xref: /openbmc/linux/kernel/sched/sched.h (revision 6774def6)
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 struct cpuidle_state;
18 
19 /* task_struct::on_rq states: */
20 #define TASK_ON_RQ_QUEUED	1
21 #define TASK_ON_RQ_MIGRATING	2
22 
23 extern __read_mostly int scheduler_running;
24 
25 extern unsigned long calc_load_update;
26 extern atomic_long_t calc_load_tasks;
27 
28 extern long calc_load_fold_active(struct rq *this_rq);
29 extern void update_cpu_load_active(struct rq *this_rq);
30 
31 /*
32  * Helpers for converting nanosecond timing to jiffy resolution
33  */
34 #define NS_TO_JIFFIES(TIME)	((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
35 
36 /*
37  * Increase resolution of nice-level calculations for 64-bit architectures.
38  * The extra resolution improves shares distribution and load balancing of
39  * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
40  * hierarchies, especially on larger systems. This is not a user-visible change
41  * and does not change the user-interface for setting shares/weights.
42  *
43  * We increase resolution only if we have enough bits to allow this increased
44  * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
45  * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
46  * increased costs.
47  */
48 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load  */
49 # define SCHED_LOAD_RESOLUTION	10
50 # define scale_load(w)		((w) << SCHED_LOAD_RESOLUTION)
51 # define scale_load_down(w)	((w) >> SCHED_LOAD_RESOLUTION)
52 #else
53 # define SCHED_LOAD_RESOLUTION	0
54 # define scale_load(w)		(w)
55 # define scale_load_down(w)	(w)
56 #endif
57 
58 #define SCHED_LOAD_SHIFT	(10 + SCHED_LOAD_RESOLUTION)
59 #define SCHED_LOAD_SCALE	(1L << SCHED_LOAD_SHIFT)
60 
61 #define NICE_0_LOAD		SCHED_LOAD_SCALE
62 #define NICE_0_SHIFT		SCHED_LOAD_SHIFT
63 
64 /*
65  * Single value that decides SCHED_DEADLINE internal math precision.
66  * 10 -> just above 1us
67  * 9  -> just above 0.5us
68  */
69 #define DL_SCALE (10)
70 
71 /*
72  * These are the 'tuning knobs' of the scheduler:
73  */
74 
75 /*
76  * single value that denotes runtime == period, ie unlimited time.
77  */
78 #define RUNTIME_INF	((u64)~0ULL)
79 
80 static inline int fair_policy(int policy)
81 {
82 	return policy == SCHED_NORMAL || policy == SCHED_BATCH;
83 }
84 
85 static inline int rt_policy(int policy)
86 {
87 	return policy == SCHED_FIFO || policy == SCHED_RR;
88 }
89 
90 static inline int dl_policy(int policy)
91 {
92 	return policy == SCHED_DEADLINE;
93 }
94 
95 static inline int task_has_rt_policy(struct task_struct *p)
96 {
97 	return rt_policy(p->policy);
98 }
99 
100 static inline int task_has_dl_policy(struct task_struct *p)
101 {
102 	return dl_policy(p->policy);
103 }
104 
105 static inline bool dl_time_before(u64 a, u64 b)
106 {
107 	return (s64)(a - b) < 0;
108 }
109 
110 /*
111  * Tells if entity @a should preempt entity @b.
112  */
113 static inline bool
114 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
115 {
116 	return dl_time_before(a->deadline, b->deadline);
117 }
118 
119 /*
120  * This is the priority-queue data structure of the RT scheduling class:
121  */
122 struct rt_prio_array {
123 	DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
124 	struct list_head queue[MAX_RT_PRIO];
125 };
126 
127 struct rt_bandwidth {
128 	/* nests inside the rq lock: */
129 	raw_spinlock_t		rt_runtime_lock;
130 	ktime_t			rt_period;
131 	u64			rt_runtime;
132 	struct hrtimer		rt_period_timer;
133 };
134 
135 void __dl_clear_params(struct task_struct *p);
136 
137 /*
138  * To keep the bandwidth of -deadline tasks and groups under control
139  * we need some place where:
140  *  - store the maximum -deadline bandwidth of the system (the group);
141  *  - cache the fraction of that bandwidth that is currently allocated.
142  *
143  * This is all done in the data structure below. It is similar to the
144  * one used for RT-throttling (rt_bandwidth), with the main difference
145  * that, since here we are only interested in admission control, we
146  * do not decrease any runtime while the group "executes", neither we
147  * need a timer to replenish it.
148  *
149  * With respect to SMP, the bandwidth is given on a per-CPU basis,
150  * meaning that:
151  *  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
152  *  - dl_total_bw array contains, in the i-eth element, the currently
153  *    allocated bandwidth on the i-eth CPU.
154  * Moreover, groups consume bandwidth on each CPU, while tasks only
155  * consume bandwidth on the CPU they're running on.
156  * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
157  * that will be shown the next time the proc or cgroup controls will
158  * be red. It on its turn can be changed by writing on its own
159  * control.
