Loading...
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
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * Scheduler internal types and methods:
4 */
5#include <linux/sched.h>
6
7#include <linux/sched/autogroup.h>
8#include <linux/sched/clock.h>
9#include <linux/sched/coredump.h>
10#include <linux/sched/cpufreq.h>
11#include <linux/sched/cputime.h>
12#include <linux/sched/deadline.h>
13#include <linux/sched/debug.h>
14#include <linux/sched/hotplug.h>
15#include <linux/sched/idle.h>
16#include <linux/sched/init.h>
17#include <linux/sched/isolation.h>
18#include <linux/sched/jobctl.h>
19#include <linux/sched/loadavg.h>
20#include <linux/sched/mm.h>
21#include <linux/sched/nohz.h>
22#include <linux/sched/numa_balancing.h>
23#include <linux/sched/prio.h>
24#include <linux/sched/rt.h>
25#include <linux/sched/signal.h>
26#include <linux/sched/smt.h>
27#include <linux/sched/stat.h>
28#include <linux/sched/sysctl.h>
29#include <linux/sched/task.h>
30#include <linux/sched/task_stack.h>
31#include <linux/sched/topology.h>
32#include <linux/sched/user.h>
33#include <linux/sched/wake_q.h>
34#include <linux/sched/xacct.h>
35
36#include <uapi/linux/sched/types.h>
37
38#include <linux/binfmts.h>
39#include <linux/bitops.h>
40#include <linux/blkdev.h>
41#include <linux/compat.h>
42#include <linux/context_tracking.h>
43#include <linux/cpufreq.h>
44#include <linux/cpuidle.h>
45#include <linux/cpuset.h>
46#include <linux/ctype.h>
47#include <linux/debugfs.h>
48#include <linux/delayacct.h>
49#include <linux/energy_model.h>
50#include <linux/init_task.h>
51#include <linux/kprobes.h>
52#include <linux/kthread.h>
53#include <linux/membarrier.h>
54#include <linux/migrate.h>
55#include <linux/mmu_context.h>
56#include <linux/nmi.h>
57#include <linux/proc_fs.h>
58#include <linux/prefetch.h>
59#include <linux/profile.h>
60#include <linux/psi.h>
61#include <linux/ratelimit.h>
62#include <linux/rcupdate_wait.h>
63#include <linux/security.h>
64#include <linux/stop_machine.h>
65#include <linux/suspend.h>
66#include <linux/swait.h>
67#include <linux/syscalls.h>
68#include <linux/task_work.h>
69#include <linux/tsacct_kern.h>
70
71#include <asm/tlb.h>
72
73#ifdef CONFIG_PARAVIRT
74# include <asm/paravirt.h>
75#endif
76
77#include "cpupri.h"
78#include "cpudeadline.h"
79
80#include <trace/events/sched.h>
81
82#ifdef CONFIG_SCHED_DEBUG
83# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
84#else
85# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
86#endif
87
88struct rq;
89struct cpuidle_state;
90
91/* task_struct::on_rq states: */
92#define TASK_ON_RQ_QUEUED 1
93#define TASK_ON_RQ_MIGRATING 2
94
95extern __read_mostly int scheduler_running;
96
97extern unsigned long calc_load_update;
98extern atomic_long_t calc_load_tasks;
99
100extern void calc_global_load_tick(struct rq *this_rq);
101extern long calc_load_fold_active(struct rq *this_rq, long adjust);
102
103extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
104/*
105 * Helpers for converting nanosecond timing to jiffy resolution
106 */
107#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108
109/*
110 * Increase resolution of nice-level calculations for 64-bit architectures.
111 * The extra resolution improves shares distribution and load balancing of
112 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
113 * hierarchies, especially on larger systems. This is not a user-visible change
114 * and does not change the user-interface for setting shares/weights.
115 *
116 * We increase resolution only if we have enough bits to allow this increased
117 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
118 * are pretty high and the returns do not justify the increased costs.
119 *
120 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
121 * increase coverage and consistency always enable it on 64-bit platforms.
122 */
123#ifdef CONFIG_64BIT
124# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
125# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
126# define scale_load_down(w) \
127({ \
128 unsigned long __w = (w); \
129 if (__w) \
130 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
131 __w; \
132})
133#else
134# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
135# define scale_load(w) (w)
136# define scale_load_down(w) (w)
137#endif
138
139/*
140 * Task weight (visible to users) and its load (invisible to users) have
141 * independent resolution, but they should be well calibrated. We use
142 * scale_load() and scale_load_down(w) to convert between them. The
143 * following must be true:
144 *
145 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
146 *
147 */
148#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
149
150/*
151 * Single value that decides SCHED_DEADLINE internal math precision.
152 * 10 -> just above 1us
153 * 9 -> just above 0.5us
154 */
155#define DL_SCALE 10
156
157/*
158 * Single value that denotes runtime == period, ie unlimited time.
159 */
160#define RUNTIME_INF ((u64)~0ULL)
161
162static inline int idle_policy(int policy)
163{
164 return policy == SCHED_IDLE;
165}
166static inline int fair_policy(int policy)
167{
168 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
169}
170
171static inline int rt_policy(int policy)
172{
173 return policy == SCHED_FIFO || policy == SCHED_RR;
174}
175
176static inline int dl_policy(int policy)
177{
178 return policy == SCHED_DEADLINE;
179}
180static inline bool valid_policy(int policy)
181{
182 return idle_policy(policy) || fair_policy(policy) ||
183 rt_policy(policy) || dl_policy(policy);
184}
185
186static inline int task_has_idle_policy(struct task_struct *p)
187{
188 return idle_policy(p->policy);
189}
190
191static inline int task_has_rt_policy(struct task_struct *p)
192{
193 return rt_policy(p->policy);
194}
195
196static inline int task_has_dl_policy(struct task_struct *p)
197{
198 return dl_policy(p->policy);
199}
200
201#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
202
203static inline void update_avg(u64 *avg, u64 sample)
204{
205 s64 diff = sample - *avg;
206 *avg += diff / 8;
207}
208
209/*
210 * Shifting a value by an exponent greater *or equal* to the size of said value
211 * is UB; cap at size-1.
212 */
213#define shr_bound(val, shift) \
214 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
215
216/*
217 * !! For sched_setattr_nocheck() (kernel) only !!
218 *
219 * This is actually gross. :(
220 *
221 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
222 * tasks, but still be able to sleep. We need this on platforms that cannot
223 * atomically change clock frequency. Remove once fast switching will be
224 * available on such platforms.
225 *
226 * SUGOV stands for SchedUtil GOVernor.
227 */
228#define SCHED_FLAG_SUGOV 0x10000000
229
230#define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
231
232static inline bool dl_entity_is_special(struct sched_dl_entity *dl_se)
233{
234#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
235 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
236#else
237 return false;
238#endif
239}
240
241/*
242 * Tells if entity @a should preempt entity @b.
243 */
244static inline bool
245dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
246{
247 return dl_entity_is_special(a) ||
248 dl_time_before(a->deadline, b->deadline);
249}
250
251/*
252 * This is the priority-queue data structure of the RT scheduling class:
253 */
254struct rt_prio_array {
255 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
256 struct list_head queue[MAX_RT_PRIO];
257};
258
259struct rt_bandwidth {
260 /* nests inside the rq lock: */
261 raw_spinlock_t rt_runtime_lock;
262 ktime_t rt_period;
263 u64 rt_runtime;
264 struct hrtimer rt_period_timer;
265 unsigned int rt_period_active;
266};
267
268void __dl_clear_params(struct task_struct *p);
269
270struct dl_bandwidth {
271 raw_spinlock_t dl_runtime_lock;
272 u64 dl_runtime;
273 u64 dl_period;
274};
275
276static inline int dl_bandwidth_enabled(void)
277{
278 return sysctl_sched_rt_runtime >= 0;
279}
280
281/*
282 * To keep the bandwidth of -deadline tasks under control
283 * we need some place where:
284 * - store the maximum -deadline bandwidth of each cpu;
285 * - cache the fraction of bandwidth that is currently allocated in
286 * each root domain;
287 *
288 * This is all done in the data structure below. It is similar to the
289 * one used for RT-throttling (rt_bandwidth), with the main difference
290 * that, since here we are only interested in admission control, we
291 * do not decrease any runtime while the group "executes", neither we
292 * need a timer to replenish it.
293 *
294 * With respect to SMP, bandwidth is given on a per root domain basis,
295 * meaning that:
296 * - bw (< 100%) is the deadline bandwidth of each CPU;
297 * - total_bw is the currently allocated bandwidth in each root domain;
298 */
299struct dl_bw {
300 raw_spinlock_t lock;
301 u64 bw;
302 u64 total_bw;
303};
304
305static inline void __dl_update(struct dl_bw *dl_b, s64 bw);
306
307static inline
308void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
309{
310 dl_b->total_bw -= tsk_bw;
311 __dl_update(dl_b, (s32)tsk_bw / cpus);
312}
313
314static inline
315void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
316{
317 dl_b->total_bw += tsk_bw;
318 __dl_update(dl_b, -((s32)tsk_bw / cpus));
319}
320
321static inline bool __dl_overflow(struct dl_bw *dl_b, unsigned long cap,
322 u64 old_bw, u64 new_bw)
323{
324 return dl_b->bw != -1 &&
325 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
326}
327
328/*
329 * Verify the fitness of task @p to run on @cpu taking into account the
330 * CPU original capacity and the runtime/deadline ratio of the task.
331 *
332 * The function will return true if the CPU original capacity of the
333 * @cpu scaled by SCHED_CAPACITY_SCALE >= runtime/deadline ratio of the
334 * task and false otherwise.
335 */
336static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
337{
338 unsigned long cap = arch_scale_cpu_capacity(cpu);
339
340 return cap_scale(p->dl.dl_deadline, cap) >= p->dl.dl_runtime;
341}
342
343extern void init_dl_bw(struct dl_bw *dl_b);
344extern int sched_dl_global_validate(void);
345extern void sched_dl_do_global(void);
346extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
347extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
348extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
349extern bool __checkparam_dl(const struct sched_attr *attr);
350extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
351extern int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed);
352extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
353extern bool dl_cpu_busy(unsigned int cpu);
354
355#ifdef CONFIG_CGROUP_SCHED
356
357#include <linux/cgroup.h>
358#include <linux/psi.h>
359
360struct cfs_rq;
361struct rt_rq;
362
363extern struct list_head task_groups;
364
365struct cfs_bandwidth {
366#ifdef CONFIG_CFS_BANDWIDTH
367 raw_spinlock_t lock;
368 ktime_t period;
369 u64 quota;
370 u64 runtime;
371 u64 burst;
372 s64 hierarchical_quota;
373
374 u8 idle;
375 u8 period_active;
376 u8 slack_started;
377 struct hrtimer period_timer;
378 struct hrtimer slack_timer;
379 struct list_head throttled_cfs_rq;
380
381 /* Statistics: */
382 int nr_periods;
383 int nr_throttled;
384 u64 throttled_time;
385#endif
386};
387
388/* Task group related information */
389struct task_group {
390 struct cgroup_subsys_state css;
391
392#ifdef CONFIG_FAIR_GROUP_SCHED
393 /* schedulable entities of this group on each CPU */
394 struct sched_entity **se;
395 /* runqueue "owned" by this group on each CPU */
396 struct cfs_rq **cfs_rq;
397 unsigned long shares;
398
399#ifdef CONFIG_SMP
400 /*
401 * load_avg can be heavily contended at clock tick time, so put
402 * it in its own cacheline separated from the fields above which
403 * will also be accessed at each tick.
