xref: /openbmc/linux/kernel/sched/core.c (revision 3e5f1ff7e4f297b928fb579fa1e8ec65a400162e)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  kernel/sched/core.c
4  *
5  *  Core kernel scheduler code and related syscalls
6  *
7  *  Copyright (C) 1991-2002  Linus Torvalds
8  */
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
12 
13 #include "sched.h"
14 
15 #include <linux/nospec.h>
16 #include <linux/blkdev.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
19 
20 #include <asm/switch_to.h>
21 #include <asm/tlb.h>
22 
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
26 
27 #include "pelt.h"
28 #include "smp.h"
29 
30 /*
31  * Export tracepoints that act as a bare tracehook (ie: have no trace event
32  * associated with them) to allow external modules to probe them.
33  */
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
44 
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
46 
47 #ifdef CONFIG_SCHED_DEBUG
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 
62 /*
63  * Print a warning if need_resched is set for the given duration (if
64  * LATENCY_WARN is enabled).
65  *
66  * If sysctl_resched_latency_warn_once is set, only one warning will be shown
67  * per boot.
68  */
69 __read_mostly int sysctl_resched_latency_warn_ms = 100;
70 __read_mostly int sysctl_resched_latency_warn_once = 1;
71 #endif /* CONFIG_SCHED_DEBUG */
72 
73 /*
74  * Number of tasks to iterate in a single balance run.
75  * Limited because this is done with IRQs disabled.
76  */
77 #ifdef CONFIG_PREEMPT_RT
78 const_debug unsigned int sysctl_sched_nr_migrate = 8;
79 #else
80 const_debug unsigned int sysctl_sched_nr_migrate = 32;
81 #endif
82 
83 /*
84  * period over which we measure -rt task CPU usage in us.
85  * default: 1s
86  */
87 unsigned int sysctl_sched_rt_period = 1000000;
88 
89 __read_mostly int scheduler_running;
90 
91 #ifdef CONFIG_SCHED_CORE
92 
93 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
94 
95 /* kernel prio, less is more */
96 static inline int __task_prio(struct task_struct *p)
97 {
98 	if (p->sched_class == &stop_sched_class) /* trumps deadline */
99 		return -2;
100 
101 	if (rt_prio(p->prio)) /* includes deadline */
102 		return p->prio; /* [-1, 99] */
103 
104 	if (p->sched_class == &idle_sched_class)
105 		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
106 
107 	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
108 }
109 
110 /*
111  * l(a,b)
112  * le(a,b) := !l(b,a)
113  * g(a,b)  := l(b,a)
114  * ge(a,b) := !l(a,b)
115  */
116 
117 /* real prio, less is less */
118 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
119 {
120 
121 	int pa = __task_prio(a), pb = __task_prio(b);
122 
123 	if (-pa < -pb)
124 		return true;
125 
126 	if (-pb < -pa)
127 		return false;
128 
129 	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
130 		return !dl_time_before(a->dl.deadline, b->dl.deadline);
131 
132 	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
133 		return cfs_prio_less(a, b, in_fi);
134 
135 	return false;
136 }
137 
138 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
139 {
140 	if (a->core_cookie < b->core_cookie)
141 		return true;
142 
143 	if (a->core_cookie > b->core_cookie)
144 		return false;
145 
146 	/* flip prio, so high prio is leftmost */
147 	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
148 		return true;
149 
150 	return false;
151 }
152 
153 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
154 
155 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
156 {
157 	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
158 }
159 
160 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
161 {
162 	const struct task_struct *p = __node_2_sc(node);
163 	unsigned long cookie = (unsigned long)key;
164 
165 	if (cookie < p->core_cookie)
166 		return -1;
167 
168 	if (cookie > p->core_cookie)
169 		return 1;
170 
171 	return 0;
172 }
173 
174 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
175 {
176 	rq->core->core_task_seq++;
177 
178 	if (!p->core_cookie)
179 		return;
180 
181 	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
182 }
183 
184 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
185 {
186 	rq->core->core_task_seq++;
187 
188 	if (sched_core_enqueued(p)) {
189 		rb_erase(&p->core_node, &rq->core_tree);
190 		RB_CLEAR_NODE(&p->core_node);
191 	}
192 
193 	/*
194 	 * Migrating the last task off the cpu, with the cpu in forced idle
195 	 * state. Reschedule to create an accounting edge for forced idle,
196 	 * and re-examine whether the core is still in forced idle state.
197 	 */
198 	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
199 	    rq->core->core_forceidle_count && rq->curr == rq->idle)
200 		resched_curr(rq);
201 }
202 
203 /*
204  * Find left-most (aka, highest priority) task matching @cookie.
205  */
206 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
207 {
208 	struct rb_node *node;
209 
210 	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
211 	/*
212 	 * The idle task always matches any cookie!
213 	 */
214 	if (!node)
215 		return idle_sched_class.pick_task(rq);
216 
217 	return __node_2_sc(node);
218 }
219 
220 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
221 {
222 	struct rb_node *node = &p->core_node;
223 
224 	node = rb_next(node);
225 	if (!node)
226 		return NULL;
227 
228 	p = container_of(node, struct task_struct, core_node);
229 	if (p->core_cookie != cookie)
230 		return NULL;
231 
232 	return p;
233 }
234 
235 /*
236  * Magic required such that:
237  *
238  *	raw_spin_rq_lock(rq);
239  *	...
240  *	raw_spin_rq_unlock(rq);
241  *
242  * ends up locking and unlocking the _same_ lock, and all CPUs
243  * always agree on what rq has what lock.
244  *
245  * XXX entirely possible to selectively enable cores, don't bother for now.
246  */
247 
248 static DEFINE_MUTEX(sched_core_mutex);
249 static atomic_t sched_core_count;
250 static struct cpumask sched_core_mask;
251 
252 static void sched_core_lock(int cpu, unsigned long *flags)
253 {
254 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
255 	int t, i = 0;
256 
257 	local_irq_save(*flags);
258 	for_each_cpu(t, smt_mask)
259 		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
260 }
261 
262 static void sched_core_unlock(int cpu, unsigned long *flags)
263 {
264 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
265 	int t;
266 
267 	for_each_cpu(t, smt_mask)
268 		raw_spin_unlock(&cpu_rq(t)->__lock);
269 	local_irq_restore(*flags);
270 }
271 
272 static void __sched_core_flip(bool enabled)
273 {
274 	unsigned long flags;
275 	int cpu, t;
276 
277 	cpus_read_lock();
278 
279 	/*
280 	 * Toggle the online cores, one by one.
281 	 */
282 	cpumask_copy(&sched_core_mask, cpu_online_mask);
283 	for_each_cpu(cpu, &sched_core_mask) {
284 		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
285 
286 		sched_core_lock(cpu, &flags);
287 
288 		for_each_cpu(t, smt_mask)
289 			cpu_rq(t)->core_enabled = enabled;
290 
291 		cpu_rq(cpu)->core->core_forceidle_start = 0;
292 
293 		sched_core_unlock(cpu, &flags);
294 
295 		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
296 	}
297 
298 	/*
299 	 * Toggle the offline CPUs.
300 	 */
301 	cpumask_copy(&sched_core_mask, cpu_possible_mask);
302 	cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
303 
304 	for_each_cpu(cpu, &sched_core_mask)
305 		cpu_rq(cpu)->core_enabled = enabled;
306 
307 	cpus_read_unlock();
308 }
309 
310 static void sched_core_assert_empty(void)
311 {
312 	int cpu;
313 
314 	for_each_possible_cpu(cpu)
315 		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
316 }
317 
318 static void __sched_core_enable(void)
319 {
320 	static_branch_enable(&__sched_core_enabled);
321 	/*
322 	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
323 	 * and future ones will observe !sched_core_disabled().
324 	 */
325 	synchronize_rcu();
326 	__sched_core_flip(true);
327 	sched_core_assert_empty();
328 }
329 
330 static void __sched_core_disable(void)
331 {
332 	sched_core_assert_empty();
333 	__sched_core_flip(false);
334 	static_branch_disable(&__sched_core_enabled);
335 }
336 
337 void sched_core_get(void)
338 {
339 	if (atomic_inc_not_zero(&sched_core_count))
340 		return;
341 
342 	mutex_lock(&sched_core_mutex);
343 	if (!atomic_read(&sched_core_count))
344 		__sched_core_enable();
345 
346 	smp_mb__before_atomic();
347 	atomic_inc(&sched_core_count);
348 	mutex_unlock(&sched_core_mutex);
349 }
350 
351 static void __sched_core_put(struct work_struct *work)
352 {
353 	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
354 		__sched_core_disable();
355 		mutex_unlock(&sched_core_mutex);
356 	}
357 }
358 
359 void sched_core_put(void)
360 {
361 	static DECLARE_WORK(_work, __sched_core_put);
362 
363 	/*
364 	 * "There can be only one"
365 	 *
366 	 * Either this is the last one, or we don't actually need to do any
367 	 * 'work'. If it is the last *again*, we rely on
368 	 * WORK_STRUCT_PENDING_BIT.
369 	 */
370 	if (!atomic_add_unless(&sched_core_count, -1, 1))
371 		schedule_work(&_work);
372 }
373 
374 #else /* !CONFIG_SCHED_CORE */
375 
376 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
377 static inline void
378 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
379 
380 #endif /* CONFIG_SCHED_CORE */
381 
382 /*
383  * part of the period that we allow rt tasks to run in us.
384  * default: 0.95s
385  */
386 int sysctl_sched_rt_runtime = 950000;
387 
388 
389 /*
390  * Serialization rules:
391  *
392  * Lock order:
393  *
394  *   p->pi_lock
395  *     rq->lock
396  *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
397  *
398  *  rq1->lock
399  *    rq2->lock  where: rq1 < rq2
400  *
401  * Regular state:
402  *
403  * Normal scheduling state is serialized by rq->lock. __schedule() takes the
404  * local CPU's rq->lock, it optionally removes the task from the runqueue and
405  * always looks at the local rq data structures to find the most eligible task
406  * to run next.
407  *
408  * Task enqueue is also under rq->lock, possibly taken from another CPU.
409  * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
410  * the local CPU to avoid bouncing the runqueue state around [ see
411  * ttwu_queue_wakelist() ]
412  *
413  * Task wakeup, specifically wakeups that involve migration, are horribly
414  * complicated to avoid having to take two rq->locks.
415  *
416  * Special state:
417  *
418  * System-calls and anything external will use task_rq_lock() which acquires
419  * both p->pi_lock and rq->lock. As a consequence the state they change is
420  * stable while holding either lock:
421  *
422  *  - sched_setaffinity()/
423  *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
424  *  - set_user_nice():		p->se.load, p->*prio
425  *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
426  *				p->se.load, p->rt_priority,
427  *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
428  *  - sched_setnuma():		p->numa_preferred_nid
429  *  - sched_move_task()/
430  *    cpu_cgroup_fork():	p->sched_task_group
431  *  - uclamp_update_active()	p->uclamp*
432  *
433  * p->state <- TASK_*:
434  *
435  *   is changed locklessly using set_current_state(), __set_current_state() or
436  *   set_special_state(), see their respective comments, or by
437  *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
438  *   concurrent self.
439  *
440  * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
441  *
442  *   is set by activate_task() and cleared by deactivate_task(), under
443  *   rq->lock. Non-zero indicates the task is runnable, the special
444  *   ON_RQ_MIGRATING state is used for migration without holding both
445  *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
446  *
447  * p->on_cpu <- { 0, 1 }:
448  *
449  *   is set by prepare_task() and cleared by finish_task() such that it will be
450  *   set before p is scheduled-in and cleared after p is scheduled-out, both
451  *   under rq->lock. Non-zero indicates the task is running on its CPU.
452  *
453  *   [ The astute reader will observe that it is possible for two tasks on one
454  *     CPU to have ->on_cpu = 1 at the same time. ]
455  *
456  * task_cpu(p): is changed by set_task_cpu(), the rules are:
457  *
458  *  - Don't call set_task_cpu() on a blocked task:
459  *
460  *    We don't care what CPU we're not running on, this simplifies hotplug,
461  *    the CPU assignment of blocked tasks isn't required to be valid.
462  *
463  *  - for try_to_wake_up(), called under p->pi_lock:
464  *
465  *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
466  *
467  *  - for migration called under rq->lock:
468  *    [ see task_on_rq_migrating() in task_rq_lock() ]
469  *
470  *    o move_queued_task()
471  *    o detach_task()
472  *
473  *  - for migration called under double_rq_lock():
474  *
475  *    o __migrate_swap_task()
476  *    o push_rt_task() / pull_rt_task()
477  *    o push_dl_task() / pull_dl_task()
478  *    o dl_task_offline_migration()
479  *
480  */
481 
482 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
483 {
484 	raw_spinlock_t *lock;
485 
486 	/* Matches synchronize_rcu() in __sched_core_enable() */
487 	preempt_disable();
488 	if (sched_core_disabled()) {
489 		raw_spin_lock_nested(&rq->__lock, subclass);
490 		/* preempt_count *MUST* be > 1 */
491 		preempt_enable_no_resched();
492 		return;
493 	}
494 
495 	for (;;) {
496 		lock = __rq_lockp(rq);
497 		raw_spin_lock_nested(lock, subclass);
498 		if (likely(lock == __rq_lockp(rq))) {
499 			/* preempt_count *MUST* be > 1 */
500 			preempt_enable_no_resched();
501 			return;
502 		}
503 		raw_spin_unlock(lock);
504 	}
505 }
506 
507 bool raw_spin_rq_trylock(struct rq *rq)
508 {
509 	raw_spinlock_t *lock;
510 	bool ret;
511 
512 	/* Matches synchronize_rcu() in __sched_core_enable() */
513 	preempt_disable();
514 	if (sched_core_disabled()) {
515 		ret = raw_spin_trylock(&rq->__lock);
516 		preempt_enable();
517 		return ret;
518 	}
519 
520 	for (;;) {
521 		lock = __rq_lockp(rq);
522 		ret = raw_spin_trylock(lock);
523 		if (!ret || (likely(lock == __rq_lockp(rq)))) {
524 			preempt_enable();
525 			return ret;
526 		}
527 		raw_spin_unlock(lock);
528 	}
529 }
530 
531 void raw_spin_rq_unlock(struct rq *rq)
532 {
533 	raw_spin_unlock(rq_lockp(rq));
534 }
535 
536 #ifdef CONFIG_SMP
537 /*
538  * double_rq_lock - safely lock two runqueues
539  */
540 void double_rq_lock(struct rq *rq1, struct rq *rq2)
541 {
542 	lockdep_assert_irqs_disabled();
543 
544 	if (rq_order_less(rq2, rq1))
545 		swap(rq1, rq2);
546 
547 	raw_spin_rq_lock(rq1);
548 	if (__rq_lockp(rq1) == __rq_lockp(rq2))
549 		return;
550 
551 	raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
552 }
553 #endif
554 
555 /*
556  * __task_rq_lock - lock the rq @p resides on.
557  */
558 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
559 	__acquires(rq->lock)
560 {
561 	struct rq *rq;
562 
563 	lockdep_assert_held(&p->pi_lock);
564 
565 	for (;;) {
566 		rq = task_rq(p);
567 		raw_spin_rq_lock(rq);
568 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
569 			rq_pin_lock(rq, rf);
570 			return rq;
571 		}
572 		raw_spin_rq_unlock(rq);
573 
574 		while (unlikely(task_on_rq_migrating(p)))
575 			cpu_relax();
576 	}
577 }
578 
579 /*
580  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
581  */
582 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
583 	__acquires(p->pi_lock)
584 	__acquires(rq->lock)
585 {
586 	struct rq *rq;
587 
588 	for (;;) {
589 		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
590 		rq = task_rq(p);
591 		raw_spin_rq_lock(rq);
592 		/*
593 		 *	move_queued_task()		task_rq_lock()
594 		 *
595 		 *	ACQUIRE (rq->lock)
596 		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
597 		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
598 		 *	[S] ->cpu = new_cpu		[L] task_rq()
599 		 *					[L] ->on_rq
600 		 *	RELEASE (rq->lock)
601 		 *
602 		 * If we observe the old CPU in task_rq_lock(), the acquire of
603 		 * the old rq->lock will fully serialize against the stores.
604 		 *
605 		 * If we observe the new CPU in task_rq_lock(), the address
606 		 * dependency headed by '[L] rq = task_rq()' and the acquire
607 		 * will pair with the WMB to ensure we then also see migrating.
608 		 */
609 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
610 			rq_pin_lock(rq, rf);
611 			return rq;
612 		}
613 		raw_spin_rq_unlock(rq);
614 		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
615 
616 		while (unlikely(task_on_rq_migrating(p)))
617 			cpu_relax();
618 	}
619 }
620 
621 /*
622  * RQ-clock updating methods:
623  */
624 
625 static void update_rq_clock_task(struct rq *rq, s64 delta)
626 {
627 /*
628  * In theory, the compile should just see 0 here, and optimize out the call
629  * to sched_rt_avg_update. But I don't trust it...
630  */
631 	s64 __maybe_unused steal = 0, irq_delta = 0;
632 
633 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
634 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
635 
636 	/*
637 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
638 	 * this case when a previous update_rq_clock() happened inside a
639 	 * {soft,}irq region.
640 	 *
641 	 * When this happens, we stop ->clock_task and only update the
642 	 * prev_irq_time stamp to account for the part that fit, so that a next
643 	 * update will consume the rest. This ensures ->clock_task is
644 	 * monotonic.
645 	 *
646 	 * It does however cause some slight miss-attribution of {soft,}irq
647 	 * time, a more accurate solution would be to update the irq_time using
648 	 * the current rq->clock timestamp, except that would require using
649 	 * atomic ops.
650 	 */
651 	if (irq_delta > delta)
652 		irq_delta = delta;
653 
654 	rq->prev_irq_time += irq_delta;
655 	delta -= irq_delta;
656 #endif
657 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
658 	if (static_key_false((&paravirt_steal_rq_enabled))) {
659 		steal = paravirt_steal_clock(cpu_of(rq));
660 		steal -= rq->prev_steal_time_rq;
661 
662 		if (unlikely(steal > delta))
663 			steal = delta;
664 
665 		rq->prev_steal_time_rq += steal;
666 		delta -= steal;
667 	}
668 #endif
669 
670 	rq->clock_task += delta;
671 
672 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
673 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
674 		update_irq_load_avg(rq, irq_delta + steal);
675 #endif
676 	update_rq_clock_pelt(rq, delta);
677 }
678 
679 void update_rq_clock(struct rq *rq)
680 {
681 	s64 delta;
682 
683 	lockdep_assert_rq_held(rq);
684 
685 	if (rq->clock_update_flags & RQCF_ACT_SKIP)
686 		return;
687 
688 #ifdef CONFIG_SCHED_DEBUG
689 	if (sched_feat(WARN_DOUBLE_CLOCK))
690 		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
691 	rq->clock_update_flags |= RQCF_UPDATED;
692 #endif
693 
694 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
695 	if (delta < 0)
696 		return;
697 	rq->clock += delta;
698 	update_rq_clock_task(rq, delta);
699 }
700 
701 #ifdef CONFIG_SCHED_HRTICK
702 /*
703  * Use HR-timers to deliver accurate preemption points.
704  */
705 
706 static void hrtick_clear(struct rq *rq)
707 {
708 	if (hrtimer_active(&rq->hrtick_timer))
709 		hrtimer_cancel(&rq->hrtick_timer);
710 }
711 
712 /*
713  * High-resolution timer tick.
714  * Runs from hardirq context with interrupts disabled.
715  */
716 static enum hrtimer_restart hrtick(struct hrtimer *timer)
717 {
718 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
719 	struct rq_flags rf;
720 
721 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
722 
723 	rq_lock(rq, &rf);
724 	update_rq_clock(rq);
725 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
726 	rq_unlock(rq, &rf);
727 
728 	return HRTIMER_NORESTART;
729 }
730 
731 #ifdef CONFIG_SMP
732 
733 static void __hrtick_restart(struct rq *rq)
734 {
735 	struct hrtimer *timer = &rq->hrtick_timer;
736 	ktime_t time = rq->hrtick_time;
737 
738 	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
739 }
740 
741 /*
742  * called from hardirq (IPI) context
743  */
744 static void __hrtick_start(void *arg)
745 {
746 	struct rq *rq = arg;
747 	struct rq_flags rf;
748 
749 	rq_lock(rq, &rf);
750 	__hrtick_restart(rq);
751 	rq_unlock(rq, &rf);
752 }
753 
754 /*
755  * Called to set the hrtick timer state.
756  *
757  * called with rq->lock held and irqs disabled
758  */
759 void hrtick_start(struct rq *rq, u64 delay)
760 {
761 	struct hrtimer *timer = &rq->hrtick_timer;
762 	s64 delta;
763 
764 	/*
765 	 * Don't schedule slices shorter than 10000ns, that just
766 	 * doesn't make sense and can cause timer DoS.
767 	 */
768 	delta = max_t(s64, delay, 10000LL);
769 	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
770 
771 	if (rq == this_rq())
772 		__hrtick_restart(rq);
773 	else
774 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
775 }
776 
777 #else
778 /*
779  * Called to set the hrtick timer state.
780  *
781  * called with rq->lock held and irqs disabled
782  */
783 void hrtick_start(struct rq *rq, u64 delay)
784 {
785 	/*
786 	 * Don't schedule slices shorter than 10000ns, that just
787 	 * doesn't make sense. Rely on vruntime for fairness.
788 	 */
789 	delay = max_t(u64, delay, 10000LL);
790 	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
791 		      HRTIMER_MODE_REL_PINNED_HARD);
792 }
793 
794 #endif /* CONFIG_SMP */
795 
796 static void hrtick_rq_init(struct rq *rq)
797 {
798 #ifdef CONFIG_SMP
799 	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
800 #endif
801 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
802 	rq->hrtick_timer.function = hrtick;
803 }
804 #else	/* CONFIG_SCHED_HRTICK */
805 static inline void hrtick_clear(struct rq *rq)
806 {
807 }
808 
809 static inline void hrtick_rq_init(struct rq *rq)
810 {
811 }
812 #endif	/* CONFIG_SCHED_HRTICK */
813 
814 /*
815  * cmpxchg based fetch_or, macro so it works for different integer types
816  */
817 #define fetch_or(ptr, mask)						\
818 	({								\
819 		typeof(ptr) _ptr = (ptr);				\
820 		typeof(mask) _mask = (mask);				\
821 		typeof(*_ptr) _old, _val = *_ptr;			\
822 									\
823 		for (;;) {						\
824 			_old = cmpxchg(_ptr, _val, _val | _mask);	\
825 			if (_old == _val)				\
826 				break;					\
827 			_val = _old;					\
828 		}							\
829 	_old;								\
830 })
831 
832 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
833 /*
834  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
835  * this avoids any races wrt polling state changes and thereby avoids
836  * spurious IPIs.
837  */
838 static bool set_nr_and_not_polling(struct task_struct *p)
839 {
840 	struct thread_info *ti = task_thread_info(p);
841 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
842 }
843 
844 /*
845  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
846  *
847  * If this returns true, then the idle task promises to call
848  * sched_ttwu_pending() and reschedule soon.
849  */
850 static bool set_nr_if_polling(struct task_struct *p)
851 {
852 	struct thread_info *ti = task_thread_info(p);
853 	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
854 
855 	for (;;) {
856 		if (!(val & _TIF_POLLING_NRFLAG))
857 			return false;
858 		if (val & _TIF_NEED_RESCHED)
859 			return true;
860 		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
861 		if (old == val)
862 			break;
863 		val = old;
864 	}
865 	return true;
866 }
867 
868 #else
869 static bool set_nr_and_not_polling(struct task_struct *p)
870 {
871 	set_tsk_need_resched(p);
872 	return true;
873 }
874 
875 #ifdef CONFIG_SMP
876 static bool set_nr_if_polling(struct task_struct *p)
877 {
878 	return false;
879 }
880 #endif
881 #endif
882 
883 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
884 {
885 	struct wake_q_node *node = &task->wake_q;
886 
887 	/*
888 	 * Atomically grab the task, if ->wake_q is !nil already it means
889 	 * it's already queued (either by us or someone else) and will get the
890 	 * wakeup due to that.
891 	 *
892 	 * In order to ensure that a pending wakeup will observe our pending
893 	 * state, even in the failed case, an explicit smp_mb() must be used.
894 	 */
895 	smp_mb__before_atomic();
896 	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
897 		return false;
898 
899 	/*
900 	 * The head is context local, there can be no concurrency.
901 	 */
902 	*head->lastp = node;
903 	head->lastp = &node->next;
904 	return true;
905 }
906 
907 /**
908  * wake_q_add() - queue a wakeup for 'later' waking.
909  * @head: the wake_q_head to add @task to
910  * @task: the task to queue for 'later' wakeup
911  *
912  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
913  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
914  * instantly.
915  *
916  * This function must be used as-if it were wake_up_process(); IOW the task
917  * must be ready to be woken at this location.
918  */
919 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
920 {
921 	if (__wake_q_add(head, task))
922 		get_task_struct(task);
923 }
924 
925 /**
926  * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
927  * @head: the wake_q_head to add @task to
928  * @task: the task to queue for 'later' wakeup
929  *
930  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
931  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
932  * instantly.
933  *
934  * This function must be used as-if it were wake_up_process(); IOW the task
935  * must be ready to be woken at this location.
936  *
937  * This function is essentially a task-safe equivalent to wake_q_add(). Callers
938  * that already hold reference to @task can call the 'safe' version and trust
939  * wake_q to do the right thing depending whether or not the @task is already
940  * queued for wakeup.
941  */
942 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
943 {
944 	if (!__wake_q_add(head, task))
945 		put_task_struct(task);
946 }
947 
948 void wake_up_q(struct wake_q_head *head)
949 {
950 	struct wake_q_node *node = head->first;
951 
952 	while (node != WAKE_Q_TAIL) {
953 		struct task_struct *task;
954 
955 		task = container_of(node, struct task_struct, wake_q);
956 		/* Task can safely be re-inserted now: */
957 		node = node->next;
958 		task->wake_q.next = NULL;
959 
960 		/*
961 		 * wake_up_process() executes a full barrier, which pairs with
962 		 * the queueing in wake_q_add() so as not to miss wakeups.
963 		 */
964 		wake_up_process(task);
965 		put_task_struct(task);
966 	}
967 }
968 
969 /*
970  * resched_curr - mark rq's current task 'to be rescheduled now'.
971  *
972  * On UP this means the setting of the need_resched flag, on SMP it
973  * might also involve a cross-CPU call to trigger the scheduler on
974  * the target CPU.
975  */
976 void resched_curr(struct rq *rq)
977 {
978 	struct task_struct *curr = rq->curr;
979 	int cpu;
980 
981 	lockdep_assert_rq_held(rq);
982 
983 	if (test_tsk_need_resched(curr))
984 		return;
985 
986 	cpu = cpu_of(rq);
987 
988 	if (cpu == smp_processor_id()) {
989 		set_tsk_need_resched(curr);
990 		set_preempt_need_resched();
991 		return;
992 	}
993 
994 	if (set_nr_and_not_polling(curr))
995 		smp_send_reschedule(cpu);
996 	else
997 		trace_sched_wake_idle_without_ipi(cpu);
998 }
999 
1000 void resched_cpu(int cpu)
1001 {
1002 	struct rq *rq = cpu_rq(cpu);
1003 	unsigned long flags;
1004 
1005 	raw_spin_rq_lock_irqsave(rq, flags);
1006 	if (cpu_online(cpu) || cpu == smp_processor_id())
1007 		resched_curr(rq);
1008 	raw_spin_rq_unlock_irqrestore(rq, flags);
1009 }
1010 
1011 #ifdef CONFIG_SMP
1012 #ifdef CONFIG_NO_HZ_COMMON
1013 /*
1014  * In the semi idle case, use the nearest busy CPU for migrating timers
1015  * from an idle CPU.  This is good for power-savings.
1016  *
1017  * We don't do similar optimization for completely idle system, as
1018  * selecting an idle CPU will add more delays to the timers than intended
1019  * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1020  */
1021 int get_nohz_timer_target(void)
1022 {
1023 	int i, cpu = smp_processor_id(), default_cpu = -1;
1024 	struct sched_domain *sd;
1025 	const struct cpumask *hk_mask;
1026 
1027 	if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1028 		if (!idle_cpu(cpu))
1029 			return cpu;
1030 		default_cpu = cpu;
1031 	}
1032 
1033 	hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1034 
1035 	rcu_read_lock();
1036 	for_each_domain(cpu, sd) {
1037 		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1038 			if (cpu == i)
1039 				continue;
1040 
1041 			if (!idle_cpu(i)) {
1042 				cpu = i;
1043 				goto unlock;
1044 			}
1045 		}
1046 	}
1047 
1048 	if (default_cpu == -1)
1049 		default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1050 	cpu = default_cpu;
1051 unlock:
1052 	rcu_read_unlock();
1053 	return cpu;
1054 }
1055 
1056 /*
1057  * When add_timer_on() enqueues a timer into the timer wheel of an
1058  * idle CPU then this timer might expire before the next timer event
1059  * which is scheduled to wake up that CPU. In case of a completely
1060  * idle system the next event might even be infinite time into the
1061  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1062  * leaves the inner idle loop so the newly added timer is taken into
1063  * account when the CPU goes back to idle and evaluates the timer
1064  * wheel for the next timer event.
1065  */
1066 static void wake_up_idle_cpu(int cpu)
1067 {
1068 	struct rq *rq = cpu_rq(cpu);
1069 
1070 	if (cpu == smp_processor_id())
1071 		return;
1072 
1073 	if (set_nr_and_not_polling(rq->idle))
1074 		smp_send_reschedule(cpu);
1075 	else
1076 		trace_sched_wake_idle_without_ipi(cpu);
1077 }
1078 
1079 static bool wake_up_full_nohz_cpu(int cpu)
1080 {
1081 	/*
1082 	 * We just need the target to call irq_exit() and re-evaluate
1083 	 * the next tick. The nohz full kick at least implies that.
1084 	 * If needed we can still optimize that later with an
1085 	 * empty IRQ.
1086 	 */
1087 	if (cpu_is_offline(cpu))
1088 		return true;  /* Don't try to wake offline CPUs. */
1089 	if (tick_nohz_full_cpu(cpu)) {
1090 		if (cpu != smp_processor_id() ||
1091 		    tick_nohz_tick_stopped())
1092 			tick_nohz_full_kick_cpu(cpu);
1093 		return true;
1094 	}
1095 
1096 	return false;
1097 }
1098 
1099 /*
1100  * Wake up the specified CPU.  If the CPU is going offline, it is the
1101  * caller's responsibility to deal with the lost wakeup, for example,
1102  * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1103  */
1104 void wake_up_nohz_cpu(int cpu)
1105 {
1106 	if (!wake_up_full_nohz_cpu(cpu))
1107 		wake_up_idle_cpu(cpu);
1108 }
1109 
1110 static void nohz_csd_func(void *info)
1111 {
1112 	struct rq *rq = info;
1113 	int cpu = cpu_of(rq);
1114 	unsigned int flags;
1115 
1116 	/*
1117 	 * Release the rq::nohz_csd.
1118 	 */
1119 	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1120 	WARN_ON(!(flags & NOHZ_KICK_MASK));
1121 
1122 	rq->idle_balance = idle_cpu(cpu);
1123 	if (rq->idle_balance && !need_resched()) {
1124 		rq->nohz_idle_balance = flags;
1125 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1126 	}
1127 }
1128 
1129 #endif /* CONFIG_NO_HZ_COMMON */
1130 
1131 #ifdef CONFIG_NO_HZ_FULL
1132 bool sched_can_stop_tick(struct rq *rq)
1133 {
1134 	int fifo_nr_running;
1135 
1136 	/* Deadline tasks, even if single, need the tick */
1137 	if (rq->dl.dl_nr_running)
1138 		return false;
1139 
1140 	/*
1141 	 * If there are more than one RR tasks, we need the tick to affect the
1142 	 * actual RR behaviour.
1143 	 */
1144 	if (rq->rt.rr_nr_running) {
1145 		if (rq->rt.rr_nr_running == 1)
1146 			return true;
1147 		else
1148 			return false;
1149 	}
1150 
1151 	/*
1152 	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1153 	 * forced preemption between FIFO tasks.
1154 	 */
1155 	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1156 	if (fifo_nr_running)
1157 		return true;
1158 
1159 	/*
1160 	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1161 	 * if there's more than one we need the tick for involuntary
1162 	 * preemption.
1163 	 */
1164 	if (rq->nr_running > 1)
1165 		return false;
1166 
1167 	return true;
1168 }
1169 #endif /* CONFIG_NO_HZ_FULL */
1170 #endif /* CONFIG_SMP */
1171 
1172 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1173 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1174 /*
1175  * Iterate task_group tree rooted at *from, calling @down when first entering a
1176  * node and @up when leaving it for the final time.
1177  *
1178  * Caller must hold rcu_lock or sufficient equivalent.
1179  */
1180 int walk_tg_tree_from(struct task_group *from,
1181 			     tg_visitor down, tg_visitor up, void *data)
1182 {
1183 	struct task_group *parent, *child;
1184 	int ret;
1185 
1186 	parent = from;
1187 
1188 down:
1189 	ret = (*down)(parent, data);
1190 	if (ret)
1191 		goto out;
1192 	list_for_each_entry_rcu(child, &parent->children, siblings) {
1193 		parent = child;
1194 		goto down;
1195 
1196 up:
1197 		continue;
1198 	}
1199 	ret = (*up)(parent, data);
1200 	if (ret || parent == from)
1201 		goto out;
1202 
1203 	child = parent;
1204 	parent = parent->parent;
1205 	if (parent)
1206 		goto up;
1207 out:
1208 	return ret;
1209 }
1210 
1211 int tg_nop(struct task_group *tg, void *data)
1212 {
1213 	return 0;
1214 }
1215 #endif
1216 
1217 static void set_load_weight(struct task_struct *p, bool update_load)
1218 {
1219 	int prio = p->static_prio - MAX_RT_PRIO;
1220 	struct load_weight *load = &p->se.load;
1221 
1222 	/*
1223 	 * SCHED_IDLE tasks get minimal weight:
1224 	 */
1225 	if (task_has_idle_policy(p)) {
1226 		load->weight = scale_load(WEIGHT_IDLEPRIO);
1227 		load->inv_weight = WMULT_IDLEPRIO;
1228 		return;
1229 	}
1230 
1231 	/*
1232 	 * SCHED_OTHER tasks have to update their load when changing their
1233 	 * weight
1234 	 */
1235 	if (update_load && p->sched_class == &fair_sched_class) {
1236 		reweight_task(p, prio);
1237 	} else {
1238 		load->weight = scale_load(sched_prio_to_weight[prio]);
1239 		load->inv_weight = sched_prio_to_wmult[prio];
1240 	}
1241 }
1242 
1243 #ifdef CONFIG_UCLAMP_TASK
1244 /*
1245  * Serializes updates of utilization clamp values
1246  *
1247  * The (slow-path) user-space triggers utilization clamp value updates which
1248  * can require updates on (fast-path) scheduler's data structures used to
1249  * support enqueue/dequeue operations.
1250  * While the per-CPU rq lock protects fast-path update operations, user-space
1251  * requests are serialized using a mutex to reduce the risk of conflicting
1252  * updates or API abuses.
1253  */
1254 static DEFINE_MUTEX(uclamp_mutex);
1255 
1256 /* Max allowed minimum utilization */
1257 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1258 
1259 /* Max allowed maximum utilization */
1260 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1261 
1262 /*
1263  * By default RT tasks run at the maximum performance point/capacity of the
1264  * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1265  * SCHED_CAPACITY_SCALE.
1266  *
1267  * This knob allows admins to change the default behavior when uclamp is being
1268  * used. In battery powered devices, particularly, running at the maximum
1269  * capacity and frequency will increase energy consumption and shorten the
1270  * battery life.
1271  *
1272  * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1273  *
1274  * This knob will not override the system default sched_util_clamp_min defined
1275  * above.
1276  */
1277 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1278 
1279 /* All clamps are required to be less or equal than these values */
1280 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1281 
1282 /*
1283  * This static key is used to reduce the uclamp overhead in the fast path. It
1284  * primarily disables the call to uclamp_rq_{inc, dec}() in
1285  * enqueue/dequeue_task().
1286  *
1287  * This allows users to continue to enable uclamp in their kernel config with
1288  * minimum uclamp overhead in the fast path.
1289  *
1290  * As soon as userspace modifies any of the uclamp knobs, the static key is
1291  * enabled, since we have an actual users that make use of uclamp
1292  * functionality.
1293  *
1294  * The knobs that would enable this static key are:
1295  *
1296  *   * A task modifying its uclamp value with sched_setattr().
1297  *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1298  *   * An admin modifying the cgroup cpu.uclamp.{min, max}
1299  */
1300 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1301 
1302 /* Integer rounded range for each bucket */
1303 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1304 
1305 #define for_each_clamp_id(clamp_id) \
1306 	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1307 
1308 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1309 {
1310 	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1311 }
1312 
1313 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1314 {
1315 	if (clamp_id == UCLAMP_MIN)
1316 		return 0;
1317 	return SCHED_CAPACITY_SCALE;
1318 }
1319 
1320 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1321 				 unsigned int value, bool user_defined)
1322 {
1323 	uc_se->value = value;
1324 	uc_se->bucket_id = uclamp_bucket_id(value);
1325 	uc_se->user_defined = user_defined;
1326 }
1327 
1328 static inline unsigned int
1329 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1330 		  unsigned int clamp_value)
1331 {
1332 	/*
1333 	 * Avoid blocked utilization pushing up the frequency when we go
1334 	 * idle (which drops the max-clamp) by retaining the last known
1335 	 * max-clamp.
1336 	 */
1337 	if (clamp_id == UCLAMP_MAX) {
1338 		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1339 		return clamp_value;
1340 	}
1341 
1342 	return uclamp_none(UCLAMP_MIN);
1343 }
1344 
1345 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1346 				     unsigned int clamp_value)
1347 {
1348 	/* Reset max-clamp retention only on idle exit */
1349 	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1350 		return;
1351 
1352 	WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1353 }
1354 
1355 static inline
1356 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1357 				   unsigned int clamp_value)
1358 {
1359 	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1360 	int bucket_id = UCLAMP_BUCKETS - 1;
1361 
1362 	/*
1363 	 * Since both min and max clamps are max aggregated, find the
1364 	 * top most bucket with tasks in.
1365 	 */
1366 	for ( ; bucket_id >= 0; bucket_id--) {
1367 		if (!bucket[bucket_id].tasks)
1368 			continue;
1369 		return bucket[bucket_id].value;
1370 	}
1371 
1372 	/* No tasks -- default clamp values */
1373 	return uclamp_idle_value(rq, clamp_id, clamp_value);
1374 }
1375 
1376 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1377 {
1378 	unsigned int default_util_min;
1379 	struct uclamp_se *uc_se;
1380 
1381 	lockdep_assert_held(&p->pi_lock);
1382 
1383 	uc_se = &p->uclamp_req[UCLAMP_MIN];
1384 
1385 	/* Only sync if user didn't override the default */
1386 	if (uc_se->user_defined)
1387 		return;
1388 
1389 	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1390 	uclamp_se_set(uc_se, default_util_min, false);
1391 }
1392 
1393 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1394 {
1395 	struct rq_flags rf;
1396 	struct rq *rq;
1397 
1398 	if (!rt_task(p))
1399 		return;
1400 
1401 	/* Protect updates to p->uclamp_* */
1402 	rq = task_rq_lock(p, &rf);
1403 	__uclamp_update_util_min_rt_default(p);
1404 	task_rq_unlock(rq, p, &rf);
1405 }
1406 
1407 static void uclamp_sync_util_min_rt_default(void)
1408 {
1409 	struct task_struct *g, *p;
1410 
1411 	/*
1412 	 * copy_process()			sysctl_uclamp
1413 	 *					  uclamp_min_rt = X;
1414 	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
1415 	 *   // link thread			  smp_mb__after_spinlock()
1416 	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
1417 	 *   sched_post_fork()			  for_each_process_thread()
1418 	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
1419 	 *
1420 	 * Ensures that either sched_post_fork() will observe the new
1421 	 * uclamp_min_rt or for_each_process_thread() will observe the new
1422 	 * task.
1423 	 */
1424 	read_lock(&tasklist_lock);
1425 	smp_mb__after_spinlock();
1426 	read_unlock(&tasklist_lock);
1427 
1428 	rcu_read_lock();
1429 	for_each_process_thread(g, p)
1430 		uclamp_update_util_min_rt_default(p);
1431 	rcu_read_unlock();
1432 }
1433 
1434 static inline struct uclamp_se
1435 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1436 {
1437 	/* Copy by value as we could modify it */
1438 	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1439 #ifdef CONFIG_UCLAMP_TASK_GROUP
1440 	unsigned int tg_min, tg_max, value;
1441 
1442 	/*
1443 	 * Tasks in autogroups or root task group will be
1444 	 * restricted by system defaults.
1445 	 */
1446 	if (task_group_is_autogroup(task_group(p)))
1447 		return uc_req;
1448 	if (task_group(p) == &root_task_group)
1449 		return uc_req;
1450 
1451 	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1452 	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1453 	value = uc_req.value;
1454 	value = clamp(value, tg_min, tg_max);
1455 	uclamp_se_set(&uc_req, value, false);
1456 #endif
1457 
1458 	return uc_req;
1459 }
1460 
1461 /*
1462  * The effective clamp bucket index of a task depends on, by increasing
1463  * priority:
1464  * - the task specific clamp value, when explicitly requested from userspace
1465  * - the task group effective clamp value, for tasks not either in the root
1466  *   group or in an autogroup
1467  * - the system default clamp value, defined by the sysadmin
1468  */
1469 static inline struct uclamp_se
1470 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1471 {
1472 	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1473 	struct uclamp_se uc_max = uclamp_default[clamp_id];
1474 
1475 	/* System default restrictions always apply */
1476 	if (unlikely(uc_req.value > uc_max.value))
1477 		return uc_max;
1478 
1479 	return uc_req;
1480 }
1481 
1482 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1483 {
1484 	struct uclamp_se uc_eff;
1485 
1486 	/* Task currently refcounted: use back-annotated (effective) value */
1487 	if (p->uclamp[clamp_id].active)
1488 		return (unsigned long)p->uclamp[clamp_id].value;
1489 
1490 	uc_eff = uclamp_eff_get(p, clamp_id);
1491 
1492 	return (unsigned long)uc_eff.value;
1493 }
1494 
1495 /*
1496  * When a task is enqueued on a rq, the clamp bucket currently defined by the
1497  * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1498  * updates the rq's clamp value if required.
1499  *
1500  * Tasks can have a task-specific value requested from user-space, track
1501  * within each bucket the maximum value for tasks refcounted in it.
1502  * This "local max aggregation" allows to track the exact "requested" value
1503  * for each bucket when all its RUNNABLE tasks require the same clamp.
