include/trace/events/sched.h

   1 #undef TRACE_SYSTEM
   2 #define TRACE_SYSTEM sched
   3
   4 #if !defined(_TRACE_SCHED_H) || defined(TRACE_HEADER_MULTI_READ)
   5 #define _TRACE_SCHED_H
   6
   7 #include <linux/sched.h>
   8 #include <linux/tracepoint.h>
   9 #include <linux/binfmts.h>
  10
  11 /*
  12  * Tracepoint for calling kthread_stop, performed to end a kthread:
  13  */
  14 TRACE_EVENT(sched_kthread_stop,
  15
  16         TP_PROTO(struct task_struct *t),
  17
  18         TP_ARGS(t),
  19
  20         TP_STRUCT__entry(
  21                 __array(        char,   comm,   TASK_COMM_LEN   )
  22                 __field(        pid_t,  pid                     )
  23         ),
  24
  25         TP_fast_assign(
  26                 memcpy(__entry->comm, t->comm, TASK_COMM_LEN);
  27                 __entry->pid    = t->pid;
  28         ),
  29
  30         TP_printk("comm=%s pid=%d", __entry->comm, __entry->pid)
  31 );
  32
  33 /*
  34  * Tracepoint for the return value of the kthread stopping:
  35  */
  36 TRACE_EVENT(sched_kthread_stop_ret,
  37
  38         TP_PROTO(int ret),
  39
  40         TP_ARGS(ret),
  41
  42         TP_STRUCT__entry(
  43                 __field(        int,    ret     )
  44         ),
  45
  46         TP_fast_assign(
  47                 __entry->ret    = ret;
  48         ),
  49
  50         TP_printk("ret=%d", __entry->ret)
  51 );
  52
  53 /*
  54  * Tracepoint for waking up a task:
  55  */
  56 DECLARE_EVENT_CLASS(sched_wakeup_template,
  57
  58         TP_PROTO(struct task_struct *p),
  59
  60         TP_ARGS(__perf_task(p)),
  61
  62         TP_STRUCT__entry(
  63                 __array(        char,   comm,   TASK_COMM_LEN   )
  64                 __field(        pid_t,  pid                     )
  65                 __field(        int,    prio                    )
  66                 __field(        int,    success                 )
  67                 __field(        int,    target_cpu              )
  68         ),
  69
  70         TP_fast_assign(
  71                 memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
  72                 __entry->pid            = p->pid;
  73                 __entry->prio           = p->prio;
  74                 __entry->success        = 1; /* rudiment, kill when possible */
  75                 __entry->target_cpu     = task_cpu(p);
  76         ),
  77
  78         TP_printk("comm=%s pid=%d prio=%d target_cpu=%03d",
  79                   __entry->comm, __entry->pid, __entry->prio,
  80                   __entry->target_cpu)
  81 );
  82
  83 /*
  84  * Tracepoint called when waking a task; this tracepoint is guaranteed to be
  85  * called from the waking context.
  86  */
  87 DEFINE_EVENT(sched_wakeup_template, sched_waking,
  88              TP_PROTO(struct task_struct *p),
  89              TP_ARGS(p));
  90
  91 /*
  92  * Tracepoint called when the task is actually woken; p->state == TASK_RUNNNG.
  93  * It it not always called from the waking context.
  94  */
  95 DEFINE_EVENT(sched_wakeup_template, sched_wakeup,
  96              TP_PROTO(struct task_struct *p),
  97              TP_ARGS(p));
  98
  99 /*
 100  * Tracepoint for waking up a new task:
 101  */
 102 DEFINE_EVENT(sched_wakeup_template, sched_wakeup_new,
 103              TP_PROTO(struct task_struct *p),
 104              TP_ARGS(p));
 105
 106 #ifdef CREATE_TRACE_POINTS
 107 static inline long __trace_sched_switch_state(bool preempt, struct task_struct *p)
 108 {
 109 #ifdef CONFIG_SCHED_DEBUG
 110         BUG_ON(p != current);
 111 #endif /* CONFIG_SCHED_DEBUG */
 112
 113         /*
 114          * Preemption ignores task state, therefore preempted tasks are always
 115          * RUNNING (we will not have dequeued if state != RUNNING).
