Loading...
1
2#include <linux/sched.h>
3#include <linux/sched/sysctl.h>
4#include <linux/sched/rt.h>
5#include <linux/sched/deadline.h>
6#include <linux/mutex.h>
7#include <linux/spinlock.h>
8#include <linux/stop_machine.h>
9#include <linux/tick.h>
10#include <linux/slab.h>
11
12#include "cpupri.h"
13#include "cpudeadline.h"
14#include "cpuacct.h"
15
16struct rq;
17
18extern __read_mostly int scheduler_running;
19
20extern unsigned long calc_load_update;
21extern atomic_long_t calc_load_tasks;
22
23extern long calc_load_fold_active(struct rq *this_rq);
24extern void update_cpu_load_active(struct rq *this_rq);
25
26/*
27 * Helpers for converting nanosecond timing to jiffy resolution
28 */
29#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
30
31/*
32 * Increase resolution of nice-level calculations for 64-bit architectures.
33 * The extra resolution improves shares distribution and load balancing of
34 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
35 * hierarchies, especially on larger systems. This is not a user-visible change
36 * and does not change the user-interface for setting shares/weights.
37 *
38 * We increase resolution only if we have enough bits to allow this increased
39 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
40 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
41 * increased costs.
42 */
43#if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
44# define SCHED_LOAD_RESOLUTION 10
45# define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
46# define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
47#else
48# define SCHED_LOAD_RESOLUTION 0
49# define scale_load(w) (w)
50# define scale_load_down(w) (w)
51#endif
52
53#define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
54#define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
55
56#define NICE_0_LOAD SCHED_LOAD_SCALE
57#define NICE_0_SHIFT SCHED_LOAD_SHIFT
58
59/*
60 * Single value that decides SCHED_DEADLINE internal math precision.
61 * 10 -> just above 1us
62 * 9 -> just above 0.5us
63 */
64#define DL_SCALE (10)
65
66/*
67 * These are the 'tuning knobs' of the scheduler:
68 */
69
70/*
71 * single value that denotes runtime == period, ie unlimited time.
72 */
73#define RUNTIME_INF ((u64)~0ULL)
74
75static inline int fair_policy(int policy)
76{
77 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
78}
79
80static inline int rt_policy(int policy)
81{
82 return policy == SCHED_FIFO || policy == SCHED_RR;
83}
84
85static inline int dl_policy(int policy)
86{
87 return policy == SCHED_DEADLINE;
88}
89
90static inline int task_has_rt_policy(struct task_struct *p)
91{
92 return rt_policy(p->policy);
93}
94
95static inline int task_has_dl_policy(struct task_struct *p)
96{
97 return dl_policy(p->policy);
98}
99
100static inline bool dl_time_before(u64 a, u64 b)
101{
102 return (s64)(a - b) < 0;
103}
104
105/*
106 * Tells if entity @a should preempt entity @b.
107 */
108static inline bool
109dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
110{
111 return dl_time_before(a->deadline, b->deadline);
112}
113
114/*
115 * This is the priority-queue data structure of the RT scheduling class:
116 */
117struct rt_prio_array {
118 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
119 struct list_head queue[MAX_RT_PRIO];
120};
121
122struct rt_bandwidth {
123 /* nests inside the rq lock: */
124 raw_spinlock_t rt_runtime_lock;
125 ktime_t rt_period;
126 u64 rt_runtime;
127 struct hrtimer rt_period_timer;
128};
129/*
130 * To keep the bandwidth of -deadline tasks and groups under control
131 * we need some place where:
132 * - store the maximum -deadline bandwidth of the system (the group);
133 * - cache the fraction of that bandwidth that is currently allocated.
134 *
135 * This is all done in the data structure below. It is similar to the
136 * one used for RT-throttling (rt_bandwidth), with the main difference
137 * that, since here we are only interested in admission control, we
138 * do not decrease any runtime while the group "executes", neither we
139 * need a timer to replenish it.
140 *
141 * With respect to SMP, the bandwidth is given on a per-CPU basis,
142 * meaning that:
143 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
144 * - dl_total_bw array contains, in the i-eth element, the currently
145 * allocated bandwidth on the i-eth CPU.
146 * Moreover, groups consume bandwidth on each CPU, while tasks only
147 * consume bandwidth on the CPU they're running on.
148 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
149 * that will be shown the next time the proc or cgroup controls will
150 * be red. It on its turn can be changed by writing on its own
151 * control.
