deadline.c source code [Linux/kernel/sched/deadline.c]

1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Deadline Scheduling Class (SCHED_DEADLINE)
4	*
5	* Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6	*
7	* Tasks that periodically executes their instances for less than their
8	* runtime won't miss any of their deadlines.
9	* Tasks that are not periodic or sporadic or that tries to execute more
10	* than their reserved bandwidth will be slowed down (and may potentially
11	* miss some of their deadlines), and won't affect any other task.
12	*
13	* Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14	* Juri Lelli <juri.lelli@gmail.com>,
15	* Michael Trimarchi <michael@amarulasolutions.com>,
16	* Fabio Checconi <fchecconi@gmail.com>
17	*/
18
19	#include <linux/cpuset.h>
20	#include <linux/sched/clock.h>
21	#include <uapi/linux/sched/types.h>
22	#include "sched.h"
23	#include "pelt.h"
24
25	/*
26	* Default limits for DL period; on the top end we guard against small util
27	* tasks still getting ridiculously long effective runtimes, on the bottom end we
28	* guard against timer DoS.
29	*/
30	static unsigned int sysctl_sched_dl_period_max = `1` << `22`; / ~4 seconds /
31	static unsigned int sysctl_sched_dl_period_min = `100`; / 100 us /
32	#ifdef CONFIG_SYSCTL
33	static const struct ctl_table sched_dl_sysctls[] = {
34	{
35	.procname = "sched_deadline_period_max_us",
36	.data = &sysctl_sched_dl_period_max,
37	.maxlen = sizeof(unsigned int),
38	.mode = `0644`,
39	.proc_handler = proc_douintvec_minmax,
40	.extra1 = (void *)&sysctl_sched_dl_period_min,
41	},
42	{
43	.procname = "sched_deadline_period_min_us",
44	.data = &sysctl_sched_dl_period_min,
45	.maxlen = sizeof(unsigned int),
46	.mode = `0644`,
47	.proc_handler = proc_douintvec_minmax,
48	.extra2 = (void *)&sysctl_sched_dl_period_max,
49	},
50	};
51
52	static int __init sched_dl_sysctl_init(void)
53	{
54	register_sysctl_init("kernel", sched_dl_sysctls);
55	return `0`;
56	}
57	late_initcall(sched_dl_sysctl_init);
58	#endif /* CONFIG_SYSCTL */
59
60	static bool dl_server(struct sched_dl_entity *dl_se)
61	{
62	return dl_se->dl_server;
63	}
64
65	static inline struct task_struct dl_task_of(struct* sched_dl_entity *dl_se)
66	{
67	BUG_ON(dl_server(dl_se));
68	return container_of(dl_se, struct task_struct, dl);
69	}
70
71	static inline struct rq rq_of_dl_rq(struct* dl_rq *dl_rq)
72	{
73	return container_of(dl_rq, struct rq, dl);
74	}
75
76	static inline struct rq rq_of_dl_se(struct* sched_dl_entity *dl_se)
77	{
78	struct rq *rq = dl_se->rq;
79
80	if (!dl_server(dl_se))
81	rq = task_rq(dl_task_of(dl_se));
82
83	return rq;
84	}
85
86	static inline struct dl_rq dl_rq_of_se(struct* sched_dl_entity *dl_se)
87	{
88	return &rq_of_dl_se(dl_se)->dl;
89	}
90
91	static inline int on_dl_rq(struct sched_dl_entity *dl_se)
92	{
93	return !RB_EMPTY_NODE(&dl_se->rb_node);
94	}
95
96	#ifdef CONFIG_RT_MUTEXES
97	static inline struct sched_dl_entity pi_of(struct* sched_dl_entity *dl_se)
98	{
99	return dl_se->pi_se;
100	}
101
102	static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
103	{
104	return pi_of(dl_se) != dl_se;
105	}
106	#else /* !CONFIG_RT_MUTEXES: */
107	static inline struct sched_dl_entity pi_of(struct* sched_dl_entity *dl_se)
108	{
109	return dl_se;
110	}
111
112	static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
113	{
114	return false;
115	}
116	#endif /* !CONFIG_RT_MUTEXES */
117
118	static inline struct dl_bw dl_bw_of(int* i)
119	{
120	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
121	"sched RCU must be held");
122	return &cpu_rq(i)->rd->dl_bw;
123	}
124
125	static inline int dl_bw_cpus(int i)
126	{
127	struct root_domain *rd = cpu_rq(i)->rd;
128	int cpus;
129
130	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
131	"sched RCU must be held");
132
133	if (cpumask_subset(src1p: rd->span, cpu_active_mask))
134	return cpumask_weight(srcp: rd->span);
135
136	cpus = `0`;
137
138	for_each_cpu_and(i, rd->span, cpu_active_mask)
139	cpus++;
140
141	return cpus;
142	}
143
144	static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
145	{
146	unsigned long cap = `0`;
147	int i;
148
149	for_each_cpu_and(i, mask, cpu_active_mask)
150	cap += arch_scale_cpu_capacity(cpu: i);
151
152	return cap;
153	}
154
155	/*
156	* XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
157	* of the CPU the task is running on rather rd's \Sum CPU capacity.
158	*/
159	static inline unsigned long dl_bw_capacity(int i)
160	{
161	if (!sched_asym_cpucap_active() &&
162	arch_scale_cpu_capacity(cpu: i) == SCHED_CAPACITY_SCALE) {
163	return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
164	} else {
165	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
166	"sched RCU must be held");
167
168	return __dl_bw_capacity(cpu_rq(i)->rd->span);
169	}
170	}
171
172	bool dl_bw_visited(int cpu, u64 cookie)
173	{
174	struct root_domain *rd = cpu_rq(cpu)->rd;
175
176	if (rd->visit_cookie == cookie)
177	return true;
178
179	rd->visit_cookie = cookie;
180	return false;
181	}
182
183	static inline
184	void __dl_update(struct dl_bw *dl_b, s64 bw)
185	{
186	struct root_domain rd = container_of(dl_b, struct* root_domain, dl_bw);
187	int i;
188
189	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
190	"sched RCU must be held");
191	for_each_cpu_and(i, rd->span, cpu_active_mask) {
192	struct rq *rq = cpu_rq(i);
193
194	rq->dl.extra_bw += bw;
195	}
196	}
197
198	static inline
199	void __dl_sub(struct dl_bw dl_b, u64 tsk_bw, int* cpus)
200	{
201	dl_b->total_bw -= tsk_bw;
202	__dl_update(dl_b, bw: (s32)tsk_bw / cpus);
203	}
204
205	static inline
206	void __dl_add(struct dl_bw dl_b, u64 tsk_bw, int* cpus)
207	{
208	dl_b->total_bw += tsk_bw;
209	__dl_update(dl_b, bw: -((s32)tsk_bw / cpus));
210	}
211
212	static inline bool
213	__dl_overflow(struct dl_bw dl_b, unsigned* long cap, u64 old_bw, u64 new_bw)
214	{
215	return dl_b->bw != -`1` &&
216	cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
217	}
218
219	static inline
220	void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
221	{
222	u64 old = dl_rq->running_bw;
223
224	lockdep_assert_rq_held(rq: rq_of_dl_rq(dl_rq));
225	dl_rq->running_bw += dl_bw;
226	WARN_ON_ONCE(dl_rq->running_bw < old); / overflow /
227	WARN_ON_ONCE(dl_rq->running_bw > dl_rq->this_bw);
228	/ kick cpufreq (see the comment in kernel/sched/sched.h). /
229	cpufreq_update_util(rq: rq_of_dl_rq(dl_rq), flags: `0`);
230	}
231
232	static inline
233	void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
234	{
235	u64 old = dl_rq->running_bw;
236
237	lockdep_assert_rq_held(rq: rq_of_dl_rq(dl_rq));
238	dl_rq->running_bw -= dl_bw;
239	WARN_ON_ONCE(dl_rq->running_bw > old); / underflow /
240	if (dl_rq->running_bw > old)
241	dl_rq->running_bw = `0`;
242	/ kick cpufreq (see the comment in kernel/sched/sched.h). /
243	cpufreq_update_util(rq: rq_of_dl_rq(dl_rq), flags: `0`);
244	}
245
246	static inline
247	void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
248	{
249	u64 old = dl_rq->this_bw;
250
251	lockdep_assert_rq_held(rq: rq_of_dl_rq(dl_rq));
252	dl_rq->this_bw += dl_bw;
253	WARN_ON_ONCE(dl_rq->this_bw < old); / overflow /
254	}
255
256	static inline
257	void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
258	{
259	u64 old = dl_rq->this_bw;
260
261	lockdep_assert_rq_held(rq: rq_of_dl_rq(dl_rq));
262	dl_rq->this_bw -= dl_bw;
263	WARN_ON_ONCE(dl_rq->this_bw > old); / underflow /
264	if (dl_rq->this_bw > old)
265	dl_rq->this_bw = `0`;
266	WARN_ON_ONCE(dl_rq->running_bw > dl_rq->this_bw);
267	}
268
269	static inline
270	void add_rq_bw(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
271	{
272	if (!dl_entity_is_special(dl_se))
273	__add_rq_bw(dl_bw: dl_se->dl_bw, dl_rq);
274	}
275
276	static inline
277	void sub_rq_bw(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
278	{
279	if (!dl_entity_is_special(dl_se))
280	__sub_rq_bw(dl_bw: dl_se->dl_bw, dl_rq);
281	}
282
283	static inline
284	void add_running_bw(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
285	{
286	if (!dl_entity_is_special(dl_se))
287	__add_running_bw(dl_bw: dl_se->dl_bw, dl_rq);
288	}
289
290	static inline
291	void sub_running_bw(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
292	{
293	if (!dl_entity_is_special(dl_se))
294	__sub_running_bw(dl_bw: dl_se->dl_bw, dl_rq);
295	}
296
297	static void dl_rq_change_utilization(struct rq rq, struct* sched_dl_entity *dl_se, u64 new_bw)
298	{
299	if (dl_se->dl_non_contending) {
300	sub_running_bw(dl_se, dl_rq: &rq->dl);
301	dl_se->dl_non_contending = `0`;
302
303	/*
304	* If the timer handler is currently running and the
305	* timer cannot be canceled, inactive_task_timer()
306	* will see that dl_not_contending is not set, and
307	* will not touch the rq's active utilization,
308	* so we are still safe.
309	*/
310	if (hrtimer_try_to_cancel(timer: &dl_se->inactive_timer) == `1`) {
311	if (!dl_server(dl_se))
312	put_task_struct(t: dl_task_of(dl_se));
313	}
314	}
315	__sub_rq_bw(dl_bw: dl_se->dl_bw, dl_rq: &rq->dl);
316	__add_rq_bw(dl_bw: new_bw, dl_rq: &rq->dl);
317	}
318
319	static __always_inline
320	void cancel_dl_timer(struct sched_dl_entity dl_se, struct* hrtimer *timer)
321	{
322	/*
323	* If the timer callback was running (hrtimer_try_to_cancel == -1),
324	* it will eventually call put_task_struct().
325	*/
326	if (hrtimer_try_to_cancel(timer) == `1` && !dl_server(dl_se))
327	put_task_struct(t: dl_task_of(dl_se));
328	}
329
330	static __always_inline
331	void cancel_replenish_timer(struct sched_dl_entity *dl_se)
332	{
333	cancel_dl_timer(dl_se, timer: &dl_se->dl_timer);
334	}
335
336	static __always_inline
337	void cancel_inactive_timer(struct sched_dl_entity *dl_se)
338	{
339	cancel_dl_timer(dl_se, timer: &dl_se->inactive_timer);
340	}
341
342	static void dl_change_utilization(struct task_struct *p, u64 new_bw)
343	{
344	WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
345
346	if (task_on_rq_queued(p))
347	return;
348
349	dl_rq_change_utilization(task_rq(p), dl_se: &p->dl, new_bw);
350	}
351
352	static void __dl_clear_params(struct sched_dl_entity *dl_se);
353
354	/*
355	* The utilization of a task cannot be immediately removed from
356	* the rq active utilization (running_bw) when the task blocks.
357	* Instead, we have to wait for the so called "0-lag time".
358	*
359	* If a task blocks before the "0-lag time", a timer (the inactive
360	* timer) is armed, and running_bw is decreased when the timer
361	* fires.
362	*
363	* If the task wakes up again before the inactive timer fires,
364	* the timer is canceled, whereas if the task wakes up after the
365	* inactive timer fired (and running_bw has been decreased) the
366	* task's utilization has to be added to running_bw again.
367	* A flag in the deadline scheduling entity (dl_non_contending)
368	* is used to avoid race conditions between the inactive timer handler
369	* and task wakeups.
370	*
371	* The following diagram shows how running_bw is updated. A task is
372	* "ACTIVE" when its utilization contributes to running_bw; an
373	* "ACTIVE contending" task is in the TASK_RUNNING state, while an
374	* "ACTIVE non contending" task is a blocked task for which the "0-lag time"
375	* has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
376	* time already passed, which does not contribute to running_bw anymore.
377	* +------------------+
378	* wakeup \| ACTIVE \|
379	* +------------------>+ contending \|
380	* \| add_running_bw \| \|
381	* \| +----+------+------+
382	* \| \| ^
383	* \| dequeue \| \|
384	* +--------+-------+ \| \|
385	* \| \| t >= 0-lag \| \| wakeup
386	* \| INACTIVE \|<---------------+ \|
387	* \| \| sub_running_bw \| \|
388	* +--------+-------+ \| \|
389	* ^ \| \|
390	* \| t < 0-lag \| \|
391	* \| \| \|
392	* \| V \|
393	* \| +----+------+------+
394	* \| sub_running_bw \| ACTIVE \|
395	* +-------------------+ \|
396	* inactive timer \| non contending \|
397	* fired +------------------+
398	*
399	* The task_non_contending() function is invoked when a task
400	* blocks, and checks if the 0-lag time already passed or
401	* not (in the first case, it directly updates running_bw;
402	* in the second case, it arms the inactive timer).
403	*
404	* The task_contending() function is invoked when a task wakes
405	* up, and checks if the task is still in the "ACTIVE non contending"
406	* state or not (in the second case, it updates running_bw).
407	*/
408	static void task_non_contending(struct sched_dl_entity *dl_se)
409	{
410	struct hrtimer *timer = &dl_se->inactive_timer;
411	struct rq *rq = rq_of_dl_se(dl_se);
412	struct dl_rq *dl_rq = &rq->dl;
413	s64 zerolag_time;
414
415	/*
416	* If this is a non-deadline task that has been boosted,
417	* do nothing
418	*/
419	if (dl_se->dl_runtime == `0`)
420	return;
421
422	if (dl_entity_is_special(dl_se))
423	return;
424
425	WARN_ON(dl_se->dl_non_contending);
426
427	zerolag_time = dl_se->deadline -
428	div64_long((dl_se->runtime * dl_se->dl_period),
429	dl_se->dl_runtime);
430
431	/*
432	* Using relative times instead of the absolute "0-lag time"
433	* allows to simplify the code
434	*/
435	zerolag_time -= rq_clock(rq);
436
437	/*
438	* If the "0-lag time" already passed, decrease the active
439	* utilization now, instead of starting a timer
440	*/
441	if ((zerolag_time < `0`) \|\| hrtimer_active(timer: &dl_se->inactive_timer)) {
442	if (dl_server(dl_se)) {
443	sub_running_bw(dl_se, dl_rq);
444	} else {
445	struct task_struct *p = dl_task_of(dl_se);
446
447	if (dl_task(p))
448	sub_running_bw(dl_se, dl_rq);
449
450	if (!dl_task(p) \|\| READ_ONCE(p->__state) == TASK_DEAD) {
451	struct dl_bw *dl_b = dl_bw_of(i: task_cpu(p));
452
453	if (READ_ONCE(p->__state) == TASK_DEAD)
454	sub_rq_bw(dl_se, dl_rq: &rq->dl);
455	raw_spin_lock(&dl_b->lock);
456	__dl_sub(dl_b, tsk_bw: dl_se->dl_bw, cpus: dl_bw_cpus(i: task_cpu(p)));
457	raw_spin_unlock(&dl_b->lock);
458	__dl_clear_params(dl_se);
459	}
460	}
461
462	return;
463	}
464
465	dl_se->dl_non_contending = `1`;
466	if (!dl_server(dl_se))
467	get_task_struct(t: dl_task_of(dl_se));
468
469	hrtimer_start(timer, tim: ns_to_ktime(ns: zerolag_time), mode: HRTIMER_MODE_REL_HARD);
470	}
471
472	static void task_contending(struct sched_dl_entity dl_se, int* flags)
473	{
474	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
475
476	/*
477	* If this is a non-deadline task that has been boosted,
478	* do nothing
479	*/
480	if (dl_se->dl_runtime == `0`)
481	return;
482
483	if (flags & ENQUEUE_MIGRATED)
484	add_rq_bw(dl_se, dl_rq);
485
486	if (dl_se->dl_non_contending) {
487	dl_se->dl_non_contending = `0`;
488	/*
489	* If the timer handler is currently running and the
490	* timer cannot be canceled, inactive_task_timer()
491	* will see that dl_not_contending is not set, and
492	* will not touch the rq's active utilization,
493	* so we are still safe.
