projects/linux·blob master · refs

602 lines · 15.3 KiB

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#include <linux/cpuidle.h>
10#include <linux/suspend.h>
11#include <linux/livepatch.h>
12#include "sched.h"
13#include "smp.h"
14
15/* Linker adds these: start and end of __cpuidle functions */
16extern char __cpuidle_text_start[], __cpuidle_text_end[];
17
18/**
19 * sched_idle_set_state - Record idle state for the current CPU.
20 * @idle_state: State to record.
21 */
22void sched_idle_set_state(struct cpuidle_state *idle_state)
23{
24 idle_set_state(this_rq(), idle_state);
25}
26
27static int __read_mostly cpu_idle_force_poll;
28
29void cpu_idle_poll_ctrl(bool enable)
30{
31 if (enable) {
32 cpu_idle_force_poll++;
33 } else {
34 cpu_idle_force_poll--;
35 WARN_ON_ONCE(cpu_idle_force_poll < 0);
36 }
37}
38
39#ifdef CONFIG_GENERIC_IDLE_POLL_SETUP
40static int __init cpu_idle_poll_setup(char *__unused)
41{
42 cpu_idle_force_poll = 1;
43
44 return 1;
45}
46__setup("nohlt", cpu_idle_poll_setup);
47
48static int __init cpu_idle_nopoll_setup(char *__unused)
49{
50 cpu_idle_force_poll = 0;
51
52 return 1;
53}
54__setup("hlt", cpu_idle_nopoll_setup);
55#endif /* CONFIG_GENERIC_IDLE_POLL_SETUP */
56
57static noinline int __cpuidle cpu_idle_poll(void)
58{
59 instrumentation_begin();
60 trace_cpu_idle(0, smp_processor_id());
61 stop_critical_timings();
62 ct_cpuidle_enter();
63
64 raw_local_irq_enable();
65 while (!tif_need_resched() &&
66 (cpu_idle_force_poll || tick_check_broadcast_expired()))
67 cpu_relax();
68 raw_local_irq_disable();
69
70 ct_cpuidle_exit();
71 start_critical_timings();
72 trace_cpu_idle(PWR_EVENT_EXIT, smp_processor_id());
73 local_irq_enable();
74 instrumentation_end();
75
76 return 1;
77}
78
79/* Weak implementations for optional arch specific functions */
80void __weak arch_cpu_idle_prepare(void) { }
81void __weak arch_cpu_idle_enter(void) { }
82void __weak arch_cpu_idle_exit(void) { }
83void __weak __noreturn arch_cpu_idle_dead(void) { while (1); }
84void __weak arch_cpu_idle(void)
85{
86 cpu_idle_force_poll = 1;
87}
88
89#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE
90DEFINE_STATIC_KEY_FALSE(arch_needs_tick_broadcast);
91
92static inline void cond_tick_broadcast_enter(void)
93{
94 if (static_branch_unlikely(&arch_needs_tick_broadcast))
95 tick_broadcast_enter();
96}
97
98static inline void cond_tick_broadcast_exit(void)
99{
100 if (static_branch_unlikely(&arch_needs_tick_broadcast))
101 tick_broadcast_exit();
102}
103#else /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE: */
104static inline void cond_tick_broadcast_enter(void) { }
105static inline void cond_tick_broadcast_exit(void) { }
106#endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
107
108/**
109 * default_idle_call - Default CPU idle routine.
110 *
111 * To use when the cpuidle framework cannot be used.