160  */
161 struct dl_bandwidth {
162 	raw_spinlock_t dl_runtime_lock;
163 	u64 dl_runtime;
164 	u64 dl_period;
165 };
166 
167 static inline int dl_bandwidth_enabled(void)
168 {
169 	return sysctl_sched_rt_runtime >= 0;
170 }
171 
172 extern struct dl_bw *dl_bw_of(int i);
173 
174 struct dl_bw {
175 	raw_spinlock_t lock;
176 	u64 bw, total_bw;
177 };
178 
179 extern struct mutex sched_domains_mutex;
180 
181 #ifdef CONFIG_CGROUP_SCHED
182 
183 #include <linux/cgroup.h>
184 
185 struct cfs_rq;
186 struct rt_rq;
187 
188 extern struct list_head task_groups;
189 
190 struct cfs_bandwidth {
191 #ifdef CONFIG_CFS_BANDWIDTH
192 	raw_spinlock_t lock;
193 	ktime_t period;
194 	u64 quota, runtime;
195 	s64 hierarchical_quota;
196 	u64 runtime_expires;
197 
198 	int idle, timer_active;
199 	struct hrtimer period_timer, slack_timer;
200 	struct list_head throttled_cfs_rq;
201 
202 	/* statistics */
203 	int nr_periods, nr_throttled;
204 	u64 throttled_time;
205 #endif
206 };
207 
208 /* task group related information */
209 struct task_group {
210 	struct cgroup_subsys_state css;
211 
212 #ifdef CONFIG_FAIR_GROUP_SCHED
213 	/* schedulable entities of this group on each cpu */
214 	struct sched_entity **se;
215 	/* runqueue "owned" by this group on each cpu */
216 	struct cfs_rq **cfs_rq;
217 	unsigned long shares;
218 
219 #ifdef	CONFIG_SMP
220 	atomic_long_t load_avg;
221 	atomic_t runnable_avg;
222 #endif
223 #endif
224 
225 #ifdef CONFIG_RT_GROUP_SCHED
226 	struct sched_rt_entity **rt_se;
227 	struct rt_rq **rt_rq;
228 
229 	struct rt_bandwidth rt_bandwidth;
230 #endif
231 
232 	struct rcu_head rcu;
233 	struct list_head list;
234 
235 	struct task_group *parent;
236 	struct list_head siblings;
237 	struct list_head children;
238 
239 #ifdef CONFIG_SCHED_AUTOGROUP
240 	struct autogroup *autogroup;
241 #endif
242 
243 	struct cfs_bandwidth cfs_bandwidth;
244 };
245 
246 #ifdef CONFIG_FAIR_GROUP_SCHED
247 #define ROOT_TASK_GROUP_LOAD	NICE_0_LOAD
248 
249 /*
250  * A weight of 0 or 1 can cause arithmetics problems.
251  * A weight of a cfs_rq is the sum of weights of which entities
252  * are queued on this cfs_rq, so a weight of a entity should not be
253  * too large, so as the shares value of a task group.
254  * (The default weight is 1024 - so there's no practical
255  *  limitation from this.)
256  */
257 #define MIN_SHARES	(1UL <<  1)
258 #define MAX_SHARES	(1UL << 18)
259 #endif
260 
261 typedef int (*tg_visitor)(struct task_group *, void *);
262 
263 extern int walk_tg_tree_from(struct task_group *from,
264 			     tg_visitor down, tg_visitor up, void *data);
265 
266 /*
267  * Iterate the full tree, calling @down when first entering a node and @up when
268  * leaving it for the final time.
269  *
270  * Caller must hold rcu_lock or sufficient equivalent.
271  */
272 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
273 {
274 	return walk_tg_tree_from(&root_task_group, down, up, data);
275 }
276 
277 extern int tg_nop(struct task_group *tg, void *data);
278 
279 extern void free_fair_sched_group(struct task_group *tg);
280 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
281 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
282 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
283 			struct sched_entity *se, int cpu,
284 			struct sched_entity *parent);
285 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
286 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
287 
288 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
289 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
290 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
291 
292 extern void free_rt_sched_group(struct task_group *tg);
293 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
294 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
295 		struct sched_rt_entity *rt_se, int cpu,
296 		struct sched_rt_entity *parent);
297 
298 extern struct task_group *sched_create_group(struct task_group *parent);
299 extern void sched_online_group(struct task_group *tg,
300 			       struct task_group *parent);
301 extern void sched_destroy_group(struct task_group *tg);
302 extern void sched_offline_group(struct task_group *tg);
303 
304 extern void sched_move_task(struct task_struct *tsk);
305 
306 #ifdef CONFIG_FAIR_GROUP_SCHED
307 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
308 #endif
309 
310 #else /* CONFIG_CGROUP_SCHED */
311 
312 struct cfs_bandwidth { };
313 
314 #endif	/* CONFIG_CGROUP_SCHED */
315 
316 /* CFS-related fields in a runqueue */
317 struct cfs_rq {
318 	struct load_weight load;
319 	unsigned int nr_running, h_nr_running;
320 
321 	u64 exec_clock;
322 	u64 min_vruntime;
323 #ifndef CONFIG_64BIT
324 	u64 min_vruntime_copy;
325 #endif
326 
327 	struct rb_root tasks_timeline;
328 	struct rb_node *rb_leftmost;
329 
330 	/*
331 	 * 'curr' points to currently running entity on this cfs_rq.
332 	 * It is set to NULL otherwise (i.e when none are currently running).
333 	 */
334 	struct sched_entity *curr, *next, *last, *skip;
335 
336 #ifdef	CONFIG_SCHED_DEBUG
337 	unsigned int nr_spread_over;
338 #endif
339 
340 #ifdef CONFIG_SMP
341 	/*
342 	 * CFS Load tracking
343 	 * Under CFS, load is tracked on a per-entity basis and aggregated up.
344 	 * This allows for the description of both thread and group usage (in
345 	 * the FAIR_GROUP_SCHED case).
346 	 */
347 	unsigned long runnable_load_avg, blocked_load_avg;
348 	atomic64_t decay_counter;
349 	u64 last_decay;
350 	atomic_long_t removed_load;
351 
352 #ifdef CONFIG_FAIR_GROUP_SCHED
353 	/* Required to track per-cpu representation of a task_group */
354 	u32 tg_runnable_contrib;
355 	unsigned long tg_load_contrib;
356 
357 	/*
358 	 *   h_load = weight * f(tg)
359 	 *
360 	 * Where f(tg) is the recursive weight fraction assigned to
361 	 * this group.
362 	 */
363 	unsigned long h_load;
364 	u64 last_h_load_update;
365 	struct sched_entity *h_load_next;
366 #endif /* CONFIG_FAIR_GROUP_SCHED */
367 #endif /* CONFIG_SMP */
368 
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 	struct rq *rq;	/* cpu runqueue to which this cfs_rq is attached */
371 
372 	/*
373 	 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
374 	 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
375 	 * (like users, containers etc.)
376 	 *
377 	 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
378 	 * list is used during load balance.