494	*/
495	cancel_inactive_timer(dl_se);
496	} else {
497	/*
498	* Since "dl_non_contending" is not set, the
499	* task's utilization has already been removed from
500	* active utilization (either when the task blocked,
501	* when the "inactive timer" fired).
502	* So, add it back.
503	*/
504	add_running_bw(dl_se, dl_rq);
505	}
506	}
507
508	static inline int is_leftmost(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
509	{
510	return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
511	}
512
513	static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
514
515	void init_dl_bw(struct dl_bw *dl_b)
516	{
517	raw_spin_lock_init(&dl_b->lock);
518	if (global_rt_runtime() == RUNTIME_INF)
519	dl_b->bw = -`1`;
520	else
521	dl_b->bw = to_ratio(period: global_rt_period(), runtime: global_rt_runtime());
522	dl_b->total_bw = `0`;
523	}
524
525	void init_dl_rq(struct dl_rq *dl_rq)
526	{
527	dl_rq->root = RB_ROOT_CACHED;
528
529	/ zero means no -deadline tasks /
530	dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = `0`;
531
532	dl_rq->overloaded = `0`;
533	dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
534
535	dl_rq->running_bw = `0`;
536	dl_rq->this_bw = `0`;
537	init_dl_rq_bw_ratio(dl_rq);
538	}
539
540	static inline int dl_overloaded(struct rq *rq)
541	{
542	return atomic_read(v: &rq->rd->dlo_count);
543	}
544
545	static inline void dl_set_overload(struct rq *rq)
546	{
547	if (!rq->online)
548	return;
549
550	cpumask_set_cpu(cpu: rq->cpu, dstp: rq->rd->dlo_mask);
551	/*
552	* Must be visible before the overload count is
553	* set (as in sched_rt.c).
554	*
555	* Matched by the barrier in pull_dl_task().
556	*/
557	smp_wmb();
558	atomic_inc(v: &rq->rd->dlo_count);
559	}
560
561	static inline void dl_clear_overload(struct rq *rq)
562	{
563	if (!rq->online)
564	return;
565
566	atomic_dec(v: &rq->rd->dlo_count);
567	cpumask_clear_cpu(cpu: rq->cpu, dstp: rq->rd->dlo_mask);
568	}
569
570	#define __node_2_pdl(node) \
571	rb_entry((node), struct task_struct, pushable_dl_tasks)
572
573	static inline bool __pushable_less(struct rb_node a, const* struct rb_node *b)
574	{
575	return dl_entity_preempt(a: &__node_2_pdl(a)->dl, b: &__node_2_pdl(b)->dl);
576	}
577
578	static inline int has_pushable_dl_tasks(struct rq *rq)
579	{
580	return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
581	}
582
583	/*
584	* The list of pushable -deadline task is not a plist, like in
585	* sched_rt.c, it is an rb-tree with tasks ordered by deadline.
586	*/
587	static void enqueue_pushable_dl_task(struct rq rq, struct* task_struct *p)
588	{
589	struct rb_node *leftmost;
590
591	WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
592
593	leftmost = rb_add_cached(node: &p->pushable_dl_tasks,
594	tree: &rq->dl.pushable_dl_tasks_root,
595	less: __pushable_less);
596	if (leftmost)
597	rq->dl.earliest_dl.next = p->dl.deadline;
598
599	if (!rq->dl.overloaded) {
600	dl_set_overload(rq);
601	rq->dl.overloaded = `1`;
602	}
603	}
604
605	static void dequeue_pushable_dl_task(struct rq rq, struct* task_struct *p)
606	{
607	struct dl_rq *dl_rq = &rq->dl;
608	struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
609	struct rb_node *leftmost;
610
611	if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
612	return;
613
614	leftmost = rb_erase_cached(node: &p->pushable_dl_tasks, root);
615	if (leftmost)
616	dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
617
618	RB_CLEAR_NODE(&p->pushable_dl_tasks);
619
620	if (!has_pushable_dl_tasks(rq) && rq->dl.overloaded) {
621	dl_clear_overload(rq);
622	rq->dl.overloaded = `0`;
623	}
624	}
625
626	static int push_dl_task(struct rq *rq);
627
628	static inline bool need_pull_dl_task(struct rq rq, struct* task_struct *prev)
629	{
630	return rq->online && dl_task(p: prev);
631	}
632
633	static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
634	static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
635
636	static void push_dl_tasks(struct rq *);
637	static void pull_dl_task(struct rq *);
638
639	static inline void deadline_queue_push_tasks(struct rq *rq)
640	{
641	if (!has_pushable_dl_tasks(rq))
642	return;
643
644	queue_balance_callback(rq, head: &per_cpu(dl_push_head, rq->cpu), func: push_dl_tasks);
645	}
646
647	static inline void deadline_queue_pull_task(struct rq *rq)
648	{
649	queue_balance_callback(rq, head: &per_cpu(dl_pull_head, rq->cpu), func: pull_dl_task);
650	}
651
652	static struct rq find_lock_later_rq(struct* task_struct task, struct* rq *rq);
653
654	static struct rq dl_task_offline_migration(struct* rq rq, struct* task_struct *p)
655	{
656	struct rq *later_rq = NULL;
657	struct dl_bw *dl_b;
658
659	later_rq = find_lock_later_rq(task: p, rq);
660	if (!later_rq) {
661	int cpu;
662
663	/*
664	* If we cannot preempt any rq, fall back to pick any
665	* online CPU:
666	*/
667	cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
668	if (cpu >= nr_cpu_ids) {
669	/*
670	* Failed to find any suitable CPU.
671	* The task will never come back!
672	*/
673	WARN_ON_ONCE(dl_bandwidth_enabled());
674
675	/*
676	* If admission control is disabled we
677	* try a little harder to let the task
678	* run.
679	*/
680	cpu = cpumask_any(cpu_active_mask);
681	}
682	later_rq = cpu_rq(cpu);
683	double_lock_balance(this_rq: rq, busiest: later_rq);
684	}
685
686	if (p->dl.dl_non_contending \|\| p->dl.dl_throttled) {
687	/*
688	* Inactive timer is armed (or callback is running, but
689	* waiting for us to release rq locks). In any case, when it
690	* will fire (or continue), it will see running_bw of this
691	* task migrated to later_rq (and correctly handle it).
692	*/
693	sub_running_bw(dl_se: &p->dl, dl_rq: &rq->dl);
694	sub_rq_bw(dl_se: &p->dl, dl_rq: &rq->dl);
695
696	add_rq_bw(dl_se: &p->dl, dl_rq: &later_rq->dl);
697	add_running_bw(dl_se: &p->dl, dl_rq: &later_rq->dl);
698	} else {
699	sub_rq_bw(dl_se: &p->dl, dl_rq: &rq->dl);
700	add_rq_bw(dl_se: &p->dl, dl_rq: &later_rq->dl);
701	}
702
703	/*
704	* And we finally need to fix up root_domain(s) bandwidth accounting,
705	* since p is still hanging out in the old (now moved to default) root
706	* domain.
707	*/
708	dl_b = &rq->rd->dl_bw;
709	raw_spin_lock(&dl_b->lock);
710	__dl_sub(dl_b, tsk_bw: p->dl.dl_bw, cpus: cpumask_weight(srcp: rq->rd->span));
711	raw_spin_unlock(&dl_b->lock);
712
713	dl_b = &later_rq->rd->dl_bw;
714	raw_spin_lock(&dl_b->lock);
715	__dl_add(dl_b, tsk_bw: p->dl.dl_bw, cpus: cpumask_weight(srcp: later_rq->rd->span));
716	raw_spin_unlock(&dl_b->lock);
717
718	set_task_cpu(p, cpu: later_rq->cpu);
719	double_unlock_balance(this_rq: later_rq, busiest: rq);
720
721	return later_rq;
722	}
723
724	static void
725	enqueue_dl_entity(struct sched_dl_entity dl_se, int* flags);
726	static void enqueue_task_dl(struct rq rq, struct* task_struct p, int* flags);
727	static void dequeue_dl_entity(struct sched_dl_entity dl_se, int* flags);
728	static void wakeup_preempt_dl(struct rq rq, struct* task_struct p, int* flags);
729
730	static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
731	struct rq *rq)
732	{
733	/ for non-boosted task, pi_of(dl_se) == dl_se /
734	dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
735	dl_se->runtime = pi_of(dl_se)->dl_runtime;
736
737	/*
738	* If it is a deferred reservation, and the server
739	* is not handling an starvation case, defer it.
740	*/
741	if (dl_se->dl_defer && !dl_se->dl_defer_running) {
742	dl_se->dl_throttled = `1`;
743	dl_se->dl_defer_armed = `1`;
744	}
745	}
746
747	/*
748	* We are being explicitly informed that a new instance is starting,
749	* and this means that:
750	* - the absolute deadline of the entity has to be placed at
751	* current time + relative deadline;
752	* - the runtime of the entity has to be set to the maximum value.
753	*
754	* The capability of specifying such event is useful whenever a -deadline
755	* entity wants to (try to!) synchronize its behaviour with the scheduler's
756	* one, and to (try to!) reconcile itself with its own scheduling
757	* parameters.
758	*/
759	static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
760	{
761	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
762	struct rq *rq = rq_of_dl_rq(dl_rq);
763
764	update_rq_clock(rq);
765
766	WARN_ON(is_dl_boosted(dl_se));
767	WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
768
769	/*
770	* We are racing with the deadline timer. So, do nothing because
771	* the deadline timer handler will take care of properly recharging
772	* the runtime and postponing the deadline
773	*/
774	if (dl_se->dl_throttled)
775	return;
776
777	/*
778	* We use the regular wall clock time to set deadlines in the
779	* future; in fact, we must consider execution overheads (time
780	* spent on hardirq context, etc.).
781	*/
782	replenish_dl_new_period(dl_se, rq);
783	}
784
785	static int start_dl_timer(struct sched_dl_entity *dl_se);
786	static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t);
787
788	/*
789	* Pure Earliest Deadline First (EDF) scheduling does not deal with the
790	* possibility of a entity lasting more than what it declared, and thus
791	* exhausting its runtime.
792	*
793	* Here we are interested in making runtime overrun possible, but we do
794	* not want a entity which is misbehaving to affect the scheduling of all
795	* other entities.
796	* Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
797	* is used, in order to confine each entity within its own bandwidth.
798	*
799	* This function deals exactly with that, and ensures that when the runtime
800	* of a entity is replenished, its deadline is also postponed. That ensures
801	* the overrunning entity can't interfere with other entity in the system and
802	* can't make them miss their deadlines. Reasons why this kind of overruns
803	* could happen are, typically, a entity voluntarily trying to overcome its
804	* runtime, or it just underestimated it during sched_setattr().
805	*/
806	static void replenish_dl_entity(struct sched_dl_entity *dl_se)
807	{
808	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
809	struct rq *rq = rq_of_dl_rq(dl_rq);
810
811	WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= `0`);
812
813	/*
814	* This could be the case for a !-dl task that is boosted.
815	* Just go with full inherited parameters.
816	*
817	* Or, it could be the case of a deferred reservation that
818	* was not able to consume its runtime in background and
819	* reached this point with current u > U.
820	*
821	* In both cases, set a new period.
822	*/
823	if (dl_se->dl_deadline == `0` \|\|
824	(dl_se->dl_defer_armed && dl_entity_overflow(dl_se, t: rq_clock(rq)))) {
825	dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
826	dl_se->runtime = pi_of(dl_se)->dl_runtime;
827	}
828
829	if (dl_se->dl_yielded && dl_se->runtime > `0`)
830	dl_se->runtime = `0`;
831
832	/*
833	* We keep moving the deadline away until we get some
834	* available runtime for the entity. This ensures correct
835	* handling of situations where the runtime overrun is
836	* arbitrary large.
837	*/
838	while (dl_se->runtime <= `0`) {
839	dl_se->deadline += pi_of(dl_se)->dl_period;
840	dl_se->runtime += pi_of(dl_se)->dl_runtime;
841	}
842
843	/*
844	* At this point, the deadline really should be "in
845	* the future" with respect to rq->clock. If it's
846	* not, we are, for some reason, lagging too much!
847	* Anyway, after having warn userspace abut that,
848	* we still try to keep the things running by
849	* resetting the deadline and the budget of the
850	* entity.
851	*/
852	if (dl_time_before(a: dl_se->deadline, b: rq_clock(rq))) {
853	printk_deferred_once("sched: DL replenish lagged too much\n");
854	replenish_dl_new_period(dl_se, rq);
855	}
856
857	if (dl_se->dl_yielded)
858	dl_se->dl_yielded = `0`;
859	if (dl_se->dl_throttled)
860	dl_se->dl_throttled = `0`;
861
862	/*
863	* If this is the replenishment of a deferred reservation,
864	* clear the flag and return.
865	*/
866	if (dl_se->dl_defer_armed) {
867	dl_se->dl_defer_armed = `0`;
868	return;
869	}
870
871	/*
872	* A this point, if the deferred server is not armed, and the deadline
873	* is in the future, if it is not running already, throttle the server
874	* and arm the defer timer.
875	*/
876	if (dl_se->dl_defer && !dl_se->dl_defer_running &&
877	dl_time_before(a: rq_clock(rq: dl_se->rq), b: dl_se->deadline - dl_se->runtime)) {
878	if (!is_dl_boosted(dl_se)) {
879
880	/*
881	* Set dl_se->dl_defer_armed and dl_throttled variables to
882	* inform the start_dl_timer() that this is a deferred
883	* activation.
884	*/
885	dl_se->dl_defer_armed = `1`;
886	dl_se->dl_throttled = `1`;
887	if (!start_dl_timer(dl_se)) {
888	/*
889	* If for whatever reason (delays), a previous timer was
890	* queued but not serviced, cancel it and clean the
891	* deferrable server variables intended for start_dl_timer().
892	*/
893	hrtimer_try_to_cancel(timer: &dl_se->dl_timer);
894	dl_se->dl_defer_armed = `0`;
895	dl_se->dl_throttled = `0`;
896	}
897	}
898	}
899	}
900
901	/*
902	* Here we check if --at time t-- an entity (which is probably being
903	* [re]activated or, in general, enqueued) can use its remaining runtime
904	* and its current deadline _without_ exceeding the bandwidth it is
905	* assigned (function returns true if it can't). We are in fact applying
906	* one of the CBS rules: when a task wakes up, if the residual runtime
907	* over residual deadline fits within the allocated bandwidth, then we
908	* can keep the current (absolute) deadline and residual budget without
909	* disrupting the schedulability of the system. Otherwise, we should
910	* refill the runtime and set the deadline a period in the future,
911	* because keeping the current (absolute) deadline of the task would
912	* result in breaking guarantees promised to other tasks (refer to
913	* Documentation/scheduler/sched-deadline.rst for more information).
