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#ifndef _KERNEL_SCHED_SCHED_H
6#define _KERNEL_SCHED_SCHED_H
7
8#include <linux/sched/affinity.h>
9#include <linux/sched/autogroup.h>
10#include <linux/sched/cpufreq.h>
11#include <linux/sched/deadline.h>
12#include <linux/sched.h>
13#include <linux/sched/loadavg.h>
14#include <linux/sched/mm.h>
15#include <linux/sched/rseq_api.h>
16#include <linux/sched/signal.h>
17#include <linux/sched/smt.h>
18#include <linux/sched/stat.h>
19#include <linux/sched/sysctl.h>
20#include <linux/sched/task_flags.h>
21#include <linux/sched/task.h>
22#include <linux/sched/topology.h>
23
24#include <linux/atomic.h>
25#include <linux/bitmap.h>
26#include <linux/bug.h>
27#include <linux/capability.h>
28#include <linux/cgroup_api.h>
29#include <linux/cgroup.h>
30#include <linux/context_tracking.h>
31#include <linux/cpufreq.h>
32#include <linux/cpumask_api.h>
33#include <linux/ctype.h>
34#include <linux/file.h>
35#include <linux/fs_api.h>
36#include <linux/hrtimer_api.h>
37#include <linux/interrupt.h>
38#include <linux/irq_work.h>
39#include <linux/jiffies.h>
40#include <linux/kref_api.h>
41#include <linux/kthread.h>
42#include <linux/ktime_api.h>
43#include <linux/lockdep_api.h>
44#include <linux/lockdep.h>
45#include <linux/minmax.h>
46#include <linux/mm.h>
47#include <linux/module.h>
48#include <linux/mutex_api.h>
49#include <linux/plist.h>
50#include <linux/poll.h>
51#include <linux/proc_fs.h>
52#include <linux/profile.h>
53#include <linux/psi.h>
54#include <linux/rcupdate.h>
55#include <linux/seq_file.h>
56#include <linux/seqlock.h>
57#include <linux/softirq.h>
58#include <linux/spinlock_api.h>
59#include <linux/static_key.h>
60#include <linux/stop_machine.h>
61#include <linux/syscalls_api.h>
62#include <linux/syscalls.h>
63#include <linux/tick.h>
64#include <linux/topology.h>
65#include <linux/types.h>
66#include <linux/u64_stats_sync_api.h>
67#include <linux/uaccess.h>
68#include <linux/wait_api.h>
69#include <linux/wait_bit.h>
70#include <linux/workqueue_api.h>
71#include <linux/delayacct.h>
72
73#include <trace/events/power.h>
74#include <trace/events/sched.h>
75
76#include "../workqueue_internal.h"
77
78struct rq;
79struct cfs_rq;
80struct rt_rq;
81struct sched_group;
82struct cpuidle_state;
83
84#ifdef CONFIG_PARAVIRT
85# include <asm/paravirt.h>
86# include <asm/paravirt_api_clock.h>
87#endif
88
89#include <asm/barrier.h>
90
91#include "cpupri.h"
92#include "cpudeadline.h"
93
94#ifdef CONFIG_SCHED_DEBUG
95# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
96#else
97# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
98#endif
99
100/* task_struct::on_rq states: */
101#define TASK_ON_RQ_QUEUED 1
102#define TASK_ON_RQ_MIGRATING 2
103
104extern __read_mostly int scheduler_running;
105
106extern unsigned long calc_load_update;
107extern atomic_long_t calc_load_tasks;
108
109extern void calc_global_load_tick(struct rq *this_rq);
110extern long calc_load_fold_active(struct rq *this_rq, long adjust);
111
112extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
113
114extern int sysctl_sched_rt_period;
115extern int sysctl_sched_rt_runtime;
116extern int sched_rr_timeslice;
117
118/*
119 * Asymmetric CPU capacity bits
120 */
121struct asym_cap_data {
122 struct list_head link;
123 struct rcu_head rcu;
124 unsigned long capacity;
125 unsigned long cpus[];
126};
127
128extern struct list_head asym_cap_list;
129
130#define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
131
132/*
133 * Helpers for converting nanosecond timing to jiffy resolution
134 */
135#define NS_TO_JIFFIES(time) ((unsigned long)(time) / (NSEC_PER_SEC/HZ))
136
137/*
138 * Increase resolution of nice-level calculations for 64-bit architectures.
139 * The extra resolution improves shares distribution and load balancing of
140 * low-weight task groups (eg. nice +19 on an autogroup), deeper task-group
141 * hierarchies, especially on larger systems. This is not a user-visible change
142 * and does not change the user-interface for setting shares/weights.
143 *
144 * We increase resolution only if we have enough bits to allow this increased
145 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
146 * are pretty high and the returns do not justify the increased costs.
147 *
148 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
149 * increase coverage and consistency always enable it on 64-bit platforms.
150 */
151#ifdef CONFIG_64BIT
152# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
153# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
154# define scale_load_down(w) \
155({ \
156 unsigned long __w = (w); \
157 \
158 if (__w) \
159 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
160 __w; \
161})
162#else
163# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
164# define scale_load(w) (w)
165# define scale_load_down(w) (w)
166#endif
167
168/*
169 * Task weight (visible to users) and its load (invisible to users) have
170 * independent resolution, but they should be well calibrated. We use
171 * scale_load() and scale_load_down(w) to convert between them. The
172 * following must be true:
173 *
174 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
175 *
176 */
177#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
178
179/*
180 * Single value that decides SCHED_DEADLINE internal math precision.
181 * 10 -> just above 1us
182 * 9 -> just above 0.5us
183 */
184#define DL_SCALE 10
185
186/*
187 * Single value that denotes runtime == period, ie unlimited time.
188 */
189#define RUNTIME_INF ((u64)~0ULL)
190
191static inline int idle_policy(int policy)
192{
193 return policy == SCHED_IDLE;
194}
195
196static inline int normal_policy(int policy)
197{
198#ifdef CONFIG_SCHED_CLASS_EXT
199 if (policy == SCHED_EXT)
200 return true;
201#endif
202 return policy == SCHED_NORMAL;
203}
204
205static inline int fair_policy(int policy)
206{
207 return normal_policy(policy) || policy == SCHED_BATCH;
208}
209
210static inline int rt_policy(int policy)
211{
212 return policy == SCHED_FIFO || policy == SCHED_RR;
213}
214
215static inline int dl_policy(int policy)
216{
217 return policy == SCHED_DEADLINE;
218}
219
220static inline bool valid_policy(int policy)
221{
222 return idle_policy(policy) || fair_policy(policy) ||
223 rt_policy(policy) || dl_policy(policy);
224}
225
226static inline int task_has_idle_policy(struct task_struct *p)
227{
228 return idle_policy(p->policy);
229}
230
231static inline int task_has_rt_policy(struct task_struct *p)
232{
233 return rt_policy(p->policy);
234}
235
236static inline int task_has_dl_policy(struct task_struct *p)
237{
238 return dl_policy(p->policy);
239}
240
241#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
242
243static inline void update_avg(u64 *avg, u64 sample)
244{
245 s64 diff = sample - *avg;
246
247 *avg += diff / 8;
248}
249
250/*
251 * Shifting a value by an exponent greater *or equal* to the size of said value
252 * is UB; cap at size-1.
253 */
254#define shr_bound(val, shift) \
255 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
256
257/*
258 * cgroup weight knobs should use the common MIN, DFL and MAX values which are
259 * 1, 100 and 10000 respectively. While it loses a bit of range on both ends, it
260 * maps pretty well onto the shares value used by scheduler and the round-trip
261 * conversions preserve the original value over the entire range.
262 */
263static inline unsigned long sched_weight_from_cgroup(unsigned long cgrp_weight)
264{
265 return DIV_ROUND_CLOSEST_ULL(cgrp_weight * 1024, CGROUP_WEIGHT_DFL);
266}
267
268static inline unsigned long sched_weight_to_cgroup(unsigned long weight)
269{
270 return clamp_t(unsigned long,
271 DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024),
272 CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
273}
274
275/*
276 * !! For sched_setattr_nocheck() (kernel) only !!
277 *
278 * This is actually gross. :(
279 *
280 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
281 * tasks, but still be able to sleep. We need this on platforms that cannot
282 * atomically change clock frequency. Remove once fast switching will be
283 * available on such platforms.
284 *
285 * SUGOV stands for SchedUtil GOVernor.
286 */
287#define SCHED_FLAG_SUGOV 0x10000000
288
289#define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
290
291static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
292{
293#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
294 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
295#else
296 return false;
297#endif
298}
299
300/*
301 * Tells if entity @a should preempt entity @b.
302 */
303static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
304 const struct sched_dl_entity *b)
305{
306 return dl_entity_is_special(a) ||
307 dl_time_before(a->deadline, b->deadline);
308}
309
310/*
311 * This is the priority-queue data structure of the RT scheduling class:
312 */
313struct rt_prio_array {
314 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
315 struct list_head queue[MAX_RT_PRIO];
316};
317
318struct rt_bandwidth {
319 /* nests inside the rq lock: */
320 raw_spinlock_t rt_runtime_lock;
321 ktime_t rt_period;
322 u64 rt_runtime;
323 struct hrtimer rt_period_timer;
324 unsigned int rt_period_active;
325};
326
327static inline int dl_bandwidth_enabled(void)
328{
329 return sysctl_sched_rt_runtime >= 0;
330}
331
332/*
333 * To keep the bandwidth of -deadline tasks under control
334 * we need some place where:
335 * - store the maximum -deadline bandwidth of each cpu;
336 * - cache the fraction of bandwidth that is currently allocated in
337 * each root domain;
338 *
339 * This is all done in the data structure below. It is similar to the
340 * one used for RT-throttling (rt_bandwidth), with the main difference
341 * that, since here we are only interested in admission control, we
342 * do not decrease any runtime while the group "executes", neither we
343 * need a timer to replenish it.
344 *
345 * With respect to SMP, bandwidth is given on a per root domain basis,
346 * meaning that:
347 * - bw (< 100%) is the deadline bandwidth of each CPU;
348 * - total_bw is the currently allocated bandwidth in each root domain;
349 */
350struct dl_bw {
351 raw_spinlock_t lock;
352 u64 bw;
353 u64 total_bw;
354};
355
356extern void init_dl_bw(struct dl_bw *dl_b);
357extern int sched_dl_global_validate(void);
358extern void sched_dl_do_global(void);
359extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
360extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
361extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
362extern bool __checkparam_dl(const struct sched_attr *attr);
363extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
364extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
365extern int dl_bw_check_overflow(int cpu);
366extern s64 dl_scaled_delta_exec(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec);
367/*
368 * SCHED_DEADLINE supports servers (nested scheduling) with the following
369 * interface:
370 *
371 * dl_se::rq -- runqueue we belong to.
372 *
373 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
374 * server when it runs out of tasks to run.
375 *
376 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
377 * returns NULL.
378 *
379 * dl_server_update() -- called from update_curr_common(), propagates runtime
380 * to the server.
381 *
382 * dl_server_start()
383 * dl_server_stop() -- start/stop the server when it has (no) tasks.
384 *
385 * dl_server_init() -- initializes the server.
386 */
387extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
388extern void dl_server_start(struct sched_dl_entity *dl_se);
389extern void dl_server_stop(struct sched_dl_entity *dl_se);
390extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
391 dl_server_has_tasks_f has_tasks,
392 dl_server_pick_f pick_task);
393
394extern void dl_server_update_idle_time(struct rq *rq,
395 struct task_struct *p);
396extern void fair_server_init(struct rq *rq);
397extern void __dl_server_attach_root(struct sched_dl_entity *dl_se, struct rq *rq);
398extern int dl_server_apply_params(struct sched_dl_entity *dl_se,
399 u64 runtime, u64 period, bool init);
400
401static inline bool dl_server_active(struct sched_dl_entity *dl_se)
402{
403 return dl_se->dl_server_active;
404}
405
406#ifdef CONFIG_CGROUP_SCHED
407
408extern struct list_head task_groups;
409
410struct cfs_bandwidth {
411#ifdef CONFIG_CFS_BANDWIDTH
412 raw_spinlock_t lock;
413 ktime_t period;
414 u64 quota;
415 u64 runtime;
416 u64 burst;
417 u64 runtime_snap;
418 s64 hierarchical_quota;
419
420 u8 idle;
421 u8 period_active;
422 u8 slack_started;
423 struct hrtimer period_timer;
424 struct hrtimer slack_timer;
425 struct list_head throttled_cfs_rq;
426
427 /* Statistics: */
428 int nr_periods;
429 int nr_throttled;
430 int nr_burst;
431 u64 throttled_time;
432 u64 burst_time;
433#endif
434};
435
436/* Task group related information */
437struct task_group {
438 struct cgroup_subsys_state css;
439
440#ifdef CONFIG_GROUP_SCHED_WEIGHT
441 /* A positive value indicates that this is a SCHED_IDLE group. */
442 int idle;
443#endif
444
445#ifdef CONFIG_FAIR_GROUP_SCHED
446 /* schedulable entities of this group on each CPU */
447 struct sched_entity **se;
448 /* runqueue "owned" by this group on each CPU */
449 struct cfs_rq **cfs_rq;
450 unsigned long shares;
451#ifdef CONFIG_SMP
452 /*
453 * load_avg can be heavily contended at clock tick time, so put
454 * it in its own cache-line separated from the fields above which
455 * will also be accessed at each tick.
456 */
457 atomic_long_t load_avg ____cacheline_aligned;
458#endif
459#endif
460
461#ifdef CONFIG_RT_GROUP_SCHED
462 struct sched_rt_entity **rt_se;
463 struct rt_rq **rt_rq;
464
465 struct rt_bandwidth rt_bandwidth;
466#endif
467
468#ifdef CONFIG_EXT_GROUP_SCHED
469 u32 scx_flags; /* SCX_TG_* */
470 u32 scx_weight;
471#endif
472
473 struct rcu_head rcu;
474 struct list_head list;
475
476 struct task_group *parent;
477 struct list_head siblings;
478 struct list_head children;
479
480#ifdef CONFIG_SCHED_AUTOGROUP
481 struct autogroup *autogroup;
482#endif
483
484 struct cfs_bandwidth cfs_bandwidth;
485
486#ifdef CONFIG_UCLAMP_TASK_GROUP
487 /* The two decimal precision [%] value requested from user-space */
488 unsigned int uclamp_pct[UCLAMP_CNT];
489 /* Clamp values requested for a task group */
490 struct uclamp_se uclamp_req[UCLAMP_CNT];
491 /* Effective clamp values used for a task group */
492 struct uclamp_se uclamp[UCLAMP_CNT];
493#endif
494
495};
496
497#ifdef CONFIG_GROUP_SCHED_WEIGHT
498#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
499
500/*
501 * A weight of 0 or 1 can cause arithmetics problems.
