kernel/sched/idle.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Generic entry points for the idle threads and
   4  * implementation of the idle task scheduling class.
   5  *
   6  * (NOTE: these are not related to SCHED_IDLE batch scheduled
   7  *        tasks which are handled in sched/fair.c )
   8  */
   9
  10 /* Linker adds these: start and end of __cpuidle functions */
  11 extern char __cpuidle_text_start[], __cpuidle_text_end[];
  12
  13 /**
  14  * sched_idle_set_state - Record idle state for the current CPU.
  15  * @idle_state: State to record.
  16  */
  17 void sched_idle_set_state(struct cpuidle_state *idle_state)
  18 {
  19         idle_set_state(this_rq(), idle_state);
  20 }
  21
  22 static int __read_mostly cpu_idle_force_poll;
  23
  24 void cpu_idle_poll_ctrl(bool enable)
  25 {
  26         if (enable) {
  27                 cpu_idle_force_poll++;
  28         } else {
  29                 cpu_idle_force_poll--;
  30                 WARN_ON_ONCE(cpu_idle_force_poll < 0);
  31         }
  32 }
  33
  34 #ifdef CONFIG_GENERIC_IDLE_POLL_SETUP
  35 static int __init cpu_idle_poll_setup(char *__unused)
  36 {
  37         cpu_idle_force_poll = 1;
  38
  39         return 1;
  40 }
  41 __setup("nohlt", cpu_idle_poll_setup);
  42
  43 static int __init cpu_idle_nopoll_setup(char *__unused)
  44 {
  45         cpu_idle_force_poll = 0;
  46
  47         return 1;
  48 }
  49 __setup("hlt", cpu_idle_nopoll_setup);
  50 #endif
  51
  52 static noinline int __cpuidle cpu_idle_poll(void)
  53 {
  54         instrumentation_begin();
  55         trace_cpu_idle(0, smp_processor_id());
  56         stop_critical_timings();
  57         ct_cpuidle_enter();
  58
  59         raw_local_irq_enable();
  60         while (!tif_need_resched() &&
  61                (cpu_idle_force_poll || tick_check_broadcast_expired()))
  62                 cpu_relax();
  63         raw_local_irq_disable();
  64
  65         ct_cpuidle_exit();
  66         start_critical_timings();
  67         trace_cpu_idle(PWR_EVENT_EXIT, smp_processor_id());
  68         local_irq_enable();
  69         instrumentation_end();
  70
  71         return 1;
  72 }
  73
  74 /* Weak implementations for optional arch specific functions */
  75 void __weak arch_cpu_idle_prepare(void) { }
  76 void __weak arch_cpu_idle_enter(void) { }
  77 void __weak arch_cpu_idle_exit(void) { }
  78 void __weak __noreturn arch_cpu_idle_dead(void) { while (1); }
  79 void __weak arch_cpu_idle(void)
  80 {
  81         cpu_idle_force_poll = 1;
  82 }
  83
  84 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE
  85 DEFINE_STATIC_KEY_FALSE(arch_needs_tick_broadcast);
  86
  87 static inline void cond_tick_broadcast_enter(void)
  88 {
  89         if (static_branch_unlikely(&arch_needs_tick_broadcast))
  90                 tick_broadcast_enter();
  91 }
  92
  93 static inline void cond_tick_broadcast_exit(void)
  94 {
  95         if (static_branch_unlikely(&arch_needs_tick_broadcast))
  96                 tick_broadcast_exit();
  97 }
  98 #else
  99 static inline void cond_tick_broadcast_enter(void) { }
 100 static inline void cond_tick_broadcast_exit(void) { }
 101 #endif
 102
 103 /**
 104  * default_idle_call - Default CPU idle routine.
 105  *
 106  * To use when the cpuidle framework cannot be used.