 116          */
 117         return preempt ? TASK_RUNNING | TASK_STATE_MAX : p->state;
 118 }
 119 #endif /* CREATE_TRACE_POINTS */
 120
 121 /*
 122  * Tracepoint for task switches, performed by the scheduler:
 123  */
 124 TRACE_EVENT(sched_switch,
 125
 126         TP_PROTO(bool preempt,
 127                  struct task_struct *prev,
 128                  struct task_struct *next),
 129
 130         TP_ARGS(preempt, prev, next),
 131
 132         TP_STRUCT__entry(
 133                 __array(        char,   prev_comm,      TASK_COMM_LEN   )
 134                 __field(        pid_t,  prev_pid                        )
 135                 __field(        int,    prev_prio                       )
 136                 __field(        long,   prev_state                      )
 137                 __array(        char,   next_comm,      TASK_COMM_LEN   )
 138                 __field(        pid_t,  next_pid                        )
 139                 __field(        int,    next_prio                       )
 140         ),
 141
 142         TP_fast_assign(
 143                 memcpy(__entry->next_comm, next->comm, TASK_COMM_LEN);
 144                 __entry->prev_pid       = prev->pid;
 145                 __entry->prev_prio      = prev->prio;
 146                 __entry->prev_state     = __trace_sched_switch_state(preempt, prev);
 147                 memcpy(__entry->prev_comm, prev->comm, TASK_COMM_LEN);
 148                 __entry->next_pid       = next->pid;
 149                 __entry->next_prio      = next->prio;
 150         ),
 151
 152         TP_printk("prev_comm=%s prev_pid=%d prev_prio=%d prev_state=%s%s ==> next_comm=%s next_pid=%d next_prio=%d",
 153                 __entry->prev_comm, __entry->prev_pid, __entry->prev_prio,
 154                 __entry->prev_state & (TASK_STATE_MAX-1) ?
 155                   __print_flags(__entry->prev_state & (TASK_STATE_MAX-1), "|",
 156                                 { 1, "S"} , { 2, "D" }, { 4, "T" }, { 8, "t" },
 157                                 { 16, "Z" }, { 32, "X" }, { 64, "x" },
 158                                 { 128, "K" }, { 256, "W" }, { 512, "P" },
 159                                 { 1024, "N" }) : "R",
 160                 __entry->prev_state & TASK_STATE_MAX ? "+" : "",
 161                 __entry->next_comm, __entry->next_pid, __entry->next_prio)
 162 );
 163
 164 /*
 165  * Tracepoint for a task being migrated:
 166  */
 167 TRACE_EVENT(sched_migrate_task,
 168
 169         TP_PROTO(struct task_struct *p, int dest_cpu),
 170
 171         TP_ARGS(p, dest_cpu),
 172
 173         TP_STRUCT__entry(
 174                 __array(        char,   comm,   TASK_COMM_LEN   )
 175                 __field(        pid_t,  pid                     )
 176                 __field(        int,    prio                    )
 177                 __field(        int,    orig_cpu                )
 178                 __field(        int,    dest_cpu                )
 179         ),
 180
 181         TP_fast_assign(
 182                 memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
 183                 __entry->pid            = p->pid;
 184                 __entry->prio           = p->prio;
 185                 __entry->orig_cpu       = task_cpu(p);
 186                 __entry->dest_cpu       = dest_cpu;
 187         ),
 188
 189         TP_printk("comm=%s pid=%d prio=%d orig_cpu=%d dest_cpu=%d",
 190                   __entry->comm, __entry->pid, __entry->prio,
 191                   __entry->orig_cpu, __entry->dest_cpu)
 192 );
 193
 194 DECLARE_EVENT_CLASS(sched_process_template,
 195
 196         TP_PROTO(struct task_struct *p),
 197
 198         TP_ARGS(p),
 199
 200         TP_STRUCT__entry(
 201                 __array(        char,   comm,   TASK_COMM_LEN   )
 202                 __field(        pid_t,  pid                     )
 203                 __field(        int,    prio                    )
 204         ),
