xref: /openbmc/linux/kernel/sched/core.c (revision e639c869)
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 <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/wait_bit.h>
14 #include <linux/cpuset.h>
15 #include <linux/delayacct.h>
16 #include <linux/init_task.h>
17 #include <linux/context_tracking.h>
18 #include <linux/rcupdate_wait.h>
19 #include <linux/compat.h>
20 
21 #include <linux/blkdev.h>
22 #include <linux/kprobes.h>
23 #include <linux/mmu_context.h>
24 #include <linux/module.h>
25 #include <linux/nmi.h>
26 #include <linux/prefetch.h>
27 #include <linux/profile.h>
28 #include <linux/security.h>
29 #include <linux/syscalls.h>
30 #include <linux/sched/isolation.h>
31 
32 #include <asm/switch_to.h>
33 #include <asm/tlb.h>
34 #ifdef CONFIG_PARAVIRT
35 #include <asm/paravirt.h>
36 #endif
37 
38 #include "sched.h"
39 #include "../workqueue_internal.h"
40 #include "../smpboot.h"
41 
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/sched.h>
44 
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
46 
47 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
48 /*
49  * Debugging: various feature bits
50  *
51  * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52  * sysctl_sched_features, defined in sched.h, to allow constants propagation
53  * at compile time and compiler optimization based on features default.
54  */
55 #define SCHED_FEAT(name, enabled)	\
56 	(1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
58 #include "features.h"
59 	0;
60 #undef SCHED_FEAT
61 #endif
62 
63 /*
64  * Number of tasks to iterate in a single balance run.
65  * Limited because this is done with IRQs disabled.
66  */
67 const_debug unsigned int sysctl_sched_nr_migrate = 32;
68 
69 /*
70  * period over which we average the RT time consumption, measured
71  * in ms.
72  *
73  * default: 1s
74  */
75 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
76 
77 /*
78  * period over which we measure -rt task CPU usage in us.
79  * default: 1s
80  */
81 unsigned int sysctl_sched_rt_period = 1000000;
82 
83 __read_mostly int scheduler_running;
84 
85 /*
86  * part of the period that we allow rt tasks to run in us.
87  * default: 0.95s
88  */
89 int sysctl_sched_rt_runtime = 950000;
90 
91 /*
92  * __task_rq_lock - lock the rq @p resides on.
93  */
94 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
95 	__acquires(rq->lock)
96 {
97 	struct rq *rq;
98 
99 	lockdep_assert_held(&p->pi_lock);
100 
101 	for (;;) {
102 		rq = task_rq(p);
103 		raw_spin_lock(&rq->lock);
104 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
105 			rq_pin_lock(rq, rf);
106 			return rq;
107 		}
108 		raw_spin_unlock(&rq->lock);
109 
110 		while (unlikely(task_on_rq_migrating(p)))
111 			cpu_relax();
112 	}
113 }
114 
115 /*
116  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
117  */
118 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
119 	__acquires(p->pi_lock)
120 	__acquires(rq->lock)
121 {
122 	struct rq *rq;
123 
124 	for (;;) {
125 		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
126 		rq = task_rq(p);
127 		raw_spin_lock(&rq->lock);
128 		/*
129 		 *	move_queued_task()		task_rq_lock()
130 		 *
131 		 *	ACQUIRE (rq->lock)
132 		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
133 		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
134 		 *	[S] ->cpu = new_cpu		[L] task_rq()
135 		 *					[L] ->on_rq
136 		 *	RELEASE (rq->lock)
137 		 *
138 		 * If we observe the old cpu in task_rq_lock, the acquire of
139 		 * the old rq->lock will fully serialize against the stores.
140 		 *
141 		 * If we observe the new CPU in task_rq_lock, the acquire will
142 		 * pair with the WMB to ensure we must then also see migrating.
143 		 */
144 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
145 			rq_pin_lock(rq, rf);
146 			return rq;
147 		}
148 		raw_spin_unlock(&rq->lock);
149 		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
150 
151 		while (unlikely(task_on_rq_migrating(p)))
152 			cpu_relax();
153 	}
154 }
155 
156 /*
157  * RQ-clock updating methods:
158  */
159 
160 static void update_rq_clock_task(struct rq *rq, s64 delta)
161 {
162 /*
163  * In theory, the compile should just see 0 here, and optimize out the call
164  * to sched_rt_avg_update. But I don't trust it...
165  */
166 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
167 	s64 steal = 0, irq_delta = 0;
168 #endif
169 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
170 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
171 
172 	/*
173 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
174 	 * this case when a previous update_rq_clock() happened inside a
175 	 * {soft,}irq region.
176 	 *
177 	 * When this happens, we stop ->clock_task and only update the
178 	 * prev_irq_time stamp to account for the part that fit, so that a next
179 	 * update will consume the rest. This ensures ->clock_task is
180 	 * monotonic.
181 	 *
182 	 * It does however cause some slight miss-attribution of {soft,}irq
183 	 * time, a more accurate solution would be to update the irq_time using
184 	 * the current rq->clock timestamp, except that would require using
185 	 * atomic ops.
186 	 */
187 	if (irq_delta > delta)
188 		irq_delta = delta;
189 
190 	rq->prev_irq_time += irq_delta;
191 	delta -= irq_delta;
192 #endif
193 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
194 	if (static_key_false((&paravirt_steal_rq_enabled))) {
195 		steal = paravirt_steal_clock(cpu_of(rq));
196 		steal -= rq->prev_steal_time_rq;
197 
198 		if (unlikely(steal > delta))
199 			steal = delta;
200 
201 		rq->prev_steal_time_rq += steal;
202 		delta -= steal;
203 	}
204 #endif
205 
206 	rq->clock_task += delta;
207 
208 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
209 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
210 		sched_rt_avg_update(rq, irq_delta + steal);
211 #endif
212 }
213 
214 void update_rq_clock(struct rq *rq)
215 {
216 	s64 delta;
217 
218 	lockdep_assert_held(&rq->lock);
219 
220 	if (rq->clock_update_flags & RQCF_ACT_SKIP)
221 		return;
222 
223 #ifdef CONFIG_SCHED_DEBUG
224 	if (sched_feat(WARN_DOUBLE_CLOCK))
225 		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
226 	rq->clock_update_flags |= RQCF_UPDATED;
227 #endif
228 
229 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
230 	if (delta < 0)
231 		return;
232 	rq->clock += delta;
233 	update_rq_clock_task(rq, delta);
234 }
235 
236 
237 #ifdef CONFIG_SCHED_HRTICK
238 /*
239  * Use HR-timers to deliver accurate preemption points.
240  */
241 
242 static void hrtick_clear(struct rq *rq)
243 {
244 	if (hrtimer_active(&rq->hrtick_timer))
245 		hrtimer_cancel(&rq->hrtick_timer);
246 }
247 
248 /*
249  * High-resolution timer tick.
250  * Runs from hardirq context with interrupts disabled.
251  */
252 static enum hrtimer_restart hrtick(struct hrtimer *timer)
253 {
254 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
255 	struct rq_flags rf;
256 
257 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
258 
259 	rq_lock(rq, &rf);
260 	update_rq_clock(rq);
261 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
262 	rq_unlock(rq, &rf);
263 
264 	return HRTIMER_NORESTART;
265 }
266 
267 #ifdef CONFIG_SMP
268 
269 static void __hrtick_restart(struct rq *rq)
270 {
271 	struct hrtimer *timer = &rq->hrtick_timer;
272 
273 	hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
274 }
275 
276 /*
277  * called from hardirq (IPI) context
278  */
279 static void __hrtick_start(void *arg)
280 {
281 	struct rq *rq = arg;
282 	struct rq_flags rf;
283 
284 	rq_lock(rq, &rf);
285 	__hrtick_restart(rq);
286 	rq->hrtick_csd_pending = 0;
287 	rq_unlock(rq, &rf);
288 }
289 
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 	struct hrtimer *timer = &rq->hrtick_timer;
298 	ktime_t time;
299 	s64 delta;
300 
301 	/*
302 	 * Don't schedule slices shorter than 10000ns, that just
303 	 * doesn't make sense and can cause timer DoS.
304 	 */
305 	delta = max_t(s64, delay, 10000LL);
306 	time = ktime_add_ns(timer->base->get_time(), delta);
307 
308 	hrtimer_set_expires(timer, time);
309 
310 	if (rq == this_rq()) {
311 		__hrtick_restart(rq);
312 	} else if (!rq->hrtick_csd_pending) {
313 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 		rq->hrtick_csd_pending = 1;
315 	}
316 }
317 
318 #else
319 /*
320  * Called to set the hrtick timer state.
321  *
322  * called with rq->lock held and irqs disabled
323  */
324 void hrtick_start(struct rq *rq, u64 delay)
325 {
326 	/*
327 	 * Don't schedule slices shorter than 10000ns, that just
328 	 * doesn't make sense. Rely on vruntime for fairness.
329 	 */
330 	delay = max_t(u64, delay, 10000LL);
331 	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 		      HRTIMER_MODE_REL_PINNED);
333 }
334 #endif /* CONFIG_SMP */
335 
336 static void init_rq_hrtick(struct rq *rq)
337 {
338 #ifdef CONFIG_SMP
339 	rq->hrtick_csd_pending = 0;
340 
341 	rq->hrtick_csd.flags = 0;
342 	rq->hrtick_csd.func = __hrtick_start;
343 	rq->hrtick_csd.info = rq;
344 #endif
345 
346 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 	rq->hrtick_timer.function = hrtick;
348 }
349 #else	/* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq *rq)
351 {
352 }
353 
354 static inline void init_rq_hrtick(struct rq *rq)
355 {
356 }
357 #endif	/* CONFIG_SCHED_HRTICK */
358 
359 /*
360  * cmpxchg based fetch_or, macro so it works for different integer types
361  */
362 #define fetch_or(ptr, mask)						\
363 	({								\
364 		typeof(ptr) _ptr = (ptr);				\
365 		typeof(mask) _mask = (mask);				\
366 		typeof(*_ptr) _old, _val = *_ptr;			\
367 									\
368 		for (;;) {						\
369 			_old = cmpxchg(_ptr, _val, _val | _mask);	\
370 			if (_old == _val)				\
371 				break;					\
372 			_val = _old;					\
373 		}							\
374 	_old;								\
375 })
376 
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378 /*
379  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380  * this avoids any races wrt polling state changes and thereby avoids
381  * spurious IPIs.
382  */
383 static bool set_nr_and_not_polling(struct task_struct *p)
384 {
385 	struct thread_info *ti = task_thread_info(p);
386 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
387 }
388 
389 /*
390  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391  *
392  * If this returns true, then the idle task promises to call
393  * sched_ttwu_pending() and reschedule soon.
394  */
395 static bool set_nr_if_polling(struct task_struct *p)
396 {
397 	struct thread_info *ti = task_thread_info(p);
398 	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
399 
400 	for (;;) {
401 		if (!(val & _TIF_POLLING_NRFLAG))
402 			return false;
403 		if (val & _TIF_NEED_RESCHED)
404 			return true;
405 		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
406 		if (old == val)
407 			break;
408 		val = old;
409 	}
410 	return true;
411 }
412 
413 #else
414 static bool set_nr_and_not_polling(struct task_struct *p)
415 {
416 	set_tsk_need_resched(p);
417 	return true;
418 }
419 
420 #ifdef CONFIG_SMP
421 static bool set_nr_if_polling(struct task_struct *p)
422 {
423 	return false;
424 }
425 #endif
426 #endif
427 
428 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
429 {
430 	struct wake_q_node *node = &task->wake_q;
431 
432 	/*
433 	 * Atomically grab the task, if ->wake_q is !nil already it means
434 	 * its already queued (either by us or someone else) and will get the
435 	 * wakeup due to that.
436 	 *
437 	 * This cmpxchg() implies a full barrier, which pairs with the write
438 	 * barrier implied by the wakeup in wake_up_q().
439 	 */
440 	if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
441 		return;
442 
443 	get_task_struct(task);
444 
445 	/*
446 	 * The head is context local, there can be no concurrency.
447 	 */
448 	*head->lastp = node;
449 	head->lastp = &node->next;
450 }
451 
452 void wake_up_q(struct wake_q_head *head)
453 {
454 	struct wake_q_node *node = head->first;
455 
456 	while (node != WAKE_Q_TAIL) {
457 		struct task_struct *task;
458 
459 		task = container_of(node, struct task_struct, wake_q);
460 		BUG_ON(!task);
461 		/* Task can safely be re-inserted now: */
462 		node = node->next;
463 		task->wake_q.next = NULL;
464 
465 		/*
466 		 * wake_up_process() implies a wmb() to pair with the queueing
467 		 * in wake_q_add() so as not to miss wakeups.
468 		 */
469 		wake_up_process(task);
470 		put_task_struct(task);
471 	}
472 }
473 
474 /*
475  * resched_curr - mark rq's current task 'to be rescheduled now'.
476  *
477  * On UP this means the setting of the need_resched flag, on SMP it
478  * might also involve a cross-CPU call to trigger the scheduler on
479  * the target CPU.
480  */
481 void resched_curr(struct rq *rq)
482 {
483 	struct task_struct *curr = rq->curr;
484 	int cpu;
485 
486 	lockdep_assert_held(&rq->lock);
487 
488 	if (test_tsk_need_resched(curr))
489 		return;
490 
491 	cpu = cpu_of(rq);
492 
493 	if (cpu == smp_processor_id()) {
494 		set_tsk_need_resched(curr);
495 		set_preempt_need_resched();
496 		return;
497 	}
498 
499 	if (set_nr_and_not_polling(curr))
500 		smp_send_reschedule(cpu);
501 	else
502 		trace_sched_wake_idle_without_ipi(cpu);
503 }
504 
505 void resched_cpu(int cpu)
506 {
507 	struct rq *rq = cpu_rq(cpu);
508 	unsigned long flags;
509 
510 	raw_spin_lock_irqsave(&rq->lock, flags);
511 	resched_curr(rq);
512 	raw_spin_unlock_irqrestore(&rq->lock, flags);
513 }
514 
515 #ifdef CONFIG_SMP
516 #ifdef CONFIG_NO_HZ_COMMON
517 /*
518  * In the semi idle case, use the nearest busy CPU for migrating timers
519  * from an idle CPU.  This is good for power-savings.
520  *
521  * We don't do similar optimization for completely idle system, as
522  * selecting an idle CPU will add more delays to the timers than intended
523  * (as that CPU's timer base may not be uptodate wrt jiffies etc).
524  */
525 int get_nohz_timer_target(void)
526 {
527 	int i, cpu = smp_processor_id();
528 	struct sched_domain *sd;
529 
530 	if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
531 		return cpu;
532 
533 	rcu_read_lock();
534 	for_each_domain(cpu, sd) {
535 		for_each_cpu(i, sched_domain_span(sd)) {
536 			if (cpu == i)
537 				continue;
538 
539 			if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
540 				cpu = i;
541 				goto unlock;
542 			}
543 		}
544 	}
545 
546 	if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
547 		cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
548 unlock:
549 	rcu_read_unlock();
550 	return cpu;
551 }
552 
553 /*
554  * When add_timer_on() enqueues a timer into the timer wheel of an
555  * idle CPU then this timer might expire before the next timer event
556  * which is scheduled to wake up that CPU. In case of a completely
557  * idle system the next event might even be infinite time into the
558  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559  * leaves the inner idle loop so the newly added timer is taken into
560  * account when the CPU goes back to idle and evaluates the timer
561  * wheel for the next timer event.
562  */
563 static void wake_up_idle_cpu(int cpu)
564 {
565 	struct rq *rq = cpu_rq(cpu);
566 
567 	if (cpu == smp_processor_id())
568 		return;
569 
570 	if (set_nr_and_not_polling(rq->idle))
571 		smp_send_reschedule(cpu);
572 	else
573 		trace_sched_wake_idle_without_ipi(cpu);
574 }
575 
576 static bool wake_up_full_nohz_cpu(int cpu)
577 {
578 	/*
579 	 * We just need the target to call irq_exit() and re-evaluate
580 	 * the next tick. The nohz full kick at least implies that.
581 	 * If needed we can still optimize that later with an
582 	 * empty IRQ.
583 	 */
584 	if (cpu_is_offline(cpu))
585 		return true;  /* Don't try to wake offline CPUs. */
586 	if (tick_nohz_full_cpu(cpu)) {
587 		if (cpu != smp_processor_id() ||
588 		    tick_nohz_tick_stopped())
589 			tick_nohz_full_kick_cpu(cpu);
590 		return true;
591 	}
592 
593 	return false;
594 }
595 
596 /*
597  * Wake up the specified CPU.  If the CPU is going offline, it is the
598  * caller's responsibility to deal with the lost wakeup, for example,
599  * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
600  */
601 void wake_up_nohz_cpu(int cpu)
602 {
603 	if (!wake_up_full_nohz_cpu(cpu))
604 		wake_up_idle_cpu(cpu);
605 }
606 
607 static inline bool got_nohz_idle_kick(void)
608 {
609 	int cpu = smp_processor_id();
610 
611 	if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
612 		return false;
613 
614 	if (idle_cpu(cpu) && !need_resched())
615 		return true;
616 
617 	/*
618 	 * We can't run Idle Load Balance on this CPU for this time so we
619 	 * cancel it and clear NOHZ_BALANCE_KICK
620 	 */
621 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
622 	return false;
623 }
624 
625 #else /* CONFIG_NO_HZ_COMMON */
626 
627 static inline bool got_nohz_idle_kick(void)
628 {
629 	return false;
630 }
631 
632 #endif /* CONFIG_NO_HZ_COMMON */
633 
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq *rq)
636 {
637 	int fifo_nr_running;
638 
639 	/* Deadline tasks, even if single, need the tick */
640 	if (rq->dl.dl_nr_running)
641 		return false;
642 
643 	/*
644 	 * If there are more than one RR tasks, we need the tick to effect the
645 	 * actual RR behaviour.
646 	 */
647 	if (rq->rt.rr_nr_running) {
648 		if (rq->rt.rr_nr_running == 1)
649 			return true;
650 		else
651 			return false;
652 	}
653 
654 	/*
655 	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 	 * forced preemption between FIFO tasks.
657 	 */
658 	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
659 	if (fifo_nr_running)
660 		return true;
661 
662 	/*
663 	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 	 * if there's more than one we need the tick for involuntary
665 	 * preemption.
666 	 */
667 	if (rq->nr_running > 1)
668 		return false;
669 
670 	return true;
671 }
672 #endif /* CONFIG_NO_HZ_FULL */
673 
674 void sched_avg_update(struct rq *rq)
675 {
676 	s64 period = sched_avg_period();
677 
678 	while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
679 		/*
680 		 * Inline assembly required to prevent the compiler
681 		 * optimising this loop into a divmod call.
682 		 * See __iter_div_u64_rem() for another example of this.
683 		 */
684 		asm("" : "+rm" (rq->age_stamp));
685 		rq->age_stamp += period;
686 		rq->rt_avg /= 2;
687 	}
688 }
689 
690 #endif /* CONFIG_SMP */
691 
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
694 /*
695  * Iterate task_group tree rooted at *from, calling @down when first entering a
696  * node and @up when leaving it for the final time.
697  *
698  * Caller must hold rcu_lock or sufficient equivalent.
699  */
700 int walk_tg_tree_from(struct task_group *from,
701 			     tg_visitor down, tg_visitor up, void *data)
702 {
703 	struct task_group *parent, *child;
704 	int ret;
705 
706 	parent = from;
707 
708 down:
709 	ret = (*down)(parent, data);
710 	if (ret)
711 		goto out;
712 	list_for_each_entry_rcu(child, &parent->children, siblings) {
713 		parent = child;
714 		goto down;
715 
716 up:
717 		continue;
718 	}
719 	ret = (*up)(parent, data);
720 	if (ret || parent == from)
721 		goto out;
722 
723 	child = parent;
724 	parent = parent->parent;
725 	if (parent)
726 		goto up;
727 out:
728 	return ret;
729 }
730 
731 int tg_nop(struct task_group *tg, void *data)
732 {
733 	return 0;
734 }
735 #endif
736 
737 static void set_load_weight(struct task_struct *p, bool update_load)
738 {
739 	int prio = p->static_prio - MAX_RT_PRIO;
740 	struct load_weight *load = &p->se.load;
741 
742 	/*
743 	 * SCHED_IDLE tasks get minimal weight:
744 	 */
745 	if (idle_policy(p->policy)) {
746 		load->weight = scale_load(WEIGHT_IDLEPRIO);
747 		load->inv_weight = WMULT_IDLEPRIO;
748 		return;
749 	}
750 
751 	/*
752 	 * SCHED_OTHER tasks have to update their load when changing their
753 	 * weight
754 	 */
755 	if (update_load && p->sched_class == &fair_sched_class) {
756 		reweight_task(p, prio);
757 	} else {
758 		load->weight = scale_load(sched_prio_to_weight[prio]);
759 		load->inv_weight = sched_prio_to_wmult[prio];
760 	}
761 }
762 
763 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
764 {
765 	if (!(flags & ENQUEUE_NOCLOCK))
766 		update_rq_clock(rq);
767 
768 	if (!(flags & ENQUEUE_RESTORE))
769 		sched_info_queued(rq, p);
770 
771 	p->sched_class->enqueue_task(rq, p, flags);
772 }
773 
774 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
775 {
776 	if (!(flags & DEQUEUE_NOCLOCK))
777 		update_rq_clock(rq);
778 
779 	if (!(flags & DEQUEUE_SAVE))
780 		sched_info_dequeued(rq, p);
781 
782 	p->sched_class->dequeue_task(rq, p, flags);
783 }
784 
785 void activate_task(struct rq *rq, struct task_struct *p, int flags)
786 {
787 	if (task_contributes_to_load(p))
788 		rq->nr_uninterruptible--;
789 
790 	enqueue_task(rq, p, flags);
791 }
792 
793 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
794 {
795 	if (task_contributes_to_load(p))
796 		rq->nr_uninterruptible++;
797 
798 	dequeue_task(rq, p, flags);
799 }
800 
801 /*
802  * __normal_prio - return the priority that is based on the static prio
803  */
804 static inline int __normal_prio(struct task_struct *p)
805 {
806 	return p->static_prio;
807 }
808 
809 /*
810  * Calculate the expected normal priority: i.e. priority
811  * without taking RT-inheritance into account. Might be
812  * boosted by interactivity modifiers. Changes upon fork,
813  * setprio syscalls, and whenever the interactivity
814  * estimator recalculates.
815  */
816 static inline int normal_prio(struct task_struct *p)
817 {
818 	int prio;
819 
820 	if (task_has_dl_policy(p))
821 		prio = MAX_DL_PRIO-1;
822 	else if (task_has_rt_policy(p))
823 		prio = MAX_RT_PRIO-1 - p->rt_priority;
824 	else
825 		prio = __normal_prio(p);
826 	return prio;
827 }
828 
829 /*
830  * Calculate the current priority, i.e. the priority
831  * taken into account by the scheduler. This value might
832  * be boosted by RT tasks, or might be boosted by
833  * interactivity modifiers. Will be RT if the task got
834  * RT-boosted. If not then it returns p->normal_prio.
835  */
836 static int effective_prio(struct task_struct *p)
837 {
838 	p->normal_prio = normal_prio(p);
839 	/*
840 	 * If we are RT tasks or we were boosted to RT priority,
841 	 * keep the priority unchanged. Otherwise, update priority
842 	 * to the normal priority:
843 	 */
844 	if (!rt_prio(p->prio))
845 		return p->normal_prio;
846 	return p->prio;
847 }
848 
849 /**
850  * task_curr - is this task currently executing on a CPU?
851  * @p: the task in question.
852  *
853  * Return: 1 if the task is currently executing. 0 otherwise.
854  */
855 inline int task_curr(const struct task_struct *p)
856 {
857 	return cpu_curr(task_cpu(p)) == p;
858 }
859 
860 /*
861  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
862  * use the balance_callback list if you want balancing.
863  *
864  * this means any call to check_class_changed() must be followed by a call to
865  * balance_callback().
866  */
867 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
868 				       const struct sched_class *prev_class,
869 				       int oldprio)
870 {
871 	if (prev_class != p->sched_class) {
872 		if (prev_class->switched_from)
873 			prev_class->switched_from(rq, p);
874 
875 		p->sched_class->switched_to(rq, p);
876 	} else if (oldprio != p->prio || dl_task(p))
877 		p->sched_class->prio_changed(rq, p, oldprio);
878 }
879 
880 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
881 {
882 	const struct sched_class *class;
883 
884 	if (p->sched_class == rq->curr->sched_class) {
885 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
886 	} else {
887 		for_each_class(class) {
888 			if (class == rq->curr->sched_class)
889 				break;
890 			if (class == p->sched_class) {
891 				resched_curr(rq);
892 				break;
893 			}
894 		}
895 	}
896 
897 	/*
898 	 * A queue event has occurred, and we're going to schedule.  In
899 	 * this case, we can save a useless back to back clock update.
900 	 */
901 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
902 		rq_clock_skip_update(rq, true);
903 }
904 
905 #ifdef CONFIG_SMP
906 /*
907  * This is how migration works:
908  *
909  * 1) we invoke migration_cpu_stop() on the target CPU using
910  *    stop_one_cpu().
911  * 2) stopper starts to run (implicitly forcing the migrated thread
912  *    off the CPU)
913  * 3) it checks whether the migrated task is still in the wrong runqueue.
914  * 4) if it's in the wrong runqueue then the migration thread removes
915  *    it and puts it into the right queue.
916  * 5) stopper completes and stop_one_cpu() returns and the migration
917  *    is done.
918  */
919 
920 /*
921  * move_queued_task - move a queued task to new rq.
922  *
923  * Returns (locked) new rq. Old rq's lock is released.
924  */
925 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
926 				   struct task_struct *p, int new_cpu)
927 {
928 	lockdep_assert_held(&rq->lock);
929 
930 	p->on_rq = TASK_ON_RQ_MIGRATING;
931 	dequeue_task(rq, p, DEQUEUE_NOCLOCK);
932 	set_task_cpu(p, new_cpu);
933 	rq_unlock(rq, rf);
934 
935 	rq = cpu_rq(new_cpu);
936 
937 	rq_lock(rq, rf);
938 	BUG_ON(task_cpu(p) != new_cpu);
939 	enqueue_task(rq, p, 0);
940 	p->on_rq = TASK_ON_RQ_QUEUED;
941 	check_preempt_curr(rq, p, 0);
942 
943 	return rq;
944 }
945 
946 struct migration_arg {
947 	struct task_struct *task;
948 	int dest_cpu;
949 };
950 
951 /*
952  * Move (not current) task off this CPU, onto the destination CPU. We're doing
953  * this because either it can't run here any more (set_cpus_allowed()
954  * away from this CPU, or CPU going down), or because we're
955  * attempting to rebalance this task on exec (sched_exec).
956  *
957  * So we race with normal scheduler movements, but that's OK, as long
958  * as the task is no longer on this CPU.
959  */
960 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
961 				 struct task_struct *p, int dest_cpu)
962 {
963 	if (p->flags & PF_KTHREAD) {
964 		if (unlikely(!cpu_online(dest_cpu)))
965 			return rq;
966 	} else {
967 		if (unlikely(!cpu_active(dest_cpu)))
968 			return rq;
969 	}
970 
971 	/* Affinity changed (again). */
972 	if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
973 		return rq;
974 
975 	update_rq_clock(rq);
976 	rq = move_queued_task(rq, rf, p, dest_cpu);
977 
978 	return rq;
979 }
980 
981 /*
982  * migration_cpu_stop - this will be executed by a highprio stopper thread
983  * and performs thread migration by bumping thread off CPU then
984  * 'pushing' onto another runqueue.
