instrumentation/events/lttng-module/sched.h

   1 #undef TRACE_SYSTEM
   2 #define TRACE_SYSTEM sched
   3
   4 #if !defined(LTTNG_TRACE_SCHED_H) || defined(TRACE_HEADER_MULTI_READ)
   5 #define LTTNG_TRACE_SCHED_H
   6
   7 #include <probes/lttng-tracepoint-event.h>
   8 #include <linux/sched.h>
   9 #include <linux/pid_namespace.h>
  10 #include <linux/binfmts.h>
  11 #include <linux/version.h>
  12 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,9,0))
  13 #include <linux/sched/rt.h>
  14 #endif
  15
  16 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,19,0))
  17 #define lttng_proc_inum ns.inum
  18 #else
  19 #define lttng_proc_inum proc_inum
  20 #endif
  21
  22 #define LTTNG_MAX_PID_NS_LEVEL 32
  23
  24 #ifndef _TRACE_SCHED_DEF_
  25 #define _TRACE_SCHED_DEF_
  26
  27 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,4,0))
  28
  29 static inline long __trace_sched_switch_state(bool preempt, struct task_struct *p)
  30 {
  31 #ifdef CONFIG_SCHED_DEBUG
  32         BUG_ON(p != current);
  33 #endif /* CONFIG_SCHED_DEBUG */
  34         /*
  35          * Preemption ignores task state, therefore preempted tasks are always RUNNING
  36          * (we will not have dequeued if state != RUNNING).
  37          */
  38         return preempt ? TASK_RUNNING | TASK_STATE_MAX : p->state;
  39 }
  40
  41 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(3,19,0))
  42
  43 static inline long __trace_sched_switch_state(struct task_struct *p)
  44 {
  45         long state = p->state;
  46
  47 #ifdef CONFIG_PREEMPT
  48 #ifdef CONFIG_SCHED_DEBUG
  49         BUG_ON(p != current);
  50 #endif /* CONFIG_SCHED_DEBUG */
  51         /*
  52          * For all intents and purposes a preempted task is a running task.
  53          */
  54         if (preempt_count() & PREEMPT_ACTIVE)
  55                 state = TASK_RUNNING | TASK_STATE_MAX;
  56 #endif  /* CONFIG_PREEMPT */
  57
  58         return state;
  59 }
  60
  61 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(3,13,0))
  62
  63 static inline long __trace_sched_switch_state(struct task_struct *p)
  64 {
  65         long state = p->state;
  66
  67 #ifdef CONFIG_PREEMPT
  68         /*
  69          * For all intents and purposes a preempted task is a running task.
  70          */
  71         if (task_preempt_count(p) & PREEMPT_ACTIVE)
  72                 state = TASK_RUNNING | TASK_STATE_MAX;
  73 #endif
  74
  75         return state;
  76 }
  77
  78 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(3,2,0))
  79
  80 static inline long __trace_sched_switch_state(struct task_struct *p)
  81 {
  82         long state = p->state;
  83
  84 #ifdef CONFIG_PREEMPT
  85         /*
  86          * For all intents and purposes a preempted task is a running task.
  87          */
  88         if (task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)
  89                 state = TASK_RUNNING | TASK_STATE_MAX;
  90 #endif
  91
  92         return state;
  93 }
  94
  95 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
  96
  97 static inline long __trace_sched_switch_state(struct task_struct *p)
  98 {
  99         long state = p->state;
 100
 101 #ifdef CONFIG_PREEMPT
 102         /*
 103          * For all intents and purposes a preempted task is a running task.
