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/irq_regs.h>
77	#include <asm/switch_to.h>
78	#include <asm/tlb.h>
79
80	#define CREATE_TRACE_POINTS
81	#include <linux/sched/rseq_api.h>
82	#include <trace/events/sched.h>
83	#include <trace/events/ipi.h>
84	#undef CREATE_TRACE_POINTS
85
86	#include "sched.h"
87	#include "stats.h"
88
89	#include "autogroup.h"
90	#include "pelt.h"
91	#include "smp.h"
92	#include "stats.h"
93
94	#include "../workqueue_internal.h"
95	#include "../../io_uring/io-wq.h"
96	#include "../smpboot.h"
97
98	EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
99	EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
100
101	/*
102	* Export tracepoints that act as a bare tracehook (ie: have no trace event
103	* associated with them) to allow external modules to probe them.
104	*/
105	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
106	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
107	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
108	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
109	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
110	EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
111	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
112	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
113	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
114	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
115	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
116	EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
117
118	DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
119
120	#ifdef CONFIG_SCHED_DEBUG
121	/*
122	* Debugging: various feature bits
123	*
124	* If SCHED_DEBUG is disabled, each compilation unit has its own copy of
125	* sysctl_sched_features, defined in sched.h, to allow constants propagation
126	* at compile time and compiler optimization based on features default.
127	*/
128	#define SCHED_FEAT(name, enabled) \
129	(1UL << __SCHED_FEAT_##name) * enabled |
130	const_debug unsigned int sysctl_sched_features =
131	#include "features.h"
132	0;
133	#undef SCHED_FEAT
134
135	/*
136	* Print a warning if need_resched is set for the given duration (if
137	* LATENCY_WARN is enabled).
138	*
139	* If sysctl_resched_latency_warn_once is set, only one warning will be shown
140	* per boot.
141	*/
142	__read_mostly int sysctl_resched_latency_warn_ms = 100;
143	__read_mostly int sysctl_resched_latency_warn_once = 1;
144	#endif /* CONFIG_SCHED_DEBUG */
145
146	/*
147	* Number of tasks to iterate in a single balance run.
148	* Limited because this is done with IRQs disabled.
149	*/
150	const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
151
152	__read_mostly int scheduler_running;
153
154	#ifdef CONFIG_SCHED_CORE
155
156	DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
157
158	/* kernel prio, less is more */
159	static inline int __task_prio(const struct task_struct *p)
160	{
161	if (p->sched_class == &stop_sched_class) /* trumps deadline */
162	return -2;
163
164	if (rt_prio(prio: p->prio)) /* includes deadline */
165	return p->prio; /* [-1, 99] */
166
167	if (p->sched_class == &idle_sched_class)
168	return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
169
170	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
171	}
172
173	/*
174	* l(a,b)
175	* le(a,b) := !l(b,a)
176	* g(a,b) := l(b,a)
177	* ge(a,b) := !l(a,b)
178	*/
179
180	/* real prio, less is less */
181	static inline bool prio_less(const struct task_struct *a,
182	const struct task_struct *b, bool in_fi)
183	{
184
185	int pa = __task_prio(p: a), pb = __task_prio(p: b);
186
187	if (-pa < -pb)
188	return true;
189
190	if (-pb < -pa)
191	return false;
192
193	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
194	return !dl_time_before(a: a->dl.deadline, b: b->dl.deadline);
195
196	if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
197	return cfs_prio_less(a, b, fi: in_fi);
198
199	return false;
200	}
201
202	static inline bool __sched_core_less(const struct task_struct *a,
203	const struct task_struct *b)
204	{
205	if (a->core_cookie < b->core_cookie)
206	return true;
207
208	if (a->core_cookie > b->core_cookie)
209	return false;
210
211	/* flip prio, so high prio is leftmost */
212	if (prio_less(a: b, b: a, in_fi: !!task_rq(a)->core->core_forceidle_count))
213	return true;
214
215	return false;
216	}
217
218	#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
219
220	static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
221	{
222	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
223	}
224
225	static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
226	{
227	const struct task_struct *p = __node_2_sc(node);
228	unsigned long cookie = (unsigned long)key;
229
230	if (cookie < p->core_cookie)
231	return -1;
232
233	if (cookie > p->core_cookie)
234	return 1;
235
236	return 0;
237	}
238
239	void sched_core_enqueue(struct rq *rq, struct task_struct *p)
240	{
241	rq->core->core_task_seq++;
242
243	if (!p->core_cookie)
244	return;
245
246	rb_add(node: &p->core_node, tree: &rq->core_tree, less: rb_sched_core_less);
247	}
248
249	void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
250	{
251	rq->core->core_task_seq++;
252
253	if (sched_core_enqueued(p)) {
254	rb_erase(&p->core_node, &rq->core_tree);
255	RB_CLEAR_NODE(&p->core_node);
256	}
257
258	/*
259	* Migrating the last task off the cpu, with the cpu in forced idle
260	* state. Reschedule to create an accounting edge for forced idle,
261	* and re-examine whether the core is still in forced idle state.
262	*/
263	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
264	rq->core->core_forceidle_count && rq->curr == rq->idle)
265	resched_curr(rq);
266	}
267
268	static int sched_task_is_throttled(struct task_struct *p, int cpu)
269	{
270	if (p->sched_class->task_is_throttled)
271	return p->sched_class->task_is_throttled(p, cpu);
272
273	return 0;
274	}
275
276	static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
277	{
278	struct rb_node *node = &p->core_node;
279	int cpu = task_cpu(p);
280
281	do {
282	node = rb_next(node);
283	if (!node)
284	return NULL;
285
286	p = __node_2_sc(node);
287	if (p->core_cookie != cookie)
288	return NULL;
289
290	} while (sched_task_is_throttled(p, cpu));
291
292	return p;
293	}
294
295	/*
296	* Find left-most (aka, highest priority) and unthrottled task matching @cookie.
297	* If no suitable task is found, NULL will be returned.
298	*/
299	static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
300	{
301	struct task_struct *p;
302	struct rb_node *node;
303
304	node = rb_find_first(key: (void *)cookie, tree: &rq->core_tree, cmp: rb_sched_core_cmp);
305	if (!node)
306	return NULL;
307
308	p = __node_2_sc(node);
309	if (!sched_task_is_throttled(p, cpu: rq->cpu))
310	return p;
311
312	return sched_core_next(p, cookie);
313	}
314
315	/*
316	* Magic required such that:
317	*
318	* raw_spin_rq_lock(rq);
319	* ...
320	* raw_spin_rq_unlock(rq);
321	*
322	* ends up locking and unlocking the _same_ lock, and all CPUs
323	* always agree on what rq has what lock.
324	*
325	* XXX entirely possible to selectively enable cores, don't bother for now.
326	*/
327
328	static DEFINE_MUTEX(sched_core_mutex);
329	static atomic_t sched_core_count;
330	static struct cpumask sched_core_mask;
331
332	static void sched_core_lock(int cpu, unsigned long *flags)
333	{
334	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
335	int t, i = 0;
336
337	local_irq_save(*flags);
338	for_each_cpu(t, smt_mask)
339	raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
340	}
341
342	static void sched_core_unlock(int cpu, unsigned long *flags)
343	{
344	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
345	int t;
346
347	for_each_cpu(t, smt_mask)
348	raw_spin_unlock(&cpu_rq(t)->__lock);
349	local_irq_restore(*flags);
350	}
351
352	static void __sched_core_flip(bool enabled)
353	{
354	unsigned long flags;
355	int cpu, t;
356
357	cpus_read_lock();
358
359	/*
360	* Toggle the online cores, one by one.
361	*/
362	cpumask_copy(dstp: &sched_core_mask, cpu_online_mask);
363	for_each_cpu(cpu, &sched_core_mask) {
364	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
365
366	sched_core_lock(cpu, flags: &flags);
367
368	for_each_cpu(t, smt_mask)
369	cpu_rq(t)->core_enabled = enabled;
370
371	cpu_rq(cpu)->core->core_forceidle_start = 0;
372
373	sched_core_unlock(cpu, flags: &flags);
374
375	cpumask_andnot(dstp: &sched_core_mask, src1p: &sched_core_mask, src2p: smt_mask);
376	}
377
378	/*
379	* Toggle the offline CPUs.
380	*/
381	for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
382	cpu_rq(cpu)->core_enabled = enabled;
383
384	cpus_read_unlock();
385	}
386
387	static void sched_core_assert_empty(void)
388	{
389	int cpu;
390
391	for_each_possible_cpu(cpu)
392	WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
393	}
394
395	static void __sched_core_enable(void)
396	{
397	static_branch_enable(&__sched_core_enabled);
398	/*
399	* Ensure all previous instances of raw_spin_rq_*lock() have finished
400	* and future ones will observe !sched_core_disabled().
401	*/
402	synchronize_rcu();
403	__sched_core_flip(enabled: true);
404	sched_core_assert_empty();
405	}
406
407	static void __sched_core_disable(void)
408	{
409	sched_core_assert_empty();
410	__sched_core_flip(enabled: false);
411	static_branch_disable(&__sched_core_enabled);
412	}
413
414	void sched_core_get(void)
415	{
416	if (atomic_inc_not_zero(v: &sched_core_count))
417	return;
418
419	mutex_lock(&sched_core_mutex);
420	if (!atomic_read(v: &sched_core_count))
421	__sched_core_enable();
422
423	smp_mb__before_atomic();
424	atomic_inc(v: &sched_core_count);
425	mutex_unlock(lock: &sched_core_mutex);
426	}
427
428	static void __sched_core_put(struct work_struct *work)
429	{
430	if (atomic_dec_and_mutex_lock(cnt: &sched_core_count, lock: &sched_core_mutex)) {
431	__sched_core_disable();
432	mutex_unlock(lock: &sched_core_mutex);
433	}
434	}
435
436	void sched_core_put(void)
437	{
438	static DECLARE_WORK(_work, __sched_core_put);
439
440	/*
441	* "There can be only one"
442	*
443	* Either this is the last one, or we don't actually need to do any
444	* 'work'. If it is the last *again*, we rely on
445	* WORK_STRUCT_PENDING_BIT.
446	*/
447	if (!atomic_add_unless(v: &sched_core_count, a: -1, u: 1))
448	schedule_work(work: &_work);
449	}
450
451	#else /* !CONFIG_SCHED_CORE */
452
453	static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
454	static inline void
455	sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
456
457	#endif /* CONFIG_SCHED_CORE */
458
459	/*
460	* Serialization rules:
461	*
462	* Lock order:
463	*
464	* p->pi_lock
465	* rq->lock
466	* hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
467	*
468	* rq1->lock
469	* rq2->lock where: rq1 < rq2
470	*
471	* Regular state:
472	*
473	* Normal scheduling state is serialized by rq->lock. __schedule() takes the
474	* local CPU's rq->lock, it optionally removes the task from the runqueue and
475	* always looks at the local rq data structures to find the most eligible task
476	* to run next.
477	*
478	* Task enqueue is also under rq->lock, possibly taken from another CPU.
479	* Wakeups from another LLC domain might use an IPI to transfer the enqueue to
480	* the local CPU to avoid bouncing the runqueue state around [ see
481	* ttwu_queue_wakelist() ]
482	*
483	* Task wakeup, specifically wakeups that involve migration, are horribly
484	* complicated to avoid having to take two rq->locks.
485	*
486	* Special state:
487	*
488	* System-calls and anything external will use task_rq_lock() which acquires
489	* both p->pi_lock and rq->lock. As a consequence the state they change is
490	* stable while holding either lock:
491	*
492	* - sched_setaffinity()/
493	* set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
494	* - set_user_nice(): p->se.load, p->*prio
495	* - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
496	* p->se.load, p->rt_priority,
497	* p->dl.dl_{runtime, deadline, period, flags, bw, density}
498	* - sched_setnuma(): p->numa_preferred_nid
499	* - sched_move_task(): p->sched_task_group
500	* - uclamp_update_active() p->uclamp*
501	*
502	* p->state <- TASK_*:
503	*
504	* is changed locklessly using set_current_state(), __set_current_state() or
505	* set_special_state(), see their respective comments, or by
506	* try_to_wake_up(). This latter uses p->pi_lock to serialize against
507	* concurrent self.
508	*
509	* p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
510	*
511	* is set by activate_task() and cleared by deactivate_task(), under
512	* rq->lock. Non-zero indicates the task is runnable, the special
513	* ON_RQ_MIGRATING state is used for migration without holding both
514	* rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
515	*
516	* p->on_cpu <- { 0, 1 }:
517	*
518	* is set by prepare_task() and cleared by finish_task() such that it will be
519	* set before p is scheduled-in and cleared after p is scheduled-out, both
520	* under rq->lock. Non-zero indicates the task is running on its CPU.
521	*
522	* [ The astute reader will observe that it is possible for two tasks on one
523	* CPU to have ->on_cpu = 1 at the same time. ]
524	*
525	* task_cpu(p): is changed by set_task_cpu(), the rules are:
526	*
527	* - Don't call set_task_cpu() on a blocked task:
528	*
529	* We don't care what CPU we're not running on, this simplifies hotplug,
530	* the CPU assignment of blocked tasks isn't required to be valid.
531	*
532	* - for try_to_wake_up(), called under p->pi_lock:
533	*
534	* This allows try_to_wake_up() to only take one rq->lock, see its comment.
535	*
536	* - for migration called under rq->lock:
537	* [ see task_on_rq_migrating() in task_rq_lock() ]
538	*
539	* o move_queued_task()
540	* o detach_task()
541	*
542	* - for migration called under double_rq_lock():
543	*
544	* o __migrate_swap_task()
545	* o push_rt_task() / pull_rt_task()
546	* o push_dl_task() / pull_dl_task()
547	* o dl_task_offline_migration()
548	*
549	*/
550
551	void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
552	{
553	raw_spinlock_t *lock;
554
555	/* Matches synchronize_rcu() in __sched_core_enable() */
556	preempt_disable();
557	if (sched_core_disabled()) {
558	raw_spin_lock_nested(&rq->__lock, subclass);
559	/* preempt_count *MUST* be > 1 */
560	preempt_enable_no_resched();
561	return;
562	}
563
564	for (;;) {
565	lock = __rq_lockp(rq);
566	raw_spin_lock_nested(lock, subclass);
567	if (likely(lock == __rq_lockp(rq))) {
568	/* preempt_count *MUST* be > 1 */
569	preempt_enable_no_resched();
570	return;
571	}
572	raw_spin_unlock(lock);
573	}
574	}
575
576	bool raw_spin_rq_trylock(struct rq *rq)
577	{
578	raw_spinlock_t *lock;
579	bool ret;
580
581	/* Matches synchronize_rcu() in __sched_core_enable() */
582	preempt_disable();
583	if (sched_core_disabled()) {
584	ret = raw_spin_trylock(&rq->__lock);
585	preempt_enable();
586	return ret;
587	}
588
589	for (;;) {
590	lock = __rq_lockp(rq);
591	ret = raw_spin_trylock(lock);
592	if (!ret || (likely(lock == __rq_lockp(rq)))) {
593	preempt_enable();
594	return ret;
595	}
596	raw_spin_unlock(lock);
597	}
598	}
599
600	void raw_spin_rq_unlock(struct rq *rq)
601	{
602	raw_spin_unlock(rq_lockp(rq));
603	}
604
605	#ifdef CONFIG_SMP
606	/*
607	* double_rq_lock - safely lock two runqueues
608	*/
609	void double_rq_lock(struct rq *rq1, struct rq *rq2)
610	{
611	lockdep_assert_irqs_disabled();
612
613	if (rq_order_less(rq1: rq2, rq2: rq1))
614	swap(rq1, rq2);
615
616	raw_spin_rq_lock(rq: rq1);
617	if (__rq_lockp(rq: rq1) != __rq_lockp(rq: rq2))
618	raw_spin_rq_lock_nested(rq: rq2, SINGLE_DEPTH_NESTING);
619
620	double_rq_clock_clear_update(rq1, rq2);
621	}
622	#endif
623
624	/*
625	* __task_rq_lock - lock the rq @p resides on.
626	*/
627	struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
628	__acquires(rq->lock)
629	{
630	struct rq *rq;
631
632	lockdep_assert_held(&p->pi_lock);
633
634	for (;;) {
635	rq = task_rq(p);
636	raw_spin_rq_lock(rq);
637	if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
638	rq_pin_lock(rq, rf);
639	return rq;
640	}
641	raw_spin_rq_unlock(rq);
642
643	while (unlikely(task_on_rq_migrating(p)))
644	cpu_relax();
645	}
646	}
647
648	/*
649	* task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
650	*/
651	struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
652	__acquires(p->pi_lock)
653	__acquires(rq->lock)
654	{
655	struct rq *rq;
656
657	for (;;) {
658	raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
659	rq = task_rq(p);
660	raw_spin_rq_lock(rq);
661	/*
662	* move_queued_task() task_rq_lock()
663	*
664	* ACQUIRE (rq->lock)
665	* [S] ->on_rq = MIGRATING [L] rq = task_rq()
666	* WMB (__set_task_cpu()) ACQUIRE (rq->lock);
667	* [S] ->cpu = new_cpu [L] task_rq()
668	* [L] ->on_rq
669	* RELEASE (rq->lock)
670	*
671	* If we observe the old CPU in task_rq_lock(), the acquire of
672	* the old rq->lock will fully serialize against the stores.
673	*
674	* If we observe the new CPU in task_rq_lock(), the address
675	* dependency headed by '[L] rq = task_rq()' and the acquire
676	* will pair with the WMB to ensure we then also see migrating.
677	*/
678	if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
679	rq_pin_lock(rq, rf);
680	return rq;
681	}
682	raw_spin_rq_unlock(rq);
683	raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
684
685	while (unlikely(task_on_rq_migrating(p)))
686	cpu_relax();
687	}
688	}
689
690	/*
691	* RQ-clock updating methods:
692	*/
693
694	static void update_rq_clock_task(struct rq *rq, s64 delta)
695	{
696	/*
697	* In theory, the compile should just see 0 here, and optimize out the call
698	* to sched_rt_avg_update. But I don't trust it...
699	*/
700	s64 __maybe_unused steal = 0, irq_delta = 0;
701
702	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
703	irq_delta = irq_time_read(cpu: cpu_of(rq)) - rq->prev_irq_time;
704
705	/*
706	* Since irq_time is only updated on {soft,}irq_exit, we might run into
707	* this case when a previous update_rq_clock() happened inside a
708	* {soft,}irq region.
709	*
710	* When this happens, we stop ->clock_task and only update the
711	* prev_irq_time stamp to account for the part that fit, so that a next
712	* update will consume the rest. This ensures ->clock_task is
713	* monotonic.
714	*
715	* It does however cause some slight miss-attribution of {soft,}irq
716	* time, a more accurate solution would be to update the irq_time using
717	* the current rq->clock timestamp, except that would require using
718	* atomic ops.
719	*/
720	if (irq_delta > delta)
721	irq_delta = delta;
722
723	rq->prev_irq_time += irq_delta;
724	delta -= irq_delta;
725	psi_account_irqtime(task: rq->curr, delta: irq_delta);
726	delayacct_irq(task: rq->curr, delta: irq_delta);
727	#endif
728	#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
729	if (static_key_false(key: (&paravirt_steal_rq_enabled))) {
730	steal = paravirt_steal_clock(cpu: cpu_of(rq));
731	steal -= rq->prev_steal_time_rq;
732
733	if (unlikely(steal > delta))
734	steal = delta;
735
736	rq->prev_steal_time_rq += steal;
737	delta -= steal;
738	}
739	#endif
740
741	rq->clock_task += delta;
742
743	#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
744	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
745	update_irq_load_avg(rq, running: irq_delta + steal);
746	#endif
747	update_rq_clock_pelt(rq, delta);
748	}
749
750	void update_rq_clock(struct rq *rq)
751	{
752	s64 delta;
753
754	lockdep_assert_rq_held(rq);
755
756	if (rq->clock_update_flags & RQCF_ACT_SKIP)
757	return;
758
759	#ifdef CONFIG_SCHED_DEBUG
760	if (sched_feat(WARN_DOUBLE_CLOCK))
761	SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
762	rq->clock_update_flags |= RQCF_UPDATED;
763	#endif
764
765	delta = sched_clock_cpu(cpu: cpu_of(rq)) - rq->clock;
766	if (delta < 0)
767	return;
768	rq->clock += delta;
769	update_rq_clock_task(rq, delta);
770	}
771
772	#ifdef CONFIG_SCHED_HRTICK
773	/*
774	* Use HR-timers to deliver accurate preemption points.
775	*/
776
777	static void hrtick_clear(struct rq *rq)
778	{
779	if (hrtimer_active(timer: &rq->hrtick_timer))
780	hrtimer_cancel(timer: &rq->hrtick_timer);
781	}
782
783	/*
784	* High-resolution timer tick.
785	* Runs from hardirq context with interrupts disabled.
786	*/
787	static enum hrtimer_restart hrtick(struct hrtimer *timer)
788	{
789	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
790	struct rq_flags rf;
791
792	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
793
794	rq_lock(rq, rf: &rf);
795	update_rq_clock(rq);
796	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
797	rq_unlock(rq, rf: &rf);
798
799	return HRTIMER_NORESTART;
800	}
801
802	#ifdef CONFIG_SMP
803
804	static void __hrtick_restart(struct rq *rq)
805	{
806	struct hrtimer *timer = &rq->hrtick_timer;
807	ktime_t time = rq->hrtick_time;
808
809	hrtimer_start(timer, tim: time, mode: HRTIMER_MODE_ABS_PINNED_HARD);
810	}
811
812	/*
813	* called from hardirq (IPI) context
814	*/
815	static void __hrtick_start(void *arg)
816	{
817	struct rq *rq = arg;
818	struct rq_flags rf;
819
820	rq_lock(rq, rf: &rf);
821	__hrtick_restart(rq);
822	rq_unlock(rq, rf: &rf);
823	}
824
825	/*
826	* Called to set the hrtick timer state.
827	*
828	* called with rq->lock held and irqs disabled
829	*/
830	void hrtick_start(struct rq *rq, u64 delay)
831	{
832	struct hrtimer *timer = &rq->hrtick_timer;
833	s64 delta;
834
835	/*
836	* Don't schedule slices shorter than 10000ns, that just
837	* doesn't make sense and can cause timer DoS.
838	*/
839	delta = max_t(s64, delay, 10000LL);
840	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
841
842	if (rq == this_rq())
843	__hrtick_restart(rq);
844	else
845	smp_call_function_single_async(cpu: cpu_of(rq), csd: &rq->hrtick_csd);
846	}
847
848	#else
849	/*
850	* Called to set the hrtick timer state.
851	*
852	* called with rq->lock held and irqs disabled
853	*/
854	void hrtick_start(struct rq *rq, u64 delay)
855	{
856	/*
857	* Don't schedule slices shorter than 10000ns, that just
858	* doesn't make sense. Rely on vruntime for fairness.
859	*/
860	delay = max_t(u64, delay, 10000LL);
861	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
862	HRTIMER_MODE_REL_PINNED_HARD);
863	}
864
865	#endif /* CONFIG_SMP */
866
867	static void hrtick_rq_init(struct rq *rq)
868	{
869	#ifdef CONFIG_SMP
870	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
871	#endif
872	hrtimer_init(timer: &rq->hrtick_timer, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL_HARD);
873	rq->hrtick_timer.function = hrtick;
874	}
875	#else /* CONFIG_SCHED_HRTICK */
876	static inline void hrtick_clear(struct rq *rq)
877	{
878	}
879
880	static inline void hrtick_rq_init(struct rq *rq)
881	{
882	}
883	#endif /* CONFIG_SCHED_HRTICK */
884
885	/*
886	* cmpxchg based fetch_or, macro so it works for different integer types
887	*/
888	#define fetch_or(ptr, mask) \
889	({ \
890	typeof(ptr) _ptr = (ptr); \
891	typeof(mask) _mask = (mask); \
892	typeof(*_ptr) _val = *_ptr; \
893	\
894	do { \
895	} while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
896	_val; \
897	})
898
899	#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
900	/*
901	* Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
902	* this avoids any races wrt polling state changes and thereby avoids
903	* spurious IPIs.
904	*/
905	static inline bool set_nr_and_not_polling(struct task_struct *p)
906	{
907	struct thread_info *ti = task_thread_info(p);
908	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
909	}
910
911	/*
912	* Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
913	*
914	* If this returns true, then the idle task promises to call
915	* sched_ttwu_pending() and reschedule soon.
916	*/
917	static bool set_nr_if_polling(struct task_struct *p)
918	{
919	struct thread_info *ti = task_thread_info(p);
920	typeof(ti->flags) val = READ_ONCE(ti->flags);
921
922	do {
923	if (!(val & _TIF_POLLING_NRFLAG))
924	return false;
925	if (val & _TIF_NEED_RESCHED)
926	return true;
927	} while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
928
929	return true;
930	}
931
932	#else
933	static inline bool set_nr_and_not_polling(struct task_struct *p)
934	{
935	set_tsk_need_resched(p);
936	return true;
937	}
938
939	#ifdef CONFIG_SMP
940	static inline bool set_nr_if_polling(struct task_struct *p)
941	{
942	return false;
943	}
944	#endif
945	#endif
946
947	static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
948	{
949	struct wake_q_node *node = &task->wake_q;
950
951	/*
952	* Atomically grab the task, if ->wake_q is !nil already it means
953	* it's already queued (either by us or someone else) and will get the
954	* wakeup due to that.
955	*
956	* In order to ensure that a pending wakeup will observe our pending
957	* state, even in the failed case, an explicit smp_mb() must be used.
958	*/
959	smp_mb__before_atomic();
960	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
961	return false;
962
963	/*
964	* The head is context local, there can be no concurrency.
965	*/
966	*head->lastp = node;
967	head->lastp = &node->next;
968	return true;
969	}
970
971	/**
972	* wake_q_add() - queue a wakeup for 'later' waking.
973	* @head: the wake_q_head to add @task to
974	* @task: the task to queue for 'later' wakeup
975	*
976	* Queue a task for later wakeup, most likely by the wake_up_q() call in the
977	* same context, _HOWEVER_ this is not guaranteed, the wakeup can come
978	* instantly.
979	*
980	* This function must be used as-if it were wake_up_process(); IOW the task
981	* must be ready to be woken at this location.
982	*/
983	void wake_q_add(struct wake_q_head *head, struct task_struct *task)
984	{
985	if (__wake_q_add(head, task))
986	get_task_struct(t: task);
987	}
988
989	/**
990	* wake_q_add_safe() - safely queue a wakeup for 'later' waking.
991	* @head: the wake_q_head to add @task to
992	* @task: the task to queue for 'later' wakeup
993	*
994	* Queue a task for later wakeup, most likely by the wake_up_q() call in the
995	* same context, _HOWEVER_ this is not guaranteed, the wakeup can come
996	* instantly.
997	*
998	* This function must be used as-if it were wake_up_process(); IOW the task
999	* must be ready to be woken at this location.
1000	*
1001	* This function is essentially a task-safe equivalent to wake_q_add(). Callers
1002	* that already hold reference to @task can call the 'safe' version and trust
1003	* wake_q to do the right thing depending whether or not the @task is already
1004	* queued for wakeup.
1005	*/
1006	void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1007	{
1008	if (!__wake_q_add(head, task))
1009	put_task_struct(t: task);
1010	}
1011
1012	void wake_up_q(struct wake_q_head *head)
1013	{
1014	struct wake_q_node *node = head->first;
1015
1016	while (node != WAKE_Q_TAIL) {
1017	struct task_struct *task;
1018
1019	task = container_of(node, struct task_struct, wake_q);
1020	/* Task can safely be re-inserted now: */
1021	node = node->next;
1022	task->wake_q.next = NULL;
1023
1024	/*
1025	* wake_up_process() executes a full barrier, which pairs with
1026	* the queueing in wake_q_add() so as not to miss wakeups.
1027	*/
1028	wake_up_process(tsk: task);
1029	put_task_struct(t: task);
1030	}
1031	}
1032
1033	/*
1034	* resched_curr - mark rq's current task 'to be rescheduled now'.
1035	*
1036	* On UP this means the setting of the need_resched flag, on SMP it
1037	* might also involve a cross-CPU call to trigger the scheduler on
1038	* the target CPU.
1039	*/
1040	void resched_curr(struct rq *rq)
1041	{
1042	struct task_struct *curr = rq->curr;
1043	int cpu;
1044
1045	lockdep_assert_rq_held(rq);
1046
1047	if (test_tsk_need_resched(tsk: curr))
1048	return;
1049
1050	cpu = cpu_of(rq);
1051
1052	if (cpu == smp_processor_id()) {
1053	set_tsk_need_resched(curr);
1054	set_preempt_need_resched();
1055	return;
1056	}
1057
1058	if (set_nr_and_not_polling(curr))
1059	smp_send_reschedule(cpu);
1060	else
1061	trace_sched_wake_idle_without_ipi(cpu);
1062	}
1063
1064	void resched_cpu(int cpu)
1065	{
1066	struct rq *rq = cpu_rq(cpu);
1067	unsigned long flags;
1068
1069	raw_spin_rq_lock_irqsave(rq, flags);
1070	if (cpu_online(cpu) || cpu == smp_processor_id())
1071	resched_curr(rq);
1072	raw_spin_rq_unlock_irqrestore(rq, flags);
1073	}
1074
1075	#ifdef CONFIG_SMP
1076	#ifdef CONFIG_NO_HZ_COMMON
1077	/*
1078	* In the semi idle case, use the nearest busy CPU for migrating timers
1079	* from an idle CPU. This is good for power-savings.
1080	*
1081	* We don't do similar optimization for completely idle system, as
1082	* selecting an idle CPU will add more delays to the timers than intended
1083	* (as that CPU's timer base may not be uptodate wrt jiffies etc).
1084	*/
1085	int get_nohz_timer_target(void)
1086	{
1087	int i, cpu = smp_processor_id(), default_cpu = -1;
1088	struct sched_domain *sd;
1089	const struct cpumask *hk_mask;
1090
1091	if (housekeeping_cpu(cpu, type: HK_TYPE_TIMER)) {
1092	if (!idle_cpu(cpu))
1093	return cpu;
1094	default_cpu = cpu;
1095	}
1096
1097	hk_mask = housekeeping_cpumask(type: HK_TYPE_TIMER);
1098
1099	guard(rcu)();
1100
1101	for_each_domain(cpu, sd) {
1102	for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1103	if (cpu == i)
1104	continue;
1105
1106	if (!idle_cpu(cpu: i))
1107	return i;
1108	}
1109	}
1110
1111	if (default_cpu == -1)
1112	default_cpu = housekeeping_any_cpu(type: HK_TYPE_TIMER);
1113
1114	return default_cpu;
1115	}
1116
1117	/*
1118	* When add_timer_on() enqueues a timer into the timer wheel of an
1119	* idle CPU then this timer might expire before the next timer event
1120	* which is scheduled to wake up that CPU. In case of a completely
1121	* idle system the next event might even be infinite time into the
1122	* future. wake_up_idle_cpu() ensures that the CPU is woken up and
1123	* leaves the inner idle loop so the newly added timer is taken into
1124	* account when the CPU goes back to idle and evaluates the timer
1125	* wheel for the next timer event.
1126	*/
1127	static void wake_up_idle_cpu(int cpu)
1128	{
1129	struct rq *rq = cpu_rq(cpu);
1130
1131	if (cpu == smp_processor_id())
1132	return;
1133
1134	if (set_nr_and_not_polling(rq->idle))
1135	smp_send_reschedule(cpu);
1136	else
1137	trace_sched_wake_idle_without_ipi(cpu);
1138	}
1139
1140	static bool wake_up_full_nohz_cpu(int cpu)
1141	{
1142	/*
1143	* We just need the target to call irq_exit() and re-evaluate
1144	* the next tick. The nohz full kick at least implies that.
1145	* If needed we can still optimize that later with an
1146	* empty IRQ.
1147	*/
1148	if (cpu_is_offline(cpu))
1149	return true; /* Don't try to wake offline CPUs. */
1150	if (tick_nohz_full_cpu(cpu)) {
1151	if (cpu != smp_processor_id() ||
1152	tick_nohz_tick_stopped())
1153	tick_nohz_full_kick_cpu(cpu);
1154	return true;
1155	}
1156
1157	return false;
1158	}
1159
1160	/*
1161	* Wake up the specified CPU. If the CPU is going offline, it is the
1162	* caller's responsibility to deal with the lost wakeup, for example,
1163	* by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1164	*/
1165	void wake_up_nohz_cpu(int cpu)
1166	{
1167	if (!wake_up_full_nohz_cpu(cpu))
1168	wake_up_idle_cpu(cpu);
1169	}
1170
1171	static void nohz_csd_func(void *info)
1172	{
1173	struct rq *rq = info;
1174	int cpu = cpu_of(rq);
1175	unsigned int flags;
1176
1177	/*
1178	* Release the rq::nohz_csd.
1179	*/
1180	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1181	WARN_ON(!(flags & NOHZ_KICK_MASK));
1182
1183	rq->idle_balance = idle_cpu(cpu);
1184	if (rq->idle_balance && !need_resched()) {
1185	rq->nohz_idle_balance = flags;
1186	raise_softirq_irqoff(nr: SCHED_SOFTIRQ);
1187	}
1188	}
1189
1190	#endif /* CONFIG_NO_HZ_COMMON */
1191
1192	#ifdef CONFIG_NO_HZ_FULL
1193	static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1194	{
1195	if (rq->nr_running != 1)
1196	return false;
1197
1198	if (p->sched_class != &fair_sched_class)
1199	return false;
1200
1201	if (!task_on_rq_queued(p))
1202	return false;
1203
1204	return true;
1205	}
1206
1207	bool sched_can_stop_tick(struct rq *rq)
1208	{
1209	int fifo_nr_running;
1210
1211	/* Deadline tasks, even if single, need the tick */
1212	if (rq->dl.dl_nr_running)
1213	return false;
1214
1215	/*
1216	* If there are more than one RR tasks, we need the tick to affect the
1217	* actual RR behaviour.
1218	*/
1219	if (rq->rt.rr_nr_running) {
1220	if (rq->rt.rr_nr_running == 1)
1221	return true;
1222	else
1223	return false;
1224	}
1225
1226	/*
1227	* If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1228	* forced preemption between FIFO tasks.
1229	*/
1230	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1231	if (fifo_nr_running)
1232	return true;
1233
1234	/*
1235	* If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1236	* if there's more than one we need the tick for involuntary
1237	* preemption.
1238	*/
1239	if (rq->nr_running > 1)
1240	return false;
1241
1242	/*
1243	* If there is one task and it has CFS runtime bandwidth constraints
1244	* and it's on the cpu now we don't want to stop the tick.
1245	* This check prevents clearing the bit if a newly enqueued task here is
1246	* dequeued by migrating while the constrained task continues to run.
1247	* E.g. going from 2->1 without going through pick_next_task().
1248	*/
1249	if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1250	if (cfs_task_bw_constrained(rq->curr))
1251	return false;
1252	}
1253
1254	return true;
1255	}
1256	#endif /* CONFIG_NO_HZ_FULL */
1257	#endif /* CONFIG_SMP */
1258
1259	#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1260	(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1261	/*
1262	* Iterate task_group tree rooted at *from, calling @down when first entering a
1263	* node and @up when leaving it for the final time.
1264	*
1265	* Caller must hold rcu_lock or sufficient equivalent.
1266	*/
1267	int walk_tg_tree_from(struct task_group *from,
1268	tg_visitor down, tg_visitor up, void *data)
1269	{
1270	struct task_group *parent, *child;
1271	int ret;
1272
1273	parent = from;
1274
1275	down:
1276	ret = (*down)(parent, data);
1277	if (ret)
1278	goto out;
1279	list_for_each_entry_rcu(child, &parent->children, siblings) {
1280	parent = child;
1281	goto down;
1282
1283	up:
1284	continue;
1285	}
1286	ret = (*up)(parent, data);
1287	if (ret || parent == from)
1288	goto out;
1289
1290	child = parent;
1291	parent = parent->parent;
1292	if (parent)
1293	goto up;
1294	out:
1295	return ret;
1296	}
1297
1298	int tg_nop(struct task_group *tg, void *data)
1299	{
1300	return 0;
1301	}
1302	#endif
1303
1304	static void set_load_weight(struct task_struct *p, bool update_load)
1305	{
1306	int prio = p->static_prio - MAX_RT_PRIO;
1307	struct load_weight *load = &p->se.load;
1308
1309	/*
1310	* SCHED_IDLE tasks get minimal weight:
1311	*/
1312	if (task_has_idle_policy(p)) {
1313	load->weight = scale_load(WEIGHT_IDLEPRIO);
1314	load->inv_weight = WMULT_IDLEPRIO;
1315	return;
1316	}
1317
1318	/*
1319	* SCHED_OTHER tasks have to update their load when changing their
1320	* weight
1321	*/
1322	if (update_load && p->sched_class == &fair_sched_class) {
1323	reweight_task(p, prio);
1324	} else {
1325	load->weight = scale_load(sched_prio_to_weight[prio]);
1326	load->inv_weight = sched_prio_to_wmult[prio];
1327	}
1328	}
1329
1330	#ifdef CONFIG_UCLAMP_TASK
1331	/*
1332	* Serializes updates of utilization clamp values
1333	*
1334	* The (slow-path) user-space triggers utilization clamp value updates which
1335	* can require updates on (fast-path) scheduler's data structures used to
1336	* support enqueue/dequeue operations.
1337	* While the per-CPU rq lock protects fast-path update operations, user-space
1338	* requests are serialized using a mutex to reduce the risk of conflicting
1339	* updates or API abuses.
1340	*/
1341	static DEFINE_MUTEX(uclamp_mutex);
1342
1343	/* Max allowed minimum utilization */
1344	static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1345
1346	/* Max allowed maximum utilization */
1347	static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1348
1349	/*
1350	* By default RT tasks run at the maximum performance point/capacity of the
1351	* system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1352	* SCHED_CAPACITY_SCALE.
1353	*
1354	* This knob allows admins to change the default behavior when uclamp is being
1355	* used. In battery powered devices, particularly, running at the maximum
1356	* capacity and frequency will increase energy consumption and shorten the
1357	* battery life.
1358	*
1359	* This knob only affects RT tasks that their uclamp_se->user_defined == false.
1360	*
1361	* This knob will not override the system default sched_util_clamp_min defined
1362	* above.
1363	*/
1364	static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1365
1366	/* All clamps are required to be less or equal than these values */
1367	static struct uclamp_se uclamp_default[UCLAMP_CNT];
1368
1369	/*
1370	* This static key is used to reduce the uclamp overhead in the fast path. It
1371	* primarily disables the call to uclamp_rq_{inc, dec}() in
1372	* enqueue/dequeue_task().
1373	*
1374	* This allows users to continue to enable uclamp in their kernel config with
1375	* minimum uclamp overhead in the fast path.
1376	*
1377	* As soon as userspace modifies any of the uclamp knobs, the static key is
1378	* enabled, since we have an actual users that make use of uclamp
1379	* functionality.
1380	*
1381	* The knobs that would enable this static key are:
1382	*
1383	* * A task modifying its uclamp value with sched_setattr().
1384	* * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1385	* * An admin modifying the cgroup cpu.uclamp.{min, max}
1386	*/
1387	DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1388
1389	/* Integer rounded range for each bucket */
1390	#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1391
1392	#define for_each_clamp_id(clamp_id) \
1393	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1394
1395	static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1396	{
1397	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1398	}
1399
1400	static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1401	{
1402	if (clamp_id == UCLAMP_MIN)
1403	return 0;
1404	return SCHED_CAPACITY_SCALE;
1405	}
1406
1407	static inline void uclamp_se_set(struct uclamp_se *uc_se,
1408	unsigned int value, bool user_defined)
1409	{
1410	uc_se->value = value;
1411	uc_se->bucket_id = uclamp_bucket_id(clamp_value: value);
1412	uc_se->user_defined = user_defined;
1413	}
1414
1415	static inline unsigned int
1416	uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1417	unsigned int clamp_value)
1418	{
1419	/*
1420	* Avoid blocked utilization pushing up the frequency when we go
1421	* idle (which drops the max-clamp) by retaining the last known
1422	* max-clamp.
1423	*/
1424	if (clamp_id == UCLAMP_MAX) {
1425	rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1426	return clamp_value;
1427	}
1428
1429	return uclamp_none(clamp_id: UCLAMP_MIN);
1430	}
1431
1432	static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1433	unsigned int clamp_value)
1434	{
1435	/* Reset max-clamp retention only on idle exit */
1436	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1437	return;
1438
1439	uclamp_rq_set(rq, clamp_id, value: clamp_value);
1440	}
1441
1442	static inline
1443	unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1444	unsigned int clamp_value)
1445	{
1446	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1447	int bucket_id = UCLAMP_BUCKETS - 1;
1448
1449	/*
1450	* Since both min and max clamps are max aggregated, find the
1451	* top most bucket with tasks in.
1452	*/
1453	for ( ; bucket_id >= 0; bucket_id--) {
1454	if (!bucket[bucket_id].tasks)
1455	continue;
1456	return bucket[bucket_id].value;
1457	}
1458
1459	/* No tasks -- default clamp values */
1460	return uclamp_idle_value(rq, clamp_id, clamp_value);
1461	}
1462
1463	static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1464	{
1465	unsigned int default_util_min;
1466	struct uclamp_se *uc_se;
1467
1468	lockdep_assert_held(&p->pi_lock);
1469
1470	uc_se = &p->uclamp_req[UCLAMP_MIN];
1471
1472	/* Only sync if user didn't override the default */
1473	if (uc_se->user_defined)
1474	return;
1475
1476	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1477	uclamp_se_set(uc_se, value: default_util_min, user_defined: false);
1478	}
1479
1480	static void uclamp_update_util_min_rt_default(struct task_struct *p)
1481	{
1482	if (!rt_task(p))
1483	return;
1484
1485	/* Protect updates to p->uclamp_* */
1486	guard(task_rq_lock)(l: p);
1487	__uclamp_update_util_min_rt_default(p);
1488	}
1489
1490	static inline struct uclamp_se
1491	uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1492	{
1493	/* Copy by value as we could modify it */
1494	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1495	#ifdef CONFIG_UCLAMP_TASK_GROUP
1496	unsigned int tg_min, tg_max, value;
1497
1498	/*
1499	* Tasks in autogroups or root task group will be
1500	* restricted by system defaults.
1501	*/
1502	if (task_group_is_autogroup(tg: task_group(p)))
1503	return uc_req;
1504	if (task_group(p) == &root_task_group)
1505	return uc_req;
1506
1507	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1508	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1509	value = uc_req.value;
1510	value = clamp(value, tg_min, tg_max);
1511	uclamp_se_set(uc_se: &uc_req, value, user_defined: false);
1512	#endif
1513
1514	return uc_req;
1515	}
1516
1517	/*
1518	* The effective clamp bucket index of a task depends on, by increasing
1519	* priority:
1520	* - the task specific clamp value, when explicitly requested from userspace
1521	* - the task group effective clamp value, for tasks not either in the root
1522	* group or in an autogroup
1523	* - the system default clamp value, defined by the sysadmin
1524	*/
1525	static inline struct uclamp_se
1526	uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1527	{
1528	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1529	struct uclamp_se uc_max = uclamp_default[clamp_id];
1530
1531	/* System default restrictions always apply */
1532	if (unlikely(uc_req.value > uc_max.value))
1533	return uc_max;
1534
1535	return uc_req;
1536	}
1537
1538	unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1539	{
1540	struct uclamp_se uc_eff;
1541
1542	/* Task currently refcounted: use back-annotated (effective) value */
1543	if (p->uclamp[clamp_id].active)
1544	return (unsigned long)p->uclamp[clamp_id].value;
1545
1546	uc_eff = uclamp_eff_get(p, clamp_id);
1547
1548	return (unsigned long)uc_eff.value;
1549	}
1550
1551	/*
1552	* When a task is enqueued on a rq, the clamp bucket currently defined by the
1553	* task's uclamp::bucket_id is refcounted on that rq. This also immediately
1554	* updates the rq's clamp value if required.
1555	*
1556	* Tasks can have a task-specific value requested from user-space, track
1557	* within each bucket the maximum value for tasks refcounted in it.
1558	* This "local max aggregation" allows to track the exact "requested" value
1559	* for each bucket when all its RUNNABLE tasks require the same clamp.
1560	*/
1561	static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1562	enum uclamp_id clamp_id)
1563	{
1564	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1565	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1566	struct uclamp_bucket *bucket;
1567
1568	lockdep_assert_rq_held(rq);
1569
1570	/* Update task effective clamp */
1571	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1572
1573	bucket = &uc_rq->bucket[uc_se->bucket_id];
1574	bucket->tasks++;
1575	uc_se->active = true;
1576
1577	uclamp_idle_reset(rq, clamp_id, clamp_value: uc_se->value);
1578
1579	/*
1580	* Local max aggregation: rq buckets always track the max
1581	* "requested" clamp value of its RUNNABLE tasks.
1582	*/
1583	if (bucket->tasks == 1 || uc_se->value > bucket->value)
1584	bucket->value = uc_se->value;
1585
1586	if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1587	uclamp_rq_set(rq, clamp_id, value: uc_se->value);
1588	}
1589
1590	/*
1591	* When a task is dequeued from a rq, the clamp bucket refcounted by the task
1592	* is released. If this is the last task reference counting the rq's max
1593	* active clamp value, then the rq's clamp value is updated.
1594	*
1595	* Both refcounted tasks and rq's cached clamp values are expected to be
1596	* always valid. If it's detected they are not, as defensive programming,
1597	* enforce the expected state and warn.
1598	*/
1599	static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1600	enum uclamp_id clamp_id)
1601	{
1602	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1603	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1604	struct uclamp_bucket *bucket;
1605	unsigned int bkt_clamp;
1606	unsigned int rq_clamp;
1607
1608	lockdep_assert_rq_held(rq);
1609
1610	/*
1611	* If sched_uclamp_used was enabled after task @p was enqueued,
1612	* we could end up with unbalanced call to uclamp_rq_dec_id().
1613	*
1614	* In this case the uc_se->active flag should be false since no uclamp
1615	* accounting was performed at enqueue time and we can just return
1616	* here.
1617	*
1618	* Need to be careful of the following enqueue/dequeue ordering
1619	* problem too
1620	*
1621	* enqueue(taskA)
1622	* // sched_uclamp_used gets enabled
1623	* enqueue(taskB)
1624	* dequeue(taskA)
1625	* // Must not decrement bucket->tasks here
1626	* dequeue(taskB)
1627	*
1628	* where we could end up with stale data in uc_se and
1629	* bucket[uc_se->bucket_id].
1630	*
1631	* The following check here eliminates the possibility of such race.
1632	*/
1633	if (unlikely(!uc_se->active))
1634	return;
1635
1636	bucket = &uc_rq->bucket[uc_se->bucket_id];
1637
1638	SCHED_WARN_ON(!bucket->tasks);
1639	if (likely(bucket->tasks))
1640	bucket->tasks--;
1641
1642	uc_se->active = false;
1643
1644	/*
1645	* Keep "local max aggregation" simple and accept to (possibly)
1646	* overboost some RUNNABLE tasks in the same bucket.
1647	* The rq clamp bucket value is reset to its base value whenever
1648	* there are no more RUNNABLE tasks refcounting it.
1649	*/
1650	if (likely(bucket->tasks))
1651	return;
1652
1653	rq_clamp = uclamp_rq_get(rq, clamp_id);
1654	/*
1655	* Defensive programming: this should never happen. If it happens,
1656	* e.g. due to future modification, warn and fixup the expected value.
