xref: /openbmc/linux/kernel/sched/core.c (revision 172caf19)
1 /*
2  *  kernel/sched/core.c
3  *
4  *  Core kernel scheduler code and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  */
8 #include "sched.h"
9 
10 #include <linux/nospec.h>
11 
12 #include <linux/kcov.h>
13 
14 #include <asm/switch_to.h>
15 #include <asm/tlb.h>
16 
17 #include "../workqueue_internal.h"
18 #include "../smpboot.h"
19 
20 #include "pelt.h"
21 
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/sched.h>
24 
25 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
26 
27 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
28 /*
29  * Debugging: various feature bits
30  *
31  * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
32  * sysctl_sched_features, defined in sched.h, to allow constants propagation
33  * at compile time and compiler optimization based on features default.
34  */
35 #define SCHED_FEAT(name, enabled)	\
36 	(1UL << __SCHED_FEAT_##name) * enabled |
37 const_debug unsigned int sysctl_sched_features =
38 #include "features.h"
39 	0;
40 #undef SCHED_FEAT
41 #endif
42 
43 /*
44  * Number of tasks to iterate in a single balance run.
45  * Limited because this is done with IRQs disabled.
46  */
47 const_debug unsigned int sysctl_sched_nr_migrate = 32;
48 
49 /*
50  * period over which we measure -rt task CPU usage in us.
51  * default: 1s
52  */
53 unsigned int sysctl_sched_rt_period = 1000000;
54 
55 __read_mostly int scheduler_running;
56 
57 /*
58  * part of the period that we allow rt tasks to run in us.
59  * default: 0.95s
60  */
61 int sysctl_sched_rt_runtime = 950000;
62 
63 /*
64  * __task_rq_lock - lock the rq @p resides on.
65  */
66 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
67 	__acquires(rq->lock)
68 {
69 	struct rq *rq;
70 
71 	lockdep_assert_held(&p->pi_lock);
72 
73 	for (;;) {
74 		rq = task_rq(p);
75 		raw_spin_lock(&rq->lock);
76 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
77 			rq_pin_lock(rq, rf);
78 			return rq;
79 		}
80 		raw_spin_unlock(&rq->lock);
81 
82 		while (unlikely(task_on_rq_migrating(p)))
83 			cpu_relax();
84 	}
85 }
86 
87 /*
88  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
89  */
90 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
91 	__acquires(p->pi_lock)
92 	__acquires(rq->lock)
93 {
94 	struct rq *rq;
95 
96 	for (;;) {
97 		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
98 		rq = task_rq(p);
99 		raw_spin_lock(&rq->lock);
100 		/*
101 		 *	move_queued_task()		task_rq_lock()
102 		 *
103 		 *	ACQUIRE (rq->lock)
104 		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
105 		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
106 		 *	[S] ->cpu = new_cpu		[L] task_rq()
107 		 *					[L] ->on_rq
108 		 *	RELEASE (rq->lock)
109 		 *
110 		 * If we observe the old CPU in task_rq_lock, the acquire of
111 		 * the old rq->lock will fully serialize against the stores.
112 		 *
113 		 * If we observe the new CPU in task_rq_lock, the acquire will
114 		 * pair with the WMB to ensure we must then also see migrating.
115 		 */
116 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
117 			rq_pin_lock(rq, rf);
118 			return rq;
119 		}
120 		raw_spin_unlock(&rq->lock);
121 		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
122 
123 		while (unlikely(task_on_rq_migrating(p)))
124 			cpu_relax();
125 	}
126 }
127 
128 /*
129  * RQ-clock updating methods:
130  */
131 
132 static void update_rq_clock_task(struct rq *rq, s64 delta)
133 {
134 /*
135  * In theory, the compile should just see 0 here, and optimize out the call
136  * to sched_rt_avg_update. But I don't trust it...
137  */
138 	s64 __maybe_unused steal = 0, irq_delta = 0;
139 
140 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
141 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
142 
143 	/*
144 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
145 	 * this case when a previous update_rq_clock() happened inside a
146 	 * {soft,}irq region.
147 	 *
148 	 * When this happens, we stop ->clock_task and only update the
149 	 * prev_irq_time stamp to account for the part that fit, so that a next
150 	 * update will consume the rest. This ensures ->clock_task is
151 	 * monotonic.
152 	 *
153 	 * It does however cause some slight miss-attribution of {soft,}irq
154 	 * time, a more accurate solution would be to update the irq_time using
155 	 * the current rq->clock timestamp, except that would require using
156 	 * atomic ops.
157 	 */
158 	if (irq_delta > delta)
159 		irq_delta = delta;
160 
161 	rq->prev_irq_time += irq_delta;
162 	delta -= irq_delta;
163 #endif
164 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
165 	if (static_key_false((&paravirt_steal_rq_enabled))) {
166 		steal = paravirt_steal_clock(cpu_of(rq));
167 		steal -= rq->prev_steal_time_rq;
168 
169 		if (unlikely(steal > delta))
170 			steal = delta;
171 
172 		rq->prev_steal_time_rq += steal;
173 		delta -= steal;
174 	}
175 #endif
176 
177 	rq->clock_task += delta;
178 
179 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
180 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
181 		update_irq_load_avg(rq, irq_delta + steal);
182 #endif
183 }
184 
185 void update_rq_clock(struct rq *rq)
186 {
187 	s64 delta;
188 
189 	lockdep_assert_held(&rq->lock);
190 
191 	if (rq->clock_update_flags & RQCF_ACT_SKIP)
192 		return;
193 
194 #ifdef CONFIG_SCHED_DEBUG
195 	if (sched_feat(WARN_DOUBLE_CLOCK))
196 		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
197 	rq->clock_update_flags |= RQCF_UPDATED;
198 #endif
199 
200 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
201 	if (delta < 0)
202 		return;
203 	rq->clock += delta;
204 	update_rq_clock_task(rq, delta);
205 }
206 
207 
208 #ifdef CONFIG_SCHED_HRTICK
209 /*
210  * Use HR-timers to deliver accurate preemption points.
211  */
212 
213 static void hrtick_clear(struct rq *rq)
214 {
215 	if (hrtimer_active(&rq->hrtick_timer))
216 		hrtimer_cancel(&rq->hrtick_timer);
217 }
218 
219 /*
220  * High-resolution timer tick.
221  * Runs from hardirq context with interrupts disabled.
222  */
223 static enum hrtimer_restart hrtick(struct hrtimer *timer)
224 {
225 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
226 	struct rq_flags rf;
227 
228 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
229 
230 	rq_lock(rq, &rf);
231 	update_rq_clock(rq);
232 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
233 	rq_unlock(rq, &rf);
234 
235 	return HRTIMER_NORESTART;
236 }
237 
238 #ifdef CONFIG_SMP
239 
240 static void __hrtick_restart(struct rq *rq)
241 {
242 	struct hrtimer *timer = &rq->hrtick_timer;
243 
244 	hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
245 }
246 
247 /*
248  * called from hardirq (IPI) context
249  */
250 static void __hrtick_start(void *arg)
251 {
252 	struct rq *rq = arg;
253 	struct rq_flags rf;
254 
255 	rq_lock(rq, &rf);
256 	__hrtick_restart(rq);
257 	rq->hrtick_csd_pending = 0;
258 	rq_unlock(rq, &rf);
259 }
260 
261 /*
262  * Called to set the hrtick timer state.
263  *
264  * called with rq->lock held and irqs disabled
265  */
266 void hrtick_start(struct rq *rq, u64 delay)
267 {
268 	struct hrtimer *timer = &rq->hrtick_timer;
269 	ktime_t time;
270 	s64 delta;
271 
272 	/*
273 	 * Don't schedule slices shorter than 10000ns, that just
274 	 * doesn't make sense and can cause timer DoS.
275 	 */
276 	delta = max_t(s64, delay, 10000LL);
277 	time = ktime_add_ns(timer->base->get_time(), delta);
278 
279 	hrtimer_set_expires(timer, time);
280 
281 	if (rq == this_rq()) {
282 		__hrtick_restart(rq);
283 	} else if (!rq->hrtick_csd_pending) {
284 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
285 		rq->hrtick_csd_pending = 1;
286 	}
287 }
288 
289 #else
290 /*
291  * Called to set the hrtick timer state.
292  *
293  * called with rq->lock held and irqs disabled
294  */
295 void hrtick_start(struct rq *rq, u64 delay)
296 {
297 	/*
298 	 * Don't schedule slices shorter than 10000ns, that just
299 	 * doesn't make sense. Rely on vruntime for fairness.
300 	 */
301 	delay = max_t(u64, delay, 10000LL);
302 	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
303 		      HRTIMER_MODE_REL_PINNED);
304 }
305 #endif /* CONFIG_SMP */
306 
307 static void hrtick_rq_init(struct rq *rq)
308 {
309 #ifdef CONFIG_SMP
310 	rq->hrtick_csd_pending = 0;
311 
312 	rq->hrtick_csd.flags = 0;
313 	rq->hrtick_csd.func = __hrtick_start;
314 	rq->hrtick_csd.info = rq;
315 #endif
316 
317 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
318 	rq->hrtick_timer.function = hrtick;
319 }
320 #else	/* CONFIG_SCHED_HRTICK */
321 static inline void hrtick_clear(struct rq *rq)
322 {
323 }
324 
325 static inline void hrtick_rq_init(struct rq *rq)
326 {
327 }
328 #endif	/* CONFIG_SCHED_HRTICK */
329 
330 /*
331  * cmpxchg based fetch_or, macro so it works for different integer types
332  */
333 #define fetch_or(ptr, mask)						\
334 	({								\
335 		typeof(ptr) _ptr = (ptr);				\
336 		typeof(mask) _mask = (mask);				\
337 		typeof(*_ptr) _old, _val = *_ptr;			\
338 									\
339 		for (;;) {						\
340 			_old = cmpxchg(_ptr, _val, _val | _mask);	\
341 			if (_old == _val)				\
342 				break;					\
343 			_val = _old;					\
344 		}							\
345 	_old;								\
346 })
347 
348 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
349 /*
350  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
351  * this avoids any races wrt polling state changes and thereby avoids
352  * spurious IPIs.
353  */
354 static bool set_nr_and_not_polling(struct task_struct *p)
355 {
356 	struct thread_info *ti = task_thread_info(p);
357 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
358 }
359 
360 /*
361  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
362  *
363  * If this returns true, then the idle task promises to call
364  * sched_ttwu_pending() and reschedule soon.
365  */
366 static bool set_nr_if_polling(struct task_struct *p)
367 {
368 	struct thread_info *ti = task_thread_info(p);
369 	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
370 
371 	for (;;) {
372 		if (!(val & _TIF_POLLING_NRFLAG))
373 			return false;
374 		if (val & _TIF_NEED_RESCHED)
375 			return true;
376 		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
377 		if (old == val)
378 			break;
379 		val = old;
380 	}
381 	return true;
382 }
383 
384 #else
385 static bool set_nr_and_not_polling(struct task_struct *p)
386 {
387 	set_tsk_need_resched(p);
388 	return true;
389 }
390 
391 #ifdef CONFIG_SMP
392 static bool set_nr_if_polling(struct task_struct *p)
393 {
394 	return false;
395 }
396 #endif
397 #endif
398 
399 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
400 {
401 	struct wake_q_node *node = &task->wake_q;
402 
403 	/*
404 	 * Atomically grab the task, if ->wake_q is !nil already it means
405 	 * its already queued (either by us or someone else) and will get the
406 	 * wakeup due to that.
407 	 *
408 	 * This cmpxchg() executes a full barrier, which pairs with the full
409 	 * barrier executed by the wakeup in wake_up_q().
410 	 */
411 	if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
412 		return;
413 
414 	get_task_struct(task);
415 
416 	/*
417 	 * The head is context local, there can be no concurrency.
418 	 */
419 	*head->lastp = node;
420 	head->lastp = &node->next;
421 }
422 
423 void wake_up_q(struct wake_q_head *head)
424 {
425 	struct wake_q_node *node = head->first;
426 
427 	while (node != WAKE_Q_TAIL) {
428 		struct task_struct *task;
429 
430 		task = container_of(node, struct task_struct, wake_q);
431 		BUG_ON(!task);
432 		/* Task can safely be re-inserted now: */
433 		node = node->next;
434 		task->wake_q.next = NULL;
435 
436 		/*
437 		 * wake_up_process() executes a full barrier, which pairs with
438 		 * the queueing in wake_q_add() so as not to miss wakeups.
439 		 */
440 		wake_up_process(task);
441 		put_task_struct(task);
442 	}
443 }
444 
445 /*
446  * resched_curr - mark rq's current task 'to be rescheduled now'.
447  *
448  * On UP this means the setting of the need_resched flag, on SMP it
449  * might also involve a cross-CPU call to trigger the scheduler on
450  * the target CPU.
451  */
452 void resched_curr(struct rq *rq)
453 {
454 	struct task_struct *curr = rq->curr;
455 	int cpu;
456 
457 	lockdep_assert_held(&rq->lock);
458 
459 	if (test_tsk_need_resched(curr))
460 		return;
461 
462 	cpu = cpu_of(rq);
463 
464 	if (cpu == smp_processor_id()) {
465 		set_tsk_need_resched(curr);
466 		set_preempt_need_resched();
467 		return;
468 	}
469 
470 	if (set_nr_and_not_polling(curr))
471 		smp_send_reschedule(cpu);
472 	else
473 		trace_sched_wake_idle_without_ipi(cpu);
474 }
475 
476 void resched_cpu(int cpu)
477 {
478 	struct rq *rq = cpu_rq(cpu);
479 	unsigned long flags;
480 
481 	raw_spin_lock_irqsave(&rq->lock, flags);
482 	if (cpu_online(cpu) || cpu == smp_processor_id())
483 		resched_curr(rq);
484 	raw_spin_unlock_irqrestore(&rq->lock, flags);
485 }
486 
487 #ifdef CONFIG_SMP
488 #ifdef CONFIG_NO_HZ_COMMON
489 /*
490  * In the semi idle case, use the nearest busy CPU for migrating timers
491  * from an idle CPU.  This is good for power-savings.
492  *
493  * We don't do similar optimization for completely idle system, as
494  * selecting an idle CPU will add more delays to the timers than intended
495  * (as that CPU's timer base may not be uptodate wrt jiffies etc).
496  */
497 int get_nohz_timer_target(void)
498 {
499 	int i, cpu = smp_processor_id();
500 	struct sched_domain *sd;
501 
502 	if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
503 		return cpu;
504 
505 	rcu_read_lock();
506 	for_each_domain(cpu, sd) {
507 		for_each_cpu(i, sched_domain_span(sd)) {
508 			if (cpu == i)
509 				continue;
510 
511 			if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
512 				cpu = i;
513 				goto unlock;
514 			}
515 		}
516 	}
517 
518 	if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
519 		cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
520 unlock:
521 	rcu_read_unlock();
522 	return cpu;
523 }
524 
525 /*
526  * When add_timer_on() enqueues a timer into the timer wheel of an
527  * idle CPU then this timer might expire before the next timer event
528  * which is scheduled to wake up that CPU. In case of a completely
529  * idle system the next event might even be infinite time into the
530  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
531  * leaves the inner idle loop so the newly added timer is taken into
532  * account when the CPU goes back to idle and evaluates the timer
533  * wheel for the next timer event.
534  */
535 static void wake_up_idle_cpu(int cpu)
536 {
537 	struct rq *rq = cpu_rq(cpu);
538 
539 	if (cpu == smp_processor_id())
540 		return;
541 
542 	if (set_nr_and_not_polling(rq->idle))
543 		smp_send_reschedule(cpu);
544 	else
545 		trace_sched_wake_idle_without_ipi(cpu);
546 }
547 
548 static bool wake_up_full_nohz_cpu(int cpu)
549 {
550 	/*
551 	 * We just need the target to call irq_exit() and re-evaluate
552 	 * the next tick. The nohz full kick at least implies that.
553 	 * If needed we can still optimize that later with an
554 	 * empty IRQ.
555 	 */
556 	if (cpu_is_offline(cpu))
557 		return true;  /* Don't try to wake offline CPUs. */
558 	if (tick_nohz_full_cpu(cpu)) {
559 		if (cpu != smp_processor_id() ||
560 		    tick_nohz_tick_stopped())
561 			tick_nohz_full_kick_cpu(cpu);
562 		return true;
563 	}
564 
565 	return false;
566 }
567 
568 /*
569  * Wake up the specified CPU.  If the CPU is going offline, it is the
570  * caller's responsibility to deal with the lost wakeup, for example,
571  * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
572  */
573 void wake_up_nohz_cpu(int cpu)
574 {
575 	if (!wake_up_full_nohz_cpu(cpu))
576 		wake_up_idle_cpu(cpu);
577 }
578 
579 static inline bool got_nohz_idle_kick(void)
580 {
581 	int cpu = smp_processor_id();
582 
583 	if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
584 		return false;
585 
586 	if (idle_cpu(cpu) && !need_resched())
587 		return true;
588 
589 	/*
590 	 * We can't run Idle Load Balance on this CPU for this time so we
591 	 * cancel it and clear NOHZ_BALANCE_KICK
592 	 */
593 	atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
594 	return false;
595 }
596 
597 #else /* CONFIG_NO_HZ_COMMON */
598 
599 static inline bool got_nohz_idle_kick(void)
600 {
601 	return false;
602 }
603 
604 #endif /* CONFIG_NO_HZ_COMMON */
605 
606 #ifdef CONFIG_NO_HZ_FULL
607 bool sched_can_stop_tick(struct rq *rq)
608 {
609 	int fifo_nr_running;
610 
611 	/* Deadline tasks, even if single, need the tick */
612 	if (rq->dl.dl_nr_running)
613 		return false;
614 
615 	/*
616 	 * If there are more than one RR tasks, we need the tick to effect the
617 	 * actual RR behaviour.
618 	 */
619 	if (rq->rt.rr_nr_running) {
620 		if (rq->rt.rr_nr_running == 1)
621 			return true;
622 		else
623 			return false;
624 	}
625 
626 	/*
627 	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
628 	 * forced preemption between FIFO tasks.
629 	 */
630 	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
631 	if (fifo_nr_running)
632 		return true;
633 
634 	/*
635 	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
636 	 * if there's more than one we need the tick for involuntary
637 	 * preemption.
638 	 */
639 	if (rq->nr_running > 1)
640 		return false;
641 
642 	return true;
643 }
644 #endif /* CONFIG_NO_HZ_FULL */
645 #endif /* CONFIG_SMP */
646 
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
649 /*
650  * Iterate task_group tree rooted at *from, calling @down when first entering a
651  * node and @up when leaving it for the final time.
652  *
653  * Caller must hold rcu_lock or sufficient equivalent.
654  */
655 int walk_tg_tree_from(struct task_group *from,
656 			     tg_visitor down, tg_visitor up, void *data)
657 {
658 	struct task_group *parent, *child;
659 	int ret;
660 
661 	parent = from;
662 
663 down:
664 	ret = (*down)(parent, data);
665 	if (ret)
666 		goto out;
667 	list_for_each_entry_rcu(child, &parent->children, siblings) {
668 		parent = child;
669 		goto down;
670 
671 up:
672 		continue;
673 	}
674 	ret = (*up)(parent, data);
675 	if (ret || parent == from)
676 		goto out;
677 
678 	child = parent;
679 	parent = parent->parent;
680 	if (parent)
681 		goto up;
682 out:
683 	return ret;
684 }
685 
686 int tg_nop(struct task_group *tg, void *data)
687 {
688 	return 0;
689 }
690 #endif
691 
692 static void set_load_weight(struct task_struct *p, bool update_load)
693 {
694 	int prio = p->static_prio - MAX_RT_PRIO;
695 	struct load_weight *load = &p->se.load;
696 
697 	/*
698 	 * SCHED_IDLE tasks get minimal weight:
699 	 */
700 	if (task_has_idle_policy(p)) {
701 		load->weight = scale_load(WEIGHT_IDLEPRIO);
702 		load->inv_weight = WMULT_IDLEPRIO;
703 		p->se.runnable_weight = load->weight;
704 		return;
705 	}
706 
707 	/*
708 	 * SCHED_OTHER tasks have to update their load when changing their
709 	 * weight
710 	 */
711 	if (update_load && p->sched_class == &fair_sched_class) {
712 		reweight_task(p, prio);
713 	} else {
714 		load->weight = scale_load(sched_prio_to_weight[prio]);
715 		load->inv_weight = sched_prio_to_wmult[prio];
716 		p->se.runnable_weight = load->weight;
717 	}
718 }
719 
720 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
721 {
722 	if (!(flags & ENQUEUE_NOCLOCK))
723 		update_rq_clock(rq);
724 
725 	if (!(flags & ENQUEUE_RESTORE)) {
726 		sched_info_queued(rq, p);
727 		psi_enqueue(p, flags & ENQUEUE_WAKEUP);
728 	}
729 
730 	p->sched_class->enqueue_task(rq, p, flags);
731 }
732 
733 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
734 {
735 	if (!(flags & DEQUEUE_NOCLOCK))
736 		update_rq_clock(rq);
737 
738 	if (!(flags & DEQUEUE_SAVE)) {
739 		sched_info_dequeued(rq, p);
740 		psi_dequeue(p, flags & DEQUEUE_SLEEP);
741 	}
742 
743 	p->sched_class->dequeue_task(rq, p, flags);
744 }
745 
746 void activate_task(struct rq *rq, struct task_struct *p, int flags)
747 {
748 	if (task_contributes_to_load(p))
749 		rq->nr_uninterruptible--;
750 
751 	enqueue_task(rq, p, flags);
752 }
753 
754 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
755 {
756 	if (task_contributes_to_load(p))
757 		rq->nr_uninterruptible++;
758 
759 	dequeue_task(rq, p, flags);
760 }
761 
762 /*
763  * __normal_prio - return the priority that is based on the static prio
764  */
765 static inline int __normal_prio(struct task_struct *p)
766 {
767 	return p->static_prio;
768 }
769 
770 /*
771  * Calculate the expected normal priority: i.e. priority
772  * without taking RT-inheritance into account. Might be
773  * boosted by interactivity modifiers. Changes upon fork,
774  * setprio syscalls, and whenever the interactivity
775  * estimator recalculates.
776  */
777 static inline int normal_prio(struct task_struct *p)
778 {
779 	int prio;
780 
781 	if (task_has_dl_policy(p))
782 		prio = MAX_DL_PRIO-1;
783 	else if (task_has_rt_policy(p))
784 		prio = MAX_RT_PRIO-1 - p->rt_priority;
785 	else
786 		prio = __normal_prio(p);
787 	return prio;
788 }
789 
790 /*
791  * Calculate the current priority, i.e. the priority
792  * taken into account by the scheduler. This value might
793  * be boosted by RT tasks, or might be boosted by
794  * interactivity modifiers. Will be RT if the task got
795  * RT-boosted. If not then it returns p->normal_prio.
796  */
797 static int effective_prio(struct task_struct *p)
798 {
799 	p->normal_prio = normal_prio(p);
800 	/*
801 	 * If we are RT tasks or we were boosted to RT priority,
802 	 * keep the priority unchanged. Otherwise, update priority
803 	 * to the normal priority:
804 	 */
805 	if (!rt_prio(p->prio))
806 		return p->normal_prio;
807 	return p->prio;
808 }
809 
810 /**
811  * task_curr - is this task currently executing on a CPU?
812  * @p: the task in question.
813  *
814  * Return: 1 if the task is currently executing. 0 otherwise.
815  */
816 inline int task_curr(const struct task_struct *p)
817 {
818 	return cpu_curr(task_cpu(p)) == p;
819 }
820 
821 /*
822  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
823  * use the balance_callback list if you want balancing.
824  *
825  * this means any call to check_class_changed() must be followed by a call to
826  * balance_callback().
827  */
828 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
829 				       const struct sched_class *prev_class,
830 				       int oldprio)
831 {
832 	if (prev_class != p->sched_class) {
833 		if (prev_class->switched_from)
834 			prev_class->switched_from(rq, p);
835 
836 		p->sched_class->switched_to(rq, p);
837 	} else if (oldprio != p->prio || dl_task(p))
838 		p->sched_class->prio_changed(rq, p, oldprio);
839 }
840 
841 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
842 {
843 	const struct sched_class *class;
844 
845 	if (p->sched_class == rq->curr->sched_class) {
846 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
847 	} else {
848 		for_each_class(class) {
849 			if (class == rq->curr->sched_class)
850 				break;
851 			if (class == p->sched_class) {
852 				resched_curr(rq);
853 				break;
854 			}
855 		}
856 	}
857 
858 	/*
859 	 * A queue event has occurred, and we're going to schedule.  In
860 	 * this case, we can save a useless back to back clock update.
861 	 */
862 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
863 		rq_clock_skip_update(rq);
864 }
865 
866 #ifdef CONFIG_SMP
867 
868 static inline bool is_per_cpu_kthread(struct task_struct *p)
869 {
870 	if (!(p->flags & PF_KTHREAD))
871 		return false;
872 
873 	if (p->nr_cpus_allowed != 1)
874 		return false;
875 
876 	return true;
877 }
878 
879 /*
880  * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
881  * __set_cpus_allowed_ptr() and select_fallback_rq().
882  */
883 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
884 {
885 	if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
886 		return false;
887 
888 	if (is_per_cpu_kthread(p))
889 		return cpu_online(cpu);
890 
891 	return cpu_active(cpu);
892 }
893 
894 /*
895  * This is how migration works:
896  *
897  * 1) we invoke migration_cpu_stop() on the target CPU using
898  *    stop_one_cpu().
899  * 2) stopper starts to run (implicitly forcing the migrated thread
900  *    off the CPU)
901  * 3) it checks whether the migrated task is still in the wrong runqueue.
902  * 4) if it's in the wrong runqueue then the migration thread removes
903  *    it and puts it into the right queue.
904  * 5) stopper completes and stop_one_cpu() returns and the migration
905  *    is done.
906  */
907 
908 /*
909  * move_queued_task - move a queued task to new rq.
910  *
911  * Returns (locked) new rq. Old rq's lock is released.
912  */
913 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
914 				   struct task_struct *p, int new_cpu)
915 {
916 	lockdep_assert_held(&rq->lock);
917 
918 	p->on_rq = TASK_ON_RQ_MIGRATING;
919 	dequeue_task(rq, p, DEQUEUE_NOCLOCK);
920 	set_task_cpu(p, new_cpu);
921 	rq_unlock(rq, rf);
922 
923 	rq = cpu_rq(new_cpu);
924 
925 	rq_lock(rq, rf);
926 	BUG_ON(task_cpu(p) != new_cpu);
927 	enqueue_task(rq, p, 0);
928 	p->on_rq = TASK_ON_RQ_QUEUED;
929 	check_preempt_curr(rq, p, 0);
930 
931 	return rq;
932 }
933 
934 struct migration_arg {
935 	struct task_struct *task;
936 	int dest_cpu;
937 };
938 
939 /*
940  * Move (not current) task off this CPU, onto the destination CPU. We're doing
941  * this because either it can't run here any more (set_cpus_allowed()
942  * away from this CPU, or CPU going down), or because we're
943  * attempting to rebalance this task on exec (sched_exec).
944  *
945  * So we race with normal scheduler movements, but that's OK, as long
946  * as the task is no longer on this CPU.
947  */
948 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
949 				 struct task_struct *p, int dest_cpu)
950 {
951 	/* Affinity changed (again). */
952 	if (!is_cpu_allowed(p, dest_cpu))
953 		return rq;
954 
955 	update_rq_clock(rq);
956 	rq = move_queued_task(rq, rf, p, dest_cpu);
957 
958 	return rq;
959 }
960 
961 /*
962  * migration_cpu_stop - this will be executed by a highprio stopper thread
963  * and performs thread migration by bumping thread off CPU then
964  * 'pushing' onto another runqueue.
965  */
966 static int migration_cpu_stop(void *data)
967 {
968 	struct migration_arg *arg = data;
969 	struct task_struct *p = arg->task;
970 	struct rq *rq = this_rq();
971 	struct rq_flags rf;
972 
973 	/*
974 	 * The original target CPU might have gone down and we might
975 	 * be on another CPU but it doesn't matter.