112 */
113void __cpuidle default_idle_call(void)
114{
115 instrumentation_begin();
116 if (!current_clr_polling_and_test()) {
117 cond_tick_broadcast_enter();
118 trace_cpu_idle(1, smp_processor_id());
119 stop_critical_timings();
120
121 ct_cpuidle_enter();
122 arch_cpu_idle();
123 ct_cpuidle_exit();
124
125 start_critical_timings();
126 trace_cpu_idle(PWR_EVENT_EXIT, smp_processor_id());
127 cond_tick_broadcast_exit();
128 }
129 local_irq_enable();
130 instrumentation_end();
131}
132
133static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
134 struct cpuidle_device *dev,
135 u64 max_latency_ns)
136{
137 if (current_clr_polling_and_test())
138 return -EBUSY;
139
140 return cpuidle_enter_s2idle(drv, dev, max_latency_ns);
141}
142
143static int call_cpuidle(struct cpuidle_driver *drv, struct cpuidle_device *dev,
144 int next_state)
145{
146 /*
147 * The idle task must be scheduled, it is pointless to go to idle, just
148 * update no idle residency and return.
149 */
150 if (current_clr_polling_and_test()) {
151 dev->last_residency_ns = 0;
152 local_irq_enable();
153 return -EBUSY;
154 }
155
156 /*
157 * Enter the idle state previously returned by the governor decision.
158 * This function will block until an interrupt occurs and will take
159 * care of re-enabling the local interrupts
160 */
161 return cpuidle_enter(drv, dev, next_state);
162}
163
164static void idle_call_stop_or_retain_tick(bool stop_tick)
165{
166 if (stop_tick || tick_nohz_tick_stopped())
167 tick_nohz_idle_stop_tick();
168 else
169 tick_nohz_idle_retain_tick();
170}
171
172/**
173 * cpuidle_idle_call - the main idle function
174 *
175 * NOTE: no locks or semaphores should be used here
176 *
177 * On architectures that support TIF_POLLING_NRFLAG, is called with polling
178 * set, and it returns with polling set. If it ever stops polling, it
179 * must clear the polling bit.
180 */
181static void cpuidle_idle_call(bool stop_tick)
182{
183 struct cpuidle_device *dev = cpuidle_get_device();
184 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
185 int next_state, entered_state;
186
187 /*
188 * Check if the idle task must be rescheduled. If it is the
189 * case, exit the function after re-enabling the local IRQ.
190 */
191 if (need_resched()) {
192 local_irq_enable();
193 return;
194 }
195
196 if (cpuidle_not_available(drv, dev)) {
197 idle_call_stop_or_retain_tick(stop_tick);
198
199 default_idle_call();
200 goto exit_idle;
201 }
202
203 /*
204 * Suspend-to-idle ("s2idle") is a system state in which all user space
205 * has been frozen, all I/O devices have been suspended and the only
206 * activity happens here and in interrupts (if any). In that case bypass
207 * the cpuidle governor and go straight for the deepest idle state
208 * available. Possibly also suspend the local tick and the entire
209 * timekeeping to prevent timer interrupts from kicking us out of idle
210 * until a proper wakeup interrupt happens.
211 */
212
213 if (idle_should_enter_s2idle() || dev->forced_idle_latency_limit_ns) {
214 u64 max_latency_ns;
215
216 if (idle_should_enter_s2idle()) {
217 max_latency_ns = cpu_wakeup_latency_qos_limit() *
218 NSEC_PER_USEC;
219
220 entered_state = call_cpuidle_s2idle(drv, dev,
221 max_latency_ns);
222 if (entered_state > 0)
223 goto exit_idle;
224 } else {
225 max_latency_ns = dev->forced_idle_latency_limit_ns;
226 }
227
228 tick_nohz_idle_stop_tick();
229
230 next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
231 call_cpuidle(drv, dev, next_state);
232 } else if (drv->state_count > 1) {
233 /*
234 * stop_tick is expected to be true by default by cpuidle
235 * governors, which allows them to select idle states with
236 * target residency above the tick period length.
237 */
238 stop_tick = true;
239
240 /*
241 * Ask the cpuidle framework to choose a convenient idle state.