 104          */
 105         if (task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)
 106                 state = TASK_RUNNING;
 107 #endif
 108
 109         return state;
 110 }
 111
 112 #endif
 113
 114 #endif /* _TRACE_SCHED_DEF_ */
 115
 116 /*
 117  * Tracepoint for calling kthread_stop, performed to end a kthread:
 118  */
 119 LTTNG_TRACEPOINT_EVENT(sched_kthread_stop,
 120
 121         TP_PROTO(struct task_struct *t),
 122
 123         TP_ARGS(t),
 124
 125         TP_FIELDS(
 126                 ctf_array_text(char, comm, t->comm, TASK_COMM_LEN)
 127                 ctf_integer(pid_t, tid, t->pid)
 128         )
 129 )
 130
 131 /*
 132  * Tracepoint for the return value of the kthread stopping:
 133  */
 134 LTTNG_TRACEPOINT_EVENT(sched_kthread_stop_ret,
 135
 136         TP_PROTO(int ret),
 137
 138         TP_ARGS(ret),
 139
 140         TP_FIELDS(
 141                 ctf_integer(int, ret, ret)
 142         )
 143 )
 144
 145 /*
 146  * Tracepoint for waking up a task:
 147  */
 148 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,3,0) || \
 149         LTTNG_RT_KERNEL_RANGE(4,1,10,11, 4,2,0,0))
 150 LTTNG_TRACEPOINT_EVENT_CLASS(sched_wakeup_template,
 151
 152         TP_PROTO(struct task_struct *p),
 153
 154         TP_ARGS(p),
 155
 156         TP_FIELDS(
 157                 ctf_array_text(char, comm, p->comm, TASK_COMM_LEN)
 158                 ctf_integer(pid_t, tid, p->pid)
 159                 ctf_integer(int, prio, p->prio - MAX_RT_PRIO)
 160                 ctf_integer(int, target_cpu, task_cpu(p))
 161         )
 162 )
 163 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,3,0)) */
 164 LTTNG_TRACEPOINT_EVENT_CLASS(sched_wakeup_template,
 165
 166 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
 167         TP_PROTO(struct task_struct *p, int success),
 168
 169         TP_ARGS(p, success),
 170 #else
 171         TP_PROTO(struct rq *rq, struct task_struct *p, int success),
 172
 173         TP_ARGS(rq, p, success),
 174 #endif
 175
 176         TP_FIELDS(
 177                 ctf_array_text(char, comm, p->comm, TASK_COMM_LEN)
 178                 ctf_integer(pid_t, tid, p->pid)
 179                 ctf_integer(int, prio, p->prio - MAX_RT_PRIO)
 180                 ctf_integer(int, success, success)
 181 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,32))
 182                 ctf_integer(int, target_cpu, task_cpu(p))
 183 #endif
 184         )
 185 )
 186 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,3,0)) */
 187
 188 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,3,0) || \
 189         LTTNG_RT_KERNEL_RANGE(4,1,10,11, 4,2,0,0))
 190
 191 /*
 192  * Tracepoint called when waking a task; this tracepoint is guaranteed to be
 193  * called from the waking context.
 194  */
 195 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_waking,
 196              TP_PROTO(struct task_struct *p),
 197              TP_ARGS(p))
 198
 199 /*
 200  * Tracepoint called when the task is actually woken; p->state == TASK_RUNNNG.
 201  * It it not always called from the waking context.
 202  */
 203 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup,
 204              TP_PROTO(struct task_struct *p),
 205              TP_ARGS(p))
 206
 207 /*
 208  * Tracepoint for waking up a new task:
 209  */
 210 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup_new,
 211              TP_PROTO(struct task_struct *p),
 212              TP_ARGS(p))
 213
 214 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
 215
 216 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup,
 217              TP_PROTO(struct task_struct *p, int success),
 218              TP_ARGS(p, success))
 219
 220 /*
 221  * Tracepoint for waking up a new task:
 222  */
 223 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup_new,
 224              TP_PROTO(struct task_struct *p, int success),
 225              TP_ARGS(p, success))
 226
 227 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 228
 229 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup,
 230              TP_PROTO(struct rq *rq, struct task_struct *p, int success),
 231              TP_ARGS(rq, p, success))
 232
 233 /*
 234  * Tracepoint for waking up a new task:
 235  */
 236 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_wakeup_template, sched_wakeup_new,
 237              TP_PROTO(struct rq *rq, struct task_struct *p, int success),
 238              TP_ARGS(rq, p, success))
 239
 240 #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 241
 242 /*
