Loading...
1
2#include <linux/sched.h>
3#include <linux/sched/sysctl.h>
4#include <linux/sched/rt.h>
5#include <linux/sched/deadline.h>
6#include <linux/binfmts.h>
7#include <linux/mutex.h>
8#include <linux/spinlock.h>
9#include <linux/stop_machine.h>
10#include <linux/irq_work.h>
11#include <linux/tick.h>
12#include <linux/slab.h>
13
14#include "cpupri.h"
15#include "cpudeadline.h"
16#include "cpuacct.h"
17
18struct rq;
19struct cpuidle_state;
20
21/* task_struct::on_rq states: */
22#define TASK_ON_RQ_QUEUED 1
23#define TASK_ON_RQ_MIGRATING 2
24
25extern __read_mostly int scheduler_running;
26
27extern unsigned long calc_load_update;
28extern atomic_long_t calc_load_tasks;
29
30extern void calc_global_load_tick(struct rq *this_rq);
31extern long calc_load_fold_active(struct rq *this_rq);
32
33#ifdef CONFIG_SMP
34extern void update_cpu_load_active(struct rq *this_rq);
35#else
36static inline void update_cpu_load_active(struct rq *this_rq) { }
37#endif
38
39/*
40 * Helpers for converting nanosecond timing to jiffy resolution
41 */
42#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
43
44/*
45 * Increase resolution of nice-level calculations for 64-bit architectures.
46 * The extra resolution improves shares distribution and load balancing of
47 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
48 * hierarchies, especially on larger systems. This is not a user-visible change
49 * and does not change the user-interface for setting shares/weights.
50 *
51 * We increase resolution only if we have enough bits to allow this increased
52 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
53 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
54 * increased costs.
55 */
56#if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
57# define SCHED_LOAD_RESOLUTION 10
58# define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
59# define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
60#else
61# define SCHED_LOAD_RESOLUTION 0
62# define scale_load(w) (w)
63# define scale_load_down(w) (w)
64#endif
65
66#define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
67#define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
68
69#define NICE_0_LOAD SCHED_LOAD_SCALE
70#define NICE_0_SHIFT SCHED_LOAD_SHIFT
71
72/*
73 * Single value that decides SCHED_DEADLINE internal math precision.
74 * 10 -> just above 1us
75 * 9 -> just above 0.5us
76 */
77#define DL_SCALE (10)
78
79/*
80 * These are the 'tuning knobs' of the scheduler:
81 */
82
83/*
84 * single value that denotes runtime == period, ie unlimited time.
85 */
86#define RUNTIME_INF ((u64)~0ULL)
87
88static inline int idle_policy(int policy)
89{
90 return policy == SCHED_IDLE;
91}
92static inline int fair_policy(int policy)
93{
94 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
95}
96
97static inline int rt_policy(int policy)
98{
99 return policy == SCHED_FIFO || policy == SCHED_RR;
100}
101
102static inline int dl_policy(int policy)
103{
104 return policy == SCHED_DEADLINE;
105}
106static inline bool valid_policy(int policy)
107{
108 return idle_policy(policy) || fair_policy(policy) ||
109 rt_policy(policy) || dl_policy(policy);
110}
111
112static inline int task_has_rt_policy(struct task_struct *p)
113{
114 return rt_policy(p->policy);
115}
116
117static inline int task_has_dl_policy(struct task_struct *p)
118{
119 return dl_policy(p->policy);
120}
121
122/*
123 * Tells if entity @a should preempt entity @b.
124 */
125static inline bool
126dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
127{
128 return dl_time_before(a->deadline, b->deadline);
129}
130
131/*
132 * This is the priority-queue data structure of the RT scheduling class:
133 */
134struct rt_prio_array {
135 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
136 struct list_head queue[MAX_RT_PRIO];
137};
138
139struct rt_bandwidth {
140 /* nests inside the rq lock: */
141 raw_spinlock_t rt_runtime_lock;
142 ktime_t rt_period;
143 u64 rt_runtime;
144 struct hrtimer rt_period_timer;
145 unsigned int rt_period_active;
146};
147
148void __dl_clear_params(struct task_struct *p);
149
150/*
151 * To keep the bandwidth of -deadline tasks and groups under control
152 * we need some place where:
153 * - store the maximum -deadline bandwidth of the system (the group);
154 * - cache the fraction of that bandwidth that is currently allocated.
155 *
156 * This is all done in the data structure below. It is similar to the
157 * one used for RT-throttling (rt_bandwidth), with the main difference
158 * that, since here we are only interested in admission control, we
159 * do not decrease any runtime while the group "executes", neither we
160 * need a timer to replenish it.
161 *
162 * With respect to SMP, the bandwidth is given on a per-CPU basis,
163 * meaning that:
164 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
165 * - dl_total_bw array contains, in the i-eth element, the currently
166 * allocated bandwidth on the i-eth CPU.
167 * Moreover, groups consume bandwidth on each CPU, while tasks only
168 * consume bandwidth on the CPU they're running on.
169 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
170 * that will be shown the next time the proc or cgroup controls will
171 * be red. It on its turn can be changed by writing on its own
172 * control.
173 */
174struct dl_bandwidth {
175 raw_spinlock_t dl_runtime_lock;
176 u64 dl_runtime;
177 u64 dl_period;
178};
179
180static inline int dl_bandwidth_enabled(void)
181{
182 return sysctl_sched_rt_runtime >= 0;
183}
184
185extern struct dl_bw *dl_bw_of(int i);
186
187struct dl_bw {
188 raw_spinlock_t lock;
189 u64 bw, total_bw;
190};
191
192static inline
193void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
194{
195 dl_b->total_bw -= tsk_bw;
196}
197
198static inline
199void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
200{
201 dl_b->total_bw += tsk_bw;
202}
203
204static inline
205bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
206{
207 return dl_b->bw != -1 &&
208 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
209}
210
211extern struct mutex sched_domains_mutex;
212
213#ifdef CONFIG_CGROUP_SCHED
214
215#include <linux/cgroup.h>
216
217struct cfs_rq;
218struct rt_rq;
219
220extern struct list_head task_groups;
221
222struct cfs_bandwidth {
223#ifdef CONFIG_CFS_BANDWIDTH
224 raw_spinlock_t lock;
225 ktime_t period;
226 u64 quota, runtime;
227 s64 hierarchical_quota;
228 u64 runtime_expires;
229
230 int idle, period_active;
231 struct hrtimer period_timer, slack_timer;
232 struct list_head throttled_cfs_rq;
233
234 /* statistics */
235 int nr_periods, nr_throttled;
236 u64 throttled_time;
237#endif
238};
239
240/* task group related information */
241struct task_group {
242 struct cgroup_subsys_state css;
243
244#ifdef CONFIG_FAIR_GROUP_SCHED
245 /* schedulable entities of this group on each cpu */
246 struct sched_entity **se;
247 /* runqueue "owned" by this group on each cpu */
248 struct cfs_rq **cfs_rq;
249 unsigned long shares;
250
251#ifdef CONFIG_SMP
252 /*
253 * load_avg can be heavily contended at clock tick time, so put
254 * it in its own cacheline separated from the fields above which
255 * will also be accessed at each tick.
256 */
257 atomic_long_t load_avg ____cacheline_aligned;
258#endif
259#endif
260
261#ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
264
265 struct rt_bandwidth rt_bandwidth;
266#endif
267
268 struct rcu_head rcu;
269 struct list_head list;
270
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
274
275#ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
277#endif
278
279 struct cfs_bandwidth cfs_bandwidth;
280};
281
282#ifdef CONFIG_FAIR_GROUP_SCHED
283#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
284
285/*
286 * A weight of 0 or 1 can cause arithmetics problems.
287 * A weight of a cfs_rq is the sum of weights of which entities
288 * are queued on this cfs_rq, so a weight of a entity should not be
289 * too large, so as the shares value of a task group.
290 * (The default weight is 1024 - so there's no practical
291 * limitation from this.)
292 */
293#define MIN_SHARES (1UL << 1)
294#define MAX_SHARES (1UL << 18)
295#endif
296
297typedef int (*tg_visitor)(struct task_group *, void *);
298
299extern int walk_tg_tree_from(struct task_group *from,
300 tg_visitor down, tg_visitor up, void *data);
301
302/*
303 * Iterate the full tree, calling @down when first entering a node and @up when
304 * leaving it for the final time.
305 *
306 * Caller must hold rcu_lock or sufficient equivalent.
307 */
308static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
309{
310 return walk_tg_tree_from(&root_task_group, down, up, data);
311}
312
313extern int tg_nop(struct task_group *tg, void *data);
314
315extern void free_fair_sched_group(struct task_group *tg);
316extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
317extern void unregister_fair_sched_group(struct task_group *tg);
318extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
319 struct sched_entity *se, int cpu,
320 struct sched_entity *parent);
321extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
322
323extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
324extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
325extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
326
327extern void free_rt_sched_group(struct task_group *tg);
328extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
329extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
330 struct sched_rt_entity *rt_se, int cpu,
331 struct sched_rt_entity *parent);
332
333extern struct task_group *sched_create_group(struct task_group *parent);
334extern void sched_online_group(struct task_group *tg,
335 struct task_group *parent);
336extern void sched_destroy_group(struct task_group *tg);
337extern void sched_offline_group(struct task_group *tg);
338
339extern void sched_move_task(struct task_struct *tsk);
340
341#ifdef CONFIG_FAIR_GROUP_SCHED
342extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
343
344#ifdef CONFIG_SMP
345extern void set_task_rq_fair(struct sched_entity *se,
346 struct cfs_rq *prev, struct cfs_rq *next);
347#else /* !CONFIG_SMP */
348static inline void set_task_rq_fair(struct sched_entity *se,
349 struct cfs_rq *prev, struct cfs_rq *next) { }
350#endif /* CONFIG_SMP */
351#endif /* CONFIG_FAIR_GROUP_SCHED */
352
353#else /* CONFIG_CGROUP_SCHED */
354
355struct cfs_bandwidth { };
356
357#endif /* CONFIG_CGROUP_SCHED */
358
359/* CFS-related fields in a runqueue */
360struct cfs_rq {
361 struct load_weight load;
362 unsigned int nr_running, h_nr_running;
363
364 u64 exec_clock;
365 u64 min_vruntime;
366#ifndef CONFIG_64BIT
367 u64 min_vruntime_copy;
368#endif
369
370 struct rb_root tasks_timeline;
371 struct rb_node *rb_leftmost;
372
373 /*
374 * 'curr' points to currently running entity on this cfs_rq.
375 * It is set to NULL otherwise (i.e when none are currently running).
376 */
377 struct sched_entity *curr, *next, *last, *skip;
378
379#ifdef CONFIG_SCHED_DEBUG
380 unsigned int nr_spread_over;
381#endif
382
383#ifdef CONFIG_SMP
384 /*
385 * CFS load tracking
386 */
387 struct sched_avg avg;
388 u64 runnable_load_sum;
389 unsigned long runnable_load_avg;
390#ifdef CONFIG_FAIR_GROUP_SCHED
391 unsigned long tg_load_avg_contrib;
392#endif
393 atomic_long_t removed_load_avg, removed_util_avg;
394#ifndef CONFIG_64BIT
395 u64 load_last_update_time_copy;
396#endif
397
398#ifdef CONFIG_FAIR_GROUP_SCHED
399 /*
400 * h_load = weight * f(tg)
401 *
402 * Where f(tg) is the recursive weight fraction assigned to
403 * this group.
404 */
405 unsigned long h_load;
406 u64 last_h_load_update;
407 struct sched_entity *h_load_next;
408#endif /* CONFIG_FAIR_GROUP_SCHED */
409#endif /* CONFIG_SMP */
410
411#ifdef CONFIG_FAIR_GROUP_SCHED
412 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
413
414 /*
415 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
416 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
417 * (like users, containers etc.)
418 *
419 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
420 * list is used during load balance.
421 */
422 int on_list;
423 struct list_head leaf_cfs_rq_list;
424 struct task_group *tg; /* group that "owns" this runqueue */
425
426#ifdef CONFIG_CFS_BANDWIDTH
427 int runtime_enabled;
428 u64 runtime_expires;
429 s64 runtime_remaining;
430
431 u64 throttled_clock, throttled_clock_task;
432 u64 throttled_clock_task_time;
433 int throttled, throttle_count;
434 struct list_head throttled_list;
435#endif /* CONFIG_CFS_BANDWIDTH */
436#endif /* CONFIG_FAIR_GROUP_SCHED */
437};
438
439static inline int rt_bandwidth_enabled(void)
440{
441 return sysctl_sched_rt_runtime >= 0;
442}
443
444/* RT IPI pull logic requires IRQ_WORK */
445#ifdef CONFIG_IRQ_WORK
446# define HAVE_RT_PUSH_IPI
447#endif
448
449/* Real-Time classes' related field in a runqueue: */
450struct rt_rq {
451 struct rt_prio_array active;
452 unsigned int rt_nr_running;
453 unsigned int rr_nr_running;
454#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
455 struct {
456 int curr; /* highest queued rt task prio */
457#ifdef CONFIG_SMP
458 int next; /* next highest */
459#endif
460 } highest_prio;
461#endif
462#ifdef CONFIG_SMP
463 unsigned long rt_nr_migratory;
464 unsigned long rt_nr_total;
465 int overloaded;
466 struct plist_head pushable_tasks;
467#ifdef HAVE_RT_PUSH_IPI
468 int push_flags;
469 int push_cpu;
470 struct irq_work push_work;
471 raw_spinlock_t push_lock;
472#endif
473#endif /* CONFIG_SMP */
474 int rt_queued;
475
476 int rt_throttled;
477 u64 rt_time;
478 u64 rt_runtime;
479 /* Nests inside the rq lock: */
480 raw_spinlock_t rt_runtime_lock;
481
482#ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted;
484
485 struct rq *rq;
486 struct task_group *tg;
487#endif
488};
489
490/* Deadline class' related fields in a runqueue */
491struct dl_rq {
492 /* runqueue is an rbtree, ordered by deadline */
493 struct rb_root rb_root;
494 struct rb_node *rb_leftmost;
495
496 unsigned long dl_nr_running;
497
498#ifdef CONFIG_SMP
499 /*
500 * Deadline values of the currently executing and the
501 * earliest ready task on this rq. Caching these facilitates
502 * the decision wether or not a ready but not running task
503 * should migrate somewhere else.
504 */
505 struct {
506 u64 curr;
507 u64 next;
508 } earliest_dl;
509
510 unsigned long dl_nr_migratory;
511 int overloaded;
512
513 /*
514 * Tasks on this rq that can be pushed away. They are kept in
515 * an rb-tree, ordered by tasks' deadlines, with caching
516 * of the leftmost (earliest deadline) element.
517 */
518 struct rb_root pushable_dl_tasks_root;
519 struct rb_node *pushable_dl_tasks_leftmost;
520#else
521 struct dl_bw dl_bw;
522#endif
523};
524
525#ifdef CONFIG_SMP
526
527/*
528 * We add the notion of a root-domain which will be used to define per-domain
529 * variables. Each exclusive cpuset essentially defines an island domain by
530 * fully partitioning the member cpus from any other cpuset. Whenever a new
531 * exclusive cpuset is created, we also create and attach a new root-domain
532 * object.
533 *
534 */
535struct root_domain {
536 atomic_t refcount;
537 atomic_t rto_count;
538 struct rcu_head rcu;
539 cpumask_var_t span;
540 cpumask_var_t online;
541
542 /* Indicate more than one runnable task for any CPU */
543 bool overload;
544
545 /*
546 * The bit corresponding to a CPU gets set here if such CPU has more
547 * than one runnable -deadline task (as it is below for RT tasks).
548 */
549 cpumask_var_t dlo_mask;
550 atomic_t dlo_count;
551 struct dl_bw dl_bw;
552 struct cpudl cpudl;
553
554 /*
555 * The "RT overload" flag: it gets set if a CPU has more than
556 * one runnable RT task.
557 */
558 cpumask_var_t rto_mask;
559 struct cpupri cpupri;
560};
561
562extern struct root_domain def_root_domain;
563
564#endif /* CONFIG_SMP */
565
566/*
567 * This is the main, per-CPU runqueue data structure.
