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#include <linux/highmem.h>
10#include <linux/hrtimer_api.h>
11#include <linux/ktime_api.h>
12#include <linux/sched/signal.h>
13#include <linux/syscalls_api.h>
14#include <linux/debug_locks.h>
15#include <linux/prefetch.h>
16#include <linux/capability.h>
17#include <linux/pgtable_api.h>
18#include <linux/wait_bit.h>
19#include <linux/jiffies.h>
20#include <linux/spinlock_api.h>
21#include <linux/cpumask_api.h>
22#include <linux/lockdep_api.h>
23#include <linux/hardirq.h>
24#include <linux/softirq.h>
25#include <linux/refcount_api.h>
26#include <linux/topology.h>
27#include <linux/sched/clock.h>
28#include <linux/sched/cond_resched.h>
29#include <linux/sched/cputime.h>
30#include <linux/sched/debug.h>
31#include <linux/sched/hotplug.h>
32#include <linux/sched/init.h>
33#include <linux/sched/isolation.h>
34#include <linux/sched/loadavg.h>
35#include <linux/sched/mm.h>
36#include <linux/sched/nohz.h>
37#include <linux/sched/rseq_api.h>
38#include <linux/sched/rt.h>
39
40#include <linux/blkdev.h>
41#include <linux/context_tracking.h>
42#include <linux/cpuset.h>
43#include <linux/delayacct.h>
44#include <linux/init_task.h>
45#include <linux/interrupt.h>
46#include <linux/ioprio.h>
47#include <linux/kallsyms.h>
48#include <linux/kcov.h>
49#include <linux/kprobes.h>
50#include <linux/llist_api.h>
51#include <linux/mmu_context.h>
52#include <linux/mmzone.h>
53#include <linux/mutex_api.h>
54#include <linux/nmi.h>
55#include <linux/nospec.h>
56#include <linux/perf_event_api.h>
57#include <linux/profile.h>
58#include <linux/psi.h>
59#include <linux/rcuwait_api.h>
60#include <linux/sched/wake_q.h>
61#include <linux/scs.h>
62#include <linux/slab.h>
63#include <linux/syscalls.h>
64#include <linux/vtime.h>
65#include <linux/wait_api.h>
66#include <linux/workqueue_api.h>
67
68#ifdef CONFIG_PREEMPT_DYNAMIC
69# ifdef CONFIG_GENERIC_ENTRY
70#  include <linux/entry-common.h>
71# endif
72#endif
73
74#include <uapi/linux/sched/types.h>
75
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78
79#define CREATE_TRACE_POINTS
80#include <linux/sched/rseq_api.h>
81#include <trace/events/sched.h>
82#undef CREATE_TRACE_POINTS
83
84#include "sched.h"
85#include "stats.h"
86#include "autogroup.h"
87
88#include "autogroup.h"
89#include "pelt.h"
90#include "smp.h"
91#include "stats.h"
92
93#include "../workqueue_internal.h"
94#include "../../io_uring/io-wq.h"
95#include "../smpboot.h"
96
97/*
98 * Export tracepoints that act as a bare tracehook (ie: have no trace event
99 * associated with them) to allow external modules to probe them.
100 */
101EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
102EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
103EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
104EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
105EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
106EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
107EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
108EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
109EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
110EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
111EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
112
113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114
115#ifdef CONFIG_SCHED_DEBUG
116/*
117 * Debugging: various feature bits
118 *
119 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
120 * sysctl_sched_features, defined in sched.h, to allow constants propagation
121 * at compile time and compiler optimization based on features default.
122 */
123#define SCHED_FEAT(name, enabled)	\
124	(1UL << __SCHED_FEAT_##name) * enabled |
125const_debug unsigned int sysctl_sched_features =
126#include "features.h"
127	0;
128#undef SCHED_FEAT
129
130/*
131 * Print a warning if need_resched is set for the given duration (if
132 * LATENCY_WARN is enabled).
133 *
134 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
135 * per boot.
136 */
137__read_mostly int sysctl_resched_latency_warn_ms = 100;
138__read_mostly int sysctl_resched_latency_warn_once = 1;
139#endif /* CONFIG_SCHED_DEBUG */
140
141/*
142 * Number of tasks to iterate in a single balance run.
143 * Limited because this is done with IRQs disabled.
144 */
145#ifdef CONFIG_PREEMPT_RT
146const_debug unsigned int sysctl_sched_nr_migrate = 8;
147#else
148const_debug unsigned int sysctl_sched_nr_migrate = 32;
149#endif
150
151__read_mostly int scheduler_running;
152
153#ifdef CONFIG_SCHED_CORE
154
155DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
156
157/* kernel prio, less is more */
158static inline int __task_prio(struct task_struct *p)
159{
160	if (p->sched_class == &stop_sched_class) /* trumps deadline */
161		return -2;
162
163	if (rt_prio(p->prio)) /* includes deadline */
164		return p->prio; /* [-1, 99] */
165
166	if (p->sched_class == &idle_sched_class)
167		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
168
169	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
170}
171
172/*
173 * l(a,b)
174 * le(a,b) := !l(b,a)
175 * g(a,b)  := l(b,a)
176 * ge(a,b) := !l(a,b)
177 */
178
179/* real prio, less is less */
180static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
181{
182
183	int pa = __task_prio(a), pb = __task_prio(b);
184
185	if (-pa < -pb)
186		return true;
187
188	if (-pb < -pa)
189		return false;
190
191	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
192		return !dl_time_before(a->dl.deadline, b->dl.deadline);
193
194	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
195		return cfs_prio_less(a, b, in_fi);
196
197	return false;
198}
199
200static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
201{
202	if (a->core_cookie < b->core_cookie)
203		return true;
204
205	if (a->core_cookie > b->core_cookie)
206		return false;
207
208	/* flip prio, so high prio is leftmost */
209	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
210		return true;
211
212	return false;
213}
214
215#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
216
217static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
218{
219	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
220}
221
222static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
223{
224	const struct task_struct *p = __node_2_sc(node);
225	unsigned long cookie = (unsigned long)key;
226
227	if (cookie < p->core_cookie)
228		return -1;
229
230	if (cookie > p->core_cookie)
231		return 1;
232
233	return 0;
234}
235
236void sched_core_enqueue(struct rq *rq, struct task_struct *p)
237{
238	rq->core->core_task_seq++;
239
240	if (!p->core_cookie)
241		return;
242
243	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
244}
245
246void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
247{
248	rq->core->core_task_seq++;
249
250	if (sched_core_enqueued(p)) {
251		rb_erase(&p->core_node, &rq->core_tree);
252		RB_CLEAR_NODE(&p->core_node);
253	}
254
255	/*
256	 * Migrating the last task off the cpu, with the cpu in forced idle
257	 * state. Reschedule to create an accounting edge for forced idle,
258	 * and re-examine whether the core is still in forced idle state.
259	 */
260	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
261	    rq->core->core_forceidle_count && rq->curr == rq->idle)
262		resched_curr(rq);
263}
264
265/*
266 * Find left-most (aka, highest priority) task matching @cookie.
267 */
268static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
269{
270	struct rb_node *node;
271
272	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
273	/*
274	 * The idle task always matches any cookie!
275	 */
276	if (!node)
277		return idle_sched_class.pick_task(rq);
278
279	return __node_2_sc(node);
280}
281
282static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
283{
284	struct rb_node *node = &p->core_node;
285
286	node = rb_next(node);
287	if (!node)
288		return NULL;
289
290	p = container_of(node, struct task_struct, core_node);
291	if (p->core_cookie != cookie)
292		return NULL;
293
294	return p;
295}
296
297/*
298 * Magic required such that:
299 *
300 *	raw_spin_rq_lock(rq);
301 *	...
302 *	raw_spin_rq_unlock(rq);
303 *
304 * ends up locking and unlocking the _same_ lock, and all CPUs
305 * always agree on what rq has what lock.
306 *
307 * XXX entirely possible to selectively enable cores, don't bother for now.
308 */
309
310static DEFINE_MUTEX(sched_core_mutex);
311static atomic_t sched_core_count;
312static struct cpumask sched_core_mask;
313
314static void sched_core_lock(int cpu, unsigned long *flags)
315{
316	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
317	int t, i = 0;
318
319	local_irq_save(*flags);
320	for_each_cpu(t, smt_mask)
321		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
322}
323
324static void sched_core_unlock(int cpu, unsigned long *flags)
325{
326	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
327	int t;
328
329	for_each_cpu(t, smt_mask)
330		raw_spin_unlock(&cpu_rq(t)->__lock);
331	local_irq_restore(*flags);
332}
333
334static void __sched_core_flip(bool enabled)
335{
336	unsigned long flags;
337	int cpu, t;
338
339	cpus_read_lock();
340
341	/*
342	 * Toggle the online cores, one by one.
343	 */
344	cpumask_copy(&sched_core_mask, cpu_online_mask);
345	for_each_cpu(cpu, &sched_core_mask) {
346		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
347
348		sched_core_lock(cpu, &flags);
349
350		for_each_cpu(t, smt_mask)
351			cpu_rq(t)->core_enabled = enabled;
352
353		cpu_rq(cpu)->core->core_forceidle_start = 0;
354
355		sched_core_unlock(cpu, &flags);
356
357		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
358	}
359
360	/*
361	 * Toggle the offline CPUs.
362	 */
363	cpumask_copy(&sched_core_mask, cpu_possible_mask);
364	cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
365
366	for_each_cpu(cpu, &sched_core_mask)
367		cpu_rq(cpu)->core_enabled = enabled;
368
369	cpus_read_unlock();
370}
371
372static void sched_core_assert_empty(void)
373{
374	int cpu;
375
376	for_each_possible_cpu(cpu)
377		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
378}
379
380static void __sched_core_enable(void)
381{
382	static_branch_enable(&__sched_core_enabled);
383	/*
384	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
385	 * and future ones will observe !sched_core_disabled().
386	 */
387	synchronize_rcu();
388	__sched_core_flip(true);
389	sched_core_assert_empty();
390}
391
392static void __sched_core_disable(void)
393{
394	sched_core_assert_empty();
395	__sched_core_flip(false);
396	static_branch_disable(&__sched_core_enabled);
397}
398
399void sched_core_get(void)
400{
401	if (atomic_inc_not_zero(&sched_core_count))
402		return;
403
404	mutex_lock(&sched_core_mutex);
405	if (!atomic_read(&sched_core_count))
406		__sched_core_enable();
407
408	smp_mb__before_atomic();
409	atomic_inc(&sched_core_count);
410	mutex_unlock(&sched_core_mutex);
411}
412
413static void __sched_core_put(struct work_struct *work)
414{
415	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
416		__sched_core_disable();
417		mutex_unlock(&sched_core_mutex);
418	}
419}
420
421void sched_core_put(void)
422{
423	static DECLARE_WORK(_work, __sched_core_put);
424
425	/*
426	 * "There can be only one"
427	 *
428	 * Either this is the last one, or we don't actually need to do any
429	 * 'work'. If it is the last *again*, we rely on
430	 * WORK_STRUCT_PENDING_BIT.
431	 */
432	if (!atomic_add_unless(&sched_core_count, -1, 1))
433		schedule_work(&_work);
434}
435
436#else /* !CONFIG_SCHED_CORE */
437
438static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
439static inline void
440sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
441
442#endif /* CONFIG_SCHED_CORE */
443
444/*
445 * Serialization rules:
446 *
447 * Lock order:
448 *
449 *   p->pi_lock
450 *     rq->lock
451 *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
452 *
453 *  rq1->lock
454 *    rq2->lock  where: rq1 < rq2
455 *
456 * Regular state:
457 *
458 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
459 * local CPU's rq->lock, it optionally removes the task from the runqueue and
460 * always looks at the local rq data structures to find the most eligible task
461 * to run next.
462 *
463 * Task enqueue is also under rq->lock, possibly taken from another CPU.
464 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
465 * the local CPU to avoid bouncing the runqueue state around [ see
466 * ttwu_queue_wakelist() ]
467 *
468 * Task wakeup, specifically wakeups that involve migration, are horribly
469 * complicated to avoid having to take two rq->locks.
470 *
471 * Special state:
472 *
473 * System-calls and anything external will use task_rq_lock() which acquires
474 * both p->pi_lock and rq->lock. As a consequence the state they change is
475 * stable while holding either lock:
476 *
477 *  - sched_setaffinity()/
478 *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
479 *  - set_user_nice():		p->se.load, p->*prio
480 *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
481 *				p->se.load, p->rt_priority,
482 *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
483 *  - sched_setnuma():		p->numa_preferred_nid
484 *  - sched_move_task()/
485 *    cpu_cgroup_fork():	p->sched_task_group
486 *  - uclamp_update_active()	p->uclamp*
487 *
488 * p->state <- TASK_*:
489 *
490 *   is changed locklessly using set_current_state(), __set_current_state() or
491 *   set_special_state(), see their respective comments, or by
492 *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
493 *   concurrent self.
494 *
495 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
496 *
497 *   is set by activate_task() and cleared by deactivate_task(), under
498 *   rq->lock. Non-zero indicates the task is runnable, the special
499 *   ON_RQ_MIGRATING state is used for migration without holding both
500 *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
501 *
502 * p->on_cpu <- { 0, 1 }:
503 *
504 *   is set by prepare_task() and cleared by finish_task() such that it will be
505 *   set before p is scheduled-in and cleared after p is scheduled-out, both
506 *   under rq->lock. Non-zero indicates the task is running on its CPU.
507 *
508 *   [ The astute reader will observe that it is possible for two tasks on one
509 *     CPU to have ->on_cpu = 1 at the same time. ]
510 *
511 * task_cpu(p): is changed by set_task_cpu(), the rules are:
512 *
513 *  - Don't call set_task_cpu() on a blocked task:
514 *
515 *    We don't care what CPU we're not running on, this simplifies hotplug,
516 *    the CPU assignment of blocked tasks isn't required to be valid.
517 *
518 *  - for try_to_wake_up(), called under p->pi_lock:
519 *
520 *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
521 *
522 *  - for migration called under rq->lock:
523 *    [ see task_on_rq_migrating() in task_rq_lock() ]
524 *
525 *    o move_queued_task()
526 *    o detach_task()
527 *
528 *  - for migration called under double_rq_lock():
529 *
530 *    o __migrate_swap_task()
531 *    o push_rt_task() / pull_rt_task()
532 *    o push_dl_task() / pull_dl_task()
533 *    o dl_task_offline_migration()
534 *
535 */
536
537void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
538{
539	raw_spinlock_t *lock;
540
541	/* Matches synchronize_rcu() in __sched_core_enable() */
542	preempt_disable();
543	if (sched_core_disabled()) {
544		raw_spin_lock_nested(&rq->__lock, subclass);
545		/* preempt_count *MUST* be > 1 */
546		preempt_enable_no_resched();
547		return;
548	}
549
550	for (;;) {
551		lock = __rq_lockp(rq);
552		raw_spin_lock_nested(lock, subclass);
553		if (likely(lock == __rq_lockp(rq))) {
554			/* preempt_count *MUST* be > 1 */
555			preempt_enable_no_resched();
556			return;
557		}
558		raw_spin_unlock(lock);
559	}
560}
561
562bool raw_spin_rq_trylock(struct rq *rq)
563{
564	raw_spinlock_t *lock;
565	bool ret;
566
567	/* Matches synchronize_rcu() in __sched_core_enable() */
568	preempt_disable();
569	if (sched_core_disabled()) {
570		ret = raw_spin_trylock(&rq->__lock);
571		preempt_enable();
572		return ret;
573	}
574
575	for (;;) {
576		lock = __rq_lockp(rq);
577		ret = raw_spin_trylock(lock);
578		if (!ret || (likely(lock == __rq_lockp(rq)))) {
579			preempt_enable();
580			return ret;
581		}
582		raw_spin_unlock(lock);
583	}
584}
585
586void raw_spin_rq_unlock(struct rq *rq)
587{
588	raw_spin_unlock(rq_lockp(rq));
589}
590
591#ifdef CONFIG_SMP
592/*
593 * double_rq_lock - safely lock two runqueues
594 */
595void double_rq_lock(struct rq *rq1, struct rq *rq2)
596{
597	lockdep_assert_irqs_disabled();
598
599	if (rq_order_less(rq2, rq1))
600		swap(rq1, rq2);
601
602	raw_spin_rq_lock(rq1);
603	if (__rq_lockp(rq1) != __rq_lockp(rq2))
604		raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
605
606	double_rq_clock_clear_update(rq1, rq2);
607}
608#endif
609
610/*
611 * __task_rq_lock - lock the rq @p resides on.
612 */
613struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
614	__acquires(rq->lock)
615{
616	struct rq *rq;
617
618	lockdep_assert_held(&p->pi_lock);
619
620	for (;;) {
621		rq = task_rq(p);
622		raw_spin_rq_lock(rq);
623		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
624			rq_pin_lock(rq, rf);
625			return rq;
626		}
627		raw_spin_rq_unlock(rq);
628
629		while (unlikely(task_on_rq_migrating(p)))
630			cpu_relax();
631	}
632}
633
634/*
635 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
636 */
637struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
638	__acquires(p->pi_lock)
639	__acquires(rq->lock)
640{
641	struct rq *rq;
642
643	for (;;) {
644		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
645		rq = task_rq(p);
646		raw_spin_rq_lock(rq);
647		/*
648		 *	move_queued_task()		task_rq_lock()
649		 *
650		 *	ACQUIRE (rq->lock)
651		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
652		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
653		 *	[S] ->cpu = new_cpu		[L] task_rq()
654		 *					[L] ->on_rq
655		 *	RELEASE (rq->lock)
656		 *
657		 * If we observe the old CPU in task_rq_lock(), the acquire of
658		 * the old rq->lock will fully serialize against the stores.
659		 *
660		 * If we observe the new CPU in task_rq_lock(), the address
661		 * dependency headed by '[L] rq = task_rq()' and the acquire
662		 * will pair with the WMB to ensure we then also see migrating.
663		 */
664		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
665			rq_pin_lock(rq, rf);
666			return rq;
667		}
668		raw_spin_rq_unlock(rq);
669		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
670
671		while (unlikely(task_on_rq_migrating(p)))
672			cpu_relax();
673	}
674}
675
676/*
677 * RQ-clock updating methods:
678 */
679
680static void update_rq_clock_task(struct rq *rq, s64 delta)
681{
682/*
683 * In theory, the compile should just see 0 here, and optimize out the call
684 * to sched_rt_avg_update. But I don't trust it...
685 */
686	s64 __maybe_unused steal = 0, irq_delta = 0;
687
688#ifdef CONFIG_IRQ_TIME_ACCOUNTING
689	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
690
691	/*
692	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
693	 * this case when a previous update_rq_clock() happened inside a
694	 * {soft,}irq region.
695	 *
696	 * When this happens, we stop ->clock_task and only update the
697	 * prev_irq_time stamp to account for the part that fit, so that a next
698	 * update will consume the rest. This ensures ->clock_task is
699	 * monotonic.
700	 *
701	 * It does however cause some slight miss-attribution of {soft,}irq
702	 * time, a more accurate solution would be to update the irq_time using
703	 * the current rq->clock timestamp, except that would require using
704	 * atomic ops.
705	 */
706	if (irq_delta > delta)
707		irq_delta = delta;
708
709	rq->prev_irq_time += irq_delta;
710	delta -= irq_delta;
711#endif
712#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
713	if (static_key_false((&paravirt_steal_rq_enabled))) {
714		steal = paravirt_steal_clock(cpu_of(rq));
715		steal -= rq->prev_steal_time_rq;
716
717		if (unlikely(steal > delta))
718			steal = delta;
719
720		rq->prev_steal_time_rq += steal;
721		delta -= steal;
722	}
723#endif
724
725	rq->clock_task += delta;
726
727#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
728	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
729		update_irq_load_avg(rq, irq_delta + steal);
730#endif
731	update_rq_clock_pelt(rq, delta);
732}
733
734void update_rq_clock(struct rq *rq)
735{
736	s64 delta;
737
738	lockdep_assert_rq_held(rq);
739
740	if (rq->clock_update_flags & RQCF_ACT_SKIP)
741		return;
742
743#ifdef CONFIG_SCHED_DEBUG
744	if (sched_feat(WARN_DOUBLE_CLOCK))
745		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
746	rq->clock_update_flags |= RQCF_UPDATED;
747#endif
748
749	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
750	if (delta < 0)
751		return;
752	rq->clock += delta;
753	update_rq_clock_task(rq, delta);
754}
755
756#ifdef CONFIG_SCHED_HRTICK
757/*
758 * Use HR-timers to deliver accurate preemption points.
759 */
760
761static void hrtick_clear(struct rq *rq)
762{
763	if (hrtimer_active(&rq->hrtick_timer))
764		hrtimer_cancel(&rq->hrtick_timer);
765}
766
767/*
768 * High-resolution timer tick.
769 * Runs from hardirq context with interrupts disabled.
770 */
771static enum hrtimer_restart hrtick(struct hrtimer *timer)
772{
773	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
774	struct rq_flags rf;
775
776	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
777
778	rq_lock(rq, &rf);
779	update_rq_clock(rq);
780	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
781	rq_unlock(rq, &rf);
782
783	return HRTIMER_NORESTART;
784}
785
786#ifdef CONFIG_SMP
787
788static void __hrtick_restart(struct rq *rq)
789{
790	struct hrtimer *timer = &rq->hrtick_timer;
791	ktime_t time = rq->hrtick_time;
792
793	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
794}
795
796/*
797 * called from hardirq (IPI) context
798 */
799static void __hrtick_start(void *arg)
800{
801	struct rq *rq = arg;
802	struct rq_flags rf;
803
804	rq_lock(rq, &rf);
805	__hrtick_restart(rq);
806	rq_unlock(rq, &rf);
807}
808
809/*
810 * Called to set the hrtick timer state.
811 *
812 * called with rq->lock held and irqs disabled
813 */
814void hrtick_start(struct rq *rq, u64 delay)
815{
816	struct hrtimer *timer = &rq->hrtick_timer;
817	s64 delta;
818
819	/*
820	 * Don't schedule slices shorter than 10000ns, that just
821	 * doesn't make sense and can cause timer DoS.
822	 */
823	delta = max_t(s64, delay, 10000LL);
824	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
825
826	if (rq == this_rq())
827		__hrtick_restart(rq);
828	else
829		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
830}
831
832#else
833/*
834 * Called to set the hrtick timer state.
835 *
836 * called with rq->lock held and irqs disabled
837 */
838void hrtick_start(struct rq *rq, u64 delay)
839{
840	/*
841	 * Don't schedule slices shorter than 10000ns, that just
842	 * doesn't make sense. Rely on vruntime for fairness.
843	 */
844	delay = max_t(u64, delay, 10000LL);
845	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
846		      HRTIMER_MODE_REL_PINNED_HARD);
847}
848
849#endif /* CONFIG_SMP */
850
851static void hrtick_rq_init(struct rq *rq)
852{
853#ifdef CONFIG_SMP
854	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
855#endif
856	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
857	rq->hrtick_timer.function = hrtick;
858}
859#else	/* CONFIG_SCHED_HRTICK */
860static inline void hrtick_clear(struct rq *rq)
861{
862}
863
864static inline void hrtick_rq_init(struct rq *rq)
865{
866}
867#endif	/* CONFIG_SCHED_HRTICK */
868
869/*
870 * cmpxchg based fetch_or, macro so it works for different integer types
871 */
872#define fetch_or(ptr, mask)						\
873	({								\
874		typeof(ptr) _ptr = (ptr);				\
875		typeof(mask) _mask = (mask);				\
876		typeof(*_ptr) _val = *_ptr;				\
877									\
878		do {							\
879		} while (!try_cmpxchg(_ptr, &_val, _val | _mask));	\
880	_val;								\
881})
882
883#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
884/*
885 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
886 * this avoids any races wrt polling state changes and thereby avoids
887 * spurious IPIs.
888 */
889static inline bool set_nr_and_not_polling(struct task_struct *p)
890{
891	struct thread_info *ti = task_thread_info(p);
892	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
893}
894
895/*
896 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
897 *
898 * If this returns true, then the idle task promises to call
899 * sched_ttwu_pending() and reschedule soon.
900 */
901static bool set_nr_if_polling(struct task_struct *p)
902{
903	struct thread_info *ti = task_thread_info(p);
904	typeof(ti->flags) val = READ_ONCE(ti->flags);
905
906	for (;;) {
907		if (!(val & _TIF_POLLING_NRFLAG))
908			return false;
909		if (val & _TIF_NEED_RESCHED)
910			return true;
911		if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
912			break;
913	}
914	return true;
915}
916
917#else
918static inline bool set_nr_and_not_polling(struct task_struct *p)
919{
920	set_tsk_need_resched(p);
921	return true;
922}
923
924#ifdef CONFIG_SMP
925static inline bool set_nr_if_polling(struct task_struct *p)
926{
927	return false;
928}
929#endif
930#endif
931
932static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
933{
934	struct wake_q_node *node = &task->wake_q;
935
936	/*
937	 * Atomically grab the task, if ->wake_q is !nil already it means
938	 * it's already queued (either by us or someone else) and will get the
939	 * wakeup due to that.
940	 *
941	 * In order to ensure that a pending wakeup will observe our pending
942	 * state, even in the failed case, an explicit smp_mb() must be used.
943	 */
944	smp_mb__before_atomic();
945	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
946		return false;
947
948	/*
949	 * The head is context local, there can be no concurrency.
950	 */
951	*head->lastp = node;
952	head->lastp = &node->next;
953	return true;
954}
955
956/**
957 * wake_q_add() - queue a wakeup for 'later' waking.
958 * @head: the wake_q_head to add @task to
959 * @task: the task to queue for 'later' wakeup
960 *
961 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
962 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
963 * instantly.
964 *
965 * This function must be used as-if it were wake_up_process(); IOW the task
966 * must be ready to be woken at this location.
967 */
968void wake_q_add(struct wake_q_head *head, struct task_struct *task)
969{
970	if (__wake_q_add(head, task))
971		get_task_struct(task);
972}
973
974/**
975 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
976 * @head: the wake_q_head to add @task to
977 * @task: the task to queue for 'later' wakeup
978 *
979 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
980 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
981 * instantly.
982 *
983 * This function must be used as-if it were wake_up_process(); IOW the task
984 * must be ready to be woken at this location.
985 *
986 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
987 * that already hold reference to @task can call the 'safe' version and trust
988 * wake_q to do the right thing depending whether or not the @task is already
989 * queued for wakeup.
990 */
991void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
992{
993	if (!__wake_q_add(head, task))
994		put_task_struct(task);
995}
996
997void wake_up_q(struct wake_q_head *head)
998{
999	struct wake_q_node *node = head->first;
1000
1001	while (node != WAKE_Q_TAIL) {
1002		struct task_struct *task;
1003
1004		task = container_of(node, struct task_struct, wake_q);
1005		/* Task can safely be re-inserted now: */
1006		node = node->next;
1007		task->wake_q.next = NULL;
1008
1009		/*
1010		 * wake_up_process() executes a full barrier, which pairs with
1011		 * the queueing in wake_q_add() so as not to miss wakeups.
1012		 */
1013		wake_up_process(task);
1014		put_task_struct(task);
1015	}
1016}
1017
1018/*
1019 * resched_curr - mark rq's current task 'to be rescheduled now'.
1020 *
1021 * On UP this means the setting of the need_resched flag, on SMP it
1022 * might also involve a cross-CPU call to trigger the scheduler on
1023 * the target CPU.