1657	*/
1658	SCHED_WARN_ON(bucket->value > rq_clamp);
1659	if (bucket->value >= rq_clamp) {
1660	bkt_clamp = uclamp_rq_max_value(rq, clamp_id, clamp_value: uc_se->value);
1661	uclamp_rq_set(rq, clamp_id, value: bkt_clamp);
1662	}
1663	}
1664
1665	static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1666	{
1667	enum uclamp_id clamp_id;
1668
1669	/*
1670	* Avoid any overhead until uclamp is actually used by the userspace.
1671	*
1672	* The condition is constructed such that a NOP is generated when
1673	* sched_uclamp_used is disabled.
1674	*/
1675	if (!static_branch_unlikely(&sched_uclamp_used))
1676	return;
1677
1678	if (unlikely(!p->sched_class->uclamp_enabled))
1679	return;
1680
1681	for_each_clamp_id(clamp_id)
1682	uclamp_rq_inc_id(rq, p, clamp_id);
1683
1684	/* Reset clamp idle holding when there is one RUNNABLE task */
1685	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1686	rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1687	}
1688
1689	static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1690	{
1691	enum uclamp_id clamp_id;
1692
1693	/*
1694	* Avoid any overhead until uclamp is actually used by the userspace.
1695	*
1696	* The condition is constructed such that a NOP is generated when
1697	* sched_uclamp_used is disabled.
1698	*/
1699	if (!static_branch_unlikely(&sched_uclamp_used))
1700	return;
1701
1702	if (unlikely(!p->sched_class->uclamp_enabled))
1703	return;
1704
1705	for_each_clamp_id(clamp_id)
1706	uclamp_rq_dec_id(rq, p, clamp_id);
1707	}
1708
1709	static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1710	enum uclamp_id clamp_id)
1711	{
1712	if (!p->uclamp[clamp_id].active)
1713	return;
1714
1715	uclamp_rq_dec_id(rq, p, clamp_id);
1716	uclamp_rq_inc_id(rq, p, clamp_id);
1717
1718	/*
1719	* Make sure to clear the idle flag if we've transiently reached 0
1720	* active tasks on rq.
1721	*/
1722	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1723	rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1724	}
1725
1726	static inline void
1727	uclamp_update_active(struct task_struct *p)
1728	{
1729	enum uclamp_id clamp_id;
1730	struct rq_flags rf;
1731	struct rq *rq;
1732
1733	/*
1734	* Lock the task and the rq where the task is (or was) queued.
1735	*
1736	* We might lock the (previous) rq of a !RUNNABLE task, but that's the
1737	* price to pay to safely serialize util_{min,max} updates with
1738	* enqueues, dequeues and migration operations.
1739	* This is the same locking schema used by __set_cpus_allowed_ptr().
1740	*/
1741	rq = task_rq_lock(p, rf: &rf);
1742
1743	/*
1744	* Setting the clamp bucket is serialized by task_rq_lock().
1745	* If the task is not yet RUNNABLE and its task_struct is not
1746	* affecting a valid clamp bucket, the next time it's enqueued,
1747	* it will already see the updated clamp bucket value.
1748	*/
1749	for_each_clamp_id(clamp_id)
1750	uclamp_rq_reinc_id(rq, p, clamp_id);
1751
1752	task_rq_unlock(rq, p, rf: &rf);
1753	}
1754
1755	#ifdef CONFIG_UCLAMP_TASK_GROUP
1756	static inline void
1757	uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1758	{
1759	struct css_task_iter it;
1760	struct task_struct *p;
1761
1762	css_task_iter_start(css, flags: 0, it: &it);
1763	while ((p = css_task_iter_next(it: &it)))
1764	uclamp_update_active(p);
1765	css_task_iter_end(it: &it);
1766	}
1767
1768	static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1769	#endif
1770
1771	#ifdef CONFIG_SYSCTL
1772	#ifdef CONFIG_UCLAMP_TASK
1773	#ifdef CONFIG_UCLAMP_TASK_GROUP
1774	static void uclamp_update_root_tg(void)
1775	{
1776	struct task_group *tg = &root_task_group;
1777
1778	uclamp_se_set(uc_se: &tg->uclamp_req[UCLAMP_MIN],
1779	value: sysctl_sched_uclamp_util_min, user_defined: false);
1780	uclamp_se_set(uc_se: &tg->uclamp_req[UCLAMP_MAX],
1781	value: sysctl_sched_uclamp_util_max, user_defined: false);
1782
1783	guard(rcu)();
1784	cpu_util_update_eff(css: &root_task_group.css);
1785	}
1786	#else
1787	static void uclamp_update_root_tg(void) { }
1788	#endif
1789
1790	static void uclamp_sync_util_min_rt_default(void)
1791	{
1792	struct task_struct *g, *p;
1793
1794	/*
1795	* copy_process() sysctl_uclamp
1796	* uclamp_min_rt = X;
1797	* write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1798	* // link thread smp_mb__after_spinlock()
1799	* write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1800	* sched_post_fork() for_each_process_thread()
1801	* __uclamp_sync_rt() __uclamp_sync_rt()
1802	*
1803	* Ensures that either sched_post_fork() will observe the new
1804	* uclamp_min_rt or for_each_process_thread() will observe the new
1805	* task.
1806	*/
1807	read_lock(&tasklist_lock);
1808	smp_mb__after_spinlock();
1809	read_unlock(&tasklist_lock);
1810
1811	guard(rcu)();
1812	for_each_process_thread(g, p)
1813	uclamp_update_util_min_rt_default(p);
1814	}
1815
1816	static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1817	void *buffer, size_t *lenp, loff_t *ppos)
1818	{
1819	bool update_root_tg = false;
1820	int old_min, old_max, old_min_rt;
1821	int result;
1822
1823	guard(mutex)(T: &uclamp_mutex);
1824
1825	old_min = sysctl_sched_uclamp_util_min;
1826	old_max = sysctl_sched_uclamp_util_max;
1827	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1828
1829	result = proc_dointvec(table, write, buffer, lenp, ppos);
1830	if (result)
1831	goto undo;
1832	if (!write)
1833	return 0;
1834
1835	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1836	sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1837	sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1838
1839	result = -EINVAL;
1840	goto undo;
1841	}
1842
1843	if (old_min != sysctl_sched_uclamp_util_min) {
1844	uclamp_se_set(uc_se: &uclamp_default[UCLAMP_MIN],
1845	value: sysctl_sched_uclamp_util_min, user_defined: false);
1846	update_root_tg = true;
1847	}
1848	if (old_max != sysctl_sched_uclamp_util_max) {
1849	uclamp_se_set(uc_se: &uclamp_default[UCLAMP_MAX],
1850	value: sysctl_sched_uclamp_util_max, user_defined: false);
1851	update_root_tg = true;
1852	}
1853
1854	if (update_root_tg) {
1855	static_branch_enable(&sched_uclamp_used);
1856	uclamp_update_root_tg();
1857	}
1858
1859	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1860	static_branch_enable(&sched_uclamp_used);
1861	uclamp_sync_util_min_rt_default();
1862	}
1863
1864	/*
1865	* We update all RUNNABLE tasks only when task groups are in use.
1866	* Otherwise, keep it simple and do just a lazy update at each next
1867	* task enqueue time.
1868	*/
1869	return 0;
1870
1871	undo:
1872	sysctl_sched_uclamp_util_min = old_min;
1873	sysctl_sched_uclamp_util_max = old_max;
1874	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1875	return result;
1876	}
1877	#endif
1878	#endif
1879
1880	static int uclamp_validate(struct task_struct *p,
1881	const struct sched_attr *attr)
1882	{
1883	int util_min = p->uclamp_req[UCLAMP_MIN].value;
1884	int util_max = p->uclamp_req[UCLAMP_MAX].value;
1885
1886	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1887	util_min = attr->sched_util_min;
1888
1889	if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1890	return -EINVAL;
1891	}
1892
1893	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1894	util_max = attr->sched_util_max;
1895
1896	if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1897	return -EINVAL;
1898	}
1899
1900	if (util_min != -1 && util_max != -1 && util_min > util_max)
1901	return -EINVAL;
1902
1903	/*
1904	* We have valid uclamp attributes; make sure uclamp is enabled.
1905	*
1906	* We need to do that here, because enabling static branches is a
1907	* blocking operation which obviously cannot be done while holding
1908	* scheduler locks.
1909	*/
1910	static_branch_enable(&sched_uclamp_used);
1911
1912	return 0;
1913	}
1914
1915	static bool uclamp_reset(const struct sched_attr *attr,
1916	enum uclamp_id clamp_id,
1917	struct uclamp_se *uc_se)
1918	{
1919	/* Reset on sched class change for a non user-defined clamp value. */
1920	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1921	!uc_se->user_defined)
1922	return true;
1923
1924	/* Reset on sched_util_{min,max} == -1. */
1925	if (clamp_id == UCLAMP_MIN &&
1926	attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1927	attr->sched_util_min == -1) {
1928	return true;
1929	}
1930
1931	if (clamp_id == UCLAMP_MAX &&
1932	attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1933	attr->sched_util_max == -1) {
1934	return true;
1935	}
1936
1937	return false;
1938	}
1939
1940	static void __setscheduler_uclamp(struct task_struct *p,
1941	const struct sched_attr *attr)
1942	{
1943	enum uclamp_id clamp_id;
1944
1945	for_each_clamp_id(clamp_id) {
1946	struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1947	unsigned int value;
1948
1949	if (!uclamp_reset(attr, clamp_id, uc_se))
1950	continue;
1951
1952	/*
1953	* RT by default have a 100% boost value that could be modified
1954	* at runtime.
1955	*/
1956	if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1957	value = sysctl_sched_uclamp_util_min_rt_default;
1958	else
1959	value = uclamp_none(clamp_id);
1960
1961	uclamp_se_set(uc_se, value, user_defined: false);
1962
1963	}
1964
1965	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1966	return;
1967
1968	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1969	attr->sched_util_min != -1) {
1970	uclamp_se_set(uc_se: &p->uclamp_req[UCLAMP_MIN],
1971	value: attr->sched_util_min, user_defined: true);
1972	}
1973
1974	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1975	attr->sched_util_max != -1) {
1976	uclamp_se_set(uc_se: &p->uclamp_req[UCLAMP_MAX],
1977	value: attr->sched_util_max, user_defined: true);
1978	}
1979	}
1980
1981	static void uclamp_fork(struct task_struct *p)
1982	{
1983	enum uclamp_id clamp_id;
1984
1985	/*
1986	* We don't need to hold task_rq_lock() when updating p->uclamp_* here
1987	* as the task is still at its early fork stages.
1988	*/
1989	for_each_clamp_id(clamp_id)
1990	p->uclamp[clamp_id].active = false;
1991
1992	if (likely(!p->sched_reset_on_fork))
1993	return;
1994
1995	for_each_clamp_id(clamp_id) {
1996	uclamp_se_set(uc_se: &p->uclamp_req[clamp_id],
1997	value: uclamp_none(clamp_id), user_defined: false);
1998	}
1999	}
2000
2001	static void uclamp_post_fork(struct task_struct *p)
2002	{
2003	uclamp_update_util_min_rt_default(p);
2004	}
2005
2006	static void __init init_uclamp_rq(struct rq *rq)
2007	{
2008	enum uclamp_id clamp_id;
2009	struct uclamp_rq *uc_rq = rq->uclamp;
2010
2011	for_each_clamp_id(clamp_id) {
2012	uc_rq[clamp_id] = (struct uclamp_rq) {
2013	.value = uclamp_none(clamp_id)
2014	};
2015	}
2016
2017	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2018	}
2019
2020	static void __init init_uclamp(void)
2021	{
2022	struct uclamp_se uc_max = {};
2023	enum uclamp_id clamp_id;
2024	int cpu;
2025
2026	for_each_possible_cpu(cpu)
2027	init_uclamp_rq(cpu_rq(cpu));
2028
2029	for_each_clamp_id(clamp_id) {
2030	uclamp_se_set(uc_se: &init_task.uclamp_req[clamp_id],
2031	value: uclamp_none(clamp_id), user_defined: false);
2032	}
2033
2034	/* System defaults allow max clamp values for both indexes */
2035	uclamp_se_set(uc_se: &uc_max, value: uclamp_none(clamp_id: UCLAMP_MAX), user_defined: false);
2036	for_each_clamp_id(clamp_id) {
2037	uclamp_default[clamp_id] = uc_max;
2038	#ifdef CONFIG_UCLAMP_TASK_GROUP
2039	root_task_group.uclamp_req[clamp_id] = uc_max;
2040	root_task_group.uclamp[clamp_id] = uc_max;
2041	#endif
2042	}
2043	}
2044
2045	#else /* CONFIG_UCLAMP_TASK */
2046	static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2047	static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2048	static inline int uclamp_validate(struct task_struct *p,
2049	const struct sched_attr *attr)
2050	{
2051	return -EOPNOTSUPP;
2052	}
2053	static void __setscheduler_uclamp(struct task_struct *p,
2054	const struct sched_attr *attr) { }
2055	static inline void uclamp_fork(struct task_struct *p) { }
2056	static inline void uclamp_post_fork(struct task_struct *p) { }
2057	static inline void init_uclamp(void) { }
2058	#endif /* CONFIG_UCLAMP_TASK */
2059
2060	bool sched_task_on_rq(struct task_struct *p)
2061	{
2062	return task_on_rq_queued(p);
2063	}
2064
2065	unsigned long get_wchan(struct task_struct *p)
2066	{
2067	unsigned long ip = 0;
2068	unsigned int state;
2069
2070	if (!p || p == current)
2071	return 0;
2072
2073	/* Only get wchan if task is blocked and we can keep it that way. */
2074	raw_spin_lock_irq(&p->pi_lock);
2075	state = READ_ONCE(p->__state);
2076	smp_rmb(); /* see try_to_wake_up() */
2077	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2078	ip = __get_wchan(p);
2079	raw_spin_unlock_irq(&p->pi_lock);
2080
2081	return ip;
2082	}
2083
2084	static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2085	{
2086	if (!(flags & ENQUEUE_NOCLOCK))
2087	update_rq_clock(rq);
2088
2089	if (!(flags & ENQUEUE_RESTORE)) {
2090	sched_info_enqueue(rq, t: p);
2091	psi_enqueue(p, wakeup: (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2092	}
2093
2094	uclamp_rq_inc(rq, p);
2095	p->sched_class->enqueue_task(rq, p, flags);
2096
2097	if (sched_core_enabled(rq))
2098	sched_core_enqueue(rq, p);
2099	}
2100
2101	static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2102	{
2103	if (sched_core_enabled(rq))
2104	sched_core_dequeue(rq, p, flags);
2105
2106	if (!(flags & DEQUEUE_NOCLOCK))
2107	update_rq_clock(rq);
2108
2109	if (!(flags & DEQUEUE_SAVE)) {
2110	sched_info_dequeue(rq, t: p);
2111	psi_dequeue(p, sleep: flags & DEQUEUE_SLEEP);
2112	}
2113
2114	uclamp_rq_dec(rq, p);
2115	p->sched_class->dequeue_task(rq, p, flags);
2116	}
2117
2118	void activate_task(struct rq *rq, struct task_struct *p, int flags)
2119	{
2120	if (task_on_rq_migrating(p))
2121	flags |= ENQUEUE_MIGRATED;
2122	if (flags & ENQUEUE_MIGRATED)
2123	sched_mm_cid_migrate_to(dst_rq: rq, t: p);
2124
2125	enqueue_task(rq, p, flags);
2126
2127	p->on_rq = TASK_ON_RQ_QUEUED;
2128	}
2129
2130	void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2131	{
2132	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2133
2134	dequeue_task(rq, p, flags);
2135	}
2136
2137	static inline int __normal_prio(int policy, int rt_prio, int nice)
2138	{
2139	int prio;
2140
2141	if (dl_policy(policy))
2142	prio = MAX_DL_PRIO - 1;
2143	else if (rt_policy(policy))
2144	prio = MAX_RT_PRIO - 1 - rt_prio;
2145	else
2146	prio = NICE_TO_PRIO(nice);
2147
2148	return prio;
2149	}
2150
2151	/*
2152	* Calculate the expected normal priority: i.e. priority
2153	* without taking RT-inheritance into account. Might be
2154	* boosted by interactivity modifiers. Changes upon fork,
2155	* setprio syscalls, and whenever the interactivity
2156	* estimator recalculates.
2157	*/
2158	static inline int normal_prio(struct task_struct *p)
2159	{
2160	return __normal_prio(policy: p->policy, rt_prio: p->rt_priority, PRIO_TO_NICE(p->static_prio));
2161	}
2162
2163	/*
2164	* Calculate the current priority, i.e. the priority
2165	* taken into account by the scheduler. This value might
2166	* be boosted by RT tasks, or might be boosted by
2167	* interactivity modifiers. Will be RT if the task got
2168	* RT-boosted. If not then it returns p->normal_prio.
2169	*/
2170	static int effective_prio(struct task_struct *p)
2171	{
2172	p->normal_prio = normal_prio(p);
2173	/*
2174	* If we are RT tasks or we were boosted to RT priority,
2175	* keep the priority unchanged. Otherwise, update priority
2176	* to the normal priority:
2177	*/
2178	if (!rt_prio(prio: p->prio))
2179	return p->normal_prio;
2180	return p->prio;
2181	}
2182
2183	/**
2184	* task_curr - is this task currently executing on a CPU?
2185	* @p: the task in question.
2186	*
2187	* Return: 1 if the task is currently executing. 0 otherwise.
2188	*/
2189	inline int task_curr(const struct task_struct *p)
2190	{
2191	return cpu_curr(task_cpu(p)) == p;
2192	}
2193
2194	/*
2195	* switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2196	* use the balance_callback list if you want balancing.
2197	*
2198	* this means any call to check_class_changed() must be followed by a call to
2199	* balance_callback().
2200	*/
2201	static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2202	const struct sched_class *prev_class,
2203	int oldprio)
2204	{
2205	if (prev_class != p->sched_class) {
2206	if (prev_class->switched_from)
2207	prev_class->switched_from(rq, p);
2208
2209	p->sched_class->switched_to(rq, p);
2210	} else if (oldprio != p->prio || dl_task(p))
2211	p->sched_class->prio_changed(rq, p, oldprio);
2212	}
2213
2214	void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
2215	{
2216	if (p->sched_class == rq->curr->sched_class)
2217	rq->curr->sched_class->wakeup_preempt(rq, p, flags);
2218	else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2219	resched_curr(rq);
2220
2221	/*
2222	* A queue event has occurred, and we're going to schedule. In
2223	* this case, we can save a useless back to back clock update.
2224	*/
2225	if (task_on_rq_queued(p: rq->curr) && test_tsk_need_resched(tsk: rq->curr))
2226	rq_clock_skip_update(rq);
2227	}
2228
2229	static __always_inline
2230	int __task_state_match(struct task_struct *p, unsigned int state)
2231	{
2232	if (READ_ONCE(p->__state) & state)
2233	return 1;
2234
2235	if (READ_ONCE(p->saved_state) & state)
2236	return -1;
2237
2238	return 0;
2239	}
2240
2241	static __always_inline
2242	int task_state_match(struct task_struct *p, unsigned int state)
2243	{
2244	/*
2245	* Serialize against current_save_and_set_rtlock_wait_state(),
2246	* current_restore_rtlock_saved_state(), and __refrigerator().
2247	*/
2248	guard(raw_spinlock_irq)(l: &p->pi_lock);
2249	return __task_state_match(p, state);
2250	}
2251
2252	/*
2253	* wait_task_inactive - wait for a thread to unschedule.
2254	*
2255	* Wait for the thread to block in any of the states set in @match_state.
2256	* If it changes, i.e. @p might have woken up, then return zero. When we
2257	* succeed in waiting for @p to be off its CPU, we return a positive number
2258	* (its total switch count). If a second call a short while later returns the
2259	* same number, the caller can be sure that @p has remained unscheduled the
2260	* whole time.
2261	*
2262	* The caller must ensure that the task *will* unschedule sometime soon,
2263	* else this function might spin for a *long* time. This function can't
2264	* be called with interrupts off, or it may introduce deadlock with
2265	* smp_call_function() if an IPI is sent by the same process we are
2266	* waiting to become inactive.
2267	*/
2268	unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2269	{
2270	int running, queued, match;
2271	struct rq_flags rf;
2272	unsigned long ncsw;
2273	struct rq *rq;
2274
2275	for (;;) {
2276	/*
2277	* We do the initial early heuristics without holding
2278	* any task-queue locks at all. We'll only try to get
2279	* the runqueue lock when things look like they will
2280	* work out!
2281	*/
2282	rq = task_rq(p);
2283
2284	/*
2285	* If the task is actively running on another CPU
2286	* still, just relax and busy-wait without holding
2287	* any locks.
2288	*
2289	* NOTE! Since we don't hold any locks, it's not
2290	* even sure that "rq" stays as the right runqueue!
2291	* But we don't care, since "task_on_cpu()" will
2292	* return false if the runqueue has changed and p
2293	* is actually now running somewhere else!
2294	*/
2295	while (task_on_cpu(rq, p)) {
2296	if (!task_state_match(p, state: match_state))
2297	return 0;
2298	cpu_relax();
2299	}
2300
2301	/*
2302	* Ok, time to look more closely! We need the rq
2303	* lock now, to be *sure*. If we're wrong, we'll
2304	* just go back and repeat.
2305	*/
2306	rq = task_rq_lock(p, rf: &rf);
2307	trace_sched_wait_task(p);
2308	running = task_on_cpu(rq, p);
2309	queued = task_on_rq_queued(p);
2310	ncsw = 0;
2311	if ((match = __task_state_match(p, state: match_state))) {
2312	/*
2313	* When matching on p->saved_state, consider this task
2314	* still queued so it will wait.
2315	*/
2316	if (match < 0)
2317	queued = 1;
2318	ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2319	}
2320	task_rq_unlock(rq, p, rf: &rf);
2321
2322	/*
2323	* If it changed from the expected state, bail out now.
2324	*/
2325	if (unlikely(!ncsw))
2326	break;
2327
2328	/*
2329	* Was it really running after all now that we
2330	* checked with the proper locks actually held?
2331	*
2332	* Oops. Go back and try again..
2333	*/
2334	if (unlikely(running)) {
2335	cpu_relax();
2336	continue;
2337	}
2338
2339	/*
2340	* It's not enough that it's not actively running,
2341	* it must be off the runqueue _entirely_, and not
2342	* preempted!
2343	*
2344	* So if it was still runnable (but just not actively
2345	* running right now), it's preempted, and we should
2346	* yield - it could be a while.
2347	*/
2348	if (unlikely(queued)) {
2349	ktime_t to = NSEC_PER_SEC / HZ;
2350
2351	set_current_state(TASK_UNINTERRUPTIBLE);
2352	schedule_hrtimeout(expires: &to, mode: HRTIMER_MODE_REL_HARD);
2353	continue;
2354	}
2355
2356	/*
2357	* Ahh, all good. It wasn't running, and it wasn't
2358	* runnable, which means that it will never become
2359	* running in the future either. We're all done!
2360	*/
2361	break;
2362	}
2363
2364	return ncsw;
2365	}
2366
2367	#ifdef CONFIG_SMP
2368
2369	static void
2370	__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2371
2372	static int __set_cpus_allowed_ptr(struct task_struct *p,
2373	struct affinity_context *ctx);
2374
2375	static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2376	{
2377	struct affinity_context ac = {
2378	.new_mask = cpumask_of(rq->cpu),
2379	.flags = SCA_MIGRATE_DISABLE,
2380	};
2381
2382	if (likely(!p->migration_disabled))
2383	return;
2384
2385	if (p->cpus_ptr != &p->cpus_mask)
2386	return;
2387
2388	/*
2389	* Violates locking rules! see comment in __do_set_cpus_allowed().
2390	*/
2391	__do_set_cpus_allowed(p, ctx: &ac);
2392	}
2393
2394	void migrate_disable(void)
2395	{
2396	struct task_struct *p = current;
2397
2398	if (p->migration_disabled) {
2399	p->migration_disabled++;
2400	return;
2401	}
2402
2403	guard(preempt)();
2404	this_rq()->nr_pinned++;
2405	p->migration_disabled = 1;
2406	}
2407	EXPORT_SYMBOL_GPL(migrate_disable);
2408
2409	void migrate_enable(void)
2410	{
2411	struct task_struct *p = current;
2412	struct affinity_context ac = {
2413	.new_mask = &p->cpus_mask,
2414	.flags = SCA_MIGRATE_ENABLE,
2415	};
2416
2417	if (p->migration_disabled > 1) {
2418	p->migration_disabled--;
2419	return;
2420	}
2421
2422	if (WARN_ON_ONCE(!p->migration_disabled))
2423	return;
2424
2425	/*
2426	* Ensure stop_task runs either before or after this, and that
2427	* __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2428	*/
2429	guard(preempt)();
2430	if (p->cpus_ptr != &p->cpus_mask)
2431	__set_cpus_allowed_ptr(p, ctx: &ac);
2432	/*
2433	* Mustn't clear migration_disabled() until cpus_ptr points back at the
2434	* regular cpus_mask, otherwise things that race (eg.
2435	* select_fallback_rq) get confused.
2436	*/
2437	barrier();
2438	p->migration_disabled = 0;
2439	this_rq()->nr_pinned--;
2440	}
2441	EXPORT_SYMBOL_GPL(migrate_enable);
2442
2443	static inline bool rq_has_pinned_tasks(struct rq *rq)
2444	{
2445	return rq->nr_pinned;
2446	}
2447
2448	/*
2449	* Per-CPU kthreads are allowed to run on !active && online CPUs, see
2450	* __set_cpus_allowed_ptr() and select_fallback_rq().
2451	*/
2452	static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2453	{
2454	/* When not in the task's cpumask, no point in looking further. */
2455	if (!cpumask_test_cpu(cpu, cpumask: p->cpus_ptr))
2456	return false;
2457
2458	/* migrate_disabled() must be allowed to finish. */
2459	if (is_migration_disabled(p))
2460	return cpu_online(cpu);
2461
2462	/* Non kernel threads are not allowed during either online or offline. */
2463	if (!(p->flags & PF_KTHREAD))
2464	return cpu_active(cpu) && task_cpu_possible(cpu, p);
2465
2466	/* KTHREAD_IS_PER_CPU is always allowed. */
2467	if (kthread_is_per_cpu(k: p))
2468	return cpu_online(cpu);
2469
2470	/* Regular kernel threads don't get to stay during offline. */
2471	if (cpu_dying(cpu))
2472	return false;
2473
2474	/* But are allowed during online. */
2475	return cpu_online(cpu);
2476	}
2477
2478	/*
2479	* This is how migration works:
2480	*
2481	* 1) we invoke migration_cpu_stop() on the target CPU using
2482	* stop_one_cpu().
2483	* 2) stopper starts to run (implicitly forcing the migrated thread
2484	* off the CPU)
2485	* 3) it checks whether the migrated task is still in the wrong runqueue.
2486	* 4) if it's in the wrong runqueue then the migration thread removes
2487	* it and puts it into the right queue.
2488	* 5) stopper completes and stop_one_cpu() returns and the migration
2489	* is done.
2490	*/
2491
2492	/*
2493	* move_queued_task - move a queued task to new rq.
2494	*
2495	* Returns (locked) new rq. Old rq's lock is released.
2496	*/
2497	static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2498	struct task_struct *p, int new_cpu)
2499	{
2500	lockdep_assert_rq_held(rq);
2501
2502	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2503	set_task_cpu(p, cpu: new_cpu);
2504	rq_unlock(rq, rf);
2505
2506	rq = cpu_rq(new_cpu);
2507
2508	rq_lock(rq, rf);
2509	WARN_ON_ONCE(task_cpu(p) != new_cpu);
2510	activate_task(rq, p, flags: 0);
2511	wakeup_preempt(rq, p, flags: 0);
2512
2513	return rq;
2514	}
2515
2516	struct migration_arg {
2517	struct task_struct *task;
2518	int dest_cpu;
2519	struct set_affinity_pending *pending;
2520	};
2521
2522	/*
2523	* @refs: number of wait_for_completion()
2524	* @stop_pending: is @stop_work in use
2525	*/
2526	struct set_affinity_pending {
2527	refcount_t refs;
2528	unsigned int stop_pending;
2529	struct completion done;
2530	struct cpu_stop_work stop_work;
2531	struct migration_arg arg;
2532	};
2533
2534	/*
2535	* Move (not current) task off this CPU, onto the destination CPU. We're doing
2536	* this because either it can't run here any more (set_cpus_allowed()
2537	* away from this CPU, or CPU going down), or because we're
2538	* attempting to rebalance this task on exec (sched_exec).
2539	*
2540	* So we race with normal scheduler movements, but that's OK, as long
2541	* as the task is no longer on this CPU.
2542	*/
2543	static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2544	struct task_struct *p, int dest_cpu)
2545	{
2546	/* Affinity changed (again). */
2547	if (!is_cpu_allowed(p, cpu: dest_cpu))
2548	return rq;
2549
2550	rq = move_queued_task(rq, rf, p, new_cpu: dest_cpu);
2551
2552	return rq;
2553	}
2554
2555	/*
2556	* migration_cpu_stop - this will be executed by a highprio stopper thread
2557	* and performs thread migration by bumping thread off CPU then
2558	* 'pushing' onto another runqueue.
2559	*/
2560	static int migration_cpu_stop(void *data)
2561	{
2562	struct migration_arg *arg = data;
2563	struct set_affinity_pending *pending = arg->pending;
2564	struct task_struct *p = arg->task;
2565	struct rq *rq = this_rq();
2566	bool complete = false;
2567	struct rq_flags rf;
2568
2569	/*
2570	* The original target CPU might have gone down and we might
2571	* be on another CPU but it doesn't matter.
2572	*/
2573	local_irq_save(rf.flags);
2574	/*
2575	* We need to explicitly wake pending tasks before running
2576	* __migrate_task() such that we will not miss enforcing cpus_ptr
2577	* during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2578	*/
2579	flush_smp_call_function_queue();
2580
2581	raw_spin_lock(&p->pi_lock);
2582	rq_lock(rq, rf: &rf);
2583
2584	/*
2585	* If we were passed a pending, then ->stop_pending was set, thus
2586	* p->migration_pending must have remained stable.
2587	*/
2588	WARN_ON_ONCE(pending && pending != p->migration_pending);
2589
2590	/*
2591	* If task_rq(p) != rq, it cannot be migrated here, because we're
2592	* holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2593	* we're holding p->pi_lock.
2594	*/
2595	if (task_rq(p) == rq) {
2596	if (is_migration_disabled(p))
2597	goto out;
2598
2599	if (pending) {
2600	p->migration_pending = NULL;
2601	complete = true;
2602
2603	if (cpumask_test_cpu(cpu: task_cpu(p), cpumask: &p->cpus_mask))
2604	goto out;
2605	}
2606
2607	if (task_on_rq_queued(p)) {
2608	update_rq_clock(rq);
2609	rq = __migrate_task(rq, rf: &rf, p, dest_cpu: arg->dest_cpu);
2610	} else {
2611	p->wake_cpu = arg->dest_cpu;
2612	}
2613
2614	/*
2615	* XXX __migrate_task() can fail, at which point we might end
2616	* up running on a dodgy CPU, AFAICT this can only happen
2617	* during CPU hotplug, at which point we'll get pushed out
2618	* anyway, so it's probably not a big deal.
2619	*/
2620
2621	} else if (pending) {
2622	/*
2623	* This happens when we get migrated between migrate_enable()'s
2624	* preempt_enable() and scheduling the stopper task. At that
2625	* point we're a regular task again and not current anymore.
2626	*
2627	* A !PREEMPT kernel has a giant hole here, which makes it far
2628	* more likely.
2629	*/
2630
2631	/*
2632	* The task moved before the stopper got to run. We're holding
2633	* ->pi_lock, so the allowed mask is stable - if it got
2634	* somewhere allowed, we're done.
2635	*/
2636	if (cpumask_test_cpu(cpu: task_cpu(p), cpumask: p->cpus_ptr)) {
2637	p->migration_pending = NULL;
2638	complete = true;
2639	goto out;
2640	}
2641
2642	/*
2643	* When migrate_enable() hits a rq mis-match we can't reliably
2644	* determine is_migration_disabled() and so have to chase after
2645	* it.
2646	*/
2647	WARN_ON_ONCE(!pending->stop_pending);
2648	preempt_disable();
2649	task_rq_unlock(rq, p, rf: &rf);
2650	stop_one_cpu_nowait(cpu: task_cpu(p), fn: migration_cpu_stop,
2651	arg: &pending->arg, work_buf: &pending->stop_work);
2652	preempt_enable();
2653	return 0;
2654	}
2655	out:
2656	if (pending)
2657	pending->stop_pending = false;
2658	task_rq_unlock(rq, p, rf: &rf);
2659
2660	if (complete)
2661	complete_all(&pending->done);
2662
2663	return 0;
2664	}
2665
2666	int push_cpu_stop(void *arg)
2667	{
2668	struct rq *lowest_rq = NULL, *rq = this_rq();
2669	struct task_struct *p = arg;
2670
2671	raw_spin_lock_irq(&p->pi_lock);
2672	raw_spin_rq_lock(rq);
2673
2674	if (task_rq(p) != rq)
2675	goto out_unlock;
2676
2677	if (is_migration_disabled(p)) {
2678	p->migration_flags |= MDF_PUSH;
2679	goto out_unlock;
2680	}
2681
2682	p->migration_flags &= ~MDF_PUSH;
2683
2684	if (p->sched_class->find_lock_rq)
2685	lowest_rq = p->sched_class->find_lock_rq(p, rq);
2686
2687	if (!lowest_rq)
2688	goto out_unlock;
2689
2690	// XXX validate p is still the highest prio task
2691	if (task_rq(p) == rq) {
2692	deactivate_task(rq, p, flags: 0);
2693	set_task_cpu(p, cpu: lowest_rq->cpu);
2694	activate_task(rq: lowest_rq, p, flags: 0);
2695	resched_curr(rq: lowest_rq);
2696	}
2697
2698	double_unlock_balance(this_rq: rq, busiest: lowest_rq);
2699
2700	out_unlock:
2701	rq->push_busy = false;
2702	raw_spin_rq_unlock(rq);
2703	raw_spin_unlock_irq(&p->pi_lock);
2704
2705	put_task_struct(t: p);
2706	return 0;
2707	}
2708
2709	/*
2710	* sched_class::set_cpus_allowed must do the below, but is not required to
2711	* actually call this function.
2712	*/
2713	void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2714	{
2715	if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2716	p->cpus_ptr = ctx->new_mask;
2717	return;
2718	}
2719
2720	cpumask_copy(dstp: &p->cpus_mask, srcp: ctx->new_mask);
2721	p->nr_cpus_allowed = cpumask_weight(srcp: ctx->new_mask);
2722
2723	/*
2724	* Swap in a new user_cpus_ptr if SCA_USER flag set
2725	*/
2726	if (ctx->flags & SCA_USER)
2727	swap(p->user_cpus_ptr, ctx->user_mask);
2728	}
2729
2730	static void
2731	__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2732	{
2733	struct rq *rq = task_rq(p);
2734	bool queued, running;
2735
2736	/*
2737	* This here violates the locking rules for affinity, since we're only
2738	* supposed to change these variables while holding both rq->lock and
2739	* p->pi_lock.
2740	*
2741	* HOWEVER, it magically works, because ttwu() is the only code that
2742	* accesses these variables under p->pi_lock and only does so after
2743	* smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2744	* before finish_task().
2745	*
2746	* XXX do further audits, this smells like something putrid.
2747	*/
2748	if (ctx->flags & SCA_MIGRATE_DISABLE)
2749	SCHED_WARN_ON(!p->on_cpu);
2750	else
2751	lockdep_assert_held(&p->pi_lock);
2752
2753	queued = task_on_rq_queued(p);
2754	running = task_current(rq, p);
2755
2756	if (queued) {
2757	/*
2758	* Because __kthread_bind() calls this on blocked tasks without
2759	* holding rq->lock.
2760	*/
2761	lockdep_assert_rq_held(rq);
2762	dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2763	}
2764	if (running)
2765	put_prev_task(rq, prev: p);
2766
2767	p->sched_class->set_cpus_allowed(p, ctx);
2768
2769	if (queued)
2770	enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2771	if (running)
2772	set_next_task(rq, next: p);
2773	}
2774
2775	/*
2776	* Used for kthread_bind() and select_fallback_rq(), in both cases the user
2777	* affinity (if any) should be destroyed too.
2778	*/
2779	void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2780	{
2781	struct affinity_context ac = {
2782	.new_mask = new_mask,
2783	.user_mask = NULL,
2784	.flags = SCA_USER, /* clear the user requested mask */
2785	};
2786	union cpumask_rcuhead {
2787	cpumask_t cpumask;
2788	struct rcu_head rcu;
2789	};
2790
2791	__do_set_cpus_allowed(p, ctx: &ac);
2792
2793	/*
2794	* Because this is called with p->pi_lock held, it is not possible
2795	* to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2796	* kfree_rcu().
2797	*/
2798	kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2799	}
2800
2801	static cpumask_t *alloc_user_cpus_ptr(int node)
2802	{
2803	/*
2804	* See do_set_cpus_allowed() above for the rcu_head usage.
2805	*/
2806	int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2807
2808	return kmalloc_node(size, GFP_KERNEL, node);
2809	}
2810
2811	int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2812	int node)
2813	{
2814	cpumask_t *user_mask;
2815	unsigned long flags;
2816
2817	/*
2818	* Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2819	* may differ by now due to racing.
2820	*/
2821	dst->user_cpus_ptr = NULL;
2822
2823	/*
2824	* This check is racy and losing the race is a valid situation.
2825	* It is not worth the extra overhead of taking the pi_lock on
2826	* every fork/clone.
2827	*/
2828	if (data_race(!src->user_cpus_ptr))
2829	return 0;
2830
2831	user_mask = alloc_user_cpus_ptr(node);
2832	if (!user_mask)
2833	return -ENOMEM;
2834
2835	/*
2836	* Use pi_lock to protect content of user_cpus_ptr
2837	*
2838	* Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2839	* do_set_cpus_allowed().
2840	*/
2841	raw_spin_lock_irqsave(&src->pi_lock, flags);
2842	if (src->user_cpus_ptr) {
2843	swap(dst->user_cpus_ptr, user_mask);
2844	cpumask_copy(dstp: dst->user_cpus_ptr, srcp: src->user_cpus_ptr);
2845	}
2846	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2847
2848	if (unlikely(user_mask))
2849	kfree(objp: user_mask);
2850
2851	return 0;
2852	}
2853
2854	static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2855	{
2856	struct cpumask *user_mask = NULL;
2857
2858	swap(p->user_cpus_ptr, user_mask);
2859
2860	return user_mask;
2861	}
2862
2863	void release_user_cpus_ptr(struct task_struct *p)
2864	{
2865	kfree(objp: clear_user_cpus_ptr(p));
2866	}
2867
2868	/*
2869	* This function is wildly self concurrent; here be dragons.
2870	*
2871	*
2872	* When given a valid mask, __set_cpus_allowed_ptr() must block until the
2873	* designated task is enqueued on an allowed CPU. If that task is currently
2874	* running, we have to kick it out using the CPU stopper.
2875	*
2876	* Migrate-Disable comes along and tramples all over our nice sandcastle.
2877	* Consider:
2878	*
2879	* Initial conditions: P0->cpus_mask = [0, 1]
2880	*
2881	* P0@CPU0 P1
2882	*
2883	* migrate_disable();
2884	* <preempted>
2885	* set_cpus_allowed_ptr(P0, [1]);
2886	*
2887	* P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2888	* its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2889	* This means we need the following scheme:
2890	*
2891	* P0@CPU0 P1
2892	*
2893	* migrate_disable();
2894	* <preempted>
2895	* set_cpus_allowed_ptr(P0, [1]);
2896	* <blocks>
2897	* <resumes>
2898	* migrate_enable();
2899	* __set_cpus_allowed_ptr();
2900	* <wakes local stopper>
2901	* `--> <woken on migration completion>
2902	*
2903	* Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2904	* concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2905	* task p are serialized by p->pi_lock, which we can leverage: the one that
2906	* should come into effect at the end of the Migrate-Disable region is the last
2907	* one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2908	* but we still need to properly signal those waiting tasks at the appropriate
2909	* moment.
2910	*
2911	* This is implemented using struct set_affinity_pending. The first
2912	* __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2913	* setup an instance of that struct and install it on the targeted task_struct.
2914	* Any and all further callers will reuse that instance. Those then wait for
2915	* a completion signaled at the tail of the CPU stopper callback (1), triggered
2916	* on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2917	*
2918	*
2919	* (1) In the cases covered above. There is one more where the completion is
2920	* signaled within affine_move_task() itself: when a subsequent affinity request
2921	* occurs after the stopper bailed out due to the targeted task still being
2922	* Migrate-Disable. Consider:
2923	*
2924	* Initial conditions: P0->cpus_mask = [0, 1]
2925	*
2926	* CPU0 P1 P2
2927	* <P0>
2928	* migrate_disable();
2929	* <preempted>
2930	* set_cpus_allowed_ptr(P0, [1]);
2931	* <blocks>
2932	* <migration/0>
2933	* migration_cpu_stop()
2934	* is_migration_disabled()
2935	* <bails>
2936	* set_cpus_allowed_ptr(P0, [0, 1]);
2937	* <signal completion>
2938	* <awakes>
2939	*
2940	* Note that the above is safe vs a concurrent migrate_enable(), as any
2941	* pending affinity completion is preceded by an uninstallation of
2942	* p->migration_pending done with p->pi_lock held.
2943	*/
2944	static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2945	int dest_cpu, unsigned int flags)
2946	__releases(rq->lock)
2947	__releases(p->pi_lock)
2948	{
2949	struct set_affinity_pending my_pending = { }, *pending = NULL;
2950	bool stop_pending, complete = false;
2951
2952	/* Can the task run on the task's current CPU? If so, we're done */
2953	if (cpumask_test_cpu(cpu: task_cpu(p), cpumask: &p->cpus_mask)) {
2954	struct task_struct *push_task = NULL;
2955
2956	if ((flags & SCA_MIGRATE_ENABLE) &&
2957	(p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2958	rq->push_busy = true;
2959	push_task = get_task_struct(t: p);
2960	}
2961
2962	/*
2963	* If there are pending waiters, but no pending stop_work,
2964	* then complete now.
2965	*/
2966	pending = p->migration_pending;
2967	if (pending && !pending->stop_pending) {
2968	p->migration_pending = NULL;
2969	complete = true;
2970	}
2971
2972	preempt_disable();
2973	task_rq_unlock(rq, p, rf);
2974	if (push_task) {
2975	stop_one_cpu_nowait(cpu: rq->cpu, fn: push_cpu_stop,
2976	arg: p, work_buf: &rq->push_work);
2977	}
2978	preempt_enable();
2979
2980	if (complete)
2981	complete_all(&pending->done);
2982
2983	return 0;
2984	}
2985
2986	if (!(flags & SCA_MIGRATE_ENABLE)) {
2987	/* serialized by p->pi_lock */
2988	if (!p->migration_pending) {
2989	/* Install the request */
2990	refcount_set(r: &my_pending.refs, n: 1);
2991	init_completion(x: &my_pending.done);
2992	my_pending.arg = (struct migration_arg) {
2993	.task = p,
2994	.dest_cpu = dest_cpu,
2995	.pending = &my_pending,
2996	};
2997
2998	p->migration_pending = &my_pending;
2999	} else {
3000	pending = p->migration_pending;
3001	refcount_inc(r: &pending->refs);
3002	/*
3003	* Affinity has changed, but we've already installed a
3004	* pending. migration_cpu_stop() *must* see this, else
3005	* we risk a completion of the pending despite having a
3006	* task on a disallowed CPU.
3007	*
3008	* Serialized by p->pi_lock, so this is safe.
3009	*/
3010	pending->arg.dest_cpu = dest_cpu;
3011	}
3012	}
3013	pending = p->migration_pending;
3014	/*
3015	* - !MIGRATE_ENABLE:
3016	* we'll have installed a pending if there wasn't one already.
3017	*
3018	* - MIGRATE_ENABLE:
3019	* we're here because the current CPU isn't matching anymore,
3020	* the only way that can happen is because of a concurrent
3021	* set_cpus_allowed_ptr() call, which should then still be
3022	* pending completion.
3023	*
3024	* Either way, we really should have a @pending here.
3025	*/
3026	if (WARN_ON_ONCE(!pending)) {
3027	task_rq_unlock(rq, p, rf);
3028	return -EINVAL;
3029	}
3030
3031	if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3032	/*
3033	* MIGRATE_ENABLE gets here because 'p == current', but for
3034	* anything else we cannot do is_migration_disabled(), punt
3035	* and have the stopper function handle it all race-free.
3036	*/
3037	stop_pending = pending->stop_pending;
3038	if (!stop_pending)
3039	pending->stop_pending = true;
3040
3041	if (flags & SCA_MIGRATE_ENABLE)
3042	p->migration_flags &= ~MDF_PUSH;
3043
3044	preempt_disable();
3045	task_rq_unlock(rq, p, rf);
3046	if (!stop_pending) {
3047	stop_one_cpu_nowait(cpu: cpu_of(rq), fn: migration_cpu_stop,
3048	arg: &pending->arg, work_buf: &pending->stop_work);
3049	}
3050	preempt_enable();
3051
3052	if (flags & SCA_MIGRATE_ENABLE)
3053	return 0;
3054	} else {
3055
3056	if (!is_migration_disabled(p)) {
3057	if (task_on_rq_queued(p))
3058	rq = move_queued_task(rq, rf, p, new_cpu: dest_cpu);
3059
3060	if (!pending->stop_pending) {
3061	p->migration_pending = NULL;
3062	complete = true;
3063	}
3064	}
3065	task_rq_unlock(rq, p, rf);
3066
3067	if (complete)
3068	complete_all(&pending->done);
3069	}
3070
3071	wait_for_completion(&pending->done);
3072
3073	if (refcount_dec_and_test(r: &pending->refs))
3074	wake_up_var(var: &pending->refs); /* No UaF, just an address */
3075
3076	/*
3077	* Block the original owner of &pending until all subsequent callers
3078	* have seen the completion and decremented the refcount
3079	*/
3080	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3081
3082	/* ARGH */
3083	WARN_ON_ONCE(my_pending.stop_pending);
3084
3085	return 0;
3086	}
3087
3088	/*
3089	* Called with both p->pi_lock and rq->lock held; drops both before returning.
3090	*/
3091	static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3092	struct affinity_context *ctx,
3093	struct rq *rq,
3094	struct rq_flags *rf)
3095	__releases(rq->lock)
3096	__releases(p->pi_lock)
3097	{
3098	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3099	const struct cpumask *cpu_valid_mask = cpu_active_mask;
3100	bool kthread = p->flags & PF_KTHREAD;
3101	unsigned int dest_cpu;
3102	int ret = 0;
3103
3104	update_rq_clock(rq);
3105
3106	if (kthread || is_migration_disabled(p)) {
3107	/*
3108	* Kernel threads are allowed on online && !active CPUs,
3109	* however, during cpu-hot-unplug, even these might get pushed
3110	* away if not KTHREAD_IS_PER_CPU.
3111	*
3112	* Specifically, migration_disabled() tasks must not fail the
3113	* cpumask_any_and_distribute() pick below, esp. so on
3114	* SCA_MIGRATE_ENABLE, otherwise we'll not call
3115	* set_cpus_allowed_common() and actually reset p->cpus_ptr.
3116	*/
3117	cpu_valid_mask = cpu_online_mask;
3118	}
3119
3120	if (!kthread && !cpumask_subset(src1p: ctx->new_mask, src2p: cpu_allowed_mask)) {
3121	ret = -EINVAL;
3122	goto out;
3123	}
3124
3125	/*
3126	* Must re-check here, to close a race against __kthread_bind(),
3127	* sched_setaffinity() is not guaranteed to observe the flag.