976 	 */
977 	local_irq_disable();
978 	/*
979 	 * We need to explicitly wake pending tasks before running
980 	 * __migrate_task() such that we will not miss enforcing cpus_allowed
981 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
982 	 */
983 	sched_ttwu_pending();
984 
985 	raw_spin_lock(&p->pi_lock);
986 	rq_lock(rq, &rf);
987 	/*
988 	 * If task_rq(p) != rq, it cannot be migrated here, because we're
989 	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
990 	 * we're holding p->pi_lock.
991 	 */
992 	if (task_rq(p) == rq) {
993 		if (task_on_rq_queued(p))
994 			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
995 		else
996 			p->wake_cpu = arg->dest_cpu;
997 	}
998 	rq_unlock(rq, &rf);
999 	raw_spin_unlock(&p->pi_lock);
1000 
1001 	local_irq_enable();
1002 	return 0;
1003 }
1004 
1005 /*
1006  * sched_class::set_cpus_allowed must do the below, but is not required to
1007  * actually call this function.
1008  */
1009 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1010 {
1011 	cpumask_copy(&p->cpus_allowed, new_mask);
1012 	p->nr_cpus_allowed = cpumask_weight(new_mask);
1013 }
1014 
1015 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1016 {
1017 	struct rq *rq = task_rq(p);
1018 	bool queued, running;
1019 
1020 	lockdep_assert_held(&p->pi_lock);
1021 
1022 	queued = task_on_rq_queued(p);
1023 	running = task_current(rq, p);
1024 
1025 	if (queued) {
1026 		/*
1027 		 * Because __kthread_bind() calls this on blocked tasks without
1028 		 * holding rq->lock.
1029 		 */
1030 		lockdep_assert_held(&rq->lock);
1031 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1032 	}
1033 	if (running)
1034 		put_prev_task(rq, p);
1035 
1036 	p->sched_class->set_cpus_allowed(p, new_mask);
1037 
1038 	if (queued)
1039 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1040 	if (running)
1041 		set_curr_task(rq, p);
1042 }
1043 
1044 /*
1045  * Change a given task's CPU affinity. Migrate the thread to a
1046  * proper CPU and schedule it away if the CPU it's executing on
1047  * is removed from the allowed bitmask.
1048  *
1049  * NOTE: the caller must have a valid reference to the task, the
1050  * task must not exit() & deallocate itself prematurely. The
1051  * call is not atomic; no spinlocks may be held.
1052  */
1053 static int __set_cpus_allowed_ptr(struct task_struct *p,
1054 				  const struct cpumask *new_mask, bool check)
1055 {
1056 	const struct cpumask *cpu_valid_mask = cpu_active_mask;
1057 	unsigned int dest_cpu;
1058 	struct rq_flags rf;
1059 	struct rq *rq;
1060 	int ret = 0;
1061 
1062 	rq = task_rq_lock(p, &rf);
1063 	update_rq_clock(rq);
1064 
1065 	if (p->flags & PF_KTHREAD) {
1066 		/*
1067 		 * Kernel threads are allowed on online && !active CPUs
1068 		 */
1069 		cpu_valid_mask = cpu_online_mask;
1070 	}
1071 
1072 	/*
1073 	 * Must re-check here, to close a race against __kthread_bind(),
1074 	 * sched_setaffinity() is not guaranteed to observe the flag.
1075 	 */
1076 	if (check && (p->flags & PF_NO_SETAFFINITY)) {
1077 		ret = -EINVAL;
1078 		goto out;
1079 	}
1080 
1081 	if (cpumask_equal(&p->cpus_allowed, new_mask))
1082 		goto out;
1083 
1084 	if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1085 		ret = -EINVAL;
1086 		goto out;
1087 	}
1088 
1089 	do_set_cpus_allowed(p, new_mask);
1090 
1091 	if (p->flags & PF_KTHREAD) {
1092 		/*
1093 		 * For kernel threads that do indeed end up on online &&
1094 		 * !active we want to ensure they are strict per-CPU threads.
1095 		 */
1096 		WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1097 			!cpumask_intersects(new_mask, cpu_active_mask) &&
1098 			p->nr_cpus_allowed != 1);
1099 	}
1100 
1101 	/* Can the task run on the task's current CPU? If so, we're done */
1102 	if (cpumask_test_cpu(task_cpu(p), new_mask))
1103 		goto out;
1104 
1105 	dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1106 	if (task_running(rq, p) || p->state == TASK_WAKING) {
1107 		struct migration_arg arg = { p, dest_cpu };
1108 		/* Need help from migration thread: drop lock and wait. */
1109 		task_rq_unlock(rq, p, &rf);
1110 		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1111 		tlb_migrate_finish(p->mm);
1112 		return 0;
1113 	} else if (task_on_rq_queued(p)) {
1114 		/*
1115 		 * OK, since we're going to drop the lock immediately
1116 		 * afterwards anyway.
1117 		 */
1118 		rq = move_queued_task(rq, &rf, p, dest_cpu);
1119 	}
1120 out:
1121 	task_rq_unlock(rq, p, &rf);
1122 
1123 	return ret;
1124 }
1125 
1126 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1127 {
1128 	return __set_cpus_allowed_ptr(p, new_mask, false);
1129 }
1130 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1131 
1132 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1133 {
1134 #ifdef CONFIG_SCHED_DEBUG
1135 	/*
1136 	 * We should never call set_task_cpu() on a blocked task,
1137 	 * ttwu() will sort out the placement.
1138 	 */
1139 	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1140 			!p->on_rq);
1141 
1142 	/*
1143 	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1144 	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1145 	 * time relying on p->on_rq.
1146 	 */
1147 	WARN_ON_ONCE(p->state == TASK_RUNNING &&
1148 		     p->sched_class == &fair_sched_class &&
1149 		     (p->on_rq && !task_on_rq_migrating(p)));
1150 
1151 #ifdef CONFIG_LOCKDEP
1152 	/*
1153 	 * The caller should hold either p->pi_lock or rq->lock, when changing
1154 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1155 	 *
1156 	 * sched_move_task() holds both and thus holding either pins the cgroup,
1157 	 * see task_group().
1158 	 *
1159 	 * Furthermore, all task_rq users should acquire both locks, see
1160 	 * task_rq_lock().
1161 	 */
1162 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1163 				      lockdep_is_held(&task_rq(p)->lock)));
1164 #endif
1165 	/*
1166 	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1167 	 */
1168 	WARN_ON_ONCE(!cpu_online(new_cpu));
1169 #endif
1170 
1171 	trace_sched_migrate_task(p, new_cpu);
1172 
1173 	if (task_cpu(p) != new_cpu) {
1174 		if (p->sched_class->migrate_task_rq)
1175 			p->sched_class->migrate_task_rq(p, new_cpu);
1176 		p->se.nr_migrations++;
1177 		rseq_migrate(p);
1178 		perf_event_task_migrate(p);
1179 	}
1180 
1181 	__set_task_cpu(p, new_cpu);
1182 }
1183 
1184 #ifdef CONFIG_NUMA_BALANCING
1185 static void __migrate_swap_task(struct task_struct *p, int cpu)
1186 {
1187 	if (task_on_rq_queued(p)) {
1188 		struct rq *src_rq, *dst_rq;
1189 		struct rq_flags srf, drf;
1190 
1191 		src_rq = task_rq(p);
1192 		dst_rq = cpu_rq(cpu);
1193 
1194 		rq_pin_lock(src_rq, &srf);
1195 		rq_pin_lock(dst_rq, &drf);
1196 
1197 		p->on_rq = TASK_ON_RQ_MIGRATING;
1198 		deactivate_task(src_rq, p, 0);
1199 		set_task_cpu(p, cpu);
1200 		activate_task(dst_rq, p, 0);
1201 		p->on_rq = TASK_ON_RQ_QUEUED;
1202 		check_preempt_curr(dst_rq, p, 0);
1203 
1204 		rq_unpin_lock(dst_rq, &drf);
1205 		rq_unpin_lock(src_rq, &srf);
1206 
1207 	} else {
1208 		/*
1209 		 * Task isn't running anymore; make it appear like we migrated
1210 		 * it before it went to sleep. This means on wakeup we make the
1211 		 * previous CPU our target instead of where it really is.
1212 		 */
1213 		p->wake_cpu = cpu;
1214 	}
1215 }
1216 
1217 struct migration_swap_arg {
1218 	struct task_struct *src_task, *dst_task;
1219 	int src_cpu, dst_cpu;
1220 };
1221 
1222 static int migrate_swap_stop(void *data)
1223 {
1224 	struct migration_swap_arg *arg = data;
1225 	struct rq *src_rq, *dst_rq;
1226 	int ret = -EAGAIN;
1227 
1228 	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1229 		return -EAGAIN;
1230 
1231 	src_rq = cpu_rq(arg->src_cpu);
1232 	dst_rq = cpu_rq(arg->dst_cpu);
1233 
1234 	double_raw_lock(&arg->src_task->pi_lock,
1235 			&arg->dst_task->pi_lock);
1236 	double_rq_lock(src_rq, dst_rq);
1237 
1238 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
1239 		goto unlock;
1240 
1241 	if (task_cpu(arg->src_task) != arg->src_cpu)
1242 		goto unlock;
1243 
1244 	if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1245 		goto unlock;
1246 
1247 	if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1248 		goto unlock;
1249 
1250 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
1251 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
1252 
1253 	ret = 0;
1254 
1255 unlock:
1256 	double_rq_unlock(src_rq, dst_rq);
1257 	raw_spin_unlock(&arg->dst_task->pi_lock);
1258 	raw_spin_unlock(&arg->src_task->pi_lock);
1259 
1260 	return ret;
1261 }
1262 
1263 /*
1264  * Cross migrate two tasks
1265  */
1266 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1267 		int target_cpu, int curr_cpu)
1268 {
1269 	struct migration_swap_arg arg;
1270 	int ret = -EINVAL;
1271 
1272 	arg = (struct migration_swap_arg){
1273 		.src_task = cur,
1274 		.src_cpu = curr_cpu,
1275 		.dst_task = p,
1276 		.dst_cpu = target_cpu,
1277 	};
1278 
1279 	if (arg.src_cpu == arg.dst_cpu)
1280 		goto out;
1281 
1282 	/*
1283 	 * These three tests are all lockless; this is OK since all of them
1284 	 * will be re-checked with proper locks held further down the line.
1285 	 */
1286 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1287 		goto out;
1288 
1289 	if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1290 		goto out;
1291 
1292 	if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1293 		goto out;
1294 
1295 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1296 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1297 
1298 out:
1299 	return ret;
1300 }
1301 #endif /* CONFIG_NUMA_BALANCING */
1302 
1303 /*
1304  * wait_task_inactive - wait for a thread to unschedule.
1305  *
1306  * If @match_state is nonzero, it's the @p->state value just checked and
1307  * not expected to change.  If it changes, i.e. @p might have woken up,
1308  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1309  * we return a positive number (its total switch count).  If a second call
1310  * a short while later returns the same number, the caller can be sure that
1311  * @p has remained unscheduled the whole time.
1312  *
1313  * The caller must ensure that the task *will* unschedule sometime soon,
1314  * else this function might spin for a *long* time. This function can't
1315  * be called with interrupts off, or it may introduce deadlock with
1316  * smp_call_function() if an IPI is sent by the same process we are
1317  * waiting to become inactive.
1318  */
1319 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1320 {
1321 	int running, queued;
1322 	struct rq_flags rf;
1323 	unsigned long ncsw;
1324 	struct rq *rq;
1325 
1326 	for (;;) {
1327 		/*
1328 		 * We do the initial early heuristics without holding
1329 		 * any task-queue locks at all. We'll only try to get
1330 		 * the runqueue lock when things look like they will
1331 		 * work out!
1332 		 */
1333 		rq = task_rq(p);
1334 
1335 		/*
1336 		 * If the task is actively running on another CPU
1337 		 * still, just relax and busy-wait without holding
1338 		 * any locks.
1339 		 *
1340 		 * NOTE! Since we don't hold any locks, it's not
1341 		 * even sure that "rq" stays as the right runqueue!
1342 		 * But we don't care, since "task_running()" will
1343 		 * return false if the runqueue has changed and p
1344 		 * is actually now running somewhere else!
1345 		 */
1346 		while (task_running(rq, p)) {
1347 			if (match_state && unlikely(p->state != match_state))
1348 				return 0;
1349 			cpu_relax();
1350 		}
1351 
1352 		/*
1353 		 * Ok, time to look more closely! We need the rq
1354 		 * lock now, to be *sure*. If we're wrong, we'll
1355 		 * just go back and repeat.
1356 		 */
1357 		rq = task_rq_lock(p, &rf);
1358 		trace_sched_wait_task(p);
1359 		running = task_running(rq, p);
1360 		queued = task_on_rq_queued(p);
1361 		ncsw = 0;
1362 		if (!match_state || p->state == match_state)
1363 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1364 		task_rq_unlock(rq, p, &rf);
1365 
1366 		/*
1367 		 * If it changed from the expected state, bail out now.
1368 		 */
1369 		if (unlikely(!ncsw))
1370 			break;
1371 
1372 		/*
1373 		 * Was it really running after all now that we
1374 		 * checked with the proper locks actually held?
1375 		 *
1376 		 * Oops. Go back and try again..
1377 		 */
1378 		if (unlikely(running)) {
1379 			cpu_relax();
1380 			continue;
1381 		}
1382 
1383 		/*
1384 		 * It's not enough that it's not actively running,
1385 		 * it must be off the runqueue _entirely_, and not
1386 		 * preempted!
1387 		 *
1388 		 * So if it was still runnable (but just not actively
1389 		 * running right now), it's preempted, and we should
1390 		 * yield - it could be a while.
1391 		 */
1392 		if (unlikely(queued)) {
1393 			ktime_t to = NSEC_PER_SEC / HZ;
1394 
1395 			set_current_state(TASK_UNINTERRUPTIBLE);
1396 			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1397 			continue;
1398 		}
1399 
1400 		/*
1401 		 * Ahh, all good. It wasn't running, and it wasn't
1402 		 * runnable, which means that it will never become
1403 		 * running in the future either. We're all done!
1404 		 */
1405 		break;
1406 	}
1407 
1408 	return ncsw;
1409 }
1410 
1411 /***
1412  * kick_process - kick a running thread to enter/exit the kernel
1413  * @p: the to-be-kicked thread
1414  *
1415  * Cause a process which is running on another CPU to enter
1416  * kernel-mode, without any delay. (to get signals handled.)
1417  *
1418  * NOTE: this function doesn't have to take the runqueue lock,
1419  * because all it wants to ensure is that the remote task enters
1420  * the kernel. If the IPI races and the task has been migrated
1421  * to another CPU then no harm is done and the purpose has been
1422  * achieved as well.
1423  */
1424 void kick_process(struct task_struct *p)
1425 {
1426 	int cpu;
1427 
1428 	preempt_disable();
1429 	cpu = task_cpu(p);
1430 	if ((cpu != smp_processor_id()) && task_curr(p))
1431 		smp_send_reschedule(cpu);
1432 	preempt_enable();
1433 }
1434 EXPORT_SYMBOL_GPL(kick_process);
1435 
1436 /*
1437  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1438  *
1439  * A few notes on cpu_active vs cpu_online:
1440  *
1441  *  - cpu_active must be a subset of cpu_online
1442  *
1443  *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1444  *    see __set_cpus_allowed_ptr(). At this point the newly online
1445  *    CPU isn't yet part of the sched domains, and balancing will not
1446  *    see it.
1447  *
1448  *  - on CPU-down we clear cpu_active() to mask the sched domains and
1449  *    avoid the load balancer to place new tasks on the to be removed
1450  *    CPU. Existing tasks will remain running there and will be taken
1451  *    off.
1452  *
1453  * This means that fallback selection must not select !active CPUs.
1454  * And can assume that any active CPU must be online. Conversely
1455  * select_task_rq() below may allow selection of !active CPUs in order
1456  * to satisfy the above rules.
1457  */
1458 static int select_fallback_rq(int cpu, struct task_struct *p)
1459 {
1460 	int nid = cpu_to_node(cpu);
1461 	const struct cpumask *nodemask = NULL;
1462 	enum { cpuset, possible, fail } state = cpuset;
1463 	int dest_cpu;
1464 
1465 	/*
1466 	 * If the node that the CPU is on has been offlined, cpu_to_node()
1467 	 * will return -1. There is no CPU on the node, and we should
1468 	 * select the CPU on the other node.
1469 	 */
1470 	if (nid != -1) {
1471 		nodemask = cpumask_of_node(nid);
1472 
1473 		/* Look for allowed, online CPU in same node. */
1474 		for_each_cpu(dest_cpu, nodemask) {
1475 			if (!cpu_active(dest_cpu))
1476 				continue;
1477 			if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1478 				return dest_cpu;
1479 		}
1480 	}
1481 
1482 	for (;;) {
1483 		/* Any allowed, online CPU? */
1484 		for_each_cpu(dest_cpu, &p->cpus_allowed) {
1485 			if (!is_cpu_allowed(p, dest_cpu))
1486 				continue;
1487 
1488 			goto out;
1489 		}
1490 
1491 		/* No more Mr. Nice Guy. */
1492 		switch (state) {
1493 		case cpuset:
1494 			if (IS_ENABLED(CONFIG_CPUSETS)) {
1495 				cpuset_cpus_allowed_fallback(p);
1496 				state = possible;
1497 				break;
1498 			}
1499 			/* Fall-through */
1500 		case possible:
1501 			do_set_cpus_allowed(p, cpu_possible_mask);
1502 			state = fail;
1503 			break;
1504 
1505 		case fail:
1506 			BUG();
1507 			break;
1508 		}
1509 	}
1510 
1511 out:
1512 	if (state != cpuset) {
1513 		/*
1514 		 * Don't tell them about moving exiting tasks or
1515 		 * kernel threads (both mm NULL), since they never
1516 		 * leave kernel.
1517 		 */
1518 		if (p->mm && printk_ratelimit()) {
1519 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1520 					task_pid_nr(p), p->comm, cpu);
1521 		}
1522 	}
1523 
1524 	return dest_cpu;
1525 }
1526 
1527 /*
1528  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1529  */
1530 static inline
1531 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1532 {
1533 	lockdep_assert_held(&p->pi_lock);
1534 
1535 	if (p->nr_cpus_allowed > 1)
1536 		cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1537 	else
1538 		cpu = cpumask_any(&p->cpus_allowed);
1539 
1540 	/*
1541 	 * In order not to call set_task_cpu() on a blocking task we need
1542 	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1543 	 * CPU.
1544 	 *
1545 	 * Since this is common to all placement strategies, this lives here.
1546 	 *
1547 	 * [ this allows ->select_task() to simply return task_cpu(p) and
1548 	 *   not worry about this generic constraint ]
1549 	 */
1550 	if (unlikely(!is_cpu_allowed(p, cpu)))
1551 		cpu = select_fallback_rq(task_cpu(p), p);
1552 
1553 	return cpu;
1554 }
1555 
1556 static void update_avg(u64 *avg, u64 sample)
1557 {
1558 	s64 diff = sample - *avg;
1559 	*avg += diff >> 3;
1560 }
1561 
1562 void sched_set_stop_task(int cpu, struct task_struct *stop)
1563 {
1564 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1565 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
1566 
1567 	if (stop) {
1568 		/*
1569 		 * Make it appear like a SCHED_FIFO task, its something
1570 		 * userspace knows about and won't get confused about.
1571 		 *
1572 		 * Also, it will make PI more or less work without too
1573 		 * much confusion -- but then, stop work should not
1574 		 * rely on PI working anyway.
1575 		 */
1576 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
1577 
1578 		stop->sched_class = &stop_sched_class;
1579 	}
1580 
1581 	cpu_rq(cpu)->stop = stop;
1582 
1583 	if (old_stop) {
1584 		/*
1585 		 * Reset it back to a normal scheduling class so that
1586 		 * it can die in pieces.
1587 		 */
1588 		old_stop->sched_class = &rt_sched_class;
1589 	}
1590 }
1591 
1592 #else
1593 
1594 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1595 					 const struct cpumask *new_mask, bool check)
1596 {
1597 	return set_cpus_allowed_ptr(p, new_mask);
1598 }
1599 
1600 #endif /* CONFIG_SMP */
1601 
1602 static void
1603 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1604 {
1605 	struct rq *rq;
1606 
1607 	if (!schedstat_enabled())
1608 		return;
1609 
1610 	rq = this_rq();
1611 
1612 #ifdef CONFIG_SMP
1613 	if (cpu == rq->cpu) {
1614 		__schedstat_inc(rq->ttwu_local);
1615 		__schedstat_inc(p->se.statistics.nr_wakeups_local);
1616 	} else {
1617 		struct sched_domain *sd;
1618 
1619 		__schedstat_inc(p->se.statistics.nr_wakeups_remote);
1620 		rcu_read_lock();
1621 		for_each_domain(rq->cpu, sd) {
1622 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1623 				__schedstat_inc(sd->ttwu_wake_remote);
1624 				break;
1625 			}
1626 		}
1627 		rcu_read_unlock();
1628 	}
1629 
1630 	if (wake_flags & WF_MIGRATED)
1631 		__schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1632 #endif /* CONFIG_SMP */
1633 
1634 	__schedstat_inc(rq->ttwu_count);
1635 	__schedstat_inc(p->se.statistics.nr_wakeups);
1636 
1637 	if (wake_flags & WF_SYNC)
1638 		__schedstat_inc(p->se.statistics.nr_wakeups_sync);
1639 }
1640 
1641 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1642 {
1643 	activate_task(rq, p, en_flags);
1644 	p->on_rq = TASK_ON_RQ_QUEUED;
1645 
1646 	/* If a worker is waking up, notify the workqueue: */
1647 	if (p->flags & PF_WQ_WORKER)
1648 		wq_worker_waking_up(p, cpu_of(rq));
1649 }
1650 
1651 /*
1652  * Mark the task runnable and perform wakeup-preemption.
1653  */
1654 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1655 			   struct rq_flags *rf)
1656 {
1657 	check_preempt_curr(rq, p, wake_flags);
1658 	p->state = TASK_RUNNING;
1659 	trace_sched_wakeup(p);
1660 
1661 #ifdef CONFIG_SMP
1662 	if (p->sched_class->task_woken) {
1663 		/*
1664 		 * Our task @p is fully woken up and running; so its safe to
1665 		 * drop the rq->lock, hereafter rq is only used for statistics.
1666 		 */
1667 		rq_unpin_lock(rq, rf);
1668 		p->sched_class->task_woken(rq, p);
1669 		rq_repin_lock(rq, rf);
1670 	}
1671 
1672 	if (rq->idle_stamp) {
1673 		u64 delta = rq_clock(rq) - rq->idle_stamp;
1674 		u64 max = 2*rq->max_idle_balance_cost;
1675 
1676 		update_avg(&rq->avg_idle, delta);
1677 
1678 		if (rq->avg_idle > max)
1679 			rq->avg_idle = max;
1680 
1681 		rq->idle_stamp = 0;
1682 	}
1683 #endif
1684 }
1685 
1686 static void
1687 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1688 		 struct rq_flags *rf)
1689 {
1690 	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1691 
1692 	lockdep_assert_held(&rq->lock);
1693 
1694 #ifdef CONFIG_SMP
1695 	if (p->sched_contributes_to_load)
1696 		rq->nr_uninterruptible--;
1697 
1698 	if (wake_flags & WF_MIGRATED)
1699 		en_flags |= ENQUEUE_MIGRATED;
1700 #endif
1701 
1702 	ttwu_activate(rq, p, en_flags);
1703 	ttwu_do_wakeup(rq, p, wake_flags, rf);
1704 }
1705 
1706 /*
1707  * Called in case the task @p isn't fully descheduled from its runqueue,
1708  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1709  * since all we need to do is flip p->state to TASK_RUNNING, since
1710  * the task is still ->on_rq.
1711  */
1712 static int ttwu_remote(struct task_struct *p, int wake_flags)
1713 {
1714 	struct rq_flags rf;
1715 	struct rq *rq;
1716 	int ret = 0;
1717 
1718 	rq = __task_rq_lock(p, &rf);
1719 	if (task_on_rq_queued(p)) {
1720 		/* check_preempt_curr() may use rq clock */
1721 		update_rq_clock(rq);
1722 		ttwu_do_wakeup(rq, p, wake_flags, &rf);
1723 		ret = 1;
1724 	}
1725 	__task_rq_unlock(rq, &rf);
1726 
1727 	return ret;
1728 }
1729 
1730 #ifdef CONFIG_SMP
1731 void sched_ttwu_pending(void)
1732 {
1733 	struct rq *rq = this_rq();
1734 	struct llist_node *llist = llist_del_all(&rq->wake_list);
1735 	struct task_struct *p, *t;
1736 	struct rq_flags rf;
1737 
1738 	if (!llist)
1739 		return;
1740 
1741 	rq_lock_irqsave(rq, &rf);
1742 	update_rq_clock(rq);
1743 
1744 	llist_for_each_entry_safe(p, t, llist, wake_entry)
1745 		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1746 
1747 	rq_unlock_irqrestore(rq, &rf);
1748 }
1749 
1750 void scheduler_ipi(void)
1751 {
1752 	/*
1753 	 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1754 	 * TIF_NEED_RESCHED remotely (for the first time) will also send
1755 	 * this IPI.
1756 	 */
1757 	preempt_fold_need_resched();
1758 
1759 	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1760 		return;
1761 
1762 	/*
1763 	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1764 	 * traditionally all their work was done from the interrupt return
1765 	 * path. Now that we actually do some work, we need to make sure
1766 	 * we do call them.
1767 	 *
1768 	 * Some archs already do call them, luckily irq_enter/exit nest
1769 	 * properly.
1770 	 *
1771 	 * Arguably we should visit all archs and update all handlers,
1772 	 * however a fair share of IPIs are still resched only so this would
1773 	 * somewhat pessimize the simple resched case.
1774 	 */
1775 	irq_enter();
1776 	sched_ttwu_pending();
1777 
1778 	/*
1779 	 * Check if someone kicked us for doing the nohz idle load balance.
1780 	 */
1781 	if (unlikely(got_nohz_idle_kick())) {
1782 		this_rq()->idle_balance = 1;
1783 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1784 	}
1785 	irq_exit();
1786 }
1787 
1788 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1789 {
1790 	struct rq *rq = cpu_rq(cpu);
1791 
1792 	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1793 
1794 	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1795 		if (!set_nr_if_polling(rq->idle))
1796 			smp_send_reschedule(cpu);
1797 		else
1798 			trace_sched_wake_idle_without_ipi(cpu);
1799 	}
1800 }
1801 
1802 void wake_up_if_idle(int cpu)
1803 {
1804 	struct rq *rq = cpu_rq(cpu);
1805 	struct rq_flags rf;
1806 
1807 	rcu_read_lock();
1808 
1809 	if (!is_idle_task(rcu_dereference(rq->curr)))
1810 		goto out;
1811 
1812 	if (set_nr_if_polling(rq->idle)) {
1813 		trace_sched_wake_idle_without_ipi(cpu);
1814 	} else {
1815 		rq_lock_irqsave(rq, &rf);
1816 		if (is_idle_task(rq->curr))
1817 			smp_send_reschedule(cpu);
1818 		/* Else CPU is not idle, do nothing here: */
1819 		rq_unlock_irqrestore(rq, &rf);
1820 	}
1821 
1822 out:
1823 	rcu_read_unlock();
1824 }
1825 
1826 bool cpus_share_cache(int this_cpu, int that_cpu)
1827 {
1828 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1829 }
1830 #endif /* CONFIG_SMP */
1831 
1832 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1833 {
1834 	struct rq *rq = cpu_rq(cpu);
1835 	struct rq_flags rf;
1836 
1837 #if defined(CONFIG_SMP)
1838 	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1839 		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1840 		ttwu_queue_remote(p, cpu, wake_flags);
1841 		return;
1842 	}
1843 #endif
1844 
1845 	rq_lock(rq, &rf);
1846 	update_rq_clock(rq);
1847 	ttwu_do_activate(rq, p, wake_flags, &rf);
1848 	rq_unlock(rq, &rf);
1849 }
1850 
1851 /*
1852  * Notes on Program-Order guarantees on SMP systems.