242 */
243 next_state = cpuidle_select(drv, dev, &stop_tick);
244
245 idle_call_stop_or_retain_tick(stop_tick);
246
247 entered_state = call_cpuidle(drv, dev, next_state);
248 /*
249 * Give the governor an opportunity to reflect on the outcome
250 */
251 cpuidle_reflect(dev, entered_state);
252 } else {
253 idle_call_stop_or_retain_tick(stop_tick);
254
255 /*
256 * If there is only a single idle state (or none), there is
257 * nothing meaningful for the governor to choose. Skip the
258 * governor and always use state 0.
259 */
260 call_cpuidle(drv, dev, 0);
261 }
262
263exit_idle:
264 __current_set_polling();
265
266 /*
267 * It is up to the idle functions to re-enable local interrupts
268 */
269 if (WARN_ON_ONCE(irqs_disabled()))
270 local_irq_enable();
271}
272
273/*
274 * Generic idle loop implementation
275 *
276 * Called with polling cleared.
277 */
278static void do_idle(void)
279{
280 int cpu = smp_processor_id();
281 bool got_tick = false;
282
283 if (cpu_is_offline(cpu)) {
284 local_irq_disable();
285 /* All per-CPU kernel threads should be done by now. */
286 WARN_ON_ONCE(need_resched());
287 cpuhp_report_idle_dead();
288 arch_cpu_idle_dead();
289 }
290
291 /*
292 * Check if we need to update blocked load
293 */
294 nohz_run_idle_balance(cpu);
295
296 /*
297 * If the arch has a polling bit, we maintain an invariant:
298 *
299 * Our polling bit is clear if we're not scheduled (i.e. if rq->curr !=
300 * rq->idle). This means that, if rq->idle has the polling bit set,
301 * then setting need_resched is guaranteed to cause the CPU to
302 * reschedule.
303 */
304
305 __current_set_polling();
306 tick_nohz_idle_enter();
307
308 while (!need_resched()) {
309
310 /*
311 * Interrupts shouldn't be re-enabled from that point on until
312 * the CPU sleeping instruction is reached. Otherwise an interrupt
313 * may fire and queue a timer that would be ignored until the CPU
314 * wakes from the sleeping instruction. And testing need_resched()
315 * doesn't tell about pending needed timer reprogram.
316 *
317 * Several cases to consider:
318 *
319 * - SLEEP-UNTIL-PENDING-INTERRUPT based instructions such as
320 * "wfi" or "mwait" are fine because they can be entered with
321 * interrupt disabled.
322 *
323 * - sti;mwait() couple is fine because the interrupts are
324 * re-enabled only upon the execution of mwait, leaving no gap
325 * in-between.
326 *
327 * - ROLLBACK based idle handlers with the sleeping instruction
328 * called with interrupts enabled are NOT fine. In this scheme
329 * when the interrupt detects it has interrupted an idle handler,
330 * it rolls back to its beginning which performs the
331 * need_resched() check before re-executing the sleeping
332 * instruction. This can leak a pending needed timer reprogram.
333 * If such a scheme is really mandatory due to the lack of an
334 * appropriate CPU sleeping instruction, then a FAST-FORWARD
335 * must instead be applied: when the interrupt detects it has
336 * interrupted an idle handler, it must resume to the end of
337 * this idle handler so that the generic idle loop is iterated
338 * again to reprogram the tick.
339 */
340 local_irq_disable();
341
342 arch_cpu_idle_enter();
343 rcu_nocb_flush_deferred_wakeup();
344
345 /*
346 * In poll mode we re-enable interrupts and spin. Also if we
347 * detected in the wakeup from idle path that the tick
348 * broadcast device expired for us, we don't want to go deep
349 * idle as we know that the IPI is going to arrive right away.
350 */
351 if (cpu_idle_force_poll || tick_check_broadcast_expired()) {
352 tick_nohz_idle_restart_tick();
353 cpu_idle_poll();
354 } else {
355 cpuidle_idle_call(got_tick);
356 }
357 got_tick = tick_nohz_idle_got_tick();
358 arch_cpu_idle_exit();
359 }
360
361 /*
362 * Since we fell out of the loop above, we know TIF_NEED_RESCHED must
363 * be set, propagate it into PREEMPT_NEED_RESCHED.