914	*
915	* This function returns true if:
916	*
917	* runtime / (deadline - t) > dl_runtime / dl_deadline ,
918	*
919	* IOW we can't recycle current parameters.
920	*
921	* Notice that the bandwidth check is done against the deadline. For
922	* task with deadline equal to period this is the same of using
923	* dl_period instead of dl_deadline in the equation above.
924	*/
925	static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
926	{
927	u64 left, right;
928
929	/*
930	* left and right are the two sides of the equation above,
931	* after a bit of shuffling to use multiplications instead
932	* of divisions.
933	*
934	* Note that none of the time values involved in the two
935	* multiplications are absolute: dl_deadline and dl_runtime
936	* are the relative deadline and the maximum runtime of each
937	* instance, runtime is the runtime left for the last instance
938	* and (deadline - t), since t is rq->clock, is the time left
939	* to the (absolute) deadline. Even if overflowing the u64 type
940	* is very unlikely to occur in both cases, here we scale down
941	* as we want to avoid that risk at all. Scaling down by 10
942	* means that we reduce granularity to 1us. We are fine with it,
943	* since this is only a true/false check and, anyway, thinking
944	* of anything below microseconds resolution is actually fiction
945	* (but still we want to give the user that illusion >;).
946	*/
947	left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
948	right = ((dl_se->deadline - t) >> DL_SCALE) *
949	(pi_of(dl_se)->dl_runtime >> DL_SCALE);
950
951	return dl_time_before(a: right, b: left);
952	}
953
954	/*
955	* Revised wakeup rule [1]: For self-suspending tasks, rather then
956	* re-initializing task's runtime and deadline, the revised wakeup
957	* rule adjusts the task's runtime to avoid the task to overrun its
958	* density.
959	*
960	* Reasoning: a task may overrun the density if:
961	* runtime / (deadline - t) > dl_runtime / dl_deadline
962	*
963	* Therefore, runtime can be adjusted to:
964	* runtime = (dl_runtime / dl_deadline) * (deadline - t)
965	*
966	* In such way that runtime will be equal to the maximum density
967	* the task can use without breaking any rule.
968	*
969	* [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
970	* bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
971	*/
972	static void
973	update_dl_revised_wakeup(struct sched_dl_entity dl_se, struct* rq *rq)
974	{
975	u64 laxity = dl_se->deadline - rq_clock(rq);
976
977	/*
978	* If the task has deadline < period, and the deadline is in the past,
979	* it should already be throttled before this check.
980	*
981	* See update_dl_entity() comments for further details.
982	*/
983	WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
984
985	dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
986	}
987
988	/*
989	* Regarding the deadline, a task with implicit deadline has a relative
990	* deadline == relative period. A task with constrained deadline has a
991	* relative deadline <= relative period.
992	*
993	* We support constrained deadline tasks. However, there are some restrictions
994	* applied only for tasks which do not have an implicit deadline. See
995	* update_dl_entity() to know more about such restrictions.
996	*
997	* The dl_is_implicit() returns true if the task has an implicit deadline.
998	*/
999	static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1000	{
1001	return dl_se->dl_deadline == dl_se->dl_period;
1002	}
1003
1004	/*
1005	* When a deadline entity is placed in the runqueue, its runtime and deadline
1006	* might need to be updated. This is done by a CBS wake up rule. There are two
1007	* different rules: 1) the original CBS; and 2) the Revisited CBS.
1008	*
1009	* When the task is starting a new period, the Original CBS is used. In this
1010	* case, the runtime is replenished and a new absolute deadline is set.
1011	*
1012	* When a task is queued before the begin of the next period, using the
1013	* remaining runtime and deadline could make the entity to overflow, see
1014	* dl_entity_overflow() to find more about runtime overflow. When such case
1015	* is detected, the runtime and deadline need to be updated.
1016	*
1017	* If the task has an implicit deadline, i.e., deadline == period, the Original
1018	* CBS is applied. The runtime is replenished and a new absolute deadline is
1019	* set, as in the previous cases.
1020	*
1021	* However, the Original CBS does not work properly for tasks with
1022	* deadline < period, which are said to have a constrained deadline. By
1023	* applying the Original CBS, a constrained deadline task would be able to run
1024	* runtime/deadline in a period. With deadline < period, the task would
1025	* overrun the runtime/period allowed bandwidth, breaking the admission test.
1026	*
1027	* In order to prevent this misbehave, the Revisited CBS is used for
1028	* constrained deadline tasks when a runtime overflow is detected. In the
1029	* Revisited CBS, rather than replenishing & setting a new absolute deadline,
1030	* the remaining runtime of the task is reduced to avoid runtime overflow.
1031	* Please refer to the comments update_dl_revised_wakeup() function to find
1032	* more about the Revised CBS rule.
1033	*/
1034	static void update_dl_entity(struct sched_dl_entity *dl_se)
1035	{
1036	struct rq *rq = rq_of_dl_se(dl_se);
1037
1038	if (dl_time_before(a: dl_se->deadline, b: rq_clock(rq)) \|\|
1039	dl_entity_overflow(dl_se, t: rq_clock(rq))) {
1040
1041	if (unlikely(!dl_is_implicit(dl_se) &&
1042	!dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1043	!is_dl_boosted(dl_se))) {
1044	update_dl_revised_wakeup(dl_se, rq);
1045	return;
1046	}
1047
1048	replenish_dl_new_period(dl_se, rq);
1049	} else if (dl_server(dl_se) && dl_se->dl_defer) {
1050	/*
1051	* The server can still use its previous deadline, so check if
1052	* it left the dl_defer_running state.
1053	*/
1054	if (!dl_se->dl_defer_running) {
1055	dl_se->dl_defer_armed = `1`;
1056	dl_se->dl_throttled = `1`;
1057	}
1058	}
1059	}
1060
1061	static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1062	{
1063	return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1064	}
1065
1066	/*
1067	* If the entity depleted all its runtime, and if we want it to sleep
1068	* while waiting for some new execution time to become available, we
1069	* set the bandwidth replenishment timer to the replenishment instant
1070	* and try to activate it.
1071	*
1072	* Notice that it is important for the caller to know if the timer
1073	* actually started or not (i.e., the replenishment instant is in
1074	* the future or in the past).
1075	*/
1076	static int start_dl_timer(struct sched_dl_entity *dl_se)
1077	{
1078	struct hrtimer *timer = &dl_se->dl_timer;
1079	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1080	struct rq *rq = rq_of_dl_rq(dl_rq);
1081	ktime_t now, act;
1082	s64 delta;
1083
1084	lockdep_assert_rq_held(rq);
1085
1086	/*
1087	* We want the timer to fire at the deadline, but considering
1088	* that it is actually coming from rq->clock and not from
1089	* hrtimer's time base reading.
1090	*
1091	* The deferred reservation will have its timer set to
1092	* (deadline - runtime). At that point, the CBS rule will decide
1093	* if the current deadline can be used, or if a replenishment is
1094	* required to avoid add too much pressure on the system
1095	* (current u > U).
1096	*/
1097	if (dl_se->dl_defer_armed) {
1098	WARN_ON_ONCE(!dl_se->dl_throttled);
1099	act = ns_to_ktime(ns: dl_se->deadline - dl_se->runtime);
1100	} else {
1101	/ act = deadline - rel-deadline + period /
1102	act = ns_to_ktime(ns: dl_next_period(dl_se));
1103	}
1104
1105	now = hrtimer_cb_get_time(timer);
1106	delta = ktime_to_ns(kt: now) - rq_clock(rq);
1107	act = ktime_add_ns(act, delta);
1108
1109	/*
1110	* If the expiry time already passed, e.g., because the value
1111	* chosen as the deadline is too small, don't even try to
1112	* start the timer in the past!
1113	*/
1114	if (ktime_us_delta(later: act, earlier: now) < `0`)
1115	return `0`;
1116
1117	/*
1118	* !enqueued will guarantee another callback; even if one is already in
1119	* progress. This ensures a balanced {get,put}_task_struct().
1120	*
1121	* The race against __run_timer() clearing the enqueued state is
1122	* harmless because we're holding task_rq()->lock, therefore the timer
1123	* expiring after we've done the check will wait on its task_rq_lock()
1124	* and observe our state.
1125	*/
1126	if (!hrtimer_is_queued(timer)) {
1127	if (!dl_server(dl_se))
1128	get_task_struct(t: dl_task_of(dl_se));
1129	hrtimer_start(timer, tim: act, mode: HRTIMER_MODE_ABS_HARD);
1130	}
1131
1132	return `1`;
1133	}
1134
1135	static void __push_dl_task(struct rq rq, struct* rq_flags *rf)
1136	{
1137	/*
1138	* Queueing this task back might have overloaded rq, check if we need
1139	* to kick someone away.
1140	*/
1141	if (has_pushable_dl_tasks(rq)) {
1142	/*
1143	* Nothing relies on rq->lock after this, so its safe to drop
1144	* rq->lock.
1145	*/
1146	rq_unpin_lock(rq, rf);
1147	push_dl_task(rq);
1148	rq_repin_lock(rq, rf);
1149	}
1150	}
1151
1152	/ a defer timer will not be reset if the runtime consumed was < dl_server_min_res /
1153	static const u64 dl_server_min_res = `1` * NSEC_PER_MSEC;
1154
1155	static enum hrtimer_restart dl_server_timer(struct hrtimer timer, struct* sched_dl_entity *dl_se)
1156	{
1157	struct rq *rq = rq_of_dl_se(dl_se);
1158	u64 fw;
1159
1160	scoped_guard (rq_lock, rq) {
1161	struct rq_flags *rf = &scope.rf;
1162
1163	if (!dl_se->dl_throttled \|\| !dl_se->dl_runtime)
1164	return HRTIMER_NORESTART;
1165
1166	sched_clock_tick();
1167	update_rq_clock(rq);
1168
1169	if (!dl_se->dl_runtime)
1170	return HRTIMER_NORESTART;
1171
1172	if (dl_se->dl_defer_armed) {
1173	/*
1174	* First check if the server could consume runtime in background.
1175	* If so, it is possible to push the defer timer for this amount
1176	* of time. The dl_server_min_res serves as a limit to avoid
1177	* forwarding the timer for a too small amount of time.
1178	*/
1179	if (dl_time_before(a: rq_clock(rq: dl_se->rq),
1180	b: (dl_se->deadline - dl_se->runtime - dl_server_min_res))) {
1181
1182	/ reset the defer timer /
1183	fw = dl_se->deadline - rq_clock(rq: dl_se->rq) - dl_se->runtime;
1184
1185	hrtimer_forward_now(timer, interval: ns_to_ktime(ns: fw));
1186	return HRTIMER_RESTART;
1187	}
1188
1189	dl_se->dl_defer_running = `1`;
1190	}
1191
1192	enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1193
1194	if (!dl_task(p: dl_se->rq->curr) \|\| dl_entity_preempt(a: dl_se, b: &dl_se->rq->curr->dl))
1195	resched_curr(rq);
1196
1197	__push_dl_task(rq, rf);
1198	}
1199
1200	return HRTIMER_NORESTART;
1201	}
1202
1203	/*
1204	* This is the bandwidth enforcement timer callback. If here, we know
1205	* a task is not on its dl_rq, since the fact that the timer was running
1206	* means the task is throttled and needs a runtime replenishment.
1207	*
1208	* However, what we actually do depends on the fact the task is active,
1209	* (it is on its rq) or has been removed from there by a call to
1210	* dequeue_task_dl(). In the former case we must issue the runtime
1211	* replenishment and add the task back to the dl_rq; in the latter, we just
1212	* do nothing but clearing dl_throttled, so that runtime and deadline
1213	* updating (and the queueing back to dl_rq) will be done by the
1214	* next call to enqueue_task_dl().
1215	*/
1216	static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1217	{
1218	struct sched_dl_entity *dl_se = container_of(timer,
1219	struct sched_dl_entity,
1220	dl_timer);
1221	struct task_struct *p;
1222	struct rq_flags rf;
1223	struct rq *rq;
1224
1225	if (dl_server(dl_se))
1226	return dl_server_timer(timer, dl_se);
1227
1228	p = dl_task_of(dl_se);
1229	rq = task_rq_lock(p, rf: &rf);
1230
1231	/*
1232	* The task might have changed its scheduling policy to something
1233	* different than SCHED_DEADLINE (through switched_from_dl()).
1234	*/
1235	if (!dl_task(p))
1236	goto unlock;
1237
1238	/*
1239	* The task might have been boosted by someone else and might be in the
1240	* boosting/deboosting path, its not throttled.
1241	*/
1242	if (is_dl_boosted(dl_se))
1243	goto unlock;
1244
1245	/*
1246	* Spurious timer due to start_dl_timer() race; or we already received
1247	* a replenishment from rt_mutex_setprio().
1248	*/
1249	if (!dl_se->dl_throttled)
1250	goto unlock;
1251
1252	sched_clock_tick();
1253	update_rq_clock(rq);
1254
1255	/*
1256	* If the throttle happened during sched-out; like:
1257	*
1258	* schedule()
1259	* deactivate_task()
1260	* dequeue_task_dl()
1261	* update_curr_dl()
1262	* start_dl_timer()
1263	* __dequeue_task_dl()
1264	* prev->on_rq = 0;
1265	*
1266	* We can be both throttled and !queued. Replenish the counter
1267	* but do not enqueue -- wait for our wakeup to do that.
1268	*/
1269	if (!task_on_rq_queued(p)) {
1270	replenish_dl_entity(dl_se);
1271	goto unlock;
1272	}
1273
1274	if (unlikely(!rq->online)) {
1275	/*
1276	* If the runqueue is no longer available, migrate the
1277	* task elsewhere. This necessarily changes rq.
1278	*/
1279	lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1280	rq = dl_task_offline_migration(rq, p);
1281	rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1282	update_rq_clock(rq);
1283
1284	/*
1285	* Now that the task has been migrated to the new RQ and we
1286	* have that locked, proceed as normal and enqueue the task
1287	* there.
1288	*/
1289	}
1290
1291	enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1292	if (dl_task(p: rq->donor))
1293	wakeup_preempt_dl(rq, p, flags: `0`);
1294	else
1295	resched_curr(rq);
1296
1297	__push_dl_task(rq, rf: &rf);
1298
1299	unlock:
1300	task_rq_unlock(rq, p, rf: &rf);
1301
1302	/*
1303	* This can free the task_struct, including this hrtimer, do not touch
1304	* anything related to that after this.
1305	*/
1306	put_task_struct(t: p);
1307
1308	return HRTIMER_NORESTART;
1309	}
1310
1311	static void init_dl_task_timer(struct sched_dl_entity *dl_se)
1312	{
1313	struct hrtimer *timer = &dl_se->dl_timer;
1314
1315	hrtimer_setup(timer, function: dl_task_timer, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL_HARD);
1316	}
1317
1318	/*
1319	* During the activation, CBS checks if it can reuse the current task's
1320	* runtime and period. If the deadline of the task is in the past, CBS
1321	* cannot use the runtime, and so it replenishes the task. This rule
1322	* works fine for implicit deadline tasks (deadline == period), and the
1323	* CBS was designed for implicit deadline tasks. However, a task with
1324	* constrained deadline (deadline < period) might be awakened after the
1325	* deadline, but before the next period. In this case, replenishing the
1326	* task would allow it to run for runtime / deadline. As in this case
1327	* deadline < period, CBS enables a task to run for more than the
1328	* runtime / period. In a very loaded system, this can cause a domino
1329	* effect, making other tasks miss their deadlines.
1330	*
1331	* To avoid this problem, in the activation of a constrained deadline
1332	* task after the deadline but before the next period, throttle the
1333	* task and set the replenishing timer to the begin of the next period,
1334	* unless it is boosted.