1853  *
1854  *  MIGRATION
1855  *
1856  * The basic program-order guarantee on SMP systems is that when a task [t]
1857  * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1858  * execution on its new CPU [c1].
1859  *
1860  * For migration (of runnable tasks) this is provided by the following means:
1861  *
1862  *  A) UNLOCK of the rq(c0)->lock scheduling out task t
1863  *  B) migration for t is required to synchronize *both* rq(c0)->lock and
1864  *     rq(c1)->lock (if not at the same time, then in that order).
1865  *  C) LOCK of the rq(c1)->lock scheduling in task
1866  *
1867  * Release/acquire chaining guarantees that B happens after A and C after B.
1868  * Note: the CPU doing B need not be c0 or c1
1869  *
1870  * Example:
1871  *
1872  *   CPU0            CPU1            CPU2
1873  *
1874  *   LOCK rq(0)->lock
1875  *   sched-out X
1876  *   sched-in Y
1877  *   UNLOCK rq(0)->lock
1878  *
1879  *                                   LOCK rq(0)->lock // orders against CPU0
1880  *                                   dequeue X
1881  *                                   UNLOCK rq(0)->lock
1882  *
1883  *                                   LOCK rq(1)->lock
1884  *                                   enqueue X
1885  *                                   UNLOCK rq(1)->lock
1886  *
1887  *                   LOCK rq(1)->lock // orders against CPU2
1888  *                   sched-out Z
1889  *                   sched-in X
1890  *                   UNLOCK rq(1)->lock
1891  *
1892  *
1893  *  BLOCKING -- aka. SLEEP + WAKEUP
1894  *
1895  * For blocking we (obviously) need to provide the same guarantee as for
1896  * migration. However the means are completely different as there is no lock
1897  * chain to provide order. Instead we do:
1898  *
1899  *   1) smp_store_release(X->on_cpu, 0)
1900  *   2) smp_cond_load_acquire(!X->on_cpu)
1901  *
1902  * Example:
1903  *
1904  *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
1905  *
1906  *   LOCK rq(0)->lock LOCK X->pi_lock
1907  *   dequeue X
1908  *   sched-out X
1909  *   smp_store_release(X->on_cpu, 0);
1910  *
1911  *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
1912  *                    X->state = WAKING
1913  *                    set_task_cpu(X,2)
1914  *
1915  *                    LOCK rq(2)->lock
1916  *                    enqueue X
1917  *                    X->state = RUNNING
1918  *                    UNLOCK rq(2)->lock
1919  *
1920  *                                          LOCK rq(2)->lock // orders against CPU1
1921  *                                          sched-out Z
1922  *                                          sched-in X
1923  *                                          UNLOCK rq(2)->lock
1924  *
1925  *                    UNLOCK X->pi_lock
1926  *   UNLOCK rq(0)->lock
1927  *
1928  *
1929  * However, for wakeups there is a second guarantee we must provide, namely we
1930  * must ensure that CONDITION=1 done by the caller can not be reordered with
1931  * accesses to the task state; see try_to_wake_up() and set_current_state().
1932  */
1933 
1934 /**
1935  * try_to_wake_up - wake up a thread
1936  * @p: the thread to be awakened
1937  * @state: the mask of task states that can be woken
1938  * @wake_flags: wake modifier flags (WF_*)
1939  *
1940  * If (@state & @p->state) @p->state = TASK_RUNNING.
1941  *
1942  * If the task was not queued/runnable, also place it back on a runqueue.
1943  *
1944  * Atomic against schedule() which would dequeue a task, also see
1945  * set_current_state().
1946  *
1947  * This function executes a full memory barrier before accessing the task
1948  * state; see set_current_state().
1949  *
1950  * Return: %true if @p->state changes (an actual wakeup was done),
1951  *	   %false otherwise.
1952  */
1953 static int
1954 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1955 {
1956 	unsigned long flags;
1957 	int cpu, success = 0;
1958 
1959 	/*
1960 	 * If we are going to wake up a thread waiting for CONDITION we
1961 	 * need to ensure that CONDITION=1 done by the caller can not be
1962 	 * reordered with p->state check below. This pairs with mb() in
1963 	 * set_current_state() the waiting thread does.
1964 	 */
1965 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1966 	smp_mb__after_spinlock();
1967 	if (!(p->state & state))
1968 		goto out;
1969 
1970 	trace_sched_waking(p);
1971 
1972 	/* We're going to change ->state: */
1973 	success = 1;
1974 	cpu = task_cpu(p);
1975 
1976 	/*
1977 	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1978 	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1979 	 * in smp_cond_load_acquire() below.
1980 	 *
1981 	 * sched_ttwu_pending()			try_to_wake_up()
1982 	 *   STORE p->on_rq = 1			  LOAD p->state
1983 	 *   UNLOCK rq->lock
1984 	 *
1985 	 * __schedule() (switch to task 'p')
1986 	 *   LOCK rq->lock			  smp_rmb();
1987 	 *   smp_mb__after_spinlock();
1988 	 *   UNLOCK rq->lock
1989 	 *
1990 	 * [task p]
1991 	 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
1992 	 *
1993 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1994 	 * __schedule().  See the comment for smp_mb__after_spinlock().
1995 	 */
1996 	smp_rmb();
1997 	if (p->on_rq && ttwu_remote(p, wake_flags))
1998 		goto stat;
1999 
2000 #ifdef CONFIG_SMP
2001 	/*
2002 	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2003 	 * possible to, falsely, observe p->on_cpu == 0.
2004 	 *
2005 	 * One must be running (->on_cpu == 1) in order to remove oneself
2006 	 * from the runqueue.
2007 	 *
2008 	 * __schedule() (switch to task 'p')	try_to_wake_up()
2009 	 *   STORE p->on_cpu = 1		  LOAD p->on_rq
2010 	 *   UNLOCK rq->lock
2011 	 *
2012 	 * __schedule() (put 'p' to sleep)
2013 	 *   LOCK rq->lock			  smp_rmb();
2014 	 *   smp_mb__after_spinlock();
2015 	 *   STORE p->on_rq = 0			  LOAD p->on_cpu
2016 	 *
2017 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2018 	 * __schedule().  See the comment for smp_mb__after_spinlock().
2019 	 */
2020 	smp_rmb();
2021 
2022 	/*
2023 	 * If the owning (remote) CPU is still in the middle of schedule() with
2024 	 * this task as prev, wait until its done referencing the task.
2025 	 *
2026 	 * Pairs with the smp_store_release() in finish_task().
2027 	 *
2028 	 * This ensures that tasks getting woken will be fully ordered against
2029 	 * their previous state and preserve Program Order.
2030 	 */
2031 	smp_cond_load_acquire(&p->on_cpu, !VAL);
2032 
2033 	p->sched_contributes_to_load = !!task_contributes_to_load(p);
2034 	p->state = TASK_WAKING;
2035 
2036 	if (p->in_iowait) {
2037 		delayacct_blkio_end(p);
2038 		atomic_dec(&task_rq(p)->nr_iowait);
2039 	}
2040 
2041 	cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2042 	if (task_cpu(p) != cpu) {
2043 		wake_flags |= WF_MIGRATED;
2044 		psi_ttwu_dequeue(p);
2045 		set_task_cpu(p, cpu);
2046 	}
2047 
2048 #else /* CONFIG_SMP */
2049 
2050 	if (p->in_iowait) {
2051 		delayacct_blkio_end(p);
2052 		atomic_dec(&task_rq(p)->nr_iowait);
2053 	}
2054 
2055 #endif /* CONFIG_SMP */
2056 
2057 	ttwu_queue(p, cpu, wake_flags);
2058 stat:
2059 	ttwu_stat(p, cpu, wake_flags);
2060 out:
2061 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2062 
2063 	return success;
2064 }
2065 
2066 /**
2067  * try_to_wake_up_local - try to wake up a local task with rq lock held
2068  * @p: the thread to be awakened
2069  * @rf: request-queue flags for pinning
2070  *
2071  * Put @p on the run-queue if it's not already there. The caller must
2072  * ensure that this_rq() is locked, @p is bound to this_rq() and not
2073  * the current task.
2074  */
2075 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2076 {
2077 	struct rq *rq = task_rq(p);
2078 
2079 	if (WARN_ON_ONCE(rq != this_rq()) ||
2080 	    WARN_ON_ONCE(p == current))
2081 		return;
2082 
2083 	lockdep_assert_held(&rq->lock);
2084 
2085 	if (!raw_spin_trylock(&p->pi_lock)) {
2086 		/*
2087 		 * This is OK, because current is on_cpu, which avoids it being
2088 		 * picked for load-balance and preemption/IRQs are still
2089 		 * disabled avoiding further scheduler activity on it and we've
2090 		 * not yet picked a replacement task.
2091 		 */
2092 		rq_unlock(rq, rf);
2093 		raw_spin_lock(&p->pi_lock);
2094 		rq_relock(rq, rf);
2095 	}
2096 
2097 	if (!(p->state & TASK_NORMAL))
2098 		goto out;
2099 
2100 	trace_sched_waking(p);
2101 
2102 	if (!task_on_rq_queued(p)) {
2103 		if (p->in_iowait) {
2104 			delayacct_blkio_end(p);
2105 			atomic_dec(&rq->nr_iowait);
2106 		}
2107 		ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2108 	}
2109 
2110 	ttwu_do_wakeup(rq, p, 0, rf);
2111 	ttwu_stat(p, smp_processor_id(), 0);
2112 out:
2113 	raw_spin_unlock(&p->pi_lock);
2114 }
2115 
2116 /**
2117  * wake_up_process - Wake up a specific process
2118  * @p: The process to be woken up.
2119  *
2120  * Attempt to wake up the nominated process and move it to the set of runnable
2121  * processes.
2122  *
2123  * Return: 1 if the process was woken up, 0 if it was already running.
2124  *
2125  * This function executes a full memory barrier before accessing the task state.
2126  */
2127 int wake_up_process(struct task_struct *p)
2128 {
2129 	return try_to_wake_up(p, TASK_NORMAL, 0);
2130 }
2131 EXPORT_SYMBOL(wake_up_process);
2132 
2133 int wake_up_state(struct task_struct *p, unsigned int state)
2134 {
2135 	return try_to_wake_up(p, state, 0);
2136 }
2137 
2138 /*
2139  * Perform scheduler related setup for a newly forked process p.
2140  * p is forked by current.
2141  *
2142  * __sched_fork() is basic setup used by init_idle() too:
2143  */
2144 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2145 {
2146 	p->on_rq			= 0;
2147 
2148 	p->se.on_rq			= 0;
2149 	p->se.exec_start		= 0;
2150 	p->se.sum_exec_runtime		= 0;
2151 	p->se.prev_sum_exec_runtime	= 0;
2152 	p->se.nr_migrations		= 0;
2153 	p->se.vruntime			= 0;
2154 	INIT_LIST_HEAD(&p->se.group_node);
2155 
2156 #ifdef CONFIG_FAIR_GROUP_SCHED
2157 	p->se.cfs_rq			= NULL;
2158 #endif
2159 
2160 #ifdef CONFIG_SCHEDSTATS
2161 	/* Even if schedstat is disabled, there should not be garbage */
2162 	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2163 #endif
2164 
2165 	RB_CLEAR_NODE(&p->dl.rb_node);
2166 	init_dl_task_timer(&p->dl);
2167 	init_dl_inactive_task_timer(&p->dl);
2168 	__dl_clear_params(p);
2169 
2170 	INIT_LIST_HEAD(&p->rt.run_list);
2171 	p->rt.timeout		= 0;
2172 	p->rt.time_slice	= sched_rr_timeslice;
2173 	p->rt.on_rq		= 0;
2174 	p->rt.on_list		= 0;
2175 
2176 #ifdef CONFIG_PREEMPT_NOTIFIERS
2177 	INIT_HLIST_HEAD(&p->preempt_notifiers);
2178 #endif
2179 
2180 	init_numa_balancing(clone_flags, p);
2181 }
2182 
2183 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2184 
2185 #ifdef CONFIG_NUMA_BALANCING
2186 
2187 void set_numabalancing_state(bool enabled)
2188 {
2189 	if (enabled)
2190 		static_branch_enable(&sched_numa_balancing);
2191 	else
2192 		static_branch_disable(&sched_numa_balancing);
2193 }
2194 
2195 #ifdef CONFIG_PROC_SYSCTL
2196 int sysctl_numa_balancing(struct ctl_table *table, int write,
2197 			 void __user *buffer, size_t *lenp, loff_t *ppos)
2198 {
2199 	struct ctl_table t;
2200 	int err;
2201 	int state = static_branch_likely(&sched_numa_balancing);
2202 
2203 	if (write && !capable(CAP_SYS_ADMIN))
2204 		return -EPERM;
2205 
2206 	t = *table;
2207 	t.data = &state;
2208 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2209 	if (err < 0)
2210 		return err;
2211 	if (write)
2212 		set_numabalancing_state(state);
2213 	return err;
2214 }
2215 #endif
2216 #endif
2217 
2218 #ifdef CONFIG_SCHEDSTATS
2219 
2220 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2221 static bool __initdata __sched_schedstats = false;
2222 
2223 static void set_schedstats(bool enabled)
2224 {
2225 	if (enabled)
2226 		static_branch_enable(&sched_schedstats);
2227 	else
2228 		static_branch_disable(&sched_schedstats);
2229 }
2230 
2231 void force_schedstat_enabled(void)
2232 {
2233 	if (!schedstat_enabled()) {
2234 		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2235 		static_branch_enable(&sched_schedstats);
2236 	}
2237 }
2238 
2239 static int __init setup_schedstats(char *str)
2240 {
2241 	int ret = 0;
2242 	if (!str)
2243 		goto out;
2244 
2245 	/*
2246 	 * This code is called before jump labels have been set up, so we can't
2247 	 * change the static branch directly just yet.  Instead set a temporary
2248 	 * variable so init_schedstats() can do it later.
2249 	 */
2250 	if (!strcmp(str, "enable")) {
2251 		__sched_schedstats = true;
2252 		ret = 1;
2253 	} else if (!strcmp(str, "disable")) {
2254 		__sched_schedstats = false;
2255 		ret = 1;
2256 	}
2257 out:
2258 	if (!ret)
2259 		pr_warn("Unable to parse schedstats=\n");
2260 
2261 	return ret;
2262 }
2263 __setup("schedstats=", setup_schedstats);
2264 
2265 static void __init init_schedstats(void)
2266 {
2267 	set_schedstats(__sched_schedstats);
2268 }
2269 
2270 #ifdef CONFIG_PROC_SYSCTL
2271 int sysctl_schedstats(struct ctl_table *table, int write,
2272 			 void __user *buffer, size_t *lenp, loff_t *ppos)
2273 {
2274 	struct ctl_table t;
2275 	int err;
2276 	int state = static_branch_likely(&sched_schedstats);
2277 
2278 	if (write && !capable(CAP_SYS_ADMIN))
2279 		return -EPERM;
2280 
2281 	t = *table;
2282 	t.data = &state;
2283 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2284 	if (err < 0)
2285 		return err;
2286 	if (write)
2287 		set_schedstats(state);
2288 	return err;
2289 }
2290 #endif /* CONFIG_PROC_SYSCTL */
2291 #else  /* !CONFIG_SCHEDSTATS */
2292 static inline void init_schedstats(void) {}
2293 #endif /* CONFIG_SCHEDSTATS */
2294 
2295 /*
2296  * fork()/clone()-time setup:
2297  */
2298 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2299 {
2300 	unsigned long flags;
2301 
2302 	__sched_fork(clone_flags, p);
2303 	/*
2304 	 * We mark the process as NEW here. This guarantees that
2305 	 * nobody will actually run it, and a signal or other external
2306 	 * event cannot wake it up and insert it on the runqueue either.
2307 	 */
2308 	p->state = TASK_NEW;
2309 
2310 	/*
2311 	 * Make sure we do not leak PI boosting priority to the child.
2312 	 */
2313 	p->prio = current->normal_prio;
2314 
2315 	/*
2316 	 * Revert to default priority/policy on fork if requested.
2317 	 */
2318 	if (unlikely(p->sched_reset_on_fork)) {
2319 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2320 			p->policy = SCHED_NORMAL;
2321 			p->static_prio = NICE_TO_PRIO(0);
2322 			p->rt_priority = 0;
2323 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
2324 			p->static_prio = NICE_TO_PRIO(0);
2325 
2326 		p->prio = p->normal_prio = __normal_prio(p);
2327 		set_load_weight(p, false);
2328 
2329 		/*
2330 		 * We don't need the reset flag anymore after the fork. It has
2331 		 * fulfilled its duty:
2332 		 */
2333 		p->sched_reset_on_fork = 0;
2334 	}
2335 
2336 	if (dl_prio(p->prio))
2337 		return -EAGAIN;
2338 	else if (rt_prio(p->prio))
2339 		p->sched_class = &rt_sched_class;
2340 	else
2341 		p->sched_class = &fair_sched_class;
2342 
2343 	init_entity_runnable_average(&p->se);
2344 
2345 	/*
2346 	 * The child is not yet in the pid-hash so no cgroup attach races,
2347 	 * and the cgroup is pinned to this child due to cgroup_fork()
2348 	 * is ran before sched_fork().
2349 	 *
2350 	 * Silence PROVE_RCU.
2351 	 */
2352 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2353 	/*
2354 	 * We're setting the CPU for the first time, we don't migrate,
2355 	 * so use __set_task_cpu().
2356 	 */
2357 	__set_task_cpu(p, smp_processor_id());
2358 	if (p->sched_class->task_fork)
2359 		p->sched_class->task_fork(p);
2360 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2361 
2362 #ifdef CONFIG_SCHED_INFO
2363 	if (likely(sched_info_on()))
2364 		memset(&p->sched_info, 0, sizeof(p->sched_info));
2365 #endif
2366 #if defined(CONFIG_SMP)
2367 	p->on_cpu = 0;
2368 #endif
2369 	init_task_preempt_count(p);
2370 #ifdef CONFIG_SMP
2371 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
2372 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
2373 #endif
2374 	return 0;
2375 }
2376 
2377 unsigned long to_ratio(u64 period, u64 runtime)
2378 {
2379 	if (runtime == RUNTIME_INF)
2380 		return BW_UNIT;
2381 
2382 	/*
2383 	 * Doing this here saves a lot of checks in all
2384 	 * the calling paths, and returning zero seems
2385 	 * safe for them anyway.
2386 	 */
2387 	if (period == 0)
2388 		return 0;
2389 
2390 	return div64_u64(runtime << BW_SHIFT, period);
2391 }
2392 
2393 /*
2394  * wake_up_new_task - wake up a newly created task for the first time.
2395  *
2396  * This function will do some initial scheduler statistics housekeeping
2397  * that must be done for every newly created context, then puts the task
2398  * on the runqueue and wakes it.
2399  */
2400 void wake_up_new_task(struct task_struct *p)
2401 {
2402 	struct rq_flags rf;
2403 	struct rq *rq;
2404 
2405 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2406 	p->state = TASK_RUNNING;
2407 #ifdef CONFIG_SMP
2408 	/*
2409 	 * Fork balancing, do it here and not earlier because:
2410 	 *  - cpus_allowed can change in the fork path
2411 	 *  - any previously selected CPU might disappear through hotplug
2412 	 *
2413 	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2414 	 * as we're not fully set-up yet.
2415 	 */
2416 	p->recent_used_cpu = task_cpu(p);
2417 	__set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2418 #endif
2419 	rq = __task_rq_lock(p, &rf);
2420 	update_rq_clock(rq);
2421 	post_init_entity_util_avg(&p->se);
2422 
2423 	activate_task(rq, p, ENQUEUE_NOCLOCK);
2424 	p->on_rq = TASK_ON_RQ_QUEUED;
2425 	trace_sched_wakeup_new(p);
2426 	check_preempt_curr(rq, p, WF_FORK);
2427 #ifdef CONFIG_SMP
2428 	if (p->sched_class->task_woken) {
2429 		/*
2430 		 * Nothing relies on rq->lock after this, so its fine to
2431 		 * drop it.
2432 		 */
2433 		rq_unpin_lock(rq, &rf);
2434 		p->sched_class->task_woken(rq, p);
2435 		rq_repin_lock(rq, &rf);
2436 	}
2437 #endif
2438 	task_rq_unlock(rq, p, &rf);
2439 }
2440 
2441 #ifdef CONFIG_PREEMPT_NOTIFIERS
2442 
2443 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2444 
2445 void preempt_notifier_inc(void)
2446 {
2447 	static_branch_inc(&preempt_notifier_key);
2448 }
2449 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2450 
2451 void preempt_notifier_dec(void)
2452 {
2453 	static_branch_dec(&preempt_notifier_key);
2454 }
2455 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2456 
2457 /**
2458  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2459  * @notifier: notifier struct to register
2460  */
2461 void preempt_notifier_register(struct preempt_notifier *notifier)
2462 {
2463 	if (!static_branch_unlikely(&preempt_notifier_key))
2464 		WARN(1, "registering preempt_notifier while notifiers disabled\n");
2465 
2466 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
2467 }
2468 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2469 
2470 /**
2471  * preempt_notifier_unregister - no longer interested in preemption notifications
2472  * @notifier: notifier struct to unregister
2473  *
2474  * This is *not* safe to call from within a preemption notifier.
2475  */
2476 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2477 {
2478 	hlist_del(&notifier->link);
2479 }
2480 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2481 
2482 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2483 {
2484 	struct preempt_notifier *notifier;
2485 
2486 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2487 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
2488 }
2489 
2490 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2491 {
2492 	if (static_branch_unlikely(&preempt_notifier_key))
2493 		__fire_sched_in_preempt_notifiers(curr);
2494 }
2495 
2496 static void
2497 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2498 				   struct task_struct *next)
2499 {
2500 	struct preempt_notifier *notifier;
2501 
2502 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 		notifier->ops->sched_out(notifier, next);
2504 }
2505 
2506 static __always_inline void
2507 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2508 				 struct task_struct *next)
2509 {
2510 	if (static_branch_unlikely(&preempt_notifier_key))
2511 		__fire_sched_out_preempt_notifiers(curr, next);
2512 }
2513 
2514 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2515 
2516 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2517 {
2518 }
2519 
2520 static inline void
2521 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2522 				 struct task_struct *next)
2523 {
2524 }
2525 
2526 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2527 
2528 static inline void prepare_task(struct task_struct *next)
2529 {
2530 #ifdef CONFIG_SMP
2531 	/*
2532 	 * Claim the task as running, we do this before switching to it
2533 	 * such that any running task will have this set.
2534 	 */
2535 	next->on_cpu = 1;
2536 #endif
2537 }
2538 
2539 static inline void finish_task(struct task_struct *prev)
2540 {
2541 #ifdef CONFIG_SMP
2542 	/*
2543 	 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2544 	 * We must ensure this doesn't happen until the switch is completely
2545 	 * finished.
2546 	 *
2547 	 * In particular, the load of prev->state in finish_task_switch() must
2548 	 * happen before this.
2549 	 *
2550 	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2551 	 */
2552 	smp_store_release(&prev->on_cpu, 0);
2553 #endif
2554 }
2555 
2556 static inline void
2557 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2558 {
2559 	/*
2560 	 * Since the runqueue lock will be released by the next
2561 	 * task (which is an invalid locking op but in the case
2562 	 * of the scheduler it's an obvious special-case), so we
2563 	 * do an early lockdep release here:
2564 	 */
2565 	rq_unpin_lock(rq, rf);
2566 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2567 #ifdef CONFIG_DEBUG_SPINLOCK
2568 	/* this is a valid case when another task releases the spinlock */
2569 	rq->lock.owner = next;
2570 #endif
2571 }
2572 
2573 static inline void finish_lock_switch(struct rq *rq)
2574 {
2575 	/*
2576 	 * If we are tracking spinlock dependencies then we have to
2577 	 * fix up the runqueue lock - which gets 'carried over' from
2578 	 * prev into current:
2579 	 */
2580 	spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2581 	raw_spin_unlock_irq(&rq->lock);
2582 }
2583 
2584 /*
2585  * NOP if the arch has not defined these:
2586  */
2587 
2588 #ifndef prepare_arch_switch
2589 # define prepare_arch_switch(next)	do { } while (0)
2590 #endif
2591 
2592 #ifndef finish_arch_post_lock_switch
2593 # define finish_arch_post_lock_switch()	do { } while (0)
2594 #endif
2595 
2596 /**
2597  * prepare_task_switch - prepare to switch tasks
2598  * @rq: the runqueue preparing to switch
2599  * @prev: the current task that is being switched out
2600  * @next: the task we are going to switch to.
2601  *
2602  * This is called with the rq lock held and interrupts off. It must
2603  * be paired with a subsequent finish_task_switch after the context
2604  * switch.
2605  *
2606  * prepare_task_switch sets up locking and calls architecture specific
2607  * hooks.
2608  */
2609 static inline void
2610 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2611 		    struct task_struct *next)
2612 {
2613 	kcov_prepare_switch(prev);
2614 	sched_info_switch(rq, prev, next);
2615 	perf_event_task_sched_out(prev, next);
2616 	rseq_preempt(prev);
2617 	fire_sched_out_preempt_notifiers(prev, next);
2618 	prepare_task(next);
2619 	prepare_arch_switch(next);
2620 }
2621 
2622 /**
2623  * finish_task_switch - clean up after a task-switch
2624  * @prev: the thread we just switched away from.
2625  *
2626  * finish_task_switch must be called after the context switch, paired
2627  * with a prepare_task_switch call before the context switch.
2628  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2629  * and do any other architecture-specific cleanup actions.
2630  *
2631  * Note that we may have delayed dropping an mm in context_switch(). If
2632  * so, we finish that here outside of the runqueue lock. (Doing it
2633  * with the lock held can cause deadlocks; see schedule() for
2634  * details.)
2635  *
2636  * The context switch have flipped the stack from under us and restored the
2637  * local variables which were saved when this task called schedule() in the
2638  * past. prev == current is still correct but we need to recalculate this_rq
2639  * because prev may have moved to another CPU.
2640  */
2641 static struct rq *finish_task_switch(struct task_struct *prev)
2642 	__releases(rq->lock)
2643 {
2644 	struct rq *rq = this_rq();
2645 	struct mm_struct *mm = rq->prev_mm;
2646 	long prev_state;
2647 
2648 	/*
2649 	 * The previous task will have left us with a preempt_count of 2
2650 	 * because it left us after:
2651 	 *
2652 	 *	schedule()
2653 	 *	  preempt_disable();			// 1
2654 	 *	  __schedule()
2655 	 *	    raw_spin_lock_irq(&rq->lock)	// 2
2656 	 *
2657 	 * Also, see FORK_PREEMPT_COUNT.
2658 	 */
2659 	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2660 		      "corrupted preempt_count: %s/%d/0x%x\n",
2661 		      current->comm, current->pid, preempt_count()))
2662 		preempt_count_set(FORK_PREEMPT_COUNT);
2663 
2664 	rq->prev_mm = NULL;
2665 
2666 	/*
2667 	 * A task struct has one reference for the use as "current".
2668 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2669 	 * schedule one last time. The schedule call will never return, and
2670 	 * the scheduled task must drop that reference.
2671 	 *
2672 	 * We must observe prev->state before clearing prev->on_cpu (in
2673 	 * finish_task), otherwise a concurrent wakeup can get prev
2674 	 * running on another CPU and we could rave with its RUNNING -> DEAD
2675 	 * transition, resulting in a double drop.