 107  */
 108 void __cpuidle default_idle_call(void)
 109 {
 110         instrumentation_begin();
 111         if (!current_clr_polling_and_test()) {
 112                 cond_tick_broadcast_enter();
 113                 trace_cpu_idle(1, smp_processor_id());
 114                 stop_critical_timings();
 115
 116                 ct_cpuidle_enter();
 117                 arch_cpu_idle();
 118                 ct_cpuidle_exit();
 119
 120                 start_critical_timings();
 121                 trace_cpu_idle(PWR_EVENT_EXIT, smp_processor_id());
 122                 cond_tick_broadcast_exit();
 123         }
 124         local_irq_enable();
 125         instrumentation_end();
 126 }
 127
 128 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
 129                                struct cpuidle_device *dev)
 130 {
 131         if (current_clr_polling_and_test())
 132                 return -EBUSY;
 133
 134         return cpuidle_enter_s2idle(drv, dev);
 135 }
 136
 137 static int call_cpuidle(struct cpuidle_driver *drv, struct cpuidle_device *dev,
 138                       int next_state)
 139 {
 140         /*
 141          * The idle task must be scheduled, it is pointless to go to idle, just
 142          * update no idle residency and return.
 143          */
 144         if (current_clr_polling_and_test()) {
 145                 dev->last_residency_ns = 0;
 146                 local_irq_enable();
 147                 return -EBUSY;
 148         }
 149
 150         /*
 151          * Enter the idle state previously returned by the governor decision.
 152          * This function will block until an interrupt occurs and will take
 153          * care of re-enabling the local interrupts
 154          */
 155         return cpuidle_enter(drv, dev, next_state);
 156 }
 157
 158 /**
 159  * cpuidle_idle_call - the main idle function
 160  *
 161  * NOTE: no locks or semaphores should be used here
 162  *
 163  * On architectures that support TIF_POLLING_NRFLAG, is called with polling
 164  * set, and it returns with polling set.  If it ever stops polling, it
 165  * must clear the polling bit.
 166  */
 167 static void cpuidle_idle_call(void)
 168 {
 169         struct cpuidle_device *dev = cpuidle_get_device();
 170         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 171         int next_state, entered_state;
 172
 173         /*
 174          * Check if the idle task must be rescheduled. If it is the
 175          * case, exit the function after re-enabling the local irq.
 176          */
 177         if (need_resched()) {
 178                 local_irq_enable();
 179                 return;
 180         }
 181
 182         /*
 183          * The RCU framework needs to be told that we are entering an idle
 184          * section, so no more rcu read side critical sections and one more
 185          * step to the grace period
 186          */
 187
 188         if (cpuidle_not_available(drv, dev)) {
 189                 tick_nohz_idle_stop_tick();
 190
 191                 default_idle_call();
 192                 goto exit_idle;
 193         }
 194
 195         /*
 196          * Suspend-to-idle ("s2idle") is a system state in which all user space
 197          * has been frozen, all I/O devices have been suspended and the only
 198          * activity happens here and in interrupts (if any). In that case bypass
 199          * the cpuidle governor and go straight for the deepest idle state
 200          * available.  Possibly also suspend the local tick and the entire
 201          * timekeeping to prevent timer interrupts from kicking us out of idle
 202          * until a proper wakeup interrupt happens.
 203          */
 204
 205         if (idle_should_enter_s2idle() || dev->forced_idle_latency_limit_ns) {
 206                 u64 max_latency_ns;
 207
 208                 if (idle_should_enter_s2idle()) {
 209
 210                         entered_state = call_cpuidle_s2idle(drv, dev);
 211                         if (entered_state > 0)
 212                                 goto exit_idle;
 213
 214                         max_latency_ns = U64_MAX;
 215                 } else {
 216                         max_latency_ns = dev->forced_idle_latency_limit_ns;
 217                 }
 218
 219                 tick_nohz_idle_stop_tick();
 220
 221                 next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
 222                 call_cpuidle(drv, dev, next_state);
 223         } else {
 224                 bool stop_tick = true;
 225
 226                 /*
 227                  * Ask the cpuidle framework to choose a convenient idle state.