404 */
405 atomic_long_t load_avg ____cacheline_aligned;
406#endif
407#endif
408
409#ifdef CONFIG_RT_GROUP_SCHED
410 struct sched_rt_entity **rt_se;
411 struct rt_rq **rt_rq;
412
413 struct rt_bandwidth rt_bandwidth;
414#endif
415
416 struct rcu_head rcu;
417 struct list_head list;
418
419 struct task_group *parent;
420 struct list_head siblings;
421 struct list_head children;
422
423#ifdef CONFIG_SCHED_AUTOGROUP
424 struct autogroup *autogroup;
425#endif
426
427 struct cfs_bandwidth cfs_bandwidth;
428
429#ifdef CONFIG_UCLAMP_TASK_GROUP
430 /* The two decimal precision [%] value requested from user-space */
431 unsigned int uclamp_pct[UCLAMP_CNT];
432 /* Clamp values requested for a task group */
433 struct uclamp_se uclamp_req[UCLAMP_CNT];
434 /* Effective clamp values used for a task group */
435 struct uclamp_se uclamp[UCLAMP_CNT];
436#endif
437
438};
439
440#ifdef CONFIG_FAIR_GROUP_SCHED
441#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
442
443/*
444 * A weight of 0 or 1 can cause arithmetics problems.
445 * A weight of a cfs_rq is the sum of weights of which entities
446 * are queued on this cfs_rq, so a weight of a entity should not be
447 * too large, so as the shares value of a task group.
448 * (The default weight is 1024 - so there's no practical
449 * limitation from this.)
450 */
451#define MIN_SHARES (1UL << 1)
452#define MAX_SHARES (1UL << 18)
453#endif
454
455typedef int (*tg_visitor)(struct task_group *, void *);
456
457extern int walk_tg_tree_from(struct task_group *from,
458 tg_visitor down, tg_visitor up, void *data);
459
460/*
461 * Iterate the full tree, calling @down when first entering a node and @up when
462 * leaving it for the final time.
463 *
464 * Caller must hold rcu_lock or sufficient equivalent.
465 */
466static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
467{
468 return walk_tg_tree_from(&root_task_group, down, up, data);
469}
470
471extern int tg_nop(struct task_group *tg, void *data);
472
473extern void free_fair_sched_group(struct task_group *tg);
474extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
475extern void online_fair_sched_group(struct task_group *tg);
476extern void unregister_fair_sched_group(struct task_group *tg);
477extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
478 struct sched_entity *se, int cpu,
479 struct sched_entity *parent);
480extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
481
482extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
483extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
484extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
485
486extern void free_rt_sched_group(struct task_group *tg);
487extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
488extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
489 struct sched_rt_entity *rt_se, int cpu,
490 struct sched_rt_entity *parent);
491extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
492extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
493extern long sched_group_rt_runtime(struct task_group *tg);
494extern long sched_group_rt_period(struct task_group *tg);
495extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
496
497extern struct task_group *sched_create_group(struct task_group *parent);
498extern void sched_online_group(struct task_group *tg,
499 struct task_group *parent);
500extern void sched_destroy_group(struct task_group *tg);
501extern void sched_offline_group(struct task_group *tg);
502
503extern void sched_move_task(struct task_struct *tsk);
504
505#ifdef CONFIG_FAIR_GROUP_SCHED
506extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
507
508#ifdef CONFIG_SMP
509extern void set_task_rq_fair(struct sched_entity *se,
510 struct cfs_rq *prev, struct cfs_rq *next);
511#else /* !CONFIG_SMP */
512static inline void set_task_rq_fair(struct sched_entity *se,
513 struct cfs_rq *prev, struct cfs_rq *next) { }
514#endif /* CONFIG_SMP */
515#endif /* CONFIG_FAIR_GROUP_SCHED */
516
517#else /* CONFIG_CGROUP_SCHED */
518
519struct cfs_bandwidth { };
520
521#endif /* CONFIG_CGROUP_SCHED */
522
523/* CFS-related fields in a runqueue */
524struct cfs_rq {
525 struct load_weight load;
526 unsigned int nr_running;
527 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
528 unsigned int idle_h_nr_running; /* SCHED_IDLE */
529
530 u64 exec_clock;
531 u64 min_vruntime;
532#ifdef CONFIG_SCHED_CORE
533 unsigned int forceidle_seq;
534 u64 min_vruntime_fi;
535#endif
536
537#ifndef CONFIG_64BIT
538 u64 min_vruntime_copy;
539#endif
540
541 struct rb_root_cached tasks_timeline;
542
543 /*
544 * 'curr' points to currently running entity on this cfs_rq.
545 * It is set to NULL otherwise (i.e when none are currently running).
546 */
547 struct sched_entity *curr;
548 struct sched_entity *next;
549 struct sched_entity *last;
550 struct sched_entity *skip;
551
552#ifdef CONFIG_SCHED_DEBUG
553 unsigned int nr_spread_over;
554#endif
555
556#ifdef CONFIG_SMP
557 /*
558 * CFS load tracking
559 */
560 struct sched_avg avg;
561#ifndef CONFIG_64BIT
562 u64 load_last_update_time_copy;
563#endif
564 struct {
565 raw_spinlock_t lock ____cacheline_aligned;
566 int nr;
567 unsigned long load_avg;
568 unsigned long util_avg;
569 unsigned long runnable_avg;
570 } removed;
571
572#ifdef CONFIG_FAIR_GROUP_SCHED
573 unsigned long tg_load_avg_contrib;
574 long propagate;
575 long prop_runnable_sum;
576
577 /*
578 * h_load = weight * f(tg)
579 *
580 * Where f(tg) is the recursive weight fraction assigned to
581 * this group.
582 */
583 unsigned long h_load;
584 u64 last_h_load_update;
585 struct sched_entity *h_load_next;
586#endif /* CONFIG_FAIR_GROUP_SCHED */
587#endif /* CONFIG_SMP */
588
589#ifdef CONFIG_FAIR_GROUP_SCHED
590 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
591
592 /*
593 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
594 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
595 * (like users, containers etc.)
596 *
597 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
598 * This list is used during load balance.
599 */
600 int on_list;
601 struct list_head leaf_cfs_rq_list;
602 struct task_group *tg; /* group that "owns" this runqueue */
603
604#ifdef CONFIG_CFS_BANDWIDTH
605 int runtime_enabled;
606 s64 runtime_remaining;
607
608 u64 throttled_clock;
609 u64 throttled_clock_task;
610 u64 throttled_clock_task_time;
611 int throttled;
612 int throttle_count;
613 struct list_head throttled_list;
614#endif /* CONFIG_CFS_BANDWIDTH */
615#endif /* CONFIG_FAIR_GROUP_SCHED */
616};
617
618static inline int rt_bandwidth_enabled(void)
619{
620 return sysctl_sched_rt_runtime >= 0;
621}
622
623/* RT IPI pull logic requires IRQ_WORK */
624#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
625# define HAVE_RT_PUSH_IPI
626#endif
627
628/* Real-Time classes' related field in a runqueue: */
629struct rt_rq {
630 struct rt_prio_array active;
631 unsigned int rt_nr_running;
632 unsigned int rr_nr_running;
633#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
634 struct {
635 int curr; /* highest queued rt task prio */
636#ifdef CONFIG_SMP
637 int next; /* next highest */
638#endif
639 } highest_prio;
640#endif
641#ifdef CONFIG_SMP
642 unsigned int rt_nr_migratory;
643 unsigned int rt_nr_total;
644 int overloaded;
645 struct plist_head pushable_tasks;
646
647#endif /* CONFIG_SMP */
648 int rt_queued;
649
650 int rt_throttled;
651 u64 rt_time;
652 u64 rt_runtime;
653 /* Nests inside the rq lock: */
654 raw_spinlock_t rt_runtime_lock;
655
656#ifdef CONFIG_RT_GROUP_SCHED
657 unsigned int rt_nr_boosted;
658
659 struct rq *rq;
660 struct task_group *tg;
661#endif
662};
663
664static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
665{
666 return rt_rq->rt_queued && rt_rq->rt_nr_running;
667}
668
669/* Deadline class' related fields in a runqueue */
670struct dl_rq {
671 /* runqueue is an rbtree, ordered by deadline */
672 struct rb_root_cached root;
673
674 unsigned int dl_nr_running;
675
676#ifdef CONFIG_SMP
677 /*
678 * Deadline values of the currently executing and the
679 * earliest ready task on this rq. Caching these facilitates
680 * the decision whether or not a ready but not running task
681 * should migrate somewhere else.
682 */
683 struct {
684 u64 curr;
685 u64 next;
686 } earliest_dl;
687
688 unsigned int dl_nr_migratory;
689 int overloaded;
690
691 /*
692 * Tasks on this rq that can be pushed away. They are kept in
693 * an rb-tree, ordered by tasks' deadlines, with caching
694 * of the leftmost (earliest deadline) element.
695 */
696 struct rb_root_cached pushable_dl_tasks_root;
697#else
698 struct dl_bw dl_bw;
699#endif
700 /*
701 * "Active utilization" for this runqueue: increased when a
702 * task wakes up (becomes TASK_RUNNING) and decreased when a
703 * task blocks
704 */
705 u64 running_bw;
706
707 /*
708 * Utilization of the tasks "assigned" to this runqueue (including
709 * the tasks that are in runqueue and the tasks that executed on this
710 * CPU and blocked). Increased when a task moves to this runqueue, and
711 * decreased when the task moves away (migrates, changes scheduling
712 * policy, or terminates).
713 * This is needed to compute the "inactive utilization" for the
714 * runqueue (inactive utilization = this_bw - running_bw).
715 */
716 u64 this_bw;
717 u64 extra_bw;
718
719 /*
720 * Inverse of the fraction of CPU utilization that can be reclaimed
721 * by the GRUB algorithm.
722 */
723 u64 bw_ratio;
724};
725
726#ifdef CONFIG_FAIR_GROUP_SCHED
727/* An entity is a task if it doesn't "own" a runqueue */
728#define entity_is_task(se) (!se->my_q)
729
730static inline void se_update_runnable(struct sched_entity *se)
731{
732 if (!entity_is_task(se))
733 se->runnable_weight = se->my_q->h_nr_running;
734}
735
736static inline long se_runnable(struct sched_entity *se)
737{
738 if (entity_is_task(se))
739 return !!se->on_rq;
740 else
741 return se->runnable_weight;
742}
743
744#else
745#define entity_is_task(se) 1
746
747static inline void se_update_runnable(struct sched_entity *se) {}
748
749static inline long se_runnable(struct sched_entity *se)
750{
751 return !!se->on_rq;
752}
753#endif
754
755#ifdef CONFIG_SMP
756/*
757 * XXX we want to get rid of these helpers and use the full load resolution.