985  */
986 static int migration_cpu_stop(void *data)
987 {
988 	struct migration_arg *arg = data;
989 	struct task_struct *p = arg->task;
990 	struct rq *rq = this_rq();
991 	struct rq_flags rf;
992 
993 	/*
994 	 * The original target CPU might have gone down and we might
995 	 * be on another CPU but it doesn't matter.
996 	 */
997 	local_irq_disable();
998 	/*
999 	 * We need to explicitly wake pending tasks before running
1000 	 * __migrate_task() such that we will not miss enforcing cpus_allowed
1001 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1002 	 */
1003 	sched_ttwu_pending();
1004 
1005 	raw_spin_lock(&p->pi_lock);
1006 	rq_lock(rq, &rf);
1007 	/*
1008 	 * If task_rq(p) != rq, it cannot be migrated here, because we're
1009 	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1010 	 * we're holding p->pi_lock.
1011 	 */
1012 	if (task_rq(p) == rq) {
1013 		if (task_on_rq_queued(p))
1014 			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1015 		else
1016 			p->wake_cpu = arg->dest_cpu;
1017 	}
1018 	rq_unlock(rq, &rf);
1019 	raw_spin_unlock(&p->pi_lock);
1020 
1021 	local_irq_enable();
1022 	return 0;
1023 }
1024 
1025 /*
1026  * sched_class::set_cpus_allowed must do the below, but is not required to
1027  * actually call this function.
1028  */
1029 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1030 {
1031 	cpumask_copy(&p->cpus_allowed, new_mask);
1032 	p->nr_cpus_allowed = cpumask_weight(new_mask);
1033 }
1034 
1035 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1036 {
1037 	struct rq *rq = task_rq(p);
1038 	bool queued, running;
1039 
1040 	lockdep_assert_held(&p->pi_lock);
1041 
1042 	queued = task_on_rq_queued(p);
1043 	running = task_current(rq, p);
1044 
1045 	if (queued) {
1046 		/*
1047 		 * Because __kthread_bind() calls this on blocked tasks without
1048 		 * holding rq->lock.
1049 		 */
1050 		lockdep_assert_held(&rq->lock);
1051 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1052 	}
1053 	if (running)
1054 		put_prev_task(rq, p);
1055 
1056 	p->sched_class->set_cpus_allowed(p, new_mask);
1057 
1058 	if (queued)
1059 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1060 	if (running)
1061 		set_curr_task(rq, p);
1062 }
1063 
1064 /*
1065  * Change a given task's CPU affinity. Migrate the thread to a
1066  * proper CPU and schedule it away if the CPU it's executing on
1067  * is removed from the allowed bitmask.
1068  *
1069  * NOTE: the caller must have a valid reference to the task, the
1070  * task must not exit() & deallocate itself prematurely. The
1071  * call is not atomic; no spinlocks may be held.
1072  */
1073 static int __set_cpus_allowed_ptr(struct task_struct *p,
1074 				  const struct cpumask *new_mask, bool check)
1075 {
1076 	const struct cpumask *cpu_valid_mask = cpu_active_mask;
1077 	unsigned int dest_cpu;
1078 	struct rq_flags rf;
1079 	struct rq *rq;
1080 	int ret = 0;
1081 
1082 	rq = task_rq_lock(p, &rf);
1083 	update_rq_clock(rq);
1084 
1085 	if (p->flags & PF_KTHREAD) {
1086 		/*
1087 		 * Kernel threads are allowed on online && !active CPUs
1088 		 */
1089 		cpu_valid_mask = cpu_online_mask;
1090 	}
1091 
1092 	/*
1093 	 * Must re-check here, to close a race against __kthread_bind(),
1094 	 * sched_setaffinity() is not guaranteed to observe the flag.
1095 	 */
1096 	if (check && (p->flags & PF_NO_SETAFFINITY)) {
1097 		ret = -EINVAL;
1098 		goto out;
1099 	}
1100 
1101 	if (cpumask_equal(&p->cpus_allowed, new_mask))
1102 		goto out;
1103 
1104 	if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1105 		ret = -EINVAL;
1106 		goto out;
1107 	}
1108 
1109 	do_set_cpus_allowed(p, new_mask);
1110 
1111 	if (p->flags & PF_KTHREAD) {
1112 		/*
1113 		 * For kernel threads that do indeed end up on online &&
1114 		 * !active we want to ensure they are strict per-CPU threads.
1115 		 */
1116 		WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1117 			!cpumask_intersects(new_mask, cpu_active_mask) &&
1118 			p->nr_cpus_allowed != 1);
1119 	}
1120 
1121 	/* Can the task run on the task's current CPU? If so, we're done */
1122 	if (cpumask_test_cpu(task_cpu(p), new_mask))
1123 		goto out;
1124 
1125 	dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1126 	if (task_running(rq, p) || p->state == TASK_WAKING) {
1127 		struct migration_arg arg = { p, dest_cpu };
1128 		/* Need help from migration thread: drop lock and wait. */
1129 		task_rq_unlock(rq, p, &rf);
1130 		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1131 		tlb_migrate_finish(p->mm);
1132 		return 0;
1133 	} else if (task_on_rq_queued(p)) {
1134 		/*
1135 		 * OK, since we're going to drop the lock immediately
1136 		 * afterwards anyway.
1137 		 */
1138 		rq = move_queued_task(rq, &rf, p, dest_cpu);
1139 	}
1140 out:
1141 	task_rq_unlock(rq, p, &rf);
1142 
1143 	return ret;
1144 }
1145 
1146 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1147 {
1148 	return __set_cpus_allowed_ptr(p, new_mask, false);
1149 }
1150 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1151 
1152 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1153 {
1154 #ifdef CONFIG_SCHED_DEBUG
1155 	/*
1156 	 * We should never call set_task_cpu() on a blocked task,
1157 	 * ttwu() will sort out the placement.
1158 	 */
1159 	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1160 			!p->on_rq);
1161 
1162 	/*
1163 	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1164 	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1165 	 * time relying on p->on_rq.
1166 	 */
1167 	WARN_ON_ONCE(p->state == TASK_RUNNING &&
1168 		     p->sched_class == &fair_sched_class &&
1169 		     (p->on_rq && !task_on_rq_migrating(p)));
1170 
1171 #ifdef CONFIG_LOCKDEP
1172 	/*
1173 	 * The caller should hold either p->pi_lock or rq->lock, when changing
1174 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1175 	 *
1176 	 * sched_move_task() holds both and thus holding either pins the cgroup,
1177 	 * see task_group().
1178 	 *
1179 	 * Furthermore, all task_rq users should acquire both locks, see
1180 	 * task_rq_lock().
1181 	 */
1182 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1183 				      lockdep_is_held(&task_rq(p)->lock)));
1184 #endif
1185 	/*
1186 	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1187 	 */
1188 	WARN_ON_ONCE(!cpu_online(new_cpu));
1189 #endif
1190 
1191 	trace_sched_migrate_task(p, new_cpu);
1192 
1193 	if (task_cpu(p) != new_cpu) {
1194 		if (p->sched_class->migrate_task_rq)
1195 			p->sched_class->migrate_task_rq(p);
1196 		p->se.nr_migrations++;
1197 		perf_event_task_migrate(p);
1198 	}
1199 
1200 	__set_task_cpu(p, new_cpu);
1201 }
1202 
1203 static void __migrate_swap_task(struct task_struct *p, int cpu)
1204 {
1205 	if (task_on_rq_queued(p)) {
1206 		struct rq *src_rq, *dst_rq;
1207 		struct rq_flags srf, drf;
1208 
1209 		src_rq = task_rq(p);
1210 		dst_rq = cpu_rq(cpu);
1211 
1212 		rq_pin_lock(src_rq, &srf);
1213 		rq_pin_lock(dst_rq, &drf);
1214 
1215 		p->on_rq = TASK_ON_RQ_MIGRATING;
1216 		deactivate_task(src_rq, p, 0);
1217 		set_task_cpu(p, cpu);
1218 		activate_task(dst_rq, p, 0);
1219 		p->on_rq = TASK_ON_RQ_QUEUED;
1220 		check_preempt_curr(dst_rq, p, 0);
1221 
1222 		rq_unpin_lock(dst_rq, &drf);
1223 		rq_unpin_lock(src_rq, &srf);
1224 
1225 	} else {
1226 		/*
1227 		 * Task isn't running anymore; make it appear like we migrated
1228 		 * it before it went to sleep. This means on wakeup we make the
1229 		 * previous CPU our target instead of where it really is.
1230 		 */
1231 		p->wake_cpu = cpu;
1232 	}
1233 }
1234 
1235 struct migration_swap_arg {
1236 	struct task_struct *src_task, *dst_task;
1237 	int src_cpu, dst_cpu;
1238 };
1239 
1240 static int migrate_swap_stop(void *data)
1241 {
1242 	struct migration_swap_arg *arg = data;
1243 	struct rq *src_rq, *dst_rq;
1244 	int ret = -EAGAIN;
1245 
1246 	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1247 		return -EAGAIN;
1248 
1249 	src_rq = cpu_rq(arg->src_cpu);
1250 	dst_rq = cpu_rq(arg->dst_cpu);
1251 
1252 	double_raw_lock(&arg->src_task->pi_lock,
1253 			&arg->dst_task->pi_lock);
1254 	double_rq_lock(src_rq, dst_rq);
1255 
1256 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
1257 		goto unlock;
1258 
1259 	if (task_cpu(arg->src_task) != arg->src_cpu)
1260 		goto unlock;
1261 
1262 	if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1263 		goto unlock;
1264 
1265 	if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1266 		goto unlock;
1267 
1268 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
1269 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
1270 
1271 	ret = 0;
1272 
1273 unlock:
1274 	double_rq_unlock(src_rq, dst_rq);
1275 	raw_spin_unlock(&arg->dst_task->pi_lock);
1276 	raw_spin_unlock(&arg->src_task->pi_lock);
1277 
1278 	return ret;
1279 }
1280 
1281 /*
1282  * Cross migrate two tasks
1283  */
1284 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1285 {
1286 	struct migration_swap_arg arg;
1287 	int ret = -EINVAL;
1288 
1289 	arg = (struct migration_swap_arg){
1290 		.src_task = cur,
1291 		.src_cpu = task_cpu(cur),
1292 		.dst_task = p,
1293 		.dst_cpu = task_cpu(p),
1294 	};
1295 
1296 	if (arg.src_cpu == arg.dst_cpu)
1297 		goto out;
1298 
1299 	/*
1300 	 * These three tests are all lockless; this is OK since all of them
1301 	 * will be re-checked with proper locks held further down the line.
1302 	 */
1303 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1304 		goto out;
1305 
1306 	if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1307 		goto out;
1308 
1309 	if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1310 		goto out;
1311 
1312 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1313 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1314 
1315 out:
1316 	return ret;
1317 }
1318 
1319 /*
1320  * wait_task_inactive - wait for a thread to unschedule.
1321  *
1322  * If @match_state is nonzero, it's the @p->state value just checked and
1323  * not expected to change.  If it changes, i.e. @p might have woken up,
1324  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1325  * we return a positive number (its total switch count).  If a second call
1326  * a short while later returns the same number, the caller can be sure that
1327  * @p has remained unscheduled the whole time.
1328  *
1329  * The caller must ensure that the task *will* unschedule sometime soon,
1330  * else this function might spin for a *long* time. This function can't
1331  * be called with interrupts off, or it may introduce deadlock with
1332  * smp_call_function() if an IPI is sent by the same process we are
1333  * waiting to become inactive.
1334  */
1335 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1336 {
1337 	int running, queued;
1338 	struct rq_flags rf;
1339 	unsigned long ncsw;
1340 	struct rq *rq;
1341 
1342 	for (;;) {
1343 		/*
1344 		 * We do the initial early heuristics without holding
1345 		 * any task-queue locks at all. We'll only try to get
1346 		 * the runqueue lock when things look like they will
1347 		 * work out!
1348 		 */
1349 		rq = task_rq(p);
1350 
1351 		/*
1352 		 * If the task is actively running on another CPU
1353 		 * still, just relax and busy-wait without holding
1354 		 * any locks.
1355 		 *
1356 		 * NOTE! Since we don't hold any locks, it's not
1357 		 * even sure that "rq" stays as the right runqueue!
1358 		 * But we don't care, since "task_running()" will
1359 		 * return false if the runqueue has changed and p
1360 		 * is actually now running somewhere else!
1361 		 */
1362 		while (task_running(rq, p)) {
1363 			if (match_state && unlikely(p->state != match_state))
1364 				return 0;
1365 			cpu_relax();
1366 		}
1367 
1368 		/*
1369 		 * Ok, time to look more closely! We need the rq
1370 		 * lock now, to be *sure*. If we're wrong, we'll
1371 		 * just go back and repeat.
1372 		 */
1373 		rq = task_rq_lock(p, &rf);
1374 		trace_sched_wait_task(p);
1375 		running = task_running(rq, p);
1376 		queued = task_on_rq_queued(p);
1377 		ncsw = 0;
1378 		if (!match_state || p->state == match_state)
1379 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1380 		task_rq_unlock(rq, p, &rf);
1381 
1382 		/*
1383 		 * If it changed from the expected state, bail out now.
1384 		 */
1385 		if (unlikely(!ncsw))
1386 			break;
1387 
1388 		/*
1389 		 * Was it really running after all now that we
1390 		 * checked with the proper locks actually held?
1391 		 *
1392 		 * Oops. Go back and try again..
1393 		 */
1394 		if (unlikely(running)) {
1395 			cpu_relax();
1396 			continue;
1397 		}
1398 
1399 		/*
1400 		 * It's not enough that it's not actively running,
1401 		 * it must be off the runqueue _entirely_, and not
1402 		 * preempted!
1403 		 *
1404 		 * So if it was still runnable (but just not actively
1405 		 * running right now), it's preempted, and we should
1406 		 * yield - it could be a while.
1407 		 */
1408 		if (unlikely(queued)) {
1409 			ktime_t to = NSEC_PER_SEC / HZ;
1410 
1411 			set_current_state(TASK_UNINTERRUPTIBLE);
1412 			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1413 			continue;
1414 		}
1415 
1416 		/*
1417 		 * Ahh, all good. It wasn't running, and it wasn't
1418 		 * runnable, which means that it will never become
1419 		 * running in the future either. We're all done!
1420 		 */
1421 		break;
1422 	}
1423 
1424 	return ncsw;
1425 }
1426 
1427 /***
1428  * kick_process - kick a running thread to enter/exit the kernel
1429  * @p: the to-be-kicked thread
1430  *
1431  * Cause a process which is running on another CPU to enter
1432  * kernel-mode, without any delay. (to get signals handled.)
1433  *
1434  * NOTE: this function doesn't have to take the runqueue lock,
1435  * because all it wants to ensure is that the remote task enters
1436  * the kernel. If the IPI races and the task has been migrated
1437  * to another CPU then no harm is done and the purpose has been
1438  * achieved as well.
1439  */
1440 void kick_process(struct task_struct *p)
1441 {
1442 	int cpu;
1443 
1444 	preempt_disable();
1445 	cpu = task_cpu(p);
1446 	if ((cpu != smp_processor_id()) && task_curr(p))
1447 		smp_send_reschedule(cpu);
1448 	preempt_enable();
1449 }
1450 EXPORT_SYMBOL_GPL(kick_process);
1451 
1452 /*
1453  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1454  *
1455  * A few notes on cpu_active vs cpu_online:
1456  *
1457  *  - cpu_active must be a subset of cpu_online
1458  *
1459  *  - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1460  *    see __set_cpus_allowed_ptr(). At this point the newly online
1461  *    CPU isn't yet part of the sched domains, and balancing will not
1462  *    see it.
1463  *
1464  *  - on CPU-down we clear cpu_active() to mask the sched domains and
1465  *    avoid the load balancer to place new tasks on the to be removed
1466  *    CPU. Existing tasks will remain running there and will be taken
1467  *    off.
1468  *
1469  * This means that fallback selection must not select !active CPUs.
1470  * And can assume that any active CPU must be online. Conversely
1471  * select_task_rq() below may allow selection of !active CPUs in order
1472  * to satisfy the above rules.
1473  */
1474 static int select_fallback_rq(int cpu, struct task_struct *p)
1475 {
1476 	int nid = cpu_to_node(cpu);
1477 	const struct cpumask *nodemask = NULL;
1478 	enum { cpuset, possible, fail } state = cpuset;
1479 	int dest_cpu;
1480 
1481 	/*
1482 	 * If the node that the CPU is on has been offlined, cpu_to_node()
1483 	 * will return -1. There is no CPU on the node, and we should
1484 	 * select the CPU on the other node.
1485 	 */
1486 	if (nid != -1) {
1487 		nodemask = cpumask_of_node(nid);
1488 
1489 		/* Look for allowed, online CPU in same node. */
1490 		for_each_cpu(dest_cpu, nodemask) {
1491 			if (!cpu_active(dest_cpu))
1492 				continue;
1493 			if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1494 				return dest_cpu;
1495 		}
1496 	}
1497 
1498 	for (;;) {
1499 		/* Any allowed, online CPU? */
1500 		for_each_cpu(dest_cpu, &p->cpus_allowed) {
1501 			if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1502 				continue;
1503 			if (!cpu_online(dest_cpu))
1504 				continue;
1505 			goto out;
1506 		}
1507 
1508 		/* No more Mr. Nice Guy. */
1509 		switch (state) {
1510 		case cpuset:
1511 			if (IS_ENABLED(CONFIG_CPUSETS)) {
1512 				cpuset_cpus_allowed_fallback(p);
1513 				state = possible;
1514 				break;
1515 			}
1516 			/* Fall-through */
1517 		case possible:
1518 			do_set_cpus_allowed(p, cpu_possible_mask);
1519 			state = fail;
1520 			break;
1521 
1522 		case fail:
1523 			BUG();
1524 			break;
1525 		}
1526 	}
1527 
1528 out:
1529 	if (state != cpuset) {
1530 		/*
1531 		 * Don't tell them about moving exiting tasks or
1532 		 * kernel threads (both mm NULL), since they never
1533 		 * leave kernel.
1534 		 */
1535 		if (p->mm && printk_ratelimit()) {
1536 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1537 					task_pid_nr(p), p->comm, cpu);
1538 		}
1539 	}
1540 
1541 	return dest_cpu;
1542 }
1543 
1544 /*
1545  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1546  */
1547 static inline
1548 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1549 {
1550 	lockdep_assert_held(&p->pi_lock);
1551 
1552 	if (p->nr_cpus_allowed > 1)
1553 		cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1554 	else
1555 		cpu = cpumask_any(&p->cpus_allowed);
1556 
1557 	/*
1558 	 * In order not to call set_task_cpu() on a blocking task we need
1559 	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1560 	 * CPU.
1561 	 *
1562 	 * Since this is common to all placement strategies, this lives here.
1563 	 *
1564 	 * [ this allows ->select_task() to simply return task_cpu(p) and
1565 	 *   not worry about this generic constraint ]
1566 	 */
1567 	if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1568 		     !cpu_online(cpu)))
1569 		cpu = select_fallback_rq(task_cpu(p), p);
1570 
1571 	return cpu;
1572 }
1573 
1574 static void update_avg(u64 *avg, u64 sample)
1575 {
1576 	s64 diff = sample - *avg;
1577 	*avg += diff >> 3;
1578 }
1579 
1580 void sched_set_stop_task(int cpu, struct task_struct *stop)
1581 {
1582 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1583 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
1584 
1585 	if (stop) {
1586 		/*
1587 		 * Make it appear like a SCHED_FIFO task, its something
1588 		 * userspace knows about and won't get confused about.
1589 		 *
1590 		 * Also, it will make PI more or less work without too
1591 		 * much confusion -- but then, stop work should not
1592 		 * rely on PI working anyway.
1593 		 */
1594 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
1595 
1596 		stop->sched_class = &stop_sched_class;
1597 	}
1598 
1599 	cpu_rq(cpu)->stop = stop;
1600 
1601 	if (old_stop) {
1602 		/*
1603 		 * Reset it back to a normal scheduling class so that
1604 		 * it can die in pieces.
1605 		 */
1606 		old_stop->sched_class = &rt_sched_class;
1607 	}
1608 }
1609 
1610 #else
1611 
1612 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1613 					 const struct cpumask *new_mask, bool check)
1614 {
1615 	return set_cpus_allowed_ptr(p, new_mask);
1616 }
1617 
1618 #endif /* CONFIG_SMP */
1619 
1620 static void
1621 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1622 {
1623 	struct rq *rq;
1624 
1625 	if (!schedstat_enabled())
1626 		return;
1627 
1628 	rq = this_rq();
1629 
1630 #ifdef CONFIG_SMP
1631 	if (cpu == rq->cpu) {
1632 		schedstat_inc(rq->ttwu_local);
1633 		schedstat_inc(p->se.statistics.nr_wakeups_local);
1634 	} else {
1635 		struct sched_domain *sd;
1636 
1637 		schedstat_inc(p->se.statistics.nr_wakeups_remote);
1638 		rcu_read_lock();
1639 		for_each_domain(rq->cpu, sd) {
1640 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1641 				schedstat_inc(sd->ttwu_wake_remote);
1642 				break;
1643 			}
1644 		}
1645 		rcu_read_unlock();
1646 	}
1647 
1648 	if (wake_flags & WF_MIGRATED)
1649 		schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1650 #endif /* CONFIG_SMP */
1651 
1652 	schedstat_inc(rq->ttwu_count);
1653 	schedstat_inc(p->se.statistics.nr_wakeups);
1654 
1655 	if (wake_flags & WF_SYNC)
1656 		schedstat_inc(p->se.statistics.nr_wakeups_sync);
1657 }
1658 
1659 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1660 {
1661 	activate_task(rq, p, en_flags);
1662 	p->on_rq = TASK_ON_RQ_QUEUED;
1663 
1664 	/* If a worker is waking up, notify the workqueue: */
1665 	if (p->flags & PF_WQ_WORKER)
1666 		wq_worker_waking_up(p, cpu_of(rq));
1667 }
1668 
1669 /*
1670  * Mark the task runnable and perform wakeup-preemption.
1671  */
1672 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1673 			   struct rq_flags *rf)
1674 {
1675 	check_preempt_curr(rq, p, wake_flags);
1676 	p->state = TASK_RUNNING;
1677 	trace_sched_wakeup(p);
1678 
1679 #ifdef CONFIG_SMP
1680 	if (p->sched_class->task_woken) {
1681 		/*
1682 		 * Our task @p is fully woken up and running; so its safe to
1683 		 * drop the rq->lock, hereafter rq is only used for statistics.
1684 		 */
1685 		rq_unpin_lock(rq, rf);
1686 		p->sched_class->task_woken(rq, p);
1687 		rq_repin_lock(rq, rf);
1688 	}
1689 
1690 	if (rq->idle_stamp) {
1691 		u64 delta = rq_clock(rq) - rq->idle_stamp;
1692 		u64 max = 2*rq->max_idle_balance_cost;
1693 
1694 		update_avg(&rq->avg_idle, delta);
1695 
1696 		if (rq->avg_idle > max)
1697 			rq->avg_idle = max;
1698 
1699 		rq->idle_stamp = 0;
1700 	}
1701 #endif
1702 }
1703 
1704 static void
1705 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1706 		 struct rq_flags *rf)
1707 {
1708 	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1709 
1710 	lockdep_assert_held(&rq->lock);
1711 
1712 #ifdef CONFIG_SMP
1713 	if (p->sched_contributes_to_load)
1714 		rq->nr_uninterruptible--;
1715 
1716 	if (wake_flags & WF_MIGRATED)
1717 		en_flags |= ENQUEUE_MIGRATED;
1718 #endif
1719 
1720 	ttwu_activate(rq, p, en_flags);
1721 	ttwu_do_wakeup(rq, p, wake_flags, rf);
1722 }
1723 
1724 /*
1725  * Called in case the task @p isn't fully descheduled from its runqueue,
1726  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1727  * since all we need to do is flip p->state to TASK_RUNNING, since
1728  * the task is still ->on_rq.
1729  */
1730 static int ttwu_remote(struct task_struct *p, int wake_flags)
1731 {
1732 	struct rq_flags rf;
1733 	struct rq *rq;
1734 	int ret = 0;
1735 
1736 	rq = __task_rq_lock(p, &rf);
1737 	if (task_on_rq_queued(p)) {
1738 		/* check_preempt_curr() may use rq clock */
1739 		update_rq_clock(rq);
1740 		ttwu_do_wakeup(rq, p, wake_flags, &rf);
1741 		ret = 1;
1742 	}
1743 	__task_rq_unlock(rq, &rf);
1744 
1745 	return ret;
1746 }
1747 
1748 #ifdef CONFIG_SMP
1749 void sched_ttwu_pending(void)
1750 {
1751 	struct rq *rq = this_rq();
1752 	struct llist_node *llist = llist_del_all(&rq->wake_list);
1753 	struct task_struct *p, *t;
1754 	struct rq_flags rf;
1755 
1756 	if (!llist)
1757 		return;
1758 
1759 	rq_lock_irqsave(rq, &rf);
1760 	update_rq_clock(rq);
1761 
1762 	llist_for_each_entry_safe(p, t, llist, wake_entry)
1763 		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1764 
1765 	rq_unlock_irqrestore(rq, &rf);
1766 }
1767 
1768 void scheduler_ipi(void)
1769 {
1770 	/*
1771 	 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1772 	 * TIF_NEED_RESCHED remotely (for the first time) will also send
1773 	 * this IPI.
1774 	 */
1775 	preempt_fold_need_resched();
1776 
1777 	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1778 		return;
1779 
1780 	/*
1781 	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1782 	 * traditionally all their work was done from the interrupt return
1783 	 * path. Now that we actually do some work, we need to make sure
1784 	 * we do call them.
1785 	 *
1786 	 * Some archs already do call them, luckily irq_enter/exit nest
1787 	 * properly.
1788 	 *
1789 	 * Arguably we should visit all archs and update all handlers,
1790 	 * however a fair share of IPIs are still resched only so this would
1791 	 * somewhat pessimize the simple resched case.
1792 	 */
1793 	irq_enter();
1794 	sched_ttwu_pending();
1795 
1796 	/*
1797 	 * Check if someone kicked us for doing the nohz idle load balance.
