xref: /linux/kernel/sched/sched.h (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 
2 #include <linux/sched.h>
3 #include <linux/sched/sysctl.h>
4 #include <linux/sched/rt.h>
5 #include <linux/sched/deadline.h>
6 #include <linux/mutex.h>
7 #include <linux/spinlock.h>
8 #include <linux/stop_machine.h>
9 #include <linux/tick.h>
10 #include <linux/slab.h>
11 
12 #include "cpupri.h"
13 #include "cpudeadline.h"
14 #include "cpuacct.h"
15 
16 struct rq;
17 
18 extern __read_mostly int scheduler_running;
19 
20 extern unsigned long calc_load_update;
21 extern atomic_long_t calc_load_tasks;
22 
23 extern long calc_load_fold_active(struct rq *this_rq);
24 extern void update_cpu_load_active(struct rq *this_rq);
25 
26 /*
27  * Helpers for converting nanosecond timing to jiffy resolution
28  */
29 #define NS_TO_JIFFIES(TIME)	((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
30 
31 /*
32  * Increase resolution of nice-level calculations for 64-bit architectures.
33  * The extra resolution improves shares distribution and load balancing of
34  * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
35  * hierarchies, especially on larger systems. This is not a user-visible change
36  * and does not change the user-interface for setting shares/weights.
37  *
38  * We increase resolution only if we have enough bits to allow this increased
39  * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
40  * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
41  * increased costs.
42  */
43 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load  */
44 # define SCHED_LOAD_RESOLUTION	10
45 # define scale_load(w)		((w) << SCHED_LOAD_RESOLUTION)
46 # define scale_load_down(w)	((w) >> SCHED_LOAD_RESOLUTION)
47 #else
48 # define SCHED_LOAD_RESOLUTION	0
49 # define scale_load(w)		(w)
50 # define scale_load_down(w)	(w)
51 #endif
52 
53 #define SCHED_LOAD_SHIFT	(10 + SCHED_LOAD_RESOLUTION)
54 #define SCHED_LOAD_SCALE	(1L << SCHED_LOAD_SHIFT)
55 
56 #define NICE_0_LOAD		SCHED_LOAD_SCALE
57 #define NICE_0_SHIFT		SCHED_LOAD_SHIFT
58 
59 /*
60  * Single value that decides SCHED_DEADLINE internal math precision.
61  * 10 -> just above 1us
62  * 9  -> just above 0.5us
63  */
64 #define DL_SCALE (10)
65 
66 /*
67  * These are the 'tuning knobs' of the scheduler:
68  */
69 
70 /*
71  * single value that denotes runtime == period, ie unlimited time.
72  */
73 #define RUNTIME_INF	((u64)~0ULL)
74 
75 static inline int fair_policy(int policy)
76 {
77 	return policy == SCHED_NORMAL || policy == SCHED_BATCH;
78 }
79 
80 static inline int rt_policy(int policy)
81 {
82 	return policy == SCHED_FIFO || policy == SCHED_RR;
83 }
84 
85 static inline int dl_policy(int policy)
86 {
87 	return policy == SCHED_DEADLINE;
88 }
89 
90 static inline int task_has_rt_policy(struct task_struct *p)
91 {
92 	return rt_policy(p->policy);
93 }
94 
95 static inline int task_has_dl_policy(struct task_struct *p)
96 {
97 	return dl_policy(p->policy);
98 }
99 
100 static inline bool dl_time_before(u64 a, u64 b)
101 {
102 	return (s64)(a - b) < 0;
103 }
104 
105 /*
106  * Tells if entity @a should preempt entity @b.
107  */
108 static inline bool
109 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
110 {
111 	return dl_time_before(a->deadline, b->deadline);
112 }
113 
114 /*
115  * This is the priority-queue data structure of the RT scheduling class:
116  */
117 struct rt_prio_array {
118 	DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
119 	struct list_head queue[MAX_RT_PRIO];
120 };
121 
122 struct rt_bandwidth {
123 	/* nests inside the rq lock: */
124 	raw_spinlock_t		rt_runtime_lock;
125 	ktime_t			rt_period;
126 	u64			rt_runtime;
127 	struct hrtimer		rt_period_timer;
128 };
129 /*
130  * To keep the bandwidth of -deadline tasks and groups under control
131  * we need some place where:
132  *  - store the maximum -deadline bandwidth of the system (the group);
133  *  - cache the fraction of that bandwidth that is currently allocated.
134  *
135  * This is all done in the data structure below. It is similar to the
136  * one used for RT-throttling (rt_bandwidth), with the main difference
137  * that, since here we are only interested in admission control, we
138  * do not decrease any runtime while the group "executes", neither we
139  * need a timer to replenish it.
140  *
141  * With respect to SMP, the bandwidth is given on a per-CPU basis,
142  * meaning that:
143  *  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
144  *  - dl_total_bw array contains, in the i-eth element, the currently
145  *    allocated bandwidth on the i-eth CPU.
146  * Moreover, groups consume bandwidth on each CPU, while tasks only
147  * consume bandwidth on the CPU they're running on.
148  * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
149  * that will be shown the next time the proc or cgroup controls will
150  * be red. It on its turn can be changed by writing on its own
151  * control.