1335	*/
1336	static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1337	{
1338	struct rq *rq = rq_of_dl_se(dl_se);
1339
1340	if (dl_time_before(a: dl_se->deadline, b: rq_clock(rq)) &&
1341	dl_time_before(a: rq_clock(rq), b: dl_next_period(dl_se))) {
1342	if (unlikely(is_dl_boosted(dl_se) \|\| !start_dl_timer(dl_se)))
1343	return;
1344	dl_se->dl_throttled = `1`;
1345	if (dl_se->runtime > `0`)
1346	dl_se->runtime = `0`;
1347	}
1348	}
1349
1350	static
1351	int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1352	{
1353	return (dl_se->runtime <= `0`);
1354	}
1355
1356	/*
1357	* This function implements the GRUB accounting rule. According to the
1358	* GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1359	* but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1360	* where u is the utilization of the task, Umax is the maximum reclaimable
1361	* utilization, Uinact is the (per-runqueue) inactive utilization, computed
1362	* as the difference between the "total runqueue utilization" and the
1363	* "runqueue active utilization", and Uextra is the (per runqueue) extra
1364	* reclaimable utilization.
1365	* Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1366	* by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1367	* Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1368	* is multiplied by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1369	* Since delta is a 64 bit variable, to have an overflow its value should be
1370	* larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1371	* not an issue here.
1372	*/
1373	static u64 grub_reclaim(u64 delta, struct rq rq, struct* sched_dl_entity *dl_se)
1374	{
1375	u64 u_act;
1376	u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; / Utot - Uact /
1377
1378	/*
1379	* Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1380	* compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1381	* can be larger than u_max. So, u_max - u_inact - u_extra would be
1382	* negative leading to wrong results.
1383	*/
1384	if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1385	u_act = dl_se->dl_bw;
1386	else
1387	u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1388
1389	u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1390	return (delta * u_act) >> BW_SHIFT;
1391	}
1392
1393	s64 dl_scaled_delta_exec(struct rq rq, struct* sched_dl_entity *dl_se, s64 delta_exec)
1394	{
1395	s64 scaled_delta_exec;
1396
1397	/*
1398	* For tasks that participate in GRUB, we implement GRUB-PA: the
1399	* spare reclaimed bandwidth is used to clock down frequency.
1400	*
1401	* For the others, we still need to scale reservation parameters
1402	* according to current frequency and CPU maximum capacity.
1403	*/
1404	if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1405	scaled_delta_exec = grub_reclaim(delta: delta_exec, rq, dl_se);
1406	} else {
1407	int cpu = cpu_of(rq);
1408	unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1409	unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1410
1411	scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1412	scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1413	}
1414
1415	return scaled_delta_exec;
1416	}
1417
1418	static inline void
1419	update_stats_dequeue_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se,
1420	int flags);
1421	static void update_curr_dl_se(struct rq rq, struct* sched_dl_entity *dl_se, s64 delta_exec)
1422	{
1423	s64 scaled_delta_exec;
1424
1425	if (unlikely(delta_exec <= `0`)) {
1426	if (unlikely(dl_se->dl_yielded))
1427	goto throttle;
1428	return;
1429	}
1430
1431	if (dl_server(dl_se) && dl_se->dl_throttled && !dl_se->dl_defer)
1432	return;
1433
1434	if (dl_entity_is_special(dl_se))
1435	return;
1436
1437	scaled_delta_exec = delta_exec;
1438	if (!dl_server(dl_se))
1439	scaled_delta_exec = dl_scaled_delta_exec(rq, dl_se, delta_exec);
1440
1441	dl_se->runtime -= scaled_delta_exec;
1442
1443	/*
1444	* The fair server can consume its runtime while throttled (not queued/
1445	* running as regular CFS).
1446	*
1447	* If the server consumes its entire runtime in this state. The server
1448	* is not required for the current period. Thus, reset the server by
1449	* starting a new period, pushing the activation.
1450	*/
1451	if (dl_se->dl_defer && dl_se->dl_throttled && dl_runtime_exceeded(dl_se)) {
1452	/*
1453	* If the server was previously activated - the starving condition
1454	* took place, it this point it went away because the fair scheduler
1455	* was able to get runtime in background. So return to the initial
1456	* state.
1457	*/
1458	dl_se->dl_defer_running = `0`;
1459
1460	hrtimer_try_to_cancel(timer: &dl_se->dl_timer);
1461
1462	replenish_dl_new_period(dl_se, rq: dl_se->rq);
1463
1464	/*
1465	* Not being able to start the timer seems problematic. If it could not
1466	* be started for whatever reason, we need to "unthrottle" the DL server
1467	* and queue right away. Otherwise nothing might queue it. That's similar
1468	* to what enqueue_dl_entity() does on start_dl_timer==0. For now, just warn.
1469	*/
1470	WARN_ON_ONCE(!start_dl_timer(dl_se));
1471
1472	return;
1473	}
1474
1475	throttle:
1476	if (dl_runtime_exceeded(dl_se) \|\| dl_se->dl_yielded) {
1477	dl_se->dl_throttled = `1`;
1478
1479	/ If requested, inform the user about runtime overruns. /
1480	if (dl_runtime_exceeded(dl_se) &&
1481	(dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1482	dl_se->dl_overrun = `1`;
1483
1484	dequeue_dl_entity(dl_se, flags: `0`);
1485	if (!dl_server(dl_se)) {
1486	update_stats_dequeue_dl(dl_rq: &rq->dl, dl_se, flags: `0`);
1487	dequeue_pushable_dl_task(rq, p: dl_task_of(dl_se));
1488	}
1489
1490	if (unlikely(is_dl_boosted(dl_se) \|\| !start_dl_timer(dl_se))) {
1491	if (dl_server(dl_se)) {
1492	replenish_dl_new_period(dl_se, rq);
1493	start_dl_timer(dl_se);
1494	} else {
1495	enqueue_task_dl(rq, p: dl_task_of(dl_se), ENQUEUE_REPLENISH);
1496	}
1497	}
1498
1499	if (!is_leftmost(dl_se, dl_rq: &rq->dl))
1500	resched_curr(rq);
1501	}
1502
1503	/*
1504	* The fair server (sole dl_server) does not account for real-time
1505	* workload because it is running fair work.
1506	*/
1507	if (dl_se == &rq->fair_server)
1508	return;
1509
1510	#ifdef CONFIG_RT_GROUP_SCHED
1511	/*
1512	* Because -- for now -- we share the rt bandwidth, we need to
1513	* account our runtime there too, otherwise actual rt tasks
1514	* would be able to exceed the shared quota.
1515	*
1516	* Account to the root rt group for now.
1517	*
1518	* The solution we're working towards is having the RT groups scheduled
1519	* using deadline servers -- however there's a few nasties to figure
1520	* out before that can happen.
1521	*/
1522	if (rt_bandwidth_enabled()) {
1523	struct rt_rq *rt_rq = &rq->rt;
1524
1525	raw_spin_lock(&rt_rq->rt_runtime_lock);
1526	/*
1527	* We'll let actual RT tasks worry about the overflow here, we
1528	* have our own CBS to keep us inline; only account when RT
1529	* bandwidth is relevant.
1530	*/
1531	if (sched_rt_bandwidth_account(rt_rq))
1532	rt_rq->rt_time += delta_exec;
1533	raw_spin_unlock(&rt_rq->rt_runtime_lock);
1534	}
1535	#endif /* CONFIG_RT_GROUP_SCHED */
1536	}
1537
1538	/*
1539	* In the non-defer mode, the idle time is not accounted, as the
1540	* server provides a guarantee.
1541	*
1542	* If the dl_server is in defer mode, the idle time is also considered
1543	* as time available for the fair server, avoiding a penalty for the
1544	* rt scheduler that did not consumed that time.
1545	*/
1546	void dl_server_update_idle_time(struct rq rq, struct* task_struct *p)
1547	{
1548	s64 delta_exec;
1549
1550	if (!rq->fair_server.dl_defer)
1551	return;
1552
1553	/ no need to discount more /
1554	if (rq->fair_server.runtime < `0`)
1555	return;
1556
1557	delta_exec = rq_clock_task(rq) - p->se.exec_start;
1558	if (delta_exec < `0`)
1559	return;
1560
1561	rq->fair_server.runtime -= delta_exec;
1562
1563	if (rq->fair_server.runtime < `0`) {
1564	rq->fair_server.dl_defer_running = `0`;
1565	rq->fair_server.runtime = `0`;
1566	}
1567
1568	p->se.exec_start = rq_clock_task(rq);
1569	}
1570
1571	void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec)
1572	{
1573	/ 0 runtime = fair server disabled /
1574	if (dl_se->dl_runtime)
1575	update_curr_dl_se(rq: dl_se->rq, dl_se, delta_exec);
1576	}
1577
1578	void dl_server_start(struct sched_dl_entity *dl_se)
1579	{
1580	struct rq *rq = dl_se->rq;
1581
1582	if (!dl_server(dl_se) \|\| dl_se->dl_server_active)
1583	return;
1584
1585	dl_se->dl_server_active = `1`;
1586	enqueue_dl_entity(dl_se, ENQUEUE_WAKEUP);
1587	if (!dl_task(p: dl_se->rq->curr) \|\| dl_entity_preempt(a: dl_se, b: &rq->curr->dl))
1588	resched_curr(rq: dl_se->rq);
1589	}
1590
1591	void dl_server_stop(struct sched_dl_entity *dl_se)
1592	{
1593	if (!dl_server(dl_se) \|\| !dl_server_active(dl_se))
1594	return;
1595
1596	dequeue_dl_entity(dl_se, DEQUEUE_SLEEP);
1597	hrtimer_try_to_cancel(timer: &dl_se->dl_timer);
1598	dl_se->dl_defer_armed = `0`;
1599	dl_se->dl_throttled = `0`;
1600	dl_se->dl_server_active = `0`;
1601	}
1602
1603	void dl_server_init(struct sched_dl_entity dl_se, struct* rq *rq,
1604	dl_server_pick_f pick_task)
1605	{
1606	dl_se->rq = rq;
1607	dl_se->server_pick_task = pick_task;
1608	}
1609
1610	void sched_init_dl_servers(void)
1611	{
1612	int cpu;
1613	struct rq *rq;
1614	struct sched_dl_entity *dl_se;
1615
1616	for_each_online_cpu(cpu) {
1617	u64 runtime = `50` * NSEC_PER_MSEC;
1618	u64 period = `1000` * NSEC_PER_MSEC;
1619
1620	rq = cpu_rq(cpu);
1621
1622	guard(rq_lock_irq)(l: rq);
1623
1624	dl_se = &rq->fair_server;
1625
1626	WARN_ON(dl_server(dl_se));
1627
1628	dl_server_apply_params(dl_se, runtime, period, init: `1`);
1629
1630	dl_se->dl_server = `1`;
1631	dl_se->dl_defer = `1`;
1632	setup_new_dl_entity(dl_se);
1633	}
1634	}
1635
1636	void __dl_server_attach_root(struct sched_dl_entity dl_se, struct* rq *rq)
1637	{
1638	u64 new_bw = dl_se->dl_bw;
1639	int cpu = cpu_of(rq);
1640	struct dl_bw *dl_b;
1641
1642	dl_b = dl_bw_of(i: cpu_of(rq));
1643	guard(raw_spinlock)(l: &dl_b->lock);
1644
1645	if (!dl_bw_cpus(i: cpu))
1646	return;
1647
1648	__dl_add(dl_b, tsk_bw: new_bw, cpus: dl_bw_cpus(i: cpu));
1649	}
1650
1651	int dl_server_apply_params(struct sched_dl_entity *dl_se, u64 runtime, u64 period, bool init)
1652	{
1653	u64 old_bw = init ? `0` : to_ratio(period: dl_se->dl_period, runtime: dl_se->dl_runtime);
1654	u64 new_bw = to_ratio(period, runtime);
1655	struct rq *rq = dl_se->rq;
1656	int cpu = cpu_of(rq);
1657	struct dl_bw *dl_b;
1658	unsigned long cap;
1659	int retval = `0`;
1660	int cpus;
1661
1662	dl_b = dl_bw_of(i: cpu);
1663	guard(raw_spinlock)(l: &dl_b->lock);
1664
1665	cpus = dl_bw_cpus(i: cpu);
1666	cap = dl_bw_capacity(i: cpu);
1667
1668	if (__dl_overflow(dl_b, cap, old_bw, new_bw))
1669	return -EBUSY;
1670
1671	if (init) {
1672	__add_rq_bw(dl_bw: new_bw, dl_rq: &rq->dl);
1673	__dl_add(dl_b, tsk_bw: new_bw, cpus);
1674	} else {
1675	__dl_sub(dl_b, tsk_bw: dl_se->dl_bw, cpus);
1676	__dl_add(dl_b, tsk_bw: new_bw, cpus);
1677
1678	dl_rq_change_utilization(rq, dl_se, new_bw);
1679	}
1680
1681	dl_se->dl_runtime = runtime;
1682	dl_se->dl_deadline = period;
1683	dl_se->dl_period = period;
1684
1685	dl_se->runtime = `0`;
1686	dl_se->deadline = `0`;
1687
1688	dl_se->dl_bw = to_ratio(period: dl_se->dl_period, runtime: dl_se->dl_runtime);
1689	dl_se->dl_density = to_ratio(period: dl_se->dl_deadline, runtime: dl_se->dl_runtime);
1690
1691	return retval;
1692	}
1693
1694	/*
1695	* Update the current task's runtime statistics (provided it is still
1696	* a -deadline task and has not been removed from the dl_rq).
1697	*/
1698	static void update_curr_dl(struct rq *rq)
1699	{
1700	struct task_struct *donor = rq->donor;
1701	struct sched_dl_entity *dl_se = &donor->dl;
1702	s64 delta_exec;
1703
1704	if (!dl_task(p: donor) \|\| !on_dl_rq(dl_se))
1705	return;
1706
1707	/*
1708	* Consumed budget is computed considering the time as
1709	* observed by schedulable tasks (excluding time spent
1710	* in hardirq context, etc.). Deadlines are instead
1711	* computed using hard walltime. This seems to be the more
1712	* natural solution, but the full ramifications of this
1713	* approach need further study.
1714	*/
1715	delta_exec = update_curr_common(rq);
1716	update_curr_dl_se(rq, dl_se, delta_exec);
1717	}
1718
1719	static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1720	{
1721	struct sched_dl_entity *dl_se = container_of(timer,
1722	struct sched_dl_entity,
1723	inactive_timer);
1724	struct task_struct *p = NULL;
1725	struct rq_flags rf;
1726	struct rq *rq;
1727
1728	if (!dl_server(dl_se)) {
1729	p = dl_task_of(dl_se);
1730	rq = task_rq_lock(p, rf: &rf);
1731	} else {
1732	rq = dl_se->rq;
1733	rq_lock(rq, rf: &rf);
1734	}
1735
1736	sched_clock_tick();
1737	update_rq_clock(rq);
1738
1739	if (dl_server(dl_se))
1740	goto no_task;
1741
1742	if (!dl_task(p) \|\| READ_ONCE(p->__state) == TASK_DEAD) {
1743	struct dl_bw *dl_b = dl_bw_of(i: task_cpu(p));
1744
1745	if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1746	sub_running_bw(dl_se: &p->dl, dl_rq: dl_rq_of_se(dl_se: &p->dl));
1747	sub_rq_bw(dl_se: &p->dl, dl_rq: dl_rq_of_se(dl_se: &p->dl));
1748	dl_se->dl_non_contending = `0`;
1749	}
1750
1751	raw_spin_lock(&dl_b->lock);
1752	__dl_sub(dl_b, tsk_bw: p->dl.dl_bw, cpus: dl_bw_cpus(i: task_cpu(p)));
1753	raw_spin_unlock(&dl_b->lock);
1754	__dl_clear_params(dl_se);
1755
1756	goto unlock;
1757	}
1758
1759	no_task:
1760	if (dl_se->dl_non_contending == `0`)
1761	goto unlock;
1762
1763	sub_running_bw(dl_se, dl_rq: &rq->dl);
1764	dl_se->dl_non_contending = `0`;
1765	unlock:
1766
1767	if (!dl_server(dl_se)) {
1768	task_rq_unlock(rq, p, rf: &rf);
1769	put_task_struct(t: p);
1770	} else {
1771	rq_unlock(rq, rf: &rf);
1772	}
1773
1774	return HRTIMER_NORESTART;
1775	}
1776
1777	static void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1778	{
1779	struct hrtimer *timer = &dl_se->inactive_timer;
1780
1781	hrtimer_setup(timer, function: inactive_task_timer, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL_HARD);
1782	}
1783
1784	#define __node_2_dle(node) \
1785	rb_entry((node), struct sched_dl_entity, rb_node)
1786
1787	static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1788	{
1789	struct rq *rq = rq_of_dl_rq(dl_rq);
1790
1791	if (dl_rq->earliest_dl.curr == `0` \|\|
1792	dl_time_before(a: deadline, b: dl_rq->earliest_dl.curr)) {
1793	if (dl_rq->earliest_dl.curr == `0`)
1794	cpupri_set(cp: &rq->rd->cpupri, cpu: rq->cpu, CPUPRI_HIGHER);
1795	dl_rq->earliest_dl.curr = deadline;
1796	cpudl_set(cp: &rq->rd->cpudl, cpu: rq->cpu, dl: deadline);
1797	}
1798	}
1799
1800	static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1801	{
1802	struct rq *rq = rq_of_dl_rq(dl_rq);
1803
1804	/*
1805	* Since we may have removed our earliest (and/or next earliest)
1806	* task we must recompute them.
