 0 == swapper == idle task */
2138         if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2139                 return evsel__intval(evsel, sample, "prev_pid") == 0;
2140
2141         return sample->pid == 0;
2142 }
2143
2144 static void save_task_callchain(struct perf_sched *sched,
2145                                 struct perf_sample *sample,
2146                                 struct evsel *evsel,
2147                                 struct machine *machine)
2148 {
2149         struct callchain_cursor *cursor = &callchain_cursor;
2150         struct thread *thread;
2151
2152         /* want main thread for process - has maps */
2153         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2154         if (thread == NULL) {
2155                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2156                 return;
2157         }
2158
2159         if (!sched->show_callchain || sample->callchain == NULL)
2160                 return;
2161
2162         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2163                                       NULL, NULL, sched->max_stack + 2) != 0) {
2164                 if (verbose > 0)
2165                         pr_err("Failed to resolve callchain. Skipping\n");
2166
2167                 return;
2168         }
2169
2170         callchain_cursor_commit(cursor);
2171
2172         while (true) {
2173                 struct callchain_cursor_node *node;
2174                 struct symbol *sym;
2175
2176                 node = callchain_cursor_current(cursor);
2177                 if (node == NULL)
2178                         break;
2179
2180                 sym = node->ms.sym;
2181                 if (sym) {
2182                         if (!strcmp(sym->name, "schedule") ||
2183                             !strcmp(sym->name, "__schedule") ||
2184                             !strcmp(sym->name, "preempt_schedule"))
2185                                 sym->ignore = 1;
2186                 }
2187
2188                 callchain_cursor_advance(cursor);
2189         }
2190 }
2191
2192 static int init_idle_thread(struct thread *thread)
2193 {
2194         struct idle_thread_runtime *itr;
2195
2196         thread__set_comm(thread, idle_comm, 0);
2197
2198         itr = zalloc(sizeof(*itr));
2199         if (itr == NULL)
2200                 return -ENOMEM;
2201
2202         init_stats(&itr->tr.run_stats);
2203         callchain_init(&itr->callchain);
2204         callchain_cursor_reset(&itr->cursor);
2205         thread__set_priv(thread, itr);
2206
2207         return 0;
2208 }
2209
2210 /*
2211  * Track idle stats per cpu by maintaining a local thread
2212  * struct for the idle task on each cpu.
2213  */
2214 static int init_idle_threads(int ncpu)
2215 {
2216         int i, ret;
2217
2218         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2219         if (!idle_threads)
2220                 return -ENOMEM;
2221
2222         idle_max_cpu = ncpu;
2223
2224         /* allocate the actual thread struct if needed */
2225         for (i = 0; i < ncpu; ++i) {
2226                 idle_threads[i] = thread__new(0, 0);
2227                 if (idle_threads[i] == NULL)
2228                         return -ENOMEM;
2229
2230                 ret = init_idle_thread(idle_threads[i]);
2231                 if (ret < 0)
2232                         return ret;
2233         }
2234
2235         return 0;
2236 }
2237
2238 static void free_idle_threads(void)
2239 {
2240         int i;
2241
2242         if (idle_threads == NULL)
2243                 return;
2244
2245         for (i = 0; i < idle_max_cpu; ++i) {
2246                 if ((idle_threads[i]))
2247                         thread__delete(idle_threads[i]);
2248         }
2249
2250         free(idle_threads);
2251 }
2252
2253 static struct thread *get_idle_thread(int cpu)
2254 {
2255         /*
2256          * expand/allocate array of pointers to local thread
2257          * structs if needed
2258          */
2259         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2260                 int i, j = __roundup_pow_of_two(cpu+1);
2261                 void *p;
2262
2263                 p = realloc(idle_threads, j * sizeof(struct thread *));
2264                 if (!p)
2265                         return NULL;
2266
2267                 idle_threads = (struct thread **) p;
2268                 for (i = idle_max_cpu; i < j; ++i)
2269                         idle_threads[i] = NULL;
2270
2271                 idle_max_cpu = j;
2272         }
2273
2274         /* allocate a new thread struct if needed */
2275         if (idle_threads[cpu] == NULL) {
2276                 idle_threads[cpu] = thread__new(0, 0);
2277
2278                 if (idle_threads[cpu]) {
2279                         if (init_idle_thread(idle_threads[cpu]) < 0)
2280                                 return NULL;
2281                 }
2282         }
2283
2284         return idle_threads[cpu];
2285 }
2286
2287 static void save_idle_callchain(struct perf_sched *sched,
2288                                 struct idle_thread_runtime *itr,
2289                                 struct perf_sample *sample)
2290 {
2291         if (!sched->show_callchain || sample->callchain == NULL)
2292                 return;
2293
2294         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2295 }
2296
2297 static struct thread *timehist_get_thread(struct perf_sched *sched,
2298                                           struct perf_sample *sample,
2299                                           struct machine *machine,
2300                                           struct evsel *evsel)
2301 {
2302         struct thread *thread;
2303
2304         if (is_idle_sample(sample, evsel)) {
2305                 thread = get_idle_thread(sample->cpu);
2306                 if (thread == NULL)
2307                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2308
2309         } else {
2310                 /* there were samples with tid 0 but non-zero pid */
2311                 thread = machine__findnew_thread(machine, sample->pid,
2312                                                  sample->tid ?: sample->pid);
2313                 if (thread == NULL) {
2314                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2315                                  sample->tid);
2316                 }
2317
2318                 save_task_callchain(sched, sample, evsel, machine);
2319                 if (sched->idle_hist) {
2320                         struct thread *idle;
2321                         struct idle_thread_runtime *itr;
2322
2323                         idle = get_idle_thread(sample->cpu);
2324                         if (idle == NULL) {
2325                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2326                                 return NULL;
2327                         }
2328
2329                         itr = thread__priv(idle);
2330                         if (itr == NULL)