1504  */
1505 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1506 				    enum uclamp_id clamp_id)
1507 {
1508 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1509 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1510 	struct uclamp_bucket *bucket;
1511 
1512 	lockdep_assert_rq_held(rq);
1513 
1514 	/* Update task effective clamp */
1515 	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1516 
1517 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1518 	bucket->tasks++;
1519 	uc_se->active = true;
1520 
1521 	uclamp_idle_reset(rq, clamp_id, uc_se->value);
1522 
1523 	/*
1524 	 * Local max aggregation: rq buckets always track the max
1525 	 * "requested" clamp value of its RUNNABLE tasks.
1526 	 */
1527 	if (bucket->tasks == 1 || uc_se->value > bucket->value)
1528 		bucket->value = uc_se->value;
1529 
1530 	if (uc_se->value > READ_ONCE(uc_rq->value))
1531 		WRITE_ONCE(uc_rq->value, uc_se->value);
1532 }
1533 
1534 /*
1535  * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1536  * is released. If this is the last task reference counting the rq's max
1537  * active clamp value, then the rq's clamp value is updated.
1538  *
1539  * Both refcounted tasks and rq's cached clamp values are expected to be
1540  * always valid. If it's detected they are not, as defensive programming,
1541  * enforce the expected state and warn.
1542  */
1543 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1544 				    enum uclamp_id clamp_id)
1545 {
1546 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1547 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1548 	struct uclamp_bucket *bucket;
1549 	unsigned int bkt_clamp;
1550 	unsigned int rq_clamp;
1551 
1552 	lockdep_assert_rq_held(rq);
1553 
1554 	/*
1555 	 * If sched_uclamp_used was enabled after task @p was enqueued,
1556 	 * we could end up with unbalanced call to uclamp_rq_dec_id().
1557 	 *
1558 	 * In this case the uc_se->active flag should be false since no uclamp
1559 	 * accounting was performed at enqueue time and we can just return
1560 	 * here.
1561 	 *
1562 	 * Need to be careful of the following enqueue/dequeue ordering
1563 	 * problem too
1564 	 *
1565 	 *	enqueue(taskA)
1566 	 *	// sched_uclamp_used gets enabled
1567 	 *	enqueue(taskB)
1568 	 *	dequeue(taskA)
1569 	 *	// Must not decrement bucket->tasks here
1570 	 *	dequeue(taskB)
1571 	 *
1572 	 * where we could end up with stale data in uc_se and
1573 	 * bucket[uc_se->bucket_id].
1574 	 *
1575 	 * The following check here eliminates the possibility of such race.
1576 	 */
1577 	if (unlikely(!uc_se->active))
1578 		return;
1579 
1580 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1581 
1582 	SCHED_WARN_ON(!bucket->tasks);
1583 	if (likely(bucket->tasks))
1584 		bucket->tasks--;
1585 
1586 	uc_se->active = false;
1587 
1588 	/*
1589 	 * Keep "local max aggregation" simple and accept to (possibly)
1590 	 * overboost some RUNNABLE tasks in the same bucket.
1591 	 * The rq clamp bucket value is reset to its base value whenever
1592 	 * there are no more RUNNABLE tasks refcounting it.
1593 	 */
1594 	if (likely(bucket->tasks))
1595 		return;
1596 
1597 	rq_clamp = READ_ONCE(uc_rq->value);
1598 	/*
1599 	 * Defensive programming: this should never happen. If it happens,
1600 	 * e.g. due to future modification, warn and fixup the expected value.
1601 	 */
1602 	SCHED_WARN_ON(bucket->value > rq_clamp);
1603 	if (bucket->value >= rq_clamp) {
1604 		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1605 		WRITE_ONCE(uc_rq->value, bkt_clamp);
1606 	}
1607 }
1608 
1609 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1610 {
1611 	enum uclamp_id clamp_id;
1612 
1613 	/*
1614 	 * Avoid any overhead until uclamp is actually used by the userspace.
1615 	 *
1616 	 * The condition is constructed such that a NOP is generated when
1617 	 * sched_uclamp_used is disabled.
1618 	 */
1619 	if (!static_branch_unlikely(&sched_uclamp_used))
1620 		return;
1621 
1622 	if (unlikely(!p->sched_class->uclamp_enabled))
1623 		return;
1624 
1625 	for_each_clamp_id(clamp_id)
1626 		uclamp_rq_inc_id(rq, p, clamp_id);
1627 
1628 	/* Reset clamp idle holding when there is one RUNNABLE task */
1629 	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1630 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1631 }
1632 
1633 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1634 {
1635 	enum uclamp_id clamp_id;
1636 
1637 	/*
1638 	 * Avoid any overhead until uclamp is actually used by the userspace.
1639 	 *
1640 	 * The condition is constructed such that a NOP is generated when
1641 	 * sched_uclamp_used is disabled.
1642 	 */
1643 	if (!static_branch_unlikely(&sched_uclamp_used))
1644 		return;
1645 
1646 	if (unlikely(!p->sched_class->uclamp_enabled))
1647 		return;
1648 
1649 	for_each_clamp_id(clamp_id)
1650 		uclamp_rq_dec_id(rq, p, clamp_id);
1651 }
1652 
1653 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1654 				      enum uclamp_id clamp_id)
1655 {
1656 	if (!p->uclamp[clamp_id].active)
1657 		return;
1658 
1659 	uclamp_rq_dec_id(rq, p, clamp_id);
1660 	uclamp_rq_inc_id(rq, p, clamp_id);
1661 
1662 	/*
1663 	 * Make sure to clear the idle flag if we've transiently reached 0
1664 	 * active tasks on rq.
1665 	 */
1666 	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1667 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1668 }
1669 
1670 static inline void
1671 uclamp_update_active(struct task_struct *p)
1672 {
1673 	enum uclamp_id clamp_id;
1674 	struct rq_flags rf;
1675 	struct rq *rq;
1676 
1677 	/*
1678 	 * Lock the task and the rq where the task is (or was) queued.
1679 	 *
1680 	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1681 	 * price to pay to safely serialize util_{min,max} updates with
1682 	 * enqueues, dequeues and migration operations.
1683 	 * This is the same locking schema used by __set_cpus_allowed_ptr().
1684 	 */
1685 	rq = task_rq_lock(p, &rf);
1686 
1687 	/*
1688 	 * Setting the clamp bucket is serialized by task_rq_lock().
1689 	 * If the task is not yet RUNNABLE and its task_struct is not
1690 	 * affecting a valid clamp bucket, the next time it's enqueued,
1691 	 * it will already see the updated clamp bucket value.
1692 	 */
1693 	for_each_clamp_id(clamp_id)
1694 		uclamp_rq_reinc_id(rq, p, clamp_id);
1695 
1696 	task_rq_unlock(rq, p, &rf);
1697 }
1698 
1699 #ifdef CONFIG_UCLAMP_TASK_GROUP
1700 static inline void
1701 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1702 {
1703 	struct css_task_iter it;
1704 	struct task_struct *p;
1705 
1706 	css_task_iter_start(css, 0, &it);
1707 	while ((p = css_task_iter_next(&it)))
1708 		uclamp_update_active(p);
1709 	css_task_iter_end(&it);
1710 }
1711 
1712 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1713 static void uclamp_update_root_tg(void)
1714 {
1715 	struct task_group *tg = &root_task_group;
1716 
1717 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1718 		      sysctl_sched_uclamp_util_min, false);
1719 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1720 		      sysctl_sched_uclamp_util_max, false);
1721 
1722 	rcu_read_lock();
1723 	cpu_util_update_eff(&root_task_group.css);
1724 	rcu_read_unlock();
1725 }
1726 #else
1727 static void uclamp_update_root_tg(void) { }
1728 #endif
1729 
1730 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1731 				void *buffer, size_t *lenp, loff_t *ppos)
1732 {
1733 	bool update_root_tg = false;
1734 	int old_min, old_max, old_min_rt;
1735 	int result;
1736 
1737 	mutex_lock(&uclamp_mutex);
1738 	old_min = sysctl_sched_uclamp_util_min;
1739 	old_max = sysctl_sched_uclamp_util_max;
1740 	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1741 
1742 	result = proc_dointvec(table, write, buffer, lenp, ppos);
1743 	if (result)
1744 		goto undo;
1745 	if (!write)
1746 		goto done;
1747 
1748 	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1749 	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
1750 	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1751 
1752 		result = -EINVAL;
1753 		goto undo;
1754 	}
1755 
1756 	if (old_min != sysctl_sched_uclamp_util_min) {
1757 		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1758 			      sysctl_sched_uclamp_util_min, false);
1759 		update_root_tg = true;
1760 	}
1761 	if (old_max != sysctl_sched_uclamp_util_max) {
1762 		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1763 			      sysctl_sched_uclamp_util_max, false);
1764 		update_root_tg = true;
1765 	}
1766 
1767 	if (update_root_tg) {
1768 		static_branch_enable(&sched_uclamp_used);
1769 		uclamp_update_root_tg();
1770 	}
1771 
1772 	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1773 		static_branch_enable(&sched_uclamp_used);
1774 		uclamp_sync_util_min_rt_default();
1775 	}
1776 
1777 	/*
1778 	 * We update all RUNNABLE tasks only when task groups are in use.
1779 	 * Otherwise, keep it simple and do just a lazy update at each next
1780 	 * task enqueue time.
1781 	 */
1782 
1783 	goto done;
1784 
1785 undo:
1786 	sysctl_sched_uclamp_util_min = old_min;
1787 	sysctl_sched_uclamp_util_max = old_max;
1788 	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1789 done:
1790 	mutex_unlock(&uclamp_mutex);
1791 
1792 	return result;
1793 }
1794 
1795 static int uclamp_validate(struct task_struct *p,
1796 			   const struct sched_attr *attr)
1797 {
1798 	int util_min = p->uclamp_req[UCLAMP_MIN].value;
1799 	int util_max = p->uclamp_req[UCLAMP_MAX].value;
1800 
1801 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1802 		util_min = attr->sched_util_min;
1803 
1804 		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1805 			return -EINVAL;
1806 	}
1807 
1808 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1809 		util_max = attr->sched_util_max;
1810 
1811 		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1812 			return -EINVAL;
1813 	}
1814 
1815 	if (util_min != -1 && util_max != -1 && util_min > util_max)
1816 		return -EINVAL;
1817 
1818 	/*
1819 	 * We have valid uclamp attributes; make sure uclamp is enabled.
1820 	 *
1821 	 * We need to do that here, because enabling static branches is a
1822 	 * blocking operation which obviously cannot be done while holding
1823 	 * scheduler locks.
1824 	 */
1825 	static_branch_enable(&sched_uclamp_used);
1826 
1827 	return 0;
1828 }
1829 
1830 static bool uclamp_reset(const struct sched_attr *attr,
1831 			 enum uclamp_id clamp_id,
1832 			 struct uclamp_se *uc_se)
1833 {
1834 	/* Reset on sched class change for a non user-defined clamp value. */
1835 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1836 	    !uc_se->user_defined)
1837 		return true;
1838 
1839 	/* Reset on sched_util_{min,max} == -1. */
1840 	if (clamp_id == UCLAMP_MIN &&
1841 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1842 	    attr->sched_util_min == -1) {
1843 		return true;
1844 	}
1845 
1846 	if (clamp_id == UCLAMP_MAX &&
1847 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1848 	    attr->sched_util_max == -1) {
1849 		return true;
1850 	}
1851 
1852 	return false;
1853 }
1854 
1855 static void __setscheduler_uclamp(struct task_struct *p,
1856 				  const struct sched_attr *attr)
1857 {
1858 	enum uclamp_id clamp_id;
1859 
1860 	for_each_clamp_id(clamp_id) {
1861 		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1862 		unsigned int value;
1863 
1864 		if (!uclamp_reset(attr, clamp_id, uc_se))
1865 			continue;
1866 
1867 		/*
1868 		 * RT by default have a 100% boost value that could be modified
1869 		 * at runtime.
1870 		 */
1871 		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1872 			value = sysctl_sched_uclamp_util_min_rt_default;
1873 		else
1874 			value = uclamp_none(clamp_id);
1875 
1876 		uclamp_se_set(uc_se, value, false);
1877 
1878 	}
1879 
1880 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1881 		return;
1882 
1883 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1884 	    attr->sched_util_min != -1) {
1885 		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1886 			      attr->sched_util_min, true);
1887 	}
1888 
1889 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1890 	    attr->sched_util_max != -1) {
1891 		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1892 			      attr->sched_util_max, true);
1893 	}
1894 }
1895 
1896 static void uclamp_fork(struct task_struct *p)
1897 {
1898 	enum uclamp_id clamp_id;
1899 
1900 	/*
1901 	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1902 	 * as the task is still at its early fork stages.
1903 	 */
1904 	for_each_clamp_id(clamp_id)
1905 		p->uclamp[clamp_id].active = false;
1906 
1907 	if (likely(!p->sched_reset_on_fork))
1908 		return;
1909 
1910 	for_each_clamp_id(clamp_id) {
1911 		uclamp_se_set(&p->uclamp_req[clamp_id],
1912 			      uclamp_none(clamp_id), false);
1913 	}
1914 }
1915 
1916 static void uclamp_post_fork(struct task_struct *p)
1917 {
1918 	uclamp_update_util_min_rt_default(p);
1919 }
1920 
1921 static void __init init_uclamp_rq(struct rq *rq)
1922 {
1923 	enum uclamp_id clamp_id;
1924 	struct uclamp_rq *uc_rq = rq->uclamp;
1925 
1926 	for_each_clamp_id(clamp_id) {
1927 		uc_rq[clamp_id] = (struct uclamp_rq) {
1928 			.value = uclamp_none(clamp_id)
1929 		};
1930 	}
1931 
1932 	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1933 }
1934 
1935 static void __init init_uclamp(void)
1936 {
1937 	struct uclamp_se uc_max = {};
1938 	enum uclamp_id clamp_id;
1939 	int cpu;
1940 
1941 	for_each_possible_cpu(cpu)
1942 		init_uclamp_rq(cpu_rq(cpu));
1943 
1944 	for_each_clamp_id(clamp_id) {
1945 		uclamp_se_set(&init_task.uclamp_req[clamp_id],
1946 			      uclamp_none(clamp_id), false);
1947 	}
1948 
1949 	/* System defaults allow max clamp values for both indexes */
1950 	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1951 	for_each_clamp_id(clamp_id) {
1952 		uclamp_default[clamp_id] = uc_max;
1953 #ifdef CONFIG_UCLAMP_TASK_GROUP
1954 		root_task_group.uclamp_req[clamp_id] = uc_max;
1955 		root_task_group.uclamp[clamp_id] = uc_max;
1956 #endif
1957 	}
1958 }
1959 
1960 #else /* CONFIG_UCLAMP_TASK */
1961 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1962 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1963 static inline int uclamp_validate(struct task_struct *p,
1964 				  const struct sched_attr *attr)
1965 {
1966 	return -EOPNOTSUPP;
1967 }
1968 static void __setscheduler_uclamp(struct task_struct *p,
1969 				  const struct sched_attr *attr) { }
1970 static inline void uclamp_fork(struct task_struct *p) { }
1971 static inline void uclamp_post_fork(struct task_struct *p) { }
1972 static inline void init_uclamp(void) { }
1973 #endif /* CONFIG_UCLAMP_TASK */
1974 
1975 bool sched_task_on_rq(struct task_struct *p)
1976 {
1977 	return task_on_rq_queued(p);
1978 }
1979 
1980 unsigned long get_wchan(struct task_struct *p)
1981 {
1982 	unsigned long ip = 0;
1983 	unsigned int state;
1984 
1985 	if (!p || p == current)
1986 		return 0;
1987 
1988 	/* Only get wchan if task is blocked and we can keep it that way. */
1989 	raw_spin_lock_irq(&p->pi_lock);
1990 	state = READ_ONCE(p->__state);
1991 	smp_rmb(); /* see try_to_wake_up() */
1992 	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
1993 		ip = __get_wchan(p);
1994 	raw_spin_unlock_irq(&p->pi_lock);
1995 
1996 	return ip;
1997 }
1998 
1999 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2000 {
2001 	if (!(flags & ENQUEUE_NOCLOCK))
2002 		update_rq_clock(rq);
2003 
2004 	if (!(flags & ENQUEUE_RESTORE)) {
2005 		sched_info_enqueue(rq, p);
2006 		psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2007 	}
2008 
2009 	uclamp_rq_inc(rq, p);
2010 	p->sched_class->enqueue_task(rq, p, flags);
2011 
2012 	if (sched_core_enabled(rq))
2013 		sched_core_enqueue(rq, p);
2014 }
2015 
2016 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2017 {
2018 	if (sched_core_enabled(rq))
2019 		sched_core_dequeue(rq, p, flags);
2020 
2021 	if (!(flags & DEQUEUE_NOCLOCK))
2022 		update_rq_clock(rq);
2023 
2024 	if (!(flags & DEQUEUE_SAVE)) {
2025 		sched_info_dequeue(rq, p);
2026 		psi_dequeue(p, flags & DEQUEUE_SLEEP);
2027 	}
2028 
2029 	uclamp_rq_dec(rq, p);
2030 	p->sched_class->dequeue_task(rq, p, flags);
2031 }
2032 
2033 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2034 {
2035 	enqueue_task(rq, p, flags);
2036 
2037 	p->on_rq = TASK_ON_RQ_QUEUED;
2038 }
2039 
2040 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2041 {
2042 	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2043 
2044 	dequeue_task(rq, p, flags);
2045 }
2046 
2047 static inline int __normal_prio(int policy, int rt_prio, int nice)
2048 {
2049 	int prio;
2050 
2051 	if (dl_policy(policy))
2052 		prio = MAX_DL_PRIO - 1;
2053 	else if (rt_policy(policy))
2054 		prio = MAX_RT_PRIO - 1 - rt_prio;
2055 	else
2056 		prio = NICE_TO_PRIO(nice);
2057 
2058 	return prio;
2059 }
2060 
2061 /*
2062  * Calculate the expected normal priority: i.e. priority
2063  * without taking RT-inheritance into account. Might be
2064  * boosted by interactivity modifiers. Changes upon fork,
2065  * setprio syscalls, and whenever the interactivity
2066  * estimator recalculates.
2067  */
2068 static inline int normal_prio(struct task_struct *p)
2069 {
2070 	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2071 }
2072 
2073 /*
2074  * Calculate the current priority, i.e. the priority
2075  * taken into account by the scheduler. This value might
2076  * be boosted by RT tasks, or might be boosted by
2077  * interactivity modifiers. Will be RT if the task got
2078  * RT-boosted. If not then it returns p->normal_prio.
2079  */
2080 static int effective_prio(struct task_struct *p)
2081 {
2082 	p->normal_prio = normal_prio(p);
2083 	/*
2084 	 * If we are RT tasks or we were boosted to RT priority,
2085 	 * keep the priority unchanged. Otherwise, update priority
2086 	 * to the normal priority:
2087 	 */
2088 	if (!rt_prio(p->prio))
2089 		return p->normal_prio;
2090 	return p->prio;
2091 }
2092 
2093 /**
2094  * task_curr - is this task currently executing on a CPU?
2095  * @p: the task in question.
2096  *
2097  * Return: 1 if the task is currently executing. 0 otherwise.
2098  */
2099 inline int task_curr(const struct task_struct *p)
2100 {
2101 	return cpu_curr(task_cpu(p)) == p;
2102 }
2103 
2104 /*
2105  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2106  * use the balance_callback list if you want balancing.
2107  *
2108  * this means any call to check_class_changed() must be followed by a call to
2109  * balance_callback().
2110  */
2111 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2112 				       const struct sched_class *prev_class,
2113 				       int oldprio)
2114 {
2115 	if (prev_class != p->sched_class) {
2116 		if (prev_class->switched_from)
2117 			prev_class->switched_from(rq, p);
2118 
2119 		p->sched_class->switched_to(rq, p);
2120 	} else if (oldprio != p->prio || dl_task(p))
2121 		p->sched_class->prio_changed(rq, p, oldprio);
2122 }
2123 
2124 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2125 {
2126 	if (p->sched_class == rq->curr->sched_class)
2127 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2128 	else if (p->sched_class > rq->curr->sched_class)
2129 		resched_curr(rq);
2130 
2131 	/*
2132 	 * A queue event has occurred, and we're going to schedule.  In
2133 	 * this case, we can save a useless back to back clock update.
2134 	 */
2135 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2136 		rq_clock_skip_update(rq);
2137 }
2138 
2139 #ifdef CONFIG_SMP
2140 
2141 static void
2142 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2143 
2144 static int __set_cpus_allowed_ptr(struct task_struct *p,
2145 				  const struct cpumask *new_mask,
2146 				  u32 flags);
2147 
2148 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2149 {
2150 	if (likely(!p->migration_disabled))
2151 		return;
2152 
2153 	if (p->cpus_ptr != &p->cpus_mask)
2154 		return;
2155 
2156 	/*
2157 	 * Violates locking rules! see comment in __do_set_cpus_allowed().
2158 	 */
2159 	__do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2160 }
2161 
2162 void migrate_disable(void)
2163 {
2164 	struct task_struct *p = current;
2165 
2166 	if (p->migration_disabled) {
2167 		p->migration_disabled++;
2168 		return;
2169 	}
2170 
2171 	preempt_disable();
2172 	this_rq()->nr_pinned++;
2173 	p->migration_disabled = 1;
2174 	preempt_enable();
2175 }
2176 EXPORT_SYMBOL_GPL(migrate_disable);
2177 
2178 void migrate_enable(void)
2179 {
2180 	struct task_struct *p = current;
2181 
2182 	if (p->migration_disabled > 1) {
2183 		p->migration_disabled--;
2184 		return;
2185 	}
2186 
2187 	if (WARN_ON_ONCE(!p->migration_disabled))
2188 		return;
2189 
2190 	/*
2191 	 * Ensure stop_task runs either before or after this, and that
2192 	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2193 	 */
2194 	preempt_disable();
2195 	if (p->cpus_ptr != &p->cpus_mask)
2196 		__set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2197 	/*
2198 	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2199 	 * regular cpus_mask, otherwise things that race (eg.
2200 	 * select_fallback_rq) get confused.
2201 	 */
2202 	barrier();
2203 	p->migration_disabled = 0;
2204 	this_rq()->nr_pinned--;
2205 	preempt_enable();
2206 }
2207 EXPORT_SYMBOL_GPL(migrate_enable);
2208 
2209 static inline bool rq_has_pinned_tasks(struct rq *rq)
2210 {
2211 	return rq->nr_pinned;
2212 }
2213 
2214 /*
2215  * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2216  * __set_cpus_allowed_ptr() and select_fallback_rq().
2217  */
2218 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2219 {
2220 	/* When not in the task's cpumask, no point in looking further. */
2221 	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2222 		return false;
2223 
2224 	/* migrate_disabled() must be allowed to finish. */
2225 	if (is_migration_disabled(p))
2226 		return cpu_online(cpu);
2227 
2228 	/* Non kernel threads are not allowed during either online or offline. */
2229 	if (!(p->flags & PF_KTHREAD))
2230 		return cpu_active(cpu) && task_cpu_possible(cpu, p);
2231 
2232 	/* KTHREAD_IS_PER_CPU is always allowed. */
2233 	if (kthread_is_per_cpu(p))
2234 		return cpu_online(cpu);
2235 
2236 	/* Regular kernel threads don't get to stay during offline. */
2237 	if (cpu_dying(cpu))
2238 		return false;
2239 
2240 	/* But are allowed during online. */
2241 	return cpu_online(cpu);
2242 }
2243 
2244 /*
2245  * This is how migration works:
2246  *
2247  * 1) we invoke migration_cpu_stop() on the target CPU using
2248  *    stop_one_cpu().
2249  * 2) stopper starts to run (implicitly forcing the migrated thread
2250  *    off the CPU)
2251  * 3) it checks whether the migrated task is still in the wrong runqueue.
2252  * 4) if it's in the wrong runqueue then the migration thread removes
2253  *    it and puts it into the right queue.
2254  * 5) stopper completes and stop_one_cpu() returns and the migration
2255  *    is done.
2256  */
2257 
2258 /*
2259  * move_queued_task - move a queued task to new rq.
2260  *
2261  * Returns (locked) new rq. Old rq's lock is released.
2262  */
2263 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2264 				   struct task_struct *p, int new_cpu)
2265 {
2266 	lockdep_assert_rq_held(rq);
2267 
2268 	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2269 	set_task_cpu(p, new_cpu);
2270 	rq_unlock(rq, rf);
2271 
2272 	rq = cpu_rq(new_cpu);
2273 
2274 	rq_lock(rq, rf);
2275 	BUG_ON(task_cpu(p) != new_cpu);
2276 	activate_task(rq, p, 0);
2277 	check_preempt_curr(rq, p, 0);
2278 
2279 	return rq;
2280 }
2281 
2282 struct migration_arg {
2283 	struct task_struct		*task;
2284 	int				dest_cpu;
2285 	struct set_affinity_pending	*pending;
2286 };
2287 
2288 /*
2289  * @refs: number of wait_for_completion()
2290  * @stop_pending: is @stop_work in use
2291  */
2292 struct set_affinity_pending {
2293 	refcount_t		refs;
2294 	unsigned int		stop_pending;
2295 	struct completion	done;
2296 	struct cpu_stop_work	stop_work;
2297 	struct migration_arg	arg;
2298 };
2299 
2300 /*
2301  * Move (not current) task off this CPU, onto the destination CPU. We're doing
2302  * this because either it can't run here any more (set_cpus_allowed()
2303  * away from this CPU, or CPU going down), or because we're
2304  * attempting to rebalance this task on exec (sched_exec).
2305  *
2306  * So we race with normal scheduler movements, but that's OK, as long
2307  * as the task is no longer on this CPU.
2308  */
2309 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2310 				 struct task_struct *p, int dest_cpu)
2311 {
2312 	/* Affinity changed (again). */
2313 	if (!is_cpu_allowed(p, dest_cpu))
2314 		return rq;
2315 
2316 	update_rq_clock(rq);
2317 	rq = move_queued_task(rq, rf, p, dest_cpu);
2318 
2319 	return rq;
2320 }
2321 
2322 /*
2323  * migration_cpu_stop - this will be executed by a highprio stopper thread
2324  * and performs thread migration by bumping thread off CPU then
2325  * 'pushing' onto another runqueue.
2326  */
2327 static int migration_cpu_stop(void *data)
2328 {
2329 	struct migration_arg *arg = data;
2330 	struct set_affinity_pending *pending = arg->pending;
2331 	struct task_struct *p = arg->task;
2332 	struct rq *rq = this_rq();
2333 	bool complete = false;
2334 	struct rq_flags rf;
2335 
2336 	/*
2337 	 * The original target CPU might have gone down and we might
2338 	 * be on another CPU but it doesn't matter.
2339 	 */
2340 	local_irq_save(rf.flags);
2341 	/*
2342 	 * We need to explicitly wake pending tasks before running
2343 	 * __migrate_task() such that we will not miss enforcing cpus_ptr
2344 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2345 	 */
2346 	flush_smp_call_function_from_idle();
2347 
2348 	raw_spin_lock(&p->pi_lock);
2349 	rq_lock(rq, &rf);
2350 
2351 	/*
2352 	 * If we were passed a pending, then ->stop_pending was set, thus
2353 	 * p->migration_pending must have remained stable.
2354 	 */
2355 	WARN_ON_ONCE(pending && pending != p->migration_pending);
2356 
2357 	/*
2358 	 * If task_rq(p) != rq, it cannot be migrated here, because we're
2359 	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2360 	 * we're holding p->pi_lock.
2361 	 */
2362 	if (task_rq(p) == rq) {
2363 		if (is_migration_disabled(p))
2364 			goto out;
2365 
2366 		if (pending) {
2367 			p->migration_pending = NULL;
2368 			complete = true;
2369 
2370 			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2371 				goto out;
2372 		}
2373 
2374 		if (task_on_rq_queued(p))
2375 			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2376 		else
2377 			p->wake_cpu = arg->dest_cpu;
2378 
2379 		/*
2380 		 * XXX __migrate_task() can fail, at which point we might end
2381 		 * up running on a dodgy CPU, AFAICT this can only happen
2382 		 * during CPU hotplug, at which point we'll get pushed out
2383 		 * anyway, so it's probably not a big deal.
2384 		 */
2385 
2386 	} else if (pending) {
2387 		/*
2388 		 * This happens when we get migrated between migrate_enable()'s
2389 		 * preempt_enable() and scheduling the stopper task. At that
2390 		 * point we're a regular task again and not current anymore.
2391 		 *
2392 		 * A !PREEMPT kernel has a giant hole here, which makes it far
2393 		 * more likely.
2394 		 */
2395 
2396 		/*
2397 		 * The task moved before the stopper got to run. We're holding
2398 		 * ->pi_lock, so the allowed mask is stable - if it got
2399 		 * somewhere allowed, we're done.
2400 		 */
2401 		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2402 			p->migration_pending = NULL;
2403 			complete = true;
2404 			goto out;
2405 		}
2406 
2407 		/*
2408 		 * When migrate_enable() hits a rq mis-match we can't reliably
2409 		 * determine is_migration_disabled() and so have to chase after
2410 		 * it.
2411 		 */
2412 		WARN_ON_ONCE(!pending->stop_pending);
2413 		task_rq_unlock(rq, p, &rf);
2414 		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2415 				    &pending->arg, &pending->stop_work);
2416 		return 0;
2417 	}
2418 out:
2419 	if (pending)
2420 		pending->stop_pending = false;
2421 	task_rq_unlock(rq, p, &rf);
2422 
2423 	if (complete)
2424 		complete_all(&pending->done);
2425 
2426 	return 0;
2427 }
2428 
2429 int push_cpu_stop(void *arg)
2430 {
2431 	struct rq *lowest_rq = NULL, *rq = this_rq();
2432 	struct task_struct *p = arg;
2433 
2434 	raw_spin_lock_irq(&p->pi_lock);
2435 	raw_spin_rq_lock(rq);
2436 
2437 	if (task_rq(p) != rq)
2438 		goto out_unlock;
2439 
2440 	if (is_migration_disabled(p)) {
2441 		p->migration_flags |= MDF_PUSH;
2442 		goto out_unlock;
2443 	}
2444 
2445 	p->migration_flags &= ~MDF_PUSH;
2446 
2447 	if (p->sched_class->find_lock_rq)
2448 		lowest_rq = p->sched_class->find_lock_rq(p, rq);
2449 
2450 	if (!lowest_rq)
2451 		goto out_unlock;
2452 
2453 	// XXX validate p is still the highest prio task
2454 	if (task_rq(p) == rq) {
2455 		deactivate_task(rq, p, 0);
2456 		set_task_cpu(p, lowest_rq->cpu);
2457 		activate_task(lowest_rq, p, 0);
2458 		resched_curr(lowest_rq);
2459 	}
2460 
2461 	double_unlock_balance(rq, lowest_rq);
2462 
2463 out_unlock:
2464 	rq->push_busy = false;
2465 	raw_spin_rq_unlock(rq);
2466 	raw_spin_unlock_irq(&p->pi_lock);
2467 
2468 	put_task_struct(p);
2469 	return 0;
2470 }
2471 
2472 /*
2473  * sched_class::set_cpus_allowed must do the below, but is not required to
2474  * actually call this function.
2475  */
2476 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2477 {
2478 	if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2479 		p->cpus_ptr = new_mask;
2480 		return;
2481 	}
2482 
2483 	cpumask_copy(&p->cpus_mask, new_mask);
2484 	p->nr_cpus_allowed = cpumask_weight(new_mask);
2485 }
2486 
2487 static void
2488 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2489 {
2490 	struct rq *rq = task_rq(p);
2491 	bool queued, running;
2492 
2493 	/*
2494 	 * This here violates the locking rules for affinity, since we're only
2495 	 * supposed to change these variables while holding both rq->lock and
2496 	 * p->pi_lock.
2497 	 *
2498 	 * HOWEVER, it magically works, because ttwu() is the only code that
2499 	 * accesses these variables under p->pi_lock and only does so after
2500 	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2501 	 * before finish_task().
2502 	 *
2503 	 * XXX do further audits, this smells like something putrid.
2504 	 */
2505 	if (flags & SCA_MIGRATE_DISABLE)
2506 		SCHED_WARN_ON(!p->on_cpu);
2507 	else
2508 		lockdep_assert_held(&p->pi_lock);
2509 
2510 	queued = task_on_rq_queued(p);
2511 	running = task_current(rq, p);
2512 
2513 	if (queued) {
2514 		/*
2515 		 * Because __kthread_bind() calls this on blocked tasks without
2516 		 * holding rq->lock.
2517 		 */
2518 		lockdep_assert_rq_held(rq);
2519 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2520 	}
2521 	if (running)
2522 		put_prev_task(rq, p);
2523 
2524 	p->sched_class->set_cpus_allowed(p, new_mask, flags);
2525 
2526 	if (queued)
2527 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2528 	if (running)
2529 		set_next_task(rq, p);
2530 }
2531 
2532 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2533 {
2534 	__do_set_cpus_allowed(p, new_mask, 0);
2535 }
2536 
2537 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2538 		      int node)
2539 {
2540 	if (!src->user_cpus_ptr)
2541 		return 0;
2542 
2543 	dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2544 	if (!dst->user_cpus_ptr)
2545 		return -ENOMEM;
2546 
2547 	cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2548 	return 0;
2549 }
2550 
2551 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2552 {
2553 	struct cpumask *user_mask = NULL;
2554 
2555 	swap(p->user_cpus_ptr, user_mask);
2556 
2557 	return user_mask;
2558 }
2559 
2560 void release_user_cpus_ptr(struct task_struct *p)
2561 {
2562 	kfree(clear_user_cpus_ptr(p));
2563 }
2564 
2565 /*
2566  * This function is wildly self concurrent; here be dragons.
2567  *
2568  *
2569  * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2570  * designated task is enqueued on an allowed CPU. If that task is currently
2571  * running, we have to kick it out using the CPU stopper.
2572  *
2573  * Migrate-Disable comes along and tramples all over our nice sandcastle.
2574  * Consider:
2575  *
2576  *     Initial conditions: P0->cpus_mask = [0, 1]
2577  *
2578  *     P0@CPU0                  P1
2579  *
2580  *     migrate_disable();
2581  *     <preempted>
2582  *                              set_cpus_allowed_ptr(P0, [1]);
2583  *
2584  * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2585  * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2586  * This means we need the following scheme:
2587  *
2588  *     P0@CPU0                  P1
2589  *
2590  *     migrate_disable();
2591  *     <preempted>
2592  *                              set_cpus_allowed_ptr(P0, [1]);
2593  *                                <blocks>
2594  *     <resumes>
2595  *     migrate_enable();
2596  *       __set_cpus_allowed_ptr();
2597  *       <wakes local stopper>
2598  *                         `--> <woken on migration completion>
2599  *
2600  * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2601  * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2602  * task p are serialized by p->pi_lock, which we can leverage: the one that
2603  * should come into effect at the end of the Migrate-Disable region is the last
2604  * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2605  * but we still need to properly signal those waiting tasks at the appropriate
2606  * moment.
2607  *
2608  * This is implemented using struct set_affinity_pending. The first
2609  * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2610  * setup an instance of that struct and install it on the targeted task_struct.
2611  * Any and all further callers will reuse that instance. Those then wait for
2612  * a completion signaled at the tail of the CPU stopper callback (1), triggered
2613  * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2614  *
2615  *
2616  * (1) In the cases covered above. There is one more where the completion is
2617  * signaled within affine_move_task() itself: when a subsequent affinity request
2618  * occurs after the stopper bailed out due to the targeted task still being
2619  * Migrate-Disable. Consider:
2620  *
2621  *     Initial conditions: P0->cpus_mask = [0, 1]
2622  *
2623  *     CPU0		  P1				P2
2624  *     <P0>
2625  *       migrate_disable();
2626  *       <preempted>
2627  *                        set_cpus_allowed_ptr(P0, [1]);
2628  *                          <blocks>
2629  *     <migration/0>
2630  *       migration_cpu_stop()
2631  *         is_migration_disabled()
2632  *           <bails>
2633  *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
2634  *                                                         <signal completion>
2635  *                          <awakes>
2636  *
2637  * Note that the above is safe vs a concurrent migrate_enable(), as any
2638  * pending affinity completion is preceded by an uninstallation of
2639  * p->migration_pending done with p->pi_lock held.
2640  */
2641 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2642 			    int dest_cpu, unsigned int flags)
2643 {
2644 	struct set_affinity_pending my_pending = { }, *pending = NULL;
2645 	bool stop_pending, complete = false;
2646 
2647 	/* Can the task run on the task's current CPU? If so, we're done */
2648 	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2649 		struct task_struct *push_task = NULL;
2650 
2651 		if ((flags & SCA_MIGRATE_ENABLE) &&
2652 		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2653 			rq->push_busy = true;
2654 			push_task = get_task_struct(p);
2655 		}
2656 
2657 		/*
2658 		 * If there are pending waiters, but no pending stop_work,
2659 		 * then complete now.
2660 		 */
2661 		pending = p->migration_pending;
2662 		if (pending && !pending->stop_pending) {
2663 			p->migration_pending = NULL;
2664 			complete = true;
2665 		}
2666 
2667 		task_rq_unlock(rq, p, rf);
2668 
2669 		if (push_task) {
2670 			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2671 					    p, &rq->push_work);
2672 		}
2673 
2674 		if (complete)
2675 			complete_all(&pending->done);
2676 
2677 		return 0;
2678 	}
2679 
2680 	if (!(flags & SCA_MIGRATE_ENABLE)) {
2681 		/* serialized by p->pi_lock */
2682 		if (!p->migration_pending) {
2683 			/* Install the request */
2684 			refcount_set(&my_pending.refs, 1);
2685 			init_completion(&my_pending.done);
2686 			my_pending.arg = (struct migration_arg) {
2687 				.task = p,
2688 				.dest_cpu = dest_cpu,
2689 				.pending = &my_pending,
2690 			};
2691 
2692 			p->migration_pending = &my_pending;
2693 		} else {
2694 			pending = p->migration_pending;
2695 			refcount_inc(&pending->refs);
2696 			/*
2697 			 * Affinity has changed, but we've already installed a
2698 			 * pending. migration_cpu_stop() *must* see this, else
2699 			 * we risk a completion of the pending despite having a
2700 			 * task on a disallowed CPU.
2701 			 *
2702 			 * Serialized by p->pi_lock, so this is safe.
2703 			 */
2704 			pending->arg.dest_cpu = dest_cpu;
2705 		}
2706 	}
2707 	pending = p->migration_pending;
2708 	/*
2709 	 * - !MIGRATE_ENABLE:
2710 	 *   we'll have installed a pending if there wasn't one already.
2711 	 *
2712 	 * - MIGRATE_ENABLE:
2713 	 *   we're here because the current CPU isn't matching anymore,
2714 	 *   the only way that can happen is because of a concurrent
2715 	 *   set_cpus_allowed_ptr() call, which should then still be
2716 	 *   pending completion.
2717 	 *
2718 	 * Either way, we really should have a @pending here.
2719 	 */
2720 	if (WARN_ON_ONCE(!pending)) {
2721 		task_rq_unlock(rq, p, rf);
2722 		return -EINVAL;
2723 	}
2724 
2725 	if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2726 		/*
2727 		 * MIGRATE_ENABLE gets here because 'p == current', but for
2728 		 * anything else we cannot do is_migration_disabled(), punt
2729 		 * and have the stopper function handle it all race-free.
2730 		 */
2731 		stop_pending = pending->stop_pending;
2732 		if (!stop_pending)
2733 			pending->stop_pending = true;
2734 
2735 		if (flags & SCA_MIGRATE_ENABLE)
2736 			p->migration_flags &= ~MDF_PUSH;
2737 
2738 		task_rq_unlock(rq, p, rf);
2739 
2740 		if (!stop_pending) {
2741 			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2742 					    &pending->arg, &pending->stop_work);
2743 		}
2744 
2745 		if (flags & SCA_MIGRATE_ENABLE)
2746 			return 0;
2747 	} else {
2748 
2749 		if (!is_migration_disabled(p)) {
2750 			if (task_on_rq_queued(p))
2751 				rq = move_queued_task(rq, rf, p, dest_cpu);
2752 
2753 			if (!pending->stop_pending) {
2754 				p->migration_pending = NULL;
2755 				complete = true;
2756 			}
2757 		}
2758 		task_rq_unlock(rq, p, rf);
2759 
2760 		if (complete)
2761 			complete_all(&pending->done);
2762 	}
2763 
2764 	wait_for_completion(&pending->done);
2765 
2766 	if (refcount_dec_and_test(&pending->refs))
2767 		wake_up_var(&pending->refs); /* No UaF, just an address */
2768 
2769 	/*
2770 	 * Block the original owner of &pending until all subsequent callers
2771 	 * have seen the completion and decremented the refcount
2772 	 */
2773 	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2774 
2775 	/* ARGH */
2776 	WARN_ON_ONCE(my_pending.stop_pending);
2777 
2778 	return 0;
2779 }
2780 
2781 /*
2782  * Called with both p->pi_lock and rq->lock held; drops both before returning.
2783  */
2784 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2785 					 const struct cpumask *new_mask,
2786 					 u32 flags,
2787 					 struct rq *rq,
2788 					 struct rq_flags *rf)
2789 	__releases(rq->lock)
2790 	__releases(p->pi_lock)
2791 {
2792 	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2793 	const struct cpumask *cpu_valid_mask = cpu_active_mask;
2794 	bool kthread = p->flags & PF_KTHREAD;
2795 	struct cpumask *user_mask = NULL;
2796 	unsigned int dest_cpu;
2797 	int ret = 0;
2798 
2799 	update_rq_clock(rq);
2800 
2801 	if (kthread || is_migration_disabled(p)) {
2802 		/*
2803 		 * Kernel threads are allowed on online && !active CPUs,
2804 		 * however, during cpu-hot-unplug, even these might get pushed
2805 		 * away if not KTHREAD_IS_PER_CPU.
2806 		 *
2807 		 * Specifically, migration_disabled() tasks must not fail the
2808 		 * cpumask_any_and_distribute() pick below, esp. so on
2809 		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2810 		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2811 		 */
2812 		cpu_valid_mask = cpu_online_mask;
2813 	}
2814 
2815 	if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2816 		ret = -EINVAL;
2817 		goto out;
2818 	}
2819 
2820 	/*
2821 	 * Must re-check here, to close a race against __kthread_bind(),
2822 	 * sched_setaffinity() is not guaranteed to observe the flag.
2823 	 */
2824 	if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2825 		ret = -EINVAL;
2826 		goto out;
2827 	}
2828 
2829 	if (!(flags & SCA_MIGRATE_ENABLE)) {
2830 		if (cpumask_equal(&p->cpus_mask, new_mask))
2831 			goto out;
2832 
2833 		if (WARN_ON_ONCE(p == current &&
2834 				 is_migration_disabled(p) &&
2835 				 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2836 			ret = -EBUSY;
2837 			goto out;
2838 		}
2839 	}
2840 
2841 	/*
2842 	 * Picking a ~random cpu helps in cases where we are changing affinity
2843 	 * for groups of tasks (ie. cpuset), so that load balancing is not
2844 	 * immediately required to distribute the tasks within their new mask.