2676 	 */
2677 	prev_state = prev->state;
2678 	vtime_task_switch(prev);
2679 	perf_event_task_sched_in(prev, current);
2680 	finish_task(prev);
2681 	finish_lock_switch(rq);
2682 	finish_arch_post_lock_switch();
2683 	kcov_finish_switch(current);
2684 
2685 	fire_sched_in_preempt_notifiers(current);
2686 	/*
2687 	 * When switching through a kernel thread, the loop in
2688 	 * membarrier_{private,global}_expedited() may have observed that
2689 	 * kernel thread and not issued an IPI. It is therefore possible to
2690 	 * schedule between user->kernel->user threads without passing though
2691 	 * switch_mm(). Membarrier requires a barrier after storing to
2692 	 * rq->curr, before returning to userspace, so provide them here:
2693 	 *
2694 	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2695 	 *   provided by mmdrop(),
2696 	 * - a sync_core for SYNC_CORE.
2697 	 */
2698 	if (mm) {
2699 		membarrier_mm_sync_core_before_usermode(mm);
2700 		mmdrop(mm);
2701 	}
2702 	if (unlikely(prev_state == TASK_DEAD)) {
2703 		if (prev->sched_class->task_dead)
2704 			prev->sched_class->task_dead(prev);
2705 
2706 		/*
2707 		 * Remove function-return probe instances associated with this
2708 		 * task and put them back on the free list.
2709 		 */
2710 		kprobe_flush_task(prev);
2711 
2712 		/* Task is done with its stack. */
2713 		put_task_stack(prev);
2714 
2715 		put_task_struct(prev);
2716 	}
2717 
2718 	tick_nohz_task_switch();
2719 	return rq;
2720 }
2721 
2722 #ifdef CONFIG_SMP
2723 
2724 /* rq->lock is NOT held, but preemption is disabled */
2725 static void __balance_callback(struct rq *rq)
2726 {
2727 	struct callback_head *head, *next;
2728 	void (*func)(struct rq *rq);
2729 	unsigned long flags;
2730 
2731 	raw_spin_lock_irqsave(&rq->lock, flags);
2732 	head = rq->balance_callback;
2733 	rq->balance_callback = NULL;
2734 	while (head) {
2735 		func = (void (*)(struct rq *))head->func;
2736 		next = head->next;
2737 		head->next = NULL;
2738 		head = next;
2739 
2740 		func(rq);
2741 	}
2742 	raw_spin_unlock_irqrestore(&rq->lock, flags);
2743 }
2744 
2745 static inline void balance_callback(struct rq *rq)
2746 {
2747 	if (unlikely(rq->balance_callback))
2748 		__balance_callback(rq);
2749 }
2750 
2751 #else
2752 
2753 static inline void balance_callback(struct rq *rq)
2754 {
2755 }
2756 
2757 #endif
2758 
2759 /**
2760  * schedule_tail - first thing a freshly forked thread must call.
2761  * @prev: the thread we just switched away from.
2762  */
2763 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2764 	__releases(rq->lock)
2765 {
2766 	struct rq *rq;
2767 
2768 	/*
2769 	 * New tasks start with FORK_PREEMPT_COUNT, see there and
2770 	 * finish_task_switch() for details.
2771 	 *
2772 	 * finish_task_switch() will drop rq->lock() and lower preempt_count
2773 	 * and the preempt_enable() will end up enabling preemption (on
2774 	 * PREEMPT_COUNT kernels).
2775 	 */
2776 
2777 	rq = finish_task_switch(prev);
2778 	balance_callback(rq);
2779 	preempt_enable();
2780 
2781 	if (current->set_child_tid)
2782 		put_user(task_pid_vnr(current), current->set_child_tid);
2783 
2784 	calculate_sigpending();
2785 }
2786 
2787 /*
2788  * context_switch - switch to the new MM and the new thread's register state.
2789  */
2790 static __always_inline struct rq *
2791 context_switch(struct rq *rq, struct task_struct *prev,
2792 	       struct task_struct *next, struct rq_flags *rf)
2793 {
2794 	struct mm_struct *mm, *oldmm;
2795 
2796 	prepare_task_switch(rq, prev, next);
2797 
2798 	mm = next->mm;
2799 	oldmm = prev->active_mm;
2800 	/*
2801 	 * For paravirt, this is coupled with an exit in switch_to to
2802 	 * combine the page table reload and the switch backend into
2803 	 * one hypercall.
2804 	 */
2805 	arch_start_context_switch(prev);
2806 
2807 	/*
2808 	 * If mm is non-NULL, we pass through switch_mm(). If mm is
2809 	 * NULL, we will pass through mmdrop() in finish_task_switch().
2810 	 * Both of these contain the full memory barrier required by
2811 	 * membarrier after storing to rq->curr, before returning to
2812 	 * user-space.
2813 	 */
2814 	if (!mm) {
2815 		next->active_mm = oldmm;
2816 		mmgrab(oldmm);
2817 		enter_lazy_tlb(oldmm, next);
2818 	} else
2819 		switch_mm_irqs_off(oldmm, mm, next);
2820 
2821 	if (!prev->mm) {
2822 		prev->active_mm = NULL;
2823 		rq->prev_mm = oldmm;
2824 	}
2825 
2826 	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2827 
2828 	prepare_lock_switch(rq, next, rf);
2829 
2830 	/* Here we just switch the register state and the stack. */
2831 	switch_to(prev, next, prev);
2832 	barrier();
2833 
2834 	return finish_task_switch(prev);
2835 }
2836 
2837 /*
2838  * nr_running and nr_context_switches:
2839  *
2840  * externally visible scheduler statistics: current number of runnable
2841  * threads, total number of context switches performed since bootup.
2842  */
2843 unsigned long nr_running(void)
2844 {
2845 	unsigned long i, sum = 0;
2846 
2847 	for_each_online_cpu(i)
2848 		sum += cpu_rq(i)->nr_running;
2849 
2850 	return sum;
2851 }
2852 
2853 /*
2854  * Check if only the current task is running on the CPU.
2855  *
2856  * Caution: this function does not check that the caller has disabled
2857  * preemption, thus the result might have a time-of-check-to-time-of-use
2858  * race.  The caller is responsible to use it correctly, for example:
2859  *
2860  * - from a non-preemptible section (of course)
2861  *
2862  * - from a thread that is bound to a single CPU
2863  *
2864  * - in a loop with very short iterations (e.g. a polling loop)
2865  */
2866 bool single_task_running(void)
2867 {
2868 	return raw_rq()->nr_running == 1;
2869 }
2870 EXPORT_SYMBOL(single_task_running);
2871 
2872 unsigned long long nr_context_switches(void)
2873 {
2874 	int i;
2875 	unsigned long long sum = 0;
2876 
2877 	for_each_possible_cpu(i)
2878 		sum += cpu_rq(i)->nr_switches;
2879 
2880 	return sum;
2881 }
2882 
2883 /*
2884  * Consumers of these two interfaces, like for example the cpuidle menu
2885  * governor, are using nonsensical data. Preferring shallow idle state selection
2886  * for a CPU that has IO-wait which might not even end up running the task when
2887  * it does become runnable.
2888  */
2889 
2890 unsigned long nr_iowait_cpu(int cpu)
2891 {
2892 	return atomic_read(&cpu_rq(cpu)->nr_iowait);
2893 }
2894 
2895 /*
2896  * IO-wait accounting, and how its mostly bollocks (on SMP).
2897  *
2898  * The idea behind IO-wait account is to account the idle time that we could
2899  * have spend running if it were not for IO. That is, if we were to improve the
2900  * storage performance, we'd have a proportional reduction in IO-wait time.
2901  *
2902  * This all works nicely on UP, where, when a task blocks on IO, we account
2903  * idle time as IO-wait, because if the storage were faster, it could've been
2904  * running and we'd not be idle.
2905  *
2906  * This has been extended to SMP, by doing the same for each CPU. This however
2907  * is broken.
2908  *
2909  * Imagine for instance the case where two tasks block on one CPU, only the one
2910  * CPU will have IO-wait accounted, while the other has regular idle. Even
2911  * though, if the storage were faster, both could've ran at the same time,
2912  * utilising both CPUs.
2913  *
2914  * This means, that when looking globally, the current IO-wait accounting on
2915  * SMP is a lower bound, by reason of under accounting.
2916  *
2917  * Worse, since the numbers are provided per CPU, they are sometimes
2918  * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2919  * associated with any one particular CPU, it can wake to another CPU than it
2920  * blocked on. This means the per CPU IO-wait number is meaningless.
2921  *
2922  * Task CPU affinities can make all that even more 'interesting'.
2923  */
2924 
2925 unsigned long nr_iowait(void)
2926 {
2927 	unsigned long i, sum = 0;
2928 
2929 	for_each_possible_cpu(i)
2930 		sum += nr_iowait_cpu(i);
2931 
2932 	return sum;
2933 }
2934 
2935 #ifdef CONFIG_SMP
2936 
2937 /*
2938  * sched_exec - execve() is a valuable balancing opportunity, because at
2939  * this point the task has the smallest effective memory and cache footprint.
2940  */
2941 void sched_exec(void)
2942 {
2943 	struct task_struct *p = current;
2944 	unsigned long flags;
2945 	int dest_cpu;
2946 
2947 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2948 	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2949 	if (dest_cpu == smp_processor_id())
2950 		goto unlock;
2951 
2952 	if (likely(cpu_active(dest_cpu))) {
2953 		struct migration_arg arg = { p, dest_cpu };
2954 
2955 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2956 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2957 		return;
2958 	}
2959 unlock:
2960 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2961 }
2962 
2963 #endif
2964 
2965 DEFINE_PER_CPU(struct kernel_stat, kstat);
2966 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2967 
2968 EXPORT_PER_CPU_SYMBOL(kstat);
2969 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2970 
2971 /*
2972  * The function fair_sched_class.update_curr accesses the struct curr
2973  * and its field curr->exec_start; when called from task_sched_runtime(),
2974  * we observe a high rate of cache misses in practice.
2975  * Prefetching this data results in improved performance.
2976  */
2977 static inline void prefetch_curr_exec_start(struct task_struct *p)
2978 {
2979 #ifdef CONFIG_FAIR_GROUP_SCHED
2980 	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2981 #else
2982 	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2983 #endif
2984 	prefetch(curr);
2985 	prefetch(&curr->exec_start);
2986 }
2987 
2988 /*
2989  * Return accounted runtime for the task.
2990  * In case the task is currently running, return the runtime plus current's
2991  * pending runtime that have not been accounted yet.
2992  */
2993 unsigned long long task_sched_runtime(struct task_struct *p)
2994 {
2995 	struct rq_flags rf;
2996 	struct rq *rq;
2997 	u64 ns;
2998 
2999 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3000 	/*
3001 	 * 64-bit doesn't need locks to atomically read a 64-bit value.
3002 	 * So we have a optimization chance when the task's delta_exec is 0.
3003 	 * Reading ->on_cpu is racy, but this is ok.
3004 	 *
3005 	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3006 	 * If we race with it entering CPU, unaccounted time is 0. This is
3007 	 * indistinguishable from the read occurring a few cycles earlier.
3008 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3009 	 * been accounted, so we're correct here as well.
3010 	 */
3011 	if (!p->on_cpu || !task_on_rq_queued(p))
3012 		return p->se.sum_exec_runtime;
3013 #endif
3014 
3015 	rq = task_rq_lock(p, &rf);
3016 	/*
3017 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
3018 	 * project cycles that may never be accounted to this
3019 	 * thread, breaking clock_gettime().
3020 	 */
3021 	if (task_current(rq, p) && task_on_rq_queued(p)) {
3022 		prefetch_curr_exec_start(p);
3023 		update_rq_clock(rq);
3024 		p->sched_class->update_curr(rq);
3025 	}
3026 	ns = p->se.sum_exec_runtime;
3027 	task_rq_unlock(rq, p, &rf);
3028 
3029 	return ns;
3030 }
3031 
3032 /*
3033  * This function gets called by the timer code, with HZ frequency.
3034  * We call it with interrupts disabled.
3035  */
3036 void scheduler_tick(void)
3037 {
3038 	int cpu = smp_processor_id();
3039 	struct rq *rq = cpu_rq(cpu);
3040 	struct task_struct *curr = rq->curr;
3041 	struct rq_flags rf;
3042 
3043 	sched_clock_tick();
3044 
3045 	rq_lock(rq, &rf);
3046 
3047 	update_rq_clock(rq);
3048 	curr->sched_class->task_tick(rq, curr, 0);
3049 	cpu_load_update_active(rq);
3050 	calc_global_load_tick(rq);
3051 	psi_task_tick(rq);
3052 
3053 	rq_unlock(rq, &rf);
3054 
3055 	perf_event_task_tick();
3056 
3057 #ifdef CONFIG_SMP
3058 	rq->idle_balance = idle_cpu(cpu);
3059 	trigger_load_balance(rq);
3060 #endif
3061 }
3062 
3063 #ifdef CONFIG_NO_HZ_FULL
3064 
3065 struct tick_work {
3066 	int			cpu;
3067 	struct delayed_work	work;
3068 };
3069 
3070 static struct tick_work __percpu *tick_work_cpu;
3071 
3072 static void sched_tick_remote(struct work_struct *work)
3073 {
3074 	struct delayed_work *dwork = to_delayed_work(work);
3075 	struct tick_work *twork = container_of(dwork, struct tick_work, work);
3076 	int cpu = twork->cpu;
3077 	struct rq *rq = cpu_rq(cpu);
3078 	struct task_struct *curr;
3079 	struct rq_flags rf;
3080 	u64 delta;
3081 
3082 	/*
3083 	 * Handle the tick only if it appears the remote CPU is running in full
3084 	 * dynticks mode. The check is racy by nature, but missing a tick or
3085 	 * having one too much is no big deal because the scheduler tick updates
3086 	 * statistics and checks timeslices in a time-independent way, regardless
3087 	 * of when exactly it is running.
3088 	 */
3089 	if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3090 		goto out_requeue;
3091 
3092 	rq_lock_irq(rq, &rf);
3093 	curr = rq->curr;
3094 	if (is_idle_task(curr))
3095 		goto out_unlock;
3096 
3097 	update_rq_clock(rq);
3098 	delta = rq_clock_task(rq) - curr->se.exec_start;
3099 
3100 	/*
3101 	 * Make sure the next tick runs within a reasonable
3102 	 * amount of time.
3103 	 */
3104 	WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3105 	curr->sched_class->task_tick(rq, curr, 0);
3106 
3107 out_unlock:
3108 	rq_unlock_irq(rq, &rf);
3109 
3110 out_requeue:
3111 	/*
3112 	 * Run the remote tick once per second (1Hz). This arbitrary
3113 	 * frequency is large enough to avoid overload but short enough
3114 	 * to keep scheduler internal stats reasonably up to date.
3115 	 */
3116 	queue_delayed_work(system_unbound_wq, dwork, HZ);
3117 }
3118 
3119 static void sched_tick_start(int cpu)
3120 {
3121 	struct tick_work *twork;
3122 
3123 	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3124 		return;
3125 
3126 	WARN_ON_ONCE(!tick_work_cpu);
3127 
3128 	twork = per_cpu_ptr(tick_work_cpu, cpu);
3129 	twork->cpu = cpu;
3130 	INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3131 	queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3132 }
3133 
3134 #ifdef CONFIG_HOTPLUG_CPU
3135 static void sched_tick_stop(int cpu)
3136 {
3137 	struct tick_work *twork;
3138 
3139 	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3140 		return;
3141 
3142 	WARN_ON_ONCE(!tick_work_cpu);
3143 
3144 	twork = per_cpu_ptr(tick_work_cpu, cpu);
3145 	cancel_delayed_work_sync(&twork->work);
3146 }
3147 #endif /* CONFIG_HOTPLUG_CPU */
3148 
3149 int __init sched_tick_offload_init(void)
3150 {
3151 	tick_work_cpu = alloc_percpu(struct tick_work);
3152 	BUG_ON(!tick_work_cpu);
3153 
3154 	return 0;
3155 }
3156 
3157 #else /* !CONFIG_NO_HZ_FULL */
3158 static inline void sched_tick_start(int cpu) { }
3159 static inline void sched_tick_stop(int cpu) { }
3160 #endif
3161 
3162 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3163 				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3164 /*
3165  * If the value passed in is equal to the current preempt count
3166  * then we just disabled preemption. Start timing the latency.
3167  */
3168 static inline void preempt_latency_start(int val)
3169 {
3170 	if (preempt_count() == val) {
3171 		unsigned long ip = get_lock_parent_ip();
3172 #ifdef CONFIG_DEBUG_PREEMPT
3173 		current->preempt_disable_ip = ip;
3174 #endif
3175 		trace_preempt_off(CALLER_ADDR0, ip);
3176 	}
3177 }
3178 
3179 void preempt_count_add(int val)
3180 {
3181 #ifdef CONFIG_DEBUG_PREEMPT
3182 	/*
3183 	 * Underflow?
3184 	 */
3185 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3186 		return;
3187 #endif
3188 	__preempt_count_add(val);
3189 #ifdef CONFIG_DEBUG_PREEMPT
3190 	/*
3191 	 * Spinlock count overflowing soon?
3192 	 */
3193 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3194 				PREEMPT_MASK - 10);
3195 #endif
3196 	preempt_latency_start(val);
3197 }
3198 EXPORT_SYMBOL(preempt_count_add);
3199 NOKPROBE_SYMBOL(preempt_count_add);
3200 
3201 /*
3202  * If the value passed in equals to the current preempt count
3203  * then we just enabled preemption. Stop timing the latency.
3204  */
3205 static inline void preempt_latency_stop(int val)
3206 {
3207 	if (preempt_count() == val)
3208 		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3209 }
3210 
3211 void preempt_count_sub(int val)
3212 {
3213 #ifdef CONFIG_DEBUG_PREEMPT
3214 	/*
3215 	 * Underflow?
3216 	 */
3217 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3218 		return;
3219 	/*
3220 	 * Is the spinlock portion underflowing?
3221 	 */
3222 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3223 			!(preempt_count() & PREEMPT_MASK)))
3224 		return;
3225 #endif
3226 
3227 	preempt_latency_stop(val);
3228 	__preempt_count_sub(val);
3229 }
3230 EXPORT_SYMBOL(preempt_count_sub);
3231 NOKPROBE_SYMBOL(preempt_count_sub);
3232 
3233 #else
3234 static inline void preempt_latency_start(int val) { }
3235 static inline void preempt_latency_stop(int val) { }
3236 #endif
3237 
3238 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3239 {
3240 #ifdef CONFIG_DEBUG_PREEMPT
3241 	return p->preempt_disable_ip;
3242 #else
3243 	return 0;
3244 #endif
3245 }
3246 
3247 /*
3248  * Print scheduling while atomic bug:
3249  */
3250 static noinline void __schedule_bug(struct task_struct *prev)
3251 {
3252 	/* Save this before calling printk(), since that will clobber it */
3253 	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3254 
3255 	if (oops_in_progress)
3256 		return;
3257 
3258 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3259 		prev->comm, prev->pid, preempt_count());
3260 
3261 	debug_show_held_locks(prev);
3262 	print_modules();
3263 	if (irqs_disabled())
3264 		print_irqtrace_events(prev);
3265 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3266 	    && in_atomic_preempt_off()) {
3267 		pr_err("Preemption disabled at:");
3268 		print_ip_sym(preempt_disable_ip);
3269 		pr_cont("\n");
3270 	}
3271 	if (panic_on_warn)
3272 		panic("scheduling while atomic\n");
3273 
3274 	dump_stack();
3275 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3276 }
3277 
3278 /*
3279  * Various schedule()-time debugging checks and statistics:
3280  */
3281 static inline void schedule_debug(struct task_struct *prev)
3282 {
3283 #ifdef CONFIG_SCHED_STACK_END_CHECK
3284 	if (task_stack_end_corrupted(prev))
3285 		panic("corrupted stack end detected inside scheduler\n");
3286 #endif
3287 
3288 	if (unlikely(in_atomic_preempt_off())) {
3289 		__schedule_bug(prev);
3290 		preempt_count_set(PREEMPT_DISABLED);
3291 	}
3292 	rcu_sleep_check();
3293 
3294 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3295 
3296 	schedstat_inc(this_rq()->sched_count);
3297 }
3298 
3299 /*
3300  * Pick up the highest-prio task:
3301  */
3302 static inline struct task_struct *
3303 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3304 {
3305 	const struct sched_class *class;
3306 	struct task_struct *p;
3307 
3308 	/*
3309 	 * Optimization: we know that if all tasks are in the fair class we can
3310 	 * call that function directly, but only if the @prev task wasn't of a
3311 	 * higher scheduling class, because otherwise those loose the
3312 	 * opportunity to pull in more work from other CPUs.
3313 	 */
3314 	if (likely((prev->sched_class == &idle_sched_class ||
3315 		    prev->sched_class == &fair_sched_class) &&
3316 		   rq->nr_running == rq->cfs.h_nr_running)) {
3317 
3318 		p = fair_sched_class.pick_next_task(rq, prev, rf);
3319 		if (unlikely(p == RETRY_TASK))
3320 			goto again;
3321 
3322 		/* Assumes fair_sched_class->next == idle_sched_class */
3323 		if (unlikely(!p))
3324 			p = idle_sched_class.pick_next_task(rq, prev, rf);
3325 
3326 		return p;
3327 	}
3328 
3329 again:
3330 	for_each_class(class) {
3331 		p = class->pick_next_task(rq, prev, rf);
3332 		if (p) {
3333 			if (unlikely(p == RETRY_TASK))
3334 				goto again;
3335 			return p;
3336 		}
3337 	}
3338 
3339 	/* The idle class should always have a runnable task: */
3340 	BUG();
3341 }
3342 
3343 /*
3344  * __schedule() is the main scheduler function.
3345  *
3346  * The main means of driving the scheduler and thus entering this function are:
3347  *
3348  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3349  *
3350  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3351  *      paths. For example, see arch/x86/entry_64.S.
3352  *
3353  *      To drive preemption between tasks, the scheduler sets the flag in timer
3354  *      interrupt handler scheduler_tick().
3355  *
3356  *   3. Wakeups don't really cause entry into schedule(). They add a
3357  *      task to the run-queue and that's it.
3358  *
3359  *      Now, if the new task added to the run-queue preempts the current
3360  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3361  *      called on the nearest possible occasion:
3362  *
3363  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
3364  *
3365  *         - in syscall or exception context, at the next outmost
3366  *           preempt_enable(). (this might be as soon as the wake_up()'s
3367  *           spin_unlock()!)
3368  *
3369  *         - in IRQ context, return from interrupt-handler to
3370  *           preemptible context
3371  *
3372  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3373  *         then at the next:
3374  *
3375  *          - cond_resched() call
3376  *          - explicit schedule() call
3377  *          - return from syscall or exception to user-space
3378  *          - return from interrupt-handler to user-space
3379  *
3380  * WARNING: must be called with preemption disabled!
3381  */
3382 static void __sched notrace __schedule(bool preempt)
3383 {
3384 	struct task_struct *prev, *next;
3385 	unsigned long *switch_count;
3386 	struct rq_flags rf;
3387 	struct rq *rq;
3388 	int cpu;
3389 
3390 	cpu = smp_processor_id();
3391 	rq = cpu_rq(cpu);
3392 	prev = rq->curr;
3393 
3394 	schedule_debug(prev);
3395 
3396 	if (sched_feat(HRTICK))
3397 		hrtick_clear(rq);
3398 
3399 	local_irq_disable();
3400 	rcu_note_context_switch(preempt);
3401 
3402 	/*
3403 	 * Make sure that signal_pending_state()->signal_pending() below
3404 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3405 	 * done by the caller to avoid the race with signal_wake_up().
3406 	 *
3407 	 * The membarrier system call requires a full memory barrier
3408 	 * after coming from user-space, before storing to rq->curr.
3409 	 */
3410 	rq_lock(rq, &rf);
3411 	smp_mb__after_spinlock();
3412 
3413 	/* Promote REQ to ACT */
3414 	rq->clock_update_flags <<= 1;
3415 	update_rq_clock(rq);
3416 
3417 	switch_count = &prev->nivcsw;
3418 	if (!preempt && prev->state) {
3419 		if (signal_pending_state(prev->state, prev)) {
3420 			prev->state = TASK_RUNNING;
3421 		} else {
3422 			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3423 			prev->on_rq = 0;
3424 
3425 			if (prev->in_iowait) {
3426 				atomic_inc(&rq->nr_iowait);
3427 				delayacct_blkio_start();
3428 			}
3429 
3430 			/*
3431 			 * If a worker went to sleep, notify and ask workqueue
3432 			 * whether it wants to wake up a task to maintain
3433 			 * concurrency.
3434 			 */
3435 			if (prev->flags & PF_WQ_WORKER) {
3436 				struct task_struct *to_wakeup;
3437 
3438 				to_wakeup = wq_worker_sleeping(prev);
3439 				if (to_wakeup)
3440 					try_to_wake_up_local(to_wakeup, &rf);
3441 			}
3442 		}
3443 		switch_count = &prev->nvcsw;
3444 	}
3445 
3446 	next = pick_next_task(rq, prev, &rf);
3447 	clear_tsk_need_resched(prev);
3448 	clear_preempt_need_resched();
3449 
3450 	if (likely(prev != next)) {
3451 		rq->nr_switches++;
3452 		rq->curr = next;
3453 		/*
3454 		 * The membarrier system call requires each architecture
3455 		 * to have a full memory barrier after updating
3456 		 * rq->curr, before returning to user-space.
3457 		 *
3458 		 * Here are the schemes providing that barrier on the
3459 		 * various architectures:
3460 		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3461 		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3462 		 * - finish_lock_switch() for weakly-ordered
3463 		 *   architectures where spin_unlock is a full barrier,
3464 		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3465 		 *   is a RELEASE barrier),
3466 		 */
3467 		++*switch_count;
3468 
3469 		trace_sched_switch(preempt, prev, next);
3470 
3471 		/* Also unlocks the rq: */
3472 		rq = context_switch(rq, prev, next, &rf);
3473 	} else {
3474 		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3475 		rq_unlock_irq(rq, &rf);
3476 	}
3477 
3478 	balance_callback(rq);
3479 }
3480 
3481 void __noreturn do_task_dead(void)
3482 {
3483 	/* Causes final put_task_struct in finish_task_switch(): */
3484 	set_special_state(TASK_DEAD);
3485 
3486 	/* Tell freezer to ignore us: */
3487 	current->flags |= PF_NOFREEZE;
3488 
3489 	__schedule(false);
3490 	BUG();
3491 
3492 	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3493 	for (;;)
3494 		cpu_relax();
3495 }
3496 
3497 static inline void sched_submit_work(struct task_struct *tsk)
3498 {
3499 	if (!tsk->state || tsk_is_pi_blocked(tsk))
3500 		return;
3501 	/*
3502 	 * If we are going to sleep and we have plugged IO queued,
3503 	 * make sure to submit it to avoid deadlocks.
3504 	 */
3505 	if (blk_needs_flush_plug(tsk))
3506 		blk_schedule_flush_plug(tsk);
3507 }
3508 
3509 asmlinkage __visible void __sched schedule(void)
3510 {
3511 	struct task_struct *tsk = current;
3512 
3513 	sched_submit_work(tsk);
3514 	do {
3515 		preempt_disable();
3516 		__schedule(false);
3517 		sched_preempt_enable_no_resched();
3518 	} while (need_resched());
3519 }
3520 EXPORT_SYMBOL(schedule);
3521 
3522 /*
3523  * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3524  * state (have scheduled out non-voluntarily) by making sure that all
3525  * tasks have either left the run queue or have gone into user space.
3526  * As idle tasks do not do either, they must not ever be preempted
3527  * (schedule out non-voluntarily).
3528  *
3529  * schedule_idle() is similar to schedule_preempt_disable() except that it
3530  * never enables preemption because it does not call sched_submit_work().
3531  */
3532 void __sched schedule_idle(void)
3533 {
3534 	/*
3535 	 * As this skips calling sched_submit_work(), which the idle task does
3536 	 * regardless because that function is a nop when the task is in a
3537 	 * TASK_RUNNING state, make sure this isn't used someplace that the
3538 	 * current task can be in any other state. Note, idle is always in the
3539 	 * TASK_RUNNING state.