502 * A weight of a cfs_rq is the sum of weights of which entities
503 * are queued on this cfs_rq, so a weight of a entity should not be
504 * too large, so as the shares value of a task group.
505 * (The default weight is 1024 - so there's no practical
506 * limitation from this.)
507 */
508#define MIN_SHARES (1UL << 1)
509#define MAX_SHARES (1UL << 18)
510#endif
511
512typedef int (*tg_visitor)(struct task_group *, void *);
513
514extern int walk_tg_tree_from(struct task_group *from,
515 tg_visitor down, tg_visitor up, void *data);
516
517/*
518 * Iterate the full tree, calling @down when first entering a node and @up when
519 * leaving it for the final time.
520 *
521 * Caller must hold rcu_lock or sufficient equivalent.
522 */
523static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
524{
525 return walk_tg_tree_from(&root_task_group, down, up, data);
526}
527
528static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
529{
530 return css ? container_of(css, struct task_group, css) : NULL;
531}
532
533extern int tg_nop(struct task_group *tg, void *data);
534
535#ifdef CONFIG_FAIR_GROUP_SCHED
536extern void free_fair_sched_group(struct task_group *tg);
537extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
538extern void online_fair_sched_group(struct task_group *tg);
539extern void unregister_fair_sched_group(struct task_group *tg);
540#else
541static inline void free_fair_sched_group(struct task_group *tg) { }
542static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
543{
544 return 1;
545}
546static inline void online_fair_sched_group(struct task_group *tg) { }
547static inline void unregister_fair_sched_group(struct task_group *tg) { }
548#endif
549
550extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
551 struct sched_entity *se, int cpu,
552 struct sched_entity *parent);
553extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
554
555extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
556extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
557extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
558extern bool cfs_task_bw_constrained(struct task_struct *p);
559
560extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
561 struct sched_rt_entity *rt_se, int cpu,
562 struct sched_rt_entity *parent);
563extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
564extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
565extern long sched_group_rt_runtime(struct task_group *tg);
566extern long sched_group_rt_period(struct task_group *tg);
567extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
568
569extern struct task_group *sched_create_group(struct task_group *parent);
570extern void sched_online_group(struct task_group *tg,
571 struct task_group *parent);
572extern void sched_destroy_group(struct task_group *tg);
573extern void sched_release_group(struct task_group *tg);
574
575extern void sched_move_task(struct task_struct *tsk, bool for_autogroup);
576
577#ifdef CONFIG_FAIR_GROUP_SCHED
578extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
579
580extern int sched_group_set_idle(struct task_group *tg, long idle);
581
582#ifdef CONFIG_SMP
583extern void set_task_rq_fair(struct sched_entity *se,
584 struct cfs_rq *prev, struct cfs_rq *next);
585#else /* !CONFIG_SMP */
586static inline void set_task_rq_fair(struct sched_entity *se,
587 struct cfs_rq *prev, struct cfs_rq *next) { }
588#endif /* CONFIG_SMP */
589#else /* !CONFIG_FAIR_GROUP_SCHED */
590static inline int sched_group_set_shares(struct task_group *tg, unsigned long shares) { return 0; }
591static inline int sched_group_set_idle(struct task_group *tg, long idle) { return 0; }
592#endif /* CONFIG_FAIR_GROUP_SCHED */
593
594#else /* CONFIG_CGROUP_SCHED */
595
596struct cfs_bandwidth { };
597
598static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
599
600#endif /* CONFIG_CGROUP_SCHED */
601
602extern void unregister_rt_sched_group(struct task_group *tg);
603extern void free_rt_sched_group(struct task_group *tg);
604extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
605
606/*
607 * u64_u32_load/u64_u32_store
608 *
609 * Use a copy of a u64 value to protect against data race. This is only
610 * applicable for 32-bits architectures.
611 */
612#ifdef CONFIG_64BIT
613# define u64_u32_load_copy(var, copy) var
614# define u64_u32_store_copy(var, copy, val) (var = val)
615#else
616# define u64_u32_load_copy(var, copy) \
617({ \
618 u64 __val, __val_copy; \
619 do { \
620 __val_copy = copy; \
621 /* \
622 * paired with u64_u32_store_copy(), ordering access \
623 * to var and copy. \
624 */ \
625 smp_rmb(); \
626 __val = var; \
627 } while (__val != __val_copy); \
628 __val; \
629})
630# define u64_u32_store_copy(var, copy, val) \
631do { \
632 typeof(val) __val = (val); \
633 var = __val; \
634 /* \
635 * paired with u64_u32_load_copy(), ordering access to var and \
636 * copy. \
637 */ \
638 smp_wmb(); \
639 copy = __val; \
640} while (0)
641#endif
642# define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
643# define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
644
645struct balance_callback {
646 struct balance_callback *next;
647 void (*func)(struct rq *rq);
648};
649
650/* CFS-related fields in a runqueue */
651struct cfs_rq {
652 struct load_weight load;
653 unsigned int nr_running;
654 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
655 unsigned int idle_nr_running; /* SCHED_IDLE */
656 unsigned int idle_h_nr_running; /* SCHED_IDLE */
657 unsigned int h_nr_delayed;
658
659 s64 avg_vruntime;
660 u64 avg_load;
661
662 u64 min_vruntime;
663#ifdef CONFIG_SCHED_CORE
664 unsigned int forceidle_seq;
665 u64 min_vruntime_fi;
666#endif
667
668 struct rb_root_cached tasks_timeline;
669
670 /*
671 * 'curr' points to currently running entity on this cfs_rq.
672 * It is set to NULL otherwise (i.e when none are currently running).
673 */
674 struct sched_entity *curr;
675 struct sched_entity *next;
676
677#ifdef CONFIG_SMP
678 /*
679 * CFS load tracking
680 */
681 struct sched_avg avg;
682#ifndef CONFIG_64BIT
683 u64 last_update_time_copy;
684#endif
685 struct {
686 raw_spinlock_t lock ____cacheline_aligned;
687 int nr;
688 unsigned long load_avg;
689 unsigned long util_avg;
690 unsigned long runnable_avg;
691 } removed;
692
693#ifdef CONFIG_FAIR_GROUP_SCHED
694 u64 last_update_tg_load_avg;
695 unsigned long tg_load_avg_contrib;
696 long propagate;
697 long prop_runnable_sum;
698
699 /*
700 * h_load = weight * f(tg)
701 *
702 * Where f(tg) is the recursive weight fraction assigned to
703 * this group.
704 */
705 unsigned long h_load;
706 u64 last_h_load_update;
707 struct sched_entity *h_load_next;
708#endif /* CONFIG_FAIR_GROUP_SCHED */
709#endif /* CONFIG_SMP */
710
711#ifdef CONFIG_FAIR_GROUP_SCHED
712 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
713
714 /*
715 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
716 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
717 * (like users, containers etc.)
718 *
719 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
720 * This list is used during load balance.
721 */
722 int on_list;
723 struct list_head leaf_cfs_rq_list;
724 struct task_group *tg; /* group that "owns" this runqueue */
725
726 /* Locally cached copy of our task_group's idle value */
727 int idle;
728
729#ifdef CONFIG_CFS_BANDWIDTH
730 int runtime_enabled;
731 s64 runtime_remaining;
732
733 u64 throttled_pelt_idle;
734#ifndef CONFIG_64BIT
735 u64 throttled_pelt_idle_copy;
736#endif
737 u64 throttled_clock;
738 u64 throttled_clock_pelt;
739 u64 throttled_clock_pelt_time;
740 u64 throttled_clock_self;
741 u64 throttled_clock_self_time;
742 int throttled;
743 int throttle_count;
744 struct list_head throttled_list;
745 struct list_head throttled_csd_list;
746#endif /* CONFIG_CFS_BANDWIDTH */
747#endif /* CONFIG_FAIR_GROUP_SCHED */
748};
749
750#ifdef CONFIG_SCHED_CLASS_EXT
751/* scx_rq->flags, protected by the rq lock */
752enum scx_rq_flags {
753 /*
754 * A hotplugged CPU starts scheduling before rq_online_scx(). Track
755 * ops.cpu_on/offline() state so that ops.enqueue/dispatch() are called
756 * only while the BPF scheduler considers the CPU to be online.
757 */
758 SCX_RQ_ONLINE = 1 << 0,
759 SCX_RQ_CAN_STOP_TICK = 1 << 1,
760 SCX_RQ_BAL_PENDING = 1 << 2, /* balance hasn't run yet */
761 SCX_RQ_BAL_KEEP = 1 << 3, /* balance decided to keep current */
762 SCX_RQ_BYPASSING = 1 << 4,
763
764 SCX_RQ_IN_WAKEUP = 1 << 16,
765 SCX_RQ_IN_BALANCE = 1 << 17,
766};
767
768struct scx_rq {
769 struct scx_dispatch_q local_dsq;
770 struct list_head runnable_list; /* runnable tasks on this rq */
771 struct list_head ddsp_deferred_locals; /* deferred ddsps from enq */
772 unsigned long ops_qseq;
773 u64 extra_enq_flags; /* see move_task_to_local_dsq() */
774 u32 nr_running;
775 u32 flags;
776 u32 cpuperf_target; /* [0, SCHED_CAPACITY_SCALE] */
777 bool cpu_released;
778 cpumask_var_t cpus_to_kick;
779 cpumask_var_t cpus_to_kick_if_idle;
780 cpumask_var_t cpus_to_preempt;
781 cpumask_var_t cpus_to_wait;
782 unsigned long pnt_seq;
783 struct balance_callback deferred_bal_cb;
784 struct irq_work deferred_irq_work;
785 struct irq_work kick_cpus_irq_work;
786};
787#endif /* CONFIG_SCHED_CLASS_EXT */
788
789static inline int rt_bandwidth_enabled(void)
790{
791 return sysctl_sched_rt_runtime >= 0;
792}
793
794/* RT IPI pull logic requires IRQ_WORK */
795#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
796# define HAVE_RT_PUSH_IPI
797#endif
798
799/* Real-Time classes' related field in a runqueue: */
800struct rt_rq {
801 struct rt_prio_array active;
802 unsigned int rt_nr_running;
803 unsigned int rr_nr_running;
804#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
805 struct {
806 int curr; /* highest queued rt task prio */
807#ifdef CONFIG_SMP
808 int next; /* next highest */
809#endif
810 } highest_prio;
811#endif
812#ifdef CONFIG_SMP
813 bool overloaded;
814 struct plist_head pushable_tasks;
815
816#endif /* CONFIG_SMP */
817 int rt_queued;
818
819#ifdef CONFIG_RT_GROUP_SCHED
820 int rt_throttled;
821 u64 rt_time;
822 u64 rt_runtime;
823 /* Nests inside the rq lock: */
824 raw_spinlock_t rt_runtime_lock;
825
826 unsigned int rt_nr_boosted;
827
828 struct rq *rq;
829 struct task_group *tg;
830#endif
831};
832
833static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
834{
835 return rt_rq->rt_queued && rt_rq->rt_nr_running;
836}
837
838/* Deadline class' related fields in a runqueue */
839struct dl_rq {
840 /* runqueue is an rbtree, ordered by deadline */
841 struct rb_root_cached root;
842
843 unsigned int dl_nr_running;
844
845#ifdef CONFIG_SMP
846 /*
847 * Deadline values of the currently executing and the
848 * earliest ready task on this rq. Caching these facilitates
849 * the decision whether or not a ready but not running task
850 * should migrate somewhere else.
851 */
852 struct {
853 u64 curr;
854 u64 next;
855 } earliest_dl;
856
857 bool overloaded;
858
859 /*
860 * Tasks on this rq that can be pushed away. They are kept in
861 * an rb-tree, ordered by tasks' deadlines, with caching
862 * of the leftmost (earliest deadline) element.
863 */
864 struct rb_root_cached pushable_dl_tasks_root;
865#else
866 struct dl_bw dl_bw;
867#endif
868 /*
869 * "Active utilization" for this runqueue: increased when a
870 * task wakes up (becomes TASK_RUNNING) and decreased when a
871 * task blocks
872 */
873 u64 running_bw;
874
875 /*
876 * Utilization of the tasks "assigned" to this runqueue (including
877 * the tasks that are in runqueue and the tasks that executed on this
878 * CPU and blocked). Increased when a task moves to this runqueue, and
879 * decreased when the task moves away (migrates, changes scheduling
880 * policy, or terminates).
881 * This is needed to compute the "inactive utilization" for the
882 * runqueue (inactive utilization = this_bw - running_bw).
883 */
884 u64 this_bw;
885 u64 extra_bw;
886
887 /*
888 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
889 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
890 */
891 u64 max_bw;
892
893 /*
894 * Inverse of the fraction of CPU utilization that can be reclaimed
895 * by the GRUB algorithm.
896 */
897 u64 bw_ratio;
898};
899
900#ifdef CONFIG_FAIR_GROUP_SCHED
901
902/* An entity is a task if it doesn't "own" a runqueue */
903#define entity_is_task(se) (!se->my_q)
904
905static inline void se_update_runnable(struct sched_entity *se)
906{
907 if (!entity_is_task(se)) {
908 struct cfs_rq *cfs_rq = se->my_q;
909
910 se->runnable_weight = cfs_rq->h_nr_running - cfs_rq->h_nr_delayed;
911 }
912}
913
914static inline long se_runnable(struct sched_entity *se)
915{
916 if (se->sched_delayed)
917 return false;
918
919 if (entity_is_task(se))
920 return !!se->on_rq;
921 else
922 return se->runnable_weight;
923}
924
925#else /* !CONFIG_FAIR_GROUP_SCHED: */
926
927#define entity_is_task(se) 1
928
929static inline void se_update_runnable(struct sched_entity *se) { }
930
931static inline long se_runnable(struct sched_entity *se)
932{
933 if (se->sched_delayed)
934 return false;
935
936 return !!se->on_rq;
937}
938
939#endif /* !CONFIG_FAIR_GROUP_SCHED */
940
941#ifdef CONFIG_SMP
942/*
943 * XXX we want to get rid of these helpers and use the full load resolution.
944 */
945static inline long se_weight(struct sched_entity *se)
946{
947 return scale_load_down(se->load.weight);
948}
949
950
951static inline bool sched_asym_prefer(int a, int b)
952{
953 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
954}
955
956struct perf_domain {
957 struct em_perf_domain *em_pd;
958 struct perf_domain *next;
959 struct rcu_head rcu;
960};
961
962/*
963 * We add the notion of a root-domain which will be used to define per-domain
964 * variables. Each exclusive cpuset essentially defines an island domain by
965 * fully partitioning the member CPUs from any other cpuset. Whenever a new
966 * exclusive cpuset is created, we also create and attach a new root-domain
967 * object.