1807	*/
1808	if (!dl_rq->dl_nr_running) {
1809	dl_rq->earliest_dl.curr = `0`;
1810	dl_rq->earliest_dl.next = `0`;
1811	cpudl_clear(cp: &rq->rd->cpudl, cpu: rq->cpu);
1812	cpupri_set(cp: &rq->rd->cpupri, cpu: rq->cpu, pri: rq->rt.highest_prio.curr);
1813	} else {
1814	struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1815	struct sched_dl_entity *entry = __node_2_dle(leftmost);
1816
1817	dl_rq->earliest_dl.curr = entry->deadline;
1818	cpudl_set(cp: &rq->rd->cpudl, cpu: rq->cpu, dl: entry->deadline);
1819	}
1820	}
1821
1822	static inline
1823	void inc_dl_tasks(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
1824	{
1825	u64 deadline = dl_se->deadline;
1826
1827	dl_rq->dl_nr_running++;
1828
1829	if (!dl_server(dl_se))
1830	add_nr_running(rq: rq_of_dl_rq(dl_rq), count: `1`);
1831
1832	inc_dl_deadline(dl_rq, deadline);
1833	}
1834
1835	static inline
1836	void dec_dl_tasks(struct sched_dl_entity dl_se, struct* dl_rq *dl_rq)
1837	{
1838	WARN_ON(!dl_rq->dl_nr_running);
1839	dl_rq->dl_nr_running--;
1840
1841	if (!dl_server(dl_se))
1842	sub_nr_running(rq: rq_of_dl_rq(dl_rq), count: `1`);
1843
1844	dec_dl_deadline(dl_rq, deadline: dl_se->deadline);
1845	}
1846
1847	static inline bool __dl_less(struct rb_node a, const* struct rb_node *b)
1848	{
1849	return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1850	}
1851
1852	static __always_inline struct sched_statistics *
1853	__schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1854	{
1855	if (!schedstat_enabled())
1856	return NULL;
1857
1858	if (dl_server(dl_se))
1859	return NULL;
1860
1861	return &dl_task_of(dl_se)->stats;
1862	}
1863
1864	static inline void
1865	update_stats_wait_start_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se)
1866	{
1867	struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1868	if (stats)
1869	__update_stats_wait_start(rq: rq_of_dl_rq(dl_rq), p: dl_task_of(dl_se), stats);
1870	}
1871
1872	static inline void
1873	update_stats_wait_end_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se)
1874	{
1875	struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1876	if (stats)
1877	__update_stats_wait_end(rq: rq_of_dl_rq(dl_rq), p: dl_task_of(dl_se), stats);
1878	}
1879
1880	static inline void
1881	update_stats_enqueue_sleeper_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se)
1882	{
1883	struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1884	if (stats)
1885	__update_stats_enqueue_sleeper(rq: rq_of_dl_rq(dl_rq), p: dl_task_of(dl_se), stats);
1886	}
1887
1888	static inline void
1889	update_stats_enqueue_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se,
1890	int flags)
1891	{
1892	if (!schedstat_enabled())
1893	return;
1894
1895	if (flags & ENQUEUE_WAKEUP)
1896	update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1897	}
1898
1899	static inline void
1900	update_stats_dequeue_dl(struct dl_rq dl_rq, struct* sched_dl_entity *dl_se,
1901	int flags)
1902	{
1903	struct task_struct *p = dl_task_of(dl_se);
1904
1905	if (!schedstat_enabled())
1906	return;
1907
1908	if ((flags & DEQUEUE_SLEEP)) {
1909	unsigned int state;
1910
1911	state = READ_ONCE(p->__state);
1912	if (state & TASK_INTERRUPTIBLE)
1913	__schedstat_set(p->stats.sleep_start,
1914	rq_clock(rq_of_dl_rq(dl_rq)));
1915
1916	if (state & TASK_UNINTERRUPTIBLE)
1917	__schedstat_set(p->stats.block_start,
1918	rq_clock(rq_of_dl_rq(dl_rq)));
1919	}
1920	}
1921
1922	static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1923	{
1924	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1925
1926	WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1927
1928	rb_add_cached(node: &dl_se->rb_node, tree: &dl_rq->root, less: __dl_less);
1929
1930	inc_dl_tasks(dl_se, dl_rq);
1931	}
1932
1933	static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1934	{
1935	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1936
1937	if (RB_EMPTY_NODE(&dl_se->rb_node))
1938	return;
1939
1940	rb_erase_cached(node: &dl_se->rb_node, root: &dl_rq->root);
1941
1942	RB_CLEAR_NODE(&dl_se->rb_node);
1943
1944	dec_dl_tasks(dl_se, dl_rq);
1945	}
1946
1947	static void
1948	enqueue_dl_entity(struct sched_dl_entity dl_se, int* flags)
1949	{
1950	WARN_ON_ONCE(on_dl_rq(dl_se));
1951
1952	update_stats_enqueue_dl(dl_rq: dl_rq_of_se(dl_se), dl_se, flags);
1953
1954	/*
1955	* Check if a constrained deadline task was activated
1956	* after the deadline but before the next period.
1957	* If that is the case, the task will be throttled and
1958	* the replenishment timer will be set to the next period.
1959	*/
1960	if (!dl_se->dl_throttled && !dl_is_implicit(dl_se))
1961	dl_check_constrained_dl(dl_se);
1962
1963	if (flags & (ENQUEUE_RESTORE\|ENQUEUE_MIGRATING)) {
1964	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1965
1966	add_rq_bw(dl_se, dl_rq);
1967	add_running_bw(dl_se, dl_rq);
1968	}
1969
1970	/*
1971	* If p is throttled, we do not enqueue it. In fact, if it exhausted
1972	* its budget it needs a replenishment and, since it now is on
1973	* its rq, the bandwidth timer callback (which clearly has not
1974	* run yet) will take care of this.
1975	* However, the active utilization does not depend on the fact
1976	* that the task is on the runqueue or not (but depends on the
1977	* task's state - in GRUB parlance, "inactive" vs "active contending").
1978	* In other words, even if a task is throttled its utilization must
1979	* be counted in the active utilization; hence, we need to call
1980	* add_running_bw().
1981	*/
1982	if (!dl_se->dl_defer && dl_se->dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1983	if (flags & ENQUEUE_WAKEUP)
1984	task_contending(dl_se, flags);
1985
1986	return;
1987	}
1988
1989	/*
1990	* If this is a wakeup or a new instance, the scheduling
1991	* parameters of the task might need updating. Otherwise,
1992	* we want a replenishment of its runtime.
1993	*/
1994	if (flags & ENQUEUE_WAKEUP) {
1995	task_contending(dl_se, flags);
1996	update_dl_entity(dl_se);
1997	} else if (flags & ENQUEUE_REPLENISH) {
1998	replenish_dl_entity(dl_se);
1999	} else if ((flags & ENQUEUE_RESTORE) &&
2000	!is_dl_boosted(dl_se) &&
2001	dl_time_before(a: dl_se->deadline, b: rq_clock(rq: rq_of_dl_se(dl_se)))) {
2002	setup_new_dl_entity(dl_se);
2003	}
2004
2005	/*
2006	* If the reservation is still throttled, e.g., it got replenished but is a
2007	* deferred task and still got to wait, don't enqueue.
2008	*/
2009	if (dl_se->dl_throttled && start_dl_timer(dl_se))
2010	return;
2011
2012	/*
2013	* We're about to enqueue, make sure we're not ->dl_throttled!
2014	* In case the timer was not started, say because the defer time
2015	* has passed, mark as not throttled and mark unarmed.
2016	* Also cancel earlier timers, since letting those run is pointless.
2017	*/
2018	if (dl_se->dl_throttled) {
2019	hrtimer_try_to_cancel(timer: &dl_se->dl_timer);
2020	dl_se->dl_defer_armed = `0`;
2021	dl_se->dl_throttled = `0`;
2022	}
2023
2024	__enqueue_dl_entity(dl_se);
2025	}
2026
2027	static void dequeue_dl_entity(struct sched_dl_entity dl_se, int* flags)
2028	{
2029	__dequeue_dl_entity(dl_se);
2030
2031	if (flags & (DEQUEUE_SAVE\|DEQUEUE_MIGRATING)) {
2032	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2033
2034	sub_running_bw(dl_se, dl_rq);
2035	sub_rq_bw(dl_se, dl_rq);
2036	}
2037
2038	/*
2039	* This check allows to start the inactive timer (or to immediately
2040	* decrease the active utilization, if needed) in two cases:
2041	* when the task blocks and when it is terminating
2042	* (p->state == TASK_DEAD). We can handle the two cases in the same
2043	* way, because from GRUB's point of view the same thing is happening
2044	* (the task moves from "active contending" to "active non contending"
2045	* or "inactive")
2046	*/
2047	if (flags & DEQUEUE_SLEEP)
2048	task_non_contending(dl_se);
2049	}
2050
2051	static void enqueue_task_dl(struct rq rq, struct* task_struct p, int* flags)
2052	{
2053	if (is_dl_boosted(dl_se: &p->dl)) {
2054	/*
2055	* Because of delays in the detection of the overrun of a
2056	* thread's runtime, it might be the case that a thread
2057	* goes to sleep in a rt mutex with negative runtime. As
2058	* a consequence, the thread will be throttled.
2059	*
2060	* While waiting for the mutex, this thread can also be
2061	* boosted via PI, resulting in a thread that is throttled
2062	* and boosted at the same time.
2063	*
2064	* In this case, the boost overrides the throttle.
2065	*/
2066	if (p->dl.dl_throttled) {
2067	/*
2068	* The replenish timer needs to be canceled. No
2069	* problem if it fires concurrently: boosted threads
2070	* are ignored in dl_task_timer().
2071	*/
2072	cancel_replenish_timer(dl_se: &p->dl);
2073	p->dl.dl_throttled = `0`;
2074	}
2075	} else if (!dl_prio(prio: p->normal_prio)) {
2076	/*
2077	* Special case in which we have a !SCHED_DEADLINE task that is going
2078	* to be deboosted, but exceeds its runtime while doing so. No point in
2079	* replenishing it, as it's going to return back to its original
2080	* scheduling class after this. If it has been throttled, we need to
2081	* clear the flag, otherwise the task may wake up as throttled after
2082	* being boosted again with no means to replenish the runtime and clear
2083	* the throttle.
2084	*/
2085	p->dl.dl_throttled = `0`;
2086	if (!(flags & ENQUEUE_REPLENISH))
2087	printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
2088	task_pid_nr(p));
2089
2090	return;
2091	}
2092
2093	check_schedstat_required();
2094	update_stats_wait_start_dl(dl_rq: dl_rq_of_se(dl_se: &p->dl), dl_se: &p->dl);
2095
2096	if (p->on_rq == TASK_ON_RQ_MIGRATING)
2097	flags \|= ENQUEUE_MIGRATING;
2098
2099	enqueue_dl_entity(dl_se: &p->dl, flags);
2100
2101	if (dl_server(dl_se: &p->dl))
2102	return;
2103
2104	if (task_is_blocked(p))
2105	return;
2106
2107	if (!task_current(rq, p) && !p->dl.dl_throttled && p->nr_cpus_allowed > `1`)
2108	enqueue_pushable_dl_task(rq, p);
2109	}
2110
2111	static bool dequeue_task_dl(struct rq rq, struct* task_struct p, int* flags)
2112	{
2113	update_curr_dl(rq);
2114
2115	if (p->on_rq == TASK_ON_RQ_MIGRATING)
2116	flags \|= DEQUEUE_MIGRATING;
2117
2118	dequeue_dl_entity(dl_se: &p->dl, flags);
2119	if (!p->dl.dl_throttled && !dl_server(dl_se: &p->dl))
2120	dequeue_pushable_dl_task(rq, p);
2121
2122	return true;
2123	}
2124
2125	/*
2126	* Yield task semantic for -deadline tasks is:
2127	*
2128	* get off from the CPU until our next instance, with
2129	* a new runtime. This is of little use now, since we
2130	* don't have a bandwidth reclaiming mechanism. Anyway,
2131	* bandwidth reclaiming is planned for the future, and
2132	* yield_task_dl will indicate that some spare budget
2133	* is available for other task instances to use it.
2134	*/
2135	static void yield_task_dl(struct rq *rq)
2136	{
2137	/*
2138	* We make the task go to sleep until its current deadline by
2139	* forcing its runtime to zero. This way, update_curr_dl() stops
2140	* it and the bandwidth timer will wake it up and will give it
2141	* new scheduling parameters (thanks to dl_yielded=1).
2142	*/
2143	rq->curr->dl.dl_yielded = `1`;
2144
2145	update_rq_clock(rq);
2146	update_curr_dl(rq);
2147	/*
2148	* Tell update_rq_clock() that we've just updated,
2149	* so we don't do microscopic update in schedule()
2150	* and double the fastpath cost.
2151	*/
2152	rq_clock_skip_update(rq);
2153	}
2154
2155	static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
2156	struct rq *rq)
2157	{
2158	return (!rq->dl.dl_nr_running \|\|
2159	dl_time_before(a: p->dl.deadline,
2160	b: rq->dl.earliest_dl.curr));
2161	}
2162
2163	static int find_later_rq(struct task_struct *task);
2164
2165	static int
2166	select_task_rq_dl(struct task_struct p, int* cpu, int flags)
2167	{
2168	struct task_struct curr, donor;
2169	bool select_rq;
2170	struct rq *rq;
2171
2172	if (!(flags & WF_TTWU))
2173	goto out;
2174
2175	rq = cpu_rq(cpu);
2176
2177	rcu_read_lock();
2178	curr = READ_ONCE(rq->curr); / unlocked access /
2179	donor = READ_ONCE(rq->donor);
2180
2181	/*
2182	* If we are dealing with a -deadline task, we must
2183	* decide where to wake it up.
2184	* If it has a later deadline and the current task
2185	* on this rq can't move (provided the waking task
2186	* can!) we prefer to send it somewhere else. On the
2187	* other hand, if it has a shorter deadline, we
2188	* try to make it stay here, it might be important.
2189	*/
2190	select_rq = unlikely(dl_task(donor)) &&
2191	(curr->nr_cpus_allowed < `2` \|\|
2192	!dl_entity_preempt(a: &p->dl, b: &donor->dl)) &&
2193	p->nr_cpus_allowed > `1`;
2194
2195	/*
2196	* Take the capacity of the CPU into account to
2197	* ensure it fits the requirement of the task.