3540 	 */
3541 	WARN_ON_ONCE(current->state);
3542 	do {
3543 		__schedule(false);
3544 	} while (need_resched());
3545 }
3546 
3547 #ifdef CONFIG_CONTEXT_TRACKING
3548 asmlinkage __visible void __sched schedule_user(void)
3549 {
3550 	/*
3551 	 * If we come here after a random call to set_need_resched(),
3552 	 * or we have been woken up remotely but the IPI has not yet arrived,
3553 	 * we haven't yet exited the RCU idle mode. Do it here manually until
3554 	 * we find a better solution.
3555 	 *
3556 	 * NB: There are buggy callers of this function.  Ideally we
3557 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
3558 	 * too frequently to make sense yet.
3559 	 */
3560 	enum ctx_state prev_state = exception_enter();
3561 	schedule();
3562 	exception_exit(prev_state);
3563 }
3564 #endif
3565 
3566 /**
3567  * schedule_preempt_disabled - called with preemption disabled
3568  *
3569  * Returns with preemption disabled. Note: preempt_count must be 1
3570  */
3571 void __sched schedule_preempt_disabled(void)
3572 {
3573 	sched_preempt_enable_no_resched();
3574 	schedule();
3575 	preempt_disable();
3576 }
3577 
3578 static void __sched notrace preempt_schedule_common(void)
3579 {
3580 	do {
3581 		/*
3582 		 * Because the function tracer can trace preempt_count_sub()
3583 		 * and it also uses preempt_enable/disable_notrace(), if
3584 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3585 		 * by the function tracer will call this function again and
3586 		 * cause infinite recursion.
3587 		 *
3588 		 * Preemption must be disabled here before the function
3589 		 * tracer can trace. Break up preempt_disable() into two
3590 		 * calls. One to disable preemption without fear of being
3591 		 * traced. The other to still record the preemption latency,
3592 		 * which can also be traced by the function tracer.
3593 		 */
3594 		preempt_disable_notrace();
3595 		preempt_latency_start(1);
3596 		__schedule(true);
3597 		preempt_latency_stop(1);
3598 		preempt_enable_no_resched_notrace();
3599 
3600 		/*
3601 		 * Check again in case we missed a preemption opportunity
3602 		 * between schedule and now.
3603 		 */
3604 	} while (need_resched());
3605 }
3606 
3607 #ifdef CONFIG_PREEMPT
3608 /*
3609  * this is the entry point to schedule() from in-kernel preemption
3610  * off of preempt_enable. Kernel preemptions off return from interrupt
3611  * occur there and call schedule directly.
3612  */
3613 asmlinkage __visible void __sched notrace preempt_schedule(void)
3614 {
3615 	/*
3616 	 * If there is a non-zero preempt_count or interrupts are disabled,
3617 	 * we do not want to preempt the current task. Just return..
3618 	 */
3619 	if (likely(!preemptible()))
3620 		return;
3621 
3622 	preempt_schedule_common();
3623 }
3624 NOKPROBE_SYMBOL(preempt_schedule);
3625 EXPORT_SYMBOL(preempt_schedule);
3626 
3627 /**
3628  * preempt_schedule_notrace - preempt_schedule called by tracing
3629  *
3630  * The tracing infrastructure uses preempt_enable_notrace to prevent
3631  * recursion and tracing preempt enabling caused by the tracing
3632  * infrastructure itself. But as tracing can happen in areas coming
3633  * from userspace or just about to enter userspace, a preempt enable
3634  * can occur before user_exit() is called. This will cause the scheduler
3635  * to be called when the system is still in usermode.
3636  *
3637  * To prevent this, the preempt_enable_notrace will use this function
3638  * instead of preempt_schedule() to exit user context if needed before
3639  * calling the scheduler.
3640  */
3641 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3642 {
3643 	enum ctx_state prev_ctx;
3644 
3645 	if (likely(!preemptible()))
3646 		return;
3647 
3648 	do {
3649 		/*
3650 		 * Because the function tracer can trace preempt_count_sub()
3651 		 * and it also uses preempt_enable/disable_notrace(), if
3652 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3653 		 * by the function tracer will call this function again and
3654 		 * cause infinite recursion.
3655 		 *
3656 		 * Preemption must be disabled here before the function
3657 		 * tracer can trace. Break up preempt_disable() into two
3658 		 * calls. One to disable preemption without fear of being
3659 		 * traced. The other to still record the preemption latency,
3660 		 * which can also be traced by the function tracer.
3661 		 */
3662 		preempt_disable_notrace();
3663 		preempt_latency_start(1);
3664 		/*
3665 		 * Needs preempt disabled in case user_exit() is traced
3666 		 * and the tracer calls preempt_enable_notrace() causing
3667 		 * an infinite recursion.
3668 		 */
3669 		prev_ctx = exception_enter();
3670 		__schedule(true);
3671 		exception_exit(prev_ctx);
3672 
3673 		preempt_latency_stop(1);
3674 		preempt_enable_no_resched_notrace();
3675 	} while (need_resched());
3676 }
3677 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3678 
3679 #endif /* CONFIG_PREEMPT */
3680 
3681 /*
3682  * this is the entry point to schedule() from kernel preemption
3683  * off of irq context.
3684  * Note, that this is called and return with irqs disabled. This will
3685  * protect us against recursive calling from irq.
3686  */
3687 asmlinkage __visible void __sched preempt_schedule_irq(void)
3688 {
3689 	enum ctx_state prev_state;
3690 
3691 	/* Catch callers which need to be fixed */
3692 	BUG_ON(preempt_count() || !irqs_disabled());
3693 
3694 	prev_state = exception_enter();
3695 
3696 	do {
3697 		preempt_disable();
3698 		local_irq_enable();
3699 		__schedule(true);
3700 		local_irq_disable();
3701 		sched_preempt_enable_no_resched();
3702 	} while (need_resched());
3703 
3704 	exception_exit(prev_state);
3705 }
3706 
3707 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3708 			  void *key)
3709 {
3710 	return try_to_wake_up(curr->private, mode, wake_flags);
3711 }
3712 EXPORT_SYMBOL(default_wake_function);
3713 
3714 #ifdef CONFIG_RT_MUTEXES
3715 
3716 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3717 {
3718 	if (pi_task)
3719 		prio = min(prio, pi_task->prio);
3720 
3721 	return prio;
3722 }
3723 
3724 static inline int rt_effective_prio(struct task_struct *p, int prio)
3725 {
3726 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
3727 
3728 	return __rt_effective_prio(pi_task, prio);
3729 }
3730 
3731 /*
3732  * rt_mutex_setprio - set the current priority of a task
3733  * @p: task to boost
3734  * @pi_task: donor task
3735  *
3736  * This function changes the 'effective' priority of a task. It does
3737  * not touch ->normal_prio like __setscheduler().
3738  *
3739  * Used by the rt_mutex code to implement priority inheritance
3740  * logic. Call site only calls if the priority of the task changed.
3741  */
3742 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3743 {
3744 	int prio, oldprio, queued, running, queue_flag =
3745 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3746 	const struct sched_class *prev_class;
3747 	struct rq_flags rf;
3748 	struct rq *rq;
3749 
3750 	/* XXX used to be waiter->prio, not waiter->task->prio */
3751 	prio = __rt_effective_prio(pi_task, p->normal_prio);
3752 
3753 	/*
3754 	 * If nothing changed; bail early.
3755 	 */
3756 	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3757 		return;
3758 
3759 	rq = __task_rq_lock(p, &rf);
3760 	update_rq_clock(rq);
3761 	/*
3762 	 * Set under pi_lock && rq->lock, such that the value can be used under
3763 	 * either lock.
3764 	 *
3765 	 * Note that there is loads of tricky to make this pointer cache work
3766 	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3767 	 * ensure a task is de-boosted (pi_task is set to NULL) before the
3768 	 * task is allowed to run again (and can exit). This ensures the pointer
3769 	 * points to a blocked task -- which guaratees the task is present.
3770 	 */
3771 	p->pi_top_task = pi_task;
3772 
3773 	/*
3774 	 * For FIFO/RR we only need to set prio, if that matches we're done.
3775 	 */
3776 	if (prio == p->prio && !dl_prio(prio))
3777 		goto out_unlock;
3778 
3779 	/*
3780 	 * Idle task boosting is a nono in general. There is one
3781 	 * exception, when PREEMPT_RT and NOHZ is active:
3782 	 *
3783 	 * The idle task calls get_next_timer_interrupt() and holds
3784 	 * the timer wheel base->lock on the CPU and another CPU wants
3785 	 * to access the timer (probably to cancel it). We can safely
3786 	 * ignore the boosting request, as the idle CPU runs this code
3787 	 * with interrupts disabled and will complete the lock
3788 	 * protected section without being interrupted. So there is no
3789 	 * real need to boost.
3790 	 */
3791 	if (unlikely(p == rq->idle)) {
3792 		WARN_ON(p != rq->curr);
3793 		WARN_ON(p->pi_blocked_on);
3794 		goto out_unlock;
3795 	}
3796 
3797 	trace_sched_pi_setprio(p, pi_task);
3798 	oldprio = p->prio;
3799 
3800 	if (oldprio == prio)
3801 		queue_flag &= ~DEQUEUE_MOVE;
3802 
3803 	prev_class = p->sched_class;
3804 	queued = task_on_rq_queued(p);
3805 	running = task_current(rq, p);
3806 	if (queued)
3807 		dequeue_task(rq, p, queue_flag);
3808 	if (running)
3809 		put_prev_task(rq, p);
3810 
3811 	/*
3812 	 * Boosting condition are:
3813 	 * 1. -rt task is running and holds mutex A
3814 	 *      --> -dl task blocks on mutex A
3815 	 *
3816 	 * 2. -dl task is running and holds mutex A
3817 	 *      --> -dl task blocks on mutex A and could preempt the
3818 	 *          running task
3819 	 */
3820 	if (dl_prio(prio)) {
3821 		if (!dl_prio(p->normal_prio) ||
3822 		    (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3823 			p->dl.dl_boosted = 1;
3824 			queue_flag |= ENQUEUE_REPLENISH;
3825 		} else
3826 			p->dl.dl_boosted = 0;
3827 		p->sched_class = &dl_sched_class;
3828 	} else if (rt_prio(prio)) {
3829 		if (dl_prio(oldprio))
3830 			p->dl.dl_boosted = 0;
3831 		if (oldprio < prio)
3832 			queue_flag |= ENQUEUE_HEAD;
3833 		p->sched_class = &rt_sched_class;
3834 	} else {
3835 		if (dl_prio(oldprio))
3836 			p->dl.dl_boosted = 0;
3837 		if (rt_prio(oldprio))
3838 			p->rt.timeout = 0;
3839 		p->sched_class = &fair_sched_class;
3840 	}
3841 
3842 	p->prio = prio;
3843 
3844 	if (queued)
3845 		enqueue_task(rq, p, queue_flag);
3846 	if (running)
3847 		set_curr_task(rq, p);
3848 
3849 	check_class_changed(rq, p, prev_class, oldprio);
3850 out_unlock:
3851 	/* Avoid rq from going away on us: */
3852 	preempt_disable();
3853 	__task_rq_unlock(rq, &rf);
3854 
3855 	balance_callback(rq);
3856 	preempt_enable();
3857 }
3858 #else
3859 static inline int rt_effective_prio(struct task_struct *p, int prio)
3860 {
3861 	return prio;
3862 }
3863 #endif
3864 
3865 void set_user_nice(struct task_struct *p, long nice)
3866 {
3867 	bool queued, running;
3868 	int old_prio, delta;
3869 	struct rq_flags rf;
3870 	struct rq *rq;
3871 
3872 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3873 		return;
3874 	/*
3875 	 * We have to be careful, if called from sys_setpriority(),
3876 	 * the task might be in the middle of scheduling on another CPU.
3877 	 */
3878 	rq = task_rq_lock(p, &rf);
3879 	update_rq_clock(rq);
3880 
3881 	/*
3882 	 * The RT priorities are set via sched_setscheduler(), but we still
3883 	 * allow the 'normal' nice value to be set - but as expected
3884 	 * it wont have any effect on scheduling until the task is
3885 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3886 	 */
3887 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3888 		p->static_prio = NICE_TO_PRIO(nice);
3889 		goto out_unlock;
3890 	}
3891 	queued = task_on_rq_queued(p);
3892 	running = task_current(rq, p);
3893 	if (queued)
3894 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3895 	if (running)
3896 		put_prev_task(rq, p);
3897 
3898 	p->static_prio = NICE_TO_PRIO(nice);
3899 	set_load_weight(p, true);
3900 	old_prio = p->prio;
3901 	p->prio = effective_prio(p);
3902 	delta = p->prio - old_prio;
3903 
3904 	if (queued) {
3905 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3906 		/*
3907 		 * If the task increased its priority or is running and
3908 		 * lowered its priority, then reschedule its CPU:
3909 		 */
3910 		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3911 			resched_curr(rq);
3912 	}
3913 	if (running)
3914 		set_curr_task(rq, p);
3915 out_unlock:
3916 	task_rq_unlock(rq, p, &rf);
3917 }
3918 EXPORT_SYMBOL(set_user_nice);
3919 
3920 /*
3921  * can_nice - check if a task can reduce its nice value
3922  * @p: task
3923  * @nice: nice value
3924  */
3925 int can_nice(const struct task_struct *p, const int nice)
3926 {
3927 	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
3928 	int nice_rlim = nice_to_rlimit(nice);
3929 
3930 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3931 		capable(CAP_SYS_NICE));
3932 }
3933 
3934 #ifdef __ARCH_WANT_SYS_NICE
3935 
3936 /*
3937  * sys_nice - change the priority of the current process.
3938  * @increment: priority increment
3939  *
3940  * sys_setpriority is a more generic, but much slower function that
3941  * does similar things.
3942  */
3943 SYSCALL_DEFINE1(nice, int, increment)
3944 {
3945 	long nice, retval;
3946 
3947 	/*
3948 	 * Setpriority might change our priority at the same moment.
3949 	 * We don't have to worry. Conceptually one call occurs first
3950 	 * and we have a single winner.
3951 	 */
3952 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3953 	nice = task_nice(current) + increment;
3954 
3955 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3956 	if (increment < 0 && !can_nice(current, nice))
3957 		return -EPERM;
3958 
3959 	retval = security_task_setnice(current, nice);
3960 	if (retval)
3961 		return retval;
3962 
3963 	set_user_nice(current, nice);
3964 	return 0;
3965 }
3966 
3967 #endif
3968 
3969 /**
3970  * task_prio - return the priority value of a given task.
3971  * @p: the task in question.
3972  *
3973  * Return: The priority value as seen by users in /proc.
3974  * RT tasks are offset by -200. Normal tasks are centered
3975  * around 0, value goes from -16 to +15.
3976  */
3977 int task_prio(const struct task_struct *p)
3978 {
3979 	return p->prio - MAX_RT_PRIO;
3980 }
3981 
3982 /**
3983  * idle_cpu - is a given CPU idle currently?
3984  * @cpu: the processor in question.
3985  *
3986  * Return: 1 if the CPU is currently idle. 0 otherwise.
3987  */
3988 int idle_cpu(int cpu)
3989 {
3990 	struct rq *rq = cpu_rq(cpu);
3991 
3992 	if (rq->curr != rq->idle)
3993 		return 0;
3994 
3995 	if (rq->nr_running)
3996 		return 0;
3997 
3998 #ifdef CONFIG_SMP
3999 	if (!llist_empty(&rq->wake_list))
4000 		return 0;
4001 #endif
4002 
4003 	return 1;
4004 }
4005 
4006 /**
4007  * available_idle_cpu - is a given CPU idle for enqueuing work.
4008  * @cpu: the CPU in question.
4009  *
4010  * Return: 1 if the CPU is currently idle. 0 otherwise.
4011  */
4012 int available_idle_cpu(int cpu)
4013 {
4014 	if (!idle_cpu(cpu))
4015 		return 0;
4016 
4017 	if (vcpu_is_preempted(cpu))
4018 		return 0;
4019 
4020 	return 1;
4021 }
4022 
4023 /**
4024  * idle_task - return the idle task for a given CPU.
4025  * @cpu: the processor in question.
4026  *
4027  * Return: The idle task for the CPU @cpu.
4028  */
4029 struct task_struct *idle_task(int cpu)
4030 {
4031 	return cpu_rq(cpu)->idle;
4032 }
4033 
4034 /**
4035  * find_process_by_pid - find a process with a matching PID value.
4036  * @pid: the pid in question.
4037  *
4038  * The task of @pid, if found. %NULL otherwise.
4039  */
4040 static struct task_struct *find_process_by_pid(pid_t pid)
4041 {
4042 	return pid ? find_task_by_vpid(pid) : current;
4043 }
4044 
4045 /*
4046  * sched_setparam() passes in -1 for its policy, to let the functions
4047  * it calls know not to change it.
4048  */
4049 #define SETPARAM_POLICY	-1
4050 
4051 static void __setscheduler_params(struct task_struct *p,
4052 		const struct sched_attr *attr)
4053 {
4054 	int policy = attr->sched_policy;
4055 
4056 	if (policy == SETPARAM_POLICY)
4057 		policy = p->policy;
4058 
4059 	p->policy = policy;
4060 
4061 	if (dl_policy(policy))
4062 		__setparam_dl(p, attr);
4063 	else if (fair_policy(policy))
4064 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4065 
4066 	/*
4067 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4068 	 * !rt_policy. Always setting this ensures that things like
4069 	 * getparam()/getattr() don't report silly values for !rt tasks.
4070 	 */
4071 	p->rt_priority = attr->sched_priority;
4072 	p->normal_prio = normal_prio(p);
4073 	set_load_weight(p, true);
4074 }
4075 
4076 /* Actually do priority change: must hold pi & rq lock. */
4077 static void __setscheduler(struct rq *rq, struct task_struct *p,
4078 			   const struct sched_attr *attr, bool keep_boost)
4079 {
4080 	__setscheduler_params(p, attr);
4081 
4082 	/*
4083 	 * Keep a potential priority boosting if called from
4084 	 * sched_setscheduler().
4085 	 */
4086 	p->prio = normal_prio(p);
4087 	if (keep_boost)
4088 		p->prio = rt_effective_prio(p, p->prio);
4089 
4090 	if (dl_prio(p->prio))
4091 		p->sched_class = &dl_sched_class;
4092 	else if (rt_prio(p->prio))
4093 		p->sched_class = &rt_sched_class;
4094 	else
4095 		p->sched_class = &fair_sched_class;
4096 }
4097 
4098 /*
4099  * Check the target process has a UID that matches the current process's:
4100  */
4101 static bool check_same_owner(struct task_struct *p)
4102 {
4103 	const struct cred *cred = current_cred(), *pcred;
4104 	bool match;
4105 
4106 	rcu_read_lock();
4107 	pcred = __task_cred(p);
4108 	match = (uid_eq(cred->euid, pcred->euid) ||
4109 		 uid_eq(cred->euid, pcred->uid));
4110 	rcu_read_unlock();
4111 	return match;
4112 }
4113 
4114 static int __sched_setscheduler(struct task_struct *p,
4115 				const struct sched_attr *attr,
4116 				bool user, bool pi)
4117 {
4118 	int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4119 		      MAX_RT_PRIO - 1 - attr->sched_priority;
4120 	int retval, oldprio, oldpolicy = -1, queued, running;
4121 	int new_effective_prio, policy = attr->sched_policy;
4122 	const struct sched_class *prev_class;
4123 	struct rq_flags rf;
4124 	int reset_on_fork;
4125 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4126 	struct rq *rq;
4127 
4128 	/* The pi code expects interrupts enabled */
4129 	BUG_ON(pi && in_interrupt());
4130 recheck:
4131 	/* Double check policy once rq lock held: */
4132 	if (policy < 0) {
4133 		reset_on_fork = p->sched_reset_on_fork;
4134 		policy = oldpolicy = p->policy;
4135 	} else {
4136 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4137 
4138 		if (!valid_policy(policy))
4139 			return -EINVAL;
4140 	}
4141 
4142 	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4143 		return -EINVAL;
4144 
4145 	/*
4146 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
4147 	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4148 	 * SCHED_BATCH and SCHED_IDLE is 0.
4149 	 */
4150 	if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4151 	    (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4152 		return -EINVAL;
4153 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4154 	    (rt_policy(policy) != (attr->sched_priority != 0)))
4155 		return -EINVAL;
4156 
4157 	/*
4158 	 * Allow unprivileged RT tasks to decrease priority:
4159 	 */
4160 	if (user && !capable(CAP_SYS_NICE)) {
4161 		if (fair_policy(policy)) {
4162 			if (attr->sched_nice < task_nice(p) &&
4163 			    !can_nice(p, attr->sched_nice))
4164 				return -EPERM;
4165 		}
4166 
4167 		if (rt_policy(policy)) {
4168 			unsigned long rlim_rtprio =
4169 					task_rlimit(p, RLIMIT_RTPRIO);
4170 
4171 			/* Can't set/change the rt policy: */
4172 			if (policy != p->policy && !rlim_rtprio)
4173 				return -EPERM;
4174 
4175 			/* Can't increase priority: */
4176 			if (attr->sched_priority > p->rt_priority &&
4177 			    attr->sched_priority > rlim_rtprio)
4178 				return -EPERM;
4179 		}
4180 
4181 		 /*
4182 		  * Can't set/change SCHED_DEADLINE policy at all for now
4183 		  * (safest behavior); in the future we would like to allow
4184 		  * unprivileged DL tasks to increase their relative deadline
4185 		  * or reduce their runtime (both ways reducing utilization)
4186 		  */
4187 		if (dl_policy(policy))
4188 			return -EPERM;
4189 
4190 		/*
4191 		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4192 		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4193 		 */
4194 		if (task_has_idle_policy(p) && !idle_policy(policy)) {
4195 			if (!can_nice(p, task_nice(p)))
4196 				return -EPERM;
4197 		}
4198 
4199 		/* Can't change other user's priorities: */
4200 		if (!check_same_owner(p))
4201 			return -EPERM;
4202 
4203 		/* Normal users shall not reset the sched_reset_on_fork flag: */
4204 		if (p->sched_reset_on_fork && !reset_on_fork)
4205 			return -EPERM;
4206 	}
4207 
4208 	if (user) {
4209 		if (attr->sched_flags & SCHED_FLAG_SUGOV)
4210 			return -EINVAL;
4211 
4212 		retval = security_task_setscheduler(p);
4213 		if (retval)
4214 			return retval;
4215 	}
4216 
4217 	/*
4218 	 * Make sure no PI-waiters arrive (or leave) while we are
4219 	 * changing the priority of the task:
4220 	 *
4221 	 * To be able to change p->policy safely, the appropriate
4222 	 * runqueue lock must be held.
4223 	 */
4224 	rq = task_rq_lock(p, &rf);
4225 	update_rq_clock(rq);
4226 
4227 	/*
4228 	 * Changing the policy of the stop threads its a very bad idea:
4229 	 */
4230 	if (p == rq->stop) {
4231 		task_rq_unlock(rq, p, &rf);
4232 		return -EINVAL;
4233 	}
4234 
4235 	/*
4236 	 * If not changing anything there's no need to proceed further,
4237 	 * but store a possible modification of reset_on_fork.
4238 	 */
4239 	if (unlikely(policy == p->policy)) {
4240 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4241 			goto change;
4242 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4243 			goto change;
4244 		if (dl_policy(policy) && dl_param_changed(p, attr))
4245 			goto change;
4246 
4247 		p->sched_reset_on_fork = reset_on_fork;
4248 		task_rq_unlock(rq, p, &rf);
4249 		return 0;
4250 	}
4251 change:
4252 
4253 	if (user) {
4254 #ifdef CONFIG_RT_GROUP_SCHED
4255 		/*
4256 		 * Do not allow realtime tasks into groups that have no runtime
4257 		 * assigned.
4258 		 */
4259 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
4260 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4261 				!task_group_is_autogroup(task_group(p))) {
4262 			task_rq_unlock(rq, p, &rf);
4263 			return -EPERM;
4264 		}
4265 #endif
4266 #ifdef CONFIG_SMP
4267 		if (dl_bandwidth_enabled() && dl_policy(policy) &&
4268 				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4269 			cpumask_t *span = rq->rd->span;
4270 
4271 			/*
4272 			 * Don't allow tasks with an affinity mask smaller than
4273 			 * the entire root_domain to become SCHED_DEADLINE. We
4274 			 * will also fail if there's no bandwidth available.
4275 			 */
4276 			if (!cpumask_subset(span, &p->cpus_allowed) ||
4277 			    rq->rd->dl_bw.bw == 0) {
4278 				task_rq_unlock(rq, p, &rf);
4279 				return -EPERM;
4280 			}
4281 		}
4282 #endif
4283 	}
4284 
4285 	/* Re-check policy now with rq lock held: */
4286 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4287 		policy = oldpolicy = -1;
4288 		task_rq_unlock(rq, p, &rf);
4289 		goto recheck;
4290 	}
4291 
4292 	/*
4293 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4294 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4295 	 * is available.
4296 	 */
4297 	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4298 		task_rq_unlock(rq, p, &rf);
4299 		return -EBUSY;
4300 	}
4301 
4302 	p->sched_reset_on_fork = reset_on_fork;
4303 	oldprio = p->prio;
4304 
4305 	if (pi) {
4306 		/*
4307 		 * Take priority boosted tasks into account. If the new
4308 		 * effective priority is unchanged, we just store the new
4309 		 * normal parameters and do not touch the scheduler class and
4310 		 * the runqueue. This will be done when the task deboost
4311 		 * itself.
4312 		 */
4313 		new_effective_prio = rt_effective_prio(p, newprio);
4314 		if (new_effective_prio == oldprio)
4315 			queue_flags &= ~DEQUEUE_MOVE;
4316 	}
4317 
4318 	queued = task_on_rq_queued(p);
4319 	running = task_current(rq, p);
4320 	if (queued)
4321 		dequeue_task(rq, p, queue_flags);
4322 	if (running)
4323 		put_prev_task(rq, p);
4324 
4325 	prev_class = p->sched_class;
4326 	__setscheduler(rq, p, attr, pi);
4327 
4328 	if (queued) {
4329 		/*
4330 		 * We enqueue to tail when the priority of a task is
4331 		 * increased (user space view).
4332 		 */
4333 		if (oldprio < p->prio)
4334 			queue_flags |= ENQUEUE_HEAD;
4335 
4336 		enqueue_task(rq, p, queue_flags);
4337 	}
4338 	if (running)
4339 		set_curr_task(rq, p);
4340 
4341 	check_class_changed(rq, p, prev_class, oldprio);
4342 
4343 	/* Avoid rq from going away on us: */
4344 	preempt_disable();
4345 	task_rq_unlock(rq, p, &rf);
4346 
4347 	if (pi)
4348 		rt_mutex_adjust_pi(p);
4349 
4350 	/* Run balance callbacks after we've adjusted the PI chain: */
4351 	balance_callback(rq);
4352 	preempt_enable();
4353 
4354 	return 0;
4355 }
4356 
4357 static int _sched_setscheduler(struct task_struct *p, int policy,
4358 			       const struct sched_param *param, bool check)
4359 {
4360 	struct sched_attr attr = {
4361 		.sched_policy   = policy,
4362 		.sched_priority = param->sched_priority,
4363 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
4364 	};
4365 
4366 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4367 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4368 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4369 		policy &= ~SCHED_RESET_ON_FORK;
4370 		attr.sched_policy = policy;
4371 	}
4372 
4373 	return __sched_setscheduler(p, &attr, check, true);
4374 }
4375 /**
4376  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4377  * @p: the task in question.
4378  * @policy: new policy.
4379  * @param: structure containing the new RT priority.
4380  *
4381  * Return: 0 on success. An error code otherwise.
4382  *
4383  * NOTE that the task may be already dead.