968 *
969 */
970struct root_domain {
971 atomic_t refcount;
972 atomic_t rto_count;
973 struct rcu_head rcu;
974 cpumask_var_t span;
975 cpumask_var_t online;
976
977 /*
978 * Indicate pullable load on at least one CPU, e.g:
979 * - More than one runnable task
980 * - Running task is misfit
981 */
982 bool overloaded;
983
984 /* Indicate one or more CPUs over-utilized (tipping point) */
985 bool overutilized;
986
987 /*
988 * The bit corresponding to a CPU gets set here if such CPU has more
989 * than one runnable -deadline task (as it is below for RT tasks).
990 */
991 cpumask_var_t dlo_mask;
992 atomic_t dlo_count;
993 struct dl_bw dl_bw;
994 struct cpudl cpudl;
995
996 /*
997 * Indicate whether a root_domain's dl_bw has been checked or
998 * updated. It's monotonously increasing value.
999 *
1000 * Also, some corner cases, like 'wrap around' is dangerous, but given
1001 * that u64 is 'big enough'. So that shouldn't be a concern.
1002 */
1003 u64 visit_gen;
1004
1005#ifdef HAVE_RT_PUSH_IPI
1006 /*
1007 * For IPI pull requests, loop across the rto_mask.
1008 */
1009 struct irq_work rto_push_work;
1010 raw_spinlock_t rto_lock;
1011 /* These are only updated and read within rto_lock */
1012 int rto_loop;
1013 int rto_cpu;
1014 /* These atomics are updated outside of a lock */
1015 atomic_t rto_loop_next;
1016 atomic_t rto_loop_start;
1017#endif
1018 /*
1019 * The "RT overload" flag: it gets set if a CPU has more than
1020 * one runnable RT task.
1021 */
1022 cpumask_var_t rto_mask;
1023 struct cpupri cpupri;
1024
1025 /*
1026 * NULL-terminated list of performance domains intersecting with the
1027 * CPUs of the rd. Protected by RCU.
1028 */
1029 struct perf_domain __rcu *pd;
1030};
1031
1032extern void init_defrootdomain(void);
1033extern int sched_init_domains(const struct cpumask *cpu_map);
1034extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
1035extern void sched_get_rd(struct root_domain *rd);
1036extern void sched_put_rd(struct root_domain *rd);
1037
1038static inline int get_rd_overloaded(struct root_domain *rd)
1039{
1040 return READ_ONCE(rd->overloaded);
1041}
1042
1043static inline void set_rd_overloaded(struct root_domain *rd, int status)
1044{
1045 if (get_rd_overloaded(rd) != status)
1046 WRITE_ONCE(rd->overloaded, status);
1047}
1048
1049#ifdef HAVE_RT_PUSH_IPI
1050extern void rto_push_irq_work_func(struct irq_work *work);
1051#endif
1052#endif /* CONFIG_SMP */
1053
1054#ifdef CONFIG_UCLAMP_TASK
1055/*
1056 * struct uclamp_bucket - Utilization clamp bucket
1057 * @value: utilization clamp value for tasks on this clamp bucket
1058 * @tasks: number of RUNNABLE tasks on this clamp bucket
1059 *
1060 * Keep track of how many tasks are RUNNABLE for a given utilization
1061 * clamp value.
1062 */
1063struct uclamp_bucket {
1064 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
1065 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
1066};
1067
1068/*
1069 * struct uclamp_rq - rq's utilization clamp
1070 * @value: currently active clamp values for a rq
1071 * @bucket: utilization clamp buckets affecting a rq
1072 *
1073 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
1074 * A clamp value is affecting a rq when there is at least one task RUNNABLE
1075 * (or actually running) with that value.
1076 *
1077 * There are up to UCLAMP_CNT possible different clamp values, currently there
1078 * are only two: minimum utilization and maximum utilization.
1079 *
1080 * All utilization clamping values are MAX aggregated, since:
1081 * - for util_min: we want to run the CPU at least at the max of the minimum
1082 * utilization required by its currently RUNNABLE tasks.
1083 * - for util_max: we want to allow the CPU to run up to the max of the
1084 * maximum utilization allowed by its currently RUNNABLE tasks.
1085 *
1086 * Since on each system we expect only a limited number of different
1087 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
1088 * the metrics required to compute all the per-rq utilization clamp values.
1089 */
1090struct uclamp_rq {
1091 unsigned int value;
1092 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
1093};
1094
1095DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
1096#endif /* CONFIG_UCLAMP_TASK */
1097
1098/*
1099 * This is the main, per-CPU runqueue data structure.
1100 *
1101 * Locking rule: those places that want to lock multiple runqueues
1102 * (such as the load balancing or the thread migration code), lock
1103 * acquire operations must be ordered by ascending &runqueue.
1104 */
1105struct rq {
1106 /* runqueue lock: */
1107 raw_spinlock_t __lock;
1108
1109 unsigned int nr_running;
1110#ifdef CONFIG_NUMA_BALANCING
1111 unsigned int nr_numa_running;
1112 unsigned int nr_preferred_running;
1113 unsigned int numa_migrate_on;
1114#endif
1115#ifdef CONFIG_NO_HZ_COMMON
1116#ifdef CONFIG_SMP
1117 unsigned long last_blocked_load_update_tick;
1118 unsigned int has_blocked_load;
1119 call_single_data_t nohz_csd;
1120#endif /* CONFIG_SMP */
1121 unsigned int nohz_tick_stopped;
1122 atomic_t nohz_flags;
1123#endif /* CONFIG_NO_HZ_COMMON */
1124
1125#ifdef CONFIG_SMP
1126 unsigned int ttwu_pending;
1127#endif
1128 u64 nr_switches;
1129
1130#ifdef CONFIG_UCLAMP_TASK
1131 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1132 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1133 unsigned int uclamp_flags;
1134#define UCLAMP_FLAG_IDLE 0x01
1135#endif
1136
1137 struct cfs_rq cfs;
1138 struct rt_rq rt;
1139 struct dl_rq dl;
1140#ifdef CONFIG_SCHED_CLASS_EXT
1141 struct scx_rq scx;
1142#endif
1143
1144 struct sched_dl_entity fair_server;
1145
1146#ifdef CONFIG_FAIR_GROUP_SCHED
1147 /* list of leaf cfs_rq on this CPU: */
1148 struct list_head leaf_cfs_rq_list;
1149 struct list_head *tmp_alone_branch;
1150#endif /* CONFIG_FAIR_GROUP_SCHED */
1151
1152 /*
1153 * This is part of a global counter where only the total sum
1154 * over all CPUs matters. A task can increase this counter on
1155 * one CPU and if it got migrated afterwards it may decrease
1156 * it on another CPU. Always updated under the runqueue lock:
1157 */
1158 unsigned int nr_uninterruptible;
1159
1160 union {
1161 struct task_struct __rcu *donor; /* Scheduler context */
1162 struct task_struct __rcu *curr; /* Execution context */
1163 };
1164 struct sched_dl_entity *dl_server;
1165 struct task_struct *idle;
1166 struct task_struct *stop;
1167 unsigned long next_balance;
1168 struct mm_struct *prev_mm;
1169
1170 unsigned int clock_update_flags;
1171 u64 clock;
1172 /* Ensure that all clocks are in the same cache line */
1173 u64 clock_task ____cacheline_aligned;
1174 u64 clock_pelt;
1175 unsigned long lost_idle_time;
1176 u64 clock_pelt_idle;
1177 u64 clock_idle;
1178#ifndef CONFIG_64BIT
1179 u64 clock_pelt_idle_copy;
1180 u64 clock_idle_copy;
1181#endif
1182
1183 atomic_t nr_iowait;
1184
1185#ifdef CONFIG_SCHED_DEBUG
1186 u64 last_seen_need_resched_ns;
1187 int ticks_without_resched;
1188#endif
1189
1190#ifdef CONFIG_MEMBARRIER
1191 int membarrier_state;
1192#endif
1193
1194#ifdef CONFIG_SMP
1195 struct root_domain *rd;
1196 struct sched_domain __rcu *sd;
1197
1198 unsigned long cpu_capacity;
1199
1200 struct balance_callback *balance_callback;
1201
1202 unsigned char nohz_idle_balance;
1203 unsigned char idle_balance;
1204
1205 unsigned long misfit_task_load;
1206
1207 /* For active balancing */
1208 int active_balance;
1209 int push_cpu;
1210 struct cpu_stop_work active_balance_work;
1211
1212 /* CPU of this runqueue: */
1213 int cpu;
1214 int online;
1215
1216 struct list_head cfs_tasks;
1217
1218 struct sched_avg avg_rt;
1219 struct sched_avg avg_dl;
1220#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1221 struct sched_avg avg_irq;
1222#endif
1223#ifdef CONFIG_SCHED_HW_PRESSURE
1224 struct sched_avg avg_hw;
1225#endif
1226 u64 idle_stamp;
1227 u64 avg_idle;
1228
1229 /* This is used to determine avg_idle's max value */
1230 u64 max_idle_balance_cost;
1231
1232#ifdef CONFIG_HOTPLUG_CPU
1233 struct rcuwait hotplug_wait;
1234#endif
1235#endif /* CONFIG_SMP */
1236
1237#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1238 u64 prev_irq_time;
1239 u64 psi_irq_time;
1240#endif
1241#ifdef CONFIG_PARAVIRT
1242 u64 prev_steal_time;
1243#endif
1244#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1245 u64 prev_steal_time_rq;
1246#endif
1247
1248 /* calc_load related fields */
1249 unsigned long calc_load_update;
1250 long calc_load_active;
1251
1252#ifdef CONFIG_SCHED_HRTICK
1253#ifdef CONFIG_SMP
1254 call_single_data_t hrtick_csd;
1255#endif
1256 struct hrtimer hrtick_timer;
1257 ktime_t hrtick_time;
1258#endif
1259
1260#ifdef CONFIG_SCHEDSTATS
1261 /* latency stats */
1262 struct sched_info rq_sched_info;
1263 unsigned long long rq_cpu_time;
1264
1265 /* sys_sched_yield() stats */
1266 unsigned int yld_count;
1267
1268 /* schedule() stats */
1269 unsigned int sched_count;
1270 unsigned int sched_goidle;
1271
1272 /* try_to_wake_up() stats */
1273 unsigned int ttwu_count;
1274 unsigned int ttwu_local;
1275#endif
1276
1277#ifdef CONFIG_CPU_IDLE
1278 /* Must be inspected within a RCU lock section */
1279 struct cpuidle_state *idle_state;
1280#endif
1281
1282#ifdef CONFIG_SMP
1283 unsigned int nr_pinned;
1284#endif
1285 unsigned int push_busy;
1286 struct cpu_stop_work push_work;
1287
1288#ifdef CONFIG_SCHED_CORE
1289 /* per rq */
1290 struct rq *core;
1291 struct task_struct *core_pick;
1292 struct sched_dl_entity *core_dl_server;
1293 unsigned int core_enabled;
1294 unsigned int core_sched_seq;
1295 struct rb_root core_tree;
1296
1297 /* shared state -- careful with sched_core_cpu_deactivate() */
1298 unsigned int core_task_seq;
1299 unsigned int core_pick_seq;
1300 unsigned long core_cookie;
1301 unsigned int core_forceidle_count;
1302 unsigned int core_forceidle_seq;
1303 unsigned int core_forceidle_occupation;
1304 u64 core_forceidle_start;
1305#endif
1306
1307 /* Scratch cpumask to be temporarily used under rq_lock */
1308 cpumask_var_t scratch_mask;
1309
1310#if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1311 call_single_data_t cfsb_csd;
1312 struct list_head cfsb_csd_list;
1313#endif
1314};
1315
1316#ifdef CONFIG_FAIR_GROUP_SCHED
1317
1318/* CPU runqueue to which this cfs_rq is attached */
1319static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1320{
1321 return cfs_rq->rq;
1322}
1323
1324#else
1325
1326static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1327{
1328 return container_of(cfs_rq, struct rq, cfs);
1329}
1330#endif
1331
1332static inline int cpu_of(struct rq *rq)
1333{
1334#ifdef CONFIG_SMP
1335 return rq->cpu;
1336#else
1337 return 0;
1338#endif
1339}
1340
1341#define MDF_PUSH 0x01
1342
1343static inline bool is_migration_disabled(struct task_struct *p)
1344{
1345#ifdef CONFIG_SMP
1346 return p->migration_disabled;
1347#else
1348 return false;
1349#endif
1350}
1351
1352DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1353
1354#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1355#define this_rq() this_cpu_ptr(&runqueues)
1356#define task_rq(p) cpu_rq(task_cpu(p))
1357#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1358#define raw_rq() raw_cpu_ptr(&runqueues)
1359
1360static inline void rq_set_donor(struct rq *rq, struct task_struct *t)
1361{
1362 /* Do nothing */
1363}
1364
1365#ifdef CONFIG_SCHED_CORE
1366static inline struct cpumask *sched_group_span(struct sched_group *sg);
1367
1368DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1369
1370static inline bool sched_core_enabled(struct rq *rq)
1371{
1372 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1373}
1374
1375static inline bool sched_core_disabled(void)
1376{
1377 return !static_branch_unlikely(&__sched_core_enabled);
1378}
1379
1380/*
1381 * Be careful with this function; not for general use. The return value isn't
1382 * stable unless you actually hold a relevant rq->__lock.
1383 */
1384static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1385{
1386 if (sched_core_enabled(rq))
1387 return &rq->core->__lock;
1388
1389 return &rq->__lock;
1390}
1391
1392static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1393{
1394 if (rq->core_enabled)
1395 return &rq->core->__lock;
1396
1397 return &rq->__lock;
1398}
1399
1400extern bool
1401cfs_prio_less(const struct task_struct *a, const struct task_struct *b, bool fi);
1402
1403extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1404
1405/*
1406 * Helpers to check if the CPU's core cookie matches with the task's cookie
1407 * when core scheduling is enabled.
1408 * A special case is that the task's cookie always matches with CPU's core
1409 * cookie if the CPU is in an idle core.
1410 */
1411static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1412{
1413 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1414 if (!sched_core_enabled(rq))
1415 return true;
1416
1417 return rq->core->core_cookie == p->core_cookie;
1418}
1419
1420static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1421{
1422 bool idle_core = true;
1423 int cpu;
1424
1425 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1426 if (!sched_core_enabled(rq))
1427 return true;
1428
1429 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1430 if (!available_idle_cpu(cpu)) {
1431 idle_core = false;
1432 break;
1433 }
1434 }
1435
1436 /*
1437 * A CPU in an idle core is always the best choice for tasks with
1438 * cookies.