152 */
153struct dl_bandwidth {
154 raw_spinlock_t dl_runtime_lock;
155 u64 dl_runtime;
156 u64 dl_period;
157};
158
159static inline int dl_bandwidth_enabled(void)
160{
161 return sysctl_sched_rt_runtime >= 0;
162}
163
164extern struct dl_bw *dl_bw_of(int i);
165
166struct dl_bw {
167 raw_spinlock_t lock;
168 u64 bw, total_bw;
169};
170
171extern struct mutex sched_domains_mutex;
172
173#ifdef CONFIG_CGROUP_SCHED
174
175#include <linux/cgroup.h>
176
177struct cfs_rq;
178struct rt_rq;
179
180extern struct list_head task_groups;
181
182struct cfs_bandwidth {
183#ifdef CONFIG_CFS_BANDWIDTH
184 raw_spinlock_t lock;
185 ktime_t period;
186 u64 quota, runtime;
187 s64 hierarchal_quota;
188 u64 runtime_expires;
189
190 int idle, timer_active;
191 struct hrtimer period_timer, slack_timer;
192 struct list_head throttled_cfs_rq;
193
194 /* statistics */
195 int nr_periods, nr_throttled;
196 u64 throttled_time;
197#endif
198};
199
200/* task group related information */
201struct task_group {
202 struct cgroup_subsys_state css;
203
204#ifdef CONFIG_FAIR_GROUP_SCHED
205 /* schedulable entities of this group on each cpu */
206 struct sched_entity **se;
207 /* runqueue "owned" by this group on each cpu */
208 struct cfs_rq **cfs_rq;
209 unsigned long shares;
210
211#ifdef CONFIG_SMP
212 atomic_long_t load_avg;
213 atomic_t runnable_avg;
214#endif
215#endif
216
217#ifdef CONFIG_RT_GROUP_SCHED
218 struct sched_rt_entity **rt_se;
219 struct rt_rq **rt_rq;
220
221 struct rt_bandwidth rt_bandwidth;
222#endif
223
224 struct rcu_head rcu;
225 struct list_head list;
226
227 struct task_group *parent;
228 struct list_head siblings;
229 struct list_head children;
230
231#ifdef CONFIG_SCHED_AUTOGROUP
232 struct autogroup *autogroup;
233#endif
234
235 struct cfs_bandwidth cfs_bandwidth;
236};
237
238#ifdef CONFIG_FAIR_GROUP_SCHED
239#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
240
241/*
242 * A weight of 0 or 1 can cause arithmetics problems.
243 * A weight of a cfs_rq is the sum of weights of which entities
244 * are queued on this cfs_rq, so a weight of a entity should not be
245 * too large, so as the shares value of a task group.
246 * (The default weight is 1024 - so there's no practical
247 * limitation from this.)
248 */
249#define MIN_SHARES (1UL << 1)
250#define MAX_SHARES (1UL << 18)
251#endif
252
253typedef int (*tg_visitor)(struct task_group *, void *);
254
255extern int walk_tg_tree_from(struct task_group *from,
256 tg_visitor down, tg_visitor up, void *data);
257
258/*
259 * Iterate the full tree, calling @down when first entering a node and @up when
260 * leaving it for the final time.
261 *
262 * Caller must hold rcu_lock or sufficient equivalent.
263 */
264static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
265{
266 return walk_tg_tree_from(&root_task_group, down, up, data);
267}
268
269extern int tg_nop(struct task_group *tg, void *data);
270
271extern void free_fair_sched_group(struct task_group *tg);
272extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
273extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
274extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
275 struct sched_entity *se, int cpu,
276 struct sched_entity *parent);
277extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
278extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
279
280extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
281extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
282extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
283
284extern void free_rt_sched_group(struct task_group *tg);
285extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
286extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
287 struct sched_rt_entity *rt_se, int cpu,
288 struct sched_rt_entity *parent);
289
290extern struct task_group *sched_create_group(struct task_group *parent);
291extern void sched_online_group(struct task_group *tg,
292 struct task_group *parent);
293extern void sched_destroy_group(struct task_group *tg);
294extern void sched_offline_group(struct task_group *tg);
295
296extern void sched_move_task(struct task_struct *tsk);
297
298#ifdef CONFIG_FAIR_GROUP_SCHED
299extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
300#endif
301
302#else /* CONFIG_CGROUP_SCHED */
303
304struct cfs_bandwidth { };
305
306#endif /* CONFIG_CGROUP_SCHED */
307
308/* CFS-related fields in a runqueue */
309struct cfs_rq {
310 struct load_weight load;
311 unsigned int nr_running, h_nr_running;
312
313 u64 exec_clock;
314 u64 min_vruntime;
315#ifndef CONFIG_64BIT
316 u64 min_vruntime_copy;
317#endif
318
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
321
322 /*
323 * 'curr' points to currently running entity on this cfs_rq.