568 *
569 * Locking rule: those places that want to lock multiple runqueues
570 * (such as the load balancing or the thread migration code), lock
571 * acquire operations must be ordered by ascending &runqueue.
572 */
573struct rq {
574 /* runqueue lock: */
575 raw_spinlock_t lock;
576
577 /*
578 * nr_running and cpu_load should be in the same cacheline because
579 * remote CPUs use both these fields when doing load calculation.
580 */
581 unsigned int nr_running;
582#ifdef CONFIG_NUMA_BALANCING
583 unsigned int nr_numa_running;
584 unsigned int nr_preferred_running;
585#endif
586 #define CPU_LOAD_IDX_MAX 5
587 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
588 unsigned long last_load_update_tick;
589#ifdef CONFIG_NO_HZ_COMMON
590 u64 nohz_stamp;
591 unsigned long nohz_flags;
592#endif
593#ifdef CONFIG_NO_HZ_FULL
594 unsigned long last_sched_tick;
595#endif
596 /* capture load from *all* tasks on this cpu: */
597 struct load_weight load;
598 unsigned long nr_load_updates;
599 u64 nr_switches;
600
601 struct cfs_rq cfs;
602 struct rt_rq rt;
603 struct dl_rq dl;
604
605#ifdef CONFIG_FAIR_GROUP_SCHED
606 /* list of leaf cfs_rq on this cpu: */
607 struct list_head leaf_cfs_rq_list;
608#endif /* CONFIG_FAIR_GROUP_SCHED */
609
610 /*
611 * This is part of a global counter where only the total sum
612 * over all CPUs matters. A task can increase this counter on
613 * one CPU and if it got migrated afterwards it may decrease
614 * it on another CPU. Always updated under the runqueue lock:
615 */
616 unsigned long nr_uninterruptible;
617
618 struct task_struct *curr, *idle, *stop;
619 unsigned long next_balance;
620 struct mm_struct *prev_mm;
621
622 unsigned int clock_skip_update;
623 u64 clock;
624 u64 clock_task;
625
626 atomic_t nr_iowait;
627
628#ifdef CONFIG_SMP
629 struct root_domain *rd;
630 struct sched_domain *sd;
631
632 unsigned long cpu_capacity;
633 unsigned long cpu_capacity_orig;
634
635 struct callback_head *balance_callback;
636
637 unsigned char idle_balance;
638 /* For active balancing */
639 int active_balance;
640 int push_cpu;
641 struct cpu_stop_work active_balance_work;
642 /* cpu of this runqueue: */
643 int cpu;
644 int online;
645
646 struct list_head cfs_tasks;
647
648 u64 rt_avg;
649 u64 age_stamp;
650 u64 idle_stamp;
651 u64 avg_idle;
652
653 /* This is used to determine avg_idle's max value */
654 u64 max_idle_balance_cost;
655#endif
656
657#ifdef CONFIG_IRQ_TIME_ACCOUNTING
658 u64 prev_irq_time;
659#endif
660#ifdef CONFIG_PARAVIRT
661 u64 prev_steal_time;
662#endif
663#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
664 u64 prev_steal_time_rq;
665#endif
666
667 /* calc_load related fields */
668 unsigned long calc_load_update;
669 long calc_load_active;
670
671#ifdef CONFIG_SCHED_HRTICK
672#ifdef CONFIG_SMP
673 int hrtick_csd_pending;
674 struct call_single_data hrtick_csd;
675#endif
676 struct hrtimer hrtick_timer;
677#endif
678
679#ifdef CONFIG_SCHEDSTATS
680 /* latency stats */
681 struct sched_info rq_sched_info;
682 unsigned long long rq_cpu_time;
683 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
684
685 /* sys_sched_yield() stats */
686 unsigned int yld_count;
687
688 /* schedule() stats */
689 unsigned int sched_count;
690 unsigned int sched_goidle;
691
692 /* try_to_wake_up() stats */
693 unsigned int ttwu_count;
694 unsigned int ttwu_local;
695#endif
696
697#ifdef CONFIG_SMP
698 struct llist_head wake_list;
699#endif
700
701#ifdef CONFIG_CPU_IDLE
702 /* Must be inspected within a rcu lock section */
703 struct cpuidle_state *idle_state;
704#endif
705};
706
707static inline int cpu_of(struct rq *rq)
708{
709#ifdef CONFIG_SMP
710 return rq->cpu;
711#else
712 return 0;
713#endif
714}
715
716DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
717
718#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
719#define this_rq() this_cpu_ptr(&runqueues)
720#define task_rq(p) cpu_rq(task_cpu(p))
721#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
722#define raw_rq() raw_cpu_ptr(&runqueues)
723
724static inline u64 __rq_clock_broken(struct rq *rq)
725{
726 return READ_ONCE(rq->clock);
727}
728
729static inline u64 rq_clock(struct rq *rq)
730{
731 lockdep_assert_held(&rq->lock);
732 return rq->clock;
733}
734
735static inline u64 rq_clock_task(struct rq *rq)
736{
737 lockdep_assert_held(&rq->lock);
738 return rq->clock_task;
739}
740
741#define RQCF_REQ_SKIP 0x01
742#define RQCF_ACT_SKIP 0x02
743
744static inline void rq_clock_skip_update(struct rq *rq, bool skip)
745{
746 lockdep_assert_held(&rq->lock);
747 if (skip)
748 rq->clock_skip_update |= RQCF_REQ_SKIP;
749 else
750 rq->clock_skip_update &= ~RQCF_REQ_SKIP;
751}
752
753#ifdef CONFIG_NUMA
754enum numa_topology_type {
755 NUMA_DIRECT,
756 NUMA_GLUELESS_MESH,
757 NUMA_BACKPLANE,
758};
759extern enum numa_topology_type sched_numa_topology_type;
760extern int sched_max_numa_distance;
761extern bool find_numa_distance(int distance);
762#endif
763
764#ifdef CONFIG_NUMA_BALANCING
765/* The regions in numa_faults array from task_struct */
766enum numa_faults_stats {
767 NUMA_MEM = 0,
768 NUMA_CPU,
769 NUMA_MEMBUF,
770 NUMA_CPUBUF
771};
772extern void sched_setnuma(struct task_struct *p, int node);
773extern int migrate_task_to(struct task_struct *p, int cpu);
774extern int migrate_swap(struct task_struct *, struct task_struct *);
775#endif /* CONFIG_NUMA_BALANCING */
776
777#ifdef CONFIG_SMP
778
779static inline void
780queue_balance_callback(struct rq *rq,
781 struct callback_head *head,
782 void (*func)(struct rq *rq))
783{
784 lockdep_assert_held(&rq->lock);
785
786 if (unlikely(head->next))
787 return;
788
789 head->func = (void (*)(struct callback_head *))func;
790 head->next = rq->balance_callback;
791 rq->balance_callback = head;
792}
793
794extern void sched_ttwu_pending(void);
795
796#define rcu_dereference_check_sched_domain(p) \
797 rcu_dereference_check((p), \
798 lockdep_is_held(&sched_domains_mutex))
799
800/*
801 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
802 * See detach_destroy_domains: synchronize_sched for details.
803 *
804 * The domain tree of any CPU may only be accessed from within
805 * preempt-disabled sections.
806 */
807#define for_each_domain(cpu, __sd) \
808 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
809 __sd; __sd = __sd->parent)
810
811#define for_each_lower_domain(sd) for (; sd; sd = sd->child)
812
813/**
814 * highest_flag_domain - Return highest sched_domain containing flag.
815 * @cpu: The cpu whose highest level of sched domain is to
816 * be returned.
817 * @flag: The flag to check for the highest sched_domain
818 * for the given cpu.
819 *
820 * Returns the highest sched_domain of a cpu which contains the given flag.
821 */
822static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
823{
824 struct sched_domain *sd, *hsd = NULL;
825
826 for_each_domain(cpu, sd) {
827 if (!(sd->flags & flag))
828 break;
829 hsd = sd;
830 }
831
832 return hsd;
833}
834
835static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
836{
837 struct sched_domain *sd;
838
839 for_each_domain(cpu, sd) {
840 if (sd->flags & flag)
841 break;
842 }
843
844 return sd;
845}
846
847DECLARE_PER_CPU(struct sched_domain *, sd_llc);
848DECLARE_PER_CPU(int, sd_llc_size);
849DECLARE_PER_CPU(int, sd_llc_id);
850DECLARE_PER_CPU(struct sched_domain *, sd_numa);
851DECLARE_PER_CPU(struct sched_domain *, sd_busy);
852DECLARE_PER_CPU(struct sched_domain *, sd_asym);
853
854struct sched_group_capacity {
855 atomic_t ref;
856 /*
857 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
858 * for a single CPU.
859 */
860 unsigned int capacity;
861 unsigned long next_update;
862 int imbalance; /* XXX unrelated to capacity but shared group state */
863 /*
864 * Number of busy cpus in this group.
865 */
866 atomic_t nr_busy_cpus;
867
868 unsigned long cpumask[0]; /* iteration mask */
869};
870
871struct sched_group {
872 struct sched_group *next; /* Must be a circular list */
873 atomic_t ref;
874
875 unsigned int group_weight;
876 struct sched_group_capacity *sgc;
877
878 /*
879 * The CPUs this group covers.
880 *
881 * NOTE: this field is variable length. (Allocated dynamically
882 * by attaching extra space to the end of the structure,
883 * depending on how many CPUs the kernel has booted up with)
884 */
885 unsigned long cpumask[0];
886};
887
888static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
889{
890 return to_cpumask(sg->cpumask);
891}
892
893/*
894 * cpumask masking which cpus in the group are allowed to iterate up the domain
895 * tree.
896 */
897static inline struct cpumask *sched_group_mask(struct sched_group *sg)
898{
899 return to_cpumask(sg->sgc->cpumask);
900}
901
902/**
903 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
904 * @group: The group whose first cpu is to be returned.
905 */
906static inline unsigned int group_first_cpu(struct sched_group *group)
907{
908 return cpumask_first(sched_group_cpus(group));
909}
910
911extern int group_balance_cpu(struct sched_group *sg);
912
913#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
914void register_sched_domain_sysctl(void);
915void unregister_sched_domain_sysctl(void);
916#else
917static inline void register_sched_domain_sysctl(void)
918{
919}
920static inline void unregister_sched_domain_sysctl(void)
921{
922}
923#endif
924
925#else
926
927static inline void sched_ttwu_pending(void) { }
928
929#endif /* CONFIG_SMP */
930
931#include "stats.h"
932#include "auto_group.h"
933
934#ifdef CONFIG_CGROUP_SCHED
935
936/*
937 * Return the group to which this tasks belongs.
938 *
939 * We cannot use task_css() and friends because the cgroup subsystem
940 * changes that value before the cgroup_subsys::attach() method is called,
941 * therefore we cannot pin it and might observe the wrong value.
942 *
943 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
944 * core changes this before calling sched_move_task().
945 *
946 * Instead we use a 'copy' which is updated from sched_move_task() while
947 * holding both task_struct::pi_lock and rq::lock.
948 */
949static inline struct task_group *task_group(struct task_struct *p)
950{
951 return p->sched_task_group;
952}
953
954/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
955static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
956{
957#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
958 struct task_group *tg = task_group(p);
959#endif
960
961#ifdef CONFIG_FAIR_GROUP_SCHED
962 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
963 p->se.cfs_rq = tg->cfs_rq[cpu];
964 p->se.parent = tg->se[cpu];
965#endif
966
967#ifdef CONFIG_RT_GROUP_SCHED
968 p->rt.rt_rq = tg->rt_rq[cpu];
969 p->rt.parent = tg->rt_se[cpu];
970#endif
971}
972
973#else /* CONFIG_CGROUP_SCHED */
974
975static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
976static inline struct task_group *task_group(struct task_struct *p)
977{
978 return NULL;
979}
980
981#endif /* CONFIG_CGROUP_SCHED */
982
983static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
984{
985 set_task_rq(p, cpu);
986#ifdef CONFIG_SMP
987 /*
988 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
989 * successfuly executed on another CPU. We must ensure that updates of
990 * per-task data have been completed by this moment.
991 */
992 smp_wmb();
993 task_thread_info(p)->cpu = cpu;
994 p->wake_cpu = cpu;
995#endif
996}
997
998/*
999 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
1000 */
1001#ifdef CONFIG_SCHED_DEBUG
1002# include <linux/static_key.h>
1003# define const_debug __read_mostly
1004#else
1005# define const_debug const
1006#endif
1007
1008extern const_debug unsigned int sysctl_sched_features;
1009
1010#define SCHED_FEAT(name, enabled) \
1011 __SCHED_FEAT_##name ,
1012
1013enum {
1014#include "features.h"
1015 __SCHED_FEAT_NR,
1016};
1017
1018#undef SCHED_FEAT
1019
1020#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
1021#define SCHED_FEAT(name, enabled) \
1022static __always_inline bool static_branch_##name(struct static_key *key) \
1023{ \
1024 return static_key_##enabled(key); \
1025}
1026
1027#include "features.h"
1028
1029#undef SCHED_FEAT
1030
1031extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
1032#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
1033#else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
1034#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
1035#endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
1036
1037extern struct static_key_false sched_numa_balancing;
1038extern struct static_key_false sched_schedstats;
1039
1040static inline u64 global_rt_period(void)
1041{
1042 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
1043}
1044
1045static inline u64 global_rt_runtime(void)
1046{
1047 if (sysctl_sched_rt_runtime < 0)
1048 return RUNTIME_INF;
1049
1050 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
1051}
1052
1053static inline int task_current(struct rq *rq, struct task_struct *p)
1054{
1055 return rq->curr == p;
1056}
1057
1058static inline int task_running(struct rq *rq, struct task_struct *p)
1059{
1060#ifdef CONFIG_SMP
1061 return p->on_cpu;
1062#else
1063 return task_current(rq, p);
1064#endif
1065}
1066
1067static inline int task_on_rq_queued(struct task_struct *p)
1068{
1069 return p->on_rq == TASK_ON_RQ_QUEUED;
1070}
1071
1072static inline int task_on_rq_migrating(struct task_struct *p)
1073{
1074 return p->on_rq == TASK_ON_RQ_MIGRATING;
1075}
1076
1077#ifndef prepare_arch_switch
1078# define prepare_arch_switch(next) do { } while (0)
1079#endif
1080#ifndef finish_arch_post_lock_switch
1081# define finish_arch_post_lock_switch() do { } while (0)
1082#endif
1083
1084static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1085{
1086#ifdef CONFIG_SMP
1087 /*
1088 * We can optimise this out completely for !SMP, because the
1089 * SMP rebalancing from interrupt is the only thing that cares
1090 * here.
1091 */
1092 next->on_cpu = 1;
1093#endif
1094}
1095
1096static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1097{
1098#ifdef CONFIG_SMP
1099 /*
1100 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1101 * We must ensure this doesn't happen until the switch is completely
1102 * finished.
1103 *
1104 * In particular, the load of prev->state in finish_task_switch() must
1105 * happen before this.
1106 *
1107 * Pairs with the smp_cond_acquire() in try_to_wake_up().
1108 */
1109 smp_store_release(&prev->on_cpu, 0);
1110#endif
1111#ifdef CONFIG_DEBUG_SPINLOCK
1112 /* this is a valid case when another task releases the spinlock */
1113 rq->lock.owner = current;
1114#endif
1115 /*
1116 * If we are tracking spinlock dependencies then we have to
1117 * fix up the runqueue lock - which gets 'carried over' from
1118 * prev into current:
1119 */
1120 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1121
1122 raw_spin_unlock_irq(&rq->lock);
1123}
1124
1125/*
1126 * wake flags
1127 */
1128#define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1129#define WF_FORK 0x02 /* child wakeup after fork */
1130#define WF_MIGRATED 0x4 /* internal use, task got migrated */
1131
1132/*
1133 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1134 * of tasks with abnormal "nice" values across CPUs the contribution that
1135 * each task makes to its run queue's load is weighted according to its
1136 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1137 * scaled version of the new time slice allocation that they receive on time
1138 * slice expiry etc.