2331                                 return NULL;
2332
2333                         itr->last_thread = thread;
2334
2335                         /* copy task callchain when entering to idle */
2336                         if (evsel__intval(evsel, sample, "next_pid") == 0)
2337                                 save_idle_callchain(sched, itr, sample);
2338                 }
2339         }
2340
2341         return thread;
2342 }
2343
2344 static bool timehist_skip_sample(struct perf_sched *sched,
2345                                  struct thread *thread,
2346                                  struct evsel *evsel,
2347                                  struct perf_sample *sample)
2348 {
2349         bool rc = false;
2350
2351         if (thread__is_filtered(thread)) {
2352                 rc = true;
2353                 sched->skipped_samples++;
2354         }
2355
2356         if (sched->idle_hist) {
2357                 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2358                         rc = true;
2359                 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2360                          evsel__intval(evsel, sample, "next_pid") != 0)
2361                         rc = true;
2362         }
2363
2364         return rc;
2365 }
2366
2367 static void timehist_print_wakeup_event(struct perf_sched *sched,
2368                                         struct evsel *evsel,
2369                                         struct perf_sample *sample,
2370                                         struct machine *machine,
2371                                         struct thread *awakened)
2372 {
2373         struct thread *thread;
2374         char tstr[64];
2375
2376         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2377         if (thread == NULL)
2378                 return;
2379
2380         /* show wakeup unless both awakee and awaker are filtered */
2381         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2382             timehist_skip_sample(sched, awakened, evsel, sample)) {
2383                 return;
2384         }
2385
2386         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2387         printf("%15s [%04d] ", tstr, sample->cpu);
2388         if (sched->show_cpu_visual)
2389                 printf(" %*s ", sched->max_cpu + 1, "");
2390
2391         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2392
2393         /* dt spacer */
2394         printf("  %9s  %9s  %9s ", "", "", "");
2395
2396         printf("awakened: %s", timehist_get_commstr(awakened));
2397
2398         printf("\n");
2399 }
2400
2401 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2402                                        union perf_event *event __maybe_unused,
2403                                        struct evsel *evsel,
2404                                        struct perf_sample *sample,
2405                                        struct machine *machine)
2406 {
2407         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2408         struct thread *thread;
2409         struct thread_runtime *tr = NULL;
2410         /* want pid of awakened task not pid in sample */
2411         const u32 pid = evsel__intval(evsel, sample, "pid");
2412
2413         thread = machine__findnew_thread(machine, 0, pid);
2414         if (thread == NULL)
2415                 return -1;
2416
2417         tr = thread__get_runtime(thread);
2418         if (tr == NULL)
2419                 return -1;
2420
2421         if (tr->ready_to_run == 0)
2422                 tr->ready_to_run = sample->time;
2423
2424         /* show wakeups if requested */
2425         if (sched->show_wakeups &&
2426             !perf_time__skip_sample(&sched->ptime, sample->time))
2427                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2428
2429         return 0;
2430 }
2431
2432 static void timehist_print_migration_event(struct perf_sched *sched,
2433                                         struct evsel *evsel,
2434                                         struct perf_sample *sample,
2435                                         struct machine *machine,
2436                                         struct thread *migrated)
2437 {
2438         struct thread *thread;
2439         char tstr[64];
2440         u32 max_cpus = sched->max_cpu + 1;
2441         u32 ocpu, dcpu;
2442
2443         if (sched->summary_only)
2444                 return;
2445
2446         max_cpus = sched->max_cpu + 1;
2447         ocpu = evsel__intval(evsel, sample, "orig_cpu");
2448         dcpu = evsel__intval(evsel, sample, "dest_cpu");
2449
2450         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2451         if (thread == NULL)
2452                 return;
2453
2454         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2455             timehist_skip_sample(sched, migrated, evsel, sample)) {
2456                 return;
2457         }
2458
2459         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2460         printf("%15s [%04d] ", tstr, sample->cpu);
2461
2462         if (sched->show_cpu_visual) {
2463                 u32 i;
2464                 char c;
2465
2466                 printf("  ");
2467                 for (i = 0; i < max_cpus; ++i) {
2468                         c = (i == sample->cpu) ? 'm' : ' ';
2469                         printf("%c", c);
2470                 }
2471                 printf("  ");
2472         }
2473
2474         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2475
2476         /* dt spacer */
2477         printf("  %9s  %9s  %9s ", "", "", "");
2478
2479         printf("migrated: %s", timehist_get_commstr(migrated));
2480         printf(" cpu %d => %d", ocpu, dcpu);
2481
2482         printf("\n");
2483 }
2484
2485 static int timehist_migrate_task_event(struct perf_tool *tool,
2486                                        union perf_event *event __maybe_unused,
2487                                        struct evsel *evsel,
2488                                        struct perf_sample *sample,
2489                                        struct machine *machine)
2490 {
2491         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2492         struct thread *thread;
2493         struct thread_runtime *tr = NULL;
2494         /* want pid of migrated task not pid in sample */
2495         const u32 pid = evsel__intval(evsel, sample, "pid");
2496
2497         thread = machine__findnew_thread(machine, 0, pid);
2498         if (thread == NULL)
2499                 return -1;
2500
2501         tr = thread__get_runtime(thread);
2502         if (tr == NULL)
2503                 return -1;
2504
2505         tr->migrations++;
2506
2507         /* show migrations if requested */
2508         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2509
2510         return 0;
2511 }
2512
2513 static int timehist_sched_change_event(struct perf_tool *tool,
2514                                        union perf_event *event,