324 * It is set to NULL otherwise (i.e when none are currently running).
325 */
326 struct sched_entity *curr, *next, *last, *skip;
327
328#ifdef CONFIG_SCHED_DEBUG
329 unsigned int nr_spread_over;
330#endif
331
332#ifdef CONFIG_SMP
333 /*
334 * CFS Load tracking
335 * Under CFS, load is tracked on a per-entity basis and aggregated up.
336 * This allows for the description of both thread and group usage (in
337 * the FAIR_GROUP_SCHED case).
338 */
339 unsigned long runnable_load_avg, blocked_load_avg;
340 atomic64_t decay_counter;
341 u64 last_decay;
342 atomic_long_t removed_load;
343
344#ifdef CONFIG_FAIR_GROUP_SCHED
345 /* Required to track per-cpu representation of a task_group */
346 u32 tg_runnable_contrib;
347 unsigned long tg_load_contrib;
348
349 /*
350 * h_load = weight * f(tg)
351 *
352 * Where f(tg) is the recursive weight fraction assigned to
353 * this group.
354 */
355 unsigned long h_load;
356 u64 last_h_load_update;
357 struct sched_entity *h_load_next;
358#endif /* CONFIG_FAIR_GROUP_SCHED */
359#endif /* CONFIG_SMP */
360
361#ifdef CONFIG_FAIR_GROUP_SCHED
362 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
363
364 /*
365 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
366 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
367 * (like users, containers etc.)
368 *
369 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
370 * list is used during load balance.
371 */
372 int on_list;
373 struct list_head leaf_cfs_rq_list;
374 struct task_group *tg; /* group that "owns" this runqueue */
375
376#ifdef CONFIG_CFS_BANDWIDTH
377 int runtime_enabled;
378 u64 runtime_expires;
379 s64 runtime_remaining;
380
381 u64 throttled_clock, throttled_clock_task;
382 u64 throttled_clock_task_time;
383 int throttled, throttle_count;
384 struct list_head throttled_list;
385#endif /* CONFIG_CFS_BANDWIDTH */
386#endif /* CONFIG_FAIR_GROUP_SCHED */
387};
388
389static inline int rt_bandwidth_enabled(void)
390{
391 return sysctl_sched_rt_runtime >= 0;
392}
393
394/* Real-Time classes' related field in a runqueue: */
395struct rt_rq {
396 struct rt_prio_array active;
397 unsigned int rt_nr_running;
398#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
399 struct {
400 int curr; /* highest queued rt task prio */
401#ifdef CONFIG_SMP
402 int next; /* next highest */
403#endif
404 } highest_prio;
405#endif
406#ifdef CONFIG_SMP
407 unsigned long rt_nr_migratory;
408 unsigned long rt_nr_total;
409 int overloaded;
410 struct plist_head pushable_tasks;
411#endif
412 int rt_throttled;
413 u64 rt_time;
414 u64 rt_runtime;
415 /* Nests inside the rq lock: */
416 raw_spinlock_t rt_runtime_lock;
417
418#ifdef CONFIG_RT_GROUP_SCHED
419 unsigned long rt_nr_boosted;
420
421 struct rq *rq;
422 struct task_group *tg;
423#endif
424};
425
426#ifdef CONFIG_RT_GROUP_SCHED
427static inline int rt_rq_throttled(struct rt_rq *rt_rq)
428{
429 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
430}
431#else
432static inline int rt_rq_throttled(struct rt_rq *rt_rq)
433{
434 return rt_rq->rt_throttled;
435}
436#endif
437
438/* Deadline class' related fields in a runqueue */
439struct dl_rq {
440 /* runqueue is an rbtree, ordered by deadline */
441 struct rb_root rb_root;
442 struct rb_node *rb_leftmost;
443
444 unsigned long dl_nr_running;
445
446#ifdef CONFIG_SMP
447 /*
448 * Deadline values of the currently executing and the
449 * earliest ready task on this rq. Caching these facilitates
450 * the decision wether or not a ready but not running task
451 * should migrate somewhere else.