379 	 */
380 	int on_list;
381 	struct list_head leaf_cfs_rq_list;
382 	struct task_group *tg;	/* group that "owns" this runqueue */
383 
384 #ifdef CONFIG_CFS_BANDWIDTH
385 	int runtime_enabled;
386 	u64 runtime_expires;
387 	s64 runtime_remaining;
388 
389 	u64 throttled_clock, throttled_clock_task;
390 	u64 throttled_clock_task_time;
391 	int throttled, throttle_count;
392 	struct list_head throttled_list;
393 #endif /* CONFIG_CFS_BANDWIDTH */
394 #endif /* CONFIG_FAIR_GROUP_SCHED */
395 };
396 
397 static inline int rt_bandwidth_enabled(void)
398 {
399 	return sysctl_sched_rt_runtime >= 0;
400 }
401 
402 /* Real-Time classes' related field in a runqueue: */
403 struct rt_rq {
404 	struct rt_prio_array active;
405 	unsigned int rt_nr_running;
406 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
407 	struct {
408 		int curr; /* highest queued rt task prio */
409 #ifdef CONFIG_SMP
410 		int next; /* next highest */
411 #endif
412 	} highest_prio;
413 #endif
414 #ifdef CONFIG_SMP
415 	unsigned long rt_nr_migratory;
416 	unsigned long rt_nr_total;
417 	int overloaded;
418 	struct plist_head pushable_tasks;
419 #endif
420 	int rt_queued;
421 
422 	int rt_throttled;
423 	u64 rt_time;
424 	u64 rt_runtime;
425 	/* Nests inside the rq lock: */
426 	raw_spinlock_t rt_runtime_lock;
427 
428 #ifdef CONFIG_RT_GROUP_SCHED
429 	unsigned long rt_nr_boosted;
430 
431 	struct rq *rq;
432 	struct task_group *tg;
433 #endif
434 };
435 
436 /* Deadline class' related fields in a runqueue */
437 struct dl_rq {
438 	/* runqueue is an rbtree, ordered by deadline */
439 	struct rb_root rb_root;
440 	struct rb_node *rb_leftmost;
441 
442 	unsigned long dl_nr_running;
443 
444 #ifdef CONFIG_SMP
445 	/*
446 	 * Deadline values of the currently executing and the
447 	 * earliest ready task on this rq. Caching these facilitates
448 	 * the decision wether or not a ready but not running task
449 	 * should migrate somewhere else.
450 	 */
451 	struct {
452 		u64 curr;
453 		u64 next;
454 	} earliest_dl;
455 
456 	unsigned long dl_nr_migratory;
457 	int overloaded;
458 
459 	/*
460 	 * Tasks on this rq that can be pushed away. They are kept in
461 	 * an rb-tree, ordered by tasks' deadlines, with caching
462 	 * of the leftmost (earliest deadline) element.
463 	 */
464 	struct rb_root pushable_dl_tasks_root;
465 	struct rb_node *pushable_dl_tasks_leftmost;
466 #else
467 	struct dl_bw dl_bw;
468 #endif
469 };
470 
471 #ifdef CONFIG_SMP
472 
473 /*
474  * We add the notion of a root-domain which will be used to define per-domain
475  * variables. Each exclusive cpuset essentially defines an island domain by
476  * fully partitioning the member cpus from any other cpuset. Whenever a new
477  * exclusive cpuset is created, we also create and attach a new root-domain
478  * object.
479  *
480  */
481 struct root_domain {
482 	atomic_t refcount;
483 	atomic_t rto_count;
484 	struct rcu_head rcu;
485 	cpumask_var_t span;
486 	cpumask_var_t online;
487 
488 	/* Indicate more than one runnable task for any CPU */
489 	bool overload;
490 
491 	/*
492 	 * The bit corresponding to a CPU gets set here if such CPU has more
493 	 * than one runnable -deadline task (as it is below for RT tasks).
494 	 */
495 	cpumask_var_t dlo_mask;
496 	atomic_t dlo_count;
497 	struct dl_bw dl_bw;
498 	struct cpudl cpudl;
499 
500 	/*
501 	 * The "RT overload" flag: it gets set if a CPU has more than
502 	 * one runnable RT task.
503 	 */
504 	cpumask_var_t rto_mask;
505 	struct cpupri cpupri;
506 };
507 
508 extern struct root_domain def_root_domain;
509 
510 #endif /* CONFIG_SMP */
511 
512 /*
513  * This is the main, per-CPU runqueue data structure.
514  *
515  * Locking rule: those places that want to lock multiple runqueues
516  * (such as the load balancing or the thread migration code), lock
517  * acquire operations must be ordered by ascending &runqueue.
518  */
519 struct rq {
520 	/* runqueue lock: */
521 	raw_spinlock_t lock;
522 
523 	/*
524 	 * nr_running and cpu_load should be in the same cacheline because
525 	 * remote CPUs use both these fields when doing load calculation.