758 */
759static inline long se_weight(struct sched_entity *se)
760{
761 return scale_load_down(se->load.weight);
762}
763
764
765static inline bool sched_asym_prefer(int a, int b)
766{
767 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
768}
769
770struct perf_domain {
771 struct em_perf_domain *em_pd;
772 struct perf_domain *next;
773 struct rcu_head rcu;
774};
775
776/* Scheduling group status flags */
777#define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
778#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
779
780/*
781 * We add the notion of a root-domain which will be used to define per-domain
782 * variables. Each exclusive cpuset essentially defines an island domain by
783 * fully partitioning the member CPUs from any other cpuset. Whenever a new
784 * exclusive cpuset is created, we also create and attach a new root-domain
785 * object.
786 *
787 */
788struct root_domain {
789 atomic_t refcount;
790 atomic_t rto_count;
791 struct rcu_head rcu;
792 cpumask_var_t span;
793 cpumask_var_t online;
794
795 /*
796 * Indicate pullable load on at least one CPU, e.g:
797 * - More than one runnable task
798 * - Running task is misfit
799 */
800 int overload;
801
802 /* Indicate one or more cpus over-utilized (tipping point) */
803 int overutilized;
804
805 /*
806 * The bit corresponding to a CPU gets set here if such CPU has more
807 * than one runnable -deadline task (as it is below for RT tasks).
808 */
809 cpumask_var_t dlo_mask;
810 atomic_t dlo_count;
811 struct dl_bw dl_bw;
812 struct cpudl cpudl;
813
814 /*
815 * Indicate whether a root_domain's dl_bw has been checked or
816 * updated. It's monotonously increasing value.
817 *
818 * Also, some corner cases, like 'wrap around' is dangerous, but given
819 * that u64 is 'big enough'. So that shouldn't be a concern.
820 */
821 u64 visit_gen;
822
823#ifdef HAVE_RT_PUSH_IPI
824 /*
825 * For IPI pull requests, loop across the rto_mask.
826 */
827 struct irq_work rto_push_work;
828 raw_spinlock_t rto_lock;
829 /* These are only updated and read within rto_lock */
830 int rto_loop;
831 int rto_cpu;
832 /* These atomics are updated outside of a lock */
833 atomic_t rto_loop_next;
834 atomic_t rto_loop_start;
835#endif
836 /*
837 * The "RT overload" flag: it gets set if a CPU has more than
838 * one runnable RT task.
839 */
840 cpumask_var_t rto_mask;
841 struct cpupri cpupri;
842
843 unsigned long max_cpu_capacity;
844
845 /*
846 * NULL-terminated list of performance domains intersecting with the
847 * CPUs of the rd. Protected by RCU.
848 */
849 struct perf_domain __rcu *pd;
850};
851
852extern void init_defrootdomain(void);
853extern int sched_init_domains(const struct cpumask *cpu_map);
854extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
855extern void sched_get_rd(struct root_domain *rd);
856extern void sched_put_rd(struct root_domain *rd);
857
858#ifdef HAVE_RT_PUSH_IPI
859extern void rto_push_irq_work_func(struct irq_work *work);
860#endif
861#endif /* CONFIG_SMP */
862
863#ifdef CONFIG_UCLAMP_TASK
864/*
865 * struct uclamp_bucket - Utilization clamp bucket
866 * @value: utilization clamp value for tasks on this clamp bucket
867 * @tasks: number of RUNNABLE tasks on this clamp bucket
868 *
869 * Keep track of how many tasks are RUNNABLE for a given utilization
870 * clamp value.
871 */
872struct uclamp_bucket {
873 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
874 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
875};
876
877/*
878 * struct uclamp_rq - rq's utilization clamp
879 * @value: currently active clamp values for a rq
880 * @bucket: utilization clamp buckets affecting a rq
881 *
882 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
883 * A clamp value is affecting a rq when there is at least one task RUNNABLE
884 * (or actually running) with that value.
885 *
886 * There are up to UCLAMP_CNT possible different clamp values, currently there
887 * are only two: minimum utilization and maximum utilization.
888 *
889 * All utilization clamping values are MAX aggregated, since:
890 * - for util_min: we want to run the CPU at least at the max of the minimum
891 * utilization required by its currently RUNNABLE tasks.
892 * - for util_max: we want to allow the CPU to run up to the max of the
893 * maximum utilization allowed by its currently RUNNABLE tasks.
894 *
895 * Since on each system we expect only a limited number of different
896 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
897 * the metrics required to compute all the per-rq utilization clamp values.
898 */
899struct uclamp_rq {
900 unsigned int value;
901 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
902};
903
904DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
905#endif /* CONFIG_UCLAMP_TASK */
906
907/*
908 * This is the main, per-CPU runqueue data structure.
909 *
910 * Locking rule: those places that want to lock multiple runqueues
911 * (such as the load balancing or the thread migration code), lock
912 * acquire operations must be ordered by ascending &runqueue.
913 */
914struct rq {
915 /* runqueue lock: */
916 raw_spinlock_t __lock;
917
918 /*
919 * nr_running and cpu_load should be in the same cacheline because
920 * remote CPUs use both these fields when doing load calculation.
921 */
922 unsigned int nr_running;
923#ifdef CONFIG_NUMA_BALANCING
924 unsigned int nr_numa_running;
925 unsigned int nr_preferred_running;
926 unsigned int numa_migrate_on;
927#endif
928#ifdef CONFIG_NO_HZ_COMMON
929#ifdef CONFIG_SMP
930 unsigned long last_blocked_load_update_tick;
931 unsigned int has_blocked_load;
932 call_single_data_t nohz_csd;
933#endif /* CONFIG_SMP */
934 unsigned int nohz_tick_stopped;
935 atomic_t nohz_flags;
936#endif /* CONFIG_NO_HZ_COMMON */
937
938#ifdef CONFIG_SMP
939 unsigned int ttwu_pending;
940#endif
941 u64 nr_switches;
942
943#ifdef CONFIG_UCLAMP_TASK
944 /* Utilization clamp values based on CPU's RUNNABLE tasks */
945 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
946 unsigned int uclamp_flags;
947#define UCLAMP_FLAG_IDLE 0x01
948#endif
949
950 struct cfs_rq cfs;
951 struct rt_rq rt;
952 struct dl_rq dl;
953
954#ifdef CONFIG_FAIR_GROUP_SCHED
955 /* list of leaf cfs_rq on this CPU: */
956 struct list_head leaf_cfs_rq_list;
957 struct list_head *tmp_alone_branch;
958#endif /* CONFIG_FAIR_GROUP_SCHED */
959
960 /*
961 * This is part of a global counter where only the total sum
962 * over all CPUs matters. A task can increase this counter on
963 * one CPU and if it got migrated afterwards it may decrease
964 * it on another CPU. Always updated under the runqueue lock:
965 */
966 unsigned int nr_uninterruptible;
967
968 struct task_struct __rcu *curr;
969 struct task_struct *idle;
970 struct task_struct *stop;
971 unsigned long next_balance;
972 struct mm_struct *prev_mm;
973
974 unsigned int clock_update_flags;
975 u64 clock;
976 /* Ensure that all clocks are in the same cache line */
977 u64 clock_task ____cacheline_aligned;
978 u64 clock_pelt;
979 unsigned long lost_idle_time;
980
981 atomic_t nr_iowait;
982
983#ifdef CONFIG_SCHED_DEBUG
984 u64 last_seen_need_resched_ns;
985 int ticks_without_resched;
986#endif
987
988#ifdef CONFIG_MEMBARRIER
989 int membarrier_state;
990#endif
991
992#ifdef CONFIG_SMP
993 struct root_domain *rd;
994 struct sched_domain __rcu *sd;
995
996 unsigned long cpu_capacity;
997 unsigned long cpu_capacity_orig;
998
999 struct callback_head *balance_callback;
1000
1001 unsigned char nohz_idle_balance;
1002 unsigned char idle_balance;
1003
1004 unsigned long misfit_task_load;
1005
1006 /* For active balancing */
1007 int active_balance;
1008 int push_cpu;
1009 struct cpu_stop_work active_balance_work;
1010
1011 /* CPU of this runqueue: */
1012 int cpu;
1013 int online;
1014
1015 struct list_head cfs_tasks;
1016
1017 struct sched_avg avg_rt;
1018 struct sched_avg avg_dl;
1019#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1020 struct sched_avg avg_irq;
1021#endif
1022#ifdef CONFIG_SCHED_THERMAL_PRESSURE
1023 struct sched_avg avg_thermal;
1024#endif
1025 u64 idle_stamp;
1026 u64 avg_idle;
1027
1028 unsigned long wake_stamp;
1029 u64 wake_avg_idle;
1030
1031 /* This is used to determine avg_idle's max value */
1032 u64 max_idle_balance_cost;
1033
1034#ifdef CONFIG_HOTPLUG_CPU
1035 struct rcuwait hotplug_wait;
1036#endif
1037#endif /* CONFIG_SMP */
1038
1039#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1040 u64 prev_irq_time;
1041#endif
1042#ifdef CONFIG_PARAVIRT
1043 u64 prev_steal_time;
1044#endif
1045#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1046 u64 prev_steal_time_rq;
1047#endif
1048
1049 /* calc_load related fields */
1050 unsigned long calc_load_update;
1051 long calc_load_active;
1052
1053#ifdef CONFIG_SCHED_HRTICK
1054#ifdef CONFIG_SMP
1055 call_single_data_t hrtick_csd;
1056#endif
1057 struct hrtimer hrtick_timer;
1058 ktime_t hrtick_time;
1059#endif
1060
1061#ifdef CONFIG_SCHEDSTATS
1062 /* latency stats */
1063 struct sched_info rq_sched_info;
1064 unsigned long long rq_cpu_time;
1065 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1066
1067 /* sys_sched_yield() stats */
1068 unsigned int yld_count;
1069
1070 /* schedule() stats */
1071 unsigned int sched_count;
1072 unsigned int sched_goidle;
1073
1074 /* try_to_wake_up() stats */
1075 unsigned int ttwu_count;
1076 unsigned int ttwu_local;
1077#endif
1078
1079#ifdef CONFIG_CPU_IDLE
1080 /* Must be inspected within a rcu lock section */
1081 struct cpuidle_state *idle_state;
1082#endif
1083
1084#ifdef CONFIG_SMP
1085 unsigned int nr_pinned;
1086#endif
1087 unsigned int push_busy;
1088 struct cpu_stop_work push_work;
1089
1090#ifdef CONFIG_SCHED_CORE
1091 /* per rq */
1092 struct rq *core;
1093 struct task_struct *core_pick;
1094 unsigned int core_enabled;
1095 unsigned int core_sched_seq;
1096 struct rb_root core_tree;
1097
1098 /* shared state -- careful with sched_core_cpu_deactivate() */
1099 unsigned int core_task_seq;
1100 unsigned int core_pick_seq;
1101 unsigned long core_cookie;
1102 unsigned char core_forceidle;
1103 unsigned int core_forceidle_seq;
1104#endif
1105};
1106
1107#ifdef CONFIG_FAIR_GROUP_SCHED
1108
1109/* CPU runqueue to which this cfs_rq is attached */
1110static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1111{
1112 return cfs_rq->rq;
1113}
1114
1115#else
1116
1117static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1118{
1119 return container_of(cfs_rq, struct rq, cfs);
1120}
1121#endif
1122
1123static inline int cpu_of(struct rq *rq)
1124{
1125#ifdef CONFIG_SMP
1126 return rq->cpu;
1127#else
1128 return 0;
1129#endif
1130}
1131
1132#define MDF_PUSH 0x01
1133
1134static inline bool is_migration_disabled(struct task_struct *p)
1135{
1136#ifdef CONFIG_SMP
1137 return p->migration_disabled;
1138#else
1139 return false;
1140#endif
1141}
1142
1143struct sched_group;
1144#ifdef CONFIG_SCHED_CORE
1145static inline struct cpumask *sched_group_span(struct sched_group *sg);
1146
1147DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1148
1149static inline bool sched_core_enabled(struct rq *rq)
1150{
1151 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1152}
1153
1154static inline bool sched_core_disabled(void)
1155{
1156 return !static_branch_unlikely(&__sched_core_enabled);
1157}
1158
1159/*
1160 * Be careful with this function; not for general use. The return value isn't
1161 * stable unless you actually hold a relevant rq->__lock.