1024 */
1025void resched_curr(struct rq *rq)
1026{
1027	struct task_struct *curr = rq->curr;
1028	int cpu;
1029
1030	lockdep_assert_rq_held(rq);
1031
1032	if (test_tsk_need_resched(curr))
1033		return;
1034
1035	cpu = cpu_of(rq);
1036
1037	if (cpu == smp_processor_id()) {
1038		set_tsk_need_resched(curr);
1039		set_preempt_need_resched();
1040		return;
1041	}
1042
1043	if (set_nr_and_not_polling(curr))
1044		smp_send_reschedule(cpu);
1045	else
1046		trace_sched_wake_idle_without_ipi(cpu);
1047}
1048
1049void resched_cpu(int cpu)
1050{
1051	struct rq *rq = cpu_rq(cpu);
1052	unsigned long flags;
1053
1054	raw_spin_rq_lock_irqsave(rq, flags);
1055	if (cpu_online(cpu) || cpu == smp_processor_id())
1056		resched_curr(rq);
1057	raw_spin_rq_unlock_irqrestore(rq, flags);
1058}
1059
1060#ifdef CONFIG_SMP
1061#ifdef CONFIG_NO_HZ_COMMON
1062/*
1063 * In the semi idle case, use the nearest busy CPU for migrating timers
1064 * from an idle CPU.  This is good for power-savings.
1065 *
1066 * We don't do similar optimization for completely idle system, as
1067 * selecting an idle CPU will add more delays to the timers than intended
1068 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1069 */
1070int get_nohz_timer_target(void)
1071{
1072	int i, cpu = smp_processor_id(), default_cpu = -1;
1073	struct sched_domain *sd;
1074	const struct cpumask *hk_mask;
1075
1076	if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1077		if (!idle_cpu(cpu))
1078			return cpu;
1079		default_cpu = cpu;
1080	}
1081
1082	hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1083
1084	rcu_read_lock();
1085	for_each_domain(cpu, sd) {
1086		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1087			if (cpu == i)
1088				continue;
1089
1090			if (!idle_cpu(i)) {
1091				cpu = i;
1092				goto unlock;
1093			}
1094		}
1095	}
1096
1097	if (default_cpu == -1)
1098		default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1099	cpu = default_cpu;
1100unlock:
1101	rcu_read_unlock();
1102	return cpu;
1103}
1104
1105/*
1106 * When add_timer_on() enqueues a timer into the timer wheel of an
1107 * idle CPU then this timer might expire before the next timer event
1108 * which is scheduled to wake up that CPU. In case of a completely
1109 * idle system the next event might even be infinite time into the
1110 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1111 * leaves the inner idle loop so the newly added timer is taken into
1112 * account when the CPU goes back to idle and evaluates the timer
1113 * wheel for the next timer event.
1114 */
1115static void wake_up_idle_cpu(int cpu)
1116{
1117	struct rq *rq = cpu_rq(cpu);
1118
1119	if (cpu == smp_processor_id())
1120		return;
1121
1122	if (set_nr_and_not_polling(rq->idle))
1123		smp_send_reschedule(cpu);
1124	else
1125		trace_sched_wake_idle_without_ipi(cpu);
1126}
1127
1128static bool wake_up_full_nohz_cpu(int cpu)
1129{
1130	/*
1131	 * We just need the target to call irq_exit() and re-evaluate
1132	 * the next tick. The nohz full kick at least implies that.
1133	 * If needed we can still optimize that later with an
1134	 * empty IRQ.
1135	 */
1136	if (cpu_is_offline(cpu))
1137		return true;  /* Don't try to wake offline CPUs. */
1138	if (tick_nohz_full_cpu(cpu)) {
1139		if (cpu != smp_processor_id() ||
1140		    tick_nohz_tick_stopped())
1141			tick_nohz_full_kick_cpu(cpu);
1142		return true;
1143	}
1144
1145	return false;
1146}
1147
1148/*
1149 * Wake up the specified CPU.  If the CPU is going offline, it is the
1150 * caller's responsibility to deal with the lost wakeup, for example,
1151 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1152 */
1153void wake_up_nohz_cpu(int cpu)
1154{
1155	if (!wake_up_full_nohz_cpu(cpu))
1156		wake_up_idle_cpu(cpu);
1157}
1158
1159static void nohz_csd_func(void *info)
1160{
1161	struct rq *rq = info;
1162	int cpu = cpu_of(rq);
1163	unsigned int flags;
1164
1165	/*
1166	 * Release the rq::nohz_csd.
1167	 */
1168	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1169	WARN_ON(!(flags & NOHZ_KICK_MASK));
1170
1171	rq->idle_balance = idle_cpu(cpu);
1172	if (rq->idle_balance && !need_resched()) {
1173		rq->nohz_idle_balance = flags;
1174		raise_softirq_irqoff(SCHED_SOFTIRQ);
1175	}
1176}
1177
1178#endif /* CONFIG_NO_HZ_COMMON */
1179
1180#ifdef CONFIG_NO_HZ_FULL
1181bool sched_can_stop_tick(struct rq *rq)
1182{
1183	int fifo_nr_running;
1184
1185	/* Deadline tasks, even if single, need the tick */
1186	if (rq->dl.dl_nr_running)
1187		return false;
1188
1189	/*
1190	 * If there are more than one RR tasks, we need the tick to affect the
1191	 * actual RR behaviour.
1192	 */
1193	if (rq->rt.rr_nr_running) {
1194		if (rq->rt.rr_nr_running == 1)
1195			return true;
1196		else
1197			return false;
1198	}
1199
1200	/*
1201	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1202	 * forced preemption between FIFO tasks.
1203	 */
1204	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1205	if (fifo_nr_running)
1206		return true;
1207
1208	/*
1209	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1210	 * if there's more than one we need the tick for involuntary
1211	 * preemption.
1212	 */
1213	if (rq->nr_running > 1)
1214		return false;
1215
1216	return true;
1217}
1218#endif /* CONFIG_NO_HZ_FULL */
1219#endif /* CONFIG_SMP */
1220
1221#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1222			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1223/*
1224 * Iterate task_group tree rooted at *from, calling @down when first entering a
1225 * node and @up when leaving it for the final time.
1226 *
1227 * Caller must hold rcu_lock or sufficient equivalent.
1228 */
1229int walk_tg_tree_from(struct task_group *from,
1230			     tg_visitor down, tg_visitor up, void *data)
1231{
1232	struct task_group *parent, *child;
1233	int ret;
1234
1235	parent = from;
1236
1237down:
1238	ret = (*down)(parent, data);
1239	if (ret)
1240		goto out;
1241	list_for_each_entry_rcu(child, &parent->children, siblings) {
1242		parent = child;
1243		goto down;
1244
1245up:
1246		continue;
1247	}
1248	ret = (*up)(parent, data);
1249	if (ret || parent == from)
1250		goto out;
1251
1252	child = parent;
1253	parent = parent->parent;
1254	if (parent)
1255		goto up;
1256out:
1257	return ret;
1258}
1259
1260int tg_nop(struct task_group *tg, void *data)
1261{
1262	return 0;
1263}
1264#endif
1265
1266static void set_load_weight(struct task_struct *p, bool update_load)
1267{
1268	int prio = p->static_prio - MAX_RT_PRIO;
1269	struct load_weight *load = &p->se.load;
1270
1271	/*
1272	 * SCHED_IDLE tasks get minimal weight:
1273	 */
1274	if (task_has_idle_policy(p)) {
1275		load->weight = scale_load(WEIGHT_IDLEPRIO);
1276		load->inv_weight = WMULT_IDLEPRIO;
1277		return;
1278	}
1279
1280	/*
1281	 * SCHED_OTHER tasks have to update their load when changing their
1282	 * weight
1283	 */
1284	if (update_load && p->sched_class == &fair_sched_class) {
1285		reweight_task(p, prio);
1286	} else {
1287		load->weight = scale_load(sched_prio_to_weight[prio]);
1288		load->inv_weight = sched_prio_to_wmult[prio];
1289	}
1290}
1291
1292#ifdef CONFIG_UCLAMP_TASK
1293/*
1294 * Serializes updates of utilization clamp values
1295 *
1296 * The (slow-path) user-space triggers utilization clamp value updates which
1297 * can require updates on (fast-path) scheduler's data structures used to
1298 * support enqueue/dequeue operations.
1299 * While the per-CPU rq lock protects fast-path update operations, user-space
1300 * requests are serialized using a mutex to reduce the risk of conflicting
1301 * updates or API abuses.
1302 */
1303static DEFINE_MUTEX(uclamp_mutex);
1304
1305/* Max allowed minimum utilization */
1306static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1307
1308/* Max allowed maximum utilization */
1309static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1310
1311/*
1312 * By default RT tasks run at the maximum performance point/capacity of the
1313 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1314 * SCHED_CAPACITY_SCALE.
1315 *
1316 * This knob allows admins to change the default behavior when uclamp is being
1317 * used. In battery powered devices, particularly, running at the maximum
1318 * capacity and frequency will increase energy consumption and shorten the
1319 * battery life.
1320 *
1321 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1322 *
1323 * This knob will not override the system default sched_util_clamp_min defined
1324 * above.
1325 */
1326static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1327
1328/* All clamps are required to be less or equal than these values */
1329static struct uclamp_se uclamp_default[UCLAMP_CNT];
1330
1331/*
1332 * This static key is used to reduce the uclamp overhead in the fast path. It
1333 * primarily disables the call to uclamp_rq_{inc, dec}() in
1334 * enqueue/dequeue_task().
1335 *
1336 * This allows users to continue to enable uclamp in their kernel config with
1337 * minimum uclamp overhead in the fast path.
1338 *
1339 * As soon as userspace modifies any of the uclamp knobs, the static key is
1340 * enabled, since we have an actual users that make use of uclamp
1341 * functionality.
1342 *
1343 * The knobs that would enable this static key are:
1344 *
1345 *   * A task modifying its uclamp value with sched_setattr().
1346 *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1347 *   * An admin modifying the cgroup cpu.uclamp.{min, max}
1348 */
1349DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1350
1351/* Integer rounded range for each bucket */
1352#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1353
1354#define for_each_clamp_id(clamp_id) \
1355	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1356
1357static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1358{
1359	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1360}
1361
1362static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1363{
1364	if (clamp_id == UCLAMP_MIN)
1365		return 0;
1366	return SCHED_CAPACITY_SCALE;
1367}
1368
1369static inline void uclamp_se_set(struct uclamp_se *uc_se,
1370				 unsigned int value, bool user_defined)
1371{
1372	uc_se->value = value;
1373	uc_se->bucket_id = uclamp_bucket_id(value);
1374	uc_se->user_defined = user_defined;
1375}
1376
1377static inline unsigned int
1378uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1379		  unsigned int clamp_value)
1380{
1381	/*
1382	 * Avoid blocked utilization pushing up the frequency when we go
1383	 * idle (which drops the max-clamp) by retaining the last known
1384	 * max-clamp.
1385	 */
1386	if (clamp_id == UCLAMP_MAX) {
1387		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1388		return clamp_value;
1389	}
1390
1391	return uclamp_none(UCLAMP_MIN);
1392}
1393
1394static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1395				     unsigned int clamp_value)
1396{
1397	/* Reset max-clamp retention only on idle exit */
1398	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1399		return;
1400
1401	WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1402}
1403
1404static inline
1405unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1406				   unsigned int clamp_value)
1407{
1408	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1409	int bucket_id = UCLAMP_BUCKETS - 1;
1410
1411	/*
1412	 * Since both min and max clamps are max aggregated, find the
1413	 * top most bucket with tasks in.
1414	 */
1415	for ( ; bucket_id >= 0; bucket_id--) {
1416		if (!bucket[bucket_id].tasks)
1417			continue;
1418		return bucket[bucket_id].value;
1419	}
1420
1421	/* No tasks -- default clamp values */
1422	return uclamp_idle_value(rq, clamp_id, clamp_value);
1423}
1424
1425static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1426{
1427	unsigned int default_util_min;
1428	struct uclamp_se *uc_se;
1429
1430	lockdep_assert_held(&p->pi_lock);
1431
1432	uc_se = &p->uclamp_req[UCLAMP_MIN];
1433
1434	/* Only sync if user didn't override the default */
1435	if (uc_se->user_defined)
1436		return;
1437
1438	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1439	uclamp_se_set(uc_se, default_util_min, false);
1440}
1441
1442static void uclamp_update_util_min_rt_default(struct task_struct *p)
1443{
1444	struct rq_flags rf;
1445	struct rq *rq;
1446
1447	if (!rt_task(p))
1448		return;
1449
1450	/* Protect updates to p->uclamp_* */
1451	rq = task_rq_lock(p, &rf);
1452	__uclamp_update_util_min_rt_default(p);
1453	task_rq_unlock(rq, p, &rf);
1454}
1455
1456static inline struct uclamp_se
1457uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1458{
1459	/* Copy by value as we could modify it */
1460	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1461#ifdef CONFIG_UCLAMP_TASK_GROUP
1462	unsigned int tg_min, tg_max, value;
1463
1464	/*
1465	 * Tasks in autogroups or root task group will be
1466	 * restricted by system defaults.
1467	 */
1468	if (task_group_is_autogroup(task_group(p)))
1469		return uc_req;
1470	if (task_group(p) == &root_task_group)
1471		return uc_req;
1472
1473	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1474	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1475	value = uc_req.value;
1476	value = clamp(value, tg_min, tg_max);
1477	uclamp_se_set(&uc_req, value, false);
1478#endif
1479
1480	return uc_req;
1481}
1482
1483/*
1484 * The effective clamp bucket index of a task depends on, by increasing
1485 * priority:
1486 * - the task specific clamp value, when explicitly requested from userspace
1487 * - the task group effective clamp value, for tasks not either in the root
1488 *   group or in an autogroup
1489 * - the system default clamp value, defined by the sysadmin
1490 */
1491static inline struct uclamp_se
1492uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1493{
1494	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1495	struct uclamp_se uc_max = uclamp_default[clamp_id];
1496
1497	/* System default restrictions always apply */
1498	if (unlikely(uc_req.value > uc_max.value))
1499		return uc_max;
1500
1501	return uc_req;
1502}
1503
1504unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1505{
1506	struct uclamp_se uc_eff;
1507
1508	/* Task currently refcounted: use back-annotated (effective) value */
1509	if (p->uclamp[clamp_id].active)
1510		return (unsigned long)p->uclamp[clamp_id].value;
1511
1512	uc_eff = uclamp_eff_get(p, clamp_id);
1513
1514	return (unsigned long)uc_eff.value;
1515}
1516
1517/*
1518 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1519 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1520 * updates the rq's clamp value if required.
1521 *
1522 * Tasks can have a task-specific value requested from user-space, track
1523 * within each bucket the maximum value for tasks refcounted in it.
1524 * This "local max aggregation" allows to track the exact "requested" value
1525 * for each bucket when all its RUNNABLE tasks require the same clamp.
1526 */
1527static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1528				    enum uclamp_id clamp_id)
1529{
1530	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1531	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1532	struct uclamp_bucket *bucket;
1533
1534	lockdep_assert_rq_held(rq);
1535
1536	/* Update task effective clamp */
1537	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1538
1539	bucket = &uc_rq->bucket[uc_se->bucket_id];
1540	bucket->tasks++;
1541	uc_se->active = true;
1542
1543	uclamp_idle_reset(rq, clamp_id, uc_se->value);
1544
1545	/*
1546	 * Local max aggregation: rq buckets always track the max
1547	 * "requested" clamp value of its RUNNABLE tasks.
1548	 */
1549	if (bucket->tasks == 1 || uc_se->value > bucket->value)
1550		bucket->value = uc_se->value;
1551
1552	if (uc_se->value > READ_ONCE(uc_rq->value))
1553		WRITE_ONCE(uc_rq->value, uc_se->value);
1554}
1555
1556/*
1557 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1558 * is released. If this is the last task reference counting the rq's max
1559 * active clamp value, then the rq's clamp value is updated.
1560 *
1561 * Both refcounted tasks and rq's cached clamp values are expected to be
1562 * always valid. If it's detected they are not, as defensive programming,
1563 * enforce the expected state and warn.
1564 */
1565static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1566				    enum uclamp_id clamp_id)
1567{
1568	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1569	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1570	struct uclamp_bucket *bucket;
1571	unsigned int bkt_clamp;
1572	unsigned int rq_clamp;
1573
1574	lockdep_assert_rq_held(rq);
1575
1576	/*
1577	 * If sched_uclamp_used was enabled after task @p was enqueued,
1578	 * we could end up with unbalanced call to uclamp_rq_dec_id().
1579	 *
1580	 * In this case the uc_se->active flag should be false since no uclamp
1581	 * accounting was performed at enqueue time and we can just return
1582	 * here.
1583	 *
1584	 * Need to be careful of the following enqueue/dequeue ordering
1585	 * problem too
1586	 *
1587	 *	enqueue(taskA)
1588	 *	// sched_uclamp_used gets enabled
1589	 *	enqueue(taskB)
1590	 *	dequeue(taskA)
1591	 *	// Must not decrement bucket->tasks here
1592	 *	dequeue(taskB)
1593	 *
1594	 * where we could end up with stale data in uc_se and
1595	 * bucket[uc_se->bucket_id].
1596	 *
1597	 * The following check here eliminates the possibility of such race.
1598	 */
1599	if (unlikely(!uc_se->active))
1600		return;
1601
1602	bucket = &uc_rq->bucket[uc_se->bucket_id];
1603
1604	SCHED_WARN_ON(!bucket->tasks);
1605	if (likely(bucket->tasks))
1606		bucket->tasks--;
1607
1608	uc_se->active = false;
1609
1610	/*
1611	 * Keep "local max aggregation" simple and accept to (possibly)
1612	 * overboost some RUNNABLE tasks in the same bucket.
1613	 * The rq clamp bucket value is reset to its base value whenever
1614	 * there are no more RUNNABLE tasks refcounting it.
1615	 */
1616	if (likely(bucket->tasks))
1617		return;
1618
1619	rq_clamp = READ_ONCE(uc_rq->value);
1620	/*
1621	 * Defensive programming: this should never happen. If it happens,
1622	 * e.g. due to future modification, warn and fixup the expected value.
1623	 */
1624	SCHED_WARN_ON(bucket->value > rq_clamp);
1625	if (bucket->value >= rq_clamp) {
1626		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1627		WRITE_ONCE(uc_rq->value, bkt_clamp);
1628	}
1629}
1630
1631static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1632{
1633	enum uclamp_id clamp_id;
1634
1635	/*
1636	 * Avoid any overhead until uclamp is actually used by the userspace.
1637	 *
1638	 * The condition is constructed such that a NOP is generated when
1639	 * sched_uclamp_used is disabled.
1640	 */
1641	if (!static_branch_unlikely(&sched_uclamp_used))
1642		return;
1643
1644	if (unlikely(!p->sched_class->uclamp_enabled))
1645		return;
1646
1647	for_each_clamp_id(clamp_id)
1648		uclamp_rq_inc_id(rq, p, clamp_id);
1649
1650	/* Reset clamp idle holding when there is one RUNNABLE task */
1651	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1652		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1653}
1654
1655static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1656{
1657	enum uclamp_id clamp_id;
1658
1659	/*
1660	 * Avoid any overhead until uclamp is actually used by the userspace.
1661	 *
1662	 * The condition is constructed such that a NOP is generated when
1663	 * sched_uclamp_used is disabled.
1664	 */
1665	if (!static_branch_unlikely(&sched_uclamp_used))
1666		return;
1667
1668	if (unlikely(!p->sched_class->uclamp_enabled))
1669		return;
1670
1671	for_each_clamp_id(clamp_id)
1672		uclamp_rq_dec_id(rq, p, clamp_id);
1673}
1674
1675static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1676				      enum uclamp_id clamp_id)
1677{
1678	if (!p->uclamp[clamp_id].active)
1679		return;
1680
1681	uclamp_rq_dec_id(rq, p, clamp_id);
1682	uclamp_rq_inc_id(rq, p, clamp_id);
1683
1684	/*
1685	 * Make sure to clear the idle flag if we've transiently reached 0
1686	 * active tasks on rq.
1687	 */
1688	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1689		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1690}
1691
1692static inline void
1693uclamp_update_active(struct task_struct *p)
1694{
1695	enum uclamp_id clamp_id;
1696	struct rq_flags rf;
1697	struct rq *rq;
1698
1699	/*
1700	 * Lock the task and the rq where the task is (or was) queued.
1701	 *
1702	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1703	 * price to pay to safely serialize util_{min,max} updates with
1704	 * enqueues, dequeues and migration operations.
1705	 * This is the same locking schema used by __set_cpus_allowed_ptr().
1706	 */
1707	rq = task_rq_lock(p, &rf);
1708
1709	/*
1710	 * Setting the clamp bucket is serialized by task_rq_lock().
1711	 * If the task is not yet RUNNABLE and its task_struct is not
1712	 * affecting a valid clamp bucket, the next time it's enqueued,
1713	 * it will already see the updated clamp bucket value.
1714	 */
1715	for_each_clamp_id(clamp_id)
1716		uclamp_rq_reinc_id(rq, p, clamp_id);
1717
1718	task_rq_unlock(rq, p, &rf);
1719}
1720
1721#ifdef CONFIG_UCLAMP_TASK_GROUP
1722static inline void
1723uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1724{
1725	struct css_task_iter it;
1726	struct task_struct *p;
1727
1728	css_task_iter_start(css, 0, &it);
1729	while ((p = css_task_iter_next(&it)))
1730		uclamp_update_active(p);
1731	css_task_iter_end(&it);
1732}
1733
1734static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1735#endif
1736
1737#ifdef CONFIG_SYSCTL
1738#ifdef CONFIG_UCLAMP_TASK
1739#ifdef CONFIG_UCLAMP_TASK_GROUP
1740static void uclamp_update_root_tg(void)
1741{
1742	struct task_group *tg = &root_task_group;
1743
1744	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1745		      sysctl_sched_uclamp_util_min, false);
1746	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1747		      sysctl_sched_uclamp_util_max, false);
1748
1749	rcu_read_lock();
1750	cpu_util_update_eff(&root_task_group.css);
1751	rcu_read_unlock();
1752}
1753#else
1754static void uclamp_update_root_tg(void) { }
1755#endif
1756
1757static void uclamp_sync_util_min_rt_default(void)
1758{
1759	struct task_struct *g, *p;
1760
1761	/*
1762	 * copy_process()			sysctl_uclamp
1763	 *					  uclamp_min_rt = X;
1764	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
1765	 *   // link thread			  smp_mb__after_spinlock()
1766	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
1767	 *   sched_post_fork()			  for_each_process_thread()
1768	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
1769	 *
1770	 * Ensures that either sched_post_fork() will observe the new
1771	 * uclamp_min_rt or for_each_process_thread() will observe the new
1772	 * task.
1773	 */
1774	read_lock(&tasklist_lock);
1775	smp_mb__after_spinlock();
1776	read_unlock(&tasklist_lock);
1777
1778	rcu_read_lock();
1779	for_each_process_thread(g, p)
1780		uclamp_update_util_min_rt_default(p);
1781	rcu_read_unlock();
1782}
1783
1784static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1785				void *buffer, size_t *lenp, loff_t *ppos)
1786{
1787	bool update_root_tg = false;
1788	int old_min, old_max, old_min_rt;
1789	int result;
1790
1791	mutex_lock(&uclamp_mutex);
1792	old_min = sysctl_sched_uclamp_util_min;
1793	old_max = sysctl_sched_uclamp_util_max;
1794	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1795
1796	result = proc_dointvec(table, write, buffer, lenp, ppos);
1797	if (result)
1798		goto undo;
1799	if (!write)
1800		goto done;
1801
1802	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1803	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
1804	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1805
1806		result = -EINVAL;
1807		goto undo;
1808	}
1809
1810	if (old_min != sysctl_sched_uclamp_util_min) {
1811		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1812			      sysctl_sched_uclamp_util_min, false);
1813		update_root_tg = true;
1814	}
1815	if (old_max != sysctl_sched_uclamp_util_max) {
1816		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1817			      sysctl_sched_uclamp_util_max, false);
1818		update_root_tg = true;
1819	}
1820
1821	if (update_root_tg) {
1822		static_branch_enable(&sched_uclamp_used);
1823		uclamp_update_root_tg();
1824	}
1825
1826	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1827		static_branch_enable(&sched_uclamp_used);
1828		uclamp_sync_util_min_rt_default();
1829	}
1830
1831	/*
1832	 * We update all RUNNABLE tasks only when task groups are in use.
1833	 * Otherwise, keep it simple and do just a lazy update at each next
1834	 * task enqueue time.
1835	 */
1836
1837	goto done;
1838
1839undo:
1840	sysctl_sched_uclamp_util_min = old_min;
1841	sysctl_sched_uclamp_util_max = old_max;
1842	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1843done:
1844	mutex_unlock(&uclamp_mutex);
1845
1846	return result;
1847}
1848#endif
1849#endif
1850
1851static int uclamp_validate(struct task_struct *p,
1852			   const struct sched_attr *attr)
1853{
1854	int util_min = p->uclamp_req[UCLAMP_MIN].value;
1855	int util_max = p->uclamp_req[UCLAMP_MAX].value;
1856
1857	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1858		util_min = attr->sched_util_min;
1859
1860		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1861			return -EINVAL;
1862	}
1863
1864	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1865		util_max = attr->sched_util_max;
1866
1867		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1868			return -EINVAL;
1869	}
1870
1871	if (util_min != -1 && util_max != -1 && util_min > util_max)
1872		return -EINVAL;
1873
1874	/*
1875	 * We have valid uclamp attributes; make sure uclamp is enabled.
1876	 *
1877	 * We need to do that here, because enabling static branches is a
1878	 * blocking operation which obviously cannot be done while holding
1879	 * scheduler locks.
1880	 */
1881	static_branch_enable(&sched_uclamp_used);
1882
1883	return 0;
1884}
1885
1886static bool uclamp_reset(const struct sched_attr *attr,
1887			 enum uclamp_id clamp_id,
1888			 struct uclamp_se *uc_se)
1889{
1890	/* Reset on sched class change for a non user-defined clamp value. */
1891	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1892	    !uc_se->user_defined)
1893		return true;
1894
1895	/* Reset on sched_util_{min,max} == -1. */
1896	if (clamp_id == UCLAMP_MIN &&
1897	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1898	    attr->sched_util_min == -1) {
1899		return true;
1900	}
1901
1902	if (clamp_id == UCLAMP_MAX &&
1903	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1904	    attr->sched_util_max == -1) {
1905		return true;
1906	}
1907
1908	return false;
1909}
1910
1911static void __setscheduler_uclamp(struct task_struct *p,
1912				  const struct sched_attr *attr)
1913{
1914	enum uclamp_id clamp_id;
1915
1916	for_each_clamp_id(clamp_id) {
1917		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1918		unsigned int value;
1919
1920		if (!uclamp_reset(attr, clamp_id, uc_se))
1921			continue;
1922
1923		/*
1924		 * RT by default have a 100% boost value that could be modified
1925		 * at runtime.
1926		 */
1927		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1928			value = sysctl_sched_uclamp_util_min_rt_default;
1929		else
1930			value = uclamp_none(clamp_id);
1931
1932		uclamp_se_set(uc_se, value, false);
1933
1934	}
1935
1936	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1937		return;
1938
1939	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1940	    attr->sched_util_min != -1) {
1941		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1942			      attr->sched_util_min, true);
1943	}
1944
1945	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1946	    attr->sched_util_max != -1) {
1947		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1948			      attr->sched_util_max, true);
1949	}
1950}
1951
1952static void uclamp_fork(struct task_struct *p)
1953{
1954	enum uclamp_id clamp_id;
1955
1956	/*
1957	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1958	 * as the task is still at its early fork stages.