4384  */
4385 int sched_setscheduler(struct task_struct *p, int policy,
4386 		       const struct sched_param *param)
4387 {
4388 	return _sched_setscheduler(p, policy, param, true);
4389 }
4390 EXPORT_SYMBOL_GPL(sched_setscheduler);
4391 
4392 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4393 {
4394 	return __sched_setscheduler(p, attr, true, true);
4395 }
4396 EXPORT_SYMBOL_GPL(sched_setattr);
4397 
4398 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4399 {
4400 	return __sched_setscheduler(p, attr, false, true);
4401 }
4402 
4403 /**
4404  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4405  * @p: the task in question.
4406  * @policy: new policy.
4407  * @param: structure containing the new RT priority.
4408  *
4409  * Just like sched_setscheduler, only don't bother checking if the
4410  * current context has permission.  For example, this is needed in
4411  * stop_machine(): we create temporary high priority worker threads,
4412  * but our caller might not have that capability.
4413  *
4414  * Return: 0 on success. An error code otherwise.
4415  */
4416 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4417 			       const struct sched_param *param)
4418 {
4419 	return _sched_setscheduler(p, policy, param, false);
4420 }
4421 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4422 
4423 static int
4424 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4425 {
4426 	struct sched_param lparam;
4427 	struct task_struct *p;
4428 	int retval;
4429 
4430 	if (!param || pid < 0)
4431 		return -EINVAL;
4432 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4433 		return -EFAULT;
4434 
4435 	rcu_read_lock();
4436 	retval = -ESRCH;
4437 	p = find_process_by_pid(pid);
4438 	if (p != NULL)
4439 		retval = sched_setscheduler(p, policy, &lparam);
4440 	rcu_read_unlock();
4441 
4442 	return retval;
4443 }
4444 
4445 /*
4446  * Mimics kernel/events/core.c perf_copy_attr().
4447  */
4448 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4449 {
4450 	u32 size;
4451 	int ret;
4452 
4453 	if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4454 		return -EFAULT;
4455 
4456 	/* Zero the full structure, so that a short copy will be nice: */
4457 	memset(attr, 0, sizeof(*attr));
4458 
4459 	ret = get_user(size, &uattr->size);
4460 	if (ret)
4461 		return ret;
4462 
4463 	/* Bail out on silly large: */
4464 	if (size > PAGE_SIZE)
4465 		goto err_size;
4466 
4467 	/* ABI compatibility quirk: */
4468 	if (!size)
4469 		size = SCHED_ATTR_SIZE_VER0;
4470 
4471 	if (size < SCHED_ATTR_SIZE_VER0)
4472 		goto err_size;
4473 
4474 	/*
4475 	 * If we're handed a bigger struct than we know of,
4476 	 * ensure all the unknown bits are 0 - i.e. new
4477 	 * user-space does not rely on any kernel feature
4478 	 * extensions we dont know about yet.
4479 	 */
4480 	if (size > sizeof(*attr)) {
4481 		unsigned char __user *addr;
4482 		unsigned char __user *end;
4483 		unsigned char val;
4484 
4485 		addr = (void __user *)uattr + sizeof(*attr);
4486 		end  = (void __user *)uattr + size;
4487 
4488 		for (; addr < end; addr++) {
4489 			ret = get_user(val, addr);
4490 			if (ret)
4491 				return ret;
4492 			if (val)
4493 				goto err_size;
4494 		}
4495 		size = sizeof(*attr);
4496 	}
4497 
4498 	ret = copy_from_user(attr, uattr, size);
4499 	if (ret)
4500 		return -EFAULT;
4501 
4502 	/*
4503 	 * XXX: Do we want to be lenient like existing syscalls; or do we want
4504 	 * to be strict and return an error on out-of-bounds values?
4505 	 */
4506 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4507 
4508 	return 0;
4509 
4510 err_size:
4511 	put_user(sizeof(*attr), &uattr->size);
4512 	return -E2BIG;
4513 }
4514 
4515 /**
4516  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4517  * @pid: the pid in question.
4518  * @policy: new policy.
4519  * @param: structure containing the new RT priority.
4520  *
4521  * Return: 0 on success. An error code otherwise.
4522  */
4523 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4524 {
4525 	if (policy < 0)
4526 		return -EINVAL;
4527 
4528 	return do_sched_setscheduler(pid, policy, param);
4529 }
4530 
4531 /**
4532  * sys_sched_setparam - set/change the RT priority of a thread
4533  * @pid: the pid in question.
4534  * @param: structure containing the new RT priority.
4535  *
4536  * Return: 0 on success. An error code otherwise.
4537  */
4538 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4539 {
4540 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4541 }
4542 
4543 /**
4544  * sys_sched_setattr - same as above, but with extended sched_attr
4545  * @pid: the pid in question.
4546  * @uattr: structure containing the extended parameters.
4547  * @flags: for future extension.
4548  */
4549 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4550 			       unsigned int, flags)
4551 {
4552 	struct sched_attr attr;
4553 	struct task_struct *p;
4554 	int retval;
4555 
4556 	if (!uattr || pid < 0 || flags)
4557 		return -EINVAL;
4558 
4559 	retval = sched_copy_attr(uattr, &attr);
4560 	if (retval)
4561 		return retval;
4562 
4563 	if ((int)attr.sched_policy < 0)
4564 		return -EINVAL;
4565 
4566 	rcu_read_lock();
4567 	retval = -ESRCH;
4568 	p = find_process_by_pid(pid);
4569 	if (p != NULL)
4570 		retval = sched_setattr(p, &attr);
4571 	rcu_read_unlock();
4572 
4573 	return retval;
4574 }
4575 
4576 /**
4577  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4578  * @pid: the pid in question.
4579  *
4580  * Return: On success, the policy of the thread. Otherwise, a negative error
4581  * code.
4582  */
4583 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4584 {
4585 	struct task_struct *p;
4586 	int retval;
4587 
4588 	if (pid < 0)
4589 		return -EINVAL;
4590 
4591 	retval = -ESRCH;
4592 	rcu_read_lock();
4593 	p = find_process_by_pid(pid);
4594 	if (p) {
4595 		retval = security_task_getscheduler(p);
4596 		if (!retval)
4597 			retval = p->policy
4598 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4599 	}
4600 	rcu_read_unlock();
4601 	return retval;
4602 }
4603 
4604 /**
4605  * sys_sched_getparam - get the RT priority of a thread
4606  * @pid: the pid in question.
4607  * @param: structure containing the RT priority.
4608  *
4609  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4610  * code.
4611  */
4612 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4613 {
4614 	struct sched_param lp = { .sched_priority = 0 };
4615 	struct task_struct *p;
4616 	int retval;
4617 
4618 	if (!param || pid < 0)
4619 		return -EINVAL;
4620 
4621 	rcu_read_lock();
4622 	p = find_process_by_pid(pid);
4623 	retval = -ESRCH;
4624 	if (!p)
4625 		goto out_unlock;
4626 
4627 	retval = security_task_getscheduler(p);
4628 	if (retval)
4629 		goto out_unlock;
4630 
4631 	if (task_has_rt_policy(p))
4632 		lp.sched_priority = p->rt_priority;
4633 	rcu_read_unlock();
4634 
4635 	/*
4636 	 * This one might sleep, we cannot do it with a spinlock held ...
4637 	 */
4638 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4639 
4640 	return retval;
4641 
4642 out_unlock:
4643 	rcu_read_unlock();
4644 	return retval;
4645 }
4646 
4647 static int sched_read_attr(struct sched_attr __user *uattr,
4648 			   struct sched_attr *attr,
4649 			   unsigned int usize)
4650 {
4651 	int ret;
4652 
4653 	if (!access_ok(uattr, usize))
4654 		return -EFAULT;
4655 
4656 	/*
4657 	 * If we're handed a smaller struct than we know of,
4658 	 * ensure all the unknown bits are 0 - i.e. old
4659 	 * user-space does not get uncomplete information.
4660 	 */
4661 	if (usize < sizeof(*attr)) {
4662 		unsigned char *addr;
4663 		unsigned char *end;
4664 
4665 		addr = (void *)attr + usize;
4666 		end  = (void *)attr + sizeof(*attr);
4667 
4668 		for (; addr < end; addr++) {
4669 			if (*addr)
4670 				return -EFBIG;
4671 		}
4672 
4673 		attr->size = usize;
4674 	}
4675 
4676 	ret = copy_to_user(uattr, attr, attr->size);
4677 	if (ret)
4678 		return -EFAULT;
4679 
4680 	return 0;
4681 }
4682 
4683 /**
4684  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4685  * @pid: the pid in question.
4686  * @uattr: structure containing the extended parameters.
4687  * @size: sizeof(attr) for fwd/bwd comp.
4688  * @flags: for future extension.
4689  */
4690 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4691 		unsigned int, size, unsigned int, flags)
4692 {
4693 	struct sched_attr attr = {
4694 		.size = sizeof(struct sched_attr),
4695 	};
4696 	struct task_struct *p;
4697 	int retval;
4698 
4699 	if (!uattr || pid < 0 || size > PAGE_SIZE ||
4700 	    size < SCHED_ATTR_SIZE_VER0 || flags)
4701 		return -EINVAL;
4702 
4703 	rcu_read_lock();
4704 	p = find_process_by_pid(pid);
4705 	retval = -ESRCH;
4706 	if (!p)
4707 		goto out_unlock;
4708 
4709 	retval = security_task_getscheduler(p);
4710 	if (retval)
4711 		goto out_unlock;
4712 
4713 	attr.sched_policy = p->policy;
4714 	if (p->sched_reset_on_fork)
4715 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4716 	if (task_has_dl_policy(p))
4717 		__getparam_dl(p, &attr);
4718 	else if (task_has_rt_policy(p))
4719 		attr.sched_priority = p->rt_priority;
4720 	else
4721 		attr.sched_nice = task_nice(p);
4722 
4723 	rcu_read_unlock();
4724 
4725 	retval = sched_read_attr(uattr, &attr, size);
4726 	return retval;
4727 
4728 out_unlock:
4729 	rcu_read_unlock();
4730 	return retval;
4731 }
4732 
4733 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4734 {
4735 	cpumask_var_t cpus_allowed, new_mask;
4736 	struct task_struct *p;
4737 	int retval;
4738 
4739 	rcu_read_lock();
4740 
4741 	p = find_process_by_pid(pid);
4742 	if (!p) {
4743 		rcu_read_unlock();
4744 		return -ESRCH;
4745 	}
4746 
4747 	/* Prevent p going away */
4748 	get_task_struct(p);
4749 	rcu_read_unlock();
4750 
4751 	if (p->flags & PF_NO_SETAFFINITY) {
4752 		retval = -EINVAL;
4753 		goto out_put_task;
4754 	}
4755 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4756 		retval = -ENOMEM;
4757 		goto out_put_task;
4758 	}
4759 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4760 		retval = -ENOMEM;
4761 		goto out_free_cpus_allowed;
4762 	}
4763 	retval = -EPERM;
4764 	if (!check_same_owner(p)) {
4765 		rcu_read_lock();
4766 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4767 			rcu_read_unlock();
4768 			goto out_free_new_mask;
4769 		}
4770 		rcu_read_unlock();
4771 	}
4772 
4773 	retval = security_task_setscheduler(p);
4774 	if (retval)
4775 		goto out_free_new_mask;
4776 
4777 
4778 	cpuset_cpus_allowed(p, cpus_allowed);
4779 	cpumask_and(new_mask, in_mask, cpus_allowed);
4780 
4781 	/*
4782 	 * Since bandwidth control happens on root_domain basis,
4783 	 * if admission test is enabled, we only admit -deadline
4784 	 * tasks allowed to run on all the CPUs in the task's
4785 	 * root_domain.
4786 	 */
4787 #ifdef CONFIG_SMP
4788 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4789 		rcu_read_lock();
4790 		if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4791 			retval = -EBUSY;
4792 			rcu_read_unlock();
4793 			goto out_free_new_mask;
4794 		}
4795 		rcu_read_unlock();
4796 	}
4797 #endif
4798 again:
4799 	retval = __set_cpus_allowed_ptr(p, new_mask, true);
4800 
4801 	if (!retval) {
4802 		cpuset_cpus_allowed(p, cpus_allowed);
4803 		if (!cpumask_subset(new_mask, cpus_allowed)) {
4804 			/*
4805 			 * We must have raced with a concurrent cpuset
4806 			 * update. Just reset the cpus_allowed to the
4807 			 * cpuset's cpus_allowed
4808 			 */
4809 			cpumask_copy(new_mask, cpus_allowed);
4810 			goto again;
4811 		}
4812 	}
4813 out_free_new_mask:
4814 	free_cpumask_var(new_mask);
4815 out_free_cpus_allowed:
4816 	free_cpumask_var(cpus_allowed);
4817 out_put_task:
4818 	put_task_struct(p);
4819 	return retval;
4820 }
4821 
4822 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4823 			     struct cpumask *new_mask)
4824 {
4825 	if (len < cpumask_size())
4826 		cpumask_clear(new_mask);
4827 	else if (len > cpumask_size())
4828 		len = cpumask_size();
4829 
4830 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4831 }
4832 
4833 /**
4834  * sys_sched_setaffinity - set the CPU affinity of a process
4835  * @pid: pid of the process
4836  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4837  * @user_mask_ptr: user-space pointer to the new CPU mask
4838  *
4839  * Return: 0 on success. An error code otherwise.
4840  */
4841 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4842 		unsigned long __user *, user_mask_ptr)
4843 {
4844 	cpumask_var_t new_mask;
4845 	int retval;
4846 
4847 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4848 		return -ENOMEM;
4849 
4850 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4851 	if (retval == 0)
4852 		retval = sched_setaffinity(pid, new_mask);
4853 	free_cpumask_var(new_mask);
4854 	return retval;
4855 }
4856 
4857 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4858 {
4859 	struct task_struct *p;
4860 	unsigned long flags;
4861 	int retval;
4862 
4863 	rcu_read_lock();
4864 
4865 	retval = -ESRCH;
4866 	p = find_process_by_pid(pid);
4867 	if (!p)
4868 		goto out_unlock;
4869 
4870 	retval = security_task_getscheduler(p);
4871 	if (retval)
4872 		goto out_unlock;
4873 
4874 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4875 	cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4876 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4877 
4878 out_unlock:
4879 	rcu_read_unlock();
4880 
4881 	return retval;
4882 }
4883 
4884 /**
4885  * sys_sched_getaffinity - get the CPU affinity of a process
4886  * @pid: pid of the process
4887  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4888  * @user_mask_ptr: user-space pointer to hold the current CPU mask
4889  *
4890  * Return: size of CPU mask copied to user_mask_ptr on success. An
4891  * error code otherwise.
4892  */
4893 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4894 		unsigned long __user *, user_mask_ptr)
4895 {
4896 	int ret;
4897 	cpumask_var_t mask;
4898 
4899 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4900 		return -EINVAL;
4901 	if (len & (sizeof(unsigned long)-1))
4902 		return -EINVAL;
4903 
4904 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4905 		return -ENOMEM;
4906 
4907 	ret = sched_getaffinity(pid, mask);
4908 	if (ret == 0) {
4909 		unsigned int retlen = min(len, cpumask_size());
4910 
4911 		if (copy_to_user(user_mask_ptr, mask, retlen))
4912 			ret = -EFAULT;
4913 		else
4914 			ret = retlen;
4915 	}
4916 	free_cpumask_var(mask);
4917 
4918 	return ret;
4919 }
4920 
4921 /**
4922  * sys_sched_yield - yield the current processor to other threads.
4923  *
4924  * This function yields the current CPU to other tasks. If there are no
4925  * other threads running on this CPU then this function will return.
4926  *
4927  * Return: 0.
4928  */
4929 static void do_sched_yield(void)
4930 {
4931 	struct rq_flags rf;
4932 	struct rq *rq;
4933 
4934 	rq = this_rq_lock_irq(&rf);
4935 
4936 	schedstat_inc(rq->yld_count);
4937 	current->sched_class->yield_task(rq);
4938 
4939 	/*
4940 	 * Since we are going to call schedule() anyway, there's
4941 	 * no need to preempt or enable interrupts:
4942 	 */
4943 	preempt_disable();
4944 	rq_unlock(rq, &rf);
4945 	sched_preempt_enable_no_resched();
4946 
4947 	schedule();
4948 }
4949 
4950 SYSCALL_DEFINE0(sched_yield)
4951 {
4952 	do_sched_yield();
4953 	return 0;
4954 }
4955 
4956 #ifndef CONFIG_PREEMPT
4957 int __sched _cond_resched(void)
4958 {
4959 	if (should_resched(0)) {
4960 		preempt_schedule_common();
4961 		return 1;
4962 	}
4963 	rcu_all_qs();
4964 	return 0;
4965 }
4966 EXPORT_SYMBOL(_cond_resched);
4967 #endif
4968 
4969 /*
4970  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4971  * call schedule, and on return reacquire the lock.
4972  *
4973  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4974  * operations here to prevent schedule() from being called twice (once via
4975  * spin_unlock(), once by hand).
4976  */
4977 int __cond_resched_lock(spinlock_t *lock)
4978 {
4979 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
4980 	int ret = 0;
4981 
4982 	lockdep_assert_held(lock);
4983 
4984 	if (spin_needbreak(lock) || resched) {
4985 		spin_unlock(lock);
4986 		if (resched)
4987 			preempt_schedule_common();
4988 		else
4989 			cpu_relax();
4990 		ret = 1;
4991 		spin_lock(lock);
4992 	}
4993 	return ret;
4994 }
4995 EXPORT_SYMBOL(__cond_resched_lock);
4996 
4997 /**
4998  * yield - yield the current processor to other threads.
4999  *
5000  * Do not ever use this function, there's a 99% chance you're doing it wrong.
5001  *
5002  * The scheduler is at all times free to pick the calling task as the most
5003  * eligible task to run, if removing the yield() call from your code breaks
5004  * it, its already broken.
5005  *
5006  * Typical broken usage is:
5007  *
5008  * while (!event)
5009  *	yield();
5010  *
5011  * where one assumes that yield() will let 'the other' process run that will
5012  * make event true. If the current task is a SCHED_FIFO task that will never
5013  * happen. Never use yield() as a progress guarantee!!
5014  *
5015  * If you want to use yield() to wait for something, use wait_event().
5016  * If you want to use yield() to be 'nice' for others, use cond_resched().
5017  * If you still want to use yield(), do not!
5018  */
5019 void __sched yield(void)
5020 {
5021 	set_current_state(TASK_RUNNING);
5022 	do_sched_yield();
5023 }
5024 EXPORT_SYMBOL(yield);
5025 
5026 /**
5027  * yield_to - yield the current processor to another thread in
5028  * your thread group, or accelerate that thread toward the
5029  * processor it's on.
5030  * @p: target task
5031  * @preempt: whether task preemption is allowed or not
5032  *
5033  * It's the caller's job to ensure that the target task struct
5034  * can't go away on us before we can do any checks.
5035  *
5036  * Return:
5037  *	true (>0) if we indeed boosted the target task.
5038  *	false (0) if we failed to boost the target.
5039  *	-ESRCH if there's no task to yield to.
5040  */
5041 int __sched yield_to(struct task_struct *p, bool preempt)
5042 {
5043 	struct task_struct *curr = current;
5044 	struct rq *rq, *p_rq;
5045 	unsigned long flags;
5046 	int yielded = 0;
5047 
5048 	local_irq_save(flags);
5049 	rq = this_rq();
5050 
5051 again:
5052 	p_rq = task_rq(p);
5053 	/*
5054 	 * If we're the only runnable task on the rq and target rq also
5055 	 * has only one task, there's absolutely no point in yielding.
5056 	 */
5057 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5058 		yielded = -ESRCH;
5059 		goto out_irq;
5060 	}
5061 
5062 	double_rq_lock(rq, p_rq);
5063 	if (task_rq(p) != p_rq) {
5064 		double_rq_unlock(rq, p_rq);
5065 		goto again;
5066 	}
5067 
5068 	if (!curr->sched_class->yield_to_task)
5069 		goto out_unlock;
5070 
5071 	if (curr->sched_class != p->sched_class)
5072 		goto out_unlock;
5073 
5074 	if (task_running(p_rq, p) || p->state)
5075 		goto out_unlock;
5076 
5077 	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5078 	if (yielded) {
5079 		schedstat_inc(rq->yld_count);
5080 		/*
5081 		 * Make p's CPU reschedule; pick_next_entity takes care of
5082 		 * fairness.
5083 		 */
5084 		if (preempt && rq != p_rq)
5085 			resched_curr(p_rq);
5086 	}
5087 
5088 out_unlock:
5089 	double_rq_unlock(rq, p_rq);
5090 out_irq:
5091 	local_irq_restore(flags);
5092 
5093 	if (yielded > 0)
5094 		schedule();
5095 
5096 	return yielded;
5097 }
5098 EXPORT_SYMBOL_GPL(yield_to);
5099 
5100 int io_schedule_prepare(void)
5101 {
5102 	int old_iowait = current->in_iowait;
5103 
5104 	current->in_iowait = 1;
5105 	blk_schedule_flush_plug(current);
5106 
5107 	return old_iowait;
5108 }
5109 
5110 void io_schedule_finish(int token)
5111 {
5112 	current->in_iowait = token;
5113 }
5114 
5115 /*
5116  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5117  * that process accounting knows that this is a task in IO wait state.
5118  */
5119 long __sched io_schedule_timeout(long timeout)
5120 {
5121 	int token;
5122 	long ret;
5123 
5124 	token = io_schedule_prepare();
5125 	ret = schedule_timeout(timeout);
5126 	io_schedule_finish(token);
5127 
5128 	return ret;
5129 }
5130 EXPORT_SYMBOL(io_schedule_timeout);
5131 
5132 void io_schedule(void)
5133 {
5134 	int token;
5135 
5136 	token = io_schedule_prepare();
5137 	schedule();
5138 	io_schedule_finish(token);
5139 }
5140 EXPORT_SYMBOL(io_schedule);
5141 
5142 /**
5143  * sys_sched_get_priority_max - return maximum RT priority.
5144  * @policy: scheduling class.
5145  *
5146  * Return: On success, this syscall returns the maximum
5147  * rt_priority that can be used by a given scheduling class.
5148  * On failure, a negative error code is returned.
5149  */
5150 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5151 {
5152 	int ret = -EINVAL;
5153 
5154 	switch (policy) {
5155 	case SCHED_FIFO:
5156 	case SCHED_RR:
5157 		ret = MAX_USER_RT_PRIO-1;
5158 		break;
5159 	case SCHED_DEADLINE:
5160 	case SCHED_NORMAL:
5161 	case SCHED_BATCH:
5162 	case SCHED_IDLE:
5163 		ret = 0;
5164 		break;
5165 	}
5166 	return ret;
5167 }
5168 
5169 /**
5170  * sys_sched_get_priority_min - return minimum RT priority.
5171  * @policy: scheduling class.
5172  *
5173  * Return: On success, this syscall returns the minimum
5174  * rt_priority that can be used by a given scheduling class.
5175  * On failure, a negative error code is returned.
5176  */
5177 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5178 {
5179 	int ret = -EINVAL;
5180 
5181 	switch (policy) {
5182 	case SCHED_FIFO:
5183 	case SCHED_RR:
5184 		ret = 1;
5185 		break;
5186 	case SCHED_DEADLINE:
5187 	case SCHED_NORMAL:
5188 	case SCHED_BATCH:
5189 	case SCHED_IDLE:
5190 		ret = 0;
5191 	}
5192 	return ret;
5193 }
5194 
5195 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5196 {
5197 	struct task_struct *p;
5198 	unsigned int time_slice;
5199 	struct rq_flags rf;
5200 	struct rq *rq;
5201 	int retval;
5202 
5203 	if (pid < 0)
5204 		return -EINVAL;
5205 
5206 	retval = -ESRCH;
5207 	rcu_read_lock();
5208 	p = find_process_by_pid(pid);
5209 	if (!p)
5210 		goto out_unlock;
5211 
5212 	retval = security_task_getscheduler(p);
5213 	if (retval)
5214 		goto out_unlock;
5215 
5216 	rq = task_rq_lock(p, &rf);
5217 	time_slice = 0;
5218 	if (p->sched_class->get_rr_interval)
5219 		time_slice = p->sched_class->get_rr_interval(rq, p);
5220 	task_rq_unlock(rq, p, &rf);
5221 
5222 	rcu_read_unlock();
5223 	jiffies_to_timespec64(time_slice, t);
5224 	return 0;
5225 
5226 out_unlock:
5227 	rcu_read_unlock();
5228 	return retval;
5229 }
5230 
5231 /**
5232  * sys_sched_rr_get_interval - return the default timeslice of a process.
5233  * @pid: pid of the process.
5234  * @interval: userspace pointer to the timeslice value.
5235  *
5236  * this syscall writes the default timeslice value of a given process
5237  * into the user-space timespec buffer. A value of '0' means infinity.
5238  *
5239  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5240  * an error code.
5241  */
5242 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5243 		struct __kernel_timespec __user *, interval)
5244 {
5245 	struct timespec64 t;
5246 	int retval = sched_rr_get_interval(pid, &t);
5247 
5248 	if (retval == 0)
5249 		retval = put_timespec64(&t, interval);
5250 
5251 	return retval;
5252 }
5253 
5254 #ifdef CONFIG_COMPAT_32BIT_TIME
5255 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5256 		       compat_pid_t, pid,
5257 		       struct old_timespec32 __user *, interval)
5258 {
5259 	struct timespec64 t;
5260 	int retval = sched_rr_get_interval(pid, &t);
5261 
5262 	if (retval == 0)
5263 		retval = put_old_timespec32(&t, interval);
5264 	return retval;
5265 }
5266 #endif
5267 
5268 void sched_show_task(struct task_struct *p)
5269 {
5270 	unsigned long free = 0;
5271 	int ppid;
5272 
5273 	if (!try_get_task_stack(p))
5274 		return;
5275 
5276 	printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5277 
5278 	if (p->state == TASK_RUNNING)
5279 		printk(KERN_CONT "  running task    ");
5280 #ifdef CONFIG_DEBUG_STACK_USAGE
5281 	free = stack_not_used(p);
5282 #endif
5283 	ppid = 0;
5284 	rcu_read_lock();
5285 	if (pid_alive(p))
5286 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
5287 	rcu_read_unlock();
5288 	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5289 		task_pid_nr(p), ppid,
5290 		(unsigned long)task_thread_info(p)->flags);
5291 
5292 	print_worker_info(KERN_INFO, p);
5293 	show_stack(p, NULL);
5294 	put_task_stack(p);
5295 }
5296 EXPORT_SYMBOL_GPL(sched_show_task);
5297 
5298 static inline bool
5299 state_filter_match(unsigned long state_filter, struct task_struct *p)
5300 {
5301 	/* no filter, everything matches */
5302 	if (!state_filter)
5303 		return true;
5304 
5305 	/* filter, but doesn't match */
5306 	if (!(p->state & state_filter))
5307 		return false;
5308 
5309 	/*
5310 	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5311 	 * TASK_KILLABLE).
5312 	 */
5313 	if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5314 		return false;
5315 
5316 	return true;
5317 }
5318 
5319 
5320 void show_state_filter(unsigned long state_filter)
5321 {
5322 	struct task_struct *g, *p;
5323 
5324 #if BITS_PER_LONG == 32
5325 	printk(KERN_INFO
5326 		"  task                PC stack   pid father\n");
5327 #else
5328 	printk(KERN_INFO
5329 		"  task                        PC stack   pid father\n");
5330 #endif
5331 	rcu_read_lock();
5332 	for_each_process_thread(g, p) {
5333 		/*
5334 		 * reset the NMI-timeout, listing all files on a slow
5335 		 * console might take a lot of time:
5336 		 * Also, reset softlockup watchdogs on all CPUs, because
5337 		 * another CPU might be blocked waiting for us to process
5338 		 * an IPI.