1798 	 */
1799 	if (unlikely(got_nohz_idle_kick())) {
1800 		this_rq()->idle_balance = 1;
1801 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1802 	}
1803 	irq_exit();
1804 }
1805 
1806 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1807 {
1808 	struct rq *rq = cpu_rq(cpu);
1809 
1810 	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1811 
1812 	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1813 		if (!set_nr_if_polling(rq->idle))
1814 			smp_send_reschedule(cpu);
1815 		else
1816 			trace_sched_wake_idle_without_ipi(cpu);
1817 	}
1818 }
1819 
1820 void wake_up_if_idle(int cpu)
1821 {
1822 	struct rq *rq = cpu_rq(cpu);
1823 	struct rq_flags rf;
1824 
1825 	rcu_read_lock();
1826 
1827 	if (!is_idle_task(rcu_dereference(rq->curr)))
1828 		goto out;
1829 
1830 	if (set_nr_if_polling(rq->idle)) {
1831 		trace_sched_wake_idle_without_ipi(cpu);
1832 	} else {
1833 		rq_lock_irqsave(rq, &rf);
1834 		if (is_idle_task(rq->curr))
1835 			smp_send_reschedule(cpu);
1836 		/* Else CPU is not idle, do nothing here: */
1837 		rq_unlock_irqrestore(rq, &rf);
1838 	}
1839 
1840 out:
1841 	rcu_read_unlock();
1842 }
1843 
1844 bool cpus_share_cache(int this_cpu, int that_cpu)
1845 {
1846 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1847 }
1848 #endif /* CONFIG_SMP */
1849 
1850 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1851 {
1852 	struct rq *rq = cpu_rq(cpu);
1853 	struct rq_flags rf;
1854 
1855 #if defined(CONFIG_SMP)
1856 	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1857 		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1858 		ttwu_queue_remote(p, cpu, wake_flags);
1859 		return;
1860 	}
1861 #endif
1862 
1863 	rq_lock(rq, &rf);
1864 	update_rq_clock(rq);
1865 	ttwu_do_activate(rq, p, wake_flags, &rf);
1866 	rq_unlock(rq, &rf);
1867 }
1868 
1869 /*
1870  * Notes on Program-Order guarantees on SMP systems.
1871  *
1872  *  MIGRATION
1873  *
1874  * The basic program-order guarantee on SMP systems is that when a task [t]
1875  * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1876  * execution on its new CPU [c1].
1877  *
1878  * For migration (of runnable tasks) this is provided by the following means:
1879  *
1880  *  A) UNLOCK of the rq(c0)->lock scheduling out task t
1881  *  B) migration for t is required to synchronize *both* rq(c0)->lock and
1882  *     rq(c1)->lock (if not at the same time, then in that order).
1883  *  C) LOCK of the rq(c1)->lock scheduling in task
1884  *
1885  * Transitivity guarantees that B happens after A and C after B.
1886  * Note: we only require RCpc transitivity.
1887  * Note: the CPU doing B need not be c0 or c1
1888  *
1889  * Example:
1890  *
1891  *   CPU0            CPU1            CPU2
1892  *
1893  *   LOCK rq(0)->lock
1894  *   sched-out X
1895  *   sched-in Y
1896  *   UNLOCK rq(0)->lock
1897  *
1898  *                                   LOCK rq(0)->lock // orders against CPU0
1899  *                                   dequeue X
1900  *                                   UNLOCK rq(0)->lock
1901  *
1902  *                                   LOCK rq(1)->lock
1903  *                                   enqueue X
1904  *                                   UNLOCK rq(1)->lock
1905  *
1906  *                   LOCK rq(1)->lock // orders against CPU2
1907  *                   sched-out Z
1908  *                   sched-in X
1909  *                   UNLOCK rq(1)->lock
1910  *
1911  *
1912  *  BLOCKING -- aka. SLEEP + WAKEUP
1913  *
1914  * For blocking we (obviously) need to provide the same guarantee as for
1915  * migration. However the means are completely different as there is no lock
1916  * chain to provide order. Instead we do:
1917  *
1918  *   1) smp_store_release(X->on_cpu, 0)
1919  *   2) smp_cond_load_acquire(!X->on_cpu)
1920  *
1921  * Example:
1922  *
1923  *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
1924  *
1925  *   LOCK rq(0)->lock LOCK X->pi_lock
1926  *   dequeue X
1927  *   sched-out X
1928  *   smp_store_release(X->on_cpu, 0);
1929  *
1930  *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
1931  *                    X->state = WAKING
1932  *                    set_task_cpu(X,2)
1933  *
1934  *                    LOCK rq(2)->lock
1935  *                    enqueue X
1936  *                    X->state = RUNNING
1937  *                    UNLOCK rq(2)->lock
1938  *
1939  *                                          LOCK rq(2)->lock // orders against CPU1
1940  *                                          sched-out Z
1941  *                                          sched-in X
1942  *                                          UNLOCK rq(2)->lock
1943  *
1944  *                    UNLOCK X->pi_lock
1945  *   UNLOCK rq(0)->lock
1946  *
1947  *
1948  * However; for wakeups there is a second guarantee we must provide, namely we
1949  * must observe the state that lead to our wakeup. That is, not only must our
1950  * task observe its own prior state, it must also observe the stores prior to
1951  * its wakeup.
1952  *
1953  * This means that any means of doing remote wakeups must order the CPU doing
1954  * the wakeup against the CPU the task is going to end up running on. This,
1955  * however, is already required for the regular Program-Order guarantee above,
1956  * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1957  *
1958  */
1959 
1960 /**
1961  * try_to_wake_up - wake up a thread
1962  * @p: the thread to be awakened
1963  * @state: the mask of task states that can be woken
1964  * @wake_flags: wake modifier flags (WF_*)
1965  *
1966  * If (@state & @p->state) @p->state = TASK_RUNNING.
1967  *
1968  * If the task was not queued/runnable, also place it back on a runqueue.
1969  *
1970  * Atomic against schedule() which would dequeue a task, also see
1971  * set_current_state().
1972  *
1973  * Return: %true if @p->state changes (an actual wakeup was done),
1974  *	   %false otherwise.
1975  */
1976 static int
1977 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1978 {
1979 	unsigned long flags;
1980 	int cpu, success = 0;
1981 
1982 	/*
1983 	 * If we are going to wake up a thread waiting for CONDITION we
1984 	 * need to ensure that CONDITION=1 done by the caller can not be
1985 	 * reordered with p->state check below. This pairs with mb() in
1986 	 * set_current_state() the waiting thread does.
1987 	 */
1988 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1989 	smp_mb__after_spinlock();
1990 	if (!(p->state & state))
1991 		goto out;
1992 
1993 	trace_sched_waking(p);
1994 
1995 	/* We're going to change ->state: */
1996 	success = 1;
1997 	cpu = task_cpu(p);
1998 
1999 	/*
2000 	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2001 	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2002 	 * in smp_cond_load_acquire() below.
2003 	 *
2004 	 * sched_ttwu_pending()                 try_to_wake_up()
2005 	 *   [S] p->on_rq = 1;                  [L] P->state
2006 	 *       UNLOCK rq->lock  -----.
2007 	 *                              \
2008 	 *				 +---   RMB
2009 	 * schedule()                   /
2010 	 *       LOCK rq->lock    -----'
2011 	 *       UNLOCK rq->lock
2012 	 *
2013 	 * [task p]
2014 	 *   [S] p->state = UNINTERRUPTIBLE     [L] p->on_rq
2015 	 *
2016 	 * Pairs with the UNLOCK+LOCK on rq->lock from the
2017 	 * last wakeup of our task and the schedule that got our task
2018 	 * current.
2019 	 */
2020 	smp_rmb();
2021 	if (p->on_rq && ttwu_remote(p, wake_flags))
2022 		goto stat;
2023 
2024 #ifdef CONFIG_SMP
2025 	/*
2026 	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2027 	 * possible to, falsely, observe p->on_cpu == 0.
2028 	 *
2029 	 * One must be running (->on_cpu == 1) in order to remove oneself
2030 	 * from the runqueue.
2031 	 *
2032 	 *  [S] ->on_cpu = 1;	[L] ->on_rq
2033 	 *      UNLOCK rq->lock
2034 	 *			RMB
2035 	 *      LOCK   rq->lock
2036 	 *  [S] ->on_rq = 0;    [L] ->on_cpu
2037 	 *
2038 	 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2039 	 * from the consecutive calls to schedule(); the first switching to our
2040 	 * task, the second putting it to sleep.
2041 	 */
2042 	smp_rmb();
2043 
2044 	/*
2045 	 * If the owning (remote) CPU is still in the middle of schedule() with
2046 	 * this task as prev, wait until its done referencing the task.
2047 	 *
2048 	 * Pairs with the smp_store_release() in finish_lock_switch().
2049 	 *
2050 	 * This ensures that tasks getting woken will be fully ordered against
2051 	 * their previous state and preserve Program Order.
2052 	 */
2053 	smp_cond_load_acquire(&p->on_cpu, !VAL);
2054 
2055 	p->sched_contributes_to_load = !!task_contributes_to_load(p);
2056 	p->state = TASK_WAKING;
2057 
2058 	if (p->in_iowait) {
2059 		delayacct_blkio_end();
2060 		atomic_dec(&task_rq(p)->nr_iowait);
2061 	}
2062 
2063 	cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2064 	if (task_cpu(p) != cpu) {
2065 		wake_flags |= WF_MIGRATED;
2066 		set_task_cpu(p, cpu);
2067 	}
2068 
2069 #else /* CONFIG_SMP */
2070 
2071 	if (p->in_iowait) {
2072 		delayacct_blkio_end();
2073 		atomic_dec(&task_rq(p)->nr_iowait);
2074 	}
2075 
2076 #endif /* CONFIG_SMP */
2077 
2078 	ttwu_queue(p, cpu, wake_flags);
2079 stat:
2080 	ttwu_stat(p, cpu, wake_flags);
2081 out:
2082 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2083 
2084 	return success;
2085 }
2086 
2087 /**
2088  * try_to_wake_up_local - try to wake up a local task with rq lock held
2089  * @p: the thread to be awakened
2090  * @rf: request-queue flags for pinning
2091  *
2092  * Put @p on the run-queue if it's not already there. The caller must
2093  * ensure that this_rq() is locked, @p is bound to this_rq() and not
2094  * the current task.
2095  */
2096 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2097 {
2098 	struct rq *rq = task_rq(p);
2099 
2100 	if (WARN_ON_ONCE(rq != this_rq()) ||
2101 	    WARN_ON_ONCE(p == current))
2102 		return;
2103 
2104 	lockdep_assert_held(&rq->lock);
2105 
2106 	if (!raw_spin_trylock(&p->pi_lock)) {
2107 		/*
2108 		 * This is OK, because current is on_cpu, which avoids it being
2109 		 * picked for load-balance and preemption/IRQs are still
2110 		 * disabled avoiding further scheduler activity on it and we've
2111 		 * not yet picked a replacement task.
2112 		 */
2113 		rq_unlock(rq, rf);
2114 		raw_spin_lock(&p->pi_lock);
2115 		rq_relock(rq, rf);
2116 	}
2117 
2118 	if (!(p->state & TASK_NORMAL))
2119 		goto out;
2120 
2121 	trace_sched_waking(p);
2122 
2123 	if (!task_on_rq_queued(p)) {
2124 		if (p->in_iowait) {
2125 			delayacct_blkio_end();
2126 			atomic_dec(&rq->nr_iowait);
2127 		}
2128 		ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2129 	}
2130 
2131 	ttwu_do_wakeup(rq, p, 0, rf);
2132 	ttwu_stat(p, smp_processor_id(), 0);
2133 out:
2134 	raw_spin_unlock(&p->pi_lock);
2135 }
2136 
2137 /**
2138  * wake_up_process - Wake up a specific process
2139  * @p: The process to be woken up.
2140  *
2141  * Attempt to wake up the nominated process and move it to the set of runnable
2142  * processes.
2143  *
2144  * Return: 1 if the process was woken up, 0 if it was already running.
2145  *
2146  * It may be assumed that this function implies a write memory barrier before
2147  * changing the task state if and only if any tasks are woken up.
2148  */
2149 int wake_up_process(struct task_struct *p)
2150 {
2151 	return try_to_wake_up(p, TASK_NORMAL, 0);
2152 }
2153 EXPORT_SYMBOL(wake_up_process);
2154 
2155 int wake_up_state(struct task_struct *p, unsigned int state)
2156 {
2157 	return try_to_wake_up(p, state, 0);
2158 }
2159 
2160 /*
2161  * Perform scheduler related setup for a newly forked process p.
2162  * p is forked by current.
2163  *
2164  * __sched_fork() is basic setup used by init_idle() too:
2165  */
2166 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2167 {
2168 	p->on_rq			= 0;
2169 
2170 	p->se.on_rq			= 0;
2171 	p->se.exec_start		= 0;
2172 	p->se.sum_exec_runtime		= 0;
2173 	p->se.prev_sum_exec_runtime	= 0;
2174 	p->se.nr_migrations		= 0;
2175 	p->se.vruntime			= 0;
2176 	INIT_LIST_HEAD(&p->se.group_node);
2177 
2178 #ifdef CONFIG_FAIR_GROUP_SCHED
2179 	p->se.cfs_rq			= NULL;
2180 #endif
2181 
2182 #ifdef CONFIG_SCHEDSTATS
2183 	/* Even if schedstat is disabled, there should not be garbage */
2184 	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2185 #endif
2186 
2187 	RB_CLEAR_NODE(&p->dl.rb_node);
2188 	init_dl_task_timer(&p->dl);
2189 	init_dl_inactive_task_timer(&p->dl);
2190 	__dl_clear_params(p);
2191 
2192 	INIT_LIST_HEAD(&p->rt.run_list);
2193 	p->rt.timeout		= 0;
2194 	p->rt.time_slice	= sched_rr_timeslice;
2195 	p->rt.on_rq		= 0;
2196 	p->rt.on_list		= 0;
2197 
2198 #ifdef CONFIG_PREEMPT_NOTIFIERS
2199 	INIT_HLIST_HEAD(&p->preempt_notifiers);
2200 #endif
2201 
2202 #ifdef CONFIG_NUMA_BALANCING
2203 	if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2204 		p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2205 		p->mm->numa_scan_seq = 0;
2206 	}
2207 
2208 	if (clone_flags & CLONE_VM)
2209 		p->numa_preferred_nid = current->numa_preferred_nid;
2210 	else
2211 		p->numa_preferred_nid = -1;
2212 
2213 	p->node_stamp = 0ULL;
2214 	p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2215 	p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2216 	p->numa_work.next = &p->numa_work;
2217 	p->numa_faults = NULL;
2218 	p->last_task_numa_placement = 0;
2219 	p->last_sum_exec_runtime = 0;
2220 
2221 	p->numa_group = NULL;
2222 #endif /* CONFIG_NUMA_BALANCING */
2223 }
2224 
2225 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2226 
2227 #ifdef CONFIG_NUMA_BALANCING
2228 
2229 void set_numabalancing_state(bool enabled)
2230 {
2231 	if (enabled)
2232 		static_branch_enable(&sched_numa_balancing);
2233 	else
2234 		static_branch_disable(&sched_numa_balancing);
2235 }
2236 
2237 #ifdef CONFIG_PROC_SYSCTL
2238 int sysctl_numa_balancing(struct ctl_table *table, int write,
2239 			 void __user *buffer, size_t *lenp, loff_t *ppos)
2240 {
2241 	struct ctl_table t;
2242 	int err;
2243 	int state = static_branch_likely(&sched_numa_balancing);
2244 
2245 	if (write && !capable(CAP_SYS_ADMIN))
2246 		return -EPERM;
2247 
2248 	t = *table;
2249 	t.data = &state;
2250 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2251 	if (err < 0)
2252 		return err;
2253 	if (write)
2254 		set_numabalancing_state(state);
2255 	return err;
2256 }
2257 #endif
2258 #endif
2259 
2260 #ifdef CONFIG_SCHEDSTATS
2261 
2262 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2263 static bool __initdata __sched_schedstats = false;
2264 
2265 static void set_schedstats(bool enabled)
2266 {
2267 	if (enabled)
2268 		static_branch_enable(&sched_schedstats);
2269 	else
2270 		static_branch_disable(&sched_schedstats);
2271 }
2272 
2273 void force_schedstat_enabled(void)
2274 {
2275 	if (!schedstat_enabled()) {
2276 		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2277 		static_branch_enable(&sched_schedstats);
2278 	}
2279 }
2280 
2281 static int __init setup_schedstats(char *str)
2282 {
2283 	int ret = 0;
2284 	if (!str)
2285 		goto out;
2286 
2287 	/*
2288 	 * This code is called before jump labels have been set up, so we can't
2289 	 * change the static branch directly just yet.  Instead set a temporary
2290 	 * variable so init_schedstats() can do it later.
2291 	 */
2292 	if (!strcmp(str, "enable")) {
2293 		__sched_schedstats = true;
2294 		ret = 1;
2295 	} else if (!strcmp(str, "disable")) {
2296 		__sched_schedstats = false;
2297 		ret = 1;
2298 	}
2299 out:
2300 	if (!ret)
2301 		pr_warn("Unable to parse schedstats=\n");
2302 
2303 	return ret;
2304 }
2305 __setup("schedstats=", setup_schedstats);
2306 
2307 static void __init init_schedstats(void)
2308 {
2309 	set_schedstats(__sched_schedstats);
2310 }
2311 
2312 #ifdef CONFIG_PROC_SYSCTL
2313 int sysctl_schedstats(struct ctl_table *table, int write,
2314 			 void __user *buffer, size_t *lenp, loff_t *ppos)
2315 {
2316 	struct ctl_table t;
2317 	int err;
2318 	int state = static_branch_likely(&sched_schedstats);
2319 
2320 	if (write && !capable(CAP_SYS_ADMIN))
2321 		return -EPERM;
2322 
2323 	t = *table;
2324 	t.data = &state;
2325 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2326 	if (err < 0)
2327 		return err;
2328 	if (write)
2329 		set_schedstats(state);
2330 	return err;
2331 }
2332 #endif /* CONFIG_PROC_SYSCTL */
2333 #else  /* !CONFIG_SCHEDSTATS */
2334 static inline void init_schedstats(void) {}
2335 #endif /* CONFIG_SCHEDSTATS */
2336 
2337 /*
2338  * fork()/clone()-time setup:
2339  */
2340 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2341 {
2342 	unsigned long flags;
2343 	int cpu = get_cpu();
2344 
2345 	__sched_fork(clone_flags, p);
2346 	/*
2347 	 * We mark the process as NEW here. This guarantees that
2348 	 * nobody will actually run it, and a signal or other external
2349 	 * event cannot wake it up and insert it on the runqueue either.
2350 	 */
2351 	p->state = TASK_NEW;
2352 
2353 	/*
2354 	 * Make sure we do not leak PI boosting priority to the child.
2355 	 */
2356 	p->prio = current->normal_prio;
2357 
2358 	/*
2359 	 * Revert to default priority/policy on fork if requested.
2360 	 */
2361 	if (unlikely(p->sched_reset_on_fork)) {
2362 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2363 			p->policy = SCHED_NORMAL;
2364 			p->static_prio = NICE_TO_PRIO(0);
2365 			p->rt_priority = 0;
2366 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
2367 			p->static_prio = NICE_TO_PRIO(0);
2368 
2369 		p->prio = p->normal_prio = __normal_prio(p);
2370 		set_load_weight(p, false);
2371 
2372 		/*
2373 		 * We don't need the reset flag anymore after the fork. It has
2374 		 * fulfilled its duty:
2375 		 */
2376 		p->sched_reset_on_fork = 0;
2377 	}
2378 
2379 	if (dl_prio(p->prio)) {
2380 		put_cpu();
2381 		return -EAGAIN;
2382 	} else if (rt_prio(p->prio)) {
2383 		p->sched_class = &rt_sched_class;
2384 	} else {
2385 		p->sched_class = &fair_sched_class;
2386 	}
2387 
2388 	init_entity_runnable_average(&p->se);
2389 
2390 	/*
2391 	 * The child is not yet in the pid-hash so no cgroup attach races,
2392 	 * and the cgroup is pinned to this child due to cgroup_fork()
2393 	 * is ran before sched_fork().
2394 	 *
2395 	 * Silence PROVE_RCU.
2396 	 */
2397 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2398 	/*
2399 	 * We're setting the CPU for the first time, we don't migrate,
2400 	 * so use __set_task_cpu().
2401 	 */
2402 	__set_task_cpu(p, cpu);
2403 	if (p->sched_class->task_fork)
2404 		p->sched_class->task_fork(p);
2405 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2406 
2407 #ifdef CONFIG_SCHED_INFO
2408 	if (likely(sched_info_on()))
2409 		memset(&p->sched_info, 0, sizeof(p->sched_info));
2410 #endif
2411 #if defined(CONFIG_SMP)
2412 	p->on_cpu = 0;
2413 #endif
2414 	init_task_preempt_count(p);
2415 #ifdef CONFIG_SMP
2416 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
2417 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
2418 #endif
2419 
2420 	put_cpu();
2421 	return 0;
2422 }
2423 
2424 unsigned long to_ratio(u64 period, u64 runtime)
2425 {
2426 	if (runtime == RUNTIME_INF)
2427 		return BW_UNIT;
2428 
2429 	/*
2430 	 * Doing this here saves a lot of checks in all
2431 	 * the calling paths, and returning zero seems
2432 	 * safe for them anyway.
2433 	 */
2434 	if (period == 0)
2435 		return 0;
2436 
2437 	return div64_u64(runtime << BW_SHIFT, period);
2438 }
2439 
2440 /*
2441  * wake_up_new_task - wake up a newly created task for the first time.
2442  *
2443  * This function will do some initial scheduler statistics housekeeping
2444  * that must be done for every newly created context, then puts the task
2445  * on the runqueue and wakes it.
2446  */
2447 void wake_up_new_task(struct task_struct *p)
2448 {
2449 	struct rq_flags rf;
2450 	struct rq *rq;
2451 
2452 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2453 	p->state = TASK_RUNNING;
2454 #ifdef CONFIG_SMP
2455 	/*
2456 	 * Fork balancing, do it here and not earlier because:
2457 	 *  - cpus_allowed can change in the fork path
2458 	 *  - any previously selected CPU might disappear through hotplug
2459 	 *
2460 	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2461 	 * as we're not fully set-up yet.
2462 	 */
2463 	__set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2464 #endif
2465 	rq = __task_rq_lock(p, &rf);
2466 	update_rq_clock(rq);
2467 	post_init_entity_util_avg(&p->se);
2468 
2469 	activate_task(rq, p, ENQUEUE_NOCLOCK);
2470 	p->on_rq = TASK_ON_RQ_QUEUED;
2471 	trace_sched_wakeup_new(p);
2472 	check_preempt_curr(rq, p, WF_FORK);
2473 #ifdef CONFIG_SMP
2474 	if (p->sched_class->task_woken) {
2475 		/*
2476 		 * Nothing relies on rq->lock after this, so its fine to
2477 		 * drop it.
2478 		 */
2479 		rq_unpin_lock(rq, &rf);
2480 		p->sched_class->task_woken(rq, p);
2481 		rq_repin_lock(rq, &rf);
2482 	}
2483 #endif
2484 	task_rq_unlock(rq, p, &rf);
2485 }
2486 
2487 #ifdef CONFIG_PREEMPT_NOTIFIERS
2488 
2489 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2490 
2491 void preempt_notifier_inc(void)
2492 {
2493 	static_key_slow_inc(&preempt_notifier_key);
2494 }
2495 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2496 
2497 void preempt_notifier_dec(void)
2498 {
2499 	static_key_slow_dec(&preempt_notifier_key);
2500 }
2501 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2502 
2503 /**
2504  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2505  * @notifier: notifier struct to register
2506  */
2507 void preempt_notifier_register(struct preempt_notifier *notifier)
2508 {
2509 	if (!static_key_false(&preempt_notifier_key))
2510 		WARN(1, "registering preempt_notifier while notifiers disabled\n");
2511 
2512 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
2513 }
2514 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2515 
2516 /**
2517  * preempt_notifier_unregister - no longer interested in preemption notifications
2518  * @notifier: notifier struct to unregister
2519  *
2520  * This is *not* safe to call from within a preemption notifier.
2521  */
2522 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2523 {
2524 	hlist_del(&notifier->link);
2525 }
2526 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2527 
2528 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2529 {
2530 	struct preempt_notifier *notifier;
2531 
2532 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2533 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
2534 }
2535 
2536 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2537 {
2538 	if (static_key_false(&preempt_notifier_key))
2539 		__fire_sched_in_preempt_notifiers(curr);
2540 }
2541 
2542 static void
2543 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2544 				   struct task_struct *next)
2545 {
2546 	struct preempt_notifier *notifier;
2547 
2548 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2549 		notifier->ops->sched_out(notifier, next);
2550 }
2551 
2552 static __always_inline void
2553 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2554 				 struct task_struct *next)
2555 {
2556 	if (static_key_false(&preempt_notifier_key))
2557 		__fire_sched_out_preempt_notifiers(curr, next);
2558 }
2559 
2560 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2561 
2562 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2563 {
2564 }
2565 
2566 static inline void
2567 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2568 				 struct task_struct *next)
2569 {
2570 }
2571 
2572 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2573 
2574 /**
2575  * prepare_task_switch - prepare to switch tasks
2576  * @rq: the runqueue preparing to switch
2577  * @prev: the current task that is being switched out
2578  * @next: the task we are going to switch to.
2579  *
2580  * This is called with the rq lock held and interrupts off. It must
2581  * be paired with a subsequent finish_task_switch after the context
2582  * switch.
2583  *
2584  * prepare_task_switch sets up locking and calls architecture specific
2585  * hooks.
2586  */
2587 static inline void
2588 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2589 		    struct task_struct *next)
2590 {
2591 	sched_info_switch(rq, prev, next);
2592 	perf_event_task_sched_out(prev, next);
2593 	fire_sched_out_preempt_notifiers(prev, next);
2594 	prepare_lock_switch(rq, next);
2595 	prepare_arch_switch(next);
2596 }
2597 
2598 /**
2599  * finish_task_switch - clean up after a task-switch
2600  * @prev: the thread we just switched away from.
2601  *
2602  * finish_task_switch must be called after the context switch, paired
2603  * with a prepare_task_switch call before the context switch.
2604  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2605  * and do any other architecture-specific cleanup actions.
2606  *
2607  * Note that we may have delayed dropping an mm in context_switch(). If
2608  * so, we finish that here outside of the runqueue lock. (Doing it
2609  * with the lock held can cause deadlocks; see schedule() for
2610  * details.)
2611  *
2612  * The context switch have flipped the stack from under us and restored the
2613  * local variables which were saved when this task called schedule() in the
2614  * past. prev == current is still correct but we need to recalculate this_rq
2615  * because prev may have moved to another CPU.
2616  */
2617 static struct rq *finish_task_switch(struct task_struct *prev)
2618 	__releases(rq->lock)
2619 {
2620 	struct rq *rq = this_rq();
2621 	struct mm_struct *mm = rq->prev_mm;
2622 	long prev_state;
2623 
2624 	/*
2625 	 * The previous task will have left us with a preempt_count of 2
2626 	 * because it left us after:
2627 	 *
2628 	 *	schedule()
2629 	 *	  preempt_disable();			// 1
2630 	 *	  __schedule()
2631 	 *	    raw_spin_lock_irq(&rq->lock)	// 2
2632 	 *
2633 	 * Also, see FORK_PREEMPT_COUNT.
2634 	 */
2635 	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2636 		      "corrupted preempt_count: %s/%d/0x%x\n",
2637 		      current->comm, current->pid, preempt_count()))
2638 		preempt_count_set(FORK_PREEMPT_COUNT);
2639 
2640 	rq->prev_mm = NULL;
2641 
2642 	/*
2643 	 * A task struct has one reference for the use as "current".
2644 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2645 	 * schedule one last time. The schedule call will never return, and
2646 	 * the scheduled task must drop that reference.
2647 	 *
2648 	 * We must observe prev->state before clearing prev->on_cpu (in
2649 	 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2650 	 * running on another CPU and we could rave with its RUNNING -> DEAD
2651 	 * transition, resulting in a double drop.
2652 	 */
2653 	prev_state = prev->state;
2654 	vtime_task_switch(prev);
2655 	perf_event_task_sched_in(prev, current);
2656 	/*
2657 	 * The membarrier system call requires a full memory barrier
2658 	 * after storing to rq->curr, before going back to user-space.