2515                                        struct evsel *evsel,
2516                                        struct perf_sample *sample,
2517                                        struct machine *machine)
2518 {
2519         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2520         struct perf_time_interval *ptime = &sched->ptime;
2521         struct addr_location al;
2522         struct thread *thread;
2523         struct thread_runtime *tr = NULL;
2524         u64 tprev, t = sample->time;
2525         int rc = 0;
2526         int state = evsel__intval(evsel, sample, "prev_state");
2527
2528         if (machine__resolve(machine, &al, sample) < 0) {
2529                 pr_err("problem processing %d event. skipping it\n",
2530                        event->header.type);
2531                 rc = -1;
2532                 goto out;
2533         }
2534
2535         thread = timehist_get_thread(sched, sample, machine, evsel);
2536         if (thread == NULL) {
2537                 rc = -1;
2538                 goto out;
2539         }
2540
2541         if (timehist_skip_sample(sched, thread, evsel, sample))
2542                 goto out;
2543
2544         tr = thread__get_runtime(thread);
2545         if (tr == NULL) {
2546                 rc = -1;
2547                 goto out;
2548         }
2549
2550         tprev = evsel__get_time(evsel, sample->cpu);
2551
2552         /*
2553          * If start time given:
2554          * - sample time is under window user cares about - skip sample
2555          * - tprev is under window user cares about  - reset to start of window
2556          */
2557         if (ptime->start && ptime->start > t)
2558                 goto out;
2559
2560         if (tprev && ptime->start > tprev)
2561                 tprev = ptime->start;
2562
2563         /*
2564          * If end time given:
2565          * - previous sched event is out of window - we are done
2566          * - sample time is beyond window user cares about - reset it
2567          *   to close out stats for time window interest
2568          */
2569         if (ptime->end) {
2570                 if (tprev > ptime->end)
2571                         goto out;
2572
2573                 if (t > ptime->end)
2574                         t = ptime->end;
2575         }
2576
2577         if (!sched->idle_hist || thread->tid == 0) {
2578                 timehist_update_runtime_stats(tr, t, tprev);
2579
2580                 if (sched->idle_hist) {
2581                         struct idle_thread_runtime *itr = (void *)tr;
2582                         struct thread_runtime *last_tr;
2583
2584                         BUG_ON(thread->tid != 0);
2585
2586                         if (itr->last_thread == NULL)
2587                                 goto out;
2588
2589                         /* add current idle time as last thread's runtime */
2590                         last_tr = thread__get_runtime(itr->last_thread);
2591                         if (last_tr == NULL)
2592                                 goto out;
2593
2594                         timehist_update_runtime_stats(last_tr, t, tprev);
2595                         /*
2596                          * remove delta time of last thread as it's not updated
2597                          * and otherwise it will show an invalid value next
2598                          * time.  we only care total run time and run stat.
2599                          */
2600                         last_tr->dt_run = 0;
2601                         last_tr->dt_delay = 0;
2602                         last_tr->dt_sleep = 0;
2603                         last_tr->dt_iowait = 0;
2604                         last_tr->dt_preempt = 0;
2605
2606                         if (itr->cursor.nr)
2607                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2608
2609                         itr->last_thread = NULL;
2610                 }
2611         }
2612
2613         if (!sched->summary_only)
2614                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2615
2616 out:
2617         if (sched->hist_time.start == 0 && t >= ptime->start)
2618                 sched->hist_time.start = t;
2619         if (ptime->end == 0 || t <= ptime->end)
2620                 sched->hist_time.end = t;
2621
2622         if (tr) {
2623                 /* time of this sched_switch event becomes last time task seen */
2624                 tr->last_time = sample->time;
2625
2626                 /* last state is used to determine where to account wait time */
2627                 tr->last_state = state;
2628
2629                 /* sched out event for task so reset ready to run time */
2630                 tr->ready_to_run = 0;
2631         }
2632
2633         evsel__save_time(evsel, sample->time, sample->cpu);
2634
2635         return rc;
2636 }
2637
2638 static int timehist_sched_switch_event(struct perf_tool *tool,
2639                              union perf_event *event,
2640                              struct evsel *evsel,
2641                              struct perf_sample *sample,
2642                              struct machine *machine __maybe_unused)
2643 {
2644         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2645 }
2646
2647 static int process_lost(struct perf_tool *tool __maybe_unused,
2648                         union perf_event *event,
2649                         struct perf_sample *sample,
2650                         struct machine *machine __maybe_unused)
2651 {
2652         char tstr[64];
2653
2654         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2655         printf("%15s ", tstr);
2656         printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2657
2658         return 0;
2659 }
2660
2661
2662 static void print_thread_runtime(struct thread *t,
2663                                  struct thread_runtime *r)
2664 {
2665         double mean = avg_stats(&r->run_stats);
2666         float stddev;
2667
2668         printf("%*s   %5d  %9" PRIu64 " ",
2669                comm_width, timehist_get_commstr(t), t->ppid,
2670                (u64) r->run_stats.n);
2671
2672         print_sched_time(r->total_run_time, 8);
2673         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2674         print_sched_time(r->run_stats.min, 6);
2675         printf(" ");
2676         print_sched_time((u64) mean, 6);
2677         printf(" ");
2678         print_sched_time(r->run_stats.max, 6);
2679         printf("  ");
2680         printf("%5.2f", stddev);
2681         printf("   %5" PRIu64, r->migrations);
2682         printf("\n");
2683 }
2684
2685 static void print_thread_waittime(struct thread *t,
2686                                   struct thread_runtime *r)
2687 {
2688         printf("%*s   %5d  %9" PRIu64 " ",
2689                comm_width, timehist_get_commstr(t), t->ppid,
2690                (u64) r->run_stats.n);
2691
2692         print_sched_time(r->total_run_time, 8);
2693         print_sched_time(r->total_sleep_time, 6);