 243  * Tracepoint for task switches, performed by the scheduler:
 244  */
 245 LTTNG_TRACEPOINT_EVENT(sched_switch,
 246
 247 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,4,0))
 248         TP_PROTO(bool preempt,
 249                  struct task_struct *prev,
 250                  struct task_struct *next),
 251
 252         TP_ARGS(preempt, prev, next),
 253 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
 254         TP_PROTO(struct task_struct *prev,
 255                  struct task_struct *next),
 256
 257         TP_ARGS(prev, next),
 258 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 259         TP_PROTO(struct rq *rq, struct task_struct *prev,
 260                  struct task_struct *next),
 261
 262         TP_ARGS(rq, prev, next),
 263 #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 264
 265         TP_FIELDS(
 266                 ctf_array_text(char, prev_comm, prev->comm, TASK_COMM_LEN)
 267                 ctf_integer(pid_t, prev_tid, prev->pid)
 268                 ctf_integer(int, prev_prio, prev->prio - MAX_RT_PRIO)
 269 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,4,0))
 270                 ctf_integer(long, prev_state, __trace_sched_switch_state(preempt, prev))
 271 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
 272                 ctf_integer(long, prev_state, __trace_sched_switch_state(prev))
 273 #else
 274                 ctf_integer(long, prev_state, prev->state)
 275 #endif
 276                 ctf_array_text(char, next_comm, next->comm, TASK_COMM_LEN)
 277                 ctf_integer(pid_t, next_tid, next->pid)
 278                 ctf_integer(int, next_prio, next->prio - MAX_RT_PRIO)
 279         )
 280 )
 281
 282 /*
 283  * Tracepoint for a task being migrated:
 284  */
 285 LTTNG_TRACEPOINT_EVENT(sched_migrate_task,
 286
 287         TP_PROTO(struct task_struct *p, int dest_cpu),
 288
 289         TP_ARGS(p, dest_cpu),
 290
 291         TP_FIELDS(
 292                 ctf_array_text(char, comm, p->comm, TASK_COMM_LEN)
 293                 ctf_integer(pid_t, tid, p->pid)
 294                 ctf_integer(int, prio, p->prio - MAX_RT_PRIO)
 295                 ctf_integer(int, orig_cpu, task_cpu(p))
 296                 ctf_integer(int, dest_cpu, dest_cpu)
 297         )
 298 )
 299
 300 LTTNG_TRACEPOINT_EVENT_CLASS(sched_process_template,
 301
 302         TP_PROTO(struct task_struct *p),
 303
 304         TP_ARGS(p),
 305
 306         TP_FIELDS(
 307                 ctf_array_text(char, comm, p->comm, TASK_COMM_LEN)
 308                 ctf_integer(pid_t, tid, p->pid)
 309                 ctf_integer(int, prio, p->prio - MAX_RT_PRIO)
 310         )
 311 )
 312
 313 /*
 314  * Tracepoint for freeing a task:
 315  */
 316 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_process_template, sched_process_free,
 317              TP_PROTO(struct task_struct *p),
 318              TP_ARGS(p))
 319
 320
 321 /*
 322  * Tracepoint for a task exiting:
 323  */
 324 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_process_template, sched_process_exit,
 325              TP_PROTO(struct task_struct *p),
 326              TP_ARGS(p))
 327
 328 /*
 329  * Tracepoint for waiting on task to unschedule:
 330  */
 331 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35))
 332 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_process_template, sched_wait_task,
 333         TP_PROTO(struct task_struct *p),
 334         TP_ARGS(p))
 335 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 336 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_process_template, sched_wait_task,
 337         TP_PROTO(struct rq *rq, struct task_struct *p),
 338         TP_ARGS(rq, p))
 339 #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,35)) */
 340
 341 /*
 342  * Tracepoint for a waiting task:
 343  */
 344 LTTNG_TRACEPOINT_EVENT(sched_process_wait,
 345
 346         TP_PROTO(struct pid *pid),
 347
 348         TP_ARGS(pid),
 349
 350         TP_FIELDS(
 351                 ctf_array_text(char, comm, current->comm, TASK_COMM_LEN)
 352                 ctf_integer(pid_t, tid, pid_nr(pid))
 353                 ctf_integer(int, prio, current->prio - MAX_RT_PRIO)
 354         )
 355 )
 356
 357 /*
 358  * Tracepoint for do_fork.
 359  * Saving both TID and PID information, especially for the child, allows
 360  * trace analyzers to distinguish between creation of a new process and
 361  * creation of a new thread. Newly created processes will have child_tid
 362  * == child_pid, while creation of a thread yields to child_tid !=
 363  * child_pid.