2198	*/
2199	if (sched_asym_cpucap_active())
2200	select_rq \|= !dl_task_fits_capacity(p, cpu);
2201
2202	if (select_rq) {
2203	int target = find_later_rq(task: p);
2204
2205	if (target != -`1` &&
2206	dl_task_is_earliest_deadline(p, cpu_rq(target)))
2207	cpu = target;
2208	}
2209	rcu_read_unlock();
2210
2211	out:
2212	return cpu;
2213	}
2214
2215	static void migrate_task_rq_dl(struct task_struct p, int* new_cpu __maybe_unused)
2216	{
2217	struct rq_flags rf;
2218	struct rq *rq;
2219
2220	if (READ_ONCE(p->__state) != TASK_WAKING)
2221	return;
2222
2223	rq = task_rq(p);
2224	/*
2225	* Since p->state == TASK_WAKING, set_task_cpu() has been called
2226	* from try_to_wake_up(). Hence, p->pi_lock is locked, but
2227	* rq->lock is not... So, lock it
2228	*/
2229	rq_lock(rq, rf: &rf);
2230	if (p->dl.dl_non_contending) {
2231	update_rq_clock(rq);
2232	sub_running_bw(dl_se: &p->dl, dl_rq: &rq->dl);
2233	p->dl.dl_non_contending = `0`;
2234	/*
2235	* If the timer handler is currently running and the
2236	* timer cannot be canceled, inactive_task_timer()
2237	* will see that dl_not_contending is not set, and
2238	* will not touch the rq's active utilization,
2239	* so we are still safe.
2240	*/
2241	cancel_inactive_timer(dl_se: &p->dl);
2242	}
2243	sub_rq_bw(dl_se: &p->dl, dl_rq: &rq->dl);
2244	rq_unlock(rq, rf: &rf);
2245	}
2246
2247	static void check_preempt_equal_dl(struct rq rq, struct* task_struct *p)
2248	{
2249	/*
2250	* Current can't be migrated, useless to reschedule,
2251	* let's hope p can move out.
2252	*/
2253	if (rq->curr->nr_cpus_allowed == `1` \|\|
2254	!cpudl_find(cp: &rq->rd->cpudl, p: rq->donor, NULL))
2255	return;
2256
2257	/*
2258	* p is migratable, so let's not schedule it and
2259	* see if it is pushed or pulled somewhere else.
2260	*/
2261	if (p->nr_cpus_allowed != `1` &&
2262	cpudl_find(cp: &rq->rd->cpudl, p, NULL))
2263	return;
2264
2265	resched_curr(rq);
2266	}
2267
2268	static int balance_dl(struct rq rq, struct* task_struct p, struct* rq_flags *rf)
2269	{
2270	if (!on_dl_rq(dl_se: &p->dl) && need_pull_dl_task(rq, prev: p)) {
2271	/*
2272	* This is OK, because current is on_cpu, which avoids it being
2273	* picked for load-balance and preemption/IRQs are still
2274	* disabled avoiding further scheduler activity on it and we've
2275	* not yet started the picking loop.
2276	*/
2277	rq_unpin_lock(rq, rf);
2278	pull_dl_task(rq);
2279	rq_repin_lock(rq, rf);
2280	}
2281
2282	return sched_stop_runnable(rq) \|\| sched_dl_runnable(rq);
2283	}
2284
2285	/*
2286	* Only called when both the current and waking task are -deadline
2287	* tasks.
2288	*/
2289	static void wakeup_preempt_dl(struct rq rq, struct* task_struct *p,
2290	int flags)
2291	{
2292	if (dl_entity_preempt(a: &p->dl, b: &rq->donor->dl)) {
2293	resched_curr(rq);
2294	return;
2295	}
2296
2297	/*
2298	* In the unlikely case current and p have the same deadline
2299	* let us try to decide what's the best thing to do...
2300	*/
2301	if ((p->dl.deadline == rq->donor->dl.deadline) &&
2302	!test_tsk_need_resched(tsk: rq->curr))
2303	check_preempt_equal_dl(rq, p);
2304	}
2305
2306	#ifdef CONFIG_SCHED_HRTICK
2307	static void start_hrtick_dl(struct rq rq, struct* sched_dl_entity *dl_se)
2308	{
2309	hrtick_start(rq, delay: dl_se->runtime);
2310	}
2311	#else /* !CONFIG_SCHED_HRTICK: */
2312	static void start_hrtick_dl(struct rq rq, struct* sched_dl_entity *dl_se)
2313	{
2314	}
2315	#endif /* !CONFIG_SCHED_HRTICK */
2316
2317	static void set_next_task_dl(struct rq rq, struct* task_struct *p, bool first)
2318	{
2319	struct sched_dl_entity *dl_se = &p->dl;
2320	struct dl_rq *dl_rq = &rq->dl;
2321
2322	p->se.exec_start = rq_clock_task(rq);
2323	if (on_dl_rq(dl_se: &p->dl))
2324	update_stats_wait_end_dl(dl_rq, dl_se);
2325
2326	/ You can't push away the running task /
2327	dequeue_pushable_dl_task(rq, p);
2328
2329	if (!first)
2330	return;
2331
2332	if (rq->donor->sched_class != &dl_sched_class)
2333	update_dl_rq_load_avg(now: rq_clock_pelt(rq), rq, running: `0`);
2334
2335	deadline_queue_push_tasks(rq);
2336
2337	if (hrtick_enabled_dl(rq))
2338	start_hrtick_dl(rq, dl_se: &p->dl);
2339	}
2340
2341	static struct sched_dl_entity pick_next_dl_entity(struct* dl_rq *dl_rq)
2342	{
2343	struct rb_node *left = rb_first_cached(&dl_rq->root);
2344
2345	if (!left)
2346	return NULL;
2347
2348	return __node_2_dle(left);
2349	}
2350
2351	/*
2352	* __pick_next_task_dl - Helper to pick the next -deadline task to run.
2353	* @rq: The runqueue to pick the next task from.
2354	*/
2355	static struct task_struct __pick_task_dl(struct* rq *rq)
2356	{
2357	struct sched_dl_entity *dl_se;
2358	struct dl_rq *dl_rq = &rq->dl;
2359	struct task_struct *p;
2360
2361	again:
2362	if (!sched_dl_runnable(rq))
2363	return NULL;
2364
2365	dl_se = pick_next_dl_entity(dl_rq);
2366	WARN_ON_ONCE(!dl_se);
2367
2368	if (dl_server(dl_se)) {
2369	p = dl_se->server_pick_task(dl_se);
2370	if (!p) {
2371	dl_server_stop(dl_se);
2372	goto again;
2373	}
2374	rq->dl_server = dl_se;
2375	} else {
2376	p = dl_task_of(dl_se);
2377	}
2378
2379	return p;
2380	}
2381
2382	static struct task_struct pick_task_dl(struct* rq *rq)
2383	{
2384	return __pick_task_dl(rq);
2385	}
2386
2387	static void put_prev_task_dl(struct rq rq, struct* task_struct p, struct* task_struct *next)
2388	{
2389	struct sched_dl_entity *dl_se = &p->dl;
2390	struct dl_rq *dl_rq = &rq->dl;
2391
2392	if (on_dl_rq(dl_se: &p->dl))
2393	update_stats_wait_start_dl(dl_rq, dl_se);
2394
2395	update_curr_dl(rq);
2396
2397	update_dl_rq_load_avg(now: rq_clock_pelt(rq), rq, running: `1`);
2398
2399	if (task_is_blocked(p))
2400	return;
2401
2402	if (on_dl_rq(dl_se: &p->dl) && p->nr_cpus_allowed > `1`)
2403	enqueue_pushable_dl_task(rq, p);
2404	}
2405
2406	/*
2407	* scheduler tick hitting a task of our scheduling class.
2408	*
2409	* NOTE: This function can be called remotely by the tick offload that
2410	* goes along full dynticks. Therefore no local assumption can be made
2411	* and everything must be accessed through the @rq and @curr passed in
2412	* parameters.
2413	*/
2414	static void task_tick_dl(struct rq rq, struct* task_struct p, int* queued)
2415	{
2416	update_curr_dl(rq);
2417
2418	update_dl_rq_load_avg(now: rq_clock_pelt(rq), rq, running: `1`);
2419	/*
2420	* Even when we have runtime, update_curr_dl() might have resulted in us
2421	* not being the leftmost task anymore. In that case NEED_RESCHED will
2422	* be set and schedule() will start a new hrtick for the next task.
2423	*/
2424	if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > `0` &&
2425	is_leftmost(dl_se: &p->dl, dl_rq: &rq->dl))
2426	start_hrtick_dl(rq, dl_se: &p->dl);
2427	}
2428
2429	static void task_fork_dl(struct task_struct *p)
2430	{
2431	/*
2432	* SCHED_DEADLINE tasks cannot fork and this is achieved through
2433	* sched_fork()
2434	*/
2435	}
2436
2437	/ Only try algorithms three times /
2438	#define DL_MAX_TRIES 3
2439
2440	/*
2441	* Return the earliest pushable rq's task, which is suitable to be executed
2442	* on the CPU, NULL otherwise:
2443	*/
2444	static struct task_struct pick_earliest_pushable_dl_task(struct* rq rq, int* cpu)
2445	{
2446	struct task_struct *p = NULL;
2447	struct rb_node *next_node;
2448
2449	if (!has_pushable_dl_tasks(rq))
2450	return NULL;
2451
2452	next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2453	while (next_node) {
2454	p = __node_2_pdl(next_node);
2455
2456	if (task_is_pushable(rq, p, cpu))
2457	return p;
2458
2459	next_node = rb_next(next_node);
2460	}
2461
2462	return NULL;
2463	}
2464
2465	static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2466
2467	static int find_later_rq(struct task_struct *task)
2468	{
2469	struct sched_domain *sd;
2470	struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2471	int this_cpu = smp_processor_id();
2472	int cpu = task_cpu(p: task);
2473
2474	/ Make sure the mask is initialized first /
2475	if (unlikely(!later_mask))
2476	return -`1`;
2477
2478	if (task->nr_cpus_allowed == `1`)
2479	return -`1`;
2480
2481	/*
2482	* We have to consider system topology and task affinity
2483	* first, then we can look for a suitable CPU.
2484	*/
2485	if (!cpudl_find(cp: &task_rq(task)->rd->cpudl, p: task, later_mask))
2486	return -`1`;
2487
2488	/*
2489	* If we are here, some targets have been found, including
2490	* the most suitable which is, among the runqueues where the
2491	* current tasks have later deadlines than the task's one, the
2492	* rq with the latest possible one.
2493	*
2494	* Now we check how well this matches with task's
2495	* affinity and system topology.
2496	*
2497	* The last CPU where the task run is our first
2498	* guess, since it is most likely cache-hot there.
2499	*/
2500	if (cpumask_test_cpu(cpu, cpumask: later_mask))
2501	return cpu;
2502	/*
2503	* Check if this_cpu is to be skipped (i.e., it is
2504	* not in the mask) or not.
2505	*/
2506	if (!cpumask_test_cpu(cpu: this_cpu, cpumask: later_mask))
2507	this_cpu = -`1`;
2508
2509	rcu_read_lock();
2510	for_each_domain(cpu, sd) {
2511	if (sd->flags & SD_WAKE_AFFINE) {
2512	int best_cpu;
2513
2514	/*
2515	* If possible, preempting this_cpu is
2516	* cheaper than migrating.
2517	*/
2518	if (this_cpu != -`1` &&
2519	cpumask_test_cpu(cpu: this_cpu, cpumask: sched_domain_span(sd))) {
2520	rcu_read_unlock();
2521	return this_cpu;
2522	}
2523
2524	best_cpu = cpumask_any_and_distribute(src1p: later_mask,
2525	src2p: sched_domain_span(sd));
2526	/*
2527	* Last chance: if a CPU being in both later_mask
2528	* and current sd span is valid, that becomes our
2529	* choice. Of course, the latest possible CPU is
2530	* already under consideration through later_mask.
2531	*/
2532	if (best_cpu < nr_cpu_ids) {
2533	rcu_read_unlock();
2534	return best_cpu;
2535	}
2536	}
2537	}
2538	rcu_read_unlock();
2539
2540	/*
2541	* At this point, all our guesses failed, we just return
2542	* 'something', and let the caller sort the things out.
2543	*/
2544	if (this_cpu != -`1`)
2545	return this_cpu;
2546
2547	cpu = cpumask_any_distribute(srcp: later_mask);
2548	if (cpu < nr_cpu_ids)
2549	return cpu;
2550
2551	return -`1`;
2552	}
2553
2554	static struct task_struct pick_next_pushable_dl_task(struct* rq *rq)
2555	{
2556	struct task_struct *p;
2557
2558	if (!has_pushable_dl_tasks(rq))
2559	return NULL;
2560
2561	p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2562
2563	WARN_ON_ONCE(rq->cpu != task_cpu(p));
2564	WARN_ON_ONCE(task_current(rq, p));
2565	WARN_ON_ONCE(p->nr_cpus_allowed <= `1`);
2566
2567	WARN_ON_ONCE(!task_on_rq_queued(p));
2568	WARN_ON_ONCE(!dl_task(p));
2569
2570	return p;
2571	}
2572
2573	/ Locks the rq it finds /
2574	static struct rq find_lock_later_rq(struct* task_struct task, struct* rq *rq)
2575	{
2576	struct rq *later_rq = NULL;
2577	int tries;
2578	int cpu;
2579
2580	for (tries = `0`; tries < DL_MAX_TRIES; tries++) {
2581	cpu = find_later_rq(task);
2582
2583	if ((cpu == -`1`) \|\| (cpu == rq->cpu))
2584	break;
2585
2586	later_rq = cpu_rq(cpu);
2587
2588	if (!dl_task_is_earliest_deadline(p: task, rq: later_rq)) {
2589	/*
2590	* Target rq has tasks of equal or earlier deadline,
2591	* retrying does not release any lock and is unlikely
2592	* to yield a different result.
2593	*/
2594	later_rq = NULL;
2595	break;
2596	}
2597
2598	/ Retry if something changed. /
2599	if (double_lock_balance(this_rq: rq, busiest: later_rq)) {
2600	/*
2601	* double_lock_balance had to release rq->lock, in the
2602	* meantime, task may no longer be fit to be migrated.
2603	* Check the following to ensure that the task is
2604	* still suitable for migration:
2605	* 1. It is possible the task was scheduled,
2606	* migrate_disabled was set and then got preempted,
2607	* so we must check the task migration disable
2608	* flag.
2609	* 2. The CPU picked is in the task's affinity.
2610	* 3. For throttled task (dl_task_offline_migration),
2611	* check the following:
2612	* - the task is not on the rq anymore (it was
2613	* migrated)
2614	* - the task is not on CPU anymore
2615	* - the task is still a dl task
2616	* - the task is not queued on the rq anymore
2617	* 4. For the non-throttled task (push_dl_task), the
2618	* check to ensure that this task is still at the
2619	* head of the pushable tasks list is enough.
2620	*/
2621	if (unlikely(is_migration_disabled(task) \|\|
2622	!cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) \|\|
2623	(task->dl.dl_throttled &&
2624	(task_rq(task) != rq \|\|
2625	task_on_cpu(rq, task) \|\|
2626	!dl_task(task) \|\|
2627	!task_on_rq_queued(task))) \|\|
2628	(!task->dl.dl_throttled &&
2629	task != pick_next_pushable_dl_task(rq)))) {
2630
2631	double_unlock_balance(this_rq: rq, busiest: later_rq);
2632	later_rq = NULL;
2633	break;
2634	}
2635	}
2636
2637	/*
2638	* If the rq we found has no -deadline task, or
2639	* its earliest one has a later deadline than our
2640	* task, the rq is a good one.
2641	*/
2642	if (dl_task_is_earliest_deadline(p: task, rq: later_rq))
2643	break;
2644
2645	/ Otherwise we try again. /
2646	double_unlock_balance(this_rq: rq, busiest: later_rq);
2647	later_rq = NULL;
2648	}
2649
2650	return later_rq;
2651	}
2652
2653	/*
2654	* See if the non running -deadline tasks on this rq
2655	* can be sent to some other CPU where they can preempt
2656	* and start executing.