2694         printf(" ");
2695         print_sched_time(r->total_iowait_time, 6);
2696         printf(" ");
2697         print_sched_time(r->total_preempt_time, 6);
2698         printf(" ");
2699         print_sched_time(r->total_delay_time, 6);
2700         printf("\n");
2701 }
2702
2703 struct total_run_stats {
2704         struct perf_sched *sched;
2705         u64  sched_count;
2706         u64  task_count;
2707         u64  total_run_time;
2708 };
2709
2710 static int __show_thread_runtime(struct thread *t, void *priv)
2711 {
2712         struct total_run_stats *stats = priv;
2713         struct thread_runtime *r;
2714
2715         if (thread__is_filtered(t))
2716                 return 0;
2717
2718         r = thread__priv(t);
2719         if (r && r->run_stats.n) {
2720                 stats->task_count++;
2721                 stats->sched_count += r->run_stats.n;
2722                 stats->total_run_time += r->total_run_time;
2723
2724                 if (stats->sched->show_state)
2725                         print_thread_waittime(t, r);
2726                 else
2727                         print_thread_runtime(t, r);
2728         }
2729
2730         return 0;
2731 }
2732
2733 static int show_thread_runtime(struct thread *t, void *priv)
2734 {
2735         if (t->dead)
2736                 return 0;
2737
2738         return __show_thread_runtime(t, priv);
2739 }
2740
2741 static int show_deadthread_runtime(struct thread *t, void *priv)
2742 {
2743         if (!t->dead)
2744                 return 0;
2745
2746         return __show_thread_runtime(t, priv);
2747 }
2748
2749 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2750 {
2751         const char *sep = " <- ";
2752         struct callchain_list *chain;
2753         size_t ret = 0;
2754         char bf[1024];
2755         bool first;
2756
2757         if (node == NULL)
2758                 return 0;
2759
2760         ret = callchain__fprintf_folded(fp, node->parent);
2761         first = (ret == 0);
2762
2763         list_for_each_entry(chain, &node->val, list) {
2764                 if (chain->ip >= PERF_CONTEXT_MAX)
2765                         continue;
2766                 if (chain->ms.sym && chain->ms.sym->ignore)
2767                         continue;
2768                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2769                                callchain_list__sym_name(chain, bf, sizeof(bf),
2770                                                         false));
2771                 first = false;
2772         }
2773
2774         return ret;
2775 }
2776
2777 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2778 {
2779         size_t ret = 0;
2780         FILE *fp = stdout;
2781         struct callchain_node *chain;
2782         struct rb_node *rb_node = rb_first_cached(root);
2783
2784         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2785         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2786                graph_dotted_line);
2787
2788         while (rb_node) {
2789                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2790                 rb_node = rb_next(rb_node);
2791
2792                 ret += fprintf(fp, "  ");
2793                 print_sched_time(chain->hit, 12);
2794                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2795                 ret += fprintf(fp, " %8d  ", chain->count);
2796                 ret += callchain__fprintf_folded(fp, chain);
2797                 ret += fprintf(fp, "\n");
2798         }
2799
2800         return ret;
2801 }
2802
2803 static void timehist_print_summary(struct perf_sched *sched,
2804                                    struct perf_session *session)
2805 {
2806         struct machine *m = &session->machines.host;
2807         struct total_run_stats totals;
2808         u64 task_count;
2809         struct thread *t;
2810         struct thread_runtime *r;
2811         int i;
2812         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2813
2814         memset(&totals, 0, sizeof(totals));
2815         totals.sched = sched;
2816
2817         if (sched->idle_hist) {
2818                 printf("\nIdle-time summary\n");
2819                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2820                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2821         } else if (sched->show_state) {
2822                 printf("\nWait-time summary\n");
2823                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2824                 printf("   run-time      sleep      iowait     preempt       delay\n");
2825         } else {
2826                 printf("\nRuntime summary\n");
2827                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2828                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2829         }
2830         printf("%*s            (count)  ", comm_width, "");
2831         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2832                sched->show_state ? "(msec)" : "%");
2833         printf("%.117s\n", graph_dotted_line);
2834
2835         machine__for_each_thread(m, show_thread_runtime, &totals);
2836         task_count = totals.task_count;
2837         if (!task_count)
2838                 printf("<no still running tasks>\n");
2839
2840         printf("\nTerminated tasks:\n");
2841         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2842         if (task_count == totals.task_count)
2843                 printf("<no terminated tasks>\n");
2844
2845         /* CPU idle stats not tracked when samples were skipped */
2846         if (sched->skipped_samples && !sched->idle_hist)
2847                 return;
2848
2849         printf("\nIdle stats:\n");
2850         for (i = 0; i < idle_max_cpu; ++i) {
2851                 t = idle_threads[i];
2852                 if (!t)
2853                         continue;
2854
2855                 r = thread__priv(t);
2856                 if (r && r->run_stats.n) {
2857                         totals.sched_count += r->run_stats.n;
2858                         printf("    CPU %2d idle for ", i);
2859                         print_sched_time(r->total_run_time, 6);
2860                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2861                 } else
2862                         printf("    CPU %2d idle entire time window\n", i);
2863         }
2864
2865         if (sched->idle_hist && sched->show_callchain) {
2866                 callchain_param.mode  = CHAIN_FOLDED;
2867                 callchain_param.value = CCVAL_PERIOD;
2868
2869                 callchain_register_param(&callchain_param);
2870
2871                 printf("\nIdle stats by callchain:\n");
2872                 for (i = 0; i < idle_max_cpu; ++i) {
2873                         struct idle_thread_runtime *itr;
2874
2875                         t = idle_threads[i];
2876                         if (!t)
2877                                 continue;