 364  */
 365 LTTNG_TRACEPOINT_EVENT_CODE(sched_process_fork,
 366
 367         TP_PROTO(struct task_struct *parent, struct task_struct *child),
 368
 369         TP_ARGS(parent, child),
 370
 371         TP_locvar(
 372                 pid_t vtids[LTTNG_MAX_PID_NS_LEVEL];
 373                 unsigned int ns_level;
 374         ),
 375
 376         TP_code_pre(
 377                 if (child) {
 378                         struct pid *child_pid;
 379                         unsigned int i;
 380
 381                         child_pid = task_pid(child);
 382                         tp_locvar->ns_level =
 383                                 min_t(unsigned int, child_pid->level + 1,
 384                                         LTTNG_MAX_PID_NS_LEVEL);
 385                         for (i = 0; i < tp_locvar->ns_level; i++)
 386                                 tp_locvar->vtids[i] = child_pid->numbers[i].nr;
 387                 }
 388         ),
 389
 390         TP_FIELDS(
 391                 ctf_array_text(char, parent_comm, parent->comm, TASK_COMM_LEN)
 392                 ctf_integer(pid_t, parent_tid, parent->pid)
 393                 ctf_integer(pid_t, parent_pid, parent->tgid)
 394 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,8,0))
 395                 ctf_integer(unsigned int, parent_ns_inum,
 396                         ({
 397                                 unsigned int parent_ns_inum = 0;
 398
 399                                 if (parent) {
 400                                         struct pid_namespace *pid_ns;
 401
 402                                         pid_ns = task_active_pid_ns(parent);
 403                                         if (pid_ns)
 404                                                 parent_ns_inum =
 405                                                         pid_ns->lttng_proc_inum;
 406                                 }
 407                                 parent_ns_inum;
 408                         }))
 409 #endif
 410                 ctf_array_text(char, child_comm, child->comm, TASK_COMM_LEN)
 411                 ctf_integer(pid_t, child_tid, child->pid)
 412                 ctf_sequence(pid_t, vtids, tp_locvar->vtids, u8, tp_locvar->ns_level)
 413                 ctf_integer(pid_t, child_pid, child->tgid)
 414 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,8,0))
 415                 ctf_integer(unsigned int, child_ns_inum,
 416                         ({
 417                                 unsigned int child_ns_inum = 0;
 418
 419                                 if (child) {
 420                                         struct pid_namespace *pid_ns;
 421
 422                                         pid_ns = task_active_pid_ns(child);
 423                                         if (pid_ns)
 424                                                 child_ns_inum =
 425                                                         pid_ns->lttng_proc_inum;
 426                                 }
 427                                 child_ns_inum;
 428                         }))
 429 #endif
 430         ),
 431
 432         TP_code_post()
 433 )
 434
 435 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,33))
 436 /*
 437  * Tracepoint for sending a signal:
 438  */
 439 LTTNG_TRACEPOINT_EVENT(sched_signal_send,
 440
 441         TP_PROTO(int sig, struct task_struct *p),
 442
 443         TP_ARGS(sig, p),
 444
 445         TP_FIELDS(
 446                 ctf_integer(int, sig, sig)
 447                 ctf_array_text(char, comm, p->comm, TASK_COMM_LEN)
 448                 ctf_integer(pid_t, tid, p->pid)
 449         )
 450 )
 451 #endif
 452
 453 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,4,0))
 454 /*
 455  * Tracepoint for exec:
 456  */
 457 LTTNG_TRACEPOINT_EVENT(sched_process_exec,
 458
 459         TP_PROTO(struct task_struct *p, pid_t old_pid,
 460                  struct linux_binprm *bprm),
 461
 462         TP_ARGS(p, old_pid, bprm),
 463
 464         TP_FIELDS(
 465                 ctf_string(filename, bprm->filename)
 466                 ctf_integer(pid_t, tid, p->pid)
 467                 ctf_integer(pid_t, old_tid, old_pid)
 468         )
 469 )
 470 #endif
 471
 472 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,32))
 473 /*
 474  * XXX the below sched_stat tracepoints only apply to SCHED_OTHER/BATCH/IDLE
 475  *     adding sched_stat support to SCHED_FIFO/RR would be welcome.
 476  */
 477 LTTNG_TRACEPOINT_EVENT_CLASS(sched_stat_template,
 478
 479         TP_PROTO(struct task_struct *tsk, u64 delay),
 480
 481         TP_ARGS(tsk, delay),
 482
 483         TP_FIELDS(
 484                 ctf_array_text(char, comm, tsk->comm, TASK_COMM_LEN)
 485                 ctf_integer(pid_t, tid, tsk->pid)
 486                 ctf_integer(u64, delay, delay)
 487         )
 488 )
 489
 490
 491 /*
 492  * Tracepoint for accounting wait time (time the task is runnable
 493  * but not actually running due to scheduler contention).