2845 	 */
2846 	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2847 	if (dest_cpu >= nr_cpu_ids) {
2848 		ret = -EINVAL;
2849 		goto out;
2850 	}
2851 
2852 	__do_set_cpus_allowed(p, new_mask, flags);
2853 
2854 	if (flags & SCA_USER)
2855 		user_mask = clear_user_cpus_ptr(p);
2856 
2857 	ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2858 
2859 	kfree(user_mask);
2860 
2861 	return ret;
2862 
2863 out:
2864 	task_rq_unlock(rq, p, rf);
2865 
2866 	return ret;
2867 }
2868 
2869 /*
2870  * Change a given task's CPU affinity. Migrate the thread to a
2871  * proper CPU and schedule it away if the CPU it's executing on
2872  * is removed from the allowed bitmask.
2873  *
2874  * NOTE: the caller must have a valid reference to the task, the
2875  * task must not exit() & deallocate itself prematurely. The
2876  * call is not atomic; no spinlocks may be held.
2877  */
2878 static int __set_cpus_allowed_ptr(struct task_struct *p,
2879 				  const struct cpumask *new_mask, u32 flags)
2880 {
2881 	struct rq_flags rf;
2882 	struct rq *rq;
2883 
2884 	rq = task_rq_lock(p, &rf);
2885 	return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2886 }
2887 
2888 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2889 {
2890 	return __set_cpus_allowed_ptr(p, new_mask, 0);
2891 }
2892 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2893 
2894 /*
2895  * Change a given task's CPU affinity to the intersection of its current
2896  * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2897  * and pointing @p->user_cpus_ptr to a copy of the old mask.
2898  * If the resulting mask is empty, leave the affinity unchanged and return
2899  * -EINVAL.
2900  */
2901 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2902 				     struct cpumask *new_mask,
2903 				     const struct cpumask *subset_mask)
2904 {
2905 	struct cpumask *user_mask = NULL;
2906 	struct rq_flags rf;
2907 	struct rq *rq;
2908 	int err;
2909 
2910 	if (!p->user_cpus_ptr) {
2911 		user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2912 		if (!user_mask)
2913 			return -ENOMEM;
2914 	}
2915 
2916 	rq = task_rq_lock(p, &rf);
2917 
2918 	/*
2919 	 * Forcefully restricting the affinity of a deadline task is
2920 	 * likely to cause problems, so fail and noisily override the
2921 	 * mask entirely.
2922 	 */
2923 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2924 		err = -EPERM;
2925 		goto err_unlock;
2926 	}
2927 
2928 	if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2929 		err = -EINVAL;
2930 		goto err_unlock;
2931 	}
2932 
2933 	/*
2934 	 * We're about to butcher the task affinity, so keep track of what
2935 	 * the user asked for in case we're able to restore it later on.
2936 	 */
2937 	if (user_mask) {
2938 		cpumask_copy(user_mask, p->cpus_ptr);
2939 		p->user_cpus_ptr = user_mask;
2940 	}
2941 
2942 	return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2943 
2944 err_unlock:
2945 	task_rq_unlock(rq, p, &rf);
2946 	kfree(user_mask);
2947 	return err;
2948 }
2949 
2950 /*
2951  * Restrict the CPU affinity of task @p so that it is a subset of
2952  * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2953  * old affinity mask. If the resulting mask is empty, we warn and walk
2954  * up the cpuset hierarchy until we find a suitable mask.
2955  */
2956 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2957 {
2958 	cpumask_var_t new_mask;
2959 	const struct cpumask *override_mask = task_cpu_possible_mask(p);
2960 
2961 	alloc_cpumask_var(&new_mask, GFP_KERNEL);
2962 
2963 	/*
2964 	 * __migrate_task() can fail silently in the face of concurrent
2965 	 * offlining of the chosen destination CPU, so take the hotplug
2966 	 * lock to ensure that the migration succeeds.
2967 	 */
2968 	cpus_read_lock();
2969 	if (!cpumask_available(new_mask))
2970 		goto out_set_mask;
2971 
2972 	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2973 		goto out_free_mask;
2974 
2975 	/*
2976 	 * We failed to find a valid subset of the affinity mask for the
2977 	 * task, so override it based on its cpuset hierarchy.
2978 	 */
2979 	cpuset_cpus_allowed(p, new_mask);
2980 	override_mask = new_mask;
2981 
2982 out_set_mask:
2983 	if (printk_ratelimit()) {
2984 		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2985 				task_pid_nr(p), p->comm,
2986 				cpumask_pr_args(override_mask));
2987 	}
2988 
2989 	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2990 out_free_mask:
2991 	cpus_read_unlock();
2992 	free_cpumask_var(new_mask);
2993 }
2994 
2995 static int
2996 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
2997 
2998 /*
2999  * Restore the affinity of a task @p which was previously restricted by a
3000  * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3001  * @p->user_cpus_ptr.
3002  *
3003  * It is the caller's responsibility to serialise this with any calls to
3004  * force_compatible_cpus_allowed_ptr(@p).
3005  */
3006 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3007 {
3008 	struct cpumask *user_mask = p->user_cpus_ptr;
3009 	unsigned long flags;
3010 
3011 	/*
3012 	 * Try to restore the old affinity mask. If this fails, then
3013 	 * we free the mask explicitly to avoid it being inherited across
3014 	 * a subsequent fork().
3015 	 */
3016 	if (!user_mask || !__sched_setaffinity(p, user_mask))
3017 		return;
3018 
3019 	raw_spin_lock_irqsave(&p->pi_lock, flags);
3020 	user_mask = clear_user_cpus_ptr(p);
3021 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3022 
3023 	kfree(user_mask);
3024 }
3025 
3026 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3027 {
3028 #ifdef CONFIG_SCHED_DEBUG
3029 	unsigned int state = READ_ONCE(p->__state);
3030 
3031 	/*
3032 	 * We should never call set_task_cpu() on a blocked task,
3033 	 * ttwu() will sort out the placement.
3034 	 */
3035 	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3036 
3037 	/*
3038 	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3039 	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3040 	 * time relying on p->on_rq.
3041 	 */
3042 	WARN_ON_ONCE(state == TASK_RUNNING &&
3043 		     p->sched_class == &fair_sched_class &&
3044 		     (p->on_rq && !task_on_rq_migrating(p)));
3045 
3046 #ifdef CONFIG_LOCKDEP
3047 	/*
3048 	 * The caller should hold either p->pi_lock or rq->lock, when changing
3049 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3050 	 *
3051 	 * sched_move_task() holds both and thus holding either pins the cgroup,
3052 	 * see task_group().
3053 	 *
3054 	 * Furthermore, all task_rq users should acquire both locks, see
3055 	 * task_rq_lock().
3056 	 */
3057 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3058 				      lockdep_is_held(__rq_lockp(task_rq(p)))));
3059 #endif
3060 	/*
3061 	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3062 	 */
3063 	WARN_ON_ONCE(!cpu_online(new_cpu));
3064 
3065 	WARN_ON_ONCE(is_migration_disabled(p));
3066 #endif
3067 
3068 	trace_sched_migrate_task(p, new_cpu);
3069 
3070 	if (task_cpu(p) != new_cpu) {
3071 		if (p->sched_class->migrate_task_rq)
3072 			p->sched_class->migrate_task_rq(p, new_cpu);
3073 		p->se.nr_migrations++;
3074 		rseq_migrate(p);
3075 		perf_event_task_migrate(p);
3076 	}
3077 
3078 	__set_task_cpu(p, new_cpu);
3079 }
3080 
3081 #ifdef CONFIG_NUMA_BALANCING
3082 static void __migrate_swap_task(struct task_struct *p, int cpu)
3083 {
3084 	if (task_on_rq_queued(p)) {
3085 		struct rq *src_rq, *dst_rq;
3086 		struct rq_flags srf, drf;
3087 
3088 		src_rq = task_rq(p);
3089 		dst_rq = cpu_rq(cpu);
3090 
3091 		rq_pin_lock(src_rq, &srf);
3092 		rq_pin_lock(dst_rq, &drf);
3093 
3094 		deactivate_task(src_rq, p, 0);
3095 		set_task_cpu(p, cpu);
3096 		activate_task(dst_rq, p, 0);
3097 		check_preempt_curr(dst_rq, p, 0);
3098 
3099 		rq_unpin_lock(dst_rq, &drf);
3100 		rq_unpin_lock(src_rq, &srf);
3101 
3102 	} else {
3103 		/*
3104 		 * Task isn't running anymore; make it appear like we migrated
3105 		 * it before it went to sleep. This means on wakeup we make the
3106 		 * previous CPU our target instead of where it really is.
3107 		 */
3108 		p->wake_cpu = cpu;
3109 	}
3110 }
3111 
3112 struct migration_swap_arg {
3113 	struct task_struct *src_task, *dst_task;
3114 	int src_cpu, dst_cpu;
3115 };
3116 
3117 static int migrate_swap_stop(void *data)
3118 {
3119 	struct migration_swap_arg *arg = data;
3120 	struct rq *src_rq, *dst_rq;
3121 	int ret = -EAGAIN;
3122 
3123 	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3124 		return -EAGAIN;
3125 
3126 	src_rq = cpu_rq(arg->src_cpu);
3127 	dst_rq = cpu_rq(arg->dst_cpu);
3128 
3129 	double_raw_lock(&arg->src_task->pi_lock,
3130 			&arg->dst_task->pi_lock);
3131 	double_rq_lock(src_rq, dst_rq);
3132 
3133 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
3134 		goto unlock;
3135 
3136 	if (task_cpu(arg->src_task) != arg->src_cpu)
3137 		goto unlock;
3138 
3139 	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3140 		goto unlock;
3141 
3142 	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3143 		goto unlock;
3144 
3145 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
3146 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
3147 
3148 	ret = 0;
3149 
3150 unlock:
3151 	double_rq_unlock(src_rq, dst_rq);
3152 	raw_spin_unlock(&arg->dst_task->pi_lock);
3153 	raw_spin_unlock(&arg->src_task->pi_lock);
3154 
3155 	return ret;
3156 }
3157 
3158 /*
3159  * Cross migrate two tasks
3160  */
3161 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3162 		int target_cpu, int curr_cpu)
3163 {
3164 	struct migration_swap_arg arg;
3165 	int ret = -EINVAL;
3166 
3167 	arg = (struct migration_swap_arg){
3168 		.src_task = cur,
3169 		.src_cpu = curr_cpu,
3170 		.dst_task = p,
3171 		.dst_cpu = target_cpu,
3172 	};
3173 
3174 	if (arg.src_cpu == arg.dst_cpu)
3175 		goto out;
3176 
3177 	/*
3178 	 * These three tests are all lockless; this is OK since all of them
3179 	 * will be re-checked with proper locks held further down the line.
3180 	 */
3181 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3182 		goto out;
3183 
3184 	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3185 		goto out;
3186 
3187 	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3188 		goto out;
3189 
3190 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3191 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3192 
3193 out:
3194 	return ret;
3195 }
3196 #endif /* CONFIG_NUMA_BALANCING */
3197 
3198 /*
3199  * wait_task_inactive - wait for a thread to unschedule.
3200  *
3201  * If @match_state is nonzero, it's the @p->state value just checked and
3202  * not expected to change.  If it changes, i.e. @p might have woken up,
3203  * then return zero.  When we succeed in waiting for @p to be off its CPU,
3204  * we return a positive number (its total switch count).  If a second call
3205  * a short while later returns the same number, the caller can be sure that
3206  * @p has remained unscheduled the whole time.
3207  *
3208  * The caller must ensure that the task *will* unschedule sometime soon,
3209  * else this function might spin for a *long* time. This function can't
3210  * be called with interrupts off, or it may introduce deadlock with
3211  * smp_call_function() if an IPI is sent by the same process we are
3212  * waiting to become inactive.
3213  */
3214 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3215 {
3216 	int running, queued;
3217 	struct rq_flags rf;
3218 	unsigned long ncsw;
3219 	struct rq *rq;
3220 
3221 	for (;;) {
3222 		/*
3223 		 * We do the initial early heuristics without holding
3224 		 * any task-queue locks at all. We'll only try to get
3225 		 * the runqueue lock when things look like they will
3226 		 * work out!
3227 		 */
3228 		rq = task_rq(p);
3229 
3230 		/*
3231 		 * If the task is actively running on another CPU
3232 		 * still, just relax and busy-wait without holding
3233 		 * any locks.
3234 		 *
3235 		 * NOTE! Since we don't hold any locks, it's not
3236 		 * even sure that "rq" stays as the right runqueue!
3237 		 * But we don't care, since "task_running()" will
3238 		 * return false if the runqueue has changed and p
3239 		 * is actually now running somewhere else!
3240 		 */
3241 		while (task_running(rq, p)) {
3242 			if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3243 				return 0;
3244 			cpu_relax();
3245 		}
3246 
3247 		/*
3248 		 * Ok, time to look more closely! We need the rq
3249 		 * lock now, to be *sure*. If we're wrong, we'll
3250 		 * just go back and repeat.
3251 		 */
3252 		rq = task_rq_lock(p, &rf);
3253 		trace_sched_wait_task(p);
3254 		running = task_running(rq, p);
3255 		queued = task_on_rq_queued(p);
3256 		ncsw = 0;
3257 		if (!match_state || READ_ONCE(p->__state) == match_state)
3258 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3259 		task_rq_unlock(rq, p, &rf);
3260 
3261 		/*
3262 		 * If it changed from the expected state, bail out now.
3263 		 */
3264 		if (unlikely(!ncsw))
3265 			break;
3266 
3267 		/*
3268 		 * Was it really running after all now that we
3269 		 * checked with the proper locks actually held?
3270 		 *
3271 		 * Oops. Go back and try again..
3272 		 */
3273 		if (unlikely(running)) {
3274 			cpu_relax();
3275 			continue;
3276 		}
3277 
3278 		/*
3279 		 * It's not enough that it's not actively running,
3280 		 * it must be off the runqueue _entirely_, and not
3281 		 * preempted!
3282 		 *
3283 		 * So if it was still runnable (but just not actively
3284 		 * running right now), it's preempted, and we should
3285 		 * yield - it could be a while.
3286 		 */
3287 		if (unlikely(queued)) {
3288 			ktime_t to = NSEC_PER_SEC / HZ;
3289 
3290 			set_current_state(TASK_UNINTERRUPTIBLE);
3291 			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3292 			continue;
3293 		}
3294 
3295 		/*
3296 		 * Ahh, all good. It wasn't running, and it wasn't
3297 		 * runnable, which means that it will never become
3298 		 * running in the future either. We're all done!
3299 		 */
3300 		break;
3301 	}
3302 
3303 	return ncsw;
3304 }
3305 
3306 /***
3307  * kick_process - kick a running thread to enter/exit the kernel
3308  * @p: the to-be-kicked thread
3309  *
3310  * Cause a process which is running on another CPU to enter
3311  * kernel-mode, without any delay. (to get signals handled.)
3312  *
3313  * NOTE: this function doesn't have to take the runqueue lock,
3314  * because all it wants to ensure is that the remote task enters
3315  * the kernel. If the IPI races and the task has been migrated
3316  * to another CPU then no harm is done and the purpose has been
3317  * achieved as well.
3318  */
3319 void kick_process(struct task_struct *p)
3320 {
3321 	int cpu;
3322 
3323 	preempt_disable();
3324 	cpu = task_cpu(p);
3325 	if ((cpu != smp_processor_id()) && task_curr(p))
3326 		smp_send_reschedule(cpu);
3327 	preempt_enable();
3328 }
3329 EXPORT_SYMBOL_GPL(kick_process);
3330 
3331 /*
3332  * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3333  *
3334  * A few notes on cpu_active vs cpu_online:
3335  *
3336  *  - cpu_active must be a subset of cpu_online
3337  *
3338  *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3339  *    see __set_cpus_allowed_ptr(). At this point the newly online
3340  *    CPU isn't yet part of the sched domains, and balancing will not
3341  *    see it.
3342  *
3343  *  - on CPU-down we clear cpu_active() to mask the sched domains and
3344  *    avoid the load balancer to place new tasks on the to be removed
3345  *    CPU. Existing tasks will remain running there and will be taken
3346  *    off.
3347  *
3348  * This means that fallback selection must not select !active CPUs.
3349  * And can assume that any active CPU must be online. Conversely
3350  * select_task_rq() below may allow selection of !active CPUs in order
3351  * to satisfy the above rules.
3352  */
3353 static int select_fallback_rq(int cpu, struct task_struct *p)
3354 {
3355 	int nid = cpu_to_node(cpu);
3356 	const struct cpumask *nodemask = NULL;
3357 	enum { cpuset, possible, fail } state = cpuset;
3358 	int dest_cpu;
3359 
3360 	/*
3361 	 * If the node that the CPU is on has been offlined, cpu_to_node()
3362 	 * will return -1. There is no CPU on the node, and we should
3363 	 * select the CPU on the other node.
3364 	 */
3365 	if (nid != -1) {
3366 		nodemask = cpumask_of_node(nid);
3367 
3368 		/* Look for allowed, online CPU in same node. */
3369 		for_each_cpu(dest_cpu, nodemask) {
3370 			if (is_cpu_allowed(p, dest_cpu))
3371 				return dest_cpu;
3372 		}
3373 	}
3374 
3375 	for (;;) {
3376 		/* Any allowed, online CPU? */
3377 		for_each_cpu(dest_cpu, p->cpus_ptr) {
3378 			if (!is_cpu_allowed(p, dest_cpu))
3379 				continue;
3380 
3381 			goto out;
3382 		}
3383 
3384 		/* No more Mr. Nice Guy. */
3385 		switch (state) {
3386 		case cpuset:
3387 			if (cpuset_cpus_allowed_fallback(p)) {
3388 				state = possible;
3389 				break;
3390 			}
3391 			fallthrough;
3392 		case possible:
3393 			/*
3394 			 * XXX When called from select_task_rq() we only
3395 			 * hold p->pi_lock and again violate locking order.
3396 			 *
3397 			 * More yuck to audit.
3398 			 */
3399 			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3400 			state = fail;
3401 			break;
3402 		case fail:
3403 			BUG();
3404 			break;
3405 		}
3406 	}
3407 
3408 out:
3409 	if (state != cpuset) {
3410 		/*
3411 		 * Don't tell them about moving exiting tasks or
3412 		 * kernel threads (both mm NULL), since they never
3413 		 * leave kernel.
3414 		 */
3415 		if (p->mm && printk_ratelimit()) {
3416 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3417 					task_pid_nr(p), p->comm, cpu);
3418 		}
3419 	}
3420 
3421 	return dest_cpu;
3422 }
3423 
3424 /*
3425  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3426  */
3427 static inline
3428 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3429 {
3430 	lockdep_assert_held(&p->pi_lock);
3431 
3432 	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3433 		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3434 	else
3435 		cpu = cpumask_any(p->cpus_ptr);
3436 
3437 	/*
3438 	 * In order not to call set_task_cpu() on a blocking task we need
3439 	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3440 	 * CPU.
3441 	 *
3442 	 * Since this is common to all placement strategies, this lives here.
3443 	 *
3444 	 * [ this allows ->select_task() to simply return task_cpu(p) and
3445 	 *   not worry about this generic constraint ]
3446 	 */
3447 	if (unlikely(!is_cpu_allowed(p, cpu)))
3448 		cpu = select_fallback_rq(task_cpu(p), p);
3449 
3450 	return cpu;
3451 }
3452 
3453 void sched_set_stop_task(int cpu, struct task_struct *stop)
3454 {
3455 	static struct lock_class_key stop_pi_lock;
3456 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3457 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
3458 
3459 	if (stop) {
3460 		/*
3461 		 * Make it appear like a SCHED_FIFO task, its something
3462 		 * userspace knows about and won't get confused about.
3463 		 *
3464 		 * Also, it will make PI more or less work without too
3465 		 * much confusion -- but then, stop work should not
3466 		 * rely on PI working anyway.
3467 		 */
3468 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3469 
3470 		stop->sched_class = &stop_sched_class;
3471 
3472 		/*
3473 		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3474 		 * adjust the effective priority of a task. As a result,
3475 		 * rt_mutex_setprio() can trigger (RT) balancing operations,
3476 		 * which can then trigger wakeups of the stop thread to push
3477 		 * around the current task.
3478 		 *
3479 		 * The stop task itself will never be part of the PI-chain, it
3480 		 * never blocks, therefore that ->pi_lock recursion is safe.
3481 		 * Tell lockdep about this by placing the stop->pi_lock in its
3482 		 * own class.
3483 		 */
3484 		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3485 	}
3486 
3487 	cpu_rq(cpu)->stop = stop;
3488 
3489 	if (old_stop) {
3490 		/*
3491 		 * Reset it back to a normal scheduling class so that
3492 		 * it can die in pieces.
3493 		 */
3494 		old_stop->sched_class = &rt_sched_class;
3495 	}
3496 }
3497 
3498 #else /* CONFIG_SMP */
3499 
3500 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3501 					 const struct cpumask *new_mask,
3502 					 u32 flags)
3503 {
3504 	return set_cpus_allowed_ptr(p, new_mask);
3505 }
3506 
3507 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3508 
3509 static inline bool rq_has_pinned_tasks(struct rq *rq)
3510 {
3511 	return false;
3512 }
3513 
3514 #endif /* !CONFIG_SMP */
3515 
3516 static void
3517 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3518 {
3519 	struct rq *rq;
3520 
3521 	if (!schedstat_enabled())
3522 		return;
3523 
3524 	rq = this_rq();
3525 
3526 #ifdef CONFIG_SMP
3527 	if (cpu == rq->cpu) {
3528 		__schedstat_inc(rq->ttwu_local);
3529 		__schedstat_inc(p->stats.nr_wakeups_local);
3530 	} else {
3531 		struct sched_domain *sd;
3532 
3533 		__schedstat_inc(p->stats.nr_wakeups_remote);
3534 		rcu_read_lock();
3535 		for_each_domain(rq->cpu, sd) {
3536 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3537 				__schedstat_inc(sd->ttwu_wake_remote);
3538 				break;
3539 			}
3540 		}
3541 		rcu_read_unlock();
3542 	}
3543 
3544 	if (wake_flags & WF_MIGRATED)
3545 		__schedstat_inc(p->stats.nr_wakeups_migrate);
3546 #endif /* CONFIG_SMP */
3547 
3548 	__schedstat_inc(rq->ttwu_count);
3549 	__schedstat_inc(p->stats.nr_wakeups);
3550 
3551 	if (wake_flags & WF_SYNC)
3552 		__schedstat_inc(p->stats.nr_wakeups_sync);
3553 }
3554 
3555 /*
3556  * Mark the task runnable and perform wakeup-preemption.
3557  */
3558 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3559 			   struct rq_flags *rf)
3560 {
3561 	check_preempt_curr(rq, p, wake_flags);
3562 	WRITE_ONCE(p->__state, TASK_RUNNING);
3563 	trace_sched_wakeup(p);
3564 
3565 #ifdef CONFIG_SMP
3566 	if (p->sched_class->task_woken) {
3567 		/*
3568 		 * Our task @p is fully woken up and running; so it's safe to
3569 		 * drop the rq->lock, hereafter rq is only used for statistics.
3570 		 */
3571 		rq_unpin_lock(rq, rf);
3572 		p->sched_class->task_woken(rq, p);
3573 		rq_repin_lock(rq, rf);
3574 	}
3575 
3576 	if (rq->idle_stamp) {
3577 		u64 delta = rq_clock(rq) - rq->idle_stamp;
3578 		u64 max = 2*rq->max_idle_balance_cost;
3579 
3580 		update_avg(&rq->avg_idle, delta);
3581 
3582 		if (rq->avg_idle > max)
3583 			rq->avg_idle = max;
3584 
3585 		rq->wake_stamp = jiffies;
3586 		rq->wake_avg_idle = rq->avg_idle / 2;
3587 
3588 		rq->idle_stamp = 0;
3589 	}
3590 #endif
3591 }
3592 
3593 static void
3594 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3595 		 struct rq_flags *rf)
3596 {
3597 	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3598 
3599 	lockdep_assert_rq_held(rq);
3600 
3601 	if (p->sched_contributes_to_load)
3602 		rq->nr_uninterruptible--;
3603 
3604 #ifdef CONFIG_SMP
3605 	if (wake_flags & WF_MIGRATED)
3606 		en_flags |= ENQUEUE_MIGRATED;
3607 	else
3608 #endif
3609 	if (p->in_iowait) {
3610 		delayacct_blkio_end(p);
3611 		atomic_dec(&task_rq(p)->nr_iowait);
3612 	}
3613 
3614 	activate_task(rq, p, en_flags);
3615 	ttwu_do_wakeup(rq, p, wake_flags, rf);
3616 }
3617 
3618 /*
3619  * Consider @p being inside a wait loop:
3620  *
3621  *   for (;;) {
3622  *      set_current_state(TASK_UNINTERRUPTIBLE);
3623  *
3624  *      if (CONDITION)
3625  *         break;
3626  *
3627  *      schedule();
3628  *   }
3629  *   __set_current_state(TASK_RUNNING);
3630  *
3631  * between set_current_state() and schedule(). In this case @p is still
3632  * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3633  * an atomic manner.
3634  *
3635  * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3636  * then schedule() must still happen and p->state can be changed to
3637  * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3638  * need to do a full wakeup with enqueue.
3639  *
3640  * Returns: %true when the wakeup is done,
3641  *          %false otherwise.
3642  */
3643 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3644 {
3645 	struct rq_flags rf;
3646 	struct rq *rq;
3647 	int ret = 0;
3648 
3649 	rq = __task_rq_lock(p, &rf);
3650 	if (task_on_rq_queued(p)) {
3651 		/* check_preempt_curr() may use rq clock */
3652 		update_rq_clock(rq);
3653 		ttwu_do_wakeup(rq, p, wake_flags, &rf);
3654 		ret = 1;
3655 	}
3656 	__task_rq_unlock(rq, &rf);
3657 
3658 	return ret;
3659 }
3660 
3661 #ifdef CONFIG_SMP
3662 void sched_ttwu_pending(void *arg)
3663 {
3664 	struct llist_node *llist = arg;
3665 	struct rq *rq = this_rq();
3666 	struct task_struct *p, *t;
3667 	struct rq_flags rf;
3668 
3669 	if (!llist)
3670 		return;
3671 
3672 	/*
3673 	 * rq::ttwu_pending racy indication of out-standing wakeups.
3674 	 * Races such that false-negatives are possible, since they
3675 	 * are shorter lived that false-positives would be.
3676 	 */
3677 	WRITE_ONCE(rq->ttwu_pending, 0);
3678 
3679 	rq_lock_irqsave(rq, &rf);
3680 	update_rq_clock(rq);
3681 
3682 	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3683 		if (WARN_ON_ONCE(p->on_cpu))
3684 			smp_cond_load_acquire(&p->on_cpu, !VAL);
3685 
3686 		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3687 			set_task_cpu(p, cpu_of(rq));
3688 
3689 		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3690 	}
3691 
3692 	rq_unlock_irqrestore(rq, &rf);
3693 }
3694 
3695 void send_call_function_single_ipi(int cpu)
3696 {
3697 	struct rq *rq = cpu_rq(cpu);
3698 
3699 	if (!set_nr_if_polling(rq->idle))
3700 		arch_send_call_function_single_ipi(cpu);
3701 	else
3702 		trace_sched_wake_idle_without_ipi(cpu);
3703 }
3704 
3705 /*
3706  * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3707  * necessary. The wakee CPU on receipt of the IPI will queue the task
3708  * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3709  * of the wakeup instead of the waker.
3710  */
3711 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3712 {
3713 	struct rq *rq = cpu_rq(cpu);
3714 
3715 	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3716 
3717 	WRITE_ONCE(rq->ttwu_pending, 1);
3718 	__smp_call_single_queue(cpu, &p->wake_entry.llist);
3719 }
3720 
3721 void wake_up_if_idle(int cpu)
3722 {
3723 	struct rq *rq = cpu_rq(cpu);
3724 	struct rq_flags rf;
3725 
3726 	rcu_read_lock();
3727 
3728 	if (!is_idle_task(rcu_dereference(rq->curr)))
3729 		goto out;
3730 
3731 	rq_lock_irqsave(rq, &rf);
3732 	if (is_idle_task(rq->curr))
3733 		resched_curr(rq);
3734 	/* Else CPU is not idle, do nothing here: */
3735 	rq_unlock_irqrestore(rq, &rf);
3736 
3737 out:
3738 	rcu_read_unlock();
3739 }
3740 
3741 bool cpus_share_cache(int this_cpu, int that_cpu)
3742 {
3743 	if (this_cpu == that_cpu)
3744 		return true;
3745 
3746 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3747 }
3748 
3749 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3750 {
3751 	/*
3752 	 * Do not complicate things with the async wake_list while the CPU is
3753 	 * in hotplug state.
3754 	 */
3755 	if (!cpu_active(cpu))
3756 		return false;
3757 
3758 	/*
3759 	 * If the CPU does not share cache, then queue the task on the
3760 	 * remote rqs wakelist to avoid accessing remote data.
3761 	 */
3762 	if (!cpus_share_cache(smp_processor_id(), cpu))
3763 		return true;
3764 
3765 	/*
3766 	 * If the task is descheduling and the only running task on the
3767 	 * CPU then use the wakelist to offload the task activation to
3768 	 * the soon-to-be-idle CPU as the current CPU is likely busy.
3769 	 * nr_running is checked to avoid unnecessary task stacking.
3770 	 */
3771 	if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3772 		return true;
3773 
3774 	return false;
3775 }
3776 
3777 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3778 {
3779 	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3780 		if (WARN_ON_ONCE(cpu == smp_processor_id()))
3781 			return false;
3782 
3783 		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3784 		__ttwu_queue_wakelist(p, cpu, wake_flags);
3785 		return true;
3786 	}
3787 
3788 	return false;
3789 }
3790 
3791 #else /* !CONFIG_SMP */
3792 
3793 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3794 {
3795 	return false;
3796 }
3797 
3798 #endif /* CONFIG_SMP */
3799 
3800 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3801 {
3802 	struct rq *rq = cpu_rq(cpu);
3803 	struct rq_flags rf;
3804 
3805 	if (ttwu_queue_wakelist(p, cpu, wake_flags))
3806 		return;
3807 
3808 	rq_lock(rq, &rf);
3809 	update_rq_clock(rq);
3810 	ttwu_do_activate(rq, p, wake_flags, &rf);
3811 	rq_unlock(rq, &rf);
3812 }
3813 
3814 /*
3815  * Invoked from try_to_wake_up() to check whether the task can be woken up.
3816  *
3817  * The caller holds p::pi_lock if p != current or has preemption
3818  * disabled when p == current.
3819  *
3820  * The rules of PREEMPT_RT saved_state:
3821  *
3822  *   The related locking code always holds p::pi_lock when updating
3823  *   p::saved_state, which means the code is fully serialized in both cases.
3824  *
3825  *   The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3826  *   bits set. This allows to distinguish all wakeup scenarios.
3827  */
3828 static __always_inline
3829 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3830 {
3831 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3832 		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3833 			     state != TASK_RTLOCK_WAIT);
3834 	}
3835 
3836 	if (READ_ONCE(p->__state) & state) {
3837 		*success = 1;
3838 		return true;
3839 	}
3840 
3841 #ifdef CONFIG_PREEMPT_RT
3842 	/*
3843 	 * Saved state preserves the task state across blocking on
3844 	 * an RT lock.  If the state matches, set p::saved_state to
3845 	 * TASK_RUNNING, but do not wake the task because it waits
3846 	 * for a lock wakeup. Also indicate success because from
3847 	 * the regular waker's point of view this has succeeded.
3848 	 *
3849 	 * After acquiring the lock the task will restore p::__state
3850 	 * from p::saved_state which ensures that the regular
3851 	 * wakeup is not lost. The restore will also set
3852 	 * p::saved_state to TASK_RUNNING so any further tests will
3853 	 * not result in false positives vs. @success
3854 	 */
3855 	if (p->saved_state & state) {
3856 		p->saved_state = TASK_RUNNING;
3857 		*success = 1;
3858 	}
3859 #endif
3860 	return false;
3861 }
3862 
3863 /*
3864  * Notes on Program-Order guarantees on SMP systems.
3865  *
3866  *  MIGRATION
3867  *
3868  * The basic program-order guarantee on SMP systems is that when a task [t]
3869  * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3870  * execution on its new CPU [c1].
3871  *
3872  * For migration (of runnable tasks) this is provided by the following means:
3873  *
3874  *  A) UNLOCK of the rq(c0)->lock scheduling out task t
3875  *  B) migration for t is required to synchronize *both* rq(c0)->lock and
3876  *     rq(c1)->lock (if not at the same time, then in that order).
3877  *  C) LOCK of the rq(c1)->lock scheduling in task
3878  *
3879  * Release/acquire chaining guarantees that B happens after A and C after B.
3880  * Note: the CPU doing B need not be c0 or c1
3881  *
3882  * Example:
3883  *
3884  *   CPU0            CPU1            CPU2
3885  *
3886  *   LOCK rq(0)->lock
3887  *   sched-out X
3888  *   sched-in Y
3889  *   UNLOCK rq(0)->lock
3890  *
3891  *                                   LOCK rq(0)->lock // orders against CPU0
3892  *                                   dequeue X
3893  *                                   UNLOCK rq(0)->lock
3894  *
3895  *                                   LOCK rq(1)->lock
3896  *                                   enqueue X
3897  *                                   UNLOCK rq(1)->lock
3898  *
3899  *                   LOCK rq(1)->lock // orders against CPU2
3900  *                   sched-out Z
3901  *                   sched-in X
3902  *                   UNLOCK rq(1)->lock
3903  *
3904  *
3905  *  BLOCKING -- aka. SLEEP + WAKEUP
3906  *
3907  * For blocking we (obviously) need to provide the same guarantee as for
3908  * migration. However the means are completely different as there is no lock
3909  * chain to provide order. Instead we do:
3910  *
3911  *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
3912  *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3913  *
3914  * Example:
3915  *
3916  *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
3917  *
3918  *   LOCK rq(0)->lock LOCK X->pi_lock
3919  *   dequeue X
3920  *   sched-out X
3921  *   smp_store_release(X->on_cpu, 0);
3922  *
3923  *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
3924  *                    X->state = WAKING
3925  *                    set_task_cpu(X,2)
3926  *
3927  *                    LOCK rq(2)->lock
3928  *                    enqueue X
3929  *                    X->state = RUNNING
3930  *                    UNLOCK rq(2)->lock
3931  *
3932  *                                          LOCK rq(2)->lock // orders against CPU1
3933  *                                          sched-out Z
3934  *                                          sched-in X
3935  *                                          UNLOCK rq(2)->lock
3936  *
3937  *                    UNLOCK X->pi_lock
3938  *   UNLOCK rq(0)->lock
3939  *
3940  *
3941  * However, for wakeups there is a second guarantee we must provide, namely we
3942  * must ensure that CONDITION=1 done by the caller can not be reordered with
3943  * accesses to the task state; see try_to_wake_up() and set_current_state().
3944  */
3945 
3946 /**
3947  * try_to_wake_up - wake up a thread
3948  * @p: the thread to be awakened
3949  * @state: the mask of task states that can be woken
3950  * @wake_flags: wake modifier flags (WF_*)
3951  *
3952  * Conceptually does:
3953  *
3954  *   If (@state & @p->state) @p->state = TASK_RUNNING.
3955  *
3956  * If the task was not queued/runnable, also place it back on a runqueue.
3957  *
3958  * This function is atomic against schedule() which would dequeue the task.
3959  *
3960  * It issues a full memory barrier before accessing @p->state, see the comment
3961  * with set_current_state().
3962  *
3963  * Uses p->pi_lock to serialize against concurrent wake-ups.
3964  *
3965  * Relies on p->pi_lock stabilizing:
3966  *  - p->sched_class
3967  *  - p->cpus_ptr
3968  *  - p->sched_task_group
3969  * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3970  *
3971  * Tries really hard to only take one task_rq(p)->lock for performance.
3972  * Takes rq->lock in:
3973  *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
3974  *  - ttwu_queue()       -- new rq, for enqueue of the task;
3975  *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3976  *
3977  * As a consequence we race really badly with just about everything. See the
3978  * many memory barriers and their comments for details.
3979  *
3980  * Return: %true if @p->state changes (an actual wakeup was done),
3981  *	   %false otherwise.
3982  */
3983 static int
3984 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3985 {
3986 	unsigned long flags;
3987 	int cpu, success = 0;
3988 
3989 	preempt_disable();
3990 	if (p == current) {
3991 		/*
3992 		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3993 		 * == smp_processor_id()'. Together this means we can special
3994 		 * case the whole 'p->on_rq && ttwu_runnable()' case below
3995 		 * without taking any locks.
3996 		 *
3997 		 * In particular:
3998 		 *  - we rely on Program-Order guarantees for all the ordering,
3999 		 *  - we're serialized against set_special_state() by virtue of
4000 		 *    it disabling IRQs (this allows not taking ->pi_lock).
4001 		 */
4002 		if (!ttwu_state_match(p, state, &success))
4003 			goto out;
4004 
4005 		trace_sched_waking(p);
4006 		WRITE_ONCE(p->__state, TASK_RUNNING);
4007 		trace_sched_wakeup(p);
4008 		goto out;
4009 	}
4010 
4011 	/*
4012 	 * If we are going to wake up a thread waiting for CONDITION we
4013 	 * need to ensure that CONDITION=1 done by the caller can not be
4014 	 * reordered with p->state check below. This pairs with smp_store_mb()
4015 	 * in set_current_state() that the waiting thread does.
4016 	 */
4017 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4018 	smp_mb__after_spinlock();
4019 	if (!ttwu_state_match(p, state, &success))
4020 		goto unlock;
4021 
4022 	trace_sched_waking(p);
4023 
4024 	/*
4025 	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4026 	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4027 	 * in smp_cond_load_acquire() below.
4028 	 *
4029 	 * sched_ttwu_pending()			try_to_wake_up()
4030 	 *   STORE p->on_rq = 1			  LOAD p->state
4031 	 *   UNLOCK rq->lock
4032 	 *
4033 	 * __schedule() (switch to task 'p')
4034 	 *   LOCK rq->lock			  smp_rmb();
4035 	 *   smp_mb__after_spinlock();
4036 	 *   UNLOCK rq->lock
4037 	 *
4038 	 * [task p]
4039 	 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
4040 	 *
4041 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4042 	 * __schedule().  See the comment for smp_mb__after_spinlock().
4043 	 *
4044 	 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4045 	 */
4046 	smp_rmb();
4047 	if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4048 		goto unlock;
4049 
4050 #ifdef CONFIG_SMP
4051 	/*
4052 	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4053 	 * possible to, falsely, observe p->on_cpu == 0.
4054 	 *
4055 	 * One must be running (->on_cpu == 1) in order to remove oneself
4056 	 * from the runqueue.
4057 	 *
4058 	 * __schedule() (switch to task 'p')	try_to_wake_up()
4059 	 *   STORE p->on_cpu = 1		  LOAD p->on_rq
4060 	 *   UNLOCK rq->lock
4061 	 *
4062 	 * __schedule() (put 'p' to sleep)
4063 	 *   LOCK rq->lock			  smp_rmb();
4064 	 *   smp_mb__after_spinlock();
4065 	 *   STORE p->on_rq = 0			  LOAD p->on_cpu
4066 	 *
4067 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4068 	 * __schedule().  See the comment for smp_mb__after_spinlock().
4069 	 *
4070 	 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4071 	 * schedule()'s deactivate_task() has 'happened' and p will no longer
4072 	 * care about it's own p->state. See the comment in __schedule().
4073 	 */
4074 	smp_acquire__after_ctrl_dep();
4075 
4076 	/*
4077 	 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4078 	 * == 0), which means we need to do an enqueue, change p->state to
4079 	 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4080 	 * enqueue, such as ttwu_queue_wakelist().
4081 	 */
4082 	WRITE_ONCE(p->__state, TASK_WAKING);
4083 
4084 	/*
4085 	 * If the owning (remote) CPU is still in the middle of schedule() with
4086 	 * this task as prev, considering queueing p on the remote CPUs wake_list
4087 	 * which potentially sends an IPI instead of spinning on p->on_cpu to
4088 	 * let the waker make forward progress. This is safe because IRQs are
4089 	 * disabled and the IPI will deliver after on_cpu is cleared.
4090 	 *
4091 	 * Ensure we load task_cpu(p) after p->on_cpu:
4092 	 *
4093 	 * set_task_cpu(p, cpu);
4094 	 *   STORE p->cpu = @cpu
4095 	 * __schedule() (switch to task 'p')
4096 	 *   LOCK rq->lock
4097 	 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
4098 	 *   STORE p->on_cpu = 1		LOAD p->cpu
4099 	 *
4100 	 * to ensure we observe the correct CPU on which the task is currently
4101 	 * scheduling.
4102 	 */
4103 	if (smp_load_acquire(&p->on_cpu) &&
4104 	    ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4105 		goto unlock;
4106 
4107 	/*
4108 	 * If the owning (remote) CPU is still in the middle of schedule() with
4109 	 * this task as prev, wait until it's done referencing the task.
4110 	 *
4111 	 * Pairs with the smp_store_release() in finish_task().
4112 	 *
4113 	 * This ensures that tasks getting woken will be fully ordered against
4114 	 * their previous state and preserve Program Order.
4115 	 */
4116 	smp_cond_load_acquire(&p->on_cpu, !VAL);
4117 
4118 	cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4119 	if (task_cpu(p) != cpu) {
4120 		if (p->in_iowait) {
4121 			delayacct_blkio_end(p);
4122 			atomic_dec(&task_rq(p)->nr_iowait);
4123 		}
4124 
4125 		wake_flags |= WF_MIGRATED;
4126 		psi_ttwu_dequeue(p);
4127 		set_task_cpu(p, cpu);
4128 	}
4129 #else
4130 	cpu = task_cpu(p);
4131 #endif /* CONFIG_SMP */
4132 
4133 	ttwu_queue(p, cpu, wake_flags);
4134 unlock:
4135 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4136 out:
4137 	if (success)
4138 		ttwu_stat(p, task_cpu(p), wake_flags);
4139 	preempt_enable();
4140 
4141 	return success;
4142 }
4143 
4144 /**
4145  * task_call_func - Invoke a function on task in fixed state
4146  * @p: Process for which the function is to be invoked, can be @current.
4147  * @func: Function to invoke.
4148  * @arg: Argument to function.
4149  *
4150  * Fix the task in it's current state by avoiding wakeups and or rq operations
4151  * and call @func(@arg) on it.  This function can use ->on_rq and task_curr()
4152  * to work out what the state is, if required.  Given that @func can be invoked
4153  * with a runqueue lock held, it had better be quite lightweight.