452 */
453 struct {
454 u64 curr;
455 u64 next;
456 } earliest_dl;
457
458 unsigned long dl_nr_migratory;
459 int overloaded;
460
461 /*
462 * Tasks on this rq that can be pushed away. They are kept in
463 * an rb-tree, ordered by tasks' deadlines, with caching
464 * of the leftmost (earliest deadline) element.
465 */
466 struct rb_root pushable_dl_tasks_root;
467 struct rb_node *pushable_dl_tasks_leftmost;
468#else
469 struct dl_bw dl_bw;
470#endif
471};
472
473#ifdef CONFIG_SMP
474
475/*
476 * We add the notion of a root-domain which will be used to define per-domain
477 * variables. Each exclusive cpuset essentially defines an island domain by
478 * fully partitioning the member cpus from any other cpuset. Whenever a new
479 * exclusive cpuset is created, we also create and attach a new root-domain
480 * object.
481 *
482 */
483struct root_domain {
484 atomic_t refcount;
485 atomic_t rto_count;
486 struct rcu_head rcu;
487 cpumask_var_t span;
488 cpumask_var_t online;
489
490 /*
491 * The bit corresponding to a CPU gets set here if such CPU has more
492 * than one runnable -deadline task (as it is below for RT tasks).
493 */
494 cpumask_var_t dlo_mask;
495 atomic_t dlo_count;
496 struct dl_bw dl_bw;
497 struct cpudl cpudl;
498
499 /*
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
502 */
503 cpumask_var_t rto_mask;
504 struct cpupri cpupri;
505};
506
507extern struct root_domain def_root_domain;
508
509#endif /* CONFIG_SMP */
510
511/*
512 * This is the main, per-CPU runqueue data structure.
513 *
514 * Locking rule: those places that want to lock multiple runqueues
515 * (such as the load balancing or the thread migration code), lock
516 * acquire operations must be ordered by ascending &runqueue.
517 */
518struct rq {
519 /* runqueue lock: */
520 raw_spinlock_t lock;
521
522 /*
523 * nr_running and cpu_load should be in the same cacheline because
524 * remote CPUs use both these fields when doing load calculation.
525 */
526 unsigned int nr_running;
527#ifdef CONFIG_NUMA_BALANCING
528 unsigned int nr_numa_running;
529 unsigned int nr_preferred_running;
530#endif
531 #define CPU_LOAD_IDX_MAX 5
532 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
533 unsigned long last_load_update_tick;
534#ifdef CONFIG_NO_HZ_COMMON
535 u64 nohz_stamp;
536 unsigned long nohz_flags;
537#endif
538#ifdef CONFIG_NO_HZ_FULL
539 unsigned long last_sched_tick;
540#endif
541 int skip_clock_update;
542
543 /* capture load from *all* tasks on this cpu: */
544 struct load_weight load;
545 unsigned long nr_load_updates;
546 u64 nr_switches;
547
548 struct cfs_rq cfs;
549 struct rt_rq rt;
550 struct dl_rq dl;
551
552#ifdef CONFIG_FAIR_GROUP_SCHED
553 /* list of leaf cfs_rq on this cpu: */
554 struct list_head leaf_cfs_rq_list;
555
556 struct sched_avg avg;
557#endif /* CONFIG_FAIR_GROUP_SCHED */
558
559 /*
560 * This is part of a global counter where only the total sum
561 * over all CPUs matters. A task can increase this counter on
562 * one CPU and if it got migrated afterwards it may decrease
563 * it on another CPU. Always updated under the runqueue lock:
564 */
565 unsigned long nr_uninterruptible;
566
567 struct task_struct *curr, *idle, *stop;
568 unsigned long next_balance;
569 struct mm_struct *prev_mm;
570
571 u64 clock;
572 u64 clock_task;
573
574 atomic_t nr_iowait;
575
576#ifdef CONFIG_SMP
577 struct root_domain *rd;
578 struct sched_domain *sd;
579
580 unsigned long cpu_power;
581
582 unsigned char idle_balance;
583 /* For active balancing */
584 int post_schedule;
585 int active_balance;
586 int push_cpu;
587 struct cpu_stop_work active_balance_work;
588 /* cpu of this runqueue: */
589 int cpu;
590 int online;
591
592 struct list_head cfs_tasks;
593
594 u64 rt_avg;
595 u64 age_stamp;
596 u64 idle_stamp;
597 u64 avg_idle;
598