5339 		 */
5340 		touch_nmi_watchdog();
5341 		touch_all_softlockup_watchdogs();
5342 		if (state_filter_match(state_filter, p))
5343 			sched_show_task(p);
5344 	}
5345 
5346 #ifdef CONFIG_SCHED_DEBUG
5347 	if (!state_filter)
5348 		sysrq_sched_debug_show();
5349 #endif
5350 	rcu_read_unlock();
5351 	/*
5352 	 * Only show locks if all tasks are dumped:
5353 	 */
5354 	if (!state_filter)
5355 		debug_show_all_locks();
5356 }
5357 
5358 /**
5359  * init_idle - set up an idle thread for a given CPU
5360  * @idle: task in question
5361  * @cpu: CPU the idle task belongs to
5362  *
5363  * NOTE: this function does not set the idle thread's NEED_RESCHED
5364  * flag, to make booting more robust.
5365  */
5366 void init_idle(struct task_struct *idle, int cpu)
5367 {
5368 	struct rq *rq = cpu_rq(cpu);
5369 	unsigned long flags;
5370 
5371 	raw_spin_lock_irqsave(&idle->pi_lock, flags);
5372 	raw_spin_lock(&rq->lock);
5373 
5374 	__sched_fork(0, idle);
5375 	idle->state = TASK_RUNNING;
5376 	idle->se.exec_start = sched_clock();
5377 	idle->flags |= PF_IDLE;
5378 
5379 	kasan_unpoison_task_stack(idle);
5380 
5381 #ifdef CONFIG_SMP
5382 	/*
5383 	 * Its possible that init_idle() gets called multiple times on a task,
5384 	 * in that case do_set_cpus_allowed() will not do the right thing.
5385 	 *
5386 	 * And since this is boot we can forgo the serialization.
5387 	 */
5388 	set_cpus_allowed_common(idle, cpumask_of(cpu));
5389 #endif
5390 	/*
5391 	 * We're having a chicken and egg problem, even though we are
5392 	 * holding rq->lock, the CPU isn't yet set to this CPU so the
5393 	 * lockdep check in task_group() will fail.
5394 	 *
5395 	 * Similar case to sched_fork(). / Alternatively we could
5396 	 * use task_rq_lock() here and obtain the other rq->lock.
5397 	 *
5398 	 * Silence PROVE_RCU
5399 	 */
5400 	rcu_read_lock();
5401 	__set_task_cpu(idle, cpu);
5402 	rcu_read_unlock();
5403 
5404 	rq->curr = rq->idle = idle;
5405 	idle->on_rq = TASK_ON_RQ_QUEUED;
5406 #ifdef CONFIG_SMP
5407 	idle->on_cpu = 1;
5408 #endif
5409 	raw_spin_unlock(&rq->lock);
5410 	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5411 
5412 	/* Set the preempt count _outside_ the spinlocks! */
5413 	init_idle_preempt_count(idle, cpu);
5414 
5415 	/*
5416 	 * The idle tasks have their own, simple scheduling class:
5417 	 */
5418 	idle->sched_class = &idle_sched_class;
5419 	ftrace_graph_init_idle_task(idle, cpu);
5420 	vtime_init_idle(idle, cpu);
5421 #ifdef CONFIG_SMP
5422 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5423 #endif
5424 }
5425 
5426 #ifdef CONFIG_SMP
5427 
5428 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5429 			      const struct cpumask *trial)
5430 {
5431 	int ret = 1;
5432 
5433 	if (!cpumask_weight(cur))
5434 		return ret;
5435 
5436 	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5437 
5438 	return ret;
5439 }
5440 
5441 int task_can_attach(struct task_struct *p,
5442 		    const struct cpumask *cs_cpus_allowed)
5443 {
5444 	int ret = 0;
5445 
5446 	/*
5447 	 * Kthreads which disallow setaffinity shouldn't be moved
5448 	 * to a new cpuset; we don't want to change their CPU
5449 	 * affinity and isolating such threads by their set of
5450 	 * allowed nodes is unnecessary.  Thus, cpusets are not
5451 	 * applicable for such threads.  This prevents checking for
5452 	 * success of set_cpus_allowed_ptr() on all attached tasks
5453 	 * before cpus_allowed may be changed.
5454 	 */
5455 	if (p->flags & PF_NO_SETAFFINITY) {
5456 		ret = -EINVAL;
5457 		goto out;
5458 	}
5459 
5460 	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5461 					      cs_cpus_allowed))
5462 		ret = dl_task_can_attach(p, cs_cpus_allowed);
5463 
5464 out:
5465 	return ret;
5466 }
5467 
5468 bool sched_smp_initialized __read_mostly;
5469 
5470 #ifdef CONFIG_NUMA_BALANCING
5471 /* Migrate current task p to target_cpu */
5472 int migrate_task_to(struct task_struct *p, int target_cpu)
5473 {
5474 	struct migration_arg arg = { p, target_cpu };
5475 	int curr_cpu = task_cpu(p);
5476 
5477 	if (curr_cpu == target_cpu)
5478 		return 0;
5479 
5480 	if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5481 		return -EINVAL;
5482 
5483 	/* TODO: This is not properly updating schedstats */
5484 
5485 	trace_sched_move_numa(p, curr_cpu, target_cpu);
5486 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5487 }
5488 
5489 /*
5490  * Requeue a task on a given node and accurately track the number of NUMA
5491  * tasks on the runqueues
5492  */
5493 void sched_setnuma(struct task_struct *p, int nid)
5494 {
5495 	bool queued, running;
5496 	struct rq_flags rf;
5497 	struct rq *rq;
5498 
5499 	rq = task_rq_lock(p, &rf);
5500 	queued = task_on_rq_queued(p);
5501 	running = task_current(rq, p);
5502 
5503 	if (queued)
5504 		dequeue_task(rq, p, DEQUEUE_SAVE);
5505 	if (running)
5506 		put_prev_task(rq, p);
5507 
5508 	p->numa_preferred_nid = nid;
5509 
5510 	if (queued)
5511 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5512 	if (running)
5513 		set_curr_task(rq, p);
5514 	task_rq_unlock(rq, p, &rf);
5515 }
5516 #endif /* CONFIG_NUMA_BALANCING */
5517 
5518 #ifdef CONFIG_HOTPLUG_CPU
5519 /*
5520  * Ensure that the idle task is using init_mm right before its CPU goes
5521  * offline.
5522  */
5523 void idle_task_exit(void)
5524 {
5525 	struct mm_struct *mm = current->active_mm;
5526 
5527 	BUG_ON(cpu_online(smp_processor_id()));
5528 
5529 	if (mm != &init_mm) {
5530 		switch_mm(mm, &init_mm, current);
5531 		current->active_mm = &init_mm;
5532 		finish_arch_post_lock_switch();
5533 	}
5534 	mmdrop(mm);
5535 }
5536 
5537 /*
5538  * Since this CPU is going 'away' for a while, fold any nr_active delta
5539  * we might have. Assumes we're called after migrate_tasks() so that the
5540  * nr_active count is stable. We need to take the teardown thread which
5541  * is calling this into account, so we hand in adjust = 1 to the load
5542  * calculation.
5543  *
5544  * Also see the comment "Global load-average calculations".
5545  */
5546 static void calc_load_migrate(struct rq *rq)
5547 {
5548 	long delta = calc_load_fold_active(rq, 1);
5549 	if (delta)
5550 		atomic_long_add(delta, &calc_load_tasks);
5551 }
5552 
5553 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5554 {
5555 }
5556 
5557 static const struct sched_class fake_sched_class = {
5558 	.put_prev_task = put_prev_task_fake,
5559 };
5560 
5561 static struct task_struct fake_task = {
5562 	/*
5563 	 * Avoid pull_{rt,dl}_task()
5564 	 */
5565 	.prio = MAX_PRIO + 1,
5566 	.sched_class = &fake_sched_class,
5567 };
5568 
5569 /*
5570  * Migrate all tasks from the rq, sleeping tasks will be migrated by
5571  * try_to_wake_up()->select_task_rq().
5572  *
5573  * Called with rq->lock held even though we'er in stop_machine() and
5574  * there's no concurrency possible, we hold the required locks anyway
5575  * because of lock validation efforts.
5576  */
5577 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5578 {
5579 	struct rq *rq = dead_rq;
5580 	struct task_struct *next, *stop = rq->stop;
5581 	struct rq_flags orf = *rf;
5582 	int dest_cpu;
5583 
5584 	/*
5585 	 * Fudge the rq selection such that the below task selection loop
5586 	 * doesn't get stuck on the currently eligible stop task.
5587 	 *
5588 	 * We're currently inside stop_machine() and the rq is either stuck
5589 	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5590 	 * either way we should never end up calling schedule() until we're
5591 	 * done here.
5592 	 */
5593 	rq->stop = NULL;
5594 
5595 	/*
5596 	 * put_prev_task() and pick_next_task() sched
5597 	 * class method both need to have an up-to-date
5598 	 * value of rq->clock[_task]
5599 	 */
5600 	update_rq_clock(rq);
5601 
5602 	for (;;) {
5603 		/*
5604 		 * There's this thread running, bail when that's the only
5605 		 * remaining thread:
5606 		 */
5607 		if (rq->nr_running == 1)
5608 			break;
5609 
5610 		/*
5611 		 * pick_next_task() assumes pinned rq->lock:
5612 		 */
5613 		next = pick_next_task(rq, &fake_task, rf);
5614 		BUG_ON(!next);
5615 		put_prev_task(rq, next);
5616 
5617 		/*
5618 		 * Rules for changing task_struct::cpus_allowed are holding
5619 		 * both pi_lock and rq->lock, such that holding either
5620 		 * stabilizes the mask.
5621 		 *
5622 		 * Drop rq->lock is not quite as disastrous as it usually is
5623 		 * because !cpu_active at this point, which means load-balance
5624 		 * will not interfere. Also, stop-machine.
5625 		 */
5626 		rq_unlock(rq, rf);
5627 		raw_spin_lock(&next->pi_lock);
5628 		rq_relock(rq, rf);
5629 
5630 		/*
5631 		 * Since we're inside stop-machine, _nothing_ should have
5632 		 * changed the task, WARN if weird stuff happened, because in
5633 		 * that case the above rq->lock drop is a fail too.
5634 		 */
5635 		if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5636 			raw_spin_unlock(&next->pi_lock);
5637 			continue;
5638 		}
5639 
5640 		/* Find suitable destination for @next, with force if needed. */
5641 		dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5642 		rq = __migrate_task(rq, rf, next, dest_cpu);
5643 		if (rq != dead_rq) {
5644 			rq_unlock(rq, rf);
5645 			rq = dead_rq;
5646 			*rf = orf;
5647 			rq_relock(rq, rf);
5648 		}
5649 		raw_spin_unlock(&next->pi_lock);
5650 	}
5651 
5652 	rq->stop = stop;
5653 }
5654 #endif /* CONFIG_HOTPLUG_CPU */
5655 
5656 void set_rq_online(struct rq *rq)
5657 {
5658 	if (!rq->online) {
5659 		const struct sched_class *class;
5660 
5661 		cpumask_set_cpu(rq->cpu, rq->rd->online);
5662 		rq->online = 1;
5663 
5664 		for_each_class(class) {
5665 			if (class->rq_online)
5666 				class->rq_online(rq);
5667 		}
5668 	}
5669 }
5670 
5671 void set_rq_offline(struct rq *rq)
5672 {
5673 	if (rq->online) {
5674 		const struct sched_class *class;
5675 
5676 		for_each_class(class) {
5677 			if (class->rq_offline)
5678 				class->rq_offline(rq);
5679 		}
5680 
5681 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5682 		rq->online = 0;
5683 	}
5684 }
5685 
5686 /*
5687  * used to mark begin/end of suspend/resume:
5688  */
5689 static int num_cpus_frozen;
5690 
5691 /*
5692  * Update cpusets according to cpu_active mask.  If cpusets are
5693  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5694  * around partition_sched_domains().
5695  *
5696  * If we come here as part of a suspend/resume, don't touch cpusets because we
5697  * want to restore it back to its original state upon resume anyway.
5698  */
5699 static void cpuset_cpu_active(void)
5700 {
5701 	if (cpuhp_tasks_frozen) {
5702 		/*
5703 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
5704 		 * resume sequence. As long as this is not the last online
5705 		 * operation in the resume sequence, just build a single sched
5706 		 * domain, ignoring cpusets.
5707 		 */
5708 		partition_sched_domains(1, NULL, NULL);
5709 		if (--num_cpus_frozen)
5710 			return;
5711 		/*
5712 		 * This is the last CPU online operation. So fall through and
5713 		 * restore the original sched domains by considering the
5714 		 * cpuset configurations.
5715 		 */
5716 		cpuset_force_rebuild();
5717 	}
5718 	cpuset_update_active_cpus();
5719 }
5720 
5721 static int cpuset_cpu_inactive(unsigned int cpu)
5722 {
5723 	if (!cpuhp_tasks_frozen) {
5724 		if (dl_cpu_busy(cpu))
5725 			return -EBUSY;
5726 		cpuset_update_active_cpus();
5727 	} else {
5728 		num_cpus_frozen++;
5729 		partition_sched_domains(1, NULL, NULL);
5730 	}
5731 	return 0;
5732 }
5733 
5734 int sched_cpu_activate(unsigned int cpu)
5735 {
5736 	struct rq *rq = cpu_rq(cpu);
5737 	struct rq_flags rf;
5738 
5739 #ifdef CONFIG_SCHED_SMT
5740 	/*
5741 	 * When going up, increment the number of cores with SMT present.
5742 	 */
5743 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5744 		static_branch_inc_cpuslocked(&sched_smt_present);
5745 #endif
5746 	set_cpu_active(cpu, true);
5747 
5748 	if (sched_smp_initialized) {
5749 		sched_domains_numa_masks_set(cpu);
5750 		cpuset_cpu_active();
5751 	}
5752 
5753 	/*
5754 	 * Put the rq online, if not already. This happens:
5755 	 *
5756 	 * 1) In the early boot process, because we build the real domains
5757 	 *    after all CPUs have been brought up.
5758 	 *
5759 	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5760 	 *    domains.
5761 	 */
5762 	rq_lock_irqsave(rq, &rf);
5763 	if (rq->rd) {
5764 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5765 		set_rq_online(rq);
5766 	}
5767 	rq_unlock_irqrestore(rq, &rf);
5768 
5769 	update_max_interval();
5770 
5771 	return 0;
5772 }
5773 
5774 int sched_cpu_deactivate(unsigned int cpu)
5775 {
5776 	int ret;
5777 
5778 	set_cpu_active(cpu, false);
5779 	/*
5780 	 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5781 	 * users of this state to go away such that all new such users will
5782 	 * observe it.
5783 	 *
5784 	 * Do sync before park smpboot threads to take care the rcu boost case.
5785 	 */
5786 	synchronize_rcu();
5787 
5788 #ifdef CONFIG_SCHED_SMT
5789 	/*
5790 	 * When going down, decrement the number of cores with SMT present.
5791 	 */
5792 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5793 		static_branch_dec_cpuslocked(&sched_smt_present);
5794 #endif
5795 
5796 	if (!sched_smp_initialized)
5797 		return 0;
5798 
5799 	ret = cpuset_cpu_inactive(cpu);
5800 	if (ret) {
5801 		set_cpu_active(cpu, true);
5802 		return ret;
5803 	}
5804 	sched_domains_numa_masks_clear(cpu);
5805 	return 0;
5806 }
5807 
5808 static void sched_rq_cpu_starting(unsigned int cpu)
5809 {
5810 	struct rq *rq = cpu_rq(cpu);
5811 
5812 	rq->calc_load_update = calc_load_update;
5813 	update_max_interval();
5814 }
5815 
5816 int sched_cpu_starting(unsigned int cpu)
5817 {
5818 	sched_rq_cpu_starting(cpu);
5819 	sched_tick_start(cpu);
5820 	return 0;
5821 }
5822 
5823 #ifdef CONFIG_HOTPLUG_CPU
5824 int sched_cpu_dying(unsigned int cpu)
5825 {
5826 	struct rq *rq = cpu_rq(cpu);
5827 	struct rq_flags rf;
5828 
5829 	/* Handle pending wakeups and then migrate everything off */
5830 	sched_ttwu_pending();
5831 	sched_tick_stop(cpu);
5832 
5833 	rq_lock_irqsave(rq, &rf);
5834 	if (rq->rd) {
5835 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5836 		set_rq_offline(rq);
5837 	}
5838 	migrate_tasks(rq, &rf);
5839 	BUG_ON(rq->nr_running != 1);
5840 	rq_unlock_irqrestore(rq, &rf);
5841 
5842 	calc_load_migrate(rq);
5843 	update_max_interval();
5844 	nohz_balance_exit_idle(rq);
5845 	hrtick_clear(rq);
5846 	return 0;
5847 }
5848 #endif
5849 
5850 void __init sched_init_smp(void)
5851 {
5852 	sched_init_numa();
5853 
5854 	/*
5855 	 * There's no userspace yet to cause hotplug operations; hence all the
5856 	 * CPU masks are stable and all blatant races in the below code cannot
5857 	 * happen. The hotplug lock is nevertheless taken to satisfy lockdep,
5858 	 * but there won't be any contention on it.
5859 	 */
5860 	cpus_read_lock();
5861 	mutex_lock(&sched_domains_mutex);
5862 	sched_init_domains(cpu_active_mask);
5863 	mutex_unlock(&sched_domains_mutex);
5864 	cpus_read_unlock();
5865 
5866 	/* Move init over to a non-isolated CPU */
5867 	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5868 		BUG();
5869 	sched_init_granularity();
5870 
5871 	init_sched_rt_class();
5872 	init_sched_dl_class();
5873 
5874 	sched_smp_initialized = true;
5875 }
5876 
5877 static int __init migration_init(void)
5878 {
5879 	sched_rq_cpu_starting(smp_processor_id());
5880 	return 0;
5881 }
5882 early_initcall(migration_init);
5883 
5884 #else
5885 void __init sched_init_smp(void)
5886 {
5887 	sched_init_granularity();
5888 }
5889 #endif /* CONFIG_SMP */
5890 
5891 int in_sched_functions(unsigned long addr)
5892 {
5893 	return in_lock_functions(addr) ||
5894 		(addr >= (unsigned long)__sched_text_start
5895 		&& addr < (unsigned long)__sched_text_end);
5896 }
5897 
5898 #ifdef CONFIG_CGROUP_SCHED
5899 /*
5900  * Default task group.
5901  * Every task in system belongs to this group at bootup.
5902  */
5903 struct task_group root_task_group;
5904 LIST_HEAD(task_groups);
5905 
5906 /* Cacheline aligned slab cache for task_group */
5907 static struct kmem_cache *task_group_cache __read_mostly;
5908 #endif
5909 
5910 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5911 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5912 
5913 void __init sched_init(void)
5914 {
5915 	int i, j;
5916 	unsigned long alloc_size = 0, ptr;
5917 
5918 	wait_bit_init();
5919 
5920 #ifdef CONFIG_FAIR_GROUP_SCHED
5921 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5922 #endif
5923 #ifdef CONFIG_RT_GROUP_SCHED
5924 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5925 #endif
5926 	if (alloc_size) {
5927 		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5928 
5929 #ifdef CONFIG_FAIR_GROUP_SCHED
5930 		root_task_group.se = (struct sched_entity **)ptr;
5931 		ptr += nr_cpu_ids * sizeof(void **);
5932 
5933 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5934 		ptr += nr_cpu_ids * sizeof(void **);
5935 
5936 #endif /* CONFIG_FAIR_GROUP_SCHED */
5937 #ifdef CONFIG_RT_GROUP_SCHED
5938 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5939 		ptr += nr_cpu_ids * sizeof(void **);
5940 
5941 		root_task_group.rt_rq = (struct rt_rq **)ptr;
5942 		ptr += nr_cpu_ids * sizeof(void **);
5943 
5944 #endif /* CONFIG_RT_GROUP_SCHED */
5945 	}
5946 #ifdef CONFIG_CPUMASK_OFFSTACK
5947 	for_each_possible_cpu(i) {
5948 		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5949 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5950 		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5951 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5952 	}
5953 #endif /* CONFIG_CPUMASK_OFFSTACK */
5954 
5955 	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5956 	init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5957 
5958 #ifdef CONFIG_SMP
5959 	init_defrootdomain();
5960 #endif
5961 
5962 #ifdef CONFIG_RT_GROUP_SCHED
5963 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
5964 			global_rt_period(), global_rt_runtime());
5965 #endif /* CONFIG_RT_GROUP_SCHED */
5966 
5967 #ifdef CONFIG_CGROUP_SCHED
5968 	task_group_cache = KMEM_CACHE(task_group, 0);
5969 
5970 	list_add(&root_task_group.list, &task_groups);
5971 	INIT_LIST_HEAD(&root_task_group.children);
5972 	INIT_LIST_HEAD(&root_task_group.siblings);
5973 	autogroup_init(&init_task);
5974 #endif /* CONFIG_CGROUP_SCHED */
5975 
5976 	for_each_possible_cpu(i) {
5977 		struct rq *rq;
5978 
5979 		rq = cpu_rq(i);
5980 		raw_spin_lock_init(&rq->lock);
5981 		rq->nr_running = 0;
5982 		rq->calc_load_active = 0;
5983 		rq->calc_load_update = jiffies + LOAD_FREQ;
5984 		init_cfs_rq(&rq->cfs);
5985 		init_rt_rq(&rq->rt);
5986 		init_dl_rq(&rq->dl);
5987 #ifdef CONFIG_FAIR_GROUP_SCHED
5988 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5989 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5990 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5991 		/*
5992 		 * How much CPU bandwidth does root_task_group get?
5993 		 *
5994 		 * In case of task-groups formed thr' the cgroup filesystem, it
5995 		 * gets 100% of the CPU resources in the system. This overall
5996 		 * system CPU resource is divided among the tasks of
5997 		 * root_task_group and its child task-groups in a fair manner,
5998 		 * based on each entity's (task or task-group's) weight
5999 		 * (se->load.weight).
6000 		 *
6001 		 * In other words, if root_task_group has 10 tasks of weight
6002 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6003 		 * then A0's share of the CPU resource is:
6004 		 *
6005 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6006 		 *
6007 		 * We achieve this by letting root_task_group's tasks sit
6008 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6009 		 */
6010 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6011 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6012 #endif /* CONFIG_FAIR_GROUP_SCHED */
6013 
6014 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6015 #ifdef CONFIG_RT_GROUP_SCHED
6016 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6017 #endif
6018 
6019 		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6020 			rq->cpu_load[j] = 0;
6021 
6022 #ifdef CONFIG_SMP
6023 		rq->sd = NULL;
6024 		rq->rd = NULL;
6025 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6026 		rq->balance_callback = NULL;
6027 		rq->active_balance = 0;
6028 		rq->next_balance = jiffies;
6029 		rq->push_cpu = 0;
6030 		rq->cpu = i;
6031 		rq->online = 0;
6032 		rq->idle_stamp = 0;
6033 		rq->avg_idle = 2*sysctl_sched_migration_cost;
6034 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6035 
6036 		INIT_LIST_HEAD(&rq->cfs_tasks);
6037 
6038 		rq_attach_root(rq, &def_root_domain);
6039 #ifdef CONFIG_NO_HZ_COMMON
6040 		rq->last_load_update_tick = jiffies;
6041 		rq->last_blocked_load_update_tick = jiffies;
6042 		atomic_set(&rq->nohz_flags, 0);
6043 #endif
6044 #endif /* CONFIG_SMP */
6045 		hrtick_rq_init(rq);
6046 		atomic_set(&rq->nr_iowait, 0);
6047 	}
6048 
6049 	set_load_weight(&init_task, false);
6050 
6051 	/*
6052 	 * The boot idle thread does lazy MMU switching as well:
6053 	 */
6054 	mmgrab(&init_mm);
6055 	enter_lazy_tlb(&init_mm, current);
6056 
6057 	/*
6058 	 * Make us the idle thread. Technically, schedule() should not be
6059 	 * called from this thread, however somewhere below it might be,
6060 	 * but because we are the idle thread, we just pick up running again
6061 	 * when this runqueue becomes "idle".
6062 	 */
6063 	init_idle(current, smp_processor_id());
6064 
6065 	calc_load_update = jiffies + LOAD_FREQ;
6066 
6067 #ifdef CONFIG_SMP
6068 	idle_thread_set_boot_cpu();
6069 #endif
6070 	init_sched_fair_class();
6071 
6072 	init_schedstats();
6073 
6074 	psi_init();
6075 
6076 	scheduler_running = 1;
6077 }
6078 
6079 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6080 static inline int preempt_count_equals(int preempt_offset)
6081 {
6082 	int nested = preempt_count() + rcu_preempt_depth();
6083 
6084 	return (nested == preempt_offset);
6085 }
6086 
6087 void __might_sleep(const char *file, int line, int preempt_offset)
6088 {
6089 	/*
6090 	 * Blocking primitives will set (and therefore destroy) current->state,
6091 	 * since we will exit with TASK_RUNNING make sure we enter with it,
6092 	 * otherwise we will destroy state.
6093 	 */
6094 	WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6095 			"do not call blocking ops when !TASK_RUNNING; "
6096 			"state=%lx set at [<%p>] %pS\n",
6097 			current->state,
6098 			(void *)current->task_state_change,
6099 			(void *)current->task_state_change);
6100 
6101 	___might_sleep(file, line, preempt_offset);
6102 }
6103 EXPORT_SYMBOL(__might_sleep);
6104 
6105 void ___might_sleep(const char *file, int line, int preempt_offset)
6106 {
6107 	/* Ratelimiting timestamp: */
6108 	static unsigned long prev_jiffy;
6109 
6110 	unsigned long preempt_disable_ip;
6111 
6112 	/* WARN_ON_ONCE() by default, no rate limit required: */
6113 	rcu_sleep_check();
6114 
6115 	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6116 	     !is_idle_task(current)) ||
6117 	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6118 	    oops_in_progress)
6119 		return;
6120 
6121 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6122 		return;
6123 	prev_jiffy = jiffies;
6124 
6125 	/* Save this before calling printk(), since that will clobber it: */
6126 	preempt_disable_ip = get_preempt_disable_ip(current);
6127 
6128 	printk(KERN_ERR
6129 		"BUG: sleeping function called from invalid context at %s:%d\n",
6130 			file, line);
6131 	printk(KERN_ERR
6132 		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6133 			in_atomic(), irqs_disabled(),
6134 			current->pid, current->comm);
6135 
6136 	if (task_stack_end_corrupted(current))
6137 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6138 
6139 	debug_show_held_locks(current);
6140 	if (irqs_disabled())
6141 		print_irqtrace_events(current);
6142 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6143 	    && !preempt_count_equals(preempt_offset)) {
6144 		pr_err("Preemption disabled at:");
6145 		print_ip_sym(preempt_disable_ip);
6146 		pr_cont("\n");
6147 	}
6148 	dump_stack();
6149 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6150 }
6151 EXPORT_SYMBOL(___might_sleep);
6152 #endif
6153 
6154 #ifdef CONFIG_MAGIC_SYSRQ
6155 void normalize_rt_tasks(void)
6156 {
6157 	struct task_struct *g, *p;
6158 	struct sched_attr attr = {
6159 		.sched_policy = SCHED_NORMAL,
6160 	};
6161 
6162 	read_lock(&tasklist_lock);
6163 	for_each_process_thread(g, p) {
6164 		/*
6165 		 * Only normalize user tasks:
6166 		 */
6167 		if (p->flags & PF_KTHREAD)
6168 			continue;
6169 
6170 		p->se.exec_start = 0;
6171 		schedstat_set(p->se.statistics.wait_start,  0);
6172 		schedstat_set(p->se.statistics.sleep_start, 0);
6173 		schedstat_set(p->se.statistics.block_start, 0);
6174 
6175 		if (!dl_task(p) && !rt_task(p)) {
6176 			/*
6177 			 * Renice negative nice level userspace
6178 			 * tasks back to 0:
6179 			 */
6180 			if (task_nice(p) < 0)
6181 				set_user_nice(p, 0);
6182 			continue;
6183 		}
6184 
6185 		__sched_setscheduler(p, &attr, false, false);
6186 	}
6187 	read_unlock(&tasklist_lock);
6188 }
6189 
6190 #endif /* CONFIG_MAGIC_SYSRQ */
6191 
6192 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6193 /*
6194  * These functions are only useful for the IA64 MCA handling, or kdb.