3128	*/
3129	if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3130	ret = -EINVAL;
3131	goto out;
3132	}
3133
3134	if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3135	if (cpumask_equal(src1p: &p->cpus_mask, src2p: ctx->new_mask)) {
3136	if (ctx->flags & SCA_USER)
3137	swap(p->user_cpus_ptr, ctx->user_mask);
3138	goto out;
3139	}
3140
3141	if (WARN_ON_ONCE(p == current &&
3142	is_migration_disabled(p) &&
3143	!cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3144	ret = -EBUSY;
3145	goto out;
3146	}
3147	}
3148
3149	/*
3150	* Picking a ~random cpu helps in cases where we are changing affinity
3151	* for groups of tasks (ie. cpuset), so that load balancing is not
3152	* immediately required to distribute the tasks within their new mask.
3153	*/
3154	dest_cpu = cpumask_any_and_distribute(src1p: cpu_valid_mask, src2p: ctx->new_mask);
3155	if (dest_cpu >= nr_cpu_ids) {
3156	ret = -EINVAL;
3157	goto out;
3158	}
3159
3160	__do_set_cpus_allowed(p, ctx);
3161
3162	return affine_move_task(rq, p, rf, dest_cpu, flags: ctx->flags);
3163
3164	out:
3165	task_rq_unlock(rq, p, rf);
3166
3167	return ret;
3168	}
3169
3170	/*
3171	* Change a given task's CPU affinity. Migrate the thread to a
3172	* proper CPU and schedule it away if the CPU it's executing on
3173	* is removed from the allowed bitmask.
3174	*
3175	* NOTE: the caller must have a valid reference to the task, the
3176	* task must not exit() & deallocate itself prematurely. The
3177	* call is not atomic; no spinlocks may be held.
3178	*/
3179	static int __set_cpus_allowed_ptr(struct task_struct *p,
3180	struct affinity_context *ctx)
3181	{
3182	struct rq_flags rf;
3183	struct rq *rq;
3184
3185	rq = task_rq_lock(p, rf: &rf);
3186	/*
3187	* Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3188	* flags are set.
3189	*/
3190	if (p->user_cpus_ptr &&
3191	!(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3192	cpumask_and(dstp: rq->scratch_mask, src1p: ctx->new_mask, src2p: p->user_cpus_ptr))
3193	ctx->new_mask = rq->scratch_mask;
3194
3195	return __set_cpus_allowed_ptr_locked(p, ctx, rq, rf: &rf);
3196	}
3197
3198	int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3199	{
3200	struct affinity_context ac = {
3201	.new_mask = new_mask,
3202	.flags = 0,
3203	};
3204
3205	return __set_cpus_allowed_ptr(p, ctx: &ac);
3206	}
3207	EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3208
3209	/*
3210	* Change a given task's CPU affinity to the intersection of its current
3211	* affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3212	* If user_cpus_ptr is defined, use it as the basis for restricting CPU
3213	* affinity or use cpu_online_mask instead.
3214	*
3215	* If the resulting mask is empty, leave the affinity unchanged and return
3216	* -EINVAL.
3217	*/
3218	static int restrict_cpus_allowed_ptr(struct task_struct *p,
3219	struct cpumask *new_mask,
3220	const struct cpumask *subset_mask)
3221	{
3222	struct affinity_context ac = {
3223	.new_mask = new_mask,
3224	.flags = 0,
3225	};
3226	struct rq_flags rf;
3227	struct rq *rq;
3228	int err;
3229
3230	rq = task_rq_lock(p, rf: &rf);
3231
3232	/*
3233	* Forcefully restricting the affinity of a deadline task is
3234	* likely to cause problems, so fail and noisily override the
3235	* mask entirely.
3236	*/
3237	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3238	err = -EPERM;
3239	goto err_unlock;
3240	}
3241
3242	if (!cpumask_and(dstp: new_mask, src1p: task_user_cpus(p), src2p: subset_mask)) {
3243	err = -EINVAL;
3244	goto err_unlock;
3245	}
3246
3247	return __set_cpus_allowed_ptr_locked(p, ctx: &ac, rq, rf: &rf);
3248
3249	err_unlock:
3250	task_rq_unlock(rq, p, rf: &rf);
3251	return err;
3252	}
3253
3254	/*
3255	* Restrict the CPU affinity of task @p so that it is a subset of
3256	* task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3257	* old affinity mask. If the resulting mask is empty, we warn and walk
3258	* up the cpuset hierarchy until we find a suitable mask.
3259	*/
3260	void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3261	{
3262	cpumask_var_t new_mask;
3263	const struct cpumask *override_mask = task_cpu_possible_mask(p);
3264
3265	alloc_cpumask_var(mask: &new_mask, GFP_KERNEL);
3266
3267	/*
3268	* __migrate_task() can fail silently in the face of concurrent
3269	* offlining of the chosen destination CPU, so take the hotplug
3270	* lock to ensure that the migration succeeds.
3271	*/
3272	cpus_read_lock();
3273	if (!cpumask_available(mask: new_mask))
3274	goto out_set_mask;
3275
3276	if (!restrict_cpus_allowed_ptr(p, new_mask, subset_mask: override_mask))
3277	goto out_free_mask;
3278
3279	/*
3280	* We failed to find a valid subset of the affinity mask for the
3281	* task, so override it based on its cpuset hierarchy.
3282	*/
3283	cpuset_cpus_allowed(p, mask: new_mask);
3284	override_mask = new_mask;
3285
3286	out_set_mask:
3287	if (printk_ratelimit()) {
3288	printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3289	task_pid_nr(p), p->comm,
3290	cpumask_pr_args(override_mask));
3291	}
3292
3293	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3294	out_free_mask:
3295	cpus_read_unlock();
3296	free_cpumask_var(mask: new_mask);
3297	}
3298
3299	static int
3300	__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3301
3302	/*
3303	* Restore the affinity of a task @p which was previously restricted by a
3304	* call to force_compatible_cpus_allowed_ptr().
3305	*
3306	* It is the caller's responsibility to serialise this with any calls to
3307	* force_compatible_cpus_allowed_ptr(@p).
3308	*/
3309	void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3310	{
3311	struct affinity_context ac = {
3312	.new_mask = task_user_cpus(p),
3313	.flags = 0,
3314	};
3315	int ret;
3316
3317	/*
3318	* Try to restore the old affinity mask with __sched_setaffinity().
3319	* Cpuset masking will be done there too.
3320	*/
3321	ret = __sched_setaffinity(p, ctx: &ac);
3322	WARN_ON_ONCE(ret);
3323	}
3324
3325	void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3326	{
3327	#ifdef CONFIG_SCHED_DEBUG
3328	unsigned int state = READ_ONCE(p->__state);
3329
3330	/*
3331	* We should never call set_task_cpu() on a blocked task,
3332	* ttwu() will sort out the placement.
3333	*/
3334	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3335
3336	/*
3337	* Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3338	* because schedstat_wait_{start,end} rebase migrating task's wait_start
3339	* time relying on p->on_rq.
3340	*/
3341	WARN_ON_ONCE(state == TASK_RUNNING &&
3342	p->sched_class == &fair_sched_class &&
3343	(p->on_rq && !task_on_rq_migrating(p)));
3344
3345	#ifdef CONFIG_LOCKDEP
3346	/*
3347	* The caller should hold either p->pi_lock or rq->lock, when changing
3348	* a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3349	*
3350	* sched_move_task() holds both and thus holding either pins the cgroup,
3351	* see task_group().
3352	*
3353	* Furthermore, all task_rq users should acquire both locks, see
3354	* task_rq_lock().
3355	*/
3356	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3357	lockdep_is_held(__rq_lockp(task_rq(p)))));
3358	#endif
3359	/*
3360	* Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3361	*/
3362	WARN_ON_ONCE(!cpu_online(new_cpu));
3363
3364	WARN_ON_ONCE(is_migration_disabled(p));
3365	#endif
3366
3367	trace_sched_migrate_task(p, dest_cpu: new_cpu);
3368
3369	if (task_cpu(p) != new_cpu) {
3370	if (p->sched_class->migrate_task_rq)
3371	p->sched_class->migrate_task_rq(p, new_cpu);
3372	p->se.nr_migrations++;
3373	rseq_migrate(t: p);
3374	sched_mm_cid_migrate_from(t: p);
3375	perf_event_task_migrate(task: p);
3376	}
3377
3378	__set_task_cpu(p, cpu: new_cpu);
3379	}
3380
3381	#ifdef CONFIG_NUMA_BALANCING
3382	static void __migrate_swap_task(struct task_struct *p, int cpu)
3383	{
3384	if (task_on_rq_queued(p)) {
3385	struct rq *src_rq, *dst_rq;
3386	struct rq_flags srf, drf;
3387
3388	src_rq = task_rq(p);
3389	dst_rq = cpu_rq(cpu);
3390
3391	rq_pin_lock(rq: src_rq, rf: &srf);
3392	rq_pin_lock(rq: dst_rq, rf: &drf);
3393
3394	deactivate_task(rq: src_rq, p, flags: 0);
3395	set_task_cpu(p, new_cpu: cpu);
3396	activate_task(rq: dst_rq, p, flags: 0);
3397	wakeup_preempt(rq: dst_rq, p, flags: 0);
3398
3399	rq_unpin_lock(rq: dst_rq, rf: &drf);
3400	rq_unpin_lock(rq: src_rq, rf: &srf);
3401
3402	} else {
3403	/*
3404	* Task isn't running anymore; make it appear like we migrated
3405	* it before it went to sleep. This means on wakeup we make the
3406	* previous CPU our target instead of where it really is.
3407	*/
3408	p->wake_cpu = cpu;
3409	}
3410	}
3411
3412	struct migration_swap_arg {
3413	struct task_struct *src_task, *dst_task;
3414	int src_cpu, dst_cpu;
3415	};
3416
3417	static int migrate_swap_stop(void *data)
3418	{
3419	struct migration_swap_arg *arg = data;
3420	struct rq *src_rq, *dst_rq;
3421
3422	if (!cpu_active(cpu: arg->src_cpu) || !cpu_active(cpu: arg->dst_cpu))
3423	return -EAGAIN;
3424
3425	src_rq = cpu_rq(arg->src_cpu);
3426	dst_rq = cpu_rq(arg->dst_cpu);
3427
3428	guard(double_raw_spinlock)(lock: &arg->src_task->pi_lock, lock2: &arg->dst_task->pi_lock);
3429	guard(double_rq_lock)(lock: src_rq, lock2: dst_rq);
3430
3431	if (task_cpu(p: arg->dst_task) != arg->dst_cpu)
3432	return -EAGAIN;
3433
3434	if (task_cpu(p: arg->src_task) != arg->src_cpu)
3435	return -EAGAIN;
3436
3437	if (!cpumask_test_cpu(cpu: arg->dst_cpu, cpumask: arg->src_task->cpus_ptr))
3438	return -EAGAIN;
3439
3440	if (!cpumask_test_cpu(cpu: arg->src_cpu, cpumask: arg->dst_task->cpus_ptr))
3441	return -EAGAIN;
3442
3443	__migrate_swap_task(p: arg->src_task, cpu: arg->dst_cpu);
3444	__migrate_swap_task(p: arg->dst_task, cpu: arg->src_cpu);
3445
3446	return 0;
3447	}
3448
3449	/*
3450	* Cross migrate two tasks
3451	*/
3452	int migrate_swap(struct task_struct *cur, struct task_struct *p,
3453	int target_cpu, int curr_cpu)
3454	{
3455	struct migration_swap_arg arg;
3456	int ret = -EINVAL;
3457
3458	arg = (struct migration_swap_arg){
3459	.src_task = cur,
3460	.src_cpu = curr_cpu,
3461	.dst_task = p,
3462	.dst_cpu = target_cpu,
3463	};
3464
3465	if (arg.src_cpu == arg.dst_cpu)
3466	goto out;
3467
3468	/*
3469	* These three tests are all lockless; this is OK since all of them
3470	* will be re-checked with proper locks held further down the line.
3471	*/
3472	if (!cpu_active(cpu: arg.src_cpu) || !cpu_active(cpu: arg.dst_cpu))
3473	goto out;
3474
3475	if (!cpumask_test_cpu(cpu: arg.dst_cpu, cpumask: arg.src_task->cpus_ptr))
3476	goto out;
3477
3478	if (!cpumask_test_cpu(cpu: arg.src_cpu, cpumask: arg.dst_task->cpus_ptr))
3479	goto out;
3480
3481	trace_sched_swap_numa(src_tsk: cur, src_cpu: arg.src_cpu, dst_tsk: p, dst_cpu: arg.dst_cpu);
3482	ret = stop_two_cpus(cpu1: arg.dst_cpu, cpu2: arg.src_cpu, fn: migrate_swap_stop, arg: &arg);
3483
3484	out:
3485	return ret;
3486	}
3487	#endif /* CONFIG_NUMA_BALANCING */
3488
3489	/***
3490	* kick_process - kick a running thread to enter/exit the kernel
3491	* @p: the to-be-kicked thread
3492	*
3493	* Cause a process which is running on another CPU to enter
3494	* kernel-mode, without any delay. (to get signals handled.)
3495	*
3496	* NOTE: this function doesn't have to take the runqueue lock,
3497	* because all it wants to ensure is that the remote task enters
3498	* the kernel. If the IPI races and the task has been migrated
3499	* to another CPU then no harm is done and the purpose has been
3500	* achieved as well.
3501	*/
3502	void kick_process(struct task_struct *p)
3503	{
3504	guard(preempt)();
3505	int cpu = task_cpu(p);
3506
3507	if ((cpu != smp_processor_id()) && task_curr(p))
3508	smp_send_reschedule(cpu);
3509	}
3510	EXPORT_SYMBOL_GPL(kick_process);
3511
3512	/*
3513	* ->cpus_ptr is protected by both rq->lock and p->pi_lock
3514	*
3515	* A few notes on cpu_active vs cpu_online:
3516	*
3517	* - cpu_active must be a subset of cpu_online
3518	*
3519	* - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3520	* see __set_cpus_allowed_ptr(). At this point the newly online
3521	* CPU isn't yet part of the sched domains, and balancing will not
3522	* see it.
3523	*
3524	* - on CPU-down we clear cpu_active() to mask the sched domains and
3525	* avoid the load balancer to place new tasks on the to be removed
3526	* CPU. Existing tasks will remain running there and will be taken
3527	* off.
3528	*
3529	* This means that fallback selection must not select !active CPUs.
3530	* And can assume that any active CPU must be online. Conversely
3531	* select_task_rq() below may allow selection of !active CPUs in order
3532	* to satisfy the above rules.
3533	*/
3534	static int select_fallback_rq(int cpu, struct task_struct *p)
3535	{
3536	int nid = cpu_to_node(cpu);
3537	const struct cpumask *nodemask = NULL;
3538	enum { cpuset, possible, fail } state = cpuset;
3539	int dest_cpu;
3540
3541	/*
3542	* If the node that the CPU is on has been offlined, cpu_to_node()
3543	* will return -1. There is no CPU on the node, and we should
3544	* select the CPU on the other node.
3545	*/
3546	if (nid != -1) {
3547	nodemask = cpumask_of_node(node: nid);
3548
3549	/* Look for allowed, online CPU in same node. */
3550	for_each_cpu(dest_cpu, nodemask) {
3551	if (is_cpu_allowed(p, cpu: dest_cpu))
3552	return dest_cpu;
3553	}
3554	}
3555
3556	for (;;) {
3557	/* Any allowed, online CPU? */
3558	for_each_cpu(dest_cpu, p->cpus_ptr) {
3559	if (!is_cpu_allowed(p, cpu: dest_cpu))
3560	continue;
3561
3562	goto out;
3563	}
3564
3565	/* No more Mr. Nice Guy. */
3566	switch (state) {
3567	case cpuset:
3568	if (cpuset_cpus_allowed_fallback(p)) {
3569	state = possible;
3570	break;
3571	}
3572	fallthrough;
3573	case possible:
3574	/*
3575	* XXX When called from select_task_rq() we only
3576	* hold p->pi_lock and again violate locking order.
3577	*
3578	* More yuck to audit.
3579	*/
3580	do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3581	state = fail;
3582	break;
3583	case fail:
3584	BUG();
3585	break;
3586	}
3587	}
3588
3589	out:
3590	if (state != cpuset) {
3591	/*
3592	* Don't tell them about moving exiting tasks or
3593	* kernel threads (both mm NULL), since they never
3594	* leave kernel.
3595	*/
3596	if (p->mm && printk_ratelimit()) {
3597	printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3598	task_pid_nr(p), p->comm, cpu);
3599	}
3600	}
3601
3602	return dest_cpu;
3603	}
3604
3605	/*
3606	* The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3607	*/
3608	static inline
3609	int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3610	{
3611	lockdep_assert_held(&p->pi_lock);
3612
3613	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3614	cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3615	else
3616	cpu = cpumask_any(p->cpus_ptr);
3617
3618	/*
3619	* In order not to call set_task_cpu() on a blocking task we need
3620	* to rely on ttwu() to place the task on a valid ->cpus_ptr
3621	* CPU.
3622	*
3623	* Since this is common to all placement strategies, this lives here.
3624	*
3625	* [ this allows ->select_task() to simply return task_cpu(p) and
3626	* not worry about this generic constraint ]
3627	*/
3628	if (unlikely(!is_cpu_allowed(p, cpu)))
3629	cpu = select_fallback_rq(cpu: task_cpu(p), p);
3630
3631	return cpu;
3632	}
3633
3634	void sched_set_stop_task(int cpu, struct task_struct *stop)
3635	{
3636	static struct lock_class_key stop_pi_lock;
3637	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3638	struct task_struct *old_stop = cpu_rq(cpu)->stop;
3639
3640	if (stop) {
3641	/*
3642	* Make it appear like a SCHED_FIFO task, its something
3643	* userspace knows about and won't get confused about.
3644	*
3645	* Also, it will make PI more or less work without too
3646	* much confusion -- but then, stop work should not
3647	* rely on PI working anyway.
3648	*/
3649	sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3650
3651	stop->sched_class = &stop_sched_class;
3652
3653	/*
3654	* The PI code calls rt_mutex_setprio() with ->pi_lock held to
3655	* adjust the effective priority of a task. As a result,
3656	* rt_mutex_setprio() can trigger (RT) balancing operations,
3657	* which can then trigger wakeups of the stop thread to push
3658	* around the current task.
3659	*
3660	* The stop task itself will never be part of the PI-chain, it
3661	* never blocks, therefore that ->pi_lock recursion is safe.
3662	* Tell lockdep about this by placing the stop->pi_lock in its
3663	* own class.
3664	*/
3665	lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3666	}
3667
3668	cpu_rq(cpu)->stop = stop;
3669
3670	if (old_stop) {
3671	/*
3672	* Reset it back to a normal scheduling class so that
3673	* it can die in pieces.
3674	*/
3675	old_stop->sched_class = &rt_sched_class;
3676	}
3677	}
3678
3679	#else /* CONFIG_SMP */
3680
3681	static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3682	struct affinity_context *ctx)
3683	{
3684	return set_cpus_allowed_ptr(p, ctx->new_mask);
3685	}
3686
3687	static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3688
3689	static inline bool rq_has_pinned_tasks(struct rq *rq)
3690	{
3691	return false;
3692	}
3693
3694	static inline cpumask_t *alloc_user_cpus_ptr(int node)
3695	{
3696	return NULL;
3697	}
3698
3699	#endif /* !CONFIG_SMP */
3700
3701	static void
3702	ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3703	{
3704	struct rq *rq;
3705
3706	if (!schedstat_enabled())
3707	return;
3708
3709	rq = this_rq();
3710
3711	#ifdef CONFIG_SMP
3712	if (cpu == rq->cpu) {
3713	__schedstat_inc(rq->ttwu_local);
3714	__schedstat_inc(p->stats.nr_wakeups_local);
3715	} else {
3716	struct sched_domain *sd;
3717
3718	__schedstat_inc(p->stats.nr_wakeups_remote);
3719
3720	guard(rcu)();
3721	for_each_domain(rq->cpu, sd) {
3722	if (cpumask_test_cpu(cpu, cpumask: sched_domain_span(sd))) {
3723	__schedstat_inc(sd->ttwu_wake_remote);
3724	break;
3725	}
3726	}
3727	}
3728
3729	if (wake_flags & WF_MIGRATED)
3730	__schedstat_inc(p->stats.nr_wakeups_migrate);
3731	#endif /* CONFIG_SMP */
3732
3733	__schedstat_inc(rq->ttwu_count);
3734	__schedstat_inc(p->stats.nr_wakeups);
3735
3736	if (wake_flags & WF_SYNC)
3737	__schedstat_inc(p->stats.nr_wakeups_sync);
3738	}
3739
3740	/*
3741	* Mark the task runnable.
3742	*/
3743	static inline void ttwu_do_wakeup(struct task_struct *p)
3744	{
3745	WRITE_ONCE(p->__state, TASK_RUNNING);
3746	trace_sched_wakeup(p);
3747	}
3748
3749	static void
3750	ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3751	struct rq_flags *rf)
3752	{
3753	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3754
3755	lockdep_assert_rq_held(rq);
3756
3757	if (p->sched_contributes_to_load)
3758	rq->nr_uninterruptible--;
3759
3760	#ifdef CONFIG_SMP
3761	if (wake_flags & WF_MIGRATED)
3762	en_flags |= ENQUEUE_MIGRATED;
3763	else
3764	#endif
3765	if (p->in_iowait) {
3766	delayacct_blkio_end(p);
3767	atomic_dec(v: &task_rq(p)->nr_iowait);
3768	}
3769
3770	activate_task(rq, p, flags: en_flags);
3771	wakeup_preempt(rq, p, flags: wake_flags);
3772
3773	ttwu_do_wakeup(p);
3774
3775	#ifdef CONFIG_SMP
3776	if (p->sched_class->task_woken) {
3777	/*
3778	* Our task @p is fully woken up and running; so it's safe to
3779	* drop the rq->lock, hereafter rq is only used for statistics.
3780	*/
3781	rq_unpin_lock(rq, rf);
3782	p->sched_class->task_woken(rq, p);
3783	rq_repin_lock(rq, rf);
3784	}
3785
3786	if (rq->idle_stamp) {
3787	u64 delta = rq_clock(rq) - rq->idle_stamp;
3788	u64 max = 2*rq->max_idle_balance_cost;
3789
3790	update_avg(avg: &rq->avg_idle, sample: delta);
3791
3792	if (rq->avg_idle > max)
3793	rq->avg_idle = max;
3794
3795	rq->idle_stamp = 0;
3796	}
3797	#endif
3798	}
3799
3800	/*
3801	* Consider @p being inside a wait loop:
3802	*
3803	* for (;;) {
3804	* set_current_state(TASK_UNINTERRUPTIBLE);
3805	*
3806	* if (CONDITION)
3807	* break;
3808	*
3809	* schedule();
3810	* }
3811	* __set_current_state(TASK_RUNNING);
3812	*
3813	* between set_current_state() and schedule(). In this case @p is still
3814	* runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3815	* an atomic manner.
3816	*
3817	* By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3818	* then schedule() must still happen and p->state can be changed to
3819	* TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3820	* need to do a full wakeup with enqueue.
3821	*
3822	* Returns: %true when the wakeup is done,
3823	* %false otherwise.
3824	*/
3825	static int ttwu_runnable(struct task_struct *p, int wake_flags)
3826	{
3827	struct rq_flags rf;
3828	struct rq *rq;
3829	int ret = 0;
3830
3831	rq = __task_rq_lock(p, rf: &rf);
3832	if (task_on_rq_queued(p)) {
3833	if (!task_on_cpu(rq, p)) {
3834	/*
3835	* When on_rq && !on_cpu the task is preempted, see if
3836	* it should preempt the task that is current now.
3837	*/
3838	update_rq_clock(rq);
3839	wakeup_preempt(rq, p, flags: wake_flags);
3840	}
3841	ttwu_do_wakeup(p);
3842	ret = 1;
3843	}
3844	__task_rq_unlock(rq, rf: &rf);
3845
3846	return ret;
3847	}
3848
3849	#ifdef CONFIG_SMP
3850	void sched_ttwu_pending(void *arg)
3851	{
3852	struct llist_node *llist = arg;
3853	struct rq *rq = this_rq();
3854	struct task_struct *p, *t;
3855	struct rq_flags rf;
3856
3857	if (!llist)
3858	return;
3859
3860	rq_lock_irqsave(rq, rf: &rf);
3861	update_rq_clock(rq);
3862
3863	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3864	if (WARN_ON_ONCE(p->on_cpu))
3865	smp_cond_load_acquire(&p->on_cpu, !VAL);
3866
3867	if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3868	set_task_cpu(p, new_cpu: cpu_of(rq));
3869
3870	ttwu_do_activate(rq, p, wake_flags: p->sched_remote_wakeup ? WF_MIGRATED : 0, rf: &rf);
3871	}
3872
3873	/*
3874	* Must be after enqueueing at least once task such that
3875	* idle_cpu() does not observe a false-negative -- if it does,
3876	* it is possible for select_idle_siblings() to stack a number
3877	* of tasks on this CPU during that window.
3878	*
3879	* It is ok to clear ttwu_pending when another task pending.
3880	* We will receive IPI after local irq enabled and then enqueue it.
3881	* Since now nr_running > 0, idle_cpu() will always get correct result.
3882	*/
3883	WRITE_ONCE(rq->ttwu_pending, 0);
3884	rq_unlock_irqrestore(rq, rf: &rf);
3885	}
3886
3887	/*
3888	* Prepare the scene for sending an IPI for a remote smp_call
3889	*
3890	* Returns true if the caller can proceed with sending the IPI.
3891	* Returns false otherwise.
3892	*/
3893	bool call_function_single_prep_ipi(int cpu)
3894	{
3895	if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3896	trace_sched_wake_idle_without_ipi(cpu);
3897	return false;
3898	}
3899
3900	return true;
3901	}
3902
3903	/*
3904	* Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3905	* necessary. The wakee CPU on receipt of the IPI will queue the task
3906	* via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3907	* of the wakeup instead of the waker.
3908	*/
3909	static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3910	{
3911	struct rq *rq = cpu_rq(cpu);
3912
3913	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3914
3915	WRITE_ONCE(rq->ttwu_pending, 1);
3916	__smp_call_single_queue(cpu, node: &p->wake_entry.llist);
3917	}
3918
3919	void wake_up_if_idle(int cpu)
3920	{
3921	struct rq *rq = cpu_rq(cpu);
3922
3923	guard(rcu)();
3924	if (is_idle_task(rcu_dereference(rq->curr))) {
3925	guard(rq_lock_irqsave)(l: rq);
3926	if (is_idle_task(p: rq->curr))
3927	resched_curr(rq);
3928	}
3929	}
3930
3931	bool cpus_share_cache(int this_cpu, int that_cpu)
3932	{
3933	if (this_cpu == that_cpu)
3934	return true;
3935
3936	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3937	}
3938
3939	/*
3940	* Whether CPUs are share cache resources, which means LLC on non-cluster
3941	* machines and LLC tag or L2 on machines with clusters.
3942	*/
3943	bool cpus_share_resources(int this_cpu, int that_cpu)
3944	{
3945	if (this_cpu == that_cpu)
3946	return true;
3947
3948	return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
3949	}
3950
3951	static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3952	{
3953	/*
3954	* Do not complicate things with the async wake_list while the CPU is
3955	* in hotplug state.
3956	*/
3957	if (!cpu_active(cpu))
3958	return false;
3959
3960	/* Ensure the task will still be allowed to run on the CPU. */
3961	if (!cpumask_test_cpu(cpu, cpumask: p->cpus_ptr))
3962	return false;
3963
3964	/*
3965	* If the CPU does not share cache, then queue the task on the
3966	* remote rqs wakelist to avoid accessing remote data.
3967	*/
3968	if (!cpus_share_cache(smp_processor_id(), that_cpu: cpu))
3969	return true;
3970
3971	if (cpu == smp_processor_id())
3972	return false;
3973
3974	/*
3975	* If the wakee cpu is idle, or the task is descheduling and the
3976	* only running task on the CPU, then use the wakelist to offload
3977	* the task activation to the idle (or soon-to-be-idle) CPU as
3978	* the current CPU is likely busy. nr_running is checked to
3979	* avoid unnecessary task stacking.
3980	*
3981	* Note that we can only get here with (wakee) p->on_rq=0,
3982	* p->on_cpu can be whatever, we've done the dequeue, so
3983	* the wakee has been accounted out of ->nr_running.
3984	*/
3985	if (!cpu_rq(cpu)->nr_running)
3986	return true;
3987
3988	return false;
3989	}
3990
3991	static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3992	{
3993	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3994	sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3995	__ttwu_queue_wakelist(p, cpu, wake_flags);
3996	return true;
3997	}
3998
3999	return false;
4000	}
4001
4002	#else /* !CONFIG_SMP */
4003
4004	static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4005	{
4006	return false;
4007	}
4008
4009	#endif /* CONFIG_SMP */
4010
4011	static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4012	{
4013	struct rq *rq = cpu_rq(cpu);
4014	struct rq_flags rf;
4015
4016	if (ttwu_queue_wakelist(p, cpu, wake_flags))
4017	return;
4018
4019	rq_lock(rq, rf: &rf);
4020	update_rq_clock(rq);
4021	ttwu_do_activate(rq, p, wake_flags, rf: &rf);
4022	rq_unlock(rq, rf: &rf);
4023	}
4024
4025	/*
4026	* Invoked from try_to_wake_up() to check whether the task can be woken up.
4027	*
4028	* The caller holds p::pi_lock if p != current or has preemption
4029	* disabled when p == current.
4030	*
4031	* The rules of saved_state:
4032	*
4033	* The related locking code always holds p::pi_lock when updating
4034	* p::saved_state, which means the code is fully serialized in both cases.
4035	*
4036	* For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
4037	* No other bits set. This allows to distinguish all wakeup scenarios.
4038	*
4039	* For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
4040	* allows us to prevent early wakeup of tasks before they can be run on
4041	* asymmetric ISA architectures (eg ARMv9).
4042	*/
4043	static __always_inline
4044	bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4045	{
4046	int match;
4047
4048	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4049	WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4050	state != TASK_RTLOCK_WAIT);
4051	}
4052
4053	*success = !!(match = __task_state_match(p, state));
4054
4055	/*
4056	* Saved state preserves the task state across blocking on
4057	* an RT lock or TASK_FREEZABLE tasks. If the state matches,
4058	* set p::saved_state to TASK_RUNNING, but do not wake the task
4059	* because it waits for a lock wakeup or __thaw_task(). Also
4060	* indicate success because from the regular waker's point of
4061	* view this has succeeded.
4062	*
4063	* After acquiring the lock the task will restore p::__state
4064	* from p::saved_state which ensures that the regular
4065	* wakeup is not lost. The restore will also set
4066	* p::saved_state to TASK_RUNNING so any further tests will
4067	* not result in false positives vs. @success
4068	*/
4069	if (match < 0)
4070	p->saved_state = TASK_RUNNING;
4071
4072	return match > 0;
4073	}
4074
4075	/*
4076	* Notes on Program-Order guarantees on SMP systems.
4077	*
4078	* MIGRATION
4079	*
4080	* The basic program-order guarantee on SMP systems is that when a task [t]
4081	* migrates, all its activity on its old CPU [c0] happens-before any subsequent
4082	* execution on its new CPU [c1].
4083	*
4084	* For migration (of runnable tasks) this is provided by the following means:
4085	*
4086	* A) UNLOCK of the rq(c0)->lock scheduling out task t
4087	* B) migration for t is required to synchronize *both* rq(c0)->lock and
4088	* rq(c1)->lock (if not at the same time, then in that order).
4089	* C) LOCK of the rq(c1)->lock scheduling in task
4090	*
4091	* Release/acquire chaining guarantees that B happens after A and C after B.
4092	* Note: the CPU doing B need not be c0 or c1
4093	*
4094	* Example:
4095	*
4096	* CPU0 CPU1 CPU2
4097	*
4098	* LOCK rq(0)->lock
4099	* sched-out X
4100	* sched-in Y
4101	* UNLOCK rq(0)->lock
4102	*
4103	* LOCK rq(0)->lock // orders against CPU0
4104	* dequeue X
4105	* UNLOCK rq(0)->lock
4106	*
4107	* LOCK rq(1)->lock
4108	* enqueue X
4109	* UNLOCK rq(1)->lock
4110	*
4111	* LOCK rq(1)->lock // orders against CPU2
4112	* sched-out Z
4113	* sched-in X
4114	* UNLOCK rq(1)->lock
4115	*
4116	*
4117	* BLOCKING -- aka. SLEEP + WAKEUP
4118	*
4119	* For blocking we (obviously) need to provide the same guarantee as for
4120	* migration. However the means are completely different as there is no lock
4121	* chain to provide order. Instead we do:
4122	*
4123	* 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4124	* 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4125	*
4126	* Example:
4127	*
4128	* CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4129	*
4130	* LOCK rq(0)->lock LOCK X->pi_lock
4131	* dequeue X
4132	* sched-out X
4133	* smp_store_release(X->on_cpu, 0);
4134	*
4135	* smp_cond_load_acquire(&X->on_cpu, !VAL);
4136	* X->state = WAKING
4137	* set_task_cpu(X,2)
4138	*
4139	* LOCK rq(2)->lock
4140	* enqueue X
4141	* X->state = RUNNING
4142	* UNLOCK rq(2)->lock
4143	*
4144	* LOCK rq(2)->lock // orders against CPU1
4145	* sched-out Z
4146	* sched-in X
4147	* UNLOCK rq(2)->lock
4148	*
4149	* UNLOCK X->pi_lock
4150	* UNLOCK rq(0)->lock
4151	*
4152	*
4153	* However, for wakeups there is a second guarantee we must provide, namely we
4154	* must ensure that CONDITION=1 done by the caller can not be reordered with
4155	* accesses to the task state; see try_to_wake_up() and set_current_state().
4156	*/
4157
4158	/**
4159	* try_to_wake_up - wake up a thread
4160	* @p: the thread to be awakened
4161	* @state: the mask of task states that can be woken
4162	* @wake_flags: wake modifier flags (WF_*)
4163	*
4164	* Conceptually does:
4165	*
4166	* If (@state & @p->state) @p->state = TASK_RUNNING.
4167	*
4168	* If the task was not queued/runnable, also place it back on a runqueue.
4169	*
4170	* This function is atomic against schedule() which would dequeue the task.
4171	*
4172	* It issues a full memory barrier before accessing @p->state, see the comment
4173	* with set_current_state().
4174	*
4175	* Uses p->pi_lock to serialize against concurrent wake-ups.
4176	*
4177	* Relies on p->pi_lock stabilizing:
4178	* - p->sched_class
4179	* - p->cpus_ptr
4180	* - p->sched_task_group
4181	* in order to do migration, see its use of select_task_rq()/set_task_cpu().
4182	*
4183	* Tries really hard to only take one task_rq(p)->lock for performance.
4184	* Takes rq->lock in:
4185	* - ttwu_runnable() -- old rq, unavoidable, see comment there;
4186	* - ttwu_queue() -- new rq, for enqueue of the task;
4187	* - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4188	*
4189	* As a consequence we race really badly with just about everything. See the
4190	* many memory barriers and their comments for details.
4191	*
4192	* Return: %true if @p->state changes (an actual wakeup was done),
4193	* %false otherwise.
4194	*/
4195	int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4196	{
4197	guard(preempt)();
4198	int cpu, success = 0;
4199
4200	if (p == current) {
4201	/*
4202	* We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4203	* == smp_processor_id()'. Together this means we can special
4204	* case the whole 'p->on_rq && ttwu_runnable()' case below
4205	* without taking any locks.
4206	*
4207	* In particular:
4208	* - we rely on Program-Order guarantees for all the ordering,
4209	* - we're serialized against set_special_state() by virtue of
4210	* it disabling IRQs (this allows not taking ->pi_lock).
4211	*/
4212	if (!ttwu_state_match(p, state, success: &success))
4213	goto out;
4214
4215	trace_sched_waking(p);
4216	ttwu_do_wakeup(p);
4217	goto out;
4218	}
4219
4220	/*
4221	* If we are going to wake up a thread waiting for CONDITION we
4222	* need to ensure that CONDITION=1 done by the caller can not be
4223	* reordered with p->state check below. This pairs with smp_store_mb()
4224	* in set_current_state() that the waiting thread does.
4225	*/
4226	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4227	smp_mb__after_spinlock();
4228	if (!ttwu_state_match(p, state, success: &success))
4229	break;
4230
4231	trace_sched_waking(p);
4232
4233	/*
4234	* Ensure we load p->on_rq _after_ p->state, otherwise it would
4235	* be possible to, falsely, observe p->on_rq == 0 and get stuck
4236	* in smp_cond_load_acquire() below.
4237	*
4238	* sched_ttwu_pending() try_to_wake_up()
4239	* STORE p->on_rq = 1 LOAD p->state
4240	* UNLOCK rq->lock
4241	*
4242	* __schedule() (switch to task 'p')
4243	* LOCK rq->lock smp_rmb();
4244	* smp_mb__after_spinlock();
4245	* UNLOCK rq->lock
4246	*
4247	* [task p]
4248	* STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4249	*
4250	* Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4251	* __schedule(). See the comment for smp_mb__after_spinlock().
4252	*
4253	* A similar smp_rmb() lives in __task_needs_rq_lock().
4254	*/
4255	smp_rmb();
4256	if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4257	break;
4258
4259	#ifdef CONFIG_SMP
4260	/*
4261	* Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4262	* possible to, falsely, observe p->on_cpu == 0.
4263	*
4264	* One must be running (->on_cpu == 1) in order to remove oneself
4265	* from the runqueue.
4266	*
4267	* __schedule() (switch to task 'p') try_to_wake_up()
4268	* STORE p->on_cpu = 1 LOAD p->on_rq
4269	* UNLOCK rq->lock
4270	*
4271	* __schedule() (put 'p' to sleep)
4272	* LOCK rq->lock smp_rmb();
4273	* smp_mb__after_spinlock();
4274	* STORE p->on_rq = 0 LOAD p->on_cpu
4275	*
4276	* Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4277	* __schedule(). See the comment for smp_mb__after_spinlock().
4278	*
4279	* Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4280	* schedule()'s deactivate_task() has 'happened' and p will no longer
4281	* care about it's own p->state. See the comment in __schedule().
4282	*/
4283	smp_acquire__after_ctrl_dep();
4284
4285	/*
4286	* We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4287	* == 0), which means we need to do an enqueue, change p->state to
4288	* TASK_WAKING such that we can unlock p->pi_lock before doing the
4289	* enqueue, such as ttwu_queue_wakelist().
4290	*/
4291	WRITE_ONCE(p->__state, TASK_WAKING);
4292
4293	/*
4294	* If the owning (remote) CPU is still in the middle of schedule() with
4295	* this task as prev, considering queueing p on the remote CPUs wake_list
4296	* which potentially sends an IPI instead of spinning on p->on_cpu to
4297	* let the waker make forward progress. This is safe because IRQs are
4298	* disabled and the IPI will deliver after on_cpu is cleared.
4299	*
4300	* Ensure we load task_cpu(p) after p->on_cpu:
4301	*
4302	* set_task_cpu(p, cpu);
4303	* STORE p->cpu = @cpu
4304	* __schedule() (switch to task 'p')
4305	* LOCK rq->lock
4306	* smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4307	* STORE p->on_cpu = 1 LOAD p->cpu
4308	*
4309	* to ensure we observe the correct CPU on which the task is currently
4310	* scheduling.
4311	*/
4312	if (smp_load_acquire(&p->on_cpu) &&
4313	ttwu_queue_wakelist(p, cpu: task_cpu(p), wake_flags))
4314	break;
4315
4316	/*
4317	* If the owning (remote) CPU is still in the middle of schedule() with
4318	* this task as prev, wait until it's done referencing the task.
4319	*
4320	* Pairs with the smp_store_release() in finish_task().
4321	*
4322	* This ensures that tasks getting woken will be fully ordered against
4323	* their previous state and preserve Program Order.
4324	*/
4325	smp_cond_load_acquire(&p->on_cpu, !VAL);
4326
4327	cpu = select_task_rq(p, cpu: p->wake_cpu, wake_flags: wake_flags | WF_TTWU);
4328	if (task_cpu(p) != cpu) {
4329	if (p->in_iowait) {
4330	delayacct_blkio_end(p);
4331	atomic_dec(v: &task_rq(p)->nr_iowait);
4332	}
4333
4334	wake_flags |= WF_MIGRATED;
4335	psi_ttwu_dequeue(p);
4336	set_task_cpu(p, new_cpu: cpu);
4337	}
4338	#else
4339	cpu = task_cpu(p);
4340	#endif /* CONFIG_SMP */
4341
4342	ttwu_queue(p, cpu, wake_flags);
4343	}
4344	out:
4345	if (success)
4346	ttwu_stat(p, cpu: task_cpu(p), wake_flags);
4347
4348	return success;
4349	}
4350
4351	static bool __task_needs_rq_lock(struct task_struct *p)
4352	{
4353	unsigned int state = READ_ONCE(p->__state);
4354
4355	/*
4356	* Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4357	* the task is blocked. Make sure to check @state since ttwu() can drop
4358	* locks at the end, see ttwu_queue_wakelist().
4359	*/
4360	if (state == TASK_RUNNING || state == TASK_WAKING)
4361	return true;
4362
4363	/*
4364	* Ensure we load p->on_rq after p->__state, otherwise it would be
4365	* possible to, falsely, observe p->on_rq == 0.
4366	*
4367	* See try_to_wake_up() for a longer comment.
4368	*/
4369	smp_rmb();
4370	if (p->on_rq)
4371	return true;
4372
4373	#ifdef CONFIG_SMP
4374	/*
4375	* Ensure the task has finished __schedule() and will not be referenced
4376	* anymore. Again, see try_to_wake_up() for a longer comment.
4377	*/
4378	smp_rmb();
4379	smp_cond_load_acquire(&p->on_cpu, !VAL);
4380	#endif
4381
4382	return false;
4383	}
4384
4385	/**
4386	* task_call_func - Invoke a function on task in fixed state
4387	* @p: Process for which the function is to be invoked, can be @current.
4388	* @func: Function to invoke.
4389	* @arg: Argument to function.
4390	*
4391	* Fix the task in it's current state by avoiding wakeups and or rq operations
4392	* and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4393	* to work out what the state is, if required. Given that @func can be invoked
4394	* with a runqueue lock held, it had better be quite lightweight.
4395	*
4396	* Returns:
4397	* Whatever @func returns
4398	*/
4399	int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4400	{
4401	struct rq *rq = NULL;
4402	struct rq_flags rf;
4403	int ret;
4404
4405	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4406
4407	if (__task_needs_rq_lock(p))
4408	rq = __task_rq_lock(p, rf: &rf);
4409
4410	/*
4411	* At this point the task is pinned; either:
4412	* - blocked and we're holding off wakeups (pi->lock)
4413	* - woken, and we're holding off enqueue (rq->lock)
4414	* - queued, and we're holding off schedule (rq->lock)
4415	* - running, and we're holding off de-schedule (rq->lock)
4416	*
4417	* The called function (@func) can use: task_curr(), p->on_rq and
4418	* p->__state to differentiate between these states.
4419	*/
4420	ret = func(p, arg);
4421
4422	if (rq)
4423	rq_unlock(rq, rf: &rf);
4424
4425	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4426	return ret;
4427	}
4428
4429	/**
4430	* cpu_curr_snapshot - Return a snapshot of the currently running task
4431	* @cpu: The CPU on which to snapshot the task.
4432	*
4433	* Returns the task_struct pointer of the task "currently" running on
4434	* the specified CPU. If the same task is running on that CPU throughout,
4435	* the return value will be a pointer to that task's task_struct structure.
4436	* If the CPU did any context switches even vaguely concurrently with the
4437	* execution of this function, the return value will be a pointer to the
4438	* task_struct structure of a randomly chosen task that was running on
4439	* that CPU somewhere around the time that this function was executing.
4440	*
4441	* If the specified CPU was offline, the return value is whatever it
4442	* is, perhaps a pointer to the task_struct structure of that CPU's idle
4443	* task, but there is no guarantee. Callers wishing a useful return
4444	* value must take some action to ensure that the specified CPU remains
4445	* online throughout.
4446	*
4447	* This function executes full memory barriers before and after fetching
4448	* the pointer, which permits the caller to confine this function's fetch
4449	* with respect to the caller's accesses to other shared variables.
4450	*/
4451	struct task_struct *cpu_curr_snapshot(int cpu)
4452	{
4453	struct task_struct *t;
4454
4455	smp_mb(); /* Pairing determined by caller's synchronization design. */
4456	t = rcu_dereference(cpu_curr(cpu));
4457	smp_mb(); /* Pairing determined by caller's synchronization design. */
4458	return t;
4459	}
4460
4461	/**
4462	* wake_up_process - Wake up a specific process
4463	* @p: The process to be woken up.
4464	*
4465	* Attempt to wake up the nominated process and move it to the set of runnable
4466	* processes.
4467	*
4468	* Return: 1 if the process was woken up, 0 if it was already running.
4469	*
4470	* This function executes a full memory barrier before accessing the task state.
4471	*/
4472	int wake_up_process(struct task_struct *p)
4473	{
4474	return try_to_wake_up(p, TASK_NORMAL, wake_flags: 0);
4475	}
4476	EXPORT_SYMBOL(wake_up_process);
4477
4478	int wake_up_state(struct task_struct *p, unsigned int state)
4479	{
4480	return try_to_wake_up(p, state, wake_flags: 0);
4481	}
4482
4483	/*
4484	* Perform scheduler related setup for a newly forked process p.
4485	* p is forked by current.