2657	*/
2658	static int push_dl_task(struct rq *rq)
2659	{
2660	struct task_struct *next_task;
2661	struct rq *later_rq;
2662	int ret = `0`;
2663
2664	next_task = pick_next_pushable_dl_task(rq);
2665	if (!next_task)
2666	return `0`;
2667
2668	retry:
2669	/*
2670	* If next_task preempts rq->curr, and rq->curr
2671	* can move away, it makes sense to just reschedule
2672	* without going further in pushing next_task.
2673	*/
2674	if (dl_task(p: rq->donor) &&
2675	dl_time_before(a: next_task->dl.deadline, b: rq->donor->dl.deadline) &&
2676	rq->curr->nr_cpus_allowed > `1`) {
2677	resched_curr(rq);
2678	return `0`;
2679	}
2680
2681	if (is_migration_disabled(p: next_task))
2682	return `0`;
2683
2684	if (WARN_ON(next_task == rq->curr))
2685	return `0`;
2686
2687	/ We might release rq lock /
2688	get_task_struct(t: next_task);
2689
2690	/ Will lock the rq it'll find /
2691	later_rq = find_lock_later_rq(task: next_task, rq);
2692	if (!later_rq) {
2693	struct task_struct *task;
2694
2695	/*
2696	* We must check all this again, since
2697	* find_lock_later_rq releases rq->lock and it is
2698	* then possible that next_task has migrated.
2699	*/
2700	task = pick_next_pushable_dl_task(rq);
2701	if (task == next_task) {
2702	/*
2703	* The task is still there. We don't try
2704	* again, some other CPU will pull it when ready.
2705	*/
2706	goto out;
2707	}
2708
2709	if (!task)
2710	/ No more tasks /
2711	goto out;
2712
2713	put_task_struct(t: next_task);
2714	next_task = task;
2715	goto retry;
2716	}
2717
2718	move_queued_task_locked(src_rq: rq, dst_rq: later_rq, task: next_task);
2719	ret = `1`;
2720
2721	resched_curr(rq: later_rq);
2722
2723	double_unlock_balance(this_rq: rq, busiest: later_rq);
2724
2725	out:
2726	put_task_struct(t: next_task);
2727
2728	return ret;
2729	}
2730
2731	static void push_dl_tasks(struct rq *rq)
2732	{
2733	/ push_dl_task() will return true if it moved a -deadline task /
2734	while (push_dl_task(rq))
2735	;
2736	}
2737
2738	static void pull_dl_task(struct rq *this_rq)
2739	{
2740	int this_cpu = this_rq->cpu, cpu;
2741	struct task_struct p, push_task;
2742	bool resched = false;
2743	struct rq *src_rq;
2744	u64 dmin = LONG_MAX;
2745
2746	if (likely(!dl_overloaded(this_rq)))
2747	return;
2748
2749	/*
2750	* Match the barrier from dl_set_overloaded; this guarantees that if we
2751	* see overloaded we must also see the dlo_mask bit.
2752	*/
2753	smp_rmb();
2754
2755	for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2756	if (this_cpu == cpu)
2757	continue;
2758
2759	src_rq = cpu_rq(cpu);
2760
2761	/*
2762	* It looks racy, and it is! However, as in sched_rt.c,
2763	* we are fine with this.
2764	*/
2765	if (this_rq->dl.dl_nr_running &&
2766	dl_time_before(a: this_rq->dl.earliest_dl.curr,
2767	b: src_rq->dl.earliest_dl.next))
2768	continue;
2769
2770	/ Might drop this_rq->lock /
2771	push_task = NULL;
2772	double_lock_balance(this_rq, busiest: src_rq);
2773
2774	/*
2775	* If there are no more pullable tasks on the
2776	* rq, we're done with it.
2777	*/
2778	if (src_rq->dl.dl_nr_running <= `1`)
2779	goto skip;
2780
2781	p = pick_earliest_pushable_dl_task(rq: src_rq, cpu: this_cpu);
2782
2783	/*
2784	* We found a task to be pulled if:
2785	* - it preempts our current (if there's one),
2786	* - it will preempt the last one we pulled (if any).
2787	*/
2788	if (p && dl_time_before(a: p->dl.deadline, b: dmin) &&
2789	dl_task_is_earliest_deadline(p, rq: this_rq)) {
2790	WARN_ON(p == src_rq->curr);
2791	WARN_ON(!task_on_rq_queued(p));
2792
2793	/*
2794	* Then we pull iff p has actually an earlier
2795	* deadline than the current task of its runqueue.
2796	*/
2797	if (dl_time_before(a: p->dl.deadline,
2798	b: src_rq->donor->dl.deadline))
2799	goto skip;
2800
2801	if (is_migration_disabled(p)) {
2802	push_task = get_push_task(rq: src_rq);
2803	} else {
2804	move_queued_task_locked(src_rq, dst_rq: this_rq, task: p);
2805	dmin = p->dl.deadline;
2806	resched = true;
2807	}
2808
2809	/ Is there any other task even earlier? /
2810	}
2811	skip:
2812	double_unlock_balance(this_rq, busiest: src_rq);
2813
2814	if (push_task) {
2815	preempt_disable();
2816	raw_spin_rq_unlock(rq: this_rq);
2817	stop_one_cpu_nowait(cpu: src_rq->cpu, fn: push_cpu_stop,
2818	arg: push_task, work_buf: &src_rq->push_work);
2819	preempt_enable();
2820	raw_spin_rq_lock(rq: this_rq);
2821	}
2822	}
2823
2824	if (resched)
2825	resched_curr(rq: this_rq);
2826	}
2827
2828	/*
2829	* Since the task is not running and a reschedule is not going to happen
2830	* anytime soon on its runqueue, we try pushing it away now.
2831	*/
2832	static void task_woken_dl(struct rq rq, struct* task_struct *p)
2833	{
2834	if (!task_on_cpu(rq, p) &&
2835	!test_tsk_need_resched(tsk: rq->curr) &&
2836	p->nr_cpus_allowed > `1` &&
2837	dl_task(p: rq->donor) &&
2838	(rq->curr->nr_cpus_allowed < `2` \|\|
2839	!dl_entity_preempt(a: &p->dl, b: &rq->donor->dl))) {
2840	push_dl_tasks(rq);
2841	}
2842	}
2843
2844	static void set_cpus_allowed_dl(struct task_struct *p,
2845	struct affinity_context *ctx)
2846	{
2847	struct root_domain *src_rd;
2848	struct rq *rq;
2849
2850	WARN_ON_ONCE(!dl_task(p));
2851
2852	rq = task_rq(p);
2853	src_rd = rq->rd;
2854	/*
2855	* Migrating a SCHED_DEADLINE task between exclusive
2856	* cpusets (different root_domains) entails a bandwidth
2857	* update. We already made space for us in the destination
2858	* domain (see cpuset_can_attach()).
2859	*/
2860	if (!cpumask_intersects(src1p: src_rd->span, src2p: ctx->new_mask)) {
2861	struct dl_bw *src_dl_b;
2862
2863	src_dl_b = dl_bw_of(i: cpu_of(rq));
2864	/*
2865	* We now free resources of the root_domain we are migrating
2866	* off. In the worst case, sched_setattr() may temporary fail
2867	* until we complete the update.
2868	*/
2869	raw_spin_lock(&src_dl_b->lock);
2870	__dl_sub(dl_b: src_dl_b, tsk_bw: p->dl.dl_bw, cpus: dl_bw_cpus(i: task_cpu(p)));
2871	raw_spin_unlock(&src_dl_b->lock);
2872	}
2873
2874	set_cpus_allowed_common(p, ctx);
2875	}
2876
2877	/ Assumes rq->lock is held /
2878	static void rq_online_dl(struct rq *rq)
2879	{
2880	if (rq->dl.overloaded)
2881	dl_set_overload(rq);
2882
2883	cpudl_set_freecpu(cp: &rq->rd->cpudl, cpu: rq->cpu);
2884	if (rq->dl.dl_nr_running > `0`)
2885	cpudl_set(cp: &rq->rd->cpudl, cpu: rq->cpu, dl: rq->dl.earliest_dl.curr);
2886	}
2887
2888	/ Assumes rq->lock is held /
2889	static void rq_offline_dl(struct rq *rq)
2890	{
2891	if (rq->dl.overloaded)
2892	dl_clear_overload(rq);
2893
2894	cpudl_clear(cp: &rq->rd->cpudl, cpu: rq->cpu);
2895	cpudl_clear_freecpu(cp: &rq->rd->cpudl, cpu: rq->cpu);
2896	}
2897
2898	void __init init_sched_dl_class(void)
2899	{
2900	unsigned int i;
2901
2902	for_each_possible_cpu(i)
2903	zalloc_cpumask_var_node(mask: &per_cpu(local_cpu_mask_dl, i),
2904	GFP_KERNEL, node: cpu_to_node(cpu: i));
2905	}
2906
2907	void dl_add_task_root_domain(struct task_struct *p)
2908	{
2909	struct rq_flags rf;
2910	struct rq *rq;
2911	struct dl_bw *dl_b;
2912
2913	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2914	if (!dl_task(p) \|\| dl_entity_is_special(dl_se: &p->dl)) {
2915	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2916	return;
2917	}
2918
2919	rq = __task_rq_lock(p, rf: &rf);
2920
2921	dl_b = &rq->rd->dl_bw;
2922	raw_spin_lock(&dl_b->lock);
2923
2924	__dl_add(dl_b, tsk_bw: p->dl.dl_bw, cpus: cpumask_weight(srcp: rq->rd->span));
2925
2926	raw_spin_unlock(&dl_b->lock);
2927
2928	task_rq_unlock(rq, p, rf: &rf);
2929	}
2930
2931	void dl_clear_root_domain(struct root_domain *rd)
2932	{
2933	int i;
2934
2935	guard(raw_spinlock_irqsave)(l: &rd->dl_bw.lock);
2936
2937	/*
2938	* Reset total_bw to zero and extra_bw to max_bw so that next
2939	* loop will add dl-servers contributions back properly,
2940	*/
2941	rd->dl_bw.total_bw = `0`;
2942	for_each_cpu(i, rd->span)
2943	cpu_rq(i)->dl.extra_bw = cpu_rq(i)->dl.max_bw;
2944
2945	/*
2946	* dl_servers are not tasks. Since dl_add_task_root_domain ignores
2947	* them, we need to account for them here explicitly.
2948	*/
2949	for_each_cpu(i, rd->span) {
2950	struct sched_dl_entity *dl_se = &cpu_rq(i)->fair_server;
2951
2952	if (dl_server(dl_se) && cpu_active(cpu: i))
2953	__dl_add(dl_b: &rd->dl_bw, tsk_bw: dl_se->dl_bw, cpus: dl_bw_cpus(i));
2954	}
2955	}
2956
2957	void dl_clear_root_domain_cpu(int cpu)
2958	{
2959	dl_clear_root_domain(cpu_rq(cpu)->rd);
2960	}
2961
2962	static void switched_from_dl(struct rq rq, struct* task_struct *p)
2963	{
2964	/*
2965	* task_non_contending() can start the "inactive timer" (if the 0-lag
2966	* time is in the future). If the task switches back to dl before
2967	* the "inactive timer" fires, it can continue to consume its current
2968	* runtime using its current deadline. If it stays outside of
2969	* SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2970	* will reset the task parameters.
2971	*/
2972	if (task_on_rq_queued(p) && p->dl.dl_runtime)
2973	task_non_contending(dl_se: &p->dl);
2974
2975	/*
2976	* In case a task is setscheduled out from SCHED_DEADLINE we need to
2977	* keep track of that on its cpuset (for correct bandwidth tracking).
2978	*/
2979	dec_dl_tasks_cs(task: p);
2980
2981	if (!task_on_rq_queued(p)) {
2982	/*
2983	* Inactive timer is armed. However, p is leaving DEADLINE and
2984	* might migrate away from this rq while continuing to run on
2985	* some other class. We need to remove its contribution from
2986	* this rq running_bw now, or sub_rq_bw (below) will complain.
2987	*/
2988	if (p->dl.dl_non_contending)
2989	sub_running_bw(dl_se: &p->dl, dl_rq: &rq->dl);
2990	sub_rq_bw(dl_se: &p->dl, dl_rq: &rq->dl);
2991	}
2992
2993	/*
2994	* We cannot use inactive_task_timer() to invoke sub_running_bw()
2995	* at the 0-lag time, because the task could have been migrated
2996	* while SCHED_OTHER in the meanwhile.
2997	*/
2998	if (p->dl.dl_non_contending)
2999	p->dl.dl_non_contending = `0`;
3000
3001	/*
3002	* Since this might be the only -deadline task on the rq,
3003	* this is the right place to try to pull some other one
3004	* from an overloaded CPU, if any.
3005	*/
3006	if (!task_on_rq_queued(p) \|\| rq->dl.dl_nr_running)
3007	return;
3008
3009	deadline_queue_pull_task(rq);
3010	}
3011
3012	/*
3013	* When switching to -deadline, we may overload the rq, then
3014	* we try to push someone off, if possible.
3015	*/
3016	static void switched_to_dl(struct rq rq, struct* task_struct *p)
3017	{
3018	cancel_inactive_timer(dl_se: &p->dl);
3019
3020	/*
3021	* In case a task is setscheduled to SCHED_DEADLINE we need to keep
3022	* track of that on its cpuset (for correct bandwidth tracking).
3023	*/
3024	inc_dl_tasks_cs(task: p);
3025
3026	/ If p is not queued we will update its parameters at next wakeup. /
3027	if (!task_on_rq_queued(p)) {
3028	add_rq_bw(dl_se: &p->dl, dl_rq: &rq->dl);
3029
3030	return;
3031	}
3032
3033	if (rq->donor != p) {
3034	if (p->nr_cpus_allowed > `1` && rq->dl.overloaded)
3035	deadline_queue_push_tasks(rq);
3036	if (dl_task(p: rq->donor))
3037	wakeup_preempt_dl(rq, p, flags: `0`);
3038	else
3039	resched_curr(rq);
3040	} else {
3041	update_dl_rq_load_avg(now: rq_clock_pelt(rq), rq, running: `0`);
3042	}
3043	}
3044
3045	/*
3046	* If the scheduling parameters of a -deadline task changed,
3047	* a push or pull operation might be needed.
3048	*/
3049	static void prio_changed_dl(struct rq rq, struct* task_struct *p,
3050	int oldprio)
3051	{
3052	if (!task_on_rq_queued(p))
3053	return;
3054
3055	/*
3056	* This might be too much, but unfortunately
3057	* we don't have the old deadline value, and
3058	* we can't argue if the task is increasing
3059	* or lowering its prio, so...
3060	*/
3061	if (!rq->dl.overloaded)
3062	deadline_queue_pull_task(rq);
3063
3064	if (task_current_donor(rq, p)) {
3065	/*
3066	* If we now have a earlier deadline task than p,
3067	* then reschedule, provided p is still on this
3068	* runqueue.
3069	*/
3070	if (dl_time_before(a: rq->dl.earliest_dl.curr, b: p->dl.deadline))
3071	resched_curr(rq);
3072	} else {
3073	/*
3074	* Current may not be deadline in case p was throttled but we
3075	* have just replenished it (e.g. rt_mutex_setprio()).
3076	*
3077	* Otherwise, if p was given an earlier deadline, reschedule.
