tools/perf/builtin-sched.c

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