4486	*
4487	* __sched_fork() is basic setup used by init_idle() too:
4488	*/
4489	static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4490	{
4491	p->on_rq = 0;
4492
4493	p->se.on_rq = 0;
4494	p->se.exec_start = 0;
4495	p->se.sum_exec_runtime = 0;
4496	p->se.prev_sum_exec_runtime = 0;
4497	p->se.nr_migrations = 0;
4498	p->se.vruntime = 0;
4499	p->se.vlag = 0;
4500	p->se.slice = sysctl_sched_base_slice;
4501	INIT_LIST_HEAD(list: &p->se.group_node);
4502
4503	#ifdef CONFIG_FAIR_GROUP_SCHED
4504	p->se.cfs_rq = NULL;
4505	#endif
4506
4507	#ifdef CONFIG_SCHEDSTATS
4508	/* Even if schedstat is disabled, there should not be garbage */
4509	memset(&p->stats, 0, sizeof(p->stats));
4510	#endif
4511
4512	RB_CLEAR_NODE(&p->dl.rb_node);
4513	init_dl_task_timer(dl_se: &p->dl);
4514	init_dl_inactive_task_timer(dl_se: &p->dl);
4515	__dl_clear_params(p);
4516
4517	INIT_LIST_HEAD(list: &p->rt.run_list);
4518	p->rt.timeout = 0;
4519	p->rt.time_slice = sched_rr_timeslice;
4520	p->rt.on_rq = 0;
4521	p->rt.on_list = 0;
4522
4523	#ifdef CONFIG_PREEMPT_NOTIFIERS
4524	INIT_HLIST_HEAD(&p->preempt_notifiers);
4525	#endif
4526
4527	#ifdef CONFIG_COMPACTION
4528	p->capture_control = NULL;
4529	#endif
4530	init_numa_balancing(clone_flags, p);
4531	#ifdef CONFIG_SMP
4532	p->wake_entry.u_flags = CSD_TYPE_TTWU;
4533	p->migration_pending = NULL;
4534	#endif
4535	init_sched_mm_cid(t: p);
4536	}
4537
4538	DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4539
4540	#ifdef CONFIG_NUMA_BALANCING
4541
4542	int sysctl_numa_balancing_mode;
4543
4544	static void __set_numabalancing_state(bool enabled)
4545	{
4546	if (enabled)
4547	static_branch_enable(&sched_numa_balancing);
4548	else
4549	static_branch_disable(&sched_numa_balancing);
4550	}
4551
4552	void set_numabalancing_state(bool enabled)
4553	{
4554	if (enabled)
4555	sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4556	else
4557	sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4558	__set_numabalancing_state(enabled);
4559	}
4560
4561	#ifdef CONFIG_PROC_SYSCTL
4562	static void reset_memory_tiering(void)
4563	{
4564	struct pglist_data *pgdat;
4565
4566	for_each_online_pgdat(pgdat) {
4567	pgdat->nbp_threshold = 0;
4568	pgdat->nbp_th_nr_cand = node_page_state(pgdat, item: PGPROMOTE_CANDIDATE);
4569	pgdat->nbp_th_start = jiffies_to_msecs(j: jiffies);
4570	}
4571	}
4572
4573	static int sysctl_numa_balancing(struct ctl_table *table, int write,
4574	void *buffer, size_t *lenp, loff_t *ppos)
4575	{
4576	struct ctl_table t;
4577	int err;
4578	int state = sysctl_numa_balancing_mode;
4579
4580	if (write && !capable(CAP_SYS_ADMIN))
4581	return -EPERM;
4582
4583	t = *table;
4584	t.data = &state;
4585	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4586	if (err < 0)
4587	return err;
4588	if (write) {
4589	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4590	(state & NUMA_BALANCING_MEMORY_TIERING))
4591	reset_memory_tiering();
4592	sysctl_numa_balancing_mode = state;
4593	__set_numabalancing_state(enabled: state);
4594	}
4595	return err;
4596	}
4597	#endif
4598	#endif
4599
4600	#ifdef CONFIG_SCHEDSTATS
4601
4602	DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4603
4604	static void set_schedstats(bool enabled)
4605	{
4606	if (enabled)
4607	static_branch_enable(&sched_schedstats);
4608	else
4609	static_branch_disable(&sched_schedstats);
4610	}
4611
4612	void force_schedstat_enabled(void)
4613	{
4614	if (!schedstat_enabled()) {
4615	pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4616	static_branch_enable(&sched_schedstats);
4617	}
4618	}
4619
4620	static int __init setup_schedstats(char *str)
4621	{
4622	int ret = 0;
4623	if (!str)
4624	goto out;
4625
4626	if (!strcmp(str, "enable")) {
4627	set_schedstats(true);
4628	ret = 1;
4629	} else if (!strcmp(str, "disable")) {
4630	set_schedstats(false);
4631	ret = 1;
4632	}
4633	out:
4634	if (!ret)
4635	pr_warn("Unable to parse schedstats=\n");
4636
4637	return ret;
4638	}
4639	__setup("schedstats=", setup_schedstats);
4640
4641	#ifdef CONFIG_PROC_SYSCTL
4642	static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4643	size_t *lenp, loff_t *ppos)
4644	{
4645	struct ctl_table t;
4646	int err;
4647	int state = static_branch_likely(&sched_schedstats);
4648
4649	if (write && !capable(CAP_SYS_ADMIN))
4650	return -EPERM;
4651
4652	t = *table;
4653	t.data = &state;
4654	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4655	if (err < 0)
4656	return err;
4657	if (write)
4658	set_schedstats(state);
4659	return err;
4660	}
4661	#endif /* CONFIG_PROC_SYSCTL */
4662	#endif /* CONFIG_SCHEDSTATS */
4663
4664	#ifdef CONFIG_SYSCTL
4665	static struct ctl_table sched_core_sysctls[] = {
4666	#ifdef CONFIG_SCHEDSTATS
4667	{
4668	.procname = "sched_schedstats",
4669	.data = NULL,
4670	.maxlen = sizeof(unsigned int),
4671	.mode = 0644,
4672	.proc_handler = sysctl_schedstats,
4673	.extra1 = SYSCTL_ZERO,
4674	.extra2 = SYSCTL_ONE,
4675	},
4676	#endif /* CONFIG_SCHEDSTATS */
4677	#ifdef CONFIG_UCLAMP_TASK
4678	{
4679	.procname = "sched_util_clamp_min",
4680	.data = &sysctl_sched_uclamp_util_min,
4681	.maxlen = sizeof(unsigned int),
4682	.mode = 0644,
4683	.proc_handler = sysctl_sched_uclamp_handler,
4684	},
4685	{
4686	.procname = "sched_util_clamp_max",
4687	.data = &sysctl_sched_uclamp_util_max,
4688	.maxlen = sizeof(unsigned int),
4689	.mode = 0644,
4690	.proc_handler = sysctl_sched_uclamp_handler,
4691	},
4692	{
4693	.procname = "sched_util_clamp_min_rt_default",
4694	.data = &sysctl_sched_uclamp_util_min_rt_default,
4695	.maxlen = sizeof(unsigned int),
4696	.mode = 0644,
4697	.proc_handler = sysctl_sched_uclamp_handler,
4698	},
4699	#endif /* CONFIG_UCLAMP_TASK */
4700	#ifdef CONFIG_NUMA_BALANCING
4701	{
4702	.procname = "numa_balancing",
4703	.data = NULL, /* filled in by handler */
4704	.maxlen = sizeof(unsigned int),
4705	.mode = 0644,
4706	.proc_handler = sysctl_numa_balancing,
4707	.extra1 = SYSCTL_ZERO,
4708	.extra2 = SYSCTL_FOUR,
4709	},
4710	#endif /* CONFIG_NUMA_BALANCING */
4711	{}
4712	};
4713	static int __init sched_core_sysctl_init(void)
4714	{
4715	register_sysctl_init("kernel", sched_core_sysctls);
4716	return 0;
4717	}
4718	late_initcall(sched_core_sysctl_init);
4719	#endif /* CONFIG_SYSCTL */
4720
4721	/*
4722	* fork()/clone()-time setup:
4723	*/
4724	int sched_fork(unsigned long clone_flags, struct task_struct *p)
4725	{
4726	__sched_fork(clone_flags, p);
4727	/*
4728	* We mark the process as NEW here. This guarantees that
4729	* nobody will actually run it, and a signal or other external
4730	* event cannot wake it up and insert it on the runqueue either.
4731	*/
4732	p->__state = TASK_NEW;
4733
4734	/*
4735	* Make sure we do not leak PI boosting priority to the child.
4736	*/
4737	p->prio = current->normal_prio;
4738
4739	uclamp_fork(p);
4740
4741	/*
4742	* Revert to default priority/policy on fork if requested.
4743	*/
4744	if (unlikely(p->sched_reset_on_fork)) {
4745	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4746	p->policy = SCHED_NORMAL;
4747	p->static_prio = NICE_TO_PRIO(0);
4748	p->rt_priority = 0;
4749	} else if (PRIO_TO_NICE(p->static_prio) < 0)
4750	p->static_prio = NICE_TO_PRIO(0);
4751
4752	p->prio = p->normal_prio = p->static_prio;
4753	set_load_weight(p, update_load: false);
4754
4755	/*
4756	* We don't need the reset flag anymore after the fork. It has
4757	* fulfilled its duty:
4758	*/
4759	p->sched_reset_on_fork = 0;
4760	}
4761
4762	if (dl_prio(prio: p->prio))
4763	return -EAGAIN;
4764	else if (rt_prio(prio: p->prio))
4765	p->sched_class = &rt_sched_class;
4766	else
4767	p->sched_class = &fair_sched_class;
4768
4769	init_entity_runnable_average(se: &p->se);
4770
4771
4772	#ifdef CONFIG_SCHED_INFO
4773	if (likely(sched_info_on()))
4774	memset(&p->sched_info, 0, sizeof(p->sched_info));
4775	#endif
4776	#if defined(CONFIG_SMP)
4777	p->on_cpu = 0;
4778	#endif
4779	init_task_preempt_count(p);
4780	#ifdef CONFIG_SMP
4781	plist_node_init(node: &p->pushable_tasks, MAX_PRIO);
4782	RB_CLEAR_NODE(&p->pushable_dl_tasks);
4783	#endif
4784	return 0;
4785	}
4786
4787	void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4788	{
4789	unsigned long flags;
4790
4791	/*
4792	* Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4793	* required yet, but lockdep gets upset if rules are violated.
4794	*/
4795	raw_spin_lock_irqsave(&p->pi_lock, flags);
4796	#ifdef CONFIG_CGROUP_SCHED
4797	if (1) {
4798	struct task_group *tg;
4799	tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4800	struct task_group, css);
4801	tg = autogroup_task_group(p, tg);
4802	p->sched_task_group = tg;
4803	}
4804	#endif
4805	rseq_migrate(t: p);
4806	/*
4807	* We're setting the CPU for the first time, we don't migrate,
4808	* so use __set_task_cpu().
4809	*/
4810	__set_task_cpu(p, smp_processor_id());
4811	if (p->sched_class->task_fork)
4812	p->sched_class->task_fork(p);
4813	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4814	}
4815
4816	void sched_post_fork(struct task_struct *p)
4817	{
4818	uclamp_post_fork(p);
4819	}
4820
4821	unsigned long to_ratio(u64 period, u64 runtime)
4822	{
4823	if (runtime == RUNTIME_INF)
4824	return BW_UNIT;
4825
4826	/*
4827	* Doing this here saves a lot of checks in all
4828	* the calling paths, and returning zero seems
4829	* safe for them anyway.
4830	*/
4831	if (period == 0)
4832	return 0;
4833
4834	return div64_u64(dividend: runtime << BW_SHIFT, divisor: period);
4835	}
4836
4837	/*
4838	* wake_up_new_task - wake up a newly created task for the first time.
4839	*
4840	* This function will do some initial scheduler statistics housekeeping
4841	* that must be done for every newly created context, then puts the task
4842	* on the runqueue and wakes it.
4843	*/
4844	void wake_up_new_task(struct task_struct *p)
4845	{
4846	struct rq_flags rf;
4847	struct rq *rq;
4848
4849	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4850	WRITE_ONCE(p->__state, TASK_RUNNING);
4851	#ifdef CONFIG_SMP
4852	/*
4853	* Fork balancing, do it here and not earlier because:
4854	* - cpus_ptr can change in the fork path
4855	* - any previously selected CPU might disappear through hotplug
4856	*
4857	* Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4858	* as we're not fully set-up yet.
4859	*/
4860	p->recent_used_cpu = task_cpu(p);
4861	rseq_migrate(t: p);
4862	__set_task_cpu(p, cpu: select_task_rq(p, cpu: task_cpu(p), WF_FORK));
4863	#endif
4864	rq = __task_rq_lock(p, rf: &rf);
4865	update_rq_clock(rq);
4866	post_init_entity_util_avg(p);
4867
4868	activate_task(rq, p, ENQUEUE_NOCLOCK);
4869	trace_sched_wakeup_new(p);
4870	wakeup_preempt(rq, p, WF_FORK);
4871	#ifdef CONFIG_SMP
4872	if (p->sched_class->task_woken) {
4873	/*
4874	* Nothing relies on rq->lock after this, so it's fine to
4875	* drop it.
4876	*/
4877	rq_unpin_lock(rq, rf: &rf);
4878	p->sched_class->task_woken(rq, p);
4879	rq_repin_lock(rq, rf: &rf);
4880	}
4881	#endif
4882	task_rq_unlock(rq, p, rf: &rf);
4883	}
4884
4885	#ifdef CONFIG_PREEMPT_NOTIFIERS
4886
4887	static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4888
4889	void preempt_notifier_inc(void)
4890	{
4891	static_branch_inc(&preempt_notifier_key);
4892	}
4893	EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4894
4895	void preempt_notifier_dec(void)
4896	{
4897	static_branch_dec(&preempt_notifier_key);
4898	}
4899	EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4900
4901	/**
4902	* preempt_notifier_register - tell me when current is being preempted & rescheduled
4903	* @notifier: notifier struct to register
4904	*/
4905	void preempt_notifier_register(struct preempt_notifier *notifier)
4906	{
4907	if (!static_branch_unlikely(&preempt_notifier_key))
4908	WARN(1, "registering preempt_notifier while notifiers disabled\n");
4909
4910	hlist_add_head(n: &notifier->link, h: &current->preempt_notifiers);
4911	}
4912	EXPORT_SYMBOL_GPL(preempt_notifier_register);
4913
4914	/**
4915	* preempt_notifier_unregister - no longer interested in preemption notifications
4916	* @notifier: notifier struct to unregister
4917	*
4918	* This is *not* safe to call from within a preemption notifier.
4919	*/
4920	void preempt_notifier_unregister(struct preempt_notifier *notifier)
4921	{
4922	hlist_del(n: &notifier->link);
4923	}
4924	EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4925
4926	static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4927	{
4928	struct preempt_notifier *notifier;
4929
4930	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4931	notifier->ops->sched_in(notifier, raw_smp_processor_id());
4932	}
4933
4934	static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4935	{
4936	if (static_branch_unlikely(&preempt_notifier_key))
4937	__fire_sched_in_preempt_notifiers(curr);
4938	}
4939
4940	static void
4941	__fire_sched_out_preempt_notifiers(struct task_struct *curr,
4942	struct task_struct *next)
4943	{
4944	struct preempt_notifier *notifier;
4945
4946	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4947	notifier->ops->sched_out(notifier, next);
4948	}
4949
4950	static __always_inline void
4951	fire_sched_out_preempt_notifiers(struct task_struct *curr,
4952	struct task_struct *next)
4953	{
4954	if (static_branch_unlikely(&preempt_notifier_key))
4955	__fire_sched_out_preempt_notifiers(curr, next);
4956	}
4957
4958	#else /* !CONFIG_PREEMPT_NOTIFIERS */
4959
4960	static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4961	{
4962	}
4963
4964	static inline void
4965	fire_sched_out_preempt_notifiers(struct task_struct *curr,
4966	struct task_struct *next)
4967	{
4968	}
4969
4970	#endif /* CONFIG_PREEMPT_NOTIFIERS */
4971
4972	static inline void prepare_task(struct task_struct *next)
4973	{
4974	#ifdef CONFIG_SMP
4975	/*
4976	* Claim the task as running, we do this before switching to it
4977	* such that any running task will have this set.
4978	*
4979	* See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4980	* its ordering comment.
4981	*/
4982	WRITE_ONCE(next->on_cpu, 1);
4983	#endif
4984	}
4985
4986	static inline void finish_task(struct task_struct *prev)
4987	{
4988	#ifdef CONFIG_SMP
4989	/*
4990	* This must be the very last reference to @prev from this CPU. After
4991	* p->on_cpu is cleared, the task can be moved to a different CPU. We
4992	* must ensure this doesn't happen until the switch is completely
4993	* finished.
4994	*
4995	* In particular, the load of prev->state in finish_task_switch() must
4996	* happen before this.
4997	*
4998	* Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4999	*/
5000	smp_store_release(&prev->on_cpu, 0);
5001	#endif
5002	}
5003
5004	#ifdef CONFIG_SMP
5005
5006	static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5007	{
5008	void (*func)(struct rq *rq);
5009	struct balance_callback *next;
5010
5011	lockdep_assert_rq_held(rq);
5012
5013	while (head) {
5014	func = (void (*)(struct rq *))head->func;
5015	next = head->next;
5016	head->next = NULL;
5017	head = next;
5018
5019	func(rq);
5020	}
5021	}
5022
5023	static void balance_push(struct rq *rq);
5024
5025	/*
5026	* balance_push_callback is a right abuse of the callback interface and plays
5027	* by significantly different rules.
5028	*
5029	* Where the normal balance_callback's purpose is to be ran in the same context
5030	* that queued it (only later, when it's safe to drop rq->lock again),
5031	* balance_push_callback is specifically targeted at __schedule().
5032	*
5033	* This abuse is tolerated because it places all the unlikely/odd cases behind
5034	* a single test, namely: rq->balance_callback == NULL.
5035	*/
5036	struct balance_callback balance_push_callback = {
5037	.next = NULL,
5038	.func = balance_push,
5039	};
5040
5041	static inline struct balance_callback *
5042	__splice_balance_callbacks(struct rq *rq, bool split)
5043	{
5044	struct balance_callback *head = rq->balance_callback;
5045
5046	if (likely(!head))
5047	return NULL;
5048
5049	lockdep_assert_rq_held(rq);
5050	/*
5051	* Must not take balance_push_callback off the list when
5052	* splice_balance_callbacks() and balance_callbacks() are not
5053	* in the same rq->lock section.
5054	*
5055	* In that case it would be possible for __schedule() to interleave
5056	* and observe the list empty.
5057	*/
5058	if (split && head == &balance_push_callback)
5059	head = NULL;
5060	else
5061	rq->balance_callback = NULL;
5062
5063	return head;
5064	}
5065
5066	static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5067	{
5068	return __splice_balance_callbacks(rq, split: true);
5069	}
5070
5071	static void __balance_callbacks(struct rq *rq)
5072	{
5073	do_balance_callbacks(rq, head: __splice_balance_callbacks(rq, split: false));
5074	}
5075
5076	static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5077	{
5078	unsigned long flags;
5079
5080	if (unlikely(head)) {
5081	raw_spin_rq_lock_irqsave(rq, flags);
5082	do_balance_callbacks(rq, head);
5083	raw_spin_rq_unlock_irqrestore(rq, flags);
5084	}
5085	}
5086
5087	#else
5088
5089	static inline void __balance_callbacks(struct rq *rq)
5090	{
5091	}
5092
5093	static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5094	{
5095	return NULL;
5096	}
5097
5098	static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5099	{
5100	}
5101
5102	#endif
5103
5104	static inline void
5105	prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5106	{
5107	/*
5108	* Since the runqueue lock will be released by the next
5109	* task (which is an invalid locking op but in the case
5110	* of the scheduler it's an obvious special-case), so we
5111	* do an early lockdep release here:
5112	*/
5113	rq_unpin_lock(rq, rf);
5114	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5115	#ifdef CONFIG_DEBUG_SPINLOCK
5116	/* this is a valid case when another task releases the spinlock */
5117	rq_lockp(rq)->owner = next;
5118	#endif
5119	}
5120
5121	static inline void finish_lock_switch(struct rq *rq)
5122	{
5123	/*
5124	* If we are tracking spinlock dependencies then we have to
5125	* fix up the runqueue lock - which gets 'carried over' from
5126	* prev into current:
5127	*/
5128	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5129	__balance_callbacks(rq);
5130	raw_spin_rq_unlock_irq(rq);
5131	}
5132
5133	/*
5134	* NOP if the arch has not defined these:
5135	*/
5136
5137	#ifndef prepare_arch_switch
5138	# define prepare_arch_switch(next) do { } while (0)
5139	#endif
5140
5141	#ifndef finish_arch_post_lock_switch
5142	# define finish_arch_post_lock_switch() do { } while (0)
5143	#endif
5144
5145	static inline void kmap_local_sched_out(void)
5146	{
5147	#ifdef CONFIG_KMAP_LOCAL
5148	if (unlikely(current->kmap_ctrl.idx))
5149	__kmap_local_sched_out();
5150	#endif
5151	}
5152
5153	static inline void kmap_local_sched_in(void)
5154	{
5155	#ifdef CONFIG_KMAP_LOCAL
5156	if (unlikely(current->kmap_ctrl.idx))
5157	__kmap_local_sched_in();
5158	#endif
5159	}
5160
5161	/**
5162	* prepare_task_switch - prepare to switch tasks
5163	* @rq: the runqueue preparing to switch
5164	* @prev: the current task that is being switched out
5165	* @next: the task we are going to switch to.
5166	*
5167	* This is called with the rq lock held and interrupts off. It must
5168	* be paired with a subsequent finish_task_switch after the context
5169	* switch.
5170	*
5171	* prepare_task_switch sets up locking and calls architecture specific
5172	* hooks.
5173	*/
5174	static inline void
5175	prepare_task_switch(struct rq *rq, struct task_struct *prev,
5176	struct task_struct *next)
5177	{
5178	kcov_prepare_switch(prev);
5179	sched_info_switch(rq, prev, next);
5180	perf_event_task_sched_out(prev, next);
5181	rseq_preempt(t: prev);
5182	fire_sched_out_preempt_notifiers(curr: prev, next);
5183	kmap_local_sched_out();
5184	prepare_task(next);
5185	prepare_arch_switch(next);
5186	}
5187
5188	/**
5189	* finish_task_switch - clean up after a task-switch
5190	* @prev: the thread we just switched away from.
5191	*
5192	* finish_task_switch must be called after the context switch, paired
5193	* with a prepare_task_switch call before the context switch.
5194	* finish_task_switch will reconcile locking set up by prepare_task_switch,
5195	* and do any other architecture-specific cleanup actions.
5196	*
5197	* Note that we may have delayed dropping an mm in context_switch(). If
5198	* so, we finish that here outside of the runqueue lock. (Doing it
5199	* with the lock held can cause deadlocks; see schedule() for
5200	* details.)
5201	*
5202	* The context switch have flipped the stack from under us and restored the
5203	* local variables which were saved when this task called schedule() in the
5204	* past. prev == current is still correct but we need to recalculate this_rq
5205	* because prev may have moved to another CPU.
5206	*/
5207	static struct rq *finish_task_switch(struct task_struct *prev)
5208	__releases(rq->lock)
5209	{
5210	struct rq *rq = this_rq();
5211	struct mm_struct *mm = rq->prev_mm;
5212	unsigned int prev_state;
5213
5214	/*
5215	* The previous task will have left us with a preempt_count of 2
5216	* because it left us after:
5217	*
5218	* schedule()
5219	* preempt_disable(); // 1
5220	* __schedule()
5221	* raw_spin_lock_irq(&rq->lock) // 2
5222	*
5223	* Also, see FORK_PREEMPT_COUNT.
5224	*/
5225	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5226	"corrupted preempt_count: %s/%d/0x%x\n",
5227	current->comm, current->pid, preempt_count()))
5228	preempt_count_set(FORK_PREEMPT_COUNT);
5229
5230	rq->prev_mm = NULL;
5231
5232	/*
5233	* A task struct has one reference for the use as "current".
5234	* If a task dies, then it sets TASK_DEAD in tsk->state and calls
5235	* schedule one last time. The schedule call will never return, and
5236	* the scheduled task must drop that reference.
5237	*
5238	* We must observe prev->state before clearing prev->on_cpu (in
5239	* finish_task), otherwise a concurrent wakeup can get prev
5240	* running on another CPU and we could rave with its RUNNING -> DEAD
5241	* transition, resulting in a double drop.
5242	*/
5243	prev_state = READ_ONCE(prev->__state);
5244	vtime_task_switch(prev);
5245	perf_event_task_sched_in(prev, current);
5246	finish_task(prev);
5247	tick_nohz_task_switch();
5248	finish_lock_switch(rq);
5249	finish_arch_post_lock_switch();
5250	kcov_finish_switch(current);
5251	/*
5252	* kmap_local_sched_out() is invoked with rq::lock held and
5253	* interrupts disabled. There is no requirement for that, but the
5254	* sched out code does not have an interrupt enabled section.
5255	* Restoring the maps on sched in does not require interrupts being
5256	* disabled either.
5257	*/
5258	kmap_local_sched_in();
5259
5260	fire_sched_in_preempt_notifiers(current);
5261	/*
5262	* When switching through a kernel thread, the loop in
5263	* membarrier_{private,global}_expedited() may have observed that
5264	* kernel thread and not issued an IPI. It is therefore possible to
5265	* schedule between user->kernel->user threads without passing though
5266	* switch_mm(). Membarrier requires a barrier after storing to
5267	* rq->curr, before returning to userspace, so provide them here:
5268	*
5269	* - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5270	* provided by mmdrop_lazy_tlb(),
5271	* - a sync_core for SYNC_CORE.
5272	*/
5273	if (mm) {
5274	membarrier_mm_sync_core_before_usermode(mm);
5275	mmdrop_lazy_tlb_sched(mm);
5276	}
5277
5278	if (unlikely(prev_state == TASK_DEAD)) {
5279	if (prev->sched_class->task_dead)
5280	prev->sched_class->task_dead(prev);
5281
5282	/* Task is done with its stack. */
5283	put_task_stack(tsk: prev);
5284
5285	put_task_struct_rcu_user(task: prev);
5286	}
5287
5288	return rq;
5289	}
5290
5291	/**
5292	* schedule_tail - first thing a freshly forked thread must call.
5293	* @prev: the thread we just switched away from.
5294	*/
5295	asmlinkage __visible void schedule_tail(struct task_struct *prev)
5296	__releases(rq->lock)
5297	{
5298	/*
5299	* New tasks start with FORK_PREEMPT_COUNT, see there and
5300	* finish_task_switch() for details.
5301	*
5302	* finish_task_switch() will drop rq->lock() and lower preempt_count
5303	* and the preempt_enable() will end up enabling preemption (on
5304	* PREEMPT_COUNT kernels).
5305	*/
5306
5307	finish_task_switch(prev);
5308	preempt_enable();
5309
5310	if (current->set_child_tid)
5311	put_user(task_pid_vnr(current), current->set_child_tid);
5312
5313	calculate_sigpending();
5314	}
5315
5316	/*
5317	* context_switch - switch to the new MM and the new thread's register state.
5318	*/
5319	static __always_inline struct rq *
5320	context_switch(struct rq *rq, struct task_struct *prev,
5321	struct task_struct *next, struct rq_flags *rf)
5322	{
5323	prepare_task_switch(rq, prev, next);
5324
5325	/*
5326	* For paravirt, this is coupled with an exit in switch_to to
5327	* combine the page table reload and the switch backend into
5328	* one hypercall.
5329	*/
5330	arch_start_context_switch(prev);
5331
5332	/*
5333	* kernel -> kernel lazy + transfer active
5334	* user -> kernel lazy + mmgrab_lazy_tlb() active
5335	*
5336	* kernel -> user switch + mmdrop_lazy_tlb() active
5337	* user -> user switch
5338	*
5339	* switch_mm_cid() needs to be updated if the barriers provided
5340	* by context_switch() are modified.
5341	*/
5342	if (!next->mm) { // to kernel
5343	enter_lazy_tlb(mm: prev->active_mm, tsk: next);
5344
5345	next->active_mm = prev->active_mm;
5346	if (prev->mm) // from user
5347	mmgrab_lazy_tlb(mm: prev->active_mm);
5348	else
5349	prev->active_mm = NULL;
5350	} else { // to user
5351	membarrier_switch_mm(rq, prev_mm: prev->active_mm, next_mm: next->mm);
5352	/*
5353	* sys_membarrier() requires an smp_mb() between setting
5354	* rq->curr / membarrier_switch_mm() and returning to userspace.
5355	*
5356	* The below provides this either through switch_mm(), or in
5357	* case 'prev->active_mm == next->mm' through
5358	* finish_task_switch()'s mmdrop().
5359	*/
5360	switch_mm_irqs_off(prev: prev->active_mm, next: next->mm, tsk: next);
5361	lru_gen_use_mm(mm: next->mm);
5362
5363	if (!prev->mm) { // from kernel
5364	/* will mmdrop_lazy_tlb() in finish_task_switch(). */
5365	rq->prev_mm = prev->active_mm;
5366	prev->active_mm = NULL;
5367	}
5368	}
5369
5370	/* switch_mm_cid() requires the memory barriers above. */
5371	switch_mm_cid(rq, prev, next);
5372
5373	prepare_lock_switch(rq, next, rf);
5374
5375	/* Here we just switch the register state and the stack. */
5376	switch_to(prev, next, prev);
5377	barrier();
5378
5379	return finish_task_switch(prev);
5380	}
5381
5382	/*
5383	* nr_running and nr_context_switches:
5384	*
5385	* externally visible scheduler statistics: current number of runnable
5386	* threads, total number of context switches performed since bootup.
5387	*/
5388	unsigned int nr_running(void)
5389	{
5390	unsigned int i, sum = 0;
5391
5392	for_each_online_cpu(i)
5393	sum += cpu_rq(i)->nr_running;
5394
5395	return sum;
5396	}
5397
5398	/*
5399	* Check if only the current task is running on the CPU.
5400	*
5401	* Caution: this function does not check that the caller has disabled
5402	* preemption, thus the result might have a time-of-check-to-time-of-use
5403	* race. The caller is responsible to use it correctly, for example:
5404	*
5405	* - from a non-preemptible section (of course)
5406	*
5407	* - from a thread that is bound to a single CPU
5408	*
5409	* - in a loop with very short iterations (e.g. a polling loop)
5410	*/
5411	bool single_task_running(void)
5412	{
5413	return raw_rq()->nr_running == 1;
5414	}
5415	EXPORT_SYMBOL(single_task_running);
5416
5417	unsigned long long nr_context_switches_cpu(int cpu)
5418	{
5419	return cpu_rq(cpu)->nr_switches;
5420	}
5421
5422	unsigned long long nr_context_switches(void)
5423	{
5424	int i;
5425	unsigned long long sum = 0;
5426
5427	for_each_possible_cpu(i)
5428	sum += cpu_rq(i)->nr_switches;
5429
5430	return sum;
5431	}
5432
5433	/*
5434	* Consumers of these two interfaces, like for example the cpuidle menu
5435	* governor, are using nonsensical data. Preferring shallow idle state selection
5436	* for a CPU that has IO-wait which might not even end up running the task when
5437	* it does become runnable.
5438	*/
5439
5440	unsigned int nr_iowait_cpu(int cpu)
5441	{
5442	return atomic_read(v: &cpu_rq(cpu)->nr_iowait);
5443	}
5444
5445	/*
5446	* IO-wait accounting, and how it's mostly bollocks (on SMP).
5447	*
5448	* The idea behind IO-wait account is to account the idle time that we could
5449	* have spend running if it were not for IO. That is, if we were to improve the
5450	* storage performance, we'd have a proportional reduction in IO-wait time.
5451	*
5452	* This all works nicely on UP, where, when a task blocks on IO, we account
5453	* idle time as IO-wait, because if the storage were faster, it could've been
5454	* running and we'd not be idle.
5455	*
5456	* This has been extended to SMP, by doing the same for each CPU. This however
5457	* is broken.
5458	*
5459	* Imagine for instance the case where two tasks block on one CPU, only the one
5460	* CPU will have IO-wait accounted, while the other has regular idle. Even
5461	* though, if the storage were faster, both could've ran at the same time,
5462	* utilising both CPUs.
5463	*
5464	* This means, that when looking globally, the current IO-wait accounting on
5465	* SMP is a lower bound, by reason of under accounting.
5466	*
5467	* Worse, since the numbers are provided per CPU, they are sometimes
5468	* interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5469	* associated with any one particular CPU, it can wake to another CPU than it
5470	* blocked on. This means the per CPU IO-wait number is meaningless.
5471	*
5472	* Task CPU affinities can make all that even more 'interesting'.
5473	*/
5474
5475	unsigned int nr_iowait(void)
5476	{
5477	unsigned int i, sum = 0;
5478
5479	for_each_possible_cpu(i)
5480	sum += nr_iowait_cpu(cpu: i);
5481
5482	return sum;
5483	}
5484
5485	#ifdef CONFIG_SMP
5486
5487	/*
5488	* sched_exec - execve() is a valuable balancing opportunity, because at
5489	* this point the task has the smallest effective memory and cache footprint.
5490	*/
5491	void sched_exec(void)
5492	{
5493	struct task_struct *p = current;
5494	struct migration_arg arg;
5495	int dest_cpu;
5496
5497	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5498	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5499	if (dest_cpu == smp_processor_id())
5500	return;
5501
5502	if (unlikely(!cpu_active(dest_cpu)))
5503	return;
5504
5505	arg = (struct migration_arg){ p, dest_cpu };
5506	}
5507	stop_one_cpu(cpu: task_cpu(p), fn: migration_cpu_stop, arg: &arg);
5508	}
5509
5510	#endif
5511
5512	DEFINE_PER_CPU(struct kernel_stat, kstat);
5513	DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5514
5515	EXPORT_PER_CPU_SYMBOL(kstat);
5516	EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5517
5518	/*
5519	* The function fair_sched_class.update_curr accesses the struct curr
5520	* and its field curr->exec_start; when called from task_sched_runtime(),
5521	* we observe a high rate of cache misses in practice.
5522	* Prefetching this data results in improved performance.
5523	*/
5524	static inline void prefetch_curr_exec_start(struct task_struct *p)
5525	{
5526	#ifdef CONFIG_FAIR_GROUP_SCHED
5527	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5528	#else
5529	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5530	#endif
5531	prefetch(curr);
5532	prefetch(&curr->exec_start);
5533	}
5534
5535	/*
5536	* Return accounted runtime for the task.
5537	* In case the task is currently running, return the runtime plus current's
5538	* pending runtime that have not been accounted yet.
5539	*/
5540	unsigned long long task_sched_runtime(struct task_struct *p)
5541	{
5542	struct rq_flags rf;
5543	struct rq *rq;
5544	u64 ns;
5545
5546	#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5547	/*
5548	* 64-bit doesn't need locks to atomically read a 64-bit value.
5549	* So we have a optimization chance when the task's delta_exec is 0.
5550	* Reading ->on_cpu is racy, but this is ok.
5551	*
5552	* If we race with it leaving CPU, we'll take a lock. So we're correct.
5553	* If we race with it entering CPU, unaccounted time is 0. This is
5554	* indistinguishable from the read occurring a few cycles earlier.
5555	* If we see ->on_cpu without ->on_rq, the task is leaving, and has
5556	* been accounted, so we're correct here as well.
5557	*/
5558	if (!p->on_cpu || !task_on_rq_queued(p))
5559	return p->se.sum_exec_runtime;
5560	#endif
5561
5562	rq = task_rq_lock(p, rf: &rf);
5563	/*
5564	* Must be ->curr _and_ ->on_rq. If dequeued, we would
5565	* project cycles that may never be accounted to this
5566	* thread, breaking clock_gettime().
5567	*/
5568	if (task_current(rq, p) && task_on_rq_queued(p)) {
5569	prefetch_curr_exec_start(p);
5570	update_rq_clock(rq);
5571	p->sched_class->update_curr(rq);
5572	}
5573	ns = p->se.sum_exec_runtime;
5574	task_rq_unlock(rq, p, rf: &rf);
5575
5576	return ns;
5577	}
5578
5579	#ifdef CONFIG_SCHED_DEBUG
5580	static u64 cpu_resched_latency(struct rq *rq)
5581	{
5582	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5583	u64 resched_latency, now = rq_clock(rq);
5584	static bool warned_once;
5585
5586	if (sysctl_resched_latency_warn_once && warned_once)
5587	return 0;
5588
5589	if (!need_resched() || !latency_warn_ms)
5590	return 0;
5591
5592	if (system_state == SYSTEM_BOOTING)
5593	return 0;
5594
5595	if (!rq->last_seen_need_resched_ns) {
5596	rq->last_seen_need_resched_ns = now;
5597	rq->ticks_without_resched = 0;
5598	return 0;
5599	}
5600
5601	rq->ticks_without_resched++;
5602	resched_latency = now - rq->last_seen_need_resched_ns;
5603	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5604	return 0;
5605
5606	warned_once = true;
5607
5608	return resched_latency;
5609	}
5610
5611	static int __init setup_resched_latency_warn_ms(char *str)
5612	{
5613	long val;
5614
5615	if ((kstrtol(s: str, base: 0, res: &val))) {
5616	pr_warn("Unable to set resched_latency_warn_ms\n");
5617	return 1;
5618	}
5619
5620	sysctl_resched_latency_warn_ms = val;
5621	return 1;
5622	}
5623	__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5624	#else
5625	static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5626	#endif /* CONFIG_SCHED_DEBUG */
5627
5628	/*
5629	* This function gets called by the timer code, with HZ frequency.
5630	* We call it with interrupts disabled.
5631	*/
5632	void scheduler_tick(void)
5633	{
5634	int cpu = smp_processor_id();
5635	struct rq *rq = cpu_rq(cpu);
5636	struct task_struct *curr = rq->curr;
5637	struct rq_flags rf;
5638	unsigned long thermal_pressure;
5639	u64 resched_latency;
5640
5641	if (housekeeping_cpu(cpu, type: HK_TYPE_TICK))
5642	arch_scale_freq_tick();
5643
5644	sched_clock_tick();
5645
5646	rq_lock(rq, rf: &rf);
5647
5648	update_rq_clock(rq);
5649	thermal_pressure = arch_scale_thermal_pressure(cpu: cpu_of(rq));
5650	update_thermal_load_avg(now: rq_clock_thermal(rq), rq, capacity: thermal_pressure);
5651	curr->sched_class->task_tick(rq, curr, 0);
5652	if (sched_feat(LATENCY_WARN))
5653	resched_latency = cpu_resched_latency(rq);
5654	calc_global_load_tick(this_rq: rq);
5655	sched_core_tick(rq);
5656	task_tick_mm_cid(rq, curr);
5657
5658	rq_unlock(rq, rf: &rf);
5659
5660	if (sched_feat(LATENCY_WARN) && resched_latency)
5661	resched_latency_warn(cpu, latency: resched_latency);
5662
5663	perf_event_task_tick();
5664
5665	if (curr->flags & PF_WQ_WORKER)
5666	wq_worker_tick(task: curr);
5667
5668	#ifdef CONFIG_SMP
5669	rq->idle_balance = idle_cpu(cpu);
5670	trigger_load_balance(rq);
5671	#endif
5672	}
5673
5674	#ifdef CONFIG_NO_HZ_FULL
5675
5676	struct tick_work {
5677	int cpu;
5678	atomic_t state;
5679	struct delayed_work work;
5680	};
5681	/* Values for ->state, see diagram below. */
5682	#define TICK_SCHED_REMOTE_OFFLINE 0
5683	#define TICK_SCHED_REMOTE_OFFLINING 1
5684	#define TICK_SCHED_REMOTE_RUNNING 2
5685
5686	/*
5687	* State diagram for ->state:
5688	*
5689	*
5690	* TICK_SCHED_REMOTE_OFFLINE
5691	* | ^
5692	* | |
5693	* | | sched_tick_remote()
5694	* | |
5695	* | |
5696	* +--TICK_SCHED_REMOTE_OFFLINING
5697	* | ^
5698	* | |
5699	* sched_tick_start() | | sched_tick_stop()
5700	* | |
5701	* V |
5702	* TICK_SCHED_REMOTE_RUNNING
5703	*
5704	*
5705	* Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5706	* and sched_tick_start() are happy to leave the state in RUNNING.
5707	*/
5708
5709	static struct tick_work __percpu *tick_work_cpu;
5710
5711	static void sched_tick_remote(struct work_struct *work)
5712	{
5713	struct delayed_work *dwork = to_delayed_work(work);
5714	struct tick_work *twork = container_of(dwork, struct tick_work, work);
5715	int cpu = twork->cpu;
5716	struct rq *rq = cpu_rq(cpu);
5717	int os;
5718
5719	/*
5720	* Handle the tick only if it appears the remote CPU is running in full
5721	* dynticks mode. The check is racy by nature, but missing a tick or
5722	* having one too much is no big deal because the scheduler tick updates
5723	* statistics and checks timeslices in a time-independent way, regardless
5724	* of when exactly it is running.
5725	*/
5726	if (tick_nohz_tick_stopped_cpu(cpu)) {
5727	guard(rq_lock_irq)(rq);
5728	struct task_struct *curr = rq->curr;
5729
5730	if (cpu_online(cpu)) {
5731	update_rq_clock(rq);
5732
5733	if (!is_idle_task(curr)) {
5734	/*
5735	* Make sure the next tick runs within a
5736	* reasonable amount of time.
5737	*/
5738	u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5739	WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5740	}
5741	curr->sched_class->task_tick(rq, curr, 0);
5742
5743	calc_load_nohz_remote(rq);
5744	}
5745	}
5746
5747	/*
5748	* Run the remote tick once per second (1Hz). This arbitrary
5749	* frequency is large enough to avoid overload but short enough
5750	* to keep scheduler internal stats reasonably up to date. But
5751	* first update state to reflect hotplug activity if required.
5752	*/
5753	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5754	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5755	if (os == TICK_SCHED_REMOTE_RUNNING)
5756	queue_delayed_work(system_unbound_wq, dwork, HZ);
5757	}
5758
5759	static void sched_tick_start(int cpu)
5760	{
5761	int os;
5762	struct tick_work *twork;
5763
5764	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5765	return;
5766
5767	WARN_ON_ONCE(!tick_work_cpu);
5768
5769	twork = per_cpu_ptr(tick_work_cpu, cpu);
5770	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5771	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5772	if (os == TICK_SCHED_REMOTE_OFFLINE) {
5773	twork->cpu = cpu;
5774	INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5775	queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5776	}
5777	}
5778
5779	#ifdef CONFIG_HOTPLUG_CPU
5780	static void sched_tick_stop(int cpu)
5781	{
5782	struct tick_work *twork;
5783	int os;
5784
5785	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5786	return;
5787
5788	WARN_ON_ONCE(!tick_work_cpu);
5789
5790	twork = per_cpu_ptr(tick_work_cpu, cpu);
5791	/* There cannot be competing actions, but don't rely on stop-machine. */
5792	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5793	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5794	/* Don't cancel, as this would mess up the state machine. */
5795	}
5796	#endif /* CONFIG_HOTPLUG_CPU */
5797
5798	int __init sched_tick_offload_init(void)
5799	{
5800	tick_work_cpu = alloc_percpu(struct tick_work);
5801	BUG_ON(!tick_work_cpu);
5802	return 0;
5803	}
5804
5805	#else /* !CONFIG_NO_HZ_FULL */
5806	static inline void sched_tick_start(int cpu) { }
5807	static inline void sched_tick_stop(int cpu) { }
5808	#endif
5809
5810	#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5811	defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5812	/*
5813	* If the value passed in is equal to the current preempt count
5814	* then we just disabled preemption. Start timing the latency.
5815	*/
5816	static inline void preempt_latency_start(int val)
5817	{
5818	if (preempt_count() == val) {
5819	unsigned long ip = get_lock_parent_ip();
5820	#ifdef CONFIG_DEBUG_PREEMPT
5821	current->preempt_disable_ip = ip;
5822	#endif
5823	trace_preempt_off(CALLER_ADDR0, a1: ip);
5824	}
5825	}
5826
5827	void preempt_count_add(int val)
5828	{
5829	#ifdef CONFIG_DEBUG_PREEMPT
5830	/*
5831	* Underflow?
5832	*/
5833	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5834	return;
5835	#endif
5836	__preempt_count_add(val);
5837	#ifdef CONFIG_DEBUG_PREEMPT
5838	/*
5839	* Spinlock count overflowing soon?
5840	*/
5841	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5842	PREEMPT_MASK - 10);
5843	#endif
5844	preempt_latency_start(val);
5845	}
5846	EXPORT_SYMBOL(preempt_count_add);
5847	NOKPROBE_SYMBOL(preempt_count_add);
5848
5849	/*
5850	* If the value passed in equals to the current preempt count
5851	* then we just enabled preemption. Stop timing the latency.
5852	*/
5853	static inline void preempt_latency_stop(int val)
5854	{
5855	if (preempt_count() == val)
5856	trace_preempt_on(CALLER_ADDR0, a1: get_lock_parent_ip());
5857	}
5858
5859	void preempt_count_sub(int val)
5860	{
5861	#ifdef CONFIG_DEBUG_PREEMPT
5862	/*
5863	* Underflow?
5864	*/
5865	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5866	return;
5867	/*
5868	* Is the spinlock portion underflowing?
5869	*/
5870	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5871	!(preempt_count() & PREEMPT_MASK)))
5872	return;
5873	#endif
5874
5875	preempt_latency_stop(val);
5876	__preempt_count_sub(val);
5877	}
5878	EXPORT_SYMBOL(preempt_count_sub);
5879	NOKPROBE_SYMBOL(preempt_count_sub);
5880
5881	#else
5882	static inline void preempt_latency_start(int val) { }
5883	static inline void preempt_latency_stop(int val) { }
5884	#endif
5885
5886	static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5887	{
5888	#ifdef CONFIG_DEBUG_PREEMPT
5889	return p->preempt_disable_ip;
5890	#else
5891	return 0;
5892	#endif
5893	}
5894
5895	/*
5896	* Print scheduling while atomic bug:
5897	*/
5898	static noinline void __schedule_bug(struct task_struct *prev)
5899	{
5900	/* Save this before calling printk(), since that will clobber it */
5901	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5902
5903	if (oops_in_progress)
5904	return;
5905
5906	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5907	prev->comm, prev->pid, preempt_count());
5908
5909	debug_show_held_locks(task: prev);
5910	print_modules();
5911	if (irqs_disabled())
5912	print_irqtrace_events(curr: prev);
5913	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
5914	pr_err("Preemption disabled at:");
5915	print_ip_sym(KERN_ERR, ip: preempt_disable_ip);
5916	}
5917	check_panic_on_warn(origin: "scheduling while atomic");
5918
5919	dump_stack();
5920	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5921	}
5922
5923	/*
5924	* Various schedule()-time debugging checks and statistics:
5925	*/
5926	static inline void schedule_debug(struct task_struct *prev, bool preempt)
5927	{
5928	#ifdef CONFIG_SCHED_STACK_END_CHECK
5929	if (task_stack_end_corrupted(prev))
5930	panic(fmt: "corrupted stack end detected inside scheduler\n");
5931
5932	if (task_scs_end_corrupted(tsk: prev))
5933	panic(fmt: "corrupted shadow stack detected inside scheduler\n");
5934	#endif
5935
5936	#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5937	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5938	printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5939	prev->comm, prev->pid, prev->non_block_count);
5940	dump_stack();
5941	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5942	}
5943	#endif
5944
5945	if (unlikely(in_atomic_preempt_off())) {
5946	__schedule_bug(prev);
5947	preempt_count_set(PREEMPT_DISABLED);
5948	}
5949	rcu_sleep_check();
5950	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5951
5952	profile_hit(SCHED_PROFILING, ip: __builtin_return_address(0));
5953
5954	schedstat_inc(this_rq()->sched_count);
5955	}
5956
5957	static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5958	struct rq_flags *rf)
5959	{
5960	#ifdef CONFIG_SMP
5961	const struct sched_class *class;
5962	/*
5963	* We must do the balancing pass before put_prev_task(), such
5964	* that when we release the rq->lock the task is in the same
5965	* state as before we took rq->lock.
5966	*
5967	* We can terminate the balance pass as soon as we know there is
5968	* a runnable task of @class priority or higher.
5969	*/
5970	for_class_range(class, prev->sched_class, &idle_sched_class) {
5971	if (class->balance(rq, prev, rf))
5972	break;
5973	}
5974	#endif
5975
5976	put_prev_task(rq, prev);
5977	}
5978
5979	/*
5980	* Pick up the highest-prio task:
5981	*/
5982	static inline struct task_struct *
5983	__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5984	{
5985	const struct sched_class *class;
5986	struct task_struct *p;
5987
5988	/*
5989	* Optimization: we know that if all tasks are in the fair class we can
5990	* call that function directly, but only if the @prev task wasn't of a
5991	* higher scheduling class, because otherwise those lose the
5992	* opportunity to pull in more work from other CPUs.