1439 */
1440 return idle_core || rq->core->core_cookie == p->core_cookie;
1441}
1442
1443static inline bool sched_group_cookie_match(struct rq *rq,
1444 struct task_struct *p,
1445 struct sched_group *group)
1446{
1447 int cpu;
1448
1449 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1450 if (!sched_core_enabled(rq))
1451 return true;
1452
1453 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1454 if (sched_core_cookie_match(cpu_rq(cpu), p))
1455 return true;
1456 }
1457 return false;
1458}
1459
1460static inline bool sched_core_enqueued(struct task_struct *p)
1461{
1462 return !RB_EMPTY_NODE(&p->core_node);
1463}
1464
1465extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1466extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1467
1468extern void sched_core_get(void);
1469extern void sched_core_put(void);
1470
1471#else /* !CONFIG_SCHED_CORE: */
1472
1473static inline bool sched_core_enabled(struct rq *rq)
1474{
1475 return false;
1476}
1477
1478static inline bool sched_core_disabled(void)
1479{
1480 return true;
1481}
1482
1483static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1484{
1485 return &rq->__lock;
1486}
1487
1488static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1489{
1490 return &rq->__lock;
1491}
1492
1493static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1494{
1495 return true;
1496}
1497
1498static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1499{
1500 return true;
1501}
1502
1503static inline bool sched_group_cookie_match(struct rq *rq,
1504 struct task_struct *p,
1505 struct sched_group *group)
1506{
1507 return true;
1508}
1509
1510#endif /* !CONFIG_SCHED_CORE */
1511
1512static inline void lockdep_assert_rq_held(struct rq *rq)
1513{
1514 lockdep_assert_held(__rq_lockp(rq));
1515}
1516
1517extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1518extern bool raw_spin_rq_trylock(struct rq *rq);
1519extern void raw_spin_rq_unlock(struct rq *rq);
1520
1521static inline void raw_spin_rq_lock(struct rq *rq)
1522{
1523 raw_spin_rq_lock_nested(rq, 0);
1524}
1525
1526static inline void raw_spin_rq_lock_irq(struct rq *rq)
1527{
1528 local_irq_disable();
1529 raw_spin_rq_lock(rq);
1530}
1531
1532static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1533{
1534 raw_spin_rq_unlock(rq);
1535 local_irq_enable();
1536}
1537
1538static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1539{
1540 unsigned long flags;
1541
1542 local_irq_save(flags);
1543 raw_spin_rq_lock(rq);
1544
1545 return flags;
1546}
1547
1548static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1549{
1550 raw_spin_rq_unlock(rq);
1551 local_irq_restore(flags);
1552}
1553
1554#define raw_spin_rq_lock_irqsave(rq, flags) \
1555do { \
1556 flags = _raw_spin_rq_lock_irqsave(rq); \
1557} while (0)
1558
1559#ifdef CONFIG_SCHED_SMT
1560extern void __update_idle_core(struct rq *rq);
1561
1562static inline void update_idle_core(struct rq *rq)
1563{
1564 if (static_branch_unlikely(&sched_smt_present))
1565 __update_idle_core(rq);
1566}
1567
1568#else
1569static inline void update_idle_core(struct rq *rq) { }
1570#endif
1571
1572#ifdef CONFIG_FAIR_GROUP_SCHED
1573
1574static inline struct task_struct *task_of(struct sched_entity *se)
1575{
1576 SCHED_WARN_ON(!entity_is_task(se));
1577 return container_of(se, struct task_struct, se);
1578}
1579
1580static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1581{
1582 return p->se.cfs_rq;
1583}
1584
1585/* runqueue on which this entity is (to be) queued */
1586static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1587{
1588 return se->cfs_rq;
1589}
1590
1591/* runqueue "owned" by this group */
1592static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1593{
1594 return grp->my_q;
1595}
1596
1597#else /* !CONFIG_FAIR_GROUP_SCHED: */
1598
1599#define task_of(_se) container_of(_se, struct task_struct, se)
1600
1601static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1602{
1603 return &task_rq(p)->cfs;
1604}
1605
1606static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1607{
1608 const struct task_struct *p = task_of(se);
1609 struct rq *rq = task_rq(p);
1610
1611 return &rq->cfs;
1612}
1613
1614/* runqueue "owned" by this group */
1615static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1616{
1617 return NULL;
1618}
1619
1620#endif /* !CONFIG_FAIR_GROUP_SCHED */
1621
1622extern void update_rq_clock(struct rq *rq);
1623
1624/*
1625 * rq::clock_update_flags bits
1626 *
1627 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1628 * call to __schedule(). This is an optimisation to avoid
1629 * neighbouring rq clock updates.
1630 *
1631 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1632 * in effect and calls to update_rq_clock() are being ignored.
1633 *
1634 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1635 * made to update_rq_clock() since the last time rq::lock was pinned.
1636 *
1637 * If inside of __schedule(), clock_update_flags will have been
1638 * shifted left (a left shift is a cheap operation for the fast path
1639 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1640 *
1641 * if (rq-clock_update_flags >= RQCF_UPDATED)
1642 *
1643 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1644 * one position though, because the next rq_unpin_lock() will shift it
1645 * back.
1646 */
1647#define RQCF_REQ_SKIP 0x01
1648#define RQCF_ACT_SKIP 0x02
1649#define RQCF_UPDATED 0x04
1650
1651static inline void assert_clock_updated(struct rq *rq)
1652{
1653 /*
1654 * The only reason for not seeing a clock update since the
1655 * last rq_pin_lock() is if we're currently skipping updates.
1656 */
1657 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1658}
1659
1660static inline u64 rq_clock(struct rq *rq)
1661{
1662 lockdep_assert_rq_held(rq);
1663 assert_clock_updated(rq);
1664
1665 return rq->clock;
1666}
1667
1668static inline u64 rq_clock_task(struct rq *rq)
1669{
1670 lockdep_assert_rq_held(rq);
1671 assert_clock_updated(rq);
1672
1673 return rq->clock_task;
1674}
1675
1676static inline void rq_clock_skip_update(struct rq *rq)
1677{
1678 lockdep_assert_rq_held(rq);
1679 rq->clock_update_flags |= RQCF_REQ_SKIP;
1680}
1681
1682/*
1683 * See rt task throttling, which is the only time a skip
1684 * request is canceled.
1685 */
1686static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1687{
1688 lockdep_assert_rq_held(rq);
1689 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1690}
1691
1692/*
1693 * During cpu offlining and rq wide unthrottling, we can trigger
1694 * an update_rq_clock() for several cfs and rt runqueues (Typically
1695 * when using list_for_each_entry_*)
1696 * rq_clock_start_loop_update() can be called after updating the clock
1697 * once and before iterating over the list to prevent multiple update.
1698 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1699 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1700 */
1701static inline void rq_clock_start_loop_update(struct rq *rq)
1702{
1703 lockdep_assert_rq_held(rq);
1704 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1705 rq->clock_update_flags |= RQCF_ACT_SKIP;
1706}
1707
1708static inline void rq_clock_stop_loop_update(struct rq *rq)
1709{
1710 lockdep_assert_rq_held(rq);
1711 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1712}
1713
1714struct rq_flags {
1715 unsigned long flags;
1716 struct pin_cookie cookie;
1717#ifdef CONFIG_SCHED_DEBUG
1718 /*
1719 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1720 * current pin context is stashed here in case it needs to be
1721 * restored in rq_repin_lock().
1722 */
1723 unsigned int clock_update_flags;
1724#endif
1725};
1726
1727extern struct balance_callback balance_push_callback;
1728
1729/*
1730 * Lockdep annotation that avoids accidental unlocks; it's like a
1731 * sticky/continuous lockdep_assert_held().
1732 *
1733 * This avoids code that has access to 'struct rq *rq' (basically everything in
1734 * the scheduler) from accidentally unlocking the rq if they do not also have a
1735 * copy of the (on-stack) 'struct rq_flags rf'.
1736 *
1737 * Also see Documentation/locking/lockdep-design.rst.
1738 */
1739static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1740{
1741 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1742
1743#ifdef CONFIG_SCHED_DEBUG
1744 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1745 rf->clock_update_flags = 0;
1746# ifdef CONFIG_SMP
1747 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1748# endif
1749#endif
1750}
1751
1752static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1753{
1754#ifdef CONFIG_SCHED_DEBUG
1755 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1756 rf->clock_update_flags = RQCF_UPDATED;
1757#endif
1758
1759 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1760}
1761
1762static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1763{
1764 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1765
1766#ifdef CONFIG_SCHED_DEBUG
1767 /*
1768 * Restore the value we stashed in @rf for this pin context.
1769 */
1770 rq->clock_update_flags |= rf->clock_update_flags;
1771#endif
1772}
1773
1774extern
1775struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1776 __acquires(rq->lock);
1777
1778extern
1779struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1780 __acquires(p->pi_lock)
1781 __acquires(rq->lock);
1782
1783static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1784 __releases(rq->lock)
1785{
1786 rq_unpin_lock(rq, rf);
1787 raw_spin_rq_unlock(rq);
1788}
1789
1790static inline void
1791task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1792 __releases(rq->lock)
1793 __releases(p->pi_lock)
1794{
1795 rq_unpin_lock(rq, rf);
1796 raw_spin_rq_unlock(rq);
1797 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1798}
1799
1800DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1801 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1802 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1803 struct rq *rq; struct rq_flags rf)
1804
1805static inline void rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1806 __acquires(rq->lock)
1807{
1808 raw_spin_rq_lock_irqsave(rq, rf->flags);
1809 rq_pin_lock(rq, rf);
1810}
1811
1812static inline void rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1813 __acquires(rq->lock)
1814{
1815 raw_spin_rq_lock_irq(rq);
1816 rq_pin_lock(rq, rf);
1817}
1818
1819static inline void rq_lock(struct rq *rq, struct rq_flags *rf)
1820 __acquires(rq->lock)
1821{
1822 raw_spin_rq_lock(rq);
1823 rq_pin_lock(rq, rf);
1824}
1825
1826static inline void rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1827 __releases(rq->lock)
1828{
1829 rq_unpin_lock(rq, rf);
1830 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1831}
1832
1833static inline void rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1834 __releases(rq->lock)
1835{
1836 rq_unpin_lock(rq, rf);
1837 raw_spin_rq_unlock_irq(rq);
1838}
1839
1840static inline void rq_unlock(struct rq *rq, struct rq_flags *rf)
1841 __releases(rq->lock)
1842{
1843 rq_unpin_lock(rq, rf);
1844 raw_spin_rq_unlock(rq);
1845}
1846
1847DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1848 rq_lock(_T->lock, &_T->rf),
1849 rq_unlock(_T->lock, &_T->rf),
1850 struct rq_flags rf)
1851
1852DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1853 rq_lock_irq(_T->lock, &_T->rf),
1854 rq_unlock_irq(_T->lock, &_T->rf),
1855 struct rq_flags rf)
1856
1857DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1858 rq_lock_irqsave(_T->lock, &_T->rf),
1859 rq_unlock_irqrestore(_T->lock, &_T->rf),
1860 struct rq_flags rf)
1861
1862static inline struct rq *this_rq_lock_irq(struct rq_flags *rf)
1863 __acquires(rq->lock)
1864{
1865 struct rq *rq;
1866
1867 local_irq_disable();
1868 rq = this_rq();
1869 rq_lock(rq, rf);
1870
1871 return rq;
1872}
1873
1874#ifdef CONFIG_NUMA
1875
1876enum numa_topology_type {
1877 NUMA_DIRECT,
1878 NUMA_GLUELESS_MESH,
1879 NUMA_BACKPLANE,
1880};
1881
1882extern enum numa_topology_type sched_numa_topology_type;
1883extern int sched_max_numa_distance;
1884extern bool find_numa_distance(int distance);
1885extern void sched_init_numa(int offline_node);
1886extern void sched_update_numa(int cpu, bool online);
1887extern void sched_domains_numa_masks_set(unsigned int cpu);
1888extern void sched_domains_numa_masks_clear(unsigned int cpu);
1889extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1890
1891#else /* !CONFIG_NUMA: */
1892
1893static inline void sched_init_numa(int offline_node) { }
1894static inline void sched_update_numa(int cpu, bool online) { }
1895static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1896static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1897
1898static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1899{
1900 return nr_cpu_ids;
1901}
1902
1903#endif /* !CONFIG_NUMA */
1904
1905#ifdef CONFIG_NUMA_BALANCING
1906
1907/* The regions in numa_faults array from task_struct */
1908enum numa_faults_stats {
1909 NUMA_MEM = 0,
1910 NUMA_CPU,
1911 NUMA_MEMBUF,
1912 NUMA_CPUBUF
1913};
1914
1915extern void sched_setnuma(struct task_struct *p, int node);
1916extern int migrate_task_to(struct task_struct *p, int cpu);
1917extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1918 int cpu, int scpu);
1919extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1920
1921#else /* !CONFIG_NUMA_BALANCING: */
1922
1923static inline void
1924init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1925{
1926}
1927
1928#endif /* !CONFIG_NUMA_BALANCING */
1929
1930#ifdef CONFIG_SMP
1931
1932static inline void
1933queue_balance_callback(struct rq *rq,
1934 struct balance_callback *head,
1935 void (*func)(struct rq *rq))
1936{
1937 lockdep_assert_rq_held(rq);
1938
1939 /*
1940 * Don't (re)queue an already queued item; nor queue anything when
1941 * balance_push() is active, see the comment with
1942 * balance_push_callback.
1943 */
1944 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1945 return;
1946
1947 head->func = func;
1948 head->next = rq->balance_callback;
1949 rq->balance_callback = head;
1950}
1951
1952#define rcu_dereference_check_sched_domain(p) \
1953 rcu_dereference_check((p), lockdep_is_held(&sched_domains_mutex))
1954
1955/*
1956 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1957 * See destroy_sched_domains: call_rcu for details.
1958 *
1959 * The domain tree of any CPU may only be accessed from within
1960 * preempt-disabled sections.
1961 */
1962#define for_each_domain(cpu, __sd) \
1963 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1964 __sd; __sd = __sd->parent)
1965
1966/* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1967#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1968static const unsigned int SD_SHARED_CHILD_MASK =
1969#include <linux/sched/sd_flags.h>
19700;
1971#undef SD_FLAG
1972
1973/**
1974 * highest_flag_domain - Return highest sched_domain containing flag.
1975 * @cpu: The CPU whose highest level of sched domain is to
1976 * be returned.
1977 * @flag: The flag to check for the highest sched_domain
1978 * for the given CPU.
1979 *
1980 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1981 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1982 */
1983static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1984{
1985 struct sched_domain *sd, *hsd = NULL;
1986
1987 for_each_domain(cpu, sd) {
1988 if (sd->flags & flag) {
1989 hsd = sd;
1990 continue;
1991 }
1992
1993 /*
1994 * Stop the search if @flag is known to be shared at lower
1995 * levels. It will not be found further up.