 494  */
 495 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_stat_template, sched_stat_wait,
 496              TP_PROTO(struct task_struct *tsk, u64 delay),
 497              TP_ARGS(tsk, delay))
 498
 499 /*
 500  * Tracepoint for accounting sleep time (time the task is not runnable,
 501  * including iowait, see below).
 502  */
 503 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_stat_template, sched_stat_sleep,
 504              TP_PROTO(struct task_struct *tsk, u64 delay),
 505              TP_ARGS(tsk, delay))
 506
 507 /*
 508  * Tracepoint for accounting iowait time (time the task is not runnable
 509  * due to waiting on IO to complete).
 510  */
 511 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_stat_template, sched_stat_iowait,
 512              TP_PROTO(struct task_struct *tsk, u64 delay),
 513              TP_ARGS(tsk, delay))
 514
 515 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,3,0))
 516 /*
 517  * Tracepoint for accounting blocked time (time the task is in uninterruptible).
 518  */
 519 LTTNG_TRACEPOINT_EVENT_INSTANCE(sched_stat_template, sched_stat_blocked,
 520              TP_PROTO(struct task_struct *tsk, u64 delay),
 521              TP_ARGS(tsk, delay))
 522 #endif
 523
 524 /*
 525  * Tracepoint for accounting runtime (time the task is executing
 526  * on a CPU).
 527  */
 528 LTTNG_TRACEPOINT_EVENT(sched_stat_runtime,
 529
 530         TP_PROTO(struct task_struct *tsk, u64 runtime, u64 vruntime),
 531
 532         TP_ARGS(tsk, runtime, vruntime),
 533
 534         TP_FIELDS(
 535                 ctf_array_text(char, comm, tsk->comm, TASK_COMM_LEN)
 536                 ctf_integer(pid_t, tid, tsk->pid)
 537                 ctf_integer(u64, runtime, runtime)
 538                 ctf_integer(u64, vruntime, vruntime)
 539         )
 540 )
 541 #endif
 542
 543 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,12,0) || \
 544         LTTNG_RT_KERNEL_RANGE(4,9,27,18, 4,10,0,0) || \
 545         LTTNG_RT_KERNEL_RANGE(4,11,5,1, 4,12,0,0))
 546 /*
 547  * Tracepoint for showing priority inheritance modifying a tasks
 548  * priority.
 549  */
 550 LTTNG_TRACEPOINT_EVENT(sched_pi_setprio,
 551
 552         TP_PROTO(struct task_struct *tsk, struct task_struct *pi_task),
 553
 554         TP_ARGS(tsk, pi_task),
 555
 556         TP_FIELDS(
 557                 ctf_array_text(char, comm, tsk->comm, TASK_COMM_LEN)
 558                 ctf_integer(pid_t, tid, tsk->pid)
 559                 ctf_integer(int, oldprio, tsk->prio - MAX_RT_PRIO)
 560                 ctf_integer(int, newprio, pi_task ? pi_task->prio - MAX_RT_PRIO : tsk->prio - MAX_RT_PRIO)
 561         )
 562 )
 563 #elif (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,37))
 564 /*
 565  * Tracepoint for showing priority inheritance modifying a tasks
 566  * priority.
 567  */
 568 LTTNG_TRACEPOINT_EVENT(sched_pi_setprio,
 569
 570         TP_PROTO(struct task_struct *tsk, int newprio),
 571
 572         TP_ARGS(tsk, newprio),
 573
 574         TP_FIELDS(
 575                 ctf_array_text(char, comm, tsk->comm, TASK_COMM_LEN)
 576                 ctf_integer(pid_t, tid, tsk->pid)
 577                 ctf_integer(int, oldprio, tsk->prio - MAX_RT_PRIO)
 578                 ctf_integer(int, newprio, newprio - MAX_RT_PRIO)
 579         )
 580 )
 581 #endif
 582
 583 #endif /* LTTNG_TRACE_SCHED_H */
 584
 585 /* This part must be outside protection */
 586 #include <probes/define_trace.h>