5993	*/
5994	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5995	rq->nr_running == rq->cfs.h_nr_running)) {
5996
5997	p = pick_next_task_fair(rq, prev, rf);
5998	if (unlikely(p == RETRY_TASK))
5999	goto restart;
6000
6001	/* Assume the next prioritized class is idle_sched_class */
6002	if (!p) {
6003	put_prev_task(rq, prev);
6004	p = pick_next_task_idle(rq);
6005	}
6006
6007	return p;
6008	}
6009
6010	restart:
6011	put_prev_task_balance(rq, prev, rf);
6012
6013	for_each_class(class) {
6014	p = class->pick_next_task(rq);
6015	if (p)
6016	return p;
6017	}
6018
6019	BUG(); /* The idle class should always have a runnable task. */
6020	}
6021
6022	#ifdef CONFIG_SCHED_CORE
6023	static inline bool is_task_rq_idle(struct task_struct *t)
6024	{
6025	return (task_rq(t)->idle == t);
6026	}
6027
6028	static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6029	{
6030	return is_task_rq_idle(t: a) || (a->core_cookie == cookie);
6031	}
6032
6033	static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6034	{
6035	if (is_task_rq_idle(t: a) || is_task_rq_idle(t: b))
6036	return true;
6037
6038	return a->core_cookie == b->core_cookie;
6039	}
6040
6041	static inline struct task_struct *pick_task(struct rq *rq)
6042	{
6043	const struct sched_class *class;
6044	struct task_struct *p;
6045
6046	for_each_class(class) {
6047	p = class->pick_task(rq);
6048	if (p)
6049	return p;
6050	}
6051
6052	BUG(); /* The idle class should always have a runnable task. */
6053	}
6054
6055	extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6056
6057	static void queue_core_balance(struct rq *rq);
6058
6059	static struct task_struct *
6060	pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6061	{
6062	struct task_struct *next, *p, *max = NULL;
6063	const struct cpumask *smt_mask;
6064	bool fi_before = false;
6065	bool core_clock_updated = (rq == rq->core);
6066	unsigned long cookie;
6067	int i, cpu, occ = 0;
6068	struct rq *rq_i;
6069	bool need_sync;
6070
6071	if (!sched_core_enabled(rq))
6072	return __pick_next_task(rq, prev, rf);
6073
6074	cpu = cpu_of(rq);
6075
6076	/* Stopper task is switching into idle, no need core-wide selection. */
6077	if (cpu_is_offline(cpu)) {
6078	/*
6079	* Reset core_pick so that we don't enter the fastpath when
6080	* coming online. core_pick would already be migrated to
6081	* another cpu during offline.
6082	*/
6083	rq->core_pick = NULL;
6084	return __pick_next_task(rq, prev, rf);
6085	}
6086
6087	/*
6088	* If there were no {en,de}queues since we picked (IOW, the task
6089	* pointers are all still valid), and we haven't scheduled the last
6090	* pick yet, do so now.
6091	*
6092	* rq->core_pick can be NULL if no selection was made for a CPU because
6093	* it was either offline or went offline during a sibling's core-wide
6094	* selection. In this case, do a core-wide selection.
6095	*/
6096	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6097	rq->core->core_pick_seq != rq->core_sched_seq &&
6098	rq->core_pick) {
6099	WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6100
6101	next = rq->core_pick;
6102	if (next != prev) {
6103	put_prev_task(rq, prev);
6104	set_next_task(rq, next);
6105	}
6106
6107	rq->core_pick = NULL;
6108	goto out;
6109	}
6110
6111	put_prev_task_balance(rq, prev, rf);
6112
6113	smt_mask = cpu_smt_mask(cpu);
6114	need_sync = !!rq->core->core_cookie;
6115
6116	/* reset state */
6117	rq->core->core_cookie = 0UL;
6118	if (rq->core->core_forceidle_count) {
6119	if (!core_clock_updated) {
6120	update_rq_clock(rq: rq->core);
6121	core_clock_updated = true;
6122	}
6123	sched_core_account_forceidle(rq);
6124	/* reset after accounting force idle */
6125	rq->core->core_forceidle_start = 0;
6126	rq->core->core_forceidle_count = 0;
6127	rq->core->core_forceidle_occupation = 0;
6128	need_sync = true;
6129	fi_before = true;
6130	}
6131
6132	/*
6133	* core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6134	*
6135	* @task_seq guards the task state ({en,de}queues)
6136	* @pick_seq is the @task_seq we did a selection on
6137	* @sched_seq is the @pick_seq we scheduled
6138	*
6139	* However, preemptions can cause multiple picks on the same task set.
6140	* 'Fix' this by also increasing @task_seq for every pick.
6141	*/
6142	rq->core->core_task_seq++;
6143
6144	/*
6145	* Optimize for common case where this CPU has no cookies
6146	* and there are no cookied tasks running on siblings.
6147	*/
6148	if (!need_sync) {
6149	next = pick_task(rq);
6150	if (!next->core_cookie) {
6151	rq->core_pick = NULL;
6152	/*
6153	* For robustness, update the min_vruntime_fi for
6154	* unconstrained picks as well.
6155	*/
6156	WARN_ON_ONCE(fi_before);
6157	task_vruntime_update(rq, p: next, in_fi: false);
6158	goto out_set_next;
6159	}
6160	}
6161
6162	/*
6163	* For each thread: do the regular task pick and find the max prio task
6164	* amongst them.
6165	*
6166	* Tie-break prio towards the current CPU
6167	*/
6168	for_each_cpu_wrap(i, smt_mask, cpu) {
6169	rq_i = cpu_rq(i);
6170
6171	/*
6172	* Current cpu always has its clock updated on entrance to
6173	* pick_next_task(). If the current cpu is not the core,
6174	* the core may also have been updated above.
6175	*/
6176	if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6177	update_rq_clock(rq: rq_i);
6178
6179	p = rq_i->core_pick = pick_task(rq: rq_i);
6180	if (!max || prio_less(a: max, b: p, in_fi: fi_before))
6181	max = p;
6182	}
6183
6184	cookie = rq->core->core_cookie = max->core_cookie;
6185
6186	/*
6187	* For each thread: try and find a runnable task that matches @max or
6188	* force idle.
6189	*/
6190	for_each_cpu(i, smt_mask) {
6191	rq_i = cpu_rq(i);
6192	p = rq_i->core_pick;
6193
6194	if (!cookie_equals(a: p, cookie)) {
6195	p = NULL;
6196	if (cookie)
6197	p = sched_core_find(rq: rq_i, cookie);
6198	if (!p)
6199	p = idle_sched_class.pick_task(rq_i);
6200	}
6201
6202	rq_i->core_pick = p;
6203
6204	if (p == rq_i->idle) {
6205	if (rq_i->nr_running) {
6206	rq->core->core_forceidle_count++;
6207	if (!fi_before)
6208	rq->core->core_forceidle_seq++;
6209	}
6210	} else {
6211	occ++;
6212	}
6213	}
6214
6215	if (schedstat_enabled() && rq->core->core_forceidle_count) {
6216	rq->core->core_forceidle_start = rq_clock(rq: rq->core);
6217	rq->core->core_forceidle_occupation = occ;
6218	}
6219
6220	rq->core->core_pick_seq = rq->core->core_task_seq;
6221	next = rq->core_pick;
6222	rq->core_sched_seq = rq->core->core_pick_seq;
6223
6224	/* Something should have been selected for current CPU */
6225	WARN_ON_ONCE(!next);
6226
6227	/*
6228	* Reschedule siblings
6229	*
6230	* NOTE: L1TF -- at this point we're no longer running the old task and
6231	* sending an IPI (below) ensures the sibling will no longer be running
6232	* their task. This ensures there is no inter-sibling overlap between
6233	* non-matching user state.
6234	*/
6235	for_each_cpu(i, smt_mask) {
6236	rq_i = cpu_rq(i);
6237
6238	/*
6239	* An online sibling might have gone offline before a task
6240	* could be picked for it, or it might be offline but later
6241	* happen to come online, but its too late and nothing was
6242	* picked for it. That's Ok - it will pick tasks for itself,
6243	* so ignore it.
6244	*/
6245	if (!rq_i->core_pick)
6246	continue;
6247
6248	/*
6249	* Update for new !FI->FI transitions, or if continuing to be in !FI:
6250	* fi_before fi update?
6251	* 0 0 1
6252	* 0 1 1
6253	* 1 0 1
6254	* 1 1 0
6255	*/
6256	if (!(fi_before && rq->core->core_forceidle_count))
6257	task_vruntime_update(rq: rq_i, p: rq_i->core_pick, in_fi: !!rq->core->core_forceidle_count);
6258
6259	rq_i->core_pick->core_occupation = occ;
6260
6261	if (i == cpu) {
6262	rq_i->core_pick = NULL;
6263	continue;
6264	}
6265
6266	/* Did we break L1TF mitigation requirements? */
6267	WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6268
6269	if (rq_i->curr == rq_i->core_pick) {
6270	rq_i->core_pick = NULL;
6271	continue;
6272	}
6273
6274	resched_curr(rq: rq_i);
6275	}
6276
6277	out_set_next:
6278	set_next_task(rq, next);
6279	out:
6280	if (rq->core->core_forceidle_count && next == rq->idle)
6281	queue_core_balance(rq);
6282
6283	return next;
6284	}
6285
6286	static bool try_steal_cookie(int this, int that)
6287	{
6288	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6289	struct task_struct *p;
6290	unsigned long cookie;
6291	bool success = false;
6292
6293	guard(irq)();
6294	guard(double_rq_lock)(lock: dst, lock2: src);
6295
6296	cookie = dst->core->core_cookie;
6297	if (!cookie)
6298	return false;
6299
6300	if (dst->curr != dst->idle)
6301	return false;
6302
6303	p = sched_core_find(rq: src, cookie);
6304	if (!p)
6305	return false;
6306
6307	do {
6308	if (p == src->core_pick || p == src->curr)
6309	goto next;
6310
6311	if (!is_cpu_allowed(p, cpu: this))
6312	goto next;
6313
6314	if (p->core_occupation > dst->idle->core_occupation)
6315	goto next;
6316	/*
6317	* sched_core_find() and sched_core_next() will ensure
6318	* that task @p is not throttled now, we also need to
6319	* check whether the runqueue of the destination CPU is
6320	* being throttled.
6321	*/
6322	if (sched_task_is_throttled(p, cpu: this))
6323	goto next;
6324
6325	deactivate_task(rq: src, p, flags: 0);
6326	set_task_cpu(p, new_cpu: this);
6327	activate_task(rq: dst, p, flags: 0);
6328
6329	resched_curr(rq: dst);
6330
6331	success = true;
6332	break;
6333
6334	next:
6335	p = sched_core_next(p, cookie);
6336	} while (p);
6337
6338	return success;
6339	}
6340
6341	static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6342	{
6343	int i;
6344
6345	for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6346	if (i == cpu)
6347	continue;
6348
6349	if (need_resched())
6350	break;
6351
6352	if (try_steal_cookie(this: cpu, that: i))
6353	return true;
6354	}
6355
6356	return false;
6357	}
6358
6359	static void sched_core_balance(struct rq *rq)
6360	{
6361	struct sched_domain *sd;
6362	int cpu = cpu_of(rq);
6363
6364	guard(preempt)();
6365	guard(rcu)();
6366
6367	raw_spin_rq_unlock_irq(rq);
6368	for_each_domain(cpu, sd) {
6369	if (need_resched())
6370	break;
6371
6372	if (steal_cookie_task(cpu, sd))
6373	break;
6374	}
6375	raw_spin_rq_lock_irq(rq);
6376	}
6377
6378	static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6379
6380	static void queue_core_balance(struct rq *rq)
6381	{
6382	if (!sched_core_enabled(rq))
6383	return;
6384
6385	if (!rq->core->core_cookie)
6386	return;
6387
6388	if (!rq->nr_running) /* not forced idle */
6389	return;
6390
6391	queue_balance_callback(rq, head: &per_cpu(core_balance_head, rq->cpu), func: sched_core_balance);
6392	}
6393
6394	DEFINE_LOCK_GUARD_1(core_lock, int,
6395	sched_core_lock(*_T->lock, &_T->flags),
6396	sched_core_unlock(*_T->lock, &_T->flags),
6397	unsigned long flags)
6398
6399	static void sched_core_cpu_starting(unsigned int cpu)
6400	{
6401	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6402	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6403	int t;
6404
6405	guard(core_lock)(l: &cpu);
6406
6407	WARN_ON_ONCE(rq->core != rq);
6408
6409	/* if we're the first, we'll be our own leader */
6410	if (cpumask_weight(srcp: smt_mask) == 1)
6411	return;
6412
6413	/* find the leader */
6414	for_each_cpu(t, smt_mask) {
6415	if (t == cpu)
6416	continue;
6417	rq = cpu_rq(t);
6418	if (rq->core == rq) {
6419	core_rq = rq;
6420	break;
6421	}
6422	}
6423
6424	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6425	return;
6426
6427	/* install and validate core_rq */
6428	for_each_cpu(t, smt_mask) {
6429	rq = cpu_rq(t);
6430
6431	if (t == cpu)
6432	rq->core = core_rq;
6433
6434	WARN_ON_ONCE(rq->core != core_rq);
6435	}
6436	}
6437
6438	static void sched_core_cpu_deactivate(unsigned int cpu)
6439	{
6440	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6441	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6442	int t;
6443
6444	guard(core_lock)(l: &cpu);
6445
6446	/* if we're the last man standing, nothing to do */
6447	if (cpumask_weight(srcp: smt_mask) == 1) {
6448	WARN_ON_ONCE(rq->core != rq);
6449	return;
6450	}
6451
6452	/* if we're not the leader, nothing to do */
6453	if (rq->core != rq)
6454	return;
6455
6456	/* find a new leader */
6457	for_each_cpu(t, smt_mask) {
6458	if (t == cpu)
6459	continue;
6460	core_rq = cpu_rq(t);
6461	break;
6462	}
6463
6464	if (WARN_ON_ONCE(!core_rq)) /* impossible */
6465	return;
6466
6467	/* copy the shared state to the new leader */
6468	core_rq->core_task_seq = rq->core_task_seq;
6469	core_rq->core_pick_seq = rq->core_pick_seq;
6470	core_rq->core_cookie = rq->core_cookie;
6471	core_rq->core_forceidle_count = rq->core_forceidle_count;
6472	core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6473	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6474
6475	/*
6476	* Accounting edge for forced idle is handled in pick_next_task().
6477	* Don't need another one here, since the hotplug thread shouldn't
6478	* have a cookie.
6479	*/
6480	core_rq->core_forceidle_start = 0;
6481
6482	/* install new leader */
6483	for_each_cpu(t, smt_mask) {
6484	rq = cpu_rq(t);
6485	rq->core = core_rq;
6486	}
6487	}
6488
6489	static inline void sched_core_cpu_dying(unsigned int cpu)
6490	{
6491	struct rq *rq = cpu_rq(cpu);
6492
6493	if (rq->core != rq)
6494	rq->core = rq;
6495	}
6496
6497	#else /* !CONFIG_SCHED_CORE */
6498
6499	static inline void sched_core_cpu_starting(unsigned int cpu) {}
6500	static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6501	static inline void sched_core_cpu_dying(unsigned int cpu) {}
6502
6503	static struct task_struct *
6504	pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6505	{
6506	return __pick_next_task(rq, prev, rf);
6507	}
6508
6509	#endif /* CONFIG_SCHED_CORE */
6510
6511	/*
6512	* Constants for the sched_mode argument of __schedule().
6513	*
6514	* The mode argument allows RT enabled kernels to differentiate a
6515	* preemption from blocking on an 'sleeping' spin/rwlock. Note that
6516	* SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6517	* optimize the AND operation out and just check for zero.
6518	*/
6519	#define SM_NONE 0x0
6520	#define SM_PREEMPT 0x1
6521	#define SM_RTLOCK_WAIT 0x2
6522
6523	#ifndef CONFIG_PREEMPT_RT
6524	# define SM_MASK_PREEMPT (~0U)
6525	#else
6526	# define SM_MASK_PREEMPT SM_PREEMPT
6527	#endif
6528
6529	/*
6530	* __schedule() is the main scheduler function.
6531	*
6532	* The main means of driving the scheduler and thus entering this function are:
6533	*
6534	* 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6535	*
6536	* 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6537	* paths. For example, see arch/x86/entry_64.S.
6538	*
6539	* To drive preemption between tasks, the scheduler sets the flag in timer
6540	* interrupt handler scheduler_tick().
6541	*
6542	* 3. Wakeups don't really cause entry into schedule(). They add a
6543	* task to the run-queue and that's it.
6544	*
6545	* Now, if the new task added to the run-queue preempts the current
6546	* task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6547	* called on the nearest possible occasion:
6548	*
6549	* - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6550	*
6551	* - in syscall or exception context, at the next outmost
6552	* preempt_enable(). (this might be as soon as the wake_up()'s
6553	* spin_unlock()!)
6554	*
6555	* - in IRQ context, return from interrupt-handler to
6556	* preemptible context
6557	*
6558	* - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6559	* then at the next:
6560	*
6561	* - cond_resched() call
6562	* - explicit schedule() call
6563	* - return from syscall or exception to user-space
6564	* - return from interrupt-handler to user-space
6565	*
6566	* WARNING: must be called with preemption disabled!
6567	*/
6568	static void __sched notrace __schedule(unsigned int sched_mode)
6569	{
6570	struct task_struct *prev, *next;
6571	unsigned long *switch_count;
6572	unsigned long prev_state;
6573	struct rq_flags rf;
6574	struct rq *rq;
6575	int cpu;
6576
6577	cpu = smp_processor_id();
6578	rq = cpu_rq(cpu);
6579	prev = rq->curr;
6580
6581	schedule_debug(prev, preempt: !!sched_mode);
6582
6583	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6584	hrtick_clear(rq);
6585
6586	local_irq_disable();
6587	rcu_note_context_switch(preempt: !!sched_mode);
6588
6589	/*
6590	* Make sure that signal_pending_state()->signal_pending() below
6591	* can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6592	* done by the caller to avoid the race with signal_wake_up():
6593	*
6594	* __set_current_state(@state) signal_wake_up()
6595	* schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6596	* wake_up_state(p, state)
6597	* LOCK rq->lock LOCK p->pi_state
6598	* smp_mb__after_spinlock() smp_mb__after_spinlock()
6599	* if (signal_pending_state()) if (p->state & @state)
6600	*
6601	* Also, the membarrier system call requires a full memory barrier
6602	* after coming from user-space, before storing to rq->curr.
6603	*/
6604	rq_lock(rq, rf: &rf);
6605	smp_mb__after_spinlock();
6606
6607	/* Promote REQ to ACT */
6608	rq->clock_update_flags <<= 1;
6609	update_rq_clock(rq);
6610	rq->clock_update_flags = RQCF_UPDATED;
6611
6612	switch_count = &prev->nivcsw;
6613
6614	/*
6615	* We must load prev->state once (task_struct::state is volatile), such
6616	* that we form a control dependency vs deactivate_task() below.
6617	*/
6618	prev_state = READ_ONCE(prev->__state);
6619	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6620	if (signal_pending_state(state: prev_state, p: prev)) {
6621	WRITE_ONCE(prev->__state, TASK_RUNNING);
6622	} else {
6623	prev->sched_contributes_to_load =
6624	(prev_state & TASK_UNINTERRUPTIBLE) &&
6625	!(prev_state & TASK_NOLOAD) &&
6626	!(prev_state & TASK_FROZEN);
6627
6628	if (prev->sched_contributes_to_load)
6629	rq->nr_uninterruptible++;
6630
6631	/*
6632	* __schedule() ttwu()
6633	* prev_state = prev->state; if (p->on_rq && ...)
6634	* if (prev_state) goto out;
6635	* p->on_rq = 0; smp_acquire__after_ctrl_dep();
6636	* p->state = TASK_WAKING
6637	*
6638	* Where __schedule() and ttwu() have matching control dependencies.
6639	*
6640	* After this, schedule() must not care about p->state any more.
6641	*/
6642	deactivate_task(rq, p: prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6643
6644	if (prev->in_iowait) {
6645	atomic_inc(v: &rq->nr_iowait);
6646	delayacct_blkio_start();
6647	}
6648	}
6649	switch_count = &prev->nvcsw;
6650	}
6651
6652	next = pick_next_task(rq, prev, rf: &rf);
6653	clear_tsk_need_resched(tsk: prev);
6654	clear_preempt_need_resched();
6655	#ifdef CONFIG_SCHED_DEBUG
6656	rq->last_seen_need_resched_ns = 0;
6657	#endif
6658
6659	if (likely(prev != next)) {
6660	rq->nr_switches++;
6661	/*
6662	* RCU users of rcu_dereference(rq->curr) may not see
6663	* changes to task_struct made by pick_next_task().
6664	*/
6665	RCU_INIT_POINTER(rq->curr, next);
6666	/*
6667	* The membarrier system call requires each architecture
6668	* to have a full memory barrier after updating
6669	* rq->curr, before returning to user-space.
6670	*
6671	* Here are the schemes providing that barrier on the
6672	* various architectures:
6673	* - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6674	* switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6675	* - finish_lock_switch() for weakly-ordered
6676	* architectures where spin_unlock is a full barrier,
6677	* - switch_to() for arm64 (weakly-ordered, spin_unlock
6678	* is a RELEASE barrier),
6679	*/
6680	++*switch_count;
6681
6682	migrate_disable_switch(rq, p: prev);
6683	psi_sched_switch(prev, next, sleep: !task_on_rq_queued(p: prev));
6684
6685	trace_sched_switch(preempt: sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6686
6687	/* Also unlocks the rq: */
6688	rq = context_switch(rq, prev, next, rf: &rf);
6689	} else {
6690	rq_unpin_lock(rq, rf: &rf);
6691	__balance_callbacks(rq);
6692	raw_spin_rq_unlock_irq(rq);
6693	}
6694	}
6695
6696	void __noreturn do_task_dead(void)
6697	{
6698	/* Causes final put_task_struct in finish_task_switch(): */
6699	set_special_state(TASK_DEAD);
6700
6701	/* Tell freezer to ignore us: */
6702	current->flags |= PF_NOFREEZE;
6703
6704	__schedule(SM_NONE);
6705	BUG();
6706
6707	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6708	for (;;)
6709	cpu_relax();
6710	}
6711
6712	static inline void sched_submit_work(struct task_struct *tsk)
6713	{
6714	static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
6715	unsigned int task_flags;
6716
6717	/*
6718	* Establish LD_WAIT_CONFIG context to ensure none of the code called
6719	* will use a blocking primitive -- which would lead to recursion.
6720	*/
6721	lock_map_acquire_try(&sched_map);
6722
6723	task_flags = tsk->flags;
6724	/*
6725	* If a worker goes to sleep, notify and ask workqueue whether it
6726	* wants to wake up a task to maintain concurrency.
6727	*/
6728	if (task_flags & PF_WQ_WORKER)
6729	wq_worker_sleeping(task: tsk);
6730	else if (task_flags & PF_IO_WORKER)
6731	io_wq_worker_sleeping(tsk);
6732
6733	/*
6734	* spinlock and rwlock must not flush block requests. This will
6735	* deadlock if the callback attempts to acquire a lock which is
6736	* already acquired.
6737	*/
6738	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6739
6740	/*
6741	* If we are going to sleep and we have plugged IO queued,
6742	* make sure to submit it to avoid deadlocks.
6743	*/
6744	blk_flush_plug(plug: tsk->plug, async: true);
6745
6746	lock_map_release(&sched_map);
6747	}
6748
6749	static void sched_update_worker(struct task_struct *tsk)
6750	{
6751	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6752	if (tsk->flags & PF_WQ_WORKER)
6753	wq_worker_running(task: tsk);
6754	else
6755	io_wq_worker_running(tsk);
6756	}
6757	}
6758
6759	static __always_inline void __schedule_loop(unsigned int sched_mode)
6760	{
6761	do {
6762	preempt_disable();
6763	__schedule(sched_mode);
6764	sched_preempt_enable_no_resched();
6765	} while (need_resched());
6766	}
6767
6768	asmlinkage __visible void __sched schedule(void)
6769	{
6770	struct task_struct *tsk = current;
6771
6772	#ifdef CONFIG_RT_MUTEXES
6773	lockdep_assert(!tsk->sched_rt_mutex);
6774	#endif
6775
6776	if (!task_is_running(tsk))
6777	sched_submit_work(tsk);
6778	__schedule_loop(SM_NONE);
6779	sched_update_worker(tsk);
6780	}
6781	EXPORT_SYMBOL(schedule);
6782
6783	/*
6784	* synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6785	* state (have scheduled out non-voluntarily) by making sure that all
6786	* tasks have either left the run queue or have gone into user space.
6787	* As idle tasks do not do either, they must not ever be preempted
6788	* (schedule out non-voluntarily).
6789	*
6790	* schedule_idle() is similar to schedule_preempt_disable() except that it
6791	* never enables preemption because it does not call sched_submit_work().
6792	*/
6793	void __sched schedule_idle(void)
6794	{
6795	/*
6796	* As this skips calling sched_submit_work(), which the idle task does
6797	* regardless because that function is a nop when the task is in a
6798	* TASK_RUNNING state, make sure this isn't used someplace that the
6799	* current task can be in any other state. Note, idle is always in the
6800	* TASK_RUNNING state.
6801	*/
6802	WARN_ON_ONCE(current->__state);
6803	do {
6804	__schedule(SM_NONE);
6805	} while (need_resched());
6806	}
6807
6808	#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6809	asmlinkage __visible void __sched schedule_user(void)
6810	{
6811	/*
6812	* If we come here after a random call to set_need_resched(),
6813	* or we have been woken up remotely but the IPI has not yet arrived,
6814	* we haven't yet exited the RCU idle mode. Do it here manually until
6815	* we find a better solution.
6816	*
6817	* NB: There are buggy callers of this function. Ideally we
6818	* should warn if prev_state != CONTEXT_USER, but that will trigger
6819	* too frequently to make sense yet.
6820	*/
6821	enum ctx_state prev_state = exception_enter();
6822	schedule();
6823	exception_exit(prev_state);
6824	}
6825	#endif
6826
6827	/**
6828	* schedule_preempt_disabled - called with preemption disabled
6829	*
6830	* Returns with preemption disabled. Note: preempt_count must be 1
6831	*/
6832	void __sched schedule_preempt_disabled(void)
6833	{
6834	sched_preempt_enable_no_resched();
6835	schedule();
6836	preempt_disable();
6837	}
6838
6839	#ifdef CONFIG_PREEMPT_RT
6840	void __sched notrace schedule_rtlock(void)
6841	{
6842	__schedule_loop(SM_RTLOCK_WAIT);
6843	}
6844	NOKPROBE_SYMBOL(schedule_rtlock);
6845	#endif
6846
6847	static void __sched notrace preempt_schedule_common(void)
6848	{
6849	do {
6850	/*
6851	* Because the function tracer can trace preempt_count_sub()
6852	* and it also uses preempt_enable/disable_notrace(), if
6853	* NEED_RESCHED is set, the preempt_enable_notrace() called
6854	* by the function tracer will call this function again and
6855	* cause infinite recursion.
6856	*
6857	* Preemption must be disabled here before the function
6858	* tracer can trace. Break up preempt_disable() into two
6859	* calls. One to disable preemption without fear of being
6860	* traced. The other to still record the preemption latency,
6861	* which can also be traced by the function tracer.
6862	*/
6863	preempt_disable_notrace();
6864	preempt_latency_start(val: 1);
6865	__schedule(SM_PREEMPT);
6866	preempt_latency_stop(val: 1);
6867	preempt_enable_no_resched_notrace();
6868
6869	/*
6870	* Check again in case we missed a preemption opportunity
6871	* between schedule and now.
6872	*/
6873	} while (need_resched());
6874	}
6875
6876	#ifdef CONFIG_PREEMPTION
6877	/*
6878	* This is the entry point to schedule() from in-kernel preemption
6879	* off of preempt_enable.
6880	*/
6881	asmlinkage __visible void __sched notrace preempt_schedule(void)
6882	{
6883	/*
6884	* If there is a non-zero preempt_count or interrupts are disabled,
6885	* we do not want to preempt the current task. Just return..
6886	*/
6887	if (likely(!preemptible()))
6888	return;
6889	preempt_schedule_common();
6890	}
6891	NOKPROBE_SYMBOL(preempt_schedule);
6892	EXPORT_SYMBOL(preempt_schedule);
6893
6894	#ifdef CONFIG_PREEMPT_DYNAMIC
6895	#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6896	#ifndef preempt_schedule_dynamic_enabled
6897	#define preempt_schedule_dynamic_enabled preempt_schedule
6898	#define preempt_schedule_dynamic_disabled NULL
6899	#endif
6900	DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6901	EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6902	#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6903	static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6904	void __sched notrace dynamic_preempt_schedule(void)
6905	{
6906	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6907	return;
6908	preempt_schedule();
6909	}
6910	NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6911	EXPORT_SYMBOL(dynamic_preempt_schedule);
6912	#endif
6913	#endif
6914
6915	/**
6916	* preempt_schedule_notrace - preempt_schedule called by tracing
6917	*
6918	* The tracing infrastructure uses preempt_enable_notrace to prevent
6919	* recursion and tracing preempt enabling caused by the tracing
6920	* infrastructure itself. But as tracing can happen in areas coming
6921	* from userspace or just about to enter userspace, a preempt enable
6922	* can occur before user_exit() is called. This will cause the scheduler
6923	* to be called when the system is still in usermode.
6924	*
6925	* To prevent this, the preempt_enable_notrace will use this function
6926	* instead of preempt_schedule() to exit user context if needed before
6927	* calling the scheduler.
6928	*/
6929	asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6930	{
6931	enum ctx_state prev_ctx;
6932
6933	if (likely(!preemptible()))
6934	return;
6935
6936	do {
6937	/*
6938	* Because the function tracer can trace preempt_count_sub()
6939	* and it also uses preempt_enable/disable_notrace(), if
6940	* NEED_RESCHED is set, the preempt_enable_notrace() called
6941	* by the function tracer will call this function again and
6942	* cause infinite recursion.
6943	*
6944	* Preemption must be disabled here before the function
6945	* tracer can trace. Break up preempt_disable() into two
6946	* calls. One to disable preemption without fear of being
6947	* traced. The other to still record the preemption latency,
6948	* which can also be traced by the function tracer.
6949	*/
6950	preempt_disable_notrace();
6951	preempt_latency_start(val: 1);
6952	/*
6953	* Needs preempt disabled in case user_exit() is traced
6954	* and the tracer calls preempt_enable_notrace() causing
6955	* an infinite recursion.
6956	*/
6957	prev_ctx = exception_enter();
6958	__schedule(SM_PREEMPT);
6959	exception_exit(prev_ctx);
6960
6961	preempt_latency_stop(val: 1);
6962	preempt_enable_no_resched_notrace();
6963	} while (need_resched());
6964	}
6965	EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6966
6967	#ifdef CONFIG_PREEMPT_DYNAMIC
6968	#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6969	#ifndef preempt_schedule_notrace_dynamic_enabled
6970	#define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6971	#define preempt_schedule_notrace_dynamic_disabled NULL
6972	#endif
6973	DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6974	EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6975	#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6976	static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6977	void __sched notrace dynamic_preempt_schedule_notrace(void)
6978	{
6979	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6980	return;
6981	preempt_schedule_notrace();
6982	}
6983	NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6984	EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6985	#endif
6986	#endif
6987
6988	#endif /* CONFIG_PREEMPTION */
6989
6990	/*
6991	* This is the entry point to schedule() from kernel preemption
6992	* off of irq context.
6993	* Note, that this is called and return with irqs disabled. This will
6994	* protect us against recursive calling from irq.
6995	*/
6996	asmlinkage __visible void __sched preempt_schedule_irq(void)
6997	{
6998	enum ctx_state prev_state;
6999
7000	/* Catch callers which need to be fixed */
7001	BUG_ON(preempt_count() || !irqs_disabled());
7002
7003	prev_state = exception_enter();
7004
7005	do {
7006	preempt_disable();
7007	local_irq_enable();
7008	__schedule(SM_PREEMPT);
7009	local_irq_disable();
7010	sched_preempt_enable_no_resched();
7011	} while (need_resched());
7012
7013	exception_exit(prev_ctx: prev_state);
7014	}
7015
7016	int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7017	void *key)
7018	{
7019	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7020	return try_to_wake_up(p: curr->private, state: mode, wake_flags);
7021	}
7022	EXPORT_SYMBOL(default_wake_function);
7023
7024	static void __setscheduler_prio(struct task_struct *p, int prio)
7025	{
7026	if (dl_prio(prio))
7027	p->sched_class = &dl_sched_class;
7028	else if (rt_prio(prio))
7029	p->sched_class = &rt_sched_class;
7030	else
7031	p->sched_class = &fair_sched_class;
7032
7033	p->prio = prio;
7034	}
7035
7036	#ifdef CONFIG_RT_MUTEXES
7037
7038	/*
7039	* Would be more useful with typeof()/auto_type but they don't mix with
7040	* bit-fields. Since it's a local thing, use int. Keep the generic sounding
7041	* name such that if someone were to implement this function we get to compare
7042	* notes.
7043	*/
7044	#define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
7045
7046	void rt_mutex_pre_schedule(void)
7047	{
7048	lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
7049	sched_submit_work(current);
7050	}
7051
7052	void rt_mutex_schedule(void)
7053	{
7054	lockdep_assert(current->sched_rt_mutex);
7055	__schedule_loop(SM_NONE);
7056	}
7057
7058	void rt_mutex_post_schedule(void)
7059	{
7060	sched_update_worker(current);
7061	lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
7062	}
7063
7064	static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7065	{
7066	if (pi_task)
7067	prio = min(prio, pi_task->prio);
7068
7069	return prio;
7070	}
7071
7072	static inline int rt_effective_prio(struct task_struct *p, int prio)
7073	{
7074	struct task_struct *pi_task = rt_mutex_get_top_task(p);
7075
7076	return __rt_effective_prio(pi_task, prio);
7077	}
7078
7079	/*
7080	* rt_mutex_setprio - set the current priority of a task
7081	* @p: task to boost
7082	* @pi_task: donor task
7083	*
7084	* This function changes the 'effective' priority of a task. It does
7085	* not touch ->normal_prio like __setscheduler().
7086	*
7087	* Used by the rt_mutex code to implement priority inheritance
7088	* logic. Call site only calls if the priority of the task changed.
7089	*/
7090	void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7091	{
7092	int prio, oldprio, queued, running, queue_flag =
7093	DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7094	const struct sched_class *prev_class;
7095	struct rq_flags rf;
7096	struct rq *rq;
7097
7098	/* XXX used to be waiter->prio, not waiter->task->prio */
7099	prio = __rt_effective_prio(pi_task, prio: p->normal_prio);
7100
7101	/*
7102	* If nothing changed; bail early.
7103	*/
7104	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7105	return;
7106
7107	rq = __task_rq_lock(p, rf: &rf);
7108	update_rq_clock(rq);
7109	/*
7110	* Set under pi_lock && rq->lock, such that the value can be used under
7111	* either lock.
7112	*
7113	* Note that there is loads of tricky to make this pointer cache work
7114	* right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7115	* ensure a task is de-boosted (pi_task is set to NULL) before the
7116	* task is allowed to run again (and can exit). This ensures the pointer
7117	* points to a blocked task -- which guarantees the task is present.
7118	*/
7119	p->pi_top_task = pi_task;
7120
7121	/*
7122	* For FIFO/RR we only need to set prio, if that matches we're done.
7123	*/
7124	if (prio == p->prio && !dl_prio(prio))
7125	goto out_unlock;
7126
7127	/*
7128	* Idle task boosting is a nono in general. There is one
7129	* exception, when PREEMPT_RT and NOHZ is active:
7130	*
7131	* The idle task calls get_next_timer_interrupt() and holds
7132	* the timer wheel base->lock on the CPU and another CPU wants
7133	* to access the timer (probably to cancel it). We can safely
7134	* ignore the boosting request, as the idle CPU runs this code
7135	* with interrupts disabled and will complete the lock
7136	* protected section without being interrupted. So there is no
7137	* real need to boost.
7138	*/
7139	if (unlikely(p == rq->idle)) {
7140	WARN_ON(p != rq->curr);
7141	WARN_ON(p->pi_blocked_on);
7142	goto out_unlock;
7143	}
7144
7145	trace_sched_pi_setprio(tsk: p, pi_task);
7146	oldprio = p->prio;
7147
7148	if (oldprio == prio)
7149	queue_flag &= ~DEQUEUE_MOVE;
7150
7151	prev_class = p->sched_class;
7152	queued = task_on_rq_queued(p);
7153	running = task_current(rq, p);
7154	if (queued)
7155	dequeue_task(rq, p, flags: queue_flag);
7156	if (running)
7157	put_prev_task(rq, prev: p);
7158
7159	/*
7160	* Boosting condition are:
7161	* 1. -rt task is running and holds mutex A
7162	* --> -dl task blocks on mutex A
7163	*
7164	* 2. -dl task is running and holds mutex A
7165	* --> -dl task blocks on mutex A and could preempt the
7166	* running task
7167	*/
7168	if (dl_prio(prio)) {
7169	if (!dl_prio(prio: p->normal_prio) ||
7170	(pi_task && dl_prio(prio: pi_task->prio) &&
7171	dl_entity_preempt(a: &pi_task->dl, b: &p->dl))) {
7172	p->dl.pi_se = pi_task->dl.pi_se;
7173	queue_flag |= ENQUEUE_REPLENISH;
7174	} else {
7175	p->dl.pi_se = &p->dl;
7176	}
7177	} else if (rt_prio(prio)) {
7178	if (dl_prio(prio: oldprio))
7179	p->dl.pi_se = &p->dl;
7180	if (oldprio < prio)
7181	queue_flag |= ENQUEUE_HEAD;
7182	} else {
7183	if (dl_prio(prio: oldprio))
7184	p->dl.pi_se = &p->dl;
7185	if (rt_prio(prio: oldprio))
7186	p->rt.timeout = 0;
7187	}
7188
7189	__setscheduler_prio(p, prio);
7190
7191	if (queued)
7192	enqueue_task(rq, p, flags: queue_flag);
7193	if (running)
7194	set_next_task(rq, next: p);
7195
7196	check_class_changed(rq, p, prev_class, oldprio);
7197	out_unlock:
7198	/* Avoid rq from going away on us: */
7199	preempt_disable();
7200
7201	rq_unpin_lock(rq, rf: &rf);
7202	__balance_callbacks(rq);
7203	raw_spin_rq_unlock(rq);
7204
7205	preempt_enable();
7206	}
7207	#else
7208	static inline int rt_effective_prio(struct task_struct *p, int prio)
7209	{
7210	return prio;
7211	}
7212	#endif
7213
7214	void set_user_nice(struct task_struct *p, long nice)
7215	{
7216	bool queued, running;
7217	struct rq *rq;
7218	int old_prio;
7219
7220	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7221	return;
7222	/*
7223	* We have to be careful, if called from sys_setpriority(),
7224	* the task might be in the middle of scheduling on another CPU.
7225	*/
7226	CLASS(task_rq_lock, rq_guard)(l: p);
7227	rq = rq_guard.rq;
7228
7229	update_rq_clock(rq);
7230
7231	/*
7232	* The RT priorities are set via sched_setscheduler(), but we still
7233	* allow the 'normal' nice value to be set - but as expected
7234	* it won't have any effect on scheduling until the task is
7235	* SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7236	*/
7237	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7238	p->static_prio = NICE_TO_PRIO(nice);
7239	return;
7240	}
7241
7242	queued = task_on_rq_queued(p);
7243	running = task_current(rq, p);
7244	if (queued)
7245	dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7246	if (running)
7247	put_prev_task(rq, prev: p);
7248
7249	p->static_prio = NICE_TO_PRIO(nice);
7250	set_load_weight(p, update_load: true);
7251	old_prio = p->prio;
7252	p->prio = effective_prio(p);
7253
7254	if (queued)
7255	enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7256	if (running)
7257	set_next_task(rq, next: p);
7258
7259	/*
7260	* If the task increased its priority or is running and
7261	* lowered its priority, then reschedule its CPU:
7262	*/
7263	p->sched_class->prio_changed(rq, p, old_prio);
7264	}
7265	EXPORT_SYMBOL(set_user_nice);
7266
7267	/*
7268	* is_nice_reduction - check if nice value is an actual reduction
7269	*
7270	* Similar to can_nice() but does not perform a capability check.
7271	*
7272	* @p: task
7273	* @nice: nice value
7274	*/
7275	static bool is_nice_reduction(const struct task_struct *p, const int nice)
7276	{
7277	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
7278	int nice_rlim = nice_to_rlimit(nice);
7279
7280	return (nice_rlim <= task_rlimit(task: p, RLIMIT_NICE));
7281	}
7282
7283	/*
7284	* can_nice - check if a task can reduce its nice value
7285	* @p: task
7286	* @nice: nice value
7287	*/
7288	int can_nice(const struct task_struct *p, const int nice)
7289	{
7290	return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7291	}
7292
7293	#ifdef __ARCH_WANT_SYS_NICE
7294
7295	/*
7296	* sys_nice - change the priority of the current process.
7297	* @increment: priority increment
7298	*
7299	* sys_setpriority is a more generic, but much slower function that
7300	* does similar things.
7301	*/
7302	SYSCALL_DEFINE1(nice, int, increment)
7303	{
7304	long nice, retval;
7305
7306	/*
7307	* Setpriority might change our priority at the same moment.
7308	* We don't have to worry. Conceptually one call occurs first
7309	* and we have a single winner.
7310	*/
7311	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7312	nice = task_nice(current) + increment;
7313
7314	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7315	if (increment < 0 && !can_nice(current, nice))
7316	return -EPERM;
7317
7318	retval = security_task_setnice(current, nice);
7319	if (retval)
7320	return retval;
7321
7322	set_user_nice(current, nice);
7323	return 0;
7324	}
7325
7326	#endif
7327
7328	/**
7329	* task_prio - return the priority value of a given task.
7330	* @p: the task in question.
7331	*
7332	* Return: The priority value as seen by users in /proc.
7333	*
7334	* sched policy return value kernel prio user prio/nice
7335	*
7336	* normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7337	* fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7338	* deadline -101 -1 0
7339	*/
7340	int task_prio(const struct task_struct *p)
7341	{
7342	return p->prio - MAX_RT_PRIO;
7343	}
7344
7345	/**
7346	* idle_cpu - is a given CPU idle currently?
7347	* @cpu: the processor in question.
7348	*
7349	* Return: 1 if the CPU is currently idle. 0 otherwise.
7350	*/
7351	int idle_cpu(int cpu)
7352	{
7353	struct rq *rq = cpu_rq(cpu);
7354
7355	if (rq->curr != rq->idle)
7356	return 0;
7357
7358	if (rq->nr_running)
7359	return 0;
7360
7361	#ifdef CONFIG_SMP
7362	if (rq->ttwu_pending)
7363	return 0;
7364	#endif
7365
7366	return 1;
7367	}
7368
7369	/**
7370	* available_idle_cpu - is a given CPU idle for enqueuing work.
7371	* @cpu: the CPU in question.
7372	*
7373	* Return: 1 if the CPU is currently idle. 0 otherwise.
7374	*/
7375	int available_idle_cpu(int cpu)
7376	{
7377	if (!idle_cpu(cpu))
7378	return 0;
7379
7380	if (vcpu_is_preempted(cpu))
7381	return 0;
7382
7383	return 1;
7384	}
7385
7386	/**
7387	* idle_task - return the idle task for a given CPU.
7388	* @cpu: the processor in question.
7389	*
7390	* Return: The idle task for the CPU @cpu.
7391	*/
7392	struct task_struct *idle_task(int cpu)
7393	{
7394	return cpu_rq(cpu)->idle;
7395	}
7396
7397	#ifdef CONFIG_SCHED_CORE
7398	int sched_core_idle_cpu(int cpu)
7399	{
7400	struct rq *rq = cpu_rq(cpu);
7401
7402	if (sched_core_enabled(rq) && rq->curr == rq->idle)
7403	return 1;
7404
7405	return idle_cpu(cpu);
7406	}
7407
7408	#endif
7409
7410	#ifdef CONFIG_SMP
7411	/*
7412	* This function computes an effective utilization for the given CPU, to be
7413	* used for frequency selection given the linear relation: f = u * f_max.
7414	*
7415	* The scheduler tracks the following metrics:
7416	*
7417	* cpu_util_{cfs,rt,dl,irq}()
7418	* cpu_bw_dl()
7419	*
7420	* Where the cfs,rt and dl util numbers are tracked with the same metric and
7421	* synchronized windows and are thus directly comparable.
7422	*
7423	* The cfs,rt,dl utilization are the running times measured with rq->clock_task
7424	* which excludes things like IRQ and steal-time. These latter are then accrued
7425	* in the irq utilization.
7426	*
7427	* The DL bandwidth number otoh is not a measured metric but a value computed
7428	* based on the task model parameters and gives the minimal utilization
7429	* required to meet deadlines.
7430	*/
7431	unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7432	enum cpu_util_type type,
7433	struct task_struct *p)
7434	{
7435	unsigned long dl_util, util, irq, max;
7436	struct rq *rq = cpu_rq(cpu);
7437
7438	max = arch_scale_cpu_capacity(cpu);
7439
7440	if (!uclamp_is_used() &&
7441	type == FREQUENCY_UTIL && rt_rq_is_runnable(rt_rq: &rq->rt)) {
7442	return max;
7443	}
7444
7445	/*
7446	* Early check to see if IRQ/steal time saturates the CPU, can be
7447	* because of inaccuracies in how we track these -- see
7448	* update_irq_load_avg().
7449	*/
7450	irq = cpu_util_irq(rq);
7451	if (unlikely(irq >= max))
7452	return max;
7453
7454	/*
7455	* Because the time spend on RT/DL tasks is visible as 'lost' time to
7456	* CFS tasks and we use the same metric to track the effective
7457	* utilization (PELT windows are synchronized) we can directly add them
7458	* to obtain the CPU's actual utilization.
7459	*
7460	* CFS and RT utilization can be boosted or capped, depending on
7461	* utilization clamp constraints requested by currently RUNNABLE
7462	* tasks.
7463	* When there are no CFS RUNNABLE tasks, clamps are released and
7464	* frequency will be gracefully reduced with the utilization decay.
7465	*/
7466	util = util_cfs + cpu_util_rt(rq);
7467	if (type == FREQUENCY_UTIL)
7468	util = uclamp_rq_util_with(rq, util, p);
7469
7470	dl_util = cpu_util_dl(rq);
7471
7472	/*
7473	* For frequency selection we do not make cpu_util_dl() a permanent part
7474	* of this sum because we want to use cpu_bw_dl() later on, but we need
7475	* to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7476	* that we select f_max when there is no idle time.
7477	*
7478	* NOTE: numerical errors or stop class might cause us to not quite hit
7479	* saturation when we should -- something for later.
7480	*/
7481	if (util + dl_util >= max)
7482	return max;
7483
7484	/*
7485	* OTOH, for energy computation we need the estimated running time, so
7486	* include util_dl and ignore dl_bw.
7487	*/
7488	if (type == ENERGY_UTIL)
7489	util += dl_util;
7490
7491	/*
7492	* There is still idle time; further improve the number by using the
7493	* irq metric. Because IRQ/steal time is hidden from the task clock we
7494	* need to scale the task numbers:
7495	*
7496	* max - irq
7497	* U' = irq + --------- * U
7498	* max
7499	*/
7500	util = scale_irq_capacity(util, irq, max);
7501	util += irq;
7502
7503	/*
7504	* Bandwidth required by DEADLINE must always be granted while, for
7505	* FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7506	* to gracefully reduce the frequency when no tasks show up for longer
7507	* periods of time.
7508	*
7509	* Ideally we would like to set bw_dl as min/guaranteed freq and util +
7510	* bw_dl as requested freq. However, cpufreq is not yet ready for such
7511	* an interface. So, we only do the latter for now.
7512	*/
7513	if (type == FREQUENCY_UTIL)
7514	util += cpu_bw_dl(rq);
7515
7516	return min(max, util);
7517	}
7518
7519	unsigned long sched_cpu_util(int cpu)
7520	{
7521	return effective_cpu_util(cpu, util_cfs: cpu_util_cfs(cpu), type: ENERGY_UTIL, NULL);
7522	}
7523	#endif /* CONFIG_SMP */
7524
7525	/**
7526	* find_process_by_pid - find a process with a matching PID value.
7527	* @pid: the pid in question.
7528	*
7529	* The task of @pid, if found. %NULL otherwise.