4154  *
4155  * Returns:
4156  *   Whatever @func returns
4157  */
4158 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4159 {
4160 	struct rq *rq = NULL;
4161 	unsigned int state;
4162 	struct rq_flags rf;
4163 	int ret;
4164 
4165 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4166 
4167 	state = READ_ONCE(p->__state);
4168 
4169 	/*
4170 	 * Ensure we load p->on_rq after p->__state, otherwise it would be
4171 	 * possible to, falsely, observe p->on_rq == 0.
4172 	 *
4173 	 * See try_to_wake_up() for a longer comment.
4174 	 */
4175 	smp_rmb();
4176 
4177 	/*
4178 	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4179 	 * the task is blocked. Make sure to check @state since ttwu() can drop
4180 	 * locks at the end, see ttwu_queue_wakelist().
4181 	 */
4182 	if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4183 		rq = __task_rq_lock(p, &rf);
4184 
4185 	/*
4186 	 * At this point the task is pinned; either:
4187 	 *  - blocked and we're holding off wakeups	 (pi->lock)
4188 	 *  - woken, and we're holding off enqueue	 (rq->lock)
4189 	 *  - queued, and we're holding off schedule	 (rq->lock)
4190 	 *  - running, and we're holding off de-schedule (rq->lock)
4191 	 *
4192 	 * The called function (@func) can use: task_curr(), p->on_rq and
4193 	 * p->__state to differentiate between these states.
4194 	 */
4195 	ret = func(p, arg);
4196 
4197 	if (rq)
4198 		rq_unlock(rq, &rf);
4199 
4200 	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4201 	return ret;
4202 }
4203 
4204 /**
4205  * wake_up_process - Wake up a specific process
4206  * @p: The process to be woken up.
4207  *
4208  * Attempt to wake up the nominated process and move it to the set of runnable
4209  * processes.
4210  *
4211  * Return: 1 if the process was woken up, 0 if it was already running.
4212  *
4213  * This function executes a full memory barrier before accessing the task state.
4214  */
4215 int wake_up_process(struct task_struct *p)
4216 {
4217 	return try_to_wake_up(p, TASK_NORMAL, 0);
4218 }
4219 EXPORT_SYMBOL(wake_up_process);
4220 
4221 int wake_up_state(struct task_struct *p, unsigned int state)
4222 {
4223 	return try_to_wake_up(p, state, 0);
4224 }
4225 
4226 /*
4227  * Perform scheduler related setup for a newly forked process p.
4228  * p is forked by current.
4229  *
4230  * __sched_fork() is basic setup used by init_idle() too:
4231  */
4232 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4233 {
4234 	p->on_rq			= 0;
4235 
4236 	p->se.on_rq			= 0;
4237 	p->se.exec_start		= 0;
4238 	p->se.sum_exec_runtime		= 0;
4239 	p->se.prev_sum_exec_runtime	= 0;
4240 	p->se.nr_migrations		= 0;
4241 	p->se.vruntime			= 0;
4242 	INIT_LIST_HEAD(&p->se.group_node);
4243 
4244 #ifdef CONFIG_FAIR_GROUP_SCHED
4245 	p->se.cfs_rq			= NULL;
4246 #endif
4247 
4248 #ifdef CONFIG_SCHEDSTATS
4249 	/* Even if schedstat is disabled, there should not be garbage */
4250 	memset(&p->stats, 0, sizeof(p->stats));
4251 #endif
4252 
4253 	RB_CLEAR_NODE(&p->dl.rb_node);
4254 	init_dl_task_timer(&p->dl);
4255 	init_dl_inactive_task_timer(&p->dl);
4256 	__dl_clear_params(p);
4257 
4258 	INIT_LIST_HEAD(&p->rt.run_list);
4259 	p->rt.timeout		= 0;
4260 	p->rt.time_slice	= sched_rr_timeslice;
4261 	p->rt.on_rq		= 0;
4262 	p->rt.on_list		= 0;
4263 
4264 #ifdef CONFIG_PREEMPT_NOTIFIERS
4265 	INIT_HLIST_HEAD(&p->preempt_notifiers);
4266 #endif
4267 
4268 #ifdef CONFIG_COMPACTION
4269 	p->capture_control = NULL;
4270 #endif
4271 	init_numa_balancing(clone_flags, p);
4272 #ifdef CONFIG_SMP
4273 	p->wake_entry.u_flags = CSD_TYPE_TTWU;
4274 	p->migration_pending = NULL;
4275 #endif
4276 }
4277 
4278 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4279 
4280 #ifdef CONFIG_NUMA_BALANCING
4281 
4282 void set_numabalancing_state(bool enabled)
4283 {
4284 	if (enabled)
4285 		static_branch_enable(&sched_numa_balancing);
4286 	else
4287 		static_branch_disable(&sched_numa_balancing);
4288 }
4289 
4290 #ifdef CONFIG_PROC_SYSCTL
4291 int sysctl_numa_balancing(struct ctl_table *table, int write,
4292 			  void *buffer, size_t *lenp, loff_t *ppos)
4293 {
4294 	struct ctl_table t;
4295 	int err;
4296 	int state = static_branch_likely(&sched_numa_balancing);
4297 
4298 	if (write && !capable(CAP_SYS_ADMIN))
4299 		return -EPERM;
4300 
4301 	t = *table;
4302 	t.data = &state;
4303 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4304 	if (err < 0)
4305 		return err;
4306 	if (write)
4307 		set_numabalancing_state(state);
4308 	return err;
4309 }
4310 #endif
4311 #endif
4312 
4313 #ifdef CONFIG_SCHEDSTATS
4314 
4315 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4316 
4317 static void set_schedstats(bool enabled)
4318 {
4319 	if (enabled)
4320 		static_branch_enable(&sched_schedstats);
4321 	else
4322 		static_branch_disable(&sched_schedstats);
4323 }
4324 
4325 void force_schedstat_enabled(void)
4326 {
4327 	if (!schedstat_enabled()) {
4328 		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4329 		static_branch_enable(&sched_schedstats);
4330 	}
4331 }
4332 
4333 static int __init setup_schedstats(char *str)
4334 {
4335 	int ret = 0;
4336 	if (!str)
4337 		goto out;
4338 
4339 	if (!strcmp(str, "enable")) {
4340 		set_schedstats(true);
4341 		ret = 1;
4342 	} else if (!strcmp(str, "disable")) {
4343 		set_schedstats(false);
4344 		ret = 1;
4345 	}
4346 out:
4347 	if (!ret)
4348 		pr_warn("Unable to parse schedstats=\n");
4349 
4350 	return ret;
4351 }
4352 __setup("schedstats=", setup_schedstats);
4353 
4354 #ifdef CONFIG_PROC_SYSCTL
4355 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4356 		size_t *lenp, loff_t *ppos)
4357 {
4358 	struct ctl_table t;
4359 	int err;
4360 	int state = static_branch_likely(&sched_schedstats);
4361 
4362 	if (write && !capable(CAP_SYS_ADMIN))
4363 		return -EPERM;
4364 
4365 	t = *table;
4366 	t.data = &state;
4367 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4368 	if (err < 0)
4369 		return err;
4370 	if (write)
4371 		set_schedstats(state);
4372 	return err;
4373 }
4374 #endif /* CONFIG_PROC_SYSCTL */
4375 #endif /* CONFIG_SCHEDSTATS */
4376 
4377 /*
4378  * fork()/clone()-time setup:
4379  */
4380 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4381 {
4382 	__sched_fork(clone_flags, p);
4383 	/*
4384 	 * We mark the process as NEW here. This guarantees that
4385 	 * nobody will actually run it, and a signal or other external
4386 	 * event cannot wake it up and insert it on the runqueue either.
4387 	 */
4388 	p->__state = TASK_NEW;
4389 
4390 	/*
4391 	 * Make sure we do not leak PI boosting priority to the child.
4392 	 */
4393 	p->prio = current->normal_prio;
4394 
4395 	uclamp_fork(p);
4396 
4397 	/*
4398 	 * Revert to default priority/policy on fork if requested.
4399 	 */
4400 	if (unlikely(p->sched_reset_on_fork)) {
4401 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4402 			p->policy = SCHED_NORMAL;
4403 			p->static_prio = NICE_TO_PRIO(0);
4404 			p->rt_priority = 0;
4405 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
4406 			p->static_prio = NICE_TO_PRIO(0);
4407 
4408 		p->prio = p->normal_prio = p->static_prio;
4409 		set_load_weight(p, false);
4410 
4411 		/*
4412 		 * We don't need the reset flag anymore after the fork. It has
4413 		 * fulfilled its duty:
4414 		 */
4415 		p->sched_reset_on_fork = 0;
4416 	}
4417 
4418 	if (dl_prio(p->prio))
4419 		return -EAGAIN;
4420 	else if (rt_prio(p->prio))
4421 		p->sched_class = &rt_sched_class;
4422 	else
4423 		p->sched_class = &fair_sched_class;
4424 
4425 	init_entity_runnable_average(&p->se);
4426 
4427 #ifdef CONFIG_SCHED_INFO
4428 	if (likely(sched_info_on()))
4429 		memset(&p->sched_info, 0, sizeof(p->sched_info));
4430 #endif
4431 #if defined(CONFIG_SMP)
4432 	p->on_cpu = 0;
4433 #endif
4434 	init_task_preempt_count(p);
4435 #ifdef CONFIG_SMP
4436 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
4437 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
4438 #endif
4439 	return 0;
4440 }
4441 
4442 void sched_post_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4443 {
4444 	unsigned long flags;
4445 #ifdef CONFIG_CGROUP_SCHED
4446 	struct task_group *tg;
4447 #endif
4448 
4449 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4450 #ifdef CONFIG_CGROUP_SCHED
4451 	tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4452 			  struct task_group, css);
4453 	p->sched_task_group = autogroup_task_group(p, tg);
4454 #endif
4455 	rseq_migrate(p);
4456 	/*
4457 	 * We're setting the CPU for the first time, we don't migrate,
4458 	 * so use __set_task_cpu().
4459 	 */
4460 	__set_task_cpu(p, smp_processor_id());
4461 	if (p->sched_class->task_fork)
4462 		p->sched_class->task_fork(p);
4463 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4464 
4465 	uclamp_post_fork(p);
4466 }
4467 
4468 unsigned long to_ratio(u64 period, u64 runtime)
4469 {
4470 	if (runtime == RUNTIME_INF)
4471 		return BW_UNIT;
4472 
4473 	/*
4474 	 * Doing this here saves a lot of checks in all
4475 	 * the calling paths, and returning zero seems
4476 	 * safe for them anyway.
4477 	 */
4478 	if (period == 0)
4479 		return 0;
4480 
4481 	return div64_u64(runtime << BW_SHIFT, period);
4482 }
4483 
4484 /*
4485  * wake_up_new_task - wake up a newly created task for the first time.
4486  *
4487  * This function will do some initial scheduler statistics housekeeping
4488  * that must be done for every newly created context, then puts the task
4489  * on the runqueue and wakes it.
4490  */
4491 void wake_up_new_task(struct task_struct *p)
4492 {
4493 	struct rq_flags rf;
4494 	struct rq *rq;
4495 
4496 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4497 	WRITE_ONCE(p->__state, TASK_RUNNING);
4498 #ifdef CONFIG_SMP
4499 	/*
4500 	 * Fork balancing, do it here and not earlier because:
4501 	 *  - cpus_ptr can change in the fork path
4502 	 *  - any previously selected CPU might disappear through hotplug
4503 	 *
4504 	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4505 	 * as we're not fully set-up yet.
4506 	 */
4507 	p->recent_used_cpu = task_cpu(p);
4508 	rseq_migrate(p);
4509 	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4510 #endif
4511 	rq = __task_rq_lock(p, &rf);
4512 	update_rq_clock(rq);
4513 	post_init_entity_util_avg(p);
4514 
4515 	activate_task(rq, p, ENQUEUE_NOCLOCK);
4516 	trace_sched_wakeup_new(p);
4517 	check_preempt_curr(rq, p, WF_FORK);
4518 #ifdef CONFIG_SMP
4519 	if (p->sched_class->task_woken) {
4520 		/*
4521 		 * Nothing relies on rq->lock after this, so it's fine to
4522 		 * drop it.
4523 		 */
4524 		rq_unpin_lock(rq, &rf);
4525 		p->sched_class->task_woken(rq, p);
4526 		rq_repin_lock(rq, &rf);
4527 	}
4528 #endif
4529 	task_rq_unlock(rq, p, &rf);
4530 }
4531 
4532 #ifdef CONFIG_PREEMPT_NOTIFIERS
4533 
4534 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4535 
4536 void preempt_notifier_inc(void)
4537 {
4538 	static_branch_inc(&preempt_notifier_key);
4539 }
4540 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4541 
4542 void preempt_notifier_dec(void)
4543 {
4544 	static_branch_dec(&preempt_notifier_key);
4545 }
4546 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4547 
4548 /**
4549  * preempt_notifier_register - tell me when current is being preempted & rescheduled
4550  * @notifier: notifier struct to register
4551  */
4552 void preempt_notifier_register(struct preempt_notifier *notifier)
4553 {
4554 	if (!static_branch_unlikely(&preempt_notifier_key))
4555 		WARN(1, "registering preempt_notifier while notifiers disabled\n");
4556 
4557 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
4558 }
4559 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4560 
4561 /**
4562  * preempt_notifier_unregister - no longer interested in preemption notifications
4563  * @notifier: notifier struct to unregister
4564  *
4565  * This is *not* safe to call from within a preemption notifier.
4566  */
4567 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4568 {
4569 	hlist_del(&notifier->link);
4570 }
4571 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4572 
4573 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4574 {
4575 	struct preempt_notifier *notifier;
4576 
4577 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4578 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
4579 }
4580 
4581 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4582 {
4583 	if (static_branch_unlikely(&preempt_notifier_key))
4584 		__fire_sched_in_preempt_notifiers(curr);
4585 }
4586 
4587 static void
4588 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4589 				   struct task_struct *next)
4590 {
4591 	struct preempt_notifier *notifier;
4592 
4593 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4594 		notifier->ops->sched_out(notifier, next);
4595 }
4596 
4597 static __always_inline void
4598 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4599 				 struct task_struct *next)
4600 {
4601 	if (static_branch_unlikely(&preempt_notifier_key))
4602 		__fire_sched_out_preempt_notifiers(curr, next);
4603 }
4604 
4605 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4606 
4607 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4608 {
4609 }
4610 
4611 static inline void
4612 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4613 				 struct task_struct *next)
4614 {
4615 }
4616 
4617 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4618 
4619 static inline void prepare_task(struct task_struct *next)
4620 {
4621 #ifdef CONFIG_SMP
4622 	/*
4623 	 * Claim the task as running, we do this before switching to it
4624 	 * such that any running task will have this set.
4625 	 *
4626 	 * See the ttwu() WF_ON_CPU case and its ordering comment.
4627 	 */
4628 	WRITE_ONCE(next->on_cpu, 1);
4629 #endif
4630 }
4631 
4632 static inline void finish_task(struct task_struct *prev)
4633 {
4634 #ifdef CONFIG_SMP
4635 	/*
4636 	 * This must be the very last reference to @prev from this CPU. After
4637 	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4638 	 * must ensure this doesn't happen until the switch is completely
4639 	 * finished.
4640 	 *
4641 	 * In particular, the load of prev->state in finish_task_switch() must
4642 	 * happen before this.
4643 	 *
4644 	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4645 	 */
4646 	smp_store_release(&prev->on_cpu, 0);
4647 #endif
4648 }
4649 
4650 #ifdef CONFIG_SMP
4651 
4652 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4653 {
4654 	void (*func)(struct rq *rq);
4655 	struct callback_head *next;
4656 
4657 	lockdep_assert_rq_held(rq);
4658 
4659 	while (head) {
4660 		func = (void (*)(struct rq *))head->func;
4661 		next = head->next;
4662 		head->next = NULL;
4663 		head = next;
4664 
4665 		func(rq);
4666 	}
4667 }
4668 
4669 static void balance_push(struct rq *rq);
4670 
4671 struct callback_head balance_push_callback = {
4672 	.next = NULL,
4673 	.func = (void (*)(struct callback_head *))balance_push,
4674 };
4675 
4676 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4677 {
4678 	struct callback_head *head = rq->balance_callback;
4679 
4680 	lockdep_assert_rq_held(rq);
4681 	if (head)
4682 		rq->balance_callback = NULL;
4683 
4684 	return head;
4685 }
4686 
4687 static void __balance_callbacks(struct rq *rq)
4688 {
4689 	do_balance_callbacks(rq, splice_balance_callbacks(rq));
4690 }
4691 
4692 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4693 {
4694 	unsigned long flags;
4695 
4696 	if (unlikely(head)) {
4697 		raw_spin_rq_lock_irqsave(rq, flags);
4698 		do_balance_callbacks(rq, head);
4699 		raw_spin_rq_unlock_irqrestore(rq, flags);
4700 	}
4701 }
4702 
4703 #else
4704 
4705 static inline void __balance_callbacks(struct rq *rq)
4706 {
4707 }
4708 
4709 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4710 {
4711 	return NULL;
4712 }
4713 
4714 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4715 {
4716 }
4717 
4718 #endif
4719 
4720 static inline void
4721 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4722 {
4723 	/*
4724 	 * Since the runqueue lock will be released by the next
4725 	 * task (which is an invalid locking op but in the case
4726 	 * of the scheduler it's an obvious special-case), so we
4727 	 * do an early lockdep release here:
4728 	 */
4729 	rq_unpin_lock(rq, rf);
4730 	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4731 #ifdef CONFIG_DEBUG_SPINLOCK
4732 	/* this is a valid case when another task releases the spinlock */
4733 	rq_lockp(rq)->owner = next;
4734 #endif
4735 }
4736 
4737 static inline void finish_lock_switch(struct rq *rq)
4738 {
4739 	/*
4740 	 * If we are tracking spinlock dependencies then we have to
4741 	 * fix up the runqueue lock - which gets 'carried over' from
4742 	 * prev into current:
4743 	 */
4744 	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4745 	__balance_callbacks(rq);
4746 	raw_spin_rq_unlock_irq(rq);
4747 }
4748 
4749 /*
4750  * NOP if the arch has not defined these:
4751  */
4752 
4753 #ifndef prepare_arch_switch
4754 # define prepare_arch_switch(next)	do { } while (0)
4755 #endif
4756 
4757 #ifndef finish_arch_post_lock_switch
4758 # define finish_arch_post_lock_switch()	do { } while (0)
4759 #endif
4760 
4761 static inline void kmap_local_sched_out(void)
4762 {
4763 #ifdef CONFIG_KMAP_LOCAL
4764 	if (unlikely(current->kmap_ctrl.idx))
4765 		__kmap_local_sched_out();
4766 #endif
4767 }
4768 
4769 static inline void kmap_local_sched_in(void)
4770 {
4771 #ifdef CONFIG_KMAP_LOCAL
4772 	if (unlikely(current->kmap_ctrl.idx))
4773 		__kmap_local_sched_in();
4774 #endif
4775 }
4776 
4777 /**
4778  * prepare_task_switch - prepare to switch tasks
4779  * @rq: the runqueue preparing to switch
4780  * @prev: the current task that is being switched out
4781  * @next: the task we are going to switch to.
4782  *
4783  * This is called with the rq lock held and interrupts off. It must
4784  * be paired with a subsequent finish_task_switch after the context
4785  * switch.
4786  *
4787  * prepare_task_switch sets up locking and calls architecture specific
4788  * hooks.
4789  */
4790 static inline void
4791 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4792 		    struct task_struct *next)
4793 {
4794 	kcov_prepare_switch(prev);
4795 	sched_info_switch(rq, prev, next);
4796 	perf_event_task_sched_out(prev, next);
4797 	rseq_preempt(prev);
4798 	fire_sched_out_preempt_notifiers(prev, next);
4799 	kmap_local_sched_out();
4800 	prepare_task(next);
4801 	prepare_arch_switch(next);
4802 }
4803 
4804 /**
4805  * finish_task_switch - clean up after a task-switch
4806  * @prev: the thread we just switched away from.
4807  *
4808  * finish_task_switch must be called after the context switch, paired
4809  * with a prepare_task_switch call before the context switch.
4810  * finish_task_switch will reconcile locking set up by prepare_task_switch,
4811  * and do any other architecture-specific cleanup actions.
4812  *
4813  * Note that we may have delayed dropping an mm in context_switch(). If
4814  * so, we finish that here outside of the runqueue lock. (Doing it
4815  * with the lock held can cause deadlocks; see schedule() for
4816  * details.)
4817  *
4818  * The context switch have flipped the stack from under us and restored the
4819  * local variables which were saved when this task called schedule() in the
4820  * past. prev == current is still correct but we need to recalculate this_rq
4821  * because prev may have moved to another CPU.
4822  */
4823 static struct rq *finish_task_switch(struct task_struct *prev)
4824 	__releases(rq->lock)
4825 {
4826 	struct rq *rq = this_rq();
4827 	struct mm_struct *mm = rq->prev_mm;
4828 	long prev_state;
4829 
4830 	/*
4831 	 * The previous task will have left us with a preempt_count of 2
4832 	 * because it left us after:
4833 	 *
4834 	 *	schedule()
4835 	 *	  preempt_disable();			// 1
4836 	 *	  __schedule()
4837 	 *	    raw_spin_lock_irq(&rq->lock)	// 2
4838 	 *
4839 	 * Also, see FORK_PREEMPT_COUNT.
4840 	 */
4841 	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4842 		      "corrupted preempt_count: %s/%d/0x%x\n",
4843 		      current->comm, current->pid, preempt_count()))
4844 		preempt_count_set(FORK_PREEMPT_COUNT);
4845 
4846 	rq->prev_mm = NULL;
4847 
4848 	/*
4849 	 * A task struct has one reference for the use as "current".
4850 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4851 	 * schedule one last time. The schedule call will never return, and
4852 	 * the scheduled task must drop that reference.
4853 	 *
4854 	 * We must observe prev->state before clearing prev->on_cpu (in
4855 	 * finish_task), otherwise a concurrent wakeup can get prev
4856 	 * running on another CPU and we could rave with its RUNNING -> DEAD
4857 	 * transition, resulting in a double drop.
4858 	 */
4859 	prev_state = READ_ONCE(prev->__state);
4860 	vtime_task_switch(prev);
4861 	perf_event_task_sched_in(prev, current);
4862 	finish_task(prev);
4863 	tick_nohz_task_switch();
4864 	finish_lock_switch(rq);
4865 	finish_arch_post_lock_switch();
4866 	kcov_finish_switch(current);
4867 	/*
4868 	 * kmap_local_sched_out() is invoked with rq::lock held and
4869 	 * interrupts disabled. There is no requirement for that, but the
4870 	 * sched out code does not have an interrupt enabled section.
4871 	 * Restoring the maps on sched in does not require interrupts being
4872 	 * disabled either.
4873 	 */
4874 	kmap_local_sched_in();
4875 
4876 	fire_sched_in_preempt_notifiers(current);
4877 	/*
4878 	 * When switching through a kernel thread, the loop in
4879 	 * membarrier_{private,global}_expedited() may have observed that
4880 	 * kernel thread and not issued an IPI. It is therefore possible to
4881 	 * schedule between user->kernel->user threads without passing though
4882 	 * switch_mm(). Membarrier requires a barrier after storing to
4883 	 * rq->curr, before returning to userspace, so provide them here:
4884 	 *
4885 	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4886 	 *   provided by mmdrop(),
4887 	 * - a sync_core for SYNC_CORE.
4888 	 */
4889 	if (mm) {
4890 		membarrier_mm_sync_core_before_usermode(mm);
4891 		mmdrop_sched(mm);
4892 	}
4893 	if (unlikely(prev_state == TASK_DEAD)) {
4894 		if (prev->sched_class->task_dead)
4895 			prev->sched_class->task_dead(prev);
4896 
4897 		/* Task is done with its stack. */
4898 		put_task_stack(prev);
4899 
4900 		put_task_struct_rcu_user(prev);
4901 	}
4902 
4903 	return rq;
4904 }
4905 
4906 /**
4907  * schedule_tail - first thing a freshly forked thread must call.
4908  * @prev: the thread we just switched away from.
4909  */
4910 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4911 	__releases(rq->lock)
4912 {
4913 	/*
4914 	 * New tasks start with FORK_PREEMPT_COUNT, see there and
4915 	 * finish_task_switch() for details.
4916 	 *
4917 	 * finish_task_switch() will drop rq->lock() and lower preempt_count
4918 	 * and the preempt_enable() will end up enabling preemption (on
4919 	 * PREEMPT_COUNT kernels).
4920 	 */
4921 
4922 	finish_task_switch(prev);
4923 	preempt_enable();
4924 
4925 	if (current->set_child_tid)
4926 		put_user(task_pid_vnr(current), current->set_child_tid);
4927 
4928 	calculate_sigpending();
4929 }
4930 
4931 /*
4932  * context_switch - switch to the new MM and the new thread's register state.
4933  */
4934 static __always_inline struct rq *
4935 context_switch(struct rq *rq, struct task_struct *prev,
4936 	       struct task_struct *next, struct rq_flags *rf)
4937 {
4938 	prepare_task_switch(rq, prev, next);
4939 
4940 	/*
4941 	 * For paravirt, this is coupled with an exit in switch_to to
4942 	 * combine the page table reload and the switch backend into
4943 	 * one hypercall.
4944 	 */
4945 	arch_start_context_switch(prev);
4946 
4947 	/*
4948 	 * kernel -> kernel   lazy + transfer active
4949 	 *   user -> kernel   lazy + mmgrab() active
4950 	 *
4951 	 * kernel ->   user   switch + mmdrop() active
4952 	 *   user ->   user   switch
4953 	 */
4954 	if (!next->mm) {                                // to kernel
4955 		enter_lazy_tlb(prev->active_mm, next);
4956 
4957 		next->active_mm = prev->active_mm;
4958 		if (prev->mm)                           // from user
4959 			mmgrab(prev->active_mm);
4960 		else
4961 			prev->active_mm = NULL;
4962 	} else {                                        // to user
4963 		membarrier_switch_mm(rq, prev->active_mm, next->mm);
4964 		/*
4965 		 * sys_membarrier() requires an smp_mb() between setting
4966 		 * rq->curr / membarrier_switch_mm() and returning to userspace.
4967 		 *
4968 		 * The below provides this either through switch_mm(), or in
4969 		 * case 'prev->active_mm == next->mm' through
4970 		 * finish_task_switch()'s mmdrop().
4971 		 */
4972 		switch_mm_irqs_off(prev->active_mm, next->mm, next);
4973 
4974 		if (!prev->mm) {                        // from kernel
4975 			/* will mmdrop() in finish_task_switch(). */
4976 			rq->prev_mm = prev->active_mm;
4977 			prev->active_mm = NULL;
4978 		}
4979 	}
4980 
4981 	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4982 
4983 	prepare_lock_switch(rq, next, rf);
4984 
4985 	/* Here we just switch the register state and the stack. */
4986 	switch_to(prev, next, prev);
4987 	barrier();
4988 
4989 	return finish_task_switch(prev);
4990 }
4991 
4992 /*
4993  * nr_running and nr_context_switches:
4994  *
4995  * externally visible scheduler statistics: current number of runnable
4996  * threads, total number of context switches performed since bootup.
4997  */
4998 unsigned int nr_running(void)
4999 {
5000 	unsigned int i, sum = 0;
5001 
5002 	for_each_online_cpu(i)
5003 		sum += cpu_rq(i)->nr_running;
5004 
5005 	return sum;
5006 }
5007 
5008 /*
5009  * Check if only the current task is running on the CPU.
5010  *
5011  * Caution: this function does not check that the caller has disabled
5012  * preemption, thus the result might have a time-of-check-to-time-of-use
5013  * race.  The caller is responsible to use it correctly, for example:
5014  *
5015  * - from a non-preemptible section (of course)
5016  *
5017  * - from a thread that is bound to a single CPU
5018  *
5019  * - in a loop with very short iterations (e.g. a polling loop)
5020  */
5021 bool single_task_running(void)
5022 {
5023 	return raw_rq()->nr_running == 1;
5024 }
5025 EXPORT_SYMBOL(single_task_running);
5026 
5027 unsigned long long nr_context_switches(void)
5028 {
5029 	int i;
5030 	unsigned long long sum = 0;
5031 
5032 	for_each_possible_cpu(i)
5033 		sum += cpu_rq(i)->nr_switches;
5034 
5035 	return sum;
5036 }
5037 
5038 /*
5039  * Consumers of these two interfaces, like for example the cpuidle menu
5040  * governor, are using nonsensical data. Preferring shallow idle state selection
5041  * for a CPU that has IO-wait which might not even end up running the task when
5042  * it does become runnable.
5043  */
5044 
5045 unsigned int nr_iowait_cpu(int cpu)
5046 {
5047 	return atomic_read(&cpu_rq(cpu)->nr_iowait);
5048 }
5049 
5050 /*
5051  * IO-wait accounting, and how it's mostly bollocks (on SMP).
5052  *
5053  * The idea behind IO-wait account is to account the idle time that we could
5054  * have spend running if it were not for IO. That is, if we were to improve the
5055  * storage performance, we'd have a proportional reduction in IO-wait time.
5056  *
5057  * This all works nicely on UP, where, when a task blocks on IO, we account
5058  * idle time as IO-wait, because if the storage were faster, it could've been
5059  * running and we'd not be idle.
5060  *
5061  * This has been extended to SMP, by doing the same for each CPU. This however
5062  * is broken.
5063  *
5064  * Imagine for instance the case where two tasks block on one CPU, only the one
5065  * CPU will have IO-wait accounted, while the other has regular idle. Even
5066  * though, if the storage were faster, both could've ran at the same time,
5067  * utilising both CPUs.
5068  *
5069  * This means, that when looking globally, the current IO-wait accounting on
5070  * SMP is a lower bound, by reason of under accounting.
5071  *
5072  * Worse, since the numbers are provided per CPU, they are sometimes
5073  * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5074  * associated with any one particular CPU, it can wake to another CPU than it
5075  * blocked on. This means the per CPU IO-wait number is meaningless.
5076  *
5077  * Task CPU affinities can make all that even more 'interesting'.
5078  */
5079 
5080 unsigned int nr_iowait(void)
5081 {
5082 	unsigned int i, sum = 0;
5083 
5084 	for_each_possible_cpu(i)
5085 		sum += nr_iowait_cpu(i);
5086 
5087 	return sum;
5088 }
5089 
5090 #ifdef CONFIG_SMP
5091 
5092 /*
5093  * sched_exec - execve() is a valuable balancing opportunity, because at
5094  * this point the task has the smallest effective memory and cache footprint.
5095  */
5096 void sched_exec(void)
5097 {
5098 	struct task_struct *p = current;
5099 	unsigned long flags;
5100 	int dest_cpu;
5101 
5102 	raw_spin_lock_irqsave(&p->pi_lock, flags);
5103 	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5104 	if (dest_cpu == smp_processor_id())
5105 		goto unlock;
5106 
5107 	if (likely(cpu_active(dest_cpu))) {
5108 		struct migration_arg arg = { p, dest_cpu };
5109 
5110 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5111 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5112 		return;
5113 	}
5114 unlock:
5115 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5116 }
5117 
5118 #endif
5119 
5120 DEFINE_PER_CPU(struct kernel_stat, kstat);
5121 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5122 
5123 EXPORT_PER_CPU_SYMBOL(kstat);
5124 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5125 
5126 /*
5127  * The function fair_sched_class.update_curr accesses the struct curr
5128  * and its field curr->exec_start; when called from task_sched_runtime(),
5129  * we observe a high rate of cache misses in practice.
5130  * Prefetching this data results in improved performance.
5131  */
5132 static inline void prefetch_curr_exec_start(struct task_struct *p)
5133 {
5134 #ifdef CONFIG_FAIR_GROUP_SCHED
5135 	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5136 #else
5137 	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5138 #endif
5139 	prefetch(curr);
5140 	prefetch(&curr->exec_start);
5141 }
5142 
5143 /*
5144  * Return accounted runtime for the task.
5145  * In case the task is currently running, return the runtime plus current's
5146  * pending runtime that have not been accounted yet.
5147  */
5148 unsigned long long task_sched_runtime(struct task_struct *p)
5149 {
5150 	struct rq_flags rf;
5151 	struct rq *rq;
5152 	u64 ns;
5153 
5154 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5155 	/*
5156 	 * 64-bit doesn't need locks to atomically read a 64-bit value.
5157 	 * So we have a optimization chance when the task's delta_exec is 0.
5158 	 * Reading ->on_cpu is racy, but this is ok.
5159 	 *
5160 	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5161 	 * If we race with it entering CPU, unaccounted time is 0. This is
5162 	 * indistinguishable from the read occurring a few cycles earlier.
5163 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5164 	 * been accounted, so we're correct here as well.
5165 	 */
5166 	if (!p->on_cpu || !task_on_rq_queued(p))
5167 		return p->se.sum_exec_runtime;
5168 #endif
5169 
5170 	rq = task_rq_lock(p, &rf);
5171 	/*
5172 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
5173 	 * project cycles that may never be accounted to this
5174 	 * thread, breaking clock_gettime().
5175 	 */
5176 	if (task_current(rq, p) && task_on_rq_queued(p)) {
5177 		prefetch_curr_exec_start(p);
5178 		update_rq_clock(rq);
5179 		p->sched_class->update_curr(rq);
5180 	}
5181 	ns = p->se.sum_exec_runtime;
5182 	task_rq_unlock(rq, p, &rf);
5183 
5184 	return ns;
5185 }
5186 
5187 #ifdef CONFIG_SCHED_DEBUG
5188 static u64 cpu_resched_latency(struct rq *rq)
5189 {
5190 	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5191 	u64 resched_latency, now = rq_clock(rq);
5192 	static bool warned_once;
5193 
5194 	if (sysctl_resched_latency_warn_once && warned_once)
5195 		return 0;
5196 
5197 	if (!need_resched() || !latency_warn_ms)
5198 		return 0;
5199 
5200 	if (system_state == SYSTEM_BOOTING)
5201 		return 0;
5202 
5203 	if (!rq->last_seen_need_resched_ns) {
5204 		rq->last_seen_need_resched_ns = now;
5205 		rq->ticks_without_resched = 0;
5206 		return 0;
5207 	}
5208 
5209 	rq->ticks_without_resched++;
5210 	resched_latency = now - rq->last_seen_need_resched_ns;
5211 	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5212 		return 0;
5213 
5214 	warned_once = true;
5215 
5216 	return resched_latency;
5217 }
5218 
5219 static int __init setup_resched_latency_warn_ms(char *str)
5220 {
5221 	long val;
5222 
5223 	if ((kstrtol(str, 0, &val))) {
5224 		pr_warn("Unable to set resched_latency_warn_ms\n");
5225 		return 1;
5226 	}
5227 
5228 	sysctl_resched_latency_warn_ms = val;
5229 	return 1;
5230 }
5231 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5232 #else
5233 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5234 #endif /* CONFIG_SCHED_DEBUG */
5235 
5236 /*
5237  * This function gets called by the timer code, with HZ frequency.
5238  * We call it with interrupts disabled.
5239  */
5240 void scheduler_tick(void)
5241 {
5242 	int cpu = smp_processor_id();
5243 	struct rq *rq = cpu_rq(cpu);
5244 	struct task_struct *curr = rq->curr;
5245 	struct rq_flags rf;
5246 	unsigned long thermal_pressure;
5247 	u64 resched_latency;
5248 
5249 	arch_scale_freq_tick();
5250 	sched_clock_tick();
5251 
5252 	rq_lock(rq, &rf);
5253 
5254 	update_rq_clock(rq);
5255 	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5256 	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5257 	curr->sched_class->task_tick(rq, curr, 0);
5258 	if (sched_feat(LATENCY_WARN))
5259 		resched_latency = cpu_resched_latency(rq);
5260 	calc_global_load_tick(rq);
5261 	sched_core_tick(rq);
5262 
5263 	rq_unlock(rq, &rf);
5264 
5265 	if (sched_feat(LATENCY_WARN) && resched_latency)
5266 		resched_latency_warn(cpu, resched_latency);
5267 
5268 	perf_event_task_tick();
5269 
5270 #ifdef CONFIG_SMP
5271 	rq->idle_balance = idle_cpu(cpu);
5272 	trigger_load_balance(rq);
5273 #endif
5274 }
5275 
5276 #ifdef CONFIG_NO_HZ_FULL
5277 
5278 struct tick_work {
5279 	int			cpu;
5280 	atomic_t		state;
5281 	struct delayed_work	work;
5282 };
5283 /* Values for ->state, see diagram below. */
5284 #define TICK_SCHED_REMOTE_OFFLINE	0
5285 #define TICK_SCHED_REMOTE_OFFLINING	1
5286 #define TICK_SCHED_REMOTE_RUNNING	2
5287 
5288 /*
5289  * State diagram for ->state:
5290  *
5291  *
5292  *          TICK_SCHED_REMOTE_OFFLINE
5293  *                    |   ^
5294  *                    |   |
5295  *                    |   | sched_tick_remote()
5296  *                    |   |
5297  *                    |   |
5298  *                    +--TICK_SCHED_REMOTE_OFFLINING
5299  *                    |   ^
5300  *                    |   |
5301  * sched_tick_start() |   | sched_tick_stop()
5302  *                    |   |
5303  *                    V   |
5304  *          TICK_SCHED_REMOTE_RUNNING
5305  *
5306  *
5307  * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5308  * and sched_tick_start() are happy to leave the state in RUNNING.
5309  */
5310 
5311 static struct tick_work __percpu *tick_work_cpu;
5312 
5313 static void sched_tick_remote(struct work_struct *work)
5314 {
5315 	struct delayed_work *dwork = to_delayed_work(work);
5316 	struct tick_work *twork = container_of(dwork, struct tick_work, work);
5317 	int cpu = twork->cpu;
5318 	struct rq *rq = cpu_rq(cpu);
5319 	struct task_struct *curr;
5320 	struct rq_flags rf;
5321 	u64 delta;
5322 	int os;
5323 
5324 	/*
5325 	 * Handle the tick only if it appears the remote CPU is running in full
5326 	 * dynticks mode. The check is racy by nature, but missing a tick or
5327 	 * having one too much is no big deal because the scheduler tick updates
5328 	 * statistics and checks timeslices in a time-independent way, regardless
5329 	 * of when exactly it is running.
5330 	 */
5331 	if (!tick_nohz_tick_stopped_cpu(cpu))
5332 		goto out_requeue;
5333 
5334 	rq_lock_irq(rq, &rf);
5335 	curr = rq->curr;
5336 	if (cpu_is_offline(cpu))
5337 		goto out_unlock;
5338 
5339 	update_rq_clock(rq);
5340 
5341 	if (!is_idle_task(curr)) {
5342 		/*
5343 		 * Make sure the next tick runs within a reasonable
5344 		 * amount of time.
5345 		 */
5346 		delta = rq_clock_task(rq) - curr->se.exec_start;
5347 		WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5348 	}
5349 	curr->sched_class->task_tick(rq, curr, 0);
5350 
5351 	calc_load_nohz_remote(rq);
5352 out_unlock:
5353 	rq_unlock_irq(rq, &rf);
5354 out_requeue:
5355 
5356 	/*
5357 	 * Run the remote tick once per second (1Hz). This arbitrary
5358 	 * frequency is large enough to avoid overload but short enough
5359 	 * to keep scheduler internal stats reasonably up to date.  But
5360 	 * first update state to reflect hotplug activity if required.
5361 	 */
5362 	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5363 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5364 	if (os == TICK_SCHED_REMOTE_RUNNING)
5365 		queue_delayed_work(system_unbound_wq, dwork, HZ);
5366 }
5367 
5368 static void sched_tick_start(int cpu)
5369 {
5370 	int os;
5371 	struct tick_work *twork;
5372 
5373 	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5374 		return;
5375 
5376 	WARN_ON_ONCE(!tick_work_cpu);
5377 
5378 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5379 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5380 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5381 	if (os == TICK_SCHED_REMOTE_OFFLINE) {
5382 		twork->cpu = cpu;
5383 		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5384 		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5385 	}
5386 }
5387 
5388 #ifdef CONFIG_HOTPLUG_CPU
5389 static void sched_tick_stop(int cpu)
5390 {
5391 	struct tick_work *twork;
5392 	int os;
5393 
5394 	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5395 		return;
5396 
5397 	WARN_ON_ONCE(!tick_work_cpu);
5398 
5399 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5400 	/* There cannot be competing actions, but don't rely on stop-machine. */
5401 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5402 	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5403 	/* Don't cancel, as this would mess up the state machine. */
5404 }
5405 #endif /* CONFIG_HOTPLUG_CPU */
5406 
5407 int __init sched_tick_offload_init(void)
5408 {
5409 	tick_work_cpu = alloc_percpu(struct tick_work);
5410 	BUG_ON(!tick_work_cpu);
5411 	return 0;
5412 }
5413 
5414 #else /* !CONFIG_NO_HZ_FULL */
5415 static inline void sched_tick_start(int cpu) { }
5416 static inline void sched_tick_stop(int cpu) { }
5417 #endif
5418 
5419 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5420 				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5421 /*
5422  * If the value passed in is equal to the current preempt count
5423  * then we just disabled preemption. Start timing the latency.
5424  */
5425 static inline void preempt_latency_start(int val)
5426 {
5427 	if (preempt_count() == val) {
5428 		unsigned long ip = get_lock_parent_ip();
5429 #ifdef CONFIG_DEBUG_PREEMPT
5430 		current->preempt_disable_ip = ip;
5431 #endif
5432 		trace_preempt_off(CALLER_ADDR0, ip);
5433 	}
5434 }
5435 
5436 void preempt_count_add(int val)
5437 {
5438 #ifdef CONFIG_DEBUG_PREEMPT
5439 	/*
5440 	 * Underflow?
5441 	 */
5442 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5443 		return;
5444 #endif
5445 	__preempt_count_add(val);
5446 #ifdef CONFIG_DEBUG_PREEMPT
5447 	/*
5448 	 * Spinlock count overflowing soon?
5449 	 */
5450 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5451 				PREEMPT_MASK - 10);
5452 #endif
5453 	preempt_latency_start(val);
5454 }
5455 EXPORT_SYMBOL(preempt_count_add);
5456 NOKPROBE_SYMBOL(preempt_count_add);
5457 
5458 /*
5459  * If the value passed in equals to the current preempt count
5460  * then we just enabled preemption. Stop timing the latency.
5461  */
5462 static inline void preempt_latency_stop(int val)
5463 {
5464 	if (preempt_count() == val)
5465 		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5466 }
5467 
5468 void preempt_count_sub(int val)
5469 {
5470 #ifdef CONFIG_DEBUG_PREEMPT
5471 	/*
5472 	 * Underflow?
5473 	 */
5474 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5475 		return;
5476 	/*
5477 	 * Is the spinlock portion underflowing?