599 /* This is used to determine avg_idle's max value */
600 u64 max_idle_balance_cost;
601#endif
602
603#ifdef CONFIG_IRQ_TIME_ACCOUNTING
604 u64 prev_irq_time;
605#endif
606#ifdef CONFIG_PARAVIRT
607 u64 prev_steal_time;
608#endif
609#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
610 u64 prev_steal_time_rq;
611#endif
612
613 /* calc_load related fields */
614 unsigned long calc_load_update;
615 long calc_load_active;
616
617#ifdef CONFIG_SCHED_HRTICK
618#ifdef CONFIG_SMP
619 int hrtick_csd_pending;
620 struct call_single_data hrtick_csd;
621#endif
622 struct hrtimer hrtick_timer;
623#endif
624
625#ifdef CONFIG_SCHEDSTATS
626 /* latency stats */
627 struct sched_info rq_sched_info;
628 unsigned long long rq_cpu_time;
629 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
630
631 /* sys_sched_yield() stats */
632 unsigned int yld_count;
633
634 /* schedule() stats */
635 unsigned int sched_count;
636 unsigned int sched_goidle;
637
638 /* try_to_wake_up() stats */
639 unsigned int ttwu_count;
640 unsigned int ttwu_local;
641#endif
642
643#ifdef CONFIG_SMP
644 struct llist_head wake_list;
645#endif
646};
647
648static inline int cpu_of(struct rq *rq)
649{
650#ifdef CONFIG_SMP
651 return rq->cpu;
652#else
653 return 0;
654#endif
655}
656
657DECLARE_PER_CPU(struct rq, runqueues);
658
659#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660#define this_rq() (&__get_cpu_var(runqueues))
661#define task_rq(p) cpu_rq(task_cpu(p))
662#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663#define raw_rq() (&__raw_get_cpu_var(runqueues))
664
665static inline u64 rq_clock(struct rq *rq)
666{
667 return rq->clock;
668}
669
670static inline u64 rq_clock_task(struct rq *rq)
671{
672 return rq->clock_task;
673}
674
675#ifdef CONFIG_NUMA_BALANCING
676extern void sched_setnuma(struct task_struct *p, int node);
677extern int migrate_task_to(struct task_struct *p, int cpu);
678extern int migrate_swap(struct task_struct *, struct task_struct *);
679#endif /* CONFIG_NUMA_BALANCING */
680
681#ifdef CONFIG_SMP
682
683#define rcu_dereference_check_sched_domain(p) \
684 rcu_dereference_check((p), \
685 lockdep_is_held(&sched_domains_mutex))
686
687/*
688 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
689 * See detach_destroy_domains: synchronize_sched for details.
690 *
691 * The domain tree of any CPU may only be accessed from within
692 * preempt-disabled sections.
693 */
694#define for_each_domain(cpu, __sd) \
695 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
696 __sd; __sd = __sd->parent)
697
698#define for_each_lower_domain(sd) for (; sd; sd = sd->child)
699
700/**
701 * highest_flag_domain - Return highest sched_domain containing flag.
702 * @cpu: The cpu whose highest level of sched domain is to
703 * be returned.
704 * @flag: The flag to check for the highest sched_domain
705 * for the given cpu.
706 *
707 * Returns the highest sched_domain of a cpu which contains the given flag.
708 */
709static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
710{
711 struct sched_domain *sd, *hsd = NULL;
712
713 for_each_domain(cpu, sd) {
714 if (!(sd->flags & flag))
715 break;
716 hsd = sd;
717 }
718
719 return hsd;
720}
721
722static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
723{
724 struct sched_domain *sd;
725
726 for_each_domain(cpu, sd) {
727 if (sd->flags & flag)
728 break;
729 }
730
731 return sd;
732}
733
734DECLARE_PER_CPU(struct sched_domain *, sd_llc);
735DECLARE_PER_CPU(int, sd_llc_size);
736DECLARE_PER_CPU(int, sd_llc_id);
737DECLARE_PER_CPU(struct sched_domain *, sd_numa);
738DECLARE_PER_CPU(struct sched_domain *, sd_busy);
739DECLARE_PER_CPU(struct sched_domain *, sd_asym);
740
741struct sched_group_power {
742 atomic_t ref;
743 /*
744 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
745 * single CPU.