2878
2879                         itr = thread__priv(t);
2880                         if (itr == NULL)
2881                                 continue;
2882
2883                         callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2884                                              0, &callchain_param);
2885
2886                         printf("  CPU %2d:", i);
2887                         print_sched_time(itr->tr.total_run_time, 6);
2888                         printf(" msec\n");
2889                         timehist_print_idlehist_callchain(&itr->sorted_root);
2890                         printf("\n");
2891                 }
2892         }
2893
2894         printf("\n"
2895                "    Total number of unique tasks: %" PRIu64 "\n"
2896                "Total number of context switches: %" PRIu64 "\n",
2897                totals.task_count, totals.sched_count);
2898
2899         printf("           Total run time (msec): ");
2900         print_sched_time(totals.total_run_time, 2);
2901         printf("\n");
2902
2903         printf("    Total scheduling time (msec): ");
2904         print_sched_time(hist_time, 2);
2905         printf(" (x %d)\n", sched->max_cpu);
2906 }
2907
2908 typedef int (*sched_handler)(struct perf_tool *tool,
2909                           union perf_event *event,
2910                           struct evsel *evsel,
2911                           struct perf_sample *sample,
2912                           struct machine *machine);
2913
2914 static int perf_timehist__process_sample(struct perf_tool *tool,
2915                                          union perf_event *event,
2916                                          struct perf_sample *sample,
2917                                          struct evsel *evsel,
2918                                          struct machine *machine)
2919 {
2920         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2921         int err = 0;
2922         int this_cpu = sample->cpu;
2923
2924         if (this_cpu > sched->max_cpu)
2925                 sched->max_cpu = this_cpu;
2926
2927         if (evsel->handler != NULL) {
2928                 sched_handler f = evsel->handler;
2929
2930                 err = f(tool, event, evsel, sample, machine);
2931         }
2932
2933         return err;
2934 }
2935
2936 static int timehist_check_attr(struct perf_sched *sched,
2937                                struct evlist *evlist)
2938 {
2939         struct evsel *evsel;
2940         struct evsel_runtime *er;
2941
2942         list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2943                 er = evsel__get_runtime(evsel);
2944                 if (er == NULL) {
2945                         pr_err("Failed to allocate memory for evsel runtime data\n");
2946                         return -1;
2947                 }
2948
2949                 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2950                         pr_info("Samples do not have callchains.\n");
2951                         sched->show_callchain = 0;
2952                         symbol_conf.use_callchain = 0;
2953                 }
2954         }
2955
2956         return 0;
2957 }
2958
2959 static int perf_sched__timehist(struct perf_sched *sched)
2960 {
2961         const struct evsel_str_handler handlers[] = {
2962                 { "sched:sched_switch",       timehist_sched_switch_event, },
2963                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2964                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2965         };
2966         const struct evsel_str_handler migrate_handlers[] = {
2967                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2968         };
2969         struct perf_data data = {
2970                 .path  = input_name,
2971                 .mode  = PERF_DATA_MODE_READ,
2972                 .force = sched->force,
2973         };
2974
2975         struct perf_session *session;
2976         struct evlist *evlist;
2977         int err = -1;
2978
2979         /*
2980          * event handlers for timehist option
2981          */
2982         sched->tool.sample       = perf_timehist__process_sample;
2983         sched->tool.mmap         = perf_event__process_mmap;
2984         sched->tool.comm         = perf_event__process_comm;
2985         sched->tool.exit         = perf_event__process_exit;
2986         sched->tool.fork         = perf_event__process_fork;
2987         sched->tool.lost         = process_lost;
2988         sched->tool.attr         = perf_event__process_attr;
2989         sched->tool.tracing_data = perf_event__process_tracing_data;
2990         sched->tool.build_id     = perf_event__process_build_id;
2991
2992         sched->tool.ordered_events = true;
2993         sched->tool.ordering_requires_timestamps = true;
2994
2995         symbol_conf.use_callchain = sched->show_callchain;
2996
2997         session = perf_session__new(&data, false, &sched->tool);
2998         if (IS_ERR(session))
2999                 return PTR_ERR(session);
3000
3001         if (cpu_list) {
3002                 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3003                 if (err < 0)
3004                         goto out;
3005         }
3006
3007         evlist = session->evlist;
3008
3009         symbol__init(&session->header.env);
3010
3011         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3012                 pr_err("Invalid time string\n");
3013                 return -EINVAL;
3014         }
3015
3016         if (timehist_check_attr(sched, evlist) != 0)
3017                 goto out;
3018
3019         setup_pager();
3020
3021         /* setup per-evsel handlers */
3022         if (perf_session__set_tracepoints_handlers(session, handlers))
3023                 goto out;
3024
3025         /* sched_switch event at a minimum needs to exist */
3026         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3027                                                   "sched:sched_switch")) {
3028                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3029                 goto out;
3030         }
3031
3032         if (sched->show_migrations &&
3033             perf_session__set_tracepoints_handlers(session, migrate_handlers))
3034                 goto out;
3035
3036         /* pre-allocate struct for per-CPU idle stats */
3037         sched->max_cpu = session->header.env.nr_cpus_online;
3038         if (sched->max_cpu == 0)
3039                 sched->max_cpu = 4;
3040         if (init_idle_threads(sched->max_cpu))
3041                 goto out;
3042
3043         /* summary_only implies summary option, but don't overwrite summary if set */
3044         if (sched->summary_only)
3045                 sched->summary = sched->summary_only;
3046
3047         if (!sched->summary_only)
3048                 timehist_header(sched);
3049
3050         err = perf_session__process_events(session);
3051         if (err) {