1959	 */
1960	for_each_clamp_id(clamp_id)
1961		p->uclamp[clamp_id].active = false;
1962
1963	if (likely(!p->sched_reset_on_fork))
1964		return;
1965
1966	for_each_clamp_id(clamp_id) {
1967		uclamp_se_set(&p->uclamp_req[clamp_id],
1968			      uclamp_none(clamp_id), false);
1969	}
1970}
1971
1972static void uclamp_post_fork(struct task_struct *p)
1973{
1974	uclamp_update_util_min_rt_default(p);
1975}
1976
1977static void __init init_uclamp_rq(struct rq *rq)
1978{
1979	enum uclamp_id clamp_id;
1980	struct uclamp_rq *uc_rq = rq->uclamp;
1981
1982	for_each_clamp_id(clamp_id) {
1983		uc_rq[clamp_id] = (struct uclamp_rq) {
1984			.value = uclamp_none(clamp_id)
1985		};
1986	}
1987
1988	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1989}
1990
1991static void __init init_uclamp(void)
1992{
1993	struct uclamp_se uc_max = {};
1994	enum uclamp_id clamp_id;
1995	int cpu;
1996
1997	for_each_possible_cpu(cpu)
1998		init_uclamp_rq(cpu_rq(cpu));
1999
2000	for_each_clamp_id(clamp_id) {
2001		uclamp_se_set(&init_task.uclamp_req[clamp_id],
2002			      uclamp_none(clamp_id), false);
2003	}
2004
2005	/* System defaults allow max clamp values for both indexes */
2006	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2007	for_each_clamp_id(clamp_id) {
2008		uclamp_default[clamp_id] = uc_max;
2009#ifdef CONFIG_UCLAMP_TASK_GROUP
2010		root_task_group.uclamp_req[clamp_id] = uc_max;
2011		root_task_group.uclamp[clamp_id] = uc_max;
2012#endif
2013	}
2014}
2015
2016#else /* CONFIG_UCLAMP_TASK */
2017static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2018static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2019static inline int uclamp_validate(struct task_struct *p,
2020				  const struct sched_attr *attr)
2021{
2022	return -EOPNOTSUPP;
2023}
2024static void __setscheduler_uclamp(struct task_struct *p,
2025				  const struct sched_attr *attr) { }
2026static inline void uclamp_fork(struct task_struct *p) { }
2027static inline void uclamp_post_fork(struct task_struct *p) { }
2028static inline void init_uclamp(void) { }
2029#endif /* CONFIG_UCLAMP_TASK */
2030
2031bool sched_task_on_rq(struct task_struct *p)
2032{
2033	return task_on_rq_queued(p);
2034}
2035
2036unsigned long get_wchan(struct task_struct *p)
2037{
2038	unsigned long ip = 0;
2039	unsigned int state;
2040
2041	if (!p || p == current)
2042		return 0;
2043
2044	/* Only get wchan if task is blocked and we can keep it that way. */
2045	raw_spin_lock_irq(&p->pi_lock);
2046	state = READ_ONCE(p->__state);
2047	smp_rmb(); /* see try_to_wake_up() */
2048	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2049		ip = __get_wchan(p);
2050	raw_spin_unlock_irq(&p->pi_lock);
2051
2052	return ip;
2053}
2054
2055static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2056{
2057	if (!(flags & ENQUEUE_NOCLOCK))
2058		update_rq_clock(rq);
2059
2060	if (!(flags & ENQUEUE_RESTORE)) {
2061		sched_info_enqueue(rq, p);
2062		psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2063	}
2064
2065	uclamp_rq_inc(rq, p);
2066	p->sched_class->enqueue_task(rq, p, flags);
2067
2068	if (sched_core_enabled(rq))
2069		sched_core_enqueue(rq, p);
2070}
2071
2072static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2073{
2074	if (sched_core_enabled(rq))
2075		sched_core_dequeue(rq, p, flags);
2076
2077	if (!(flags & DEQUEUE_NOCLOCK))
2078		update_rq_clock(rq);
2079
2080	if (!(flags & DEQUEUE_SAVE)) {
2081		sched_info_dequeue(rq, p);
2082		psi_dequeue(p, flags & DEQUEUE_SLEEP);
2083	}
2084
2085	uclamp_rq_dec(rq, p);
2086	p->sched_class->dequeue_task(rq, p, flags);
2087}
2088
2089void activate_task(struct rq *rq, struct task_struct *p, int flags)
2090{
2091	enqueue_task(rq, p, flags);
2092
2093	p->on_rq = TASK_ON_RQ_QUEUED;
2094}
2095
2096void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2097{
2098	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2099
2100	dequeue_task(rq, p, flags);
2101}
2102
2103static inline int __normal_prio(int policy, int rt_prio, int nice)
2104{
2105	int prio;
2106
2107	if (dl_policy(policy))
2108		prio = MAX_DL_PRIO - 1;
2109	else if (rt_policy(policy))
2110		prio = MAX_RT_PRIO - 1 - rt_prio;
2111	else
2112		prio = NICE_TO_PRIO(nice);
2113
2114	return prio;
2115}
2116
2117/*
2118 * Calculate the expected normal priority: i.e. priority
2119 * without taking RT-inheritance into account. Might be
2120 * boosted by interactivity modifiers. Changes upon fork,
2121 * setprio syscalls, and whenever the interactivity
2122 * estimator recalculates.
2123 */
2124static inline int normal_prio(struct task_struct *p)
2125{
2126	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2127}
2128
2129/*
2130 * Calculate the current priority, i.e. the priority
2131 * taken into account by the scheduler. This value might
2132 * be boosted by RT tasks, or might be boosted by
2133 * interactivity modifiers. Will be RT if the task got
2134 * RT-boosted. If not then it returns p->normal_prio.
2135 */
2136static int effective_prio(struct task_struct *p)
2137{
2138	p->normal_prio = normal_prio(p);
2139	/*
2140	 * If we are RT tasks or we were boosted to RT priority,
2141	 * keep the priority unchanged. Otherwise, update priority
2142	 * to the normal priority:
2143	 */
2144	if (!rt_prio(p->prio))
2145		return p->normal_prio;
2146	return p->prio;
2147}
2148
2149/**
2150 * task_curr - is this task currently executing on a CPU?
2151 * @p: the task in question.
2152 *
2153 * Return: 1 if the task is currently executing. 0 otherwise.
2154 */
2155inline int task_curr(const struct task_struct *p)
2156{
2157	return cpu_curr(task_cpu(p)) == p;
2158}
2159
2160/*
2161 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2162 * use the balance_callback list if you want balancing.
2163 *
2164 * this means any call to check_class_changed() must be followed by a call to
2165 * balance_callback().
2166 */
2167static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2168				       const struct sched_class *prev_class,
2169				       int oldprio)
2170{
2171	if (prev_class != p->sched_class) {
2172		if (prev_class->switched_from)
2173			prev_class->switched_from(rq, p);
2174
2175		p->sched_class->switched_to(rq, p);
2176	} else if (oldprio != p->prio || dl_task(p))
2177		p->sched_class->prio_changed(rq, p, oldprio);
2178}
2179
2180void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2181{
2182	if (p->sched_class == rq->curr->sched_class)
2183		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2184	else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2185		resched_curr(rq);
2186
2187	/*
2188	 * A queue event has occurred, and we're going to schedule.  In
2189	 * this case, we can save a useless back to back clock update.
2190	 */
2191	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2192		rq_clock_skip_update(rq);
2193}
2194
2195#ifdef CONFIG_SMP
2196
2197static void
2198__do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2199
2200static int __set_cpus_allowed_ptr(struct task_struct *p,
2201				  const struct cpumask *new_mask,
2202				  u32 flags);
2203
2204static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2205{
2206	if (likely(!p->migration_disabled))
2207		return;
2208
2209	if (p->cpus_ptr != &p->cpus_mask)
2210		return;
2211
2212	/*
2213	 * Violates locking rules! see comment in __do_set_cpus_allowed().
2214	 */
2215	__do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2216}
2217
2218void migrate_disable(void)
2219{
2220	struct task_struct *p = current;
2221
2222	if (p->migration_disabled) {
2223		p->migration_disabled++;
2224		return;
2225	}
2226
2227	preempt_disable();
2228	this_rq()->nr_pinned++;
2229	p->migration_disabled = 1;
2230	preempt_enable();
2231}
2232EXPORT_SYMBOL_GPL(migrate_disable);
2233
2234void migrate_enable(void)
2235{
2236	struct task_struct *p = current;
2237
2238	if (p->migration_disabled > 1) {
2239		p->migration_disabled--;
2240		return;
2241	}
2242
2243	if (WARN_ON_ONCE(!p->migration_disabled))
2244		return;
2245
2246	/*
2247	 * Ensure stop_task runs either before or after this, and that
2248	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2249	 */
2250	preempt_disable();
2251	if (p->cpus_ptr != &p->cpus_mask)
2252		__set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2253	/*
2254	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2255	 * regular cpus_mask, otherwise things that race (eg.
2256	 * select_fallback_rq) get confused.
2257	 */
2258	barrier();
2259	p->migration_disabled = 0;
2260	this_rq()->nr_pinned--;
2261	preempt_enable();
2262}
2263EXPORT_SYMBOL_GPL(migrate_enable);
2264
2265static inline bool rq_has_pinned_tasks(struct rq *rq)
2266{
2267	return rq->nr_pinned;
2268}
2269
2270/*
2271 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2272 * __set_cpus_allowed_ptr() and select_fallback_rq().
2273 */
2274static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2275{
2276	/* When not in the task's cpumask, no point in looking further. */
2277	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2278		return false;
2279
2280	/* migrate_disabled() must be allowed to finish. */
2281	if (is_migration_disabled(p))
2282		return cpu_online(cpu);
2283
2284	/* Non kernel threads are not allowed during either online or offline. */
2285	if (!(p->flags & PF_KTHREAD))
2286		return cpu_active(cpu) && task_cpu_possible(cpu, p);
2287
2288	/* KTHREAD_IS_PER_CPU is always allowed. */
2289	if (kthread_is_per_cpu(p))
2290		return cpu_online(cpu);
2291
2292	/* Regular kernel threads don't get to stay during offline. */
2293	if (cpu_dying(cpu))
2294		return false;
2295
2296	/* But are allowed during online. */
2297	return cpu_online(cpu);
2298}
2299
2300/*
2301 * This is how migration works:
2302 *
2303 * 1) we invoke migration_cpu_stop() on the target CPU using
2304 *    stop_one_cpu().
2305 * 2) stopper starts to run (implicitly forcing the migrated thread
2306 *    off the CPU)
2307 * 3) it checks whether the migrated task is still in the wrong runqueue.
2308 * 4) if it's in the wrong runqueue then the migration thread removes
2309 *    it and puts it into the right queue.
2310 * 5) stopper completes and stop_one_cpu() returns and the migration
2311 *    is done.
2312 */
2313
2314/*
2315 * move_queued_task - move a queued task to new rq.
2316 *
2317 * Returns (locked) new rq. Old rq's lock is released.
2318 */
2319static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2320				   struct task_struct *p, int new_cpu)
2321{
2322	lockdep_assert_rq_held(rq);
2323
2324	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2325	set_task_cpu(p, new_cpu);
2326	rq_unlock(rq, rf);
2327
2328	rq = cpu_rq(new_cpu);
2329
2330	rq_lock(rq, rf);
2331	BUG_ON(task_cpu(p) != new_cpu);
2332	activate_task(rq, p, 0);
2333	check_preempt_curr(rq, p, 0);
2334
2335	return rq;
2336}
2337
2338struct migration_arg {
2339	struct task_struct		*task;
2340	int				dest_cpu;
2341	struct set_affinity_pending	*pending;
2342};
2343
2344/*
2345 * @refs: number of wait_for_completion()
2346 * @stop_pending: is @stop_work in use
2347 */
2348struct set_affinity_pending {
2349	refcount_t		refs;
2350	unsigned int		stop_pending;
2351	struct completion	done;
2352	struct cpu_stop_work	stop_work;
2353	struct migration_arg	arg;
2354};
2355
2356/*
2357 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2358 * this because either it can't run here any more (set_cpus_allowed()
2359 * away from this CPU, or CPU going down), or because we're
2360 * attempting to rebalance this task on exec (sched_exec).
2361 *
2362 * So we race with normal scheduler movements, but that's OK, as long
2363 * as the task is no longer on this CPU.
2364 */
2365static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2366				 struct task_struct *p, int dest_cpu)
2367{
2368	/* Affinity changed (again). */
2369	if (!is_cpu_allowed(p, dest_cpu))
2370		return rq;
2371
2372	update_rq_clock(rq);
2373	rq = move_queued_task(rq, rf, p, dest_cpu);
2374
2375	return rq;
2376}
2377
2378/*
2379 * migration_cpu_stop - this will be executed by a highprio stopper thread
2380 * and performs thread migration by bumping thread off CPU then
2381 * 'pushing' onto another runqueue.
2382 */
2383static int migration_cpu_stop(void *data)
2384{
2385	struct migration_arg *arg = data;
2386	struct set_affinity_pending *pending = arg->pending;
2387	struct task_struct *p = arg->task;
2388	struct rq *rq = this_rq();
2389	bool complete = false;
2390	struct rq_flags rf;
2391
2392	/*
2393	 * The original target CPU might have gone down and we might
2394	 * be on another CPU but it doesn't matter.
2395	 */
2396	local_irq_save(rf.flags);
2397	/*
2398	 * We need to explicitly wake pending tasks before running
2399	 * __migrate_task() such that we will not miss enforcing cpus_ptr
2400	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2401	 */
2402	flush_smp_call_function_queue();
2403
2404	raw_spin_lock(&p->pi_lock);
2405	rq_lock(rq, &rf);
2406
2407	/*
2408	 * If we were passed a pending, then ->stop_pending was set, thus
2409	 * p->migration_pending must have remained stable.
2410	 */
2411	WARN_ON_ONCE(pending && pending != p->migration_pending);
2412
2413	/*
2414	 * If task_rq(p) != rq, it cannot be migrated here, because we're
2415	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2416	 * we're holding p->pi_lock.
2417	 */
2418	if (task_rq(p) == rq) {
2419		if (is_migration_disabled(p))
2420			goto out;
2421
2422		if (pending) {
2423			p->migration_pending = NULL;
2424			complete = true;
2425
2426			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2427				goto out;
2428		}
2429
2430		if (task_on_rq_queued(p))
2431			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2432		else
2433			p->wake_cpu = arg->dest_cpu;
2434
2435		/*
2436		 * XXX __migrate_task() can fail, at which point we might end
2437		 * up running on a dodgy CPU, AFAICT this can only happen
2438		 * during CPU hotplug, at which point we'll get pushed out
2439		 * anyway, so it's probably not a big deal.
2440		 */
2441
2442	} else if (pending) {
2443		/*
2444		 * This happens when we get migrated between migrate_enable()'s
2445		 * preempt_enable() and scheduling the stopper task. At that
2446		 * point we're a regular task again and not current anymore.
2447		 *
2448		 * A !PREEMPT kernel has a giant hole here, which makes it far
2449		 * more likely.
2450		 */
2451
2452		/*
2453		 * The task moved before the stopper got to run. We're holding
2454		 * ->pi_lock, so the allowed mask is stable - if it got
2455		 * somewhere allowed, we're done.
2456		 */
2457		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2458			p->migration_pending = NULL;
2459			complete = true;
2460			goto out;
2461		}
2462
2463		/*
2464		 * When migrate_enable() hits a rq mis-match we can't reliably
2465		 * determine is_migration_disabled() and so have to chase after
2466		 * it.
2467		 */
2468		WARN_ON_ONCE(!pending->stop_pending);
2469		task_rq_unlock(rq, p, &rf);
2470		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2471				    &pending->arg, &pending->stop_work);
2472		return 0;
2473	}
2474out:
2475	if (pending)
2476		pending->stop_pending = false;
2477	task_rq_unlock(rq, p, &rf);
2478
2479	if (complete)
2480		complete_all(&pending->done);
2481
2482	return 0;
2483}
2484
2485int push_cpu_stop(void *arg)
2486{
2487	struct rq *lowest_rq = NULL, *rq = this_rq();
2488	struct task_struct *p = arg;
2489
2490	raw_spin_lock_irq(&p->pi_lock);
2491	raw_spin_rq_lock(rq);
2492
2493	if (task_rq(p) != rq)
2494		goto out_unlock;
2495
2496	if (is_migration_disabled(p)) {
2497		p->migration_flags |= MDF_PUSH;
2498		goto out_unlock;
2499	}
2500
2501	p->migration_flags &= ~MDF_PUSH;
2502
2503	if (p->sched_class->find_lock_rq)
2504		lowest_rq = p->sched_class->find_lock_rq(p, rq);
2505
2506	if (!lowest_rq)
2507		goto out_unlock;
2508
2509	// XXX validate p is still the highest prio task
2510	if (task_rq(p) == rq) {
2511		deactivate_task(rq, p, 0);
2512		set_task_cpu(p, lowest_rq->cpu);
2513		activate_task(lowest_rq, p, 0);
2514		resched_curr(lowest_rq);
2515	}
2516
2517	double_unlock_balance(rq, lowest_rq);
2518
2519out_unlock:
2520	rq->push_busy = false;
2521	raw_spin_rq_unlock(rq);
2522	raw_spin_unlock_irq(&p->pi_lock);
2523
2524	put_task_struct(p);
2525	return 0;
2526}
2527
2528/*
2529 * sched_class::set_cpus_allowed must do the below, but is not required to
2530 * actually call this function.
2531 */
2532void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2533{
2534	if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2535		p->cpus_ptr = new_mask;
2536		return;
2537	}
2538
2539	cpumask_copy(&p->cpus_mask, new_mask);
2540	p->nr_cpus_allowed = cpumask_weight(new_mask);
2541}
2542
2543static void
2544__do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2545{
2546	struct rq *rq = task_rq(p);
2547	bool queued, running;
2548
2549	/*
2550	 * This here violates the locking rules for affinity, since we're only
2551	 * supposed to change these variables while holding both rq->lock and
2552	 * p->pi_lock.
2553	 *
2554	 * HOWEVER, it magically works, because ttwu() is the only code that
2555	 * accesses these variables under p->pi_lock and only does so after
2556	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2557	 * before finish_task().
2558	 *
2559	 * XXX do further audits, this smells like something putrid.
2560	 */
2561	if (flags & SCA_MIGRATE_DISABLE)
2562		SCHED_WARN_ON(!p->on_cpu);
2563	else
2564		lockdep_assert_held(&p->pi_lock);
2565
2566	queued = task_on_rq_queued(p);
2567	running = task_current(rq, p);
2568
2569	if (queued) {
2570		/*
2571		 * Because __kthread_bind() calls this on blocked tasks without
2572		 * holding rq->lock.
2573		 */
2574		lockdep_assert_rq_held(rq);
2575		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2576	}
2577	if (running)
2578		put_prev_task(rq, p);
2579
2580	p->sched_class->set_cpus_allowed(p, new_mask, flags);
2581
2582	if (queued)
2583		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2584	if (running)
2585		set_next_task(rq, p);
2586}
2587
2588void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2589{
2590	__do_set_cpus_allowed(p, new_mask, 0);
2591}
2592
2593int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2594		      int node)
2595{
2596	if (!src->user_cpus_ptr)
2597		return 0;
2598
2599	dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2600	if (!dst->user_cpus_ptr)
2601		return -ENOMEM;
2602
2603	cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2604	return 0;
2605}
2606
2607static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2608{
2609	struct cpumask *user_mask = NULL;
2610
2611	swap(p->user_cpus_ptr, user_mask);
2612
2613	return user_mask;
2614}
2615
2616void release_user_cpus_ptr(struct task_struct *p)
2617{
2618	kfree(clear_user_cpus_ptr(p));
2619}
2620
2621/*
2622 * This function is wildly self concurrent; here be dragons.
2623 *
2624 *
2625 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2626 * designated task is enqueued on an allowed CPU. If that task is currently
2627 * running, we have to kick it out using the CPU stopper.
2628 *
2629 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2630 * Consider:
2631 *
2632 *     Initial conditions: P0->cpus_mask = [0, 1]
2633 *
2634 *     P0@CPU0                  P1
2635 *
2636 *     migrate_disable();
2637 *     <preempted>
2638 *                              set_cpus_allowed_ptr(P0, [1]);
2639 *
2640 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2641 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2642 * This means we need the following scheme:
2643 *
2644 *     P0@CPU0                  P1
2645 *
2646 *     migrate_disable();
2647 *     <preempted>
2648 *                              set_cpus_allowed_ptr(P0, [1]);
2649 *                                <blocks>
2650 *     <resumes>
2651 *     migrate_enable();
2652 *       __set_cpus_allowed_ptr();
2653 *       <wakes local stopper>
2654 *                         `--> <woken on migration completion>
2655 *
2656 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2657 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2658 * task p are serialized by p->pi_lock, which we can leverage: the one that
2659 * should come into effect at the end of the Migrate-Disable region is the last
2660 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2661 * but we still need to properly signal those waiting tasks at the appropriate
2662 * moment.
2663 *
2664 * This is implemented using struct set_affinity_pending. The first
2665 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2666 * setup an instance of that struct and install it on the targeted task_struct.
2667 * Any and all further callers will reuse that instance. Those then wait for
2668 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2669 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2670 *
2671 *
2672 * (1) In the cases covered above. There is one more where the completion is
2673 * signaled within affine_move_task() itself: when a subsequent affinity request
2674 * occurs after the stopper bailed out due to the targeted task still being
2675 * Migrate-Disable. Consider:
2676 *
2677 *     Initial conditions: P0->cpus_mask = [0, 1]
2678 *
2679 *     CPU0		  P1				P2
2680 *     <P0>
2681 *       migrate_disable();
2682 *       <preempted>
2683 *                        set_cpus_allowed_ptr(P0, [1]);
2684 *                          <blocks>
2685 *     <migration/0>
2686 *       migration_cpu_stop()
2687 *         is_migration_disabled()
2688 *           <bails>
2689 *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
2690 *                                                         <signal completion>
2691 *                          <awakes>
2692 *
2693 * Note that the above is safe vs a concurrent migrate_enable(), as any
2694 * pending affinity completion is preceded by an uninstallation of
2695 * p->migration_pending done with p->pi_lock held.
2696 */
2697static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2698			    int dest_cpu, unsigned int flags)
2699{
2700	struct set_affinity_pending my_pending = { }, *pending = NULL;
2701	bool stop_pending, complete = false;
2702
2703	/* Can the task run on the task's current CPU? If so, we're done */
2704	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2705		struct task_struct *push_task = NULL;
2706
2707		if ((flags & SCA_MIGRATE_ENABLE) &&
2708		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2709			rq->push_busy = true;
2710			push_task = get_task_struct(p);
2711		}
2712
2713		/*
2714		 * If there are pending waiters, but no pending stop_work,
2715		 * then complete now.
2716		 */
2717		pending = p->migration_pending;
2718		if (pending && !pending->stop_pending) {
2719			p->migration_pending = NULL;
2720			complete = true;
2721		}
2722
2723		task_rq_unlock(rq, p, rf);
2724
2725		if (push_task) {
2726			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2727					    p, &rq->push_work);
2728		}
2729
2730		if (complete)
2731			complete_all(&pending->done);
2732
2733		return 0;
2734	}
2735
2736	if (!(flags & SCA_MIGRATE_ENABLE)) {
2737		/* serialized by p->pi_lock */
2738		if (!p->migration_pending) {
2739			/* Install the request */
2740			refcount_set(&my_pending.refs, 1);
2741			init_completion(&my_pending.done);
2742			my_pending.arg = (struct migration_arg) {
2743				.task = p,
2744				.dest_cpu = dest_cpu,
2745				.pending = &my_pending,
2746			};
2747
2748			p->migration_pending = &my_pending;
2749		} else {
2750			pending = p->migration_pending;
2751			refcount_inc(&pending->refs);
2752			/*
2753			 * Affinity has changed, but we've already installed a
2754			 * pending. migration_cpu_stop() *must* see this, else
2755			 * we risk a completion of the pending despite having a
2756			 * task on a disallowed CPU.
2757			 *
2758			 * Serialized by p->pi_lock, so this is safe.
2759			 */
2760			pending->arg.dest_cpu = dest_cpu;
2761		}
2762	}
2763	pending = p->migration_pending;
2764	/*
2765	 * - !MIGRATE_ENABLE:
2766	 *   we'll have installed a pending if there wasn't one already.
2767	 *
2768	 * - MIGRATE_ENABLE:
2769	 *   we're here because the current CPU isn't matching anymore,
2770	 *   the only way that can happen is because of a concurrent
2771	 *   set_cpus_allowed_ptr() call, which should then still be
2772	 *   pending completion.
2773	 *
2774	 * Either way, we really should have a @pending here.
2775	 */
2776	if (WARN_ON_ONCE(!pending)) {
2777		task_rq_unlock(rq, p, rf);
2778		return -EINVAL;
2779	}
2780
2781	if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2782		/*
2783		 * MIGRATE_ENABLE gets here because 'p == current', but for
2784		 * anything else we cannot do is_migration_disabled(), punt
2785		 * and have the stopper function handle it all race-free.
2786		 */
2787		stop_pending = pending->stop_pending;
2788		if (!stop_pending)
2789			pending->stop_pending = true;
2790
2791		if (flags & SCA_MIGRATE_ENABLE)
2792			p->migration_flags &= ~MDF_PUSH;
2793
2794		task_rq_unlock(rq, p, rf);
2795
2796		if (!stop_pending) {
2797			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2798					    &pending->arg, &pending->stop_work);
2799		}
2800
2801		if (flags & SCA_MIGRATE_ENABLE)
2802			return 0;
2803	} else {
2804
2805		if (!is_migration_disabled(p)) {
2806			if (task_on_rq_queued(p))
2807				rq = move_queued_task(rq, rf, p, dest_cpu);
2808
2809			if (!pending->stop_pending) {
2810				p->migration_pending = NULL;
2811				complete = true;
2812			}
2813		}
2814		task_rq_unlock(rq, p, rf);
2815
2816		if (complete)
2817			complete_all(&pending->done);
2818	}
2819
2820	wait_for_completion(&pending->done);
2821
2822	if (refcount_dec_and_test(&pending->refs))
2823		wake_up_var(&pending->refs); /* No UaF, just an address */
2824
2825	/*
2826	 * Block the original owner of &pending until all subsequent callers
2827	 * have seen the completion and decremented the refcount
2828	 */
2829	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2830
2831	/* ARGH */
2832	WARN_ON_ONCE(my_pending.stop_pending);
2833
2834	return 0;
2835}
2836
2837/*
2838 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2839 */
2840static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2841					 const struct cpumask *new_mask,
2842					 u32 flags,
2843					 struct rq *rq,
2844					 struct rq_flags *rf)
2845	__releases(rq->lock)
2846	__releases(p->pi_lock)
2847{
2848	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2849	const struct cpumask *cpu_valid_mask = cpu_active_mask;
2850	bool kthread = p->flags & PF_KTHREAD;
2851	struct cpumask *user_mask = NULL;
2852	unsigned int dest_cpu;
2853	int ret = 0;
2854
2855	update_rq_clock(rq);
2856
2857	if (kthread || is_migration_disabled(p)) {
2858		/*
2859		 * Kernel threads are allowed on online && !active CPUs,
2860		 * however, during cpu-hot-unplug, even these might get pushed
2861		 * away if not KTHREAD_IS_PER_CPU.
