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