1139 */
1140
1141#define WEIGHT_IDLEPRIO 3
1142#define WMULT_IDLEPRIO 1431655765
1143
1144extern const int sched_prio_to_weight[40];
1145extern const u32 sched_prio_to_wmult[40];
1146
1147/*
1148 * {de,en}queue flags:
1149 *
1150 * DEQUEUE_SLEEP - task is no longer runnable
1151 * ENQUEUE_WAKEUP - task just became runnable
1152 *
1153 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
1154 * are in a known state which allows modification. Such pairs
1155 * should preserve as much state as possible.
1156 *
1157 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
1158 * in the runqueue.
1159 *
1160 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
1161 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
1162 * ENQUEUE_WAKING - sched_class::task_waking was called
1163 *
1164 */
1165
1166#define DEQUEUE_SLEEP 0x01
1167#define DEQUEUE_SAVE 0x02 /* matches ENQUEUE_RESTORE */
1168#define DEQUEUE_MOVE 0x04 /* matches ENQUEUE_MOVE */
1169
1170#define ENQUEUE_WAKEUP 0x01
1171#define ENQUEUE_RESTORE 0x02
1172#define ENQUEUE_MOVE 0x04
1173
1174#define ENQUEUE_HEAD 0x08
1175#define ENQUEUE_REPLENISH 0x10
1176#ifdef CONFIG_SMP
1177#define ENQUEUE_WAKING 0x20
1178#else
1179#define ENQUEUE_WAKING 0x00
1180#endif
1181
1182#define RETRY_TASK ((void *)-1UL)
1183
1184struct sched_class {
1185 const struct sched_class *next;
1186
1187 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1188 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1189 void (*yield_task) (struct rq *rq);
1190 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1191
1192 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1193
1194 /*
1195 * It is the responsibility of the pick_next_task() method that will
1196 * return the next task to call put_prev_task() on the @prev task or
1197 * something equivalent.
1198 *
1199 * May return RETRY_TASK when it finds a higher prio class has runnable
1200 * tasks.
1201 */
1202 struct task_struct * (*pick_next_task) (struct rq *rq,
1203 struct task_struct *prev);
1204 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1205
1206#ifdef CONFIG_SMP
1207 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1208 void (*migrate_task_rq)(struct task_struct *p);
1209
1210 void (*task_waking) (struct task_struct *task);
1211 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1212
1213 void (*set_cpus_allowed)(struct task_struct *p,
1214 const struct cpumask *newmask);
1215
1216 void (*rq_online)(struct rq *rq);
1217 void (*rq_offline)(struct rq *rq);
1218#endif
1219
1220 void (*set_curr_task) (struct rq *rq);
1221 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1222 void (*task_fork) (struct task_struct *p);
1223 void (*task_dead) (struct task_struct *p);
1224
1225 /*
1226 * The switched_from() call is allowed to drop rq->lock, therefore we
1227 * cannot assume the switched_from/switched_to pair is serliazed by
1228 * rq->lock. They are however serialized by p->pi_lock.
1229 */
1230 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1231 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1232 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1233 int oldprio);
1234
1235 unsigned int (*get_rr_interval) (struct rq *rq,
1236 struct task_struct *task);
1237
1238 void (*update_curr) (struct rq *rq);
1239
1240#ifdef CONFIG_FAIR_GROUP_SCHED
1241 void (*task_move_group) (struct task_struct *p);
1242#endif
1243};
1244
1245static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1246{
1247 prev->sched_class->put_prev_task(rq, prev);
1248}
1249
1250#define sched_class_highest (&stop_sched_class)
1251#define for_each_class(class) \
1252 for (class = sched_class_highest; class; class = class->next)
1253
1254extern const struct sched_class stop_sched_class;
1255extern const struct sched_class dl_sched_class;
1256extern const struct sched_class rt_sched_class;
1257extern const struct sched_class fair_sched_class;
1258extern const struct sched_class idle_sched_class;
1259
1260
1261#ifdef CONFIG_SMP
1262
1263extern void update_group_capacity(struct sched_domain *sd, int cpu);
1264
1265extern void trigger_load_balance(struct rq *rq);
1266
1267extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask);
1268
1269#endif
1270
1271#ifdef CONFIG_CPU_IDLE
1272static inline void idle_set_state(struct rq *rq,
1273 struct cpuidle_state *idle_state)
1274{
1275 rq->idle_state = idle_state;
1276}
1277
1278static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1279{
1280 WARN_ON(!rcu_read_lock_held());
1281 return rq->idle_state;
1282}
1283#else
1284static inline void idle_set_state(struct rq *rq,
1285 struct cpuidle_state *idle_state)
1286{
1287}
1288
1289static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1290{
1291 return NULL;
1292}
1293#endif
1294
1295extern void sysrq_sched_debug_show(void);
1296extern void sched_init_granularity(void);
1297extern void update_max_interval(void);
1298
1299extern void init_sched_dl_class(void);
1300extern void init_sched_rt_class(void);
1301extern void init_sched_fair_class(void);
1302
1303extern void resched_curr(struct rq *rq);
1304extern void resched_cpu(int cpu);
1305
1306extern struct rt_bandwidth def_rt_bandwidth;
1307extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1308
1309extern struct dl_bandwidth def_dl_bandwidth;
1310extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1311extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1312
1313unsigned long to_ratio(u64 period, u64 runtime);
1314
1315extern void init_entity_runnable_average(struct sched_entity *se);
1316
1317#ifdef CONFIG_NO_HZ_FULL
1318extern bool sched_can_stop_tick(struct rq *rq);
1319
1320/*
1321 * Tick may be needed by tasks in the runqueue depending on their policy and
1322 * requirements. If tick is needed, lets send the target an IPI to kick it out of
1323 * nohz mode if necessary.
1324 */
1325static inline void sched_update_tick_dependency(struct rq *rq)
1326{
1327 int cpu;
1328
1329 if (!tick_nohz_full_enabled())
1330 return;
1331
1332 cpu = cpu_of(rq);
1333
1334 if (!tick_nohz_full_cpu(cpu))
1335 return;
1336
1337 if (sched_can_stop_tick(rq))
1338 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
1339 else
1340 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
1341}
1342#else
1343static inline void sched_update_tick_dependency(struct rq *rq) { }
1344#endif
1345
1346static inline void add_nr_running(struct rq *rq, unsigned count)
1347{
1348 unsigned prev_nr = rq->nr_running;
1349
1350 rq->nr_running = prev_nr + count;
1351
1352 if (prev_nr < 2 && rq->nr_running >= 2) {
1353#ifdef CONFIG_SMP
1354 if (!rq->rd->overload)
1355 rq->rd->overload = true;
1356#endif
1357 }
1358
1359 sched_update_tick_dependency(rq);
1360}
1361
1362static inline void sub_nr_running(struct rq *rq, unsigned count)
1363{
1364 rq->nr_running -= count;
1365 /* Check if we still need preemption */
1366 sched_update_tick_dependency(rq);
1367}
1368
1369static inline void rq_last_tick_reset(struct rq *rq)
1370{
1371#ifdef CONFIG_NO_HZ_FULL
1372 rq->last_sched_tick = jiffies;
1373#endif
1374}
1375
1376extern void update_rq_clock(struct rq *rq);
1377
1378extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1379extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1380
1381extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1382
1383extern const_debug unsigned int sysctl_sched_time_avg;
1384extern const_debug unsigned int sysctl_sched_nr_migrate;
1385extern const_debug unsigned int sysctl_sched_migration_cost;
1386
1387static inline u64 sched_avg_period(void)
1388{
1389 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1390}
1391
1392#ifdef CONFIG_SCHED_HRTICK
1393
1394/*
1395 * Use hrtick when:
1396 * - enabled by features
1397 * - hrtimer is actually high res
1398 */
1399static inline int hrtick_enabled(struct rq *rq)
1400{
1401 if (!sched_feat(HRTICK))
1402 return 0;
1403 if (!cpu_active(cpu_of(rq)))
1404 return 0;
1405 return hrtimer_is_hres_active(&rq->hrtick_timer);
1406}
1407
1408void hrtick_start(struct rq *rq, u64 delay);
1409
1410#else
1411
1412static inline int hrtick_enabled(struct rq *rq)
1413{
1414 return 0;
1415}
1416
1417#endif /* CONFIG_SCHED_HRTICK */
1418
1419#ifdef CONFIG_SMP
1420extern void sched_avg_update(struct rq *rq);
1421
1422#ifndef arch_scale_freq_capacity
1423static __always_inline
1424unsigned long arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
1425{
1426 return SCHED_CAPACITY_SCALE;
1427}
1428#endif
1429
1430#ifndef arch_scale_cpu_capacity
1431static __always_inline
1432unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
1433{
1434 if (sd && (sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
1435 return sd->smt_gain / sd->span_weight;
1436
1437 return SCHED_CAPACITY_SCALE;
1438}
1439#endif
1440
1441static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1442{
1443 rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq));
1444 sched_avg_update(rq);
1445}
1446#else
1447static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1448static inline void sched_avg_update(struct rq *rq) { }
1449#endif
1450
1451/*
1452 * __task_rq_lock - lock the rq @p resides on.
1453 */
1454static inline struct rq *__task_rq_lock(struct task_struct *p)
1455 __acquires(rq->lock)
1456{
1457 struct rq *rq;
1458
1459 lockdep_assert_held(&p->pi_lock);
1460
1461 for (;;) {
1462 rq = task_rq(p);
1463 raw_spin_lock(&rq->lock);
1464 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
1465 lockdep_pin_lock(&rq->lock);
1466 return rq;
1467 }
1468 raw_spin_unlock(&rq->lock);
1469
1470 while (unlikely(task_on_rq_migrating(p)))
1471 cpu_relax();
1472 }
1473}
1474
1475/*
1476 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1477 */
1478static inline struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1479 __acquires(p->pi_lock)
1480 __acquires(rq->lock)
1481{
1482 struct rq *rq;
1483
1484 for (;;) {
1485 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1486 rq = task_rq(p);
1487 raw_spin_lock(&rq->lock);
1488 /*
1489 * move_queued_task() task_rq_lock()
1490 *
1491 * ACQUIRE (rq->lock)
1492 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
1493 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
1494 * [S] ->cpu = new_cpu [L] task_rq()
1495 * [L] ->on_rq
1496 * RELEASE (rq->lock)
1497 *
1498 * If we observe the old cpu in task_rq_lock, the acquire of
1499 * the old rq->lock will fully serialize against the stores.
1500 *
1501 * If we observe the new cpu in task_rq_lock, the acquire will
1502 * pair with the WMB to ensure we must then also see migrating.
1503 */
1504 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
1505 lockdep_pin_lock(&rq->lock);
1506 return rq;
1507 }
1508 raw_spin_unlock(&rq->lock);
1509 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1510
1511 while (unlikely(task_on_rq_migrating(p)))
1512 cpu_relax();
1513 }
1514}
1515
1516static inline void __task_rq_unlock(struct rq *rq)
1517 __releases(rq->lock)
1518{
1519 lockdep_unpin_lock(&rq->lock);
1520 raw_spin_unlock(&rq->lock);
1521}
1522
1523static inline void
1524task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1525 __releases(rq->lock)
1526 __releases(p->pi_lock)
1527{
1528 lockdep_unpin_lock(&rq->lock);
1529 raw_spin_unlock(&rq->lock);
1530 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1531}
1532
1533#ifdef CONFIG_SMP
1534#ifdef CONFIG_PREEMPT
1535
1536static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1537
1538/*
1539 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1540 * way at the expense of forcing extra atomic operations in all
1541 * invocations. This assures that the double_lock is acquired using the
1542 * same underlying policy as the spinlock_t on this architecture, which
1543 * reduces latency compared to the unfair variant below. However, it
1544 * also adds more overhead and therefore may reduce throughput.
1545 */
1546static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1547 __releases(this_rq->lock)
1548 __acquires(busiest->lock)
1549 __acquires(this_rq->lock)
1550{
1551 raw_spin_unlock(&this_rq->lock);
1552 double_rq_lock(this_rq, busiest);
1553
1554 return 1;
1555}
1556
1557#else
1558/*
1559 * Unfair double_lock_balance: Optimizes throughput at the expense of
1560 * latency by eliminating extra atomic operations when the locks are
1561 * already in proper order on entry. This favors lower cpu-ids and will
1562 * grant the double lock to lower cpus over higher ids under contention,
1563 * regardless of entry order into the function.
1564 */
1565static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1566 __releases(this_rq->lock)
1567 __acquires(busiest->lock)
1568 __acquires(this_rq->lock)
1569{
1570 int ret = 0;
1571
1572 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1573 if (busiest < this_rq) {
1574 raw_spin_unlock(&this_rq->lock);
1575 raw_spin_lock(&busiest->lock);
1576 raw_spin_lock_nested(&this_rq->lock,
1577 SINGLE_DEPTH_NESTING);
1578 ret = 1;
1579 } else
1580 raw_spin_lock_nested(&busiest->lock,
1581 SINGLE_DEPTH_NESTING);
1582 }
1583 return ret;
1584}
1585
1586#endif /* CONFIG_PREEMPT */
1587
1588/*
1589 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1590 */
1591static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592{
1593 if (unlikely(!irqs_disabled())) {
1594 /* printk() doesn't work good under rq->lock */
1595 raw_spin_unlock(&this_rq->lock);
1596 BUG_ON(1);
1597 }
1598
1599 return _double_lock_balance(this_rq, busiest);
1600}
1601
1602static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1603 __releases(busiest->lock)
1604{
1605 raw_spin_unlock(&busiest->lock);
1606 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1607}
1608
1609static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1610{
1611 if (l1 > l2)
1612 swap(l1, l2);
1613
1614 spin_lock(l1);
1615 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1616}
1617
1618static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1619{
1620 if (l1 > l2)
1621 swap(l1, l2);
1622
1623 spin_lock_irq(l1);
1624 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1625}
1626
1627static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1628{
1629 if (l1 > l2)
1630 swap(l1, l2);
1631
1632 raw_spin_lock(l1);
1633 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1634}
1635
1636/*
1637 * double_rq_lock - safely lock two runqueues
1638 *
1639 * Note this does not disable interrupts like task_rq_lock,
1640 * you need to do so manually before calling.
1641 */
1642static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1643 __acquires(rq1->lock)
1644 __acquires(rq2->lock)
1645{
1646 BUG_ON(!irqs_disabled());
1647 if (rq1 == rq2) {
1648 raw_spin_lock(&rq1->lock);
1649 __acquire(rq2->lock); /* Fake it out ;) */
1650 } else {
1651 if (rq1 < rq2) {
1652 raw_spin_lock(&rq1->lock);
1653 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1654 } else {
1655 raw_spin_lock(&rq2->lock);
1656 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1657 }
1658 }
1659}
1660
1661/*
1662 * double_rq_unlock - safely unlock two runqueues
1663 *
1664 * Note this does not restore interrupts like task_rq_unlock,
1665 * you need to do so manually after calling.
1666 */
1667static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1668 __releases(rq1->lock)
1669 __releases(rq2->lock)
1670{
1671 raw_spin_unlock(&rq1->lock);
1672 if (rq1 != rq2)
1673 raw_spin_unlock(&rq2->lock);
1674 else
1675 __release(rq2->lock);
1676}
1677
1678#else /* CONFIG_SMP */
1679
1680/*
1681 * double_rq_lock - safely lock two runqueues
1682 *
1683 * Note this does not disable interrupts like task_rq_lock,
1684 * you need to do so manually before calling.
1685 */
1686static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1687 __acquires(rq1->lock)
1688 __acquires(rq2->lock)
1689{
1690 BUG_ON(!irqs_disabled());
1691 BUG_ON(rq1 != rq2);
1692 raw_spin_lock(&rq1->lock);
1693 __acquire(rq2->lock); /* Fake it out ;) */
1694}
1695
1696/*
1697 * double_rq_unlock - safely unlock two runqueues
1698 *
1699 * Note this does not restore interrupts like task_rq_unlock,
1700 * you need to do so manually after calling.