2862		 *
2863		 * Specifically, migration_disabled() tasks must not fail the
2864		 * cpumask_any_and_distribute() pick below, esp. so on
2865		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2866		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2867		 */
2868		cpu_valid_mask = cpu_online_mask;
2869	}
2870
2871	if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2872		ret = -EINVAL;
2873		goto out;
2874	}
2875
2876	/*
2877	 * Must re-check here, to close a race against __kthread_bind(),
2878	 * sched_setaffinity() is not guaranteed to observe the flag.
2879	 */
2880	if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2881		ret = -EINVAL;
2882		goto out;
2883	}
2884
2885	if (!(flags & SCA_MIGRATE_ENABLE)) {
2886		if (cpumask_equal(&p->cpus_mask, new_mask))
2887			goto out;
2888
2889		if (WARN_ON_ONCE(p == current &&
2890				 is_migration_disabled(p) &&
2891				 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2892			ret = -EBUSY;
2893			goto out;
2894		}
2895	}
2896
2897	/*
2898	 * Picking a ~random cpu helps in cases where we are changing affinity
2899	 * for groups of tasks (ie. cpuset), so that load balancing is not
2900	 * immediately required to distribute the tasks within their new mask.
2901	 */
2902	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2903	if (dest_cpu >= nr_cpu_ids) {
2904		ret = -EINVAL;
2905		goto out;
2906	}
2907
2908	__do_set_cpus_allowed(p, new_mask, flags);
2909
2910	if (flags & SCA_USER)
2911		user_mask = clear_user_cpus_ptr(p);
2912
2913	ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2914
2915	kfree(user_mask);
2916
2917	return ret;
2918
2919out:
2920	task_rq_unlock(rq, p, rf);
2921
2922	return ret;
2923}
2924
2925/*
2926 * Change a given task's CPU affinity. Migrate the thread to a
2927 * proper CPU and schedule it away if the CPU it's executing on
2928 * is removed from the allowed bitmask.
2929 *
2930 * NOTE: the caller must have a valid reference to the task, the
2931 * task must not exit() & deallocate itself prematurely. The
2932 * call is not atomic; no spinlocks may be held.
2933 */
2934static int __set_cpus_allowed_ptr(struct task_struct *p,
2935				  const struct cpumask *new_mask, u32 flags)
2936{
2937	struct rq_flags rf;
2938	struct rq *rq;
2939
2940	rq = task_rq_lock(p, &rf);
2941	return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2942}
2943
2944int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2945{
2946	return __set_cpus_allowed_ptr(p, new_mask, 0);
2947}
2948EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2949
2950/*
2951 * Change a given task's CPU affinity to the intersection of its current
2952 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2953 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2954 * If the resulting mask is empty, leave the affinity unchanged and return
2955 * -EINVAL.
2956 */
2957static int restrict_cpus_allowed_ptr(struct task_struct *p,
2958				     struct cpumask *new_mask,
2959				     const struct cpumask *subset_mask)
2960{
2961	struct cpumask *user_mask = NULL;
2962	struct rq_flags rf;
2963	struct rq *rq;
2964	int err;
2965
2966	if (!p->user_cpus_ptr) {
2967		user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2968		if (!user_mask)
2969			return -ENOMEM;
2970	}
2971
2972	rq = task_rq_lock(p, &rf);
2973
2974	/*
2975	 * Forcefully restricting the affinity of a deadline task is
2976	 * likely to cause problems, so fail and noisily override the
2977	 * mask entirely.
2978	 */
2979	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2980		err = -EPERM;
2981		goto err_unlock;
2982	}
2983
2984	if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2985		err = -EINVAL;
2986		goto err_unlock;
2987	}
2988
2989	/*
2990	 * We're about to butcher the task affinity, so keep track of what
2991	 * the user asked for in case we're able to restore it later on.
2992	 */
2993	if (user_mask) {
2994		cpumask_copy(user_mask, p->cpus_ptr);
2995		p->user_cpus_ptr = user_mask;
2996	}
2997
2998	return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2999
3000err_unlock:
3001	task_rq_unlock(rq, p, &rf);
3002	kfree(user_mask);
3003	return err;
3004}
3005
3006/*
3007 * Restrict the CPU affinity of task @p so that it is a subset of
3008 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3009 * old affinity mask. If the resulting mask is empty, we warn and walk
3010 * up the cpuset hierarchy until we find a suitable mask.
3011 */
3012void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3013{
3014	cpumask_var_t new_mask;
3015	const struct cpumask *override_mask = task_cpu_possible_mask(p);
3016
3017	alloc_cpumask_var(&new_mask, GFP_KERNEL);
3018
3019	/*
3020	 * __migrate_task() can fail silently in the face of concurrent
3021	 * offlining of the chosen destination CPU, so take the hotplug
3022	 * lock to ensure that the migration succeeds.
3023	 */
3024	cpus_read_lock();
3025	if (!cpumask_available(new_mask))
3026		goto out_set_mask;
3027
3028	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3029		goto out_free_mask;
3030
3031	/*
3032	 * We failed to find a valid subset of the affinity mask for the
3033	 * task, so override it based on its cpuset hierarchy.
3034	 */
3035	cpuset_cpus_allowed(p, new_mask);
3036	override_mask = new_mask;
3037
3038out_set_mask:
3039	if (printk_ratelimit()) {
3040		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3041				task_pid_nr(p), p->comm,
3042				cpumask_pr_args(override_mask));
3043	}
3044
3045	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3046out_free_mask:
3047	cpus_read_unlock();
3048	free_cpumask_var(new_mask);
3049}
3050
3051static int
3052__sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3053
3054/*
3055 * Restore the affinity of a task @p which was previously restricted by a
3056 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3057 * @p->user_cpus_ptr.
3058 *
3059 * It is the caller's responsibility to serialise this with any calls to
3060 * force_compatible_cpus_allowed_ptr(@p).
3061 */
3062void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3063{
3064	struct cpumask *user_mask = p->user_cpus_ptr;
3065	unsigned long flags;
3066
3067	/*
3068	 * Try to restore the old affinity mask. If this fails, then
3069	 * we free the mask explicitly to avoid it being inherited across
3070	 * a subsequent fork().
3071	 */
3072	if (!user_mask || !__sched_setaffinity(p, user_mask))
3073		return;
3074
3075	raw_spin_lock_irqsave(&p->pi_lock, flags);
3076	user_mask = clear_user_cpus_ptr(p);
3077	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3078
3079	kfree(user_mask);
3080}
3081
3082void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3083{
3084#ifdef CONFIG_SCHED_DEBUG
3085	unsigned int state = READ_ONCE(p->__state);
3086
3087	/*
3088	 * We should never call set_task_cpu() on a blocked task,
3089	 * ttwu() will sort out the placement.
3090	 */
3091	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3092
3093	/*
3094	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3095	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3096	 * time relying on p->on_rq.
3097	 */
3098	WARN_ON_ONCE(state == TASK_RUNNING &&
3099		     p->sched_class == &fair_sched_class &&
3100		     (p->on_rq && !task_on_rq_migrating(p)));
3101
3102#ifdef CONFIG_LOCKDEP
3103	/*
3104	 * The caller should hold either p->pi_lock or rq->lock, when changing
3105	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3106	 *
3107	 * sched_move_task() holds both and thus holding either pins the cgroup,
3108	 * see task_group().
3109	 *
3110	 * Furthermore, all task_rq users should acquire both locks, see
3111	 * task_rq_lock().
3112	 */
3113	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3114				      lockdep_is_held(__rq_lockp(task_rq(p)))));
3115#endif
3116	/*
3117	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3118	 */
3119	WARN_ON_ONCE(!cpu_online(new_cpu));
3120
3121	WARN_ON_ONCE(is_migration_disabled(p));
3122#endif
3123
3124	trace_sched_migrate_task(p, new_cpu);
3125
3126	if (task_cpu(p) != new_cpu) {
3127		if (p->sched_class->migrate_task_rq)
3128			p->sched_class->migrate_task_rq(p, new_cpu);
3129		p->se.nr_migrations++;
3130		rseq_migrate(p);
3131		perf_event_task_migrate(p);
3132	}
3133
3134	__set_task_cpu(p, new_cpu);
3135}
3136
3137#ifdef CONFIG_NUMA_BALANCING
3138static void __migrate_swap_task(struct task_struct *p, int cpu)
3139{
3140	if (task_on_rq_queued(p)) {
3141		struct rq *src_rq, *dst_rq;
3142		struct rq_flags srf, drf;
3143
3144		src_rq = task_rq(p);
3145		dst_rq = cpu_rq(cpu);
3146
3147		rq_pin_lock(src_rq, &srf);
3148		rq_pin_lock(dst_rq, &drf);
3149
3150		deactivate_task(src_rq, p, 0);
3151		set_task_cpu(p, cpu);
3152		activate_task(dst_rq, p, 0);
3153		check_preempt_curr(dst_rq, p, 0);
3154
3155		rq_unpin_lock(dst_rq, &drf);
3156		rq_unpin_lock(src_rq, &srf);
3157
3158	} else {
3159		/*
3160		 * Task isn't running anymore; make it appear like we migrated
3161		 * it before it went to sleep. This means on wakeup we make the
3162		 * previous CPU our target instead of where it really is.
3163		 */
3164		p->wake_cpu = cpu;
3165	}
3166}
3167
3168struct migration_swap_arg {
3169	struct task_struct *src_task, *dst_task;
3170	int src_cpu, dst_cpu;
3171};
3172
3173static int migrate_swap_stop(void *data)
3174{
3175	struct migration_swap_arg *arg = data;
3176	struct rq *src_rq, *dst_rq;
3177	int ret = -EAGAIN;
3178
3179	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3180		return -EAGAIN;
3181
3182	src_rq = cpu_rq(arg->src_cpu);
3183	dst_rq = cpu_rq(arg->dst_cpu);
3184
3185	double_raw_lock(&arg->src_task->pi_lock,
3186			&arg->dst_task->pi_lock);
3187	double_rq_lock(src_rq, dst_rq);
3188
3189	if (task_cpu(arg->dst_task) != arg->dst_cpu)
3190		goto unlock;
3191
3192	if (task_cpu(arg->src_task) != arg->src_cpu)
3193		goto unlock;
3194
3195	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3196		goto unlock;
3197
3198	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3199		goto unlock;
3200
3201	__migrate_swap_task(arg->src_task, arg->dst_cpu);
3202	__migrate_swap_task(arg->dst_task, arg->src_cpu);
3203
3204	ret = 0;
3205
3206unlock:
3207	double_rq_unlock(src_rq, dst_rq);
3208	raw_spin_unlock(&arg->dst_task->pi_lock);
3209	raw_spin_unlock(&arg->src_task->pi_lock);
3210
3211	return ret;
3212}
3213
3214/*
3215 * Cross migrate two tasks
3216 */
3217int migrate_swap(struct task_struct *cur, struct task_struct *p,
3218		int target_cpu, int curr_cpu)
3219{
3220	struct migration_swap_arg arg;
3221	int ret = -EINVAL;
3222
3223	arg = (struct migration_swap_arg){
3224		.src_task = cur,
3225		.src_cpu = curr_cpu,
3226		.dst_task = p,
3227		.dst_cpu = target_cpu,
3228	};
3229
3230	if (arg.src_cpu == arg.dst_cpu)
3231		goto out;
3232
3233	/*
3234	 * These three tests are all lockless; this is OK since all of them
3235	 * will be re-checked with proper locks held further down the line.
3236	 */
3237	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3238		goto out;
3239
3240	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3241		goto out;
3242
3243	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3244		goto out;
3245
3246	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3247	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3248
3249out:
3250	return ret;
3251}
3252#endif /* CONFIG_NUMA_BALANCING */
3253
3254/*
3255 * wait_task_inactive - wait for a thread to unschedule.
3256 *
3257 * If @match_state is nonzero, it's the @p->state value just checked and
3258 * not expected to change.  If it changes, i.e. @p might have woken up,
3259 * then return zero.  When we succeed in waiting for @p to be off its CPU,
3260 * we return a positive number (its total switch count).  If a second call
3261 * a short while later returns the same number, the caller can be sure that
3262 * @p has remained unscheduled the whole time.
3263 *
3264 * The caller must ensure that the task *will* unschedule sometime soon,
3265 * else this function might spin for a *long* time. This function can't
3266 * be called with interrupts off, or it may introduce deadlock with
3267 * smp_call_function() if an IPI is sent by the same process we are
3268 * waiting to become inactive.
3269 */
3270unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3271{
3272	int running, queued;
3273	struct rq_flags rf;
3274	unsigned long ncsw;
3275	struct rq *rq;
3276
3277	for (;;) {
3278		/*
3279		 * We do the initial early heuristics without holding
3280		 * any task-queue locks at all. We'll only try to get
3281		 * the runqueue lock when things look like they will
3282		 * work out!
3283		 */
3284		rq = task_rq(p);
3285
3286		/*
3287		 * If the task is actively running on another CPU
3288		 * still, just relax and busy-wait without holding
3289		 * any locks.
3290		 *
3291		 * NOTE! Since we don't hold any locks, it's not
3292		 * even sure that "rq" stays as the right runqueue!
3293		 * But we don't care, since "task_running()" will
3294		 * return false if the runqueue has changed and p
3295		 * is actually now running somewhere else!
3296		 */
3297		while (task_running(rq, p)) {
3298			if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3299				return 0;
3300			cpu_relax();
3301		}
3302
3303		/*
3304		 * Ok, time to look more closely! We need the rq
3305		 * lock now, to be *sure*. If we're wrong, we'll
3306		 * just go back and repeat.
3307		 */
3308		rq = task_rq_lock(p, &rf);
3309		trace_sched_wait_task(p);
3310		running = task_running(rq, p);
3311		queued = task_on_rq_queued(p);
3312		ncsw = 0;
3313		if (!match_state || READ_ONCE(p->__state) == match_state)
3314			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3315		task_rq_unlock(rq, p, &rf);
3316
3317		/*
3318		 * If it changed from the expected state, bail out now.
3319		 */
3320		if (unlikely(!ncsw))
3321			break;
3322
3323		/*
3324		 * Was it really running after all now that we
3325		 * checked with the proper locks actually held?
3326		 *
3327		 * Oops. Go back and try again..
3328		 */
3329		if (unlikely(running)) {
3330			cpu_relax();
3331			continue;
3332		}
3333
3334		/*
3335		 * It's not enough that it's not actively running,
3336		 * it must be off the runqueue _entirely_, and not
3337		 * preempted!
3338		 *
3339		 * So if it was still runnable (but just not actively
3340		 * running right now), it's preempted, and we should
3341		 * yield - it could be a while.
3342		 */
3343		if (unlikely(queued)) {
3344			ktime_t to = NSEC_PER_SEC / HZ;
3345
3346			set_current_state(TASK_UNINTERRUPTIBLE);
3347			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3348			continue;
3349		}
3350
3351		/*
3352		 * Ahh, all good. It wasn't running, and it wasn't
3353		 * runnable, which means that it will never become
3354		 * running in the future either. We're all done!
3355		 */
3356		break;
3357	}
3358
3359	return ncsw;
3360}
3361
3362/***
3363 * kick_process - kick a running thread to enter/exit the kernel
3364 * @p: the to-be-kicked thread
3365 *
3366 * Cause a process which is running on another CPU to enter
3367 * kernel-mode, without any delay. (to get signals handled.)
3368 *
3369 * NOTE: this function doesn't have to take the runqueue lock,
3370 * because all it wants to ensure is that the remote task enters
3371 * the kernel. If the IPI races and the task has been migrated
3372 * to another CPU then no harm is done and the purpose has been
3373 * achieved as well.
3374 */
3375void kick_process(struct task_struct *p)
3376{
3377	int cpu;
3378
3379	preempt_disable();
3380	cpu = task_cpu(p);
3381	if ((cpu != smp_processor_id()) && task_curr(p))
3382		smp_send_reschedule(cpu);
3383	preempt_enable();
3384}
3385EXPORT_SYMBOL_GPL(kick_process);
3386
3387/*
3388 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3389 *
3390 * A few notes on cpu_active vs cpu_online:
3391 *
3392 *  - cpu_active must be a subset of cpu_online
3393 *
3394 *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3395 *    see __set_cpus_allowed_ptr(). At this point the newly online
3396 *    CPU isn't yet part of the sched domains, and balancing will not
3397 *    see it.
3398 *
3399 *  - on CPU-down we clear cpu_active() to mask the sched domains and
3400 *    avoid the load balancer to place new tasks on the to be removed
3401 *    CPU. Existing tasks will remain running there and will be taken
3402 *    off.
3403 *
3404 * This means that fallback selection must not select !active CPUs.
3405 * And can assume that any active CPU must be online. Conversely
3406 * select_task_rq() below may allow selection of !active CPUs in order
3407 * to satisfy the above rules.
3408 */
3409static int select_fallback_rq(int cpu, struct task_struct *p)
3410{
3411	int nid = cpu_to_node(cpu);
3412	const struct cpumask *nodemask = NULL;
3413	enum { cpuset, possible, fail } state = cpuset;
3414	int dest_cpu;
3415
3416	/*
3417	 * If the node that the CPU is on has been offlined, cpu_to_node()
3418	 * will return -1. There is no CPU on the node, and we should
3419	 * select the CPU on the other node.
3420	 */
3421	if (nid != -1) {
3422		nodemask = cpumask_of_node(nid);
3423
3424		/* Look for allowed, online CPU in same node. */
3425		for_each_cpu(dest_cpu, nodemask) {
3426			if (is_cpu_allowed(p, dest_cpu))
3427				return dest_cpu;
3428		}
3429	}
3430
3431	for (;;) {
3432		/* Any allowed, online CPU? */
3433		for_each_cpu(dest_cpu, p->cpus_ptr) {
3434			if (!is_cpu_allowed(p, dest_cpu))
3435				continue;
3436
3437			goto out;
3438		}
3439
3440		/* No more Mr. Nice Guy. */
3441		switch (state) {
3442		case cpuset:
3443			if (cpuset_cpus_allowed_fallback(p)) {
3444				state = possible;
3445				break;
3446			}
3447			fallthrough;
3448		case possible:
3449			/*
3450			 * XXX When called from select_task_rq() we only
3451			 * hold p->pi_lock and again violate locking order.
3452			 *
3453			 * More yuck to audit.
3454			 */
3455			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3456			state = fail;
3457			break;
3458		case fail:
3459			BUG();
3460			break;
3461		}
3462	}
3463
3464out:
3465	if (state != cpuset) {
3466		/*
3467		 * Don't tell them about moving exiting tasks or
3468		 * kernel threads (both mm NULL), since they never
3469		 * leave kernel.
3470		 */
3471		if (p->mm && printk_ratelimit()) {
3472			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3473					task_pid_nr(p), p->comm, cpu);
3474		}
3475	}
3476
3477	return dest_cpu;
3478}
3479
3480/*
3481 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3482 */
3483static inline
3484int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3485{
3486	lockdep_assert_held(&p->pi_lock);
3487
3488	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3489		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3490	else
3491		cpu = cpumask_any(p->cpus_ptr);
3492
3493	/*
3494	 * In order not to call set_task_cpu() on a blocking task we need
3495	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3496	 * CPU.
3497	 *
3498	 * Since this is common to all placement strategies, this lives here.
3499	 *
3500	 * [ this allows ->select_task() to simply return task_cpu(p) and
3501	 *   not worry about this generic constraint ]
3502	 */
3503	if (unlikely(!is_cpu_allowed(p, cpu)))
3504		cpu = select_fallback_rq(task_cpu(p), p);
3505
3506	return cpu;
3507}
3508
3509void sched_set_stop_task(int cpu, struct task_struct *stop)
3510{
3511	static struct lock_class_key stop_pi_lock;
3512	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3513	struct task_struct *old_stop = cpu_rq(cpu)->stop;
3514
3515	if (stop) {
3516		/*
3517		 * Make it appear like a SCHED_FIFO task, its something
3518		 * userspace knows about and won't get confused about.
3519		 *
3520		 * Also, it will make PI more or less work without too
3521		 * much confusion -- but then, stop work should not
3522		 * rely on PI working anyway.
3523		 */
3524		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3525
3526		stop->sched_class = &stop_sched_class;
3527
3528		/*
3529		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3530		 * adjust the effective priority of a task. As a result,
3531		 * rt_mutex_setprio() can trigger (RT) balancing operations,
3532		 * which can then trigger wakeups of the stop thread to push
3533		 * around the current task.
3534		 *
3535		 * The stop task itself will never be part of the PI-chain, it
3536		 * never blocks, therefore that ->pi_lock recursion is safe.
3537		 * Tell lockdep about this by placing the stop->pi_lock in its
3538		 * own class.
3539		 */
3540		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3541	}
3542
3543	cpu_rq(cpu)->stop = stop;
3544
3545	if (old_stop) {
3546		/*
3547		 * Reset it back to a normal scheduling class so that
3548		 * it can die in pieces.
3549		 */
3550		old_stop->sched_class = &rt_sched_class;
3551	}
3552}
3553
3554#else /* CONFIG_SMP */
3555
3556static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3557					 const struct cpumask *new_mask,
3558					 u32 flags)
3559{
3560	return set_cpus_allowed_ptr(p, new_mask);
3561}
3562
3563static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3564
3565static inline bool rq_has_pinned_tasks(struct rq *rq)
3566{
3567	return false;
3568}
3569
3570#endif /* !CONFIG_SMP */
3571
3572static void
3573ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3574{
3575	struct rq *rq;
3576
3577	if (!schedstat_enabled())
3578		return;
3579
3580	rq = this_rq();
3581
3582#ifdef CONFIG_SMP
3583	if (cpu == rq->cpu) {
3584		__schedstat_inc(rq->ttwu_local);
3585		__schedstat_inc(p->stats.nr_wakeups_local);
3586	} else {
3587		struct sched_domain *sd;
3588
3589		__schedstat_inc(p->stats.nr_wakeups_remote);
3590		rcu_read_lock();
3591		for_each_domain(rq->cpu, sd) {
3592			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3593				__schedstat_inc(sd->ttwu_wake_remote);
3594				break;
3595			}
3596		}
3597		rcu_read_unlock();
3598	}
3599
3600	if (wake_flags & WF_MIGRATED)
3601		__schedstat_inc(p->stats.nr_wakeups_migrate);
3602#endif /* CONFIG_SMP */
3603
3604	__schedstat_inc(rq->ttwu_count);
3605	__schedstat_inc(p->stats.nr_wakeups);
3606
3607	if (wake_flags & WF_SYNC)
3608		__schedstat_inc(p->stats.nr_wakeups_sync);
3609}
3610
3611/*
3612 * Mark the task runnable and perform wakeup-preemption.
3613 */
3614static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3615			   struct rq_flags *rf)
3616{
3617	check_preempt_curr(rq, p, wake_flags);
3618	WRITE_ONCE(p->__state, TASK_RUNNING);
3619	trace_sched_wakeup(p);
3620
3621#ifdef CONFIG_SMP
3622	if (p->sched_class->task_woken) {
3623		/*
3624		 * Our task @p is fully woken up and running; so it's safe to
3625		 * drop the rq->lock, hereafter rq is only used for statistics.
3626		 */
3627		rq_unpin_lock(rq, rf);
3628		p->sched_class->task_woken(rq, p);
3629		rq_repin_lock(rq, rf);
3630	}
3631
3632	if (rq->idle_stamp) {
3633		u64 delta = rq_clock(rq) - rq->idle_stamp;
3634		u64 max = 2*rq->max_idle_balance_cost;
3635
3636		update_avg(&rq->avg_idle, delta);
3637
3638		if (rq->avg_idle > max)
3639			rq->avg_idle = max;
3640
3641		rq->wake_stamp = jiffies;
3642		rq->wake_avg_idle = rq->avg_idle / 2;
3643
3644		rq->idle_stamp = 0;
3645	}
3646#endif
3647}
3648
3649static void
3650ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3651		 struct rq_flags *rf)
3652{
3653	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3654
3655	lockdep_assert_rq_held(rq);
3656
3657	if (p->sched_contributes_to_load)
3658		rq->nr_uninterruptible--;
3659
3660#ifdef CONFIG_SMP
3661	if (wake_flags & WF_MIGRATED)
3662		en_flags |= ENQUEUE_MIGRATED;
3663	else
3664#endif
3665	if (p->in_iowait) {
3666		delayacct_blkio_end(p);
3667		atomic_dec(&task_rq(p)->nr_iowait);
3668	}
3669
3670	activate_task(rq, p, en_flags);
3671	ttwu_do_wakeup(rq, p, wake_flags, rf);
3672}
3673
3674/*
3675 * Consider @p being inside a wait loop:
3676 *
3677 *   for (;;) {
3678 *      set_current_state(TASK_UNINTERRUPTIBLE);
3679 *
3680 *      if (CONDITION)
3681 *         break;
3682 *
3683 *      schedule();
3684 *   }
3685 *   __set_current_state(TASK_RUNNING);
3686 *
3687 * between set_current_state() and schedule(). In this case @p is still
3688 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3689 * an atomic manner.
3690 *
3691 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3692 * then schedule() must still happen and p->state can be changed to
3693 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3694 * need to do a full wakeup with enqueue.
3695 *
3696 * Returns: %true when the wakeup is done,
3697 *          %false otherwise.
3698 */
3699static int ttwu_runnable(struct task_struct *p, int wake_flags)
3700{
3701	struct rq_flags rf;
3702	struct rq *rq;
3703	int ret = 0;
3704
3705	rq = __task_rq_lock(p, &rf);
3706	if (task_on_rq_queued(p)) {
3707		/* check_preempt_curr() may use rq clock */
3708		update_rq_clock(rq);
3709		ttwu_do_wakeup(rq, p, wake_flags, &rf);
3710		ret = 1;
3711	}
3712	__task_rq_unlock(rq, &rf);
3713
3714	return ret;
3715}
3716
3717#ifdef CONFIG_SMP
3718void sched_ttwu_pending(void *arg)
3719{
3720	struct llist_node *llist = arg;
3721	struct rq *rq = this_rq();
3722	struct task_struct *p, *t;
3723	struct rq_flags rf;
3724
3725	if (!llist)
3726		return;
3727
3728	/*
3729	 * rq::ttwu_pending racy indication of out-standing wakeups.
3730	 * Races such that false-negatives are possible, since they
3731	 * are shorter lived that false-positives would be.
3732	 */
3733	WRITE_ONCE(rq->ttwu_pending, 0);
3734
3735	rq_lock_irqsave(rq, &rf);
3736	update_rq_clock(rq);
3737
3738	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3739		if (WARN_ON_ONCE(p->on_cpu))
3740			smp_cond_load_acquire(&p->on_cpu, !VAL);
3741
3742		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3743			set_task_cpu(p, cpu_of(rq));
3744
3745		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3746	}
3747
3748	rq_unlock_irqrestore(rq, &rf);
3749}
3750
3751void send_call_function_single_ipi(int cpu)
3752{
3753	struct rq *rq = cpu_rq(cpu);
3754
3755	if (!set_nr_if_polling(rq->idle))
3756		arch_send_call_function_single_ipi(cpu);
3757	else
3758		trace_sched_wake_idle_without_ipi(cpu);
3759}
3760
3761/*
3762 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3763 * necessary. The wakee CPU on receipt of the IPI will queue the task
3764 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3765 * of the wakeup instead of the waker.