1162 */
1163static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1164{
1165 if (sched_core_enabled(rq))
1166 return &rq->core->__lock;
1167
1168 return &rq->__lock;
1169}
1170
1171static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1172{
1173 if (rq->core_enabled)
1174 return &rq->core->__lock;
1175
1176 return &rq->__lock;
1177}
1178
1179bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool fi);
1180
1181/*
1182 * Helpers to check if the CPU's core cookie matches with the task's cookie
1183 * when core scheduling is enabled.
1184 * A special case is that the task's cookie always matches with CPU's core
1185 * cookie if the CPU is in an idle core.
1186 */
1187static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1188{
1189 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1190 if (!sched_core_enabled(rq))
1191 return true;
1192
1193 return rq->core->core_cookie == p->core_cookie;
1194}
1195
1196static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1197{
1198 bool idle_core = true;
1199 int cpu;
1200
1201 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1202 if (!sched_core_enabled(rq))
1203 return true;
1204
1205 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1206 if (!available_idle_cpu(cpu)) {
1207 idle_core = false;
1208 break;
1209 }
1210 }
1211
1212 /*
1213 * A CPU in an idle core is always the best choice for tasks with
1214 * cookies.
1215 */
1216 return idle_core || rq->core->core_cookie == p->core_cookie;
1217}
1218
1219static inline bool sched_group_cookie_match(struct rq *rq,
1220 struct task_struct *p,
1221 struct sched_group *group)
1222{
1223 int cpu;
1224
1225 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1226 if (!sched_core_enabled(rq))
1227 return true;
1228
1229 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1230 if (sched_core_cookie_match(rq, p))
1231 return true;
1232 }
1233 return false;
1234}
1235
1236extern void queue_core_balance(struct rq *rq);
1237
1238static inline bool sched_core_enqueued(struct task_struct *p)
1239{
1240 return !RB_EMPTY_NODE(&p->core_node);
1241}
1242
1243extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1244extern void sched_core_dequeue(struct rq *rq, struct task_struct *p);
1245
1246extern void sched_core_get(void);
1247extern void sched_core_put(void);
1248
1249extern unsigned long sched_core_alloc_cookie(void);
1250extern void sched_core_put_cookie(unsigned long cookie);
1251extern unsigned long sched_core_get_cookie(unsigned long cookie);
1252extern unsigned long sched_core_update_cookie(struct task_struct *p, unsigned long cookie);
1253
1254#else /* !CONFIG_SCHED_CORE */
1255
1256static inline bool sched_core_enabled(struct rq *rq)
1257{
1258 return false;
1259}
1260
1261static inline bool sched_core_disabled(void)
1262{
1263 return true;
1264}
1265
1266static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1267{
1268 return &rq->__lock;
1269}
1270
1271static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1272{
1273 return &rq->__lock;
1274}
1275
1276static inline void queue_core_balance(struct rq *rq)
1277{
1278}
1279
1280static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1281{
1282 return true;
1283}
1284
1285static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1286{
1287 return true;
1288}
1289
1290static inline bool sched_group_cookie_match(struct rq *rq,
1291 struct task_struct *p,
1292 struct sched_group *group)
1293{
1294 return true;
1295}
1296#endif /* CONFIG_SCHED_CORE */
1297
1298static inline void lockdep_assert_rq_held(struct rq *rq)
1299{
1300 lockdep_assert_held(__rq_lockp(rq));
1301}
1302
1303extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1304extern bool raw_spin_rq_trylock(struct rq *rq);
1305extern void raw_spin_rq_unlock(struct rq *rq);
1306
1307static inline void raw_spin_rq_lock(struct rq *rq)
1308{
1309 raw_spin_rq_lock_nested(rq, 0);
1310}
1311
1312static inline void raw_spin_rq_lock_irq(struct rq *rq)
1313{
1314 local_irq_disable();
1315 raw_spin_rq_lock(rq);
1316}
1317
1318static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1319{
1320 raw_spin_rq_unlock(rq);
1321 local_irq_enable();
1322}
1323
1324static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1325{
1326 unsigned long flags;
1327 local_irq_save(flags);
1328 raw_spin_rq_lock(rq);
1329 return flags;
1330}
1331
1332static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1333{
1334 raw_spin_rq_unlock(rq);
1335 local_irq_restore(flags);
1336}
1337
1338#define raw_spin_rq_lock_irqsave(rq, flags) \
1339do { \
1340 flags = _raw_spin_rq_lock_irqsave(rq); \
1341} while (0)
1342
1343#ifdef CONFIG_SCHED_SMT
1344extern void __update_idle_core(struct rq *rq);
1345
1346static inline void update_idle_core(struct rq *rq)
1347{
1348 if (static_branch_unlikely(&sched_smt_present))
1349 __update_idle_core(rq);
1350}
1351
1352#else
1353static inline void update_idle_core(struct rq *rq) { }
1354#endif
1355
1356DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1357
1358#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1359#define this_rq() this_cpu_ptr(&runqueues)
1360#define task_rq(p) cpu_rq(task_cpu(p))
1361#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1362#define raw_rq() raw_cpu_ptr(&runqueues)
1363
1364#ifdef CONFIG_FAIR_GROUP_SCHED
1365static inline struct task_struct *task_of(struct sched_entity *se)
1366{
1367 SCHED_WARN_ON(!entity_is_task(se));
1368 return container_of(se, struct task_struct, se);
1369}
1370
1371static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1372{
1373 return p->se.cfs_rq;
1374}
1375
1376/* runqueue on which this entity is (to be) queued */
1377static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
1378{
1379 return se->cfs_rq;
1380}
1381
1382/* runqueue "owned" by this group */
1383static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1384{
1385 return grp->my_q;
1386}
1387
1388#else
1389
1390static inline struct task_struct *task_of(struct sched_entity *se)
1391{
1392 return container_of(se, struct task_struct, se);
1393}
1394
1395static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1396{
1397 return &task_rq(p)->cfs;
1398}
1399
1400static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
1401{
1402 struct task_struct *p = task_of(se);
1403 struct rq *rq = task_rq(p);
1404
1405 return &rq->cfs;
1406}
1407
1408/* runqueue "owned" by this group */
1409static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1410{
1411 return NULL;
1412}
1413#endif
1414
1415extern void update_rq_clock(struct rq *rq);
1416
1417static inline u64 __rq_clock_broken(struct rq *rq)
1418{
1419 return READ_ONCE(rq->clock);
1420}
1421
1422/*
1423 * rq::clock_update_flags bits
1424 *
1425 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1426 * call to __schedule(). This is an optimisation to avoid
1427 * neighbouring rq clock updates.
1428 *
1429 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1430 * in effect and calls to update_rq_clock() are being ignored.
1431 *
1432 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1433 * made to update_rq_clock() since the last time rq::lock was pinned.
1434 *
1435 * If inside of __schedule(), clock_update_flags will have been
1436 * shifted left (a left shift is a cheap operation for the fast path
1437 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1438 *
1439 * if (rq-clock_update_flags >= RQCF_UPDATED)
1440 *
1441 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1442 * one position though, because the next rq_unpin_lock() will shift it
1443 * back.
1444 */
1445#define RQCF_REQ_SKIP 0x01
1446#define RQCF_ACT_SKIP 0x02
1447#define RQCF_UPDATED 0x04
1448
1449static inline void assert_clock_updated(struct rq *rq)
1450{
1451 /*
1452 * The only reason for not seeing a clock update since the
1453 * last rq_pin_lock() is if we're currently skipping updates.
1454 */
1455 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1456}
1457
1458static inline u64 rq_clock(struct rq *rq)
1459{
1460 lockdep_assert_rq_held(rq);
1461 assert_clock_updated(rq);
1462
1463 return rq->clock;
1464}
1465
1466static inline u64 rq_clock_task(struct rq *rq)
1467{
1468 lockdep_assert_rq_held(rq);
1469 assert_clock_updated(rq);
1470
1471 return rq->clock_task;
1472}
1473
1474/**
1475 * By default the decay is the default pelt decay period.
1476 * The decay shift can change the decay period in
1477 * multiples of 32.
1478 * Decay shift Decay period(ms)
1479 * 0 32
1480 * 1 64
1481 * 2 128
1482 * 3 256
1483 * 4 512
1484 */
1485extern int sched_thermal_decay_shift;
1486
1487static inline u64 rq_clock_thermal(struct rq *rq)
1488{
1489 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1490}
1491
1492static inline void rq_clock_skip_update(struct rq *rq)
1493{
1494 lockdep_assert_rq_held(rq);
1495 rq->clock_update_flags |= RQCF_REQ_SKIP;
1496}
1497
1498/*
1499 * See rt task throttling, which is the only time a skip
1500 * request is canceled.
1501 */
1502static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1503{
1504 lockdep_assert_rq_held(rq);
1505 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1506}
1507
1508struct rq_flags {
1509 unsigned long flags;
1510 struct pin_cookie cookie;
1511#ifdef CONFIG_SCHED_DEBUG
1512 /*
1513 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1514 * current pin context is stashed here in case it needs to be
1515 * restored in rq_repin_lock().