1701 */
1702static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1703 __releases(rq1->lock)
1704 __releases(rq2->lock)
1705{
1706 BUG_ON(rq1 != rq2);
1707 raw_spin_unlock(&rq1->lock);
1708 __release(rq2->lock);
1709}
1710
1711#endif
1712
1713extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1714extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1715
1716#ifdef CONFIG_SCHED_DEBUG
1717extern void print_cfs_stats(struct seq_file *m, int cpu);
1718extern void print_rt_stats(struct seq_file *m, int cpu);
1719extern void print_dl_stats(struct seq_file *m, int cpu);
1720extern void
1721print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
1722
1723#ifdef CONFIG_NUMA_BALANCING
1724extern void
1725show_numa_stats(struct task_struct *p, struct seq_file *m);
1726extern void
1727print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
1728 unsigned long tpf, unsigned long gsf, unsigned long gpf);
1729#endif /* CONFIG_NUMA_BALANCING */
1730#endif /* CONFIG_SCHED_DEBUG */
1731
1732extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1733extern void init_rt_rq(struct rt_rq *rt_rq);
1734extern void init_dl_rq(struct dl_rq *dl_rq);
1735
1736extern void cfs_bandwidth_usage_inc(void);
1737extern void cfs_bandwidth_usage_dec(void);
1738
1739#ifdef CONFIG_NO_HZ_COMMON
1740enum rq_nohz_flag_bits {
1741 NOHZ_TICK_STOPPED,
1742 NOHZ_BALANCE_KICK,
1743};
1744
1745#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1746#endif
1747
1748#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1749
1750DECLARE_PER_CPU(u64, cpu_hardirq_time);
1751DECLARE_PER_CPU(u64, cpu_softirq_time);
1752
1753#ifndef CONFIG_64BIT
1754DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1755
1756static inline void irq_time_write_begin(void)
1757{
1758 __this_cpu_inc(irq_time_seq.sequence);
1759 smp_wmb();
1760}
1761
1762static inline void irq_time_write_end(void)
1763{
1764 smp_wmb();
1765 __this_cpu_inc(irq_time_seq.sequence);
1766}
1767
1768static inline u64 irq_time_read(int cpu)
1769{
1770 u64 irq_time;
1771 unsigned seq;
1772
1773 do {
1774 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1775 irq_time = per_cpu(cpu_softirq_time, cpu) +
1776 per_cpu(cpu_hardirq_time, cpu);
1777 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1778
1779 return irq_time;
1780}
1781#else /* CONFIG_64BIT */
1782static inline void irq_time_write_begin(void)
1783{
1784}
1785
1786static inline void irq_time_write_end(void)
1787{
1788}
1789
1790static inline u64 irq_time_read(int cpu)
1791{
1792 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1793}
1794#endif /* CONFIG_64BIT */
1795#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1796
1797#ifdef CONFIG_CPU_FREQ
1798DECLARE_PER_CPU(struct update_util_data *, cpufreq_update_util_data);
1799
1800/**
1801 * cpufreq_update_util - Take a note about CPU utilization changes.
1802 * @time: Current time.
1803 * @util: Current utilization.
1804 * @max: Utilization ceiling.
1805 *
1806 * This function is called by the scheduler on every invocation of
1807 * update_load_avg() on the CPU whose utilization is being updated.
1808 *
1809 * It can only be called from RCU-sched read-side critical sections.
1810 */
1811static inline void cpufreq_update_util(u64 time, unsigned long util, unsigned long max)
1812{
1813 struct update_util_data *data;
1814
1815 data = rcu_dereference_sched(*this_cpu_ptr(&cpufreq_update_util_data));
1816 if (data)
1817 data->func(data, time, util, max);
1818}
1819
1820/**
1821 * cpufreq_trigger_update - Trigger CPU performance state evaluation if needed.
1822 * @time: Current time.
1823 *
1824 * The way cpufreq is currently arranged requires it to evaluate the CPU
1825 * performance state (frequency/voltage) on a regular basis to prevent it from
1826 * being stuck in a completely inadequate performance level for too long.
1827 * That is not guaranteed to happen if the updates are only triggered from CFS,
1828 * though, because they may not be coming in if RT or deadline tasks are active
1829 * all the time (or there are RT and DL tasks only).
1830 *
1831 * As a workaround for that issue, this function is called by the RT and DL
1832 * sched classes to trigger extra cpufreq updates to prevent it from stalling,
1833 * but that really is a band-aid. Going forward it should be replaced with
1834 * solutions targeted more specifically at RT and DL tasks.
1835 */
1836static inline void cpufreq_trigger_update(u64 time)
1837{
1838 cpufreq_update_util(time, ULONG_MAX, 0);
1839}
1840#else
1841static inline void cpufreq_update_util(u64 time, unsigned long util, unsigned long max) {}
1842static inline void cpufreq_trigger_update(u64 time) {}
1843#endif /* CONFIG_CPU_FREQ */
1844
1845static inline void account_reset_rq(struct rq *rq)
1846{
1847#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1848 rq->prev_irq_time = 0;
1849#endif
1850#ifdef CONFIG_PARAVIRT
1851 rq->prev_steal_time = 0;
1852#endif
1853#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1854 rq->prev_steal_time_rq = 0;
1855#endif
1856}
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/blkdev.h>
40#include <linux/compat.h>
41#include <linux/context_tracking.h>
42#include <linux/cpufreq.h>
43#include <linux/cpuidle.h>
44#include <linux/cpuset.h>
45#include <linux/ctype.h>
46#include <linux/debugfs.h>
47#include <linux/delayacct.h>
48#include <linux/energy_model.h>
49#include <linux/init_task.h>
50#include <linux/kprobes.h>
51#include <linux/kthread.h>
52#include <linux/membarrier.h>
53#include <linux/migrate.h>
54#include <linux/mmu_context.h>
55#include <linux/nmi.h>
56#include <linux/proc_fs.h>
57#include <linux/prefetch.h>
58#include <linux/profile.h>
59#include <linux/psi.h>
60#include <linux/rcupdate_wait.h>
61#include <linux/security.h>
62#include <linux/stop_machine.h>
63#include <linux/suspend.h>
64#include <linux/swait.h>
65#include <linux/syscalls.h>
66#include <linux/task_work.h>
67#include <linux/tsacct_kern.h>
68
69#include <asm/tlb.h>
70#include <asm-generic/vmlinux.lds.h>
71
72#ifdef CONFIG_PARAVIRT
73# include <asm/paravirt.h>
74#endif
75
76#include "cpupri.h"
77#include "cpudeadline.h"
78
79#include <trace/events/sched.h>
80
81#ifdef CONFIG_SCHED_DEBUG
82# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
83#else
84# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
85#endif
86
87struct rq;
88struct cpuidle_state;
89
90/* task_struct::on_rq states: */
91#define TASK_ON_RQ_QUEUED 1
92#define TASK_ON_RQ_MIGRATING 2
93
94extern __read_mostly int scheduler_running;
95
96extern unsigned long calc_load_update;
97extern atomic_long_t calc_load_tasks;
98
99extern void calc_global_load_tick(struct rq *this_rq);
100extern long calc_load_fold_active(struct rq *this_rq, long adjust);
101
102extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
103/*
104 * Helpers for converting nanosecond timing to jiffy resolution
105 */
106#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107
108/*
109 * Increase resolution of nice-level calculations for 64-bit architectures.
110 * The extra resolution improves shares distribution and load balancing of
111 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
112 * hierarchies, especially on larger systems. This is not a user-visible change
113 * and does not change the user-interface for setting shares/weights.
114 *
115 * We increase resolution only if we have enough bits to allow this increased
116 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
117 * are pretty high and the returns do not justify the increased costs.
118 *
119 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
120 * increase coverage and consistency always enable it on 64-bit platforms.
121 */
122#ifdef CONFIG_64BIT
123# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
124# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
125# define scale_load_down(w) \
126({ \
127 unsigned long __w = (w); \
128 if (__w) \
129 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
130 __w; \
131})
132#else
133# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
134# define scale_load(w) (w)
135# define scale_load_down(w) (w)
136#endif
137
138/*
139 * Task weight (visible to users) and its load (invisible to users) have
140 * independent resolution, but they should be well calibrated. We use
141 * scale_load() and scale_load_down(w) to convert between them. The
142 * following must be true:
143 *
144 * scale_load(sched_prio_to_weight[USER_PRIO(NICE_TO_PRIO(0))]) == NICE_0_LOAD
145 *
146 */
147#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
148
149/*
150 * Single value that decides SCHED_DEADLINE internal math precision.
151 * 10 -> just above 1us
152 * 9 -> just above 0.5us
153 */
154#define DL_SCALE 10
155
156/*
157 * Single value that denotes runtime == period, ie unlimited time.
158 */
159#define RUNTIME_INF ((u64)~0ULL)
160
161static inline int idle_policy(int policy)
162{
163 return policy == SCHED_IDLE;
164}
165static inline int fair_policy(int policy)
166{
167 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
168}
169
170static inline int rt_policy(int policy)
171{
172 return policy == SCHED_FIFO || policy == SCHED_RR;
173}
174
175static inline int dl_policy(int policy)
176{
177 return policy == SCHED_DEADLINE;
178}
179static inline bool valid_policy(int policy)
180{
181 return idle_policy(policy) || fair_policy(policy) ||
182 rt_policy(policy) || dl_policy(policy);
183}
184
185static inline int task_has_idle_policy(struct task_struct *p)
186{
187 return idle_policy(p->policy);
188}
189
190static inline int task_has_rt_policy(struct task_struct *p)
191{
192 return rt_policy(p->policy);
193}
194
195static inline int task_has_dl_policy(struct task_struct *p)
196{
197 return dl_policy(p->policy);
198}
199
200#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
201
202static inline void update_avg(u64 *avg, u64 sample)
203{
204 s64 diff = sample - *avg;
205 *avg += diff / 8;
206}
207
208/*
209 * !! For sched_setattr_nocheck() (kernel) only !!
210 *
211 * This is actually gross. :(
212 *
213 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
214 * tasks, but still be able to sleep. We need this on platforms that cannot
215 * atomically change clock frequency. Remove once fast switching will be
216 * available on such platforms.
217 *
218 * SUGOV stands for SchedUtil GOVernor.
219 */
220#define SCHED_FLAG_SUGOV 0x10000000
221
222static inline bool dl_entity_is_special(struct sched_dl_entity *dl_se)
223{
224#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
225 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
226#else
227 return false;
228#endif
229}
230
231/*
232 * Tells if entity @a should preempt entity @b.
233 */
234static inline bool
235dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
236{
237 return dl_entity_is_special(a) ||
238 dl_time_before(a->deadline, b->deadline);
239}
240
241/*
242 * This is the priority-queue data structure of the RT scheduling class:
243 */
244struct rt_prio_array {
245 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
246 struct list_head queue[MAX_RT_PRIO];
247};
248
249struct rt_bandwidth {
250 /* nests inside the rq lock: */
251 raw_spinlock_t rt_runtime_lock;
252 ktime_t rt_period;
253 u64 rt_runtime;
254 struct hrtimer rt_period_timer;
255 unsigned int rt_period_active;
256};
257
258void __dl_clear_params(struct task_struct *p);
259
260/*
261 * To keep the bandwidth of -deadline tasks and groups under control
262 * we need some place where:
263 * - store the maximum -deadline bandwidth of the system (the group);
264 * - cache the fraction of that bandwidth that is currently allocated.
265 *
266 * This is all done in the data structure below. It is similar to the
267 * one used for RT-throttling (rt_bandwidth), with the main difference
268 * that, since here we are only interested in admission control, we
269 * do not decrease any runtime while the group "executes", neither we
270 * need a timer to replenish it.
271 *
272 * With respect to SMP, the bandwidth is given on a per-CPU basis,
273 * meaning that:
274 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
275 * - dl_total_bw array contains, in the i-eth element, the currently
276 * allocated bandwidth on the i-eth CPU.
277 * Moreover, groups consume bandwidth on each CPU, while tasks only
278 * consume bandwidth on the CPU they're running on.
279 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
280 * that will be shown the next time the proc or cgroup controls will
281 * be red. It on its turn can be changed by writing on its own
282 * control.
283 */
284struct dl_bandwidth {
285 raw_spinlock_t dl_runtime_lock;
286 u64 dl_runtime;
287 u64 dl_period;
288};
289
290static inline int dl_bandwidth_enabled(void)
291{
292 return sysctl_sched_rt_runtime >= 0;
293}
294
295struct dl_bw {
296 raw_spinlock_t lock;
297 u64 bw;
298 u64 total_bw;
299};
300
301static inline void __dl_update(struct dl_bw *dl_b, s64 bw);
302
303static inline
304void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
305{
306 dl_b->total_bw -= tsk_bw;
307 __dl_update(dl_b, (s32)tsk_bw / cpus);
308}
309
310static inline
311void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
312{
313 dl_b->total_bw += tsk_bw;
314 __dl_update(dl_b, -((s32)tsk_bw / cpus));
315}
316
317static inline bool __dl_overflow(struct dl_bw *dl_b, unsigned long cap,
318 u64 old_bw, u64 new_bw)
319{
320 return dl_b->bw != -1 &&
321 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
322}
323
324/*
325 * Verify the fitness of task @p to run on @cpu taking into account the
326 * CPU original capacity and the runtime/deadline ratio of the task.
327 *
328 * The function will return true if the CPU original capacity of the
329 * @cpu scaled by SCHED_CAPACITY_SCALE >= runtime/deadline ratio of the
330 * task and false otherwise.
331 */
332static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
333{
334 unsigned long cap = arch_scale_cpu_capacity(cpu);
335
336 return cap_scale(p->dl.dl_deadline, cap) >= p->dl.dl_runtime;
337}
338
339extern void init_dl_bw(struct dl_bw *dl_b);
340extern int sched_dl_global_validate(void);
341extern void sched_dl_do_global(void);
342extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
343extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
344extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
345extern bool __checkparam_dl(const struct sched_attr *attr);
346extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
347extern int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed);
348extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
349extern bool dl_cpu_busy(unsigned int cpu);
350
351#ifdef CONFIG_CGROUP_SCHED
352
353#include <linux/cgroup.h>
354#include <linux/psi.h>
355
356struct cfs_rq;
357struct rt_rq;
358
359extern struct list_head task_groups;
360
361struct cfs_bandwidth {
362#ifdef CONFIG_CFS_BANDWIDTH
363 raw_spinlock_t lock;
364 ktime_t period;
365 u64 quota;
366 u64 runtime;
367 s64 hierarchical_quota;
368
369 u8 idle;
370 u8 period_active;
371 u8 slack_started;
372 struct hrtimer period_timer;
373 struct hrtimer slack_timer;
374 struct list_head throttled_cfs_rq;
375
376 /* Statistics: */
377 int nr_periods;
378 int nr_throttled;
379 u64 throttled_time;
380#endif
381};
382
383/* Task group related information */
384struct task_group {
385 struct cgroup_subsys_state css;
386
387#ifdef CONFIG_FAIR_GROUP_SCHED
388 /* schedulable entities of this group on each CPU */
389 struct sched_entity **se;
390 /* runqueue "owned" by this group on each CPU */
391 struct cfs_rq **cfs_rq;
392 unsigned long shares;
393
394#ifdef CONFIG_SMP
395 /*
396 * load_avg can be heavily contended at clock tick time, so put
397 * it in its own cacheline separated from the fields above which
398 * will also be accessed at each tick.
399 */
400 atomic_long_t load_avg ____cacheline_aligned;
401#endif
402#endif
403
404#ifdef CONFIG_RT_GROUP_SCHED
405 struct sched_rt_entity **rt_se;
406 struct rt_rq **rt_rq;
407
408 struct rt_bandwidth rt_bandwidth;
409#endif
410
411 struct rcu_head rcu;
412 struct list_head list;
413
414 struct task_group *parent;
415 struct list_head siblings;
416 struct list_head children;
417
418#ifdef CONFIG_SCHED_AUTOGROUP
419 struct autogroup *autogroup;
420#endif
421
422 struct cfs_bandwidth cfs_bandwidth;
423
424#ifdef CONFIG_UCLAMP_TASK_GROUP
425 /* The two decimal precision [%] value requested from user-space */
426 unsigned int uclamp_pct[UCLAMP_CNT];
427 /* Clamp values requested for a task group */
428 struct uclamp_se uclamp_req[UCLAMP_CNT];
429 /* Effective clamp values used for a task group */
430 struct uclamp_se uclamp[UCLAMP_CNT];
431#endif
432
433};
434
435#ifdef CONFIG_FAIR_GROUP_SCHED
436#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
437
438/*
439 * A weight of 0 or 1 can cause arithmetics problems.