526 	 */
527 	unsigned int nr_running;
528 #ifdef CONFIG_NUMA_BALANCING
529 	unsigned int nr_numa_running;
530 	unsigned int nr_preferred_running;
531 #endif
532 	#define CPU_LOAD_IDX_MAX 5
533 	unsigned long cpu_load[CPU_LOAD_IDX_MAX];
534 	unsigned long last_load_update_tick;
535 #ifdef CONFIG_NO_HZ_COMMON
536 	u64 nohz_stamp;
537 	unsigned long nohz_flags;
538 #endif
539 #ifdef CONFIG_NO_HZ_FULL
540 	unsigned long last_sched_tick;
541 #endif
542 	int skip_clock_update;
543 
544 	/* capture load from *all* tasks on this cpu: */
545 	struct load_weight load;
546 	unsigned long nr_load_updates;
547 	u64 nr_switches;
548 
549 	struct cfs_rq cfs;
550 	struct rt_rq rt;
551 	struct dl_rq dl;
552 
553 #ifdef CONFIG_FAIR_GROUP_SCHED
554 	/* list of leaf cfs_rq on this cpu: */
555 	struct list_head leaf_cfs_rq_list;
556 
557 	struct sched_avg avg;
558 #endif /* CONFIG_FAIR_GROUP_SCHED */
559 
560 	/*
561 	 * This is part of a global counter where only the total sum
562 	 * over all CPUs matters. A task can increase this counter on
563 	 * one CPU and if it got migrated afterwards it may decrease
564 	 * it on another CPU. Always updated under the runqueue lock:
565 	 */
566 	unsigned long nr_uninterruptible;
567 
568 	struct task_struct *curr, *idle, *stop;
569 	unsigned long next_balance;
570 	struct mm_struct *prev_mm;
571 
572 	u64 clock;
573 	u64 clock_task;
574 
575 	atomic_t nr_iowait;
576 
577 #ifdef CONFIG_SMP
578 	struct root_domain *rd;
579 	struct sched_domain *sd;
580 
581 	unsigned long cpu_capacity;
582 
583 	unsigned char idle_balance;
584 	/* For active balancing */
585 	int post_schedule;
586 	int active_balance;
587 	int push_cpu;
588 	struct cpu_stop_work active_balance_work;
589 	/* cpu of this runqueue: */
590 	int cpu;
591 	int online;
592 
593 	struct list_head cfs_tasks;
594 
595 	u64 rt_avg;
596 	u64 age_stamp;
597 	u64 idle_stamp;
598 	u64 avg_idle;
599 
600 	/* This is used to determine avg_idle's max value */
601 	u64 max_idle_balance_cost;
602 #endif
603 
604 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
605 	u64 prev_irq_time;
606 #endif
607 #ifdef CONFIG_PARAVIRT
608 	u64 prev_steal_time;
609 #endif
610 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
611 	u64 prev_steal_time_rq;
612 #endif
613 
614 	/* calc_load related fields */
615 	unsigned long calc_load_update;
616 	long calc_load_active;
617 
618 #ifdef CONFIG_SCHED_HRTICK
619 #ifdef CONFIG_SMP
620 	int hrtick_csd_pending;
621 	struct call_single_data hrtick_csd;
622 #endif
623 	struct hrtimer hrtick_timer;
624 #endif
625 
626 #ifdef CONFIG_SCHEDSTATS
627 	/* latency stats */
628 	struct sched_info rq_sched_info;
629 	unsigned long long rq_cpu_time;
630 	/* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
631 
632 	/* sys_sched_yield() stats */
633 	unsigned int yld_count;
634 
635 	/* schedule() stats */
636 	unsigned int sched_count;
637 	unsigned int sched_goidle;
638 
639 	/* try_to_wake_up() stats */
640 	unsigned int ttwu_count;
641 	unsigned int ttwu_local;
642 #endif
643 
644 #ifdef CONFIG_SMP
645 	struct llist_head wake_list;
646 #endif
647 
648 #ifdef CONFIG_CPU_IDLE
649 	/* Must be inspected within a rcu lock section */
650 	struct cpuidle_state *idle_state;
651 #endif
652 };
653 
654 static inline int cpu_of(struct rq *rq)
655 {
656 #ifdef CONFIG_SMP
657 	return rq->cpu;
658 #else
659 	return 0;
660 #endif
661 }
662 
663 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
664 
665 #define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
666 #define this_rq()		this_cpu_ptr(&runqueues)
667 #define task_rq(p)		cpu_rq(task_cpu(p))
668 #define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
669 #define raw_rq()		raw_cpu_ptr(&runqueues)
670 
671 static inline u64 rq_clock(struct rq *rq)
672 {
673 	return rq->clock;
674 }
675 
676 static inline u64 rq_clock_task(struct rq *rq)
677 {
678 	return rq->clock_task;
679 }
680 
681 #ifdef CONFIG_NUMA_BALANCING
682 extern void sched_setnuma(struct task_struct *p, int node);
683 extern int migrate_task_to(struct task_struct *p, int cpu);
684 extern int migrate_swap(struct task_struct *, struct task_struct *);
685 #endif /* CONFIG_NUMA_BALANCING */
686 
687 #ifdef CONFIG_SMP
688 
689 extern void sched_ttwu_pending(void);
690 
691 #define rcu_dereference_check_sched_domain(p) \
692 	rcu_dereference_check((p), \
693 			      lockdep_is_held(&sched_domains_mutex))
694 
695 /*
696  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
697  * See detach_destroy_domains: synchronize_sched for details.
698  *
699  * The domain tree of any CPU may only be accessed from within
700  * preempt-disabled sections.
701  */
702 #define for_each_domain(cpu, __sd) \
703 	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
704 			__sd; __sd = __sd->parent)
705 
706 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
707 
708 /**
709  * highest_flag_domain - Return highest sched_domain containing flag.
710  * @cpu:	The cpu whose highest level of sched domain is to
711  *		be returned.
712  * @flag:	The flag to check for the highest sched_domain
713  *		for the given cpu.
714  *
715  * Returns the highest sched_domain of a cpu which contains the given flag.
716  */
717 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
718 {
719 	struct sched_domain *sd, *hsd = NULL;
720 
721 	for_each_domain(cpu, sd) {
722 		if (!(sd->flags & flag))
723 			break;
724 		hsd = sd;
725 	}
726 
727 	return hsd;
728 }
729 
730 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
731 {
732 	struct sched_domain *sd;
733 
734 	for_each_domain(cpu, sd) {
735 		if (sd->flags & flag)
736 			break;
737 	}
738 
739 	return sd;
740 }
741 
742 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
743 DECLARE_PER_CPU(int, sd_llc_size);
744 DECLARE_PER_CPU(int, sd_llc_id);
745 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
746 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
747 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
748 
749 struct sched_group_capacity {
750 	atomic_t ref;
751 	/*
752 	 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
753 	 * for a single CPU.
754 	 */
755 	unsigned int capacity, capacity_orig;
756 	unsigned long next_update;
757 	int imbalance; /* XXX unrelated to capacity but shared group state */
758 	/*
759 	 * Number of busy cpus in this group.
760 	 */
761 	atomic_t nr_busy_cpus;
762 
763 	unsigned long cpumask[0]; /* iteration mask */
764 };
765 
766 struct sched_group {
767 	struct sched_group *next;	/* Must be a circular list */
768 	atomic_t ref;
769 
770 	unsigned int group_weight;
771 	struct sched_group_capacity *sgc;
772 
773 	/*
774 	 * The CPUs this group covers.