364 *
365 * This is required because for polling idle loops we will not have had
366 * an IPI to fold the state for us.
367 */
368 preempt_set_need_resched();
369 tick_nohz_idle_exit();
370 __current_clr_polling();
371
372 /*
373 * We promise to call sched_ttwu_pending() and reschedule if
374 * need_resched() is set while polling is set. That means that clearing
375 * polling needs to be visible before doing these things.
376 */
377 smp_mb__after_atomic();
378
379 /*
380 * RCU relies on this call to be done outside of an RCU read-side
381 * critical section.
382 */
383 flush_smp_call_function_queue();
384 schedule_idle();
385
386 if (unlikely(klp_patch_pending(current)))
387 klp_update_patch_state(current);
388}
389
390bool cpu_in_idle(unsigned long pc)
391{
392 return pc >= (unsigned long)__cpuidle_text_start &&
393 pc < (unsigned long)__cpuidle_text_end;
394}
395
396struct idle_timer {
397 struct hrtimer timer;
398 int done;
399};
400
401static enum hrtimer_restart idle_inject_timer_fn(struct hrtimer *timer)
402{
403 struct idle_timer *it = container_of(timer, struct idle_timer, timer);
404
405 WRITE_ONCE(it->done, 1);
406 set_tsk_need_resched(current);
407
408 return HRTIMER_NORESTART;
409}
410
411void play_idle_precise(u64 duration_ns, u64 latency_ns)
412{
413 struct idle_timer it;
414
415 /*
416 * Only FIFO tasks can disable the tick since they don't need the forced
417 * preemption.
418 */
419 WARN_ON_ONCE(current->policy != SCHED_FIFO);
420 WARN_ON_ONCE(current->nr_cpus_allowed != 1);
421 WARN_ON_ONCE(!(current->flags & PF_KTHREAD));
422 WARN_ON_ONCE(!(current->flags & PF_NO_SETAFFINITY));
423 WARN_ON_ONCE(!duration_ns);
424 WARN_ON_ONCE(current->mm);
425
426 rcu_sleep_check();
427 preempt_disable();
428 current->flags |= PF_IDLE;
429 cpuidle_use_deepest_state(latency_ns);
430
431 it.done = 0;
432 hrtimer_setup_on_stack(&it.timer, idle_inject_timer_fn, CLOCK_MONOTONIC,
433 HRTIMER_MODE_REL_HARD);
434 hrtimer_start(&it.timer, ns_to_ktime(duration_ns),
435 HRTIMER_MODE_REL_PINNED_HARD);
436
437 while (!READ_ONCE(it.done))
438 do_idle();
439
440 cpuidle_use_deepest_state(0);
441 current->flags &= ~PF_IDLE;
442
443 preempt_fold_need_resched();
444 preempt_enable();
445}
446EXPORT_SYMBOL_GPL(play_idle_precise);
447
448void cpu_startup_entry(enum cpuhp_state state)
449{
450 current->flags |= PF_IDLE;
451 arch_cpu_idle_prepare();
452 cpuhp_online_idle(state);
453 while (1)
454 do_idle();
455}
456
457/*
458 * idle-task scheduling class.