3052                 pr_err("Failed to process events, error %d", err);
3053                 goto out;
3054         }
3055
3056         sched->nr_events      = evlist->stats.nr_events[0];
3057         sched->nr_lost_events = evlist->stats.total_lost;
3058         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3059
3060         if (sched->summary)
3061                 timehist_print_summary(sched, session);
3062
3063 out:
3064         free_idle_threads();
3065         perf_session__delete(session);
3066
3067         return err;
3068 }
3069
3070
3071 static void print_bad_events(struct perf_sched *sched)
3072 {
3073         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3074                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3075                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3076                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3077         }
3078         if (sched->nr_lost_events && sched->nr_events) {
3079                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3080                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3081                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3082         }
3083         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3084                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3085                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3086                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3087                 if (sched->nr_lost_events)
3088                         printf(" (due to lost events?)");
3089                 printf("\n");
3090         }
3091 }
3092
3093 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3094 {
3095         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3096         struct work_atoms *this;
3097         const char *comm = thread__comm_str(data->thread), *this_comm;
3098         bool leftmost = true;
3099
3100         while (*new) {
3101                 int cmp;
3102
3103                 this = container_of(*new, struct work_atoms, node);
3104                 parent = *new;
3105
3106                 this_comm = thread__comm_str(this->thread);
3107                 cmp = strcmp(comm, this_comm);
3108                 if (cmp > 0) {
3109                         new = &((*new)->rb_left);
3110                 } else if (cmp < 0) {
3111                         new = &((*new)->rb_right);
3112                         leftmost = false;
3113                 } else {
3114                         this->num_merged++;
3115                         this->total_runtime += data->total_runtime;
3116                         this->nb_atoms += data->nb_atoms;
3117                         this->total_lat += data->total_lat;
3118                         list_splice(&data->work_list, &this->work_list);
3119                         if (this->max_lat < data->max_lat) {
3120                                 this->max_lat = data->max_lat;
3121                                 this->max_lat_at = data->max_lat_at;
3122                         }
3123                         zfree(&data);
3124                         return;
3125                 }
3126         }
3127
3128         data->num_merged++;
3129         rb_link_node(&data->node, parent, new);
3130         rb_insert_color_cached(&data->node, root, leftmost);
3131 }
3132
3133 static void perf_sched__merge_lat(struct perf_sched *sched)
3134 {
3135         struct work_atoms *data;
3136         struct rb_node *node;
3137
3138         if (sched->skip_merge)
3139                 return;
3140
3141         while ((node = rb_first_cached(&sched->atom_root))) {
3142                 rb_erase_cached(node, &sched->atom_root);
3143                 data = rb_entry(node, struct work_atoms, node);
3144                 __merge_work_atoms(&sched->merged_atom_root, data);
3145         }
3146 }
3147
3148 static int perf_sched__lat(struct perf_sched *sched)
3149 {
3150         struct rb_node *next;
3151
3152         setup_pager();
3153
3154         if (perf_sched__read_events(sched))
3155                 return -1;
3156
3157         perf_sched__merge_lat(sched);
3158         perf_sched__sort_lat(sched);
3159
3160         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3161         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3162         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3163
3164         next = rb_first_cached(&sched->sorted_atom_root);
3165
3166         while (next) {
3167                 struct work_atoms *work_list;
3168
3169                 work_list = rb_entry(next, struct work_atoms, node);
3170                 output_lat_thread(sched, work_list);
3171                 next = rb_next(next);
3172                 thread__zput(work_list->thread);
3173         }
3174
3175         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3176         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3177                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3178
3179         printf(" ---------------------------------------------------\n");
3180
3181         print_bad_events(sched);
3182         printf("\n");
3183
3184         return 0;
3185 }
3186
3187 static int setup_map_cpus(struct perf_sched *sched)
3188 {
3189         struct perf_cpu_map *map;
3190
3191         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3192
3193         if (sched->map.comp) {
3194                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3195                 if (!sched->map.comp_cpus)
3196                         return -1;
3197         }
3198
3199         if (!sched->map.cpus_str)
3200                 return 0;
3201
3202         map = perf_cpu_map__new(sched->map.cpus_str);
3203         if (!map) {
3204                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3205                 return -1;
3206         }
3207
3208         sched->map.cpus = map;
3209         return 0;
3210 }
3211
3212 static int setup_color_pids(struct perf_sched *sched)
3213 {
3214         struct perf_thread_map *map;
3215
3216         if (!sched->map.color_pids_str)
3217                 return 0;
3218
3219         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3220         if (!map) {
3221                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3222                 return -1;
3223         }
3224
3225         sched->map.color_pids = map;
3226         return 0;
3227 }
3228
3229 static int setup_color_cpus(struct perf_sched *sched)
3230 {
3231         struct perf_cpu_map *map;
3232
3233         if (!sched->map.color_cpus_str)
3234                 return 0;
3235
3236         map = perf_cpu_map__new(sched->map.color_cpus_str);
3237         if (!map) {
3238                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3239                 return -1;
3240         }
3241
3242         sched->map.color_cpus = map;
3243         return 0;
3244 }
3245
3246 static int perf_sched__map(struct perf_sched *sched)