3766 */
3767static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3768{
3769	struct rq *rq = cpu_rq(cpu);
3770
3771	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3772
3773	WRITE_ONCE(rq->ttwu_pending, 1);
3774	__smp_call_single_queue(cpu, &p->wake_entry.llist);
3775}
3776
3777void wake_up_if_idle(int cpu)
3778{
3779	struct rq *rq = cpu_rq(cpu);
3780	struct rq_flags rf;
3781
3782	rcu_read_lock();
3783
3784	if (!is_idle_task(rcu_dereference(rq->curr)))
3785		goto out;
3786
3787	rq_lock_irqsave(rq, &rf);
3788	if (is_idle_task(rq->curr))
3789		resched_curr(rq);
3790	/* Else CPU is not idle, do nothing here: */
3791	rq_unlock_irqrestore(rq, &rf);
3792
3793out:
3794	rcu_read_unlock();
3795}
3796
3797bool cpus_share_cache(int this_cpu, int that_cpu)
3798{
3799	if (this_cpu == that_cpu)
3800		return true;
3801
3802	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3803}
3804
3805static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3806{
3807	/*
3808	 * Do not complicate things with the async wake_list while the CPU is
3809	 * in hotplug state.
3810	 */
3811	if (!cpu_active(cpu))
3812		return false;
3813
3814	/* Ensure the task will still be allowed to run on the CPU. */
3815	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3816		return false;
3817
3818	/*
3819	 * If the CPU does not share cache, then queue the task on the
3820	 * remote rqs wakelist to avoid accessing remote data.
3821	 */
3822	if (!cpus_share_cache(smp_processor_id(), cpu))
3823		return true;
3824
3825	if (cpu == smp_processor_id())
3826		return false;
3827
3828	/*
3829	 * If the wakee cpu is idle, or the task is descheduling and the
3830	 * only running task on the CPU, then use the wakelist to offload
3831	 * the task activation to the idle (or soon-to-be-idle) CPU as
3832	 * the current CPU is likely busy. nr_running is checked to
3833	 * avoid unnecessary task stacking.
3834	 *
3835	 * Note that we can only get here with (wakee) p->on_rq=0,
3836	 * p->on_cpu can be whatever, we've done the dequeue, so
3837	 * the wakee has been accounted out of ->nr_running.
3838	 */
3839	if (!cpu_rq(cpu)->nr_running)
3840		return true;
3841
3842	return false;
3843}
3844
3845static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3846{
3847	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3848		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3849		__ttwu_queue_wakelist(p, cpu, wake_flags);
3850		return true;
3851	}
3852
3853	return false;
3854}
3855
3856#else /* !CONFIG_SMP */
3857
3858static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3859{
3860	return false;
3861}
3862
3863#endif /* CONFIG_SMP */
3864
3865static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3866{
3867	struct rq *rq = cpu_rq(cpu);
3868	struct rq_flags rf;
3869
3870	if (ttwu_queue_wakelist(p, cpu, wake_flags))
3871		return;
3872
3873	rq_lock(rq, &rf);
3874	update_rq_clock(rq);
3875	ttwu_do_activate(rq, p, wake_flags, &rf);
3876	rq_unlock(rq, &rf);
3877}
3878
3879/*
3880 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3881 *
3882 * The caller holds p::pi_lock if p != current or has preemption
3883 * disabled when p == current.
3884 *
3885 * The rules of PREEMPT_RT saved_state:
3886 *
3887 *   The related locking code always holds p::pi_lock when updating
3888 *   p::saved_state, which means the code is fully serialized in both cases.
3889 *
3890 *   The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3891 *   bits set. This allows to distinguish all wakeup scenarios.
3892 */
3893static __always_inline
3894bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3895{
3896	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3897		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3898			     state != TASK_RTLOCK_WAIT);
3899	}
3900
3901	if (READ_ONCE(p->__state) & state) {
3902		*success = 1;
3903		return true;
3904	}
3905
3906#ifdef CONFIG_PREEMPT_RT
3907	/*
3908	 * Saved state preserves the task state across blocking on
3909	 * an RT lock.  If the state matches, set p::saved_state to
3910	 * TASK_RUNNING, but do not wake the task because it waits
3911	 * for a lock wakeup. Also indicate success because from
3912	 * the regular waker's point of view this has succeeded.
3913	 *
3914	 * After acquiring the lock the task will restore p::__state
3915	 * from p::saved_state which ensures that the regular
3916	 * wakeup is not lost. The restore will also set
3917	 * p::saved_state to TASK_RUNNING so any further tests will
3918	 * not result in false positives vs. @success
3919	 */
3920	if (p->saved_state & state) {
3921		p->saved_state = TASK_RUNNING;
3922		*success = 1;
3923	}
3924#endif
3925	return false;
3926}
3927
3928/*
3929 * Notes on Program-Order guarantees on SMP systems.
3930 *
3931 *  MIGRATION
3932 *
3933 * The basic program-order guarantee on SMP systems is that when a task [t]
3934 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3935 * execution on its new CPU [c1].
3936 *
3937 * For migration (of runnable tasks) this is provided by the following means:
3938 *
3939 *  A) UNLOCK of the rq(c0)->lock scheduling out task t
3940 *  B) migration for t is required to synchronize *both* rq(c0)->lock and
3941 *     rq(c1)->lock (if not at the same time, then in that order).
3942 *  C) LOCK of the rq(c1)->lock scheduling in task
3943 *
3944 * Release/acquire chaining guarantees that B happens after A and C after B.
3945 * Note: the CPU doing B need not be c0 or c1
3946 *
3947 * Example:
3948 *
3949 *   CPU0            CPU1            CPU2
3950 *
3951 *   LOCK rq(0)->lock
3952 *   sched-out X
3953 *   sched-in Y
3954 *   UNLOCK rq(0)->lock
3955 *
3956 *                                   LOCK rq(0)->lock // orders against CPU0
3957 *                                   dequeue X
3958 *                                   UNLOCK rq(0)->lock
3959 *
3960 *                                   LOCK rq(1)->lock
3961 *                                   enqueue X
3962 *                                   UNLOCK rq(1)->lock
3963 *
3964 *                   LOCK rq(1)->lock // orders against CPU2
3965 *                   sched-out Z
3966 *                   sched-in X
3967 *                   UNLOCK rq(1)->lock
3968 *
3969 *
3970 *  BLOCKING -- aka. SLEEP + WAKEUP
3971 *
3972 * For blocking we (obviously) need to provide the same guarantee as for
3973 * migration. However the means are completely different as there is no lock
3974 * chain to provide order. Instead we do:
3975 *
3976 *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
3977 *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3978 *
3979 * Example:
3980 *
3981 *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
3982 *
3983 *   LOCK rq(0)->lock LOCK X->pi_lock
3984 *   dequeue X
3985 *   sched-out X
3986 *   smp_store_release(X->on_cpu, 0);
3987 *
3988 *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
3989 *                    X->state = WAKING
3990 *                    set_task_cpu(X,2)
3991 *
3992 *                    LOCK rq(2)->lock
3993 *                    enqueue X
3994 *                    X->state = RUNNING
3995 *                    UNLOCK rq(2)->lock
3996 *
3997 *                                          LOCK rq(2)->lock // orders against CPU1
3998 *                                          sched-out Z
3999 *                                          sched-in X
4000 *                                          UNLOCK rq(2)->lock
4001 *
4002 *                    UNLOCK X->pi_lock
4003 *   UNLOCK rq(0)->lock
4004 *
4005 *
4006 * However, for wakeups there is a second guarantee we must provide, namely we
4007 * must ensure that CONDITION=1 done by the caller can not be reordered with
4008 * accesses to the task state; see try_to_wake_up() and set_current_state().
4009 */
4010
4011/**
4012 * try_to_wake_up - wake up a thread
4013 * @p: the thread to be awakened
4014 * @state: the mask of task states that can be woken
4015 * @wake_flags: wake modifier flags (WF_*)
4016 *
4017 * Conceptually does:
4018 *
4019 *   If (@state & @p->state) @p->state = TASK_RUNNING.
4020 *
4021 * If the task was not queued/runnable, also place it back on a runqueue.
4022 *
4023 * This function is atomic against schedule() which would dequeue the task.
4024 *
4025 * It issues a full memory barrier before accessing @p->state, see the comment
4026 * with set_current_state().
4027 *
4028 * Uses p->pi_lock to serialize against concurrent wake-ups.
4029 *
4030 * Relies on p->pi_lock stabilizing:
4031 *  - p->sched_class
4032 *  - p->cpus_ptr
4033 *  - p->sched_task_group
4034 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4035 *
4036 * Tries really hard to only take one task_rq(p)->lock for performance.
4037 * Takes rq->lock in:
4038 *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
4039 *  - ttwu_queue()       -- new rq, for enqueue of the task;
4040 *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4041 *
4042 * As a consequence we race really badly with just about everything. See the
4043 * many memory barriers and their comments for details.
4044 *
4045 * Return: %true if @p->state changes (an actual wakeup was done),
4046 *	   %false otherwise.
4047 */
4048static int
4049try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4050{
4051	unsigned long flags;
4052	int cpu, success = 0;
4053
4054	preempt_disable();
4055	if (p == current) {
4056		/*
4057		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4058		 * == smp_processor_id()'. Together this means we can special
4059		 * case the whole 'p->on_rq && ttwu_runnable()' case below
4060		 * without taking any locks.
4061		 *
4062		 * In particular:
4063		 *  - we rely on Program-Order guarantees for all the ordering,
4064		 *  - we're serialized against set_special_state() by virtue of
4065		 *    it disabling IRQs (this allows not taking ->pi_lock).
4066		 */
4067		if (!ttwu_state_match(p, state, &success))
4068			goto out;
4069
4070		trace_sched_waking(p);
4071		WRITE_ONCE(p->__state, TASK_RUNNING);
4072		trace_sched_wakeup(p);
4073		goto out;
4074	}
4075
4076	/*
4077	 * If we are going to wake up a thread waiting for CONDITION we
4078	 * need to ensure that CONDITION=1 done by the caller can not be
4079	 * reordered with p->state check below. This pairs with smp_store_mb()
4080	 * in set_current_state() that the waiting thread does.
4081	 */
4082	raw_spin_lock_irqsave(&p->pi_lock, flags);
4083	smp_mb__after_spinlock();
4084	if (!ttwu_state_match(p, state, &success))
4085		goto unlock;
4086
4087	trace_sched_waking(p);
4088
4089	/*
4090	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4091	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4092	 * in smp_cond_load_acquire() below.
4093	 *
4094	 * sched_ttwu_pending()			try_to_wake_up()
4095	 *   STORE p->on_rq = 1			  LOAD p->state
4096	 *   UNLOCK rq->lock
4097	 *
4098	 * __schedule() (switch to task 'p')
4099	 *   LOCK rq->lock			  smp_rmb();
4100	 *   smp_mb__after_spinlock();
4101	 *   UNLOCK rq->lock
4102	 *
4103	 * [task p]
4104	 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
4105	 *
4106	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4107	 * __schedule().  See the comment for smp_mb__after_spinlock().
4108	 *
4109	 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4110	 */
4111	smp_rmb();
4112	if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4113		goto unlock;
4114
4115#ifdef CONFIG_SMP
4116	/*
4117	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4118	 * possible to, falsely, observe p->on_cpu == 0.
4119	 *
4120	 * One must be running (->on_cpu == 1) in order to remove oneself
4121	 * from the runqueue.
4122	 *
4123	 * __schedule() (switch to task 'p')	try_to_wake_up()
4124	 *   STORE p->on_cpu = 1		  LOAD p->on_rq
4125	 *   UNLOCK rq->lock
4126	 *
4127	 * __schedule() (put 'p' to sleep)
4128	 *   LOCK rq->lock			  smp_rmb();
4129	 *   smp_mb__after_spinlock();
4130	 *   STORE p->on_rq = 0			  LOAD p->on_cpu
4131	 *
4132	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4133	 * __schedule().  See the comment for smp_mb__after_spinlock().
4134	 *
4135	 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4136	 * schedule()'s deactivate_task() has 'happened' and p will no longer
4137	 * care about it's own p->state. See the comment in __schedule().
4138	 */
4139	smp_acquire__after_ctrl_dep();
4140
4141	/*
4142	 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4143	 * == 0), which means we need to do an enqueue, change p->state to
4144	 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4145	 * enqueue, such as ttwu_queue_wakelist().
4146	 */
4147	WRITE_ONCE(p->__state, TASK_WAKING);
4148
4149	/*
4150	 * If the owning (remote) CPU is still in the middle of schedule() with
4151	 * this task as prev, considering queueing p on the remote CPUs wake_list
4152	 * which potentially sends an IPI instead of spinning on p->on_cpu to
4153	 * let the waker make forward progress. This is safe because IRQs are
4154	 * disabled and the IPI will deliver after on_cpu is cleared.
4155	 *
4156	 * Ensure we load task_cpu(p) after p->on_cpu:
4157	 *
4158	 * set_task_cpu(p, cpu);
4159	 *   STORE p->cpu = @cpu
4160	 * __schedule() (switch to task 'p')
4161	 *   LOCK rq->lock
4162	 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
4163	 *   STORE p->on_cpu = 1		LOAD p->cpu
4164	 *
4165	 * to ensure we observe the correct CPU on which the task is currently
4166	 * scheduling.
4167	 */
4168	if (smp_load_acquire(&p->on_cpu) &&
4169	    ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4170		goto unlock;
4171
4172	/*
4173	 * If the owning (remote) CPU is still in the middle of schedule() with
4174	 * this task as prev, wait until it's done referencing the task.
4175	 *
4176	 * Pairs with the smp_store_release() in finish_task().
4177	 *
4178	 * This ensures that tasks getting woken will be fully ordered against
4179	 * their previous state and preserve Program Order.
4180	 */
4181	smp_cond_load_acquire(&p->on_cpu, !VAL);
4182
4183	cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4184	if (task_cpu(p) != cpu) {
4185		if (p->in_iowait) {
4186			delayacct_blkio_end(p);
4187			atomic_dec(&task_rq(p)->nr_iowait);
4188		}
4189
4190		wake_flags |= WF_MIGRATED;
4191		psi_ttwu_dequeue(p);
4192		set_task_cpu(p, cpu);
4193	}
4194#else
4195	cpu = task_cpu(p);
4196#endif /* CONFIG_SMP */
4197
4198	ttwu_queue(p, cpu, wake_flags);
4199unlock:
4200	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4201out:
4202	if (success)
4203		ttwu_stat(p, task_cpu(p), wake_flags);
4204	preempt_enable();
4205
4206	return success;
4207}
4208
4209/**
4210 * task_call_func - Invoke a function on task in fixed state
4211 * @p: Process for which the function is to be invoked, can be @current.
4212 * @func: Function to invoke.
4213 * @arg: Argument to function.
4214 *
4215 * Fix the task in it's current state by avoiding wakeups and or rq operations
4216 * and call @func(@arg) on it.  This function can use ->on_rq and task_curr()
4217 * to work out what the state is, if required.  Given that @func can be invoked
4218 * with a runqueue lock held, it had better be quite lightweight.
4219 *
4220 * Returns:
4221 *   Whatever @func returns
4222 */
4223int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4224{
4225	struct rq *rq = NULL;
4226	unsigned int state;
4227	struct rq_flags rf;
4228	int ret;
4229
4230	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4231
4232	state = READ_ONCE(p->__state);
4233
4234	/*
4235	 * Ensure we load p->on_rq after p->__state, otherwise it would be
4236	 * possible to, falsely, observe p->on_rq == 0.
4237	 *
4238	 * See try_to_wake_up() for a longer comment.
4239	 */
4240	smp_rmb();
4241
4242	/*
4243	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4244	 * the task is blocked. Make sure to check @state since ttwu() can drop
4245	 * locks at the end, see ttwu_queue_wakelist().
4246	 */
4247	if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4248		rq = __task_rq_lock(p, &rf);
4249
4250	/*
4251	 * At this point the task is pinned; either:
4252	 *  - blocked and we're holding off wakeups	 (pi->lock)
4253	 *  - woken, and we're holding off enqueue	 (rq->lock)
4254	 *  - queued, and we're holding off schedule	 (rq->lock)
4255	 *  - running, and we're holding off de-schedule (rq->lock)
4256	 *
4257	 * The called function (@func) can use: task_curr(), p->on_rq and
4258	 * p->__state to differentiate between these states.
4259	 */
4260	ret = func(p, arg);
4261
4262	if (rq)
4263		rq_unlock(rq, &rf);
4264
4265	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4266	return ret;
4267}
4268
4269/**
4270 * cpu_curr_snapshot - Return a snapshot of the currently running task
4271 * @cpu: The CPU on which to snapshot the task.
4272 *
4273 * Returns the task_struct pointer of the task "currently" running on
4274 * the specified CPU.  If the same task is running on that CPU throughout,
4275 * the return value will be a pointer to that task's task_struct structure.
4276 * If the CPU did any context switches even vaguely concurrently with the
4277 * execution of this function, the return value will be a pointer to the
4278 * task_struct structure of a randomly chosen task that was running on
4279 * that CPU somewhere around the time that this function was executing.
4280 *
4281 * If the specified CPU was offline, the return value is whatever it
4282 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4283 * task, but there is no guarantee.  Callers wishing a useful return
4284 * value must take some action to ensure that the specified CPU remains
4285 * online throughout.
4286 *
4287 * This function executes full memory barriers before and after fetching
4288 * the pointer, which permits the caller to confine this function's fetch
4289 * with respect to the caller's accesses to other shared variables.
4290 */
4291struct task_struct *cpu_curr_snapshot(int cpu)
4292{
4293	struct task_struct *t;
4294
4295	smp_mb(); /* Pairing determined by caller's synchronization design. */
4296	t = rcu_dereference(cpu_curr(cpu));
4297	smp_mb(); /* Pairing determined by caller's synchronization design. */
4298	return t;
4299}
4300
4301/**
4302 * wake_up_process - Wake up a specific process
4303 * @p: The process to be woken up.
4304 *
4305 * Attempt to wake up the nominated process and move it to the set of runnable
4306 * processes.
4307 *
4308 * Return: 1 if the process was woken up, 0 if it was already running.
4309 *
4310 * This function executes a full memory barrier before accessing the task state.
4311 */
4312int wake_up_process(struct task_struct *p)
4313{
4314	return try_to_wake_up(p, TASK_NORMAL, 0);
4315}
4316EXPORT_SYMBOL(wake_up_process);
4317
4318int wake_up_state(struct task_struct *p, unsigned int state)
4319{
4320	return try_to_wake_up(p, state, 0);
4321}
4322
4323/*
4324 * Perform scheduler related setup for a newly forked process p.
4325 * p is forked by current.
4326 *
4327 * __sched_fork() is basic setup used by init_idle() too:
4328 */
4329static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4330{
4331	p->on_rq			= 0;
4332
4333	p->se.on_rq			= 0;
4334	p->se.exec_start		= 0;
4335	p->se.sum_exec_runtime		= 0;
4336	p->se.prev_sum_exec_runtime	= 0;
4337	p->se.nr_migrations		= 0;
4338	p->se.vruntime			= 0;
4339	INIT_LIST_HEAD(&p->se.group_node);
4340
4341#ifdef CONFIG_FAIR_GROUP_SCHED
4342	p->se.cfs_rq			= NULL;
4343#endif
4344
4345#ifdef CONFIG_SCHEDSTATS
4346	/* Even if schedstat is disabled, there should not be garbage */
4347	memset(&p->stats, 0, sizeof(p->stats));
4348#endif
4349
4350	RB_CLEAR_NODE(&p->dl.rb_node);
4351	init_dl_task_timer(&p->dl);
4352	init_dl_inactive_task_timer(&p->dl);
4353	__dl_clear_params(p);
4354
4355	INIT_LIST_HEAD(&p->rt.run_list);
4356	p->rt.timeout		= 0;
4357	p->rt.time_slice	= sched_rr_timeslice;
4358	p->rt.on_rq		= 0;
4359	p->rt.on_list		= 0;
4360
4361#ifdef CONFIG_PREEMPT_NOTIFIERS
4362	INIT_HLIST_HEAD(&p->preempt_notifiers);
4363#endif
4364
4365#ifdef CONFIG_COMPACTION
4366	p->capture_control = NULL;
4367#endif
4368	init_numa_balancing(clone_flags, p);
4369#ifdef CONFIG_SMP
4370	p->wake_entry.u_flags = CSD_TYPE_TTWU;
4371	p->migration_pending = NULL;
4372#endif
4373}
4374
4375DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4376
4377#ifdef CONFIG_NUMA_BALANCING
4378
4379int sysctl_numa_balancing_mode;
4380
4381static void __set_numabalancing_state(bool enabled)
4382{
4383	if (enabled)
4384		static_branch_enable(&sched_numa_balancing);
4385	else
4386		static_branch_disable(&sched_numa_balancing);
4387}
4388
4389void set_numabalancing_state(bool enabled)
4390{
4391	if (enabled)
4392		sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4393	else
4394		sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4395	__set_numabalancing_state(enabled);
4396}
4397
4398#ifdef CONFIG_PROC_SYSCTL
4399int sysctl_numa_balancing(struct ctl_table *table, int write,
4400			  void *buffer, size_t *lenp, loff_t *ppos)
4401{
4402	struct ctl_table t;
4403	int err;
4404	int state = sysctl_numa_balancing_mode;
4405
4406	if (write && !capable(CAP_SYS_ADMIN))
4407		return -EPERM;
4408
4409	t = *table;
4410	t.data = &state;
4411	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4412	if (err < 0)
4413		return err;
4414	if (write) {
4415		sysctl_numa_balancing_mode = state;
4416		__set_numabalancing_state(state);
4417	}
4418	return err;
4419}
4420#endif
4421#endif
4422
4423#ifdef CONFIG_SCHEDSTATS
4424
4425DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4426
4427static void set_schedstats(bool enabled)
4428{
4429	if (enabled)
4430		static_branch_enable(&sched_schedstats);
4431	else
4432		static_branch_disable(&sched_schedstats);
4433}
4434
4435void force_schedstat_enabled(void)
4436{
4437	if (!schedstat_enabled()) {
4438		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4439		static_branch_enable(&sched_schedstats);
4440	}
4441}
4442
4443static int __init setup_schedstats(char *str)
4444{
4445	int ret = 0;
4446	if (!str)
4447		goto out;
4448
4449	if (!strcmp(str, "enable")) {
4450		set_schedstats(true);
4451		ret = 1;
4452	} else if (!strcmp(str, "disable")) {
4453		set_schedstats(false);
4454		ret = 1;
4455	}
4456out:
4457	if (!ret)
4458		pr_warn("Unable to parse schedstats=\n");
4459
4460	return ret;
4461}
4462__setup("schedstats=", setup_schedstats);
4463
4464#ifdef CONFIG_PROC_SYSCTL
4465static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4466		size_t *lenp, loff_t *ppos)
4467{
4468	struct ctl_table t;
4469	int err;
4470	int state = static_branch_likely(&sched_schedstats);
4471
4472	if (write && !capable(CAP_SYS_ADMIN))
4473		return -EPERM;
4474
4475	t = *table;
4476	t.data = &state;
4477	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4478	if (err < 0)
4479		return err;
4480	if (write)
4481		set_schedstats(state);
4482	return err;
4483}
4484#endif /* CONFIG_PROC_SYSCTL */
4485#endif /* CONFIG_SCHEDSTATS */
4486
4487#ifdef CONFIG_SYSCTL
4488static struct ctl_table sched_core_sysctls[] = {
4489#ifdef CONFIG_SCHEDSTATS
4490	{
4491		.procname       = "sched_schedstats",
4492		.data           = NULL,
4493		.maxlen         = sizeof(unsigned int),
4494		.mode           = 0644,
4495		.proc_handler   = sysctl_schedstats,
4496		.extra1         = SYSCTL_ZERO,
4497		.extra2         = SYSCTL_ONE,
4498	},
4499#endif /* CONFIG_SCHEDSTATS */
4500#ifdef CONFIG_UCLAMP_TASK
4501	{
4502		.procname       = "sched_util_clamp_min",
4503		.data           = &sysctl_sched_uclamp_util_min,
4504		.maxlen         = sizeof(unsigned int),
4505		.mode           = 0644,
4506		.proc_handler   = sysctl_sched_uclamp_handler,
4507	},
4508	{
4509		.procname       = "sched_util_clamp_max",
4510		.data           = &sysctl_sched_uclamp_util_max,
4511		.maxlen         = sizeof(unsigned int),
4512		.mode           = 0644,
4513		.proc_handler   = sysctl_sched_uclamp_handler,
4514	},
4515	{
4516		.procname       = "sched_util_clamp_min_rt_default",
4517		.data           = &sysctl_sched_uclamp_util_min_rt_default,
4518		.maxlen         = sizeof(unsigned int),
4519		.mode           = 0644,
4520		.proc_handler   = sysctl_sched_uclamp_handler,
4521	},
4522#endif /* CONFIG_UCLAMP_TASK */
4523	{}
4524};
4525static int __init sched_core_sysctl_init(void)
4526{
4527	register_sysctl_init("kernel", sched_core_sysctls);
4528	return 0;
4529}
4530late_initcall(sched_core_sysctl_init);
4531#endif /* CONFIG_SYSCTL */
4532
4533/*
4534 * fork()/clone()-time setup:
4535 */
4536int sched_fork(unsigned long clone_flags, struct task_struct *p)
4537{
4538	__sched_fork(clone_flags, p);
4539	/*
4540	 * We mark the process as NEW here. This guarantees that
4541	 * nobody will actually run it, and a signal or other external
4542	 * event cannot wake it up and insert it on the runqueue either.
4543	 */
4544	p->__state = TASK_NEW;
4545
4546	/*
4547	 * Make sure we do not leak PI boosting priority to the child.
4548	 */
4549	p->prio = current->normal_prio;
4550
4551	uclamp_fork(p);
4552
4553	/*
4554	 * Revert to default priority/policy on fork if requested.
4555	 */
4556	if (unlikely(p->sched_reset_on_fork)) {
4557		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4558			p->policy = SCHED_NORMAL;
4559			p->static_prio = NICE_TO_PRIO(0);
4560			p->rt_priority = 0;
4561		} else if (PRIO_TO_NICE(p->static_prio) < 0)
4562			p->static_prio = NICE_TO_PRIO(0);
4563
4564		p->prio = p->normal_prio = p->static_prio;
4565		set_load_weight(p, false);
4566
4567		/*
4568		 * We don't need the reset flag anymore after the fork. It has
4569		 * fulfilled its duty:
4570		 */
4571		p->sched_reset_on_fork = 0;
4572	}
4573
4574	if (dl_prio(p->prio))
4575		return -EAGAIN;
4576	else if (rt_prio(p->prio))
4577		p->sched_class = &rt_sched_class;
4578	else
4579		p->sched_class = &fair_sched_class;
4580
4581	init_entity_runnable_average(&p->se);
4582
4583
4584#ifdef CONFIG_SCHED_INFO
4585	if (likely(sched_info_on()))
4586		memset(&p->sched_info, 0, sizeof(p->sched_info));
4587#endif
4588#if defined(CONFIG_SMP)
4589	p->on_cpu = 0;
4590#endif
4591	init_task_preempt_count(p);
4592#ifdef CONFIG_SMP
4593	plist_node_init(&p->pushable_tasks, MAX_PRIO);
4594	RB_CLEAR_NODE(&p->pushable_dl_tasks);
4595#endif
4596	return 0;
4597}
4598
4599void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4600{
4601	unsigned long flags;
4602
4603	/*
4604	 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4605	 * required yet, but lockdep gets upset if rules are violated.
4606	 */
4607	raw_spin_lock_irqsave(&p->pi_lock, flags);
4608#ifdef CONFIG_CGROUP_SCHED
4609	if (1) {
4610		struct task_group *tg;
4611		tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4612				  struct task_group, css);
4613		tg = autogroup_task_group(p, tg);
4614		p->sched_task_group = tg;
4615	}
4616#endif
4617	rseq_migrate(p);
4618	/*
4619	 * We're setting the CPU for the first time, we don't migrate,
4620	 * so use __set_task_cpu().