2659 	 *
2660 	 * TODO: This smp_mb__after_unlock_lock can go away if PPC end
2661 	 * up adding a full barrier to switch_mm(), or we should figure
2662 	 * out if a smp_mb__after_unlock_lock is really the proper API
2663 	 * to use.
2664 	 */
2665 	smp_mb__after_unlock_lock();
2666 	finish_lock_switch(rq, prev);
2667 	finish_arch_post_lock_switch();
2668 
2669 	fire_sched_in_preempt_notifiers(current);
2670 	if (mm)
2671 		mmdrop(mm);
2672 	if (unlikely(prev_state == TASK_DEAD)) {
2673 		if (prev->sched_class->task_dead)
2674 			prev->sched_class->task_dead(prev);
2675 
2676 		/*
2677 		 * Remove function-return probe instances associated with this
2678 		 * task and put them back on the free list.
2679 		 */
2680 		kprobe_flush_task(prev);
2681 
2682 		/* Task is done with its stack. */
2683 		put_task_stack(prev);
2684 
2685 		put_task_struct(prev);
2686 	}
2687 
2688 	tick_nohz_task_switch();
2689 	return rq;
2690 }
2691 
2692 #ifdef CONFIG_SMP
2693 
2694 /* rq->lock is NOT held, but preemption is disabled */
2695 static void __balance_callback(struct rq *rq)
2696 {
2697 	struct callback_head *head, *next;
2698 	void (*func)(struct rq *rq);
2699 	unsigned long flags;
2700 
2701 	raw_spin_lock_irqsave(&rq->lock, flags);
2702 	head = rq->balance_callback;
2703 	rq->balance_callback = NULL;
2704 	while (head) {
2705 		func = (void (*)(struct rq *))head->func;
2706 		next = head->next;
2707 		head->next = NULL;
2708 		head = next;
2709 
2710 		func(rq);
2711 	}
2712 	raw_spin_unlock_irqrestore(&rq->lock, flags);
2713 }
2714 
2715 static inline void balance_callback(struct rq *rq)
2716 {
2717 	if (unlikely(rq->balance_callback))
2718 		__balance_callback(rq);
2719 }
2720 
2721 #else
2722 
2723 static inline void balance_callback(struct rq *rq)
2724 {
2725 }
2726 
2727 #endif
2728 
2729 /**
2730  * schedule_tail - first thing a freshly forked thread must call.
2731  * @prev: the thread we just switched away from.
2732  */
2733 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2734 	__releases(rq->lock)
2735 {
2736 	struct rq *rq;
2737 
2738 	/*
2739 	 * New tasks start with FORK_PREEMPT_COUNT, see there and
2740 	 * finish_task_switch() for details.
2741 	 *
2742 	 * finish_task_switch() will drop rq->lock() and lower preempt_count
2743 	 * and the preempt_enable() will end up enabling preemption (on
2744 	 * PREEMPT_COUNT kernels).
2745 	 */
2746 
2747 	rq = finish_task_switch(prev);
2748 	balance_callback(rq);
2749 	preempt_enable();
2750 
2751 	if (current->set_child_tid)
2752 		put_user(task_pid_vnr(current), current->set_child_tid);
2753 }
2754 
2755 /*
2756  * context_switch - switch to the new MM and the new thread's register state.
2757  */
2758 static __always_inline struct rq *
2759 context_switch(struct rq *rq, struct task_struct *prev,
2760 	       struct task_struct *next, struct rq_flags *rf)
2761 {
2762 	struct mm_struct *mm, *oldmm;
2763 
2764 	prepare_task_switch(rq, prev, next);
2765 
2766 	mm = next->mm;
2767 	oldmm = prev->active_mm;
2768 	/*
2769 	 * For paravirt, this is coupled with an exit in switch_to to
2770 	 * combine the page table reload and the switch backend into
2771 	 * one hypercall.
2772 	 */
2773 	arch_start_context_switch(prev);
2774 
2775 	if (!mm) {
2776 		next->active_mm = oldmm;
2777 		mmgrab(oldmm);
2778 		enter_lazy_tlb(oldmm, next);
2779 	} else
2780 		switch_mm_irqs_off(oldmm, mm, next);
2781 
2782 	if (!prev->mm) {
2783 		prev->active_mm = NULL;
2784 		rq->prev_mm = oldmm;
2785 	}
2786 
2787 	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2788 
2789 	/*
2790 	 * Since the runqueue lock will be released by the next
2791 	 * task (which is an invalid locking op but in the case
2792 	 * of the scheduler it's an obvious special-case), so we
2793 	 * do an early lockdep release here:
2794 	 */
2795 	rq_unpin_lock(rq, rf);
2796 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2797 
2798 	/* Here we just switch the register state and the stack. */
2799 	switch_to(prev, next, prev);
2800 	barrier();
2801 
2802 	return finish_task_switch(prev);
2803 }
2804 
2805 /*
2806  * nr_running and nr_context_switches:
2807  *
2808  * externally visible scheduler statistics: current number of runnable
2809  * threads, total number of context switches performed since bootup.
2810  */
2811 unsigned long nr_running(void)
2812 {
2813 	unsigned long i, sum = 0;
2814 
2815 	for_each_online_cpu(i)
2816 		sum += cpu_rq(i)->nr_running;
2817 
2818 	return sum;
2819 }
2820 
2821 /*
2822  * Check if only the current task is running on the CPU.
2823  *
2824  * Caution: this function does not check that the caller has disabled
2825  * preemption, thus the result might have a time-of-check-to-time-of-use
2826  * race.  The caller is responsible to use it correctly, for example:
2827  *
2828  * - from a non-preemptable section (of course)
2829  *
2830  * - from a thread that is bound to a single CPU
2831  *
2832  * - in a loop with very short iterations (e.g. a polling loop)
2833  */
2834 bool single_task_running(void)
2835 {
2836 	return raw_rq()->nr_running == 1;
2837 }
2838 EXPORT_SYMBOL(single_task_running);
2839 
2840 unsigned long long nr_context_switches(void)
2841 {
2842 	int i;
2843 	unsigned long long sum = 0;
2844 
2845 	for_each_possible_cpu(i)
2846 		sum += cpu_rq(i)->nr_switches;
2847 
2848 	return sum;
2849 }
2850 
2851 /*
2852  * IO-wait accounting, and how its mostly bollocks (on SMP).
2853  *
2854  * The idea behind IO-wait account is to account the idle time that we could
2855  * have spend running if it were not for IO. That is, if we were to improve the
2856  * storage performance, we'd have a proportional reduction in IO-wait time.
2857  *
2858  * This all works nicely on UP, where, when a task blocks on IO, we account
2859  * idle time as IO-wait, because if the storage were faster, it could've been
2860  * running and we'd not be idle.
2861  *
2862  * This has been extended to SMP, by doing the same for each CPU. This however
2863  * is broken.
2864  *
2865  * Imagine for instance the case where two tasks block on one CPU, only the one
2866  * CPU will have IO-wait accounted, while the other has regular idle. Even
2867  * though, if the storage were faster, both could've ran at the same time,
2868  * utilising both CPUs.
2869  *
2870  * This means, that when looking globally, the current IO-wait accounting on
2871  * SMP is a lower bound, by reason of under accounting.
2872  *
2873  * Worse, since the numbers are provided per CPU, they are sometimes
2874  * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2875  * associated with any one particular CPU, it can wake to another CPU than it
2876  * blocked on. This means the per CPU IO-wait number is meaningless.
2877  *
2878  * Task CPU affinities can make all that even more 'interesting'.
2879  */
2880 
2881 unsigned long nr_iowait(void)
2882 {
2883 	unsigned long i, sum = 0;
2884 
2885 	for_each_possible_cpu(i)
2886 		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2887 
2888 	return sum;
2889 }
2890 
2891 /*
2892  * Consumers of these two interfaces, like for example the cpufreq menu
2893  * governor are using nonsensical data. Boosting frequency for a CPU that has
2894  * IO-wait which might not even end up running the task when it does become
2895  * runnable.
2896  */
2897 
2898 unsigned long nr_iowait_cpu(int cpu)
2899 {
2900 	struct rq *this = cpu_rq(cpu);
2901 	return atomic_read(&this->nr_iowait);
2902 }
2903 
2904 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2905 {
2906 	struct rq *rq = this_rq();
2907 	*nr_waiters = atomic_read(&rq->nr_iowait);
2908 	*load = rq->load.weight;
2909 }
2910 
2911 #ifdef CONFIG_SMP
2912 
2913 /*
2914  * sched_exec - execve() is a valuable balancing opportunity, because at
2915  * this point the task has the smallest effective memory and cache footprint.
2916  */
2917 void sched_exec(void)
2918 {
2919 	struct task_struct *p = current;
2920 	unsigned long flags;
2921 	int dest_cpu;
2922 
2923 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2924 	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2925 	if (dest_cpu == smp_processor_id())
2926 		goto unlock;
2927 
2928 	if (likely(cpu_active(dest_cpu))) {
2929 		struct migration_arg arg = { p, dest_cpu };
2930 
2931 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2932 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2933 		return;
2934 	}
2935 unlock:
2936 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2937 }
2938 
2939 #endif
2940 
2941 DEFINE_PER_CPU(struct kernel_stat, kstat);
2942 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2943 
2944 EXPORT_PER_CPU_SYMBOL(kstat);
2945 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2946 
2947 /*
2948  * The function fair_sched_class.update_curr accesses the struct curr
2949  * and its field curr->exec_start; when called from task_sched_runtime(),
2950  * we observe a high rate of cache misses in practice.
2951  * Prefetching this data results in improved performance.
2952  */
2953 static inline void prefetch_curr_exec_start(struct task_struct *p)
2954 {
2955 #ifdef CONFIG_FAIR_GROUP_SCHED
2956 	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2957 #else
2958 	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2959 #endif
2960 	prefetch(curr);
2961 	prefetch(&curr->exec_start);
2962 }
2963 
2964 /*
2965  * Return accounted runtime for the task.
2966  * In case the task is currently running, return the runtime plus current's
2967  * pending runtime that have not been accounted yet.
2968  */
2969 unsigned long long task_sched_runtime(struct task_struct *p)
2970 {
2971 	struct rq_flags rf;
2972 	struct rq *rq;
2973 	u64 ns;
2974 
2975 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2976 	/*
2977 	 * 64-bit doesn't need locks to atomically read a 64bit value.
2978 	 * So we have a optimization chance when the task's delta_exec is 0.
2979 	 * Reading ->on_cpu is racy, but this is ok.
2980 	 *
2981 	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
2982 	 * If we race with it entering CPU, unaccounted time is 0. This is
2983 	 * indistinguishable from the read occurring a few cycles earlier.
2984 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2985 	 * been accounted, so we're correct here as well.
2986 	 */
2987 	if (!p->on_cpu || !task_on_rq_queued(p))
2988 		return p->se.sum_exec_runtime;
2989 #endif
2990 
2991 	rq = task_rq_lock(p, &rf);
2992 	/*
2993 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2994 	 * project cycles that may never be accounted to this
2995 	 * thread, breaking clock_gettime().
2996 	 */
2997 	if (task_current(rq, p) && task_on_rq_queued(p)) {
2998 		prefetch_curr_exec_start(p);
2999 		update_rq_clock(rq);
3000 		p->sched_class->update_curr(rq);
3001 	}
3002 	ns = p->se.sum_exec_runtime;
3003 	task_rq_unlock(rq, p, &rf);
3004 
3005 	return ns;
3006 }
3007 
3008 /*
3009  * This function gets called by the timer code, with HZ frequency.
3010  * We call it with interrupts disabled.
3011  */
3012 void scheduler_tick(void)
3013 {
3014 	int cpu = smp_processor_id();
3015 	struct rq *rq = cpu_rq(cpu);
3016 	struct task_struct *curr = rq->curr;
3017 	struct rq_flags rf;
3018 
3019 	sched_clock_tick();
3020 
3021 	rq_lock(rq, &rf);
3022 
3023 	update_rq_clock(rq);
3024 	curr->sched_class->task_tick(rq, curr, 0);
3025 	cpu_load_update_active(rq);
3026 	calc_global_load_tick(rq);
3027 
3028 	rq_unlock(rq, &rf);
3029 
3030 	perf_event_task_tick();
3031 
3032 #ifdef CONFIG_SMP
3033 	rq->idle_balance = idle_cpu(cpu);
3034 	trigger_load_balance(rq);
3035 #endif
3036 	rq_last_tick_reset(rq);
3037 }
3038 
3039 #ifdef CONFIG_NO_HZ_FULL
3040 /**
3041  * scheduler_tick_max_deferment
3042  *
3043  * Keep at least one tick per second when a single
3044  * active task is running because the scheduler doesn't
3045  * yet completely support full dynticks environment.
3046  *
3047  * This makes sure that uptime, CFS vruntime, load
3048  * balancing, etc... continue to move forward, even
3049  * with a very low granularity.
3050  *
3051  * Return: Maximum deferment in nanoseconds.
3052  */
3053 u64 scheduler_tick_max_deferment(void)
3054 {
3055 	struct rq *rq = this_rq();
3056 	unsigned long next, now = READ_ONCE(jiffies);
3057 
3058 	next = rq->last_sched_tick + HZ;
3059 
3060 	if (time_before_eq(next, now))
3061 		return 0;
3062 
3063 	return jiffies_to_nsecs(next - now);
3064 }
3065 #endif
3066 
3067 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3068 				defined(CONFIG_PREEMPT_TRACER))
3069 /*
3070  * If the value passed in is equal to the current preempt count
3071  * then we just disabled preemption. Start timing the latency.
3072  */
3073 static inline void preempt_latency_start(int val)
3074 {
3075 	if (preempt_count() == val) {
3076 		unsigned long ip = get_lock_parent_ip();
3077 #ifdef CONFIG_DEBUG_PREEMPT
3078 		current->preempt_disable_ip = ip;
3079 #endif
3080 		trace_preempt_off(CALLER_ADDR0, ip);
3081 	}
3082 }
3083 
3084 void preempt_count_add(int val)
3085 {
3086 #ifdef CONFIG_DEBUG_PREEMPT
3087 	/*
3088 	 * Underflow?
3089 	 */
3090 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3091 		return;
3092 #endif
3093 	__preempt_count_add(val);
3094 #ifdef CONFIG_DEBUG_PREEMPT
3095 	/*
3096 	 * Spinlock count overflowing soon?
3097 	 */
3098 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3099 				PREEMPT_MASK - 10);
3100 #endif
3101 	preempt_latency_start(val);
3102 }
3103 EXPORT_SYMBOL(preempt_count_add);
3104 NOKPROBE_SYMBOL(preempt_count_add);
3105 
3106 /*
3107  * If the value passed in equals to the current preempt count
3108  * then we just enabled preemption. Stop timing the latency.
3109  */
3110 static inline void preempt_latency_stop(int val)
3111 {
3112 	if (preempt_count() == val)
3113 		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3114 }
3115 
3116 void preempt_count_sub(int val)
3117 {
3118 #ifdef CONFIG_DEBUG_PREEMPT
3119 	/*
3120 	 * Underflow?
3121 	 */
3122 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3123 		return;
3124 	/*
3125 	 * Is the spinlock portion underflowing?
3126 	 */
3127 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3128 			!(preempt_count() & PREEMPT_MASK)))
3129 		return;
3130 #endif
3131 
3132 	preempt_latency_stop(val);
3133 	__preempt_count_sub(val);
3134 }
3135 EXPORT_SYMBOL(preempt_count_sub);
3136 NOKPROBE_SYMBOL(preempt_count_sub);
3137 
3138 #else
3139 static inline void preempt_latency_start(int val) { }
3140 static inline void preempt_latency_stop(int val) { }
3141 #endif
3142 
3143 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3144 {
3145 #ifdef CONFIG_DEBUG_PREEMPT
3146 	return p->preempt_disable_ip;
3147 #else
3148 	return 0;
3149 #endif
3150 }
3151 
3152 /*
3153  * Print scheduling while atomic bug:
3154  */
3155 static noinline void __schedule_bug(struct task_struct *prev)
3156 {
3157 	/* Save this before calling printk(), since that will clobber it */
3158 	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3159 
3160 	if (oops_in_progress)
3161 		return;
3162 
3163 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3164 		prev->comm, prev->pid, preempt_count());
3165 
3166 	debug_show_held_locks(prev);
3167 	print_modules();
3168 	if (irqs_disabled())
3169 		print_irqtrace_events(prev);
3170 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3171 	    && in_atomic_preempt_off()) {
3172 		pr_err("Preemption disabled at:");
3173 		print_ip_sym(preempt_disable_ip);
3174 		pr_cont("\n");
3175 	}
3176 	if (panic_on_warn)
3177 		panic("scheduling while atomic\n");
3178 
3179 	dump_stack();
3180 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3181 }
3182 
3183 /*
3184  * Various schedule()-time debugging checks and statistics:
3185  */
3186 static inline void schedule_debug(struct task_struct *prev)
3187 {
3188 #ifdef CONFIG_SCHED_STACK_END_CHECK
3189 	if (task_stack_end_corrupted(prev))
3190 		panic("corrupted stack end detected inside scheduler\n");
3191 #endif
3192 
3193 	if (unlikely(in_atomic_preempt_off())) {
3194 		__schedule_bug(prev);
3195 		preempt_count_set(PREEMPT_DISABLED);
3196 	}
3197 	rcu_sleep_check();
3198 
3199 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3200 
3201 	schedstat_inc(this_rq()->sched_count);
3202 }
3203 
3204 /*
3205  * Pick up the highest-prio task:
3206  */
3207 static inline struct task_struct *
3208 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3209 {
3210 	const struct sched_class *class;
3211 	struct task_struct *p;
3212 
3213 	/*
3214 	 * Optimization: we know that if all tasks are in the fair class we can
3215 	 * call that function directly, but only if the @prev task wasn't of a
3216 	 * higher scheduling class, because otherwise those loose the
3217 	 * opportunity to pull in more work from other CPUs.
3218 	 */
3219 	if (likely((prev->sched_class == &idle_sched_class ||
3220 		    prev->sched_class == &fair_sched_class) &&
3221 		   rq->nr_running == rq->cfs.h_nr_running)) {
3222 
3223 		p = fair_sched_class.pick_next_task(rq, prev, rf);
3224 		if (unlikely(p == RETRY_TASK))
3225 			goto again;
3226 
3227 		/* Assumes fair_sched_class->next == idle_sched_class */
3228 		if (unlikely(!p))
3229 			p = idle_sched_class.pick_next_task(rq, prev, rf);
3230 
3231 		return p;
3232 	}
3233 
3234 again:
3235 	for_each_class(class) {
3236 		p = class->pick_next_task(rq, prev, rf);
3237 		if (p) {
3238 			if (unlikely(p == RETRY_TASK))
3239 				goto again;
3240 			return p;
3241 		}
3242 	}
3243 
3244 	/* The idle class should always have a runnable task: */
3245 	BUG();
3246 }
3247 
3248 /*
3249  * __schedule() is the main scheduler function.
3250  *
3251  * The main means of driving the scheduler and thus entering this function are:
3252  *
3253  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3254  *
3255  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3256  *      paths. For example, see arch/x86/entry_64.S.
3257  *
3258  *      To drive preemption between tasks, the scheduler sets the flag in timer
3259  *      interrupt handler scheduler_tick().
3260  *
3261  *   3. Wakeups don't really cause entry into schedule(). They add a
3262  *      task to the run-queue and that's it.
3263  *
3264  *      Now, if the new task added to the run-queue preempts the current
3265  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3266  *      called on the nearest possible occasion:
3267  *
3268  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
3269  *
3270  *         - in syscall or exception context, at the next outmost
3271  *           preempt_enable(). (this might be as soon as the wake_up()'s
3272  *           spin_unlock()!)
3273  *
3274  *         - in IRQ context, return from interrupt-handler to
3275  *           preemptible context
3276  *
3277  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3278  *         then at the next:
3279  *
3280  *          - cond_resched() call
3281  *          - explicit schedule() call
3282  *          - return from syscall or exception to user-space
3283  *          - return from interrupt-handler to user-space
3284  *
3285  * WARNING: must be called with preemption disabled!
3286  */
3287 static void __sched notrace __schedule(bool preempt)
3288 {
3289 	struct task_struct *prev, *next;
3290 	unsigned long *switch_count;
3291 	struct rq_flags rf;
3292 	struct rq *rq;
3293 	int cpu;
3294 
3295 	cpu = smp_processor_id();
3296 	rq = cpu_rq(cpu);
3297 	prev = rq->curr;
3298 
3299 	schedule_debug(prev);
3300 
3301 	if (sched_feat(HRTICK))
3302 		hrtick_clear(rq);
3303 
3304 	local_irq_disable();
3305 	rcu_note_context_switch(preempt);
3306 
3307 	/*
3308 	 * Make sure that signal_pending_state()->signal_pending() below
3309 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3310 	 * done by the caller to avoid the race with signal_wake_up().
3311 	 */
3312 	rq_lock(rq, &rf);
3313 	smp_mb__after_spinlock();
3314 
3315 	/* Promote REQ to ACT */
3316 	rq->clock_update_flags <<= 1;
3317 	update_rq_clock(rq);
3318 
3319 	switch_count = &prev->nivcsw;
3320 	if (!preempt && prev->state) {
3321 		if (unlikely(signal_pending_state(prev->state, prev))) {
3322 			prev->state = TASK_RUNNING;
3323 		} else {
3324 			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3325 			prev->on_rq = 0;
3326 
3327 			if (prev->in_iowait) {
3328 				atomic_inc(&rq->nr_iowait);
3329 				delayacct_blkio_start();
3330 			}
3331 
3332 			/*
3333 			 * If a worker went to sleep, notify and ask workqueue
3334 			 * whether it wants to wake up a task to maintain
3335 			 * concurrency.
3336 			 */
3337 			if (prev->flags & PF_WQ_WORKER) {
3338 				struct task_struct *to_wakeup;
3339 
3340 				to_wakeup = wq_worker_sleeping(prev);
3341 				if (to_wakeup)
3342 					try_to_wake_up_local(to_wakeup, &rf);
3343 			}
3344 		}
3345 		switch_count = &prev->nvcsw;
3346 	}
3347 
3348 	next = pick_next_task(rq, prev, &rf);
3349 	clear_tsk_need_resched(prev);
3350 	clear_preempt_need_resched();
3351 
3352 	if (likely(prev != next)) {
3353 		rq->nr_switches++;
3354 		rq->curr = next;
3355 		/*
3356 		 * The membarrier system call requires each architecture
3357 		 * to have a full memory barrier after updating
3358 		 * rq->curr, before returning to user-space. For TSO
3359 		 * (e.g. x86), the architecture must provide its own
3360 		 * barrier in switch_mm(). For weakly ordered machines
3361 		 * for which spin_unlock() acts as a full memory
3362 		 * barrier, finish_lock_switch() in common code takes
3363 		 * care of this barrier. For weakly ordered machines for
3364 		 * which spin_unlock() acts as a RELEASE barrier (only
3365 		 * arm64 and PowerPC), arm64 has a full barrier in
3366 		 * switch_to(), and PowerPC has
3367 		 * smp_mb__after_unlock_lock() before
3368 		 * finish_lock_switch().
3369 		 */
3370 		++*switch_count;
3371 
3372 		trace_sched_switch(preempt, prev, next);
3373 
3374 		/* Also unlocks the rq: */
3375 		rq = context_switch(rq, prev, next, &rf);
3376 	} else {
3377 		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3378 		rq_unlock_irq(rq, &rf);
3379 	}
3380 
3381 	balance_callback(rq);
3382 }
3383 
3384 void __noreturn do_task_dead(void)
3385 {
3386 	/*
3387 	 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3388 	 * when the following two conditions become true.
3389 	 *   - There is race condition of mmap_sem (It is acquired by
3390 	 *     exit_mm()), and
3391 	 *   - SMI occurs before setting TASK_RUNINNG.
3392 	 *     (or hypervisor of virtual machine switches to other guest)
3393 	 *  As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3394 	 *
3395 	 * To avoid it, we have to wait for releasing tsk->pi_lock which
3396 	 * is held by try_to_wake_up()
3397 	 */
3398 	raw_spin_lock_irq(&current->pi_lock);
3399 	raw_spin_unlock_irq(&current->pi_lock);
3400 
3401 	/* Causes final put_task_struct in finish_task_switch(): */
3402 	__set_current_state(TASK_DEAD);
3403 
3404 	/* Tell freezer to ignore us: */
3405 	current->flags |= PF_NOFREEZE;
3406 
3407 	__schedule(false);
3408 	BUG();
3409 
3410 	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3411 	for (;;)
3412 		cpu_relax();
3413 }
3414 
3415 static inline void sched_submit_work(struct task_struct *tsk)
3416 {
3417 	if (!tsk->state || tsk_is_pi_blocked(tsk))
3418 		return;
3419 	/*
3420 	 * If we are going to sleep and we have plugged IO queued,
3421 	 * make sure to submit it to avoid deadlocks.
3422 	 */
3423 	if (blk_needs_flush_plug(tsk))
3424 		blk_schedule_flush_plug(tsk);
3425 }
3426 
3427 asmlinkage __visible void __sched schedule(void)
3428 {
3429 	struct task_struct *tsk = current;
3430 
3431 	sched_submit_work(tsk);
3432 	do {
3433 		preempt_disable();
3434 		__schedule(false);
3435 		sched_preempt_enable_no_resched();
3436 	} while (need_resched());
3437 }
3438 EXPORT_SYMBOL(schedule);
3439 
3440 /*
3441  * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3442  * state (have scheduled out non-voluntarily) by making sure that all
3443  * tasks have either left the run queue or have gone into user space.
3444  * As idle tasks do not do either, they must not ever be preempted
3445  * (schedule out non-voluntarily).
3446  *
3447  * schedule_idle() is similar to schedule_preempt_disable() except that it
3448  * never enables preemption because it does not call sched_submit_work().
3449  */
3450 void __sched schedule_idle(void)
3451 {
3452 	/*
3453 	 * As this skips calling sched_submit_work(), which the idle task does
3454 	 * regardless because that function is a nop when the task is in a
3455 	 * TASK_RUNNING state, make sure this isn't used someplace that the
3456 	 * current task can be in any other state. Note, idle is always in the
3457 	 * TASK_RUNNING state.
3458 	 */
3459 	WARN_ON_ONCE(current->state);
3460 	do {
3461 		__schedule(false);
3462 	} while (need_resched());
3463 }
3464 
3465 #ifdef CONFIG_CONTEXT_TRACKING
3466 asmlinkage __visible void __sched schedule_user(void)
3467 {
3468 	/*
3469 	 * If we come here after a random call to set_need_resched(),
3470 	 * or we have been woken up remotely but the IPI has not yet arrived,
3471 	 * we haven't yet exited the RCU idle mode. Do it here manually until
3472 	 * we find a better solution.
3473 	 *
3474 	 * NB: There are buggy callers of this function.  Ideally we
3475 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
3476 	 * too frequently to make sense yet.
3477 	 */
3478 	enum ctx_state prev_state = exception_enter();
3479 	schedule();
3480 	exception_exit(prev_state);
3481 }
3482 #endif
3483 
3484 /**
3485  * schedule_preempt_disabled - called with preemption disabled
3486  *
3487  * Returns with preemption disabled. Note: preempt_count must be 1
3488  */
3489 void __sched schedule_preempt_disabled(void)
3490 {
3491 	sched_preempt_enable_no_resched();
3492 	schedule();
3493 	preempt_disable();
3494 }
3495 
3496 static void __sched notrace preempt_schedule_common(void)
3497 {
3498 	do {
3499 		/*
3500 		 * Because the function tracer can trace preempt_count_sub()
3501 		 * and it also uses preempt_enable/disable_notrace(), if
3502 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3503 		 * by the function tracer will call this function again and
3504 		 * cause infinite recursion.