775 	 *
776 	 * NOTE: this field is variable length. (Allocated dynamically
777 	 * by attaching extra space to the end of the structure,
778 	 * depending on how many CPUs the kernel has booted up with)
779 	 */
780 	unsigned long cpumask[0];
781 };
782 
783 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
784 {
785 	return to_cpumask(sg->cpumask);
786 }
787 
788 /*
789  * cpumask masking which cpus in the group are allowed to iterate up the domain
790  * tree.
791  */
792 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
793 {
794 	return to_cpumask(sg->sgc->cpumask);
795 }
796 
797 /**
798  * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
799  * @group: The group whose first cpu is to be returned.
800  */
801 static inline unsigned int group_first_cpu(struct sched_group *group)
802 {
803 	return cpumask_first(sched_group_cpus(group));
804 }
805 
806 extern int group_balance_cpu(struct sched_group *sg);
807 
808 #else
809 
810 static inline void sched_ttwu_pending(void) { }
811 
812 #endif /* CONFIG_SMP */
813 
814 #include "stats.h"
815 #include "auto_group.h"
816 
817 #ifdef CONFIG_CGROUP_SCHED
818 
819 /*
820  * Return the group to which this tasks belongs.
821  *
822  * We cannot use task_css() and friends because the cgroup subsystem
823  * changes that value before the cgroup_subsys::attach() method is called,
824  * therefore we cannot pin it and might observe the wrong value.
825  *
826  * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
827  * core changes this before calling sched_move_task().
828  *
829  * Instead we use a 'copy' which is updated from sched_move_task() while
830  * holding both task_struct::pi_lock and rq::lock.
831  */
832 static inline struct task_group *task_group(struct task_struct *p)
833 {
834 	return p->sched_task_group;
835 }
836 
837 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
838 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
839 {
840 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
841 	struct task_group *tg = task_group(p);
842 #endif
843 
844 #ifdef CONFIG_FAIR_GROUP_SCHED
845 	p->se.cfs_rq = tg->cfs_rq[cpu];
846 	p->se.parent = tg->se[cpu];
847 #endif
848 
849 #ifdef CONFIG_RT_GROUP_SCHED
850 	p->rt.rt_rq  = tg->rt_rq[cpu];
851 	p->rt.parent = tg->rt_se[cpu];
852 #endif
853 }
854 
855 #else /* CONFIG_CGROUP_SCHED */
856 
857 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
858 static inline struct task_group *task_group(struct task_struct *p)
859 {
860 	return NULL;
861 }
862 
863 #endif /* CONFIG_CGROUP_SCHED */
864 
865 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
866 {
867 	set_task_rq(p, cpu);
868 #ifdef CONFIG_SMP
869 	/*
870 	 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
871 	 * successfuly executed on another CPU. We must ensure that updates of
872 	 * per-task data have been completed by this moment.
873 	 */
874 	smp_wmb();
875 	task_thread_info(p)->cpu = cpu;
876 	p->wake_cpu = cpu;
877 #endif
878 }
879 
880 /*
881  * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
882  */
883 #ifdef CONFIG_SCHED_DEBUG
884 # include <linux/static_key.h>
885 # define const_debug __read_mostly
886 #else
887 # define const_debug const
888 #endif
889 
890 extern const_debug unsigned int sysctl_sched_features;
891 
892 #define SCHED_FEAT(name, enabled)	\
893 	__SCHED_FEAT_##name ,
894 
895 enum {
896 #include "features.h"
897 	__SCHED_FEAT_NR,
898 };
899 
900 #undef SCHED_FEAT
901 
902 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
903 #define SCHED_FEAT(name, enabled)					\
904 static __always_inline bool static_branch_##name(struct static_key *key) \
905 {									\
906 	return static_key_##enabled(key);				\
907 }
908 
909 #include "features.h"
910 
911 #undef SCHED_FEAT
912 
913 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
914 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
915 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
916 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
917 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
918 
919 #ifdef CONFIG_NUMA_BALANCING
920 #define sched_feat_numa(x) sched_feat(x)
921 #ifdef CONFIG_SCHED_DEBUG
922 #define numabalancing_enabled sched_feat_numa(NUMA)
923 #else
924 extern bool numabalancing_enabled;
925 #endif /* CONFIG_SCHED_DEBUG */
926 #else
927 #define sched_feat_numa(x) (0)
928 #define numabalancing_enabled (0)
929 #endif /* CONFIG_NUMA_BALANCING */
930 
931 static inline u64 global_rt_period(void)
932 {
933 	return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
934 }
935 
936 static inline u64 global_rt_runtime(void)
937 {
938 	if (sysctl_sched_rt_runtime < 0)
939 		return RUNTIME_INF;
940 
941 	return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
942 }
943 
944 static inline int task_current(struct rq *rq, struct task_struct *p)
945 {
946 	return rq->curr == p;
947 }
948 
949 static inline int task_running(struct rq *rq, struct task_struct *p)
950 {
951 #ifdef CONFIG_SMP
952 	return p->on_cpu;
953 #else
954 	return task_current(rq, p);
955 #endif
956 }
957 
958 static inline int task_on_rq_queued(struct task_struct *p)
959 {
960 	return p->on_rq == TASK_ON_RQ_QUEUED;
961 }
962 
963 static inline int task_on_rq_migrating(struct task_struct *p)
964 {
965 	return p->on_rq == TASK_ON_RQ_MIGRATING;
966 }
967 
968 #ifndef prepare_arch_switch
969 # define prepare_arch_switch(next)	do { } while (0)
970 #endif
971 #ifndef finish_arch_switch
972 # define finish_arch_switch(prev)	do { } while (0)
973 #endif
974 #ifndef finish_arch_post_lock_switch
975 # define finish_arch_post_lock_switch()	do { } while (0)
976 #endif
977 
978 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
979 {
980 #ifdef CONFIG_SMP
981 	/*
982 	 * We can optimise this out completely for !SMP, because the
983 	 * SMP rebalancing from interrupt is the only thing that cares
984 	 * here.
985 	 */
986 	next->on_cpu = 1;
987 #endif
988 }
989 
990 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
991 {
992 #ifdef CONFIG_SMP
993 	/*
994 	 * After ->on_cpu is cleared, the task can be moved to a different CPU.