4621	 */
4622	__set_task_cpu(p, smp_processor_id());
4623	if (p->sched_class->task_fork)
4624		p->sched_class->task_fork(p);
4625	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4626}
4627
4628void sched_post_fork(struct task_struct *p)
4629{
4630	uclamp_post_fork(p);
4631}
4632
4633unsigned long to_ratio(u64 period, u64 runtime)
4634{
4635	if (runtime == RUNTIME_INF)
4636		return BW_UNIT;
4637
4638	/*
4639	 * Doing this here saves a lot of checks in all
4640	 * the calling paths, and returning zero seems
4641	 * safe for them anyway.
4642	 */
4643	if (period == 0)
4644		return 0;
4645
4646	return div64_u64(runtime << BW_SHIFT, period);
4647}
4648
4649/*
4650 * wake_up_new_task - wake up a newly created task for the first time.
4651 *
4652 * This function will do some initial scheduler statistics housekeeping
4653 * that must be done for every newly created context, then puts the task
4654 * on the runqueue and wakes it.
4655 */
4656void wake_up_new_task(struct task_struct *p)
4657{
4658	struct rq_flags rf;
4659	struct rq *rq;
4660
4661	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4662	WRITE_ONCE(p->__state, TASK_RUNNING);
4663#ifdef CONFIG_SMP
4664	/*
4665	 * Fork balancing, do it here and not earlier because:
4666	 *  - cpus_ptr can change in the fork path
4667	 *  - any previously selected CPU might disappear through hotplug
4668	 *
4669	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4670	 * as we're not fully set-up yet.
4671	 */
4672	p->recent_used_cpu = task_cpu(p);
4673	rseq_migrate(p);
4674	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4675#endif
4676	rq = __task_rq_lock(p, &rf);
4677	update_rq_clock(rq);
4678	post_init_entity_util_avg(p);
4679
4680	activate_task(rq, p, ENQUEUE_NOCLOCK);
4681	trace_sched_wakeup_new(p);
4682	check_preempt_curr(rq, p, WF_FORK);
4683#ifdef CONFIG_SMP
4684	if (p->sched_class->task_woken) {
4685		/*
4686		 * Nothing relies on rq->lock after this, so it's fine to
4687		 * drop it.
4688		 */
4689		rq_unpin_lock(rq, &rf);
4690		p->sched_class->task_woken(rq, p);
4691		rq_repin_lock(rq, &rf);
4692	}
4693#endif
4694	task_rq_unlock(rq, p, &rf);
4695}
4696
4697#ifdef CONFIG_PREEMPT_NOTIFIERS
4698
4699static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4700
4701void preempt_notifier_inc(void)
4702{
4703	static_branch_inc(&preempt_notifier_key);
4704}
4705EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4706
4707void preempt_notifier_dec(void)
4708{
4709	static_branch_dec(&preempt_notifier_key);
4710}
4711EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4712
4713/**
4714 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4715 * @notifier: notifier struct to register
4716 */
4717void preempt_notifier_register(struct preempt_notifier *notifier)
4718{
4719	if (!static_branch_unlikely(&preempt_notifier_key))
4720		WARN(1, "registering preempt_notifier while notifiers disabled\n");
4721
4722	hlist_add_head(&notifier->link, &current->preempt_notifiers);
4723}
4724EXPORT_SYMBOL_GPL(preempt_notifier_register);
4725
4726/**
4727 * preempt_notifier_unregister - no longer interested in preemption notifications
4728 * @notifier: notifier struct to unregister
4729 *
4730 * This is *not* safe to call from within a preemption notifier.
4731 */
4732void preempt_notifier_unregister(struct preempt_notifier *notifier)
4733{
4734	hlist_del(&notifier->link);
4735}
4736EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4737
4738static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4739{
4740	struct preempt_notifier *notifier;
4741
4742	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4743		notifier->ops->sched_in(notifier, raw_smp_processor_id());
4744}
4745
4746static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4747{
4748	if (static_branch_unlikely(&preempt_notifier_key))
4749		__fire_sched_in_preempt_notifiers(curr);
4750}
4751
4752static void
4753__fire_sched_out_preempt_notifiers(struct task_struct *curr,
4754				   struct task_struct *next)
4755{
4756	struct preempt_notifier *notifier;
4757
4758	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4759		notifier->ops->sched_out(notifier, next);
4760}
4761
4762static __always_inline void
4763fire_sched_out_preempt_notifiers(struct task_struct *curr,
4764				 struct task_struct *next)
4765{
4766	if (static_branch_unlikely(&preempt_notifier_key))
4767		__fire_sched_out_preempt_notifiers(curr, next);
4768}
4769
4770#else /* !CONFIG_PREEMPT_NOTIFIERS */
4771
4772static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4773{
4774}
4775
4776static inline void
4777fire_sched_out_preempt_notifiers(struct task_struct *curr,
4778				 struct task_struct *next)
4779{
4780}
4781
4782#endif /* CONFIG_PREEMPT_NOTIFIERS */
4783
4784static inline void prepare_task(struct task_struct *next)
4785{
4786#ifdef CONFIG_SMP
4787	/*
4788	 * Claim the task as running, we do this before switching to it
4789	 * such that any running task will have this set.
4790	 *
4791	 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4792	 * its ordering comment.
4793	 */
4794	WRITE_ONCE(next->on_cpu, 1);
4795#endif
4796}
4797
4798static inline void finish_task(struct task_struct *prev)
4799{
4800#ifdef CONFIG_SMP
4801	/*
4802	 * This must be the very last reference to @prev from this CPU. After
4803	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4804	 * must ensure this doesn't happen until the switch is completely
4805	 * finished.
4806	 *
4807	 * In particular, the load of prev->state in finish_task_switch() must
4808	 * happen before this.
4809	 *
4810	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4811	 */
4812	smp_store_release(&prev->on_cpu, 0);
4813#endif
4814}
4815
4816#ifdef CONFIG_SMP
4817
4818static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4819{
4820	void (*func)(struct rq *rq);
4821	struct callback_head *next;
4822
4823	lockdep_assert_rq_held(rq);
4824
4825	while (head) {
4826		func = (void (*)(struct rq *))head->func;
4827		next = head->next;
4828		head->next = NULL;
4829		head = next;
4830
4831		func(rq);
4832	}
4833}
4834
4835static void balance_push(struct rq *rq);
4836
4837/*
4838 * balance_push_callback is a right abuse of the callback interface and plays
4839 * by significantly different rules.
4840 *
4841 * Where the normal balance_callback's purpose is to be ran in the same context
4842 * that queued it (only later, when it's safe to drop rq->lock again),
4843 * balance_push_callback is specifically targeted at __schedule().
4844 *
4845 * This abuse is tolerated because it places all the unlikely/odd cases behind
4846 * a single test, namely: rq->balance_callback == NULL.
4847 */
4848struct callback_head balance_push_callback = {
4849	.next = NULL,
4850	.func = (void (*)(struct callback_head *))balance_push,
4851};
4852
4853static inline struct callback_head *
4854__splice_balance_callbacks(struct rq *rq, bool split)
4855{
4856	struct callback_head *head = rq->balance_callback;
4857
4858	if (likely(!head))
4859		return NULL;
4860
4861	lockdep_assert_rq_held(rq);
4862	/*
4863	 * Must not take balance_push_callback off the list when
4864	 * splice_balance_callbacks() and balance_callbacks() are not
4865	 * in the same rq->lock section.
4866	 *
4867	 * In that case it would be possible for __schedule() to interleave
4868	 * and observe the list empty.
4869	 */
4870	if (split && head == &balance_push_callback)
4871		head = NULL;
4872	else
4873		rq->balance_callback = NULL;
4874
4875	return head;
4876}
4877
4878static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4879{
4880	return __splice_balance_callbacks(rq, true);
4881}
4882
4883static void __balance_callbacks(struct rq *rq)
4884{
4885	do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4886}
4887
4888static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4889{
4890	unsigned long flags;
4891
4892	if (unlikely(head)) {
4893		raw_spin_rq_lock_irqsave(rq, flags);
4894		do_balance_callbacks(rq, head);
4895		raw_spin_rq_unlock_irqrestore(rq, flags);
4896	}
4897}
4898
4899#else
4900
4901static inline void __balance_callbacks(struct rq *rq)
4902{
4903}
4904
4905static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4906{
4907	return NULL;
4908}
4909
4910static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4911{
4912}
4913
4914#endif
4915
4916static inline void
4917prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4918{
4919	/*
4920	 * Since the runqueue lock will be released by the next
4921	 * task (which is an invalid locking op but in the case
4922	 * of the scheduler it's an obvious special-case), so we
4923	 * do an early lockdep release here:
4924	 */
4925	rq_unpin_lock(rq, rf);
4926	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4927#ifdef CONFIG_DEBUG_SPINLOCK
4928	/* this is a valid case when another task releases the spinlock */
4929	rq_lockp(rq)->owner = next;
4930#endif
4931}
4932
4933static inline void finish_lock_switch(struct rq *rq)
4934{
4935	/*
4936	 * If we are tracking spinlock dependencies then we have to
4937	 * fix up the runqueue lock - which gets 'carried over' from
4938	 * prev into current:
4939	 */
4940	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4941	__balance_callbacks(rq);
4942	raw_spin_rq_unlock_irq(rq);
4943}
4944
4945/*
4946 * NOP if the arch has not defined these:
4947 */
4948
4949#ifndef prepare_arch_switch
4950# define prepare_arch_switch(next)	do { } while (0)
4951#endif
4952
4953#ifndef finish_arch_post_lock_switch
4954# define finish_arch_post_lock_switch()	do { } while (0)
4955#endif
4956
4957static inline void kmap_local_sched_out(void)
4958{
4959#ifdef CONFIG_KMAP_LOCAL
4960	if (unlikely(current->kmap_ctrl.idx))
4961		__kmap_local_sched_out();
4962#endif
4963}
4964
4965static inline void kmap_local_sched_in(void)
4966{
4967#ifdef CONFIG_KMAP_LOCAL
4968	if (unlikely(current->kmap_ctrl.idx))
4969		__kmap_local_sched_in();
4970#endif
4971}
4972
4973/**
4974 * prepare_task_switch - prepare to switch tasks
4975 * @rq: the runqueue preparing to switch
4976 * @prev: the current task that is being switched out
4977 * @next: the task we are going to switch to.
4978 *
4979 * This is called with the rq lock held and interrupts off. It must
4980 * be paired with a subsequent finish_task_switch after the context
4981 * switch.
4982 *
4983 * prepare_task_switch sets up locking and calls architecture specific
4984 * hooks.
4985 */
4986static inline void
4987prepare_task_switch(struct rq *rq, struct task_struct *prev,
4988		    struct task_struct *next)
4989{
4990	kcov_prepare_switch(prev);
4991	sched_info_switch(rq, prev, next);
4992	perf_event_task_sched_out(prev, next);
4993	rseq_preempt(prev);
4994	fire_sched_out_preempt_notifiers(prev, next);
4995	kmap_local_sched_out();
4996	prepare_task(next);
4997	prepare_arch_switch(next);
4998}
4999
5000/**
5001 * finish_task_switch - clean up after a task-switch
5002 * @prev: the thread we just switched away from.
5003 *
5004 * finish_task_switch must be called after the context switch, paired
5005 * with a prepare_task_switch call before the context switch.
5006 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5007 * and do any other architecture-specific cleanup actions.
5008 *
5009 * Note that we may have delayed dropping an mm in context_switch(). If
5010 * so, we finish that here outside of the runqueue lock. (Doing it
5011 * with the lock held can cause deadlocks; see schedule() for
5012 * details.)
5013 *
5014 * The context switch have flipped the stack from under us and restored the
5015 * local variables which were saved when this task called schedule() in the
5016 * past. prev == current is still correct but we need to recalculate this_rq
5017 * because prev may have moved to another CPU.
5018 */
5019static struct rq *finish_task_switch(struct task_struct *prev)
5020	__releases(rq->lock)
5021{
5022	struct rq *rq = this_rq();
5023	struct mm_struct *mm = rq->prev_mm;
5024	unsigned int prev_state;
5025
5026	/*
5027	 * The previous task will have left us with a preempt_count of 2
5028	 * because it left us after:
5029	 *
5030	 *	schedule()
5031	 *	  preempt_disable();			// 1
5032	 *	  __schedule()
5033	 *	    raw_spin_lock_irq(&rq->lock)	// 2
5034	 *
5035	 * Also, see FORK_PREEMPT_COUNT.
5036	 */
5037	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5038		      "corrupted preempt_count: %s/%d/0x%x\n",
5039		      current->comm, current->pid, preempt_count()))
5040		preempt_count_set(FORK_PREEMPT_COUNT);
5041
5042	rq->prev_mm = NULL;
5043
5044	/*
5045	 * A task struct has one reference for the use as "current".
5046	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5047	 * schedule one last time. The schedule call will never return, and
5048	 * the scheduled task must drop that reference.
5049	 *
5050	 * We must observe prev->state before clearing prev->on_cpu (in
5051	 * finish_task), otherwise a concurrent wakeup can get prev
5052	 * running on another CPU and we could rave with its RUNNING -> DEAD
5053	 * transition, resulting in a double drop.
5054	 */
5055	prev_state = READ_ONCE(prev->__state);
5056	vtime_task_switch(prev);
5057	perf_event_task_sched_in(prev, current);
5058	finish_task(prev);
5059	tick_nohz_task_switch();
5060	finish_lock_switch(rq);
5061	finish_arch_post_lock_switch();
5062	kcov_finish_switch(current);
5063	/*
5064	 * kmap_local_sched_out() is invoked with rq::lock held and
5065	 * interrupts disabled. There is no requirement for that, but the
5066	 * sched out code does not have an interrupt enabled section.
5067	 * Restoring the maps on sched in does not require interrupts being
5068	 * disabled either.
5069	 */
5070	kmap_local_sched_in();
5071
5072	fire_sched_in_preempt_notifiers(current);
5073	/*
5074	 * When switching through a kernel thread, the loop in
5075	 * membarrier_{private,global}_expedited() may have observed that
5076	 * kernel thread and not issued an IPI. It is therefore possible to
5077	 * schedule between user->kernel->user threads without passing though
5078	 * switch_mm(). Membarrier requires a barrier after storing to
5079	 * rq->curr, before returning to userspace, so provide them here:
5080	 *
5081	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5082	 *   provided by mmdrop(),
5083	 * - a sync_core for SYNC_CORE.
5084	 */
5085	if (mm) {
5086		membarrier_mm_sync_core_before_usermode(mm);
5087		mmdrop_sched(mm);
5088	}
5089	if (unlikely(prev_state == TASK_DEAD)) {
5090		if (prev->sched_class->task_dead)
5091			prev->sched_class->task_dead(prev);
5092
5093		/* Task is done with its stack. */
5094		put_task_stack(prev);
5095
5096		put_task_struct_rcu_user(prev);
5097	}
5098
5099	return rq;
5100}
5101
5102/**
5103 * schedule_tail - first thing a freshly forked thread must call.
5104 * @prev: the thread we just switched away from.
5105 */
5106asmlinkage __visible void schedule_tail(struct task_struct *prev)
5107	__releases(rq->lock)
5108{
5109	/*
5110	 * New tasks start with FORK_PREEMPT_COUNT, see there and
5111	 * finish_task_switch() for details.
5112	 *
5113	 * finish_task_switch() will drop rq->lock() and lower preempt_count
5114	 * and the preempt_enable() will end up enabling preemption (on
5115	 * PREEMPT_COUNT kernels).
5116	 */
5117
5118	finish_task_switch(prev);
5119	preempt_enable();
5120
5121	if (current->set_child_tid)
5122		put_user(task_pid_vnr(current), current->set_child_tid);
5123
5124	calculate_sigpending();
5125}
5126
5127/*
5128 * context_switch - switch to the new MM and the new thread's register state.
5129 */
5130static __always_inline struct rq *
5131context_switch(struct rq *rq, struct task_struct *prev,
5132	       struct task_struct *next, struct rq_flags *rf)
5133{
5134	prepare_task_switch(rq, prev, next);
5135
5136	/*
5137	 * For paravirt, this is coupled with an exit in switch_to to
5138	 * combine the page table reload and the switch backend into
5139	 * one hypercall.
5140	 */
5141	arch_start_context_switch(prev);
5142
5143	/*
5144	 * kernel -> kernel   lazy + transfer active
5145	 *   user -> kernel   lazy + mmgrab() active
5146	 *
5147	 * kernel ->   user   switch + mmdrop() active
5148	 *   user ->   user   switch
5149	 */
5150	if (!next->mm) {                                // to kernel
5151		enter_lazy_tlb(prev->active_mm, next);
5152
5153		next->active_mm = prev->active_mm;
5154		if (prev->mm)                           // from user
5155			mmgrab(prev->active_mm);
5156		else
5157			prev->active_mm = NULL;
5158	} else {                                        // to user
5159		membarrier_switch_mm(rq, prev->active_mm, next->mm);
5160		/*
5161		 * sys_membarrier() requires an smp_mb() between setting
5162		 * rq->curr / membarrier_switch_mm() and returning to userspace.
5163		 *
5164		 * The below provides this either through switch_mm(), or in
5165		 * case 'prev->active_mm == next->mm' through
5166		 * finish_task_switch()'s mmdrop().
5167		 */
5168		switch_mm_irqs_off(prev->active_mm, next->mm, next);
5169
5170		if (!prev->mm) {                        // from kernel
5171			/* will mmdrop() in finish_task_switch(). */
5172			rq->prev_mm = prev->active_mm;
5173			prev->active_mm = NULL;
5174		}
5175	}
5176
5177	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5178
5179	prepare_lock_switch(rq, next, rf);
5180
5181	/* Here we just switch the register state and the stack. */
5182	switch_to(prev, next, prev);
5183	barrier();
5184
5185	return finish_task_switch(prev);
5186}
5187
5188/*
5189 * nr_running and nr_context_switches:
5190 *
5191 * externally visible scheduler statistics: current number of runnable
5192 * threads, total number of context switches performed since bootup.
5193 */
5194unsigned int nr_running(void)
5195{
5196	unsigned int i, sum = 0;
5197
5198	for_each_online_cpu(i)
5199		sum += cpu_rq(i)->nr_running;
5200
5201	return sum;
5202}
5203
5204/*
5205 * Check if only the current task is running on the CPU.
5206 *
5207 * Caution: this function does not check that the caller has disabled
5208 * preemption, thus the result might have a time-of-check-to-time-of-use
5209 * race.  The caller is responsible to use it correctly, for example:
5210 *
5211 * - from a non-preemptible section (of course)
5212 *
5213 * - from a thread that is bound to a single CPU
5214 *
5215 * - in a loop with very short iterations (e.g. a polling loop)
5216 */
5217bool single_task_running(void)
5218{
5219	return raw_rq()->nr_running == 1;
5220}
5221EXPORT_SYMBOL(single_task_running);
5222
5223unsigned long long nr_context_switches(void)
5224{
5225	int i;
5226	unsigned long long sum = 0;
5227
5228	for_each_possible_cpu(i)
5229		sum += cpu_rq(i)->nr_switches;
5230
5231	return sum;
5232}
5233
5234/*
5235 * Consumers of these two interfaces, like for example the cpuidle menu
5236 * governor, are using nonsensical data. Preferring shallow idle state selection
5237 * for a CPU that has IO-wait which might not even end up running the task when
5238 * it does become runnable.
5239 */
5240
5241unsigned int nr_iowait_cpu(int cpu)
5242{
5243	return atomic_read(&cpu_rq(cpu)->nr_iowait);
5244}
5245
5246/*
5247 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5248 *
5249 * The idea behind IO-wait account is to account the idle time that we could
5250 * have spend running if it were not for IO. That is, if we were to improve the
5251 * storage performance, we'd have a proportional reduction in IO-wait time.
5252 *
5253 * This all works nicely on UP, where, when a task blocks on IO, we account
5254 * idle time as IO-wait, because if the storage were faster, it could've been
5255 * running and we'd not be idle.
5256 *
5257 * This has been extended to SMP, by doing the same for each CPU. This however
5258 * is broken.
5259 *
5260 * Imagine for instance the case where two tasks block on one CPU, only the one
5261 * CPU will have IO-wait accounted, while the other has regular idle. Even
5262 * though, if the storage were faster, both could've ran at the same time,
5263 * utilising both CPUs.
5264 *
5265 * This means, that when looking globally, the current IO-wait accounting on
5266 * SMP is a lower bound, by reason of under accounting.
5267 *
5268 * Worse, since the numbers are provided per CPU, they are sometimes
5269 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5270 * associated with any one particular CPU, it can wake to another CPU than it
5271 * blocked on. This means the per CPU IO-wait number is meaningless.
5272 *
5273 * Task CPU affinities can make all that even more 'interesting'.
5274 */
5275
5276unsigned int nr_iowait(void)
5277{
5278	unsigned int i, sum = 0;
5279
5280	for_each_possible_cpu(i)
5281		sum += nr_iowait_cpu(i);
5282
5283	return sum;
5284}
5285
5286#ifdef CONFIG_SMP
5287
5288/*
5289 * sched_exec - execve() is a valuable balancing opportunity, because at
5290 * this point the task has the smallest effective memory and cache footprint.
5291 */
5292void sched_exec(void)
5293{
5294	struct task_struct *p = current;
5295	unsigned long flags;
5296	int dest_cpu;
5297
5298	raw_spin_lock_irqsave(&p->pi_lock, flags);
5299	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5300	if (dest_cpu == smp_processor_id())
5301		goto unlock;
5302
5303	if (likely(cpu_active(dest_cpu))) {
5304		struct migration_arg arg = { p, dest_cpu };
5305
5306		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5307		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5308		return;
5309	}
5310unlock:
5311	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5312}
5313
5314#endif
5315
5316DEFINE_PER_CPU(struct kernel_stat, kstat);
5317DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5318
5319EXPORT_PER_CPU_SYMBOL(kstat);
5320EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5321
5322/*
5323 * The function fair_sched_class.update_curr accesses the struct curr
5324 * and its field curr->exec_start; when called from task_sched_runtime(),
5325 * we observe a high rate of cache misses in practice.
5326 * Prefetching this data results in improved performance.
5327 */
5328static inline void prefetch_curr_exec_start(struct task_struct *p)
5329{
5330#ifdef CONFIG_FAIR_GROUP_SCHED
5331	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5332#else
5333	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5334#endif
5335	prefetch(curr);
5336	prefetch(&curr->exec_start);
5337}
5338
5339/*
5340 * Return accounted runtime for the task.
5341 * In case the task is currently running, return the runtime plus current's
5342 * pending runtime that have not been accounted yet.
5343 */
5344unsigned long long task_sched_runtime(struct task_struct *p)
5345{
5346	struct rq_flags rf;
5347	struct rq *rq;
5348	u64 ns;
5349
5350#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5351	/*
5352	 * 64-bit doesn't need locks to atomically read a 64-bit value.
5353	 * So we have a optimization chance when the task's delta_exec is 0.
5354	 * Reading ->on_cpu is racy, but this is ok.
5355	 *
5356	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5357	 * If we race with it entering CPU, unaccounted time is 0. This is
5358	 * indistinguishable from the read occurring a few cycles earlier.
5359	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5360	 * been accounted, so we're correct here as well.
5361	 */
5362	if (!p->on_cpu || !task_on_rq_queued(p))
5363		return p->se.sum_exec_runtime;
5364#endif
5365
5366	rq = task_rq_lock(p, &rf);
5367	/*
5368	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
5369	 * project cycles that may never be accounted to this
5370	 * thread, breaking clock_gettime().
5371	 */
5372	if (task_current(rq, p) && task_on_rq_queued(p)) {
5373		prefetch_curr_exec_start(p);
5374		update_rq_clock(rq);
5375		p->sched_class->update_curr(rq);
5376	}
5377	ns = p->se.sum_exec_runtime;
5378	task_rq_unlock(rq, p, &rf);
5379
5380	return ns;
5381}
5382
5383#ifdef CONFIG_SCHED_DEBUG
5384static u64 cpu_resched_latency(struct rq *rq)
5385{
5386	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5387	u64 resched_latency, now = rq_clock(rq);
5388	static bool warned_once;
5389
5390	if (sysctl_resched_latency_warn_once && warned_once)
5391		return 0;
5392
5393	if (!need_resched() || !latency_warn_ms)
5394		return 0;
5395
5396	if (system_state == SYSTEM_BOOTING)
5397		return 0;
5398
5399	if (!rq->last_seen_need_resched_ns) {
5400		rq->last_seen_need_resched_ns = now;
5401		rq->ticks_without_resched = 0;
5402		return 0;
5403	}
5404
5405	rq->ticks_without_resched++;
5406	resched_latency = now - rq->last_seen_need_resched_ns;
5407	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5408		return 0;
5409
5410	warned_once = true;
5411
5412	return resched_latency;
5413}
5414
5415static int __init setup_resched_latency_warn_ms(char *str)
5416{
5417	long val;
5418
5419	if ((kstrtol(str, 0, &val))) {
5420		pr_warn("Unable to set resched_latency_warn_ms\n");
5421		return 1;
5422	}
5423
5424	sysctl_resched_latency_warn_ms = val;
5425	return 1;
5426}
5427__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5428#else
5429static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5430#endif /* CONFIG_SCHED_DEBUG */
5431
5432/*
5433 * This function gets called by the timer code, with HZ frequency.
5434 * We call it with interrupts disabled.
5435 */
5436void scheduler_tick(void)
5437{
5438	int cpu = smp_processor_id();
5439	struct rq *rq = cpu_rq(cpu);
5440	struct task_struct *curr = rq->curr;
5441	struct rq_flags rf;
5442	unsigned long thermal_pressure;
5443	u64 resched_latency;
5444
5445	arch_scale_freq_tick();
5446	sched_clock_tick();
5447
5448	rq_lock(rq, &rf);
5449
5450	update_rq_clock(rq);
5451	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5452	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5453	curr->sched_class->task_tick(rq, curr, 0);
5454	if (sched_feat(LATENCY_WARN))
5455		resched_latency = cpu_resched_latency(rq);
5456	calc_global_load_tick(rq);
5457	sched_core_tick(rq);
5458
5459	rq_unlock(rq, &rf);
5460
5461	if (sched_feat(LATENCY_WARN) && resched_latency)
5462		resched_latency_warn(cpu, resched_latency);
5463
5464	perf_event_task_tick();
5465
5466#ifdef CONFIG_SMP
5467	rq->idle_balance = idle_cpu(cpu);
5468	trigger_load_balance(rq);
5469#endif
5470}
5471
5472#ifdef CONFIG_NO_HZ_FULL
5473
5474struct tick_work {
5475	int			cpu;
5476	atomic_t		state;
5477	struct delayed_work	work;
5478};
5479/* Values for ->state, see diagram below. */
5480#define TICK_SCHED_REMOTE_OFFLINE	0
5481#define TICK_SCHED_REMOTE_OFFLINING	1
5482#define TICK_SCHED_REMOTE_RUNNING	2
5483
5484/*
5485 * State diagram for ->state:
5486 *
5487 *
5488 *          TICK_SCHED_REMOTE_OFFLINE
5489 *                    |   ^
5490 *                    |   |
5491 *                    |   | sched_tick_remote()
5492 *                    |   |
5493 *                    |   |
5494 *                    +--TICK_SCHED_REMOTE_OFFLINING
5495 *                    |   ^
5496 *                    |   |
5497 * sched_tick_start() |   | sched_tick_stop()
5498 *                    |   |
5499 *                    V   |
5500 *          TICK_SCHED_REMOTE_RUNNING
5501 *
5502 *
5503 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5504 * and sched_tick_start() are happy to leave the state in RUNNING.