440 * A weight of a cfs_rq is the sum of weights of which entities
441 * are queued on this cfs_rq, so a weight of a entity should not be
442 * too large, so as the shares value of a task group.
443 * (The default weight is 1024 - so there's no practical
444 * limitation from this.)
445 */
446#define MIN_SHARES (1UL << 1)
447#define MAX_SHARES (1UL << 18)
448#endif
449
450typedef int (*tg_visitor)(struct task_group *, void *);
451
452extern int walk_tg_tree_from(struct task_group *from,
453 tg_visitor down, tg_visitor up, void *data);
454
455/*
456 * Iterate the full tree, calling @down when first entering a node and @up when
457 * leaving it for the final time.
458 *
459 * Caller must hold rcu_lock or sufficient equivalent.
460 */
461static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
462{
463 return walk_tg_tree_from(&root_task_group, down, up, data);
464}
465
466extern int tg_nop(struct task_group *tg, void *data);
467
468extern void free_fair_sched_group(struct task_group *tg);
469extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
470extern void online_fair_sched_group(struct task_group *tg);
471extern void unregister_fair_sched_group(struct task_group *tg);
472extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
473 struct sched_entity *se, int cpu,
474 struct sched_entity *parent);
475extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
476
477extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
478extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
479extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
480
481extern void free_rt_sched_group(struct task_group *tg);
482extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
483extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
484 struct sched_rt_entity *rt_se, int cpu,
485 struct sched_rt_entity *parent);
486extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
487extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
488extern long sched_group_rt_runtime(struct task_group *tg);
489extern long sched_group_rt_period(struct task_group *tg);
490extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
491
492extern struct task_group *sched_create_group(struct task_group *parent);
493extern void sched_online_group(struct task_group *tg,
494 struct task_group *parent);
495extern void sched_destroy_group(struct task_group *tg);
496extern void sched_offline_group(struct task_group *tg);
497
498extern void sched_move_task(struct task_struct *tsk);
499
500#ifdef CONFIG_FAIR_GROUP_SCHED
501extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
502
503#ifdef CONFIG_SMP
504extern void set_task_rq_fair(struct sched_entity *se,
505 struct cfs_rq *prev, struct cfs_rq *next);
506#else /* !CONFIG_SMP */
507static inline void set_task_rq_fair(struct sched_entity *se,
508 struct cfs_rq *prev, struct cfs_rq *next) { }
509#endif /* CONFIG_SMP */
510#endif /* CONFIG_FAIR_GROUP_SCHED */
511
512#else /* CONFIG_CGROUP_SCHED */
513
514struct cfs_bandwidth { };
515
516#endif /* CONFIG_CGROUP_SCHED */
517
518/* CFS-related fields in a runqueue */
519struct cfs_rq {
520 struct load_weight load;
521 unsigned int nr_running;
522 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
523 unsigned int idle_h_nr_running; /* SCHED_IDLE */
524
525 u64 exec_clock;
526 u64 min_vruntime;
527#ifndef CONFIG_64BIT
528 u64 min_vruntime_copy;
529#endif
530
531 struct rb_root_cached tasks_timeline;
532
533 /*
534 * 'curr' points to currently running entity on this cfs_rq.
535 * It is set to NULL otherwise (i.e when none are currently running).
536 */
537 struct sched_entity *curr;
538 struct sched_entity *next;
539 struct sched_entity *last;
540 struct sched_entity *skip;
541
542#ifdef CONFIG_SCHED_DEBUG
543 unsigned int nr_spread_over;
544#endif
545
546#ifdef CONFIG_SMP
547 /*
548 * CFS load tracking
549 */
550 struct sched_avg avg;
551#ifndef CONFIG_64BIT
552 u64 load_last_update_time_copy;
553#endif
554 struct {
555 raw_spinlock_t lock ____cacheline_aligned;
556 int nr;
557 unsigned long load_avg;
558 unsigned long util_avg;
559 unsigned long runnable_avg;
560 } removed;
561
562#ifdef CONFIG_FAIR_GROUP_SCHED
563 unsigned long tg_load_avg_contrib;
564 long propagate;
565 long prop_runnable_sum;
566
567 /*
568 * h_load = weight * f(tg)
569 *
570 * Where f(tg) is the recursive weight fraction assigned to
571 * this group.
572 */
573 unsigned long h_load;
574 u64 last_h_load_update;
575 struct sched_entity *h_load_next;
576#endif /* CONFIG_FAIR_GROUP_SCHED */
577#endif /* CONFIG_SMP */
578
579#ifdef CONFIG_FAIR_GROUP_SCHED
580 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
581
582 /*
583 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
584 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
585 * (like users, containers etc.)
586 *
587 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
588 * This list is used during load balance.
589 */
590 int on_list;
591 struct list_head leaf_cfs_rq_list;
592 struct task_group *tg; /* group that "owns" this runqueue */
593
594#ifdef CONFIG_CFS_BANDWIDTH
595 int runtime_enabled;
596 s64 runtime_remaining;
597
598 u64 throttled_clock;
599 u64 throttled_clock_task;
600 u64 throttled_clock_task_time;
601 int throttled;
602 int throttle_count;
603 struct list_head throttled_list;
604#endif /* CONFIG_CFS_BANDWIDTH */
605#endif /* CONFIG_FAIR_GROUP_SCHED */
606};
607
608static inline int rt_bandwidth_enabled(void)
609{
610 return sysctl_sched_rt_runtime >= 0;
611}
612
613/* RT IPI pull logic requires IRQ_WORK */
614#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
615# define HAVE_RT_PUSH_IPI
616#endif
617
618/* Real-Time classes' related field in a runqueue: */
619struct rt_rq {
620 struct rt_prio_array active;
621 unsigned int rt_nr_running;
622 unsigned int rr_nr_running;
623#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
624 struct {
625 int curr; /* highest queued rt task prio */
626#ifdef CONFIG_SMP
627 int next; /* next highest */
628#endif
629 } highest_prio;
630#endif
631#ifdef CONFIG_SMP
632 unsigned long rt_nr_migratory;
633 unsigned long rt_nr_total;
634 int overloaded;
635 struct plist_head pushable_tasks;
636
637#endif /* CONFIG_SMP */
638 int rt_queued;
639
640 int rt_throttled;
641 u64 rt_time;
642 u64 rt_runtime;
643 /* Nests inside the rq lock: */
644 raw_spinlock_t rt_runtime_lock;
645
646#ifdef CONFIG_RT_GROUP_SCHED
647 unsigned long rt_nr_boosted;
648
649 struct rq *rq;
650 struct task_group *tg;
651#endif
652};
653
654static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
655{
656 return rt_rq->rt_queued && rt_rq->rt_nr_running;
657}
658
659/* Deadline class' related fields in a runqueue */
660struct dl_rq {
661 /* runqueue is an rbtree, ordered by deadline */
662 struct rb_root_cached root;
663
664 unsigned long dl_nr_running;
665
666#ifdef CONFIG_SMP
667 /*
668 * Deadline values of the currently executing and the
669 * earliest ready task on this rq. Caching these facilitates
670 * the decision whether or not a ready but not running task
671 * should migrate somewhere else.
672 */
673 struct {
674 u64 curr;
675 u64 next;
676 } earliest_dl;
677
678 unsigned long dl_nr_migratory;
679 int overloaded;
680
681 /*
682 * Tasks on this rq that can be pushed away. They are kept in
683 * an rb-tree, ordered by tasks' deadlines, with caching
684 * of the leftmost (earliest deadline) element.
685 */
686 struct rb_root_cached pushable_dl_tasks_root;
687#else
688 struct dl_bw dl_bw;
689#endif
690 /*
691 * "Active utilization" for this runqueue: increased when a
692 * task wakes up (becomes TASK_RUNNING) and decreased when a
693 * task blocks
694 */
695 u64 running_bw;
696
697 /*
698 * Utilization of the tasks "assigned" to this runqueue (including
699 * the tasks that are in runqueue and the tasks that executed on this
700 * CPU and blocked). Increased when a task moves to this runqueue, and
701 * decreased when the task moves away (migrates, changes scheduling
702 * policy, or terminates).
703 * This is needed to compute the "inactive utilization" for the
704 * runqueue (inactive utilization = this_bw - running_bw).
705 */
706 u64 this_bw;
707 u64 extra_bw;
708
709 /*
710 * Inverse of the fraction of CPU utilization that can be reclaimed
711 * by the GRUB algorithm.
712 */
713 u64 bw_ratio;
714};
715
716#ifdef CONFIG_FAIR_GROUP_SCHED
717/* An entity is a task if it doesn't "own" a runqueue */
718#define entity_is_task(se) (!se->my_q)
719
720static inline void se_update_runnable(struct sched_entity *se)
721{
722 if (!entity_is_task(se))
723 se->runnable_weight = se->my_q->h_nr_running;
724}
725
726static inline long se_runnable(struct sched_entity *se)
727{
728 if (entity_is_task(se))
729 return !!se->on_rq;
730 else
731 return se->runnable_weight;
732}
733
734#else
735#define entity_is_task(se) 1
736
737static inline void se_update_runnable(struct sched_entity *se) {}
738
739static inline long se_runnable(struct sched_entity *se)
740{
741 return !!se->on_rq;
742}
743#endif
744
745#ifdef CONFIG_SMP
746/*
747 * XXX we want to get rid of these helpers and use the full load resolution.
748 */
749static inline long se_weight(struct sched_entity *se)
750{
751 return scale_load_down(se->load.weight);
752}
753
754
755static inline bool sched_asym_prefer(int a, int b)
756{
757 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
758}
759
760struct perf_domain {
761 struct em_perf_domain *em_pd;
762 struct perf_domain *next;
763 struct rcu_head rcu;
764};
765
766/* Scheduling group status flags */
767#define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
768#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
769
770/*
771 * We add the notion of a root-domain which will be used to define per-domain
772 * variables. Each exclusive cpuset essentially defines an island domain by
773 * fully partitioning the member CPUs from any other cpuset. Whenever a new
774 * exclusive cpuset is created, we also create and attach a new root-domain
775 * object.
776 *
777 */
778struct root_domain {
779 atomic_t refcount;
780 atomic_t rto_count;
781 struct rcu_head rcu;
782 cpumask_var_t span;
783 cpumask_var_t online;
784
785 /*
786 * Indicate pullable load on at least one CPU, e.g:
787 * - More than one runnable task
788 * - Running task is misfit
789 */
790 int overload;
791
792 /* Indicate one or more cpus over-utilized (tipping point) */
793 int overutilized;
794
795 /*
796 * The bit corresponding to a CPU gets set here if such CPU has more
797 * than one runnable -deadline task (as it is below for RT tasks).
798 */
799 cpumask_var_t dlo_mask;
800 atomic_t dlo_count;
801 struct dl_bw dl_bw;
802 struct cpudl cpudl;
803
804#ifdef HAVE_RT_PUSH_IPI
805 /*
806 * For IPI pull requests, loop across the rto_mask.
807 */
808 struct irq_work rto_push_work;
809 raw_spinlock_t rto_lock;
810 /* These are only updated and read within rto_lock */
811 int rto_loop;
812 int rto_cpu;
813 /* These atomics are updated outside of a lock */
814 atomic_t rto_loop_next;
815 atomic_t rto_loop_start;
816#endif
817 /*
818 * The "RT overload" flag: it gets set if a CPU has more than
819 * one runnable RT task.
820 */
821 cpumask_var_t rto_mask;
822 struct cpupri cpupri;
823
824 unsigned long max_cpu_capacity;
825
826 /*
827 * NULL-terminated list of performance domains intersecting with the
828 * CPUs of the rd. Protected by RCU.
829 */
830 struct perf_domain __rcu *pd;
831};
832
833extern void init_defrootdomain(void);
834extern int sched_init_domains(const struct cpumask *cpu_map);
835extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
836extern void sched_get_rd(struct root_domain *rd);
837extern void sched_put_rd(struct root_domain *rd);
838
839#ifdef HAVE_RT_PUSH_IPI
840extern void rto_push_irq_work_func(struct irq_work *work);
841#endif
842#endif /* CONFIG_SMP */
843
844#ifdef CONFIG_UCLAMP_TASK
845/*
846 * struct uclamp_bucket - Utilization clamp bucket
847 * @value: utilization clamp value for tasks on this clamp bucket
848 * @tasks: number of RUNNABLE tasks on this clamp bucket
849 *
850 * Keep track of how many tasks are RUNNABLE for a given utilization
851 * clamp value.
852 */
853struct uclamp_bucket {
854 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
855 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
856};
857
858/*
859 * struct uclamp_rq - rq's utilization clamp
860 * @value: currently active clamp values for a rq
861 * @bucket: utilization clamp buckets affecting a rq
862 *
863 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
864 * A clamp value is affecting a rq when there is at least one task RUNNABLE
865 * (or actually running) with that value.
866 *
867 * There are up to UCLAMP_CNT possible different clamp values, currently there
868 * are only two: minimum utilization and maximum utilization.
869 *
870 * All utilization clamping values are MAX aggregated, since:
871 * - for util_min: we want to run the CPU at least at the max of the minimum
872 * utilization required by its currently RUNNABLE tasks.
873 * - for util_max: we want to allow the CPU to run up to the max of the
874 * maximum utilization allowed by its currently RUNNABLE tasks.
875 *
876 * Since on each system we expect only a limited number of different
877 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
878 * the metrics required to compute all the per-rq utilization clamp values.
879 */
880struct uclamp_rq {
881 unsigned int value;
882 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
883};
884
885DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
886#endif /* CONFIG_UCLAMP_TASK */
887
888/*
889 * This is the main, per-CPU runqueue data structure.
890 *
891 * Locking rule: those places that want to lock multiple runqueues
892 * (such as the load balancing or the thread migration code), lock
893 * acquire operations must be ordered by ascending &runqueue.
894 */
895struct rq {
896 /* runqueue lock: */
897 raw_spinlock_t lock;
898
899 /*
900 * nr_running and cpu_load should be in the same cacheline because
901 * remote CPUs use both these fields when doing load calculation.