 205
 206         TP_fast_assign(
 207                 memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
 208                 __entry->pid            = p->pid;
 209                 __entry->prio           = p->prio;
 210         ),
 211
 212         TP_printk("comm=%s pid=%d prio=%d",
 213                   __entry->comm, __entry->pid, __entry->prio)
 214 );
 215
 216 /*
 217  * Tracepoint for freeing a task:
 218  */
 219 DEFINE_EVENT(sched_process_template, sched_process_free,
 220              TP_PROTO(struct task_struct *p),
 221              TP_ARGS(p));
 222
 223
 224 /*
 225  * Tracepoint for a task exiting:
 226  */
 227 DEFINE_EVENT(sched_process_template, sched_process_exit,
 228              TP_PROTO(struct task_struct *p),
 229              TP_ARGS(p));
 230
 231 /*
 232  * Tracepoint for waiting on task to unschedule:
 233  */
 234 DEFINE_EVENT(sched_process_template, sched_wait_task,
 235         TP_PROTO(struct task_struct *p),
 236         TP_ARGS(p));
 237
 238 /*
 239  * Tracepoint for a waiting task:
 240  */
 241 TRACE_EVENT(sched_process_wait,
 242
 243         TP_PROTO(struct pid *pid),
 244
 245         TP_ARGS(pid),
 246
 247         TP_STRUCT__entry(
 248                 __array(        char,   comm,   TASK_COMM_LEN   )
 249                 __field(        pid_t,  pid                     )
 250                 __field(        int,    prio                    )
 251         ),
 252
 253         TP_fast_assign(
 254                 memcpy(__entry->comm, current->comm, TASK_COMM_LEN);
 255                 __entry->pid            = pid_nr(pid);
 256                 __entry->prio           = current->prio;
 257         ),
 258
 259         TP_printk("comm=%s pid=%d prio=%d",
 260                   __entry->comm, __entry->pid, __entry->prio)
 261 );
 262
 263 /*
 264  * Tracepoint for do_fork:
 265  */
 266 TRACE_EVENT(sched_process_fork,
 267
 268         TP_PROTO(struct task_struct *parent, struct task_struct *child),
 269
 270         TP_ARGS(parent, child),
 271
 272         TP_STRUCT__entry(
 273                 __array(        char,   parent_comm,    TASK_COMM_LEN   )
 274                 __field(        pid_t,  parent_pid                      )
 275                 __array(        char,   child_comm,     TASK_COMM_LEN   )
 276                 __field(        pid_t,  child_pid                       )
 277         ),
 278
 279         TP_fast_assign(
 280                 memcpy(__entry->parent_comm, parent->comm, TASK_COMM_LEN);
 281                 __entry->parent_pid     = parent->pid;
 282                 memcpy(__entry->child_comm, child->comm, TASK_COMM_LEN);
 283                 __entry->child_pid      = child->pid;
 284         ),
 285
 286         TP_printk("comm=%s pid=%d child_comm=%s child_pid=%d",
 287                 __entry->parent_comm, __entry->parent_pid,
 288                 __entry->child_comm, __entry->child_pid)
 289 );
 290
 291 /*
 292  * Tracepoint for exec:
 293  */
 294 TRACE_EVENT(sched_process_exec,
 295
 296         TP_PROTO(struct task_struct *p, pid_t old_pid,
 297                  struct linux_binprm *bprm),
 298
 299         TP_ARGS(p, old_pid, bprm),
 300
 301         TP_STRUCT__entry(
 302                 __string(       filename,       bprm->filename  )
 303                 __field(        pid_t,          pid             )
 304                 __field(        pid_t,          old_pid         )
 305         ),
 306
 307         TP_fast_assign(
 308                 __assign_str(filename, bprm->filename);
 309                 __entry->pid            = p->pid;
 310                 __entry->old_pid        = old_pid;
 311         ),
 312
 313         TP_printk("filename=%s pid=%d old_pid=%d", __get_str(filename),
 314                   __entry->pid, __entry->old_pid)
 315 );
 316
 317 /*
 318  * XXX the below sched_stat tracepoints only apply to SCHED_OTHER/BATCH/IDLE
 319  *     adding sched_stat support to SCHED_FIFO/RR would be welcome.