 228                  */
 229                 next_state = cpuidle_select(drv, dev, &stop_tick);
 230
 231                 if (stop_tick || tick_nohz_tick_stopped())
 232                         tick_nohz_idle_stop_tick();
 233                 else
 234                         tick_nohz_idle_retain_tick();
 235
 236                 entered_state = call_cpuidle(drv, dev, next_state);
 237                 /*
 238                  * Give the governor an opportunity to reflect on the outcome
 239                  */
 240                 cpuidle_reflect(dev, entered_state);
 241         }
 242
 243 exit_idle:
 244         __current_set_polling();
 245
 246         /*
 247          * It is up to the idle functions to reenable local interrupts
 248          */
 249         if (WARN_ON_ONCE(irqs_disabled()))
 250                 local_irq_enable();
 251 }
 252
 253 /*
 254  * Generic idle loop implementation
 255  *
 256  * Called with polling cleared.
 257  */
 258 static void do_idle(void)
 259 {
 260         int cpu = smp_processor_id();
 261
 262         /*
 263          * Check if we need to update blocked load
 264          */
 265         nohz_run_idle_balance(cpu);
 266
 267         /*
 268          * If the arch has a polling bit, we maintain an invariant:
 269          *
 270          * Our polling bit is clear if we're not scheduled (i.e. if rq->curr !=
 271          * rq->idle). This means that, if rq->idle has the polling bit set,
 272          * then setting need_resched is guaranteed to cause the CPU to
 273          * reschedule.
 274          */
 275
 276         __current_set_polling();
 277         tick_nohz_idle_enter();
 278
 279         while (!need_resched()) {
 280                 rmb();
 281
 282                 /*
 283                  * Interrupts shouldn't be re-enabled from that point on until
 284                  * the CPU sleeping instruction is reached. Otherwise an interrupt
 285                  * may fire and queue a timer that would be ignored until the CPU
 286                  * wakes from the sleeping instruction. And testing need_resched()
 287                  * doesn't tell about pending needed timer reprogram.
 288                  *
 289                  * Several cases to consider:
 290                  *
 291                  * - SLEEP-UNTIL-PENDING-INTERRUPT based instructions such as
 292                  *   "wfi" or "mwait" are fine because they can be entered with
 293                  *   interrupt disabled.
 294                  *
 295                  * - sti;mwait() couple is fine because the interrupts are
 296                  *   re-enabled only upon the execution of mwait, leaving no gap
 297                  *   in-between.
 298                  *
 299                  * - ROLLBACK based idle handlers with the sleeping instruction
 300                  *   called with interrupts enabled are NOT fine. In this scheme
 301                  *   when the interrupt detects it has interrupted an idle handler,
 302                  *   it rolls back to its beginning which performs the
 303                  *   need_resched() check before re-executing the sleeping
 304                  *   instruction. This can leak a pending needed timer reprogram.
 305                  *   If such a scheme is really mandatory due to the lack of an
 306                  *   appropriate CPU sleeping instruction, then a FAST-FORWARD
 307                  *   must instead be applied: when the interrupt detects it has
 308                  *   interrupted an idle handler, it must resume to the end of
 309                  *   this idle handler so that the generic idle loop is iterated
 310                  *   again to reprogram the tick.
 311                  */
 312                 local_irq_disable();
 313
 314                 if (cpu_is_offline(cpu)) {
 315                         cpuhp_report_idle_dead();
 316                         arch_cpu_idle_dead();
 317                 }
 318
 319                 arch_cpu_idle_enter();
 320                 rcu_nocb_flush_deferred_wakeup();
 321
 322                 /*
 323                  * In poll mode we reenable interrupts and spin. Also if we
 324                  * detected in the wakeup from idle path that the tick
 325                  * broadcast device expired for us, we don't want to go deep
 326                  * idle as we know that the IPI is going to arrive right away.