995 	 * We must ensure this doesn't happen until the switch is completely
996 	 * finished.
997 	 */
998 	smp_wmb();
999 	prev->on_cpu = 0;
1000 #endif
1001 #ifdef CONFIG_DEBUG_SPINLOCK
1002 	/* this is a valid case when another task releases the spinlock */
1003 	rq->lock.owner = current;
1004 #endif
1005 	/*
1006 	 * If we are tracking spinlock dependencies then we have to
1007 	 * fix up the runqueue lock - which gets 'carried over' from
1008 	 * prev into current:
1009 	 */
1010 	spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1011 
1012 	raw_spin_unlock_irq(&rq->lock);
1013 }
1014 
1015 /*
1016  * wake flags
1017  */
1018 #define WF_SYNC		0x01		/* waker goes to sleep after wakeup */
1019 #define WF_FORK		0x02		/* child wakeup after fork */
1020 #define WF_MIGRATED	0x4		/* internal use, task got migrated */
1021 
1022 /*
1023  * To aid in avoiding the subversion of "niceness" due to uneven distribution
1024  * of tasks with abnormal "nice" values across CPUs the contribution that
1025  * each task makes to its run queue's load is weighted according to its
1026  * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1027  * scaled version of the new time slice allocation that they receive on time
1028  * slice expiry etc.
1029  */
1030 
1031 #define WEIGHT_IDLEPRIO                3
1032 #define WMULT_IDLEPRIO         1431655765
1033 
1034 /*
1035  * Nice levels are multiplicative, with a gentle 10% change for every
1036  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1037  * nice 1, it will get ~10% less CPU time than another CPU-bound task
1038  * that remained on nice 0.
1039  *
1040  * The "10% effect" is relative and cumulative: from _any_ nice level,
1041  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1042  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1043  * If a task goes up by ~10% and another task goes down by ~10% then
1044  * the relative distance between them is ~25%.)
1045  */
1046 static const int prio_to_weight[40] = {
1047  /* -20 */     88761,     71755,     56483,     46273,     36291,
1048  /* -15 */     29154,     23254,     18705,     14949,     11916,
1049  /* -10 */      9548,      7620,      6100,      4904,      3906,
1050  /*  -5 */      3121,      2501,      1991,      1586,      1277,
1051  /*   0 */      1024,       820,       655,       526,       423,
1052  /*   5 */       335,       272,       215,       172,       137,
1053  /*  10 */       110,        87,        70,        56,        45,
1054  /*  15 */        36,        29,        23,        18,        15,
1055 };
1056 
1057 /*
1058  * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1059  *
1060  * In cases where the weight does not change often, we can use the
1061  * precalculated inverse to speed up arithmetics by turning divisions
1062  * into multiplications:
1063  */
1064 static const u32 prio_to_wmult[40] = {
1065  /* -20 */     48388,     59856,     76040,     92818,    118348,
1066  /* -15 */    147320,    184698,    229616,    287308,    360437,
1067  /* -10 */    449829,    563644,    704093,    875809,   1099582,
1068  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1069  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1070  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1071  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1072  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1073 };
1074 
1075 #define ENQUEUE_WAKEUP		1
1076 #define ENQUEUE_HEAD		2
1077 #ifdef CONFIG_SMP
1078 #define ENQUEUE_WAKING		4	/* sched_class::task_waking was called */
1079 #else
1080 #define ENQUEUE_WAKING		0
1081 #endif
1082 #define ENQUEUE_REPLENISH	8
1083 
1084 #define DEQUEUE_SLEEP		1
1085 
1086 #define RETRY_TASK		((void *)-1UL)
1087 
1088 struct sched_class {
1089 	const struct sched_class *next;
1090 
1091 	void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1092 	void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1093 	void (*yield_task) (struct rq *rq);
1094 	bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1095 
1096 	void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1097 
1098 	/*
1099 	 * It is the responsibility of the pick_next_task() method that will
1100 	 * return the next task to call put_prev_task() on the @prev task or
1101 	 * something equivalent.
1102 	 *
1103 	 * May return RETRY_TASK when it finds a higher prio class has runnable
1104 	 * tasks.
1105 	 */
1106 	struct task_struct * (*pick_next_task) (struct rq *rq,
1107 						struct task_struct *prev);
1108 	void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1109 
1110 #ifdef CONFIG_SMP
1111 	int  (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1112 	void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1113 
1114 	void (*post_schedule) (struct rq *this_rq);
1115 	void (*task_waking) (struct task_struct *task);
1116 	void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1117 
1118 	void (*set_cpus_allowed)(struct task_struct *p,
1119 				 const struct cpumask *newmask);
1120 
1121 	void (*rq_online)(struct rq *rq);
1122 	void (*rq_offline)(struct rq *rq);
1123 #endif
1124 
1125 	void (*set_curr_task) (struct rq *rq);
1126 	void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1127 	void (*task_fork) (struct task_struct *p);
1128 	void (*task_dead) (struct task_struct *p);
1129 
1130 	void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1131 	void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1132 	void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1133 			     int oldprio);
1134 
1135 	unsigned int (*get_rr_interval) (struct rq *rq,
1136 					 struct task_struct *task);
1137 
1138 #ifdef CONFIG_FAIR_GROUP_SCHED
1139 	void (*task_move_group) (struct task_struct *p, int on_rq);
1140 #endif
1141 };
1142 
1143 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1144 {
1145 	prev->sched_class->put_prev_task(rq, prev);
1146 }
1147 
1148 #define sched_class_highest (&stop_sched_class)
1149 #define for_each_class(class) \
1150    for (class = sched_class_highest; class; class = class->next)