152  */
153 struct dl_bandwidth {
154 	raw_spinlock_t dl_runtime_lock;
155 	u64 dl_runtime;
156 	u64 dl_period;
157 };
158 
159 static inline int dl_bandwidth_enabled(void)
160 {
161 	return sysctl_sched_rt_runtime >= 0;
162 }
163 
164 extern struct dl_bw *dl_bw_of(int i);
165 
166 struct dl_bw {
167 	raw_spinlock_t lock;
168 	u64 bw, total_bw;
169 };
170 
171 extern struct mutex sched_domains_mutex;
172 
173 #ifdef CONFIG_CGROUP_SCHED
174 
175 #include <linux/cgroup.h>
176 
177 struct cfs_rq;
178 struct rt_rq;
179 
180 extern struct list_head task_groups;
181 
182 struct cfs_bandwidth {
183 #ifdef CONFIG_CFS_BANDWIDTH
184 	raw_spinlock_t lock;
185 	ktime_t period;
186 	u64 quota, runtime;
187 	s64 hierarchal_quota;
188 	u64 runtime_expires;
189 
190 	int idle, timer_active;
191 	struct hrtimer period_timer, slack_timer;
192 	struct list_head throttled_cfs_rq;
193 
194 	/* statistics */
195 	int nr_periods, nr_throttled;
196 	u64 throttled_time;
197 #endif
198 };
199 
200 /* task group related information */
201 struct task_group {
202 	struct cgroup_subsys_state css;
203 
204 #ifdef CONFIG_FAIR_GROUP_SCHED
205 	/* schedulable entities of this group on each cpu */
206 	struct sched_entity **se;
207 	/* runqueue "owned" by this group on each cpu */
208 	struct cfs_rq **cfs_rq;
209 	unsigned long shares;
210 
211 #ifdef	CONFIG_SMP
212 	atomic_long_t load_avg;
213 	atomic_t runnable_avg;
214 #endif
215 #endif
216 
217 #ifdef CONFIG_RT_GROUP_SCHED
218 	struct sched_rt_entity **rt_se;
219 	struct rt_rq **rt_rq;
220 
221 	struct rt_bandwidth rt_bandwidth;
222 #endif
223 
224 	struct rcu_head rcu;
225 	struct list_head list;
226 
227 	struct task_group *parent;
228 	struct list_head siblings;
229 	struct list_head children;
230 
231 #ifdef CONFIG_SCHED_AUTOGROUP
232 	struct autogroup *autogroup;
233 #endif
234 
235 	struct cfs_bandwidth cfs_bandwidth;
236 };
237 
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 #define ROOT_TASK_GROUP_LOAD	NICE_0_LOAD
240 
241 /*
242  * A weight of 0 or 1 can cause arithmetics problems.
243  * A weight of a cfs_rq is the sum of weights of which entities
244  * are queued on this cfs_rq, so a weight of a entity should not be
245  * too large, so as the shares value of a task group.
246  * (The default weight is 1024 - so there's no practical
247  *  limitation from this.)
248  */
249 #define MIN_SHARES	(1UL <<  1)
250 #define MAX_SHARES	(1UL << 18)
251 #endif
252 
253 typedef int (*tg_visitor)(struct task_group *, void *);
254 
255 extern int walk_tg_tree_from(struct task_group *from,
256 			     tg_visitor down, tg_visitor up, void *data);
257 
258 /*
259  * Iterate the full tree, calling @down when first entering a node and @up when
260  * leaving it for the final time.
261  *
262  * Caller must hold rcu_lock or sufficient equivalent.
263  */
264 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
265 {
266 	return walk_tg_tree_from(&root_task_group, down, up, data);
267 }
268 
269 extern int tg_nop(struct task_group *tg, void *data);
270 
271 extern void free_fair_sched_group(struct task_group *tg);
272 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
273 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
274 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
275 			struct sched_entity *se, int cpu,
276 			struct sched_entity *parent);
277 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
278 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
279 
280 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
281 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
282 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
283 
284 extern void free_rt_sched_group(struct task_group *tg);
285 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
286 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
287 		struct sched_rt_entity *rt_se, int cpu,
288 		struct sched_rt_entity *parent);
289 
290 extern struct task_group *sched_create_group(struct task_group *parent);
291 extern void sched_online_group(struct task_group *tg,
292 			       struct task_group *parent);
293 extern void sched_destroy_group(struct task_group *tg);
294 extern void sched_offline_group(struct task_group *tg);
295 
296 extern void sched_move_task(struct task_struct *tsk);
297 
298 #ifdef CONFIG_FAIR_GROUP_SCHED
299 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
300 #endif
301 
302 #else /* CONFIG_CGROUP_SCHED */
303 
304 struct cfs_bandwidth { };
305 
306 #endif	/* CONFIG_CGROUP_SCHED */
307 
308 /* CFS-related fields in a runqueue */
309 struct cfs_rq {
310 	struct load_weight load;
311 	unsigned int nr_running, h_nr_running;
312 
313 	u64 exec_clock;
314 	u64 min_vruntime;
315 #ifndef CONFIG_64BIT
316 	u64 min_vruntime_copy;
317 #endif
318 
319 	struct rb_root tasks_timeline;
320 	struct rb_node *rb_leftmost;
321 
322 	/*
323 	 * 'curr' points to currently running entity on this cfs_rq.
324 	 * It is set to NULL otherwise (i.e when none are currently running).
325 	 */
326 	struct sched_entity *curr, *next, *last, *skip;
327 
328 #ifdef	CONFIG_SCHED_DEBUG
329 	unsigned int nr_spread_over;
330 #endif
331 
332 #ifdef CONFIG_SMP
333 	/*
334 	 * CFS Load tracking
335 	 * Under CFS, load is tracked on a per-entity basis and aggregated up.
336 	 * This allows for the description of both thread and group usage (in
337 	 * the FAIR_GROUP_SCHED case).
338 	 */
339 	unsigned long runnable_load_avg, blocked_load_avg;
340 	atomic64_t decay_counter;
341 	u64 last_decay;
342 	atomic_long_t removed_load;
343 
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 	/* Required to track per-cpu representation of a task_group */
346 	u32 tg_runnable_contrib;
347 	unsigned long tg_load_contrib;
348 
349 	/*
350 	 *   h_load = weight * f(tg)
351 	 *
352 	 * Where f(tg) is the recursive weight fraction assigned to
353 	 * this group.
354 	 */
355 	unsigned long h_load;
356 	u64 last_h_load_update;
357 	struct sched_entity *h_load_next;
358 #endif /* CONFIG_FAIR_GROUP_SCHED */
359 #endif /* CONFIG_SMP */
360 
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 	struct rq *rq;	/* cpu runqueue to which this cfs_rq is attached */
363 
364 	/*
365 	 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
366 	 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
367 	 * (like users, containers etc.)
368 	 *
369 	 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
370 	 * list is used during load balance.