5478 	 */
5479 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5480 			!(preempt_count() & PREEMPT_MASK)))
5481 		return;
5482 #endif
5483 
5484 	preempt_latency_stop(val);
5485 	__preempt_count_sub(val);
5486 }
5487 EXPORT_SYMBOL(preempt_count_sub);
5488 NOKPROBE_SYMBOL(preempt_count_sub);
5489 
5490 #else
5491 static inline void preempt_latency_start(int val) { }
5492 static inline void preempt_latency_stop(int val) { }
5493 #endif
5494 
5495 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5496 {
5497 #ifdef CONFIG_DEBUG_PREEMPT
5498 	return p->preempt_disable_ip;
5499 #else
5500 	return 0;
5501 #endif
5502 }
5503 
5504 /*
5505  * Print scheduling while atomic bug:
5506  */
5507 static noinline void __schedule_bug(struct task_struct *prev)
5508 {
5509 	/* Save this before calling printk(), since that will clobber it */
5510 	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5511 
5512 	if (oops_in_progress)
5513 		return;
5514 
5515 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5516 		prev->comm, prev->pid, preempt_count());
5517 
5518 	debug_show_held_locks(prev);
5519 	print_modules();
5520 	if (irqs_disabled())
5521 		print_irqtrace_events(prev);
5522 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5523 	    && in_atomic_preempt_off()) {
5524 		pr_err("Preemption disabled at:");
5525 		print_ip_sym(KERN_ERR, preempt_disable_ip);
5526 	}
5527 	if (panic_on_warn)
5528 		panic("scheduling while atomic\n");
5529 
5530 	dump_stack();
5531 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5532 }
5533 
5534 /*
5535  * Various schedule()-time debugging checks and statistics:
5536  */
5537 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5538 {
5539 #ifdef CONFIG_SCHED_STACK_END_CHECK
5540 	if (task_stack_end_corrupted(prev))
5541 		panic("corrupted stack end detected inside scheduler\n");
5542 
5543 	if (task_scs_end_corrupted(prev))
5544 		panic("corrupted shadow stack detected inside scheduler\n");
5545 #endif
5546 
5547 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5548 	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5549 		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5550 			prev->comm, prev->pid, prev->non_block_count);
5551 		dump_stack();
5552 		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5553 	}
5554 #endif
5555 
5556 	if (unlikely(in_atomic_preempt_off())) {
5557 		__schedule_bug(prev);
5558 		preempt_count_set(PREEMPT_DISABLED);
5559 	}
5560 	rcu_sleep_check();
5561 	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5562 
5563 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5564 
5565 	schedstat_inc(this_rq()->sched_count);
5566 }
5567 
5568 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5569 				  struct rq_flags *rf)
5570 {
5571 #ifdef CONFIG_SMP
5572 	const struct sched_class *class;
5573 	/*
5574 	 * We must do the balancing pass before put_prev_task(), such
5575 	 * that when we release the rq->lock the task is in the same
5576 	 * state as before we took rq->lock.
5577 	 *
5578 	 * We can terminate the balance pass as soon as we know there is
5579 	 * a runnable task of @class priority or higher.
5580 	 */
5581 	for_class_range(class, prev->sched_class, &idle_sched_class) {
5582 		if (class->balance(rq, prev, rf))
5583 			break;
5584 	}
5585 #endif
5586 
5587 	put_prev_task(rq, prev);
5588 }
5589 
5590 /*
5591  * Pick up the highest-prio task:
5592  */
5593 static inline struct task_struct *
5594 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5595 {
5596 	const struct sched_class *class;
5597 	struct task_struct *p;
5598 
5599 	/*
5600 	 * Optimization: we know that if all tasks are in the fair class we can
5601 	 * call that function directly, but only if the @prev task wasn't of a
5602 	 * higher scheduling class, because otherwise those lose the
5603 	 * opportunity to pull in more work from other CPUs.
5604 	 */
5605 	if (likely(prev->sched_class <= &fair_sched_class &&
5606 		   rq->nr_running == rq->cfs.h_nr_running)) {
5607 
5608 		p = pick_next_task_fair(rq, prev, rf);
5609 		if (unlikely(p == RETRY_TASK))
5610 			goto restart;
5611 
5612 		/* Assume the next prioritized class is idle_sched_class */
5613 		if (!p) {
5614 			put_prev_task(rq, prev);
5615 			p = pick_next_task_idle(rq);
5616 		}
5617 
5618 		return p;
5619 	}
5620 
5621 restart:
5622 	put_prev_task_balance(rq, prev, rf);
5623 
5624 	for_each_class(class) {
5625 		p = class->pick_next_task(rq);
5626 		if (p)
5627 			return p;
5628 	}
5629 
5630 	BUG(); /* The idle class should always have a runnable task. */
5631 }
5632 
5633 #ifdef CONFIG_SCHED_CORE
5634 static inline bool is_task_rq_idle(struct task_struct *t)
5635 {
5636 	return (task_rq(t)->idle == t);
5637 }
5638 
5639 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5640 {
5641 	return is_task_rq_idle(a) || (a->core_cookie == cookie);
5642 }
5643 
5644 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5645 {
5646 	if (is_task_rq_idle(a) || is_task_rq_idle(b))
5647 		return true;
5648 
5649 	return a->core_cookie == b->core_cookie;
5650 }
5651 
5652 static inline struct task_struct *pick_task(struct rq *rq)
5653 {
5654 	const struct sched_class *class;
5655 	struct task_struct *p;
5656 
5657 	for_each_class(class) {
5658 		p = class->pick_task(rq);
5659 		if (p)
5660 			return p;
5661 	}
5662 
5663 	BUG(); /* The idle class should always have a runnable task. */
5664 }
5665 
5666 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5667 
5668 static struct task_struct *
5669 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5670 {
5671 	struct task_struct *next, *p, *max = NULL;
5672 	const struct cpumask *smt_mask;
5673 	bool fi_before = false;
5674 	bool core_clock_updated = (rq == rq->core);
5675 	unsigned long cookie;
5676 	int i, cpu, occ = 0;
5677 	struct rq *rq_i;
5678 	bool need_sync;
5679 
5680 	if (!sched_core_enabled(rq))
5681 		return __pick_next_task(rq, prev, rf);
5682 
5683 	cpu = cpu_of(rq);
5684 
5685 	/* Stopper task is switching into idle, no need core-wide selection. */
5686 	if (cpu_is_offline(cpu)) {
5687 		/*
5688 		 * Reset core_pick so that we don't enter the fastpath when
5689 		 * coming online. core_pick would already be migrated to
5690 		 * another cpu during offline.
5691 		 */
5692 		rq->core_pick = NULL;
5693 		return __pick_next_task(rq, prev, rf);
5694 	}
5695 
5696 	/*
5697 	 * If there were no {en,de}queues since we picked (IOW, the task
5698 	 * pointers are all still valid), and we haven't scheduled the last
5699 	 * pick yet, do so now.
5700 	 *
5701 	 * rq->core_pick can be NULL if no selection was made for a CPU because
5702 	 * it was either offline or went offline during a sibling's core-wide
5703 	 * selection. In this case, do a core-wide selection.
5704 	 */
5705 	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5706 	    rq->core->core_pick_seq != rq->core_sched_seq &&
5707 	    rq->core_pick) {
5708 		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5709 
5710 		next = rq->core_pick;
5711 		if (next != prev) {
5712 			put_prev_task(rq, prev);
5713 			set_next_task(rq, next);
5714 		}
5715 
5716 		rq->core_pick = NULL;
5717 		return next;
5718 	}
5719 
5720 	put_prev_task_balance(rq, prev, rf);
5721 
5722 	smt_mask = cpu_smt_mask(cpu);
5723 	need_sync = !!rq->core->core_cookie;
5724 
5725 	/* reset state */
5726 	rq->core->core_cookie = 0UL;
5727 	if (rq->core->core_forceidle_count) {
5728 		if (!core_clock_updated) {
5729 			update_rq_clock(rq->core);
5730 			core_clock_updated = true;
5731 		}
5732 		sched_core_account_forceidle(rq);
5733 		/* reset after accounting force idle */
5734 		rq->core->core_forceidle_start = 0;
5735 		rq->core->core_forceidle_count = 0;
5736 		rq->core->core_forceidle_occupation = 0;
5737 		need_sync = true;
5738 		fi_before = true;
5739 	}
5740 
5741 	/*
5742 	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5743 	 *
5744 	 * @task_seq guards the task state ({en,de}queues)
5745 	 * @pick_seq is the @task_seq we did a selection on
5746 	 * @sched_seq is the @pick_seq we scheduled
5747 	 *
5748 	 * However, preemptions can cause multiple picks on the same task set.
5749 	 * 'Fix' this by also increasing @task_seq for every pick.
5750 	 */
5751 	rq->core->core_task_seq++;
5752 
5753 	/*
5754 	 * Optimize for common case where this CPU has no cookies
5755 	 * and there are no cookied tasks running on siblings.
5756 	 */
5757 	if (!need_sync) {
5758 		next = pick_task(rq);
5759 		if (!next->core_cookie) {
5760 			rq->core_pick = NULL;
5761 			/*
5762 			 * For robustness, update the min_vruntime_fi for
5763 			 * unconstrained picks as well.
5764 			 */
5765 			WARN_ON_ONCE(fi_before);
5766 			task_vruntime_update(rq, next, false);
5767 			goto done;
5768 		}
5769 	}
5770 
5771 	/*
5772 	 * For each thread: do the regular task pick and find the max prio task
5773 	 * amongst them.
5774 	 *
5775 	 * Tie-break prio towards the current CPU
5776 	 */
5777 	for_each_cpu_wrap(i, smt_mask, cpu) {
5778 		rq_i = cpu_rq(i);
5779 
5780 		/*
5781 		 * Current cpu always has its clock updated on entrance to
5782 		 * pick_next_task(). If the current cpu is not the core,
5783 		 * the core may also have been updated above.
5784 		 */
5785 		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5786 			update_rq_clock(rq_i);
5787 
5788 		p = rq_i->core_pick = pick_task(rq_i);
5789 		if (!max || prio_less(max, p, fi_before))
5790 			max = p;
5791 	}
5792 
5793 	cookie = rq->core->core_cookie = max->core_cookie;
5794 
5795 	/*
5796 	 * For each thread: try and find a runnable task that matches @max or
5797 	 * force idle.
5798 	 */
5799 	for_each_cpu(i, smt_mask) {
5800 		rq_i = cpu_rq(i);
5801 		p = rq_i->core_pick;
5802 
5803 		if (!cookie_equals(p, cookie)) {
5804 			p = NULL;
5805 			if (cookie)
5806 				p = sched_core_find(rq_i, cookie);
5807 			if (!p)
5808 				p = idle_sched_class.pick_task(rq_i);
5809 		}
5810 
5811 		rq_i->core_pick = p;
5812 
5813 		if (p == rq_i->idle) {
5814 			if (rq_i->nr_running) {
5815 				rq->core->core_forceidle_count++;
5816 				if (!fi_before)
5817 					rq->core->core_forceidle_seq++;
5818 			}
5819 		} else {
5820 			occ++;
5821 		}
5822 	}
5823 
5824 	if (schedstat_enabled() && rq->core->core_forceidle_count) {
5825 		rq->core->core_forceidle_start = rq_clock(rq->core);
5826 		rq->core->core_forceidle_occupation = occ;
5827 	}
5828 
5829 	rq->core->core_pick_seq = rq->core->core_task_seq;
5830 	next = rq->core_pick;
5831 	rq->core_sched_seq = rq->core->core_pick_seq;
5832 
5833 	/* Something should have been selected for current CPU */
5834 	WARN_ON_ONCE(!next);
5835 
5836 	/*
5837 	 * Reschedule siblings
5838 	 *
5839 	 * NOTE: L1TF -- at this point we're no longer running the old task and
5840 	 * sending an IPI (below) ensures the sibling will no longer be running
5841 	 * their task. This ensures there is no inter-sibling overlap between
5842 	 * non-matching user state.
5843 	 */
5844 	for_each_cpu(i, smt_mask) {
5845 		rq_i = cpu_rq(i);
5846 
5847 		/*
5848 		 * An online sibling might have gone offline before a task
5849 		 * could be picked for it, or it might be offline but later
5850 		 * happen to come online, but its too late and nothing was
5851 		 * picked for it.  That's Ok - it will pick tasks for itself,
5852 		 * so ignore it.
5853 		 */
5854 		if (!rq_i->core_pick)
5855 			continue;
5856 
5857 		/*
5858 		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5859 		 * fi_before     fi      update?
5860 		 *  0            0       1
5861 		 *  0            1       1
5862 		 *  1            0       1
5863 		 *  1            1       0
5864 		 */
5865 		if (!(fi_before && rq->core->core_forceidle_count))
5866 			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
5867 
5868 		rq_i->core_pick->core_occupation = occ;
5869 
5870 		if (i == cpu) {
5871 			rq_i->core_pick = NULL;
5872 			continue;
5873 		}
5874 
5875 		/* Did we break L1TF mitigation requirements? */
5876 		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5877 
5878 		if (rq_i->curr == rq_i->core_pick) {
5879 			rq_i->core_pick = NULL;
5880 			continue;
5881 		}
5882 
5883 		resched_curr(rq_i);
5884 	}
5885 
5886 done:
5887 	set_next_task(rq, next);
5888 	return next;
5889 }
5890 
5891 static bool try_steal_cookie(int this, int that)
5892 {
5893 	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5894 	struct task_struct *p;
5895 	unsigned long cookie;
5896 	bool success = false;
5897 
5898 	local_irq_disable();
5899 	double_rq_lock(dst, src);
5900 
5901 	cookie = dst->core->core_cookie;
5902 	if (!cookie)
5903 		goto unlock;
5904 
5905 	if (dst->curr != dst->idle)
5906 		goto unlock;
5907 
5908 	p = sched_core_find(src, cookie);
5909 	if (p == src->idle)
5910 		goto unlock;
5911 
5912 	do {
5913 		if (p == src->core_pick || p == src->curr)
5914 			goto next;
5915 
5916 		if (!cpumask_test_cpu(this, &p->cpus_mask))
5917 			goto next;
5918 
5919 		if (p->core_occupation > dst->idle->core_occupation)
5920 			goto next;
5921 
5922 		deactivate_task(src, p, 0);
5923 		set_task_cpu(p, this);
5924 		activate_task(dst, p, 0);
5925 
5926 		resched_curr(dst);
5927 
5928 		success = true;
5929 		break;
5930 
5931 next:
5932 		p = sched_core_next(p, cookie);
5933 	} while (p);
5934 
5935 unlock:
5936 	double_rq_unlock(dst, src);
5937 	local_irq_enable();
5938 
5939 	return success;
5940 }
5941 
5942 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5943 {
5944 	int i;
5945 
5946 	for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5947 		if (i == cpu)
5948 			continue;
5949 
5950 		if (need_resched())
5951 			break;
5952 
5953 		if (try_steal_cookie(cpu, i))
5954 			return true;
5955 	}
5956 
5957 	return false;
5958 }
5959 
5960 static void sched_core_balance(struct rq *rq)
5961 {
5962 	struct sched_domain *sd;
5963 	int cpu = cpu_of(rq);
5964 
5965 	preempt_disable();
5966 	rcu_read_lock();
5967 	raw_spin_rq_unlock_irq(rq);
5968 	for_each_domain(cpu, sd) {
5969 		if (need_resched())
5970 			break;
5971 
5972 		if (steal_cookie_task(cpu, sd))
5973 			break;
5974 	}
5975 	raw_spin_rq_lock_irq(rq);
5976 	rcu_read_unlock();
5977 	preempt_enable();
5978 }
5979 
5980 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5981 
5982 void queue_core_balance(struct rq *rq)
5983 {
5984 	if (!sched_core_enabled(rq))
5985 		return;
5986 
5987 	if (!rq->core->core_cookie)
5988 		return;
5989 
5990 	if (!rq->nr_running) /* not forced idle */
5991 		return;
5992 
5993 	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5994 }
5995 
5996 static void sched_core_cpu_starting(unsigned int cpu)
5997 {
5998 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5999 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6000 	unsigned long flags;
6001 	int t;
6002 
6003 	sched_core_lock(cpu, &flags);
6004 
6005 	WARN_ON_ONCE(rq->core != rq);
6006 
6007 	/* if we're the first, we'll be our own leader */
6008 	if (cpumask_weight(smt_mask) == 1)
6009 		goto unlock;
6010 
6011 	/* find the leader */
6012 	for_each_cpu(t, smt_mask) {
6013 		if (t == cpu)
6014 			continue;
6015 		rq = cpu_rq(t);
6016 		if (rq->core == rq) {
6017 			core_rq = rq;
6018 			break;
6019 		}
6020 	}
6021 
6022 	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6023 		goto unlock;
6024 
6025 	/* install and validate core_rq */
6026 	for_each_cpu(t, smt_mask) {
6027 		rq = cpu_rq(t);
6028 
6029 		if (t == cpu)
6030 			rq->core = core_rq;
6031 
6032 		WARN_ON_ONCE(rq->core != core_rq);
6033 	}
6034 
6035 unlock:
6036 	sched_core_unlock(cpu, &flags);
6037 }
6038 
6039 static void sched_core_cpu_deactivate(unsigned int cpu)
6040 {
6041 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6042 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6043 	unsigned long flags;
6044 	int t;
6045 
6046 	sched_core_lock(cpu, &flags);
6047 
6048 	/* if we're the last man standing, nothing to do */
6049 	if (cpumask_weight(smt_mask) == 1) {
6050 		WARN_ON_ONCE(rq->core != rq);
6051 		goto unlock;
6052 	}
6053 
6054 	/* if we're not the leader, nothing to do */
6055 	if (rq->core != rq)
6056 		goto unlock;
6057 
6058 	/* find a new leader */
6059 	for_each_cpu(t, smt_mask) {
6060 		if (t == cpu)
6061 			continue;
6062 		core_rq = cpu_rq(t);
6063 		break;
6064 	}
6065 
6066 	if (WARN_ON_ONCE(!core_rq)) /* impossible */
6067 		goto unlock;
6068 
6069 	/* copy the shared state to the new leader */
6070 	core_rq->core_task_seq             = rq->core_task_seq;
6071 	core_rq->core_pick_seq             = rq->core_pick_seq;
6072 	core_rq->core_cookie               = rq->core_cookie;
6073 	core_rq->core_forceidle_count      = rq->core_forceidle_count;
6074 	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
6075 	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6076 
6077 	/*
6078 	 * Accounting edge for forced idle is handled in pick_next_task().
6079 	 * Don't need another one here, since the hotplug thread shouldn't
6080 	 * have a cookie.
6081 	 */
6082 	core_rq->core_forceidle_start = 0;
6083 
6084 	/* install new leader */
6085 	for_each_cpu(t, smt_mask) {
6086 		rq = cpu_rq(t);
6087 		rq->core = core_rq;
6088 	}
6089 
6090 unlock:
6091 	sched_core_unlock(cpu, &flags);
6092 }
6093 
6094 static inline void sched_core_cpu_dying(unsigned int cpu)
6095 {
6096 	struct rq *rq = cpu_rq(cpu);
6097 
6098 	if (rq->core != rq)
6099 		rq->core = rq;
6100 }
6101 
6102 #else /* !CONFIG_SCHED_CORE */
6103 
6104 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6105 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6106 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6107 
6108 static struct task_struct *
6109 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6110 {
6111 	return __pick_next_task(rq, prev, rf);
6112 }
6113 
6114 #endif /* CONFIG_SCHED_CORE */
6115 
6116 /*
6117  * Constants for the sched_mode argument of __schedule().
6118  *
6119  * The mode argument allows RT enabled kernels to differentiate a
6120  * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6121  * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6122  * optimize the AND operation out and just check for zero.
6123  */
6124 #define SM_NONE			0x0
6125 #define SM_PREEMPT		0x1
6126 #define SM_RTLOCK_WAIT		0x2
6127 
6128 #ifndef CONFIG_PREEMPT_RT
6129 # define SM_MASK_PREEMPT	(~0U)
6130 #else
6131 # define SM_MASK_PREEMPT	SM_PREEMPT
6132 #endif
6133 
6134 /*
6135  * __schedule() is the main scheduler function.
6136  *
6137  * The main means of driving the scheduler and thus entering this function are:
6138  *
6139  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6140  *
6141  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6142  *      paths. For example, see arch/x86/entry_64.S.
6143  *
6144  *      To drive preemption between tasks, the scheduler sets the flag in timer
6145  *      interrupt handler scheduler_tick().
6146  *
6147  *   3. Wakeups don't really cause entry into schedule(). They add a
6148  *      task to the run-queue and that's it.
6149  *
6150  *      Now, if the new task added to the run-queue preempts the current
6151  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6152  *      called on the nearest possible occasion:
6153  *
6154  *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6155  *
6156  *         - in syscall or exception context, at the next outmost
6157  *           preempt_enable(). (this might be as soon as the wake_up()'s
6158  *           spin_unlock()!)
6159  *
6160  *         - in IRQ context, return from interrupt-handler to
6161  *           preemptible context
6162  *
6163  *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6164  *         then at the next:
6165  *
6166  *          - cond_resched() call
6167  *          - explicit schedule() call
6168  *          - return from syscall or exception to user-space
6169  *          - return from interrupt-handler to user-space
6170  *
6171  * WARNING: must be called with preemption disabled!
6172  */
6173 static void __sched notrace __schedule(unsigned int sched_mode)
6174 {
6175 	struct task_struct *prev, *next;
6176 	unsigned long *switch_count;
6177 	unsigned long prev_state;
6178 	struct rq_flags rf;
6179 	struct rq *rq;
6180 	int cpu;
6181 
6182 	cpu = smp_processor_id();
6183 	rq = cpu_rq(cpu);
6184 	prev = rq->curr;
6185 
6186 	schedule_debug(prev, !!sched_mode);
6187 
6188 	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6189 		hrtick_clear(rq);
6190 
6191 	local_irq_disable();
6192 	rcu_note_context_switch(!!sched_mode);
6193 
6194 	/*
6195 	 * Make sure that signal_pending_state()->signal_pending() below
6196 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6197 	 * done by the caller to avoid the race with signal_wake_up():
6198 	 *
6199 	 * __set_current_state(@state)		signal_wake_up()
6200 	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
6201 	 *					  wake_up_state(p, state)
6202 	 *   LOCK rq->lock			    LOCK p->pi_state
6203 	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
6204 	 *     if (signal_pending_state())	    if (p->state & @state)
6205 	 *
6206 	 * Also, the membarrier system call requires a full memory barrier
6207 	 * after coming from user-space, before storing to rq->curr.
6208 	 */
6209 	rq_lock(rq, &rf);
6210 	smp_mb__after_spinlock();
6211 
6212 	/* Promote REQ to ACT */
6213 	rq->clock_update_flags <<= 1;
6214 	update_rq_clock(rq);
6215 
6216 	switch_count = &prev->nivcsw;
6217 
6218 	/*
6219 	 * We must load prev->state once (task_struct::state is volatile), such
6220 	 * that:
6221 	 *
6222 	 *  - we form a control dependency vs deactivate_task() below.
6223 	 *  - ptrace_{,un}freeze_traced() can change ->state underneath us.
6224 	 */
6225 	prev_state = READ_ONCE(prev->__state);
6226 	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6227 		if (signal_pending_state(prev_state, prev)) {
6228 			WRITE_ONCE(prev->__state, TASK_RUNNING);
6229 		} else {
6230 			prev->sched_contributes_to_load =
6231 				(prev_state & TASK_UNINTERRUPTIBLE) &&
6232 				!(prev_state & TASK_NOLOAD) &&
6233 				!(prev->flags & PF_FROZEN);
6234 
6235 			if (prev->sched_contributes_to_load)
6236 				rq->nr_uninterruptible++;
6237 
6238 			/*
6239 			 * __schedule()			ttwu()
6240 			 *   prev_state = prev->state;    if (p->on_rq && ...)
6241 			 *   if (prev_state)		    goto out;
6242 			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
6243 			 *				  p->state = TASK_WAKING
6244 			 *
6245 			 * Where __schedule() and ttwu() have matching control dependencies.
6246 			 *
6247 			 * After this, schedule() must not care about p->state any more.
6248 			 */
6249 			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6250 
6251 			if (prev->in_iowait) {
6252 				atomic_inc(&rq->nr_iowait);
6253 				delayacct_blkio_start();
6254 			}
6255 		}
6256 		switch_count = &prev->nvcsw;
6257 	}
6258 
6259 	next = pick_next_task(rq, prev, &rf);
6260 	clear_tsk_need_resched(prev);
6261 	clear_preempt_need_resched();
6262 #ifdef CONFIG_SCHED_DEBUG
6263 	rq->last_seen_need_resched_ns = 0;
6264 #endif
6265 
6266 	if (likely(prev != next)) {
6267 		rq->nr_switches++;
6268 		/*
6269 		 * RCU users of rcu_dereference(rq->curr) may not see
6270 		 * changes to task_struct made by pick_next_task().
6271 		 */
6272 		RCU_INIT_POINTER(rq->curr, next);
6273 		/*
6274 		 * The membarrier system call requires each architecture
6275 		 * to have a full memory barrier after updating
6276 		 * rq->curr, before returning to user-space.
6277 		 *
6278 		 * Here are the schemes providing that barrier on the
6279 		 * various architectures:
6280 		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6281 		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6282 		 * - finish_lock_switch() for weakly-ordered
6283 		 *   architectures where spin_unlock is a full barrier,
6284 		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6285 		 *   is a RELEASE barrier),
6286 		 */
6287 		++*switch_count;
6288 
6289 		migrate_disable_switch(rq, prev);
6290 		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6291 
6292 		trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next);
6293 
6294 		/* Also unlocks the rq: */
6295 		rq = context_switch(rq, prev, next, &rf);
6296 	} else {
6297 		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6298 
6299 		rq_unpin_lock(rq, &rf);
6300 		__balance_callbacks(rq);
6301 		raw_spin_rq_unlock_irq(rq);
6302 	}
6303 }
6304 
6305 void __noreturn do_task_dead(void)
6306 {
6307 	/* Causes final put_task_struct in finish_task_switch(): */
6308 	set_special_state(TASK_DEAD);
6309 
6310 	/* Tell freezer to ignore us: */
6311 	current->flags |= PF_NOFREEZE;
6312 
6313 	__schedule(SM_NONE);
6314 	BUG();
6315 
6316 	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6317 	for (;;)
6318 		cpu_relax();
6319 }
6320 
6321 static inline void sched_submit_work(struct task_struct *tsk)
6322 {
6323 	unsigned int task_flags;
6324 
6325 	if (task_is_running(tsk))
6326 		return;
6327 
6328 	task_flags = tsk->flags;
6329 	/*
6330 	 * If a worker goes to sleep, notify and ask workqueue whether it
6331 	 * wants to wake up a task to maintain concurrency.
6332 	 */
6333 	if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6334 		if (task_flags & PF_WQ_WORKER)
6335 			wq_worker_sleeping(tsk);
6336 		else
6337 			io_wq_worker_sleeping(tsk);
6338 	}
6339 
6340 	if (tsk_is_pi_blocked(tsk))
6341 		return;
6342 
6343 	/*
6344 	 * If we are going to sleep and we have plugged IO queued,
6345 	 * make sure to submit it to avoid deadlocks.
6346 	 */
6347 	if (blk_needs_flush_plug(tsk))
6348 		blk_flush_plug(tsk->plug, true);
6349 }
6350 
6351 static void sched_update_worker(struct task_struct *tsk)
6352 {
6353 	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6354 		if (tsk->flags & PF_WQ_WORKER)
6355 			wq_worker_running(tsk);
6356 		else
6357 			io_wq_worker_running(tsk);
6358 	}
6359 }
6360 
6361 asmlinkage __visible void __sched schedule(void)
6362 {
6363 	struct task_struct *tsk = current;
6364 
6365 	sched_submit_work(tsk);
6366 	do {
6367 		preempt_disable();
6368 		__schedule(SM_NONE);
6369 		sched_preempt_enable_no_resched();
6370 	} while (need_resched());
6371 	sched_update_worker(tsk);
6372 }
6373 EXPORT_SYMBOL(schedule);
6374 
6375 /*
6376  * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6377  * state (have scheduled out non-voluntarily) by making sure that all
6378  * tasks have either left the run queue or have gone into user space.
6379  * As idle tasks do not do either, they must not ever be preempted
6380  * (schedule out non-voluntarily).
6381  *
6382  * schedule_idle() is similar to schedule_preempt_disable() except that it
6383  * never enables preemption because it does not call sched_submit_work().
6384  */
6385 void __sched schedule_idle(void)
6386 {
6387 	/*
6388 	 * As this skips calling sched_submit_work(), which the idle task does
6389 	 * regardless because that function is a nop when the task is in a
6390 	 * TASK_RUNNING state, make sure this isn't used someplace that the
6391 	 * current task can be in any other state. Note, idle is always in the
6392 	 * TASK_RUNNING state.
6393 	 */
6394 	WARN_ON_ONCE(current->__state);
6395 	do {
6396 		__schedule(SM_NONE);
6397 	} while (need_resched());
6398 }
6399 
6400 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6401 asmlinkage __visible void __sched schedule_user(void)
6402 {
6403 	/*
6404 	 * If we come here after a random call to set_need_resched(),
6405 	 * or we have been woken up remotely but the IPI has not yet arrived,
6406 	 * we haven't yet exited the RCU idle mode. Do it here manually until
6407 	 * we find a better solution.
6408 	 *
6409 	 * NB: There are buggy callers of this function.  Ideally we
6410 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
6411 	 * too frequently to make sense yet.
6412 	 */
6413 	enum ctx_state prev_state = exception_enter();
6414 	schedule();
6415 	exception_exit(prev_state);
6416 }
6417 #endif
6418 
6419 /**
6420  * schedule_preempt_disabled - called with preemption disabled
6421  *
6422  * Returns with preemption disabled. Note: preempt_count must be 1
6423  */
6424 void __sched schedule_preempt_disabled(void)
6425 {
6426 	sched_preempt_enable_no_resched();
6427 	schedule();
6428 	preempt_disable();
6429 }
6430 
6431 #ifdef CONFIG_PREEMPT_RT
6432 void __sched notrace schedule_rtlock(void)
6433 {
6434 	do {
6435 		preempt_disable();
6436 		__schedule(SM_RTLOCK_WAIT);
6437 		sched_preempt_enable_no_resched();
6438 	} while (need_resched());
6439 }
6440 NOKPROBE_SYMBOL(schedule_rtlock);
6441 #endif
6442 
6443 static void __sched notrace preempt_schedule_common(void)
6444 {
6445 	do {
6446 		/*
6447 		 * Because the function tracer can trace preempt_count_sub()
6448 		 * and it also uses preempt_enable/disable_notrace(), if
6449 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6450 		 * by the function tracer will call this function again and
6451 		 * cause infinite recursion.
6452 		 *
6453 		 * Preemption must be disabled here before the function
6454 		 * tracer can trace. Break up preempt_disable() into two
6455 		 * calls. One to disable preemption without fear of being
6456 		 * traced. The other to still record the preemption latency,
6457 		 * which can also be traced by the function tracer.
6458 		 */
6459 		preempt_disable_notrace();
6460 		preempt_latency_start(1);
6461 		__schedule(SM_PREEMPT);
6462 		preempt_latency_stop(1);
6463 		preempt_enable_no_resched_notrace();
6464 
6465 		/*
6466 		 * Check again in case we missed a preemption opportunity
6467 		 * between schedule and now.
6468 		 */
6469 	} while (need_resched());
6470 }
6471 
6472 #ifdef CONFIG_PREEMPTION
6473 /*
6474  * This is the entry point to schedule() from in-kernel preemption
6475  * off of preempt_enable.
6476  */
6477 asmlinkage __visible void __sched notrace preempt_schedule(void)
6478 {
6479 	/*
6480 	 * If there is a non-zero preempt_count or interrupts are disabled,
6481 	 * we do not want to preempt the current task. Just return..
6482 	 */
6483 	if (likely(!preemptible()))
6484 		return;
6485 
6486 	preempt_schedule_common();
6487 }
6488 NOKPROBE_SYMBOL(preempt_schedule);
6489 EXPORT_SYMBOL(preempt_schedule);
6490 
6491 #ifdef CONFIG_PREEMPT_DYNAMIC
6492 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6493 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6494 #endif
6495 
6496 
6497 /**
6498  * preempt_schedule_notrace - preempt_schedule called by tracing
6499  *
6500  * The tracing infrastructure uses preempt_enable_notrace to prevent
6501  * recursion and tracing preempt enabling caused by the tracing
6502  * infrastructure itself. But as tracing can happen in areas coming
6503  * from userspace or just about to enter userspace, a preempt enable
6504  * can occur before user_exit() is called. This will cause the scheduler
6505  * to be called when the system is still in usermode.
6506  *
6507  * To prevent this, the preempt_enable_notrace will use this function
6508  * instead of preempt_schedule() to exit user context if needed before
6509  * calling the scheduler.
6510  */
6511 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6512 {
6513 	enum ctx_state prev_ctx;
6514 
6515 	if (likely(!preemptible()))
6516 		return;
6517 
6518 	do {
6519 		/*
6520 		 * Because the function tracer can trace preempt_count_sub()
6521 		 * and it also uses preempt_enable/disable_notrace(), if
6522 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6523 		 * by the function tracer will call this function again and
6524 		 * cause infinite recursion.
6525 		 *
6526 		 * Preemption must be disabled here before the function
6527 		 * tracer can trace. Break up preempt_disable() into two
6528 		 * calls. One to disable preemption without fear of being
6529 		 * traced. The other to still record the preemption latency,
6530 		 * which can also be traced by the function tracer.
6531 		 */
6532 		preempt_disable_notrace();
6533 		preempt_latency_start(1);
6534 		/*
6535 		 * Needs preempt disabled in case user_exit() is traced
6536 		 * and the tracer calls preempt_enable_notrace() causing
6537 		 * an infinite recursion.
6538 		 */
6539 		prev_ctx = exception_enter();
6540 		__schedule(SM_PREEMPT);
6541 		exception_exit(prev_ctx);
6542 
6543 		preempt_latency_stop(1);
6544 		preempt_enable_no_resched_notrace();
6545 	} while (need_resched());
6546 }
6547 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6548 
6549 #ifdef CONFIG_PREEMPT_DYNAMIC
6550 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6551 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6552 #endif
6553 
6554 #endif /* CONFIG_PREEMPTION */
6555 
6556 #ifdef CONFIG_PREEMPT_DYNAMIC
6557 
6558 #include <linux/entry-common.h>
6559 
6560 /*
6561  * SC:cond_resched
6562  * SC:might_resched
6563  * SC:preempt_schedule
6564  * SC:preempt_schedule_notrace
6565  * SC:irqentry_exit_cond_resched
6566  *
6567  *
6568  * NONE:
6569  *   cond_resched               <- __cond_resched
6570  *   might_resched              <- RET0
6571  *   preempt_schedule           <- NOP
6572  *   preempt_schedule_notrace   <- NOP
6573  *   irqentry_exit_cond_resched <- NOP
6574  *
6575  * VOLUNTARY:
6576  *   cond_resched               <- __cond_resched
6577  *   might_resched              <- __cond_resched
6578  *   preempt_schedule           <- NOP
6579  *   preempt_schedule_notrace   <- NOP
6580  *   irqentry_exit_cond_resched <- NOP
6581  *
6582  * FULL:
6583  *   cond_resched               <- RET0
6584  *   might_resched              <- RET0
6585  *   preempt_schedule           <- preempt_schedule
6586  *   preempt_schedule_notrace   <- preempt_schedule_notrace
6587  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6588  */
6589 
6590 enum {
6591 	preempt_dynamic_undefined = -1,
6592 	preempt_dynamic_none,
6593 	preempt_dynamic_voluntary,
6594 	preempt_dynamic_full,
6595 };
6596 
6597 int preempt_dynamic_mode = preempt_dynamic_undefined;
6598 
6599 int sched_dynamic_mode(const char *str)
6600 {
6601 	if (!strcmp(str, "none"))
6602 		return preempt_dynamic_none;
6603 
6604 	if (!strcmp(str, "voluntary"))
6605 		return preempt_dynamic_voluntary;
6606 
6607 	if (!strcmp(str, "full"))
6608 		return preempt_dynamic_full;
6609 
6610 	return -EINVAL;
6611 }
6612 
6613 void sched_dynamic_update(int mode)
6614 {
6615 	/*
6616 	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6617 	 * the ZERO state, which is invalid.
6618 	 */
6619 	static_call_update(cond_resched, __cond_resched);
6620 	static_call_update(might_resched, __cond_resched);
6621 	static_call_update(preempt_schedule, __preempt_schedule_func);
6622 	static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6623 	static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6624 
6625 	switch (mode) {
6626 	case preempt_dynamic_none:
6627 		static_call_update(cond_resched, __cond_resched);
6628 		static_call_update(might_resched, (void *)&__static_call_return0);
6629 		static_call_update(preempt_schedule, NULL);
6630 		static_call_update(preempt_schedule_notrace, NULL);
6631 		static_call_update(irqentry_exit_cond_resched, NULL);
6632 		pr_info("Dynamic Preempt: none\n");
6633 		break;
6634 
6635 	case preempt_dynamic_voluntary:
6636 		static_call_update(cond_resched, __cond_resched);
6637 		static_call_update(might_resched, __cond_resched);
6638 		static_call_update(preempt_schedule, NULL);
6639 		static_call_update(preempt_schedule_notrace, NULL);
6640 		static_call_update(irqentry_exit_cond_resched, NULL);
6641 		pr_info("Dynamic Preempt: voluntary\n");
6642 		break;
6643 
6644 	case preempt_dynamic_full:
6645 		static_call_update(cond_resched, (void *)&__static_call_return0);
6646 		static_call_update(might_resched, (void *)&__static_call_return0);
6647 		static_call_update(preempt_schedule, __preempt_schedule_func);
6648 		static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6649 		static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6650 		pr_info("Dynamic Preempt: full\n");
6651 		break;
6652 	}
6653 
6654 	preempt_dynamic_mode = mode;
6655 }
6656 
6657 static int __init setup_preempt_mode(char *str)
6658 {
6659 	int mode = sched_dynamic_mode(str);
6660 	if (mode < 0) {
6661 		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6662 		return 0;
6663 	}
6664 
6665 	sched_dynamic_update(mode);
6666 	return 1;
6667 }
6668 __setup("preempt=", setup_preempt_mode);
6669 
6670 static void __init preempt_dynamic_init(void)
6671 {
6672 	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
6673 		if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
6674 			sched_dynamic_update(preempt_dynamic_none);
6675 		} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
6676 			sched_dynamic_update(preempt_dynamic_voluntary);
6677 		} else {
6678 			/* Default static call setting, nothing to do */
6679 			WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
6680 			preempt_dynamic_mode = preempt_dynamic_full;
6681 			pr_info("Dynamic Preempt: full\n");
6682 		}
6683 	}
6684 }
6685 
6686 #else /* !CONFIG_PREEMPT_DYNAMIC */
6687 
6688 static inline void preempt_dynamic_init(void) { }
6689 
6690 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
6691 
6692 /*
6693  * This is the entry point to schedule() from kernel preemption
6694  * off of irq context.
6695  * Note, that this is called and return with irqs disabled. This will
6696  * protect us against recursive calling from irq.
6697  */
6698 asmlinkage __visible void __sched preempt_schedule_irq(void)
6699 {
6700 	enum ctx_state prev_state;
6701 
6702 	/* Catch callers which need to be fixed */
6703 	BUG_ON(preempt_count() || !irqs_disabled());
6704 
6705 	prev_state = exception_enter();
6706 
6707 	do {
6708 		preempt_disable();
6709 		local_irq_enable();
6710 		__schedule(SM_PREEMPT);
6711 		local_irq_disable();
6712 		sched_preempt_enable_no_resched();
6713 	} while (need_resched());
6714 
6715 	exception_exit(prev_state);
6716 }
6717 
6718 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6719 			  void *key)
6720 {
6721 	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6722 	return try_to_wake_up(curr->private, mode, wake_flags);
6723 }
6724 EXPORT_SYMBOL(default_wake_function);
6725 
6726 static void __setscheduler_prio(struct task_struct *p, int prio)
6727 {
6728 	if (dl_prio(prio))
6729 		p->sched_class = &dl_sched_class;
6730 	else if (rt_prio(prio))
6731 		p->sched_class = &rt_sched_class;
6732 	else
6733 		p->sched_class = &fair_sched_class;
6734 
6735 	p->prio = prio;
6736 }
6737 
6738 #ifdef CONFIG_RT_MUTEXES
6739 
6740 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6741 {
6742 	if (pi_task)
6743 		prio = min(prio, pi_task->prio);
6744 
6745 	return prio;
6746 }
6747 
6748 static inline int rt_effective_prio(struct task_struct *p, int prio)
6749 {
6750 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
6751 
6752 	return __rt_effective_prio(pi_task, prio);
6753 }
6754 
6755 /*
6756  * rt_mutex_setprio - set the current priority of a task
6757  * @p: task to boost
6758  * @pi_task: donor task
6759  *
6760  * This function changes the 'effective' priority of a task. It does
6761  * not touch ->normal_prio like __setscheduler().
6762  *
6763  * Used by the rt_mutex code to implement priority inheritance
6764  * logic. Call site only calls if the priority of the task changed.
6765  */
6766 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6767 {
6768 	int prio, oldprio, queued, running, queue_flag =
6769 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6770 	const struct sched_class *prev_class;
6771 	struct rq_flags rf;
6772 	struct rq *rq;
6773 
6774 	/* XXX used to be waiter->prio, not waiter->task->prio */
6775 	prio = __rt_effective_prio(pi_task, p->normal_prio);
6776 
6777 	/*
6778 	 * If nothing changed; bail early.
6779 	 */
6780 	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6781 		return;
6782 
6783 	rq = __task_rq_lock(p, &rf);
6784 	update_rq_clock(rq);
6785 	/*
6786 	 * Set under pi_lock && rq->lock, such that the value can be used under
6787 	 * either lock.
6788 	 *
6789 	 * Note that there is loads of tricky to make this pointer cache work
6790 	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6791 	 * ensure a task is de-boosted (pi_task is set to NULL) before the
6792 	 * task is allowed to run again (and can exit). This ensures the pointer
6793 	 * points to a blocked task -- which guarantees the task is present.
6794 	 */
6795 	p->pi_top_task = pi_task;
6796 
6797 	/*
6798 	 * For FIFO/RR we only need to set prio, if that matches we're done.
6799 	 */
6800 	if (prio == p->prio && !dl_prio(prio))
6801 		goto out_unlock;
6802 
6803 	/*
6804 	 * Idle task boosting is a nono in general. There is one
6805 	 * exception, when PREEMPT_RT and NOHZ is active:
6806 	 *
6807 	 * The idle task calls get_next_timer_interrupt() and holds
6808 	 * the timer wheel base->lock on the CPU and another CPU wants
6809 	 * to access the timer (probably to cancel it). We can safely
6810 	 * ignore the boosting request, as the idle CPU runs this code
6811 	 * with interrupts disabled and will complete the lock
6812 	 * protected section without being interrupted. So there is no
6813 	 * real need to boost.