3505 		 *
3506 		 * Preemption must be disabled here before the function
3507 		 * tracer can trace. Break up preempt_disable() into two
3508 		 * calls. One to disable preemption without fear of being
3509 		 * traced. The other to still record the preemption latency,
3510 		 * which can also be traced by the function tracer.
3511 		 */
3512 		preempt_disable_notrace();
3513 		preempt_latency_start(1);
3514 		__schedule(true);
3515 		preempt_latency_stop(1);
3516 		preempt_enable_no_resched_notrace();
3517 
3518 		/*
3519 		 * Check again in case we missed a preemption opportunity
3520 		 * between schedule and now.
3521 		 */
3522 	} while (need_resched());
3523 }
3524 
3525 #ifdef CONFIG_PREEMPT
3526 /*
3527  * this is the entry point to schedule() from in-kernel preemption
3528  * off of preempt_enable. Kernel preemptions off return from interrupt
3529  * occur there and call schedule directly.
3530  */
3531 asmlinkage __visible void __sched notrace preempt_schedule(void)
3532 {
3533 	/*
3534 	 * If there is a non-zero preempt_count or interrupts are disabled,
3535 	 * we do not want to preempt the current task. Just return..
3536 	 */
3537 	if (likely(!preemptible()))
3538 		return;
3539 
3540 	preempt_schedule_common();
3541 }
3542 NOKPROBE_SYMBOL(preempt_schedule);
3543 EXPORT_SYMBOL(preempt_schedule);
3544 
3545 /**
3546  * preempt_schedule_notrace - preempt_schedule called by tracing
3547  *
3548  * The tracing infrastructure uses preempt_enable_notrace to prevent
3549  * recursion and tracing preempt enabling caused by the tracing
3550  * infrastructure itself. But as tracing can happen in areas coming
3551  * from userspace or just about to enter userspace, a preempt enable
3552  * can occur before user_exit() is called. This will cause the scheduler
3553  * to be called when the system is still in usermode.
3554  *
3555  * To prevent this, the preempt_enable_notrace will use this function
3556  * instead of preempt_schedule() to exit user context if needed before
3557  * calling the scheduler.
3558  */
3559 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3560 {
3561 	enum ctx_state prev_ctx;
3562 
3563 	if (likely(!preemptible()))
3564 		return;
3565 
3566 	do {
3567 		/*
3568 		 * Because the function tracer can trace preempt_count_sub()
3569 		 * and it also uses preempt_enable/disable_notrace(), if
3570 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3571 		 * by the function tracer will call this function again and
3572 		 * cause infinite recursion.
3573 		 *
3574 		 * Preemption must be disabled here before the function
3575 		 * tracer can trace. Break up preempt_disable() into two
3576 		 * calls. One to disable preemption without fear of being
3577 		 * traced. The other to still record the preemption latency,
3578 		 * which can also be traced by the function tracer.
3579 		 */
3580 		preempt_disable_notrace();
3581 		preempt_latency_start(1);
3582 		/*
3583 		 * Needs preempt disabled in case user_exit() is traced
3584 		 * and the tracer calls preempt_enable_notrace() causing
3585 		 * an infinite recursion.
3586 		 */
3587 		prev_ctx = exception_enter();
3588 		__schedule(true);
3589 		exception_exit(prev_ctx);
3590 
3591 		preempt_latency_stop(1);
3592 		preempt_enable_no_resched_notrace();
3593 	} while (need_resched());
3594 }
3595 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3596 
3597 #endif /* CONFIG_PREEMPT */
3598 
3599 /*
3600  * this is the entry point to schedule() from kernel preemption
3601  * off of irq context.
3602  * Note, that this is called and return with irqs disabled. This will
3603  * protect us against recursive calling from irq.
3604  */
3605 asmlinkage __visible void __sched preempt_schedule_irq(void)
3606 {
3607 	enum ctx_state prev_state;
3608 
3609 	/* Catch callers which need to be fixed */
3610 	BUG_ON(preempt_count() || !irqs_disabled());
3611 
3612 	prev_state = exception_enter();
3613 
3614 	do {
3615 		preempt_disable();
3616 		local_irq_enable();
3617 		__schedule(true);
3618 		local_irq_disable();
3619 		sched_preempt_enable_no_resched();
3620 	} while (need_resched());
3621 
3622 	exception_exit(prev_state);
3623 }
3624 
3625 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3626 			  void *key)
3627 {
3628 	return try_to_wake_up(curr->private, mode, wake_flags);
3629 }
3630 EXPORT_SYMBOL(default_wake_function);
3631 
3632 #ifdef CONFIG_RT_MUTEXES
3633 
3634 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3635 {
3636 	if (pi_task)
3637 		prio = min(prio, pi_task->prio);
3638 
3639 	return prio;
3640 }
3641 
3642 static inline int rt_effective_prio(struct task_struct *p, int prio)
3643 {
3644 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
3645 
3646 	return __rt_effective_prio(pi_task, prio);
3647 }
3648 
3649 /*
3650  * rt_mutex_setprio - set the current priority of a task
3651  * @p: task to boost
3652  * @pi_task: donor task
3653  *
3654  * This function changes the 'effective' priority of a task. It does
3655  * not touch ->normal_prio like __setscheduler().
3656  *
3657  * Used by the rt_mutex code to implement priority inheritance
3658  * logic. Call site only calls if the priority of the task changed.
3659  */
3660 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3661 {
3662 	int prio, oldprio, queued, running, queue_flag =
3663 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3664 	const struct sched_class *prev_class;
3665 	struct rq_flags rf;
3666 	struct rq *rq;
3667 
3668 	/* XXX used to be waiter->prio, not waiter->task->prio */
3669 	prio = __rt_effective_prio(pi_task, p->normal_prio);
3670 
3671 	/*
3672 	 * If nothing changed; bail early.
3673 	 */
3674 	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3675 		return;
3676 
3677 	rq = __task_rq_lock(p, &rf);
3678 	update_rq_clock(rq);
3679 	/*
3680 	 * Set under pi_lock && rq->lock, such that the value can be used under
3681 	 * either lock.
3682 	 *
3683 	 * Note that there is loads of tricky to make this pointer cache work
3684 	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3685 	 * ensure a task is de-boosted (pi_task is set to NULL) before the
3686 	 * task is allowed to run again (and can exit). This ensures the pointer
3687 	 * points to a blocked task -- which guaratees the task is present.
3688 	 */
3689 	p->pi_top_task = pi_task;
3690 
3691 	/*
3692 	 * For FIFO/RR we only need to set prio, if that matches we're done.
3693 	 */
3694 	if (prio == p->prio && !dl_prio(prio))
3695 		goto out_unlock;
3696 
3697 	/*
3698 	 * Idle task boosting is a nono in general. There is one
3699 	 * exception, when PREEMPT_RT and NOHZ is active:
3700 	 *
3701 	 * The idle task calls get_next_timer_interrupt() and holds
3702 	 * the timer wheel base->lock on the CPU and another CPU wants
3703 	 * to access the timer (probably to cancel it). We can safely
3704 	 * ignore the boosting request, as the idle CPU runs this code
3705 	 * with interrupts disabled and will complete the lock
3706 	 * protected section without being interrupted. So there is no
3707 	 * real need to boost.
3708 	 */
3709 	if (unlikely(p == rq->idle)) {
3710 		WARN_ON(p != rq->curr);
3711 		WARN_ON(p->pi_blocked_on);
3712 		goto out_unlock;
3713 	}
3714 
3715 	trace_sched_pi_setprio(p, pi_task);
3716 	oldprio = p->prio;
3717 
3718 	if (oldprio == prio)
3719 		queue_flag &= ~DEQUEUE_MOVE;
3720 
3721 	prev_class = p->sched_class;
3722 	queued = task_on_rq_queued(p);
3723 	running = task_current(rq, p);
3724 	if (queued)
3725 		dequeue_task(rq, p, queue_flag);
3726 	if (running)
3727 		put_prev_task(rq, p);
3728 
3729 	/*
3730 	 * Boosting condition are:
3731 	 * 1. -rt task is running and holds mutex A
3732 	 *      --> -dl task blocks on mutex A
3733 	 *
3734 	 * 2. -dl task is running and holds mutex A
3735 	 *      --> -dl task blocks on mutex A and could preempt the
3736 	 *          running task
3737 	 */
3738 	if (dl_prio(prio)) {
3739 		if (!dl_prio(p->normal_prio) ||
3740 		    (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3741 			p->dl.dl_boosted = 1;
3742 			queue_flag |= ENQUEUE_REPLENISH;
3743 		} else
3744 			p->dl.dl_boosted = 0;
3745 		p->sched_class = &dl_sched_class;
3746 	} else if (rt_prio(prio)) {
3747 		if (dl_prio(oldprio))
3748 			p->dl.dl_boosted = 0;
3749 		if (oldprio < prio)
3750 			queue_flag |= ENQUEUE_HEAD;
3751 		p->sched_class = &rt_sched_class;
3752 	} else {
3753 		if (dl_prio(oldprio))
3754 			p->dl.dl_boosted = 0;
3755 		if (rt_prio(oldprio))
3756 			p->rt.timeout = 0;
3757 		p->sched_class = &fair_sched_class;
3758 	}
3759 
3760 	p->prio = prio;
3761 
3762 	if (queued)
3763 		enqueue_task(rq, p, queue_flag);
3764 	if (running)
3765 		set_curr_task(rq, p);
3766 
3767 	check_class_changed(rq, p, prev_class, oldprio);
3768 out_unlock:
3769 	/* Avoid rq from going away on us: */
3770 	preempt_disable();
3771 	__task_rq_unlock(rq, &rf);
3772 
3773 	balance_callback(rq);
3774 	preempt_enable();
3775 }
3776 #else
3777 static inline int rt_effective_prio(struct task_struct *p, int prio)
3778 {
3779 	return prio;
3780 }
3781 #endif
3782 
3783 void set_user_nice(struct task_struct *p, long nice)
3784 {
3785 	bool queued, running;
3786 	int old_prio, delta;
3787 	struct rq_flags rf;
3788 	struct rq *rq;
3789 
3790 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3791 		return;
3792 	/*
3793 	 * We have to be careful, if called from sys_setpriority(),
3794 	 * the task might be in the middle of scheduling on another CPU.
3795 	 */
3796 	rq = task_rq_lock(p, &rf);
3797 	update_rq_clock(rq);
3798 
3799 	/*
3800 	 * The RT priorities are set via sched_setscheduler(), but we still
3801 	 * allow the 'normal' nice value to be set - but as expected
3802 	 * it wont have any effect on scheduling until the task is
3803 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3804 	 */
3805 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3806 		p->static_prio = NICE_TO_PRIO(nice);
3807 		goto out_unlock;
3808 	}
3809 	queued = task_on_rq_queued(p);
3810 	running = task_current(rq, p);
3811 	if (queued)
3812 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3813 	if (running)
3814 		put_prev_task(rq, p);
3815 
3816 	p->static_prio = NICE_TO_PRIO(nice);
3817 	set_load_weight(p, true);
3818 	old_prio = p->prio;
3819 	p->prio = effective_prio(p);
3820 	delta = p->prio - old_prio;
3821 
3822 	if (queued) {
3823 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3824 		/*
3825 		 * If the task increased its priority or is running and
3826 		 * lowered its priority, then reschedule its CPU:
3827 		 */
3828 		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3829 			resched_curr(rq);
3830 	}
3831 	if (running)
3832 		set_curr_task(rq, p);
3833 out_unlock:
3834 	task_rq_unlock(rq, p, &rf);
3835 }
3836 EXPORT_SYMBOL(set_user_nice);
3837 
3838 /*
3839  * can_nice - check if a task can reduce its nice value
3840  * @p: task
3841  * @nice: nice value
3842  */
3843 int can_nice(const struct task_struct *p, const int nice)
3844 {
3845 	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
3846 	int nice_rlim = nice_to_rlimit(nice);
3847 
3848 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3849 		capable(CAP_SYS_NICE));
3850 }
3851 
3852 #ifdef __ARCH_WANT_SYS_NICE
3853 
3854 /*
3855  * sys_nice - change the priority of the current process.
3856  * @increment: priority increment
3857  *
3858  * sys_setpriority is a more generic, but much slower function that
3859  * does similar things.
3860  */
3861 SYSCALL_DEFINE1(nice, int, increment)
3862 {
3863 	long nice, retval;
3864 
3865 	/*
3866 	 * Setpriority might change our priority at the same moment.
3867 	 * We don't have to worry. Conceptually one call occurs first
3868 	 * and we have a single winner.
3869 	 */
3870 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3871 	nice = task_nice(current) + increment;
3872 
3873 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3874 	if (increment < 0 && !can_nice(current, nice))
3875 		return -EPERM;
3876 
3877 	retval = security_task_setnice(current, nice);
3878 	if (retval)
3879 		return retval;
3880 
3881 	set_user_nice(current, nice);
3882 	return 0;
3883 }
3884 
3885 #endif
3886 
3887 /**
3888  * task_prio - return the priority value of a given task.
3889  * @p: the task in question.
3890  *
3891  * Return: The priority value as seen by users in /proc.
3892  * RT tasks are offset by -200. Normal tasks are centered
3893  * around 0, value goes from -16 to +15.
3894  */
3895 int task_prio(const struct task_struct *p)
3896 {
3897 	return p->prio - MAX_RT_PRIO;
3898 }
3899 
3900 /**
3901  * idle_cpu - is a given CPU idle currently?
3902  * @cpu: the processor in question.
3903  *
3904  * Return: 1 if the CPU is currently idle. 0 otherwise.
3905  */
3906 int idle_cpu(int cpu)
3907 {
3908 	struct rq *rq = cpu_rq(cpu);
3909 
3910 	if (rq->curr != rq->idle)
3911 		return 0;
3912 
3913 	if (rq->nr_running)
3914 		return 0;
3915 
3916 #ifdef CONFIG_SMP
3917 	if (!llist_empty(&rq->wake_list))
3918 		return 0;
3919 #endif
3920 
3921 	return 1;
3922 }
3923 
3924 /**
3925  * idle_task - return the idle task for a given CPU.
3926  * @cpu: the processor in question.
3927  *
3928  * Return: The idle task for the CPU @cpu.
3929  */
3930 struct task_struct *idle_task(int cpu)
3931 {
3932 	return cpu_rq(cpu)->idle;
3933 }
3934 
3935 /**
3936  * find_process_by_pid - find a process with a matching PID value.
3937  * @pid: the pid in question.
3938  *
3939  * The task of @pid, if found. %NULL otherwise.
3940  */
3941 static struct task_struct *find_process_by_pid(pid_t pid)
3942 {
3943 	return pid ? find_task_by_vpid(pid) : current;
3944 }
3945 
3946 /*
3947  * sched_setparam() passes in -1 for its policy, to let the functions
3948  * it calls know not to change it.
3949  */
3950 #define SETPARAM_POLICY	-1
3951 
3952 static void __setscheduler_params(struct task_struct *p,
3953 		const struct sched_attr *attr)
3954 {
3955 	int policy = attr->sched_policy;
3956 
3957 	if (policy == SETPARAM_POLICY)
3958 		policy = p->policy;
3959 
3960 	p->policy = policy;
3961 
3962 	if (dl_policy(policy))
3963 		__setparam_dl(p, attr);
3964 	else if (fair_policy(policy))
3965 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3966 
3967 	/*
3968 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3969 	 * !rt_policy. Always setting this ensures that things like
3970 	 * getparam()/getattr() don't report silly values for !rt tasks.
3971 	 */
3972 	p->rt_priority = attr->sched_priority;
3973 	p->normal_prio = normal_prio(p);
3974 	set_load_weight(p, true);
3975 }
3976 
3977 /* Actually do priority change: must hold pi & rq lock. */
3978 static void __setscheduler(struct rq *rq, struct task_struct *p,
3979 			   const struct sched_attr *attr, bool keep_boost)
3980 {
3981 	__setscheduler_params(p, attr);
3982 
3983 	/*
3984 	 * Keep a potential priority boosting if called from
3985 	 * sched_setscheduler().
3986 	 */
3987 	p->prio = normal_prio(p);
3988 	if (keep_boost)
3989 		p->prio = rt_effective_prio(p, p->prio);
3990 
3991 	if (dl_prio(p->prio))
3992 		p->sched_class = &dl_sched_class;
3993 	else if (rt_prio(p->prio))
3994 		p->sched_class = &rt_sched_class;
3995 	else
3996 		p->sched_class = &fair_sched_class;
3997 }
3998 
3999 /*
4000  * Check the target process has a UID that matches the current process's:
4001  */
4002 static bool check_same_owner(struct task_struct *p)
4003 {
4004 	const struct cred *cred = current_cred(), *pcred;
4005 	bool match;
4006 
4007 	rcu_read_lock();
4008 	pcred = __task_cred(p);
4009 	match = (uid_eq(cred->euid, pcred->euid) ||
4010 		 uid_eq(cred->euid, pcred->uid));
4011 	rcu_read_unlock();
4012 	return match;
4013 }
4014 
4015 static int __sched_setscheduler(struct task_struct *p,
4016 				const struct sched_attr *attr,
4017 				bool user, bool pi)
4018 {
4019 	int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4020 		      MAX_RT_PRIO - 1 - attr->sched_priority;
4021 	int retval, oldprio, oldpolicy = -1, queued, running;
4022 	int new_effective_prio, policy = attr->sched_policy;
4023 	const struct sched_class *prev_class;
4024 	struct rq_flags rf;
4025 	int reset_on_fork;
4026 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4027 	struct rq *rq;
4028 
4029 	/* The pi code expects interrupts enabled */
4030 	BUG_ON(pi && in_interrupt());
4031 recheck:
4032 	/* Double check policy once rq lock held: */
4033 	if (policy < 0) {
4034 		reset_on_fork = p->sched_reset_on_fork;
4035 		policy = oldpolicy = p->policy;
4036 	} else {
4037 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4038 
4039 		if (!valid_policy(policy))
4040 			return -EINVAL;
4041 	}
4042 
4043 	if (attr->sched_flags &
4044 		~(SCHED_FLAG_RESET_ON_FORK | SCHED_FLAG_RECLAIM))
4045 		return -EINVAL;
4046 
4047 	/*
4048 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
4049 	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4050 	 * SCHED_BATCH and SCHED_IDLE is 0.
4051 	 */
4052 	if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4053 	    (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4054 		return -EINVAL;
4055 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4056 	    (rt_policy(policy) != (attr->sched_priority != 0)))
4057 		return -EINVAL;
4058 
4059 	/*
4060 	 * Allow unprivileged RT tasks to decrease priority:
4061 	 */
4062 	if (user && !capable(CAP_SYS_NICE)) {
4063 		if (fair_policy(policy)) {
4064 			if (attr->sched_nice < task_nice(p) &&
4065 			    !can_nice(p, attr->sched_nice))
4066 				return -EPERM;
4067 		}
4068 
4069 		if (rt_policy(policy)) {
4070 			unsigned long rlim_rtprio =
4071 					task_rlimit(p, RLIMIT_RTPRIO);
4072 
4073 			/* Can't set/change the rt policy: */
4074 			if (policy != p->policy && !rlim_rtprio)
4075 				return -EPERM;
4076 
4077 			/* Can't increase priority: */
4078 			if (attr->sched_priority > p->rt_priority &&
4079 			    attr->sched_priority > rlim_rtprio)
4080 				return -EPERM;
4081 		}
4082 
4083 		 /*
4084 		  * Can't set/change SCHED_DEADLINE policy at all for now
4085 		  * (safest behavior); in the future we would like to allow
4086 		  * unprivileged DL tasks to increase their relative deadline
4087 		  * or reduce their runtime (both ways reducing utilization)
4088 		  */
4089 		if (dl_policy(policy))
4090 			return -EPERM;
4091 
4092 		/*
4093 		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4094 		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4095 		 */
4096 		if (idle_policy(p->policy) && !idle_policy(policy)) {
4097 			if (!can_nice(p, task_nice(p)))
4098 				return -EPERM;
4099 		}
4100 
4101 		/* Can't change other user's priorities: */
4102 		if (!check_same_owner(p))
4103 			return -EPERM;
4104 
4105 		/* Normal users shall not reset the sched_reset_on_fork flag: */
4106 		if (p->sched_reset_on_fork && !reset_on_fork)
4107 			return -EPERM;
4108 	}
4109 
4110 	if (user) {
4111 		retval = security_task_setscheduler(p);
4112 		if (retval)
4113 			return retval;
4114 	}
4115 
4116 	/*
4117 	 * Make sure no PI-waiters arrive (or leave) while we are
4118 	 * changing the priority of the task:
4119 	 *
4120 	 * To be able to change p->policy safely, the appropriate
4121 	 * runqueue lock must be held.
4122 	 */
4123 	rq = task_rq_lock(p, &rf);
4124 	update_rq_clock(rq);
4125 
4126 	/*
4127 	 * Changing the policy of the stop threads its a very bad idea:
4128 	 */
4129 	if (p == rq->stop) {
4130 		task_rq_unlock(rq, p, &rf);
4131 		return -EINVAL;
4132 	}
4133 
4134 	/*
4135 	 * If not changing anything there's no need to proceed further,
4136 	 * but store a possible modification of reset_on_fork.
4137 	 */
4138 	if (unlikely(policy == p->policy)) {
4139 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4140 			goto change;
4141 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4142 			goto change;
4143 		if (dl_policy(policy) && dl_param_changed(p, attr))
4144 			goto change;
4145 
4146 		p->sched_reset_on_fork = reset_on_fork;
4147 		task_rq_unlock(rq, p, &rf);
4148 		return 0;
4149 	}
4150 change:
4151 
4152 	if (user) {
4153 #ifdef CONFIG_RT_GROUP_SCHED
4154 		/*
4155 		 * Do not allow realtime tasks into groups that have no runtime
4156 		 * assigned.
4157 		 */
4158 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
4159 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4160 				!task_group_is_autogroup(task_group(p))) {
4161 			task_rq_unlock(rq, p, &rf);
4162 			return -EPERM;
4163 		}
4164 #endif
4165 #ifdef CONFIG_SMP
4166 		if (dl_bandwidth_enabled() && dl_policy(policy)) {
4167 			cpumask_t *span = rq->rd->span;
4168 
4169 			/*
4170 			 * Don't allow tasks with an affinity mask smaller than
4171 			 * the entire root_domain to become SCHED_DEADLINE. We
4172 			 * will also fail if there's no bandwidth available.
4173 			 */
4174 			if (!cpumask_subset(span, &p->cpus_allowed) ||
4175 			    rq->rd->dl_bw.bw == 0) {
4176 				task_rq_unlock(rq, p, &rf);
4177 				return -EPERM;
4178 			}
4179 		}
4180 #endif
4181 	}
4182 
4183 	/* Re-check policy now with rq lock held: */
4184 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4185 		policy = oldpolicy = -1;
4186 		task_rq_unlock(rq, p, &rf);
4187 		goto recheck;
4188 	}
4189 
4190 	/*
4191 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4192 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4193 	 * is available.
4194 	 */
4195 	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4196 		task_rq_unlock(rq, p, &rf);
4197 		return -EBUSY;
4198 	}
4199 
4200 	p->sched_reset_on_fork = reset_on_fork;
4201 	oldprio = p->prio;
4202 
4203 	if (pi) {
4204 		/*
4205 		 * Take priority boosted tasks into account. If the new
4206 		 * effective priority is unchanged, we just store the new
4207 		 * normal parameters and do not touch the scheduler class and
4208 		 * the runqueue. This will be done when the task deboost
4209 		 * itself.
4210 		 */
4211 		new_effective_prio = rt_effective_prio(p, newprio);
4212 		if (new_effective_prio == oldprio)
4213 			queue_flags &= ~DEQUEUE_MOVE;
4214 	}
4215 
4216 	queued = task_on_rq_queued(p);
4217 	running = task_current(rq, p);
4218 	if (queued)
4219 		dequeue_task(rq, p, queue_flags);
4220 	if (running)
4221 		put_prev_task(rq, p);
4222 
4223 	prev_class = p->sched_class;
4224 	__setscheduler(rq, p, attr, pi);
4225 
4226 	if (queued) {
4227 		/*
4228 		 * We enqueue to tail when the priority of a task is
4229 		 * increased (user space view).
4230 		 */
4231 		if (oldprio < p->prio)
4232 			queue_flags |= ENQUEUE_HEAD;
4233 
4234 		enqueue_task(rq, p, queue_flags);
4235 	}
4236 	if (running)
4237 		set_curr_task(rq, p);
4238 
4239 	check_class_changed(rq, p, prev_class, oldprio);
4240 
4241 	/* Avoid rq from going away on us: */
4242 	preempt_disable();
4243 	task_rq_unlock(rq, p, &rf);
4244 
4245 	if (pi)
4246 		rt_mutex_adjust_pi(p);
4247 
4248 	/* Run balance callbacks after we've adjusted the PI chain: */
4249 	balance_callback(rq);
4250 	preempt_enable();
4251 
4252 	return 0;
4253 }
4254 
4255 static int _sched_setscheduler(struct task_struct *p, int policy,
4256 			       const struct sched_param *param, bool check)
4257 {
4258 	struct sched_attr attr = {
4259 		.sched_policy   = policy,
4260 		.sched_priority = param->sched_priority,
4261 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
4262 	};
4263 
4264 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4265 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4266 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4267 		policy &= ~SCHED_RESET_ON_FORK;
4268 		attr.sched_policy = policy;
4269 	}
4270 
4271 	return __sched_setscheduler(p, &attr, check, true);
4272 }
4273 /**
4274  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4275  * @p: the task in question.
4276  * @policy: new policy.
4277  * @param: structure containing the new RT priority.
4278  *
4279  * Return: 0 on success. An error code otherwise.
4280  *
4281  * NOTE that the task may be already dead.
4282  */
4283 int sched_setscheduler(struct task_struct *p, int policy,
4284 		       const struct sched_param *param)
4285 {
4286 	return _sched_setscheduler(p, policy, param, true);
4287 }
4288 EXPORT_SYMBOL_GPL(sched_setscheduler);
4289 
4290 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4291 {
4292 	return __sched_setscheduler(p, attr, true, true);
4293 }
4294 EXPORT_SYMBOL_GPL(sched_setattr);
4295 
4296 /**
4297  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4298  * @p: the task in question.
4299  * @policy: new policy.
4300  * @param: structure containing the new RT priority.
4301  *
4302  * Just like sched_setscheduler, only don't bother checking if the
4303  * current context has permission.  For example, this is needed in
4304  * stop_machine(): we create temporary high priority worker threads,
4305  * but our caller might not have that capability.
4306  *
4307  * Return: 0 on success. An error code otherwise.
4308  */
4309 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4310 			       const struct sched_param *param)
4311 {
4312 	return _sched_setscheduler(p, policy, param, false);
4313 }
4314 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4315 
4316 static int
4317 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4318 {
4319 	struct sched_param lparam;
4320 	struct task_struct *p;
4321 	int retval;
4322 
4323 	if (!param || pid < 0)
4324 		return -EINVAL;
4325 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4326 		return -EFAULT;
4327 
4328 	rcu_read_lock();
4329 	retval = -ESRCH;
4330 	p = find_process_by_pid(pid);
4331 	if (p != NULL)
4332 		retval = sched_setscheduler(p, policy, &lparam);
4333 	rcu_read_unlock();
4334 
4335 	return retval;
4336 }
4337 
4338 /*
4339  * Mimics kernel/events/core.c perf_copy_attr().