1151 
1152 extern const struct sched_class stop_sched_class;
1153 extern const struct sched_class dl_sched_class;
1154 extern const struct sched_class rt_sched_class;
1155 extern const struct sched_class fair_sched_class;
1156 extern const struct sched_class idle_sched_class;
1157 
1158 
1159 #ifdef CONFIG_SMP
1160 
1161 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1162 
1163 extern void trigger_load_balance(struct rq *rq);
1164 
1165 extern void idle_enter_fair(struct rq *this_rq);
1166 extern void idle_exit_fair(struct rq *this_rq);
1167 
1168 #else
1169 
1170 static inline void idle_enter_fair(struct rq *rq) { }
1171 static inline void idle_exit_fair(struct rq *rq) { }
1172 
1173 #endif
1174 
1175 #ifdef CONFIG_CPU_IDLE
1176 static inline void idle_set_state(struct rq *rq,
1177 				  struct cpuidle_state *idle_state)
1178 {
1179 	rq->idle_state = idle_state;
1180 }
1181 
1182 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1183 {
1184 	WARN_ON(!rcu_read_lock_held());
1185 	return rq->idle_state;
1186 }
1187 #else
1188 static inline void idle_set_state(struct rq *rq,
1189 				  struct cpuidle_state *idle_state)
1190 {
1191 }
1192 
1193 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1194 {
1195 	return NULL;
1196 }
1197 #endif
1198 
1199 extern void sysrq_sched_debug_show(void);
1200 extern void sched_init_granularity(void);
1201 extern void update_max_interval(void);
1202 
1203 extern void init_sched_dl_class(void);
1204 extern void init_sched_rt_class(void);
1205 extern void init_sched_fair_class(void);
1206 extern void init_sched_dl_class(void);
1207 
1208 extern void resched_curr(struct rq *rq);
1209 extern void resched_cpu(int cpu);
1210 
1211 extern struct rt_bandwidth def_rt_bandwidth;
1212 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1213 
1214 extern struct dl_bandwidth def_dl_bandwidth;
1215 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1216 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1217 
1218 unsigned long to_ratio(u64 period, u64 runtime);
1219 
1220 extern void update_idle_cpu_load(struct rq *this_rq);
1221 
1222 extern void init_task_runnable_average(struct task_struct *p);
1223 
1224 static inline void add_nr_running(struct rq *rq, unsigned count)
1225 {
1226 	unsigned prev_nr = rq->nr_running;
1227 
1228 	rq->nr_running = prev_nr + count;
1229 
1230 	if (prev_nr < 2 && rq->nr_running >= 2) {
1231 #ifdef CONFIG_SMP
1232 		if (!rq->rd->overload)
1233 			rq->rd->overload = true;
1234 #endif
1235 
1236 #ifdef CONFIG_NO_HZ_FULL
1237 		if (tick_nohz_full_cpu(rq->cpu)) {
1238 			/*
1239 			 * Tick is needed if more than one task runs on a CPU.
1240 			 * Send the target an IPI to kick it out of nohz mode.
1241 			 *
1242 			 * We assume that IPI implies full memory barrier and the
1243 			 * new value of rq->nr_running is visible on reception
1244 			 * from the target.
1245 			 */
1246 			tick_nohz_full_kick_cpu(rq->cpu);
1247 		}
1248 #endif
1249 	}
1250 }
1251 
1252 static inline void sub_nr_running(struct rq *rq, unsigned count)
1253 {
1254 	rq->nr_running -= count;
1255 }
1256 
1257 static inline void rq_last_tick_reset(struct rq *rq)
1258 {
1259 #ifdef CONFIG_NO_HZ_FULL
1260 	rq->last_sched_tick = jiffies;
1261 #endif
1262 }
1263 
1264 extern void update_rq_clock(struct rq *rq);
1265 
1266 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1267 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1268 
1269 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1270 
1271 extern const_debug unsigned int sysctl_sched_time_avg;
1272 extern const_debug unsigned int sysctl_sched_nr_migrate;
1273 extern const_debug unsigned int sysctl_sched_migration_cost;
1274 
1275 static inline u64 sched_avg_period(void)
1276 {
1277 	return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1278 }
1279 
1280 #ifdef CONFIG_SCHED_HRTICK
1281 
1282 /*
1283  * Use hrtick when:
1284  *  - enabled by features
1285  *  - hrtimer is actually high res
1286  */
1287 static inline int hrtick_enabled(struct rq *rq)
1288 {
1289 	if (!sched_feat(HRTICK))
1290 		return 0;
1291 	if (!cpu_active(cpu_of(rq)))
1292 		return 0;
1293 	return hrtimer_is_hres_active(&rq->hrtick_timer);
1294 }
1295 
1296 void hrtick_start(struct rq *rq, u64 delay);
1297 
1298 #else
1299 
1300 static inline int hrtick_enabled(struct rq *rq)
1301 {
1302 	return 0;
1303 }
1304 
1305 #endif /* CONFIG_SCHED_HRTICK */
1306 
1307 #ifdef CONFIG_SMP
1308 extern void sched_avg_update(struct rq *rq);
1309 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1310 {
1311 	rq->rt_avg += rt_delta;
1312 	sched_avg_update(rq);
1313 }
1314 #else
1315 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1316 static inline void sched_avg_update(struct rq *rq) { }
1317 #endif
1318 
1319 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1320 
1321 #ifdef CONFIG_SMP
1322 #ifdef CONFIG_PREEMPT
1323 
1324 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1325 
1326 /*
1327  * fair double_lock_balance: Safely acquires both rq->locks in a fair
1328  * way at the expense of forcing extra atomic operations in all
1329  * invocations.  This assures that the double_lock is acquired using the
1330  * same underlying policy as the spinlock_t on this architecture, which
1331  * reduces latency compared to the unfair variant below.  However, it
1332  * also adds more overhead and therefore may reduce throughput.
1333  */
1334 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1335 	__releases(this_rq->lock)
1336 	__acquires(busiest->lock)
1337 	__acquires(this_rq->lock)
1338 {
1339 	raw_spin_unlock(&this_rq->lock);
1340 	double_rq_lock(this_rq, busiest);
1341 
1342 	return 1;
1343 }
1344 
1345 #else
1346 /*
1347  * Unfair double_lock_balance: Optimizes throughput at the expense of
1348  * latency by eliminating extra atomic operations when the locks are
1349  * already in proper order on entry.  This favors lower cpu-ids and will
1350  * grant the double lock to lower cpus over higher ids under contention,
1351  * regardless of entry order into the function.