6814 	 */
6815 	if (unlikely(p == rq->idle)) {
6816 		WARN_ON(p != rq->curr);
6817 		WARN_ON(p->pi_blocked_on);
6818 		goto out_unlock;
6819 	}
6820 
6821 	trace_sched_pi_setprio(p, pi_task);
6822 	oldprio = p->prio;
6823 
6824 	if (oldprio == prio)
6825 		queue_flag &= ~DEQUEUE_MOVE;
6826 
6827 	prev_class = p->sched_class;
6828 	queued = task_on_rq_queued(p);
6829 	running = task_current(rq, p);
6830 	if (queued)
6831 		dequeue_task(rq, p, queue_flag);
6832 	if (running)
6833 		put_prev_task(rq, p);
6834 
6835 	/*
6836 	 * Boosting condition are:
6837 	 * 1. -rt task is running and holds mutex A
6838 	 *      --> -dl task blocks on mutex A
6839 	 *
6840 	 * 2. -dl task is running and holds mutex A
6841 	 *      --> -dl task blocks on mutex A and could preempt the
6842 	 *          running task
6843 	 */
6844 	if (dl_prio(prio)) {
6845 		if (!dl_prio(p->normal_prio) ||
6846 		    (pi_task && dl_prio(pi_task->prio) &&
6847 		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
6848 			p->dl.pi_se = pi_task->dl.pi_se;
6849 			queue_flag |= ENQUEUE_REPLENISH;
6850 		} else {
6851 			p->dl.pi_se = &p->dl;
6852 		}
6853 	} else if (rt_prio(prio)) {
6854 		if (dl_prio(oldprio))
6855 			p->dl.pi_se = &p->dl;
6856 		if (oldprio < prio)
6857 			queue_flag |= ENQUEUE_HEAD;
6858 	} else {
6859 		if (dl_prio(oldprio))
6860 			p->dl.pi_se = &p->dl;
6861 		if (rt_prio(oldprio))
6862 			p->rt.timeout = 0;
6863 	}
6864 
6865 	__setscheduler_prio(p, prio);
6866 
6867 	if (queued)
6868 		enqueue_task(rq, p, queue_flag);
6869 	if (running)
6870 		set_next_task(rq, p);
6871 
6872 	check_class_changed(rq, p, prev_class, oldprio);
6873 out_unlock:
6874 	/* Avoid rq from going away on us: */
6875 	preempt_disable();
6876 
6877 	rq_unpin_lock(rq, &rf);
6878 	__balance_callbacks(rq);
6879 	raw_spin_rq_unlock(rq);
6880 
6881 	preempt_enable();
6882 }
6883 #else
6884 static inline int rt_effective_prio(struct task_struct *p, int prio)
6885 {
6886 	return prio;
6887 }
6888 #endif
6889 
6890 void set_user_nice(struct task_struct *p, long nice)
6891 {
6892 	bool queued, running;
6893 	int old_prio;
6894 	struct rq_flags rf;
6895 	struct rq *rq;
6896 
6897 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6898 		return;
6899 	/*
6900 	 * We have to be careful, if called from sys_setpriority(),
6901 	 * the task might be in the middle of scheduling on another CPU.
6902 	 */
6903 	rq = task_rq_lock(p, &rf);
6904 	update_rq_clock(rq);
6905 
6906 	/*
6907 	 * The RT priorities are set via sched_setscheduler(), but we still
6908 	 * allow the 'normal' nice value to be set - but as expected
6909 	 * it won't have any effect on scheduling until the task is
6910 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6911 	 */
6912 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6913 		p->static_prio = NICE_TO_PRIO(nice);
6914 		goto out_unlock;
6915 	}
6916 	queued = task_on_rq_queued(p);
6917 	running = task_current(rq, p);
6918 	if (queued)
6919 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6920 	if (running)
6921 		put_prev_task(rq, p);
6922 
6923 	p->static_prio = NICE_TO_PRIO(nice);
6924 	set_load_weight(p, true);
6925 	old_prio = p->prio;
6926 	p->prio = effective_prio(p);
6927 
6928 	if (queued)
6929 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6930 	if (running)
6931 		set_next_task(rq, p);
6932 
6933 	/*
6934 	 * If the task increased its priority or is running and
6935 	 * lowered its priority, then reschedule its CPU:
6936 	 */
6937 	p->sched_class->prio_changed(rq, p, old_prio);
6938 
6939 out_unlock:
6940 	task_rq_unlock(rq, p, &rf);
6941 }
6942 EXPORT_SYMBOL(set_user_nice);
6943 
6944 /*
6945  * can_nice - check if a task can reduce its nice value
6946  * @p: task
6947  * @nice: nice value
6948  */
6949 int can_nice(const struct task_struct *p, const int nice)
6950 {
6951 	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
6952 	int nice_rlim = nice_to_rlimit(nice);
6953 
6954 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6955 		capable(CAP_SYS_NICE));
6956 }
6957 
6958 #ifdef __ARCH_WANT_SYS_NICE
6959 
6960 /*
6961  * sys_nice - change the priority of the current process.
6962  * @increment: priority increment
6963  *
6964  * sys_setpriority is a more generic, but much slower function that
6965  * does similar things.
6966  */
6967 SYSCALL_DEFINE1(nice, int, increment)
6968 {
6969 	long nice, retval;
6970 
6971 	/*
6972 	 * Setpriority might change our priority at the same moment.
6973 	 * We don't have to worry. Conceptually one call occurs first
6974 	 * and we have a single winner.
6975 	 */
6976 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6977 	nice = task_nice(current) + increment;
6978 
6979 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6980 	if (increment < 0 && !can_nice(current, nice))
6981 		return -EPERM;
6982 
6983 	retval = security_task_setnice(current, nice);
6984 	if (retval)
6985 		return retval;
6986 
6987 	set_user_nice(current, nice);
6988 	return 0;
6989 }
6990 
6991 #endif
6992 
6993 /**
6994  * task_prio - return the priority value of a given task.
6995  * @p: the task in question.
6996  *
6997  * Return: The priority value as seen by users in /proc.
6998  *
6999  * sched policy         return value   kernel prio    user prio/nice
7000  *
7001  * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
7002  * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
7003  * deadline                     -101             -1           0
7004  */
7005 int task_prio(const struct task_struct *p)
7006 {
7007 	return p->prio - MAX_RT_PRIO;
7008 }
7009 
7010 /**
7011  * idle_cpu - is a given CPU idle currently?
7012  * @cpu: the processor in question.
7013  *
7014  * Return: 1 if the CPU is currently idle. 0 otherwise.
7015  */
7016 int idle_cpu(int cpu)
7017 {
7018 	struct rq *rq = cpu_rq(cpu);
7019 
7020 	if (rq->curr != rq->idle)
7021 		return 0;
7022 
7023 	if (rq->nr_running)
7024 		return 0;
7025 
7026 #ifdef CONFIG_SMP
7027 	if (rq->ttwu_pending)
7028 		return 0;
7029 #endif
7030 
7031 	return 1;
7032 }
7033 
7034 /**
7035  * available_idle_cpu - is a given CPU idle for enqueuing work.
7036  * @cpu: the CPU in question.
7037  *
7038  * Return: 1 if the CPU is currently idle. 0 otherwise.
7039  */
7040 int available_idle_cpu(int cpu)
7041 {
7042 	if (!idle_cpu(cpu))
7043 		return 0;
7044 
7045 	if (vcpu_is_preempted(cpu))
7046 		return 0;
7047 
7048 	return 1;
7049 }
7050 
7051 /**
7052  * idle_task - return the idle task for a given CPU.
7053  * @cpu: the processor in question.
7054  *
7055  * Return: The idle task for the CPU @cpu.
7056  */
7057 struct task_struct *idle_task(int cpu)
7058 {
7059 	return cpu_rq(cpu)->idle;
7060 }
7061 
7062 #ifdef CONFIG_SMP
7063 /*
7064  * This function computes an effective utilization for the given CPU, to be
7065  * used for frequency selection given the linear relation: f = u * f_max.
7066  *
7067  * The scheduler tracks the following metrics:
7068  *
7069  *   cpu_util_{cfs,rt,dl,irq}()
7070  *   cpu_bw_dl()
7071  *
7072  * Where the cfs,rt and dl util numbers are tracked with the same metric and
7073  * synchronized windows and are thus directly comparable.
7074  *
7075  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7076  * which excludes things like IRQ and steal-time. These latter are then accrued
7077  * in the irq utilization.
7078  *
7079  * The DL bandwidth number otoh is not a measured metric but a value computed
7080  * based on the task model parameters and gives the minimal utilization
7081  * required to meet deadlines.
7082  */
7083 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7084 				 unsigned long max, enum cpu_util_type type,
7085 				 struct task_struct *p)
7086 {
7087 	unsigned long dl_util, util, irq;
7088 	struct rq *rq = cpu_rq(cpu);
7089 
7090 	if (!uclamp_is_used() &&
7091 	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7092 		return max;
7093 	}
7094 
7095 	/*
7096 	 * Early check to see if IRQ/steal time saturates the CPU, can be
7097 	 * because of inaccuracies in how we track these -- see
7098 	 * update_irq_load_avg().
7099 	 */
7100 	irq = cpu_util_irq(rq);
7101 	if (unlikely(irq >= max))
7102 		return max;
7103 
7104 	/*
7105 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7106 	 * CFS tasks and we use the same metric to track the effective
7107 	 * utilization (PELT windows are synchronized) we can directly add them
7108 	 * to obtain the CPU's actual utilization.
7109 	 *
7110 	 * CFS and RT utilization can be boosted or capped, depending on
7111 	 * utilization clamp constraints requested by currently RUNNABLE
7112 	 * tasks.
7113 	 * When there are no CFS RUNNABLE tasks, clamps are released and
7114 	 * frequency will be gracefully reduced with the utilization decay.
7115 	 */
7116 	util = util_cfs + cpu_util_rt(rq);
7117 	if (type == FREQUENCY_UTIL)
7118 		util = uclamp_rq_util_with(rq, util, p);
7119 
7120 	dl_util = cpu_util_dl(rq);
7121 
7122 	/*
7123 	 * For frequency selection we do not make cpu_util_dl() a permanent part
7124 	 * of this sum because we want to use cpu_bw_dl() later on, but we need
7125 	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7126 	 * that we select f_max when there is no idle time.
7127 	 *
7128 	 * NOTE: numerical errors or stop class might cause us to not quite hit
7129 	 * saturation when we should -- something for later.
7130 	 */
7131 	if (util + dl_util >= max)
7132 		return max;
7133 
7134 	/*
7135 	 * OTOH, for energy computation we need the estimated running time, so
7136 	 * include util_dl and ignore dl_bw.
7137 	 */
7138 	if (type == ENERGY_UTIL)
7139 		util += dl_util;
7140 
7141 	/*
7142 	 * There is still idle time; further improve the number by using the
7143 	 * irq metric. Because IRQ/steal time is hidden from the task clock we
7144 	 * need to scale the task numbers:
7145 	 *
7146 	 *              max - irq
7147 	 *   U' = irq + --------- * U
7148 	 *                 max
7149 	 */
7150 	util = scale_irq_capacity(util, irq, max);
7151 	util += irq;
7152 
7153 	/*
7154 	 * Bandwidth required by DEADLINE must always be granted while, for
7155 	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7156 	 * to gracefully reduce the frequency when no tasks show up for longer
7157 	 * periods of time.
7158 	 *
7159 	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7160 	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7161 	 * an interface. So, we only do the latter for now.
7162 	 */
7163 	if (type == FREQUENCY_UTIL)
7164 		util += cpu_bw_dl(rq);
7165 
7166 	return min(max, util);
7167 }
7168 
7169 unsigned long sched_cpu_util(int cpu, unsigned long max)
7170 {
7171 	return effective_cpu_util(cpu, cpu_util_cfs(cpu), max,
7172 				  ENERGY_UTIL, NULL);
7173 }
7174 #endif /* CONFIG_SMP */
7175 
7176 /**
7177  * find_process_by_pid - find a process with a matching PID value.
7178  * @pid: the pid in question.
7179  *
7180  * The task of @pid, if found. %NULL otherwise.
7181  */
7182 static struct task_struct *find_process_by_pid(pid_t pid)
7183 {
7184 	return pid ? find_task_by_vpid(pid) : current;
7185 }
7186 
7187 /*
7188  * sched_setparam() passes in -1 for its policy, to let the functions
7189  * it calls know not to change it.
7190  */
7191 #define SETPARAM_POLICY	-1
7192 
7193 static void __setscheduler_params(struct task_struct *p,
7194 		const struct sched_attr *attr)
7195 {
7196 	int policy = attr->sched_policy;
7197 
7198 	if (policy == SETPARAM_POLICY)
7199 		policy = p->policy;
7200 
7201 	p->policy = policy;
7202 
7203 	if (dl_policy(policy))
7204 		__setparam_dl(p, attr);
7205 	else if (fair_policy(policy))
7206 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7207 
7208 	/*
7209 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7210 	 * !rt_policy. Always setting this ensures that things like
7211 	 * getparam()/getattr() don't report silly values for !rt tasks.
7212 	 */
7213 	p->rt_priority = attr->sched_priority;
7214 	p->normal_prio = normal_prio(p);
7215 	set_load_weight(p, true);
7216 }
7217 
7218 /*
7219  * Check the target process has a UID that matches the current process's:
7220  */
7221 static bool check_same_owner(struct task_struct *p)
7222 {
7223 	const struct cred *cred = current_cred(), *pcred;
7224 	bool match;
7225 
7226 	rcu_read_lock();
7227 	pcred = __task_cred(p);
7228 	match = (uid_eq(cred->euid, pcred->euid) ||
7229 		 uid_eq(cred->euid, pcred->uid));
7230 	rcu_read_unlock();
7231 	return match;
7232 }
7233 
7234 static int __sched_setscheduler(struct task_struct *p,
7235 				const struct sched_attr *attr,
7236 				bool user, bool pi)
7237 {
7238 	int oldpolicy = -1, policy = attr->sched_policy;
7239 	int retval, oldprio, newprio, queued, running;
7240 	const struct sched_class *prev_class;
7241 	struct callback_head *head;
7242 	struct rq_flags rf;
7243 	int reset_on_fork;
7244 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7245 	struct rq *rq;
7246 
7247 	/* The pi code expects interrupts enabled */
7248 	BUG_ON(pi && in_interrupt());
7249 recheck:
7250 	/* Double check policy once rq lock held: */
7251 	if (policy < 0) {
7252 		reset_on_fork = p->sched_reset_on_fork;
7253 		policy = oldpolicy = p->policy;
7254 	} else {
7255 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7256 
7257 		if (!valid_policy(policy))
7258 			return -EINVAL;
7259 	}
7260 
7261 	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7262 		return -EINVAL;
7263 
7264 	/*
7265 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
7266 	 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7267 	 * SCHED_BATCH and SCHED_IDLE is 0.
7268 	 */
7269 	if (attr->sched_priority > MAX_RT_PRIO-1)
7270 		return -EINVAL;
7271 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7272 	    (rt_policy(policy) != (attr->sched_priority != 0)))
7273 		return -EINVAL;
7274 
7275 	/*
7276 	 * Allow unprivileged RT tasks to decrease priority:
7277 	 */
7278 	if (user && !capable(CAP_SYS_NICE)) {
7279 		if (fair_policy(policy)) {
7280 			if (attr->sched_nice < task_nice(p) &&
7281 			    !can_nice(p, attr->sched_nice))
7282 				return -EPERM;
7283 		}
7284 
7285 		if (rt_policy(policy)) {
7286 			unsigned long rlim_rtprio =
7287 					task_rlimit(p, RLIMIT_RTPRIO);
7288 
7289 			/* Can't set/change the rt policy: */
7290 			if (policy != p->policy && !rlim_rtprio)
7291 				return -EPERM;
7292 
7293 			/* Can't increase priority: */
7294 			if (attr->sched_priority > p->rt_priority &&
7295 			    attr->sched_priority > rlim_rtprio)
7296 				return -EPERM;
7297 		}
7298 
7299 		 /*
7300 		  * Can't set/change SCHED_DEADLINE policy at all for now
7301 		  * (safest behavior); in the future we would like to allow
7302 		  * unprivileged DL tasks to increase their relative deadline
7303 		  * or reduce their runtime (both ways reducing utilization)
7304 		  */
7305 		if (dl_policy(policy))
7306 			return -EPERM;
7307 
7308 		/*
7309 		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7310 		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7311 		 */
7312 		if (task_has_idle_policy(p) && !idle_policy(policy)) {
7313 			if (!can_nice(p, task_nice(p)))
7314 				return -EPERM;
7315 		}
7316 
7317 		/* Can't change other user's priorities: */
7318 		if (!check_same_owner(p))
7319 			return -EPERM;
7320 
7321 		/* Normal users shall not reset the sched_reset_on_fork flag: */
7322 		if (p->sched_reset_on_fork && !reset_on_fork)
7323 			return -EPERM;
7324 	}
7325 
7326 	if (user) {
7327 		if (attr->sched_flags & SCHED_FLAG_SUGOV)
7328 			return -EINVAL;
7329 
7330 		retval = security_task_setscheduler(p);
7331 		if (retval)
7332 			return retval;
7333 	}
7334 
7335 	/* Update task specific "requested" clamps */
7336 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7337 		retval = uclamp_validate(p, attr);
7338 		if (retval)
7339 			return retval;
7340 	}
7341 
7342 	if (pi)
7343 		cpuset_read_lock();
7344 
7345 	/*
7346 	 * Make sure no PI-waiters arrive (or leave) while we are
7347 	 * changing the priority of the task:
7348 	 *
7349 	 * To be able to change p->policy safely, the appropriate
7350 	 * runqueue lock must be held.
7351 	 */
7352 	rq = task_rq_lock(p, &rf);
7353 	update_rq_clock(rq);
7354 
7355 	/*
7356 	 * Changing the policy of the stop threads its a very bad idea:
7357 	 */
7358 	if (p == rq->stop) {
7359 		retval = -EINVAL;
7360 		goto unlock;
7361 	}
7362 
7363 	/*
7364 	 * If not changing anything there's no need to proceed further,
7365 	 * but store a possible modification of reset_on_fork.
7366 	 */
7367 	if (unlikely(policy == p->policy)) {
7368 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7369 			goto change;
7370 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7371 			goto change;
7372 		if (dl_policy(policy) && dl_param_changed(p, attr))
7373 			goto change;
7374 		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7375 			goto change;
7376 
7377 		p->sched_reset_on_fork = reset_on_fork;
7378 		retval = 0;
7379 		goto unlock;
7380 	}
7381 change:
7382 
7383 	if (user) {
7384 #ifdef CONFIG_RT_GROUP_SCHED
7385 		/*
7386 		 * Do not allow realtime tasks into groups that have no runtime
7387 		 * assigned.
7388 		 */
7389 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
7390 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7391 				!task_group_is_autogroup(task_group(p))) {
7392 			retval = -EPERM;
7393 			goto unlock;
7394 		}
7395 #endif
7396 #ifdef CONFIG_SMP
7397 		if (dl_bandwidth_enabled() && dl_policy(policy) &&
7398 				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7399 			cpumask_t *span = rq->rd->span;
7400 
7401 			/*
7402 			 * Don't allow tasks with an affinity mask smaller than
7403 			 * the entire root_domain to become SCHED_DEADLINE. We
7404 			 * will also fail if there's no bandwidth available.
7405 			 */
7406 			if (!cpumask_subset(span, p->cpus_ptr) ||
7407 			    rq->rd->dl_bw.bw == 0) {
7408 				retval = -EPERM;
7409 				goto unlock;
7410 			}
7411 		}
7412 #endif
7413 	}
7414 
7415 	/* Re-check policy now with rq lock held: */
7416 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7417 		policy = oldpolicy = -1;
7418 		task_rq_unlock(rq, p, &rf);
7419 		if (pi)
7420 			cpuset_read_unlock();
7421 		goto recheck;
7422 	}
7423 
7424 	/*
7425 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7426 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7427 	 * is available.
7428 	 */
7429 	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7430 		retval = -EBUSY;
7431 		goto unlock;
7432 	}
7433 
7434 	p->sched_reset_on_fork = reset_on_fork;
7435 	oldprio = p->prio;
7436 
7437 	newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7438 	if (pi) {
7439 		/*
7440 		 * Take priority boosted tasks into account. If the new
7441 		 * effective priority is unchanged, we just store the new
7442 		 * normal parameters and do not touch the scheduler class and
7443 		 * the runqueue. This will be done when the task deboost
7444 		 * itself.
7445 		 */
7446 		newprio = rt_effective_prio(p, newprio);
7447 		if (newprio == oldprio)
7448 			queue_flags &= ~DEQUEUE_MOVE;
7449 	}
7450 
7451 	queued = task_on_rq_queued(p);
7452 	running = task_current(rq, p);
7453 	if (queued)
7454 		dequeue_task(rq, p, queue_flags);
7455 	if (running)
7456 		put_prev_task(rq, p);
7457 
7458 	prev_class = p->sched_class;
7459 
7460 	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7461 		__setscheduler_params(p, attr);
7462 		__setscheduler_prio(p, newprio);
7463 	}
7464 	__setscheduler_uclamp(p, attr);
7465 
7466 	if (queued) {
7467 		/*
7468 		 * We enqueue to tail when the priority of a task is
7469 		 * increased (user space view).
7470 		 */
7471 		if (oldprio < p->prio)
7472 			queue_flags |= ENQUEUE_HEAD;
7473 
7474 		enqueue_task(rq, p, queue_flags);
7475 	}
7476 	if (running)
7477 		set_next_task(rq, p);
7478 
7479 	check_class_changed(rq, p, prev_class, oldprio);
7480 
7481 	/* Avoid rq from going away on us: */
7482 	preempt_disable();
7483 	head = splice_balance_callbacks(rq);
7484 	task_rq_unlock(rq, p, &rf);
7485 
7486 	if (pi) {
7487 		cpuset_read_unlock();
7488 		rt_mutex_adjust_pi(p);
7489 	}
7490 
7491 	/* Run balance callbacks after we've adjusted the PI chain: */
7492 	balance_callbacks(rq, head);
7493 	preempt_enable();
7494 
7495 	return 0;
7496 
7497 unlock:
7498 	task_rq_unlock(rq, p, &rf);
7499 	if (pi)
7500 		cpuset_read_unlock();
7501 	return retval;
7502 }
7503 
7504 static int _sched_setscheduler(struct task_struct *p, int policy,
7505 			       const struct sched_param *param, bool check)
7506 {
7507 	struct sched_attr attr = {
7508 		.sched_policy   = policy,
7509 		.sched_priority = param->sched_priority,
7510 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
7511 	};
7512 
7513 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7514 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7515 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7516 		policy &= ~SCHED_RESET_ON_FORK;
7517 		attr.sched_policy = policy;
7518 	}
7519 
7520 	return __sched_setscheduler(p, &attr, check, true);
7521 }
7522 /**
7523  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7524  * @p: the task in question.
7525  * @policy: new policy.
7526  * @param: structure containing the new RT priority.
7527  *
7528  * Use sched_set_fifo(), read its comment.
7529  *
7530  * Return: 0 on success. An error code otherwise.
7531  *
7532  * NOTE that the task may be already dead.
7533  */
7534 int sched_setscheduler(struct task_struct *p, int policy,
7535 		       const struct sched_param *param)
7536 {
7537 	return _sched_setscheduler(p, policy, param, true);
7538 }
7539 
7540 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7541 {
7542 	return __sched_setscheduler(p, attr, true, true);
7543 }
7544 
7545 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7546 {
7547 	return __sched_setscheduler(p, attr, false, true);
7548 }
7549 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7550 
7551 /**
7552  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7553  * @p: the task in question.
7554  * @policy: new policy.
7555  * @param: structure containing the new RT priority.
7556  *
7557  * Just like sched_setscheduler, only don't bother checking if the
7558  * current context has permission.  For example, this is needed in
7559  * stop_machine(): we create temporary high priority worker threads,
7560  * but our caller might not have that capability.
7561  *
7562  * Return: 0 on success. An error code otherwise.
7563  */
7564 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7565 			       const struct sched_param *param)
7566 {
7567 	return _sched_setscheduler(p, policy, param, false);
7568 }
7569 
7570 /*
7571  * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7572  * incapable of resource management, which is the one thing an OS really should
7573  * be doing.
7574  *
7575  * This is of course the reason it is limited to privileged users only.
7576  *
7577  * Worse still; it is fundamentally impossible to compose static priority
7578  * workloads. You cannot take two correctly working static prio workloads
7579  * and smash them together and still expect them to work.
7580  *
7581  * For this reason 'all' FIFO tasks the kernel creates are basically at:
7582  *
7583  *   MAX_RT_PRIO / 2
7584  *
7585  * The administrator _MUST_ configure the system, the kernel simply doesn't
7586  * know enough information to make a sensible choice.
7587  */
7588 void sched_set_fifo(struct task_struct *p)
7589 {
7590 	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7591 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7592 }
7593 EXPORT_SYMBOL_GPL(sched_set_fifo);
7594 
7595 /*
7596  * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7597  */
7598 void sched_set_fifo_low(struct task_struct *p)
7599 {
7600 	struct sched_param sp = { .sched_priority = 1 };
7601 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7602 }
7603 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7604 
7605 void sched_set_normal(struct task_struct *p, int nice)
7606 {
7607 	struct sched_attr attr = {
7608 		.sched_policy = SCHED_NORMAL,
7609 		.sched_nice = nice,
7610 	};
7611 	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7612 }
7613 EXPORT_SYMBOL_GPL(sched_set_normal);
7614 
7615 static int
7616 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7617 {
7618 	struct sched_param lparam;
7619 	struct task_struct *p;
7620 	int retval;
7621 
7622 	if (!param || pid < 0)
7623 		return -EINVAL;
7624 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7625 		return -EFAULT;
7626 
7627 	rcu_read_lock();
7628 	retval = -ESRCH;
7629 	p = find_process_by_pid(pid);
7630 	if (likely(p))
7631 		get_task_struct(p);
7632 	rcu_read_unlock();
7633 
7634 	if (likely(p)) {
7635 		retval = sched_setscheduler(p, policy, &lparam);
7636 		put_task_struct(p);
7637 	}
7638 
7639 	return retval;
7640 }
7641 
7642 /*
7643  * Mimics kernel/events/core.c perf_copy_attr().
7644  */
7645 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7646 {
7647 	u32 size;
7648 	int ret;
7649 
7650 	/* Zero the full structure, so that a short copy will be nice: */
7651 	memset(attr, 0, sizeof(*attr));
7652 
7653 	ret = get_user(size, &uattr->size);
7654 	if (ret)
7655 		return ret;
7656 
7657 	/* ABI compatibility quirk: */
7658 	if (!size)
7659 		size = SCHED_ATTR_SIZE_VER0;
7660 	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7661 		goto err_size;
7662 
7663 	ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7664 	if (ret) {
7665 		if (ret == -E2BIG)
7666 			goto err_size;
7667 		return ret;
7668 	}
7669 
7670 	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7671 	    size < SCHED_ATTR_SIZE_VER1)
7672 		return -EINVAL;
7673 
7674 	/*
7675 	 * XXX: Do we want to be lenient like existing syscalls; or do we want
7676 	 * to be strict and return an error on out-of-bounds values?
7677 	 */
7678 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7679 
7680 	return 0;
7681 
7682 err_size:
7683 	put_user(sizeof(*attr), &uattr->size);
7684 	return -E2BIG;
7685 }
7686 
7687 static void get_params(struct task_struct *p, struct sched_attr *attr)
7688 {
7689 	if (task_has_dl_policy(p))
7690 		__getparam_dl(p, attr);
7691 	else if (task_has_rt_policy(p))
7692 		attr->sched_priority = p->rt_priority;
7693 	else
7694 		attr->sched_nice = task_nice(p);
7695 }
7696 
7697 /**
7698  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7699  * @pid: the pid in question.
7700  * @policy: new policy.
7701  * @param: structure containing the new RT priority.
7702  *
7703  * Return: 0 on success. An error code otherwise.
7704  */
7705 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7706 {
7707 	if (policy < 0)
7708 		return -EINVAL;
7709 
7710 	return do_sched_setscheduler(pid, policy, param);
7711 }
7712 
7713 /**
7714  * sys_sched_setparam - set/change the RT priority of a thread
7715  * @pid: the pid in question.
7716  * @param: structure containing the new RT priority.
7717  *
7718  * Return: 0 on success. An error code otherwise.
7719  */
7720 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7721 {
7722 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7723 }
7724 
7725 /**
7726  * sys_sched_setattr - same as above, but with extended sched_attr
7727  * @pid: the pid in question.
7728  * @uattr: structure containing the extended parameters.
7729  * @flags: for future extension.
7730  */
7731 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7732 			       unsigned int, flags)
7733 {
7734 	struct sched_attr attr;
7735 	struct task_struct *p;
7736 	int retval;
7737 
7738 	if (!uattr || pid < 0 || flags)
7739 		return -EINVAL;
7740 
7741 	retval = sched_copy_attr(uattr, &attr);
7742 	if (retval)
7743 		return retval;
7744 
7745 	if ((int)attr.sched_policy < 0)
7746 		return -EINVAL;
7747 	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7748 		attr.sched_policy = SETPARAM_POLICY;
7749 
7750 	rcu_read_lock();
7751 	retval = -ESRCH;
7752 	p = find_process_by_pid(pid);
7753 	if (likely(p))
7754 		get_task_struct(p);
7755 	rcu_read_unlock();
7756 
7757 	if (likely(p)) {
7758 		if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7759 			get_params(p, &attr);
7760 		retval = sched_setattr(p, &attr);
7761 		put_task_struct(p);
7762 	}
7763 
7764 	return retval;
7765 }
7766 
7767 /**
7768  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7769  * @pid: the pid in question.
7770  *
7771  * Return: On success, the policy of the thread. Otherwise, a negative error
7772  * code.
7773  */
7774 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7775 {
7776 	struct task_struct *p;
7777 	int retval;
7778 
7779 	if (pid < 0)
7780 		return -EINVAL;
7781 
7782 	retval = -ESRCH;
7783 	rcu_read_lock();
7784 	p = find_process_by_pid(pid);
7785 	if (p) {
7786 		retval = security_task_getscheduler(p);
7787 		if (!retval)
7788 			retval = p->policy
7789 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7790 	}
7791 	rcu_read_unlock();
7792 	return retval;
7793 }
7794 
7795 /**
7796  * sys_sched_getparam - get the RT priority of a thread
7797  * @pid: the pid in question.
7798  * @param: structure containing the RT priority.
7799  *
7800  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7801  * code.
7802  */
7803 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7804 {
7805 	struct sched_param lp = { .sched_priority = 0 };
7806 	struct task_struct *p;
7807 	int retval;
7808 
7809 	if (!param || pid < 0)
7810 		return -EINVAL;
7811 
7812 	rcu_read_lock();
7813 	p = find_process_by_pid(pid);
7814 	retval = -ESRCH;
7815 	if (!p)
7816 		goto out_unlock;
7817 
7818 	retval = security_task_getscheduler(p);
7819 	if (retval)
7820 		goto out_unlock;
7821 
7822 	if (task_has_rt_policy(p))
7823 		lp.sched_priority = p->rt_priority;
7824 	rcu_read_unlock();
7825 
7826 	/*
7827 	 * This one might sleep, we cannot do it with a spinlock held ...
7828 	 */
7829 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7830 
7831 	return retval;
7832 
7833 out_unlock:
7834 	rcu_read_unlock();
7835 	return retval;
7836 }
7837 
7838 /*
7839  * Copy the kernel size attribute structure (which might be larger
7840  * than what user-space knows about) to user-space.
7841  *
7842  * Note that all cases are valid: user-space buffer can be larger or
7843  * smaller than the kernel-space buffer. The usual case is that both
7844  * have the same size.
7845  */
7846 static int
7847 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7848 			struct sched_attr *kattr,
7849 			unsigned int usize)
7850 {
7851 	unsigned int ksize = sizeof(*kattr);
7852 
7853 	if (!access_ok(uattr, usize))
7854 		return -EFAULT;
7855 
7856 	/*
7857 	 * sched_getattr() ABI forwards and backwards compatibility:
7858 	 *
7859 	 * If usize == ksize then we just copy everything to user-space and all is good.
7860 	 *
7861 	 * If usize < ksize then we only copy as much as user-space has space for,
7862 	 * this keeps ABI compatibility as well. We skip the rest.
7863 	 *
7864 	 * If usize > ksize then user-space is using a newer version of the ABI,
7865 	 * which part the kernel doesn't know about. Just ignore it - tooling can
7866 	 * detect the kernel's knowledge of attributes from the attr->size value
7867 	 * which is set to ksize in this case.
7868 	 */
7869 	kattr->size = min(usize, ksize);
7870 
7871 	if (copy_to_user(uattr, kattr, kattr->size))
7872 		return -EFAULT;
7873 
7874 	return 0;
7875 }
7876 
7877 /**
7878  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7879  * @pid: the pid in question.
7880  * @uattr: structure containing the extended parameters.
7881  * @usize: sizeof(attr) for fwd/bwd comp.
7882  * @flags: for future extension.
7883  */
7884 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7885 		unsigned int, usize, unsigned int, flags)
7886 {
7887 	struct sched_attr kattr = { };
7888 	struct task_struct *p;
7889 	int retval;
7890 
7891 	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7892 	    usize < SCHED_ATTR_SIZE_VER0 || flags)
7893 		return -EINVAL;
7894 
7895 	rcu_read_lock();
7896 	p = find_process_by_pid(pid);
7897 	retval = -ESRCH;
7898 	if (!p)
7899 		goto out_unlock;
7900 
7901 	retval = security_task_getscheduler(p);
7902 	if (retval)
7903 		goto out_unlock;
7904 
7905 	kattr.sched_policy = p->policy;
7906 	if (p->sched_reset_on_fork)
7907 		kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7908 	get_params(p, &kattr);
7909 	kattr.sched_flags &= SCHED_FLAG_ALL;
7910 
7911 #ifdef CONFIG_UCLAMP_TASK
7912 	/*
7913 	 * This could race with another potential updater, but this is fine
7914 	 * because it'll correctly read the old or the new value. We don't need
7915 	 * to guarantee who wins the race as long as it doesn't return garbage.
7916 	 */
7917 	kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7918 	kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7919 #endif
7920 
7921 	rcu_read_unlock();
7922 
7923 	return sched_attr_copy_to_user(uattr, &kattr, usize);
7924 
7925 out_unlock:
7926 	rcu_read_unlock();
7927 	return retval;
7928 }
7929 
7930 #ifdef CONFIG_SMP
7931 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7932 {
7933 	int ret = 0;
7934 
7935 	/*
7936 	 * If the task isn't a deadline task or admission control is
7937 	 * disabled then we don't care about affinity changes.
7938 	 */
7939 	if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7940 		return 0;
7941 
7942 	/*
7943 	 * Since bandwidth control happens on root_domain basis,
7944 	 * if admission test is enabled, we only admit -deadline
7945 	 * tasks allowed to run on all the CPUs in the task's
7946 	 * root_domain.
7947 	 */
7948 	rcu_read_lock();
7949 	if (!cpumask_subset(task_rq(p)->rd->span, mask))
7950 		ret = -EBUSY;
7951 	rcu_read_unlock();
7952 	return ret;
7953 }
7954 #endif
7955 
7956 static int
7957 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7958 {
7959 	int retval;
7960 	cpumask_var_t cpus_allowed, new_mask;
7961 
7962 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7963 		return -ENOMEM;
7964 
7965 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7966 		retval = -ENOMEM;
7967 		goto out_free_cpus_allowed;
7968 	}
7969 
7970 	cpuset_cpus_allowed(p, cpus_allowed);
7971 	cpumask_and(new_mask, mask, cpus_allowed);
7972 
7973 	retval = dl_task_check_affinity(p, new_mask);
7974 	if (retval)
7975 		goto out_free_new_mask;
7976 again:
7977 	retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7978 	if (retval)
7979 		goto out_free_new_mask;
7980 
7981 	cpuset_cpus_allowed(p, cpus_allowed);
7982 	if (!cpumask_subset(new_mask, cpus_allowed)) {
7983 		/*
7984 		 * We must have raced with a concurrent cpuset update.
7985 		 * Just reset the cpumask to the cpuset's cpus_allowed.
7986 		 */
7987 		cpumask_copy(new_mask, cpus_allowed);
7988 		goto again;
7989 	}
7990 
7991 out_free_new_mask:
7992 	free_cpumask_var(new_mask);
7993 out_free_cpus_allowed:
7994 	free_cpumask_var(cpus_allowed);
7995 	return retval;
7996 }
7997 
7998 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7999 {
8000 	struct task_struct *p;
8001 	int retval;
8002 
8003 	rcu_read_lock();
8004 
8005 	p = find_process_by_pid(pid);
8006 	if (!p) {
8007 		rcu_read_unlock();
8008 		return -ESRCH;
8009 	}
8010 
8011 	/* Prevent p going away */
8012 	get_task_struct(p);
8013 	rcu_read_unlock();
8014 
8015 	if (p->flags & PF_NO_SETAFFINITY) {
8016 		retval = -EINVAL;
8017 		goto out_put_task;
8018 	}
8019 
8020 	if (!check_same_owner(p)) {
8021 		rcu_read_lock();
8022 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8023 			rcu_read_unlock();
8024 			retval = -EPERM;
8025 			goto out_put_task;
8026 		}
8027 		rcu_read_unlock();
8028 	}
8029 
8030 	retval = security_task_setscheduler(p);
8031 	if (retval)
8032 		goto out_put_task;
8033 
8034 	retval = __sched_setaffinity(p, in_mask);
8035 out_put_task:
8036 	put_task_struct(p);
8037 	return retval;
8038 }
8039 
8040 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8041 			     struct cpumask *new_mask)
8042 {
8043 	if (len < cpumask_size())
8044 		cpumask_clear(new_mask);
8045 	else if (len > cpumask_size())
8046 		len = cpumask_size();
8047 
8048 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8049 }
8050 
8051 /**
8052  * sys_sched_setaffinity - set the CPU affinity of a process
8053  * @pid: pid of the process
8054  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8055  * @user_mask_ptr: user-space pointer to the new CPU mask
8056  *
8057  * Return: 0 on success. An error code otherwise.
8058  */
8059 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8060 		unsigned long __user *, user_mask_ptr)
8061 {
8062 	cpumask_var_t new_mask;
8063 	int retval;
8064 
8065 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8066 		return -ENOMEM;
8067 
8068 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8069 	if (retval == 0)
8070 		retval = sched_setaffinity(pid, new_mask);
8071 	free_cpumask_var(new_mask);
8072 	return retval;
8073 }
8074 
8075 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8076 {
8077 	struct task_struct *p;
8078 	unsigned long flags;
8079 	int retval;
8080 
8081 	rcu_read_lock();
8082 
8083 	retval = -ESRCH;
8084 	p = find_process_by_pid(pid);
8085 	if (!p)
8086 		goto out_unlock;
8087 
8088 	retval = security_task_getscheduler(p);
8089 	if (retval)
8090 		goto out_unlock;
8091 
8092 	raw_spin_lock_irqsave(&p->pi_lock, flags);
8093 	cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8094 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8095 
8096 out_unlock:
8097 	rcu_read_unlock();
8098 
8099 	return retval;
8100 }
8101 
8102 /**
8103  * sys_sched_getaffinity - get the CPU affinity of a process
8104  * @pid: pid of the process
8105  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8106  * @user_mask_ptr: user-space pointer to hold the current CPU mask
8107  *
8108  * Return: size of CPU mask copied to user_mask_ptr on success. An
8109  * error code otherwise.
8110  */
8111 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8112 		unsigned long __user *, user_mask_ptr)
8113 {
8114 	int ret;
8115 	cpumask_var_t mask;
8116 
8117 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8118 		return -EINVAL;
8119 	if (len & (sizeof(unsigned long)-1))
8120 		return -EINVAL;
8121 
8122 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8123 		return -ENOMEM;
8124 
8125 	ret = sched_getaffinity(pid, mask);
8126 	if (ret == 0) {
8127 		unsigned int retlen = min(len, cpumask_size());
8128 
8129 		if (copy_to_user(user_mask_ptr, mask, retlen))
8130 			ret = -EFAULT;
8131 		else
8132 			ret = retlen;
8133 	}
8134 	free_cpumask_var(mask);
8135 
8136 	return ret;
8137 }
8138 
8139 static void do_sched_yield(void)
8140 {
8141 	struct rq_flags rf;
8142 	struct rq *rq;
8143 
8144 	rq = this_rq_lock_irq(&rf);
8145 
8146 	schedstat_inc(rq->yld_count);
8147 	current->sched_class->yield_task(rq);
8148 
8149 	preempt_disable();
8150 	rq_unlock_irq(rq, &rf);
8151 	sched_preempt_enable_no_resched();
8152 
8153 	schedule();
8154 }
8155 
8156 /**
8157  * sys_sched_yield - yield the current processor to other threads.
8158  *
8159  * This function yields the current CPU to other tasks. If there are no
8160  * other threads running on this CPU then this function will return.
8161  *
8162  * Return: 0.
8163  */
8164 SYSCALL_DEFINE0(sched_yield)
8165 {
8166 	do_sched_yield();
8167 	return 0;
8168 }
8169 
8170 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8171 int __sched __cond_resched(void)
8172 {
8173 	if (should_resched(0)) {
8174 		preempt_schedule_common();
8175 		return 1;
8176 	}
8177 	/*
8178 	 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8179 	 * whether the current CPU is in an RCU read-side critical section,
8180 	 * so the tick can report quiescent states even for CPUs looping
8181 	 * in kernel context.  In contrast, in non-preemptible kernels,
8182 	 * RCU readers leave no in-memory hints, which means that CPU-bound
8183 	 * processes executing in kernel context might never report an
8184 	 * RCU quiescent state.  Therefore, the following code causes
8185 	 * cond_resched() to report a quiescent state, but only when RCU
8186 	 * is in urgent need of one.
8187 	 */
8188 #ifndef CONFIG_PREEMPT_RCU
8189 	rcu_all_qs();
8190 #endif
8191 	return 0;
8192 }
8193 EXPORT_SYMBOL(__cond_resched);
8194 #endif
8195 
8196 #ifdef CONFIG_PREEMPT_DYNAMIC
8197 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8198 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8199 
8200 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8201 EXPORT_STATIC_CALL_TRAMP(might_resched);
8202 #endif
8203 
8204 /*
8205  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8206  * call schedule, and on return reacquire the lock.
8207  *
8208  * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8209  * operations here to prevent schedule() from being called twice (once via
8210  * spin_unlock(), once by hand).