5505 */
5506
5507static struct tick_work __percpu *tick_work_cpu;
5508
5509static void sched_tick_remote(struct work_struct *work)
5510{
5511	struct delayed_work *dwork = to_delayed_work(work);
5512	struct tick_work *twork = container_of(dwork, struct tick_work, work);
5513	int cpu = twork->cpu;
5514	struct rq *rq = cpu_rq(cpu);
5515	struct task_struct *curr;
5516	struct rq_flags rf;
5517	u64 delta;
5518	int os;
5519
5520	/*
5521	 * Handle the tick only if it appears the remote CPU is running in full
5522	 * dynticks mode. The check is racy by nature, but missing a tick or
5523	 * having one too much is no big deal because the scheduler tick updates
5524	 * statistics and checks timeslices in a time-independent way, regardless
5525	 * of when exactly it is running.
5526	 */
5527	if (!tick_nohz_tick_stopped_cpu(cpu))
5528		goto out_requeue;
5529
5530	rq_lock_irq(rq, &rf);
5531	curr = rq->curr;
5532	if (cpu_is_offline(cpu))
5533		goto out_unlock;
5534
5535	update_rq_clock(rq);
5536
5537	if (!is_idle_task(curr)) {
5538		/*
5539		 * Make sure the next tick runs within a reasonable
5540		 * amount of time.
5541		 */
5542		delta = rq_clock_task(rq) - curr->se.exec_start;
5543		WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5544	}
5545	curr->sched_class->task_tick(rq, curr, 0);
5546
5547	calc_load_nohz_remote(rq);
5548out_unlock:
5549	rq_unlock_irq(rq, &rf);
5550out_requeue:
5551
5552	/*
5553	 * Run the remote tick once per second (1Hz). This arbitrary
5554	 * frequency is large enough to avoid overload but short enough
5555	 * to keep scheduler internal stats reasonably up to date.  But
5556	 * first update state to reflect hotplug activity if required.
5557	 */
5558	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5559	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5560	if (os == TICK_SCHED_REMOTE_RUNNING)
5561		queue_delayed_work(system_unbound_wq, dwork, HZ);
5562}
5563
5564static void sched_tick_start(int cpu)
5565{
5566	int os;
5567	struct tick_work *twork;
5568
5569	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5570		return;
5571
5572	WARN_ON_ONCE(!tick_work_cpu);
5573
5574	twork = per_cpu_ptr(tick_work_cpu, cpu);
5575	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5576	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5577	if (os == TICK_SCHED_REMOTE_OFFLINE) {
5578		twork->cpu = cpu;
5579		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5580		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5581	}
5582}
5583
5584#ifdef CONFIG_HOTPLUG_CPU
5585static void sched_tick_stop(int cpu)
5586{
5587	struct tick_work *twork;
5588	int os;
5589
5590	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5591		return;
5592
5593	WARN_ON_ONCE(!tick_work_cpu);
5594
5595	twork = per_cpu_ptr(tick_work_cpu, cpu);
5596	/* There cannot be competing actions, but don't rely on stop-machine. */
5597	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5598	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5599	/* Don't cancel, as this would mess up the state machine. */
5600}
5601#endif /* CONFIG_HOTPLUG_CPU */
5602
5603int __init sched_tick_offload_init(void)
5604{
5605	tick_work_cpu = alloc_percpu(struct tick_work);
5606	BUG_ON(!tick_work_cpu);
5607	return 0;
5608}
5609
5610#else /* !CONFIG_NO_HZ_FULL */
5611static inline void sched_tick_start(int cpu) { }
5612static inline void sched_tick_stop(int cpu) { }
5613#endif
5614
5615#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5616				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5617/*
5618 * If the value passed in is equal to the current preempt count
5619 * then we just disabled preemption. Start timing the latency.
5620 */
5621static inline void preempt_latency_start(int val)
5622{
5623	if (preempt_count() == val) {
5624		unsigned long ip = get_lock_parent_ip();
5625#ifdef CONFIG_DEBUG_PREEMPT
5626		current->preempt_disable_ip = ip;
5627#endif
5628		trace_preempt_off(CALLER_ADDR0, ip);
5629	}
5630}
5631
5632void preempt_count_add(int val)
5633{
5634#ifdef CONFIG_DEBUG_PREEMPT
5635	/*
5636	 * Underflow?
5637	 */
5638	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5639		return;
5640#endif
5641	__preempt_count_add(val);
5642#ifdef CONFIG_DEBUG_PREEMPT
5643	/*
5644	 * Spinlock count overflowing soon?
5645	 */
5646	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5647				PREEMPT_MASK - 10);
5648#endif
5649	preempt_latency_start(val);
5650}
5651EXPORT_SYMBOL(preempt_count_add);
5652NOKPROBE_SYMBOL(preempt_count_add);
5653
5654/*
5655 * If the value passed in equals to the current preempt count
5656 * then we just enabled preemption. Stop timing the latency.
5657 */
5658static inline void preempt_latency_stop(int val)
5659{
5660	if (preempt_count() == val)
5661		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5662}
5663
5664void preempt_count_sub(int val)
5665{
5666#ifdef CONFIG_DEBUG_PREEMPT
5667	/*
5668	 * Underflow?
5669	 */
5670	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5671		return;
5672	/*
5673	 * Is the spinlock portion underflowing?
5674	 */
5675	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5676			!(preempt_count() & PREEMPT_MASK)))
5677		return;
5678#endif
5679
5680	preempt_latency_stop(val);
5681	__preempt_count_sub(val);
5682}
5683EXPORT_SYMBOL(preempt_count_sub);
5684NOKPROBE_SYMBOL(preempt_count_sub);
5685
5686#else
5687static inline void preempt_latency_start(int val) { }
5688static inline void preempt_latency_stop(int val) { }
5689#endif
5690
5691static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5692{
5693#ifdef CONFIG_DEBUG_PREEMPT
5694	return p->preempt_disable_ip;
5695#else
5696	return 0;
5697#endif
5698}
5699
5700/*
5701 * Print scheduling while atomic bug:
5702 */
5703static noinline void __schedule_bug(struct task_struct *prev)
5704{
5705	/* Save this before calling printk(), since that will clobber it */
5706	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5707
5708	if (oops_in_progress)
5709		return;
5710
5711	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5712		prev->comm, prev->pid, preempt_count());
5713
5714	debug_show_held_locks(prev);
5715	print_modules();
5716	if (irqs_disabled())
5717		print_irqtrace_events(prev);
5718	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5719	    && in_atomic_preempt_off()) {
5720		pr_err("Preemption disabled at:");
5721		print_ip_sym(KERN_ERR, preempt_disable_ip);
5722	}
5723	if (panic_on_warn)
5724		panic("scheduling while atomic\n");
5725
5726	dump_stack();
5727	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5728}
5729
5730/*
5731 * Various schedule()-time debugging checks and statistics:
5732 */
5733static inline void schedule_debug(struct task_struct *prev, bool preempt)
5734{
5735#ifdef CONFIG_SCHED_STACK_END_CHECK
5736	if (task_stack_end_corrupted(prev))
5737		panic("corrupted stack end detected inside scheduler\n");
5738
5739	if (task_scs_end_corrupted(prev))
5740		panic("corrupted shadow stack detected inside scheduler\n");
5741#endif
5742
5743#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5744	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5745		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5746			prev->comm, prev->pid, prev->non_block_count);
5747		dump_stack();
5748		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5749	}
5750#endif
5751
5752	if (unlikely(in_atomic_preempt_off())) {
5753		__schedule_bug(prev);
5754		preempt_count_set(PREEMPT_DISABLED);
5755	}
5756	rcu_sleep_check();
5757	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5758
5759	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5760
5761	schedstat_inc(this_rq()->sched_count);
5762}
5763
5764static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5765				  struct rq_flags *rf)
5766{
5767#ifdef CONFIG_SMP
5768	const struct sched_class *class;
5769	/*
5770	 * We must do the balancing pass before put_prev_task(), such
5771	 * that when we release the rq->lock the task is in the same
5772	 * state as before we took rq->lock.
5773	 *
5774	 * We can terminate the balance pass as soon as we know there is
5775	 * a runnable task of @class priority or higher.
5776	 */
5777	for_class_range(class, prev->sched_class, &idle_sched_class) {
5778		if (class->balance(rq, prev, rf))
5779			break;
5780	}
5781#endif
5782
5783	put_prev_task(rq, prev);
5784}
5785
5786/*
5787 * Pick up the highest-prio task:
5788 */
5789static inline struct task_struct *
5790__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5791{
5792	const struct sched_class *class;
5793	struct task_struct *p;
5794
5795	/*
5796	 * Optimization: we know that if all tasks are in the fair class we can
5797	 * call that function directly, but only if the @prev task wasn't of a
5798	 * higher scheduling class, because otherwise those lose the
5799	 * opportunity to pull in more work from other CPUs.
5800	 */
5801	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5802		   rq->nr_running == rq->cfs.h_nr_running)) {
5803
5804		p = pick_next_task_fair(rq, prev, rf);
5805		if (unlikely(p == RETRY_TASK))
5806			goto restart;
5807
5808		/* Assume the next prioritized class is idle_sched_class */
5809		if (!p) {
5810			put_prev_task(rq, prev);
5811			p = pick_next_task_idle(rq);
5812		}
5813
5814		return p;
5815	}
5816
5817restart:
5818	put_prev_task_balance(rq, prev, rf);
5819
5820	for_each_class(class) {
5821		p = class->pick_next_task(rq);
5822		if (p)
5823			return p;
5824	}
5825
5826	BUG(); /* The idle class should always have a runnable task. */
5827}
5828
5829#ifdef CONFIG_SCHED_CORE
5830static inline bool is_task_rq_idle(struct task_struct *t)
5831{
5832	return (task_rq(t)->idle == t);
5833}
5834
5835static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5836{
5837	return is_task_rq_idle(a) || (a->core_cookie == cookie);
5838}
5839
5840static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5841{
5842	if (is_task_rq_idle(a) || is_task_rq_idle(b))
5843		return true;
5844
5845	return a->core_cookie == b->core_cookie;
5846}
5847
5848static inline struct task_struct *pick_task(struct rq *rq)
5849{
5850	const struct sched_class *class;
5851	struct task_struct *p;
5852
5853	for_each_class(class) {
5854		p = class->pick_task(rq);
5855		if (p)
5856			return p;
5857	}
5858
5859	BUG(); /* The idle class should always have a runnable task. */
5860}
5861
5862extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5863
5864static void queue_core_balance(struct rq *rq);
5865
5866static struct task_struct *
5867pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5868{
5869	struct task_struct *next, *p, *max = NULL;
5870	const struct cpumask *smt_mask;
5871	bool fi_before = false;
5872	bool core_clock_updated = (rq == rq->core);
5873	unsigned long cookie;
5874	int i, cpu, occ = 0;
5875	struct rq *rq_i;
5876	bool need_sync;
5877
5878	if (!sched_core_enabled(rq))
5879		return __pick_next_task(rq, prev, rf);
5880
5881	cpu = cpu_of(rq);
5882
5883	/* Stopper task is switching into idle, no need core-wide selection. */
5884	if (cpu_is_offline(cpu)) {
5885		/*
5886		 * Reset core_pick so that we don't enter the fastpath when
5887		 * coming online. core_pick would already be migrated to
5888		 * another cpu during offline.
5889		 */
5890		rq->core_pick = NULL;
5891		return __pick_next_task(rq, prev, rf);
5892	}
5893
5894	/*
5895	 * If there were no {en,de}queues since we picked (IOW, the task
5896	 * pointers are all still valid), and we haven't scheduled the last
5897	 * pick yet, do so now.
5898	 *
5899	 * rq->core_pick can be NULL if no selection was made for a CPU because
5900	 * it was either offline or went offline during a sibling's core-wide
5901	 * selection. In this case, do a core-wide selection.
5902	 */
5903	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5904	    rq->core->core_pick_seq != rq->core_sched_seq &&
5905	    rq->core_pick) {
5906		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5907
5908		next = rq->core_pick;
5909		if (next != prev) {
5910			put_prev_task(rq, prev);
5911			set_next_task(rq, next);
5912		}
5913
5914		rq->core_pick = NULL;
5915		goto out;
5916	}
5917
5918	put_prev_task_balance(rq, prev, rf);
5919
5920	smt_mask = cpu_smt_mask(cpu);
5921	need_sync = !!rq->core->core_cookie;
5922
5923	/* reset state */
5924	rq->core->core_cookie = 0UL;
5925	if (rq->core->core_forceidle_count) {
5926		if (!core_clock_updated) {
5927			update_rq_clock(rq->core);
5928			core_clock_updated = true;
5929		}
5930		sched_core_account_forceidle(rq);
5931		/* reset after accounting force idle */
5932		rq->core->core_forceidle_start = 0;
5933		rq->core->core_forceidle_count = 0;
5934		rq->core->core_forceidle_occupation = 0;
5935		need_sync = true;
5936		fi_before = true;
5937	}
5938
5939	/*
5940	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5941	 *
5942	 * @task_seq guards the task state ({en,de}queues)
5943	 * @pick_seq is the @task_seq we did a selection on
5944	 * @sched_seq is the @pick_seq we scheduled
5945	 *
5946	 * However, preemptions can cause multiple picks on the same task set.
5947	 * 'Fix' this by also increasing @task_seq for every pick.
5948	 */
5949	rq->core->core_task_seq++;
5950
5951	/*
5952	 * Optimize for common case where this CPU has no cookies
5953	 * and there are no cookied tasks running on siblings.
5954	 */
5955	if (!need_sync) {
5956		next = pick_task(rq);
5957		if (!next->core_cookie) {
5958			rq->core_pick = NULL;
5959			/*
5960			 * For robustness, update the min_vruntime_fi for
5961			 * unconstrained picks as well.
5962			 */
5963			WARN_ON_ONCE(fi_before);
5964			task_vruntime_update(rq, next, false);
5965			goto out_set_next;
5966		}
5967	}
5968
5969	/*
5970	 * For each thread: do the regular task pick and find the max prio task
5971	 * amongst them.
5972	 *
5973	 * Tie-break prio towards the current CPU
5974	 */
5975	for_each_cpu_wrap(i, smt_mask, cpu) {
5976		rq_i = cpu_rq(i);
5977
5978		/*
5979		 * Current cpu always has its clock updated on entrance to
5980		 * pick_next_task(). If the current cpu is not the core,
5981		 * the core may also have been updated above.
5982		 */
5983		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5984			update_rq_clock(rq_i);
5985
5986		p = rq_i->core_pick = pick_task(rq_i);
5987		if (!max || prio_less(max, p, fi_before))
5988			max = p;
5989	}
5990
5991	cookie = rq->core->core_cookie = max->core_cookie;
5992
5993	/*
5994	 * For each thread: try and find a runnable task that matches @max or
5995	 * force idle.
5996	 */
5997	for_each_cpu(i, smt_mask) {
5998		rq_i = cpu_rq(i);
5999		p = rq_i->core_pick;
6000
6001		if (!cookie_equals(p, cookie)) {
6002			p = NULL;
6003			if (cookie)
6004				p = sched_core_find(rq_i, cookie);
6005			if (!p)
6006				p = idle_sched_class.pick_task(rq_i);
6007		}
6008
6009		rq_i->core_pick = p;
6010
6011		if (p == rq_i->idle) {
6012			if (rq_i->nr_running) {
6013				rq->core->core_forceidle_count++;
6014				if (!fi_before)
6015					rq->core->core_forceidle_seq++;
6016			}
6017		} else {
6018			occ++;
6019		}
6020	}
6021
6022	if (schedstat_enabled() && rq->core->core_forceidle_count) {
6023		rq->core->core_forceidle_start = rq_clock(rq->core);
6024		rq->core->core_forceidle_occupation = occ;
6025	}
6026
6027	rq->core->core_pick_seq = rq->core->core_task_seq;
6028	next = rq->core_pick;
6029	rq->core_sched_seq = rq->core->core_pick_seq;
6030
6031	/* Something should have been selected for current CPU */
6032	WARN_ON_ONCE(!next);
6033
6034	/*
6035	 * Reschedule siblings
6036	 *
6037	 * NOTE: L1TF -- at this point we're no longer running the old task and
6038	 * sending an IPI (below) ensures the sibling will no longer be running
6039	 * their task. This ensures there is no inter-sibling overlap between
6040	 * non-matching user state.
6041	 */
6042	for_each_cpu(i, smt_mask) {
6043		rq_i = cpu_rq(i);
6044
6045		/*
6046		 * An online sibling might have gone offline before a task
6047		 * could be picked for it, or it might be offline but later
6048		 * happen to come online, but its too late and nothing was
6049		 * picked for it.  That's Ok - it will pick tasks for itself,
6050		 * so ignore it.
6051		 */
6052		if (!rq_i->core_pick)
6053			continue;
6054
6055		/*
6056		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6057		 * fi_before     fi      update?
6058		 *  0            0       1
6059		 *  0            1       1
6060		 *  1            0       1
6061		 *  1            1       0
6062		 */
6063		if (!(fi_before && rq->core->core_forceidle_count))
6064			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6065
6066		rq_i->core_pick->core_occupation = occ;
6067
6068		if (i == cpu) {
6069			rq_i->core_pick = NULL;
6070			continue;
6071		}
6072
6073		/* Did we break L1TF mitigation requirements? */
6074		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6075
6076		if (rq_i->curr == rq_i->core_pick) {
6077			rq_i->core_pick = NULL;
6078			continue;
6079		}
6080
6081		resched_curr(rq_i);
6082	}
6083
6084out_set_next:
6085	set_next_task(rq, next);
6086out:
6087	if (rq->core->core_forceidle_count && next == rq->idle)
6088		queue_core_balance(rq);
6089
6090	return next;
6091}
6092
6093static bool try_steal_cookie(int this, int that)
6094{
6095	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6096	struct task_struct *p;
6097	unsigned long cookie;
6098	bool success = false;
6099
6100	local_irq_disable();
6101	double_rq_lock(dst, src);
6102
6103	cookie = dst->core->core_cookie;
6104	if (!cookie)
6105		goto unlock;
6106
6107	if (dst->curr != dst->idle)
6108		goto unlock;
6109
6110	p = sched_core_find(src, cookie);
6111	if (p == src->idle)
6112		goto unlock;
6113
6114	do {
6115		if (p == src->core_pick || p == src->curr)
6116			goto next;
6117
6118		if (!is_cpu_allowed(p, this))
6119			goto next;
6120
6121		if (p->core_occupation > dst->idle->core_occupation)
6122			goto next;
6123
6124		deactivate_task(src, p, 0);
6125		set_task_cpu(p, this);
6126		activate_task(dst, p, 0);
6127
6128		resched_curr(dst);
6129
6130		success = true;
6131		break;
6132
6133next:
6134		p = sched_core_next(p, cookie);
6135	} while (p);
6136
6137unlock:
6138	double_rq_unlock(dst, src);
6139	local_irq_enable();
6140
6141	return success;
6142}
6143
6144static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6145{
6146	int i;
6147
6148	for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6149		if (i == cpu)
6150			continue;
6151
6152		if (need_resched())
6153			break;
6154
6155		if (try_steal_cookie(cpu, i))
6156			return true;
6157	}
6158
6159	return false;
6160}
6161
6162static void sched_core_balance(struct rq *rq)
6163{
6164	struct sched_domain *sd;
6165	int cpu = cpu_of(rq);
6166
6167	preempt_disable();
6168	rcu_read_lock();
6169	raw_spin_rq_unlock_irq(rq);
6170	for_each_domain(cpu, sd) {
6171		if (need_resched())
6172			break;
6173
6174		if (steal_cookie_task(cpu, sd))
6175			break;
6176	}
6177	raw_spin_rq_lock_irq(rq);
6178	rcu_read_unlock();
6179	preempt_enable();
6180}
6181
6182static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6183
6184static void queue_core_balance(struct rq *rq)
6185{
6186	if (!sched_core_enabled(rq))
6187		return;
6188
6189	if (!rq->core->core_cookie)
6190		return;
6191
6192	if (!rq->nr_running) /* not forced idle */
6193		return;
6194
6195	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6196}
6197
6198static void sched_core_cpu_starting(unsigned int cpu)
6199{
6200	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6201	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6202	unsigned long flags;
6203	int t;
6204
6205	sched_core_lock(cpu, &flags);
6206
6207	WARN_ON_ONCE(rq->core != rq);
6208
6209	/* if we're the first, we'll be our own leader */
6210	if (cpumask_weight(smt_mask) == 1)
6211		goto unlock;
6212
6213	/* find the leader */
6214	for_each_cpu(t, smt_mask) {
6215		if (t == cpu)
6216			continue;
6217		rq = cpu_rq(t);
6218		if (rq->core == rq) {
6219			core_rq = rq;
6220			break;
6221		}
6222	}
6223
6224	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6225		goto unlock;
6226
6227	/* install and validate core_rq */
6228	for_each_cpu(t, smt_mask) {
6229		rq = cpu_rq(t);
6230
6231		if (t == cpu)
6232			rq->core = core_rq;
6233
6234		WARN_ON_ONCE(rq->core != core_rq);
6235	}
6236
6237unlock:
6238	sched_core_unlock(cpu, &flags);
6239}
6240
6241static void sched_core_cpu_deactivate(unsigned int cpu)
6242{
6243	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6244	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6245	unsigned long flags;
6246	int t;
6247
6248	sched_core_lock(cpu, &flags);
6249
6250	/* if we're the last man standing, nothing to do */
6251	if (cpumask_weight(smt_mask) == 1) {
6252		WARN_ON_ONCE(rq->core != rq);
6253		goto unlock;
6254	}
6255
6256	/* if we're not the leader, nothing to do */
6257	if (rq->core != rq)
6258		goto unlock;
6259
6260	/* find a new leader */
6261	for_each_cpu(t, smt_mask) {
6262		if (t == cpu)
6263			continue;
6264		core_rq = cpu_rq(t);
6265		break;
6266	}
6267
6268	if (WARN_ON_ONCE(!core_rq)) /* impossible */
6269		goto unlock;
6270
6271	/* copy the shared state to the new leader */
6272	core_rq->core_task_seq             = rq->core_task_seq;
6273	core_rq->core_pick_seq             = rq->core_pick_seq;
6274	core_rq->core_cookie               = rq->core_cookie;
6275	core_rq->core_forceidle_count      = rq->core_forceidle_count;
6276	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
6277	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6278
6279	/*
6280	 * Accounting edge for forced idle is handled in pick_next_task().
6281	 * Don't need another one here, since the hotplug thread shouldn't
6282	 * have a cookie.
6283	 */
6284	core_rq->core_forceidle_start = 0;
6285
6286	/* install new leader */
6287	for_each_cpu(t, smt_mask) {
6288		rq = cpu_rq(t);
6289		rq->core = core_rq;
6290	}
6291
6292unlock:
6293	sched_core_unlock(cpu, &flags);
6294}
6295
6296static inline void sched_core_cpu_dying(unsigned int cpu)
6297{
6298	struct rq *rq = cpu_rq(cpu);
6299
6300	if (rq->core != rq)
6301		rq->core = rq;
6302}
6303
6304#else /* !CONFIG_SCHED_CORE */
6305
6306static inline void sched_core_cpu_starting(unsigned int cpu) {}
6307static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6308static inline void sched_core_cpu_dying(unsigned int cpu) {}
6309
6310static struct task_struct *
6311pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6312{
6313	return __pick_next_task(rq, prev, rf);
6314}
6315
6316#endif /* CONFIG_SCHED_CORE */
6317
6318/*
6319 * Constants for the sched_mode argument of __schedule().
6320 *
6321 * The mode argument allows RT enabled kernels to differentiate a
6322 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6323 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6324 * optimize the AND operation out and just check for zero.
6325 */
6326#define SM_NONE			0x0
6327#define SM_PREEMPT		0x1
6328#define SM_RTLOCK_WAIT		0x2
6329
6330#ifndef CONFIG_PREEMPT_RT
6331# define SM_MASK_PREEMPT	(~0U)
6332#else
6333# define SM_MASK_PREEMPT	SM_PREEMPT
6334#endif
6335
6336/*
6337 * __schedule() is the main scheduler function.
6338 *
6339 * The main means of driving the scheduler and thus entering this function are:
6340 *
6341 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6342 *
6343 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6344 *      paths. For example, see arch/x86/entry_64.S.
6345 *
6346 *      To drive preemption between tasks, the scheduler sets the flag in timer
6347 *      interrupt handler scheduler_tick().
6348 *
6349 *   3. Wakeups don't really cause entry into schedule(). They add a
6350 *      task to the run-queue and that's it.
6351 *
6352 *      Now, if the new task added to the run-queue preempts the current
6353 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6354 *      called on the nearest possible occasion:
6355 *
6356 *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6357 *
6358 *         - in syscall or exception context, at the next outmost
6359 *           preempt_enable(). (this might be as soon as the wake_up()'s
6360 *           spin_unlock()!)
6361 *
6362 *         - in IRQ context, return from interrupt-handler to
6363 *           preemptible context
6364 *
6365 *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6366 *         then at the next:
6367 *
6368 *          - cond_resched() call
6369 *          - explicit schedule() call
6370 *          - return from syscall or exception to user-space
6371 *          - return from interrupt-handler to user-space
6372 *
6373 * WARNING: must be called with preemption disabled!
6374 */
6375static void __sched notrace __schedule(unsigned int sched_mode)
6376{
6377	struct task_struct *prev, *next;
6378	unsigned long *switch_count;
6379	unsigned long prev_state;
6380	struct rq_flags rf;
6381	struct rq *rq;
6382	int cpu;
6383
6384	cpu = smp_processor_id();
6385	rq = cpu_rq(cpu);
6386	prev = rq->curr;
6387
6388	schedule_debug(prev, !!sched_mode);
6389
6390	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6391		hrtick_clear(rq);
6392
6393	local_irq_disable();
6394	rcu_note_context_switch(!!sched_mode);
6395
6396	/*
6397	 * Make sure that signal_pending_state()->signal_pending() below
6398	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6399	 * done by the caller to avoid the race with signal_wake_up():
6400	 *
6401	 * __set_current_state(@state)		signal_wake_up()
6402	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
6403	 *					  wake_up_state(p, state)
6404	 *   LOCK rq->lock			    LOCK p->pi_state
6405	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
6406	 *     if (signal_pending_state())	    if (p->state & @state)
6407	 *
6408	 * Also, the membarrier system call requires a full memory barrier
6409	 * after coming from user-space, before storing to rq->curr.
6410	 */
6411	rq_lock(rq, &rf);
6412	smp_mb__after_spinlock();
6413
6414	/* Promote REQ to ACT */
6415	rq->clock_update_flags <<= 1;
6416	update_rq_clock(rq);
6417
6418	switch_count = &prev->nivcsw;
6419
6420	/*
6421	 * We must load prev->state once (task_struct::state is volatile), such
6422	 * that we form a control dependency vs deactivate_task() below.
6423	 */
6424	prev_state = READ_ONCE(prev->__state);
6425	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6426		if (signal_pending_state(prev_state, prev)) {
6427			WRITE_ONCE(prev->__state, TASK_RUNNING);
6428		} else {
6429			prev->sched_contributes_to_load =
6430				(prev_state & TASK_UNINTERRUPTIBLE) &&
6431				!(prev_state & TASK_NOLOAD) &&
6432				!(prev->flags & PF_FROZEN);
6433
6434			if (prev->sched_contributes_to_load)
6435				rq->nr_uninterruptible++;
6436
6437			/*
6438			 * __schedule()			ttwu()
6439			 *   prev_state = prev->state;    if (p->on_rq && ...)