371 	 */
372 	int on_list;
373 	struct list_head leaf_cfs_rq_list;
374 	struct task_group *tg;	/* group that "owns" this runqueue */
375 
376 #ifdef CONFIG_CFS_BANDWIDTH
377 	int runtime_enabled;
378 	u64 runtime_expires;
379 	s64 runtime_remaining;
380 
381 	u64 throttled_clock, throttled_clock_task;
382 	u64 throttled_clock_task_time;
383 	int throttled, throttle_count;
384 	struct list_head throttled_list;
385 #endif /* CONFIG_CFS_BANDWIDTH */
386 #endif /* CONFIG_FAIR_GROUP_SCHED */
387 };
388 
389 static inline int rt_bandwidth_enabled(void)
390 {
391 	return sysctl_sched_rt_runtime >= 0;
392 }
393 
394 /* Real-Time classes' related field in a runqueue: */
395 struct rt_rq {
396 	struct rt_prio_array active;
397 	unsigned int rt_nr_running;
398 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
399 	struct {
400 		int curr; /* highest queued rt task prio */
401 #ifdef CONFIG_SMP
402 		int next; /* next highest */
403 #endif
404 	} highest_prio;
405 #endif
406 #ifdef CONFIG_SMP
407 	unsigned long rt_nr_migratory;
408 	unsigned long rt_nr_total;
409 	int overloaded;
410 	struct plist_head pushable_tasks;
411 #endif
412 	int rt_queued;
413 
414 	int rt_throttled;
415 	u64 rt_time;
416 	u64 rt_runtime;
417 	/* Nests inside the rq lock: */
418 	raw_spinlock_t rt_runtime_lock;
419 
420 #ifdef CONFIG_RT_GROUP_SCHED
421 	unsigned long rt_nr_boosted;
422 
423 	struct rq *rq;
424 	struct task_group *tg;
425 #endif
426 };
427 
428 /* Deadline class' related fields in a runqueue */
429 struct dl_rq {
430 	/* runqueue is an rbtree, ordered by deadline */
431 	struct rb_root rb_root;
432 	struct rb_node *rb_leftmost;
433 
434 	unsigned long dl_nr_running;
435 
436 #ifdef CONFIG_SMP
437 	/*
438 	 * Deadline values of the currently executing and the
439 	 * earliest ready task on this rq. Caching these facilitates
440 	 * the decision wether or not a ready but not running task
441 	 * should migrate somewhere else.
442 	 */
443 	struct {
444 		u64 curr;
445 		u64 next;
446 	} earliest_dl;
447 
448 	unsigned long dl_nr_migratory;
449 	int overloaded;
450 
451 	/*
452 	 * Tasks on this rq that can be pushed away. They are kept in
453 	 * an rb-tree, ordered by tasks' deadlines, with caching
454 	 * of the leftmost (earliest deadline) element.
455 	 */
456 	struct rb_root pushable_dl_tasks_root;
457 	struct rb_node *pushable_dl_tasks_leftmost;
458 #else
459 	struct dl_bw dl_bw;
460 #endif
461 };
462 
463 #ifdef CONFIG_SMP
464 
465 /*
466  * We add the notion of a root-domain which will be used to define per-domain
467  * variables. Each exclusive cpuset essentially defines an island domain by
468  * fully partitioning the member cpus from any other cpuset. Whenever a new
469  * exclusive cpuset is created, we also create and attach a new root-domain
470  * object.
471  *
472  */
473 struct root_domain {
474 	atomic_t refcount;
475 	atomic_t rto_count;
476 	struct rcu_head rcu;
477 	cpumask_var_t span;
478 	cpumask_var_t online;
479 
480 	/* Indicate more than one runnable task for any CPU */
481 	bool overload;
482 
483 	/*
484 	 * The bit corresponding to a CPU gets set here if such CPU has more
485 	 * than one runnable -deadline task (as it is below for RT tasks).
486 	 */
487 	cpumask_var_t dlo_mask;
488 	atomic_t dlo_count;
489 	struct dl_bw dl_bw;
490 	struct cpudl cpudl;
491 
492 	/*
493 	 * The "RT overload" flag: it gets set if a CPU has more than
494 	 * one runnable RT task.
495 	 */
496 	cpumask_var_t rto_mask;
497 	struct cpupri cpupri;
498 };
499 
500 extern struct root_domain def_root_domain;
501 
502 #endif /* CONFIG_SMP */
503 
504 /*
505  * This is the main, per-CPU runqueue data structure.
506  *
507  * Locking rule: those places that want to lock multiple runqueues
508  * (such as the load balancing or the thread migration code), lock
509  * acquire operations must be ordered by ascending &runqueue.
510  */
511 struct rq {
512 	/* runqueue lock: */
513 	raw_spinlock_t lock;
514 
515 	/*
516 	 * nr_running and cpu_load should be in the same cacheline because
517 	 * remote CPUs use both these fields when doing load calculation.