902 */
903 unsigned int nr_running;
904#ifdef CONFIG_NUMA_BALANCING
905 unsigned int nr_numa_running;
906 unsigned int nr_preferred_running;
907 unsigned int numa_migrate_on;
908#endif
909#ifdef CONFIG_NO_HZ_COMMON
910#ifdef CONFIG_SMP
911 unsigned long last_blocked_load_update_tick;
912 unsigned int has_blocked_load;
913 call_single_data_t nohz_csd;
914#endif /* CONFIG_SMP */
915 unsigned int nohz_tick_stopped;
916 atomic_t nohz_flags;
917#endif /* CONFIG_NO_HZ_COMMON */
918
919#ifdef CONFIG_SMP
920 unsigned int ttwu_pending;
921#endif
922 u64 nr_switches;
923
924#ifdef CONFIG_UCLAMP_TASK
925 /* Utilization clamp values based on CPU's RUNNABLE tasks */
926 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
927 unsigned int uclamp_flags;
928#define UCLAMP_FLAG_IDLE 0x01
929#endif
930
931 struct cfs_rq cfs;
932 struct rt_rq rt;
933 struct dl_rq dl;
934
935#ifdef CONFIG_FAIR_GROUP_SCHED
936 /* list of leaf cfs_rq on this CPU: */
937 struct list_head leaf_cfs_rq_list;
938 struct list_head *tmp_alone_branch;
939#endif /* CONFIG_FAIR_GROUP_SCHED */
940
941 /*
942 * This is part of a global counter where only the total sum
943 * over all CPUs matters. A task can increase this counter on
944 * one CPU and if it got migrated afterwards it may decrease
945 * it on another CPU. Always updated under the runqueue lock:
946 */
947 unsigned long nr_uninterruptible;
948
949 struct task_struct __rcu *curr;
950 struct task_struct *idle;
951 struct task_struct *stop;
952 unsigned long next_balance;
953 struct mm_struct *prev_mm;
954
955 unsigned int clock_update_flags;
956 u64 clock;
957 /* Ensure that all clocks are in the same cache line */
958 u64 clock_task ____cacheline_aligned;
959 u64 clock_pelt;
960 unsigned long lost_idle_time;
961
962 atomic_t nr_iowait;
963
964#ifdef CONFIG_MEMBARRIER
965 int membarrier_state;
966#endif
967
968#ifdef CONFIG_SMP
969 struct root_domain *rd;
970 struct sched_domain __rcu *sd;
971
972 unsigned long cpu_capacity;
973 unsigned long cpu_capacity_orig;
974
975 struct callback_head *balance_callback;
976
977 unsigned char nohz_idle_balance;
978 unsigned char idle_balance;
979
980 unsigned long misfit_task_load;
981
982 /* For active balancing */
983 int active_balance;
984 int push_cpu;
985 struct cpu_stop_work active_balance_work;
986
987 /* CPU of this runqueue: */
988 int cpu;
989 int online;
990
991 struct list_head cfs_tasks;
992
993 struct sched_avg avg_rt;
994 struct sched_avg avg_dl;
995#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
996 struct sched_avg avg_irq;
997#endif
998#ifdef CONFIG_SCHED_THERMAL_PRESSURE
999 struct sched_avg avg_thermal;
1000#endif
1001 u64 idle_stamp;
1002 u64 avg_idle;
1003
1004 /* This is used to determine avg_idle's max value */
1005 u64 max_idle_balance_cost;
1006#endif /* CONFIG_SMP */
1007
1008#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1009 u64 prev_irq_time;
1010#endif
1011#ifdef CONFIG_PARAVIRT
1012 u64 prev_steal_time;
1013#endif
1014#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1015 u64 prev_steal_time_rq;
1016#endif
1017
1018 /* calc_load related fields */
1019 unsigned long calc_load_update;
1020 long calc_load_active;
1021
1022#ifdef CONFIG_SCHED_HRTICK
1023#ifdef CONFIG_SMP
1024 call_single_data_t hrtick_csd;
1025#endif
1026 struct hrtimer hrtick_timer;
1027#endif
1028
1029#ifdef CONFIG_SCHEDSTATS
1030 /* latency stats */
1031 struct sched_info rq_sched_info;
1032 unsigned long long rq_cpu_time;
1033 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1034
1035 /* sys_sched_yield() stats */
1036 unsigned int yld_count;
1037
1038 /* schedule() stats */
1039 unsigned int sched_count;
1040 unsigned int sched_goidle;
1041
1042 /* try_to_wake_up() stats */
1043 unsigned int ttwu_count;
1044 unsigned int ttwu_local;
1045#endif
1046
1047#ifdef CONFIG_CPU_IDLE
1048 /* Must be inspected within a rcu lock section */
1049 struct cpuidle_state *idle_state;
1050#endif
1051};
1052
1053#ifdef CONFIG_FAIR_GROUP_SCHED
1054
1055/* CPU runqueue to which this cfs_rq is attached */
1056static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1057{
1058 return cfs_rq->rq;
1059}
1060
1061#else
1062
1063static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1064{
1065 return container_of(cfs_rq, struct rq, cfs);
1066}
1067#endif
1068
1069static inline int cpu_of(struct rq *rq)
1070{
1071#ifdef CONFIG_SMP
1072 return rq->cpu;
1073#else
1074 return 0;
1075#endif
1076}
1077
1078
1079#ifdef CONFIG_SCHED_SMT
1080extern void __update_idle_core(struct rq *rq);
1081
1082static inline void update_idle_core(struct rq *rq)
1083{
1084 if (static_branch_unlikely(&sched_smt_present))
1085 __update_idle_core(rq);
1086}
1087
1088#else
1089static inline void update_idle_core(struct rq *rq) { }
1090#endif
1091
1092DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1093
1094#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1095#define this_rq() this_cpu_ptr(&runqueues)
1096#define task_rq(p) cpu_rq(task_cpu(p))
1097#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1098#define raw_rq() raw_cpu_ptr(&runqueues)
1099
1100extern void update_rq_clock(struct rq *rq);
1101
1102static inline u64 __rq_clock_broken(struct rq *rq)
1103{
1104 return READ_ONCE(rq->clock);
1105}
1106
1107/*
1108 * rq::clock_update_flags bits
1109 *
1110 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1111 * call to __schedule(). This is an optimisation to avoid
1112 * neighbouring rq clock updates.
1113 *
1114 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1115 * in effect and calls to update_rq_clock() are being ignored.
1116 *
1117 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1118 * made to update_rq_clock() since the last time rq::lock was pinned.
1119 *
1120 * If inside of __schedule(), clock_update_flags will have been
1121 * shifted left (a left shift is a cheap operation for the fast path
1122 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1123 *
1124 * if (rq-clock_update_flags >= RQCF_UPDATED)
1125 *
1126 * to check if %RQCF_UPADTED is set. It'll never be shifted more than
1127 * one position though, because the next rq_unpin_lock() will shift it
1128 * back.
1129 */
1130#define RQCF_REQ_SKIP 0x01
1131#define RQCF_ACT_SKIP 0x02
1132#define RQCF_UPDATED 0x04
1133
1134static inline void assert_clock_updated(struct rq *rq)
1135{
1136 /*
1137 * The only reason for not seeing a clock update since the
1138 * last rq_pin_lock() is if we're currently skipping updates.
1139 */
1140 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1141}
1142
1143static inline u64 rq_clock(struct rq *rq)
1144{
1145 lockdep_assert_held(&rq->lock);
1146 assert_clock_updated(rq);
1147
1148 return rq->clock;
1149}
1150
1151static inline u64 rq_clock_task(struct rq *rq)
1152{
1153 lockdep_assert_held(&rq->lock);
1154 assert_clock_updated(rq);
1155
1156 return rq->clock_task;
1157}
1158
1159/**
1160 * By default the decay is the default pelt decay period.
1161 * The decay shift can change the decay period in
1162 * multiples of 32.
1163 * Decay shift Decay period(ms)
1164 * 0 32
1165 * 1 64
1166 * 2 128
1167 * 3 256
1168 * 4 512
1169 */
1170extern int sched_thermal_decay_shift;
1171
1172static inline u64 rq_clock_thermal(struct rq *rq)
1173{
1174 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1175}
1176
1177static inline void rq_clock_skip_update(struct rq *rq)
1178{
1179 lockdep_assert_held(&rq->lock);
1180 rq->clock_update_flags |= RQCF_REQ_SKIP;
1181}
1182
1183/*
1184 * See rt task throttling, which is the only time a skip
1185 * request is cancelled.
1186 */
1187static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1188{
1189 lockdep_assert_held(&rq->lock);
1190 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1191}
1192
1193struct rq_flags {
1194 unsigned long flags;
1195 struct pin_cookie cookie;
1196#ifdef CONFIG_SCHED_DEBUG
1197 /*
1198 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1199 * current pin context is stashed here in case it needs to be
1200 * restored in rq_repin_lock().
1201 */
1202 unsigned int clock_update_flags;
1203#endif
1204};
1205
1206/*
1207 * Lockdep annotation that avoids accidental unlocks; it's like a
1208 * sticky/continuous lockdep_assert_held().
1209 *
1210 * This avoids code that has access to 'struct rq *rq' (basically everything in
1211 * the scheduler) from accidentally unlocking the rq if they do not also have a
1212 * copy of the (on-stack) 'struct rq_flags rf'.
1213 *
1214 * Also see Documentation/locking/lockdep-design.rst.
1215 */
1216static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1217{
1218 rf->cookie = lockdep_pin_lock(&rq->lock);
1219
1220#ifdef CONFIG_SCHED_DEBUG
1221 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1222 rf->clock_update_flags = 0;
1223#endif
1224}
1225
1226static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1227{
1228#ifdef CONFIG_SCHED_DEBUG
1229 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1230 rf->clock_update_flags = RQCF_UPDATED;
1231#endif
1232
1233 lockdep_unpin_lock(&rq->lock, rf->cookie);
1234}
1235
1236static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1237{
1238 lockdep_repin_lock(&rq->lock, rf->cookie);
1239
1240#ifdef CONFIG_SCHED_DEBUG
1241 /*
1242 * Restore the value we stashed in @rf for this pin context.
1243 */
1244 rq->clock_update_flags |= rf->clock_update_flags;
1245#endif
1246}
1247
1248struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1249 __acquires(rq->lock);
1250
1251struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1252 __acquires(p->pi_lock)
1253 __acquires(rq->lock);
1254
1255static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1256 __releases(rq->lock)
1257{
1258 rq_unpin_lock(rq, rf);
1259 raw_spin_unlock(&rq->lock);
1260}
1261
1262static inline void
1263task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1264 __releases(rq->lock)
1265 __releases(p->pi_lock)
1266{
1267 rq_unpin_lock(rq, rf);
1268 raw_spin_unlock(&rq->lock);
1269 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1270}
1271
1272static inline void
1273rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1274 __acquires(rq->lock)
1275{
1276 raw_spin_lock_irqsave(&rq->lock, rf->flags);
1277 rq_pin_lock(rq, rf);
1278}
1279
1280static inline void
1281rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1282 __acquires(rq->lock)
1283{
1284 raw_spin_lock_irq(&rq->lock);
1285 rq_pin_lock(rq, rf);
1286}
1287
1288static inline void
1289rq_lock(struct rq *rq, struct rq_flags *rf)
1290 __acquires(rq->lock)
1291{
1292 raw_spin_lock(&rq->lock);
1293 rq_pin_lock(rq, rf);
1294}
1295
1296static inline void
1297rq_relock(struct rq *rq, struct rq_flags *rf)
1298 __acquires(rq->lock)
1299{
1300 raw_spin_lock(&rq->lock);
1301 rq_repin_lock(rq, rf);
1302}
1303
1304static inline void
1305rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1306 __releases(rq->lock)
1307{
1308 rq_unpin_lock(rq, rf);
1309 raw_spin_unlock_irqrestore(&rq->lock, rf->flags);
1310}
1311
1312static inline void
1313rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1314 __releases(rq->lock)
1315{
1316 rq_unpin_lock(rq, rf);
1317 raw_spin_unlock_irq(&rq->lock);
1318}
1319
1320static inline void
1321rq_unlock(struct rq *rq, struct rq_flags *rf)
1322 __releases(rq->lock)
1323{
1324 rq_unpin_lock(rq, rf);
1325 raw_spin_unlock(&rq->lock);
1326}
1327
1328static inline struct rq *
1329this_rq_lock_irq(struct rq_flags *rf)
1330 __acquires(rq->lock)
1331{
1332 struct rq *rq;
1333
1334 local_irq_disable();
1335 rq = this_rq();
1336 rq_lock(rq, rf);
1337 return rq;
1338}
1339
1340#ifdef CONFIG_NUMA
1341enum numa_topology_type {
1342 NUMA_DIRECT,
1343 NUMA_GLUELESS_MESH,
1344 NUMA_BACKPLANE,
1345};
1346extern enum numa_topology_type sched_numa_topology_type;
1347extern int sched_max_numa_distance;
1348extern bool find_numa_distance(int distance);
1349extern void sched_init_numa(void);
1350extern void sched_domains_numa_masks_set(unsigned int cpu);
1351extern void sched_domains_numa_masks_clear(unsigned int cpu);
1352extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1353#else
1354static inline void sched_init_numa(void) { }
1355static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1356static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1357static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1358{
1359 return nr_cpu_ids;
1360}
1361#endif
1362
1363#ifdef CONFIG_NUMA_BALANCING
1364/* The regions in numa_faults array from task_struct */
1365enum numa_faults_stats {
1366 NUMA_MEM = 0,
1367 NUMA_CPU,
1368 NUMA_MEMBUF,
1369 NUMA_CPUBUF
1370};
1371extern void sched_setnuma(struct task_struct *p, int node);
1372extern int migrate_task_to(struct task_struct *p, int cpu);
1373extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1374 int cpu, int scpu);
1375extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1376#else
1377static inline void
1378init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1379{
1380}
1381#endif /* CONFIG_NUMA_BALANCING */
1382
1383#ifdef CONFIG_SMP
1384
1385static inline void
1386queue_balance_callback(struct rq *rq,
1387 struct callback_head *head,
1388 void (*func)(struct rq *rq))
1389{
1390 lockdep_assert_held(&rq->lock);
1391
1392 if (unlikely(head->next))
1393 return;
1394
1395 head->func = (void (*)(struct callback_head *))func;
1396 head->next = rq->balance_callback;
1397 rq->balance_callback = head;
1398}
1399
1400#define rcu_dereference_check_sched_domain(p) \
1401 rcu_dereference_check((p), \
1402 lockdep_is_held(&sched_domains_mutex))
1403
1404/*
1405 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1406 * See destroy_sched_domains: call_rcu for details.
1407 *
1408 * The domain tree of any CPU may only be accessed from within
1409 * preempt-disabled sections.
1410 */
1411#define for_each_domain(cpu, __sd) \
1412 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1413 __sd; __sd = __sd->parent)
1414
1415/**
1416 * highest_flag_domain - Return highest sched_domain containing flag.
1417 * @cpu: The CPU whose highest level of sched domain is to
1418 * be returned.
1419 * @flag: The flag to check for the highest sched_domain
1420 * for the given CPU.
1421 *
1422 * Returns the highest sched_domain of a CPU which contains the given flag.
1423 */
1424static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1425{
1426 struct sched_domain *sd, *hsd = NULL;
1427
1428 for_each_domain(cpu, sd) {
1429 if (!(sd->flags & flag))
1430 break;
1431 hsd = sd;
1432 }
1433
1434 return hsd;
1435}
1436
1437static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1438{
1439 struct sched_domain *sd;
1440
1441 for_each_domain(cpu, sd) {
1442 if (sd->flags & flag)
1443 break;
1444 }
1445
1446 return sd;
1447}
1448
1449DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1450DECLARE_PER_CPU(int, sd_llc_size);
1451DECLARE_PER_CPU(int, sd_llc_id);
1452DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1453DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1454DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1455DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1456extern struct static_key_false sched_asym_cpucapacity;
1457
1458struct sched_group_capacity {
1459 atomic_t ref;
1460 /*
1461 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1462 * for a single CPU.
1463 */
1464 unsigned long capacity;
1465 unsigned long min_capacity; /* Min per-CPU capacity in group */
1466 unsigned long max_capacity; /* Max per-CPU capacity in group */
1467 unsigned long next_update;
1468 int imbalance; /* XXX unrelated to capacity but shared group state */
1469
1470#ifdef CONFIG_SCHED_DEBUG
1471 int id;
1472#endif
1473
1474 unsigned long cpumask[0]; /* Balance mask */
1475};
1476
1477struct sched_group {
1478 struct sched_group *next; /* Must be a circular list */
1479 atomic_t ref;
1480
1481 unsigned int group_weight;
1482 struct sched_group_capacity *sgc;
1483 int asym_prefer_cpu; /* CPU of highest priority in group */
1484
1485 /*
1486 * The CPUs this group covers.
1487 *
1488 * NOTE: this field is variable length. (Allocated dynamically
1489 * by attaching extra space to the end of the structure,
1490 * depending on how many CPUs the kernel has booted up with)
1491 */
1492 unsigned long cpumask[];
1493};
1494
1495static inline struct cpumask *sched_group_span(struct sched_group *sg)
1496{
1497 return to_cpumask(sg->cpumask);
1498}
1499
1500/*
1501 * See build_balance_mask().
1502 */
1503static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1504{
1505 return to_cpumask(sg->sgc->cpumask);
1506}
1507
1508/**
1509 * group_first_cpu - Returns the first CPU in the cpumask of a sched_group.
1510 * @group: The group whose first CPU is to be returned.