4340  */
4341 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4342 {
4343 	u32 size;
4344 	int ret;
4345 
4346 	if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4347 		return -EFAULT;
4348 
4349 	/* Zero the full structure, so that a short copy will be nice: */
4350 	memset(attr, 0, sizeof(*attr));
4351 
4352 	ret = get_user(size, &uattr->size);
4353 	if (ret)
4354 		return ret;
4355 
4356 	/* Bail out on silly large: */
4357 	if (size > PAGE_SIZE)
4358 		goto err_size;
4359 
4360 	/* ABI compatibility quirk: */
4361 	if (!size)
4362 		size = SCHED_ATTR_SIZE_VER0;
4363 
4364 	if (size < SCHED_ATTR_SIZE_VER0)
4365 		goto err_size;
4366 
4367 	/*
4368 	 * If we're handed a bigger struct than we know of,
4369 	 * ensure all the unknown bits are 0 - i.e. new
4370 	 * user-space does not rely on any kernel feature
4371 	 * extensions we dont know about yet.
4372 	 */
4373 	if (size > sizeof(*attr)) {
4374 		unsigned char __user *addr;
4375 		unsigned char __user *end;
4376 		unsigned char val;
4377 
4378 		addr = (void __user *)uattr + sizeof(*attr);
4379 		end  = (void __user *)uattr + size;
4380 
4381 		for (; addr < end; addr++) {
4382 			ret = get_user(val, addr);
4383 			if (ret)
4384 				return ret;
4385 			if (val)
4386 				goto err_size;
4387 		}
4388 		size = sizeof(*attr);
4389 	}
4390 
4391 	ret = copy_from_user(attr, uattr, size);
4392 	if (ret)
4393 		return -EFAULT;
4394 
4395 	/*
4396 	 * XXX: Do we want to be lenient like existing syscalls; or do we want
4397 	 * to be strict and return an error on out-of-bounds values?
4398 	 */
4399 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4400 
4401 	return 0;
4402 
4403 err_size:
4404 	put_user(sizeof(*attr), &uattr->size);
4405 	return -E2BIG;
4406 }
4407 
4408 /**
4409  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4410  * @pid: the pid in question.
4411  * @policy: new policy.
4412  * @param: structure containing the new RT priority.
4413  *
4414  * Return: 0 on success. An error code otherwise.
4415  */
4416 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4417 {
4418 	if (policy < 0)
4419 		return -EINVAL;
4420 
4421 	return do_sched_setscheduler(pid, policy, param);
4422 }
4423 
4424 /**
4425  * sys_sched_setparam - set/change the RT priority of a thread
4426  * @pid: the pid in question.
4427  * @param: structure containing the new RT priority.
4428  *
4429  * Return: 0 on success. An error code otherwise.
4430  */
4431 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4432 {
4433 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4434 }
4435 
4436 /**
4437  * sys_sched_setattr - same as above, but with extended sched_attr
4438  * @pid: the pid in question.
4439  * @uattr: structure containing the extended parameters.
4440  * @flags: for future extension.
4441  */
4442 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4443 			       unsigned int, flags)
4444 {
4445 	struct sched_attr attr;
4446 	struct task_struct *p;
4447 	int retval;
4448 
4449 	if (!uattr || pid < 0 || flags)
4450 		return -EINVAL;
4451 
4452 	retval = sched_copy_attr(uattr, &attr);
4453 	if (retval)
4454 		return retval;
4455 
4456 	if ((int)attr.sched_policy < 0)
4457 		return -EINVAL;
4458 
4459 	rcu_read_lock();
4460 	retval = -ESRCH;
4461 	p = find_process_by_pid(pid);
4462 	if (p != NULL)
4463 		retval = sched_setattr(p, &attr);
4464 	rcu_read_unlock();
4465 
4466 	return retval;
4467 }
4468 
4469 /**
4470  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4471  * @pid: the pid in question.
4472  *
4473  * Return: On success, the policy of the thread. Otherwise, a negative error
4474  * code.
4475  */
4476 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4477 {
4478 	struct task_struct *p;
4479 	int retval;
4480 
4481 	if (pid < 0)
4482 		return -EINVAL;
4483 
4484 	retval = -ESRCH;
4485 	rcu_read_lock();
4486 	p = find_process_by_pid(pid);
4487 	if (p) {
4488 		retval = security_task_getscheduler(p);
4489 		if (!retval)
4490 			retval = p->policy
4491 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4492 	}
4493 	rcu_read_unlock();
4494 	return retval;
4495 }
4496 
4497 /**
4498  * sys_sched_getparam - get the RT priority of a thread
4499  * @pid: the pid in question.
4500  * @param: structure containing the RT priority.
4501  *
4502  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4503  * code.
4504  */
4505 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4506 {
4507 	struct sched_param lp = { .sched_priority = 0 };
4508 	struct task_struct *p;
4509 	int retval;
4510 
4511 	if (!param || pid < 0)
4512 		return -EINVAL;
4513 
4514 	rcu_read_lock();
4515 	p = find_process_by_pid(pid);
4516 	retval = -ESRCH;
4517 	if (!p)
4518 		goto out_unlock;
4519 
4520 	retval = security_task_getscheduler(p);
4521 	if (retval)
4522 		goto out_unlock;
4523 
4524 	if (task_has_rt_policy(p))
4525 		lp.sched_priority = p->rt_priority;
4526 	rcu_read_unlock();
4527 
4528 	/*
4529 	 * This one might sleep, we cannot do it with a spinlock held ...
4530 	 */
4531 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4532 
4533 	return retval;
4534 
4535 out_unlock:
4536 	rcu_read_unlock();
4537 	return retval;
4538 }
4539 
4540 static int sched_read_attr(struct sched_attr __user *uattr,
4541 			   struct sched_attr *attr,
4542 			   unsigned int usize)
4543 {
4544 	int ret;
4545 
4546 	if (!access_ok(VERIFY_WRITE, uattr, usize))
4547 		return -EFAULT;
4548 
4549 	/*
4550 	 * If we're handed a smaller struct than we know of,
4551 	 * ensure all the unknown bits are 0 - i.e. old
4552 	 * user-space does not get uncomplete information.
4553 	 */
4554 	if (usize < sizeof(*attr)) {
4555 		unsigned char *addr;
4556 		unsigned char *end;
4557 
4558 		addr = (void *)attr + usize;
4559 		end  = (void *)attr + sizeof(*attr);
4560 
4561 		for (; addr < end; addr++) {
4562 			if (*addr)
4563 				return -EFBIG;
4564 		}
4565 
4566 		attr->size = usize;
4567 	}
4568 
4569 	ret = copy_to_user(uattr, attr, attr->size);
4570 	if (ret)
4571 		return -EFAULT;
4572 
4573 	return 0;
4574 }
4575 
4576 /**
4577  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4578  * @pid: the pid in question.
4579  * @uattr: structure containing the extended parameters.
4580  * @size: sizeof(attr) for fwd/bwd comp.
4581  * @flags: for future extension.
4582  */
4583 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4584 		unsigned int, size, unsigned int, flags)
4585 {
4586 	struct sched_attr attr = {
4587 		.size = sizeof(struct sched_attr),
4588 	};
4589 	struct task_struct *p;
4590 	int retval;
4591 
4592 	if (!uattr || pid < 0 || size > PAGE_SIZE ||
4593 	    size < SCHED_ATTR_SIZE_VER0 || flags)
4594 		return -EINVAL;
4595 
4596 	rcu_read_lock();
4597 	p = find_process_by_pid(pid);
4598 	retval = -ESRCH;
4599 	if (!p)
4600 		goto out_unlock;
4601 
4602 	retval = security_task_getscheduler(p);
4603 	if (retval)
4604 		goto out_unlock;
4605 
4606 	attr.sched_policy = p->policy;
4607 	if (p->sched_reset_on_fork)
4608 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4609 	if (task_has_dl_policy(p))
4610 		__getparam_dl(p, &attr);
4611 	else if (task_has_rt_policy(p))
4612 		attr.sched_priority = p->rt_priority;
4613 	else
4614 		attr.sched_nice = task_nice(p);
4615 
4616 	rcu_read_unlock();
4617 
4618 	retval = sched_read_attr(uattr, &attr, size);
4619 	return retval;
4620 
4621 out_unlock:
4622 	rcu_read_unlock();
4623 	return retval;
4624 }
4625 
4626 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4627 {
4628 	cpumask_var_t cpus_allowed, new_mask;
4629 	struct task_struct *p;
4630 	int retval;
4631 
4632 	rcu_read_lock();
4633 
4634 	p = find_process_by_pid(pid);
4635 	if (!p) {
4636 		rcu_read_unlock();
4637 		return -ESRCH;
4638 	}
4639 
4640 	/* Prevent p going away */
4641 	get_task_struct(p);
4642 	rcu_read_unlock();
4643 
4644 	if (p->flags & PF_NO_SETAFFINITY) {
4645 		retval = -EINVAL;
4646 		goto out_put_task;
4647 	}
4648 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4649 		retval = -ENOMEM;
4650 		goto out_put_task;
4651 	}
4652 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4653 		retval = -ENOMEM;
4654 		goto out_free_cpus_allowed;
4655 	}
4656 	retval = -EPERM;
4657 	if (!check_same_owner(p)) {
4658 		rcu_read_lock();
4659 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4660 			rcu_read_unlock();
4661 			goto out_free_new_mask;
4662 		}
4663 		rcu_read_unlock();
4664 	}
4665 
4666 	retval = security_task_setscheduler(p);
4667 	if (retval)
4668 		goto out_free_new_mask;
4669 
4670 
4671 	cpuset_cpus_allowed(p, cpus_allowed);
4672 	cpumask_and(new_mask, in_mask, cpus_allowed);
4673 
4674 	/*
4675 	 * Since bandwidth control happens on root_domain basis,
4676 	 * if admission test is enabled, we only admit -deadline
4677 	 * tasks allowed to run on all the CPUs in the task's
4678 	 * root_domain.
4679 	 */
4680 #ifdef CONFIG_SMP
4681 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4682 		rcu_read_lock();
4683 		if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4684 			retval = -EBUSY;
4685 			rcu_read_unlock();
4686 			goto out_free_new_mask;
4687 		}
4688 		rcu_read_unlock();
4689 	}
4690 #endif
4691 again:
4692 	retval = __set_cpus_allowed_ptr(p, new_mask, true);
4693 
4694 	if (!retval) {
4695 		cpuset_cpus_allowed(p, cpus_allowed);
4696 		if (!cpumask_subset(new_mask, cpus_allowed)) {
4697 			/*
4698 			 * We must have raced with a concurrent cpuset
4699 			 * update. Just reset the cpus_allowed to the
4700 			 * cpuset's cpus_allowed
4701 			 */
4702 			cpumask_copy(new_mask, cpus_allowed);
4703 			goto again;
4704 		}
4705 	}
4706 out_free_new_mask:
4707 	free_cpumask_var(new_mask);
4708 out_free_cpus_allowed:
4709 	free_cpumask_var(cpus_allowed);
4710 out_put_task:
4711 	put_task_struct(p);
4712 	return retval;
4713 }
4714 
4715 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4716 			     struct cpumask *new_mask)
4717 {
4718 	if (len < cpumask_size())
4719 		cpumask_clear(new_mask);
4720 	else if (len > cpumask_size())
4721 		len = cpumask_size();
4722 
4723 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4724 }
4725 
4726 /**
4727  * sys_sched_setaffinity - set the CPU affinity of a process
4728  * @pid: pid of the process
4729  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4730  * @user_mask_ptr: user-space pointer to the new CPU mask
4731  *
4732  * Return: 0 on success. An error code otherwise.
4733  */
4734 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4735 		unsigned long __user *, user_mask_ptr)
4736 {
4737 	cpumask_var_t new_mask;
4738 	int retval;
4739 
4740 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4741 		return -ENOMEM;
4742 
4743 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4744 	if (retval == 0)
4745 		retval = sched_setaffinity(pid, new_mask);
4746 	free_cpumask_var(new_mask);
4747 	return retval;
4748 }
4749 
4750 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4751 {
4752 	struct task_struct *p;
4753 	unsigned long flags;
4754 	int retval;
4755 
4756 	rcu_read_lock();
4757 
4758 	retval = -ESRCH;
4759 	p = find_process_by_pid(pid);
4760 	if (!p)
4761 		goto out_unlock;
4762 
4763 	retval = security_task_getscheduler(p);
4764 	if (retval)
4765 		goto out_unlock;
4766 
4767 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4768 	cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4769 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4770 
4771 out_unlock:
4772 	rcu_read_unlock();
4773 
4774 	return retval;
4775 }
4776 
4777 /**
4778  * sys_sched_getaffinity - get the CPU affinity of a process
4779  * @pid: pid of the process
4780  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4781  * @user_mask_ptr: user-space pointer to hold the current CPU mask
4782  *
4783  * Return: size of CPU mask copied to user_mask_ptr on success. An
4784  * error code otherwise.
4785  */
4786 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4787 		unsigned long __user *, user_mask_ptr)
4788 {
4789 	int ret;
4790 	cpumask_var_t mask;
4791 
4792 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4793 		return -EINVAL;
4794 	if (len & (sizeof(unsigned long)-1))
4795 		return -EINVAL;
4796 
4797 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4798 		return -ENOMEM;
4799 
4800 	ret = sched_getaffinity(pid, mask);
4801 	if (ret == 0) {
4802 		size_t retlen = min_t(size_t, len, cpumask_size());
4803 
4804 		if (copy_to_user(user_mask_ptr, mask, retlen))
4805 			ret = -EFAULT;
4806 		else
4807 			ret = retlen;
4808 	}
4809 	free_cpumask_var(mask);
4810 
4811 	return ret;
4812 }
4813 
4814 /**
4815  * sys_sched_yield - yield the current processor to other threads.
4816  *
4817  * This function yields the current CPU to other tasks. If there are no
4818  * other threads running on this CPU then this function will return.
4819  *
4820  * Return: 0.
4821  */
4822 SYSCALL_DEFINE0(sched_yield)
4823 {
4824 	struct rq_flags rf;
4825 	struct rq *rq;
4826 
4827 	local_irq_disable();
4828 	rq = this_rq();
4829 	rq_lock(rq, &rf);
4830 
4831 	schedstat_inc(rq->yld_count);
4832 	current->sched_class->yield_task(rq);
4833 
4834 	/*
4835 	 * Since we are going to call schedule() anyway, there's
4836 	 * no need to preempt or enable interrupts:
4837 	 */
4838 	preempt_disable();
4839 	rq_unlock(rq, &rf);
4840 	sched_preempt_enable_no_resched();
4841 
4842 	schedule();
4843 
4844 	return 0;
4845 }
4846 
4847 #ifndef CONFIG_PREEMPT
4848 int __sched _cond_resched(void)
4849 {
4850 	if (should_resched(0)) {
4851 		preempt_schedule_common();
4852 		return 1;
4853 	}
4854 	rcu_all_qs();
4855 	return 0;
4856 }
4857 EXPORT_SYMBOL(_cond_resched);
4858 #endif
4859 
4860 /*
4861  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4862  * call schedule, and on return reacquire the lock.
4863  *
4864  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4865  * operations here to prevent schedule() from being called twice (once via
4866  * spin_unlock(), once by hand).
4867  */
4868 int __cond_resched_lock(spinlock_t *lock)
4869 {
4870 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
4871 	int ret = 0;
4872 
4873 	lockdep_assert_held(lock);
4874 
4875 	if (spin_needbreak(lock) || resched) {
4876 		spin_unlock(lock);
4877 		if (resched)
4878 			preempt_schedule_common();
4879 		else
4880 			cpu_relax();
4881 		ret = 1;
4882 		spin_lock(lock);
4883 	}
4884 	return ret;
4885 }
4886 EXPORT_SYMBOL(__cond_resched_lock);
4887 
4888 int __sched __cond_resched_softirq(void)
4889 {
4890 	BUG_ON(!in_softirq());
4891 
4892 	if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4893 		local_bh_enable();
4894 		preempt_schedule_common();
4895 		local_bh_disable();
4896 		return 1;
4897 	}
4898 	return 0;
4899 }
4900 EXPORT_SYMBOL(__cond_resched_softirq);
4901 
4902 /**
4903  * yield - yield the current processor to other threads.
4904  *
4905  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4906  *
4907  * The scheduler is at all times free to pick the calling task as the most
4908  * eligible task to run, if removing the yield() call from your code breaks
4909  * it, its already broken.
4910  *
4911  * Typical broken usage is:
4912  *
4913  * while (!event)
4914  *	yield();
4915  *
4916  * where one assumes that yield() will let 'the other' process run that will
4917  * make event true. If the current task is a SCHED_FIFO task that will never
4918  * happen. Never use yield() as a progress guarantee!!
4919  *
4920  * If you want to use yield() to wait for something, use wait_event().
4921  * If you want to use yield() to be 'nice' for others, use cond_resched().
4922  * If you still want to use yield(), do not!
4923  */
4924 void __sched yield(void)
4925 {
4926 	set_current_state(TASK_RUNNING);
4927 	sys_sched_yield();
4928 }
4929 EXPORT_SYMBOL(yield);
4930 
4931 /**
4932  * yield_to - yield the current processor to another thread in
4933  * your thread group, or accelerate that thread toward the
4934  * processor it's on.
4935  * @p: target task
4936  * @preempt: whether task preemption is allowed or not
4937  *
4938  * It's the caller's job to ensure that the target task struct
4939  * can't go away on us before we can do any checks.
4940  *
4941  * Return:
4942  *	true (>0) if we indeed boosted the target task.
4943  *	false (0) if we failed to boost the target.
4944  *	-ESRCH if there's no task to yield to.
4945  */
4946 int __sched yield_to(struct task_struct *p, bool preempt)
4947 {
4948 	struct task_struct *curr = current;
4949 	struct rq *rq, *p_rq;
4950 	unsigned long flags;
4951 	int yielded = 0;
4952 
4953 	local_irq_save(flags);
4954 	rq = this_rq();
4955 
4956 again:
4957 	p_rq = task_rq(p);
4958 	/*
4959 	 * If we're the only runnable task on the rq and target rq also
4960 	 * has only one task, there's absolutely no point in yielding.
4961 	 */
4962 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4963 		yielded = -ESRCH;
4964 		goto out_irq;
4965 	}
4966 
4967 	double_rq_lock(rq, p_rq);
4968 	if (task_rq(p) != p_rq) {
4969 		double_rq_unlock(rq, p_rq);
4970 		goto again;
4971 	}
4972 
4973 	if (!curr->sched_class->yield_to_task)
4974 		goto out_unlock;
4975 
4976 	if (curr->sched_class != p->sched_class)
4977 		goto out_unlock;
4978 
4979 	if (task_running(p_rq, p) || p->state)
4980 		goto out_unlock;
4981 
4982 	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4983 	if (yielded) {
4984 		schedstat_inc(rq->yld_count);
4985 		/*
4986 		 * Make p's CPU reschedule; pick_next_entity takes care of
4987 		 * fairness.
4988 		 */
4989 		if (preempt && rq != p_rq)
4990 			resched_curr(p_rq);
4991 	}
4992 
4993 out_unlock:
4994 	double_rq_unlock(rq, p_rq);
4995 out_irq:
4996 	local_irq_restore(flags);
4997 
4998 	if (yielded > 0)
4999 		schedule();
5000 
5001 	return yielded;
5002 }
5003 EXPORT_SYMBOL_GPL(yield_to);
5004 
5005 int io_schedule_prepare(void)
5006 {
5007 	int old_iowait = current->in_iowait;
5008 
5009 	current->in_iowait = 1;
5010 	blk_schedule_flush_plug(current);
5011 
5012 	return old_iowait;
5013 }
5014 
5015 void io_schedule_finish(int token)
5016 {
5017 	current->in_iowait = token;
5018 }
5019 
5020 /*
5021  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5022  * that process accounting knows that this is a task in IO wait state.
5023  */
5024 long __sched io_schedule_timeout(long timeout)
5025 {
5026 	int token;
5027 	long ret;
5028 
5029 	token = io_schedule_prepare();
5030 	ret = schedule_timeout(timeout);
5031 	io_schedule_finish(token);
5032 
5033 	return ret;
5034 }
5035 EXPORT_SYMBOL(io_schedule_timeout);
5036 
5037 void io_schedule(void)
5038 {
5039 	int token;
5040 
5041 	token = io_schedule_prepare();
5042 	schedule();
5043 	io_schedule_finish(token);
5044 }
5045 EXPORT_SYMBOL(io_schedule);
5046 
5047 /**
5048  * sys_sched_get_priority_max - return maximum RT priority.
5049  * @policy: scheduling class.
5050  *
5051  * Return: On success, this syscall returns the maximum
5052  * rt_priority that can be used by a given scheduling class.
5053  * On failure, a negative error code is returned.
5054  */
5055 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5056 {
5057 	int ret = -EINVAL;
5058 
5059 	switch (policy) {
5060 	case SCHED_FIFO:
5061 	case SCHED_RR:
5062 		ret = MAX_USER_RT_PRIO-1;
5063 		break;
5064 	case SCHED_DEADLINE:
5065 	case SCHED_NORMAL:
5066 	case SCHED_BATCH:
5067 	case SCHED_IDLE:
5068 		ret = 0;
5069 		break;
5070 	}
5071 	return ret;
5072 }
5073 
5074 /**
5075  * sys_sched_get_priority_min - return minimum RT priority.
5076  * @policy: scheduling class.
5077  *
5078  * Return: On success, this syscall returns the minimum
5079  * rt_priority that can be used by a given scheduling class.
5080  * On failure, a negative error code is returned.
5081  */
5082 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5083 {
5084 	int ret = -EINVAL;
5085 
5086 	switch (policy) {
5087 	case SCHED_FIFO:
5088 	case SCHED_RR:
5089 		ret = 1;
5090 		break;
5091 	case SCHED_DEADLINE:
5092 	case SCHED_NORMAL:
5093 	case SCHED_BATCH:
5094 	case SCHED_IDLE:
5095 		ret = 0;
5096 	}
5097 	return ret;
5098 }
5099 
5100 /**
5101  * sys_sched_rr_get_interval - return the default timeslice of a process.
5102  * @pid: pid of the process.
5103  * @interval: userspace pointer to the timeslice value.
5104  *
5105  * this syscall writes the default timeslice value of a given process
5106  * into the user-space timespec buffer. A value of '0' means infinity.
5107  *
5108  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5109  * an error code.
5110  */
5111 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5112 {
5113 	struct task_struct *p;
5114 	unsigned int time_slice;
5115 	struct rq_flags rf;
5116 	struct rq *rq;
5117 	int retval;
5118 
5119 	if (pid < 0)
5120 		return -EINVAL;
5121 
5122 	retval = -ESRCH;
5123 	rcu_read_lock();
5124 	p = find_process_by_pid(pid);
5125 	if (!p)
5126 		goto out_unlock;
5127 
5128 	retval = security_task_getscheduler(p);
5129 	if (retval)
5130 		goto out_unlock;
5131 
5132 	rq = task_rq_lock(p, &rf);
5133 	time_slice = 0;
5134 	if (p->sched_class->get_rr_interval)
5135 		time_slice = p->sched_class->get_rr_interval(rq, p);
5136 	task_rq_unlock(rq, p, &rf);
5137 
5138 	rcu_read_unlock();
5139 	jiffies_to_timespec64(time_slice, t);
5140 	return 0;
5141 
5142 out_unlock:
5143 	rcu_read_unlock();
5144 	return retval;
5145 }
5146 
5147 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5148 		struct timespec __user *, interval)
5149 {
5150 	struct timespec64 t;
5151 	int retval = sched_rr_get_interval(pid, &t);
5152 
5153 	if (retval == 0)
5154 		retval = put_timespec64(&t, interval);
5155 
5156 	return retval;
5157 }
5158 
5159 #ifdef CONFIG_COMPAT
5160 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5161 		       compat_pid_t, pid,
5162 		       struct compat_timespec __user *, interval)
5163 {
5164 	struct timespec64 t;
5165 	int retval = sched_rr_get_interval(pid, &t);
5166 
5167 	if (retval == 0)
5168 		retval = compat_put_timespec64(&t, interval);
5169 	return retval;
5170 }
5171 #endif
5172 
5173 void sched_show_task(struct task_struct *p)
5174 {
5175 	unsigned long free = 0;
5176 	int ppid;
5177 
5178 	if (!try_get_task_stack(p))
5179 		return;
5180 
5181 	printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5182 
5183 	if (p->state == TASK_RUNNING)
5184 		printk(KERN_CONT "  running task    ");
5185 #ifdef CONFIG_DEBUG_STACK_USAGE
5186 	free = stack_not_used(p);
5187 #endif
5188 	ppid = 0;
5189 	rcu_read_lock();
5190 	if (pid_alive(p))
5191 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
5192 	rcu_read_unlock();
5193 	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5194 		task_pid_nr(p), ppid,
5195 		(unsigned long)task_thread_info(p)->flags);
5196 
5197 	print_worker_info(KERN_INFO, p);
5198 	show_stack(p, NULL);
5199 	put_task_stack(p);
5200 }
5201 EXPORT_SYMBOL_GPL(sched_show_task);
5202 
5203 static inline bool
5204 state_filter_match(unsigned long state_filter, struct task_struct *p)
5205 {
5206 	/* no filter, everything matches */
5207 	if (!state_filter)
5208 		return true;
5209 
5210 	/* filter, but doesn't match */
5211 	if (!(p->state & state_filter))
5212 		return false;
5213 
5214 	/*
5215 	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5216 	 * TASK_KILLABLE).
5217 	 */
5218 	if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5219 		return false;
5220 
5221 	return true;
5222 }
5223 
5224 
5225 void show_state_filter(unsigned long state_filter)
5226 {
5227 	struct task_struct *g, *p;
5228 
5229 #if BITS_PER_LONG == 32
5230 	printk(KERN_INFO
5231 		"  task                PC stack   pid father\n");
5232 #else
5233 	printk(KERN_INFO
5234 		"  task                        PC stack   pid father\n");
5235 #endif
5236 	rcu_read_lock();
5237 	for_each_process_thread(g, p) {
5238 		/*
5239 		 * reset the NMI-timeout, listing all files on a slow
5240 		 * console might take a lot of time:
5241 		 * Also, reset softlockup watchdogs on all CPUs, because
5242 		 * another CPU might be blocked waiting for us to process
5243 		 * an IPI.
5244 		 */
5245 		touch_nmi_watchdog();
5246 		touch_all_softlockup_watchdogs();
5247 		if (state_filter_match(state_filter, p))
5248 			sched_show_task(p);
5249 	}
5250 
5251 #ifdef CONFIG_SCHED_DEBUG
5252 	if (!state_filter)
5253 		sysrq_sched_debug_show();
5254 #endif
5255 	rcu_read_unlock();
5256 	/*
5257 	 * Only show locks if all tasks are dumped:
5258 	 */
5259 	if (!state_filter)
5260 		debug_show_all_locks();
5261 }
5262 
5263 /**
5264  * init_idle - set up an idle thread for a given CPU
5265  * @idle: task in question
5266  * @cpu: CPU the idle task belongs to
5267  *
5268  * NOTE: this function does not set the idle thread's NEED_RESCHED
5269  * flag, to make booting more robust.