 320  */
 321 DECLARE_EVENT_CLASS(sched_stat_template,
 322
 323         TP_PROTO(struct task_struct *tsk, u64 delay),
 324
 325         TP_ARGS(__perf_task(tsk), __perf_count(delay)),
 326
 327         TP_STRUCT__entry(
 328                 __array( char,  comm,   TASK_COMM_LEN   )
 329                 __field( pid_t, pid                     )
 330                 __field( u64,   delay                   )
 331         ),
 332
 333         TP_fast_assign(
 334                 memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
 335                 __entry->pid    = tsk->pid;
 336                 __entry->delay  = delay;
 337         ),
 338
 339         TP_printk("comm=%s pid=%d delay=%Lu [ns]",
 340                         __entry->comm, __entry->pid,
 341                         (unsigned long long)__entry->delay)
 342 );
 343
 344
 345 /*
 346  * Tracepoint for accounting wait time (time the task is runnable
 347  * but not actually running due to scheduler contention).
 348  */
 349 DEFINE_EVENT(sched_stat_template, sched_stat_wait,
 350              TP_PROTO(struct task_struct *tsk, u64 delay),
 351              TP_ARGS(tsk, delay));
 352
 353 /*
 354  * Tracepoint for accounting sleep time (time the task is not runnable,
 355  * including iowait, see below).
 356  */
 357 DEFINE_EVENT(sched_stat_template, sched_stat_sleep,
 358              TP_PROTO(struct task_struct *tsk, u64 delay),
 359              TP_ARGS(tsk, delay));
 360
 361 /*
 362  * Tracepoint for accounting iowait time (time the task is not runnable
 363  * due to waiting on IO to complete).
 364  */
 365 DEFINE_EVENT(sched_stat_template, sched_stat_iowait,
 366              TP_PROTO(struct task_struct *tsk, u64 delay),
 367              TP_ARGS(tsk, delay));
 368
 369 /*
 370  * Tracepoint for accounting blocked time (time the task is in uninterruptible).
 371  */
 372 DEFINE_EVENT(sched_stat_template, sched_stat_blocked,
 373              TP_PROTO(struct task_struct *tsk, u64 delay),
 374              TP_ARGS(tsk, delay));
 375
 376 /*
 377  * Tracepoint for recording the cause of uninterruptible sleep.
 378  */
 379 TRACE_EVENT(sched_blocked_reason,
 380
 381         TP_PROTO(struct task_struct *tsk),
 382
 383         TP_ARGS(tsk),
 384
 385         TP_STRUCT__entry(
 386                 __field( pid_t, pid     )
 387                 __field( void*, caller  )
 388                 __field( bool, io_wait  )
 389         ),
 390
 391         TP_fast_assign(
 392                 __entry->pid    = tsk->pid;
 393                 __entry->caller = (void*)get_wchan(tsk);
 394                 __entry->io_wait = tsk->in_iowait;
 395         ),
 396
 397         TP_printk("pid=%d iowait=%d caller=%pS", __entry->pid, __entry->io_wait, __entry->caller)
 398 );
 399
 400 /*
 401  * Tracepoint for accounting runtime (time the task is executing
 402  * on a CPU).