7530	*/
7531	static struct task_struct *find_process_by_pid(pid_t pid)
7532	{
7533	return pid ? find_task_by_vpid(nr: pid) : current;
7534	}
7535
7536	static struct task_struct *find_get_task(pid_t pid)
7537	{
7538	struct task_struct *p;
7539	guard(rcu)();
7540
7541	p = find_process_by_pid(pid);
7542	if (likely(p))
7543	get_task_struct(t: p);
7544
7545	return p;
7546	}
7547
7548	DEFINE_CLASS(find_get_task, struct task_struct *, if (_T) put_task_struct(_T),
7549	find_get_task(pid), pid_t pid)
7550
7551	/*
7552	* sched_setparam() passes in -1 for its policy, to let the functions
7553	* it calls know not to change it.
7554	*/
7555	#define SETPARAM_POLICY -1
7556
7557	static void __setscheduler_params(struct task_struct *p,
7558	const struct sched_attr *attr)
7559	{
7560	int policy = attr->sched_policy;
7561
7562	if (policy == SETPARAM_POLICY)
7563	policy = p->policy;
7564
7565	p->policy = policy;
7566
7567	if (dl_policy(policy))
7568	__setparam_dl(p, attr);
7569	else if (fair_policy(policy))
7570	p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7571
7572	/*
7573	* __sched_setscheduler() ensures attr->sched_priority == 0 when
7574	* !rt_policy. Always setting this ensures that things like
7575	* getparam()/getattr() don't report silly values for !rt tasks.
7576	*/
7577	p->rt_priority = attr->sched_priority;
7578	p->normal_prio = normal_prio(p);
7579	set_load_weight(p, update_load: true);
7580	}
7581
7582	/*
7583	* Check the target process has a UID that matches the current process's:
7584	*/
7585	static bool check_same_owner(struct task_struct *p)
7586	{
7587	const struct cred *cred = current_cred(), *pcred;
7588	guard(rcu)();
7589
7590	pcred = __task_cred(p);
7591	return (uid_eq(left: cred->euid, right: pcred->euid) ||
7592	uid_eq(left: cred->euid, right: pcred->uid));
7593	}
7594
7595	/*
7596	* Allow unprivileged RT tasks to decrease priority.
7597	* Only issue a capable test if needed and only once to avoid an audit
7598	* event on permitted non-privileged operations:
7599	*/
7600	static int user_check_sched_setscheduler(struct task_struct *p,
7601	const struct sched_attr *attr,
7602	int policy, int reset_on_fork)
7603	{
7604	if (fair_policy(policy)) {
7605	if (attr->sched_nice < task_nice(p) &&
7606	!is_nice_reduction(p, nice: attr->sched_nice))
7607	goto req_priv;
7608	}
7609
7610	if (rt_policy(policy)) {
7611	unsigned long rlim_rtprio = task_rlimit(task: p, RLIMIT_RTPRIO);
7612
7613	/* Can't set/change the rt policy: */
7614	if (policy != p->policy && !rlim_rtprio)
7615	goto req_priv;
7616
7617	/* Can't increase priority: */
7618	if (attr->sched_priority > p->rt_priority &&
7619	attr->sched_priority > rlim_rtprio)
7620	goto req_priv;
7621	}
7622
7623	/*
7624	* Can't set/change SCHED_DEADLINE policy at all for now
7625	* (safest behavior); in the future we would like to allow
7626	* unprivileged DL tasks to increase their relative deadline
7627	* or reduce their runtime (both ways reducing utilization)
7628	*/
7629	if (dl_policy(policy))
7630	goto req_priv;
7631
7632	/*
7633	* Treat SCHED_IDLE as nice 20. Only allow a switch to
7634	* SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7635	*/
7636	if (task_has_idle_policy(p) && !idle_policy(policy)) {
7637	if (!is_nice_reduction(p, nice: task_nice(p)))
7638	goto req_priv;
7639	}
7640
7641	/* Can't change other user's priorities: */
7642	if (!check_same_owner(p))
7643	goto req_priv;
7644
7645	/* Normal users shall not reset the sched_reset_on_fork flag: */
7646	if (p->sched_reset_on_fork && !reset_on_fork)
7647	goto req_priv;
7648
7649	return 0;
7650
7651	req_priv:
7652	if (!capable(CAP_SYS_NICE))
7653	return -EPERM;
7654
7655	return 0;
7656	}
7657
7658	static int __sched_setscheduler(struct task_struct *p,
7659	const struct sched_attr *attr,
7660	bool user, bool pi)
7661	{
7662	int oldpolicy = -1, policy = attr->sched_policy;
7663	int retval, oldprio, newprio, queued, running;
7664	const struct sched_class *prev_class;
7665	struct balance_callback *head;
7666	struct rq_flags rf;
7667	int reset_on_fork;
7668	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7669	struct rq *rq;
7670	bool cpuset_locked = false;
7671
7672	/* The pi code expects interrupts enabled */
7673	BUG_ON(pi && in_interrupt());
7674	recheck:
7675	/* Double check policy once rq lock held: */
7676	if (policy < 0) {
7677	reset_on_fork = p->sched_reset_on_fork;
7678	policy = oldpolicy = p->policy;
7679	} else {
7680	reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7681
7682	if (!valid_policy(policy))
7683	return -EINVAL;
7684	}
7685
7686	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7687	return -EINVAL;
7688
7689	/*
7690	* Valid priorities for SCHED_FIFO and SCHED_RR are
7691	* 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7692	* SCHED_BATCH and SCHED_IDLE is 0.
7693	*/
7694	if (attr->sched_priority > MAX_RT_PRIO-1)
7695	return -EINVAL;
7696	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7697	(rt_policy(policy) != (attr->sched_priority != 0)))
7698	return -EINVAL;
7699
7700	if (user) {
7701	retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7702	if (retval)
7703	return retval;
7704
7705	if (attr->sched_flags & SCHED_FLAG_SUGOV)
7706	return -EINVAL;
7707
7708	retval = security_task_setscheduler(p);
7709	if (retval)
7710	return retval;
7711	}
7712
7713	/* Update task specific "requested" clamps */
7714	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7715	retval = uclamp_validate(p, attr);
7716	if (retval)
7717	return retval;
7718	}
7719
7720	/*
7721	* SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7722	* information.
7723	*/
7724	if (dl_policy(policy) || dl_policy(policy: p->policy)) {
7725	cpuset_locked = true;
7726	cpuset_lock();
7727	}
7728
7729	/*
7730	* Make sure no PI-waiters arrive (or leave) while we are
7731	* changing the priority of the task:
7732	*
7733	* To be able to change p->policy safely, the appropriate
7734	* runqueue lock must be held.
7735	*/
7736	rq = task_rq_lock(p, rf: &rf);
7737	update_rq_clock(rq);
7738
7739	/*
7740	* Changing the policy of the stop threads its a very bad idea:
7741	*/
7742	if (p == rq->stop) {
7743	retval = -EINVAL;
7744	goto unlock;
7745	}
7746
7747	/*
7748	* If not changing anything there's no need to proceed further,
7749	* but store a possible modification of reset_on_fork.
7750	*/
7751	if (unlikely(policy == p->policy)) {
7752	if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7753	goto change;
7754	if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7755	goto change;
7756	if (dl_policy(policy) && dl_param_changed(p, attr))
7757	goto change;
7758	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7759	goto change;
7760
7761	p->sched_reset_on_fork = reset_on_fork;
7762	retval = 0;
7763	goto unlock;
7764	}
7765	change:
7766
7767	if (user) {
7768	#ifdef CONFIG_RT_GROUP_SCHED
7769	/*
7770	* Do not allow realtime tasks into groups that have no runtime
7771	* assigned.
7772	*/
7773	if (rt_bandwidth_enabled() && rt_policy(policy) &&
7774	task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7775	!task_group_is_autogroup(tg: task_group(p))) {
7776	retval = -EPERM;
7777	goto unlock;
7778	}
7779	#endif
7780	#ifdef CONFIG_SMP
7781	if (dl_bandwidth_enabled() && dl_policy(policy) &&
7782	!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7783	cpumask_t *span = rq->rd->span;
7784
7785	/*
7786	* Don't allow tasks with an affinity mask smaller than
7787	* the entire root_domain to become SCHED_DEADLINE. We
7788	* will also fail if there's no bandwidth available.
7789	*/
7790	if (!cpumask_subset(src1p: span, src2p: p->cpus_ptr) ||
7791	rq->rd->dl_bw.bw == 0) {
7792	retval = -EPERM;
7793	goto unlock;
7794	}
7795	}
7796	#endif
7797	}
7798
7799	/* Re-check policy now with rq lock held: */
7800	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7801	policy = oldpolicy = -1;
7802	task_rq_unlock(rq, p, rf: &rf);
7803	if (cpuset_locked)
7804	cpuset_unlock();
7805	goto recheck;
7806	}
7807
7808	/*
7809	* If setscheduling to SCHED_DEADLINE (or changing the parameters
7810	* of a SCHED_DEADLINE task) we need to check if enough bandwidth
7811	* is available.
7812	*/
7813	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7814	retval = -EBUSY;
7815	goto unlock;
7816	}
7817
7818	p->sched_reset_on_fork = reset_on_fork;
7819	oldprio = p->prio;
7820
7821	newprio = __normal_prio(policy, rt_prio: attr->sched_priority, nice: attr->sched_nice);
7822	if (pi) {
7823	/*
7824	* Take priority boosted tasks into account. If the new
7825	* effective priority is unchanged, we just store the new
7826	* normal parameters and do not touch the scheduler class and
7827	* the runqueue. This will be done when the task deboost
7828	* itself.
7829	*/
7830	newprio = rt_effective_prio(p, prio: newprio);
7831	if (newprio == oldprio)
7832	queue_flags &= ~DEQUEUE_MOVE;
7833	}
7834
7835	queued = task_on_rq_queued(p);
7836	running = task_current(rq, p);
7837	if (queued)
7838	dequeue_task(rq, p, flags: queue_flags);
7839	if (running)
7840	put_prev_task(rq, prev: p);
7841
7842	prev_class = p->sched_class;
7843
7844	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7845	__setscheduler_params(p, attr);
7846	__setscheduler_prio(p, prio: newprio);
7847	}
7848	__setscheduler_uclamp(p, attr);
7849
7850	if (queued) {
7851	/*
7852	* We enqueue to tail when the priority of a task is
7853	* increased (user space view).
7854	*/
7855	if (oldprio < p->prio)
7856	queue_flags |= ENQUEUE_HEAD;
7857
7858	enqueue_task(rq, p, flags: queue_flags);
7859	}
7860	if (running)
7861	set_next_task(rq, next: p);
7862
7863	check_class_changed(rq, p, prev_class, oldprio);
7864
7865	/* Avoid rq from going away on us: */
7866	preempt_disable();
7867	head = splice_balance_callbacks(rq);
7868	task_rq_unlock(rq, p, rf: &rf);
7869
7870	if (pi) {
7871	if (cpuset_locked)
7872	cpuset_unlock();
7873	rt_mutex_adjust_pi(p);
7874	}
7875
7876	/* Run balance callbacks after we've adjusted the PI chain: */
7877	balance_callbacks(rq, head);
7878	preempt_enable();
7879
7880	return 0;
7881
7882	unlock:
7883	task_rq_unlock(rq, p, rf: &rf);
7884	if (cpuset_locked)
7885	cpuset_unlock();
7886	return retval;
7887	}
7888
7889	static int _sched_setscheduler(struct task_struct *p, int policy,
7890	const struct sched_param *param, bool check)
7891	{
7892	struct sched_attr attr = {
7893	.sched_policy = policy,
7894	.sched_priority = param->sched_priority,
7895	.sched_nice = PRIO_TO_NICE(p->static_prio),
7896	};
7897
7898	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7899	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7900	attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7901	policy &= ~SCHED_RESET_ON_FORK;
7902	attr.sched_policy = policy;
7903	}
7904
7905	return __sched_setscheduler(p, attr: &attr, user: check, pi: true);
7906	}
7907	/**
7908	* sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7909	* @p: the task in question.
7910	* @policy: new policy.
7911	* @param: structure containing the new RT priority.
7912	*
7913	* Use sched_set_fifo(), read its comment.
7914	*
7915	* Return: 0 on success. An error code otherwise.
7916	*
7917	* NOTE that the task may be already dead.
7918	*/
7919	int sched_setscheduler(struct task_struct *p, int policy,
7920	const struct sched_param *param)
7921	{
7922	return _sched_setscheduler(p, policy, param, check: true);
7923	}
7924
7925	int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7926	{
7927	return __sched_setscheduler(p, attr, user: true, pi: true);
7928	}
7929
7930	int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7931	{
7932	return __sched_setscheduler(p, attr, user: false, pi: true);
7933	}
7934	EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7935
7936	/**
7937	* sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7938	* @p: the task in question.
7939	* @policy: new policy.
7940	* @param: structure containing the new RT priority.
7941	*
7942	* Just like sched_setscheduler, only don't bother checking if the
7943	* current context has permission. For example, this is needed in
7944	* stop_machine(): we create temporary high priority worker threads,
7945	* but our caller might not have that capability.
7946	*
7947	* Return: 0 on success. An error code otherwise.
7948	*/
7949	int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7950	const struct sched_param *param)
7951	{
7952	return _sched_setscheduler(p, policy, param, check: false);
7953	}
7954
7955	/*
7956	* SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7957	* incapable of resource management, which is the one thing an OS really should
7958	* be doing.
7959	*
7960	* This is of course the reason it is limited to privileged users only.
7961	*
7962	* Worse still; it is fundamentally impossible to compose static priority
7963	* workloads. You cannot take two correctly working static prio workloads
7964	* and smash them together and still expect them to work.
7965	*
7966	* For this reason 'all' FIFO tasks the kernel creates are basically at:
7967	*
7968	* MAX_RT_PRIO / 2
7969	*
7970	* The administrator _MUST_ configure the system, the kernel simply doesn't
7971	* know enough information to make a sensible choice.
7972	*/
7973	void sched_set_fifo(struct task_struct *p)
7974	{
7975	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7976	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7977	}
7978	EXPORT_SYMBOL_GPL(sched_set_fifo);
7979
7980	/*
7981	* For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7982	*/
7983	void sched_set_fifo_low(struct task_struct *p)
7984	{
7985	struct sched_param sp = { .sched_priority = 1 };
7986	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7987	}
7988	EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7989
7990	void sched_set_normal(struct task_struct *p, int nice)
7991	{
7992	struct sched_attr attr = {
7993	.sched_policy = SCHED_NORMAL,
7994	.sched_nice = nice,
7995	};
7996	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7997	}
7998	EXPORT_SYMBOL_GPL(sched_set_normal);
7999
8000	static int
8001	do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
8002	{
8003	struct sched_param lparam;
8004
8005	if (!param || pid < 0)
8006	return -EINVAL;
8007	if (copy_from_user(to: &lparam, from: param, n: sizeof(struct sched_param)))
8008	return -EFAULT;
8009
8010	CLASS(find_get_task, p)(pid);
8011	if (!p)
8012	return -ESRCH;
8013
8014	return sched_setscheduler(p, policy, param: &lparam);
8015	}
8016
8017	/*
8018	* Mimics kernel/events/core.c perf_copy_attr().
8019	*/
8020	static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
8021	{
8022	u32 size;
8023	int ret;
8024
8025	/* Zero the full structure, so that a short copy will be nice: */
8026	memset(attr, 0, sizeof(*attr));
8027
8028	ret = get_user(size, &uattr->size);
8029	if (ret)
8030	return ret;
8031
8032	/* ABI compatibility quirk: */
8033	if (!size)
8034	size = SCHED_ATTR_SIZE_VER0;
8035	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8036	goto err_size;
8037
8038	ret = copy_struct_from_user(dst: attr, ksize: sizeof(*attr), src: uattr, usize: size);
8039	if (ret) {
8040	if (ret == -E2BIG)
8041	goto err_size;
8042	return ret;
8043	}
8044
8045	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8046	size < SCHED_ATTR_SIZE_VER1)
8047	return -EINVAL;
8048
8049	/*
8050	* XXX: Do we want to be lenient like existing syscalls; or do we want
8051	* to be strict and return an error on out-of-bounds values?
8052	*/
8053	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8054
8055	return 0;
8056
8057	err_size:
8058	put_user(sizeof(*attr), &uattr->size);
8059	return -E2BIG;
8060	}
8061
8062	static void get_params(struct task_struct *p, struct sched_attr *attr)
8063	{
8064	if (task_has_dl_policy(p))
8065	__getparam_dl(p, attr);
8066	else if (task_has_rt_policy(p))
8067	attr->sched_priority = p->rt_priority;
8068	else
8069	attr->sched_nice = task_nice(p);
8070	}
8071
8072	/**
8073	* sys_sched_setscheduler - set/change the scheduler policy and RT priority
8074	* @pid: the pid in question.
8075	* @policy: new policy.
8076	* @param: structure containing the new RT priority.
8077	*
8078	* Return: 0 on success. An error code otherwise.
8079	*/
8080	SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8081	{
8082	if (policy < 0)
8083	return -EINVAL;
8084
8085	return do_sched_setscheduler(pid, policy, param);
8086	}
8087
8088	/**
8089	* sys_sched_setparam - set/change the RT priority of a thread
8090	* @pid: the pid in question.
8091	* @param: structure containing the new RT priority.
8092	*
8093	* Return: 0 on success. An error code otherwise.
8094	*/
8095	SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8096	{
8097	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8098	}
8099
8100	/**
8101	* sys_sched_setattr - same as above, but with extended sched_attr
8102	* @pid: the pid in question.
8103	* @uattr: structure containing the extended parameters.
8104	* @flags: for future extension.
8105	*/
8106	SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8107	unsigned int, flags)
8108	{
8109	struct sched_attr attr;
8110	int retval;
8111
8112	if (!uattr || pid < 0 || flags)
8113	return -EINVAL;
8114
8115	retval = sched_copy_attr(uattr, attr: &attr);
8116	if (retval)
8117	return retval;
8118
8119	if ((int)attr.sched_policy < 0)
8120	return -EINVAL;
8121	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8122	attr.sched_policy = SETPARAM_POLICY;
8123
8124	CLASS(find_get_task, p)(pid);
8125	if (!p)
8126	return -ESRCH;
8127
8128	if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8129	get_params(p, attr: &attr);
8130
8131	return sched_setattr(p, attr: &attr);
8132	}
8133
8134	/**
8135	* sys_sched_getscheduler - get the policy (scheduling class) of a thread
8136	* @pid: the pid in question.
8137	*
8138	* Return: On success, the policy of the thread. Otherwise, a negative error
8139	* code.
8140	*/
8141	SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8142	{
8143	struct task_struct *p;
8144	int retval;
8145
8146	if (pid < 0)
8147	return -EINVAL;
8148
8149	guard(rcu)();
8150	p = find_process_by_pid(pid);
8151	if (!p)
8152	return -ESRCH;
8153
8154	retval = security_task_getscheduler(p);
8155	if (!retval) {
8156	retval = p->policy;
8157	if (p->sched_reset_on_fork)
8158	retval |= SCHED_RESET_ON_FORK;
8159	}
8160	return retval;
8161	}
8162
8163	/**
8164	* sys_sched_getparam - get the RT priority of a thread
8165	* @pid: the pid in question.
8166	* @param: structure containing the RT priority.
8167	*
8168	* Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8169	* code.
8170	*/
8171	SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8172	{
8173	struct sched_param lp = { .sched_priority = 0 };
8174	struct task_struct *p;
8175	int retval;
8176
8177	if (!param || pid < 0)
8178	return -EINVAL;
8179
8180	scoped_guard (rcu) {
8181	p = find_process_by_pid(pid);
8182	if (!p)
8183	return -ESRCH;
8184
8185	retval = security_task_getscheduler(p);
8186	if (retval)
8187	return retval;
8188
8189	if (task_has_rt_policy(p))
8190	lp.sched_priority = p->rt_priority;
8191	}
8192
8193	/*
8194	* This one might sleep, we cannot do it with a spinlock held ...
8195	*/
8196	return copy_to_user(to: param, from: &lp, n: sizeof(*param)) ? -EFAULT : 0;
8197	}
8198
8199	/*
8200	* Copy the kernel size attribute structure (which might be larger
8201	* than what user-space knows about) to user-space.
8202	*
8203	* Note that all cases are valid: user-space buffer can be larger or
8204	* smaller than the kernel-space buffer. The usual case is that both
8205	* have the same size.
8206	*/
8207	static int
8208	sched_attr_copy_to_user(struct sched_attr __user *uattr,
8209	struct sched_attr *kattr,
8210	unsigned int usize)
8211	{
8212	unsigned int ksize = sizeof(*kattr);
8213
8214	if (!access_ok(uattr, usize))
8215	return -EFAULT;
8216
8217	/*
8218	* sched_getattr() ABI forwards and backwards compatibility:
8219	*
8220	* If usize == ksize then we just copy everything to user-space and all is good.
8221	*
8222	* If usize < ksize then we only copy as much as user-space has space for,
8223	* this keeps ABI compatibility as well. We skip the rest.
8224	*
8225	* If usize > ksize then user-space is using a newer version of the ABI,
8226	* which part the kernel doesn't know about. Just ignore it - tooling can
8227	* detect the kernel's knowledge of attributes from the attr->size value
8228	* which is set to ksize in this case.
8229	*/
8230	kattr->size = min(usize, ksize);
8231
8232	if (copy_to_user(to: uattr, from: kattr, n: kattr->size))
8233	return -EFAULT;
8234
8235	return 0;
8236	}
8237
8238	/**
8239	* sys_sched_getattr - similar to sched_getparam, but with sched_attr
8240	* @pid: the pid in question.
8241	* @uattr: structure containing the extended parameters.
8242	* @usize: sizeof(attr) for fwd/bwd comp.
8243	* @flags: for future extension.
8244	*/
8245	SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8246	unsigned int, usize, unsigned int, flags)
8247	{
8248	struct sched_attr kattr = { };
8249	struct task_struct *p;
8250	int retval;
8251
8252	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8253	usize < SCHED_ATTR_SIZE_VER0 || flags)
8254	return -EINVAL;
8255
8256	scoped_guard (rcu) {
8257	p = find_process_by_pid(pid);
8258	if (!p)
8259	return -ESRCH;
8260
8261	retval = security_task_getscheduler(p);
8262	if (retval)
8263	return retval;
8264
8265	kattr.sched_policy = p->policy;
8266	if (p->sched_reset_on_fork)
8267	kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8268	get_params(p, attr: &kattr);
8269	kattr.sched_flags &= SCHED_FLAG_ALL;
8270
8271	#ifdef CONFIG_UCLAMP_TASK
8272	/*
8273	* This could race with another potential updater, but this is fine
8274	* because it'll correctly read the old or the new value. We don't need
8275	* to guarantee who wins the race as long as it doesn't return garbage.
8276	*/
8277	kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8278	kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8279	#endif
8280	}
8281
8282	return sched_attr_copy_to_user(uattr, kattr: &kattr, usize);
8283	}
8284
8285	#ifdef CONFIG_SMP
8286	int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8287	{
8288	/*
8289	* If the task isn't a deadline task or admission control is
8290	* disabled then we don't care about affinity changes.
8291	*/
8292	if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8293	return 0;
8294
8295	/*
8296	* Since bandwidth control happens on root_domain basis,
8297	* if admission test is enabled, we only admit -deadline
8298	* tasks allowed to run on all the CPUs in the task's
8299	* root_domain.
8300	*/
8301	guard(rcu)();
8302	if (!cpumask_subset(task_rq(p)->rd->span, src2p: mask))
8303	return -EBUSY;
8304
8305	return 0;
8306	}
8307	#endif
8308
8309	static int
8310	__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8311	{
8312	int retval;
8313	cpumask_var_t cpus_allowed, new_mask;
8314
8315	if (!alloc_cpumask_var(mask: &cpus_allowed, GFP_KERNEL))
8316	return -ENOMEM;
8317
8318	if (!alloc_cpumask_var(mask: &new_mask, GFP_KERNEL)) {
8319	retval = -ENOMEM;
8320	goto out_free_cpus_allowed;
8321	}
8322
8323	cpuset_cpus_allowed(p, mask: cpus_allowed);
8324	cpumask_and(dstp: new_mask, src1p: ctx->new_mask, src2p: cpus_allowed);
8325
8326	ctx->new_mask = new_mask;
8327	ctx->flags |= SCA_CHECK;
8328
8329	retval = dl_task_check_affinity(p, mask: new_mask);
8330	if (retval)
8331	goto out_free_new_mask;
8332
8333	retval = __set_cpus_allowed_ptr(p, ctx);
8334	if (retval)
8335	goto out_free_new_mask;
8336
8337	cpuset_cpus_allowed(p, mask: cpus_allowed);
8338	if (!cpumask_subset(src1p: new_mask, src2p: cpus_allowed)) {
8339	/*
8340	* We must have raced with a concurrent cpuset update.
8341	* Just reset the cpumask to the cpuset's cpus_allowed.
8342	*/
8343	cpumask_copy(dstp: new_mask, srcp: cpus_allowed);
8344
8345	/*
8346	* If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8347	* will restore the previous user_cpus_ptr value.
8348	*
8349	* In the unlikely event a previous user_cpus_ptr exists,
8350	* we need to further restrict the mask to what is allowed
8351	* by that old user_cpus_ptr.
8352	*/
8353	if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8354	bool empty = !cpumask_and(dstp: new_mask, src1p: new_mask,
8355	src2p: ctx->user_mask);
8356
8357	if (WARN_ON_ONCE(empty))
8358	cpumask_copy(dstp: new_mask, srcp: cpus_allowed);
8359	}
8360	__set_cpus_allowed_ptr(p, ctx);
8361	retval = -EINVAL;
8362	}
8363
8364	out_free_new_mask:
8365	free_cpumask_var(mask: new_mask);
8366	out_free_cpus_allowed:
8367	free_cpumask_var(mask: cpus_allowed);
8368	return retval;
8369	}
8370
8371	long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8372	{
8373	struct affinity_context ac;
8374	struct cpumask *user_mask;
8375	int retval;
8376
8377	CLASS(find_get_task, p)(pid);
8378	if (!p)
8379	return -ESRCH;
8380
8381	if (p->flags & PF_NO_SETAFFINITY)
8382	return -EINVAL;
8383
8384	if (!check_same_owner(p)) {
8385	guard(rcu)();
8386	if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE))
8387	return -EPERM;
8388	}
8389
8390	retval = security_task_setscheduler(p);
8391	if (retval)
8392	return retval;
8393
8394	/*
8395	* With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8396	* alloc_user_cpus_ptr() returns NULL.
8397	*/
8398	user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8399	if (user_mask) {
8400	cpumask_copy(dstp: user_mask, srcp: in_mask);
8401	} else if (IS_ENABLED(CONFIG_SMP)) {
8402	return -ENOMEM;
8403	}
8404
8405	ac = (struct affinity_context){
8406	.new_mask = in_mask,
8407	.user_mask = user_mask,
8408	.flags = SCA_USER,
8409	};
8410
8411	retval = __sched_setaffinity(p, ctx: &ac);
8412	kfree(objp: ac.user_mask);
8413
8414	return retval;
8415	}
8416
8417	static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8418	struct cpumask *new_mask)
8419	{
8420	if (len < cpumask_size())
8421	cpumask_clear(dstp: new_mask);
8422	else if (len > cpumask_size())
8423	len = cpumask_size();
8424
8425	return copy_from_user(to: new_mask, from: user_mask_ptr, n: len) ? -EFAULT : 0;
8426	}
8427
8428	/**
8429	* sys_sched_setaffinity - set the CPU affinity of a process
8430	* @pid: pid of the process
8431	* @len: length in bytes of the bitmask pointed to by user_mask_ptr
8432	* @user_mask_ptr: user-space pointer to the new CPU mask
8433	*
8434	* Return: 0 on success. An error code otherwise.
8435	*/
8436	SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8437	unsigned long __user *, user_mask_ptr)
8438	{
8439	cpumask_var_t new_mask;
8440	int retval;
8441
8442	if (!alloc_cpumask_var(mask: &new_mask, GFP_KERNEL))
8443	return -ENOMEM;
8444
8445	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8446	if (retval == 0)
8447	retval = sched_setaffinity(pid, in_mask: new_mask);
8448	free_cpumask_var(mask: new_mask);
8449	return retval;
8450	}
8451
8452	long sched_getaffinity(pid_t pid, struct cpumask *mask)
8453	{
8454	struct task_struct *p;
8455	int retval;
8456
8457	guard(rcu)();
8458	p = find_process_by_pid(pid);
8459	if (!p)
8460	return -ESRCH;
8461
8462	retval = security_task_getscheduler(p);
8463	if (retval)
8464	return retval;
8465
8466	guard(raw_spinlock_irqsave)(l: &p->pi_lock);
8467	cpumask_and(dstp: mask, src1p: &p->cpus_mask, cpu_active_mask);
8468
8469	return 0;
8470	}
8471
8472	/**
8473	* sys_sched_getaffinity - get the CPU affinity of a process
8474	* @pid: pid of the process
8475	* @len: length in bytes of the bitmask pointed to by user_mask_ptr
8476	* @user_mask_ptr: user-space pointer to hold the current CPU mask
8477	*
8478	* Return: size of CPU mask copied to user_mask_ptr on success. An
8479	* error code otherwise.
8480	*/
8481	SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8482	unsigned long __user *, user_mask_ptr)
8483	{
8484	int ret;
8485	cpumask_var_t mask;
8486
8487	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8488	return -EINVAL;
8489	if (len & (sizeof(unsigned long)-1))
8490	return -EINVAL;
8491
8492	if (!zalloc_cpumask_var(mask: &mask, GFP_KERNEL))
8493	return -ENOMEM;
8494
8495	ret = sched_getaffinity(pid, mask);
8496	if (ret == 0) {
8497	unsigned int retlen = min(len, cpumask_size());
8498
8499	if (copy_to_user(to: user_mask_ptr, cpumask_bits(mask), n: retlen))
8500	ret = -EFAULT;
8501	else
8502	ret = retlen;
8503	}
8504	free_cpumask_var(mask);
8505
8506	return ret;
8507	}
8508
8509	static void do_sched_yield(void)
8510	{
8511	struct rq_flags rf;
8512	struct rq *rq;
8513
8514	rq = this_rq_lock_irq(rf: &rf);
8515
8516	schedstat_inc(rq->yld_count);
8517	current->sched_class->yield_task(rq);
8518
8519	preempt_disable();
8520	rq_unlock_irq(rq, rf: &rf);
8521	sched_preempt_enable_no_resched();
8522
8523	schedule();
8524	}
8525
8526	/**
8527	* sys_sched_yield - yield the current processor to other threads.
8528	*
8529	* This function yields the current CPU to other tasks. If there are no
8530	* other threads running on this CPU then this function will return.
8531	*
8532	* Return: 0.
8533	*/
8534	SYSCALL_DEFINE0(sched_yield)
8535	{
8536	do_sched_yield();
8537	return 0;
8538	}
8539
8540	#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8541	int __sched __cond_resched(void)
8542	{
8543	if (should_resched(preempt_offset: 0)) {
8544	preempt_schedule_common();
8545	return 1;
8546	}
8547	/*
8548	* In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8549	* whether the current CPU is in an RCU read-side critical section,
8550	* so the tick can report quiescent states even for CPUs looping
8551	* in kernel context. In contrast, in non-preemptible kernels,
8552	* RCU readers leave no in-memory hints, which means that CPU-bound
8553	* processes executing in kernel context might never report an
8554	* RCU quiescent state. Therefore, the following code causes
8555	* cond_resched() to report a quiescent state, but only when RCU
8556	* is in urgent need of one.
8557	*/
8558	#ifndef CONFIG_PREEMPT_RCU
8559	rcu_all_qs();
8560	#endif
8561	return 0;
8562	}
8563	EXPORT_SYMBOL(__cond_resched);
8564	#endif
8565
8566	#ifdef CONFIG_PREEMPT_DYNAMIC
8567	#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8568	#define cond_resched_dynamic_enabled __cond_resched
8569	#define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8570	DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8571	EXPORT_STATIC_CALL_TRAMP(cond_resched);
8572
8573	#define might_resched_dynamic_enabled __cond_resched
8574	#define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8575	DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8576	EXPORT_STATIC_CALL_TRAMP(might_resched);
8577	#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8578	static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8579	int __sched dynamic_cond_resched(void)
8580	{
8581	klp_sched_try_switch();
8582	if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8583	return 0;
8584	return __cond_resched();
8585	}
8586	EXPORT_SYMBOL(dynamic_cond_resched);
8587
8588	static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8589	int __sched dynamic_might_resched(void)
8590	{
8591	if (!static_branch_unlikely(&sk_dynamic_might_resched))
8592	return 0;
8593	return __cond_resched();
8594	}
8595	EXPORT_SYMBOL(dynamic_might_resched);
8596	#endif
8597	#endif
8598
8599	/*
8600	* __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8601	* call schedule, and on return reacquire the lock.
8602	*
8603	* This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8604	* operations here to prevent schedule() from being called twice (once via
8605	* spin_unlock(), once by hand).
8606	*/
8607	int __cond_resched_lock(spinlock_t *lock)
8608	{
8609	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8610	int ret = 0;
8611
8612	lockdep_assert_held(lock);
8613
8614	if (spin_needbreak(lock) || resched) {
8615	spin_unlock(lock);
8616	if (!_cond_resched())
8617	cpu_relax();
8618	ret = 1;
8619	spin_lock(lock);
8620	}
8621	return ret;
8622	}
8623	EXPORT_SYMBOL(__cond_resched_lock);
8624
8625	int __cond_resched_rwlock_read(rwlock_t *lock)
8626	{
8627	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8628	int ret = 0;
8629
8630	lockdep_assert_held_read(lock);
8631
8632	if (rwlock_needbreak(lock) || resched) {
8633	read_unlock(lock);
8634	if (!_cond_resched())
8635	cpu_relax();
8636	ret = 1;
8637	read_lock(lock);
8638	}
8639	return ret;
8640	}
8641	EXPORT_SYMBOL(__cond_resched_rwlock_read);
8642
8643	int __cond_resched_rwlock_write(rwlock_t *lock)
8644	{
8645	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8646	int ret = 0;
8647
8648	lockdep_assert_held_write(lock);
8649
8650	if (rwlock_needbreak(lock) || resched) {
8651	write_unlock(lock);
8652	if (!_cond_resched())
8653	cpu_relax();
8654	ret = 1;
8655	write_lock(lock);
8656	}
8657	return ret;
8658	}
8659	EXPORT_SYMBOL(__cond_resched_rwlock_write);
8660
8661	#ifdef CONFIG_PREEMPT_DYNAMIC
8662
8663	#ifdef CONFIG_GENERIC_ENTRY
8664	#include <linux/entry-common.h>
8665	#endif
8666
8667	/*
8668	* SC:cond_resched
8669	* SC:might_resched
8670	* SC:preempt_schedule
8671	* SC:preempt_schedule_notrace
8672	* SC:irqentry_exit_cond_resched
8673	*
8674	*
8675	* NONE:
8676	* cond_resched <- __cond_resched
8677	* might_resched <- RET0
8678	* preempt_schedule <- NOP
8679	* preempt_schedule_notrace <- NOP
8680	* irqentry_exit_cond_resched <- NOP
8681	*
8682	* VOLUNTARY:
8683	* cond_resched <- __cond_resched
8684	* might_resched <- __cond_resched
8685	* preempt_schedule <- NOP
8686	* preempt_schedule_notrace <- NOP
8687	* irqentry_exit_cond_resched <- NOP
8688	*
8689	* FULL:
8690	* cond_resched <- RET0
8691	* might_resched <- RET0
8692	* preempt_schedule <- preempt_schedule
8693	* preempt_schedule_notrace <- preempt_schedule_notrace
8694	* irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8695	*/
8696
8697	enum {
8698	preempt_dynamic_undefined = -1,
8699	preempt_dynamic_none,
8700	preempt_dynamic_voluntary,
8701	preempt_dynamic_full,
8702	};
8703
8704	int preempt_dynamic_mode = preempt_dynamic_undefined;
8705
8706	int sched_dynamic_mode(const char *str)
8707	{
8708	if (!strcmp(str, "none"))
8709	return preempt_dynamic_none;
8710
8711	if (!strcmp(str, "voluntary"))
8712	return preempt_dynamic_voluntary;
8713
8714	if (!strcmp(str, "full"))
8715	return preempt_dynamic_full;
8716
8717	return -EINVAL;
8718	}
8719
8720	#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8721	#define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8722	#define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8723	#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8724	#define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8725	#define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8726	#else
8727	#error "Unsupported PREEMPT_DYNAMIC mechanism"
8728	#endif
8729
8730	static DEFINE_MUTEX(sched_dynamic_mutex);
8731	static bool klp_override;
8732
8733	static void __sched_dynamic_update(int mode)
8734	{
8735	/*
8736	* Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8737	* the ZERO state, which is invalid.
8738	*/
8739	if (!klp_override)
8740	preempt_dynamic_enable(cond_resched);
8741	preempt_dynamic_enable(might_resched);
8742	preempt_dynamic_enable(preempt_schedule);
8743	preempt_dynamic_enable(preempt_schedule_notrace);
8744	preempt_dynamic_enable(irqentry_exit_cond_resched);
8745
8746	switch (mode) {
8747	case preempt_dynamic_none:
8748	if (!klp_override)
8749	preempt_dynamic_enable(cond_resched);
8750	preempt_dynamic_disable(might_resched);
8751	preempt_dynamic_disable(preempt_schedule);
8752	preempt_dynamic_disable(preempt_schedule_notrace);
8753	preempt_dynamic_disable(irqentry_exit_cond_resched);
8754	if (mode != preempt_dynamic_mode)
8755	pr_info("Dynamic Preempt: none\n");
8756	break;
8757
8758	case preempt_dynamic_voluntary:
8759	if (!klp_override)
8760	preempt_dynamic_enable(cond_resched);
8761	preempt_dynamic_enable(might_resched);
8762	preempt_dynamic_disable(preempt_schedule);
8763	preempt_dynamic_disable(preempt_schedule_notrace);
8764	preempt_dynamic_disable(irqentry_exit_cond_resched);
8765	if (mode != preempt_dynamic_mode)
8766	pr_info("Dynamic Preempt: voluntary\n");
8767	break;
8768
8769	case preempt_dynamic_full:
8770	if (!klp_override)
8771	preempt_dynamic_disable(cond_resched);
8772	preempt_dynamic_disable(might_resched);
8773	preempt_dynamic_enable(preempt_schedule);
8774	preempt_dynamic_enable(preempt_schedule_notrace);
8775	preempt_dynamic_enable(irqentry_exit_cond_resched);
8776	if (mode != preempt_dynamic_mode)
8777	pr_info("Dynamic Preempt: full\n");
8778	break;
8779	}
8780
8781	preempt_dynamic_mode = mode;
8782	}
8783
8784	void sched_dynamic_update(int mode)
8785	{
8786	mutex_lock(&sched_dynamic_mutex);
8787	__sched_dynamic_update(mode);
8788	mutex_unlock(lock: &sched_dynamic_mutex);
8789	}
8790
8791	#ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8792
8793	static int klp_cond_resched(void)
8794	{
8795	__klp_sched_try_switch();
8796	return __cond_resched();
8797	}
8798
8799	void sched_dynamic_klp_enable(void)
8800	{
8801	mutex_lock(&sched_dynamic_mutex);
8802
8803	klp_override = true;
8804	static_call_update(cond_resched, klp_cond_resched);
8805
8806	mutex_unlock(lock: &sched_dynamic_mutex);
8807	}
8808
8809	void sched_dynamic_klp_disable(void)
8810	{
8811	mutex_lock(&sched_dynamic_mutex);
8812
8813	klp_override = false;
8814	__sched_dynamic_update(mode: preempt_dynamic_mode);
8815
8816	mutex_unlock(lock: &sched_dynamic_mutex);
8817	}
8818
8819	#endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8820
8821	static int __init setup_preempt_mode(char *str)
8822	{
8823	int mode = sched_dynamic_mode(str);
8824	if (mode < 0) {
8825	pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8826	return 0;
8827	}
8828
8829	sched_dynamic_update(mode);
8830	return 1;
8831	}
8832	__setup("preempt=", setup_preempt_mode);
8833
8834	static void __init preempt_dynamic_init(void)
8835	{
8836	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8837	if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8838	sched_dynamic_update(mode: preempt_dynamic_none);
8839	} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8840	sched_dynamic_update(mode: preempt_dynamic_voluntary);
8841	} else {
8842	/* Default static call setting, nothing to do */
8843	WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8844	preempt_dynamic_mode = preempt_dynamic_full;
8845	pr_info("Dynamic Preempt: full\n");
8846	}
8847	}
8848	}
8849
8850	#define PREEMPT_MODEL_ACCESSOR(mode) \
8851	bool preempt_model_##mode(void) \
8852	{ \
8853	WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8854	return preempt_dynamic_mode == preempt_dynamic_##mode; \
8855	} \
8856	EXPORT_SYMBOL_GPL(preempt_model_##mode)
8857
8858	PREEMPT_MODEL_ACCESSOR(none);
8859	PREEMPT_MODEL_ACCESSOR(voluntary);
8860	PREEMPT_MODEL_ACCESSOR(full);
8861
8862	#else /* !CONFIG_PREEMPT_DYNAMIC */
8863
8864	static inline void preempt_dynamic_init(void) { }
8865
8866	#endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8867
8868	/**
8869	* yield - yield the current processor to other threads.
8870	*
8871	* Do not ever use this function, there's a 99% chance you're doing it wrong.
8872	*
8873	* The scheduler is at all times free to pick the calling task as the most
8874	* eligible task to run, if removing the yield() call from your code breaks
8875	* it, it's already broken.
8876	*
8877	* Typical broken usage is:
8878	*
8879	* while (!event)
8880	* yield();
8881	*
8882	* where one assumes that yield() will let 'the other' process run that will
8883	* make event true. If the current task is a SCHED_FIFO task that will never
8884	* happen. Never use yield() as a progress guarantee!!
8885	*
8886	* If you want to use yield() to wait for something, use wait_event().
8887	* If you want to use yield() to be 'nice' for others, use cond_resched().
8888	* If you still want to use yield(), do not!
8889	*/
8890	void __sched yield(void)
8891	{
8892	set_current_state(TASK_RUNNING);
8893	do_sched_yield();
8894	}
8895	EXPORT_SYMBOL(yield);
8896
8897	/**
8898	* yield_to - yield the current processor to another thread in
8899	* your thread group, or accelerate that thread toward the
8900	* processor it's on.
8901	* @p: target task
8902	* @preempt: whether task preemption is allowed or not
8903	*
8904	* It's the caller's job to ensure that the target task struct
8905	* can't go away on us before we can do any checks.
8906	*
8907	* Return:
8908	* true (>0) if we indeed boosted the target task.
8909	* false (0) if we failed to boost the target.
8910	* -ESRCH if there's no task to yield to.
8911	*/
8912	int __sched yield_to(struct task_struct *p, bool preempt)
8913	{
8914	struct task_struct *curr = current;
8915	struct rq *rq, *p_rq;
8916	int yielded = 0;
8917
8918	scoped_guard (irqsave) {
8919	rq = this_rq();
8920
8921	again:
8922	p_rq = task_rq(p);
8923	/*
8924	* If we're the only runnable task on the rq and target rq also
8925	* has only one task, there's absolutely no point in yielding.
8926	*/
8927	if (rq->nr_running == 1 && p_rq->nr_running == 1)
8928	return -ESRCH;
8929
8930	guard(double_rq_lock)(lock: rq, lock2: p_rq);
8931	if (task_rq(p) != p_rq)
8932	goto again;
8933
8934	if (!curr->sched_class->yield_to_task)
8935	return 0;
8936
8937	if (curr->sched_class != p->sched_class)
8938	return 0;
8939
8940	if (task_on_cpu(rq: p_rq, p) || !task_is_running(p))
8941	return 0;
8942
8943	yielded = curr->sched_class->yield_to_task(rq, p);
8944	if (yielded) {
8945	schedstat_inc(rq->yld_count);
8946	/*
8947	* Make p's CPU reschedule; pick_next_entity
8948	* takes care of fairness.
8949	*/
8950	if (preempt && rq != p_rq)
8951	resched_curr(rq: p_rq);
8952	}
8953	}
8954
8955	if (yielded)
8956	schedule();
8957
8958	return yielded;
8959	}
8960	EXPORT_SYMBOL_GPL(yield_to);
8961
8962	int io_schedule_prepare(void)
8963	{
8964	int old_iowait = current->in_iowait;
8965
8966	current->in_iowait = 1;
8967	blk_flush_plug(current->plug, async: true);
8968	return old_iowait;
8969	}
8970
8971	void io_schedule_finish(int token)
8972	{
8973	current->in_iowait = token;
8974	}
8975
8976	/*
8977	* This task is about to go to sleep on IO. Increment rq->nr_iowait so
8978	* that process accounting knows that this is a task in IO wait state.
8979	*/
8980	long __sched io_schedule_timeout(long timeout)
8981	{
8982	int token;
8983	long ret;
8984
8985	token = io_schedule_prepare();
8986	ret = schedule_timeout(timeout);
8987	io_schedule_finish(token);
8988
8989	return ret;
8990	}
8991	EXPORT_SYMBOL(io_schedule_timeout);
8992
8993	void __sched io_schedule(void)
8994	{
8995	int token;
8996
8997	token = io_schedule_prepare();
8998	schedule();
8999	io_schedule_finish(token);
9000	}
9001	EXPORT_SYMBOL(io_schedule);
9002
9003	/**
9004	* sys_sched_get_priority_max - return maximum RT priority.
9005	* @policy: scheduling class.
9006	*
9007	* Return: On success, this syscall returns the maximum
9008	* rt_priority that can be used by a given scheduling class.
9009	* On failure, a negative error code is returned.
9010	*/
9011	SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9012	{
9013	int ret = -EINVAL;
9014
9015	switch (policy) {
9016	case SCHED_FIFO:
9017	case SCHED_RR:
9018	ret = MAX_RT_PRIO-1;
9019	break;
9020	case SCHED_DEADLINE:
9021	case SCHED_NORMAL:
9022	case SCHED_BATCH:
9023	case SCHED_IDLE:
9024	ret = 0;
9025	break;
9026	}
9027	return ret;
9028	}
9029
9030	/**
9031	* sys_sched_get_priority_min - return minimum RT priority.
9032	* @policy: scheduling class.
9033	*
9034	* Return: On success, this syscall returns the minimum
9035	* rt_priority that can be used by a given scheduling class.
9036	* On failure, a negative error code is returned.
9037	*/
9038	SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9039	{
9040	int ret = -EINVAL;
9041
9042	switch (policy) {
9043	case SCHED_FIFO:
9044	case SCHED_RR:
9045	ret = 1;
9046	break;
9047	case SCHED_DEADLINE:
9048	case SCHED_NORMAL:
9049	case SCHED_BATCH:
9050	case SCHED_IDLE:
9051	ret = 0;
9052	}
9053	return ret;
9054	}
9055
9056	static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9057	{
9058	unsigned int time_slice = 0;
9059	int retval;
9060
9061	if (pid < 0)
9062	return -EINVAL;
9063
9064	scoped_guard (rcu) {
9065	struct task_struct *p = find_process_by_pid(pid);
9066	if (!p)
9067	return -ESRCH;
9068
9069	retval = security_task_getscheduler(p);
9070	if (retval)
9071	return retval;
9072
9073	scoped_guard (task_rq_lock, p) {
9074	struct rq *rq = scope.rq;
9075	if (p->sched_class->get_rr_interval)
9076	time_slice = p->sched_class->get_rr_interval(rq, p);
9077	}
9078	}
9079
9080	jiffies_to_timespec64(jiffies: time_slice, value: t);
9081	return 0;
9082	}
9083
9084	/**
9085	* sys_sched_rr_get_interval - return the default timeslice of a process.