1511 */
1512static inline unsigned int group_first_cpu(struct sched_group *group)
1513{
1514 return cpumask_first(sched_group_span(group));
1515}
1516
1517extern int group_balance_cpu(struct sched_group *sg);
1518
1519#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
1520void register_sched_domain_sysctl(void);
1521void dirty_sched_domain_sysctl(int cpu);
1522void unregister_sched_domain_sysctl(void);
1523#else
1524static inline void register_sched_domain_sysctl(void)
1525{
1526}
1527static inline void dirty_sched_domain_sysctl(int cpu)
1528{
1529}
1530static inline void unregister_sched_domain_sysctl(void)
1531{
1532}
1533#endif
1534
1535extern void flush_smp_call_function_from_idle(void);
1536
1537#else /* !CONFIG_SMP: */
1538static inline void flush_smp_call_function_from_idle(void) { }
1539#endif
1540
1541#include "stats.h"
1542#include "autogroup.h"
1543
1544#ifdef CONFIG_CGROUP_SCHED
1545
1546/*
1547 * Return the group to which this tasks belongs.
1548 *
1549 * We cannot use task_css() and friends because the cgroup subsystem
1550 * changes that value before the cgroup_subsys::attach() method is called,
1551 * therefore we cannot pin it and might observe the wrong value.
1552 *
1553 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
1554 * core changes this before calling sched_move_task().
1555 *
1556 * Instead we use a 'copy' which is updated from sched_move_task() while
1557 * holding both task_struct::pi_lock and rq::lock.
1558 */
1559static inline struct task_group *task_group(struct task_struct *p)
1560{
1561 return p->sched_task_group;
1562}
1563
1564/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
1565static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
1566{
1567#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
1568 struct task_group *tg = task_group(p);
1569#endif
1570
1571#ifdef CONFIG_FAIR_GROUP_SCHED
1572 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
1573 p->se.cfs_rq = tg->cfs_rq[cpu];
1574 p->se.parent = tg->se[cpu];
1575#endif
1576
1577#ifdef CONFIG_RT_GROUP_SCHED
1578 p->rt.rt_rq = tg->rt_rq[cpu];
1579 p->rt.parent = tg->rt_se[cpu];
1580#endif
1581}
1582
1583#else /* CONFIG_CGROUP_SCHED */
1584
1585static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
1586static inline struct task_group *task_group(struct task_struct *p)
1587{
1588 return NULL;
1589}
1590
1591#endif /* CONFIG_CGROUP_SCHED */
1592
1593static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1594{
1595 set_task_rq(p, cpu);
1596#ifdef CONFIG_SMP
1597 /*
1598 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1599 * successfully executed on another CPU. We must ensure that updates of
1600 * per-task data have been completed by this moment.
1601 */
1602 smp_wmb();
1603#ifdef CONFIG_THREAD_INFO_IN_TASK
1604 WRITE_ONCE(p->cpu, cpu);
1605#else
1606 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
1607#endif
1608 p->wake_cpu = cpu;
1609#endif
1610}
1611
1612/*
1613 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
1614 */
1615#ifdef CONFIG_SCHED_DEBUG
1616# include <linux/static_key.h>
1617# define const_debug __read_mostly
1618#else
1619# define const_debug const
1620#endif
1621
1622#define SCHED_FEAT(name, enabled) \
1623 __SCHED_FEAT_##name ,
1624
1625enum {
1626#include "features.h"
1627 __SCHED_FEAT_NR,
1628};
1629
1630#undef SCHED_FEAT
1631
1632#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
1633
1634/*
1635 * To support run-time toggling of sched features, all the translation units
1636 * (but core.c) reference the sysctl_sched_features defined in core.c.
1637 */
1638extern const_debug unsigned int sysctl_sched_features;
1639
1640#define SCHED_FEAT(name, enabled) \
1641static __always_inline bool static_branch_##name(struct static_key *key) \
1642{ \
1643 return static_key_##enabled(key); \
1644}
1645
1646#include "features.h"
1647#undef SCHED_FEAT
1648
1649extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
1650#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
1651
1652#else /* !(SCHED_DEBUG && CONFIG_JUMP_LABEL) */
1653
1654/*
1655 * Each translation unit has its own copy of sysctl_sched_features to allow
1656 * constants propagation at compile time and compiler optimization based on
1657 * features default.
1658 */
1659#define SCHED_FEAT(name, enabled) \
1660 (1UL << __SCHED_FEAT_##name) * enabled |
1661static const_debug __maybe_unused unsigned int sysctl_sched_features =
1662#include "features.h"
1663 0;
1664#undef SCHED_FEAT
1665
1666#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
1667
1668#endif /* SCHED_DEBUG && CONFIG_JUMP_LABEL */
1669
1670extern struct static_key_false sched_numa_balancing;
1671extern struct static_key_false sched_schedstats;
1672
1673static inline u64 global_rt_period(void)
1674{
1675 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
1676}
1677
1678static inline u64 global_rt_runtime(void)
1679{
1680 if (sysctl_sched_rt_runtime < 0)
1681 return RUNTIME_INF;
1682
1683 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
1684}
1685
1686static inline int task_current(struct rq *rq, struct task_struct *p)
1687{
1688 return rq->curr == p;
1689}
1690
1691static inline int task_running(struct rq *rq, struct task_struct *p)
1692{
1693#ifdef CONFIG_SMP
1694 return p->on_cpu;
1695#else
1696 return task_current(rq, p);
1697#endif
1698}
1699
1700static inline int task_on_rq_queued(struct task_struct *p)
1701{
1702 return p->on_rq == TASK_ON_RQ_QUEUED;
1703}
1704
1705static inline int task_on_rq_migrating(struct task_struct *p)
1706{
1707 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
1708}
1709
1710/*
1711 * wake flags
1712 */
1713#define WF_SYNC 0x01 /* Waker goes to sleep after wakeup */
1714#define WF_FORK 0x02 /* Child wakeup after fork */
1715#define WF_MIGRATED 0x04 /* Internal use, task got migrated */
1716#define WF_ON_CPU 0x08 /* Wakee is on_cpu */
1717
1718/*
1719 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1720 * of tasks with abnormal "nice" values across CPUs the contribution that
1721 * each task makes to its run queue's load is weighted according to its
1722 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1723 * scaled version of the new time slice allocation that they receive on time
1724 * slice expiry etc.
1725 */
1726
1727#define WEIGHT_IDLEPRIO 3
1728#define WMULT_IDLEPRIO 1431655765
1729
1730extern const int sched_prio_to_weight[40];
1731extern const u32 sched_prio_to_wmult[40];
1732
1733/*
1734 * {de,en}queue flags:
1735 *
1736 * DEQUEUE_SLEEP - task is no longer runnable
1737 * ENQUEUE_WAKEUP - task just became runnable
1738 *
1739 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
1740 * are in a known state which allows modification. Such pairs
1741 * should preserve as much state as possible.
1742 *
1743 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
1744 * in the runqueue.
1745 *
1746 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
1747 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
1748 * ENQUEUE_MIGRATED - the task was migrated during wakeup
1749 *
1750 */
1751
1752#define DEQUEUE_SLEEP 0x01
1753#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
1754#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
1755#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
1756
1757#define ENQUEUE_WAKEUP 0x01
1758#define ENQUEUE_RESTORE 0x02
1759#define ENQUEUE_MOVE 0x04
1760#define ENQUEUE_NOCLOCK 0x08
1761
1762#define ENQUEUE_HEAD 0x10
1763#define ENQUEUE_REPLENISH 0x20
1764#ifdef CONFIG_SMP
1765#define ENQUEUE_MIGRATED 0x40
1766#else
1767#define ENQUEUE_MIGRATED 0x00
1768#endif
1769
1770#define RETRY_TASK ((void *)-1UL)
1771
1772struct sched_class {
1773
1774#ifdef CONFIG_UCLAMP_TASK
1775 int uclamp_enabled;
1776#endif
1777
1778 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1779 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1780 void (*yield_task) (struct rq *rq);
1781 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
1782
1783 void (*check_preempt_curr)(struct rq *rq, struct task_struct *p, int flags);
1784
1785 struct task_struct *(*pick_next_task)(struct rq *rq);
1786
1787 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
1788 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
1789
1790#ifdef CONFIG_SMP
1791 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
1792 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1793 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
1794
1795 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
1796
1797 void (*set_cpus_allowed)(struct task_struct *p,
1798 const struct cpumask *newmask);
1799
1800 void (*rq_online)(struct rq *rq);
1801 void (*rq_offline)(struct rq *rq);
1802#endif
1803
1804 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
1805 void (*task_fork)(struct task_struct *p);
1806 void (*task_dead)(struct task_struct *p);
1807
1808 /*
1809 * The switched_from() call is allowed to drop rq->lock, therefore we
1810 * cannot assume the switched_from/switched_to pair is serliazed by
1811 * rq->lock. They are however serialized by p->pi_lock.
1812 */
1813 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
1814 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1815 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1816 int oldprio);
1817
1818 unsigned int (*get_rr_interval)(struct rq *rq,
1819 struct task_struct *task);
1820
1821 void (*update_curr)(struct rq *rq);
1822
1823#define TASK_SET_GROUP 0
1824#define TASK_MOVE_GROUP 1
1825
1826#ifdef CONFIG_FAIR_GROUP_SCHED
1827 void (*task_change_group)(struct task_struct *p, int type);
1828#endif
1829} __aligned(STRUCT_ALIGNMENT); /* STRUCT_ALIGN(), vmlinux.lds.h */
1830
1831static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1832{
1833 WARN_ON_ONCE(rq->curr != prev);
1834 prev->sched_class->put_prev_task(rq, prev);
1835}
1836
1837static inline void set_next_task(struct rq *rq, struct task_struct *next)
1838{
1839 WARN_ON_ONCE(rq->curr != next);
1840 next->sched_class->set_next_task(rq, next, false);
1841}
1842
1843/* Defined in include/asm-generic/vmlinux.lds.h */
1844extern struct sched_class __begin_sched_classes[];
1845extern struct sched_class __end_sched_classes[];
1846
1847#define sched_class_highest (__end_sched_classes - 1)
1848#define sched_class_lowest (__begin_sched_classes - 1)
1849
1850#define for_class_range(class, _from, _to) \
1851 for (class = (_from); class != (_to); class--)
1852
1853#define for_each_class(class) \
1854 for_class_range(class, sched_class_highest, sched_class_lowest)
1855
1856extern const struct sched_class stop_sched_class;
1857extern const struct sched_class dl_sched_class;
1858extern const struct sched_class rt_sched_class;
1859extern const struct sched_class fair_sched_class;
1860extern const struct sched_class idle_sched_class;
1861
1862static inline bool sched_stop_runnable(struct rq *rq)
1863{
1864 return rq->stop && task_on_rq_queued(rq->stop);
1865}
1866
1867static inline bool sched_dl_runnable(struct rq *rq)
1868{
1869 return rq->dl.dl_nr_running > 0;
1870}
1871
1872static inline bool sched_rt_runnable(struct rq *rq)
1873{
1874 return rq->rt.rt_queued > 0;
1875}
1876
1877static inline bool sched_fair_runnable(struct rq *rq)
1878{
1879 return rq->cfs.nr_running > 0;
1880}
1881
1882extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
1883extern struct task_struct *pick_next_task_idle(struct rq *rq);
1884
1885#ifdef CONFIG_SMP
1886
1887extern void update_group_capacity(struct sched_domain *sd, int cpu);
1888
1889extern void trigger_load_balance(struct rq *rq);
1890
1891extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask);
1892
1893#endif
1894
1895#ifdef CONFIG_CPU_IDLE
1896static inline void idle_set_state(struct rq *rq,
1897 struct cpuidle_state *idle_state)
1898{
1899 rq->idle_state = idle_state;
1900}
1901
1902static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1903{
1904 SCHED_WARN_ON(!rcu_read_lock_held());
1905
1906 return rq->idle_state;
1907}
1908#else
1909static inline void idle_set_state(struct rq *rq,
1910 struct cpuidle_state *idle_state)
1911{
1912}
1913
1914static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1915{
1916 return NULL;
1917}
1918#endif
1919
1920extern void schedule_idle(void);
1921
1922extern void sysrq_sched_debug_show(void);
1923extern void sched_init_granularity(void);
1924extern void update_max_interval(void);
1925
1926extern void init_sched_dl_class(void);
1927extern void init_sched_rt_class(void);
1928extern void init_sched_fair_class(void);
1929
1930extern void reweight_task(struct task_struct *p, int prio);
1931
1932extern void resched_curr(struct rq *rq);
1933extern void resched_cpu(int cpu);
1934
1935extern struct rt_bandwidth def_rt_bandwidth;
1936extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1937
1938extern struct dl_bandwidth def_dl_bandwidth;
1939extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1940extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1941extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se);
1942
1943#define BW_SHIFT 20
1944#define BW_UNIT (1 << BW_SHIFT)
1945#define RATIO_SHIFT 8
1946#define MAX_BW_BITS (64 - BW_SHIFT)
1947#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
1948unsigned long to_ratio(u64 period, u64 runtime);
1949
1950extern void init_entity_runnable_average(struct sched_entity *se);
1951extern void post_init_entity_util_avg(struct task_struct *p);
1952
1953#ifdef CONFIG_NO_HZ_FULL
1954extern bool sched_can_stop_tick(struct rq *rq);
1955extern int __init sched_tick_offload_init(void);
1956
1957/*
1958 * Tick may be needed by tasks in the runqueue depending on their policy and
1959 * requirements. If tick is needed, lets send the target an IPI to kick it out of
1960 * nohz mode if necessary.
1961 */
1962static inline void sched_update_tick_dependency(struct rq *rq)
1963{
1964 int cpu = cpu_of(rq);
1965
1966 if (!tick_nohz_full_cpu(cpu))
1967 return;
1968
1969 if (sched_can_stop_tick(rq))
1970 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
1971 else
1972 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
1973}
1974#else
1975static inline int sched_tick_offload_init(void) { return 0; }
1976static inline void sched_update_tick_dependency(struct rq *rq) { }
1977#endif
1978
1979static inline void add_nr_running(struct rq *rq, unsigned count)
1980{
1981 unsigned prev_nr = rq->nr_running;
1982
1983 rq->nr_running = prev_nr + count;
1984 if (trace_sched_update_nr_running_tp_enabled()) {
1985 call_trace_sched_update_nr_running(rq, count);
1986 }
1987
1988#ifdef CONFIG_SMP
1989 if (prev_nr < 2 && rq->nr_running >= 2) {
1990 if (!READ_ONCE(rq->rd->overload))
1991 WRITE_ONCE(rq->rd->overload, 1);
1992 }
1993#endif
1994
1995 sched_update_tick_dependency(rq);
1996}
1997
1998static inline void sub_nr_running(struct rq *rq, unsigned count)
1999{
2000 rq->nr_running -= count;
2001 if (trace_sched_update_nr_running_tp_enabled()) {
2002 call_trace_sched_update_nr_running(rq, -count);
2003 }
2004
2005 /* Check if we still need preemption */
2006 sched_update_tick_dependency(rq);
2007}
2008
2009extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2010extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2011
2012extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
2013
2014extern const_debug unsigned int sysctl_sched_nr_migrate;
2015extern const_debug unsigned int sysctl_sched_migration_cost;
2016
2017#ifdef CONFIG_SCHED_HRTICK
2018
2019/*
2020 * Use hrtick when:
2021 * - enabled by features
2022 * - hrtimer is actually high res
2023 */
2024static inline int hrtick_enabled(struct rq *rq)
2025{
2026 if (!sched_feat(HRTICK))
2027 return 0;
2028 if (!cpu_active(cpu_of(rq)))
2029 return 0;
2030 return hrtimer_is_hres_active(&rq->hrtick_timer);
2031}
2032
2033void hrtick_start(struct rq *rq, u64 delay);
2034
2035#else
2036
2037static inline int hrtick_enabled(struct rq *rq)
2038{
2039 return 0;
2040}
2041
2042#endif /* CONFIG_SCHED_HRTICK */
2043
2044#ifndef arch_scale_freq_tick
2045static __always_inline
2046void arch_scale_freq_tick(void)
2047{
2048}
2049#endif
2050
2051#ifndef arch_scale_freq_capacity
2052/**
2053 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2054 * @cpu: the CPU in question.