5270  */
5271 void init_idle(struct task_struct *idle, int cpu)
5272 {
5273 	struct rq *rq = cpu_rq(cpu);
5274 	unsigned long flags;
5275 
5276 	raw_spin_lock_irqsave(&idle->pi_lock, flags);
5277 	raw_spin_lock(&rq->lock);
5278 
5279 	__sched_fork(0, idle);
5280 	idle->state = TASK_RUNNING;
5281 	idle->se.exec_start = sched_clock();
5282 	idle->flags |= PF_IDLE;
5283 
5284 	kasan_unpoison_task_stack(idle);
5285 
5286 #ifdef CONFIG_SMP
5287 	/*
5288 	 * Its possible that init_idle() gets called multiple times on a task,
5289 	 * in that case do_set_cpus_allowed() will not do the right thing.
5290 	 *
5291 	 * And since this is boot we can forgo the serialization.
5292 	 */
5293 	set_cpus_allowed_common(idle, cpumask_of(cpu));
5294 #endif
5295 	/*
5296 	 * We're having a chicken and egg problem, even though we are
5297 	 * holding rq->lock, the CPU isn't yet set to this CPU so the
5298 	 * lockdep check in task_group() will fail.
5299 	 *
5300 	 * Similar case to sched_fork(). / Alternatively we could
5301 	 * use task_rq_lock() here and obtain the other rq->lock.
5302 	 *
5303 	 * Silence PROVE_RCU
5304 	 */
5305 	rcu_read_lock();
5306 	__set_task_cpu(idle, cpu);
5307 	rcu_read_unlock();
5308 
5309 	rq->curr = rq->idle = idle;
5310 	idle->on_rq = TASK_ON_RQ_QUEUED;
5311 #ifdef CONFIG_SMP
5312 	idle->on_cpu = 1;
5313 #endif
5314 	raw_spin_unlock(&rq->lock);
5315 	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5316 
5317 	/* Set the preempt count _outside_ the spinlocks! */
5318 	init_idle_preempt_count(idle, cpu);
5319 
5320 	/*
5321 	 * The idle tasks have their own, simple scheduling class:
5322 	 */
5323 	idle->sched_class = &idle_sched_class;
5324 	ftrace_graph_init_idle_task(idle, cpu);
5325 	vtime_init_idle(idle, cpu);
5326 #ifdef CONFIG_SMP
5327 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5328 #endif
5329 }
5330 
5331 #ifdef CONFIG_SMP
5332 
5333 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5334 			      const struct cpumask *trial)
5335 {
5336 	int ret = 1;
5337 
5338 	if (!cpumask_weight(cur))
5339 		return ret;
5340 
5341 	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5342 
5343 	return ret;
5344 }
5345 
5346 int task_can_attach(struct task_struct *p,
5347 		    const struct cpumask *cs_cpus_allowed)
5348 {
5349 	int ret = 0;
5350 
5351 	/*
5352 	 * Kthreads which disallow setaffinity shouldn't be moved
5353 	 * to a new cpuset; we don't want to change their CPU
5354 	 * affinity and isolating such threads by their set of
5355 	 * allowed nodes is unnecessary.  Thus, cpusets are not
5356 	 * applicable for such threads.  This prevents checking for
5357 	 * success of set_cpus_allowed_ptr() on all attached tasks
5358 	 * before cpus_allowed may be changed.
5359 	 */
5360 	if (p->flags & PF_NO_SETAFFINITY) {
5361 		ret = -EINVAL;
5362 		goto out;
5363 	}
5364 
5365 	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5366 					      cs_cpus_allowed))
5367 		ret = dl_task_can_attach(p, cs_cpus_allowed);
5368 
5369 out:
5370 	return ret;
5371 }
5372 
5373 bool sched_smp_initialized __read_mostly;
5374 
5375 #ifdef CONFIG_NUMA_BALANCING
5376 /* Migrate current task p to target_cpu */
5377 int migrate_task_to(struct task_struct *p, int target_cpu)
5378 {
5379 	struct migration_arg arg = { p, target_cpu };
5380 	int curr_cpu = task_cpu(p);
5381 
5382 	if (curr_cpu == target_cpu)
5383 		return 0;
5384 
5385 	if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5386 		return -EINVAL;
5387 
5388 	/* TODO: This is not properly updating schedstats */
5389 
5390 	trace_sched_move_numa(p, curr_cpu, target_cpu);
5391 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5392 }
5393 
5394 /*
5395  * Requeue a task on a given node and accurately track the number of NUMA
5396  * tasks on the runqueues
5397  */
5398 void sched_setnuma(struct task_struct *p, int nid)
5399 {
5400 	bool queued, running;
5401 	struct rq_flags rf;
5402 	struct rq *rq;
5403 
5404 	rq = task_rq_lock(p, &rf);
5405 	queued = task_on_rq_queued(p);
5406 	running = task_current(rq, p);
5407 
5408 	if (queued)
5409 		dequeue_task(rq, p, DEQUEUE_SAVE);
5410 	if (running)
5411 		put_prev_task(rq, p);
5412 
5413 	p->numa_preferred_nid = nid;
5414 
5415 	if (queued)
5416 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5417 	if (running)
5418 		set_curr_task(rq, p);
5419 	task_rq_unlock(rq, p, &rf);
5420 }
5421 #endif /* CONFIG_NUMA_BALANCING */
5422 
5423 #ifdef CONFIG_HOTPLUG_CPU
5424 /*
5425  * Ensure that the idle task is using init_mm right before its CPU goes
5426  * offline.
5427  */
5428 void idle_task_exit(void)
5429 {
5430 	struct mm_struct *mm = current->active_mm;
5431 
5432 	BUG_ON(cpu_online(smp_processor_id()));
5433 
5434 	if (mm != &init_mm) {
5435 		switch_mm(mm, &init_mm, current);
5436 		finish_arch_post_lock_switch();
5437 	}
5438 	mmdrop(mm);
5439 }
5440 
5441 /*
5442  * Since this CPU is going 'away' for a while, fold any nr_active delta
5443  * we might have. Assumes we're called after migrate_tasks() so that the
5444  * nr_active count is stable. We need to take the teardown thread which
5445  * is calling this into account, so we hand in adjust = 1 to the load
5446  * calculation.
5447  *
5448  * Also see the comment "Global load-average calculations".
5449  */
5450 static void calc_load_migrate(struct rq *rq)
5451 {
5452 	long delta = calc_load_fold_active(rq, 1);
5453 	if (delta)
5454 		atomic_long_add(delta, &calc_load_tasks);
5455 }
5456 
5457 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5458 {
5459 }
5460 
5461 static const struct sched_class fake_sched_class = {
5462 	.put_prev_task = put_prev_task_fake,
5463 };
5464 
5465 static struct task_struct fake_task = {
5466 	/*
5467 	 * Avoid pull_{rt,dl}_task()
5468 	 */
5469 	.prio = MAX_PRIO + 1,
5470 	.sched_class = &fake_sched_class,
5471 };
5472 
5473 /*
5474  * Migrate all tasks from the rq, sleeping tasks will be migrated by
5475  * try_to_wake_up()->select_task_rq().
5476  *
5477  * Called with rq->lock held even though we'er in stop_machine() and
5478  * there's no concurrency possible, we hold the required locks anyway
5479  * because of lock validation efforts.
5480  */
5481 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5482 {
5483 	struct rq *rq = dead_rq;
5484 	struct task_struct *next, *stop = rq->stop;
5485 	struct rq_flags orf = *rf;
5486 	int dest_cpu;
5487 
5488 	/*
5489 	 * Fudge the rq selection such that the below task selection loop
5490 	 * doesn't get stuck on the currently eligible stop task.
5491 	 *
5492 	 * We're currently inside stop_machine() and the rq is either stuck
5493 	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5494 	 * either way we should never end up calling schedule() until we're
5495 	 * done here.
5496 	 */
5497 	rq->stop = NULL;
5498 
5499 	/*
5500 	 * put_prev_task() and pick_next_task() sched
5501 	 * class method both need to have an up-to-date
5502 	 * value of rq->clock[_task]
5503 	 */
5504 	update_rq_clock(rq);
5505 
5506 	for (;;) {
5507 		/*
5508 		 * There's this thread running, bail when that's the only
5509 		 * remaining thread:
5510 		 */
5511 		if (rq->nr_running == 1)
5512 			break;
5513 
5514 		/*
5515 		 * pick_next_task() assumes pinned rq->lock:
5516 		 */
5517 		next = pick_next_task(rq, &fake_task, rf);
5518 		BUG_ON(!next);
5519 		put_prev_task(rq, next);
5520 
5521 		/*
5522 		 * Rules for changing task_struct::cpus_allowed are holding
5523 		 * both pi_lock and rq->lock, such that holding either
5524 		 * stabilizes the mask.
5525 		 *
5526 		 * Drop rq->lock is not quite as disastrous as it usually is
5527 		 * because !cpu_active at this point, which means load-balance
5528 		 * will not interfere. Also, stop-machine.
5529 		 */
5530 		rq_unlock(rq, rf);
5531 		raw_spin_lock(&next->pi_lock);
5532 		rq_relock(rq, rf);
5533 
5534 		/*
5535 		 * Since we're inside stop-machine, _nothing_ should have
5536 		 * changed the task, WARN if weird stuff happened, because in
5537 		 * that case the above rq->lock drop is a fail too.
5538 		 */
5539 		if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5540 			raw_spin_unlock(&next->pi_lock);
5541 			continue;
5542 		}
5543 
5544 		/* Find suitable destination for @next, with force if needed. */
5545 		dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5546 		rq = __migrate_task(rq, rf, next, dest_cpu);
5547 		if (rq != dead_rq) {
5548 			rq_unlock(rq, rf);
5549 			rq = dead_rq;
5550 			*rf = orf;
5551 			rq_relock(rq, rf);
5552 		}
5553 		raw_spin_unlock(&next->pi_lock);
5554 	}
5555 
5556 	rq->stop = stop;
5557 }
5558 #endif /* CONFIG_HOTPLUG_CPU */
5559 
5560 void set_rq_online(struct rq *rq)
5561 {
5562 	if (!rq->online) {
5563 		const struct sched_class *class;
5564 
5565 		cpumask_set_cpu(rq->cpu, rq->rd->online);
5566 		rq->online = 1;
5567 
5568 		for_each_class(class) {
5569 			if (class->rq_online)
5570 				class->rq_online(rq);
5571 		}
5572 	}
5573 }
5574 
5575 void set_rq_offline(struct rq *rq)
5576 {
5577 	if (rq->online) {
5578 		const struct sched_class *class;
5579 
5580 		for_each_class(class) {
5581 			if (class->rq_offline)
5582 				class->rq_offline(rq);
5583 		}
5584 
5585 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5586 		rq->online = 0;
5587 	}
5588 }
5589 
5590 static void set_cpu_rq_start_time(unsigned int cpu)
5591 {
5592 	struct rq *rq = cpu_rq(cpu);
5593 
5594 	rq->age_stamp = sched_clock_cpu(cpu);
5595 }
5596 
5597 /*
5598  * used to mark begin/end of suspend/resume:
5599  */
5600 static int num_cpus_frozen;
5601 
5602 /*
5603  * Update cpusets according to cpu_active mask.  If cpusets are
5604  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5605  * around partition_sched_domains().
5606  *
5607  * If we come here as part of a suspend/resume, don't touch cpusets because we
5608  * want to restore it back to its original state upon resume anyway.
5609  */
5610 static void cpuset_cpu_active(void)
5611 {
5612 	if (cpuhp_tasks_frozen) {
5613 		/*
5614 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
5615 		 * resume sequence. As long as this is not the last online
5616 		 * operation in the resume sequence, just build a single sched
5617 		 * domain, ignoring cpusets.
5618 		 */
5619 		partition_sched_domains(1, NULL, NULL);
5620 		if (--num_cpus_frozen)
5621 			return;
5622 		/*
5623 		 * This is the last CPU online operation. So fall through and
5624 		 * restore the original sched domains by considering the
5625 		 * cpuset configurations.
5626 		 */
5627 		cpuset_force_rebuild();
5628 	}
5629 	cpuset_update_active_cpus();
5630 }
5631 
5632 static int cpuset_cpu_inactive(unsigned int cpu)
5633 {
5634 	if (!cpuhp_tasks_frozen) {
5635 		if (dl_cpu_busy(cpu))
5636 			return -EBUSY;
5637 		cpuset_update_active_cpus();
5638 	} else {
5639 		num_cpus_frozen++;
5640 		partition_sched_domains(1, NULL, NULL);
5641 	}
5642 	return 0;
5643 }
5644 
5645 int sched_cpu_activate(unsigned int cpu)
5646 {
5647 	struct rq *rq = cpu_rq(cpu);
5648 	struct rq_flags rf;
5649 
5650 	set_cpu_active(cpu, true);
5651 
5652 	if (sched_smp_initialized) {
5653 		sched_domains_numa_masks_set(cpu);
5654 		cpuset_cpu_active();
5655 	}
5656 
5657 	/*
5658 	 * Put the rq online, if not already. This happens:
5659 	 *
5660 	 * 1) In the early boot process, because we build the real domains
5661 	 *    after all CPUs have been brought up.
5662 	 *
5663 	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5664 	 *    domains.
5665 	 */
5666 	rq_lock_irqsave(rq, &rf);
5667 	if (rq->rd) {
5668 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5669 		set_rq_online(rq);
5670 	}
5671 	rq_unlock_irqrestore(rq, &rf);
5672 
5673 	update_max_interval();
5674 
5675 	return 0;
5676 }
5677 
5678 int sched_cpu_deactivate(unsigned int cpu)
5679 {
5680 	int ret;
5681 
5682 	set_cpu_active(cpu, false);
5683 	/*
5684 	 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5685 	 * users of this state to go away such that all new such users will
5686 	 * observe it.
5687 	 *
5688 	 * Do sync before park smpboot threads to take care the rcu boost case.
5689 	 */
5690 	synchronize_rcu_mult(call_rcu, call_rcu_sched);
5691 
5692 	if (!sched_smp_initialized)
5693 		return 0;
5694 
5695 	ret = cpuset_cpu_inactive(cpu);
5696 	if (ret) {
5697 		set_cpu_active(cpu, true);
5698 		return ret;
5699 	}
5700 	sched_domains_numa_masks_clear(cpu);
5701 	return 0;
5702 }
5703 
5704 static void sched_rq_cpu_starting(unsigned int cpu)
5705 {
5706 	struct rq *rq = cpu_rq(cpu);
5707 
5708 	rq->calc_load_update = calc_load_update;
5709 	update_max_interval();
5710 }
5711 
5712 int sched_cpu_starting(unsigned int cpu)
5713 {
5714 	set_cpu_rq_start_time(cpu);
5715 	sched_rq_cpu_starting(cpu);
5716 	return 0;
5717 }
5718 
5719 #ifdef CONFIG_HOTPLUG_CPU
5720 int sched_cpu_dying(unsigned int cpu)
5721 {
5722 	struct rq *rq = cpu_rq(cpu);
5723 	struct rq_flags rf;
5724 
5725 	/* Handle pending wakeups and then migrate everything off */
5726 	sched_ttwu_pending();
5727 
5728 	rq_lock_irqsave(rq, &rf);
5729 	if (rq->rd) {
5730 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5731 		set_rq_offline(rq);
5732 	}
5733 	migrate_tasks(rq, &rf);
5734 	BUG_ON(rq->nr_running != 1);
5735 	rq_unlock_irqrestore(rq, &rf);
5736 
5737 	calc_load_migrate(rq);
5738 	update_max_interval();
5739 	nohz_balance_exit_idle(cpu);
5740 	hrtick_clear(rq);
5741 	return 0;
5742 }
5743 #endif
5744 
5745 #ifdef CONFIG_SCHED_SMT
5746 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5747 
5748 static void sched_init_smt(void)
5749 {
5750 	/*
5751 	 * We've enumerated all CPUs and will assume that if any CPU
5752 	 * has SMT siblings, CPU0 will too.
5753 	 */
5754 	if (cpumask_weight(cpu_smt_mask(0)) > 1)
5755 		static_branch_enable(&sched_smt_present);
5756 }
5757 #else
5758 static inline void sched_init_smt(void) { }
5759 #endif
5760 
5761 void __init sched_init_smp(void)
5762 {
5763 	sched_init_numa();
5764 
5765 	/*
5766 	 * There's no userspace yet to cause hotplug operations; hence all the
5767 	 * CPU masks are stable and all blatant races in the below code cannot
5768 	 * happen.
5769 	 */
5770 	mutex_lock(&sched_domains_mutex);
5771 	sched_init_domains(cpu_active_mask);
5772 	mutex_unlock(&sched_domains_mutex);
5773 
5774 	/* Move init over to a non-isolated CPU */
5775 	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5776 		BUG();
5777 	sched_init_granularity();
5778 
5779 	init_sched_rt_class();
5780 	init_sched_dl_class();
5781 
5782 	sched_init_smt();
5783 
5784 	sched_smp_initialized = true;
5785 }
5786 
5787 static int __init migration_init(void)
5788 {
5789 	sched_rq_cpu_starting(smp_processor_id());
5790 	return 0;
5791 }
5792 early_initcall(migration_init);
5793 
5794 #else
5795 void __init sched_init_smp(void)
5796 {
5797 	sched_init_granularity();
5798 }
5799 #endif /* CONFIG_SMP */
5800 
5801 int in_sched_functions(unsigned long addr)
5802 {
5803 	return in_lock_functions(addr) ||
5804 		(addr >= (unsigned long)__sched_text_start
5805 		&& addr < (unsigned long)__sched_text_end);
5806 }
5807 
5808 #ifdef CONFIG_CGROUP_SCHED
5809 /*
5810  * Default task group.
5811  * Every task in system belongs to this group at bootup.
5812  */
5813 struct task_group root_task_group;
5814 LIST_HEAD(task_groups);
5815 
5816 /* Cacheline aligned slab cache for task_group */
5817 static struct kmem_cache *task_group_cache __read_mostly;
5818 #endif
5819 
5820 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5821 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5822 
5823 void __init sched_init(void)
5824 {
5825 	int i, j;
5826 	unsigned long alloc_size = 0, ptr;
5827 
5828 	sched_clock_init();
5829 	wait_bit_init();
5830 
5831 #ifdef CONFIG_FAIR_GROUP_SCHED
5832 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5833 #endif
5834 #ifdef CONFIG_RT_GROUP_SCHED
5835 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5836 #endif
5837 	if (alloc_size) {
5838 		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5839 
5840 #ifdef CONFIG_FAIR_GROUP_SCHED
5841 		root_task_group.se = (struct sched_entity **)ptr;
5842 		ptr += nr_cpu_ids * sizeof(void **);
5843 
5844 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5845 		ptr += nr_cpu_ids * sizeof(void **);
5846 
5847 #endif /* CONFIG_FAIR_GROUP_SCHED */
5848 #ifdef CONFIG_RT_GROUP_SCHED
5849 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5850 		ptr += nr_cpu_ids * sizeof(void **);
5851 
5852 		root_task_group.rt_rq = (struct rt_rq **)ptr;
5853 		ptr += nr_cpu_ids * sizeof(void **);
5854 
5855 #endif /* CONFIG_RT_GROUP_SCHED */
5856 	}
5857 #ifdef CONFIG_CPUMASK_OFFSTACK
5858 	for_each_possible_cpu(i) {
5859 		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5860 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5861 		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5862 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5863 	}
5864 #endif /* CONFIG_CPUMASK_OFFSTACK */
5865 
5866 	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5867 	init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5868 
5869 #ifdef CONFIG_SMP
5870 	init_defrootdomain();
5871 #endif
5872 
5873 #ifdef CONFIG_RT_GROUP_SCHED
5874 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
5875 			global_rt_period(), global_rt_runtime());
5876 #endif /* CONFIG_RT_GROUP_SCHED */
5877 
5878 #ifdef CONFIG_CGROUP_SCHED
5879 	task_group_cache = KMEM_CACHE(task_group, 0);
5880 
5881 	list_add(&root_task_group.list, &task_groups);
5882 	INIT_LIST_HEAD(&root_task_group.children);
5883 	INIT_LIST_HEAD(&root_task_group.siblings);
5884 	autogroup_init(&init_task);
5885 #endif /* CONFIG_CGROUP_SCHED */
5886 
5887 	for_each_possible_cpu(i) {
5888 		struct rq *rq;
5889 
5890 		rq = cpu_rq(i);
5891 		raw_spin_lock_init(&rq->lock);
5892 		rq->nr_running = 0;
5893 		rq->calc_load_active = 0;
5894 		rq->calc_load_update = jiffies + LOAD_FREQ;
5895 		init_cfs_rq(&rq->cfs);
5896 		init_rt_rq(&rq->rt);
5897 		init_dl_rq(&rq->dl);
5898 #ifdef CONFIG_FAIR_GROUP_SCHED
5899 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5900 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5901 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5902 		/*
5903 		 * How much CPU bandwidth does root_task_group get?
5904 		 *
5905 		 * In case of task-groups formed thr' the cgroup filesystem, it
5906 		 * gets 100% of the CPU resources in the system. This overall
5907 		 * system CPU resource is divided among the tasks of
5908 		 * root_task_group and its child task-groups in a fair manner,
5909 		 * based on each entity's (task or task-group's) weight
5910 		 * (se->load.weight).
5911 		 *
5912 		 * In other words, if root_task_group has 10 tasks of weight
5913 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5914 		 * then A0's share of the CPU resource is:
5915 		 *
5916 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5917 		 *
5918 		 * We achieve this by letting root_task_group's tasks sit
5919 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5920 		 */
5921 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
5922 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
5923 #endif /* CONFIG_FAIR_GROUP_SCHED */
5924 
5925 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
5926 #ifdef CONFIG_RT_GROUP_SCHED
5927 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
5928 #endif
5929 
5930 		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
5931 			rq->cpu_load[j] = 0;
5932 
5933 #ifdef CONFIG_SMP
5934 		rq->sd = NULL;
5935 		rq->rd = NULL;
5936 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
5937 		rq->balance_callback = NULL;
5938 		rq->active_balance = 0;
5939 		rq->next_balance = jiffies;
5940 		rq->push_cpu = 0;
5941 		rq->cpu = i;
5942 		rq->online = 0;
5943 		rq->idle_stamp = 0;
5944 		rq->avg_idle = 2*sysctl_sched_migration_cost;
5945 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
5946 
5947 		INIT_LIST_HEAD(&rq->cfs_tasks);
5948 
5949 		rq_attach_root(rq, &def_root_domain);
5950 #ifdef CONFIG_NO_HZ_COMMON
5951 		rq->last_load_update_tick = jiffies;
5952 		rq->nohz_flags = 0;
5953 #endif
5954 #ifdef CONFIG_NO_HZ_FULL
5955 		rq->last_sched_tick = 0;
5956 #endif
5957 #endif /* CONFIG_SMP */
5958 		init_rq_hrtick(rq);
5959 		atomic_set(&rq->nr_iowait, 0);
5960 	}
5961 
5962 	set_load_weight(&init_task, false);
5963 
5964 	/*
5965 	 * The boot idle thread does lazy MMU switching as well:
5966 	 */
5967 	mmgrab(&init_mm);
5968 	enter_lazy_tlb(&init_mm, current);
5969 
5970 	/*
5971 	 * Make us the idle thread. Technically, schedule() should not be
5972 	 * called from this thread, however somewhere below it might be,
5973 	 * but because we are the idle thread, we just pick up running again
5974 	 * when this runqueue becomes "idle".
5975 	 */
5976 	init_idle(current, smp_processor_id());
5977 
5978 	calc_load_update = jiffies + LOAD_FREQ;
5979 
5980 #ifdef CONFIG_SMP
5981 	idle_thread_set_boot_cpu();
5982 	set_cpu_rq_start_time(smp_processor_id());
5983 #endif
5984 	init_sched_fair_class();
5985 
5986 	init_schedstats();
5987 
5988 	scheduler_running = 1;
5989 }
5990 
5991 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5992 static inline int preempt_count_equals(int preempt_offset)
5993 {
5994 	int nested = preempt_count() + rcu_preempt_depth();
5995 
5996 	return (nested == preempt_offset);
5997 }
5998 
5999 void __might_sleep(const char *file, int line, int preempt_offset)
6000 {
6001 	/*
6002 	 * Blocking primitives will set (and therefore destroy) current->state,
6003 	 * since we will exit with TASK_RUNNING make sure we enter with it,
6004 	 * otherwise we will destroy state.
6005 	 */
6006 	WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6007 			"do not call blocking ops when !TASK_RUNNING; "
6008 			"state=%lx set at [<%p>] %pS\n",
6009 			current->state,
6010 			(void *)current->task_state_change,
6011 			(void *)current->task_state_change);
6012 
6013 	___might_sleep(file, line, preempt_offset);
6014 }
6015 EXPORT_SYMBOL(__might_sleep);
6016 
6017 void ___might_sleep(const char *file, int line, int preempt_offset)
6018 {
6019 	/* Ratelimiting timestamp: */
6020 	static unsigned long prev_jiffy;
6021 
6022 	unsigned long preempt_disable_ip;
6023 
6024 	/* WARN_ON_ONCE() by default, no rate limit required: */
6025 	rcu_sleep_check();
6026 
6027 	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6028 	     !is_idle_task(current)) ||
6029 	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6030 	    oops_in_progress)
6031 		return;
6032 
6033 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6034 		return;
6035 	prev_jiffy = jiffies;
6036 
6037 	/* Save this before calling printk(), since that will clobber it: */
6038 	preempt_disable_ip = get_preempt_disable_ip(current);
6039 
6040 	printk(KERN_ERR
6041 		"BUG: sleeping function called from invalid context at %s:%d\n",
6042 			file, line);
6043 	printk(KERN_ERR
6044 		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6045 			in_atomic(), irqs_disabled(),
6046 			current->pid, current->comm);
6047 
6048 	if (task_stack_end_corrupted(current))
6049 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6050 
6051 	debug_show_held_locks(current);
6052 	if (irqs_disabled())
6053 		print_irqtrace_events(current);
6054 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6055 	    && !preempt_count_equals(preempt_offset)) {
6056 		pr_err("Preemption disabled at:");
6057 		print_ip_sym(preempt_disable_ip);
6058 		pr_cont("\n");
6059 	}
6060 	dump_stack();
6061 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6062 }
6063 EXPORT_SYMBOL(___might_sleep);
6064 #endif
6065 
6066 #ifdef CONFIG_MAGIC_SYSRQ
6067 void normalize_rt_tasks(void)
6068 {
6069 	struct task_struct *g, *p;
6070 	struct sched_attr attr = {
6071 		.sched_policy = SCHED_NORMAL,
6072 	};
6073 
6074 	read_lock(&tasklist_lock);
6075 	for_each_process_thread(g, p) {
6076 		/*
6077 		 * Only normalize user tasks:
6078 		 */
6079 		if (p->flags & PF_KTHREAD)
6080 			continue;
6081 
6082 		p->se.exec_start = 0;
6083 		schedstat_set(p->se.statistics.wait_start,  0);
6084 		schedstat_set(p->se.statistics.sleep_start, 0);
6085 		schedstat_set(p->se.statistics.block_start, 0);
6086 
6087 		if (!dl_task(p) && !rt_task(p)) {
6088 			/*
6089 			 * Renice negative nice level userspace
6090 			 * tasks back to 0:
6091 			 */
6092 			if (task_nice(p) < 0)
6093 				set_user_nice(p, 0);
6094 			continue;
6095 		}
6096 
6097 		__sched_setscheduler(p, &attr, false, false);
6098 	}
6099 	read_unlock(&tasklist_lock);
6100 }
6101 
6102 #endif /* CONFIG_MAGIC_SYSRQ */
6103 
6104 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6105 /*
6106  * These functions are only useful for the IA64 MCA handling, or kdb.