1516 */
1517 unsigned int clock_update_flags;
1518#endif
1519};
1520
1521extern struct callback_head balance_push_callback;
1522
1523/*
1524 * Lockdep annotation that avoids accidental unlocks; it's like a
1525 * sticky/continuous lockdep_assert_held().
1526 *
1527 * This avoids code that has access to 'struct rq *rq' (basically everything in
1528 * the scheduler) from accidentally unlocking the rq if they do not also have a
1529 * copy of the (on-stack) 'struct rq_flags rf'.
1530 *
1531 * Also see Documentation/locking/lockdep-design.rst.
1532 */
1533static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1534{
1535 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1536
1537#ifdef CONFIG_SCHED_DEBUG
1538 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1539 rf->clock_update_flags = 0;
1540#ifdef CONFIG_SMP
1541 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1542#endif
1543#endif
1544}
1545
1546static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1547{
1548#ifdef CONFIG_SCHED_DEBUG
1549 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1550 rf->clock_update_flags = RQCF_UPDATED;
1551#endif
1552
1553 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1554}
1555
1556static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1557{
1558 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1559
1560#ifdef CONFIG_SCHED_DEBUG
1561 /*
1562 * Restore the value we stashed in @rf for this pin context.
1563 */
1564 rq->clock_update_flags |= rf->clock_update_flags;
1565#endif
1566}
1567
1568struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1569 __acquires(rq->lock);
1570
1571struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1572 __acquires(p->pi_lock)
1573 __acquires(rq->lock);
1574
1575static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1576 __releases(rq->lock)
1577{
1578 rq_unpin_lock(rq, rf);
1579 raw_spin_rq_unlock(rq);
1580}
1581
1582static inline void
1583task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1584 __releases(rq->lock)
1585 __releases(p->pi_lock)
1586{
1587 rq_unpin_lock(rq, rf);
1588 raw_spin_rq_unlock(rq);
1589 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1590}
1591
1592static inline void
1593rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1594 __acquires(rq->lock)
1595{
1596 raw_spin_rq_lock_irqsave(rq, rf->flags);
1597 rq_pin_lock(rq, rf);
1598}
1599
1600static inline void
1601rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1602 __acquires(rq->lock)
1603{
1604 raw_spin_rq_lock_irq(rq);
1605 rq_pin_lock(rq, rf);
1606}
1607
1608static inline void
1609rq_lock(struct rq *rq, struct rq_flags *rf)
1610 __acquires(rq->lock)
1611{
1612 raw_spin_rq_lock(rq);
1613 rq_pin_lock(rq, rf);
1614}
1615
1616static inline void
1617rq_relock(struct rq *rq, struct rq_flags *rf)
1618 __acquires(rq->lock)
1619{
1620 raw_spin_rq_lock(rq);
1621 rq_repin_lock(rq, rf);
1622}
1623
1624static inline void
1625rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1626 __releases(rq->lock)
1627{
1628 rq_unpin_lock(rq, rf);
1629 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1630}
1631
1632static inline void
1633rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1634 __releases(rq->lock)
1635{
1636 rq_unpin_lock(rq, rf);
1637 raw_spin_rq_unlock_irq(rq);
1638}
1639
1640static inline void
1641rq_unlock(struct rq *rq, struct rq_flags *rf)
1642 __releases(rq->lock)
1643{
1644 rq_unpin_lock(rq, rf);
1645 raw_spin_rq_unlock(rq);
1646}
1647
1648static inline struct rq *
1649this_rq_lock_irq(struct rq_flags *rf)
1650 __acquires(rq->lock)
1651{
1652 struct rq *rq;
1653
1654 local_irq_disable();
1655 rq = this_rq();
1656 rq_lock(rq, rf);
1657 return rq;
1658}
1659
1660#ifdef CONFIG_NUMA
1661enum numa_topology_type {
1662 NUMA_DIRECT,
1663 NUMA_GLUELESS_MESH,
1664 NUMA_BACKPLANE,
1665};
1666extern enum numa_topology_type sched_numa_topology_type;
1667extern int sched_max_numa_distance;
1668extern bool find_numa_distance(int distance);
1669extern void sched_init_numa(void);
1670extern void sched_domains_numa_masks_set(unsigned int cpu);
1671extern void sched_domains_numa_masks_clear(unsigned int cpu);
1672extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1673#else
1674static inline void sched_init_numa(void) { }
1675static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1676static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1677static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1678{
1679 return nr_cpu_ids;
1680}
1681#endif
1682
1683#ifdef CONFIG_NUMA_BALANCING
1684/* The regions in numa_faults array from task_struct */
1685enum numa_faults_stats {
1686 NUMA_MEM = 0,
1687 NUMA_CPU,
1688 NUMA_MEMBUF,
1689 NUMA_CPUBUF
1690};
1691extern void sched_setnuma(struct task_struct *p, int node);
1692extern int migrate_task_to(struct task_struct *p, int cpu);
1693extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1694 int cpu, int scpu);
1695extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1696#else
1697static inline void
1698init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1699{
1700}
1701#endif /* CONFIG_NUMA_BALANCING */
1702
1703#ifdef CONFIG_SMP
1704
1705static inline void
1706queue_balance_callback(struct rq *rq,
1707 struct callback_head *head,
1708 void (*func)(struct rq *rq))
1709{
1710 lockdep_assert_rq_held(rq);
1711
1712 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1713 return;
1714
1715 head->func = (void (*)(struct callback_head *))func;
1716 head->next = rq->balance_callback;
1717 rq->balance_callback = head;
1718}
1719
1720#define rcu_dereference_check_sched_domain(p) \
1721 rcu_dereference_check((p), \
1722 lockdep_is_held(&sched_domains_mutex))
1723
1724/*
1725 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1726 * See destroy_sched_domains: call_rcu for details.
1727 *
1728 * The domain tree of any CPU may only be accessed from within
1729 * preempt-disabled sections.
1730 */
1731#define for_each_domain(cpu, __sd) \
1732 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1733 __sd; __sd = __sd->parent)
1734
1735/**
1736 * highest_flag_domain - Return highest sched_domain containing flag.
1737 * @cpu: The CPU whose highest level of sched domain is to
1738 * be returned.
1739 * @flag: The flag to check for the highest sched_domain
1740 * for the given CPU.
1741 *
1742 * Returns the highest sched_domain of a CPU which contains the given flag.
1743 */
1744static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1745{
1746 struct sched_domain *sd, *hsd = NULL;
1747
1748 for_each_domain(cpu, sd) {
1749 if (!(sd->flags & flag))
1750 break;
1751 hsd = sd;
1752 }
1753
1754 return hsd;
1755}
1756
1757static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1758{
1759 struct sched_domain *sd;
1760
1761 for_each_domain(cpu, sd) {
1762 if (sd->flags & flag)
1763 break;
1764 }
1765
1766 return sd;
1767}
1768
1769DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1770DECLARE_PER_CPU(int, sd_llc_size);
1771DECLARE_PER_CPU(int, sd_llc_id);
1772DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1773DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1774DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1775DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1776extern struct static_key_false sched_asym_cpucapacity;
1777
1778struct sched_group_capacity {
1779 atomic_t ref;
1780 /*
1781 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1782 * for a single CPU.
1783 */
1784 unsigned long capacity;
1785 unsigned long min_capacity; /* Min per-CPU capacity in group */
1786 unsigned long max_capacity; /* Max per-CPU capacity in group */
1787 unsigned long next_update;
1788 int imbalance; /* XXX unrelated to capacity but shared group state */
1789
1790#ifdef CONFIG_SCHED_DEBUG
1791 int id;
1792#endif
1793
1794 unsigned long cpumask[]; /* Balance mask */
1795};
1796
1797struct sched_group {
1798 struct sched_group *next; /* Must be a circular list */
1799 atomic_t ref;
1800
1801 unsigned int group_weight;
1802 struct sched_group_capacity *sgc;
1803 int asym_prefer_cpu; /* CPU of highest priority in group */
1804
1805 /*
1806 * The CPUs this group covers.
1807 *
1808 * NOTE: this field is variable length. (Allocated dynamically
1809 * by attaching extra space to the end of the structure,
1810 * depending on how many CPUs the kernel has booted up with)
1811 */
1812 unsigned long cpumask[];
1813};
1814
1815static inline struct cpumask *sched_group_span(struct sched_group *sg)
1816{
1817 return to_cpumask(sg->cpumask);
1818}
1819
1820/*
1821 * See build_balance_mask().
1822 */
1823static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1824{
1825 return to_cpumask(sg->sgc->cpumask);
1826}
1827
1828/**
1829 * group_first_cpu - Returns the first CPU in the cpumask of a sched_group.
1830 * @group: The group whose first CPU is to be returned.
1831 */
1832static inline unsigned int group_first_cpu(struct sched_group *group)
1833{
1834 return cpumask_first(sched_group_span(group));
1835}
1836
1837extern int group_balance_cpu(struct sched_group *sg);
1838
1839#ifdef CONFIG_SCHED_DEBUG
1840void update_sched_domain_debugfs(void);
1841void dirty_sched_domain_sysctl(int cpu);
1842#else
1843static inline void update_sched_domain_debugfs(void)
1844{
1845}
1846static inline void dirty_sched_domain_sysctl(int cpu)
1847{
1848}
1849#endif
1850
1851extern int sched_update_scaling(void);
1852
1853extern void flush_smp_call_function_from_idle(void);
1854
1855#else /* !CONFIG_SMP: */
1856static inline void flush_smp_call_function_from_idle(void) { }
1857#endif
1858
1859#include "stats.h"
1860#include "autogroup.h"
1861
1862#ifdef CONFIG_CGROUP_SCHED
1863
1864/*
1865 * Return the group to which this tasks belongs.
1866 *
1867 * We cannot use task_css() and friends because the cgroup subsystem
1868 * changes that value before the cgroup_subsys::attach() method is called,
1869 * therefore we cannot pin it and might observe the wrong value.
1870 *
1871 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
1872 * core changes this before calling sched_move_task().
1873 *
1874 * Instead we use a 'copy' which is updated from sched_move_task() while
1875 * holding both task_struct::pi_lock and rq::lock.
1876 */
1877static inline struct task_group *task_group(struct task_struct *p)
1878{
1879 return p->sched_task_group;
1880}
1881
1882/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
1883static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
1884{
1885#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
1886 struct task_group *tg = task_group(p);
1887#endif
1888
1889#ifdef CONFIG_FAIR_GROUP_SCHED
1890 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
1891 p->se.cfs_rq = tg->cfs_rq[cpu];
1892 p->se.parent = tg->se[cpu];
1893#endif
1894
1895#ifdef CONFIG_RT_GROUP_SCHED
1896 p->rt.rt_rq = tg->rt_rq[cpu];
1897 p->rt.parent = tg->rt_se[cpu];
1898#endif
1899}
1900
1901#else /* CONFIG_CGROUP_SCHED */
1902
1903static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
1904static inline struct task_group *task_group(struct task_struct *p)
1905{
1906 return NULL;
1907}
1908
1909#endif /* CONFIG_CGROUP_SCHED */
1910
1911static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1912{
1913 set_task_rq(p, cpu);
1914#ifdef CONFIG_SMP
1915 /*
1916 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1917 * successfully executed on another CPU. We must ensure that updates of
1918 * per-task data have been completed by this moment.