1996 */
1997 if (flag & SD_SHARED_CHILD_MASK)
1998 break;
1999 }
2000
2001 return hsd;
2002}
2003
2004static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
2005{
2006 struct sched_domain *sd;
2007
2008 for_each_domain(cpu, sd) {
2009 if (sd->flags & flag)
2010 break;
2011 }
2012
2013 return sd;
2014}
2015
2016DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
2017DECLARE_PER_CPU(int, sd_llc_size);
2018DECLARE_PER_CPU(int, sd_llc_id);
2019DECLARE_PER_CPU(int, sd_share_id);
2020DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
2021DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
2022DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
2023DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
2024
2025extern struct static_key_false sched_asym_cpucapacity;
2026extern struct static_key_false sched_cluster_active;
2027
2028static __always_inline bool sched_asym_cpucap_active(void)
2029{
2030 return static_branch_unlikely(&sched_asym_cpucapacity);
2031}
2032
2033struct sched_group_capacity {
2034 atomic_t ref;
2035 /*
2036 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
2037 * for a single CPU.
2038 */
2039 unsigned long capacity;
2040 unsigned long min_capacity; /* Min per-CPU capacity in group */
2041 unsigned long max_capacity; /* Max per-CPU capacity in group */
2042 unsigned long next_update;
2043 int imbalance; /* XXX unrelated to capacity but shared group state */
2044
2045#ifdef CONFIG_SCHED_DEBUG
2046 int id;
2047#endif
2048
2049 unsigned long cpumask[]; /* Balance mask */
2050};
2051
2052struct sched_group {
2053 struct sched_group *next; /* Must be a circular list */
2054 atomic_t ref;
2055
2056 unsigned int group_weight;
2057 unsigned int cores;
2058 struct sched_group_capacity *sgc;
2059 int asym_prefer_cpu; /* CPU of highest priority in group */
2060 int flags;
2061
2062 /*
2063 * The CPUs this group covers.
2064 *
2065 * NOTE: this field is variable length. (Allocated dynamically
2066 * by attaching extra space to the end of the structure,
2067 * depending on how many CPUs the kernel has booted up with)
2068 */
2069 unsigned long cpumask[];
2070};
2071
2072static inline struct cpumask *sched_group_span(struct sched_group *sg)
2073{
2074 return to_cpumask(sg->cpumask);
2075}
2076
2077/*
2078 * See build_balance_mask().
2079 */
2080static inline struct cpumask *group_balance_mask(struct sched_group *sg)
2081{
2082 return to_cpumask(sg->sgc->cpumask);
2083}
2084
2085extern int group_balance_cpu(struct sched_group *sg);
2086
2087#ifdef CONFIG_SCHED_DEBUG
2088extern void update_sched_domain_debugfs(void);
2089extern void dirty_sched_domain_sysctl(int cpu);
2090#else
2091static inline void update_sched_domain_debugfs(void) { }
2092static inline void dirty_sched_domain_sysctl(int cpu) { }
2093#endif
2094
2095extern int sched_update_scaling(void);
2096
2097static inline const struct cpumask *task_user_cpus(struct task_struct *p)
2098{
2099 if (!p->user_cpus_ptr)
2100 return cpu_possible_mask; /* &init_task.cpus_mask */
2101 return p->user_cpus_ptr;
2102}
2103
2104#endif /* CONFIG_SMP */
2105
2106#ifdef CONFIG_CGROUP_SCHED
2107
2108/*
2109 * Return the group to which this tasks belongs.
2110 *
2111 * We cannot use task_css() and friends because the cgroup subsystem
2112 * changes that value before the cgroup_subsys::attach() method is called,
2113 * therefore we cannot pin it and might observe the wrong value.
2114 *
2115 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2116 * core changes this before calling sched_move_task().
2117 *
2118 * Instead we use a 'copy' which is updated from sched_move_task() while
2119 * holding both task_struct::pi_lock and rq::lock.
2120 */
2121static inline struct task_group *task_group(struct task_struct *p)
2122{
2123 return p->sched_task_group;
2124}
2125
2126/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2127static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2128{
2129#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2130 struct task_group *tg = task_group(p);
2131#endif
2132
2133#ifdef CONFIG_FAIR_GROUP_SCHED
2134 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2135 p->se.cfs_rq = tg->cfs_rq[cpu];
2136 p->se.parent = tg->se[cpu];
2137 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2138#endif
2139
2140#ifdef CONFIG_RT_GROUP_SCHED
2141 p->rt.rt_rq = tg->rt_rq[cpu];
2142 p->rt.parent = tg->rt_se[cpu];
2143#endif
2144}
2145
2146#else /* !CONFIG_CGROUP_SCHED: */
2147
2148static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2149
2150static inline struct task_group *task_group(struct task_struct *p)
2151{
2152 return NULL;
2153}
2154
2155#endif /* !CONFIG_CGROUP_SCHED */
2156
2157static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2158{
2159 set_task_rq(p, cpu);
2160#ifdef CONFIG_SMP
2161 /*
2162 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2163 * successfully executed on another CPU. We must ensure that updates of
2164 * per-task data have been completed by this moment.
2165 */
2166 smp_wmb();
2167 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2168 p->wake_cpu = cpu;
2169#endif
2170}
2171
2172/*
2173 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2174 */
2175#ifdef CONFIG_SCHED_DEBUG
2176# define const_debug __read_mostly
2177#else
2178# define const_debug const
2179#endif
2180
2181#define SCHED_FEAT(name, enabled) \
2182 __SCHED_FEAT_##name ,
2183
2184enum {
2185#include "features.h"
2186 __SCHED_FEAT_NR,
2187};
2188
2189#undef SCHED_FEAT
2190
2191#ifdef CONFIG_SCHED_DEBUG
2192
2193/*
2194 * To support run-time toggling of sched features, all the translation units
2195 * (but core.c) reference the sysctl_sched_features defined in core.c.
2196 */
2197extern const_debug unsigned int sysctl_sched_features;
2198
2199#ifdef CONFIG_JUMP_LABEL
2200
2201#define SCHED_FEAT(name, enabled) \
2202static __always_inline bool static_branch_##name(struct static_key *key) \
2203{ \
2204 return static_key_##enabled(key); \
2205}
2206
2207#include "features.h"
2208#undef SCHED_FEAT
2209
2210extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2211#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2212
2213#else /* !CONFIG_JUMP_LABEL: */
2214
2215#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2216
2217#endif /* !CONFIG_JUMP_LABEL */
2218
2219#else /* !SCHED_DEBUG: */
2220
2221/*
2222 * Each translation unit has its own copy of sysctl_sched_features to allow
2223 * constants propagation at compile time and compiler optimization based on
2224 * features default.
2225 */
2226#define SCHED_FEAT(name, enabled) \
2227 (1UL << __SCHED_FEAT_##name) * enabled |
2228static const_debug __maybe_unused unsigned int sysctl_sched_features =
2229#include "features.h"
2230 0;
2231#undef SCHED_FEAT
2232
2233#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2234
2235#endif /* !SCHED_DEBUG */
2236
2237extern struct static_key_false sched_numa_balancing;
2238extern struct static_key_false sched_schedstats;
2239
2240static inline u64 global_rt_period(void)
2241{
2242 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2243}
2244
2245static inline u64 global_rt_runtime(void)
2246{
2247 if (sysctl_sched_rt_runtime < 0)
2248 return RUNTIME_INF;
2249
2250 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2251}
2252
2253/*
2254 * Is p the current execution context?
2255 */
2256static inline int task_current(struct rq *rq, struct task_struct *p)
2257{
2258 return rq->curr == p;
2259}
2260
2261/*
2262 * Is p the current scheduling context?
2263 *
2264 * Note that it might be the current execution context at the same time if
2265 * rq->curr == rq->donor == p.
2266 */
2267static inline int task_current_donor(struct rq *rq, struct task_struct *p)
2268{
2269 return rq->donor == p;
2270}
2271
2272static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2273{
2274#ifdef CONFIG_SMP
2275 return p->on_cpu;
2276#else
2277 return task_current(rq, p);
2278#endif
2279}
2280
2281static inline int task_on_rq_queued(struct task_struct *p)
2282{
2283 return p->on_rq == TASK_ON_RQ_QUEUED;
2284}
2285
2286static inline int task_on_rq_migrating(struct task_struct *p)
2287{
2288 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2289}
2290
2291/* Wake flags. The first three directly map to some SD flag value */
2292#define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2293#define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2294#define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2295
2296#define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2297#define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2298#define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2299#define WF_RQ_SELECTED 0x80 /* ->select_task_rq() was called */
2300
2301#ifdef CONFIG_SMP
2302static_assert(WF_EXEC == SD_BALANCE_EXEC);
2303static_assert(WF_FORK == SD_BALANCE_FORK);
2304static_assert(WF_TTWU == SD_BALANCE_WAKE);
2305#endif
2306
2307/*
2308 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2309 * of tasks with abnormal "nice" values across CPUs the contribution that
2310 * each task makes to its run queue's load is weighted according to its
2311 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2312 * scaled version of the new time slice allocation that they receive on time
2313 * slice expiry etc.
2314 */
2315
2316#define WEIGHT_IDLEPRIO 3
2317#define WMULT_IDLEPRIO 1431655765
2318
2319extern const int sched_prio_to_weight[40];
2320extern const u32 sched_prio_to_wmult[40];
2321
2322/*
2323 * {de,en}queue flags:
2324 *
2325 * DEQUEUE_SLEEP - task is no longer runnable
2326 * ENQUEUE_WAKEUP - task just became runnable
2327 *
2328 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2329 * are in a known state which allows modification. Such pairs
2330 * should preserve as much state as possible.
2331 *
2332 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2333 * in the runqueue.
2334 *
2335 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2336 *
2337 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2338 *
2339 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2340 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2341 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2342 * ENQUEUE_RQ_SELECTED - ->select_task_rq() was called
2343 *
2344 */
2345
2346#define DEQUEUE_SLEEP 0x01 /* Matches ENQUEUE_WAKEUP */
2347#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2348#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2349#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2350#define DEQUEUE_SPECIAL 0x10
2351#define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2352#define DEQUEUE_DELAYED 0x200 /* Matches ENQUEUE_DELAYED */
2353
2354#define ENQUEUE_WAKEUP 0x01
2355#define ENQUEUE_RESTORE 0x02
2356#define ENQUEUE_MOVE 0x04
2357#define ENQUEUE_NOCLOCK 0x08
2358
2359#define ENQUEUE_HEAD 0x10
2360#define ENQUEUE_REPLENISH 0x20
2361#ifdef CONFIG_SMP
2362#define ENQUEUE_MIGRATED 0x40
2363#else
2364#define ENQUEUE_MIGRATED 0x00
2365#endif
2366#define ENQUEUE_INITIAL 0x80
2367#define ENQUEUE_MIGRATING 0x100
2368#define ENQUEUE_DELAYED 0x200
2369#define ENQUEUE_RQ_SELECTED 0x400
2370
2371#define RETRY_TASK ((void *)-1UL)
2372
2373struct affinity_context {
2374 const struct cpumask *new_mask;
2375 struct cpumask *user_mask;
2376 unsigned int flags;
2377};
2378
2379extern s64 update_curr_common(struct rq *rq);
2380
2381struct sched_class {
2382
2383#ifdef CONFIG_UCLAMP_TASK
2384 int uclamp_enabled;
2385#endif
2386
2387 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2388 bool (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2389 void (*yield_task) (struct rq *rq);
2390 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2391
2392 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2393
2394 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2395 struct task_struct *(*pick_task)(struct rq *rq);
2396 /*
2397 * Optional! When implemented pick_next_task() should be equivalent to:
2398 *
2399 * next = pick_task();
2400 * if (next) {
2401 * put_prev_task(prev);
2402 * set_next_task_first(next);
2403 * }
2404 */
2405 struct task_struct *(*pick_next_task)(struct rq *rq, struct task_struct *prev);
2406
2407 void (*put_prev_task)(struct rq *rq, struct task_struct *p, struct task_struct *next);
2408 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2409
2410#ifdef CONFIG_SMP
2411 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2412
2413 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2414
2415 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2416
2417 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2418
2419 void (*rq_online)(struct rq *rq);
2420 void (*rq_offline)(struct rq *rq);
2421
2422 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2423#endif
2424
2425 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2426 void (*task_fork)(struct task_struct *p);
2427 void (*task_dead)(struct task_struct *p);
2428
2429 /*
2430 * The switched_from() call is allowed to drop rq->lock, therefore we
2431 * cannot assume the switched_from/switched_to pair is serialized by
2432 * rq->lock. They are however serialized by p->pi_lock.
2433 */
2434 void (*switching_to) (struct rq *this_rq, struct task_struct *task);
2435 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2436 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2437 void (*reweight_task)(struct rq *this_rq, struct task_struct *task,
2438 const struct load_weight *lw);
2439 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2440 int oldprio);
2441
2442 unsigned int (*get_rr_interval)(struct rq *rq,
2443 struct task_struct *task);
2444
2445 void (*update_curr)(struct rq *rq);
2446
2447#ifdef CONFIG_FAIR_GROUP_SCHED
2448 void (*task_change_group)(struct task_struct *p);
2449#endif
2450
2451#ifdef CONFIG_SCHED_CORE
2452 int (*task_is_throttled)(struct task_struct *p, int cpu);
2453#endif
2454};
2455
2456static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2457{
2458 WARN_ON_ONCE(rq->donor != prev);
2459 prev->sched_class->put_prev_task(rq, prev, NULL);
2460}
2461
2462static inline void set_next_task(struct rq *rq, struct task_struct *next)
2463{
2464 next->sched_class->set_next_task(rq, next, false);
2465}
2466
2467static inline void
2468__put_prev_set_next_dl_server(struct rq *rq,
2469 struct task_struct *prev,
2470 struct task_struct *next)
2471{
2472 prev->dl_server = NULL;
2473 next->dl_server = rq->dl_server;
2474 rq->dl_server = NULL;
2475}
2476
2477static inline void put_prev_set_next_task(struct rq *rq,
2478 struct task_struct *prev,
2479 struct task_struct *next)
2480{
2481 WARN_ON_ONCE(rq->curr != prev);
2482
2483 __put_prev_set_next_dl_server(rq, prev, next);
2484
2485 if (next == prev)
2486 return;
2487
2488 prev->sched_class->put_prev_task(rq, prev, next);
2489 next->sched_class->set_next_task(rq, next, true);
2490}
2491
2492/*
2493 * Helper to define a sched_class instance; each one is placed in a separate
2494 * section which is ordered by the linker script:
2495 *
2496 * include/asm-generic/vmlinux.lds.h
2497 *
2498 * *CAREFUL* they are laid out in *REVERSE* order!!!
2499 *
2500 * Also enforce alignment on the instance, not the type, to guarantee layout.