518 	 */
519 	unsigned int nr_running;
520 #ifdef CONFIG_NUMA_BALANCING
521 	unsigned int nr_numa_running;
522 	unsigned int nr_preferred_running;
523 #endif
524 	#define CPU_LOAD_IDX_MAX 5
525 	unsigned long cpu_load[CPU_LOAD_IDX_MAX];
526 	unsigned long last_load_update_tick;
527 #ifdef CONFIG_NO_HZ_COMMON
528 	u64 nohz_stamp;
529 	unsigned long nohz_flags;
530 #endif
531 #ifdef CONFIG_NO_HZ_FULL
532 	unsigned long last_sched_tick;
533 #endif
534 	int skip_clock_update;
535 
536 	/* capture load from *all* tasks on this cpu: */
537 	struct load_weight load;
538 	unsigned long nr_load_updates;
539 	u64 nr_switches;
540 
541 	struct cfs_rq cfs;
542 	struct rt_rq rt;
543 	struct dl_rq dl;
544 
545 #ifdef CONFIG_FAIR_GROUP_SCHED
546 	/* list of leaf cfs_rq on this cpu: */
547 	struct list_head leaf_cfs_rq_list;
548 
549 	struct sched_avg avg;
550 #endif /* CONFIG_FAIR_GROUP_SCHED */
551 
552 	/*
553 	 * This is part of a global counter where only the total sum
554 	 * over all CPUs matters. A task can increase this counter on
555 	 * one CPU and if it got migrated afterwards it may decrease
556 	 * it on another CPU. Always updated under the runqueue lock:
557 	 */
558 	unsigned long nr_uninterruptible;
559 
560 	struct task_struct *curr, *idle, *stop;
561 	unsigned long next_balance;
562 	struct mm_struct *prev_mm;
563 
564 	u64 clock;
565 	u64 clock_task;
566 
567 	atomic_t nr_iowait;
568 
569 #ifdef CONFIG_SMP
570 	struct root_domain *rd;
571 	struct sched_domain *sd;
572 
573 	unsigned long cpu_capacity;
574 
575 	unsigned char idle_balance;
576 	/* For active balancing */
577 	int post_schedule;
578 	int active_balance;
579 	int push_cpu;
580 	struct cpu_stop_work active_balance_work;
581 	/* cpu of this runqueue: */
582 	int cpu;
583 	int online;
584 
585 	struct list_head cfs_tasks;
586 
587 	u64 rt_avg;
588 	u64 age_stamp;
589 	u64 idle_stamp;
590 	u64 avg_idle;
591 
592 	/* This is used to determine avg_idle's max value */
593 	u64 max_idle_balance_cost;
594 #endif
595 
596 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
597 	u64 prev_irq_time;
598 #endif
599 #ifdef CONFIG_PARAVIRT
600 	u64 prev_steal_time;
601 #endif
602 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
603 	u64 prev_steal_time_rq;
604 #endif
605 
606 	/* calc_load related fields */
607 	unsigned long calc_load_update;
608 	long calc_load_active;
609 
610 #ifdef CONFIG_SCHED_HRTICK
611 #ifdef CONFIG_SMP
612 	int hrtick_csd_pending;
613 	struct call_single_data hrtick_csd;
614 #endif
615 	struct hrtimer hrtick_timer;
616 #endif
617 
618 #ifdef CONFIG_SCHEDSTATS
619 	/* latency stats */
620 	struct sched_info rq_sched_info;
621 	unsigned long long rq_cpu_time;
622 	/* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
623 
624 	/* sys_sched_yield() stats */
625 	unsigned int yld_count;
626 
627 	/* schedule() stats */
628 	unsigned int sched_count;
629 	unsigned int sched_goidle;
630 
631 	/* try_to_wake_up() stats */
632 	unsigned int ttwu_count;
633 	unsigned int ttwu_local;
634 #endif
635 
636 #ifdef CONFIG_SMP
637 	struct llist_head wake_list;
638 #endif
639 };
640 
641 static inline int cpu_of(struct rq *rq)
642 {
643 #ifdef CONFIG_SMP
644 	return rq->cpu;
645 #else
646 	return 0;
647 #endif
648 }
649 
650 DECLARE_PER_CPU(struct rq, runqueues);
651 
652 #define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
653 #define this_rq()		(&__get_cpu_var(runqueues))
654 #define task_rq(p)		cpu_rq(task_cpu(p))
655 #define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
656 #define raw_rq()		(&__raw_get_cpu_var(runqueues))
657 
658 static inline u64 rq_clock(struct rq *rq)
659 {
660 	return rq->clock;
661 }
662 
663 static inline u64 rq_clock_task(struct rq *rq)
664 {
665 	return rq->clock_task;
666 }
667 
668 #ifdef CONFIG_NUMA_BALANCING
669 extern void sched_setnuma(struct task_struct *p, int node);
670 extern int migrate_task_to(struct task_struct *p, int cpu);
671 extern int migrate_swap(struct task_struct *, struct task_struct *);
672 #endif /* CONFIG_NUMA_BALANCING */
673 
674 #ifdef CONFIG_SMP
675 
676 extern void sched_ttwu_pending(void);
677 
678 #define rcu_dereference_check_sched_domain(p) \
679 	rcu_dereference_check((p), \
680 			      lockdep_is_held(&sched_domains_mutex))
681 
682 /*
683  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684  * See detach_destroy_domains: synchronize_sched for details.
685  *
686  * The domain tree of any CPU may only be accessed from within
687  * preempt-disabled sections.
688  */
689 #define for_each_domain(cpu, __sd) \
690 	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
691 			__sd; __sd = __sd->parent)
692 
693 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
694 
695 /**
696  * highest_flag_domain - Return highest sched_domain containing flag.
697  * @cpu:	The cpu whose highest level of sched domain is to
698  *		be returned.
699  * @flag:	The flag to check for the highest sched_domain
700  *		for the given cpu.
701  *
702  * Returns the highest sched_domain of a cpu which contains the given flag.
703  */
704 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
705 {
706 	struct sched_domain *sd, *hsd = NULL;
707 
708 	for_each_domain(cpu, sd) {
709 		if (!(sd->flags & flag))
710 			break;
711 		hsd = sd;
712 	}
713 
714 	return hsd;
715 }
716 
717 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
718 {
719 	struct sched_domain *sd;
720 
721 	for_each_domain(cpu, sd) {
722 		if (sd->flags & flag)
723 			break;
724 	}
725 
726 	return sd;
727 }
728 
729 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
730 DECLARE_PER_CPU(int, sd_llc_size);
731 DECLARE_PER_CPU(int, sd_llc_id);
732 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
733 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
734 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
735 
736 struct sched_group_capacity {
737 	atomic_t ref;
738 	/*
739 	 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
740 	 * for a single CPU.
741 	 */
742 	unsigned int capacity, capacity_orig;
743 	unsigned long next_update;
744 	int imbalance; /* XXX unrelated to capacity but shared group state */
745 	/*
746 	 * Number of busy cpus in this group.
747 	 */
748 	atomic_t nr_busy_cpus;
749 
750 	unsigned long cpumask[0]; /* iteration mask */
751 };
752 
753 struct sched_group {
754 	struct sched_group *next;	/* Must be a circular list */
755 	atomic_t ref;
756 
757 	unsigned int group_weight;
758 	struct sched_group_capacity *sgc;
759 
760 	/*
761 	 * The CPUs this group covers.