9086	* @pid: pid of the process.
9087	* @interval: userspace pointer to the timeslice value.
9088	*
9089	* this syscall writes the default timeslice value of a given process
9090	* into the user-space timespec buffer. A value of '0' means infinity.
9091	*
9092	* Return: On success, 0 and the timeslice is in @interval. Otherwise,
9093	* an error code.
9094	*/
9095	SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9096	struct __kernel_timespec __user *, interval)
9097	{
9098	struct timespec64 t;
9099	int retval = sched_rr_get_interval(pid, t: &t);
9100
9101	if (retval == 0)
9102	retval = put_timespec64(ts: &t, uts: interval);
9103
9104	return retval;
9105	}
9106
9107	#ifdef CONFIG_COMPAT_32BIT_TIME
9108	SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9109	struct old_timespec32 __user *, interval)
9110	{
9111	struct timespec64 t;
9112	int retval = sched_rr_get_interval(pid, t: &t);
9113
9114	if (retval == 0)
9115	retval = put_old_timespec32(&t, interval);
9116	return retval;
9117	}
9118	#endif
9119
9120	void sched_show_task(struct task_struct *p)
9121	{
9122	unsigned long free = 0;
9123	int ppid;
9124
9125	if (!try_get_task_stack(tsk: p))
9126	return;
9127
9128	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9129
9130	if (task_is_running(p))
9131	pr_cont(" running task ");
9132	#ifdef CONFIG_DEBUG_STACK_USAGE
9133	free = stack_not_used(p);
9134	#endif
9135	ppid = 0;
9136	rcu_read_lock();
9137	if (pid_alive(p))
9138	ppid = task_pid_nr(rcu_dereference(p->real_parent));
9139	rcu_read_unlock();
9140	pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
9141	free, task_pid_nr(p), task_tgid_nr(p),
9142	ppid, read_task_thread_flags(p));
9143
9144	print_worker_info(KERN_INFO, task: p);
9145	print_stop_info(KERN_INFO, task: p);
9146	show_stack(task: p, NULL, KERN_INFO);
9147	put_task_stack(tsk: p);
9148	}
9149	EXPORT_SYMBOL_GPL(sched_show_task);
9150
9151	static inline bool
9152	state_filter_match(unsigned long state_filter, struct task_struct *p)
9153	{
9154	unsigned int state = READ_ONCE(p->__state);
9155
9156	/* no filter, everything matches */
9157	if (!state_filter)
9158	return true;
9159
9160	/* filter, but doesn't match */
9161	if (!(state & state_filter))
9162	return false;
9163
9164	/*
9165	* When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9166	* TASK_KILLABLE).
9167	*/
9168	if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9169	return false;
9170
9171	return true;
9172	}
9173
9174
9175	void show_state_filter(unsigned int state_filter)
9176	{
9177	struct task_struct *g, *p;
9178
9179	rcu_read_lock();
9180	for_each_process_thread(g, p) {
9181	/*
9182	* reset the NMI-timeout, listing all files on a slow
9183	* console might take a lot of time:
9184	* Also, reset softlockup watchdogs on all CPUs, because
9185	* another CPU might be blocked waiting for us to process
9186	* an IPI.
9187	*/
9188	touch_nmi_watchdog();
9189	touch_all_softlockup_watchdogs();
9190	if (state_filter_match(state_filter, p))
9191	sched_show_task(p);
9192	}
9193
9194	#ifdef CONFIG_SCHED_DEBUG
9195	if (!state_filter)
9196	sysrq_sched_debug_show();
9197	#endif
9198	rcu_read_unlock();
9199	/*
9200	* Only show locks if all tasks are dumped:
9201	*/
9202	if (!state_filter)
9203	debug_show_all_locks();
9204	}
9205
9206	/**
9207	* init_idle - set up an idle thread for a given CPU
9208	* @idle: task in question
9209	* @cpu: CPU the idle task belongs to
9210	*
9211	* NOTE: this function does not set the idle thread's NEED_RESCHED
9212	* flag, to make booting more robust.
9213	*/
9214	void __init init_idle(struct task_struct *idle, int cpu)
9215	{
9216	#ifdef CONFIG_SMP
9217	struct affinity_context ac = (struct affinity_context) {
9218	.new_mask = cpumask_of(cpu),
9219	.flags = 0,
9220	};
9221	#endif
9222	struct rq *rq = cpu_rq(cpu);
9223	unsigned long flags;
9224
9225	__sched_fork(clone_flags: 0, p: idle);
9226
9227	raw_spin_lock_irqsave(&idle->pi_lock, flags);
9228	raw_spin_rq_lock(rq);
9229
9230	idle->__state = TASK_RUNNING;
9231	idle->se.exec_start = sched_clock();
9232	/*
9233	* PF_KTHREAD should already be set at this point; regardless, make it
9234	* look like a proper per-CPU kthread.
9235	*/
9236	idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9237	kthread_set_per_cpu(k: idle, cpu);
9238
9239	#ifdef CONFIG_SMP
9240	/*
9241	* It's possible that init_idle() gets called multiple times on a task,
9242	* in that case do_set_cpus_allowed() will not do the right thing.
9243	*
9244	* And since this is boot we can forgo the serialization.
9245	*/
9246	set_cpus_allowed_common(p: idle, ctx: &ac);
9247	#endif
9248	/*
9249	* We're having a chicken and egg problem, even though we are
9250	* holding rq->lock, the CPU isn't yet set to this CPU so the
9251	* lockdep check in task_group() will fail.
9252	*
9253	* Similar case to sched_fork(). / Alternatively we could
9254	* use task_rq_lock() here and obtain the other rq->lock.
9255	*
9256	* Silence PROVE_RCU
9257	*/
9258	rcu_read_lock();
9259	__set_task_cpu(p: idle, cpu);
9260	rcu_read_unlock();
9261
9262	rq->idle = idle;
9263	rcu_assign_pointer(rq->curr, idle);
9264	idle->on_rq = TASK_ON_RQ_QUEUED;
9265	#ifdef CONFIG_SMP
9266	idle->on_cpu = 1;
9267	#endif
9268	raw_spin_rq_unlock(rq);
9269	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9270
9271	/* Set the preempt count _outside_ the spinlocks! */
9272	init_idle_preempt_count(idle, cpu);
9273
9274	/*
9275	* The idle tasks have their own, simple scheduling class:
9276	*/
9277	idle->sched_class = &idle_sched_class;
9278	ftrace_graph_init_idle_task(t: idle, cpu);
9279	vtime_init_idle(tsk: idle, cpu);
9280	#ifdef CONFIG_SMP
9281	sprintf(buf: idle->comm, fmt: "%s/%d", INIT_TASK_COMM, cpu);
9282	#endif
9283	}
9284
9285	#ifdef CONFIG_SMP
9286
9287	int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9288	const struct cpumask *trial)
9289	{
9290	int ret = 1;
9291
9292	if (cpumask_empty(srcp: cur))
9293	return ret;
9294
9295	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9296
9297	return ret;
9298	}
9299
9300	int task_can_attach(struct task_struct *p)
9301	{
9302	int ret = 0;
9303
9304	/*
9305	* Kthreads which disallow setaffinity shouldn't be moved
9306	* to a new cpuset; we don't want to change their CPU
9307	* affinity and isolating such threads by their set of
9308	* allowed nodes is unnecessary. Thus, cpusets are not
9309	* applicable for such threads. This prevents checking for
9310	* success of set_cpus_allowed_ptr() on all attached tasks
9311	* before cpus_mask may be changed.
9312	*/
9313	if (p->flags & PF_NO_SETAFFINITY)
9314	ret = -EINVAL;
9315
9316	return ret;
9317	}
9318
9319	bool sched_smp_initialized __read_mostly;
9320
9321	#ifdef CONFIG_NUMA_BALANCING
9322	/* Migrate current task p to target_cpu */
9323	int migrate_task_to(struct task_struct *p, int target_cpu)
9324	{
9325	struct migration_arg arg = { p, target_cpu };
9326	int curr_cpu = task_cpu(p);
9327
9328	if (curr_cpu == target_cpu)
9329	return 0;
9330
9331	if (!cpumask_test_cpu(cpu: target_cpu, cpumask: p->cpus_ptr))
9332	return -EINVAL;
9333
9334	/* TODO: This is not properly updating schedstats */
9335
9336	trace_sched_move_numa(tsk: p, src_cpu: curr_cpu, dst_cpu: target_cpu);
9337	return stop_one_cpu(cpu: curr_cpu, fn: migration_cpu_stop, arg: &arg);
9338	}
9339
9340	/*
9341	* Requeue a task on a given node and accurately track the number of NUMA
9342	* tasks on the runqueues
9343	*/
9344	void sched_setnuma(struct task_struct *p, int nid)
9345	{
9346	bool queued, running;
9347	struct rq_flags rf;
9348	struct rq *rq;
9349
9350	rq = task_rq_lock(p, rf: &rf);
9351	queued = task_on_rq_queued(p);
9352	running = task_current(rq, p);
9353
9354	if (queued)
9355	dequeue_task(rq, p, DEQUEUE_SAVE);
9356	if (running)
9357	put_prev_task(rq, prev: p);
9358
9359	p->numa_preferred_nid = nid;
9360
9361	if (queued)
9362	enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9363	if (running)
9364	set_next_task(rq, next: p);
9365	task_rq_unlock(rq, p, rf: &rf);
9366	}
9367	#endif /* CONFIG_NUMA_BALANCING */
9368
9369	#ifdef CONFIG_HOTPLUG_CPU
9370	/*
9371	* Ensure that the idle task is using init_mm right before its CPU goes
9372	* offline.
9373	*/
9374	void idle_task_exit(void)
9375	{
9376	struct mm_struct *mm = current->active_mm;
9377
9378	BUG_ON(cpu_online(smp_processor_id()));
9379	BUG_ON(current != this_rq()->idle);
9380
9381	if (mm != &init_mm) {
9382	switch_mm(prev: mm, next: &init_mm, current);
9383	finish_arch_post_lock_switch();
9384	}
9385
9386	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9387	}
9388
9389	static int __balance_push_cpu_stop(void *arg)
9390	{
9391	struct task_struct *p = arg;
9392	struct rq *rq = this_rq();
9393	struct rq_flags rf;
9394	int cpu;
9395
9396	raw_spin_lock_irq(&p->pi_lock);
9397	rq_lock(rq, rf: &rf);
9398
9399	update_rq_clock(rq);
9400
9401	if (task_rq(p) == rq && task_on_rq_queued(p)) {
9402	cpu = select_fallback_rq(cpu: rq->cpu, p);
9403	rq = __migrate_task(rq, rf: &rf, p, dest_cpu: cpu);
9404	}
9405
9406	rq_unlock(rq, rf: &rf);
9407	raw_spin_unlock_irq(&p->pi_lock);
9408
9409	put_task_struct(t: p);
9410
9411	return 0;
9412	}
9413
9414	static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9415
9416	/*
9417	* Ensure we only run per-cpu kthreads once the CPU goes !active.
9418	*
9419	* This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9420	* effective when the hotplug motion is down.
9421	*/
9422	static void balance_push(struct rq *rq)
9423	{
9424	struct task_struct *push_task = rq->curr;
9425
9426	lockdep_assert_rq_held(rq);
9427
9428	/*
9429	* Ensure the thing is persistent until balance_push_set(.on = false);
9430	*/
9431	rq->balance_callback = &balance_push_callback;
9432
9433	/*
9434	* Only active while going offline and when invoked on the outgoing
9435	* CPU.
9436	*/
9437	if (!cpu_dying(cpu: rq->cpu) || rq != this_rq())
9438	return;
9439
9440	/*
9441	* Both the cpu-hotplug and stop task are in this case and are
9442	* required to complete the hotplug process.
9443	*/
9444	if (kthread_is_per_cpu(k: push_task) ||
9445	is_migration_disabled(p: push_task)) {
9446
9447	/*
9448	* If this is the idle task on the outgoing CPU try to wake
9449	* up the hotplug control thread which might wait for the
9450	* last task to vanish. The rcuwait_active() check is
9451	* accurate here because the waiter is pinned on this CPU
9452	* and can't obviously be running in parallel.
9453	*
9454	* On RT kernels this also has to check whether there are
9455	* pinned and scheduled out tasks on the runqueue. They
9456	* need to leave the migrate disabled section first.
9457	*/
9458	if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9459	rcuwait_active(w: &rq->hotplug_wait)) {
9460	raw_spin_rq_unlock(rq);
9461	rcuwait_wake_up(w: &rq->hotplug_wait);
9462	raw_spin_rq_lock(rq);
9463	}
9464	return;
9465	}
9466
9467	get_task_struct(t: push_task);
9468	/*
9469	* Temporarily drop rq->lock such that we can wake-up the stop task.
9470	* Both preemption and IRQs are still disabled.
9471	*/
9472	preempt_disable();
9473	raw_spin_rq_unlock(rq);
9474	stop_one_cpu_nowait(cpu: rq->cpu, fn: __balance_push_cpu_stop, arg: push_task,
9475	this_cpu_ptr(&push_work));
9476	preempt_enable();
9477	/*
9478	* At this point need_resched() is true and we'll take the loop in
9479	* schedule(). The next pick is obviously going to be the stop task
9480	* which kthread_is_per_cpu() and will push this task away.
9481	*/
9482	raw_spin_rq_lock(rq);
9483	}
9484
9485	static void balance_push_set(int cpu, bool on)
9486	{
9487	struct rq *rq = cpu_rq(cpu);
9488	struct rq_flags rf;
9489
9490	rq_lock_irqsave(rq, rf: &rf);
9491	if (on) {
9492	WARN_ON_ONCE(rq->balance_callback);
9493	rq->balance_callback = &balance_push_callback;
9494	} else if (rq->balance_callback == &balance_push_callback) {
9495	rq->balance_callback = NULL;
9496	}
9497	rq_unlock_irqrestore(rq, rf: &rf);
9498	}
9499
9500	/*
9501	* Invoked from a CPUs hotplug control thread after the CPU has been marked
9502	* inactive. All tasks which are not per CPU kernel threads are either
9503	* pushed off this CPU now via balance_push() or placed on a different CPU
9504	* during wakeup. Wait until the CPU is quiescent.
9505	*/
9506	static void balance_hotplug_wait(void)
9507	{
9508	struct rq *rq = this_rq();
9509
9510	rcuwait_wait_event(&rq->hotplug_wait,
9511	rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9512	TASK_UNINTERRUPTIBLE);
9513	}
9514
9515	#else
9516
9517	static inline void balance_push(struct rq *rq)
9518	{
9519	}
9520
9521	static inline void balance_push_set(int cpu, bool on)
9522	{
9523	}
9524
9525	static inline void balance_hotplug_wait(void)
9526	{
9527	}
9528
9529	#endif /* CONFIG_HOTPLUG_CPU */
9530
9531	void set_rq_online(struct rq *rq)
9532	{
9533	if (!rq->online) {
9534	const struct sched_class *class;
9535
9536	cpumask_set_cpu(cpu: rq->cpu, dstp: rq->rd->online);
9537	rq->online = 1;
9538
9539	for_each_class(class) {
9540	if (class->rq_online)
9541	class->rq_online(rq);
9542	}
9543	}
9544	}
9545
9546	void set_rq_offline(struct rq *rq)
9547	{
9548	if (rq->online) {
9549	const struct sched_class *class;
9550
9551	update_rq_clock(rq);
9552	for_each_class(class) {
9553	if (class->rq_offline)
9554	class->rq_offline(rq);
9555	}
9556
9557	cpumask_clear_cpu(cpu: rq->cpu, dstp: rq->rd->online);
9558	rq->online = 0;
9559	}
9560	}
9561
9562	/*
9563	* used to mark begin/end of suspend/resume:
9564	*/
9565	static int num_cpus_frozen;
9566
9567	/*
9568	* Update cpusets according to cpu_active mask. If cpusets are
9569	* disabled, cpuset_update_active_cpus() becomes a simple wrapper
9570	* around partition_sched_domains().
9571	*
9572	* If we come here as part of a suspend/resume, don't touch cpusets because we
9573	* want to restore it back to its original state upon resume anyway.
9574	*/
9575	static void cpuset_cpu_active(void)
9576	{
9577	if (cpuhp_tasks_frozen) {
9578	/*
9579	* num_cpus_frozen tracks how many CPUs are involved in suspend
9580	* resume sequence. As long as this is not the last online
9581	* operation in the resume sequence, just build a single sched
9582	* domain, ignoring cpusets.
9583	*/
9584	partition_sched_domains(ndoms_new: 1, NULL, NULL);
9585	if (--num_cpus_frozen)
9586	return;
9587	/*
9588	* This is the last CPU online operation. So fall through and
9589	* restore the original sched domains by considering the
9590	* cpuset configurations.
9591	*/
9592	cpuset_force_rebuild();
9593	}
9594	cpuset_update_active_cpus();
9595	}
9596
9597	static int cpuset_cpu_inactive(unsigned int cpu)
9598	{
9599	if (!cpuhp_tasks_frozen) {
9600	int ret = dl_bw_check_overflow(cpu);
9601
9602	if (ret)
9603	return ret;
9604	cpuset_update_active_cpus();
9605	} else {
9606	num_cpus_frozen++;
9607	partition_sched_domains(ndoms_new: 1, NULL, NULL);
9608	}
9609	return 0;
9610	}
9611
9612	int sched_cpu_activate(unsigned int cpu)
9613	{
9614	struct rq *rq = cpu_rq(cpu);
9615	struct rq_flags rf;
9616
9617	/*
9618	* Clear the balance_push callback and prepare to schedule
9619	* regular tasks.
9620	*/
9621	balance_push_set(cpu, on: false);
9622
9623	#ifdef CONFIG_SCHED_SMT
9624	/*
9625	* When going up, increment the number of cores with SMT present.
9626	*/
9627	if (cpumask_weight(srcp: cpu_smt_mask(cpu)) == 2)
9628	static_branch_inc_cpuslocked(&sched_smt_present);
9629	#endif
9630	set_cpu_active(cpu, active: true);
9631
9632	if (sched_smp_initialized) {
9633	sched_update_numa(cpu, online: true);
9634	sched_domains_numa_masks_set(cpu);
9635	cpuset_cpu_active();
9636	}
9637
9638	/*
9639	* Put the rq online, if not already. This happens:
9640	*
9641	* 1) In the early boot process, because we build the real domains
9642	* after all CPUs have been brought up.
9643	*
9644	* 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9645	* domains.
9646	*/
9647	rq_lock_irqsave(rq, rf: &rf);
9648	if (rq->rd) {
9649	BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9650	set_rq_online(rq);
9651	}
9652	rq_unlock_irqrestore(rq, rf: &rf);
9653
9654	return 0;
9655	}
9656
9657	int sched_cpu_deactivate(unsigned int cpu)
9658	{
9659	struct rq *rq = cpu_rq(cpu);
9660	struct rq_flags rf;
9661	int ret;
9662
9663	/*
9664	* Remove CPU from nohz.idle_cpus_mask to prevent participating in
9665	* load balancing when not active
9666	*/
9667	nohz_balance_exit_idle(rq);
9668
9669	set_cpu_active(cpu, active: false);
9670
9671	/*
9672	* From this point forward, this CPU will refuse to run any task that
9673	* is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9674	* push those tasks away until this gets cleared, see
9675	* sched_cpu_dying().
9676	*/
9677	balance_push_set(cpu, on: true);
9678
9679	/*
9680	* We've cleared cpu_active_mask / set balance_push, wait for all
9681	* preempt-disabled and RCU users of this state to go away such that
9682	* all new such users will observe it.
9683	*
9684	* Specifically, we rely on ttwu to no longer target this CPU, see
9685	* ttwu_queue_cond() and is_cpu_allowed().
9686	*
9687	* Do sync before park smpboot threads to take care the rcu boost case.
9688	*/
9689	synchronize_rcu();
9690
9691	rq_lock_irqsave(rq, rf: &rf);
9692	if (rq->rd) {
9693	BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9694	set_rq_offline(rq);
9695	}
9696	rq_unlock_irqrestore(rq, rf: &rf);
9697
9698	#ifdef CONFIG_SCHED_SMT
9699	/*
9700	* When going down, decrement the number of cores with SMT present.
9701	*/
9702	if (cpumask_weight(srcp: cpu_smt_mask(cpu)) == 2)
9703	static_branch_dec_cpuslocked(&sched_smt_present);
9704
9705	sched_core_cpu_deactivate(cpu);
9706	#endif
9707
9708	if (!sched_smp_initialized)
9709	return 0;
9710
9711	sched_update_numa(cpu, online: false);
9712	ret = cpuset_cpu_inactive(cpu);
9713	if (ret) {
9714	balance_push_set(cpu, on: false);
9715	set_cpu_active(cpu, active: true);
9716	sched_update_numa(cpu, online: true);
9717	return ret;
9718	}
9719	sched_domains_numa_masks_clear(cpu);
9720	return 0;
9721	}
9722
9723	static void sched_rq_cpu_starting(unsigned int cpu)
9724	{
9725	struct rq *rq = cpu_rq(cpu);
9726
9727	rq->calc_load_update = calc_load_update;
9728	update_max_interval();
9729	}
9730
9731	int sched_cpu_starting(unsigned int cpu)
9732	{
9733	sched_core_cpu_starting(cpu);
9734	sched_rq_cpu_starting(cpu);
9735	sched_tick_start(cpu);
9736	return 0;
9737	}
9738
9739	#ifdef CONFIG_HOTPLUG_CPU
9740
9741	/*
9742	* Invoked immediately before the stopper thread is invoked to bring the
9743	* CPU down completely. At this point all per CPU kthreads except the
9744	* hotplug thread (current) and the stopper thread (inactive) have been
9745	* either parked or have been unbound from the outgoing CPU. Ensure that
9746	* any of those which might be on the way out are gone.
9747	*
9748	* If after this point a bound task is being woken on this CPU then the
9749	* responsible hotplug callback has failed to do it's job.
9750	* sched_cpu_dying() will catch it with the appropriate fireworks.
9751	*/
9752	int sched_cpu_wait_empty(unsigned int cpu)
9753	{
9754	balance_hotplug_wait();
9755	return 0;
9756	}
9757
9758	/*
9759	* Since this CPU is going 'away' for a while, fold any nr_active delta we
9760	* might have. Called from the CPU stopper task after ensuring that the
9761	* stopper is the last running task on the CPU, so nr_active count is
9762	* stable. We need to take the teardown thread which is calling this into
9763	* account, so we hand in adjust = 1 to the load calculation.
9764	*
9765	* Also see the comment "Global load-average calculations".
9766	*/
9767	static void calc_load_migrate(struct rq *rq)
9768	{
9769	long delta = calc_load_fold_active(this_rq: rq, adjust: 1);
9770
9771	if (delta)
9772	atomic_long_add(i: delta, v: &calc_load_tasks);
9773	}
9774
9775	static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9776	{
9777	struct task_struct *g, *p;
9778	int cpu = cpu_of(rq);
9779
9780	lockdep_assert_rq_held(rq);
9781
9782	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9783	for_each_process_thread(g, p) {
9784	if (task_cpu(p) != cpu)
9785	continue;
9786
9787	if (!task_on_rq_queued(p))
9788	continue;
9789
9790	printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9791	}
9792	}
9793
9794	int sched_cpu_dying(unsigned int cpu)
9795	{
9796	struct rq *rq = cpu_rq(cpu);
9797	struct rq_flags rf;
9798
9799	/* Handle pending wakeups and then migrate everything off */
9800	sched_tick_stop(cpu);
9801
9802	rq_lock_irqsave(rq, rf: &rf);
9803	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9804	WARN(true, "Dying CPU not properly vacated!");
9805	dump_rq_tasks(rq, KERN_WARNING);
9806	}
9807	rq_unlock_irqrestore(rq, rf: &rf);
9808
9809	calc_load_migrate(rq);
9810	update_max_interval();
9811	hrtick_clear(rq);
9812	sched_core_cpu_dying(cpu);
9813	return 0;
9814	}
9815	#endif
9816
9817	void __init sched_init_smp(void)
9818	{
9819	sched_init_numa(NUMA_NO_NODE);
9820
9821	/*
9822	* There's no userspace yet to cause hotplug operations; hence all the
9823	* CPU masks are stable and all blatant races in the below code cannot
9824	* happen.
9825	*/
9826	mutex_lock(&sched_domains_mutex);
9827	sched_init_domains(cpu_active_mask);
9828	mutex_unlock(lock: &sched_domains_mutex);
9829
9830	/* Move init over to a non-isolated CPU */
9831	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(type: HK_TYPE_DOMAIN)) < 0)
9832	BUG();
9833	current->flags &= ~PF_NO_SETAFFINITY;
9834	sched_init_granularity();
9835
9836	init_sched_rt_class();
9837	init_sched_dl_class();
9838
9839	sched_smp_initialized = true;
9840	}
9841
9842	static int __init migration_init(void)
9843	{
9844	sched_cpu_starting(smp_processor_id());
9845	return 0;
9846	}
9847	early_initcall(migration_init);
9848
9849	#else
9850	void __init sched_init_smp(void)
9851	{
9852	sched_init_granularity();
9853	}
9854	#endif /* CONFIG_SMP */
9855
9856	int in_sched_functions(unsigned long addr)
9857	{
9858	return in_lock_functions(addr) ||
9859	(addr >= (unsigned long)__sched_text_start
9860	&& addr < (unsigned long)__sched_text_end);
9861	}
9862
9863	#ifdef CONFIG_CGROUP_SCHED
9864	/*
9865	* Default task group.
9866	* Every task in system belongs to this group at bootup.
9867	*/
9868	struct task_group root_task_group;
9869	LIST_HEAD(task_groups);
9870
9871	/* Cacheline aligned slab cache for task_group */
9872	static struct kmem_cache *task_group_cache __ro_after_init;
9873	#endif
9874
9875	void __init sched_init(void)
9876	{
9877	unsigned long ptr = 0;
9878	int i;
9879
9880	/* Make sure the linker didn't screw up */
9881	BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9882	&fair_sched_class != &rt_sched_class + 1 ||
9883	&rt_sched_class != &dl_sched_class + 1);
9884	#ifdef CONFIG_SMP
9885	BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9886	#endif
9887
9888	wait_bit_init();
9889
9890	#ifdef CONFIG_FAIR_GROUP_SCHED
9891	ptr += 2 * nr_cpu_ids * sizeof(void **);
9892	#endif
9893	#ifdef CONFIG_RT_GROUP_SCHED
9894	ptr += 2 * nr_cpu_ids * sizeof(void **);
9895	#endif
9896	if (ptr) {
9897	ptr = (unsigned long)kzalloc(size: ptr, GFP_NOWAIT);
9898
9899	#ifdef CONFIG_FAIR_GROUP_SCHED
9900	root_task_group.se = (struct sched_entity **)ptr;
9901	ptr += nr_cpu_ids * sizeof(void **);
9902
9903	root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9904	ptr += nr_cpu_ids * sizeof(void **);
9905
9906	root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9907	init_cfs_bandwidth(cfs_b: &root_task_group.cfs_bandwidth, NULL);
9908	#endif /* CONFIG_FAIR_GROUP_SCHED */
9909	#ifdef CONFIG_RT_GROUP_SCHED
9910	root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9911	ptr += nr_cpu_ids * sizeof(void **);
9912
9913	root_task_group.rt_rq = (struct rt_rq **)ptr;
9914	ptr += nr_cpu_ids * sizeof(void **);
9915
9916	#endif /* CONFIG_RT_GROUP_SCHED */
9917	}
9918
9919	init_rt_bandwidth(rt_b: &def_rt_bandwidth, period: global_rt_period(), runtime: global_rt_runtime());
9920
9921	#ifdef CONFIG_SMP
9922	init_defrootdomain();
9923	#endif
9924
9925	#ifdef CONFIG_RT_GROUP_SCHED
9926	init_rt_bandwidth(rt_b: &root_task_group.rt_bandwidth,
9927	period: global_rt_period(), runtime: global_rt_runtime());
9928	#endif /* CONFIG_RT_GROUP_SCHED */
9929
9930	#ifdef CONFIG_CGROUP_SCHED
9931	task_group_cache = KMEM_CACHE(task_group, 0);
9932
9933	list_add(new: &root_task_group.list, head: &task_groups);
9934	INIT_LIST_HEAD(list: &root_task_group.children);
9935	INIT_LIST_HEAD(list: &root_task_group.siblings);
9936	autogroup_init(init_task: &init_task);
9937	#endif /* CONFIG_CGROUP_SCHED */
9938
9939	for_each_possible_cpu(i) {
9940	struct rq *rq;
9941
9942	rq = cpu_rq(i);
9943	raw_spin_lock_init(&rq->__lock);
9944	rq->nr_running = 0;
9945	rq->calc_load_active = 0;
9946	rq->calc_load_update = jiffies + LOAD_FREQ;
9947	init_cfs_rq(cfs_rq: &rq->cfs);
9948	init_rt_rq(rt_rq: &rq->rt);
9949	init_dl_rq(dl_rq: &rq->dl);
9950	#ifdef CONFIG_FAIR_GROUP_SCHED
9951	INIT_LIST_HEAD(list: &rq->leaf_cfs_rq_list);
9952	rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9953	/*
9954	* How much CPU bandwidth does root_task_group get?
9955	*
9956	* In case of task-groups formed thr' the cgroup filesystem, it
9957	* gets 100% of the CPU resources in the system. This overall
9958	* system CPU resource is divided among the tasks of
9959	* root_task_group and its child task-groups in a fair manner,
9960	* based on each entity's (task or task-group's) weight
9961	* (se->load.weight).
9962	*
9963	* In other words, if root_task_group has 10 tasks of weight
9964	* 1024) and two child groups A0 and A1 (of weight 1024 each),
9965	* then A0's share of the CPU resource is:
9966	*
9967	* A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9968	*
9969	* We achieve this by letting root_task_group's tasks sit
9970	* directly in rq->cfs (i.e root_task_group->se[] = NULL).
9971	*/
9972	init_tg_cfs_entry(tg: &root_task_group, cfs_rq: &rq->cfs, NULL, cpu: i, NULL);
9973	#endif /* CONFIG_FAIR_GROUP_SCHED */
9974
9975	rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9976	#ifdef CONFIG_RT_GROUP_SCHED
9977	init_tg_rt_entry(tg: &root_task_group, rt_rq: &rq->rt, NULL, cpu: i, NULL);
9978	#endif
9979	#ifdef CONFIG_SMP
9980	rq->sd = NULL;
9981	rq->rd = NULL;
9982	rq->cpu_capacity = SCHED_CAPACITY_SCALE;
9983	rq->balance_callback = &balance_push_callback;
9984	rq->active_balance = 0;
9985	rq->next_balance = jiffies;
9986	rq->push_cpu = 0;
9987	rq->cpu = i;
9988	rq->online = 0;
9989	rq->idle_stamp = 0;
9990	rq->avg_idle = 2*sysctl_sched_migration_cost;
9991	rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9992
9993	INIT_LIST_HEAD(list: &rq->cfs_tasks);
9994
9995	rq_attach_root(rq, rd: &def_root_domain);
9996	#ifdef CONFIG_NO_HZ_COMMON
9997	rq->last_blocked_load_update_tick = jiffies;
9998	atomic_set(v: &rq->nohz_flags, i: 0);
9999
10000	INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10001	#endif
10002	#ifdef CONFIG_HOTPLUG_CPU
10003	rcuwait_init(w: &rq->hotplug_wait);
10004	#endif
10005	#endif /* CONFIG_SMP */
10006	hrtick_rq_init(rq);
10007	atomic_set(v: &rq->nr_iowait, i: 0);
10008
10009	#ifdef CONFIG_SCHED_CORE
10010	rq->core = rq;
10011	rq->core_pick = NULL;
10012	rq->core_enabled = 0;
10013	rq->core_tree = RB_ROOT;
10014	rq->core_forceidle_count = 0;
10015	rq->core_forceidle_occupation = 0;
10016	rq->core_forceidle_start = 0;
10017
10018	rq->core_cookie = 0UL;
10019	#endif
10020	zalloc_cpumask_var_node(mask: &rq->scratch_mask, GFP_KERNEL, cpu_to_node(cpu: i));
10021	}
10022
10023	set_load_weight(p: &init_task, update_load: false);
10024
10025	/*
10026	* The boot idle thread does lazy MMU switching as well:
10027	*/
10028	mmgrab_lazy_tlb(mm: &init_mm);
10029	enter_lazy_tlb(mm: &init_mm, current);
10030
10031	/*
10032	* The idle task doesn't need the kthread struct to function, but it
10033	* is dressed up as a per-CPU kthread and thus needs to play the part
10034	* if we want to avoid special-casing it in code that deals with per-CPU
10035	* kthreads.
10036	*/
10037	WARN_ON(!set_kthread_struct(current));
10038
10039	/*
10040	* Make us the idle thread. Technically, schedule() should not be
10041	* called from this thread, however somewhere below it might be,
10042	* but because we are the idle thread, we just pick up running again
10043	* when this runqueue becomes "idle".
10044	*/
10045	init_idle(current, smp_processor_id());
10046
10047	calc_load_update = jiffies + LOAD_FREQ;
10048
10049	#ifdef CONFIG_SMP
10050	idle_thread_set_boot_cpu();
10051	balance_push_set(smp_processor_id(), on: false);
10052	#endif
10053	init_sched_fair_class();
10054
10055	psi_init();
10056
10057	init_uclamp();
10058
10059	preempt_dynamic_init();
10060
10061	scheduler_running = 1;
10062	}
10063
10064	#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10065
10066	void __might_sleep(const char *file, int line)
10067	{
10068	unsigned int state = get_current_state();
10069	/*
10070	* Blocking primitives will set (and therefore destroy) current->state,
10071	* since we will exit with TASK_RUNNING make sure we enter with it,
10072	* otherwise we will destroy state.
10073	*/
10074	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10075	"do not call blocking ops when !TASK_RUNNING; "
10076	"state=%x set at [<%p>] %pS\n", state,
10077	(void *)current->task_state_change,
10078	(void *)current->task_state_change);
10079
10080	__might_resched(file, line, offsets: 0);
10081	}
10082	EXPORT_SYMBOL(__might_sleep);
10083
10084	static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10085	{
10086	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10087	return;
10088
10089	if (preempt_count() == preempt_offset)
10090	return;
10091
10092	pr_err("Preemption disabled at:");
10093	print_ip_sym(KERN_ERR, ip);
10094	}
10095
10096	static inline bool resched_offsets_ok(unsigned int offsets)
10097	{
10098	unsigned int nested = preempt_count();
10099
10100	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10101
10102	return nested == offsets;
10103	}
10104
10105	void __might_resched(const char *file, int line, unsigned int offsets)
10106	{
10107	/* Ratelimiting timestamp: */
10108	static unsigned long prev_jiffy;
10109
10110	unsigned long preempt_disable_ip;
10111
10112	/* WARN_ON_ONCE() by default, no rate limit required: */
10113	rcu_sleep_check();
10114
10115	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10116	!is_idle_task(current) && !current->non_block_count) ||
10117	system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10118	oops_in_progress)
10119	return;
10120
10121	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10122	return;
10123	prev_jiffy = jiffies;
10124
10125	/* Save this before calling printk(), since that will clobber it: */
10126	preempt_disable_ip = get_preempt_disable_ip(current);
10127
10128	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10129	file, line);
10130	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10131	in_atomic(), irqs_disabled(), current->non_block_count,
10132	current->pid, current->comm);
10133	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10134	offsets & MIGHT_RESCHED_PREEMPT_MASK);
10135
10136	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10137	pr_err("RCU nest depth: %d, expected: %u\n",
10138	rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10139	}
10140
10141	if (task_stack_end_corrupted(current))
10142	pr_emerg("Thread overran stack, or stack corrupted\n");
10143
10144	debug_show_held_locks(current);
10145	if (irqs_disabled())
10146	print_irqtrace_events(current);
10147
10148	print_preempt_disable_ip(preempt_offset: offsets & MIGHT_RESCHED_PREEMPT_MASK,
10149	ip: preempt_disable_ip);
10150
10151	dump_stack();
10152	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10153	}
10154	EXPORT_SYMBOL(__might_resched);
10155
10156	void __cant_sleep(const char *file, int line, int preempt_offset)
10157	{
10158	static unsigned long prev_jiffy;
10159
10160	if (irqs_disabled())
10161	return;
10162
10163	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10164	return;
10165
10166	if (preempt_count() > preempt_offset)
10167	return;
10168
10169	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10170	return;
10171	prev_jiffy = jiffies;
10172
10173	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10174	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10175	in_atomic(), irqs_disabled(),
10176	current->pid, current->comm);
10177
10178	debug_show_held_locks(current);
10179	dump_stack();
10180	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10181	}
10182	EXPORT_SYMBOL_GPL(__cant_sleep);
10183
10184	#ifdef CONFIG_SMP
10185	void __cant_migrate(const char *file, int line)
10186	{
10187	static unsigned long prev_jiffy;
10188
10189	if (irqs_disabled())
10190	return;
10191
10192	if (is_migration_disabled(current))
10193	return;
10194
10195	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10196	return;
10197
10198	if (preempt_count() > 0)
10199	return;
10200
10201	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10202	return;
10203	prev_jiffy = jiffies;
10204
10205	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10206	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10207	in_atomic(), irqs_disabled(), is_migration_disabled(current),
10208	current->pid, current->comm);
10209
10210	debug_show_held_locks(current);
10211	dump_stack();
10212	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10213	}
10214	EXPORT_SYMBOL_GPL(__cant_migrate);
10215	#endif
10216	#endif
10217
10218	#ifdef CONFIG_MAGIC_SYSRQ
10219	void normalize_rt_tasks(void)
10220	{
10221	struct task_struct *g, *p;
10222	struct sched_attr attr = {
10223	.sched_policy = SCHED_NORMAL,
10224	};
10225
10226	read_lock(&tasklist_lock);
10227	for_each_process_thread(g, p) {
10228	/*
10229	* Only normalize user tasks:
10230	*/
10231	if (p->flags & PF_KTHREAD)
10232	continue;
10233
10234	p->se.exec_start = 0;
10235	schedstat_set(p->stats.wait_start, 0);
10236	schedstat_set(p->stats.sleep_start, 0);
10237	schedstat_set(p->stats.block_start, 0);
10238
10239	if (!dl_task(p) && !rt_task(p)) {
10240	/*
10241	* Renice negative nice level userspace
10242	* tasks back to 0:
10243	*/
10244	if (task_nice(p) < 0)
10245	set_user_nice(p, 0);
10246	continue;
10247	}
10248
10249	__sched_setscheduler(p, attr: &attr, user: false, pi: false);
10250	}
10251	read_unlock(&tasklist_lock);
10252	}
10253
10254	#endif /* CONFIG_MAGIC_SYSRQ */
10255
10256	#if defined(CONFIG_KGDB_KDB)
10257	/*
10258	* These functions are only useful for kdb.
10259	*
10260	* They can only be called when the whole system has been
10261	* stopped - every CPU needs to be quiescent, and no scheduling
10262	* activity can take place. Using them for anything else would
10263	* be a serious bug, and as a result, they aren't even visible
10264	* under any other configuration.
10265	*/
10266
10267	/**
10268	* curr_task - return the current task for a given CPU.
10269	* @cpu: the processor in question.
10270	*
10271	* ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10272	*
10273	* Return: The current task for @cpu.
10274	*/
10275	struct task_struct *curr_task(int cpu)
10276	{
10277	return cpu_curr(cpu);
10278	}
10279
10280	#endif /* defined(CONFIG_KGDB_KDB) */
10281
10282	#ifdef CONFIG_CGROUP_SCHED
10283	/* task_group_lock serializes the addition/removal of task groups */
10284	static DEFINE_SPINLOCK(task_group_lock);
10285
10286	static inline void alloc_uclamp_sched_group(struct task_group *tg,
10287	struct task_group *parent)
10288	{
10289	#ifdef CONFIG_UCLAMP_TASK_GROUP
10290	enum uclamp_id clamp_id;
10291
10292	for_each_clamp_id(clamp_id) {
10293	uclamp_se_set(uc_se: &tg->uclamp_req[clamp_id],
10294	value: uclamp_none(clamp_id), user_defined: false);
10295	tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10296	}
10297	#endif
10298	}
10299
10300	static void sched_free_group(struct task_group *tg)
10301	{
10302	free_fair_sched_group(tg);
10303	free_rt_sched_group(tg);
10304	autogroup_free(tg);
10305	kmem_cache_free(s: task_group_cache, objp: tg);
10306	}
10307
10308	static void sched_free_group_rcu(struct rcu_head *rcu)
10309	{
10310	sched_free_group(container_of(rcu, struct task_group, rcu));
10311	}
10312
10313	static void sched_unregister_group(struct task_group *tg)
10314	{
10315	unregister_fair_sched_group(tg);
10316	unregister_rt_sched_group(tg);
10317	/*
10318	* We have to wait for yet another RCU grace period to expire, as
10319	* print_cfs_stats() might run concurrently.
10320	*/
10321	call_rcu(head: &tg->rcu, func: sched_free_group_rcu);
10322	}
10323
10324	/* allocate runqueue etc for a new task group */
10325	struct task_group *sched_create_group(struct task_group *parent)
10326	{
10327	struct task_group *tg;
10328
10329	tg = kmem_cache_alloc(cachep: task_group_cache, GFP_KERNEL | __GFP_ZERO);
10330	if (!tg)
10331	return ERR_PTR(error: -ENOMEM);
10332
10333	if (!alloc_fair_sched_group(tg, parent))
10334	goto err;
10335
10336	if (!alloc_rt_sched_group(tg, parent))
10337	goto err;
10338
10339	alloc_uclamp_sched_group(tg, parent);
10340
10341	return tg;
10342
10343	err:
10344	sched_free_group(tg);
10345	return ERR_PTR(error: -ENOMEM);
10346	}
10347
10348	void sched_online_group(struct task_group *tg, struct task_group *parent)
10349	{
10350	unsigned long flags;
10351
10352	spin_lock_irqsave(&task_group_lock, flags);
10353	list_add_rcu(new: &tg->list, head: &task_groups);
10354
10355	/* Root should already exist: */
10356	WARN_ON(!parent);
10357
10358	tg->parent = parent;
10359	INIT_LIST_HEAD(list: &tg->children);
10360	list_add_rcu(new: &tg->siblings, head: &parent->children);
10361	spin_unlock_irqrestore(lock: &task_group_lock, flags);
10362
10363	online_fair_sched_group(tg);
10364	}
10365
10366	/* rcu callback to free various structures associated with a task group */
10367	static void sched_unregister_group_rcu(struct rcu_head *rhp)
10368	{
10369	/* Now it should be safe to free those cfs_rqs: */
10370	sched_unregister_group(container_of(rhp, struct task_group, rcu));
10371	}
10372
10373	void sched_destroy_group(struct task_group *tg)
10374	{
10375	/* Wait for possible concurrent references to cfs_rqs complete: */
10376	call_rcu(head: &tg->rcu, func: sched_unregister_group_rcu);
10377	}
10378
10379	void sched_release_group(struct task_group *tg)
10380	{
10381	unsigned long flags;
10382
10383	/*
10384	* Unlink first, to avoid walk_tg_tree_from() from finding us (via
10385	* sched_cfs_period_timer()).
10386	*
10387	* For this to be effective, we have to wait for all pending users of
10388	* this task group to leave their RCU critical section to ensure no new
10389	* user will see our dying task group any more. Specifically ensure
10390	* that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10391	*
10392	* We therefore defer calling unregister_fair_sched_group() to
10393	* sched_unregister_group() which is guarantied to get called only after the
10394	* current RCU grace period has expired.
10395	*/
10396	spin_lock_irqsave(&task_group_lock, flags);
10397	list_del_rcu(entry: &tg->list);
10398	list_del_rcu(entry: &tg->siblings);
10399	spin_unlock_irqrestore(lock: &task_group_lock, flags);
10400	}
10401
10402	static struct task_group *sched_get_task_group(struct task_struct *tsk)
10403	{
10404	struct task_group *tg;
10405
10406	/*
10407	* All callers are synchronized by task_rq_lock(); we do not use RCU
10408	* which is pointless here. Thus, we pass "true" to task_css_check()
10409	* to prevent lockdep warnings.
10410	*/
10411	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10412	struct task_group, css);
10413	tg = autogroup_task_group(p: tsk, tg);
10414
10415	return tg;
10416	}
10417
10418	static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10419	{
10420	tsk->sched_task_group = group;
10421
10422	#ifdef CONFIG_FAIR_GROUP_SCHED
10423	if (tsk->sched_class->task_change_group)
10424	tsk->sched_class->task_change_group(tsk);
10425	else
10426	#endif
10427	set_task_rq(p: tsk, cpu: task_cpu(p: tsk));
10428	}
10429
10430	/*
10431	* Change task's runqueue when it moves between groups.
10432	*
10433	* The caller of this function should have put the task in its new group by
10434	* now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10435	* its new group.
10436	*/
10437	void sched_move_task(struct task_struct *tsk)
10438	{
10439	int queued, running, queue_flags =
10440	DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10441	struct task_group *group;
10442	struct rq *rq;
10443
10444	CLASS(task_rq_lock, rq_guard)(l: tsk);
10445	rq = rq_guard.rq;
10446
10447	/*
10448	* Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10449	* group changes.
10450	*/
10451	group = sched_get_task_group(tsk);
10452	if (group == tsk->sched_task_group)
10453	return;
10454
10455	update_rq_clock(rq);
10456
10457	running = task_current(rq, p: tsk);
10458	queued = task_on_rq_queued(p: tsk);
10459
10460	if (queued)
10461	dequeue_task(rq, p: tsk, flags: queue_flags);
10462	if (running)
10463	put_prev_task(rq, prev: tsk);
10464
10465	sched_change_group(tsk, group);
10466
10467	if (queued)
10468	enqueue_task(rq, p: tsk, flags: queue_flags);
10469	if (running) {
10470	set_next_task(rq, next: tsk);
10471	/*
10472	* After changing group, the running task may have joined a
10473	* throttled one but it's still the running task. Trigger a
10474	* resched to make sure that task can still run.
10475	*/
10476	resched_curr(rq);
10477	}
10478	}
10479
10480	static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10481	{
10482	return css ? container_of(css, struct task_group, css) : NULL;
10483	}
10484
10485	static struct cgroup_subsys_state *
10486	cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10487	{
10488	struct task_group *parent = css_tg(css: parent_css);
10489	struct task_group *tg;
10490
10491	if (!parent) {
10492	/* This is early initialization for the top cgroup */
10493	return &root_task_group.css;
10494	}
10495
10496	tg = sched_create_group(parent);
10497	if (IS_ERR(ptr: tg))
10498	return ERR_PTR(error: -ENOMEM);
10499
10500	return &tg->css;
10501	}
10502
10503	/* Expose task group only after completing cgroup initialization */
10504	static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10505	{
10506	struct task_group *tg = css_tg(css);
10507	struct task_group *parent = css_tg(css: css->parent);
10508
10509	if (parent)
10510	sched_online_group(tg, parent);
10511
10512	#ifdef CONFIG_UCLAMP_TASK_GROUP
10513	/* Propagate the effective uclamp value for the new group */
10514	guard(mutex)(T: &uclamp_mutex);
10515	guard(rcu)();
10516	cpu_util_update_eff(css);
10517	#endif
10518
10519	return 0;
10520	}
10521
10522	static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10523	{
10524	struct task_group *tg = css_tg(css);
10525
10526	sched_release_group(tg);
10527	}
10528
10529	static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10530	{
10531	struct task_group *tg = css_tg(css);
10532
10533	/*
10534	* Relies on the RCU grace period between css_released() and this.