 403  */
 404 DECLARE_EVENT_CLASS(sched_stat_runtime,
 405
 406         TP_PROTO(struct task_struct *tsk, u64 runtime, u64 vruntime),
 407
 408         TP_ARGS(tsk, __perf_count(runtime), vruntime),
 409
 410         TP_STRUCT__entry(
 411                 __array( char,  comm,   TASK_COMM_LEN   )
 412                 __field( pid_t, pid                     )
 413                 __field( u64,   runtime                 )
 414                 __field( u64,   vruntime                        )
 415         ),
 416
 417         TP_fast_assign(
 418                 memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
 419                 __entry->pid            = tsk->pid;
 420                 __entry->runtime        = runtime;
 421                 __entry->vruntime       = vruntime;
 422         ),
 423
 424         TP_printk("comm=%s pid=%d runtime=%Lu [ns] vruntime=%Lu [ns]",
 425                         __entry->comm, __entry->pid,
 426                         (unsigned long long)__entry->runtime,
 427                         (unsigned long long)__entry->vruntime)
 428 );
 429
 430 DEFINE_EVENT(sched_stat_runtime, sched_stat_runtime,
 431              TP_PROTO(struct task_struct *tsk, u64 runtime, u64 vruntime),
 432              TP_ARGS(tsk, runtime, vruntime));
 433
 434 /*
 435  * Tracepoint for showing priority inheritance modifying a tasks
 436  * priority.
 437  */
 438 TRACE_EVENT(sched_pi_setprio,
 439
 440         TP_PROTO(struct task_struct *tsk, int newprio),
 441
 442         TP_ARGS(tsk, newprio),
 443
 444         TP_STRUCT__entry(
 445                 __array( char,  comm,   TASK_COMM_LEN   )
 446                 __field( pid_t, pid                     )
 447                 __field( int,   oldprio                 )
 448                 __field( int,   newprio                 )
 449         ),
 450
 451         TP_fast_assign(
 452                 memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
 453                 __entry->pid            = tsk->pid;
 454                 __entry->oldprio        = tsk->prio;
 455                 __entry->newprio        = newprio;
 456         ),
 457
 458         TP_printk("comm=%s pid=%d oldprio=%d newprio=%d",
 459                         __entry->comm, __entry->pid,
 460                         __entry->oldprio, __entry->newprio)
 461 );
 462
 463 #ifdef CONFIG_DETECT_HUNG_TASK
 464 TRACE_EVENT(sched_process_hang,
 465         TP_PROTO(struct task_struct *tsk),
 466         TP_ARGS(tsk),
 467
 468         TP_STRUCT__entry(
 469                 __array( char,  comm,   TASK_COMM_LEN   )
 470                 __field( pid_t, pid                     )
 471         ),
 472
 473         TP_fast_assign(
 474                 memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
 475                 __entry->pid = tsk->pid;
 476         ),
 477
 478         TP_printk("comm=%s pid=%d", __entry->comm, __entry->pid)
 479 );
 480 #endif /* CONFIG_DETECT_HUNG_TASK */
 481
 482 DECLARE_EVENT_CLASS(sched_move_task_template,
 483
 484         TP_PROTO(struct task_struct *tsk, int src_cpu, int dst_cpu),
 485
 486         TP_ARGS(tsk, src_cpu, dst_cpu),
 487
 488         TP_STRUCT__entry(
 489                 __field( pid_t, pid                     )
 490                 __field( pid_t, tgid                    )
 491                 __field( pid_t, ngid                    )
 492                 __field( int,   src_cpu                 )
 493                 __field( int,   src_nid                 )
 494                 __field( int,   dst_cpu                 )
 495                 __field( int,   dst_nid                 )
 496         ),
 497
 498         TP_fast_assign(
 499                 __entry->pid            = task_pid_nr(tsk);
 500                 __entry->tgid           = task_tgid_nr(tsk);
 501                 __entry->ngid           = task_numa_group_id(tsk);