2501 */
2502#define DEFINE_SCHED_CLASS(name) \
2503const struct sched_class name##_sched_class \
2504 __aligned(__alignof__(struct sched_class)) \
2505 __section("__" #name "_sched_class")
2506
2507/* Defined in include/asm-generic/vmlinux.lds.h */
2508extern struct sched_class __sched_class_highest[];
2509extern struct sched_class __sched_class_lowest[];
2510
2511extern const struct sched_class stop_sched_class;
2512extern const struct sched_class dl_sched_class;
2513extern const struct sched_class rt_sched_class;
2514extern const struct sched_class fair_sched_class;
2515extern const struct sched_class idle_sched_class;
2516
2517#ifdef CONFIG_SCHED_CLASS_EXT
2518extern const struct sched_class ext_sched_class;
2519
2520DECLARE_STATIC_KEY_FALSE(__scx_ops_enabled); /* SCX BPF scheduler loaded */
2521DECLARE_STATIC_KEY_FALSE(__scx_switched_all); /* all fair class tasks on SCX */
2522
2523#define scx_enabled() static_branch_unlikely(&__scx_ops_enabled)
2524#define scx_switched_all() static_branch_unlikely(&__scx_switched_all)
2525#else /* !CONFIG_SCHED_CLASS_EXT */
2526#define scx_enabled() false
2527#define scx_switched_all() false
2528#endif /* !CONFIG_SCHED_CLASS_EXT */
2529
2530/*
2531 * Iterate only active classes. SCX can take over all fair tasks or be
2532 * completely disabled. If the former, skip fair. If the latter, skip SCX.
2533 */
2534static inline const struct sched_class *next_active_class(const struct sched_class *class)
2535{
2536 class++;
2537#ifdef CONFIG_SCHED_CLASS_EXT
2538 if (scx_switched_all() && class == &fair_sched_class)
2539 class++;
2540 if (!scx_enabled() && class == &ext_sched_class)
2541 class++;
2542#endif
2543 return class;
2544}
2545
2546#define for_class_range(class, _from, _to) \
2547 for (class = (_from); class < (_to); class++)
2548
2549#define for_each_class(class) \
2550 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2551
2552#define for_active_class_range(class, _from, _to) \
2553 for (class = (_from); class != (_to); class = next_active_class(class))
2554
2555#define for_each_active_class(class) \
2556 for_active_class_range(class, __sched_class_highest, __sched_class_lowest)
2557
2558#define sched_class_above(_a, _b) ((_a) < (_b))
2559
2560static inline bool sched_stop_runnable(struct rq *rq)
2561{
2562 return rq->stop && task_on_rq_queued(rq->stop);
2563}
2564
2565static inline bool sched_dl_runnable(struct rq *rq)
2566{
2567 return rq->dl.dl_nr_running > 0;
2568}
2569
2570static inline bool sched_rt_runnable(struct rq *rq)
2571{
2572 return rq->rt.rt_queued > 0;
2573}
2574
2575static inline bool sched_fair_runnable(struct rq *rq)
2576{
2577 return rq->cfs.nr_running > 0;
2578}
2579
2580extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2581extern struct task_struct *pick_task_idle(struct rq *rq);
2582
2583#define SCA_CHECK 0x01
2584#define SCA_MIGRATE_DISABLE 0x02
2585#define SCA_MIGRATE_ENABLE 0x04
2586#define SCA_USER 0x08
2587
2588#ifdef CONFIG_SMP
2589
2590extern void update_group_capacity(struct sched_domain *sd, int cpu);
2591
2592extern void sched_balance_trigger(struct rq *rq);
2593
2594extern int __set_cpus_allowed_ptr(struct task_struct *p, struct affinity_context *ctx);
2595extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2596
2597static inline bool task_allowed_on_cpu(struct task_struct *p, int cpu)
2598{
2599 /* When not in the task's cpumask, no point in looking further. */
2600 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2601 return false;
2602
2603 /* Can @cpu run a user thread? */
2604 if (!(p->flags & PF_KTHREAD) && !task_cpu_possible(cpu, p))
2605 return false;
2606
2607 return true;
2608}
2609
2610static inline cpumask_t *alloc_user_cpus_ptr(int node)
2611{
2612 /*
2613 * See do_set_cpus_allowed() above for the rcu_head usage.
2614 */
2615 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2616
2617 return kmalloc_node(size, GFP_KERNEL, node);
2618}
2619
2620static inline struct task_struct *get_push_task(struct rq *rq)
2621{
2622 struct task_struct *p = rq->donor;
2623
2624 lockdep_assert_rq_held(rq);
2625
2626 if (rq->push_busy)
2627 return NULL;
2628
2629 if (p->nr_cpus_allowed == 1)
2630 return NULL;
2631
2632 if (p->migration_disabled)
2633 return NULL;
2634
2635 rq->push_busy = true;
2636 return get_task_struct(p);
2637}
2638
2639extern int push_cpu_stop(void *arg);
2640
2641#else /* !CONFIG_SMP: */
2642
2643static inline bool task_allowed_on_cpu(struct task_struct *p, int cpu)
2644{
2645 return true;
2646}
2647
2648static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2649 struct affinity_context *ctx)
2650{
2651 return set_cpus_allowed_ptr(p, ctx->new_mask);
2652}
2653
2654static inline cpumask_t *alloc_user_cpus_ptr(int node)
2655{
2656 return NULL;
2657}
2658
2659#endif /* !CONFIG_SMP */
2660
2661#ifdef CONFIG_CPU_IDLE
2662
2663static inline void idle_set_state(struct rq *rq,
2664 struct cpuidle_state *idle_state)
2665{
2666 rq->idle_state = idle_state;
2667}
2668
2669static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2670{
2671 SCHED_WARN_ON(!rcu_read_lock_held());
2672
2673 return rq->idle_state;
2674}
2675
2676#else /* !CONFIG_CPU_IDLE: */
2677
2678static inline void idle_set_state(struct rq *rq,
2679 struct cpuidle_state *idle_state)
2680{
2681}
2682
2683static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2684{
2685 return NULL;
2686}
2687
2688#endif /* !CONFIG_CPU_IDLE */
2689
2690extern void schedule_idle(void);
2691asmlinkage void schedule_user(void);
2692
2693extern void sysrq_sched_debug_show(void);
2694extern void sched_init_granularity(void);
2695extern void update_max_interval(void);
2696
2697extern void init_sched_dl_class(void);
2698extern void init_sched_rt_class(void);
2699extern void init_sched_fair_class(void);
2700
2701extern void resched_curr(struct rq *rq);
2702extern void resched_curr_lazy(struct rq *rq);
2703extern void resched_cpu(int cpu);
2704
2705extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2706extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2707
2708extern void init_dl_entity(struct sched_dl_entity *dl_se);
2709
2710#define BW_SHIFT 20
2711#define BW_UNIT (1 << BW_SHIFT)
2712#define RATIO_SHIFT 8
2713#define MAX_BW_BITS (64 - BW_SHIFT)
2714#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2715
2716extern unsigned long to_ratio(u64 period, u64 runtime);
2717
2718extern void init_entity_runnable_average(struct sched_entity *se);
2719extern void post_init_entity_util_avg(struct task_struct *p);
2720
2721#ifdef CONFIG_NO_HZ_FULL
2722extern bool sched_can_stop_tick(struct rq *rq);
2723extern int __init sched_tick_offload_init(void);
2724
2725/*
2726 * Tick may be needed by tasks in the runqueue depending on their policy and
2727 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2728 * nohz mode if necessary.
2729 */
2730static inline void sched_update_tick_dependency(struct rq *rq)
2731{
2732 int cpu = cpu_of(rq);
2733
2734 if (!tick_nohz_full_cpu(cpu))
2735 return;
2736
2737 if (sched_can_stop_tick(rq))
2738 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2739 else
2740 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2741}
2742#else /* !CONFIG_NO_HZ_FULL: */
2743static inline int sched_tick_offload_init(void) { return 0; }
2744static inline void sched_update_tick_dependency(struct rq *rq) { }
2745#endif /* !CONFIG_NO_HZ_FULL */
2746
2747static inline void add_nr_running(struct rq *rq, unsigned count)
2748{
2749 unsigned prev_nr = rq->nr_running;
2750
2751 rq->nr_running = prev_nr + count;
2752 if (trace_sched_update_nr_running_tp_enabled()) {
2753 call_trace_sched_update_nr_running(rq, count);
2754 }
2755
2756#ifdef CONFIG_SMP
2757 if (prev_nr < 2 && rq->nr_running >= 2)
2758 set_rd_overloaded(rq->rd, 1);
2759#endif
2760
2761 sched_update_tick_dependency(rq);
2762}
2763
2764static inline void sub_nr_running(struct rq *rq, unsigned count)
2765{
2766 rq->nr_running -= count;
2767 if (trace_sched_update_nr_running_tp_enabled()) {
2768 call_trace_sched_update_nr_running(rq, -count);
2769 }
2770
2771 /* Check if we still need preemption */
2772 sched_update_tick_dependency(rq);
2773}
2774
2775static inline void __block_task(struct rq *rq, struct task_struct *p)
2776{
2777 if (p->sched_contributes_to_load)
2778 rq->nr_uninterruptible++;
2779
2780 if (p->in_iowait) {
2781 atomic_inc(&rq->nr_iowait);
2782 delayacct_blkio_start();
2783 }
2784
2785 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2786
2787 /*
2788 * The moment this write goes through, ttwu() can swoop in and migrate
2789 * this task, rendering our rq->__lock ineffective.
2790 *
2791 * __schedule() try_to_wake_up()
2792 * LOCK rq->__lock LOCK p->pi_lock
2793 * pick_next_task()
2794 * pick_next_task_fair()
2795 * pick_next_entity()
2796 * dequeue_entities()
2797 * __block_task()
2798 * RELEASE p->on_rq = 0 if (p->on_rq && ...)
2799 * break;
2800 *
2801 * ACQUIRE (after ctrl-dep)
2802 *
2803 * cpu = select_task_rq();
2804 * set_task_cpu(p, cpu);
2805 * ttwu_queue()
2806 * ttwu_do_activate()
2807 * LOCK rq->__lock
2808 * activate_task()
2809 * STORE p->on_rq = 1
2810 * UNLOCK rq->__lock
2811 *
2812 * Callers must ensure to not reference @p after this -- we no longer
2813 * own it.
2814 */
2815 smp_store_release(&p->on_rq, 0);
2816}
2817
2818extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2819extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2820
2821extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2822
2823#ifdef CONFIG_PREEMPT_RT
2824# define SCHED_NR_MIGRATE_BREAK 8
2825#else
2826# define SCHED_NR_MIGRATE_BREAK 32
2827#endif
2828
2829extern const_debug unsigned int sysctl_sched_nr_migrate;
2830extern const_debug unsigned int sysctl_sched_migration_cost;
2831
2832extern unsigned int sysctl_sched_base_slice;
2833
2834#ifdef CONFIG_SCHED_DEBUG
2835extern int sysctl_resched_latency_warn_ms;
2836extern int sysctl_resched_latency_warn_once;
2837
2838extern unsigned int sysctl_sched_tunable_scaling;
2839
2840extern unsigned int sysctl_numa_balancing_scan_delay;
2841extern unsigned int sysctl_numa_balancing_scan_period_min;
2842extern unsigned int sysctl_numa_balancing_scan_period_max;
2843extern unsigned int sysctl_numa_balancing_scan_size;
2844extern unsigned int sysctl_numa_balancing_hot_threshold;
2845#endif
2846
2847#ifdef CONFIG_SCHED_HRTICK
2848
2849/*
2850 * Use hrtick when:
2851 * - enabled by features
2852 * - hrtimer is actually high res
2853 */
2854static inline int hrtick_enabled(struct rq *rq)
2855{
2856 if (!cpu_active(cpu_of(rq)))
2857 return 0;
2858 return hrtimer_is_hres_active(&rq->hrtick_timer);
2859}
2860
2861static inline int hrtick_enabled_fair(struct rq *rq)
2862{
2863 if (!sched_feat(HRTICK))
2864 return 0;
2865 return hrtick_enabled(rq);
2866}
2867
2868static inline int hrtick_enabled_dl(struct rq *rq)
2869{
2870 if (!sched_feat(HRTICK_DL))
2871 return 0;
2872 return hrtick_enabled(rq);
2873}
2874
2875extern void hrtick_start(struct rq *rq, u64 delay);
2876
2877#else /* !CONFIG_SCHED_HRTICK: */
2878
2879static inline int hrtick_enabled_fair(struct rq *rq)
2880{
2881 return 0;
2882}
2883
2884static inline int hrtick_enabled_dl(struct rq *rq)
2885{
2886 return 0;
2887}
2888
2889static inline int hrtick_enabled(struct rq *rq)
2890{
2891 return 0;
2892}
2893
2894#endif /* !CONFIG_SCHED_HRTICK */
2895
2896#ifndef arch_scale_freq_tick
2897static __always_inline void arch_scale_freq_tick(void) { }
2898#endif
2899
2900#ifndef arch_scale_freq_capacity
2901/**
2902 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2903 * @cpu: the CPU in question.
2904 *
2905 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2906 *
2907 * f_curr
2908 * ------ * SCHED_CAPACITY_SCALE
2909 * f_max
2910 */
2911static __always_inline
2912unsigned long arch_scale_freq_capacity(int cpu)
2913{
2914 return SCHED_CAPACITY_SCALE;
2915}
2916#endif
2917
2918#ifdef CONFIG_SCHED_DEBUG
2919/*
2920 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2921 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2922 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2923 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2924 */
2925static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2926{
2927 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2928 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2929#ifdef CONFIG_SMP
2930 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2931#endif
2932}
2933#else
2934static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) { }
2935#endif
2936
2937#define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2938__DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2939static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2940{ class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2941 _lock; return _t; }
2942
2943#ifdef CONFIG_SMP
2944
2945static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2946{
2947#ifdef CONFIG_SCHED_CORE
2948 /*
2949 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2950 * order by core-id first and cpu-id second.
2951 *
2952 * Notably:
2953 *
2954 * double_rq_lock(0,3); will take core-0, core-1 lock
2955 * double_rq_lock(1,2); will take core-1, core-0 lock
2956 *
2957 * when only cpu-id is considered.
2958 */
2959 if (rq1->core->cpu < rq2->core->cpu)
2960 return true;
2961 if (rq1->core->cpu > rq2->core->cpu)
2962 return false;
2963
2964 /*
2965 * __sched_core_flip() relies on SMT having cpu-id lock order.
2966 */
2967#endif
2968 return rq1->cpu < rq2->cpu;
2969}
2970
2971extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2972
2973#ifdef CONFIG_PREEMPTION
2974
2975/*
2976 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2977 * way at the expense of forcing extra atomic operations in all
2978 * invocations. This assures that the double_lock is acquired using the
2979 * same underlying policy as the spinlock_t on this architecture, which
2980 * reduces latency compared to the unfair variant below. However, it
2981 * also adds more overhead and therefore may reduce throughput.