10535	*/
10536	sched_unregister_group(tg);
10537	}
10538
10539	#ifdef CONFIG_RT_GROUP_SCHED
10540	static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10541	{
10542	struct task_struct *task;
10543	struct cgroup_subsys_state *css;
10544
10545	cgroup_taskset_for_each(task, css, tset) {
10546	if (!sched_rt_can_attach(tg: css_tg(css), tsk: task))
10547	return -EINVAL;
10548	}
10549	return 0;
10550	}
10551	#endif
10552
10553	static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10554	{
10555	struct task_struct *task;
10556	struct cgroup_subsys_state *css;
10557
10558	cgroup_taskset_for_each(task, css, tset)
10559	sched_move_task(tsk: task);
10560	}
10561
10562	#ifdef CONFIG_UCLAMP_TASK_GROUP
10563	static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10564	{
10565	struct cgroup_subsys_state *top_css = css;
10566	struct uclamp_se *uc_parent = NULL;
10567	struct uclamp_se *uc_se = NULL;
10568	unsigned int eff[UCLAMP_CNT];
10569	enum uclamp_id clamp_id;
10570	unsigned int clamps;
10571
10572	lockdep_assert_held(&uclamp_mutex);
10573	SCHED_WARN_ON(!rcu_read_lock_held());
10574
10575	css_for_each_descendant_pre(css, top_css) {
10576	uc_parent = css_tg(css)->parent
10577	? css_tg(css)->parent->uclamp : NULL;
10578
10579	for_each_clamp_id(clamp_id) {
10580	/* Assume effective clamps matches requested clamps */
10581	eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10582	/* Cap effective clamps with parent's effective clamps */
10583	if (uc_parent &&
10584	eff[clamp_id] > uc_parent[clamp_id].value) {
10585	eff[clamp_id] = uc_parent[clamp_id].value;
10586	}
10587	}
10588	/* Ensure protection is always capped by limit */
10589	eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10590
10591	/* Propagate most restrictive effective clamps */
10592	clamps = 0x0;
10593	uc_se = css_tg(css)->uclamp;
10594	for_each_clamp_id(clamp_id) {
10595	if (eff[clamp_id] == uc_se[clamp_id].value)
10596	continue;
10597	uc_se[clamp_id].value = eff[clamp_id];
10598	uc_se[clamp_id].bucket_id = uclamp_bucket_id(clamp_value: eff[clamp_id]);
10599	clamps |= (0x1 << clamp_id);
10600	}
10601	if (!clamps) {
10602	css = css_rightmost_descendant(pos: css);
10603	continue;
10604	}
10605
10606	/* Immediately update descendants RUNNABLE tasks */
10607	uclamp_update_active_tasks(css);
10608	}
10609	}
10610
10611	/*
10612	* Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10613	* C expression. Since there is no way to convert a macro argument (N) into a
10614	* character constant, use two levels of macros.
10615	*/
10616	#define _POW10(exp) ((unsigned int)1e##exp)
10617	#define POW10(exp) _POW10(exp)
10618
10619	struct uclamp_request {
10620	#define UCLAMP_PERCENT_SHIFT 2
10621	#define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10622	s64 percent;
10623	u64 util;
10624	int ret;
10625	};
10626
10627	static inline struct uclamp_request
10628	capacity_from_percent(char *buf)
10629	{
10630	struct uclamp_request req = {
10631	.percent = UCLAMP_PERCENT_SCALE,
10632	.util = SCHED_CAPACITY_SCALE,
10633	.ret = 0,
10634	};
10635
10636	buf = strim(buf);
10637	if (strcmp(buf, "max")) {
10638	req.ret = cgroup_parse_float(input: buf, UCLAMP_PERCENT_SHIFT,
10639	v: &req.percent);
10640	if (req.ret)
10641	return req;
10642	if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10643	req.ret = -ERANGE;
10644	return req;
10645	}
10646
10647	req.util = req.percent << SCHED_CAPACITY_SHIFT;
10648	req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10649	}
10650
10651	return req;
10652	}
10653
10654	static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10655	size_t nbytes, loff_t off,
10656	enum uclamp_id clamp_id)
10657	{
10658	struct uclamp_request req;
10659	struct task_group *tg;
10660
10661	req = capacity_from_percent(buf);
10662	if (req.ret)
10663	return req.ret;
10664
10665	static_branch_enable(&sched_uclamp_used);
10666
10667	guard(mutex)(T: &uclamp_mutex);
10668	guard(rcu)();
10669
10670	tg = css_tg(css: of_css(of));
10671	if (tg->uclamp_req[clamp_id].value != req.util)
10672	uclamp_se_set(uc_se: &tg->uclamp_req[clamp_id], value: req.util, user_defined: false);
10673
10674	/*
10675	* Because of not recoverable conversion rounding we keep track of the
10676	* exact requested value
10677	*/
10678	tg->uclamp_pct[clamp_id] = req.percent;
10679
10680	/* Update effective clamps to track the most restrictive value */
10681	cpu_util_update_eff(css: of_css(of));
10682
10683	return nbytes;
10684	}
10685
10686	static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10687	char *buf, size_t nbytes,
10688	loff_t off)
10689	{
10690	return cpu_uclamp_write(of, buf, nbytes, off, clamp_id: UCLAMP_MIN);
10691	}
10692
10693	static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10694	char *buf, size_t nbytes,
10695	loff_t off)
10696	{
10697	return cpu_uclamp_write(of, buf, nbytes, off, clamp_id: UCLAMP_MAX);
10698	}
10699
10700	static inline void cpu_uclamp_print(struct seq_file *sf,
10701	enum uclamp_id clamp_id)
10702	{
10703	struct task_group *tg;
10704	u64 util_clamp;
10705	u64 percent;
10706	u32 rem;
10707
10708	scoped_guard (rcu) {
10709	tg = css_tg(css: seq_css(seq: sf));
10710	util_clamp = tg->uclamp_req[clamp_id].value;
10711	}
10712
10713	if (util_clamp == SCHED_CAPACITY_SCALE) {
10714	seq_puts(m: sf, s: "max\n");
10715	return;
10716	}
10717
10718	percent = tg->uclamp_pct[clamp_id];
10719	percent = div_u64_rem(dividend: percent, POW10(UCLAMP_PERCENT_SHIFT), remainder: &rem);
10720	seq_printf(m: sf, fmt: "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10721	}
10722
10723	static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10724	{
10725	cpu_uclamp_print(sf, clamp_id: UCLAMP_MIN);
10726	return 0;
10727	}
10728
10729	static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10730	{
10731	cpu_uclamp_print(sf, clamp_id: UCLAMP_MAX);
10732	return 0;
10733	}
10734	#endif /* CONFIG_UCLAMP_TASK_GROUP */
10735
10736	#ifdef CONFIG_FAIR_GROUP_SCHED
10737	static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10738	struct cftype *cftype, u64 shareval)
10739	{
10740	if (shareval > scale_load_down(ULONG_MAX))
10741	shareval = MAX_SHARES;
10742	return sched_group_set_shares(tg: css_tg(css), scale_load(shareval));
10743	}
10744
10745	static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10746	struct cftype *cft)
10747	{
10748	struct task_group *tg = css_tg(css);
10749
10750	return (u64) scale_load_down(tg->shares);
10751	}
10752
10753	#ifdef CONFIG_CFS_BANDWIDTH
10754	static DEFINE_MUTEX(cfs_constraints_mutex);
10755
10756	const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10757	static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10758	/* More than 203 days if BW_SHIFT equals 20. */
10759	static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10760
10761	static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10762
10763	static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10764	u64 burst)
10765	{
10766	int i, ret = 0, runtime_enabled, runtime_was_enabled;
10767	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10768
10769	if (tg == &root_task_group)
10770	return -EINVAL;
10771
10772	/*
10773	* Ensure we have at some amount of bandwidth every period. This is
10774	* to prevent reaching a state of large arrears when throttled via
10775	* entity_tick() resulting in prolonged exit starvation.
10776	*/
10777	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10778	return -EINVAL;
10779
10780	/*
10781	* Likewise, bound things on the other side by preventing insane quota
10782	* periods. This also allows us to normalize in computing quota
10783	* feasibility.
10784	*/
10785	if (period > max_cfs_quota_period)
10786	return -EINVAL;
10787
10788	/*
10789	* Bound quota to defend quota against overflow during bandwidth shift.
10790	*/
10791	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10792	return -EINVAL;
10793
10794	if (quota != RUNTIME_INF && (burst > quota ||
10795	burst + quota > max_cfs_runtime))
10796	return -EINVAL;
10797
10798	/*
10799	* Prevent race between setting of cfs_rq->runtime_enabled and
10800	* unthrottle_offline_cfs_rqs().
10801	*/
10802	guard(cpus_read_lock)();
10803	guard(mutex)(T: &cfs_constraints_mutex);
10804
10805	ret = __cfs_schedulable(tg, period, runtime: quota);
10806	if (ret)
10807	return ret;
10808
10809	runtime_enabled = quota != RUNTIME_INF;
10810	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10811	/*
10812	* If we need to toggle cfs_bandwidth_used, off->on must occur
10813	* before making related changes, and on->off must occur afterwards
10814	*/
10815	if (runtime_enabled && !runtime_was_enabled)
10816	cfs_bandwidth_usage_inc();
10817
10818	scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
10819	cfs_b->period = ns_to_ktime(ns: period);
10820	cfs_b->quota = quota;
10821	cfs_b->burst = burst;
10822
10823	__refill_cfs_bandwidth_runtime(cfs_b);
10824
10825	/*
10826	* Restart the period timer (if active) to handle new
10827	* period expiry:
10828	*/
10829	if (runtime_enabled)
10830	start_cfs_bandwidth(cfs_b);
10831	}
10832
10833	for_each_online_cpu(i) {
10834	struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10835	struct rq *rq = cfs_rq->rq;
10836
10837	guard(rq_lock_irq)(l: rq);
10838	cfs_rq->runtime_enabled = runtime_enabled;
10839	cfs_rq->runtime_remaining = 0;
10840
10841	if (cfs_rq->throttled)
10842	unthrottle_cfs_rq(cfs_rq);
10843	}
10844
10845	if (runtime_was_enabled && !runtime_enabled)
10846	cfs_bandwidth_usage_dec();
10847
10848	return 0;
10849	}
10850
10851	static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10852	{
10853	u64 quota, period, burst;
10854
10855	period = ktime_to_ns(kt: tg->cfs_bandwidth.period);
10856	burst = tg->cfs_bandwidth.burst;
10857	if (cfs_quota_us < 0)
10858	quota = RUNTIME_INF;
10859	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10860	quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10861	else
10862	return -EINVAL;
10863
10864	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10865	}
10866
10867	static long tg_get_cfs_quota(struct task_group *tg)
10868	{
10869	u64 quota_us;
10870
10871	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10872	return -1;
10873
10874	quota_us = tg->cfs_bandwidth.quota;
10875	do_div(quota_us, NSEC_PER_USEC);
10876
10877	return quota_us;
10878	}
10879
10880	static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10881	{
10882	u64 quota, period, burst;
10883
10884	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10885	return -EINVAL;
10886
10887	period = (u64)cfs_period_us * NSEC_PER_USEC;
10888	quota = tg->cfs_bandwidth.quota;
10889	burst = tg->cfs_bandwidth.burst;
10890
10891	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10892	}
10893
10894	static long tg_get_cfs_period(struct task_group *tg)
10895	{
10896	u64 cfs_period_us;
10897
10898	cfs_period_us = ktime_to_ns(kt: tg->cfs_bandwidth.period);
10899	do_div(cfs_period_us, NSEC_PER_USEC);
10900
10901	return cfs_period_us;
10902	}
10903
10904	static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10905	{
10906	u64 quota, period, burst;
10907
10908	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10909	return -EINVAL;
10910
10911	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10912	period = ktime_to_ns(kt: tg->cfs_bandwidth.period);
10913	quota = tg->cfs_bandwidth.quota;
10914
10915	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10916	}
10917
10918	static long tg_get_cfs_burst(struct task_group *tg)
10919	{
10920	u64 burst_us;
10921
10922	burst_us = tg->cfs_bandwidth.burst;
10923	do_div(burst_us, NSEC_PER_USEC);
10924
10925	return burst_us;
10926	}
10927
10928	static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10929	struct cftype *cft)
10930	{
10931	return tg_get_cfs_quota(tg: css_tg(css));
10932	}
10933
10934	static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10935	struct cftype *cftype, s64 cfs_quota_us)
10936	{
10937	return tg_set_cfs_quota(tg: css_tg(css), cfs_quota_us);
10938	}
10939
10940	static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10941	struct cftype *cft)
10942	{
10943	return tg_get_cfs_period(tg: css_tg(css));
10944	}
10945
10946	static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10947	struct cftype *cftype, u64 cfs_period_us)
10948	{
10949	return tg_set_cfs_period(tg: css_tg(css), cfs_period_us);
10950	}
10951
10952	static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10953	struct cftype *cft)
10954	{
10955	return tg_get_cfs_burst(tg: css_tg(css));
10956	}
10957
10958	static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10959	struct cftype *cftype, u64 cfs_burst_us)
10960	{
10961	return tg_set_cfs_burst(tg: css_tg(css), cfs_burst_us);
10962	}
10963
10964	struct cfs_schedulable_data {
10965	struct task_group *tg;
10966	u64 period, quota;
10967	};
10968
10969	/*
10970	* normalize group quota/period to be quota/max_period
10971	* note: units are usecs
10972	*/
10973	static u64 normalize_cfs_quota(struct task_group *tg,
10974	struct cfs_schedulable_data *d)
10975	{
10976	u64 quota, period;
10977
10978	if (tg == d->tg) {
10979	period = d->period;
10980	quota = d->quota;
10981	} else {
10982	period = tg_get_cfs_period(tg);
10983	quota = tg_get_cfs_quota(tg);
10984	}
10985
10986	/* note: these should typically be equivalent */
10987	if (quota == RUNTIME_INF || quota == -1)
10988	return RUNTIME_INF;
10989
10990	return to_ratio(period, runtime: quota);
10991	}
10992
10993	static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10994	{
10995	struct cfs_schedulable_data *d = data;
10996	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10997	s64 quota = 0, parent_quota = -1;
10998
10999	if (!tg->parent) {
11000	quota = RUNTIME_INF;
11001	} else {
11002	struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11003
11004	quota = normalize_cfs_quota(tg, d);
11005	parent_quota = parent_b->hierarchical_quota;
11006
11007	/*
11008	* Ensure max(child_quota) <= parent_quota. On cgroup2,
11009	* always take the non-RUNTIME_INF min. On cgroup1, only
11010	* inherit when no limit is set. In both cases this is used
11011	* by the scheduler to determine if a given CFS task has a
11012	* bandwidth constraint at some higher level.
11013	*/
11014	if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11015	if (quota == RUNTIME_INF)
11016	quota = parent_quota;
11017	else if (parent_quota != RUNTIME_INF)
11018	quota = min(quota, parent_quota);
11019	} else {
11020	if (quota == RUNTIME_INF)
11021	quota = parent_quota;
11022	else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11023	return -EINVAL;
11024	}
11025	}
11026	cfs_b->hierarchical_quota = quota;
11027
11028	return 0;
11029	}
11030
11031	static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11032	{
11033	struct cfs_schedulable_data data = {
11034	.tg = tg,
11035	.period = period,
11036	.quota = quota,
11037	};
11038
11039	if (quota != RUNTIME_INF) {
11040	do_div(data.period, NSEC_PER_USEC);
11041	do_div(data.quota, NSEC_PER_USEC);
11042	}
11043
11044	guard(rcu)();
11045	return walk_tg_tree(down: tg_cfs_schedulable_down, up: tg_nop, data: &data);
11046	}
11047
11048	static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11049	{
11050	struct task_group *tg = css_tg(css: seq_css(seq: sf));
11051	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11052
11053	seq_printf(m: sf, fmt: "nr_periods %d\n", cfs_b->nr_periods);
11054	seq_printf(m: sf, fmt: "nr_throttled %d\n", cfs_b->nr_throttled);
11055	seq_printf(m: sf, fmt: "throttled_time %llu\n", cfs_b->throttled_time);
11056
11057	if (schedstat_enabled() && tg != &root_task_group) {
11058	struct sched_statistics *stats;
11059	u64 ws = 0;
11060	int i;
11061
11062	for_each_possible_cpu(i) {
11063	stats = __schedstats_from_se(se: tg->se[i]);
11064	ws += schedstat_val(stats->wait_sum);
11065	}
11066
11067	seq_printf(m: sf, fmt: "wait_sum %llu\n", ws);
11068	}
11069
11070	seq_printf(m: sf, fmt: "nr_bursts %d\n", cfs_b->nr_burst);
11071	seq_printf(m: sf, fmt: "burst_time %llu\n", cfs_b->burst_time);
11072
11073	return 0;
11074	}
11075
11076	static u64 throttled_time_self(struct task_group *tg)
11077	{
11078	int i;
11079	u64 total = 0;
11080
11081	for_each_possible_cpu(i) {
11082	total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11083	}
11084
11085	return total;
11086	}
11087
11088	static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11089	{
11090	struct task_group *tg = css_tg(css: seq_css(seq: sf));
11091
11092	seq_printf(m: sf, fmt: "throttled_time %llu\n", throttled_time_self(tg));
11093
11094	return 0;
11095	}
11096	#endif /* CONFIG_CFS_BANDWIDTH */
11097	#endif /* CONFIG_FAIR_GROUP_SCHED */
11098
11099	#ifdef CONFIG_RT_GROUP_SCHED
11100	static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11101	struct cftype *cft, s64 val)
11102	{
11103	return sched_group_set_rt_runtime(tg: css_tg(css), rt_runtime_us: val);
11104	}
11105
11106	static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11107	struct cftype *cft)
11108	{
11109	return sched_group_rt_runtime(tg: css_tg(css));
11110	}
11111
11112	static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11113	struct cftype *cftype, u64 rt_period_us)
11114	{
11115	return sched_group_set_rt_period(tg: css_tg(css), rt_period_us);
11116	}
11117
11118	static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11119	struct cftype *cft)
11120	{
11121	return sched_group_rt_period(tg: css_tg(css));
11122	}
11123	#endif /* CONFIG_RT_GROUP_SCHED */
11124
11125	#ifdef CONFIG_FAIR_GROUP_SCHED
11126	static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11127	struct cftype *cft)
11128	{
11129	return css_tg(css)->idle;
11130	}
11131
11132	static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11133	struct cftype *cft, s64 idle)
11134	{
11135	return sched_group_set_idle(tg: css_tg(css), idle);
11136	}
11137	#endif
11138
11139	static struct cftype cpu_legacy_files[] = {
11140	#ifdef CONFIG_FAIR_GROUP_SCHED
11141	{
11142	.name = "shares",
11143	.read_u64 = cpu_shares_read_u64,
11144	.write_u64 = cpu_shares_write_u64,
11145	},
11146	{
11147	.name = "idle",
11148	.read_s64 = cpu_idle_read_s64,
11149	.write_s64 = cpu_idle_write_s64,
11150	},
11151	#endif
11152	#ifdef CONFIG_CFS_BANDWIDTH
11153	{
11154	.name = "cfs_quota_us",
11155	.read_s64 = cpu_cfs_quota_read_s64,
11156	.write_s64 = cpu_cfs_quota_write_s64,
11157	},
11158	{
11159	.name = "cfs_period_us",
11160	.read_u64 = cpu_cfs_period_read_u64,
11161	.write_u64 = cpu_cfs_period_write_u64,
11162	},
11163	{
11164	.name = "cfs_burst_us",
11165	.read_u64 = cpu_cfs_burst_read_u64,
11166	.write_u64 = cpu_cfs_burst_write_u64,
11167	},
11168	{
11169	.name = "stat",
11170	.seq_show = cpu_cfs_stat_show,
11171	},
11172	{
11173	.name = "stat.local",
11174	.seq_show = cpu_cfs_local_stat_show,
11175	},
11176	#endif
11177	#ifdef CONFIG_RT_GROUP_SCHED
11178	{
11179	.name = "rt_runtime_us",
11180	.read_s64 = cpu_rt_runtime_read,
11181	.write_s64 = cpu_rt_runtime_write,
11182	},
11183	{
11184	.name = "rt_period_us",
11185	.read_u64 = cpu_rt_period_read_uint,
11186	.write_u64 = cpu_rt_period_write_uint,
11187	},
11188	#endif
11189	#ifdef CONFIG_UCLAMP_TASK_GROUP
11190	{
11191	.name = "uclamp.min",
11192	.flags = CFTYPE_NOT_ON_ROOT,
11193	.seq_show = cpu_uclamp_min_show,
11194	.write = cpu_uclamp_min_write,
11195	},
11196	{
11197	.name = "uclamp.max",
11198	.flags = CFTYPE_NOT_ON_ROOT,
11199	.seq_show = cpu_uclamp_max_show,
11200	.write = cpu_uclamp_max_write,
11201	},
11202	#endif
11203	{ } /* Terminate */
11204	};
11205
11206	static int cpu_extra_stat_show(struct seq_file *sf,
11207	struct cgroup_subsys_state *css)
11208	{
11209	#ifdef CONFIG_CFS_BANDWIDTH
11210	{
11211	struct task_group *tg = css_tg(css);
11212	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11213	u64 throttled_usec, burst_usec;
11214
11215	throttled_usec = cfs_b->throttled_time;
11216	do_div(throttled_usec, NSEC_PER_USEC);
11217	burst_usec = cfs_b->burst_time;
11218	do_div(burst_usec, NSEC_PER_USEC);
11219
11220	seq_printf(m: sf, fmt: "nr_periods %d\n"
11221	"nr_throttled %d\n"
11222	"throttled_usec %llu\n"
11223	"nr_bursts %d\n"
11224	"burst_usec %llu\n",
11225	cfs_b->nr_periods, cfs_b->nr_throttled,
11226	throttled_usec, cfs_b->nr_burst, burst_usec);
11227	}
11228	#endif
11229	return 0;
11230	}
11231
11232	static int cpu_local_stat_show(struct seq_file *sf,
11233	struct cgroup_subsys_state *css)
11234	{
11235	#ifdef CONFIG_CFS_BANDWIDTH
11236	{
11237	struct task_group *tg = css_tg(css);
11238	u64 throttled_self_usec;
11239
11240	throttled_self_usec = throttled_time_self(tg);
11241	do_div(throttled_self_usec, NSEC_PER_USEC);
11242
11243	seq_printf(m: sf, fmt: "throttled_usec %llu\n",
11244	throttled_self_usec);
11245	}
11246	#endif
11247	return 0;
11248	}
11249
11250	#ifdef CONFIG_FAIR_GROUP_SCHED
11251	static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11252	struct cftype *cft)
11253	{
11254	struct task_group *tg = css_tg(css);
11255	u64 weight = scale_load_down(tg->shares);
11256
11257	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11258	}
11259
11260	static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11261	struct cftype *cft, u64 weight)
11262	{
11263	/*
11264	* cgroup weight knobs should use the common MIN, DFL and MAX
11265	* values which are 1, 100 and 10000 respectively. While it loses
11266	* a bit of range on both ends, it maps pretty well onto the shares
11267	* value used by scheduler and the round-trip conversions preserve
11268	* the original value over the entire range.
11269	*/
11270	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11271	return -ERANGE;
11272
11273	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11274
11275	return sched_group_set_shares(tg: css_tg(css), scale_load(weight));
11276	}
11277
11278	static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11279	struct cftype *cft)
11280	{
11281	unsigned long weight = scale_load_down(css_tg(css)->shares);
11282	int last_delta = INT_MAX;
11283	int prio, delta;
11284
11285	/* find the closest nice value to the current weight */
11286	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11287	delta = abs(sched_prio_to_weight[prio] - weight);
11288	if (delta >= last_delta)
11289	break;
11290	last_delta = delta;
11291	}
11292
11293	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11294	}
11295
11296	static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11297	struct cftype *cft, s64 nice)
11298	{
11299	unsigned long weight;
11300	int idx;
11301
11302	if (nice < MIN_NICE || nice > MAX_NICE)
11303	return -ERANGE;
11304
11305	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11306	idx = array_index_nospec(idx, 40);
11307	weight = sched_prio_to_weight[idx];
11308
11309	return sched_group_set_shares(tg: css_tg(css), scale_load(weight));
11310	}
11311	#endif
11312
11313	static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11314	long period, long quota)
11315	{
11316	if (quota < 0)
11317	seq_puts(m: sf, s: "max");
11318	else
11319	seq_printf(m: sf, fmt: "%ld", quota);
11320
11321	seq_printf(m: sf, fmt: " %ld\n", period);
11322	}
11323
11324	/* caller should put the current value in *@periodp before calling */
11325	static int __maybe_unused cpu_period_quota_parse(char *buf,
11326	u64 *periodp, u64 *quotap)
11327	{
11328	char tok[21]; /* U64_MAX */
11329
11330	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11331	return -EINVAL;
11332
11333	*periodp *= NSEC_PER_USEC;
11334
11335	if (sscanf(tok, "%llu", quotap))
11336	*quotap *= NSEC_PER_USEC;
11337	else if (!strcmp(tok, "max"))
11338	*quotap = RUNTIME_INF;
11339	else
11340	return -EINVAL;
11341
11342	return 0;
11343	}
11344
11345	#ifdef CONFIG_CFS_BANDWIDTH
11346	static int cpu_max_show(struct seq_file *sf, void *v)
11347	{
11348	struct task_group *tg = css_tg(css: seq_css(seq: sf));
11349
11350	cpu_period_quota_print(sf, period: tg_get_cfs_period(tg), quota: tg_get_cfs_quota(tg));
11351	return 0;
11352	}
11353
11354	static ssize_t cpu_max_write(struct kernfs_open_file *of,
11355	char *buf, size_t nbytes, loff_t off)
11356	{
11357	struct task_group *tg = css_tg(css: of_css(of));
11358	u64 period = tg_get_cfs_period(tg);
11359	u64 burst = tg_get_cfs_burst(tg);
11360	u64 quota;
11361	int ret;
11362
11363	ret = cpu_period_quota_parse(buf, periodp: &period, quotap: &quota);
11364	if (!ret)
11365	ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11366	return ret ?: nbytes;
11367	}
11368	#endif
11369
11370	static struct cftype cpu_files[] = {
11371	#ifdef CONFIG_FAIR_GROUP_SCHED
11372	{
11373	.name = "weight",
11374	.flags = CFTYPE_NOT_ON_ROOT,
11375	.read_u64 = cpu_weight_read_u64,
11376	.write_u64 = cpu_weight_write_u64,
11377	},
11378	{
11379	.name = "weight.nice",
11380	.flags = CFTYPE_NOT_ON_ROOT,
11381	.read_s64 = cpu_weight_nice_read_s64,
11382	.write_s64 = cpu_weight_nice_write_s64,
11383	},
11384	{
11385	.name = "idle",
11386	.flags = CFTYPE_NOT_ON_ROOT,
11387	.read_s64 = cpu_idle_read_s64,
11388	.write_s64 = cpu_idle_write_s64,
11389	},
11390	#endif
11391	#ifdef CONFIG_CFS_BANDWIDTH
11392	{
11393	.name = "max",
11394	.flags = CFTYPE_NOT_ON_ROOT,
11395	.seq_show = cpu_max_show,
11396	.write = cpu_max_write,
11397	},
11398	{
11399	.name = "max.burst",
11400	.flags = CFTYPE_NOT_ON_ROOT,
11401	.read_u64 = cpu_cfs_burst_read_u64,
11402	.write_u64 = cpu_cfs_burst_write_u64,
11403	},
11404	#endif
11405	#ifdef CONFIG_UCLAMP_TASK_GROUP
11406	{
11407	.name = "uclamp.min",
11408	.flags = CFTYPE_NOT_ON_ROOT,
11409	.seq_show = cpu_uclamp_min_show,
11410	.write = cpu_uclamp_min_write,
11411	},
11412	{
11413	.name = "uclamp.max",
11414	.flags = CFTYPE_NOT_ON_ROOT,
11415	.seq_show = cpu_uclamp_max_show,
11416	.write = cpu_uclamp_max_write,
11417	},
11418	#endif
11419	{ } /* terminate */
11420	};
11421
11422	struct cgroup_subsys cpu_cgrp_subsys = {
11423	.css_alloc = cpu_cgroup_css_alloc,
11424	.css_online = cpu_cgroup_css_online,
11425	.css_released = cpu_cgroup_css_released,
11426	.css_free = cpu_cgroup_css_free,
11427	.css_extra_stat_show = cpu_extra_stat_show,
11428	.css_local_stat_show = cpu_local_stat_show,
11429	#ifdef CONFIG_RT_GROUP_SCHED
11430	.can_attach = cpu_cgroup_can_attach,
11431	#endif
11432	.attach = cpu_cgroup_attach,
11433	.legacy_cftypes = cpu_legacy_files,
11434	.dfl_cftypes = cpu_files,
11435	.early_init = true,
11436	.threaded = true,
11437	};
11438
11439	#endif /* CONFIG_CGROUP_SCHED */
11440
11441	void dump_cpu_task(int cpu)
11442	{
11443	if (cpu == smp_processor_id() && in_hardirq()) {
11444	struct pt_regs *regs;
11445
11446	regs = get_irq_regs();
11447	if (regs) {
11448	show_regs(regs);
11449	return;
11450	}
11451	}
11452
11453	if (trigger_single_cpu_backtrace(cpu))
11454	return;
11455
11456	pr_info("Task dump for CPU %d:\n", cpu);
11457	sched_show_task(cpu_curr(cpu));
11458	}
11459
11460	/*
11461	* Nice levels are multiplicative, with a gentle 10% change for every
11462	* nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11463	* nice 1, it will get ~10% less CPU time than another CPU-bound task
11464	* that remained on nice 0.
11465	*
11466	* The "10% effect" is relative and cumulative: from _any_ nice level,
11467	* if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11468	* it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11469	* If a task goes up by ~10% and another task goes down by ~10% then
11470	* the relative distance between them is ~25%.)
11471	*/
11472	const int sched_prio_to_weight[40] = {
11473	/* -20 */ 88761, 71755, 56483, 46273, 36291,
11474	/* -15 */ 29154, 23254, 18705, 14949, 11916,
11475	/* -10 */ 9548, 7620, 6100, 4904, 3906,
11476	/* -5 */ 3121, 2501, 1991, 1586, 1277,
11477	/* 0 */ 1024, 820, 655, 526, 423,
11478	/* 5 */ 335, 272, 215, 172, 137,
11479	/* 10 */ 110, 87, 70, 56, 45,
11480	/* 15 */ 36, 29, 23, 18, 15,
11481	};
11482
11483	/*
11484	* Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11485	*
11486	* In cases where the weight does not change often, we can use the
11487	* precalculated inverse to speed up arithmetics by turning divisions
11488	* into multiplications:
11489	*/
11490	const u32 sched_prio_to_wmult[40] = {
11491	/* -20 */ 48388, 59856, 76040, 92818, 118348,
11492	/* -15 */ 147320, 184698, 229616, 287308, 360437,
11493	/* -10 */ 449829, 563644, 704093, 875809, 1099582,
11494	/* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11495	/* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11496	/* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11497	/* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11498	/* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11499	};
11500
11501	void call_trace_sched_update_nr_running(struct rq *rq, int count)
11502	{
11503	trace_sched_update_nr_running_tp(rq, change: count);
11504	}
11505
11506	#ifdef CONFIG_SCHED_MM_CID
11507
11508	/*
11509	* @cid_lock: Guarantee forward-progress of cid allocation.
11510	*
11511	* Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11512	* is only used when contention is detected by the lock-free allocation so
11513	* forward progress can be guaranteed.
11514	*/
11515	DEFINE_RAW_SPINLOCK(cid_lock);
11516
11517	/*
11518	* @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11519	*
11520	* When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11521	* detected, it is set to 1 to ensure that all newly coming allocations are
11522	* serialized by @cid_lock until the allocation which detected contention
11523	* completes and sets @use_cid_lock back to 0. This guarantees forward progress
11524	* of a cid allocation.
11525	*/
11526	int use_cid_lock;
11527
11528	/*
11529	* mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11530	* concurrently with respect to the execution of the source runqueue context
11531	* switch.
11532	*
11533	* There is one basic properties we want to guarantee here:
11534	*
11535	* (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11536	* used by a task. That would lead to concurrent allocation of the cid and
11537	* userspace corruption.
11538	*
11539	* Provide this guarantee by introducing a Dekker memory ordering to guarantee
11540	* that a pair of loads observe at least one of a pair of stores, which can be
11541	* shown as:
11542	*
11543	* X = Y = 0
11544	*
11545	* w[X]=1 w[Y]=1
11546	* MB MB
11547	* r[Y]=y r[X]=x
11548	*
11549	* Which guarantees that x==0 && y==0 is impossible. But rather than using
11550	* values 0 and 1, this algorithm cares about specific state transitions of the
11551	* runqueue current task (as updated by the scheduler context switch), and the
11552	* per-mm/cpu cid value.
11553	*
11554	* Let's introduce task (Y) which has task->mm == mm and task (N) which has
11555	* task->mm != mm for the rest of the discussion. There are two scheduler state
11556	* transitions on context switch we care about:
11557	*
11558	* (TSA) Store to rq->curr with transition from (N) to (Y)
11559	*
11560	* (TSB) Store to rq->curr with transition from (Y) to (N)
11561	*
11562	* On the remote-clear side, there is one transition we care about:
11563	*
11564	* (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11565	*
11566	* There is also a transition to UNSET state which can be performed from all
11567	* sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11568	* guarantees that only a single thread will succeed:
11569	*
11570	* (TMB) cmpxchg to *pcpu_cid to mark UNSET
11571	*
11572	* Just to be clear, what we do _not_ want to happen is a transition to UNSET
11573	* when a thread is actively using the cid (property (1)).
11574	*
11575	* Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11576	*
11577	* Scenario A) (TSA)+(TMA) (from next task perspective)
11578	*
11579	* CPU0 CPU1
11580	*
11581	* Context switch CS-1 Remote-clear
11582	* - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11583	* (implied barrier after cmpxchg)
11584	* - switch_mm_cid()
11585	* - memory barrier (see switch_mm_cid()
11586	* comment explaining how this barrier
11587	* is combined with other scheduler
11588	* barriers)
11589	* - mm_cid_get (next)
11590	* - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11591	*
11592	* This Dekker ensures that either task (Y) is observed by the
11593	* rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11594	* observed.
11595	*
11596	* If task (Y) store is observed by rcu_dereference(), it means that there is
11597	* still an active task on the cpu. Remote-clear will therefore not transition
11598	* to UNSET, which fulfills property (1).
11599	*
11600	* If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11601	* it will move its state to UNSET, which clears the percpu cid perhaps
11602	* uselessly (which is not an issue for correctness). Because task (Y) is not
11603	* observed, CPU1 can move ahead to set the state to UNSET. Because moving
11604	* state to UNSET is done with a cmpxchg expecting that the old state has the
11605	* LAZY flag set, only one thread will successfully UNSET.
11606	*
11607	* If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11608	* will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11609	* CPU1 will observe task (Y) and do nothing more, which is fine.
11610	*
11611	* What we are effectively preventing with this Dekker is a scenario where
11612	* neither LAZY flag nor store (Y) are observed, which would fail property (1)
11613	* because this would UNSET a cid which is actively used.
11614	*/
11615
11616	void sched_mm_cid_migrate_from(struct task_struct *t)
11617	{
11618	t->migrate_from_cpu = task_cpu(p: t);
11619	}
11620
11621	static
11622	int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11623	struct task_struct *t,
11624	struct mm_cid *src_pcpu_cid)
11625	{
11626	struct mm_struct *mm = t->mm;
11627	struct task_struct *src_task;
11628	int src_cid, last_mm_cid;
11629
11630	if (!mm)
11631	return -1;
11632
11633	last_mm_cid = t->last_mm_cid;
11634	/*
11635	* If the migrated task has no last cid, or if the current
11636	* task on src rq uses the cid, it means the source cid does not need
11637	* to be moved to the destination cpu.
11638	*/
11639	if (last_mm_cid == -1)
11640	return -1;
11641	src_cid = READ_ONCE(src_pcpu_cid->cid);
11642	if (!mm_cid_is_valid(cid: src_cid) || last_mm_cid != src_cid)
11643	return -1;
11644
11645	/*
11646	* If we observe an active task using the mm on this rq, it means we
11647	* are not the last task to be migrated from this cpu for this mm, so
11648	* there is no need to move src_cid to the destination cpu.
11649	*/
11650	guard(rcu)();
11651	src_task = rcu_dereference(src_rq->curr);
11652	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11653	t->last_mm_cid = -1;
11654	return -1;
11655	}
11656
11657	return src_cid;
11658	}
11659
11660	static
11661	int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11662	struct task_struct *t,
11663	struct mm_cid *src_pcpu_cid,
11664	int src_cid)
11665	{
11666	struct task_struct *src_task;
11667	struct mm_struct *mm = t->mm;
11668	int lazy_cid;
11669
11670	if (src_cid == -1)
11671	return -1;
11672
11673	/*
11674	* Attempt to clear the source cpu cid to move it to the destination
11675	* cpu.
11676	*/
11677	lazy_cid = mm_cid_set_lazy_put(cid: src_cid);
11678	if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11679	return -1;
11680
11681	/*
11682	* The implicit barrier after cmpxchg per-mm/cpu cid before loading
11683	* rq->curr->mm matches the scheduler barrier in context_switch()
11684	* between store to rq->curr and load of prev and next task's
11685	* per-mm/cpu cid.
11686	*
11687	* The implicit barrier after cmpxchg per-mm/cpu cid before loading
11688	* rq->curr->mm_cid_active matches the barrier in
11689	* sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11690	* sched_mm_cid_after_execve() between store to t->mm_cid_active and
11691	* load of per-mm/cpu cid.
11692	*/
11693
11694	/*
11695	* If we observe an active task using the mm on this rq after setting
11696	* the lazy-put flag, this task will be responsible for transitioning
11697	* from lazy-put flag set to MM_CID_UNSET.
11698	*/
11699	scoped_guard (rcu) {
11700	src_task = rcu_dereference(src_rq->curr);
11701	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11702	/*
11703	* We observed an active task for this mm, there is therefore
11704	* no point in moving this cid to the destination cpu.
11705	*/
11706	t->last_mm_cid = -1;
11707	return -1;
11708	}
11709	}
11710
11711	/*
11712	* The src_cid is unused, so it can be unset.
11713	*/
11714	if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11715	return -1;
11716	return src_cid;
11717	}
11718
11719	/*
11720	* Migration to dst cpu. Called with dst_rq lock held.
11721	* Interrupts are disabled, which keeps the window of cid ownership without the
11722	* source rq lock held small.
11723	*/
11724	void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11725	{
11726	struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11727	struct mm_struct *mm = t->mm;
11728	int src_cid, dst_cid, src_cpu;
11729	struct rq *src_rq;
11730
11731	lockdep_assert_rq_held(rq: dst_rq);
11732
11733	if (!mm)
11734	return;
11735	src_cpu = t->migrate_from_cpu;
11736	if (src_cpu == -1) {
11737	t->last_mm_cid = -1;
11738	return;
11739	}
11740	/*
11741	* Move the src cid if the dst cid is unset. This keeps id
11742	* allocation closest to 0 in cases where few threads migrate around
11743	* many cpus.
11744	*
11745	* If destination cid is already set, we may have to just clear
11746	* the src cid to ensure compactness in frequent migrations
11747	* scenarios.
11748	*
11749	* It is not useful to clear the src cid when the number of threads is
11750	* greater or equal to the number of allowed cpus, because user-space
11751	* can expect that the number of allowed cids can reach the number of
11752	* allowed cpus.
11753	*/
11754	dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11755	dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11756	if (!mm_cid_is_unset(cid: dst_cid) &&
11757	atomic_read(v: &mm->mm_users) >= t->nr_cpus_allowed)
11758	return;
11759	src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11760	src_rq = cpu_rq(src_cpu);
11761	src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11762	if (src_cid == -1)
11763	return;
11764	src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11765	src_cid);
11766	if (src_cid == -1)
11767	return;
11768	if (!mm_cid_is_unset(cid: dst_cid)) {
11769	__mm_cid_put(mm, cid: src_cid);
11770	return;
11771	}
11772	/* Move src_cid to dst cpu. */
11773	mm_cid_snapshot_time(rq: dst_rq, mm);
11774	WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11775	}
11776
11777	static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11778	int cpu)
11779	{
11780	struct rq *rq = cpu_rq(cpu);
11781	struct task_struct *t;
11782	int cid, lazy_cid;
11783
11784	cid = READ_ONCE(pcpu_cid->cid);
11785	if (!mm_cid_is_valid(cid))
11786	return;
11787
11788	/*
11789	* Clear the cpu cid if it is set to keep cid allocation compact. If
11790	* there happens to be other tasks left on the source cpu using this
11791	* mm, the next task using this mm will reallocate its cid on context
11792	* switch.
11793	*/
11794	lazy_cid = mm_cid_set_lazy_put(cid);
11795	if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11796	return;
11797
11798	/*
11799	* The implicit barrier after cmpxchg per-mm/cpu cid before loading
11800	* rq->curr->mm matches the scheduler barrier in context_switch()
11801	* between store to rq->curr and load of prev and next task's
11802	* per-mm/cpu cid.
11803	*
11804	* The implicit barrier after cmpxchg per-mm/cpu cid before loading
11805	* rq->curr->mm_cid_active matches the barrier in
11806	* sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11807	* sched_mm_cid_after_execve() between store to t->mm_cid_active and
11808	* load of per-mm/cpu cid.
11809	*/
11810
11811	/*
11812	* If we observe an active task using the mm on this rq after setting
11813	* the lazy-put flag, that task will be responsible for transitioning
11814	* from lazy-put flag set to MM_CID_UNSET.
11815	*/
11816	scoped_guard (rcu) {
11817	t = rcu_dereference(rq->curr);
11818	if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
11819	return;
11820	}
11821
11822	/*
11823	* The cid is unused, so it can be unset.
11824	* Disable interrupts to keep the window of cid ownership without rq
11825	* lock small.
11826	*/
11827	scoped_guard (irqsave) {
11828	if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11829	__mm_cid_put(mm, cid);
11830	}
11831	}
11832
11833	static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11834	{
11835	struct rq *rq = cpu_rq(cpu);
11836	struct mm_cid *pcpu_cid;
11837	struct task_struct *curr;
11838	u64 rq_clock;
11839
11840	/*
11841	* rq->clock load is racy on 32-bit but one spurious clear once in a
11842	* while is irrelevant.
11843	*/
11844	rq_clock = READ_ONCE(rq->clock);
11845	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11846
11847	/*
11848	* In order to take care of infrequently scheduled tasks, bump the time
11849	* snapshot associated with this cid if an active task using the mm is
11850	* observed on this rq.
11851	*/
11852	scoped_guard (rcu) {
11853	curr = rcu_dereference(rq->curr);
11854	if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11855	WRITE_ONCE(pcpu_cid->time, rq_clock);
11856	return;
11857	}
11858	}
11859
11860	if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11861	return;
11862	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11863	}
11864
11865	static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11866	int weight)
11867	{
11868	struct mm_cid *pcpu_cid;
11869	int cid;
11870
11871	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11872	cid = READ_ONCE(pcpu_cid->cid);
11873	if (!mm_cid_is_valid(cid) || cid < weight)
11874	return;
11875	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11876	}
11877
11878	static void task_mm_cid_work(struct callback_head *work)
11879	{
11880	unsigned long now = jiffies, old_scan, next_scan;
11881	struct task_struct *t = current;
11882	struct cpumask *cidmask;
11883	struct mm_struct *mm;
11884	int weight, cpu;
11885
11886	SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11887
11888	work->next = work; /* Prevent double-add */
11889	if (t->flags & PF_EXITING)
11890	return;
11891	mm = t->mm;
11892	if (!mm)
11893	return;
11894	old_scan = READ_ONCE(mm->mm_cid_next_scan);
11895	next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11896	if (!old_scan) {
11897	unsigned long res;
11898
11899	res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11900	if (res != old_scan)
11901	old_scan = res;
11902	else
11903	old_scan = next_scan;
11904	}
11905	if (time_before(now, old_scan))
11906	return;
11907	if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11908	return;
11909	cidmask = mm_cidmask(mm);
11910	/* Clear cids that were not recently used. */
11911	for_each_possible_cpu(cpu)
11912	sched_mm_cid_remote_clear_old(mm, cpu);
11913	weight = cpumask_weight(srcp: cidmask);
11914	/*
11915	* Clear cids that are greater or equal to the cidmask weight to
11916	* recompact it.
11917	*/
11918	for_each_possible_cpu(cpu)
11919	sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11920	}
11921
11922	void init_sched_mm_cid(struct task_struct *t)
11923	{
11924	struct mm_struct *mm = t->mm;
11925	int mm_users = 0;
11926
11927	if (mm) {
11928	mm_users = atomic_read(v: &mm->mm_users);
11929	if (mm_users == 1)
11930	mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11931	}
11932	t->cid_work.next = &t->cid_work; /* Protect against double add */
11933	init_task_work(twork: &t->cid_work, func: task_mm_cid_work);
11934	}
11935
11936	void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11937	{
11938	struct callback_head *work = &curr->cid_work;
11939	unsigned long now = jiffies;
11940
11941	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11942	work->next != work)
11943	return;
11944	if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11945	return;
11946	task_work_add(task: curr, twork: work, mode: TWA_RESUME);
11947	}
11948
11949	void sched_mm_cid_exit_signals(struct task_struct *t)
11950	{
11951	struct mm_struct *mm = t->mm;
11952	struct rq *rq;
11953
11954	if (!mm)
11955	return;
11956
11957	preempt_disable();
11958	rq = this_rq();
11959	guard(rq_lock_irqsave)(l: rq);
11960	preempt_enable_no_resched(); /* holding spinlock */
11961	WRITE_ONCE(t->mm_cid_active, 0);
11962	/*
11963	* Store t->mm_cid_active before loading per-mm/cpu cid.
11964	* Matches barrier in sched_mm_cid_remote_clear_old().
11965	*/
11966	smp_mb();
11967	mm_cid_put(mm);
11968	t->last_mm_cid = t->mm_cid = -1;
11969	}
11970
11971	void sched_mm_cid_before_execve(struct task_struct *t)
11972	{
11973	struct mm_struct *mm = t->mm;
11974	struct rq *rq;
11975
11976	if (!mm)
11977	return;
11978
11979	preempt_disable();
11980	rq = this_rq();
11981	guard(rq_lock_irqsave)(l: rq);
11982	preempt_enable_no_resched(); /* holding spinlock */
11983	WRITE_ONCE(t->mm_cid_active, 0);
11984	/*
11985	* Store t->mm_cid_active before loading per-mm/cpu cid.
11986	* Matches barrier in sched_mm_cid_remote_clear_old().
11987	*/
11988	smp_mb();
11989	mm_cid_put(mm);
11990	t->last_mm_cid = t->mm_cid = -1;
11991	}
11992
11993	void sched_mm_cid_after_execve(struct task_struct *t)
11994	{
11995	struct mm_struct *mm = t->mm;
11996	struct rq *rq;
11997
11998	if (!mm)
11999	return;
12000
12001	preempt_disable();
12002	rq = this_rq();
12003	scoped_guard (rq_lock_irqsave, rq) {
12004	preempt_enable_no_resched(); /* holding spinlock */
12005	WRITE_ONCE(t->mm_cid_active, 1);
12006	/*
12007	* Store t->mm_cid_active before loading per-mm/cpu cid.
12008	* Matches barrier in sched_mm_cid_remote_clear_old().
12009	*/
12010	smp_mb();
12011	t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12012	}
12013	rseq_set_notify_resume(t);
12014	}
12015
12016	void sched_mm_cid_fork(struct task_struct *t)
12017	{
12018	WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12019	t->mm_cid_active = 1;
12020	}
12021	#endif
12022