6440			 *   if (prev_state)		    goto out;
6441			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
6442			 *				  p->state = TASK_WAKING
6443			 *
6444			 * Where __schedule() and ttwu() have matching control dependencies.
6445			 *
6446			 * After this, schedule() must not care about p->state any more.
6447			 */
6448			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6449
6450			if (prev->in_iowait) {
6451				atomic_inc(&rq->nr_iowait);
6452				delayacct_blkio_start();
6453			}
6454		}
6455		switch_count = &prev->nvcsw;
6456	}
6457
6458	next = pick_next_task(rq, prev, &rf);
6459	clear_tsk_need_resched(prev);
6460	clear_preempt_need_resched();
6461#ifdef CONFIG_SCHED_DEBUG
6462	rq->last_seen_need_resched_ns = 0;
6463#endif
6464
6465	if (likely(prev != next)) {
6466		rq->nr_switches++;
6467		/*
6468		 * RCU users of rcu_dereference(rq->curr) may not see
6469		 * changes to task_struct made by pick_next_task().
6470		 */
6471		RCU_INIT_POINTER(rq->curr, next);
6472		/*
6473		 * The membarrier system call requires each architecture
6474		 * to have a full memory barrier after updating
6475		 * rq->curr, before returning to user-space.
6476		 *
6477		 * Here are the schemes providing that barrier on the
6478		 * various architectures:
6479		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6480		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6481		 * - finish_lock_switch() for weakly-ordered
6482		 *   architectures where spin_unlock is a full barrier,
6483		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6484		 *   is a RELEASE barrier),
6485		 */
6486		++*switch_count;
6487
6488		migrate_disable_switch(rq, prev);
6489		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6490
6491		trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6492
6493		/* Also unlocks the rq: */
6494		rq = context_switch(rq, prev, next, &rf);
6495	} else {
6496		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6497
6498		rq_unpin_lock(rq, &rf);
6499		__balance_callbacks(rq);
6500		raw_spin_rq_unlock_irq(rq);
6501	}
6502}
6503
6504void __noreturn do_task_dead(void)
6505{
6506	/* Causes final put_task_struct in finish_task_switch(): */
6507	set_special_state(TASK_DEAD);
6508
6509	/* Tell freezer to ignore us: */
6510	current->flags |= PF_NOFREEZE;
6511
6512	__schedule(SM_NONE);
6513	BUG();
6514
6515	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6516	for (;;)
6517		cpu_relax();
6518}
6519
6520static inline void sched_submit_work(struct task_struct *tsk)
6521{
6522	unsigned int task_flags;
6523
6524	if (task_is_running(tsk))
6525		return;
6526
6527	task_flags = tsk->flags;
6528	/*
6529	 * If a worker goes to sleep, notify and ask workqueue whether it
6530	 * wants to wake up a task to maintain concurrency.
6531	 */
6532	if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6533		if (task_flags & PF_WQ_WORKER)
6534			wq_worker_sleeping(tsk);
6535		else
6536			io_wq_worker_sleeping(tsk);
6537	}
6538
6539	/*
6540	 * spinlock and rwlock must not flush block requests.  This will
6541	 * deadlock if the callback attempts to acquire a lock which is
6542	 * already acquired.
6543	 */
6544	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6545
6546	/*
6547	 * If we are going to sleep and we have plugged IO queued,
6548	 * make sure to submit it to avoid deadlocks.
6549	 */
6550	blk_flush_plug(tsk->plug, true);
6551}
6552
6553static void sched_update_worker(struct task_struct *tsk)
6554{
6555	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6556		if (tsk->flags & PF_WQ_WORKER)
6557			wq_worker_running(tsk);
6558		else
6559			io_wq_worker_running(tsk);
6560	}
6561}
6562
6563asmlinkage __visible void __sched schedule(void)
6564{
6565	struct task_struct *tsk = current;
6566
6567	sched_submit_work(tsk);
6568	do {
6569		preempt_disable();
6570		__schedule(SM_NONE);
6571		sched_preempt_enable_no_resched();
6572	} while (need_resched());
6573	sched_update_worker(tsk);
6574}
6575EXPORT_SYMBOL(schedule);
6576
6577/*
6578 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6579 * state (have scheduled out non-voluntarily) by making sure that all
6580 * tasks have either left the run queue or have gone into user space.
6581 * As idle tasks do not do either, they must not ever be preempted
6582 * (schedule out non-voluntarily).
6583 *
6584 * schedule_idle() is similar to schedule_preempt_disable() except that it
6585 * never enables preemption because it does not call sched_submit_work().
6586 */
6587void __sched schedule_idle(void)
6588{
6589	/*
6590	 * As this skips calling sched_submit_work(), which the idle task does
6591	 * regardless because that function is a nop when the task is in a
6592	 * TASK_RUNNING state, make sure this isn't used someplace that the
6593	 * current task can be in any other state. Note, idle is always in the
6594	 * TASK_RUNNING state.
6595	 */
6596	WARN_ON_ONCE(current->__state);
6597	do {
6598		__schedule(SM_NONE);
6599	} while (need_resched());
6600}
6601
6602#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6603asmlinkage __visible void __sched schedule_user(void)
6604{
6605	/*
6606	 * If we come here after a random call to set_need_resched(),
6607	 * or we have been woken up remotely but the IPI has not yet arrived,
6608	 * we haven't yet exited the RCU idle mode. Do it here manually until
6609	 * we find a better solution.
6610	 *
6611	 * NB: There are buggy callers of this function.  Ideally we
6612	 * should warn if prev_state != CONTEXT_USER, but that will trigger
6613	 * too frequently to make sense yet.
6614	 */
6615	enum ctx_state prev_state = exception_enter();
6616	schedule();
6617	exception_exit(prev_state);
6618}
6619#endif
6620
6621/**
6622 * schedule_preempt_disabled - called with preemption disabled
6623 *
6624 * Returns with preemption disabled. Note: preempt_count must be 1
6625 */
6626void __sched schedule_preempt_disabled(void)
6627{
6628	sched_preempt_enable_no_resched();
6629	schedule();
6630	preempt_disable();
6631}
6632
6633#ifdef CONFIG_PREEMPT_RT
6634void __sched notrace schedule_rtlock(void)
6635{
6636	do {
6637		preempt_disable();
6638		__schedule(SM_RTLOCK_WAIT);
6639		sched_preempt_enable_no_resched();
6640	} while (need_resched());
6641}
6642NOKPROBE_SYMBOL(schedule_rtlock);
6643#endif
6644
6645static void __sched notrace preempt_schedule_common(void)
6646{
6647	do {
6648		/*
6649		 * Because the function tracer can trace preempt_count_sub()
6650		 * and it also uses preempt_enable/disable_notrace(), if
6651		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6652		 * by the function tracer will call this function again and
6653		 * cause infinite recursion.
6654		 *
6655		 * Preemption must be disabled here before the function
6656		 * tracer can trace. Break up preempt_disable() into two
6657		 * calls. One to disable preemption without fear of being
6658		 * traced. The other to still record the preemption latency,
6659		 * which can also be traced by the function tracer.
6660		 */
6661		preempt_disable_notrace();
6662		preempt_latency_start(1);
6663		__schedule(SM_PREEMPT);
6664		preempt_latency_stop(1);
6665		preempt_enable_no_resched_notrace();
6666
6667		/*
6668		 * Check again in case we missed a preemption opportunity
6669		 * between schedule and now.
6670		 */
6671	} while (need_resched());
6672}
6673
6674#ifdef CONFIG_PREEMPTION
6675/*
6676 * This is the entry point to schedule() from in-kernel preemption
6677 * off of preempt_enable.
6678 */
6679asmlinkage __visible void __sched notrace preempt_schedule(void)
6680{
6681	/*
6682	 * If there is a non-zero preempt_count or interrupts are disabled,
6683	 * we do not want to preempt the current task. Just return..
6684	 */
6685	if (likely(!preemptible()))
6686		return;
6687	preempt_schedule_common();
6688}
6689NOKPROBE_SYMBOL(preempt_schedule);
6690EXPORT_SYMBOL(preempt_schedule);
6691
6692#ifdef CONFIG_PREEMPT_DYNAMIC
6693#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6694#ifndef preempt_schedule_dynamic_enabled
6695#define preempt_schedule_dynamic_enabled	preempt_schedule
6696#define preempt_schedule_dynamic_disabled	NULL
6697#endif
6698DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6699EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6700#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6701static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6702void __sched notrace dynamic_preempt_schedule(void)
6703{
6704	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6705		return;
6706	preempt_schedule();
6707}
6708NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6709EXPORT_SYMBOL(dynamic_preempt_schedule);
6710#endif
6711#endif
6712
6713/**
6714 * preempt_schedule_notrace - preempt_schedule called by tracing
6715 *
6716 * The tracing infrastructure uses preempt_enable_notrace to prevent
6717 * recursion and tracing preempt enabling caused by the tracing
6718 * infrastructure itself. But as tracing can happen in areas coming
6719 * from userspace or just about to enter userspace, a preempt enable
6720 * can occur before user_exit() is called. This will cause the scheduler
6721 * to be called when the system is still in usermode.
6722 *
6723 * To prevent this, the preempt_enable_notrace will use this function
6724 * instead of preempt_schedule() to exit user context if needed before
6725 * calling the scheduler.
6726 */
6727asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6728{
6729	enum ctx_state prev_ctx;
6730
6731	if (likely(!preemptible()))
6732		return;
6733
6734	do {
6735		/*
6736		 * Because the function tracer can trace preempt_count_sub()
6737		 * and it also uses preempt_enable/disable_notrace(), if
6738		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6739		 * by the function tracer will call this function again and
6740		 * cause infinite recursion.
6741		 *
6742		 * Preemption must be disabled here before the function
6743		 * tracer can trace. Break up preempt_disable() into two
6744		 * calls. One to disable preemption without fear of being
6745		 * traced. The other to still record the preemption latency,
6746		 * which can also be traced by the function tracer.
6747		 */
6748		preempt_disable_notrace();
6749		preempt_latency_start(1);
6750		/*
6751		 * Needs preempt disabled in case user_exit() is traced
6752		 * and the tracer calls preempt_enable_notrace() causing
6753		 * an infinite recursion.
6754		 */
6755		prev_ctx = exception_enter();
6756		__schedule(SM_PREEMPT);
6757		exception_exit(prev_ctx);
6758
6759		preempt_latency_stop(1);
6760		preempt_enable_no_resched_notrace();
6761	} while (need_resched());
6762}
6763EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6764
6765#ifdef CONFIG_PREEMPT_DYNAMIC
6766#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6767#ifndef preempt_schedule_notrace_dynamic_enabled
6768#define preempt_schedule_notrace_dynamic_enabled	preempt_schedule_notrace
6769#define preempt_schedule_notrace_dynamic_disabled	NULL
6770#endif
6771DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6772EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6773#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6774static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6775void __sched notrace dynamic_preempt_schedule_notrace(void)
6776{
6777	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6778		return;
6779	preempt_schedule_notrace();
6780}
6781NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6782EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6783#endif
6784#endif
6785
6786#endif /* CONFIG_PREEMPTION */
6787
6788/*
6789 * This is the entry point to schedule() from kernel preemption
6790 * off of irq context.
6791 * Note, that this is called and return with irqs disabled. This will
6792 * protect us against recursive calling from irq.
6793 */
6794asmlinkage __visible void __sched preempt_schedule_irq(void)
6795{
6796	enum ctx_state prev_state;
6797
6798	/* Catch callers which need to be fixed */
6799	BUG_ON(preempt_count() || !irqs_disabled());
6800
6801	prev_state = exception_enter();
6802
6803	do {
6804		preempt_disable();
6805		local_irq_enable();
6806		__schedule(SM_PREEMPT);
6807		local_irq_disable();
6808		sched_preempt_enable_no_resched();
6809	} while (need_resched());
6810
6811	exception_exit(prev_state);
6812}
6813
6814int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6815			  void *key)
6816{
6817	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6818	return try_to_wake_up(curr->private, mode, wake_flags);
6819}
6820EXPORT_SYMBOL(default_wake_function);
6821
6822static void __setscheduler_prio(struct task_struct *p, int prio)
6823{
6824	if (dl_prio(prio))
6825		p->sched_class = &dl_sched_class;
6826	else if (rt_prio(prio))
6827		p->sched_class = &rt_sched_class;
6828	else
6829		p->sched_class = &fair_sched_class;
6830
6831	p->prio = prio;
6832}
6833
6834#ifdef CONFIG_RT_MUTEXES
6835
6836static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6837{
6838	if (pi_task)
6839		prio = min(prio, pi_task->prio);
6840
6841	return prio;
6842}
6843
6844static inline int rt_effective_prio(struct task_struct *p, int prio)
6845{
6846	struct task_struct *pi_task = rt_mutex_get_top_task(p);
6847
6848	return __rt_effective_prio(pi_task, prio);
6849}
6850
6851/*
6852 * rt_mutex_setprio - set the current priority of a task
6853 * @p: task to boost
6854 * @pi_task: donor task
6855 *
6856 * This function changes the 'effective' priority of a task. It does
6857 * not touch ->normal_prio like __setscheduler().
6858 *
6859 * Used by the rt_mutex code to implement priority inheritance
6860 * logic. Call site only calls if the priority of the task changed.
6861 */
6862void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6863{
6864	int prio, oldprio, queued, running, queue_flag =
6865		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6866	const struct sched_class *prev_class;
6867	struct rq_flags rf;
6868	struct rq *rq;
6869
6870	/* XXX used to be waiter->prio, not waiter->task->prio */
6871	prio = __rt_effective_prio(pi_task, p->normal_prio);
6872
6873	/*
6874	 * If nothing changed; bail early.
6875	 */
6876	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6877		return;
6878
6879	rq = __task_rq_lock(p, &rf);
6880	update_rq_clock(rq);
6881	/*
6882	 * Set under pi_lock && rq->lock, such that the value can be used under
6883	 * either lock.
6884	 *
6885	 * Note that there is loads of tricky to make this pointer cache work
6886	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6887	 * ensure a task is de-boosted (pi_task is set to NULL) before the
6888	 * task is allowed to run again (and can exit). This ensures the pointer
6889	 * points to a blocked task -- which guarantees the task is present.
6890	 */
6891	p->pi_top_task = pi_task;
6892
6893	/*
6894	 * For FIFO/RR we only need to set prio, if that matches we're done.
6895	 */
6896	if (prio == p->prio && !dl_prio(prio))
6897		goto out_unlock;
6898
6899	/*
6900	 * Idle task boosting is a nono in general. There is one
6901	 * exception, when PREEMPT_RT and NOHZ is active:
6902	 *
6903	 * The idle task calls get_next_timer_interrupt() and holds
6904	 * the timer wheel base->lock on the CPU and another CPU wants
6905	 * to access the timer (probably to cancel it). We can safely
6906	 * ignore the boosting request, as the idle CPU runs this code
6907	 * with interrupts disabled and will complete the lock
6908	 * protected section without being interrupted. So there is no
6909	 * real need to boost.
6910	 */
6911	if (unlikely(p == rq->idle)) {
6912		WARN_ON(p != rq->curr);
6913		WARN_ON(p->pi_blocked_on);
6914		goto out_unlock;
6915	}
6916
6917	trace_sched_pi_setprio(p, pi_task);
6918	oldprio = p->prio;
6919
6920	if (oldprio == prio)
6921		queue_flag &= ~DEQUEUE_MOVE;
6922
6923	prev_class = p->sched_class;
6924	queued = task_on_rq_queued(p);
6925	running = task_current(rq, p);
6926	if (queued)
6927		dequeue_task(rq, p, queue_flag);
6928	if (running)
6929		put_prev_task(rq, p);
6930
6931	/*
6932	 * Boosting condition are:
6933	 * 1. -rt task is running and holds mutex A
6934	 *      --> -dl task blocks on mutex A
6935	 *
6936	 * 2. -dl task is running and holds mutex A
6937	 *      --> -dl task blocks on mutex A and could preempt the
6938	 *          running task
6939	 */
6940	if (dl_prio(prio)) {
6941		if (!dl_prio(p->normal_prio) ||
6942		    (pi_task && dl_prio(pi_task->prio) &&
6943		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
6944			p->dl.pi_se = pi_task->dl.pi_se;
6945			queue_flag |= ENQUEUE_REPLENISH;
6946		} else {
6947			p->dl.pi_se = &p->dl;
6948		}
6949	} else if (rt_prio(prio)) {
6950		if (dl_prio(oldprio))
6951			p->dl.pi_se = &p->dl;
6952		if (oldprio < prio)
6953			queue_flag |= ENQUEUE_HEAD;
6954	} else {
6955		if (dl_prio(oldprio))
6956			p->dl.pi_se = &p->dl;
6957		if (rt_prio(oldprio))
6958			p->rt.timeout = 0;
6959	}
6960
6961	__setscheduler_prio(p, prio);
6962
6963	if (queued)
6964		enqueue_task(rq, p, queue_flag);
6965	if (running)
6966		set_next_task(rq, p);
6967
6968	check_class_changed(rq, p, prev_class, oldprio);
6969out_unlock:
6970	/* Avoid rq from going away on us: */
6971	preempt_disable();
6972
6973	rq_unpin_lock(rq, &rf);
6974	__balance_callbacks(rq);
6975	raw_spin_rq_unlock(rq);
6976
6977	preempt_enable();
6978}
6979#else
6980static inline int rt_effective_prio(struct task_struct *p, int prio)
6981{
6982	return prio;
6983}
6984#endif
6985
6986void set_user_nice(struct task_struct *p, long nice)
6987{
6988	bool queued, running;
6989	int old_prio;
6990	struct rq_flags rf;
6991	struct rq *rq;
6992
6993	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6994		return;
6995	/*
6996	 * We have to be careful, if called from sys_setpriority(),
6997	 * the task might be in the middle of scheduling on another CPU.
6998	 */
6999	rq = task_rq_lock(p, &rf);
7000	update_rq_clock(rq);
7001
7002	/*
7003	 * The RT priorities are set via sched_setscheduler(), but we still
7004	 * allow the 'normal' nice value to be set - but as expected
7005	 * it won't have any effect on scheduling until the task is
7006	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7007	 */
7008	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7009		p->static_prio = NICE_TO_PRIO(nice);
7010		goto out_unlock;
7011	}
7012	queued = task_on_rq_queued(p);
7013	running = task_current(rq, p);
7014	if (queued)
7015		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7016	if (running)
7017		put_prev_task(rq, p);
7018
7019	p->static_prio = NICE_TO_PRIO(nice);
7020	set_load_weight(p, true);
7021	old_prio = p->prio;
7022	p->prio = effective_prio(p);
7023
7024	if (queued)
7025		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7026	if (running)
7027		set_next_task(rq, p);
7028
7029	/*
7030	 * If the task increased its priority or is running and
7031	 * lowered its priority, then reschedule its CPU:
7032	 */
7033	p->sched_class->prio_changed(rq, p, old_prio);
7034
7035out_unlock:
7036	task_rq_unlock(rq, p, &rf);
7037}
7038EXPORT_SYMBOL(set_user_nice);
7039
7040/*
7041 * is_nice_reduction - check if nice value is an actual reduction
7042 *
7043 * Similar to can_nice() but does not perform a capability check.
7044 *
7045 * @p: task
7046 * @nice: nice value
7047 */
7048static bool is_nice_reduction(const struct task_struct *p, const int nice)
7049{
7050	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
7051	int nice_rlim = nice_to_rlimit(nice);
7052
7053	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7054}
7055
7056/*
7057 * can_nice - check if a task can reduce its nice value
7058 * @p: task
7059 * @nice: nice value
7060 */
7061int can_nice(const struct task_struct *p, const int nice)
7062{
7063	return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7064}
7065
7066#ifdef __ARCH_WANT_SYS_NICE
7067
7068/*
7069 * sys_nice - change the priority of the current process.
7070 * @increment: priority increment
7071 *
7072 * sys_setpriority is a more generic, but much slower function that
7073 * does similar things.
7074 */
7075SYSCALL_DEFINE1(nice, int, increment)
7076{
7077	long nice, retval;
7078
7079	/*
7080	 * Setpriority might change our priority at the same moment.
7081	 * We don't have to worry. Conceptually one call occurs first
7082	 * and we have a single winner.
7083	 */
7084	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7085	nice = task_nice(current) + increment;
7086
7087	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7088	if (increment < 0 && !can_nice(current, nice))
7089		return -EPERM;
7090
7091	retval = security_task_setnice(current, nice);
7092	if (retval)
7093		return retval;
7094
7095	set_user_nice(current, nice);
7096	return 0;
7097}
7098
7099#endif
7100
7101/**
7102 * task_prio - return the priority value of a given task.
7103 * @p: the task in question.
7104 *
7105 * Return: The priority value as seen by users in /proc.
7106 *
7107 * sched policy         return value   kernel prio    user prio/nice
7108 *
7109 * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
7110 * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
7111 * deadline                     -101             -1           0
7112 */
7113int task_prio(const struct task_struct *p)
7114{
7115	return p->prio - MAX_RT_PRIO;
7116}
7117
7118/**
7119 * idle_cpu - is a given CPU idle currently?
7120 * @cpu: the processor in question.
7121 *
7122 * Return: 1 if the CPU is currently idle. 0 otherwise.
7123 */
7124int idle_cpu(int cpu)
7125{
7126	struct rq *rq = cpu_rq(cpu);
7127
7128	if (rq->curr != rq->idle)
7129		return 0;
7130
7131	if (rq->nr_running)
7132		return 0;
7133
7134#ifdef CONFIG_SMP
7135	if (rq->ttwu_pending)
7136		return 0;
7137#endif
7138
7139	return 1;
7140}
7141
7142/**
7143 * available_idle_cpu - is a given CPU idle for enqueuing work.
7144 * @cpu: the CPU in question.
7145 *
7146 * Return: 1 if the CPU is currently idle. 0 otherwise.
7147 */
7148int available_idle_cpu(int cpu)
7149{
7150	if (!idle_cpu(cpu))
7151		return 0;
7152
7153	if (vcpu_is_preempted(cpu))
7154		return 0;
7155
7156	return 1;
7157}
7158
7159/**
7160 * idle_task - return the idle task for a given CPU.
7161 * @cpu: the processor in question.
7162 *
7163 * Return: The idle task for the CPU @cpu.
7164 */
7165struct task_struct *idle_task(int cpu)
7166{
7167	return cpu_rq(cpu)->idle;
7168}
7169
7170#ifdef CONFIG_SMP
7171/*
7172 * This function computes an effective utilization for the given CPU, to be
7173 * used for frequency selection given the linear relation: f = u * f_max.
7174 *
7175 * The scheduler tracks the following metrics:
7176 *
7177 *   cpu_util_{cfs,rt,dl,irq}()
7178 *   cpu_bw_dl()
7179 *
7180 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7181 * synchronized windows and are thus directly comparable.
7182 *
7183 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7184 * which excludes things like IRQ and steal-time. These latter are then accrued
7185 * in the irq utilization.
7186 *
7187 * The DL bandwidth number otoh is not a measured metric but a value computed
7188 * based on the task model parameters and gives the minimal utilization
7189 * required to meet deadlines.
7190 */
7191unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7192				 enum cpu_util_type type,
7193				 struct task_struct *p)
7194{
7195	unsigned long dl_util, util, irq, max;
7196	struct rq *rq = cpu_rq(cpu);
7197
7198	max = arch_scale_cpu_capacity(cpu);
7199
7200	if (!uclamp_is_used() &&
7201	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7202		return max;
7203	}
7204
7205	/*
7206	 * Early check to see if IRQ/steal time saturates the CPU, can be
7207	 * because of inaccuracies in how we track these -- see
7208	 * update_irq_load_avg().
7209	 */
7210	irq = cpu_util_irq(rq);
7211	if (unlikely(irq >= max))
7212		return max;
7213
7214	/*
7215	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7216	 * CFS tasks and we use the same metric to track the effective
7217	 * utilization (PELT windows are synchronized) we can directly add them
7218	 * to obtain the CPU's actual utilization.
7219	 *
7220	 * CFS and RT utilization can be boosted or capped, depending on
7221	 * utilization clamp constraints requested by currently RUNNABLE
7222	 * tasks.
7223	 * When there are no CFS RUNNABLE tasks, clamps are released and
7224	 * frequency will be gracefully reduced with the utilization decay.
7225	 */
7226	util = util_cfs + cpu_util_rt(rq);
7227	if (type == FREQUENCY_UTIL)
7228		util = uclamp_rq_util_with(rq, util, p);
7229
7230	dl_util = cpu_util_dl(rq);
7231
7232	/*
7233	 * For frequency selection we do not make cpu_util_dl() a permanent part
7234	 * of this sum because we want to use cpu_bw_dl() later on, but we need
7235	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7236	 * that we select f_max when there is no idle time.
7237	 *
7238	 * NOTE: numerical errors or stop class might cause us to not quite hit
7239	 * saturation when we should -- something for later.
7240	 */
7241	if (util + dl_util >= max)
7242		return max;
7243
7244	/*
7245	 * OTOH, for energy computation we need the estimated running time, so
7246	 * include util_dl and ignore dl_bw.
7247	 */
7248	if (type == ENERGY_UTIL)
7249		util += dl_util;
7250
7251	/*
7252	 * There is still idle time; further improve the number by using the
7253	 * irq metric. Because IRQ/steal time is hidden from the task clock we
7254	 * need to scale the task numbers:
7255	 *
7256	 *              max - irq
7257	 *   U' = irq + --------- * U
7258	 *                 max
7259	 */
7260	util = scale_irq_capacity(util, irq, max);
7261	util += irq;
7262
7263	/*
7264	 * Bandwidth required by DEADLINE must always be granted while, for
7265	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7266	 * to gracefully reduce the frequency when no tasks show up for longer
7267	 * periods of time.
7268	 *
7269	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7270	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7271	 * an interface. So, we only do the latter for now.
7272	 */
7273	if (type == FREQUENCY_UTIL)
7274		util += cpu_bw_dl(rq);
7275
7276	return min(max, util);
7277}
7278
7279unsigned long sched_cpu_util(int cpu)
7280{
7281	return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7282}
7283#endif /* CONFIG_SMP */
7284
7285/**
7286 * find_process_by_pid - find a process with a matching PID value.
7287 * @pid: the pid in question.
7288 *
7289 * The task of @pid, if found. %NULL otherwise.
7290 */
7291static struct task_struct *find_process_by_pid(pid_t pid)
7292{
7293	return pid ? find_task_by_vpid(pid) : current;
7294}
7295
7296/*
7297 * sched_setparam() passes in -1 for its policy, to let the functions
7298 * it calls know not to change it.
7299 */
7300#define SETPARAM_POLICY	-1
7301
7302static void __setscheduler_params(struct task_struct *p,
7303		const struct sched_attr *attr)
7304{
7305	int policy = attr->sched_policy;
7306
7307	if (policy == SETPARAM_POLICY)
7308		policy = p->policy;
7309
7310	p->policy = policy;
7311
7312	if (dl_policy(policy))
7313		__setparam_dl(p, attr);
7314	else if (fair_policy(policy))
7315		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7316
7317	/*
7318	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7319	 * !rt_policy. Always setting this ensures that things like
7320	 * getparam()/getattr() don't report silly values for !rt tasks.
7321	 */
7322	p->rt_priority = attr->sched_priority;
7323	p->normal_prio = normal_prio(p);
7324	set_load_weight(p, true);
7325}
7326
7327/*
7328 * Check the target process has a UID that matches the current process's:
7329 */
7330static bool check_same_owner(struct task_struct *p)
7331{
7332	const