1352  */
1353 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1354 	__releases(this_rq->lock)
1355 	__acquires(busiest->lock)
1356 	__acquires(this_rq->lock)
1357 {
1358 	int ret = 0;
1359 
1360 	if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1361 		if (busiest < this_rq) {
1362 			raw_spin_unlock(&this_rq->lock);
1363 			raw_spin_lock(&busiest->lock);
1364 			raw_spin_lock_nested(&this_rq->lock,
1365 					      SINGLE_DEPTH_NESTING);
1366 			ret = 1;
1367 		} else
1368 			raw_spin_lock_nested(&busiest->lock,
1369 					      SINGLE_DEPTH_NESTING);
1370 	}
1371 	return ret;
1372 }
1373 
1374 #endif /* CONFIG_PREEMPT */
1375 
1376 /*
1377  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1378  */
1379 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1380 {
1381 	if (unlikely(!irqs_disabled())) {
1382 		/* printk() doesn't work good under rq->lock */
1383 		raw_spin_unlock(&this_rq->lock);
1384 		BUG_ON(1);
1385 	}
1386 
1387 	return _double_lock_balance(this_rq, busiest);
1388 }
1389 
1390 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1391 	__releases(busiest->lock)
1392 {
1393 	raw_spin_unlock(&busiest->lock);
1394 	lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1395 }
1396 
1397 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1398 {
1399 	if (l1 > l2)
1400 		swap(l1, l2);
1401 
1402 	spin_lock(l1);
1403 	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1404 }
1405 
1406 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1407 {
1408 	if (l1 > l2)
1409 		swap(l1, l2);
1410 
1411 	spin_lock_irq(l1);
1412 	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1413 }
1414 
1415 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1416 {
1417 	if (l1 > l2)
1418 		swap(l1, l2);
1419 
1420 	raw_spin_lock(l1);
1421 	raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1422 }
1423 
1424 /*
1425  * double_rq_lock - safely lock two runqueues
1426  *
1427  * Note this does not disable interrupts like task_rq_lock,
1428  * you need to do so manually before calling.
1429  */
1430 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1431 	__acquires(rq1->lock)
1432 	__acquires(rq2->lock)
1433 {
1434 	BUG_ON(!irqs_disabled());
1435 	if (rq1 == rq2) {
1436 		raw_spin_lock(&rq1->lock);
1437 		__acquire(rq2->lock);	/* Fake it out ;) */
1438 	} else {
1439 		if (rq1 < rq2) {
1440 			raw_spin_lock(&rq1->lock);
1441 			raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1442 		} else {
1443 			raw_spin_lock(&rq2->lock);
1444 			raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1445 		}
1446 	}
1447 }
1448 
1449 /*
1450  * double_rq_unlock - safely unlock two runqueues
1451  *
1452  * Note this does not restore interrupts like task_rq_unlock,
1453  * you need to do so manually after calling.
1454  */
1455 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1456 	__releases(rq1->lock)
1457 	__releases(rq2->lock)
1458 {
1459 	raw_spin_unlock(&rq1->lock);
1460 	if (rq1 != rq2)
1461 		raw_spin_unlock(&rq2->lock);
1462 	else
1463 		__release(rq2->lock);
1464 }
1465 
1466 #else /* CONFIG_SMP */
1467 
1468 /*
1469  * double_rq_lock - safely lock two runqueues
1470  *
1471  * Note this does not disable interrupts like task_rq_lock,
1472  * you need to do so manually before calling.
1473  */
1474 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1475 	__acquires(rq1->lock)
1476 	__acquires(rq2->lock)
1477 {
1478 	BUG_ON(!irqs_disabled());
1479 	BUG_ON(rq1 != rq2);
1480 	raw_spin_lock(&rq1->lock);
1481 	__acquire(rq2->lock);	/* Fake it out ;) */
1482 }
1483 
1484 /*
1485  * double_rq_unlock - safely unlock two runqueues
1486  *
1487  * Note this does not restore interrupts like task_rq_unlock,
1488  * you need to do so manually after calling.
1489  */
1490 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1491 	__releases(rq1->lock)
1492 	__releases(rq2->lock)
1493 {
1494 	BUG_ON(rq1 != rq2);
1495 	raw_spin_unlock(&rq1->lock);
1496 	__release(rq2->lock);
1497 }
1498 
1499 #endif
1500 
1501 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1502 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1503 extern void print_cfs_stats(struct seq_file *m, int cpu);
1504 extern void print_rt_stats(struct seq_file *m, int cpu);
1505 
1506 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1507 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1508 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1509 
1510 extern void cfs_bandwidth_usage_inc(void);
1511 extern void cfs_bandwidth_usage_dec(void);
1512 
1513 #ifdef CONFIG_NO_HZ_COMMON
1514 enum rq_nohz_flag_bits {
1515 	NOHZ_TICK_STOPPED,
1516 	NOHZ_BALANCE_KICK,
1517 };
1518 
1519 #define nohz_flags(cpu)	(&cpu_rq(cpu)->nohz_flags)
1520 #endif
1521 
1522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1523 
1524 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1525 DECLARE_PER_CPU(u64, cpu_softirq_time);
1526 
1527 #ifndef CONFIG_64BIT
1528 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1529 
1530 static inline void irq_time_write_begin(void)
1531 {
1532 	__this_cpu_inc(irq_time_seq.sequence);
1533 	smp_wmb();
1534 }
1535 
1536 static inline void irq_time_write_end(void)
1537 {
1538 	smp_wmb();
1539 	__this_cpu_inc(irq_time_seq.sequence);
1540 }
1541 
1542 static inline u64 irq_time_read(int cpu)
1543 {
1544 	u64 irq_time;
1545 	unsigned seq;
1546 
1547 	do {
1548 		seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1549 		irq_time = per_cpu(cpu_softirq_time, cpu) +
1550 			   per_cpu(cpu_hardirq_time, cpu);
1551 	} while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1552 
1553 	return irq_time;
1554 }
1555 #else /* CONFIG_64BIT */
1556 static inline void irq_time_write_begin(void)
1557 {
1558 }
1559 
1560 static inline void irq_time_write_end(void)
1561 {
1562 }
1563 
1564 static inline u64 irq_time_read(int cpu)
1565 {
1566 	return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1567 }
1568 #endif /* CONFIG_64BIT */
1569 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1570