8211  */
8212 int __cond_resched_lock(spinlock_t *lock)
8213 {
8214 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8215 	int ret = 0;
8216 
8217 	lockdep_assert_held(lock);
8218 
8219 	if (spin_needbreak(lock) || resched) {
8220 		spin_unlock(lock);
8221 		if (!_cond_resched())
8222 			cpu_relax();
8223 		ret = 1;
8224 		spin_lock(lock);
8225 	}
8226 	return ret;
8227 }
8228 EXPORT_SYMBOL(__cond_resched_lock);
8229 
8230 int __cond_resched_rwlock_read(rwlock_t *lock)
8231 {
8232 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8233 	int ret = 0;
8234 
8235 	lockdep_assert_held_read(lock);
8236 
8237 	if (rwlock_needbreak(lock) || resched) {
8238 		read_unlock(lock);
8239 		if (!_cond_resched())
8240 			cpu_relax();
8241 		ret = 1;
8242 		read_lock(lock);
8243 	}
8244 	return ret;
8245 }
8246 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8247 
8248 int __cond_resched_rwlock_write(rwlock_t *lock)
8249 {
8250 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8251 	int ret = 0;
8252 
8253 	lockdep_assert_held_write(lock);
8254 
8255 	if (rwlock_needbreak(lock) || resched) {
8256 		write_unlock(lock);
8257 		if (!_cond_resched())
8258 			cpu_relax();
8259 		ret = 1;
8260 		write_lock(lock);
8261 	}
8262 	return ret;
8263 }
8264 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8265 
8266 /**
8267  * yield - yield the current processor to other threads.
8268  *
8269  * Do not ever use this function, there's a 99% chance you're doing it wrong.
8270  *
8271  * The scheduler is at all times free to pick the calling task as the most
8272  * eligible task to run, if removing the yield() call from your code breaks
8273  * it, it's already broken.
8274  *
8275  * Typical broken usage is:
8276  *
8277  * while (!event)
8278  *	yield();
8279  *
8280  * where one assumes that yield() will let 'the other' process run that will
8281  * make event true. If the current task is a SCHED_FIFO task that will never
8282  * happen. Never use yield() as a progress guarantee!!
8283  *
8284  * If you want to use yield() to wait for something, use wait_event().
8285  * If you want to use yield() to be 'nice' for others, use cond_resched().
8286  * If you still want to use yield(), do not!
8287  */
8288 void __sched yield(void)
8289 {
8290 	set_current_state(TASK_RUNNING);
8291 	do_sched_yield();
8292 }
8293 EXPORT_SYMBOL(yield);
8294 
8295 /**
8296  * yield_to - yield the current processor to another thread in
8297  * your thread group, or accelerate that thread toward the
8298  * processor it's on.
8299  * @p: target task
8300  * @preempt: whether task preemption is allowed or not
8301  *
8302  * It's the caller's job to ensure that the target task struct
8303  * can't go away on us before we can do any checks.
8304  *
8305  * Return:
8306  *	true (>0) if we indeed boosted the target task.
8307  *	false (0) if we failed to boost the target.
8308  *	-ESRCH if there's no task to yield to.
8309  */
8310 int __sched yield_to(struct task_struct *p, bool preempt)
8311 {
8312 	struct task_struct *curr = current;
8313 	struct rq *rq, *p_rq;
8314 	unsigned long flags;
8315 	int yielded = 0;
8316 
8317 	local_irq_save(flags);
8318 	rq = this_rq();
8319 
8320 again:
8321 	p_rq = task_rq(p);
8322 	/*
8323 	 * If we're the only runnable task on the rq and target rq also
8324 	 * has only one task, there's absolutely no point in yielding.
8325 	 */
8326 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8327 		yielded = -ESRCH;
8328 		goto out_irq;
8329 	}
8330 
8331 	double_rq_lock(rq, p_rq);
8332 	if (task_rq(p) != p_rq) {
8333 		double_rq_unlock(rq, p_rq);
8334 		goto again;
8335 	}
8336 
8337 	if (!curr->sched_class->yield_to_task)
8338 		goto out_unlock;
8339 
8340 	if (curr->sched_class != p->sched_class)
8341 		goto out_unlock;
8342 
8343 	if (task_running(p_rq, p) || !task_is_running(p))
8344 		goto out_unlock;
8345 
8346 	yielded = curr->sched_class->yield_to_task(rq, p);
8347 	if (yielded) {
8348 		schedstat_inc(rq->yld_count);
8349 		/*
8350 		 * Make p's CPU reschedule; pick_next_entity takes care of
8351 		 * fairness.
8352 		 */
8353 		if (preempt && rq != p_rq)
8354 			resched_curr(p_rq);
8355 	}
8356 
8357 out_unlock:
8358 	double_rq_unlock(rq, p_rq);
8359 out_irq:
8360 	local_irq_restore(flags);
8361 
8362 	if (yielded > 0)
8363 		schedule();
8364 
8365 	return yielded;
8366 }
8367 EXPORT_SYMBOL_GPL(yield_to);
8368 
8369 int io_schedule_prepare(void)
8370 {
8371 	int old_iowait = current->in_iowait;
8372 
8373 	current->in_iowait = 1;
8374 	if (current->plug)
8375 		blk_flush_plug(current->plug, true);
8376 
8377 	return old_iowait;
8378 }
8379 
8380 void io_schedule_finish(int token)
8381 {
8382 	current->in_iowait = token;
8383 }
8384 
8385 /*
8386  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8387  * that process accounting knows that this is a task in IO wait state.
8388  */
8389 long __sched io_schedule_timeout(long timeout)
8390 {
8391 	int token;
8392 	long ret;
8393 
8394 	token = io_schedule_prepare();
8395 	ret = schedule_timeout(timeout);
8396 	io_schedule_finish(token);
8397 
8398 	return ret;
8399 }
8400 EXPORT_SYMBOL(io_schedule_timeout);
8401 
8402 void __sched io_schedule(void)
8403 {
8404 	int token;
8405 
8406 	token = io_schedule_prepare();
8407 	schedule();
8408 	io_schedule_finish(token);
8409 }
8410 EXPORT_SYMBOL(io_schedule);
8411 
8412 /**
8413  * sys_sched_get_priority_max - return maximum RT priority.
8414  * @policy: scheduling class.
8415  *
8416  * Return: On success, this syscall returns the maximum
8417  * rt_priority that can be used by a given scheduling class.
8418  * On failure, a negative error code is returned.
8419  */
8420 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8421 {
8422 	int ret = -EINVAL;
8423 
8424 	switch (policy) {
8425 	case SCHED_FIFO:
8426 	case SCHED_RR:
8427 		ret = MAX_RT_PRIO-1;
8428 		break;
8429 	case SCHED_DEADLINE:
8430 	case SCHED_NORMAL:
8431 	case SCHED_BATCH:
8432 	case SCHED_IDLE:
8433 		ret = 0;
8434 		break;
8435 	}
8436 	return ret;
8437 }
8438 
8439 /**
8440  * sys_sched_get_priority_min - return minimum RT priority.
8441  * @policy: scheduling class.
8442  *
8443  * Return: On success, this syscall returns the minimum
8444  * rt_priority that can be used by a given scheduling class.
8445  * On failure, a negative error code is returned.
8446  */
8447 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8448 {
8449 	int ret = -EINVAL;
8450 
8451 	switch (policy) {
8452 	case SCHED_FIFO:
8453 	case SCHED_RR:
8454 		ret = 1;
8455 		break;
8456 	case SCHED_DEADLINE:
8457 	case SCHED_NORMAL:
8458 	case SCHED_BATCH:
8459 	case SCHED_IDLE:
8460 		ret = 0;
8461 	}
8462 	return ret;
8463 }
8464 
8465 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8466 {
8467 	struct task_struct *p;
8468 	unsigned int time_slice;
8469 	struct rq_flags rf;
8470 	struct rq *rq;
8471 	int retval;
8472 
8473 	if (pid < 0)
8474 		return -EINVAL;
8475 
8476 	retval = -ESRCH;
8477 	rcu_read_lock();
8478 	p = find_process_by_pid(pid);
8479 	if (!p)
8480 		goto out_unlock;
8481 
8482 	retval = security_task_getscheduler(p);
8483 	if (retval)
8484 		goto out_unlock;
8485 
8486 	rq = task_rq_lock(p, &rf);
8487 	time_slice = 0;
8488 	if (p->sched_class->get_rr_interval)
8489 		time_slice = p->sched_class->get_rr_interval(rq, p);
8490 	task_rq_unlock(rq, p, &rf);
8491 
8492 	rcu_read_unlock();
8493 	jiffies_to_timespec64(time_slice, t);
8494 	return 0;
8495 
8496 out_unlock:
8497 	rcu_read_unlock();
8498 	return retval;
8499 }
8500 
8501 /**
8502  * sys_sched_rr_get_interval - return the default timeslice of a process.
8503  * @pid: pid of the process.
8504  * @interval: userspace pointer to the timeslice value.
8505  *
8506  * this syscall writes the default timeslice value of a given process
8507  * into the user-space timespec buffer. A value of '0' means infinity.
8508  *
8509  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8510  * an error code.
8511  */
8512 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8513 		struct __kernel_timespec __user *, interval)
8514 {
8515 	struct timespec64 t;
8516 	int retval = sched_rr_get_interval(pid, &t);
8517 
8518 	if (retval == 0)
8519 		retval = put_timespec64(&t, interval);
8520 
8521 	return retval;
8522 }
8523 
8524 #ifdef CONFIG_COMPAT_32BIT_TIME
8525 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8526 		struct old_timespec32 __user *, interval)
8527 {
8528 	struct timespec64 t;
8529 	int retval = sched_rr_get_interval(pid, &t);
8530 
8531 	if (retval == 0)
8532 		retval = put_old_timespec32(&t, interval);
8533 	return retval;
8534 }
8535 #endif
8536 
8537 void sched_show_task(struct task_struct *p)
8538 {
8539 	unsigned long free = 0;
8540 	int ppid;
8541 
8542 	if (!try_get_task_stack(p))
8543 		return;
8544 
8545 	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8546 
8547 	if (task_is_running(p))
8548 		pr_cont("  running task    ");
8549 #ifdef CONFIG_DEBUG_STACK_USAGE
8550 	free = stack_not_used(p);
8551 #endif
8552 	ppid = 0;
8553 	rcu_read_lock();
8554 	if (pid_alive(p))
8555 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
8556 	rcu_read_unlock();
8557 	pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8558 		free, task_pid_nr(p), ppid,
8559 		read_task_thread_flags(p));
8560 
8561 	print_worker_info(KERN_INFO, p);
8562 	print_stop_info(KERN_INFO, p);
8563 	show_stack(p, NULL, KERN_INFO);
8564 	put_task_stack(p);
8565 }
8566 EXPORT_SYMBOL_GPL(sched_show_task);
8567 
8568 static inline bool
8569 state_filter_match(unsigned long state_filter, struct task_struct *p)
8570 {
8571 	unsigned int state = READ_ONCE(p->__state);
8572 
8573 	/* no filter, everything matches */
8574 	if (!state_filter)
8575 		return true;
8576 
8577 	/* filter, but doesn't match */
8578 	if (!(state & state_filter))
8579 		return false;
8580 
8581 	/*
8582 	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8583 	 * TASK_KILLABLE).
8584 	 */
8585 	if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8586 		return false;
8587 
8588 	return true;
8589 }
8590 
8591 
8592 void show_state_filter(unsigned int state_filter)
8593 {
8594 	struct task_struct *g, *p;
8595 
8596 	rcu_read_lock();
8597 	for_each_process_thread(g, p) {
8598 		/*
8599 		 * reset the NMI-timeout, listing all files on a slow
8600 		 * console might take a lot of time:
8601 		 * Also, reset softlockup watchdogs on all CPUs, because
8602 		 * another CPU might be blocked waiting for us to process
8603 		 * an IPI.
8604 		 */
8605 		touch_nmi_watchdog();
8606 		touch_all_softlockup_watchdogs();
8607 		if (state_filter_match(state_filter, p))
8608 			sched_show_task(p);
8609 	}
8610 
8611 #ifdef CONFIG_SCHED_DEBUG
8612 	if (!state_filter)
8613 		sysrq_sched_debug_show();
8614 #endif
8615 	rcu_read_unlock();
8616 	/*
8617 	 * Only show locks if all tasks are dumped:
8618 	 */
8619 	if (!state_filter)
8620 		debug_show_all_locks();
8621 }
8622 
8623 /**
8624  * init_idle - set up an idle thread for a given CPU
8625  * @idle: task in question
8626  * @cpu: CPU the idle task belongs to
8627  *
8628  * NOTE: this function does not set the idle thread's NEED_RESCHED
8629  * flag, to make booting more robust.
8630  */
8631 void __init init_idle(struct task_struct *idle, int cpu)
8632 {
8633 	struct rq *rq = cpu_rq(cpu);
8634 	unsigned long flags;
8635 
8636 	__sched_fork(0, idle);
8637 
8638 	raw_spin_lock_irqsave(&idle->pi_lock, flags);
8639 	raw_spin_rq_lock(rq);
8640 
8641 	idle->__state = TASK_RUNNING;
8642 	idle->se.exec_start = sched_clock();
8643 	/*
8644 	 * PF_KTHREAD should already be set at this point; regardless, make it
8645 	 * look like a proper per-CPU kthread.
8646 	 */
8647 	idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8648 	kthread_set_per_cpu(idle, cpu);
8649 
8650 #ifdef CONFIG_SMP
8651 	/*
8652 	 * It's possible that init_idle() gets called multiple times on a task,
8653 	 * in that case do_set_cpus_allowed() will not do the right thing.
8654 	 *
8655 	 * And since this is boot we can forgo the serialization.
8656 	 */
8657 	set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8658 #endif
8659 	/*
8660 	 * We're having a chicken and egg problem, even though we are
8661 	 * holding rq->lock, the CPU isn't yet set to this CPU so the
8662 	 * lockdep check in task_group() will fail.
8663 	 *
8664 	 * Similar case to sched_fork(). / Alternatively we could
8665 	 * use task_rq_lock() here and obtain the other rq->lock.
8666 	 *
8667 	 * Silence PROVE_RCU
8668 	 */
8669 	rcu_read_lock();
8670 	__set_task_cpu(idle, cpu);
8671 	rcu_read_unlock();
8672 
8673 	rq->idle = idle;
8674 	rcu_assign_pointer(rq->curr, idle);
8675 	idle->on_rq = TASK_ON_RQ_QUEUED;
8676 #ifdef CONFIG_SMP
8677 	idle->on_cpu = 1;
8678 #endif
8679 	raw_spin_rq_unlock(rq);
8680 	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8681 
8682 	/* Set the preempt count _outside_ the spinlocks! */
8683 	init_idle_preempt_count(idle, cpu);
8684 
8685 	/*
8686 	 * The idle tasks have their own, simple scheduling class:
8687 	 */
8688 	idle->sched_class = &idle_sched_class;
8689 	ftrace_graph_init_idle_task(idle, cpu);
8690 	vtime_init_idle(idle, cpu);
8691 #ifdef CONFIG_SMP
8692 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8693 #endif
8694 }
8695 
8696 #ifdef CONFIG_SMP
8697 
8698 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8699 			      const struct cpumask *trial)
8700 {
8701 	int ret = 1;
8702 
8703 	if (!cpumask_weight(cur))
8704 		return ret;
8705 
8706 	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8707 
8708 	return ret;
8709 }
8710 
8711 int task_can_attach(struct task_struct *p,
8712 		    const struct cpumask *cs_cpus_allowed)
8713 {
8714 	int ret = 0;
8715 
8716 	/*
8717 	 * Kthreads which disallow setaffinity shouldn't be moved
8718 	 * to a new cpuset; we don't want to change their CPU
8719 	 * affinity and isolating such threads by their set of
8720 	 * allowed nodes is unnecessary.  Thus, cpusets are not
8721 	 * applicable for such threads.  This prevents checking for
8722 	 * success of set_cpus_allowed_ptr() on all attached tasks
8723 	 * before cpus_mask may be changed.
8724 	 */
8725 	if (p->flags & PF_NO_SETAFFINITY) {
8726 		ret = -EINVAL;
8727 		goto out;
8728 	}
8729 
8730 	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8731 					      cs_cpus_allowed))
8732 		ret = dl_task_can_attach(p, cs_cpus_allowed);
8733 
8734 out:
8735 	return ret;
8736 }
8737 
8738 bool sched_smp_initialized __read_mostly;
8739 
8740 #ifdef CONFIG_NUMA_BALANCING
8741 /* Migrate current task p to target_cpu */
8742 int migrate_task_to(struct task_struct *p, int target_cpu)
8743 {
8744 	struct migration_arg arg = { p, target_cpu };
8745 	int curr_cpu = task_cpu(p);
8746 
8747 	if (curr_cpu == target_cpu)
8748 		return 0;
8749 
8750 	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8751 		return -EINVAL;
8752 
8753 	/* TODO: This is not properly updating schedstats */
8754 
8755 	trace_sched_move_numa(p, curr_cpu, target_cpu);
8756 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8757 }
8758 
8759 /*
8760  * Requeue a task on a given node and accurately track the number of NUMA
8761  * tasks on the runqueues
8762  */
8763 void sched_setnuma(struct task_struct *p, int nid)
8764 {
8765 	bool queued, running;
8766 	struct rq_flags rf;
8767 	struct rq *rq;
8768 
8769 	rq = task_rq_lock(p, &rf);
8770 	queued = task_on_rq_queued(p);
8771 	running = task_current(rq, p);
8772 
8773 	if (queued)
8774 		dequeue_task(rq, p, DEQUEUE_SAVE);
8775 	if (running)
8776 		put_prev_task(rq, p);
8777 
8778 	p->numa_preferred_nid = nid;
8779 
8780 	if (queued)
8781 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8782 	if (running)
8783 		set_next_task(rq, p);
8784 	task_rq_unlock(rq, p, &rf);
8785 }
8786 #endif /* CONFIG_NUMA_BALANCING */
8787 
8788 #ifdef CONFIG_HOTPLUG_CPU
8789 /*
8790  * Ensure that the idle task is using init_mm right before its CPU goes
8791  * offline.
8792  */
8793 void idle_task_exit(void)
8794 {
8795 	struct mm_struct *mm = current->active_mm;
8796 
8797 	BUG_ON(cpu_online(smp_processor_id()));
8798 	BUG_ON(current != this_rq()->idle);
8799 
8800 	if (mm != &init_mm) {
8801 		switch_mm(mm, &init_mm, current);
8802 		finish_arch_post_lock_switch();
8803 	}
8804 
8805 	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8806 }
8807 
8808 static int __balance_push_cpu_stop(void *arg)
8809 {
8810 	struct task_struct *p = arg;
8811 	struct rq *rq = this_rq();
8812 	struct rq_flags rf;
8813 	int cpu;
8814 
8815 	raw_spin_lock_irq(&p->pi_lock);
8816 	rq_lock(rq, &rf);
8817 
8818 	update_rq_clock(rq);
8819 
8820 	if (task_rq(p) == rq && task_on_rq_queued(p)) {
8821 		cpu = select_fallback_rq(rq->cpu, p);
8822 		rq = __migrate_task(rq, &rf, p, cpu);
8823 	}
8824 
8825 	rq_unlock(rq, &rf);
8826 	raw_spin_unlock_irq(&p->pi_lock);
8827 
8828 	put_task_struct(p);
8829 
8830 	return 0;
8831 }
8832 
8833 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8834 
8835 /*
8836  * Ensure we only run per-cpu kthreads once the CPU goes !active.
8837  *
8838  * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8839  * effective when the hotplug motion is down.
8840  */
8841 static void balance_push(struct rq *rq)
8842 {
8843 	struct task_struct *push_task = rq->curr;
8844 
8845 	lockdep_assert_rq_held(rq);
8846 
8847 	/*
8848 	 * Ensure the thing is persistent until balance_push_set(.on = false);
8849 	 */
8850 	rq->balance_callback = &balance_push_callback;
8851 
8852 	/*
8853 	 * Only active while going offline and when invoked on the outgoing
8854 	 * CPU.
8855 	 */
8856 	if (!cpu_dying(rq->cpu) || rq != this_rq())
8857 		return;
8858 
8859 	/*
8860 	 * Both the cpu-hotplug and stop task are in this case and are
8861 	 * required to complete the hotplug process.
8862 	 */
8863 	if (kthread_is_per_cpu(push_task) ||
8864 	    is_migration_disabled(push_task)) {
8865 
8866 		/*
8867 		 * If this is the idle task on the outgoing CPU try to wake
8868 		 * up the hotplug control thread which might wait for the
8869 		 * last task to vanish. The rcuwait_active() check is
8870 		 * accurate here because the waiter is pinned on this CPU
8871 		 * and can't obviously be running in parallel.
8872 		 *
8873 		 * On RT kernels this also has to check whether there are
8874 		 * pinned and scheduled out tasks on the runqueue. They
8875 		 * need to leave the migrate disabled section first.
8876 		 */
8877 		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8878 		    rcuwait_active(&rq->hotplug_wait)) {
8879 			raw_spin_rq_unlock(rq);
8880 			rcuwait_wake_up(&rq->hotplug_wait);
8881 			raw_spin_rq_lock(rq);
8882 		}
8883 		return;
8884 	}
8885 
8886 	get_task_struct(push_task);
8887 	/*
8888 	 * Temporarily drop rq->lock such that we can wake-up the stop task.
8889 	 * Both preemption and IRQs are still disabled.
8890 	 */
8891 	raw_spin_rq_unlock(rq);
8892 	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8893 			    this_cpu_ptr(&push_work));
8894 	/*
8895 	 * At this point need_resched() is true and we'll take the loop in
8896 	 * schedule(). The next pick is obviously going to be the stop task
8897 	 * which kthread_is_per_cpu() and will push this task away.
8898 	 */
8899 	raw_spin_rq_lock(rq);
8900 }
8901 
8902 static void balance_push_set(int cpu, bool on)
8903 {
8904 	struct rq *rq = cpu_rq(cpu);
8905 	struct rq_flags rf;
8906 
8907 	rq_lock_irqsave(rq, &rf);
8908 	if (on) {
8909 		WARN_ON_ONCE(rq->balance_callback);
8910 		rq->balance_callback = &balance_push_callback;
8911 	} else if (rq->balance_callback == &balance_push_callback) {
8912 		rq->balance_callback = NULL;
8913 	}
8914 	rq_unlock_irqrestore(rq, &rf);
8915 }
8916 
8917 /*
8918  * Invoked from a CPUs hotplug control thread after the CPU has been marked
8919  * inactive. All tasks which are not per CPU kernel threads are either
8920  * pushed off this CPU now via balance_push() or placed on a different CPU
8921  * during wakeup. Wait until the CPU is quiescent.
8922  */
8923 static void balance_hotplug_wait(void)
8924 {
8925 	struct rq *rq = this_rq();
8926 
8927 	rcuwait_wait_event(&rq->hotplug_wait,
8928 			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8929 			   TASK_UNINTERRUPTIBLE);
8930 }
8931 
8932 #else
8933 
8934 static inline void balance_push(struct rq *rq)
8935 {
8936 }
8937 
8938 static inline void balance_push_set(int cpu, bool on)
8939 {
8940 }
8941 
8942 static inline void balance_hotplug_wait(void)
8943 {
8944 }
8945 
8946 #endif /* CONFIG_HOTPLUG_CPU */
8947 
8948 void set_rq_online(struct rq *rq)
8949 {
8950 	if (!rq->online) {
8951 		const struct sched_class *class;
8952 
8953 		cpumask_set_cpu(rq->cpu, rq->rd->online);
8954 		rq->online = 1;
8955 
8956 		for_each_class(class) {
8957 			if (class->rq_online)
8958 				class->rq_online(rq);
8959 		}
8960 	}
8961 }
8962 
8963 void set_rq_offline(struct rq *rq)
8964 {
8965 	if (rq->online) {
8966 		const struct sched_class *class;
8967 
8968 		for_each_class(class) {
8969 			if (class->rq_offline)
8970 				class->rq_offline(rq);
8971 		}
8972 
8973 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
8974 		rq->online = 0;
8975 	}
8976 }
8977 
8978 /*
8979  * used to mark begin/end of suspend/resume:
8980  */
8981 static int num_cpus_frozen;
8982 
8983 /*
8984  * Update cpusets according to cpu_active mask.  If cpusets are
8985  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8986  * around partition_sched_domains().
8987  *
8988  * If we come here as part of a suspend/resume, don't touch cpusets because we
8989  * want to restore it back to its original state upon resume anyway.
8990  */
8991 static void cpuset_cpu_active(void)
8992 {
8993 	if (cpuhp_tasks_frozen) {
8994 		/*
8995 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
8996 		 * resume sequence. As long as this is not the last online
8997 		 * operation in the resume sequence, just build a single sched
8998 		 * domain, ignoring cpusets.
8999 		 */
9000 		partition_sched_domains(1, NULL, NULL);
9001 		if (--num_cpus_frozen)
9002 			return;
9003 		/*
9004 		 * This is the last CPU online operation. So fall through and
9005 		 * restore the original sched domains by considering the
9006 		 * cpuset configurations.
9007 		 */
9008 		cpuset_force_rebuild();
9009 	}
9010 	cpuset_update_active_cpus();
9011 }
9012 
9013 static int cpuset_cpu_inactive(unsigned int cpu)
9014 {
9015 	if (!cpuhp_tasks_frozen) {
9016 		if (dl_cpu_busy(cpu))
9017 			return -EBUSY;
9018 		cpuset_update_active_cpus();
9019 	} else {
9020 		num_cpus_frozen++;
9021 		partition_sched_domains(1, NULL, NULL);
9022 	}
9023 	return 0;
9024 }
9025 
9026 int sched_cpu_activate(unsigned int cpu)
9027 {
9028 	struct rq *rq = cpu_rq(cpu);
9029 	struct rq_flags rf;
9030 
9031 	/*
9032 	 * Clear the balance_push callback and prepare to schedule
9033 	 * regular tasks.
9034 	 */
9035 	balance_push_set(cpu, false);
9036 
9037 #ifdef CONFIG_SCHED_SMT
9038 	/*
9039 	 * When going up, increment the number of cores with SMT present.
9040 	 */
9041 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9042 		static_branch_inc_cpuslocked(&sched_smt_present);
9043 #endif
9044 	set_cpu_active(cpu, true);
9045 
9046 	if (sched_smp_initialized) {
9047 		sched_domains_numa_masks_set(cpu);
9048 		cpuset_cpu_active();
9049 	}
9050 
9051 	/*
9052 	 * Put the rq online, if not already. This happens:
9053 	 *
9054 	 * 1) In the early boot process, because we build the real domains
9055 	 *    after all CPUs have been brought up.
9056 	 *
9057 	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9058 	 *    domains.
9059 	 */
9060 	rq_lock_irqsave(rq, &rf);
9061 	if (rq->rd) {
9062 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9063 		set_rq_online(rq);
9064 	}
9065 	rq_unlock_irqrestore(rq, &rf);
9066 
9067 	return 0;
9068 }
9069 
9070 int sched_cpu_deactivate(unsigned int cpu)
9071 {
9072 	struct rq *rq = cpu_rq(cpu);
9073 	struct rq_flags rf;
9074 	int ret;
9075 
9076 	/*
9077 	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9078 	 * load balancing when not active
9079 	 */
9080 	nohz_balance_exit_idle(rq);
9081 
9082 	set_cpu_active(cpu, false);
9083 
9084 	/*
9085 	 * From this point forward, this CPU will refuse to run any task that
9086 	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9087 	 * push those tasks away until this gets cleared, see
9088 	 * sched_cpu_dying().
9089 	 */
9090 	balance_push_set(cpu, true);
9091 
9092 	/*
9093 	 * We've cleared cpu_active_mask / set balance_push, wait for all
9094 	 * preempt-disabled and RCU users of this state to go away such that
9095 	 * all new such users will observe it.
9096 	 *
9097 	 * Specifically, we rely on ttwu to no longer target this CPU, see
9098 	 * ttwu_queue_cond() and is_cpu_allowed().
9099 	 *
9100 	 * Do sync before park smpboot threads to take care the rcu boost case.
9101 	 */
9102 	synchronize_rcu();
9103 
9104 	rq_lock_irqsave(rq, &rf);
9105 	if (rq->rd) {
9106 		update_rq_clock(rq);
9107 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9108 		set_rq_offline(rq);
9109 	}
9110 	rq_unlock_irqrestore(rq, &rf);
9111 
9112 #ifdef CONFIG_SCHED_SMT
9113 	/*
9114 	 * When going down, decrement the number of cores with SMT present.
9115 	 */
9116 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9117 		static_branch_dec_cpuslocked(&sched_smt_present);
9118 
9119 	sched_core_cpu_deactivate(cpu);
9120 #endif
9121 
9122 	if (!sched_smp_initialized)
9123 		return 0;
9124 
9125 	ret = cpuset_cpu_inactive(cpu);
9126 	if (ret) {
9127 		balance_push_set(cpu, false);
9128 		set_cpu_active(cpu, true);
9129 		return ret;
9130 	}
9131 	sched_domains_numa_masks_clear(cpu);
9132 	return 0;
9133 }
9134 
9135 static void sched_rq_cpu_starting(unsigned int cpu)
9136 {
9137 	struct rq *rq = cpu_rq(cpu);
9138 
9139 	rq->calc_load_update = calc_load_update;
9140 	update_max_interval();
9141 }
9142 
9143 int sched_cpu_starting(unsigned int cpu)
9144 {
9145 	sched_core_cpu_starting(cpu);
9146 	sched_rq_cpu_starting(cpu);
9147 	sched_tick_start(cpu);
9148 	return 0;
9149 }
9150 
9151 #ifdef CONFIG_HOTPLUG_CPU
9152 
9153 /*
9154  * Invoked immediately before the stopper thread is invoked to bring the
9155  * CPU down completely. At this point all per CPU kthreads except the
9156  * hotplug thread (current) and the stopper thread (inactive) have been
9157  * either parked or have been unbound from the outgoing CPU. Ensure that
9158  * any of those which might be on the way out are gone.
9159  *
9160  * If after this point a bound task is being woken on this CPU then the
9161  * responsible hotplug callback has failed to do it's job.
9162  * sched_cpu_dying() will catch it with the appropriate fireworks.
9163  */
9164 int sched_cpu_wait_empty(unsigned int cpu)
9165 {
9166 	balance_hotplug_wait();
9167 	return 0;
9168 }
9169 
9170 /*
9171  * Since this CPU is going 'away' for a while, fold any nr_active delta we
9172  * might have. Called from the CPU stopper task after ensuring that the
9173  * stopper is the last running task on the CPU, so nr_active count is
9174  * stable. We need to take the teardown thread which is calling this into
9175  * account, so we hand in adjust = 1 to the load calculation.
9176  *
9177  * Also see the comment "Global load-average calculations".
9178  */
9179 static void calc_load_migrate(struct rq *rq)
9180 {
9181 	long delta = calc_load_fold_active(rq, 1);
9182 
9183 	if (delta)
9184 		atomic_long_add(delta, &calc_load_tasks);
9185 }
9186 
9187 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9188 {
9189 	struct task_struct *g, *p;
9190 	int cpu = cpu_of(rq);
9191 
9192 	lockdep_assert_rq_held(rq);
9193 
9194 	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9195 	for_each_process_thread(g, p) {
9196 		if (task_cpu(p) != cpu)
9197 			continue;
9198 
9199 		if (!task_on_rq_queued(p))
9200 			continue;
9201 
9202 		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9203 	}
9204 }
9205 
9206 int sched_cpu_dying(unsigned int cpu)
9207 {
9208 	struct rq *rq = cpu_rq(cpu);
9209 	struct rq_flags rf;
9210 
9211 	/* Handle pending wakeups and then migrate everything off */
9212 	sched_tick_stop(cpu);
9213 
9214 	rq_lock_irqsave(rq, &rf);
9215 	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9216 		WARN(true, "Dying CPU not properly vacated!");
9217 		dump_rq_tasks(rq, KERN_WARNING);
9218 	}
9219 	rq_unlock_irqrestore(rq, &rf);
9220 
9221 	calc_load_migrate(rq);
9222 	update_max_interval();
9223 	hrtick_clear(rq);
9224 	sched_core_cpu_dying(cpu);
9225 	return 0;
9226 }
9227 #endif
9228 
9229 void __init sched_init_smp(void)
9230 {
9231 	sched_init_numa();
9232 
9233 	/*
9234 	 * There's no userspace yet to cause hotplug operations; hence all the
9235 	 * CPU masks are stable and all blatant races in the below code cannot
9236 	 * happen.
9237 	 */
9238 	mutex_lock(&sched_domains_mutex);
9239 	sched_init_domains(cpu_active_mask);
9240 	mutex_unlock(&sched_domains_mutex);
9241 
9242 	/* Move init over to a non-isolated CPU */
9243 	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9244 		BUG();
9245 	current->flags &= ~PF_NO_SETAFFINITY;
9246 	sched_init_granularity();
9247 
9248 	init_sched_rt_class();
9249 	init_sched_dl_class();
9250 
9251 	sched_smp_initialized = true;
9252 }
9253 
9254 static int __init migration_init(void)
9255 {
9256 	sched_cpu_starting(smp_processor_id());
9257 	return 0;
9258 }
9259 early_initcall(migration_init);
9260 
9261 #else
9262 void __init sched_init_smp(void)
9263 {
9264 	sched_init_granularity();
9265 }
9266 #endif /* CONFIG_SMP */
9267 
9268 int in_sched_functions(unsigned long addr)
9269 {
9270 	return in_lock_functions(addr) ||
9271 		(addr >= (unsigned long)__sched_text_start
9272 		&& addr < (unsigned long)__sched_text_end);
9273 }
9274 
9275 #ifdef CONFIG_CGROUP_SCHED
9276 /*
9277  * Default task group.
9278  * Every task in system belongs to this group at bootup.
9279  */
9280 struct task_group root_task_group;
9281 LIST_HEAD(task_groups);
9282 
9283 /* Cacheline aligned slab cache for task_group */
9284 static struct kmem_cache *task_group_cache __read_mostly;
9285 #endif
9286 
9287 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9288 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9289 
9290 void __init sched_init(void)
9291 {
9292 	unsigned long ptr = 0;
9293 	int i;
9294 
9295 	/* Make sure the linker didn't screw up */
9296 	BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9297 	       &fair_sched_class + 1 != &rt_sched_class ||
9298 	       &rt_sched_class + 1   != &dl_sched_class);
9299 #ifdef CONFIG_SMP
9300 	BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9301 #endif
9302 
9303 	wait_bit_init();
9304 
9305 #ifdef CONFIG_FAIR_GROUP_SCHED
9306 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9307 #endif
9308 #ifdef CONFIG_RT_GROUP_SCHED
9309 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9310 #endif
9311 	if (ptr) {
9312 		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9313 
9314 #ifdef CONFIG_FAIR_GROUP_SCHED
9315 		root_task_group.se = (struct sched_entity **)ptr;
9316 		ptr += nr_cpu_ids * sizeof(void **);
9317 
9318 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9319 		ptr += nr_cpu_ids * sizeof(void **);
9320 
9321 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9322 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9323 #endif /* CONFIG_FAIR_GROUP_SCHED */
9324 #ifdef CONFIG_RT_GROUP_SCHED
9325 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9326 		ptr += nr_cpu_ids * sizeof(void **);
9327 
9328 		root_task_group.rt_rq = (struct rt_rq **)ptr;
9329 		ptr += nr_cpu_ids * sizeof(void **);
9330 
9331 #endif /* CONFIG_RT_GROUP_SCHED */
9332 	}
9333 #ifdef CONFIG_CPUMASK_OFFSTACK
9334 	for_each_possible_cpu(i) {
9335 		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9336 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9337 		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9338 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9339 	}
9340 #endif /* CONFIG_CPUMASK_OFFSTACK */
9341 
9342 	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9343 	init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9344 
9345 #ifdef CONFIG_SMP
9346 	init_defrootdomain();
9347 #endif
9348 
9349 #ifdef CONFIG_RT_GROUP_SCHED
9350 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
9351 			global_rt_period(), global_rt_runtime());
9352 #endif /* CONFIG_RT_GROUP_SCHED */
9353 
9354 #ifdef CONFIG_CGROUP_SCHED
9355 	task_group_cache = KMEM_CACHE(task_group, 0);
9356 
9357 	list_add(&root_task_group.list, &task_groups);
9358 	INIT_LIST_HEAD(&root_task_group.children);
9359 	INIT_LIST_HEAD(&root_task_group.siblings);
9360 	autogroup_init(&init_task);
9361 #endif /* CONFIG_CGROUP_SCHED */
9362 
9363 	for_each_possible_cpu(i) {
9364 		struct rq *rq;
9365 
9366 		rq = cpu_rq(i);
9367 		raw_spin_lock_init(&rq->__lock);
9368 		rq->nr_running = 0;
9369 		rq->calc_load_active = 0;
9370 		rq->calc_load_update = jiffies + LOAD_FREQ;
9371 		init_cfs_rq(&rq->cfs);
9372 		init_rt_rq(&rq->rt);
9373 		init_dl_rq(&rq->dl);
9374 #ifdef CONFIG_FAIR_GROUP_SCHED
9375 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9376 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9377 		/*
9378 		 * How much CPU bandwidth does root_task_group get?
9379 		 *
9380 		 * In case of task-groups formed thr' the cgroup filesystem, it
9381 		 * gets 100% of the CPU resources in the system. This overall
9382 		 * system CPU resource is divided among the tasks of
9383 		 * root_task_group and its child task-groups in a fair manner,
9384 		 * based on each entity's (task or task-group's) weight
9385 		 * (se->load.weight).
9386 		 *
9387 		 * In other words, if root_task_group has 10 tasks of weight
9388 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9389 		 * then A0's share of the CPU resource is:
9390 		 *
9391 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9392 		 *
9393 		 * We achieve this by letting root_task_group's tasks sit
9394 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9395 		 */
9396 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9397 #endif /* CONFIG_FAIR_GROUP_SCHED */
9398 
9399 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9400 #ifdef CONFIG_RT_GROUP_SCHED
9401 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9402 #endif
9403 #ifdef CONFIG_SMP
9404 		rq->sd = NULL;
9405 		rq->rd = NULL;
9406 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9407 		rq->balance_callback = &balance_push_callback;
9408 		rq->active_balance = 0;
9409 		rq->next_balance = jiffies;
9410 		rq->push_cpu = 0;
9411 		rq->cpu = i;
9412 		rq->online = 0;
9413 		rq->idle_stamp = 0;
9414 		rq->avg_idle = 2*sysctl_sched_migration_cost;
9415 		rq->wake_stamp = jiffies;
9416 		rq->wake_avg_idle = rq->avg_idle;
9417 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9418 
9419 		INIT_LIST_HEAD(&rq->cfs_tasks);
9420 
9421 		rq_attach_root(rq, &def_root_domain);
9422 #ifdef CONFIG_NO_HZ_COMMON
9423 		rq->last_blocked_load_update_tick = jiffies;
9424 		atomic_set(&rq->nohz_flags, 0);
9425 
9426 		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9427 #endif
9428 #ifdef CONFIG_HOTPLUG_CPU
9429 		rcuwait_init(&rq->hotplug_wait);
9430 #endif
9431 #endif /* CONFIG_SMP */
9432 		hrtick_rq_init(rq);
9433 		atomic_set(&rq->nr_iowait, 0);
9434 
9435 #ifdef CONFIG_SCHED_CORE
9436 		rq->core = rq;
9437 		rq->core_pick = NULL;
9438 		rq->core_enabled = 0;
9439 		rq->core_tree = RB_ROOT;
9440 		rq->core_forceidle_count = 0;
9441 		rq->core_forceidle_occupation = 0;
9442 		rq->core_forceidle_start = 0;
9443 
9444 		rq->core_cookie = 0UL;
9445 #endif
9446 	}
9447 
9448 	set_load_weight(&init_task, false);
9449 
9450 	/*
9451 	 * The boot idle thread does lazy MMU switching as well:
9452 	 */
9453 	mmgrab(&init_mm);
9454 	enter_lazy_tlb(&init_mm, current);
9455 
9456 	/*
9457 	 * The idle task doesn't need the kthread struct to function, but it
9458 	 * is dressed up as a per-CPU kthread and thus needs to play the part
9459 	 * if we want to avoid special-casing it in code that deals with per-CPU
9460 	 * kthreads.
9461 	 */
9462 	WARN_ON(!set_kthread_struct(current));
9463 
9464 	/*
9465 	 * Make us the idle thread. Technically, schedule() should not be
9466 	 * called from this thread, however somewhere below it might be,
9467 	 * but because we are the idle thread, we just pick up running again
9468 	 * when this runqueue becomes "idle".
9469 	 */
9470 	init_idle(current, smp_processor_id());
9471 
9472 	calc_load_update = jiffies + LOAD_FREQ;
9473 
9474 #ifdef CONFIG_SMP
9475 	idle_thread_set_boot_cpu();
9476 	balance_push_set(smp_processor_id(), false);
9477 #endif
9478 	init_sched_fair_class();
9479 
9480 	psi_init();
9481 
9482 	init_uclamp();
9483 
9484 	preempt_dynamic_init();
9485 
9486 	scheduler_running = 1;
9487 }
9488 
9489 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9490 
9491 void __might_sleep(const char *file, int line)
9492 {
9493 	unsigned int state = get_current_state();
9494 	/*
9495 	 * Blocking primitives will set (and therefore destroy) current->state,
9496 	 * since we will exit with TASK_RUNNING make sure we enter with it,
9497 	 * otherwise we will destroy state.