2982 */
2983static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2984 __releases(this_rq->lock)
2985 __acquires(busiest->lock)
2986 __acquires(this_rq->lock)
2987{
2988 raw_spin_rq_unlock(this_rq);
2989 double_rq_lock(this_rq, busiest);
2990
2991 return 1;
2992}
2993
2994#else /* !CONFIG_PREEMPTION: */
2995/*
2996 * Unfair double_lock_balance: Optimizes throughput at the expense of
2997 * latency by eliminating extra atomic operations when the locks are
2998 * already in proper order on entry. This favors lower CPU-ids and will
2999 * grant the double lock to lower CPUs over higher ids under contention,
3000 * regardless of entry order into the function.
3001 */
3002static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
3003 __releases(this_rq->lock)
3004 __acquires(busiest->lock)
3005 __acquires(this_rq->lock)
3006{
3007 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
3008 likely(raw_spin_rq_trylock(busiest))) {
3009 double_rq_clock_clear_update(this_rq, busiest);
3010 return 0;
3011 }
3012
3013 if (rq_order_less(this_rq, busiest)) {
3014 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
3015 double_rq_clock_clear_update(this_rq, busiest);
3016 return 0;
3017 }
3018
3019 raw_spin_rq_unlock(this_rq);
3020 double_rq_lock(this_rq, busiest);
3021
3022 return 1;
3023}
3024
3025#endif /* !CONFIG_PREEMPTION */
3026
3027/*
3028 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3029 */
3030static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
3031{
3032 lockdep_assert_irqs_disabled();
3033
3034 return _double_lock_balance(this_rq, busiest);
3035}
3036
3037static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
3038 __releases(busiest->lock)
3039{
3040 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
3041 raw_spin_rq_unlock(busiest);
3042 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
3043}
3044
3045static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
3046{
3047 if (l1 > l2)
3048 swap(l1, l2);
3049
3050 spin_lock(l1);
3051 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
3052}
3053
3054static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
3055{
3056 if (l1 > l2)
3057 swap(l1, l2);
3058
3059 spin_lock_irq(l1);
3060 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
3061}
3062
3063static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
3064{
3065 if (l1 > l2)
3066 swap(l1, l2);
3067
3068 raw_spin_lock(l1);
3069 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
3070}
3071
3072static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
3073{
3074 raw_spin_unlock(l1);
3075 raw_spin_unlock(l2);
3076}
3077
3078DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
3079 double_raw_lock(_T->lock, _T->lock2),
3080 double_raw_unlock(_T->lock, _T->lock2))
3081
3082/*
3083 * double_rq_unlock - safely unlock two runqueues
3084 *
3085 * Note this does not restore interrupts like task_rq_unlock,
3086 * you need to do so manually after calling.
3087 */
3088static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3089 __releases(rq1->lock)
3090 __releases(rq2->lock)
3091{
3092 if (__rq_lockp(rq1) != __rq_lockp(rq2))
3093 raw_spin_rq_unlock(rq2);
3094 else
3095 __release(rq2->lock);
3096 raw_spin_rq_unlock(rq1);
3097}
3098
3099extern void set_rq_online (struct rq *rq);
3100extern void set_rq_offline(struct rq *rq);
3101
3102extern bool sched_smp_initialized;
3103
3104#else /* !CONFIG_SMP: */
3105
3106/*
3107 * double_rq_lock - safely lock two runqueues
3108 *
3109 * Note this does not disable interrupts like task_rq_lock,
3110 * you need to do so manually before calling.
3111 */
3112static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
3113 __acquires(rq1->lock)
3114 __acquires(rq2->lock)
3115{
3116 WARN_ON_ONCE(!irqs_disabled());
3117 WARN_ON_ONCE(rq1 != rq2);
3118 raw_spin_rq_lock(rq1);
3119 __acquire(rq2->lock); /* Fake it out ;) */
3120 double_rq_clock_clear_update(rq1, rq2);
3121}
3122
3123/*
3124 * double_rq_unlock - safely unlock two runqueues
3125 *
3126 * Note this does not restore interrupts like task_rq_unlock,
3127 * you need to do so manually after calling.
3128 */
3129static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3130 __releases(rq1->lock)
3131 __releases(rq2->lock)
3132{
3133 WARN_ON_ONCE(rq1 != rq2);
3134 raw_spin_rq_unlock(rq1);
3135 __release(rq2->lock);
3136}
3137
3138#endif /* !CONFIG_SMP */
3139
3140DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
3141 double_rq_lock(_T->lock, _T->lock2),
3142 double_rq_unlock(_T->lock, _T->lock2))
3143
3144extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
3145extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
3146extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
3147
3148#ifdef CONFIG_SCHED_DEBUG
3149extern bool sched_debug_verbose;
3150
3151extern void print_cfs_stats(struct seq_file *m, int cpu);
3152extern void print_rt_stats(struct seq_file *m, int cpu);
3153extern void print_dl_stats(struct seq_file *m, int cpu);
3154extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
3155extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
3156extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
3157
3158extern void resched_latency_warn(int cpu, u64 latency);
3159# ifdef CONFIG_NUMA_BALANCING
3160extern void show_numa_stats(struct task_struct *p, struct seq_file *m);
3161extern void
3162print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
3163 unsigned long tpf, unsigned long gsf, unsigned long gpf);
3164# endif /* CONFIG_NUMA_BALANCING */
3165#else /* !CONFIG_SCHED_DEBUG: */
3166static inline void resched_latency_warn(int cpu, u64 latency) { }
3167#endif /* !CONFIG_SCHED_DEBUG */
3168
3169extern void init_cfs_rq(struct cfs_rq *cfs_rq);
3170extern void init_rt_rq(struct rt_rq *rt_rq);
3171extern void init_dl_rq(struct dl_rq *dl_rq);
3172
3173extern void cfs_bandwidth_usage_inc(void);
3174extern void cfs_bandwidth_usage_dec(void);
3175
3176#ifdef CONFIG_NO_HZ_COMMON
3177
3178#define NOHZ_BALANCE_KICK_BIT 0
3179#define NOHZ_STATS_KICK_BIT 1
3180#define NOHZ_NEWILB_KICK_BIT 2
3181#define NOHZ_NEXT_KICK_BIT 3
3182
3183/* Run sched_balance_domains() */
3184#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
3185/* Update blocked load */
3186#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
3187/* Update blocked load when entering idle */
3188#define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
3189/* Update nohz.next_balance */
3190#define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
3191
3192#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
3193
3194#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
3195
3196extern void nohz_balance_exit_idle(struct rq *rq);
3197#else /* !CONFIG_NO_HZ_COMMON: */
3198static inline void nohz_balance_exit_idle(struct rq *rq) { }
3199#endif /* !CONFIG_NO_HZ_COMMON */
3200
3201#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
3202extern void nohz_run_idle_balance(int cpu);
3203#else
3204static inline void nohz_run_idle_balance(int cpu) { }
3205#endif
3206
3207#include "stats.h"
3208
3209#if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
3210
3211extern void __sched_core_account_forceidle(struct rq *rq);
3212
3213static inline void sched_core_account_forceidle(struct rq *rq)
3214{
3215 if (schedstat_enabled())
3216 __sched_core_account_forceidle(rq);
3217}
3218
3219extern void __sched_core_tick(struct rq *rq);
3220
3221static inline void sched_core_tick(struct rq *rq)
3222{
3223 if (sched_core_enabled(rq) && schedstat_enabled())
3224 __sched_core_tick(rq);
3225}
3226
3227#else /* !(CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS): */
3228
3229static inline void sched_core_account_forceidle(struct rq *rq) { }
3230
3231static inline void sched_core_tick(struct rq *rq) { }
3232
3233#endif /* !(CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS) */
3234
3235#ifdef CONFIG_IRQ_TIME_ACCOUNTING
3236
3237struct irqtime {
3238 u64 total;
3239 u64 tick_delta;
3240 u64 irq_start_time;
3241 struct u64_stats_sync sync;
3242};
3243
3244DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
3245
3246/*
3247 * Returns the irqtime minus the softirq time computed by ksoftirqd.
3248 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
3249 * and never move forward.
3250 */
3251static inline u64 irq_time_read(int cpu)
3252{
3253 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
3254 unsigned int seq;
3255 u64 total;
3256
3257 do {
3258 seq = __u64_stats_fetch_begin(&irqtime->sync);
3259 total = irqtime->total;
3260 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
3261
3262 return total;
3263}
3264
3265#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3266
3267#ifdef CONFIG_CPU_FREQ
3268
3269DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
3270
3271/**
3272 * cpufreq_update_util - Take a note about CPU utilization changes.
3273 * @rq: Runqueue to carry out the update for.
3274 * @flags: Update reason flags.
3275 *
3276 * This function is called by the scheduler on the CPU whose utilization is
3277 * being updated.
3278 *
3279 * It can only be called from RCU-sched read-side critical sections.
3280 *
3281 * The way cpufreq is currently arranged requires it to evaluate the CPU
3282 * performance state (frequency/voltage) on a regular basis to prevent it from
3283 * being stuck in a completely inadequate performance level for too long.
3284 * That is not guaranteed to happen if the updates are only triggered from CFS
3285 * and DL, though, because they may not be coming in if only RT tasks are
3286 * active all the time (or there are RT tasks only).
3287 *
3288 * As a workaround for that issue, this function is called periodically by the
3289 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
3290 * but that really is a band-aid. Going forward it should be replaced with
3291 * solutions targeted more specifically at RT tasks.
3292 */
3293static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
3294{
3295 struct update_util_data *data;
3296
3297 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
3298 cpu_of(rq)));
3299 if (data)
3300 data->func(data, rq_clock(rq), flags);
3301}
3302#else /* !CONFIG_CPU_FREQ: */
3303static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) { }
3304#endif /* !CONFIG_CPU_FREQ */
3305
3306#ifdef arch_scale_freq_capacity
3307# ifndef arch_scale_freq_invariant
3308# define arch_scale_freq_invariant() true
3309# endif
3310#else
3311# define arch_scale_freq_invariant() false
3312#endif
3313
3314#ifdef CONFIG_SMP
3315
3316unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3317 unsigned long *min,
3318 unsigned long *max);
3319
3320unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3321 unsigned long min,
3322 unsigned long max);
3323
3324
3325/*
3326 * Verify the fitness of task @p to run on @cpu taking into account the
3327 * CPU original capacity and the runtime/deadline ratio of the task.
3328 *
3329 * The function will return true if the original capacity of @cpu is
3330 * greater than or equal to task's deadline density right shifted by
3331 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3332 */
3333static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3334{
3335 unsigned long cap = arch_scale_cpu_capacity(cpu);
3336
3337 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3338}
3339
3340static inline unsigned long cpu_bw_dl(struct rq *rq)
3341{
3342 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3343}
3344
3345static inline unsigned long cpu_util_dl(struct rq *rq)
3346{
3347 return READ_ONCE(rq->avg_dl.util_avg);
3348}
3349
3350
3351extern unsigned long cpu_util_cfs(int cpu);
3352extern unsigned long cpu_util_cfs_boost(int cpu);
3353
3354static inline unsigned long cpu_util_rt(struct rq *rq)
3355{
3356 return READ_ONCE(rq->avg_rt.util_avg);
3357}
3358
3359#else /* !CONFIG_SMP */
3360static inline bool update_other_load_avgs(struct rq *rq) { return false; }
3361#endif /* CONFIG_SMP */
3362
3363#ifdef CONFIG_UCLAMP_TASK
3364
3365unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3366
3367static inline unsigned long uclamp_rq_get(struct rq *rq,
3368 enum uclamp_id clamp_id)
3369{
3370 return READ_ONCE(rq->uclamp[clamp_id].value);
3371}
3372
3373static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3374 unsigned int value)
3375{
3376 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3377}
3378
3379static inline bool uclamp_rq_is_idle(struct rq *rq)
3380{
3381 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3382}
3383
3384/* Is the rq being capped/throttled by uclamp_max? */
3385static inline bool uclamp_rq_is_capped(struct rq *rq)
3386{
3387 unsigned long rq_util;
3388 unsigned long max_util;
3389
3390 if (!static_branch_likely(&sched_uclamp_used))
3391 return false;
3392
3393 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3394 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3395
3396 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3397}
3398
3399/*
3400 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3401 * by default in the fast path and only gets turned on once userspace performs
3402 * an operation that requires it.
3403 *
3404 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3405 * hence is active.
3406 */
3407static inline bool uclamp_is_used(void)
3408{
3409 return static_branch_likely(&sched_uclamp_used);
3410}
3411
3412#define for_each_clamp_id(clamp_id) \
3413 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
3414
3415extern unsigned int sysctl_sched_uclamp_util_min_rt_default;
3416
3417
3418static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
3419{
3420 if (clamp_id == UCLAMP_MIN)
3421 return 0;
3422 return SCHED_CAPACITY_SCALE;
3423}
3424
3425/* Integer rounded range for each bucket */
3426#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
3427
3428static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
3429{
3430 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
3431}
3432
3433static inline void
3434uclamp_se_set(struct uclamp_se *uc_se, unsigned int value, bool user_defined)
3435{
3436 uc_se->value = value;
3437 uc_se->bucket_id = uclamp_bucket_id(value);
3438 uc_se->user_defined = user_defined;
3439}
3440
3441#else /* !CONFIG_UCLAMP_TASK: */
3442
3443static inline unsigned long
3444uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
3445{
3446 if (clamp_id == UCLAMP_MIN)
3447 return 0;
3448
3449 return SCHED_CAPACITY_SCALE;
3450}
3451
3452static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3453
3454static inline bool uclamp_is_used(void)
3455{
3456 return false;
3457}
3458
3459static inline unsigned long
3460uclamp_rq_get(struct rq *rq, enum uclamp_id clamp_id)
3461{
3462 if (clamp_id == UCLAMP_MIN)
3463 return 0;
3464
3465 return SCHED_CAPACITY_SCALE;
3466}
3467
3468static inline void
3469uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id, unsigned int value)
3470{
3471}
3472
3473static inline bool uclamp_rq_is_idle(struct rq *rq)
3474{
3475 return false;
3476}
3477
3478#endif /* !CONFIG_UCLAMP_TASK */
3479
3480#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3481
3482static inline unsigned long cpu_util_irq(struct rq *rq)
3483{
3484 return READ_ONCE(rq->avg_irq.util_avg);
3485}
3486
3487static inline
3488unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3489{
3490 util *= (max - irq);
3491 util /= max;
3492
3493 return util;
3494
3495}
3496
3497#else /* !CONFIG_HAVE_SCHED_AVG_IRQ: */
3498
3499static inline unsigned long cpu_util_irq(struct rq *rq)
3500{
3501 return 0;
3502}
3503
3504static inline
3505unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3506{
3507 return util;
3508}
3509
3510#endif /* !CONFIG_HAVE_SCHED_AVG_IRQ */
3511
3512#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3513
3514#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3515
3516DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3517
3518static inline bool sched_energy_enabled(void)
3519{
3520 return static_branch_unlikely(&sched_energy_present);
3521}
3522
3523extern struct cpufreq_governor schedutil_gov;
3524
3525#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3526
3527#define perf_domain_span(pd) NULL
3528
3529static inline bool sched_energy_enabled(void) { return false; }
3530
3531#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3532
3533#ifdef CONFIG_MEMBARRIER
3534
3535/*
3536 * The scheduler provides memory barriers required by membarrier between:
3537 * - prior user-space memory accesses and store to rq->membarrier_state,
3538 * - store to rq->membarrier_state and following user-space memory accesses.