746 */
747 unsigned int power, power_orig;
748 unsigned long next_update;
749 int imbalance; /* XXX unrelated to power but shared group state */
750 /*
751 * Number of busy cpus in this group.
752 */
753 atomic_t nr_busy_cpus;
754
755 unsigned long cpumask[0]; /* iteration mask */
756};
757
758struct sched_group {
759 struct sched_group *next; /* Must be a circular list */
760 atomic_t ref;
761
762 unsigned int group_weight;
763 struct sched_group_power *sgp;
764
765 /*
766 * The CPUs this group covers.
767 *
768 * NOTE: this field is variable length. (Allocated dynamically
769 * by attaching extra space to the end of the structure,
770 * depending on how many CPUs the kernel has booted up with)
771 */
772 unsigned long cpumask[0];
773};
774
775static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
776{
777 return to_cpumask(sg->cpumask);
778}
779
780/*
781 * cpumask masking which cpus in the group are allowed to iterate up the domain
782 * tree.
783 */
784static inline struct cpumask *sched_group_mask(struct sched_group *sg)
785{
786 return to_cpumask(sg->sgp->cpumask);
787}
788
789/**
790 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
791 * @group: The group whose first cpu is to be returned.
792 */
793static inline unsigned int group_first_cpu(struct sched_group *group)
794{
795 return cpumask_first(sched_group_cpus(group));
796}
797
798extern int group_balance_cpu(struct sched_group *sg);
799
800#endif /* CONFIG_SMP */
801
802#include "stats.h"
803#include "auto_group.h"
804
805#ifdef CONFIG_CGROUP_SCHED
806
807/*
808 * Return the group to which this tasks belongs.
809 *
810 * We cannot use task_css() and friends because the cgroup subsystem
811 * changes that value before the cgroup_subsys::attach() method is called,
812 * therefore we cannot pin it and might observe the wrong value.
813 *
814 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
815 * core changes this before calling sched_move_task().
816 *
817 * Instead we use a 'copy' which is updated from sched_move_task() while
818 * holding both task_struct::pi_lock and rq::lock.
819 */
820static inline struct task_group *task_group(struct task_struct *p)
821{
822 return p->sched_task_group;
823}
824
825/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
826static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
827{
828#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
829 struct task_group *tg = task_group(p);
830#endif
831
832#ifdef CONFIG_FAIR_GROUP_SCHED
833 p->se.cfs_rq = tg->cfs_rq[cpu];
834 p->se.parent = tg->se[cpu];
835#endif
836
837#ifdef CONFIG_RT_GROUP_SCHED
838 p->rt.rt_rq = tg->rt_rq[cpu];
839 p->rt.parent = tg->rt_se[cpu];
840#endif
841}
842
843#else /* CONFIG_CGROUP_SCHED */
844
845static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
846static inline struct task_group *task_group(struct task_struct *p)
847{
848 return NULL;
849}
850
851#endif /* CONFIG_CGROUP_SCHED */
852
853static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
854{
855 set_task_rq(p, cpu);
856#ifdef CONFIG_SMP
857 /*
858 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
859 * successfuly executed on another CPU. We must ensure that updates of
860 * per-task data have been completed by this moment.