 327                  */
 328                 if (cpu_idle_force_poll || tick_check_broadcast_expired()) {
 329                         tick_nohz_idle_restart_tick();
 330                         cpu_idle_poll();
 331                 } else {
 332                         cpuidle_idle_call();
 333                 }
 334                 arch_cpu_idle_exit();
 335         }
 336
 337         /*
 338          * Since we fell out of the loop above, we know TIF_NEED_RESCHED must
 339          * be set, propagate it into PREEMPT_NEED_RESCHED.
 340          *
 341          * This is required because for polling idle loops we will not have had
 342          * an IPI to fold the state for us.
 343          */
 344         preempt_set_need_resched();
 345         tick_nohz_idle_exit();
 346         __current_clr_polling();
 347
 348         /*
 349          * We promise to call sched_ttwu_pending() and reschedule if
 350          * need_resched() is set while polling is set. That means that clearing
 351          * polling needs to be visible before doing these things.
 352          */
 353         smp_mb__after_atomic();
 354
 355         /*
 356          * RCU relies on this call to be done outside of an RCU read-side
 357          * critical section.
 358          */
 359         flush_smp_call_function_queue();
 360         schedule_idle();
 361
 362         if (unlikely(klp_patch_pending(current)))
 363                 klp_update_patch_state(current);
 364 }
 365
 366 bool cpu_in_idle(unsigned long pc)
 367 {
 368         return pc >= (unsigned long)__cpuidle_text_start &&
 369                 pc < (unsigned long)__cpuidle_text_end;
 370 }
 371
 372 struct idle_timer {
 373         struct hrtimer timer;
 374         int done;
 375 };
 376
 377 static enum hrtimer_restart idle_inject_timer_fn(struct hrtimer *timer)
 378 {
 379         struct idle_timer *it = container_of(timer, struct idle_timer, timer);
 380
 381         WRITE_ONCE(it->done, 1);
 382         set_tsk_need_resched(current);
 383
 384         return HRTIMER_NORESTART;
 385 }
 386
 387 void play_idle_precise(u64 duration_ns, u64 latency_ns)
 388 {
 389         struct idle_timer it;
 390
 391         /*
 392          * Only FIFO tasks can disable the tick since they don't need the forced
 393          * preemption.
 394          */
 395         WARN_ON_ONCE(current->policy != SCHED_FIFO);
 396         WARN_ON_ONCE(current->nr_cpus_allowed != 1);
 397         WARN_ON_ONCE(!(current->flags & PF_KTHREAD));
 398         WARN_ON_ONCE(!(current->flags & PF_NO_SETAFFINITY));
 399         WARN_ON_ONCE(!duration_ns);
 400         WARN_ON_ONCE(current->mm);
 401
 402         rcu_sleep_check();
 403         preempt_disable();
 404         current->flags |= PF_IDLE;
 405         cpuidle_use_deepest_state(latency_ns);
 406
 407         it.done = 0;
 408         hrtimer_init_on_stack(&it.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 409         it.timer.function = idle_inject_timer_fn;
 410         hrtimer_start(&it.timer, ns_to_ktime(duration_ns),
 411                       HRTIMER_MODE_REL_PINNED_HARD);
 412
 413         while (!READ_ONCE(it.done))
 414                 do_idle();
 415
 416         cpuidle_use_deepest_state(0);
 417         current->flags &= ~PF_IDLE;
 418
 419         preempt_fold_need_resched();
 420         preempt_enable();
 421 }
 422 EXPORT_SYMBOL_GPL(play_idle_precise);
 423
 424 void cpu_startup_entry(enum cpuhp_state state)
 425 {
 426         current->flags |= PF_IDLE;
 427         arch_cpu_idle_prepare();
 428         cpuhp_online_idle(state);
 429         while (1)
 430                 do_idle();
 431 }
 432
 433 /*
 434  * idle-task scheduling class.