459 */
460
461static int
462select_task_rq_idle(struct task_struct *p, int cpu, int flags)
463{
464 return task_cpu(p); /* IDLE tasks as never migrated */
465}
466
467static int
468balance_idle(struct rq *rq, struct rq_flags *rf)
469{
470 return WARN_ON_ONCE(1);
471}
472
473/*
474 * Idle tasks are unconditionally rescheduled:
475 */
476static void wakeup_preempt_idle(struct rq *rq, struct task_struct *p, int flags)
477{
478 resched_curr(rq);
479}
480
481static void update_curr_idle(struct rq *rq);
482
483static void put_prev_task_idle(struct rq *rq, struct task_struct *prev, struct task_struct *next)
484{
485 update_curr_idle(rq);
486 scx_update_idle(rq, false, true);
487 update_rq_avg_idle(rq);
488}
489
490static void set_next_task_idle(struct rq *rq, struct task_struct *next, enum snt_e type)
491{
492 if (type == SNT_REPICK)
493 return;
494
495 update_idle_core(rq);
496 scx_update_idle(rq, true, true);
497 schedstat_inc(rq->sched_goidle);
498 next->se.exec_start = rq_clock_task(rq);
499
500 /*
501 * rq is about to be idle, check if we need to update the
502 * lost_idle_time of clock_pelt
503 */
504 update_idle_rq_clock_pelt(rq);
505}
506
507struct task_struct *pick_task_idle(struct rq *rq, struct rq_flags *rf)
508{
509 /*
510 * Notify scx only on an idle-to-idle re-pick (the cpu was already idle).
511 * A real task->idle transition is delivered by set_next_task_idle(), so
512 * calling here too would duplicate it.
513 */
514 if (scx_enabled() && is_idle_task(rq->curr))
515 scx_update_idle(rq, true, false);
516 return rq->idle;
517}
518
519/*
520 * It is not legal to sleep in the idle task - print a warning
521 * message if some code attempts to do it:
522 */
523static bool
524dequeue_task_idle(struct rq *rq, struct task_struct *p, int flags)
525{
526 raw_spin_rq_unlock_irq(rq);
527 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
528 dump_stack();
529 raw_spin_rq_lock_irq(rq);
530 return true;
531}
532
533/*
534 * scheduler tick hitting a task of our scheduling class.
535 *
536 * NOTE: This function can be called remotely by the tick offload that
537 * goes along full dynticks. Therefore no local assumption can be made
538 * and everything must be accessed through the @rq and @curr passed in
539 * parameters.
540 */
541static void task_tick_idle(struct rq *rq, struct task_struct *curr, int queued)
542{
543 update_curr_idle(rq);
544}
545
546static void switching_to_idle(struct rq *rq, struct task_struct *p)
547{
548 BUG();
549}
550
551static void
552prio_changed_idle(struct rq *rq, struct task_struct *p, u64 oldprio)
553{
554 if (p->prio == oldprio)
555 return;
556
557 BUG();
558}
559
560static void update_curr_idle(struct rq *rq)
561{
562 struct sched_entity *se = &rq->idle->se;
563 u64 now = rq_clock_task(rq);
564 s64 delta_exec;
565
566 delta_exec = now - se->exec_start;
567 if (unlikely(delta_exec <= 0))
568 return;
569
570 se->exec_start = now;
571
572 dl_server_update_idle(&rq->fair_server, delta_exec);
573#ifdef CONFIG_SCHED_CLASS_EXT
574 dl_server_update_idle(&rq->ext_server, delta_exec);
575#endif
576}
577
578/*
579 * Simple, special scheduling class for the per-CPU idle tasks:
580 */
581DEFINE_SCHED_CLASS(idle) = {
582 /* no enqueue/yield_task for idle tasks */
583
584 /* dequeue is not valid, we print a debug message there: */
585 .dequeue_task = dequeue_task_idle,
586
587 .wakeup_preempt = wakeup_preempt_idle,
588
589 .pick_task = pick_task_idle,
590 .put_prev_task = put_prev_task_idle,
591 .set_next_task = set_next_task_idle,
592
593 .balance = balance_idle,
594 .select_task_rq = select_task_rq_idle,
595 .set_cpus_allowed = set_cpus_allowed_common,
596
597 .task_tick = task_tick_idle,
598
599 .prio_changed = prio_changed_idle,
600 .switching_to = switching_to_idle,
601 .update_curr = update_curr_idle,
602};