3078	*/
3079	if (!dl_task(p: rq->curr) \|\|
3080	dl_time_before(a: p->dl.deadline, b: rq->curr->dl.deadline))
3081	resched_curr(rq);
3082	}
3083	}
3084
3085	#ifdef CONFIG_SCHED_CORE
3086	static int task_is_throttled_dl(struct task_struct p, int* cpu)
3087	{
3088	return p->dl.dl_throttled;
3089	}
3090	#endif
3091
3092	DEFINE_SCHED_CLASS(dl) = {
3093
3094	.enqueue_task = enqueue_task_dl,
3095	.dequeue_task = dequeue_task_dl,
3096	.yield_task = yield_task_dl,
3097
3098	.wakeup_preempt = wakeup_preempt_dl,
3099
3100	.pick_task = pick_task_dl,
3101	.put_prev_task = put_prev_task_dl,
3102	.set_next_task = set_next_task_dl,
3103
3104	.balance = balance_dl,
3105	.select_task_rq = select_task_rq_dl,
3106	.migrate_task_rq = migrate_task_rq_dl,
3107	.set_cpus_allowed = set_cpus_allowed_dl,
3108	.rq_online = rq_online_dl,
3109	.rq_offline = rq_offline_dl,
3110	.task_woken = task_woken_dl,
3111	.find_lock_rq = find_lock_later_rq,
3112
3113	.task_tick = task_tick_dl,
3114	.task_fork = task_fork_dl,
3115
3116	.prio_changed = prio_changed_dl,
3117	.switched_from = switched_from_dl,
3118	.switched_to = switched_to_dl,
3119
3120	.update_curr = update_curr_dl,
3121	#ifdef CONFIG_SCHED_CORE
3122	.task_is_throttled = task_is_throttled_dl,
3123	#endif
3124	};
3125
3126	/*
3127	* Used for dl_bw check and update, used under sched_rt_handler()::mutex and
3128	* sched_domains_mutex.
3129	*/
3130	u64 dl_cookie;
3131
3132	int sched_dl_global_validate(void)
3133	{
3134	u64 runtime = global_rt_runtime();
3135	u64 period = global_rt_period();
3136	u64 new_bw = to_ratio(period, runtime);
3137	u64 cookie = ++dl_cookie;
3138	struct dl_bw *dl_b;
3139	int cpu, cpus, ret = `0`;
3140	unsigned long flags;
3141
3142	/*
3143	* Here we want to check the bandwidth not being set to some
3144	* value smaller than the currently allocated bandwidth in
3145	* any of the root_domains.
3146	*/
3147	for_each_online_cpu(cpu) {
3148	rcu_read_lock_sched();
3149
3150	if (dl_bw_visited(cpu, cookie))
3151	goto next;
3152
3153	dl_b = dl_bw_of(i: cpu);
3154	cpus = dl_bw_cpus(i: cpu);
3155
3156	raw_spin_lock_irqsave(&dl_b->lock, flags);
3157	if (new_bw * cpus < dl_b->total_bw)
3158	ret = -EBUSY;
3159	raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3160
3161	next:
3162	rcu_read_unlock_sched();
3163
3164	if (ret)
3165	break;
3166	}
3167
3168	return ret;
3169	}
3170
3171	static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
3172	{
3173	if (global_rt_runtime() == RUNTIME_INF) {
3174	dl_rq->bw_ratio = `1` << RATIO_SHIFT;
3175	dl_rq->max_bw = dl_rq->extra_bw = `1` << BW_SHIFT;
3176	} else {
3177	dl_rq->bw_ratio = to_ratio(period: global_rt_runtime(),
3178	runtime: global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
3179	dl_rq->max_bw = dl_rq->extra_bw =
3180	to_ratio(period: global_rt_period(), runtime: global_rt_runtime());
3181	}
3182	}
3183
3184	void sched_dl_do_global(void)
3185	{
3186	u64 new_bw = -`1`;
3187	u64 cookie = ++dl_cookie;
3188	struct dl_bw *dl_b;
3189	int cpu;
3190	unsigned long flags;
3191
3192	if (global_rt_runtime() != RUNTIME_INF)
3193	new_bw = to_ratio(period: global_rt_period(), runtime: global_rt_runtime());
3194
3195	for_each_possible_cpu(cpu)
3196	init_dl_rq_bw_ratio(dl_rq: &cpu_rq(cpu)->dl);
3197
3198	for_each_possible_cpu(cpu) {
3199	rcu_read_lock_sched();
3200
3201	if (dl_bw_visited(cpu, cookie)) {
3202	rcu_read_unlock_sched();
3203	continue;
3204	}
3205
3206	dl_b = dl_bw_of(i: cpu);
3207
3208	raw_spin_lock_irqsave(&dl_b->lock, flags);
3209	dl_b->bw = new_bw;
3210	raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3211
3212	rcu_read_unlock_sched();
3213	}
3214	}
3215
3216	/*
3217	* We must be sure that accepting a new task (or allowing changing the
3218	* parameters of an existing one) is consistent with the bandwidth
3219	* constraints. If yes, this function also accordingly updates the currently
3220	* allocated bandwidth to reflect the new situation.
3221	*
3222	* This function is called while holding p's rq->lock.
3223	*/
3224	int sched_dl_overflow(struct task_struct p, int* policy,
3225	const struct sched_attr *attr)
3226	{
3227	u64 period = attr->sched_period ?: attr->sched_deadline;
3228	u64 runtime = attr->sched_runtime;
3229	u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : `0`;
3230	int cpus, err = -`1`, cpu = task_cpu(p);
3231	struct dl_bw *dl_b = dl_bw_of(i: cpu);
3232	unsigned long cap;
3233
3234	if (attr->sched_flags & SCHED_FLAG_SUGOV)
3235	return `0`;
3236
3237	/ !deadline task may carry old deadline bandwidth /
3238	if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
3239	return `0`;
3240
3241	/*
3242	* Either if a task, enters, leave, or stays -deadline but changes
3243	* its parameters, we may need to update accordingly the total
3244	* allocated bandwidth of the container.
3245	*/
3246	raw_spin_lock(&dl_b->lock);
3247	cpus = dl_bw_cpus(i: cpu);
3248	cap = dl_bw_capacity(i: cpu);
3249
3250	if (dl_policy(policy) && !task_has_dl_policy(p) &&
3251	!__dl_overflow(dl_b, cap, old_bw: `0`, new_bw)) {
3252	if (hrtimer_active(timer: &p->dl.inactive_timer))
3253	__dl_sub(dl_b, tsk_bw: p->dl.dl_bw, cpus);
3254	__dl_add(dl_b, tsk_bw: new_bw, cpus);
3255	err = `0`;
3256	} else if (dl_policy(policy) && task_has_dl_policy(p) &&
3257	!__dl_overflow(dl_b, cap, old_bw: p->dl.dl_bw, new_bw)) {
3258	/*
3259	* XXX this is slightly incorrect: when the task
3260	* utilization decreases, we should delay the total
3261	* utilization change until the task's 0-lag point.
3262	* But this would require to set the task's "inactive
3263	* timer" when the task is not inactive.
3264	*/
3265	__dl_sub(dl_b, tsk_bw: p->dl.dl_bw, cpus);
3266	__dl_add(dl_b, tsk_bw: new_bw, cpus);
3267	dl_change_utilization(p, new_bw);
3268	err = `0`;
3269	} else if (!dl_policy(policy) && task_has_dl_policy(p)) {
3270	/*
3271	* Do not decrease the total deadline utilization here,
3272	* switched_from_dl() will take care to do it at the correct
3273	* (0-lag) time.
3274	*/
3275	err = `0`;
3276	}
3277	raw_spin_unlock(&dl_b->lock);
3278
3279	return err;
3280	}
3281
3282	/*
3283	* This function initializes the sched_dl_entity of a newly becoming
3284	* SCHED_DEADLINE task.
3285	*
3286	* Only the static values are considered here, the actual runtime and the
3287	* absolute deadline will be properly calculated when the task is enqueued
3288	* for the first time with its new policy.
3289	*/
3290	void __setparam_dl(struct task_struct p, const* struct sched_attr *attr)
3291	{
3292	struct sched_dl_entity *dl_se = &p->dl;
3293
3294	dl_se->dl_runtime = attr->sched_runtime;
3295	dl_se->dl_deadline = attr->sched_deadline;
3296	dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3297	dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3298	dl_se->dl_bw = to_ratio(period: dl_se->dl_period, runtime: dl_se->dl_runtime);
3299	dl_se->dl_density = to_ratio(period: dl_se->dl_deadline, runtime: dl_se->dl_runtime);
3300	}
3301
3302	void __getparam_dl(struct task_struct p, struct* sched_attr *attr)
3303	{
3304	struct sched_dl_entity *dl_se = &p->dl;
3305
3306	attr->sched_priority = p->rt_priority;
3307	attr->sched_runtime = dl_se->dl_runtime;
3308	attr->sched_deadline = dl_se->dl_deadline;
3309	attr->sched_period = dl_se->dl_period;
3310	attr->sched_flags &= ~SCHED_DL_FLAGS;
3311	attr->sched_flags \|= dl_se->flags;
3312	}
3313
3314	/*
3315	* This function validates the new parameters of a -deadline task.
3316	* We ask for the deadline not being zero, and greater or equal
3317	* than the runtime, as well as the period of being zero or
3318	* greater than deadline. Furthermore, we have to be sure that
3319	* user parameters are above the internal resolution of 1us (we
3320	* check sched_runtime only since it is always the smaller one) and
3321	* below 2^63 ns (we have to check both sched_deadline and
3322	* sched_period, as the latter can be zero).
3323	*/
3324	bool __checkparam_dl(const struct sched_attr *attr)
3325	{
3326	u64 period, max, min;
3327
3328	/ special dl tasks don't actually use any parameter /
3329	if (attr->sched_flags & SCHED_FLAG_SUGOV)
3330	return true;
3331
3332	/ deadline != 0 /
3333	if (attr->sched_deadline == `0`)
3334	return false;
3335
3336	/*
3337	* Since we truncate DL_SCALE bits, make sure we're at least
3338	* that big.
3339	*/
3340	if (attr->sched_runtime < (`1ULL` << DL_SCALE))
3341	return false;
3342
3343	/*
3344	* Since we use the MSB for wrap-around and sign issues, make
3345	* sure it's not set (mind that period can be equal to zero).
3346	*/
3347	if (attr->sched_deadline & (`1ULL` << `63`) \|\|
3348	attr->sched_period & (`1ULL` << `63`))
3349	return false;
3350
3351	period = attr->sched_period;
3352	if (!period)
3353	period = attr->sched_deadline;
3354
3355	/ runtime <= deadline <= period (if period != 0) /
3356	if (period < attr->sched_deadline \|\|
3357	attr->sched_deadline < attr->sched_runtime)
3358	return false;
3359
3360	max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3361	min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3362
3363	if (period < min \|\| period > max)
3364	return false;
3365
3366	return true;
3367	}
3368
3369	/*
3370	* This function clears the sched_dl_entity static params.
3371	*/
3372	static void __dl_clear_params(struct sched_dl_entity *dl_se)
3373	{
3374	dl_se->dl_runtime = `0`;
3375	dl_se->dl_deadline = `0`;
3376	dl_se->dl_period = `0`;
3377	dl_se->flags = `0`;
3378	dl_se->dl_bw = `0`;
3379	dl_se->dl_density = `0`;
3380
3381	dl_se->dl_throttled = `0`;
3382	dl_se->dl_yielded = `0`;
3383	dl_se->dl_non_contending = `0`;
3384	dl_se->dl_overrun = `0`;
3385	dl_se->dl_server = `0`;
3386
3387	#ifdef CONFIG_RT_MUTEXES
3388	dl_se->pi_se = dl_se;
3389	#endif
3390	}
3391
3392	void init_dl_entity(struct sched_dl_entity *dl_se)
3393	{
3394	RB_CLEAR_NODE(&dl_se->rb_node);
3395	init_dl_task_timer(dl_se);
3396	init_dl_inactive_task_timer(dl_se);
3397	__dl_clear_params(dl_se);
3398	}
3399
3400	bool dl_param_changed(struct task_struct p, const* struct sched_attr *attr)
3401	{
3402	struct sched_dl_entity *dl_se = &p->dl;
3403
3404	if (dl_se->dl_runtime != attr->sched_runtime \|\|
3405	dl_se->dl_deadline != attr->sched_deadline \|\|
3406	dl_se->dl_period != attr->sched_period \|\|
3407	dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3408	return true;
3409
3410	return false;
3411	}
3412
3413	int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3414	const struct cpumask *trial)
3415	{
3416	unsigned long flags, cap;
3417	struct dl_bw *cur_dl_b;
3418	int ret = `1`;
3419
3420	rcu_read_lock_sched();
3421	cur_dl_b = dl_bw_of(cpumask_any(cur));
3422	cap = __dl_bw_capacity(mask: trial);
3423	raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3424	if (__dl_overflow(dl_b: cur_dl_b, cap, old_bw: `0`, new_bw: `0`))
3425	ret = `0`;
3426	raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3427	rcu_read_unlock_sched();
3428
3429	return ret;
3430	}
3431
3432	enum dl_bw_request {
3433	dl_bw_req_deactivate = `0`,
3434	dl_bw_req_alloc,
3435	dl_bw_req_free
3436	};
3437
3438	static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3439	{
3440	unsigned long flags, cap;
3441	struct dl_bw *dl_b;
3442	bool overflow = `0`;
3443	u64 fair_server_bw = `0`;
3444
3445	rcu_read_lock_sched();
3446	dl_b = dl_bw_of(i: cpu);
3447	raw_spin_lock_irqsave(&dl_b->lock, flags);
3448
3449	cap = dl_bw_capacity(i: cpu);
3450	switch (req) {
3451	case dl_bw_req_free:
3452	__dl_sub(dl_b, tsk_bw: dl_bw, cpus: dl_bw_cpus(i: cpu));
3453	break;
3454	case dl_bw_req_alloc:
3455	overflow = __dl_overflow(dl_b, cap, old_bw: `0`, new_bw: dl_bw);
3456
3457	if (!overflow) {
3458	/*
3459	* We reserve space in the destination
3460	* root_domain, as we can't fail after this point.
3461	* We will free resources in the source root_domain
3462	* later on (see set_cpus_allowed_dl()).
3463	*/
3464	__dl_add(dl_b, tsk_bw: dl_bw, cpus: dl_bw_cpus(i: cpu));
3465	}
3466	break;
3467	case dl_bw_req_deactivate:
3468	/*
3469	* cpu is not off yet, but we need to do the math by
3470	* considering it off already (i.e., what would happen if we
3471	* turn cpu off?).
3472	*/
3473	cap -= arch_scale_cpu_capacity(cpu);
3474
3475	/*
3476	* cpu is going offline and NORMAL tasks will be moved away
3477	* from it. We can thus discount dl_server bandwidth
3478	* contribution as it won't need to be servicing tasks after
3479	* the cpu is off.
3480	*/
3481	if (cpu_rq(cpu)->fair_server.dl_server)
3482	fair_server_bw = cpu_rq(cpu)->fair_server.dl_bw;
3483
3484	/*
3485	* Not much to check if no DEADLINE bandwidth is present.
3486	* dl_servers we can discount, as tasks will be moved out the
3487	* offlined CPUs anyway.
3488	*/
3489	if (dl_b->total_bw - fair_server_bw > `0`) {
3490	/*
3491	* Leaving at least one CPU for DEADLINE tasks seems a
3492	* wise thing to do. As said above, cpu is not offline
3493	* yet, so account for that.
3494	*/
3495	if (dl_bw_cpus(i: cpu) - `1`)
3496	overflow = __dl_overflow(dl_b, cap, old_bw: fair_server_bw, new_bw: `0`);
3497	else
3498	overflow = `1`;
3499	}
3500
3501	break;
3502	}
3503
3504	raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3505	rcu_read_unlock_sched();
3506
3507	return overflow ? -EBUSY : `0`;
3508	}
3509
3510	int dl_bw_deactivate(int cpu)
3511	{
3512	return dl_bw_manage(req: dl_bw_req_deactivate, cpu, dl_bw: `0`);
3513	}
3514
3515	int dl_bw_alloc(int cpu, u64 dl_bw)
3516	{
3517	return dl_bw_manage(req: dl_bw_req_alloc, cpu, dl_bw);
3518	}
3519
3520	void dl_bw_free(int cpu, u64 dl_bw)
3521	{
3522	dl_bw_manage(req: dl_bw_req_free, cpu, dl_bw);
3523	}
3524
3525	void print_dl_stats(struct seq_file m, int* cpu)
3526	{
3527	print_dl_rq(m, cpu, dl_rq: &cpu_rq(cpu)->dl);
3528	}
3529

Browse the source code of Linux/kernel/sched/deadline.c