1919 */
1920 smp_wmb();
1921#ifdef CONFIG_THREAD_INFO_IN_TASK
1922 WRITE_ONCE(p->cpu, cpu);
1923#else
1924 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
1925#endif
1926 p->wake_cpu = cpu;
1927#endif
1928}
1929
1930/*
1931 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
1932 */
1933#ifdef CONFIG_SCHED_DEBUG
1934# include <linux/static_key.h>
1935# define const_debug __read_mostly
1936#else
1937# define const_debug const
1938#endif
1939
1940#define SCHED_FEAT(name, enabled) \
1941 __SCHED_FEAT_##name ,
1942
1943enum {
1944#include "features.h"
1945 __SCHED_FEAT_NR,
1946};
1947
1948#undef SCHED_FEAT
1949
1950#ifdef CONFIG_SCHED_DEBUG
1951
1952/*
1953 * To support run-time toggling of sched features, all the translation units
1954 * (but core.c) reference the sysctl_sched_features defined in core.c.
1955 */
1956extern const_debug unsigned int sysctl_sched_features;
1957
1958#ifdef CONFIG_JUMP_LABEL
1959#define SCHED_FEAT(name, enabled) \
1960static __always_inline bool static_branch_##name(struct static_key *key) \
1961{ \
1962 return static_key_##enabled(key); \
1963}
1964
1965#include "features.h"
1966#undef SCHED_FEAT
1967
1968extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
1969#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
1970
1971#else /* !CONFIG_JUMP_LABEL */
1972
1973#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
1974
1975#endif /* CONFIG_JUMP_LABEL */
1976
1977#else /* !SCHED_DEBUG */
1978
1979/*
1980 * Each translation unit has its own copy of sysctl_sched_features to allow
1981 * constants propagation at compile time and compiler optimization based on
1982 * features default.
1983 */
1984#define SCHED_FEAT(name, enabled) \
1985 (1UL << __SCHED_FEAT_##name) * enabled |
1986static const_debug __maybe_unused unsigned int sysctl_sched_features =
1987#include "features.h"
1988 0;
1989#undef SCHED_FEAT
1990
1991#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
1992
1993#endif /* SCHED_DEBUG */
1994
1995extern struct static_key_false sched_numa_balancing;
1996extern struct static_key_false sched_schedstats;
1997
1998static inline u64 global_rt_period(void)
1999{
2000 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2001}
2002
2003static inline u64 global_rt_runtime(void)
2004{
2005 if (sysctl_sched_rt_runtime < 0)
2006 return RUNTIME_INF;
2007
2008 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2009}
2010
2011static inline int task_current(struct rq *rq, struct task_struct *p)
2012{
2013 return rq->curr == p;
2014}
2015
2016static inline int task_running(struct rq *rq, struct task_struct *p)
2017{
2018#ifdef CONFIG_SMP
2019 return p->on_cpu;
2020#else
2021 return task_current(rq, p);
2022#endif
2023}
2024
2025static inline int task_on_rq_queued(struct task_struct *p)
2026{
2027 return p->on_rq == TASK_ON_RQ_QUEUED;
2028}
2029
2030static inline int task_on_rq_migrating(struct task_struct *p)
2031{
2032 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2033}
2034
2035/* Wake flags. The first three directly map to some SD flag value */
2036#define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2037#define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2038#define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2039
2040#define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2041#define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2042#define WF_ON_CPU 0x40 /* Wakee is on_cpu */
2043
2044#ifdef CONFIG_SMP
2045static_assert(WF_EXEC == SD_BALANCE_EXEC);
2046static_assert(WF_FORK == SD_BALANCE_FORK);
2047static_assert(WF_TTWU == SD_BALANCE_WAKE);
2048#endif
2049
2050/*
2051 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2052 * of tasks with abnormal "nice" values across CPUs the contribution that
2053 * each task makes to its run queue's load is weighted according to its
2054 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2055 * scaled version of the new time slice allocation that they receive on time
2056 * slice expiry etc.
2057 */
2058
2059#define WEIGHT_IDLEPRIO 3
2060#define WMULT_IDLEPRIO 1431655765
2061
2062extern const int sched_prio_to_weight[40];
2063extern const u32 sched_prio_to_wmult[40];
2064
2065/*
2066 * {de,en}queue flags:
2067 *
2068 * DEQUEUE_SLEEP - task is no longer runnable
2069 * ENQUEUE_WAKEUP - task just became runnable
2070 *
2071 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2072 * are in a known state which allows modification. Such pairs
2073 * should preserve as much state as possible.
2074 *
2075 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2076 * in the runqueue.
2077 *
2078 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2079 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2080 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2081 *
2082 */
2083
2084#define DEQUEUE_SLEEP 0x01
2085#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2086#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2087#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2088
2089#define ENQUEUE_WAKEUP 0x01
2090#define ENQUEUE_RESTORE 0x02
2091#define ENQUEUE_MOVE 0x04
2092#define ENQUEUE_NOCLOCK 0x08
2093
2094#define ENQUEUE_HEAD 0x10
2095#define ENQUEUE_REPLENISH 0x20
2096#ifdef CONFIG_SMP
2097#define ENQUEUE_MIGRATED 0x40
2098#else
2099#define ENQUEUE_MIGRATED 0x00
2100#endif
2101
2102#define RETRY_TASK ((void *)-1UL)
2103
2104struct sched_class {
2105
2106#ifdef CONFIG_UCLAMP_TASK
2107 int uclamp_enabled;
2108#endif
2109
2110 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2111 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2112 void (*yield_task) (struct rq *rq);
2113 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2114
2115 void (*check_preempt_curr)(struct rq *rq, struct task_struct *p, int flags);
2116
2117 struct task_struct *(*pick_next_task)(struct rq *rq);
2118
2119 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2120 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2121
2122#ifdef CONFIG_SMP
2123 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2124 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2125
2126 struct task_struct * (*pick_task)(struct rq *rq);
2127
2128 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2129
2130 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2131
2132 void (*set_cpus_allowed)(struct task_struct *p,
2133 const struct cpumask *newmask,
2134 u32 flags);
2135
2136 void (*rq_online)(struct rq *rq);
2137 void (*rq_offline)(struct rq *rq);
2138
2139 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2140#endif
2141
2142 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2143 void (*task_fork)(struct task_struct *p);
2144 void (*task_dead)(struct task_struct *p);
2145
2146 /*
2147 * The switched_from() call is allowed to drop rq->lock, therefore we
2148 * cannot assume the switched_from/switched_to pair is serialized by
2149 * rq->lock. They are however serialized by p->pi_lock.
2150 */
2151 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2152 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2153 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2154 int oldprio);
2155
2156 unsigned int (*get_rr_interval)(struct rq *rq,
2157 struct task_struct *task);
2158
2159 void (*update_curr)(struct rq *rq);
2160
2161#define TASK_SET_GROUP 0
2162#define TASK_MOVE_GROUP 1
2163
2164#ifdef CONFIG_FAIR_GROUP_SCHED
2165 void (*task_change_group)(struct task_struct *p, int type);
2166#endif
2167};
2168
2169static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2170{
2171 WARN_ON_ONCE(rq->curr != prev);
2172 prev->sched_class->put_prev_task(rq, prev);
2173}
2174
2175static inline void set_next_task(struct rq *rq, struct task_struct *next)
2176{
2177 next->sched_class->set_next_task(rq, next, false);
2178}
2179
2180
2181/*
2182 * Helper to define a sched_class instance; each one is placed in a separate
2183 * section which is ordered by the linker script:
2184 *
2185 * include/asm-generic/vmlinux.lds.h
2186 *
2187 * Also enforce alignment on the instance, not the type, to guarantee layout.
2188 */
2189#define DEFINE_SCHED_CLASS(name) \
2190const struct sched_class name##_sched_class \
2191 __aligned(__alignof__(struct sched_class)) \
2192 __section("__" #name "_sched_class")
2193
2194/* Defined in include/asm-generic/vmlinux.lds.h */
2195extern struct sched_class __begin_sched_classes[];
2196extern struct sched_class __end_sched_classes[];
2197
2198#define sched_class_highest (__end_sched_classes - 1)
2199#define sched_class_lowest (__begin_sched_classes - 1)
2200
2201#define for_class_range(class, _from, _to) \
2202 for (class = (_from); class != (_to); class--)
2203
2204#define for_each_class(class) \
2205 for_class_range(class, sched_class_highest, sched_class_lowest)
2206
2207extern const struct sched_class stop_sched_class;
2208extern const struct sched_class dl_sched_class;
2209extern const struct sched_class rt_sched_class;
2210extern const struct sched_class fair_sched_class;
2211extern const struct sched_class idle_sched_class;
2212
2213static inline bool sched_stop_runnable(struct rq *rq)
2214{
2215 return rq->stop && task_on_rq_queued(rq->stop);
2216}
2217
2218static inline bool sched_dl_runnable(struct rq *rq)
2219{
2220 return rq->dl.dl_nr_running > 0;
2221}
2222
2223static inline bool sched_rt_runnable(struct rq *rq)
2224{
2225 return rq->rt.rt_queued > 0;
2226}
2227
2228static inline bool sched_fair_runnable(struct rq *rq)
2229{
2230 return rq->cfs.nr_running > 0;
2231}
2232
2233extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2234extern struct task_struct *pick_next_task_idle(struct rq *rq);
2235
2236#define SCA_CHECK 0x01
2237#define SCA_MIGRATE_DISABLE 0x02
2238#define SCA_MIGRATE_ENABLE 0x04
2239
2240#ifdef CONFIG_SMP
2241
2242extern void update_group_capacity(struct sched_domain *sd, int cpu);
2243
2244extern void trigger_load_balance(struct rq *rq);
2245
2246extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2247
2248static inline struct task_struct *get_push_task(struct rq *rq)
2249{
2250 struct task_struct *p = rq->curr;
2251
2252 lockdep_assert_rq_held(rq);
2253
2254 if (rq->push_busy)
2255 return NULL;
2256
2257 if (p->nr_cpus_allowed == 1)
2258 return NULL;
2259
2260 if (p->migration_disabled)
2261 return NULL;
2262
2263 rq->push_busy = true;
2264 return get_task_struct(p);
2265}
2266
2267extern int push_cpu_stop(void *arg);
2268
2269#endif
2270
2271#ifdef CONFIG_CPU_IDLE
2272static inline void idle_set_state(struct rq *rq,
2273 struct cpuidle_state *idle_state)
2274{
2275 rq->idle_state = idle_state;
2276}
2277
2278static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2279{
2280 SCHED_WARN_ON(!rcu_read_lock_held());
2281
2282 return rq->idle_state;
2283}
2284#else
2285static inline void idle_set_state(struct rq *rq,
2286 struct cpuidle_state *idle_state)
2287{
2288}
2289
2290static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2291{
2292 return NULL;
2293}
2294#endif
2295
2296extern void schedule_idle(void);
2297
2298extern void sysrq_sched_debug_show(void);
2299extern void sched_init_granularity(void);
2300extern void update_max_interval(void);
2301
2302extern void init_sched_dl_class(void);
2303extern void init_sched_rt_class(void);
2304extern void init_sched_fair_class(void);
2305
2306extern void reweight_task(struct task_struct *p, int prio);
2307
2308extern void resched_curr(struct rq *rq);
2309extern void resched_cpu(int cpu);
2310
2311extern struct rt_bandwidth def_rt_bandwidth;
2312extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2313
2314extern struct dl_bandwidth def_dl_bandwidth;
2315extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
2316extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
2317extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se);
2318
2319#define BW_SHIFT 20
2320#define BW_UNIT (1 << BW_SHIFT)
2321#define RATIO_SHIFT 8
2322#define MAX_BW_BITS (64 - BW_SHIFT)
2323#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2324unsigned long to_ratio(u64 period, u64 runtime);
2325
2326extern void init_entity_runnable_average(struct sched_entity *se);
2327extern void post_init_entity_util_avg(struct task_struct *p);
2328
2329#ifdef CONFIG_NO_HZ_FULL
2330extern bool sched_can_stop_tick(struct rq *rq);
2331extern int __init sched_tick_offload_init(void);
2332
2333/*
2334 * Tick may be needed by tasks in the runqueue depending on their policy and
2335 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2336 * nohz mode if necessary.