6195  *
6196  * They can only be called when the whole system has been
6197  * stopped - every CPU needs to be quiescent, and no scheduling
6198  * activity can take place. Using them for anything else would
6199  * be a serious bug, and as a result, they aren't even visible
6200  * under any other configuration.
6201  */
6202 
6203 /**
6204  * curr_task - return the current task for a given CPU.
6205  * @cpu: the processor in question.
6206  *
6207  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6208  *
6209  * Return: The current task for @cpu.
6210  */
6211 struct task_struct *curr_task(int cpu)
6212 {
6213 	return cpu_curr(cpu);
6214 }
6215 
6216 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6217 
6218 #ifdef CONFIG_IA64
6219 /**
6220  * set_curr_task - set the current task for a given CPU.
6221  * @cpu: the processor in question.
6222  * @p: the task pointer to set.
6223  *
6224  * Description: This function must only be used when non-maskable interrupts
6225  * are serviced on a separate stack. It allows the architecture to switch the
6226  * notion of the current task on a CPU in a non-blocking manner. This function
6227  * must be called with all CPU's synchronized, and interrupts disabled, the
6228  * and caller must save the original value of the current task (see
6229  * curr_task() above) and restore that value before reenabling interrupts and
6230  * re-starting the system.
6231  *
6232  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6233  */
6234 void ia64_set_curr_task(int cpu, struct task_struct *p)
6235 {
6236 	cpu_curr(cpu) = p;
6237 }
6238 
6239 #endif
6240 
6241 #ifdef CONFIG_CGROUP_SCHED
6242 /* task_group_lock serializes the addition/removal of task groups */
6243 static DEFINE_SPINLOCK(task_group_lock);
6244 
6245 static void sched_free_group(struct task_group *tg)
6246 {
6247 	free_fair_sched_group(tg);
6248 	free_rt_sched_group(tg);
6249 	autogroup_free(tg);
6250 	kmem_cache_free(task_group_cache, tg);
6251 }
6252 
6253 /* allocate runqueue etc for a new task group */
6254 struct task_group *sched_create_group(struct task_group *parent)
6255 {
6256 	struct task_group *tg;
6257 
6258 	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6259 	if (!tg)
6260 		return ERR_PTR(-ENOMEM);
6261 
6262 	if (!alloc_fair_sched_group(tg, parent))
6263 		goto err;
6264 
6265 	if (!alloc_rt_sched_group(tg, parent))
6266 		goto err;
6267 
6268 	return tg;
6269 
6270 err:
6271 	sched_free_group(tg);
6272 	return ERR_PTR(-ENOMEM);
6273 }
6274 
6275 void sched_online_group(struct task_group *tg, struct task_group *parent)
6276 {
6277 	unsigned long flags;
6278 
6279 	spin_lock_irqsave(&task_group_lock, flags);
6280 	list_add_rcu(&tg->list, &task_groups);
6281 
6282 	/* Root should already exist: */
6283 	WARN_ON(!parent);
6284 
6285 	tg->parent = parent;
6286 	INIT_LIST_HEAD(&tg->children);
6287 	list_add_rcu(&tg->siblings, &parent->children);
6288 	spin_unlock_irqrestore(&task_group_lock, flags);
6289 
6290 	online_fair_sched_group(tg);
6291 }
6292 
6293 /* rcu callback to free various structures associated with a task group */
6294 static void sched_free_group_rcu(struct rcu_head *rhp)
6295 {
6296 	/* Now it should be safe to free those cfs_rqs: */
6297 	sched_free_group(container_of(rhp, struct task_group, rcu));
6298 }
6299 
6300 void sched_destroy_group(struct task_group *tg)
6301 {
6302 	/* Wait for possible concurrent references to cfs_rqs complete: */
6303 	call_rcu(&tg->rcu, sched_free_group_rcu);
6304 }
6305 
6306 void sched_offline_group(struct task_group *tg)
6307 {
6308 	unsigned long flags;
6309 
6310 	/* End participation in shares distribution: */
6311 	unregister_fair_sched_group(tg);
6312 
6313 	spin_lock_irqsave(&task_group_lock, flags);
6314 	list_del_rcu(&tg->list);
6315 	list_del_rcu(&tg->siblings);
6316 	spin_unlock_irqrestore(&task_group_lock, flags);
6317 }
6318 
6319 static void sched_change_group(struct task_struct *tsk, int type)
6320 {
6321 	struct task_group *tg;
6322 
6323 	/*
6324 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
6325 	 * which is pointless here. Thus, we pass "true" to task_css_check()
6326 	 * to prevent lockdep warnings.
6327 	 */
6328 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6329 			  struct task_group, css);
6330 	tg = autogroup_task_group(tsk, tg);
6331 	tsk->sched_task_group = tg;
6332 
6333 #ifdef CONFIG_FAIR_GROUP_SCHED
6334 	if (tsk->sched_class->task_change_group)
6335 		tsk->sched_class->task_change_group(tsk, type);
6336 	else
6337 #endif
6338 		set_task_rq(tsk, task_cpu(tsk));
6339 }
6340 
6341 /*
6342  * Change task's runqueue when it moves between groups.
6343  *
6344  * The caller of this function should have put the task in its new group by
6345  * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6346  * its new group.
6347  */
6348 void sched_move_task(struct task_struct *tsk)
6349 {
6350 	int queued, running, queue_flags =
6351 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6352 	struct rq_flags rf;
6353 	struct rq *rq;
6354 
6355 	rq = task_rq_lock(tsk, &rf);
6356 	update_rq_clock(rq);
6357 
6358 	running = task_current(rq, tsk);
6359 	queued = task_on_rq_queued(tsk);
6360 
6361 	if (queued)
6362 		dequeue_task(rq, tsk, queue_flags);
6363 	if (running)
6364 		put_prev_task(rq, tsk);
6365 
6366 	sched_change_group(tsk, TASK_MOVE_GROUP);
6367 
6368 	if (queued)
6369 		enqueue_task(rq, tsk, queue_flags);
6370 	if (running)
6371 		set_curr_task(rq, tsk);
6372 
6373 	task_rq_unlock(rq, tsk, &rf);
6374 }
6375 
6376 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6377 {
6378 	return css ? container_of(css, struct task_group, css) : NULL;
6379 }
6380 
6381 static struct cgroup_subsys_state *
6382 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6383 {
6384 	struct task_group *parent = css_tg(parent_css);
6385 	struct task_group *tg;
6386 
6387 	if (!parent) {
6388 		/* This is early initialization for the top cgroup */
6389 		return &root_task_group.css;
6390 	}
6391 
6392 	tg = sched_create_group(parent);
6393 	if (IS_ERR(tg))
6394 		return ERR_PTR(-ENOMEM);
6395 
6396 	return &tg->css;
6397 }
6398 
6399 /* Expose task group only after completing cgroup initialization */
6400 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6401 {
6402 	struct task_group *tg = css_tg(css);
6403 	struct task_group *parent = css_tg(css->parent);
6404 
6405 	if (parent)
6406 		sched_online_group(tg, parent);
6407 	return 0;
6408 }
6409 
6410 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6411 {
6412 	struct task_group *tg = css_tg(css);
6413 
6414 	sched_offline_group(tg);
6415 }
6416 
6417 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6418 {
6419 	struct task_group *tg = css_tg(css);
6420 
6421 	/*
6422 	 * Relies on the RCU grace period between css_released() and this.
6423 	 */
6424 	sched_free_group(tg);
6425 }
6426 
6427 /*
6428  * This is called before wake_up_new_task(), therefore we really only
6429  * have to set its group bits, all the other stuff does not apply.
6430  */
6431 static void cpu_cgroup_fork(struct task_struct *task)
6432 {
6433 	struct rq_flags rf;
6434 	struct rq *rq;
6435 
6436 	rq = task_rq_lock(task, &rf);
6437 
6438 	update_rq_clock(rq);
6439 	sched_change_group(task, TASK_SET_GROUP);
6440 
6441 	task_rq_unlock(rq, task, &rf);
6442 }
6443 
6444 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6445 {
6446 	struct task_struct *task;
6447 	struct cgroup_subsys_state *css;
6448 	int ret = 0;
6449 
6450 	cgroup_taskset_for_each(task, css, tset) {
6451 #ifdef CONFIG_RT_GROUP_SCHED
6452 		if (!sched_rt_can_attach(css_tg(css), task))
6453 			return -EINVAL;
6454 #else
6455 		/* We don't support RT-tasks being in separate groups */
6456 		if (task->sched_class != &fair_sched_class)
6457 			return -EINVAL;
6458 #endif
6459 		/*
6460 		 * Serialize against wake_up_new_task() such that if its
6461 		 * running, we're sure to observe its full state.
6462 		 */
6463 		raw_spin_lock_irq(&task->pi_lock);
6464 		/*
6465 		 * Avoid calling sched_move_task() before wake_up_new_task()
6466 		 * has happened. This would lead to problems with PELT, due to
6467 		 * move wanting to detach+attach while we're not attached yet.
6468 		 */
6469 		if (task->state == TASK_NEW)
6470 			ret = -EINVAL;
6471 		raw_spin_unlock_irq(&task->pi_lock);
6472 
6473 		if (ret)
6474 			break;
6475 	}
6476 	return ret;
6477 }
6478 
6479 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6480 {
6481 	struct task_struct *task;
6482 	struct cgroup_subsys_state *css;
6483 
6484 	cgroup_taskset_for_each(task, css, tset)
6485 		sched_move_task(task);
6486 }
6487 
6488 #ifdef CONFIG_FAIR_GROUP_SCHED
6489 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6490 				struct cftype *cftype, u64 shareval)
6491 {
6492 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
6493 }
6494 
6495 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6496 			       struct cftype *cft)
6497 {
6498 	struct task_group *tg = css_tg(css);
6499 
6500 	return (u64) scale_load_down(tg->shares);
6501 }
6502 
6503 #ifdef CONFIG_CFS_BANDWIDTH
6504 static DEFINE_MUTEX(cfs_constraints_mutex);
6505 
6506 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6507 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6508 
6509 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6510 
6511 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6512 {
6513 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
6514 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6515 
6516 	if (tg == &root_task_group)
6517 		return -EINVAL;
6518 
6519 	/*
6520 	 * Ensure we have at some amount of bandwidth every period.  This is
6521 	 * to prevent reaching a state of large arrears when throttled via
6522 	 * entity_tick() resulting in prolonged exit starvation.
6523 	 */
6524 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6525 		return -EINVAL;
6526 
6527 	/*
6528 	 * Likewise, bound things on the otherside by preventing insane quota
6529 	 * periods.  This also allows us to normalize in computing quota
6530 	 * feasibility.
6531 	 */
6532 	if (period > max_cfs_quota_period)
6533 		return -EINVAL;
6534 
6535 	/*
6536 	 * Prevent race between setting of cfs_rq->runtime_enabled and
6537 	 * unthrottle_offline_cfs_rqs().
6538 	 */
6539 	get_online_cpus();
6540 	mutex_lock(&cfs_constraints_mutex);
6541 	ret = __cfs_schedulable(tg, period, quota);
6542 	if (ret)
6543 		goto out_unlock;
6544 
6545 	runtime_enabled = quota != RUNTIME_INF;
6546 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6547 	/*
6548 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
6549 	 * before making related changes, and on->off must occur afterwards
6550 	 */
6551 	if (runtime_enabled && !runtime_was_enabled)
6552 		cfs_bandwidth_usage_inc();
6553 	raw_spin_lock_irq(&cfs_b->lock);
6554 	cfs_b->period = ns_to_ktime(period);
6555 	cfs_b->quota = quota;
6556 
6557 	__refill_cfs_bandwidth_runtime(cfs_b);
6558 
6559 	/* Restart the period timer (if active) to handle new period expiry: */
6560 	if (runtime_enabled)
6561 		start_cfs_bandwidth(cfs_b);
6562 
6563 	raw_spin_unlock_irq(&cfs_b->lock);
6564 
6565 	for_each_online_cpu(i) {
6566 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6567 		struct rq *rq = cfs_rq->rq;
6568 		struct rq_flags rf;
6569 
6570 		rq_lock_irq(rq, &rf);
6571 		cfs_rq->runtime_enabled = runtime_enabled;
6572 		cfs_rq->runtime_remaining = 0;
6573 
6574 		if (cfs_rq->throttled)
6575 			unthrottle_cfs_rq(cfs_rq);
6576 		rq_unlock_irq(rq, &rf);
6577 	}
6578 	if (runtime_was_enabled && !runtime_enabled)
6579 		cfs_bandwidth_usage_dec();
6580 out_unlock:
6581 	mutex_unlock(&cfs_constraints_mutex);
6582 	put_online_cpus();
6583 
6584 	return ret;
6585 }
6586 
6587 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6588 {
6589 	u64 quota, period;
6590 
6591 	period = ktime_to_ns(tg->cfs_bandwidth.period);
6592 	if (cfs_quota_us < 0)
6593 		quota = RUNTIME_INF;
6594 	else
6595 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6596 
6597 	return tg_set_cfs_bandwidth(tg, period, quota);
6598 }
6599 
6600 long tg_get_cfs_quota(struct task_group *tg)
6601 {
6602 	u64 quota_us;
6603 
6604 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6605 		return -1;
6606 
6607 	quota_us = tg->cfs_bandwidth.quota;
6608 	do_div(quota_us, NSEC_PER_USEC);
6609 
6610 	return quota_us;
6611 }
6612 
6613 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6614 {
6615 	u64 quota, period;
6616 
6617 	period = (u64)cfs_period_us * NSEC_PER_USEC;
6618 	quota = tg->cfs_bandwidth.quota;
6619 
6620 	return tg_set_cfs_bandwidth(tg, period, quota);
6621 }
6622 
6623 long tg_get_cfs_period(struct task_group *tg)
6624 {
6625 	u64 cfs_period_us;
6626 
6627 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6628 	do_div(cfs_period_us, NSEC_PER_USEC);
6629 
6630 	return cfs_period_us;
6631 }
6632 
6633 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6634 				  struct cftype *cft)
6635 {
6636 	return tg_get_cfs_quota(css_tg(css));
6637 }
6638 
6639 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6640 				   struct cftype *cftype, s64 cfs_quota_us)
6641 {
6642 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6643 }
6644 
6645 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6646 				   struct cftype *cft)
6647 {
6648 	return tg_get_cfs_period(css_tg(css));
6649 }
6650 
6651 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6652 				    struct cftype *cftype, u64 cfs_period_us)
6653 {
6654 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
6655 }
6656 
6657 struct cfs_schedulable_data {
6658 	struct task_group *tg;
6659 	u64 period, quota;
6660 };
6661 
6662 /*
6663  * normalize group quota/period to be quota/max_period
6664  * note: units are usecs
6665  */
6666 static u64 normalize_cfs_quota(struct task_group *tg,
6667 			       struct cfs_schedulable_data *d)
6668 {
6669 	u64 quota, period;
6670 
6671 	if (tg == d->tg) {
6672 		period = d->period;
6673 		quota = d->quota;
6674 	} else {
6675 		period = tg_get_cfs_period(tg);
6676 		quota = tg_get_cfs_quota(tg);
6677 	}
6678 
6679 	/* note: these should typically be equivalent */
6680 	if (quota == RUNTIME_INF || quota == -1)
6681 		return RUNTIME_INF;
6682 
6683 	return to_ratio(period, quota);
6684 }
6685 
6686 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6687 {
6688 	struct cfs_schedulable_data *d = data;
6689 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6690 	s64 quota = 0, parent_quota = -1;
6691 
6692 	if (!tg->parent) {
6693 		quota = RUNTIME_INF;
6694 	} else {
6695 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6696 
6697 		quota = normalize_cfs_quota(tg, d);
6698 		parent_quota = parent_b->hierarchical_quota;
6699 
6700 		/*
6701 		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
6702 		 * always take the min.  On cgroup1, only inherit when no
6703 		 * limit is set:
6704 		 */
6705 		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6706 			quota = min(quota, parent_quota);
6707 		} else {
6708 			if (quota == RUNTIME_INF)
6709 				quota = parent_quota;
6710 			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6711 				return -EINVAL;
6712 		}
6713 	}
6714 	cfs_b->hierarchical_quota = quota;
6715 
6716 	return 0;
6717 }
6718 
6719 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6720 {
6721 	int ret;
6722 	struct cfs_schedulable_data data = {
6723 		.tg = tg,
6724 		.period = period,
6725 		.quota = quota,
6726 	};
6727 
6728 	if (quota != RUNTIME_INF) {
6729 		do_div(data.period, NSEC_PER_USEC);
6730 		do_div(data.quota, NSEC_PER_USEC);
6731 	}
6732 
6733 	rcu_read_lock();
6734 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6735 	rcu_read_unlock();
6736 
6737 	return ret;
6738 }
6739 
6740 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6741 {
6742 	struct task_group *tg = css_tg(seq_css(sf));
6743 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6744 
6745 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6746 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6747 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6748 
6749 	if (schedstat_enabled() && tg != &root_task_group) {
6750 		u64 ws = 0;
6751 		int i;
6752 
6753 		for_each_possible_cpu(i)
6754 			ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6755 
6756 		seq_printf(sf, "wait_sum %llu\n", ws);
6757 	}
6758 
6759 	return 0;
6760 }
6761 #endif /* CONFIG_CFS_BANDWIDTH */
6762 #endif /* CONFIG_FAIR_GROUP_SCHED */
6763 
6764 #ifdef CONFIG_RT_GROUP_SCHED
6765 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6766 				struct cftype *cft, s64 val)
6767 {
6768 	return sched_group_set_rt_runtime(css_tg(css), val);
6769 }
6770 
6771 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6772 			       struct cftype *cft)
6773 {
6774 	return sched_group_rt_runtime(css_tg(css));
6775 }
6776 
6777 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6778 				    struct cftype *cftype, u64 rt_period_us)
6779 {
6780 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
6781 }
6782 
6783 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6784 				   struct cftype *cft)
6785 {
6786 	return sched_group_rt_period(css_tg(css));
6787 }
6788 #endif /* CONFIG_RT_GROUP_SCHED */
6789 
6790 static struct cftype cpu_legacy_files[] = {
6791 #ifdef CONFIG_FAIR_GROUP_SCHED
6792 	{
6793 		.name = "shares",
6794 		.read_u64 = cpu_shares_read_u64,
6795 		.write_u64 = cpu_shares_write_u64,
6796 	},
6797 #endif
6798 #ifdef CONFIG_CFS_BANDWIDTH
6799 	{
6800 		.name = "cfs_quota_us",
6801 		.read_s64 = cpu_cfs_quota_read_s64,
6802 		.write_s64 = cpu_cfs_quota_write_s64,
6803 	},
6804 	{
6805 		.name = "cfs_period_us",
6806 		.read_u64 = cpu_cfs_period_read_u64,
6807 		.write_u64 = cpu_cfs_period_write_u64,
6808 	},
6809 	{
6810 		.name = "stat",
6811 		.seq_show = cpu_cfs_stat_show,
6812 	},
6813 #endif
6814 #ifdef CONFIG_RT_GROUP_SCHED
6815 	{
6816 		.name = "rt_runtime_us",
6817 		.read_s64 = cpu_rt_runtime_read,
6818 		.write_s64 = cpu_rt_runtime_write,
6819 	},
6820 	{
6821 		.name = "rt_period_us",
6822 		.read_u64 = cpu_rt_period_read_uint,
6823 		.write_u64 = cpu_rt_period_write_uint,
6824 	},
6825 #endif
6826 	{ }	/* Terminate */
6827 };
6828 
6829 static int cpu_extra_stat_show(struct seq_file *sf,
6830 			       struct cgroup_subsys_state *css)
6831 {
6832 #ifdef CONFIG_CFS_BANDWIDTH
6833 	{
6834 		struct task_group *tg = css_tg(css);
6835 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6836 		u64 throttled_usec;
6837 
6838 		throttled_usec = cfs_b->throttled_time;
6839 		do_div(throttled_usec, NSEC_PER_USEC);
6840 
6841 		seq_printf(sf, "nr_periods %d\n"
6842 			   "nr_throttled %d\n"
6843 			   "throttled_usec %llu\n",
6844 			   cfs_b->nr_periods, cfs_b->nr_throttled,
6845 			   throttled_usec);
6846 	}
6847 #endif
6848 	return 0;
6849 }
6850 
6851 #ifdef CONFIG_FAIR_GROUP_SCHED
6852 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6853 			       struct cftype *cft)
6854 {
6855 	struct task_group *tg = css_tg(css);
6856 	u64 weight = scale_load_down(tg->shares);
6857 
6858 	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6859 }
6860 
6861 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6862 				struct cftype *cft, u64 weight)
6863 {
6864 	/*
6865 	 * cgroup weight knobs should use the common MIN, DFL and MAX
6866 	 * values which are 1, 100 and 10000 respectively.  While it loses
6867 	 * a bit of range on both ends, it maps pretty well onto the shares
6868 	 * value used by scheduler and the round-trip conversions preserve
6869 	 * the original value over the entire range.
6870 	 */
6871 	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6872 		return -ERANGE;
6873 
6874 	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6875 
6876 	return sched_group_set_shares(css_tg(css), scale_load(weight));
6877 }
6878 
6879 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6880 				    struct cftype *cft)
6881 {
6882 	unsigned long weight = scale_load_down(css_tg(css)->shares);
6883 	int last_delta = INT_MAX;
6884 	int prio, delta;
6885 
6886 	/* find the closest nice value to the current weight */
6887 	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6888 		delta = abs(sched_prio_to_weight[prio] - weight);
6889 		if (delta >= last_delta)
6890 			break;
6891 		last_delta = delta;
6892 	}
6893 
6894 	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6895 }
6896 
6897 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6898 				     struct cftype *cft, s64 nice)
6899 {
6900 	unsigned long weight;
6901 	int idx;
6902 
6903 	if (nice < MIN_NICE || nice > MAX_NICE)
6904 		return -ERANGE;
6905 
6906 	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6907 	idx = array_index_nospec(idx, 40);
6908 	weight = sched_prio_to_weight[idx];
6909 
6910 	return sched_group_set_shares(css_tg(css), scale_load(weight));
6911 }
6912 #endif
6913 
6914 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6915 						  long period, long quota)
6916 {
6917 	if (quota < 0)
6918 		seq_puts(sf, "max");
6919 	else
6920 		seq_printf(sf, "%ld", quota);
6921 
6922 	seq_printf(sf, " %ld\n", period);
6923 }
6924 
6925 /* caller should put the current value in *@periodp before calling */
6926 static int __maybe_unused cpu_period_quota_parse(char *buf,
6927 						 u64 *periodp, u64 *quotap)
6928 {
6929 	char tok[21];	/* U64_MAX */
6930 
6931 	if (!sscanf(buf, "%s %llu", tok, periodp))
6932 		return -EINVAL;
6933 
6934 	*periodp *= NSEC_PER_USEC;
6935 
6936 	if (sscanf(tok, "%llu", quotap))
6937 		*quotap *= NSEC_PER_USEC;
6938 	else if (!strcmp(tok, "max"))
6939 		*quotap = RUNTIME_INF;
6940 	else
6941 		return -EINVAL;
6942 
6943 	return 0;
6944 }
6945 
6946 #ifdef CONFIG_CFS_BANDWIDTH
6947 static int cpu_max_show(struct seq_file *sf, void *v)
6948 {
6949 	struct task_group *tg = css_tg(seq_css(sf));
6950 
6951 	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6952 	return 0;
6953 }
6954 
6955 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6956 			     char *buf, size_t nbytes, loff_t off)
6957 {
6958 	struct task_group *tg = css_tg(of_css(of));
6959 	u64 period = tg_get_cfs_period(tg);
6960 	u64 quota;
6961 	int ret;
6962 
6963 	ret = cpu_period_quota_parse(buf, &period, &quota);
6964 	if (!ret)
6965 		ret = tg_set_cfs_bandwidth(tg, period, quota);
6966 	return ret ?: nbytes;
6967 }
6968 #endif
6969 
6970 static struct cftype cpu_files[] = {
6971 #ifdef CONFIG_FAIR_GROUP_SCHED
6972 	{
6973 		.name = "weight",
6974 		.flags = CFTYPE_NOT_ON_ROOT,
6975 		.read_u64 = cpu_weight_read_u64,
6976 		.write_u64 = cpu_weight_write_u64,
6977 	},
6978 	{
6979 		.name = "weight.nice",
6980 		.flags = CFTYPE_NOT_ON_ROOT,
6981 		.read_s64 = cpu_weight_nice_read_s64,
6982 		.write_s64 = cpu_weight_nice_write_s64,
6983 	},
6984 #endif
6985 #ifdef CONFIG_CFS_BANDWIDTH
6986 	{
6987 		.name = "max",
6988 		.flags = CFTYPE_NOT_ON_ROOT,
6989 		.seq_show = cpu_max_show,
6990 		.write = cpu_max_write,
6991 	},
6992 #endif
6993 	{ }	/* terminate */
6994 };
6995 
6996 struct cgroup_subsys cpu_cgrp_subsys = {
6997 	.css_alloc	= cpu_cgroup_css_alloc,
6998 	.css_online	= cpu_cgroup_css_online,
6999 	.css_released	= cpu_cgroup_css_released,
7000 	.css_free	= cpu_cgroup_css_free,
7001 	.css_extra_stat_show = cpu_extra_stat_show,
7002 	.fork		= cpu_cgroup_fork,
7003 	.can_attach	= cpu_cgroup_can_attach,
7004 	.attach		= cpu_cgroup_attach,
7005 	.legacy_cftypes	= cpu_legacy_files,
7006 	.dfl_cftypes	= cpu_files,
7007 	.early_init	= true,
7008 	.threaded	= true,
7009 };
7010 
7011 #endif	/* CONFIG_CGROUP_SCHED */
7012 
7013 void dump_cpu_task(int cpu)
7014 {
7015 	pr_info("Task dump for CPU %d:\n", cpu);
7016 	sched_show_task(cpu_curr(cpu));
7017 }
7018 
7019 /*
7020  * Nice levels are multiplicative, with a gentle 10% change for every
7021  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7022  * nice 1, it will get ~10% less CPU time than another CPU-bound task
7023  * that remained on nice 0.
7024  *
7025  * The "10% effect" is relative and cumulative: from _any_ nice level,
7026  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7027  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7028  * If a task goes up by ~10% and another task goes down by ~10% then
7029  * the relative distance between them is ~25%.)
7030  */
7031 const int sched_prio_to_weight[40] = {
7032  /* -20 */     88761,     71755,     56483,     46273,     36291,
7033  /* -15 */     29154,     23254,     18705,     14949,     11916,
7034  /* -10 */      9548,      7620,      6100,      4904,      3906,
7035  /*  -5 */      3121,      2501,      1991,      1586,      1277,
7036  /*   0 */      1024,       820,       655,       526,       423,
7037  /*   5 */       335,       272,       215,       172,       137,
7038  /*  10 */       110,        87,        70,        56,        45,
7039  /*  15 */        36,        29,        23,        18,        15,
7040 };
7041 
7042 /*
7043  * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7044  *
7045  * In cases where the weight does not change often, we can use the
7046  * precalculated inverse to speed up arithmetics by turning divisions
7047  * into multiplications:
7048  */
7049 const u32 sched_prio_to_wmult[40] = {
7050  /* -20 */     48388,     59856,     76040,     92818,    118348,
7051  /* -15 */    147320,    184698,    229616,    287308,    360437,
7052  /* -10 */    449829,    563644,    704093,    875809,   1099582,
7053  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
7054  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
7055  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
7056  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
7057  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7058 };
7059 
7060 #undef CREATE_TRACE_POINTS
7061