9498 	 */
9499 	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9500 			"do not call blocking ops when !TASK_RUNNING; "
9501 			"state=%x set at [<%p>] %pS\n", state,
9502 			(void *)current->task_state_change,
9503 			(void *)current->task_state_change);
9504 
9505 	__might_resched(file, line, 0);
9506 }
9507 EXPORT_SYMBOL(__might_sleep);
9508 
9509 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9510 {
9511 	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9512 		return;
9513 
9514 	if (preempt_count() == preempt_offset)
9515 		return;
9516 
9517 	pr_err("Preemption disabled at:");
9518 	print_ip_sym(KERN_ERR, ip);
9519 }
9520 
9521 static inline bool resched_offsets_ok(unsigned int offsets)
9522 {
9523 	unsigned int nested = preempt_count();
9524 
9525 	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9526 
9527 	return nested == offsets;
9528 }
9529 
9530 void __might_resched(const char *file, int line, unsigned int offsets)
9531 {
9532 	/* Ratelimiting timestamp: */
9533 	static unsigned long prev_jiffy;
9534 
9535 	unsigned long preempt_disable_ip;
9536 
9537 	/* WARN_ON_ONCE() by default, no rate limit required: */
9538 	rcu_sleep_check();
9539 
9540 	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9541 	     !is_idle_task(current) && !current->non_block_count) ||
9542 	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9543 	    oops_in_progress)
9544 		return;
9545 
9546 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9547 		return;
9548 	prev_jiffy = jiffies;
9549 
9550 	/* Save this before calling printk(), since that will clobber it: */
9551 	preempt_disable_ip = get_preempt_disable_ip(current);
9552 
9553 	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9554 	       file, line);
9555 	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9556 	       in_atomic(), irqs_disabled(), current->non_block_count,
9557 	       current->pid, current->comm);
9558 	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9559 	       offsets & MIGHT_RESCHED_PREEMPT_MASK);
9560 
9561 	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9562 		pr_err("RCU nest depth: %d, expected: %u\n",
9563 		       rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9564 	}
9565 
9566 	if (task_stack_end_corrupted(current))
9567 		pr_emerg("Thread overran stack, or stack corrupted\n");
9568 
9569 	debug_show_held_locks(current);
9570 	if (irqs_disabled())
9571 		print_irqtrace_events(current);
9572 
9573 	print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9574 				 preempt_disable_ip);
9575 
9576 	dump_stack();
9577 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9578 }
9579 EXPORT_SYMBOL(__might_resched);
9580 
9581 void __cant_sleep(const char *file, int line, int preempt_offset)
9582 {
9583 	static unsigned long prev_jiffy;
9584 
9585 	if (irqs_disabled())
9586 		return;
9587 
9588 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9589 		return;
9590 
9591 	if (preempt_count() > preempt_offset)
9592 		return;
9593 
9594 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9595 		return;
9596 	prev_jiffy = jiffies;
9597 
9598 	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9599 	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9600 			in_atomic(), irqs_disabled(),
9601 			current->pid, current->comm);
9602 
9603 	debug_show_held_locks(current);
9604 	dump_stack();
9605 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9606 }
9607 EXPORT_SYMBOL_GPL(__cant_sleep);
9608 
9609 #ifdef CONFIG_SMP
9610 void __cant_migrate(const char *file, int line)
9611 {
9612 	static unsigned long prev_jiffy;
9613 
9614 	if (irqs_disabled())
9615 		return;
9616 
9617 	if (is_migration_disabled(current))
9618 		return;
9619 
9620 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9621 		return;
9622 
9623 	if (preempt_count() > 0)
9624 		return;
9625 
9626 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9627 		return;
9628 	prev_jiffy = jiffies;
9629 
9630 	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9631 	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9632 	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
9633 	       current->pid, current->comm);
9634 
9635 	debug_show_held_locks(current);
9636 	dump_stack();
9637 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9638 }
9639 EXPORT_SYMBOL_GPL(__cant_migrate);
9640 #endif
9641 #endif
9642 
9643 #ifdef CONFIG_MAGIC_SYSRQ
9644 void normalize_rt_tasks(void)
9645 {
9646 	struct task_struct *g, *p;
9647 	struct sched_attr attr = {
9648 		.sched_policy = SCHED_NORMAL,
9649 	};
9650 
9651 	read_lock(&tasklist_lock);
9652 	for_each_process_thread(g, p) {
9653 		/*
9654 		 * Only normalize user tasks:
9655 		 */
9656 		if (p->flags & PF_KTHREAD)
9657 			continue;
9658 
9659 		p->se.exec_start = 0;
9660 		schedstat_set(p->stats.wait_start,  0);
9661 		schedstat_set(p->stats.sleep_start, 0);
9662 		schedstat_set(p->stats.block_start, 0);
9663 
9664 		if (!dl_task(p) && !rt_task(p)) {
9665 			/*
9666 			 * Renice negative nice level userspace
9667 			 * tasks back to 0:
9668 			 */
9669 			if (task_nice(p) < 0)
9670 				set_user_nice(p, 0);
9671 			continue;
9672 		}
9673 
9674 		__sched_setscheduler(p, &attr, false, false);
9675 	}
9676 	read_unlock(&tasklist_lock);
9677 }
9678 
9679 #endif /* CONFIG_MAGIC_SYSRQ */
9680 
9681 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9682 /*
9683  * These functions are only useful for the IA64 MCA handling, or kdb.
9684  *
9685  * They can only be called when the whole system has been
9686  * stopped - every CPU needs to be quiescent, and no scheduling
9687  * activity can take place. Using them for anything else would
9688  * be a serious bug, and as a result, they aren't even visible
9689  * under any other configuration.
9690  */
9691 
9692 /**
9693  * curr_task - return the current task for a given CPU.
9694  * @cpu: the processor in question.
9695  *
9696  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9697  *
9698  * Return: The current task for @cpu.
9699  */
9700 struct task_struct *curr_task(int cpu)
9701 {
9702 	return cpu_curr(cpu);
9703 }
9704 
9705 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9706 
9707 #ifdef CONFIG_IA64
9708 /**
9709  * ia64_set_curr_task - set the current task for a given CPU.
9710  * @cpu: the processor in question.
9711  * @p: the task pointer to set.
9712  *
9713  * Description: This function must only be used when non-maskable interrupts
9714  * are serviced on a separate stack. It allows the architecture to switch the
9715  * notion of the current task on a CPU in a non-blocking manner. This function
9716  * must be called with all CPU's synchronized, and interrupts disabled, the
9717  * and caller must save the original value of the current task (see
9718  * curr_task() above) and restore that value before reenabling interrupts and
9719  * re-starting the system.
9720  *
9721  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9722  */
9723 void ia64_set_curr_task(int cpu, struct task_struct *p)
9724 {
9725 	cpu_curr(cpu) = p;
9726 }
9727 
9728 #endif
9729 
9730 #ifdef CONFIG_CGROUP_SCHED
9731 /* task_group_lock serializes the addition/removal of task groups */
9732 static DEFINE_SPINLOCK(task_group_lock);
9733 
9734 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9735 					    struct task_group *parent)
9736 {
9737 #ifdef CONFIG_UCLAMP_TASK_GROUP
9738 	enum uclamp_id clamp_id;
9739 
9740 	for_each_clamp_id(clamp_id) {
9741 		uclamp_se_set(&tg->uclamp_req[clamp_id],
9742 			      uclamp_none(clamp_id), false);
9743 		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9744 	}
9745 #endif
9746 }
9747 
9748 static void sched_free_group(struct task_group *tg)
9749 {
9750 	free_fair_sched_group(tg);
9751 	free_rt_sched_group(tg);
9752 	autogroup_free(tg);
9753 	kmem_cache_free(task_group_cache, tg);
9754 }
9755 
9756 static void sched_free_group_rcu(struct rcu_head *rcu)
9757 {
9758 	sched_free_group(container_of(rcu, struct task_group, rcu));
9759 }
9760 
9761 static void sched_unregister_group(struct task_group *tg)
9762 {
9763 	unregister_fair_sched_group(tg);
9764 	unregister_rt_sched_group(tg);
9765 	/*
9766 	 * We have to wait for yet another RCU grace period to expire, as
9767 	 * print_cfs_stats() might run concurrently.
9768 	 */
9769 	call_rcu(&tg->rcu, sched_free_group_rcu);
9770 }
9771 
9772 /* allocate runqueue etc for a new task group */
9773 struct task_group *sched_create_group(struct task_group *parent)
9774 {
9775 	struct task_group *tg;
9776 
9777 	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9778 	if (!tg)
9779 		return ERR_PTR(-ENOMEM);
9780 
9781 	if (!alloc_fair_sched_group(tg, parent))
9782 		goto err;
9783 
9784 	if (!alloc_rt_sched_group(tg, parent))
9785 		goto err;
9786 
9787 	alloc_uclamp_sched_group(tg, parent);
9788 
9789 	return tg;
9790 
9791 err:
9792 	sched_free_group(tg);
9793 	return ERR_PTR(-ENOMEM);
9794 }
9795 
9796 void sched_online_group(struct task_group *tg, struct task_group *parent)
9797 {
9798 	unsigned long flags;
9799 
9800 	spin_lock_irqsave(&task_group_lock, flags);
9801 	list_add_rcu(&tg->list, &task_groups);
9802 
9803 	/* Root should already exist: */
9804 	WARN_ON(!parent);
9805 
9806 	tg->parent = parent;
9807 	INIT_LIST_HEAD(&tg->children);
9808 	list_add_rcu(&tg->siblings, &parent->children);
9809 	spin_unlock_irqrestore(&task_group_lock, flags);
9810 
9811 	online_fair_sched_group(tg);
9812 }
9813 
9814 /* rcu callback to free various structures associated with a task group */
9815 static void sched_unregister_group_rcu(struct rcu_head *rhp)
9816 {
9817 	/* Now it should be safe to free those cfs_rqs: */
9818 	sched_unregister_group(container_of(rhp, struct task_group, rcu));
9819 }
9820 
9821 void sched_destroy_group(struct task_group *tg)
9822 {
9823 	/* Wait for possible concurrent references to cfs_rqs complete: */
9824 	call_rcu(&tg->rcu, sched_unregister_group_rcu);
9825 }
9826 
9827 void sched_release_group(struct task_group *tg)
9828 {
9829 	unsigned long flags;
9830 
9831 	/*
9832 	 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
9833 	 * sched_cfs_period_timer()).
9834 	 *
9835 	 * For this to be effective, we have to wait for all pending users of
9836 	 * this task group to leave their RCU critical section to ensure no new
9837 	 * user will see our dying task group any more. Specifically ensure
9838 	 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
9839 	 *
9840 	 * We therefore defer calling unregister_fair_sched_group() to
9841 	 * sched_unregister_group() which is guarantied to get called only after the
9842 	 * current RCU grace period has expired.
9843 	 */
9844 	spin_lock_irqsave(&task_group_lock, flags);
9845 	list_del_rcu(&tg->list);
9846 	list_del_rcu(&tg->siblings);
9847 	spin_unlock_irqrestore(&task_group_lock, flags);
9848 }
9849 
9850 static void sched_change_group(struct task_struct *tsk, int type)
9851 {
9852 	struct task_group *tg;
9853 
9854 	/*
9855 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
9856 	 * which is pointless here. Thus, we pass "true" to task_css_check()
9857 	 * to prevent lockdep warnings.
9858 	 */
9859 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9860 			  struct task_group, css);
9861 	tg = autogroup_task_group(tsk, tg);
9862 	tsk->sched_task_group = tg;
9863 
9864 #ifdef CONFIG_FAIR_GROUP_SCHED
9865 	if (tsk->sched_class->task_change_group)
9866 		tsk->sched_class->task_change_group(tsk, type);
9867 	else
9868 #endif
9869 		set_task_rq(tsk, task_cpu(tsk));
9870 }
9871 
9872 /*
9873  * Change task's runqueue when it moves between groups.
9874  *
9875  * The caller of this function should have put the task in its new group by
9876  * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9877  * its new group.
9878  */
9879 void sched_move_task(struct task_struct *tsk)
9880 {
9881 	int queued, running, queue_flags =
9882 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9883 	struct rq_flags rf;
9884 	struct rq *rq;
9885 
9886 	rq = task_rq_lock(tsk, &rf);
9887 	update_rq_clock(rq);
9888 
9889 	running = task_current(rq, tsk);
9890 	queued = task_on_rq_queued(tsk);
9891 
9892 	if (queued)
9893 		dequeue_task(rq, tsk, queue_flags);
9894 	if (running)
9895 		put_prev_task(rq, tsk);
9896 
9897 	sched_change_group(tsk, TASK_MOVE_GROUP);
9898 
9899 	if (queued)
9900 		enqueue_task(rq, tsk, queue_flags);
9901 	if (running) {
9902 		set_next_task(rq, tsk);
9903 		/*
9904 		 * After changing group, the running task may have joined a
9905 		 * throttled one but it's still the running task. Trigger a
9906 		 * resched to make sure that task can still run.
9907 		 */
9908 		resched_curr(rq);
9909 	}
9910 
9911 	task_rq_unlock(rq, tsk, &rf);
9912 }
9913 
9914 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9915 {
9916 	return css ? container_of(css, struct task_group, css) : NULL;
9917 }
9918 
9919 static struct cgroup_subsys_state *
9920 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9921 {
9922 	struct task_group *parent = css_tg(parent_css);
9923 	struct task_group *tg;
9924 
9925 	if (!parent) {
9926 		/* This is early initialization for the top cgroup */
9927 		return &root_task_group.css;
9928 	}
9929 
9930 	tg = sched_create_group(parent);
9931 	if (IS_ERR(tg))
9932 		return ERR_PTR(-ENOMEM);
9933 
9934 	return &tg->css;
9935 }
9936 
9937 /* Expose task group only after completing cgroup initialization */
9938 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9939 {
9940 	struct task_group *tg = css_tg(css);
9941 	struct task_group *parent = css_tg(css->parent);
9942 
9943 	if (parent)
9944 		sched_online_group(tg, parent);
9945 
9946 #ifdef CONFIG_UCLAMP_TASK_GROUP
9947 	/* Propagate the effective uclamp value for the new group */
9948 	mutex_lock(&uclamp_mutex);
9949 	rcu_read_lock();
9950 	cpu_util_update_eff(css);
9951 	rcu_read_unlock();
9952 	mutex_unlock(&uclamp_mutex);
9953 #endif
9954 
9955 	return 0;
9956 }
9957 
9958 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9959 {
9960 	struct task_group *tg = css_tg(css);
9961 
9962 	sched_release_group(tg);
9963 }
9964 
9965 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9966 {
9967 	struct task_group *tg = css_tg(css);
9968 
9969 	/*
9970 	 * Relies on the RCU grace period between css_released() and this.
9971 	 */
9972 	sched_unregister_group(tg);
9973 }
9974 
9975 /*
9976  * This is called before wake_up_new_task(), therefore we really only
9977  * have to set its group bits, all the other stuff does not apply.
9978  */
9979 static void cpu_cgroup_fork(struct task_struct *task)
9980 {
9981 	struct rq_flags rf;
9982 	struct rq *rq;
9983 
9984 	rq = task_rq_lock(task, &rf);
9985 
9986 	update_rq_clock(rq);
9987 	sched_change_group(task, TASK_SET_GROUP);
9988 
9989 	task_rq_unlock(rq, task, &rf);
9990 }
9991 
9992 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9993 {
9994 	struct task_struct *task;
9995 	struct cgroup_subsys_state *css;
9996 	int ret = 0;
9997 
9998 	cgroup_taskset_for_each(task, css, tset) {
9999 #ifdef CONFIG_RT_GROUP_SCHED
10000 		if (!sched_rt_can_attach(css_tg(css), task))
10001 			return -EINVAL;
10002 #endif
10003 		/*
10004 		 * Serialize against wake_up_new_task() such that if it's
10005 		 * running, we're sure to observe its full state.
10006 		 */
10007 		raw_spin_lock_irq(&task->pi_lock);
10008 		/*
10009 		 * Avoid calling sched_move_task() before wake_up_new_task()
10010 		 * has happened. This would lead to problems with PELT, due to
10011 		 * move wanting to detach+attach while we're not attached yet.
10012 		 */
10013 		if (READ_ONCE(task->__state) == TASK_NEW)
10014 			ret = -EINVAL;
10015 		raw_spin_unlock_irq(&task->pi_lock);
10016 
10017 		if (ret)
10018 			break;
10019 	}
10020 	return ret;
10021 }
10022 
10023 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10024 {
10025 	struct task_struct *task;
10026 	struct cgroup_subsys_state *css;
10027 
10028 	cgroup_taskset_for_each(task, css, tset)
10029 		sched_move_task(task);
10030 }
10031 
10032 #ifdef CONFIG_UCLAMP_TASK_GROUP
10033 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10034 {
10035 	struct cgroup_subsys_state *top_css = css;
10036 	struct uclamp_se *uc_parent = NULL;
10037 	struct uclamp_se *uc_se = NULL;
10038 	unsigned int eff[UCLAMP_CNT];
10039 	enum uclamp_id clamp_id;
10040 	unsigned int clamps;
10041 
10042 	lockdep_assert_held(&uclamp_mutex);
10043 	SCHED_WARN_ON(!rcu_read_lock_held());
10044 
10045 	css_for_each_descendant_pre(css, top_css) {
10046 		uc_parent = css_tg(css)->parent
10047 			? css_tg(css)->parent->uclamp : NULL;
10048 
10049 		for_each_clamp_id(clamp_id) {
10050 			/* Assume effective clamps matches requested clamps */
10051 			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10052 			/* Cap effective clamps with parent's effective clamps */
10053 			if (uc_parent &&
10054 			    eff[clamp_id] > uc_parent[clamp_id].value) {
10055 				eff[clamp_id] = uc_parent[clamp_id].value;
10056 			}
10057 		}
10058 		/* Ensure protection is always capped by limit */
10059 		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10060 
10061 		/* Propagate most restrictive effective clamps */
10062 		clamps = 0x0;
10063 		uc_se = css_tg(css)->uclamp;
10064 		for_each_clamp_id(clamp_id) {
10065 			if (eff[clamp_id] == uc_se[clamp_id].value)
10066 				continue;
10067 			uc_se[clamp_id].value = eff[clamp_id];
10068 			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10069 			clamps |= (0x1 << clamp_id);
10070 		}
10071 		if (!clamps) {
10072 			css = css_rightmost_descendant(css);
10073 			continue;
10074 		}
10075 
10076 		/* Immediately update descendants RUNNABLE tasks */
10077 		uclamp_update_active_tasks(css);
10078 	}
10079 }
10080 
10081 /*
10082  * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10083  * C expression. Since there is no way to convert a macro argument (N) into a
10084  * character constant, use two levels of macros.
10085  */
10086 #define _POW10(exp) ((unsigned int)1e##exp)
10087 #define POW10(exp) _POW10(exp)
10088 
10089 struct uclamp_request {
10090 #define UCLAMP_PERCENT_SHIFT	2
10091 #define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
10092 	s64 percent;
10093 	u64 util;
10094 	int ret;
10095 };
10096 
10097 static inline struct uclamp_request
10098 capacity_from_percent(char *buf)
10099 {
10100 	struct uclamp_request req = {
10101 		.percent = UCLAMP_PERCENT_SCALE,
10102 		.util = SCHED_CAPACITY_SCALE,
10103 		.ret = 0,
10104 	};
10105 
10106 	buf = strim(buf);
10107 	if (strcmp(buf, "max")) {
10108 		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10109 					     &req.percent);
10110 		if (req.ret)
10111 			return req;
10112 		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10113 			req.ret = -ERANGE;
10114 			return req;
10115 		}
10116 
10117 		req.util = req.percent << SCHED_CAPACITY_SHIFT;
10118 		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10119 	}
10120 
10121 	return req;
10122 }
10123 
10124 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10125 				size_t nbytes, loff_t off,
10126 				enum uclamp_id clamp_id)
10127 {
10128 	struct uclamp_request req;
10129 	struct task_group *tg;
10130 
10131 	req = capacity_from_percent(buf);
10132 	if (req.ret)
10133 		return req.ret;
10134 
10135 	static_branch_enable(&sched_uclamp_used);
10136 
10137 	mutex_lock(&uclamp_mutex);
10138 	rcu_read_lock();
10139 
10140 	tg = css_tg(of_css(of));
10141 	if (tg->uclamp_req[clamp_id].value != req.util)
10142 		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10143 
10144 	/*
10145 	 * Because of not recoverable conversion rounding we keep track of the
10146 	 * exact requested value
10147 	 */
10148 	tg->uclamp_pct[clamp_id] = req.percent;
10149 
10150 	/* Update effective clamps to track the most restrictive value */
10151 	cpu_util_update_eff(of_css(of));
10152 
10153 	rcu_read_unlock();
10154 	mutex_unlock(&uclamp_mutex);
10155 
10156 	return nbytes;
10157 }
10158 
10159 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10160 				    char *buf, size_t nbytes,
10161 				    loff_t off)
10162 {
10163 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10164 }
10165 
10166 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10167 				    char *buf, size_t nbytes,
10168 				    loff_t off)
10169 {
10170 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10171 }
10172 
10173 static inline void cpu_uclamp_print(struct seq_file *sf,
10174 				    enum uclamp_id clamp_id)
10175 {
10176 	struct task_group *tg;
10177 	u64 util_clamp;
10178 	u64 percent;
10179 	u32 rem;
10180 
10181 	rcu_read_lock();
10182 	tg = css_tg(seq_css(sf));
10183 	util_clamp = tg->uclamp_req[clamp_id].value;
10184 	rcu_read_unlock();
10185 
10186 	if (util_clamp == SCHED_CAPACITY_SCALE) {
10187 		seq_puts(sf, "max\n");
10188 		return;
10189 	}
10190 
10191 	percent = tg->uclamp_pct[clamp_id];
10192 	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10193 	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10194 }
10195 
10196 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10197 {
10198 	cpu_uclamp_print(sf, UCLAMP_MIN);
10199 	return 0;
10200 }
10201 
10202 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10203 {
10204 	cpu_uclamp_print(sf, UCLAMP_MAX);
10205 	return 0;
10206 }
10207 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10208 
10209 #ifdef CONFIG_FAIR_GROUP_SCHED
10210 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10211 				struct cftype *cftype, u64 shareval)
10212 {
10213 	if (shareval > scale_load_down(ULONG_MAX))
10214 		shareval = MAX_SHARES;
10215 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
10216 }
10217 
10218 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10219 			       struct cftype *cft)
10220 {
10221 	struct task_group *tg = css_tg(css);
10222 
10223 	return (u64) scale_load_down(tg->shares);
10224 }
10225 
10226 #ifdef CONFIG_CFS_BANDWIDTH
10227 static DEFINE_MUTEX(cfs_constraints_mutex);
10228 
10229 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10230 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10231 /* More than 203 days if BW_SHIFT equals 20. */
10232 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10233 
10234 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10235 
10236 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10237 				u64 burst)
10238 {
10239 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
10240 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10241 
10242 	if (tg == &root_task_group)
10243 		return -EINVAL;
10244 
10245 	/*
10246 	 * Ensure we have at some amount of bandwidth every period.  This is
10247 	 * to prevent reaching a state of large arrears when throttled via
10248 	 * entity_tick() resulting in prolonged exit starvation.
10249 	 */
10250 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10251 		return -EINVAL;
10252 
10253 	/*
10254 	 * Likewise, bound things on the other side by preventing insane quota
10255 	 * periods.  This also allows us to normalize in computing quota
10256 	 * feasibility.
10257 	 */
10258 	if (period > max_cfs_quota_period)
10259 		return -EINVAL;
10260 
10261 	/*
10262 	 * Bound quota to defend quota against overflow during bandwidth shift.
10263 	 */
10264 	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10265 		return -EINVAL;
10266 
10267 	if (quota != RUNTIME_INF && (burst > quota ||
10268 				     burst + quota > max_cfs_runtime))
10269 		return -EINVAL;
10270 
10271 	/*
10272 	 * Prevent race between setting of cfs_rq->runtime_enabled and
10273 	 * unthrottle_offline_cfs_rqs().
10274 	 */
10275 	cpus_read_lock();
10276 	mutex_lock(&cfs_constraints_mutex);
10277 	ret = __cfs_schedulable(tg, period, quota);
10278 	if (ret)
10279 		goto out_unlock;
10280 
10281 	runtime_enabled = quota != RUNTIME_INF;
10282 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10283 	/*
10284 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
10285 	 * before making related changes, and on->off must occur afterwards
10286 	 */
10287 	if (runtime_enabled && !runtime_was_enabled)
10288 		cfs_bandwidth_usage_inc();
10289 	raw_spin_lock_irq(&cfs_b->lock);
10290 	cfs_b->period = ns_to_ktime(period);
10291 	cfs_b->quota = quota;
10292 	cfs_b->burst = burst;
10293 
10294 	__refill_cfs_bandwidth_runtime(cfs_b);
10295 
10296 	/* Restart the period timer (if active) to handle new period expiry: */
10297 	if (runtime_enabled)
10298 		start_cfs_bandwidth(cfs_b);
10299 
10300 	raw_spin_unlock_irq(&cfs_b->lock);
10301 
10302 	for_each_online_cpu(i) {
10303 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10304 		struct rq *rq = cfs_rq->rq;
10305 		struct rq_flags rf;
10306 
10307 		rq_lock_irq(rq, &rf);
10308 		cfs_rq->runtime_enabled = runtime_enabled;
10309 		cfs_rq->runtime_remaining = 0;
10310 
10311 		if (cfs_rq->throttled)
10312 			unthrottle_cfs_rq(cfs_rq);
10313 		rq_unlock_irq(rq, &rf);
10314 	}
10315 	if (runtime_was_enabled && !runtime_enabled)
10316 		cfs_bandwidth_usage_dec();
10317 out_unlock:
10318 	mutex_unlock(&cfs_constraints_mutex);
10319 	cpus_read_unlock();
10320 
10321 	return ret;
10322 }
10323 
10324 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10325 {
10326 	u64 quota, period, burst;
10327 
10328 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10329 	burst = tg->cfs_bandwidth.burst;
10330 	if (cfs_quota_us < 0)
10331 		quota = RUNTIME_INF;
10332 	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10333 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10334 	else
10335 		return -EINVAL;
10336 
10337 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10338 }
10339 
10340 static long tg_get_cfs_quota(struct task_group *tg)
10341 {
10342 	u64 quota_us;
10343 
10344 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10345 		return -1;
10346 
10347 	quota_us = tg->cfs_bandwidth.quota;
10348 	do_div(quota_us, NSEC_PER_USEC);
10349 
10350 	return quota_us;
10351 }
10352 
10353 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10354 {
10355 	u64 quota, period, burst;
10356 
10357 	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10358 		return -EINVAL;
10359 
10360 	period = (u64)cfs_period_us * NSEC_PER_USEC;
10361 	quota = tg->cfs_bandwidth.quota;
10362 	burst = tg->cfs_bandwidth.burst;
10363 
10364 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10365 }
10366 
10367 static long tg_get_cfs_period(struct task_group *tg)
10368 {
10369 	u64 cfs_period_us;
10370 
10371 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10372 	do_div(cfs_period_us, NSEC_PER_USEC);
10373 
10374 	return cfs_period_us;
10375 }
10376 
10377 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10378 {
10379 	u64 quota, period, burst;
10380 
10381 	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10382 		return -EINVAL;
10383 
10384 	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10385 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10386 	quota = tg->cfs_bandwidth.quota;
10387 
10388 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10389 }
10390 
10391 static long tg_get_cfs_burst(struct task_group *tg)
10392 {
10393 	u64 burst_us;
10394 
10395 	burst_us = tg->cfs_bandwidth.burst;
10396 	do_div(burst_us, NSEC_PER_USEC);
10397 
10398 	return burst_us;
10399 }
10400 
10401 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10402 				  struct cftype *cft)
10403 {
10404 	return tg_get_cfs_quota(css_tg(css));
10405 }
10406 
10407 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10408 				   struct cftype *cftype, s64 cfs_quota_us)
10409 {
10410 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10411 }
10412 
10413 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10414 				   struct cftype *cft)
10415 {
10416 	return tg_get_cfs_period(css_tg(css));
10417 }
10418 
10419 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10420 				    struct cftype *cftype, u64 cfs_period_us)
10421 {
10422 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
10423 }
10424 
10425 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10426 				  struct cftype *cft)
10427 {
10428 	return tg_get_cfs_burst(css_tg(css));
10429 }
10430 
10431 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10432 				   struct cftype *cftype, u64 cfs_burst_us)
10433 {
10434 	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10435 }
10436 
10437 struct cfs_schedulable_data {
10438 	struct task_group *tg;
10439 	u64 period, quota;
10440 };
10441 
10442 /*
10443  * normalize group quota/period to be quota/max_period
10444  * note: units are usecs
10445  */
10446 static u64 normalize_cfs_quota(struct task_group *tg,
10447 			       struct cfs_schedulable_data *d)
10448 {
10449 	u64 quota, period;
10450 
10451 	if (tg == d->tg) {
10452 		period = d->period;
10453 		quota = d->quota;
10454 	} else {
10455 		period = tg_get_cfs_period(tg);
10456 		quota = tg_get_cfs_quota(tg);
10457 	}
10458 
10459 	/* note: these should typically be equivalent */
10460 	if (quota == RUNTIME_INF || quota == -1)
10461 		return RUNTIME_INF;
10462 
10463 	return to_ratio(period, quota);
10464 }
10465 
10466 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10467 {
10468 	struct cfs_schedulable_data *d = data;
10469 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10470 	s64 quota = 0, parent_quota = -1;
10471 
10472 	if (!tg->parent) {
10473 		quota = RUNTIME_INF;
10474 	} else {
10475 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10476 
10477 		quota = normalize_cfs_quota(tg, d);
10478 		parent_quota = parent_b->hierarchical_quota;
10479 
10480 		/*
10481 		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
10482 		 * always take the min.  On cgroup1, only inherit when no
10483 		 * limit is set:
10484 		 */
10485 		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10486 			quota = min(quota, parent_quota);
10487 		} else {
10488 			if (quota == RUNTIME_INF)
10489 				quota = parent_quota;
10490 			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10491 				return -EINVAL;
10492 		}
10493 	}
10494 	cfs_b->hierarchical_quota = quota;
10495 
10496 	return 0;
10497 }
10498 
10499 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10500 {
10501 	int ret;
10502 	struct cfs_schedulable_data data = {
10503 		.tg = tg,
10504 		.period = period,
10505 		.quota = quota,
10506 	};
10507 
10508 	if (quota != RUNTIME_INF) {
10509 		do_div(data.period, NSEC_PER_USEC);
10510 		do_div(data.quota, NSEC_PER_USEC);
10511 	}
10512 
10513 	rcu_read_lock();
10514 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10515 	rcu_read_unlock();
10516 
10517 	return ret;
10518 }
10519 
10520 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10521 {
10522 	struct task_group *tg = css_tg(seq_css(sf));
10523 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10524 
10525 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10526 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10527 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10528 
10529 	if (schedstat_enabled() && tg != &root_task_group) {
10530 		struct sched_statistics *stats;
10531 		u64 ws = 0;
10532 		int i;
10533 
10534 		for_each_possible_cpu(i) {
10535 			stats = __schedstats_from_se(tg->se[i]);
10536 			ws += schedstat_val(stats->wait_sum);
10537 		}
10538 
10539 		seq_printf(sf, "wait_sum %llu\n", ws);
10540 	}
10541 
10542 	seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10543 	seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10544 
10545 	return 0;
10546 }
10547 #endif /* CONFIG_CFS_BANDWIDTH */
10548 #endif /* CONFIG_FAIR_GROUP_SCHED */
10549 
10550 #ifdef CONFIG_RT_GROUP_SCHED
10551 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10552 				struct cftype *cft, s64 val)
10553 {
10554 	return sched_group_set_rt_runtime(css_tg(css), val);
10555 }
10556 
10557 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10558 			       struct cftype *cft)
10559 {
10560 	return sched_group_rt_runtime(css_tg(css));
10561 }
10562 
10563 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10564 				    struct cftype *cftype, u64 rt_period_us)
10565 {
10566 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
10567 }
10568 
10569 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10570 				   struct cftype *cft)
10571 {
10572 	return sched_group_rt_period(css_tg(css));
10573 }
10574 #endif /* CONFIG_RT_GROUP_SCHED */
10575 
10576 #ifdef CONFIG_FAIR_GROUP_SCHED
10577 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10578 			       struct cftype *cft)
10579 {
10580 	return css_tg(css)->idle;
10581 }
10582 
10583 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10584 				struct cftype *cft, s64 idle)
10585 {
10586 	return sched_group_set_idle(css_tg(css), idle);
10587 }
10588 #endif
10589 
10590 static struct cftype cpu_legacy_files[] = {
10591 #ifdef CONFIG_FAIR_GROUP_SCHED
10592 	{
10593 		.name = "shares",
10594 		.read_u64 = cpu_shares_read_u64,
10595 		.write_u64 = cpu_shares_write_u64,
10596 	},
10597 	{
10598 		.name = "idle",
10599 		.read_s64 = cpu_idle_read_s64,
10600 		.write_s64 = cpu_idle_write_s64,
10601 	},
10602 #endif
10603 #ifdef CONFIG_CFS_BANDWIDTH
10604 	{
10605 		.name = "cfs_quota_us",
10606 		.read_s64 = cpu_cfs_quota_read_s64,
10607 		.write_s64 = cpu_cfs_quota_write_s64,
10608 	},
10609 	{
10610 		.name = "cfs_period_us",
10611 		.read_u64 = cpu_cfs_period_read_u64,
10612 		.write_u64 = cpu_cfs_period_write_u64,
10613 	},
10614 	{
10615 		.name = "cfs_burst_us",
10616 		.read_u64 = cpu_cfs_burst_read_u64,
10617 		.write_u64 = cpu_cfs_burst_write_u64,
10618 	},
10619 	{
10620 		.name = "stat",
10621 		.seq_show = cpu_cfs_stat_show,
10622 	},
10623 #endif
10624 #ifdef CONFIG_RT_GROUP_SCHED
10625 	{
10626 		.name = "rt_runtime_us",
10627 		.read_s64 = cpu_rt_runtime_read,
10628 		.write_s64 = cpu_rt_runtime_write,
10629 	},
10630 	{
10631 		.name = "rt_period_us",
10632 		.read_u64 = cpu_rt_period_read_uint,
10633 		.write_u64 = cpu_rt_period_write_uint,
10634 	},
10635 #endif
10636 #ifdef CONFIG_UCLAMP_TASK_GROUP
10637 	{
10638 		.name = "uclamp.min",
10639 		.flags = CFTYPE_NOT_ON_ROOT,
10640 		.seq_show = cpu_uclamp_min_show,
10641 		.write = cpu_uclamp_min_write,
10642 	},
10643 	{
10644 		.name = "uclamp.max",
10645 		.flags = CFTYPE_NOT_ON_ROOT,
10646 		.seq_show = cpu_uclamp_max_show,
10647 		.write = cpu_uclamp_max_write,
10648 	},
10649 #endif
10650 	{ }	/* Terminate */
10651 };
10652 
10653 static int cpu_extra_stat_show(struct seq_file *sf,
10654 			       struct cgroup_subsys_state *css)
10655 {
10656 #ifdef CONFIG_CFS_BANDWIDTH
10657 	{
10658 		struct task_group *tg = css_tg(css);
10659 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10660 		u64 throttled_usec, burst_usec;
10661 
10662 		throttled_usec = cfs_b->throttled_time;
10663 		do_div(throttled_usec, NSEC_PER_USEC);
10664 		burst_usec = cfs_b->burst_time;
10665 		do_div(burst_usec, NSEC_PER_USEC);
10666 
10667 		seq_printf(sf, "nr_periods %d\n"
10668 			   "nr_throttled %d\n"
10669 			   "throttled_usec %llu\n"
10670 			   "nr_bursts %d\n"
10671 			   "burst_usec %llu\n",
10672 			   cfs_b->nr_periods, cfs_b->nr_throttled,
10673 			   throttled_usec, cfs_b->nr_burst, burst_usec);
10674 	}
10675 #endif
10676 	return 0;
10677 }
10678 
10679 #ifdef CONFIG_FAIR_GROUP_SCHED
10680 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10681 			       struct cftype *cft)
10682 {
10683 	struct task_group *tg = css_tg(css);
10684 	u64 weight = scale_load_down(tg->shares);
10685 
10686 	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10687 }
10688 
10689 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10690 				struct cftype *cft, u64 weight)
10691 {
10692 	/*
10693 	 * cgroup weight knobs should use the common MIN, DFL and MAX
10694 	 * values which are 1, 100 and 10000 respectively.  While it loses
10695 	 * a bit of range on both ends, it maps pretty well onto the shares
10696 	 * value used by scheduler and the round-trip conversions preserve
10697 	 * the original value over the entire range.
10698 	 */
10699 	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10700 		return -ERANGE;
10701 
10702 	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10703 
10704 	return sched_group_set_shares(css_tg(css), scale_load(weight));
10705 }
10706 
10707 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10708 				    struct cftype *cft)
10709 {
10710 	unsigned long weight = scale_load_down(css_tg(css)->shares);
10711 	int last_delta = INT_MAX;
10712 	int prio, delta;
10713 
10714 	/* find the closest nice value to the current weight */
10715 	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10716 		delta = abs(sched_prio_to_weight[prio] - weight);
10717 		if (delta >= last_delta)
10718 			break;
10719 		last_delta = delta;
10720 	}
10721 
10722 	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10723 }
10724 
10725 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10726 				     struct cftype *cft, s64 nice)
10727 {
10728 	unsigned long weight;
10729 	int idx;
10730 
10731 	if (nice < MIN_NICE || nice > MAX_NICE)
10732 		return -ERANGE;
10733 
10734 	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10735 	idx = array_index_nospec(idx, 40);
10736 	weight = sched_prio_to_weight[idx];
10737 
10738 	return sched_group_set_shares(css_tg(css), scale_load(weight));
10739 }
10740 #endif
10741 
10742 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10743 						  long period, long quota)
10744 {
10745 	if (quota < 0)
10746 		seq_puts(sf, "max");
10747 	else
10748 		seq_printf(sf, "%ld", quota);
10749 
10750 	seq_printf(sf, " %ld\n", period);
10751 }
10752 
10753 /* caller should put the current value in *@periodp before calling */
10754 static int __maybe_unused cpu_period_quota_parse(char *buf,
10755 						 u64 *periodp, u64 *quotap)
10756 {
10757 	char tok[21];	/* U64_MAX */
10758 
10759 	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10760 		return -EINVAL;
10761 
10762 	*periodp *= NSEC_PER_USEC;
10763 
10764 	if (sscanf(tok, "%llu", quotap))
10765 		*quotap *= NSEC_PER_USEC;
10766 	else if (!strcmp(tok, "max"))
10767 		*quotap = RUNTIME_INF;
10768 	else
10769 		return -EINVAL;
10770 
10771 	return 0;
10772 }
10773 
10774 #ifdef CONFIG_CFS_BANDWIDTH
10775 static int cpu_max_show(struct seq_file *sf, void *v)
10776 {
10777 	struct task_group *tg = css_tg(seq_css(sf));
10778 
10779 	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10780 	return 0;
10781 }
10782 
10783 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10784 			     char *buf, size_t nbytes, loff_t off)
10785 {
10786 	struct task_group *tg = css_tg(of_css(of));
10787 	u64 period = tg_get_cfs_period(tg);
10788 	u64 burst = tg_get_cfs_burst(tg);
10789 	u64 quota;
10790 	int ret;
10791 
10792 	ret = cpu_period_quota_parse(buf, &period, &quota);
10793 	if (!ret)
10794 		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10795 	return ret ?: nbytes;
10796 }
10797 #endif
10798 
10799 static struct cftype cpu_files[] = {
10800 #ifdef CONFIG_FAIR_GROUP_SCHED
10801 	{
10802 		.name = "weight",
10803 		.flags = CFTYPE_NOT_ON_ROOT,
10804 		.read_u64 = cpu_weight_read_u64,
10805 		.write_u64 = cpu_weight_write_u64,
10806 	},
10807 	{
10808 		.name = "weight.nice",
10809 		.flags = CFTYPE_NOT_ON_ROOT,
10810 		.read_s64 = cpu_weight_nice_read_s64,
10811 		.write_s64 = cpu_weight_nice_write_s64,
10812 	},
10813 	{
10814 		.name = "idle",
10815 		.flags = CFTYPE_NOT_ON_ROOT,
10816 		.read_s64 = cpu_idle_read_s64,
10817 		.write_s64 = cpu_idle_write_s64,
10818 	},
10819 #endif
10820 #ifdef CONFIG_CFS_BANDWIDTH
10821 	{
10822 		.name = "max",
10823 		.flags = CFTYPE_NOT_ON_ROOT,
10824 		.seq_show = cpu_max_show,
10825 		.write = cpu_max_write,
10826 	},
10827 	{
10828 		.name = "max.burst",
10829 		.flags = CFTYPE_NOT_ON_ROOT,
10830 		.read_u64 = cpu_cfs_burst_read_u64,
10831 		.write_u64 = cpu_cfs_burst_write_u64,
10832 	},
10833 #endif
10834 #ifdef CONFIG_UCLAMP_TASK_GROUP
10835 	{
10836 		.name = "uclamp.min",
10837 		.flags = CFTYPE_NOT_ON_ROOT,
10838 		.seq_show = cpu_uclamp_min_show,
10839 		.write = cpu_uclamp_min_write,
10840 	},
10841 	{
10842 		.name = "uclamp.max",
10843 		.flags = CFTYPE_NOT_ON_ROOT,
10844 		.seq_show = cpu_uclamp_max_show,
10845 		.write = cpu_uclamp_max_write,
10846 	},
10847 #endif
10848 	{ }	/* terminate */
10849 };
10850 
10851 struct cgroup_subsys cpu_cgrp_subsys = {
10852 	.css_alloc	= cpu_cgroup_css_alloc,
10853 	.css_online	= cpu_cgroup_css_online,
10854 	.css_released	= cpu_cgroup_css_released,
10855 	.css_free	= cpu_cgroup_css_free,
10856 	.css_extra_stat_show = cpu_extra_stat_show,
10857 	.fork		= cpu_cgroup_fork,
10858 	.can_attach	= cpu_cgroup_can_attach,
10859 	.attach		= cpu_cgroup_attach,
10860 	.legacy_cftypes	= cpu_legacy_files,
10861 	.dfl_cftypes	= cpu_files,
10862 	.early_init	= true,
10863 	.threaded	= true,
10864 };
10865 
10866 #endif	/* CONFIG_CGROUP_SCHED */
10867 
10868 void dump_cpu_task(int cpu)
10869 {
10870 	pr_info("Task dump for CPU %d:\n", cpu);
10871 	sched_show_task(cpu_curr(cpu));
10872 }
10873 
10874 /*
10875  * Nice levels are multiplicative, with a gentle 10% change for every
10876  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10877  * nice 1, it will get ~10% less CPU time than another CPU-bound task
10878  * that remained on nice 0.
10879  *
10880  * The "10% effect" is relative and cumulative: from _any_ nice level,
10881  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10882  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10883  * If a task goes up by ~10% and another task goes down by ~10% then
10884  * the relative distance between them is ~25%.)
10885  */
10886 const int sched_prio_to_weight[40] = {
10887  /* -20 */     88761,     71755,     56483,     46273,     36291,
10888  /* -15 */     29154,     23254,     18705,     14949,     11916,
10889  /* -10 */      9548,      7620,      6100,      4904,      3906,
10890  /*  -5 */      3121,      2501,      1991,      1586,      1277,
10891  /*   0 */      1024,       820,       655,       526,       423,
10892  /*   5 */       335,       272,       215,       172,       137,
10893  /*  10 */       110,        87,        70,        56,        45,
10894  /*  15 */        36,        29,        23,        18,        15,
10895 };
10896 
10897 /*
10898  * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10899  *
10900  * In cases where the weight does not change often, we can use the
10901  * precalculated inverse to speed up arithmetics by turning divisions
10902  * into multiplications:
10903  */
10904 const u32 sched_prio_to_wmult[40] = {
10905  /* -20 */     48388,     59856,     76040,     92818,    118348,
10906  /* -15 */    147320,    184698,    229616,    287308,    360437,
10907  /* -10 */    449829,    563644,    704093,    875809,   1099582,
10908  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
10909  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
10910  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
10911  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
10912  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10913 };
10914 
10915 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10916 {
10917         trace_sched_update_nr_running_tp(rq, count);
10918 }
10919