2337 */
2338static inline void sched_update_tick_dependency(struct rq *rq)
2339{
2340 int cpu = cpu_of(rq);
2341
2342 if (!tick_nohz_full_cpu(cpu))
2343 return;
2344
2345 if (sched_can_stop_tick(rq))
2346 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2347 else
2348 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2349}
2350#else
2351static inline int sched_tick_offload_init(void) { return 0; }
2352static inline void sched_update_tick_dependency(struct rq *rq) { }
2353#endif
2354
2355static inline void add_nr_running(struct rq *rq, unsigned count)
2356{
2357 unsigned prev_nr = rq->nr_running;
2358
2359 rq->nr_running = prev_nr + count;
2360 if (trace_sched_update_nr_running_tp_enabled()) {
2361 call_trace_sched_update_nr_running(rq, count);
2362 }
2363
2364#ifdef CONFIG_SMP
2365 if (prev_nr < 2 && rq->nr_running >= 2) {
2366 if (!READ_ONCE(rq->rd->overload))
2367 WRITE_ONCE(rq->rd->overload, 1);
2368 }
2369#endif
2370
2371 sched_update_tick_dependency(rq);
2372}
2373
2374static inline void sub_nr_running(struct rq *rq, unsigned count)
2375{
2376 rq->nr_running -= count;
2377 if (trace_sched_update_nr_running_tp_enabled()) {
2378 call_trace_sched_update_nr_running(rq, -count);
2379 }
2380
2381 /* Check if we still need preemption */
2382 sched_update_tick_dependency(rq);
2383}
2384
2385extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2386extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2387
2388extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
2389
2390extern const_debug unsigned int sysctl_sched_nr_migrate;
2391extern const_debug unsigned int sysctl_sched_migration_cost;
2392
2393#ifdef CONFIG_SCHED_HRTICK
2394
2395/*
2396 * Use hrtick when:
2397 * - enabled by features
2398 * - hrtimer is actually high res
2399 */
2400static inline int hrtick_enabled(struct rq *rq)
2401{
2402 if (!cpu_active(cpu_of(rq)))
2403 return 0;
2404 return hrtimer_is_hres_active(&rq->hrtick_timer);
2405}
2406
2407static inline int hrtick_enabled_fair(struct rq *rq)
2408{
2409 if (!sched_feat(HRTICK))
2410 return 0;
2411 return hrtick_enabled(rq);
2412}
2413
2414static inline int hrtick_enabled_dl(struct rq *rq)
2415{
2416 if (!sched_feat(HRTICK_DL))
2417 return 0;
2418 return hrtick_enabled(rq);
2419}
2420
2421void hrtick_start(struct rq *rq, u64 delay);
2422
2423#else
2424
2425static inline int hrtick_enabled_fair(struct rq *rq)
2426{
2427 return 0;
2428}
2429
2430static inline int hrtick_enabled_dl(struct rq *rq)
2431{
2432 return 0;
2433}
2434
2435static inline int hrtick_enabled(struct rq *rq)
2436{
2437 return 0;
2438}
2439
2440#endif /* CONFIG_SCHED_HRTICK */
2441
2442#ifndef arch_scale_freq_tick
2443static __always_inline
2444void arch_scale_freq_tick(void)
2445{
2446}
2447#endif
2448
2449#ifndef arch_scale_freq_capacity
2450/**
2451 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2452 * @cpu: the CPU in question.
2453 *
2454 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2455 *
2456 * f_curr
2457 * ------ * SCHED_CAPACITY_SCALE
2458 * f_max
2459 */
2460static __always_inline
2461unsigned long arch_scale_freq_capacity(int cpu)
2462{
2463 return SCHED_CAPACITY_SCALE;
2464}
2465#endif
2466
2467
2468#ifdef CONFIG_SMP
2469
2470static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2471{
2472#ifdef CONFIG_SCHED_CORE
2473 /*
2474 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2475 * order by core-id first and cpu-id second.
2476 *
2477 * Notably:
2478 *
2479 * double_rq_lock(0,3); will take core-0, core-1 lock
2480 * double_rq_lock(1,2); will take core-1, core-0 lock
2481 *
2482 * when only cpu-id is considered.
2483 */
2484 if (rq1->core->cpu < rq2->core->cpu)
2485 return true;
2486 if (rq1->core->cpu > rq2->core->cpu)
2487 return false;
2488
2489 /*
2490 * __sched_core_flip() relies on SMT having cpu-id lock order.
2491 */
2492#endif
2493 return rq1->cpu < rq2->cpu;
2494}
2495
2496extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2497
2498#ifdef CONFIG_PREEMPTION
2499
2500/*
2501 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2502 * way at the expense of forcing extra atomic operations in all
2503 * invocations. This assures that the double_lock is acquired using the
2504 * same underlying policy as the spinlock_t on this architecture, which
2505 * reduces latency compared to the unfair variant below. However, it
2506 * also adds more overhead and therefore may reduce throughput.
2507 */
2508static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2509 __releases(this_rq->lock)
2510 __acquires(busiest->lock)
2511 __acquires(this_rq->lock)
2512{
2513 raw_spin_rq_unlock(this_rq);
2514 double_rq_lock(this_rq, busiest);
2515
2516 return 1;
2517}
2518
2519#else
2520/*
2521 * Unfair double_lock_balance: Optimizes throughput at the expense of
2522 * latency by eliminating extra atomic operations when the locks are
2523 * already in proper order on entry. This favors lower CPU-ids and will
2524 * grant the double lock to lower CPUs over higher ids under contention,
2525 * regardless of entry order into the function.
2526 */
2527static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2528 __releases(this_rq->lock)
2529 __acquires(busiest->lock)
2530 __acquires(this_rq->lock)
2531{
2532 if (__rq_lockp(this_rq) == __rq_lockp(busiest))
2533 return 0;
2534
2535 if (likely(raw_spin_rq_trylock(busiest)))
2536 return 0;
2537
2538 if (rq_order_less(this_rq, busiest)) {
2539 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2540 return 0;
2541 }
2542
2543 raw_spin_rq_unlock(this_rq);
2544 double_rq_lock(this_rq, busiest);
2545
2546 return 1;
2547}
2548
2549#endif /* CONFIG_PREEMPTION */
2550
2551/*
2552 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2553 */
2554static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2555{
2556 lockdep_assert_irqs_disabled();
2557
2558 return _double_lock_balance(this_rq, busiest);
2559}
2560
2561static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2562 __releases(busiest->lock)
2563{
2564 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2565 raw_spin_rq_unlock(busiest);
2566 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2567}
2568
2569static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2570{
2571 if (l1 > l2)
2572 swap(l1, l2);
2573
2574 spin_lock(l1);
2575 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2576}
2577
2578static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2579{
2580 if (l1 > l2)
2581 swap(l1, l2);
2582
2583 spin_lock_irq(l1);
2584 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2585}
2586
2587static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2588{
2589 if (l1 > l2)
2590 swap(l1, l2);
2591
2592 raw_spin_lock(l1);
2593 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2594}
2595
2596/*
2597 * double_rq_unlock - safely unlock two runqueues
2598 *
2599 * Note this does not restore interrupts like task_rq_unlock,
2600 * you need to do so manually after calling.
2601 */
2602static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2603 __releases(rq1->lock)
2604 __releases(rq2->lock)
2605{
2606 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2607 raw_spin_rq_unlock(rq2);
2608 else
2609 __release(rq2->lock);
2610 raw_spin_rq_unlock(rq1);
2611}
2612
2613extern void set_rq_online (struct rq *rq);
2614extern void set_rq_offline(struct rq *rq);
2615extern bool sched_smp_initialized;
2616
2617#else /* CONFIG_SMP */
2618
2619/*
2620 * double_rq_lock - safely lock two runqueues
2621 *
2622 * Note this does not disable interrupts like task_rq_lock,
2623 * you need to do so manually before calling.
2624 */
2625static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2626 __acquires(rq1->lock)
2627 __acquires(rq2->lock)
2628{
2629 BUG_ON(!irqs_disabled());
2630 BUG_ON(rq1 != rq2);
2631 raw_spin_rq_lock(rq1);
2632 __acquire(rq2->lock); /* Fake it out ;) */
2633}
2634
2635/*
2636 * double_rq_unlock - safely unlock two runqueues
2637 *
2638 * Note this does not restore interrupts like task_rq_unlock,
2639 * you need to do so manually after calling.