3539 * In the same way it provides those guarantees around store to rq->curr.
3540 */
3541static inline void membarrier_switch_mm(struct rq *rq,
3542 struct mm_struct *prev_mm,
3543 struct mm_struct *next_mm)
3544{
3545 int membarrier_state;
3546
3547 if (prev_mm == next_mm)
3548 return;
3549
3550 membarrier_state = atomic_read(&next_mm->membarrier_state);
3551 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3552 return;
3553
3554 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3555}
3556
3557#else /* !CONFIG_MEMBARRIER :*/
3558
3559static inline void membarrier_switch_mm(struct rq *rq,
3560 struct mm_struct *prev_mm,
3561 struct mm_struct *next_mm)
3562{
3563}
3564
3565#endif /* !CONFIG_MEMBARRIER */
3566
3567#ifdef CONFIG_SMP
3568static inline bool is_per_cpu_kthread(struct task_struct *p)
3569{
3570 if (!(p->flags & PF_KTHREAD))
3571 return false;
3572
3573 if (p->nr_cpus_allowed != 1)
3574 return false;
3575
3576 return true;
3577}
3578#endif
3579
3580extern void swake_up_all_locked(struct swait_queue_head *q);
3581extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3582
3583extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3584
3585#ifdef CONFIG_PREEMPT_DYNAMIC
3586extern int preempt_dynamic_mode;
3587extern int sched_dynamic_mode(const char *str);
3588extern void sched_dynamic_update(int mode);
3589#endif
3590
3591#ifdef CONFIG_SCHED_MM_CID
3592
3593#define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3594#define MM_CID_SCAN_DELAY 100 /* 100ms */
3595
3596extern raw_spinlock_t cid_lock;
3597extern int use_cid_lock;
3598
3599extern void sched_mm_cid_migrate_from(struct task_struct *t);
3600extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3601extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3602extern void init_sched_mm_cid(struct task_struct *t);
3603
3604static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3605{
3606 if (cid < 0)
3607 return;
3608 cpumask_clear_cpu(cid, mm_cidmask(mm));
3609}
3610
3611/*
3612 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3613 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3614 * be held to transition to other states.
3615 *
3616 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3617 * consistent across CPUs, which prevents use of this_cpu_cmpxchg.
3618 */
3619static inline void mm_cid_put_lazy(struct task_struct *t)
3620{
3621 struct mm_struct *mm = t->mm;
3622 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3623 int cid;
3624
3625 lockdep_assert_irqs_disabled();
3626 cid = __this_cpu_read(pcpu_cid->cid);
3627 if (!mm_cid_is_lazy_put(cid) ||
3628 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3629 return;
3630 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3631}
3632
3633static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3634{
3635 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3636 int cid, res;
3637
3638 lockdep_assert_irqs_disabled();
3639 cid = __this_cpu_read(pcpu_cid->cid);
3640 for (;;) {
3641 if (mm_cid_is_unset(cid))
3642 return MM_CID_UNSET;
3643 /*
3644 * Attempt transition from valid or lazy-put to unset.
3645 */
3646 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3647 if (res == cid)
3648 break;
3649 cid = res;
3650 }
3651 return cid;
3652}
3653
3654static inline void mm_cid_put(struct mm_struct *mm)
3655{
3656 int cid;
3657
3658 lockdep_assert_irqs_disabled();
3659 cid = mm_cid_pcpu_unset(mm);
3660 if (cid == MM_CID_UNSET)
3661 return;
3662 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3663}
3664
3665static inline int __mm_cid_try_get(struct task_struct *t, struct mm_struct *mm)
3666{
3667 struct cpumask *cidmask = mm_cidmask(mm);
3668 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3669 int cid, max_nr_cid, allowed_max_nr_cid;
3670
3671 /*
3672 * After shrinking the number of threads or reducing the number
3673 * of allowed cpus, reduce the value of max_nr_cid so expansion
3674 * of cid allocation will preserve cache locality if the number
3675 * of threads or allowed cpus increase again.
3676 */
3677 max_nr_cid = atomic_read(&mm->max_nr_cid);
3678 while ((allowed_max_nr_cid = min_t(int, READ_ONCE(mm->nr_cpus_allowed),
3679 atomic_read(&mm->mm_users))),
3680 max_nr_cid > allowed_max_nr_cid) {
3681 /* atomic_try_cmpxchg loads previous mm->max_nr_cid into max_nr_cid. */
3682 if (atomic_try_cmpxchg(&mm->max_nr_cid, &max_nr_cid, allowed_max_nr_cid)) {
3683 max_nr_cid = allowed_max_nr_cid;
3684 break;
3685 }
3686 }
3687 /* Try to re-use recent cid. This improves cache locality. */
3688 cid = __this_cpu_read(pcpu_cid->recent_cid);
3689 if (!mm_cid_is_unset(cid) && cid < max_nr_cid &&
3690 !cpumask_test_and_set_cpu(cid, cidmask))
3691 return cid;
3692 /*
3693 * Expand cid allocation if the maximum number of concurrency
3694 * IDs allocated (max_nr_cid) is below the number cpus allowed
3695 * and number of threads. Expanding cid allocation as much as
3696 * possible improves cache locality.
3697 */
3698 cid = max_nr_cid;
3699 while (cid < READ_ONCE(mm->nr_cpus_allowed) && cid < atomic_read(&mm->mm_users)) {
3700 /* atomic_try_cmpxchg loads previous mm->max_nr_cid into cid. */
3701 if (!atomic_try_cmpxchg(&mm->max_nr_cid, &cid, cid + 1))
3702 continue;
3703 if (!cpumask_test_and_set_cpu(cid, cidmask))
3704 return cid;
3705 }
3706 /*
3707 * Find the first available concurrency id.
3708 * Retry finding first zero bit if the mask is temporarily
3709 * filled. This only happens during concurrent remote-clear
3710 * which owns a cid without holding a rq lock.
3711 */
3712 for (;;) {
3713 cid = cpumask_first_zero(cidmask);
3714 if (cid < READ_ONCE(mm->nr_cpus_allowed))
3715 break;
3716 cpu_relax();
3717 }
3718 if (cpumask_test_and_set_cpu(cid, cidmask))
3719 return -1;
3720
3721 return cid;
3722}
3723
3724/*
3725 * Save a snapshot of the current runqueue time of this cpu
3726 * with the per-cpu cid value, allowing to estimate how recently it was used.
3727 */
3728static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3729{
3730 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3731
3732 lockdep_assert_rq_held(rq);
3733 WRITE_ONCE(pcpu_cid->time, rq->clock);
3734}
3735
3736static inline int __mm_cid_get(struct rq *rq, struct task_struct *t,
3737 struct mm_struct *mm)
3738{
3739 int cid;
3740
3741 /*
3742 * All allocations (even those using the cid_lock) are lock-free. If
3743 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3744 * guarantee forward progress.
3745 */
3746 if (!READ_ONCE(use_cid_lock)) {
3747 cid = __mm_cid_try_get(t, mm);
3748 if (cid >= 0)
3749 goto end;
3750 raw_spin_lock(&cid_lock);
3751 } else {
3752 raw_spin_lock(&cid_lock);
3753 cid = __mm_cid_try_get(t, mm);
3754 if (cid >= 0)
3755 goto unlock;
3756 }
3757
3758 /*
3759 * cid concurrently allocated. Retry while forcing following
3760 * allocations to use the cid_lock to ensure forward progress.
3761 */
3762 WRITE_ONCE(use_cid_lock, 1);
3763 /*
3764 * Set use_cid_lock before allocation. Only care about program order
3765 * because this is only required for forward progress.
3766 */
3767 barrier();
3768 /*
3769 * Retry until it succeeds. It is guaranteed to eventually succeed once
3770 * all newcoming allocations observe the use_cid_lock flag set.
3771 */
3772 do {
3773 cid = __mm_cid_try_get(t, mm);
3774 cpu_relax();
3775 } while (cid < 0);
3776 /*
3777 * Allocate before clearing use_cid_lock. Only care about
3778 * program order because this is for forward progress.
3779 */
3780 barrier();
3781 WRITE_ONCE(use_cid_lock, 0);
3782unlock:
3783 raw_spin_unlock(&cid_lock);
3784end:
3785 mm_cid_snapshot_time(rq, mm);
3786
3787 return cid;
3788}
3789
3790static inline int mm_cid_get(struct rq *rq, struct task_struct *t,
3791 struct mm_struct *mm)
3792{
3793 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3794 struct cpumask *cpumask;
3795 int cid;
3796
3797 lockdep_assert_rq_held(rq);
3798 cpumask = mm_cidmask(mm);
3799 cid = __this_cpu_read(pcpu_cid->cid);
3800 if (mm_cid_is_valid(cid)) {
3801 mm_cid_snapshot_time(rq, mm);
3802 return cid;
3803 }
3804 if (mm_cid_is_lazy_put(cid)) {
3805 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3806 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3807 }
3808 cid = __mm_cid_get(rq, t, mm);
3809 __this_cpu_write(pcpu_cid->cid, cid);
3810 __this_cpu_write(pcpu_cid->recent_cid, cid);
3811
3812 return cid;
3813}
3814
3815static inline void switch_mm_cid(struct rq *rq,
3816 struct task_struct *prev,
3817 struct task_struct *next)
3818{
3819 /*
3820 * Provide a memory barrier between rq->curr store and load of
3821 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3822 *
3823 * Should be adapted if context_switch() is modified.
3824 */
3825 if (!next->mm) { // to kernel
3826 /*
3827 * user -> kernel transition does not guarantee a barrier, but
3828 * we can use the fact that it performs an atomic operation in
3829 * mmgrab().
3830 */
3831 if (prev->mm) // from user
3832 smp_mb__after_mmgrab();
3833 /*
3834 * kernel -> kernel transition does not change rq->curr->mm
3835 * state. It stays NULL.
3836 */
3837 } else { // to user
3838 /*
3839 * kernel -> user transition does not provide a barrier
3840 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3841 * Provide it here.
3842 */
3843 if (!prev->mm) { // from kernel
3844 smp_mb();
3845 } else { // from user
3846 /*
3847 * user->user transition relies on an implicit
3848 * memory barrier in switch_mm() when
3849 * current->mm changes. If the architecture
3850 * switch_mm() does not have an implicit memory
3851 * barrier, it is emitted here. If current->mm
3852 * is unchanged, no barrier is needed.
3853 */
3854 smp_mb__after_switch_mm();
3855 }
3856 }
3857 if (prev->mm_cid_active) {
3858 mm_cid_snapshot_time(rq, prev->mm);
3859 mm_cid_put_lazy(prev);
3860 prev->mm_cid = -1;
3861 }
3862 if (next->mm_cid_active)
3863 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next, next->mm);
3864}
3865
3866#else /* !CONFIG_SCHED_MM_CID: */
3867static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3868static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3869static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3870static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3871static inline void init_sched_mm_cid(struct task_struct *t) { }
3872#endif /* !CONFIG_SCHED_MM_CID */
3873
3874extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3875extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3876#ifdef CONFIG_SMP
3877static inline
3878void move_queued_task_locked(struct rq *src_rq, struct rq *dst_rq, struct task_struct *task)
3879{
3880 lockdep_assert_rq_held(src_rq);
3881 lockdep_assert_rq_held(dst_rq);
3882
3883 deactivate_task(src_rq, task, 0);
3884 set_task_cpu(task, dst_rq->cpu);
3885 activate_task(dst_rq, task, 0);
3886}
3887
3888static inline
3889bool task_is_pushable(struct rq *rq, struct task_struct *p, int cpu)
3890{
3891 if (!task_on_cpu(rq, p) &&
3892 cpumask_test_cpu(cpu, &p->cpus_mask))
3893 return true;
3894
3895 return false;
3896}
3897#endif
3898
3899#ifdef CONFIG_RT_MUTEXES
3900
3901static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3902{
3903 if (pi_task)
3904 prio = min(prio, pi_task->prio);
3905
3906 return prio;
3907}
3908
3909static inline int rt_effective_prio(struct task_struct *p, int prio)
3910{
3911 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3912
3913 return __rt_effective_prio(pi_task, prio);
3914}
3915
3916#else /* !CONFIG_RT_MUTEXES: */
3917
3918static inline int rt_effective_prio(struct task_struct *p, int prio)
3919{
3920 return prio;
3921}
3922
3923#endif /* !CONFIG_RT_MUTEXES */
3924
3925extern int __sched_setscheduler(struct task_struct *p, const struct sched_attr *attr, bool user, bool pi);
3926extern int __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3927extern const struct sched_class *__setscheduler_class(int policy, int prio);
3928extern void set_load_weight(struct task_struct *p, bool update_load);
3929extern void enqueue_task(struct rq *rq, struct task_struct *p, int flags);
3930extern bool dequeue_task(struct rq *rq, struct task_struct *p, int flags);
3931
3932extern void check_class_changing(struct rq *rq, struct task_struct *p,
3933 const struct sched_class *prev_class);
3934extern void check_class_changed(struct rq *rq, struct task_struct *p,
3935 const struct sched_class *prev_class,
3936 int oldprio);
3937
3938#ifdef CONFIG_SMP
3939extern struct balance_callback *splice_balance_callbacks(struct rq *rq);
3940extern void balance_callbacks(struct rq *rq, struct balance_callback *head);
3941#else
3942
3943static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
3944{
3945 return NULL;
3946}
3947
3948static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
3949{
3950}
3951
3952#endif
3953
3954#ifdef CONFIG_SCHED_CLASS_EXT
3955/*
3956 * Used by SCX in the enable/disable paths to move tasks between sched_classes
3957 * and establish invariants.
3958 */
3959struct sched_enq_and_set_ctx {
3960 struct task_struct *p;
3961 int queue_flags;
3962 bool queued;
3963 bool running;
3964};
3965
3966void sched_deq_and_put_task(struct task_struct *p, int queue_flags,
3967 struct sched_enq_and_set_ctx *ctx);
3968void sched_enq_and_set_task(struct sched_enq_and_set_ctx *ctx);
3969
3970#endif /* CONFIG_SCHED_CLASS_EXT */
3971
3972#include "ext.h"
3973
3974#endif /* _KERNEL_SCHED_SCHED_H */