6107  *
6108  * They can only be called when the whole system has been
6109  * stopped - every CPU needs to be quiescent, and no scheduling
6110  * activity can take place. Using them for anything else would
6111  * be a serious bug, and as a result, they aren't even visible
6112  * under any other configuration.
6113  */
6114 
6115 /**
6116  * curr_task - return the current task for a given CPU.
6117  * @cpu: the processor in question.
6118  *
6119  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6120  *
6121  * Return: The current task for @cpu.
6122  */
6123 struct task_struct *curr_task(int cpu)
6124 {
6125 	return cpu_curr(cpu);
6126 }
6127 
6128 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6129 
6130 #ifdef CONFIG_IA64
6131 /**
6132  * set_curr_task - set the current task for a given CPU.
6133  * @cpu: the processor in question.
6134  * @p: the task pointer to set.
6135  *
6136  * Description: This function must only be used when non-maskable interrupts
6137  * are serviced on a separate stack. It allows the architecture to switch the
6138  * notion of the current task on a CPU in a non-blocking manner. This function
6139  * must be called with all CPU's synchronized, and interrupts disabled, the
6140  * and caller must save the original value of the current task (see
6141  * curr_task() above) and restore that value before reenabling interrupts and
6142  * re-starting the system.
6143  *
6144  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6145  */
6146 void ia64_set_curr_task(int cpu, struct task_struct *p)
6147 {
6148 	cpu_curr(cpu) = p;
6149 }
6150 
6151 #endif
6152 
6153 #ifdef CONFIG_CGROUP_SCHED
6154 /* task_group_lock serializes the addition/removal of task groups */
6155 static DEFINE_SPINLOCK(task_group_lock);
6156 
6157 static void sched_free_group(struct task_group *tg)
6158 {
6159 	free_fair_sched_group(tg);
6160 	free_rt_sched_group(tg);
6161 	autogroup_free(tg);
6162 	kmem_cache_free(task_group_cache, tg);
6163 }
6164 
6165 /* allocate runqueue etc for a new task group */
6166 struct task_group *sched_create_group(struct task_group *parent)
6167 {
6168 	struct task_group *tg;
6169 
6170 	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6171 	if (!tg)
6172 		return ERR_PTR(-ENOMEM);
6173 
6174 	if (!alloc_fair_sched_group(tg, parent))
6175 		goto err;
6176 
6177 	if (!alloc_rt_sched_group(tg, parent))
6178 		goto err;
6179 
6180 	return tg;
6181 
6182 err:
6183 	sched_free_group(tg);
6184 	return ERR_PTR(-ENOMEM);
6185 }
6186 
6187 void sched_online_group(struct task_group *tg, struct task_group *parent)
6188 {
6189 	unsigned long flags;
6190 
6191 	spin_lock_irqsave(&task_group_lock, flags);
6192 	list_add_rcu(&tg->list, &task_groups);
6193 
6194 	/* Root should already exist: */
6195 	WARN_ON(!parent);
6196 
6197 	tg->parent = parent;
6198 	INIT_LIST_HEAD(&tg->children);
6199 	list_add_rcu(&tg->siblings, &parent->children);
6200 	spin_unlock_irqrestore(&task_group_lock, flags);
6201 
6202 	online_fair_sched_group(tg);
6203 }
6204 
6205 /* rcu callback to free various structures associated with a task group */
6206 static void sched_free_group_rcu(struct rcu_head *rhp)
6207 {
6208 	/* Now it should be safe to free those cfs_rqs: */
6209 	sched_free_group(container_of(rhp, struct task_group, rcu));
6210 }
6211 
6212 void sched_destroy_group(struct task_group *tg)
6213 {
6214 	/* Wait for possible concurrent references to cfs_rqs complete: */
6215 	call_rcu(&tg->rcu, sched_free_group_rcu);
6216 }
6217 
6218 void sched_offline_group(struct task_group *tg)
6219 {
6220 	unsigned long flags;
6221 
6222 	/* End participation in shares distribution: */
6223 	unregister_fair_sched_group(tg);
6224 
6225 	spin_lock_irqsave(&task_group_lock, flags);
6226 	list_del_rcu(&tg->list);
6227 	list_del_rcu(&tg->siblings);
6228 	spin_unlock_irqrestore(&task_group_lock, flags);
6229 }
6230 
6231 static void sched_change_group(struct task_struct *tsk, int type)
6232 {
6233 	struct task_group *tg;
6234 
6235 	/*
6236 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
6237 	 * which is pointless here. Thus, we pass "true" to task_css_check()
6238 	 * to prevent lockdep warnings.
6239 	 */
6240 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6241 			  struct task_group, css);
6242 	tg = autogroup_task_group(tsk, tg);
6243 	tsk->sched_task_group = tg;
6244 
6245 #ifdef CONFIG_FAIR_GROUP_SCHED
6246 	if (tsk->sched_class->task_change_group)
6247 		tsk->sched_class->task_change_group(tsk, type);
6248 	else
6249 #endif
6250 		set_task_rq(tsk, task_cpu(tsk));
6251 }
6252 
6253 /*
6254  * Change task's runqueue when it moves between groups.
6255  *
6256  * The caller of this function should have put the task in its new group by
6257  * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6258  * its new group.
6259  */
6260 void sched_move_task(struct task_struct *tsk)
6261 {
6262 	int queued, running, queue_flags =
6263 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6264 	struct rq_flags rf;
6265 	struct rq *rq;
6266 
6267 	rq = task_rq_lock(tsk, &rf);
6268 	update_rq_clock(rq);
6269 
6270 	running = task_current(rq, tsk);
6271 	queued = task_on_rq_queued(tsk);
6272 
6273 	if (queued)
6274 		dequeue_task(rq, tsk, queue_flags);
6275 	if (running)
6276 		put_prev_task(rq, tsk);
6277 
6278 	sched_change_group(tsk, TASK_MOVE_GROUP);
6279 
6280 	if (queued)
6281 		enqueue_task(rq, tsk, queue_flags);
6282 	if (running)
6283 		set_curr_task(rq, tsk);
6284 
6285 	task_rq_unlock(rq, tsk, &rf);
6286 }
6287 
6288 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6289 {
6290 	return css ? container_of(css, struct task_group, css) : NULL;
6291 }
6292 
6293 static struct cgroup_subsys_state *
6294 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6295 {
6296 	struct task_group *parent = css_tg(parent_css);
6297 	struct task_group *tg;
6298 
6299 	if (!parent) {
6300 		/* This is early initialization for the top cgroup */
6301 		return &root_task_group.css;
6302 	}
6303 
6304 	tg = sched_create_group(parent);
6305 	if (IS_ERR(tg))
6306 		return ERR_PTR(-ENOMEM);
6307 
6308 	return &tg->css;
6309 }
6310 
6311 /* Expose task group only after completing cgroup initialization */
6312 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6313 {
6314 	struct task_group *tg = css_tg(css);
6315 	struct task_group *parent = css_tg(css->parent);
6316 
6317 	if (parent)
6318 		sched_online_group(tg, parent);
6319 	return 0;
6320 }
6321 
6322 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6323 {
6324 	struct task_group *tg = css_tg(css);
6325 
6326 	sched_offline_group(tg);
6327 }
6328 
6329 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6330 {
6331 	struct task_group *tg = css_tg(css);
6332 
6333 	/*
6334 	 * Relies on the RCU grace period between css_released() and this.
6335 	 */
6336 	sched_free_group(tg);
6337 }
6338 
6339 /*
6340  * This is called before wake_up_new_task(), therefore we really only
6341  * have to set its group bits, all the other stuff does not apply.
6342  */
6343 static void cpu_cgroup_fork(struct task_struct *task)
6344 {
6345 	struct rq_flags rf;
6346 	struct rq *rq;
6347 
6348 	rq = task_rq_lock(task, &rf);
6349 
6350 	update_rq_clock(rq);
6351 	sched_change_group(task, TASK_SET_GROUP);
6352 
6353 	task_rq_unlock(rq, task, &rf);
6354 }
6355 
6356 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6357 {
6358 	struct task_struct *task;
6359 	struct cgroup_subsys_state *css;
6360 	int ret = 0;
6361 
6362 	cgroup_taskset_for_each(task, css, tset) {
6363 #ifdef CONFIG_RT_GROUP_SCHED
6364 		if (!sched_rt_can_attach(css_tg(css), task))
6365 			return -EINVAL;
6366 #else
6367 		/* We don't support RT-tasks being in separate groups */
6368 		if (task->sched_class != &fair_sched_class)
6369 			return -EINVAL;
6370 #endif
6371 		/*
6372 		 * Serialize against wake_up_new_task() such that if its
6373 		 * running, we're sure to observe its full state.
6374 		 */
6375 		raw_spin_lock_irq(&task->pi_lock);
6376 		/*
6377 		 * Avoid calling sched_move_task() before wake_up_new_task()
6378 		 * has happened. This would lead to problems with PELT, due to
6379 		 * move wanting to detach+attach while we're not attached yet.
6380 		 */
6381 		if (task->state == TASK_NEW)
6382 			ret = -EINVAL;
6383 		raw_spin_unlock_irq(&task->pi_lock);
6384 
6385 		if (ret)
6386 			break;
6387 	}
6388 	return ret;
6389 }
6390 
6391 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6392 {
6393 	struct task_struct *task;
6394 	struct cgroup_subsys_state *css;
6395 
6396 	cgroup_taskset_for_each(task, css, tset)
6397 		sched_move_task(task);
6398 }
6399 
6400 #ifdef CONFIG_FAIR_GROUP_SCHED
6401 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6402 				struct cftype *cftype, u64 shareval)
6403 {
6404 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
6405 }
6406 
6407 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6408 			       struct cftype *cft)
6409 {
6410 	struct task_group *tg = css_tg(css);
6411 
6412 	return (u64) scale_load_down(tg->shares);
6413 }
6414 
6415 #ifdef CONFIG_CFS_BANDWIDTH
6416 static DEFINE_MUTEX(cfs_constraints_mutex);
6417 
6418 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6419 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6420 
6421 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6422 
6423 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6424 {
6425 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
6426 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6427 
6428 	if (tg == &root_task_group)
6429 		return -EINVAL;
6430 
6431 	/*
6432 	 * Ensure we have at some amount of bandwidth every period.  This is
6433 	 * to prevent reaching a state of large arrears when throttled via
6434 	 * entity_tick() resulting in prolonged exit starvation.
6435 	 */
6436 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6437 		return -EINVAL;
6438 
6439 	/*
6440 	 * Likewise, bound things on the otherside by preventing insane quota
6441 	 * periods.  This also allows us to normalize in computing quota
6442 	 * feasibility.
6443 	 */
6444 	if (period > max_cfs_quota_period)
6445 		return -EINVAL;
6446 
6447 	/*
6448 	 * Prevent race between setting of cfs_rq->runtime_enabled and
6449 	 * unthrottle_offline_cfs_rqs().
6450 	 */
6451 	get_online_cpus();
6452 	mutex_lock(&cfs_constraints_mutex);
6453 	ret = __cfs_schedulable(tg, period, quota);
6454 	if (ret)
6455 		goto out_unlock;
6456 
6457 	runtime_enabled = quota != RUNTIME_INF;
6458 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6459 	/*
6460 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
6461 	 * before making related changes, and on->off must occur afterwards
6462 	 */
6463 	if (runtime_enabled && !runtime_was_enabled)
6464 		cfs_bandwidth_usage_inc();
6465 	raw_spin_lock_irq(&cfs_b->lock);
6466 	cfs_b->period = ns_to_ktime(period);
6467 	cfs_b->quota = quota;
6468 
6469 	__refill_cfs_bandwidth_runtime(cfs_b);
6470 
6471 	/* Restart the period timer (if active) to handle new period expiry: */
6472 	if (runtime_enabled)
6473 		start_cfs_bandwidth(cfs_b);
6474 
6475 	raw_spin_unlock_irq(&cfs_b->lock);
6476 
6477 	for_each_online_cpu(i) {
6478 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6479 		struct rq *rq = cfs_rq->rq;
6480 		struct rq_flags rf;
6481 
6482 		rq_lock_irq(rq, &rf);
6483 		cfs_rq->runtime_enabled = runtime_enabled;
6484 		cfs_rq->runtime_remaining = 0;
6485 
6486 		if (cfs_rq->throttled)
6487 			unthrottle_cfs_rq(cfs_rq);
6488 		rq_unlock_irq(rq, &rf);
6489 	}
6490 	if (runtime_was_enabled && !runtime_enabled)
6491 		cfs_bandwidth_usage_dec();
6492 out_unlock:
6493 	mutex_unlock(&cfs_constraints_mutex);
6494 	put_online_cpus();
6495 
6496 	return ret;
6497 }
6498 
6499 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6500 {
6501 	u64 quota, period;
6502 
6503 	period = ktime_to_ns(tg->cfs_bandwidth.period);
6504 	if (cfs_quota_us < 0)
6505 		quota = RUNTIME_INF;
6506 	else
6507 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6508 
6509 	return tg_set_cfs_bandwidth(tg, period, quota);
6510 }
6511 
6512 long tg_get_cfs_quota(struct task_group *tg)
6513 {
6514 	u64 quota_us;
6515 
6516 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6517 		return -1;
6518 
6519 	quota_us = tg->cfs_bandwidth.quota;
6520 	do_div(quota_us, NSEC_PER_USEC);
6521 
6522 	return quota_us;
6523 }
6524 
6525 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6526 {
6527 	u64 quota, period;
6528 
6529 	period = (u64)cfs_period_us * NSEC_PER_USEC;
6530 	quota = tg->cfs_bandwidth.quota;
6531 
6532 	return tg_set_cfs_bandwidth(tg, period, quota);
6533 }
6534 
6535 long tg_get_cfs_period(struct task_group *tg)
6536 {
6537 	u64 cfs_period_us;
6538 
6539 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6540 	do_div(cfs_period_us, NSEC_PER_USEC);
6541 
6542 	return cfs_period_us;
6543 }
6544 
6545 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6546 				  struct cftype *cft)
6547 {
6548 	return tg_get_cfs_quota(css_tg(css));
6549 }
6550 
6551 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6552 				   struct cftype *cftype, s64 cfs_quota_us)
6553 {
6554 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6555 }
6556 
6557 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6558 				   struct cftype *cft)
6559 {
6560 	return tg_get_cfs_period(css_tg(css));
6561 }
6562 
6563 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6564 				    struct cftype *cftype, u64 cfs_period_us)
6565 {
6566 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
6567 }
6568 
6569 struct cfs_schedulable_data {
6570 	struct task_group *tg;
6571 	u64 period, quota;
6572 };
6573 
6574 /*
6575  * normalize group quota/period to be quota/max_period
6576  * note: units are usecs
6577  */
6578 static u64 normalize_cfs_quota(struct task_group *tg,
6579 			       struct cfs_schedulable_data *d)
6580 {
6581 	u64 quota, period;
6582 
6583 	if (tg == d->tg) {
6584 		period = d->period;
6585 		quota = d->quota;
6586 	} else {
6587 		period = tg_get_cfs_period(tg);
6588 		quota = tg_get_cfs_quota(tg);
6589 	}
6590 
6591 	/* note: these should typically be equivalent */
6592 	if (quota == RUNTIME_INF || quota == -1)
6593 		return RUNTIME_INF;
6594 
6595 	return to_ratio(period, quota);
6596 }
6597 
6598 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6599 {
6600 	struct cfs_schedulable_data *d = data;
6601 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6602 	s64 quota = 0, parent_quota = -1;
6603 
6604 	if (!tg->parent) {
6605 		quota = RUNTIME_INF;
6606 	} else {
6607 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6608 
6609 		quota = normalize_cfs_quota(tg, d);
6610 		parent_quota = parent_b->hierarchical_quota;
6611 
6612 		/*
6613 		 * Ensure max(child_quota) <= parent_quota, inherit when no
6614 		 * limit is set:
6615 		 */
6616 		if (quota == RUNTIME_INF)
6617 			quota = parent_quota;
6618 		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6619 			return -EINVAL;
6620 	}
6621 	cfs_b->hierarchical_quota = quota;
6622 
6623 	return 0;
6624 }
6625 
6626 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6627 {
6628 	int ret;
6629 	struct cfs_schedulable_data data = {
6630 		.tg = tg,
6631 		.period = period,
6632 		.quota = quota,
6633 	};
6634 
6635 	if (quota != RUNTIME_INF) {
6636 		do_div(data.period, NSEC_PER_USEC);
6637 		do_div(data.quota, NSEC_PER_USEC);
6638 	}
6639 
6640 	rcu_read_lock();
6641 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6642 	rcu_read_unlock();
6643 
6644 	return ret;
6645 }
6646 
6647 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6648 {
6649 	struct task_group *tg = css_tg(seq_css(sf));
6650 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6651 
6652 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6653 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6654 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6655 
6656 	return 0;
6657 }
6658 #endif /* CONFIG_CFS_BANDWIDTH */
6659 #endif /* CONFIG_FAIR_GROUP_SCHED */
6660 
6661 #ifdef CONFIG_RT_GROUP_SCHED
6662 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6663 				struct cftype *cft, s64 val)
6664 {
6665 	return sched_group_set_rt_runtime(css_tg(css), val);
6666 }
6667 
6668 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6669 			       struct cftype *cft)
6670 {
6671 	return sched_group_rt_runtime(css_tg(css));
6672 }
6673 
6674 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6675 				    struct cftype *cftype, u64 rt_period_us)
6676 {
6677 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
6678 }
6679 
6680 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6681 				   struct cftype *cft)
6682 {
6683 	return sched_group_rt_period(css_tg(css));
6684 }
6685 #endif /* CONFIG_RT_GROUP_SCHED */
6686 
6687 static struct cftype cpu_legacy_files[] = {
6688 #ifdef CONFIG_FAIR_GROUP_SCHED
6689 	{
6690 		.name = "shares",
6691 		.read_u64 = cpu_shares_read_u64,
6692 		.write_u64 = cpu_shares_write_u64,
6693 	},
6694 #endif
6695 #ifdef CONFIG_CFS_BANDWIDTH
6696 	{
6697 		.name = "cfs_quota_us",
6698 		.read_s64 = cpu_cfs_quota_read_s64,
6699 		.write_s64 = cpu_cfs_quota_write_s64,
6700 	},
6701 	{
6702 		.name = "cfs_period_us",
6703 		.read_u64 = cpu_cfs_period_read_u64,
6704 		.write_u64 = cpu_cfs_period_write_u64,
6705 	},
6706 	{
6707 		.name = "stat",
6708 		.seq_show = cpu_cfs_stat_show,
6709 	},
6710 #endif
6711 #ifdef CONFIG_RT_GROUP_SCHED
6712 	{
6713 		.name = "rt_runtime_us",
6714 		.read_s64 = cpu_rt_runtime_read,
6715 		.write_s64 = cpu_rt_runtime_write,
6716 	},
6717 	{
6718 		.name = "rt_period_us",
6719 		.read_u64 = cpu_rt_period_read_uint,
6720 		.write_u64 = cpu_rt_period_write_uint,
6721 	},
6722 #endif
6723 	{ }	/* Terminate */
6724 };
6725 
6726 static int cpu_extra_stat_show(struct seq_file *sf,
6727 			       struct cgroup_subsys_state *css)
6728 {
6729 #ifdef CONFIG_CFS_BANDWIDTH
6730 	{
6731 		struct task_group *tg = css_tg(css);
6732 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6733 		u64 throttled_usec;
6734 
6735 		throttled_usec = cfs_b->throttled_time;
6736 		do_div(throttled_usec, NSEC_PER_USEC);
6737 
6738 		seq_printf(sf, "nr_periods %d\n"
6739 			   "nr_throttled %d\n"
6740 			   "throttled_usec %llu\n",
6741 			   cfs_b->nr_periods, cfs_b->nr_throttled,
6742 			   throttled_usec);
6743 	}
6744 #endif
6745 	return 0;
6746 }
6747 
6748 #ifdef CONFIG_FAIR_GROUP_SCHED
6749 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6750 			       struct cftype *cft)
6751 {
6752 	struct task_group *tg = css_tg(css);
6753 	u64 weight = scale_load_down(tg->shares);
6754 
6755 	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6756 }
6757 
6758 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6759 				struct cftype *cft, u64 weight)
6760 {
6761 	/*
6762 	 * cgroup weight knobs should use the common MIN, DFL and MAX
6763 	 * values which are 1, 100 and 10000 respectively.  While it loses
6764 	 * a bit of range on both ends, it maps pretty well onto the shares
6765 	 * value used by scheduler and the round-trip conversions preserve
6766 	 * the original value over the entire range.
6767 	 */
6768 	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6769 		return -ERANGE;
6770 
6771 	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6772 
6773 	return sched_group_set_shares(css_tg(css), scale_load(weight));
6774 }
6775 
6776 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6777 				    struct cftype *cft)
6778 {
6779 	unsigned long weight = scale_load_down(css_tg(css)->shares);
6780 	int last_delta = INT_MAX;
6781 	int prio, delta;
6782 
6783 	/* find the closest nice value to the current weight */
6784 	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6785 		delta = abs(sched_prio_to_weight[prio] - weight);
6786 		if (delta >= last_delta)
6787 			break;
6788 		last_delta = delta;
6789 	}
6790 
6791 	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6792 }
6793 
6794 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6795 				     struct cftype *cft, s64 nice)
6796 {
6797 	unsigned long weight;
6798 
6799 	if (nice < MIN_NICE || nice > MAX_NICE)
6800 		return -ERANGE;
6801 
6802 	weight = sched_prio_to_weight[NICE_TO_PRIO(nice) - MAX_RT_PRIO];
6803 	return sched_group_set_shares(css_tg(css), scale_load(weight));
6804 }
6805 #endif
6806 
6807 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6808 						  long period, long quota)
6809 {
6810 	if (quota < 0)
6811 		seq_puts(sf, "max");
6812 	else
6813 		seq_printf(sf, "%ld", quota);
6814 
6815 	seq_printf(sf, " %ld\n", period);
6816 }
6817 
6818 /* caller should put the current value in *@periodp before calling */
6819 static int __maybe_unused cpu_period_quota_parse(char *buf,
6820 						 u64 *periodp, u64 *quotap)
6821 {
6822 	char tok[21];	/* U64_MAX */
6823 
6824 	if (!sscanf(buf, "%s %llu", tok, periodp))
6825 		return -EINVAL;
6826 
6827 	*periodp *= NSEC_PER_USEC;
6828 
6829 	if (sscanf(tok, "%llu", quotap))
6830 		*quotap *= NSEC_PER_USEC;
6831 	else if (!strcmp(tok, "max"))
6832 		*quotap = RUNTIME_INF;
6833 	else
6834 		return -EINVAL;
6835 
6836 	return 0;
6837 }
6838 
6839 #ifdef CONFIG_CFS_BANDWIDTH
6840 static int cpu_max_show(struct seq_file *sf, void *v)
6841 {
6842 	struct task_group *tg = css_tg(seq_css(sf));
6843 
6844 	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6845 	return 0;
6846 }
6847 
6848 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6849 			     char *buf, size_t nbytes, loff_t off)
6850 {
6851 	struct task_group *tg = css_tg(of_css(of));
6852 	u64 period = tg_get_cfs_period(tg);
6853 	u64 quota;
6854 	int ret;
6855 
6856 	ret = cpu_period_quota_parse(buf, &period, &quota);
6857 	if (!ret)
6858 		ret = tg_set_cfs_bandwidth(tg, period, quota);
6859 	return ret ?: nbytes;
6860 }
6861 #endif
6862 
6863 static struct cftype cpu_files[] = {
6864 #ifdef CONFIG_FAIR_GROUP_SCHED
6865 	{
6866 		.name = "weight",
6867 		.flags = CFTYPE_NOT_ON_ROOT,
6868 		.read_u64 = cpu_weight_read_u64,
6869 		.write_u64 = cpu_weight_write_u64,
6870 	},
6871 	{
6872 		.name = "weight.nice",
6873 		.flags = CFTYPE_NOT_ON_ROOT,
6874 		.read_s64 = cpu_weight_nice_read_s64,
6875 		.write_s64 = cpu_weight_nice_write_s64,
6876 	},
6877 #endif
6878 #ifdef CONFIG_CFS_BANDWIDTH
6879 	{
6880 		.name = "max",
6881 		.flags = CFTYPE_NOT_ON_ROOT,
6882 		.seq_show = cpu_max_show,
6883 		.write = cpu_max_write,
6884 	},
6885 #endif
6886 	{ }	/* terminate */
6887 };
6888 
6889 struct cgroup_subsys cpu_cgrp_subsys = {
6890 	.css_alloc	= cpu_cgroup_css_alloc,
6891 	.css_online	= cpu_cgroup_css_online,
6892 	.css_released	= cpu_cgroup_css_released,
6893 	.css_free	= cpu_cgroup_css_free,
6894 	.css_extra_stat_show = cpu_extra_stat_show,
6895 	.fork		= cpu_cgroup_fork,
6896 	.can_attach	= cpu_cgroup_can_attach,
6897 	.attach		= cpu_cgroup_attach,
6898 	.legacy_cftypes	= cpu_legacy_files,
6899 	.dfl_cftypes	= cpu_files,
6900 	.early_init	= true,
6901 	.threaded	= true,
6902 };
6903 
6904 #endif	/* CONFIG_CGROUP_SCHED */
6905 
6906 void dump_cpu_task(int cpu)
6907 {
6908 	pr_info("Task dump for CPU %d:\n", cpu);
6909 	sched_show_task(cpu_curr(cpu));
6910 }
6911 
6912 /*
6913  * Nice levels are multiplicative, with a gentle 10% change for every
6914  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
6915  * nice 1, it will get ~10% less CPU time than another CPU-bound task
6916  * that remained on nice 0.
6917  *
6918  * The "10% effect" is relative and cumulative: from _any_ nice level,
6919  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
6920  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
6921  * If a task goes up by ~10% and another task goes down by ~10% then
6922  * the relative distance between them is ~25%.)
6923  */
6924 const int sched_prio_to_weight[40] = {
6925  /* -20 */     88761,     71755,     56483,     46273,     36291,
6926  /* -15 */     29154,     23254,     18705,     14949,     11916,
6927  /* -10 */      9548,      7620,      6100,      4904,      3906,
6928  /*  -5 */      3121,      2501,      1991,      1586,      1277,
6929  /*   0 */      1024,       820,       655,       526,       423,
6930  /*   5 */       335,       272,       215,       172,       137,
6931  /*  10 */       110,        87,        70,        56,        45,
6932  /*  15 */        36,        29,        23,        18,        15,
6933 };
6934 
6935 /*
6936  * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
6937  *
6938  * In cases where the weight does not change often, we can use the
6939  * precalculated inverse to speed up arithmetics by turning divisions
6940  * into multiplications:
6941  */
6942 const u32 sched_prio_to_wmult[40] = {
6943  /* -20 */     48388,     59856,     76040,     92818,    118348,
6944  /* -15 */    147320,    184698,    229616,    287308,    360437,
6945  /* -10 */    449829,    563644,    704093,    875809,   1099582,
6946  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
6947  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
6948  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
6949  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
6950  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
6951 };
6952