2640 */
2641static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2642 __releases(rq1->lock)
2643 __releases(rq2->lock)
2644{
2645 BUG_ON(rq1 != rq2);
2646 raw_spin_rq_unlock(rq1);
2647 __release(rq2->lock);
2648}
2649
2650#endif
2651
2652extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2653extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2654
2655#ifdef CONFIG_SCHED_DEBUG
2656extern bool sched_debug_verbose;
2657
2658extern void print_cfs_stats(struct seq_file *m, int cpu);
2659extern void print_rt_stats(struct seq_file *m, int cpu);
2660extern void print_dl_stats(struct seq_file *m, int cpu);
2661extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2662extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2663extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2664
2665extern void resched_latency_warn(int cpu, u64 latency);
2666#ifdef CONFIG_NUMA_BALANCING
2667extern void
2668show_numa_stats(struct task_struct *p, struct seq_file *m);
2669extern void
2670print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2671 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2672#endif /* CONFIG_NUMA_BALANCING */
2673#else
2674static inline void resched_latency_warn(int cpu, u64 latency) {}
2675#endif /* CONFIG_SCHED_DEBUG */
2676
2677extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2678extern void init_rt_rq(struct rt_rq *rt_rq);
2679extern void init_dl_rq(struct dl_rq *dl_rq);
2680
2681extern void cfs_bandwidth_usage_inc(void);
2682extern void cfs_bandwidth_usage_dec(void);
2683
2684#ifdef CONFIG_NO_HZ_COMMON
2685#define NOHZ_BALANCE_KICK_BIT 0
2686#define NOHZ_STATS_KICK_BIT 1
2687#define NOHZ_NEWILB_KICK_BIT 2
2688
2689#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2690#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2691#define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2692
2693#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK)
2694
2695#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2696
2697extern void nohz_balance_exit_idle(struct rq *rq);
2698#else
2699static inline void nohz_balance_exit_idle(struct rq *rq) { }
2700#endif
2701
2702#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2703extern void nohz_run_idle_balance(int cpu);
2704#else
2705static inline void nohz_run_idle_balance(int cpu) { }
2706#endif
2707
2708#ifdef CONFIG_SMP
2709static inline
2710void __dl_update(struct dl_bw *dl_b, s64 bw)
2711{
2712 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
2713 int i;
2714
2715 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2716 "sched RCU must be held");
2717 for_each_cpu_and(i, rd->span, cpu_active_mask) {
2718 struct rq *rq = cpu_rq(i);
2719
2720 rq->dl.extra_bw += bw;
2721 }
2722}
2723#else
2724static inline
2725void __dl_update(struct dl_bw *dl_b, s64 bw)
2726{
2727 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
2728
2729 dl->extra_bw += bw;
2730}
2731#endif
2732
2733
2734#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2735struct irqtime {
2736 u64 total;
2737 u64 tick_delta;
2738 u64 irq_start_time;
2739 struct u64_stats_sync sync;
2740};
2741
2742DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2743
2744/*
2745 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2746 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2747 * and never move forward.
2748 */
2749static inline u64 irq_time_read(int cpu)
2750{
2751 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2752 unsigned int seq;
2753 u64 total;
2754
2755 do {
2756 seq = __u64_stats_fetch_begin(&irqtime->sync);
2757 total = irqtime->total;
2758 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2759
2760 return total;
2761}
2762#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2763
2764#ifdef CONFIG_CPU_FREQ
2765DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2766
2767/**
2768 * cpufreq_update_util - Take a note about CPU utilization changes.
2769 * @rq: Runqueue to carry out the update for.
2770 * @flags: Update reason flags.
2771 *
2772 * This function is called by the scheduler on the CPU whose utilization is
2773 * being updated.
2774 *
2775 * It can only be called from RCU-sched read-side critical sections.
2776 *
2777 * The way cpufreq is currently arranged requires it to evaluate the CPU
2778 * performance state (frequency/voltage) on a regular basis to prevent it from
2779 * being stuck in a completely inadequate performance level for too long.
2780 * That is not guaranteed to happen if the updates are only triggered from CFS
2781 * and DL, though, because they may not be coming in if only RT tasks are
2782 * active all the time (or there are RT tasks only).
2783 *
2784 * As a workaround for that issue, this function is called periodically by the
2785 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2786 * but that really is a band-aid. Going forward it should be replaced with
2787 * solutions targeted more specifically at RT tasks.
2788 */
2789static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2790{
2791 struct update_util_data *data;
2792
2793 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2794 cpu_of(rq)));
2795 if (data)
2796 data->func(data, rq_clock(rq), flags);
2797}
2798#else
2799static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2800#endif /* CONFIG_CPU_FREQ */
2801
2802#ifdef CONFIG_UCLAMP_TASK
2803unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
2804
2805/**
2806 * uclamp_rq_util_with - clamp @util with @rq and @p effective uclamp values.
2807 * @rq: The rq to clamp against. Must not be NULL.
2808 * @util: The util value to clamp.
2809 * @p: The task to clamp against. Can be NULL if you want to clamp
2810 * against @rq only.
2811 *
2812 * Clamps the passed @util to the max(@rq, @p) effective uclamp values.
2813 *
2814 * If sched_uclamp_used static key is disabled, then just return the util
2815 * without any clamping since uclamp aggregation at the rq level in the fast
2816 * path is disabled, rendering this operation a NOP.
2817 *
2818 * Use uclamp_eff_value() if you don't care about uclamp values at rq level. It
2819 * will return the correct effective uclamp value of the task even if the
2820 * static key is disabled.
2821 */
2822static __always_inline
2823unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
2824 struct task_struct *p)
2825{
2826 unsigned long min_util = 0;
2827 unsigned long max_util = 0;
2828
2829 if (!static_branch_likely(&sched_uclamp_used))
2830 return util;
2831
2832 if (p) {
2833 min_util = uclamp_eff_value(p, UCLAMP_MIN);
2834 max_util = uclamp_eff_value(p, UCLAMP_MAX);
2835
2836 /*
2837 * Ignore last runnable task's max clamp, as this task will
2838 * reset it. Similarly, no need to read the rq's min clamp.
2839 */
2840 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
2841 goto out;
2842 }
2843
2844 min_util = max_t(unsigned long, min_util, READ_ONCE(rq->uclamp[UCLAMP_MIN].value));
2845 max_util = max_t(unsigned long, max_util, READ_ONCE(rq->uclamp[UCLAMP_MAX].value));
2846out:
2847 /*
2848 * Since CPU's {min,max}_util clamps are MAX aggregated considering
2849 * RUNNABLE tasks with _different_ clamps, we can end up with an
2850 * inversion. Fix it now when the clamps are applied.
2851 */
2852 if (unlikely(min_util >= max_util))
2853 return min_util;
2854
2855 return clamp(util, min_util, max_util);
2856}
2857
2858/*
2859 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
2860 * by default in the fast path and only gets turned on once userspace performs
2861 * an operation that requires it.
2862 *
2863 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
2864 * hence is active.
2865 */
2866static inline bool uclamp_is_used(void)
2867{
2868 return static_branch_likely(&sched_uclamp_used);
2869}
2870#else /* CONFIG_UCLAMP_TASK */
2871static inline
2872unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
2873 struct task_struct *p)
2874{
2875 return util;
2876}
2877
2878static inline bool uclamp_is_used(void)
2879{
2880 return false;
2881}
2882#endif /* CONFIG_UCLAMP_TASK */
2883
2884#ifdef arch_scale_freq_capacity
2885# ifndef arch_scale_freq_invariant
2886# define arch_scale_freq_invariant() true
2887# endif
2888#else
2889# define arch_scale_freq_invariant() false
2890#endif
2891
2892#ifdef CONFIG_SMP
2893static inline unsigned long capacity_orig_of(int cpu)
2894{
2895 return cpu_rq(cpu)->cpu_capacity_orig;
2896}
2897
2898/**
2899 * enum cpu_util_type - CPU utilization type
2900 * @FREQUENCY_UTIL: Utilization used to select frequency
2901 * @ENERGY_UTIL: Utilization used during energy calculation
2902 *
2903 * The utilization signals of all scheduling classes (CFS/RT/DL) and IRQ time
2904 * need to be aggregated differently depending on the usage made of them. This
2905 * enum is used within effective_cpu_util() to differentiate the types of
2906 * utilization expected by the callers, and adjust the aggregation accordingly.
2907 */
2908enum cpu_util_type {
2909 FREQUENCY_UTIL,
2910 ENERGY_UTIL,
2911};
2912
2913unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
2914 unsigned long max, enum cpu_util_type type,
2915 struct task_struct *p);
2916
2917static inline unsigned long cpu_bw_dl(struct rq *rq)
2918{
2919 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
2920}
2921
2922static inline unsigned long cpu_util_dl(struct rq *rq)
2923{
2924 return READ_ONCE(rq->avg_dl.util_avg);
2925}
2926
2927static inline unsigned long cpu_util_cfs(struct rq *rq)
2928{
2929 unsigned long util = READ_ONCE(rq->cfs.avg.util_avg);
2930
2931 if (sched_feat(UTIL_EST)) {
2932 util = max_t(unsigned long, util,
2933 READ_ONCE(rq->cfs.avg.util_est.enqueued));
2934 }
2935
2936 return util;
2937}
2938
2939static inline unsigned long cpu_util_rt(struct rq *rq)
2940{
2941 return READ_ONCE(rq->avg_rt.util_avg);
2942}
2943#endif
2944
2945#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
2946static inline unsigned long cpu_util_irq(struct rq *rq)
2947{
2948 return rq->avg_irq.util_avg;
2949}
2950
2951static inline
2952unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
2953{
2954 util *= (max - irq);
2955 util /= max;
2956
2957 return util;
2958
2959}
2960#else
2961static inline unsigned long cpu_util_irq(struct rq *rq)
2962{
2963 return 0;
2964}
2965
2966static inline
2967unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
2968{
2969 return util;
2970}
2971#endif
2972
2973#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2974
2975#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
2976
2977DECLARE_STATIC_KEY_FALSE(sched_energy_present);
2978
2979static inline bool sched_energy_enabled(void)
2980{
2981 return static_branch_unlikely(&sched_energy_present);
2982}
2983
2984#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
2985
2986#define perf_domain_span(pd) NULL
2987static inline bool sched_energy_enabled(void) { return false; }
2988
2989#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
2990
2991#ifdef CONFIG_MEMBARRIER
2992/*
2993 * The scheduler provides memory barriers required by membarrier between:
2994 * - prior user-space memory accesses and store to rq->membarrier_state,
2995 * - store to rq->membarrier_state and following user-space memory accesses.
2996 * In the same way it provides those guarantees around store to rq->curr.
2997 */
2998static inline void membarrier_switch_mm(struct rq *rq,
2999 struct mm_struct *prev_mm,
3000 struct mm_struct *next_mm)
3001{
3002 int membarrier_state;
3003
3004 if (prev_mm == next_mm)
3005 return;
3006
3007 membarrier_state = atomic_read(&next_mm->membarrier_state);
3008 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3009 return;
3010
3011 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3012}
3013#else
3014static inline void membarrier_switch_mm(struct rq *rq,
3015 struct mm_struct *prev_mm,
3016 struct mm_struct *next_mm)
3017{
3018}
3019#endif
3020
3021#ifdef CONFIG_SMP
3022static inline bool is_per_cpu_kthread(struct task_struct *p)
3023{
3024 if (!(p->flags & PF_KTHREAD))
3025 return false;
3026
3027 if (p->nr_cpus_allowed != 1)
3028 return false;
3029
3030 return true;
3031}
3032#endif
3033
3034extern void swake_up_all_locked(struct swait_queue_head *q);
3035extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3036
3037#ifdef CONFIG_PREEMPT_DYNAMIC
3038extern int preempt_dynamic_mode;
3039extern int sched_dynamic_mode(const char *str);
3040extern void sched_dynamic_update(int mode);
3041#endif
3042