 435  */
 436
 437 #ifdef CONFIG_SMP
 438 static int
 439 select_task_rq_idle(struct task_struct *p, int cpu, int flags)
 440 {
 441         return task_cpu(p); /* IDLE tasks as never migrated */
 442 }
 443
 444 static int
 445 balance_idle(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 446 {
 447         return WARN_ON_ONCE(1);
 448 }
 449 #endif
 450
 451 /*
 452  * Idle tasks are unconditionally rescheduled:
 453  */
 454 static void wakeup_preempt_idle(struct rq *rq, struct task_struct *p, int flags)
 455 {
 456         resched_curr(rq);
 457 }
 458
 459 static void put_prev_task_idle(struct rq *rq, struct task_struct *prev)
 460 {
 461 }
 462
 463 static void set_next_task_idle(struct rq *rq, struct task_struct *next, bool first)
 464 {
 465         update_idle_core(rq);
 466         schedstat_inc(rq->sched_goidle);
 467 }
 468
 469 #ifdef CONFIG_SMP
 470 static struct task_struct *pick_task_idle(struct rq *rq)
 471 {
 472         return rq->idle;
 473 }
 474 #endif
 475
 476 struct task_struct *pick_next_task_idle(struct rq *rq)
 477 {
 478         struct task_struct *next = rq->idle;
 479
 480         set_next_task_idle(rq, next, true);
 481
 482         return next;
 483 }
 484
 485 /*
 486  * It is not legal to sleep in the idle task - print a warning
 487  * message if some code attempts to do it:
 488  */
 489 static void
 490 dequeue_task_idle(struct rq *rq, struct task_struct *p, int flags)
 491 {
 492         raw_spin_rq_unlock_irq(rq);
 493         printk(KERN_ERR "bad: scheduling from the idle thread!\n");
 494         dump_stack();
 495         raw_spin_rq_lock_irq(rq);
 496 }
 497
 498 /*
 499  * scheduler tick hitting a task of our scheduling class.
 500  *
 501  * NOTE: This function can be called remotely by the tick offload that
 502  * goes along full dynticks. Therefore no local assumption can be made
 503  * and everything must be accessed through the @rq and @curr passed in
 504  * parameters.
 505  */
 506 static void task_tick_idle(struct rq *rq, struct task_struct *curr, int queued)
 507 {
 508 }
 509
 510 static void switched_to_idle(struct rq *rq, struct task_struct *p)
 511 {
 512         BUG();
 513 }
 514
 515 static void
 516 prio_changed_idle(struct rq *rq, struct task_struct *p, int oldprio)
 517 {
 518         BUG();
 519 }
 520
 521 static void update_curr_idle(struct rq *rq)
 522 {
 523 }
 524
 525 /*
 526  * Simple, special scheduling class for the per-CPU idle tasks:
 527  */
 528 DEFINE_SCHED_CLASS(idle) = {
 529
 530         /* no enqueue/yield_task for idle tasks */
 531
 532         /* dequeue is not valid, we print a debug message there: */
 533         .dequeue_task           = dequeue_task_idle,
 534
 535         .wakeup_preempt         = wakeup_preempt_idle,
 536
 537         .pick_next_task         = pick_next_task_idle,
 538         .put_prev_task          = put_prev_task_idle,
 539         .set_next_task          = set_next_task_idle,
 540
 541 #ifdef CONFIG_SMP
 542         .balance                = balance_idle,
 543         .pick_task              = pick_task_idle,
 544         .select_task_rq         = select_task_rq_idle,
 545         .set_cpus_allowed       = set_cpus_allowed_common,
 546 #endif
 547
 548         .task_tick              = task_tick_idle,
 549
 550         .prio_changed           = prio_changed_idle,
 551         .switched_to            = switched_to_idle,
 552         .update_curr            = update_curr_idle,
 553 };