762 	 *
763 	 * NOTE: this field is variable length. (Allocated dynamically
764 	 * by attaching extra space to the end of the structure,
765 	 * depending on how many CPUs the kernel has booted up with)
766 	 */
767 	unsigned long cpumask[0];
768 };
769 
770 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
771 {
772 	return to_cpumask(sg->cpumask);
773 }
774 
775 /*
776  * cpumask masking which cpus in the group are allowed to iterate up the domain
777  * tree.
778  */
779 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
780 {
781 	return to_cpumask(sg->sgc->cpumask);
782 }
783 
784 /**
785  * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
786  * @group: The group whose first cpu is to be returned.
787  */
788 static inline unsigned int group_first_cpu(struct sched_group *group)
789 {
790 	return cpumask_first(sched_group_cpus(group));
791 }
792 
793 extern int group_balance_cpu(struct sched_group *sg);
794 
795 #else
796 
797 static inline void sched_ttwu_pending(void) { }
798 
799 #endif /* CONFIG_SMP */
800 
801 #include "stats.h"
802 #include "auto_group.h"
803 
804 #ifdef CONFIG_CGROUP_SCHED
805 
806 /*
807  * Return the group to which this tasks belongs.
808  *
809  * We cannot use task_css() and friends because the cgroup subsystem
810  * changes that value before the cgroup_subsys::attach() method is called,
811  * therefore we cannot pin it and might observe the wrong value.
812  *
813  * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
814  * core changes this before calling sched_move_task().
815  *
816  * Instead we use a 'copy' which is updated from sched_move_task() while
817  * holding both task_struct::pi_lock and rq::lock.
818  */
819 static inline struct task_group *task_group(struct task_struct *p)
820 {
821 	return p->sched_task_group;
822 }
823 
824 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
825 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
826 {
827 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
828 	struct task_group *tg = task_group(p);
829 #endif
830 
831 #ifdef CONFIG_FAIR_GROUP_SCHED
832 	p->se.cfs_rq = tg->cfs_rq[cpu];
833 	p->se.parent = tg->se[cpu];
834 #endif
835 
836 #ifdef CONFIG_RT_GROUP_SCHED
837 	p->rt.rt_rq  = tg->rt_rq[cpu];
838 	p->rt.parent = tg->rt_se[cpu];
839 #endif
840 }
841 
842 #else /* CONFIG_CGROUP_SCHED */
843 
844 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
845 static inline struct task_group *task_group(struct task_struct *p)
846 {
847 	return NULL;
848 }
849 
850 #endif /* CONFIG_CGROUP_SCHED */
851 
852 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
853 {
854 	set_task_rq(p, cpu);
855 #ifdef CONFIG_SMP
856 	/*
857 	 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
858 	 * successfuly executed on another CPU. We must ensure that updates of
859 	 * per-task data have been completed by this moment.
860 	 */
861 	smp_wmb();
862 	task_thread_info(p)->cpu = cpu;
863 	p->wake_cpu = cpu;
864 #endif
865 }
866 
867 /*
868  * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
869  */
870 #ifdef CONFIG_SCHED_DEBUG
871 # include <linux/static_key.h>
872 # define const_debug __read_mostly
873 #else
874 # define const_debug const
875 #endif
876 
877 extern const_debug unsigned int sysctl_sched_features;
878 
879 #define SCHED_FEAT(name, enabled)	\
880 	__SCHED_FEAT_##name ,
881 
882 enum {
883 #include "features.h"
884 	__SCHED_FEAT_NR,
885 };
886 
887 #undef SCHED_FEAT
888 
889 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
890 #define SCHED_FEAT(name, enabled)					\
891 static __always_inline bool static_branch_##name(struct static_key *key) \
892 {									\
893 	return static_key_##enabled(key);				\
894 }
895 
896 #include "features.h"
897 
898 #undef SCHED_FEAT
899 
900 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
901 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
902 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
903 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
904 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
905 
906 #ifdef CONFIG_NUMA_BALANCING
907 #define sched_feat_numa(x) sched_feat(x)
908 #ifdef CONFIG_SCHED_DEBUG
909 #define numabalancing_enabled sched_feat_numa(NUMA)
910 #else
911 extern bool numabalancing_enabled;
912 #endif /* CONFIG_SCHED_DEBUG */
913 #else
914 #define sched_feat_numa(x) (0)
915 #define numabalancing_enabled (0)
916 #endif /* CONFIG_NUMA_BALANCING */
917 
918 static inline u64 global_rt_period(void)
919 {
920 	return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
921 }
922 
923 static inline u64 global_rt_runtime(void)
924 {
925 	if (sysctl_sched_rt_runtime < 0)
926 		return RUNTIME_INF;
927 
928 	return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
929 }
930 
931 static inline int task_current(struct rq *rq, struct task_struct *p)
932 {
933 	return rq->curr == p;
934 }
935 
936 static inline int task_running(struct rq *rq, struct task_struct *p)
937 {
938 #ifdef CONFIG_SMP
939 	return p->on_cpu;
940 #else
941 	return task_current(rq, p);
942 #endif
943 }
944 
945 
946 #ifndef prepare_arch_switch
947 # define prepare_arch_switch(next)	do { } while (0)
948 #endif
949 #ifndef finish_arch_switch
950 # define finish_arch_switch(prev)	do { } while (0)
951 #endif
952 #ifndef finish_arch_post_lock_switch
953 # define finish_arch_post_lock_switch()	do { } while (0)
954 #endif
955 
956 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
957 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
958 {
959 #ifdef CONFIG_SMP
960 	/*
961 	 * We can optimise this out completely for !SMP, because the
962 	 * SMP rebalancing from interrupt is the only thing that cares
963 	 * here.
964 	 */
965 	next->on_cpu = 1;
966 #endif
967 }
968 
969 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
970 {
971 #ifdef CONFIG_SMP
972 	/*
973 	 * After ->on_cpu is cleared, the task can be moved to a different CPU.
974 	 * We must ensure this doesn't happen until the switch is completely
975 	 * finished.