 502                 __entry->src_cpu        = src_cpu;
 503                 __entry->src_nid        = cpu_to_node(src_cpu);
 504                 __entry->dst_cpu        = dst_cpu;
 505                 __entry->dst_nid        = cpu_to_node(dst_cpu);
 506         ),
 507
 508         TP_printk("pid=%d tgid=%d ngid=%d src_cpu=%d src_nid=%d dst_cpu=%d dst_nid=%d",
 509                         __entry->pid, __entry->tgid, __entry->ngid,
 510                         __entry->src_cpu, __entry->src_nid,
 511                         __entry->dst_cpu, __entry->dst_nid)
 512 );
 513
 514 /*
 515  * Tracks migration of tasks from one runqueue to another. Can be used to
 516  * detect if automatic NUMA balancing is bouncing between nodes
 517  */
 518 DEFINE_EVENT(sched_move_task_template, sched_move_numa,
 519         TP_PROTO(struct task_struct *tsk, int src_cpu, int dst_cpu),
 520
 521         TP_ARGS(tsk, src_cpu, dst_cpu)
 522 );
 523
 524 DEFINE_EVENT(sched_move_task_template, sched_stick_numa,
 525         TP_PROTO(struct task_struct *tsk, int src_cpu, int dst_cpu),
 526
 527         TP_ARGS(tsk, src_cpu, dst_cpu)
 528 );
 529
 530 TRACE_EVENT(sched_swap_numa,
 531
 532         TP_PROTO(struct task_struct *src_tsk, int src_cpu,
 533                  struct task_struct *dst_tsk, int dst_cpu),
 534
 535         TP_ARGS(src_tsk, src_cpu, dst_tsk, dst_cpu),
 536
 537         TP_STRUCT__entry(
 538                 __field( pid_t, src_pid                 )
 539                 __field( pid_t, src_tgid                )
 540                 __field( pid_t, src_ngid                )
 541                 __field( int,   src_cpu                 )
 542                 __field( int,   src_nid                 )
 543                 __field( pid_t, dst_pid                 )
 544                 __field( pid_t, dst_tgid                )
 545                 __field( pid_t, dst_ngid                )
 546                 __field( int,   dst_cpu                 )
 547                 __field( int,   dst_nid                 )
 548         ),
 549
 550         TP_fast_assign(
 551                 __entry->src_pid        = task_pid_nr(src_tsk);
 552                 __entry->src_tgid       = task_tgid_nr(src_tsk);
 553                 __entry->src_ngid       = task_numa_group_id(src_tsk);
 554                 __entry->src_cpu        = src_cpu;
 555                 __entry->src_nid        = cpu_to_node(src_cpu);
 556                 __entry->dst_pid        = task_pid_nr(dst_tsk);
 557                 __entry->dst_tgid       = task_tgid_nr(dst_tsk);
 558                 __entry->dst_ngid       = task_numa_group_id(dst_tsk);
 559                 __entry->dst_cpu        = dst_cpu;
 560                 __entry->dst_nid        = cpu_to_node(dst_cpu);
 561         ),
 562
 563         TP_printk("src_pid=%d src_tgid=%d src_ngid=%d src_cpu=%d src_nid=%d dst_pid=%d dst_tgid=%d dst_ngid=%d dst_cpu=%d dst_nid=%d",
 564                         __entry->src_pid, __entry->src_tgid, __entry->src_ngid,
 565                         __entry->src_cpu, __entry->src_nid,
 566                         __entry->dst_pid, __entry->dst_tgid, __entry->dst_ngid,
 567                         __entry->dst_cpu, __entry->dst_nid)
 568 );
 569
 570 /*
 571  * Tracepoint for waking a polling cpu without an IPI.
 572  */
 573 TRACE_EVENT(sched_wake_idle_without_ipi,
 574
 575         TP_PROTO(int cpu),
 576
 577         TP_ARGS(cpu),
 578
 579         TP_STRUCT__entry(
 580                 __field(        int,    cpu     )
 581         ),
 582
 583         TP_fast_assign(
 584                 __entry->cpu    = cpu;
 585         ),
 586
 587         TP_printk("cpu=%d", __entry->cpu)
 588 );
 589 #endif /* _TRACE_SCHED_H */
 590
 591 /* This part must be outside protection */
 592 #include <trace/define_trace.h>