3247 {
3248         if (setup_map_cpus(sched))
3249                 return -1;
3250
3251         if (setup_color_pids(sched))
3252                 return -1;
3253
3254         if (setup_color_cpus(sched))
3255                 return -1;
3256
3257         setup_pager();
3258         if (perf_sched__read_events(sched))
3259                 return -1;
3260         print_bad_events(sched);
3261         return 0;
3262 }
3263
3264 static int perf_sched__replay(struct perf_sched *sched)
3265 {
3266         unsigned long i;
3267
3268         calibrate_run_measurement_overhead(sched);
3269         calibrate_sleep_measurement_overhead(sched);
3270
3271         test_calibrations(sched);
3272
3273         if (perf_sched__read_events(sched))
3274                 return -1;
3275
3276         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3277         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3278         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3279
3280         if (sched->targetless_wakeups)
3281                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3282         if (sched->multitarget_wakeups)
3283                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3284         if (sched->nr_run_events_optimized)
3285                 printf("run atoms optimized: %ld\n",
3286                         sched->nr_run_events_optimized);
3287
3288         print_task_traces(sched);
3289         add_cross_task_wakeups(sched);
3290
3291         create_tasks(sched);
3292         printf("------------------------------------------------------------\n");
3293         for (i = 0; i < sched->replay_repeat; i++)
3294                 run_one_test(sched);
3295
3296         return 0;
3297 }
3298
3299 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3300                           const char * const usage_msg[])
3301 {
3302         char *tmp, *tok, *str = strdup(sched->sort_order);
3303
3304         for (tok = strtok_r(str, ", ", &tmp);
3305                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3306                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3307                         usage_with_options_msg(usage_msg, options,
3308                                         "Unknown --sort key: `%s'", tok);
3309                 }
3310         }
3311
3312         free(str);
3313
3314         sort_dimension__add("pid", &sched->cmp_pid);
3315 }
3316
3317 static int __cmd_record(int argc, const char **argv)
3318 {
3319         unsigned int rec_argc, i, j;
3320         const char **rec_argv;
3321         const char * const record_args[] = {
3322                 "record",
3323                 "-a",
3324                 "-R",
3325                 "-m", "1024",
3326                 "-c", "1",
3327                 "-e", "sched:sched_switch",
3328                 "-e", "sched:sched_stat_wait",
3329                 "-e", "sched:sched_stat_sleep",
3330                 "-e", "sched:sched_stat_iowait",
3331                 "-e", "sched:sched_stat_runtime",
3332                 "-e", "sched:sched_process_fork",
3333                 "-e", "sched:sched_wakeup",
3334                 "-e", "sched:sched_wakeup_new",
3335                 "-e", "sched:sched_migrate_task",
3336         };
3337
3338         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3339         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3340
3341         if (rec_argv == NULL)
3342                 return -ENOMEM;
3343
3344         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3345                 rec_argv[i] = strdup(record_args[i]);
3346
3347         for (j = 1; j < (unsigned int)argc; j++, i++)
3348                 rec_argv[i] = argv[j];
3349
3350         BUG_ON(i != rec_argc);
3351
3352         return cmd_record(i, rec_argv);
3353 }
3354
3355 int cmd_sched(int argc, const char **argv)
3356 {
3357         static const char default_sort_order[] = "avg, max, switch, runtime";
3358         struct perf_sched sched = {
3359                 .tool = {
3360                         .sample          = perf_sched__process_tracepoint_sample,
3361                         .comm            = perf_sched__process_comm,
3362                         .namespaces      = perf_event__process_namespaces,
3363                         .lost            = perf_event__process_lost,
3364                         .fork            = perf_sched__process_fork_event,
3365                         .ordered_events = true,
3366                 },
3367                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3368                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3369                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3370                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3371                 .sort_order           = default_sort_order,
3372                 .replay_repeat        = 10,
3373                 .profile_cpu          = -1,
3374                 .next_shortname1      = 'A',
3375                 .next_shortname2      = '0',
3376                 .skip_merge           = 0,
3377                 .show_callchain       = 1,
3378                 .max_stack            = 5,
3379         };
3380         const struct option sched_options[] = {
3381         OPT_STRING('i', "input", &input_name, "file",
3382                     "input file name"),
3383         OPT_INCR('v', "verbose", &verbose,
3384                     "be more verbose (show symbol address, etc)"),
3385         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3386                     "dump raw trace in ASCII"),
3387         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3388         OPT_END()
3389         };
3390         const struct option latency_options[] = {
3391         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3392                    "sort by key(s): runtime, switch, avg, max"),
3393         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3394                     "CPU to profile on"),
3395         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3396                     "latency stats per pid instead of per comm"),
3397         OPT_PARENT(sched_options)
3398         };
3399         const struct option replay_options[] = {
3400         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3401                      "repeat the workload replay N times (-1: infinite)"),
3402         OPT_PARENT(sched_options)
3403         };
3404         const struct option map_options[] = {
3405         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3406                     "map output in compact mode"),
3407         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3408                    "highlight given pids in map"),
3409         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3410                     "highlight given CPUs in map"),
3411         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3412                     "display given CPUs in map"),