976 	 */
977 	smp_wmb();
978 	prev->on_cpu = 0;
979 #endif
980 #ifdef CONFIG_DEBUG_SPINLOCK
981 	/* this is a valid case when another task releases the spinlock */
982 	rq->lock.owner = current;
983 #endif
984 	/*
985 	 * If we are tracking spinlock dependencies then we have to
986 	 * fix up the runqueue lock - which gets 'carried over' from
987 	 * prev into current:
988 	 */
989 	spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
990 
991 	raw_spin_unlock_irq(&rq->lock);
992 }
993 
994 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
995 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
996 {
997 #ifdef CONFIG_SMP
998 	/*
999 	 * We can optimise this out completely for !SMP, because the
1000 	 * SMP rebalancing from interrupt is the only thing that cares
1001 	 * here.
1002 	 */
1003 	next->on_cpu = 1;
1004 #endif
1005 	raw_spin_unlock(&rq->lock);
1006 }
1007 
1008 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1009 {
1010 #ifdef CONFIG_SMP
1011 	/*
1012 	 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1013 	 * We must ensure this doesn't happen until the switch is completely
1014 	 * finished.
1015 	 */
1016 	smp_wmb();
1017 	prev->on_cpu = 0;
1018 #endif
1019 	local_irq_enable();
1020 }
1021 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1022 
1023 /*
1024  * wake flags
1025  */
1026 #define WF_SYNC		0x01		/* waker goes to sleep after wakeup */
1027 #define WF_FORK		0x02		/* child wakeup after fork */
1028 #define WF_MIGRATED	0x4		/* internal use, task got migrated */
1029 
1030 /*
1031  * To aid in avoiding the subversion of "niceness" due to uneven distribution
1032  * of tasks with abnormal "nice" values across CPUs the contribution that
1033  * each task makes to its run queue's load is weighted according to its
1034  * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1035  * scaled version of the new time slice allocation that they receive on time
1036  * slice expiry etc.
1037  */
1038 
1039 #define WEIGHT_IDLEPRIO                3
1040 #define WMULT_IDLEPRIO         1431655765
1041 
1042 /*
1043  * Nice levels are multiplicative, with a gentle 10% change for every
1044  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1045  * nice 1, it will get ~10% less CPU time than another CPU-bound task
1046  * that remained on nice 0.
1047  *
1048  * The "10% effect" is relative and cumulative: from _any_ nice level,
1049  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1050  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1051  * If a task goes up by ~10% and another task goes down by ~10% then
1052  * the relative distance between them is ~25%.)
1053  */
1054 static const int prio_to_weight[40] = {
1055  /* -20 */     88761,     71755,     56483,     46273,     36291,
1056  /* -15 */     29154,     23254,     18705,     14949,     11916,
1057  /* -10 */      9548,      7620,      6100,      4904,      3906,
1058  /*  -5 */      3121,      2501,      1991,      1586,      1277,
1059  /*   0 */      1024,       820,       655,       526,       423,
1060  /*   5 */       335,       272,       215,       172,       137,
1061  /*  10 */       110,        87,        70,        56,        45,
1062  /*  15 */        36,        29,        23,        18,        15,
1063 };
1064 
1065 /*
1066  * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1067  *
1068  * In cases where the weight does not change often, we can use the
1069  * precalculated inverse to speed up arithmetics by turning divisions
1070  * into multiplications:
1071  */
1072 static const u32 prio_to_wmult[40] = {
1073  /* -20 */     48388,     59856,     76040,     92818,    118348,
1074  /* -15 */    147320,    184698,    229616,    287308,    360437,
1075  /* -10 */    449829,    563644,    704093,    875809,   1099582,
1076  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1077  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1078  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1079  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1080  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1081 };
1082 
1083 #define ENQUEUE_WAKEUP		1
1084 #define ENQUEUE_HEAD		2
1085 #ifdef CONFIG_SMP
1086 #define ENQUEUE_WAKING		4	/* sched_class::task_waking was called */
1087 #else
1088 #define ENQUEUE_WAKING		0
1089 #endif
1090 #define ENQUEUE_REPLENISH	8
1091 
1092 #define DEQUEUE_SLEEP		1
1093 
1094 #define RETRY_TASK		((void *)-1UL)
1095 
1096 struct sched_class {
1097 	const struct sched_class *next;
1098 
1099 	void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1100 	void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1101 	void (*yield_task) (struct rq *rq);
1102 	bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1103 
1104 	void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1105 
1106 	/*
1107 	 * It is the responsibility of the pick_next_task() method that will
1108 	 * return the next task to call put_prev_task() on the @prev task or
1109 	 * something equivalent.
1110 	 *
1111 	 * May return RETRY_TASK when it finds a higher prio class has runnable
1112 	 * tasks.