2055 *
2056 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2057 *
2058 * f_curr
2059 * ------ * SCHED_CAPACITY_SCALE
2060 * f_max
2061 */
2062static __always_inline
2063unsigned long arch_scale_freq_capacity(int cpu)
2064{
2065 return SCHED_CAPACITY_SCALE;
2066}
2067#endif
2068
2069#ifdef CONFIG_SMP
2070#ifdef CONFIG_PREEMPTION
2071
2072static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
2073
2074/*
2075 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2076 * way at the expense of forcing extra atomic operations in all
2077 * invocations. This assures that the double_lock is acquired using the
2078 * same underlying policy as the spinlock_t on this architecture, which
2079 * reduces latency compared to the unfair variant below. However, it
2080 * also adds more overhead and therefore may reduce throughput.
2081 */
2082static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2083 __releases(this_rq->lock)
2084 __acquires(busiest->lock)
2085 __acquires(this_rq->lock)
2086{
2087 raw_spin_unlock(&this_rq->lock);
2088 double_rq_lock(this_rq, busiest);
2089
2090 return 1;
2091}
2092
2093#else
2094/*
2095 * Unfair double_lock_balance: Optimizes throughput at the expense of
2096 * latency by eliminating extra atomic operations when the locks are
2097 * already in proper order on entry. This favors lower CPU-ids and will
2098 * grant the double lock to lower CPUs over higher ids under contention,
2099 * regardless of entry order into the function.
2100 */
2101static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2102 __releases(this_rq->lock)
2103 __acquires(busiest->lock)
2104 __acquires(this_rq->lock)
2105{
2106 int ret = 0;
2107
2108 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
2109 if (busiest < this_rq) {
2110 raw_spin_unlock(&this_rq->lock);
2111 raw_spin_lock(&busiest->lock);
2112 raw_spin_lock_nested(&this_rq->lock,
2113 SINGLE_DEPTH_NESTING);
2114 ret = 1;
2115 } else
2116 raw_spin_lock_nested(&busiest->lock,
2117 SINGLE_DEPTH_NESTING);
2118 }
2119 return ret;
2120}
2121
2122#endif /* CONFIG_PREEMPTION */
2123
2124/*
2125 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2126 */
2127static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2128{
2129 if (unlikely(!irqs_disabled())) {
2130 /* printk() doesn't work well under rq->lock */
2131 raw_spin_unlock(&this_rq->lock);
2132 BUG_ON(1);
2133 }
2134
2135 return _double_lock_balance(this_rq, busiest);
2136}
2137
2138static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2139 __releases(busiest->lock)
2140{
2141 raw_spin_unlock(&busiest->lock);
2142 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2143}
2144
2145static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2146{
2147 if (l1 > l2)
2148 swap(l1, l2);
2149
2150 spin_lock(l1);
2151 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2152}
2153
2154static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2155{
2156 if (l1 > l2)
2157 swap(l1, l2);
2158
2159 spin_lock_irq(l1);
2160 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2161}
2162
2163static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2164{
2165 if (l1 > l2)
2166 swap(l1, l2);
2167
2168 raw_spin_lock(l1);
2169 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2170}
2171
2172/*
2173 * double_rq_lock - safely lock two runqueues
2174 *
2175 * Note this does not disable interrupts like task_rq_lock,
2176 * you need to do so manually before calling.
2177 */
2178static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2179 __acquires(rq1->lock)
2180 __acquires(rq2->lock)
2181{
2182 BUG_ON(!irqs_disabled());
2183 if (rq1 == rq2) {
2184 raw_spin_lock(&rq1->lock);
2185 __acquire(rq2->lock); /* Fake it out ;) */
2186 } else {
2187 if (rq1 < rq2) {
2188 raw_spin_lock(&rq1->lock);
2189 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2190 } else {
2191 raw_spin_lock(&rq2->lock);
2192 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2193 }
2194 }
2195}
2196
2197/*
2198 * double_rq_unlock - safely unlock two runqueues
2199 *
2200 * Note this does not restore interrupts like task_rq_unlock,
2201 * you need to do so manually after calling.
2202 */
2203static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2204 __releases(rq1->lock)
2205 __releases(rq2->lock)
2206{
2207 raw_spin_unlock(&rq1->lock);
2208 if (rq1 != rq2)
2209 raw_spin_unlock(&rq2->lock);
2210 else
2211 __release(rq2->lock);
2212}
2213
2214extern void set_rq_online (struct rq *rq);
2215extern void set_rq_offline(struct rq *rq);
2216extern bool sched_smp_initialized;
2217
2218#else /* CONFIG_SMP */
2219
2220/*
2221 * double_rq_lock - safely lock two runqueues
2222 *
2223 * Note this does not disable interrupts like task_rq_lock,
2224 * you need to do so manually before calling.
2225 */
2226static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2227 __acquires(rq1->lock)
2228 __acquires(rq2->lock)
2229{
2230 BUG_ON(!irqs_disabled());
2231 BUG_ON(rq1 != rq2);
2232 raw_spin_lock(&rq1->lock);
2233 __acquire(rq2->lock); /* Fake it out ;) */
2234}
2235
2236/*
2237 * double_rq_unlock - safely unlock two runqueues
2238 *
2239 * Note this does not restore interrupts like task_rq_unlock,
2240 * you need to do so manually after calling.
2241 */
2242static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2243 __releases(rq1->lock)
2244 __releases(rq2->lock)
2245{
2246 BUG_ON(rq1 != rq2);
2247 raw_spin_unlock(&rq1->lock);
2248 __release(rq2->lock);
2249}
2250
2251#endif
2252
2253extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2254extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2255
2256#ifdef CONFIG_SCHED_DEBUG
2257extern bool sched_debug_enabled;
2258
2259extern void print_cfs_stats(struct seq_file *m, int cpu);
2260extern void print_rt_stats(struct seq_file *m, int cpu);
2261extern void print_dl_stats(struct seq_file *m, int cpu);
2262extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2263extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2264extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2265#ifdef CONFIG_NUMA_BALANCING
2266extern void
2267show_numa_stats(struct task_struct *p, struct seq_file *m);
2268extern void
2269print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2270 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2271#endif /* CONFIG_NUMA_BALANCING */
2272#endif /* CONFIG_SCHED_DEBUG */
2273
2274extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2275extern void init_rt_rq(struct rt_rq *rt_rq);
2276extern void init_dl_rq(struct dl_rq *dl_rq);
2277
2278extern void cfs_bandwidth_usage_inc(void);
2279extern void cfs_bandwidth_usage_dec(void);
2280
2281#ifdef CONFIG_NO_HZ_COMMON
2282#define NOHZ_BALANCE_KICK_BIT 0
2283#define NOHZ_STATS_KICK_BIT 1
2284
2285#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2286#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2287
2288#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK)
2289
2290#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2291
2292extern void nohz_balance_exit_idle(struct rq *rq);
2293#else
2294static inline void nohz_balance_exit_idle(struct rq *rq) { }
2295#endif
2296
2297
2298#ifdef CONFIG_SMP
2299static inline
2300void __dl_update(struct dl_bw *dl_b, s64 bw)
2301{
2302 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
2303 int i;
2304
2305 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2306 "sched RCU must be held");
2307 for_each_cpu_and(i, rd->span, cpu_active_mask) {
2308 struct rq *rq = cpu_rq(i);
2309
2310 rq->dl.extra_bw += bw;
2311 }
2312}
2313#else
2314static inline
2315void __dl_update(struct dl_bw *dl_b, s64 bw)
2316{
2317 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
2318
2319 dl->extra_bw += bw;
2320}
2321#endif
2322
2323
2324#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2325struct irqtime {
2326 u64 total;
2327 u64 tick_delta;
2328 u64 irq_start_time;
2329 struct u64_stats_sync sync;
2330};
2331
2332DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2333
2334/*
2335 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2336 * Otherwise ksoftirqd's sum_exec_runtime is substracted its own runtime
2337 * and never move forward.
2338 */
2339static inline u64 irq_time_read(int cpu)
2340{
2341 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2342 unsigned int seq;
2343 u64 total;
2344
2345 do {
2346 seq = __u64_stats_fetch_begin(&irqtime->sync);
2347 total = irqtime->total;
2348 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2349
2350 return total;
2351}
2352#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2353
2354#ifdef CONFIG_CPU_FREQ
2355DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2356
2357/**
2358 * cpufreq_update_util - Take a note about CPU utilization changes.
2359 * @rq: Runqueue to carry out the update for.
2360 * @flags: Update reason flags.
2361 *
2362 * This function is called by the scheduler on the CPU whose utilization is
2363 * being updated.
2364 *
2365 * It can only be called from RCU-sched read-side critical sections.
2366 *
2367 * The way cpufreq is currently arranged requires it to evaluate the CPU
2368 * performance state (frequency/voltage) on a regular basis to prevent it from
2369 * being stuck in a completely inadequate performance level for too long.
2370 * That is not guaranteed to happen if the updates are only triggered from CFS
2371 * and DL, though, because they may not be coming in if only RT tasks are
2372 * active all the time (or there are RT tasks only).
2373 *
2374 * As a workaround for that issue, this function is called periodically by the
2375 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2376 * but that really is a band-aid. Going forward it should be replaced with
2377 * solutions targeted more specifically at RT tasks.
2378 */
2379static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2380{
2381 struct update_util_data *data;
2382
2383 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2384 cpu_of(rq)));
2385 if (data)
2386 data->func(data, rq_clock(rq), flags);
2387}
2388#else
2389static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2390#endif /* CONFIG_CPU_FREQ */
2391
2392#ifdef CONFIG_UCLAMP_TASK
2393unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
2394
2395/**
2396 * uclamp_rq_util_with - clamp @util with @rq and @p effective uclamp values.
2397 * @rq: The rq to clamp against. Must not be NULL.
2398 * @util: The util value to clamp.
2399 * @p: The task to clamp against. Can be NULL if you want to clamp
2400 * against @rq only.
2401 *
2402 * Clamps the passed @util to the max(@rq, @p) effective uclamp values.
2403 *
2404 * If sched_uclamp_used static key is disabled, then just return the util
2405 * without any clamping since uclamp aggregation at the rq level in the fast
2406 * path is disabled, rendering this operation a NOP.
2407 *
2408 * Use uclamp_eff_value() if you don't care about uclamp values at rq level. It
2409 * will return the correct effective uclamp value of the task even if the
2410 * static key is disabled.
2411 */
2412static __always_inline
2413unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
2414 struct task_struct *p)
2415{
2416 unsigned long min_util;
2417 unsigned long max_util;
2418
2419 if (!static_branch_likely(&sched_uclamp_used))
2420 return util;
2421
2422 min_util = READ_ONCE(rq->uclamp[UCLAMP_MIN].value);
2423 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
2424
2425 if (p) {
2426 min_util = max(min_util, uclamp_eff_value(p, UCLAMP_MIN));
2427 max_util = max(max_util, uclamp_eff_value(p, UCLAMP_MAX));
2428 }
2429
2430 /*
2431 * Since CPU's {min,max}_util clamps are MAX aggregated considering
2432 * RUNNABLE tasks with _different_ clamps, we can end up with an
2433 * inversion. Fix it now when the clamps are applied.
2434 */
2435 if (unlikely(min_util >= max_util))
2436 return min_util;
2437
2438 return clamp(util, min_util, max_util);
2439}
2440
2441/*
2442 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
2443 * by default in the fast path and only gets turned on once userspace performs
2444 * an operation that requires it.
2445 *
2446 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
2447 * hence is active.
2448 */
2449static inline bool uclamp_is_used(void)
2450{
2451 return static_branch_likely(&sched_uclamp_used);
2452}
2453#else /* CONFIG_UCLAMP_TASK */
2454static inline
2455unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
2456 struct task_struct *p)
2457{
2458 return util;
2459}
2460
2461static inline bool uclamp_is_used(void)
2462{
2463 return false;
2464}
2465#endif /* CONFIG_UCLAMP_TASK */
2466
2467#ifdef arch_scale_freq_capacity
2468# ifndef arch_scale_freq_invariant
2469# define arch_scale_freq_invariant() true
2470# endif
2471#else
2472# define arch_scale_freq_invariant() false
2473#endif
2474
2475#ifdef CONFIG_SMP
2476static inline unsigned long capacity_orig_of(int cpu)
2477{
2478 return cpu_rq(cpu)->cpu_capacity_orig;
2479}
2480#endif
2481
2482/**
2483 * enum schedutil_type - CPU utilization type
2484 * @FREQUENCY_UTIL: Utilization used to select frequency
2485 * @ENERGY_UTIL: Utilization used during energy calculation
2486 *
2487 * The utilization signals of all scheduling classes (CFS/RT/DL) and IRQ time
2488 * need to be aggregated differently depending on the usage made of them. This
2489 * enum is used within schedutil_freq_util() to differentiate the types of
2490 * utilization expected by the callers, and adjust the aggregation accordingly.
2491 */
2492enum schedutil_type {
2493 FREQUENCY_UTIL,
2494 ENERGY_UTIL,
2495};
2496
2497#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
2498
2499unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
2500 unsigned long max, enum schedutil_type type,
2501 struct task_struct *p);
2502
2503static inline unsigned long cpu_bw_dl(struct rq *rq)
2504{
2505 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
2506}
2507
2508static inline unsigned long cpu_util_dl(struct rq *rq)
2509{
2510 return READ_ONCE(rq->avg_dl.util_avg);
2511}
2512
2513static inline unsigned long cpu_util_cfs(struct rq *rq)
2514{
2515 unsigned long util = READ_ONCE(rq->cfs.avg.util_avg);
2516
2517 if (sched_feat(UTIL_EST)) {
2518 util = max_t(unsigned long, util,
2519 READ_ONCE(rq->cfs.avg.util_est.enqueued));
2520 }
2521
2522 return util;
2523}
2524
2525static inline unsigned long cpu_util_rt(struct rq *rq)
2526{
2527 return READ_ONCE(rq->avg_rt.util_avg);
2528}
2529#else /* CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
2530static inline unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
2531 unsigned long max, enum schedutil_type type,
2532 struct task_struct *p)
2533{
2534 return 0;
2535}
2536#endif /* CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
2537
2538#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
2539static inline unsigned long cpu_util_irq(struct rq *rq)
2540{
2541 return rq->avg_irq.util_avg;
2542}
2543
2544static inline
2545unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
2546{
2547 util *= (max - irq);
2548 util /= max;
2549
2550 return util;
2551
2552}
2553#else
2554static inline unsigned long cpu_util_irq(struct rq *rq)
2555{
2556 return 0;
2557}
2558
2559static inline
2560unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
2561{
2562 return util;
2563}
2564#endif
2565
2566#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2567
2568#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
2569
2570DECLARE_STATIC_KEY_FALSE(sched_energy_present);
2571
2572static inline bool sched_energy_enabled(void)
2573{
2574 return static_branch_unlikely(&sched_energy_present);
2575}
2576
2577#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
2578
2579#define perf_domain_span(pd) NULL
2580static inline bool sched_energy_enabled(void) { return false; }
2581
2582#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
2583
2584#ifdef CONFIG_MEMBARRIER
2585/*
2586 * The scheduler provides memory barriers required by membarrier between:
2587 * - prior user-space memory accesses and store to rq->membarrier_state,
2588 * - store to rq->membarrier_state and following user-space memory accesses.
2589 * In the same way it provides those guarantees around store to rq->curr.
2590 */
2591static inline void membarrier_switch_mm(struct rq *rq,
2592 struct mm_struct *prev_mm,
2593 struct mm_struct *next_mm)
2594{
2595 int membarrier_state;
2596
2597 if (prev_mm == next_mm)
2598 return;
2599
2600 membarrier_state = atomic_read(&next_mm->membarrier_state);
2601 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
2602 return;
2603
2604 WRITE_ONCE(rq->membarrier_state, membarrier_state);
2605}
2606#else
2607static inline void membarrier_switch_mm(struct rq *rq,
2608 struct mm_struct *prev_mm,
2609 struct mm_struct *next_mm)
2610{
2611}
2612#endif
2613
2614#ifdef CONFIG_SMP
2615static inline bool is_per_cpu_kthread(struct task_struct *p)
2616{
2617 if (!(p->flags & PF_KTHREAD))
2618 return false;
2619
2620 if (p->nr_cpus_allowed != 1)
2621 return false;
2622
2623 return true;
2624}
2625#endif
2626
2627void swake_up_all_locked(struct swait_queue_head *q);
2628void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);