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