1113 	 */
1114 	struct task_struct * (*pick_next_task) (struct rq *rq,
1115 						struct task_struct *prev);
1116 	void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1117 
1118 #ifdef CONFIG_SMP
1119 	int  (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1120 	void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1121 
1122 	void (*post_schedule) (struct rq *this_rq);
1123 	void (*task_waking) (struct task_struct *task);
1124 	void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1125 
1126 	void (*set_cpus_allowed)(struct task_struct *p,
1127 				 const struct cpumask *newmask);
1128 
1129 	void (*rq_online)(struct rq *rq);
1130 	void (*rq_offline)(struct rq *rq);
1131 #endif
1132 
1133 	void (*set_curr_task) (struct rq *rq);
1134 	void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1135 	void (*task_fork) (struct task_struct *p);
1136 	void (*task_dead) (struct task_struct *p);
1137 
1138 	void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1139 	void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1140 	void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1141 			     int oldprio);
1142 
1143 	unsigned int (*get_rr_interval) (struct rq *rq,
1144 					 struct task_struct *task);
1145 
1146 #ifdef CONFIG_FAIR_GROUP_SCHED
1147 	void (*task_move_group) (struct task_struct *p, int on_rq);
1148 #endif
1149 };
1150 
1151 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1152 {
1153 	prev->sched_class->put_prev_task(rq, prev);
1154 }
1155 
1156 #define sched_class_highest (&stop_sched_class)
1157 #define for_each_class(class) \
1158    for (class = sched_class_highest; class; class = class->next)
1159 
1160 extern const struct sched_class stop_sched_class;
1161 extern const struct sched_class dl_sched_class;
1162 extern const struct sched_class rt_sched_class;
1163 extern const struct sched_class fair_sched_class;
1164 extern const struct sched_class idle_sched_class;
1165 
1166 
1167 #ifdef CONFIG_SMP
1168 
1169 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1170 
1171 extern void trigger_load_balance(struct rq *rq);
1172 
1173 extern void idle_enter_fair(struct rq *this_rq);
1174 extern void idle_exit_fair(struct rq *this_rq);
1175 
1176 #else
1177 
1178 static inline void idle_enter_fair(struct rq *rq) { }
1179 static inline void idle_exit_fair(struct rq *rq) { }
1180 
1181 #endif
1182 
1183 extern void sysrq_sched_debug_show(void);
1184 extern void sched_init_granularity(void);
1185 extern void update_max_interval(void);
1186 
1187 extern void init_sched_dl_class(void);
1188 extern void init_sched_rt_class(void);
1189 extern void init_sched_fair_class(void);
1190 extern void init_sched_dl_class(void);
1191 
1192 extern void resched_curr(struct rq *rq);
1193 extern void resched_cpu(int cpu);
1194 
1195 extern struct rt_bandwidth def_rt_bandwidth;
1196 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1197 
1198 extern struct dl_bandwidth def_dl_bandwidth;
1199 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1200 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1201 
1202 unsigned long to_ratio(u64 period, u64 runtime);
1203 
1204 extern void update_idle_cpu_load(struct rq *this_rq);
1205 
1206 extern void init_task_runnable_average(struct task_struct *p);
1207 
1208 static inline void add_nr_running(struct rq *rq, unsigned count)
1209 {
1210 	unsigned prev_nr = rq->nr_running;
1211 
1212 	rq->nr_running = prev_nr + count;
1213 
1214 	if (prev_nr < 2 && rq->nr_running >= 2) {
1215 #ifdef CONFIG_SMP
1216 		if (!rq->rd->overload)
1217 			rq->rd->overload = true;
1218 #endif
1219 
1220 #ifdef CONFIG_NO_HZ_FULL
1221 		if (tick_nohz_full_cpu(rq->cpu)) {
1222 			/*
1223 			 * Tick is needed if more than one task runs on a CPU.
1224 			 * Send the target an IPI to kick it out of nohz mode.
1225 			 *
1226 			 * We assume that IPI implies full memory barrier and the
1227 			 * new value of rq->nr_running is visible on reception
1228 			 * from the target.
1229 			 */
1230 			tick_nohz_full_kick_cpu(rq->cpu);
1231 		}
1232 #endif
1233 	}
1234 }
1235 
1236 static inline void sub_nr_running(struct rq *rq, unsigned count)
1237 {
1238 	rq->nr_running -= count;
1239 }
1240 
1241 static inline void rq_last_tick_reset(struct rq *rq)
1242 {
1243 #ifdef CONFIG_NO_HZ_FULL
1244 	rq->last_sched_tick = jiffies;
1245 #endif
1246 }
1247 
1248 extern void update_rq_clock(struct rq *rq);
1249 
1250 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1251 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1252 
1253 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1254 
1255 extern const_debug unsigned int sysctl_sched_time_avg;
1256 extern const_debug unsigned int sysctl_sched_nr_migrate;
1257 extern const_debug unsigned int sysctl_sched_migration_cost;
1258 
1259 static inline u64 sched_avg_period(void)
1260 {
1261 	return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 }
1263 
1264 #ifdef CONFIG_SCHED_HRTICK
1265 
1266 /*
1267  * Use hrtick when:
1268  *  - enabled by features
1269  *  - hrtimer is actually high res
1270  */
1271 static inline int hrtick_enabled(struct rq *rq)
1272 {
1273 	if (!sched_feat(HRTICK))
1274 		return 0;
1275 	if (!cpu_active(cpu_of(rq)))
1276 		return 0;
1277 	return hrtimer_is_hres_active(&rq->hrtick_timer);
1278 }
1279 
1280 void hrtick_start(struct rq *rq, u64 delay);
1281 
1282 #else
1283 
1284 static inline int hrtick_enabled(struct rq *rq)
1285 {
1286 	return 0;
1287 }
1288 
1289 #endif /* CONFIG_SCHED_HRTICK */
1290 
1291 #ifdef CONFIG_SMP
1292 extern void sched_avg_update(struct rq *rq);
1293 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1294 {
1295 	rq->rt_avg += rt_delta;
1296 	sched_avg_update(rq);
1297 }
1298 #else
1299 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1300 static inline void sched_avg_update(struct rq *rq) { }
1301 #endif
1302 
1303 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1304 
1305 #ifdef CONFIG_SMP
1306 #ifdef CONFIG_PREEMPT
1307 
1308 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1309 
1310 /*
1311  * fair double_lock_balance: Safely acquires both rq->locks in a fair
1312  * way at the expense of forcing extra atomic operations in all
1313  * invocations.  This assures that the double_lock is acquired using the
1314  * same underlying policy as the spinlock_t on this architecture, which
1315  * reduces latency compared to the unfair variant below.  However, it
1316  * also adds more overhead and therefore may reduce throughput.
1317  */
1318 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1319 	__releases(this_rq->lock)
1320 	__acquires(busiest->lock)
1321 	__acquires(this_rq->lock)
1322 {
1323 	raw_spin_unlock(&this_rq->lock);
1324 	double_rq_lock(this_rq, busiest);
1325 
1326 	return 1;
1327 }
1328 
1329 #else
1330 /*
1331  * Unfair double_lock_balance: Optimizes throughput at the expense of
1332  * latency by eliminating extra atomic operations when the locks are
1333  * already in proper order on entry.  This favors lower cpu-ids and will
1334  * grant the double lock to lower cpus over higher ids under contention,
1335  * regardless of entry order into the function.