3413         OPT_PARENT(sched_options)
3414         };
3415         const struct option timehist_options[] = {
3416         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3417                    "file", "vmlinux pathname"),
3418         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3419                    "file", "kallsyms pathname"),
3420         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3421                     "Display call chains if present (default on)"),
3422         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3423                    "Maximum number of functions to display backtrace."),
3424         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3425                     "Look for files with symbols relative to this directory"),
3426         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3427                     "Show only syscall summary with statistics"),
3428         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3429                     "Show all syscalls and summary with statistics"),
3430         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3431         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3432         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3433         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3434         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3435         OPT_STRING(0, "time", &sched.time_str, "str",
3436                    "Time span for analysis (start,stop)"),
3437         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3438         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3439                    "analyze events only for given process id(s)"),
3440         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3441                    "analyze events only for given thread id(s)"),
3442         OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3443         OPT_PARENT(sched_options)
3444         };
3445
3446         const char * const latency_usage[] = {
3447                 "perf sched latency [<options>]",
3448                 NULL
3449         };
3450         const char * const replay_usage[] = {
3451                 "perf sched replay [<options>]",
3452                 NULL
3453         };
3454         const char * const map_usage[] = {
3455                 "perf sched map [<options>]",
3456                 NULL
3457         };
3458         const char * const timehist_usage[] = {
3459                 "perf sched timehist [<options>]",
3460                 NULL
3461         };
3462         const char *const sched_subcommands[] = { "record", "latency", "map",
3463                                                   "replay", "script",
3464                                                   "timehist", NULL };
3465         const char *sched_usage[] = {
3466                 NULL,
3467                 NULL
3468         };
3469         struct trace_sched_handler lat_ops  = {
3470                 .wakeup_event       = latency_wakeup_event,
3471                 .switch_event       = latency_switch_event,
3472                 .runtime_event      = latency_runtime_event,
3473                 .migrate_task_event = latency_migrate_task_event,
3474         };
3475         struct trace_sched_handler map_ops  = {
3476                 .switch_event       = map_switch_event,
3477         };
3478         struct trace_sched_handler replay_ops  = {
3479                 .wakeup_event       = replay_wakeup_event,
3480                 .switch_event       = replay_switch_event,
3481                 .fork_event         = replay_fork_event,
3482         };
3483         unsigned int i;
3484
3485         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3486                 sched.curr_pid[i] = -1;
3487
3488         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3489                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3490         if (!argc)
3491                 usage_with_options(sched_usage, sched_options);
3492
3493         /*
3494          * Aliased to 'perf script' for now:
3495          */
3496         if (!strcmp(argv[0], "script"))
3497                 return cmd_script(argc, argv);
3498
3499         if (!strncmp(argv[0], "rec", 3)) {
3500                 return __cmd_record(argc, argv);
3501         } else if (!strncmp(argv[0], "lat", 3)) {
3502                 sched.tp_handler = &lat_ops;
3503                 if (argc > 1) {
3504                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3505                         if (argc)
3506                                 usage_with_options(latency_usage, latency_options);
3507                 }
3508                 setup_sorting(&sched, latency_options, latency_usage);
3509                 return perf_sched__lat(&sched);
3510         } else if (!strcmp(argv[0], "map")) {
3511                 if (argc) {
3512                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3513                         if (argc)
3514                                 usage_with_options(map_usage, map_options);
3515                 }
3516                 sched.tp_handler = &map_ops;
3517                 setup_sorting(&sched, latency_options, latency_usage);
3518                 return perf_sched__map(&sched);
3519         } else if (!strncmp(argv[0], "rep", 3)) {
3520                 sched.tp_handler = &replay_ops;
3521                 if (argc) {
3522                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3523                         if (argc)
3524                                 usage_with_options(replay_usage, replay_options);
3525                 }
3526                 return perf_sched__replay(&sched);
3527         } else if (!strcmp(argv[0], "timehist")) {
3528                 if (argc) {
3529                         argc = parse_options(argc, argv, timehist_options,
3530                                              timehist_usage, 0);
3531                         if (argc)
3532                                 usage_with_options(timehist_usage, timehist_options);
3533                 }
3534                 if ((sched.show_wakeups || sched.show_next) &&
3535                     sched.summary_only) {
3536                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3537                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3538                         if (sched.show_wakeups)
3539                                 parse_options_usage(NULL, timehist_options, "w", true);
3540                         if (sched.show_next)
3541                                 parse_options_usage(NULL, timehist_options, "n", true);
3542                         return -EINVAL;
3543                 }
3544
3545                 return perf_sched__timehist(&sched);
3546         } else {
3547                 usage_with_options(sched_usage, sched_options);
3548         }
3549
3550         return 0;
3551 }