1336  */
1337 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1338 	__releases(this_rq->lock)
1339 	__acquires(busiest->lock)
1340 	__acquires(this_rq->lock)
1341 {
1342 	int ret = 0;
1343 
1344 	if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1345 		if (busiest < this_rq) {
1346 			raw_spin_unlock(&this_rq->lock);
1347 			raw_spin_lock(&busiest->lock);
1348 			raw_spin_lock_nested(&this_rq->lock,
1349 					      SINGLE_DEPTH_NESTING);
1350 			ret = 1;
1351 		} else
1352 			raw_spin_lock_nested(&busiest->lock,
1353 					      SINGLE_DEPTH_NESTING);
1354 	}
1355 	return ret;
1356 }
1357 
1358 #endif /* CONFIG_PREEMPT */
1359 
1360 /*
1361  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1362  */
1363 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1364 {
1365 	if (unlikely(!irqs_disabled())) {
1366 		/* printk() doesn't work good under rq->lock */
1367 		raw_spin_unlock(&this_rq->lock);
1368 		BUG_ON(1);
1369 	}
1370 
1371 	return _double_lock_balance(this_rq, busiest);
1372 }
1373 
1374 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1375 	__releases(busiest->lock)
1376 {
1377 	raw_spin_unlock(&busiest->lock);
1378 	lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1379 }
1380 
1381 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1382 {
1383 	if (l1 > l2)
1384 		swap(l1, l2);
1385 
1386 	spin_lock(l1);
1387 	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1388 }
1389 
1390 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1391 {
1392 	if (l1 > l2)
1393 		swap(l1, l2);
1394 
1395 	spin_lock_irq(l1);
1396 	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1397 }
1398 
1399 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1400 {
1401 	if (l1 > l2)
1402 		swap(l1, l2);
1403 
1404 	raw_spin_lock(l1);
1405 	raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1406 }
1407 
1408 /*
1409  * double_rq_lock - safely lock two runqueues
1410  *
1411  * Note this does not disable interrupts like task_rq_lock,
1412  * you need to do so manually before calling.
1413  */
1414 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1415 	__acquires(rq1->lock)
1416 	__acquires(rq2->lock)
1417 {
1418 	BUG_ON(!irqs_disabled());
1419 	if (rq1 == rq2) {
1420 		raw_spin_lock(&rq1->lock);
1421 		__acquire(rq2->lock);	/* Fake it out ;) */
1422 	} else {
1423 		if (rq1 < rq2) {
1424 			raw_spin_lock(&rq1->lock);
1425 			raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1426 		} else {
1427 			raw_spin_lock(&rq2->lock);
1428 			raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1429 		}
1430 	}
1431 }
1432 
1433 /*
1434  * double_rq_unlock - safely unlock two runqueues
1435  *
1436  * Note this does not restore interrupts like task_rq_unlock,
1437  * you need to do so manually after calling.
1438  */
1439 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1440 	__releases(rq1->lock)
1441 	__releases(rq2->lock)
1442 {
1443 	raw_spin_unlock(&rq1->lock);
1444 	if (rq1 != rq2)
1445 		raw_spin_unlock(&rq2->lock);
1446 	else
1447 		__release(rq2->lock);
1448 }
1449 
1450 #else /* CONFIG_SMP */
1451 
1452 /*
1453  * double_rq_lock - safely lock two runqueues
1454  *
1455  * Note this does not disable interrupts like task_rq_lock,
1456  * you need to do so manually before calling.
1457  */
1458 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1459 	__acquires(rq1->lock)
1460 	__acquires(rq2->lock)
1461 {
1462 	BUG_ON(!irqs_disabled());
1463 	BUG_ON(rq1 != rq2);
1464 	raw_spin_lock(&rq1->lock);
1465 	__acquire(rq2->lock);	/* Fake it out ;) */
1466 }
1467 
1468 /*
1469  * double_rq_unlock - safely unlock two runqueues
1470  *
1471  * Note this does not restore interrupts like task_rq_unlock,
1472  * you need to do so manually after calling.
1473  */
1474 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1475 	__releases(rq1->lock)
1476 	__releases(rq2->lock)
1477 {
1478 	BUG_ON(rq1 != rq2);
1479 	raw_spin_unlock(&rq1->lock);
1480 	__release(rq2->lock);
1481 }
1482 
1483 #endif
1484 
1485 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1486 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1487 extern void print_cfs_stats(struct seq_file *m, int cpu);
1488 extern void print_rt_stats(struct seq_file *m, int cpu);
1489 
1490 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1491 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1492 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1493 
1494 extern void cfs_bandwidth_usage_inc(void);
1495 extern void cfs_bandwidth_usage_dec(void);
1496 
1497 #ifdef CONFIG_NO_HZ_COMMON
1498 enum rq_nohz_flag_bits {
1499 	NOHZ_TICK_STOPPED,
1500 	NOHZ_BALANCE_KICK,
1501 };
1502 
1503 #define nohz_flags(cpu)	(&cpu_rq(cpu)->nohz_flags)
1504 #endif
1505 
1506 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1507 
1508 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1509 DECLARE_PER_CPU(u64, cpu_softirq_time);
1510 
1511 #ifndef CONFIG_64BIT
1512 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1513 
1514 static inline void irq_time_write_begin(void)
1515 {
1516 	__this_cpu_inc(irq_time_seq.sequence);
1517 	smp_wmb();
1518 }
1519 
1520 static inline void irq_time_write_end(void)
1521 {
1522 	smp_wmb();
1523 	__this_cpu_inc(irq_time_seq.sequence);
1524 }
1525 
1526 static inline u64 irq_time_read(int cpu)
1527 {
1528 	u64 irq_time;
1529 	unsigned seq;
1530 
1531 	do {
1532 		seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1533 		irq_time = per_cpu(cpu_softirq_time, cpu) +
1534 			   per_cpu(cpu_hardirq_time, cpu);
1535 	} while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1536 
1537 	return irq_time;
1538 }
1539 #else /* CONFIG_64BIT */
1540 static inline void irq_time_write_begin(void)
1541 {
1542 }
1543 
1544 static inline void irq_time_write_end(void)
1545 {
1546 }
1547 
1548 static inline u64 irq_time_read(int cpu)
1549 {
1550 	return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1551 }
1552 #endif /* CONFIG_64BIT */
1553 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1554