861 */
862 smp_wmb();
863 task_thread_info(p)->cpu = cpu;
864 p->wake_cpu = cpu;
865#endif
866}
867
868/*
869 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
870 */
871#ifdef CONFIG_SCHED_DEBUG
872# include <linux/static_key.h>
873# define const_debug __read_mostly
874#else
875# define const_debug const
876#endif
877
878extern const_debug unsigned int sysctl_sched_features;
879
880#define SCHED_FEAT(name, enabled) \
881 __SCHED_FEAT_##name ,
882
883enum {
884#include "features.h"
885 __SCHED_FEAT_NR,
886};
887
888#undef SCHED_FEAT
889
890#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
891static __always_inline bool static_branch__true(struct static_key *key)
892{
893 return static_key_true(key); /* Not out of line branch. */
894}
895
896static __always_inline bool static_branch__false(struct static_key *key)
897{
898 return static_key_false(key); /* Out of line branch. */
899}
900
901#define SCHED_FEAT(name, enabled) \
902static __always_inline bool static_branch_##name(struct static_key *key) \
903{ \
904 return static_branch__##enabled(key); \
905}
906
907#include "features.h"
908
909#undef SCHED_FEAT
910
911extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
912#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
913#else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
914#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
915#endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
916
917#ifdef CONFIG_NUMA_BALANCING
918#define sched_feat_numa(x) sched_feat(x)
919#ifdef CONFIG_SCHED_DEBUG
920#define numabalancing_enabled sched_feat_numa(NUMA)
921#else
922extern bool numabalancing_enabled;
923#endif /* CONFIG_SCHED_DEBUG */
924#else
925#define sched_feat_numa(x) (0)
926#define numabalancing_enabled (0)
927#endif /* CONFIG_NUMA_BALANCING */
928
929static inline u64 global_rt_period(void)
930{
931 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
932}
933
934static inline u64 global_rt_runtime(void)
935{
936 if (sysctl_sched_rt_runtime < 0)
937 return RUNTIME_INF;
938
939 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
940}
941
942static inline int task_current(struct rq *rq, struct task_struct *p)
943{
944 return rq->curr == p;
945}
946
947static inline int task_running(struct rq *rq, struct task_struct *p)
948{
949#ifdef CONFIG_SMP
950 return p->on_cpu;
951#else
952 return task_current(rq, p);
953#endif
954}
955
956
957#ifndef prepare_arch_switch
958# define prepare_arch_switch(next) do { } while (0)
959#endif
960#ifndef finish_arch_switch
961# define finish_arch_switch(prev) do { } while (0)
962#endif
963#ifndef finish_arch_post_lock_switch
964# define finish_arch_post_lock_switch() do { } while (0)
965#endif
966
967#ifndef __ARCH_WANT_UNLOCKED_CTXSW
968static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
969{
970#ifdef CONFIG_SMP
971 /*
972 * We can optimise this out completely for !SMP, because the
973 * SMP rebalancing from interrupt is the only thing that cares
974 * here.
975 */
976 next->on_cpu = 1;
977#endif
978}
979
980static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
981{
982#ifdef CONFIG_SMP
983 /*
984 * After ->on_cpu is cleared, the task can be moved to a different CPU.
985 * We must ensure this doesn't happen until the switch is completely
986 * finished.
987 */
988 smp_wmb();
989 prev->on_cpu = 0;
990#endif
991#ifdef CONFIG_DEBUG_SPINLOCK
992 /* this is a valid case when another task releases the spinlock */
993 rq->lock.owner = current;
994#endif
995 /*
996 * If we are tracking spinlock dependencies then we have to
997 * fix up the runqueue lock - which gets 'carried over' from
998 * prev into current:
999 */
1000 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1001
1002 raw_spin_unlock_irq(&rq->lock);
1003}
1004
1005#else /* __ARCH_WANT_UNLOCKED_CTXSW */
1006static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1007{
1008#ifdef CONFIG_SMP
1009 /*
1010 * We can optimise this out completely for !SMP, because the
1011 * SMP rebalancing from interrupt is the only thing that cares
1012 * here.
1013 */
1014 next->on_cpu = 1;
1015#endif
1016 raw_spin_unlock(&rq->lock);
1017}
1018
1019static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1020{
1021#ifdef CONFIG_SMP
1022 /*
1023 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1024 * We must ensure this doesn't happen until the switch is completely
1025 * finished.
1026 */
1027 smp_wmb();
1028 prev->on_cpu = 0;
1029#endif
1030 local_irq_enable();
1031}
1032#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1033
1034/*
1035 * wake flags
1036 */
1037#define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1038#define WF_FORK 0x02 /* child wakeup after fork */
1039#define WF_MIGRATED 0x4 /* internal use, task got migrated */
1040
1041/*
1042 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1043 * of tasks with abnormal "nice" values across CPUs the contribution that
1044 * each task makes to its run queue's load is weighted according to its
1045 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1046 * scaled version of the new time slice allocation that they receive on time
1047 * slice expiry etc.
1048 */
1049
1050#define WEIGHT_IDLEPRIO 3
1051#define WMULT_IDLEPRIO 1431655765
1052
1053/*
1054 * Nice levels are multiplicative, with a gentle 10% change for every
1055 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1056 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1057 * that remained on nice 0.
1058 *
1059 * The "10% effect" is relative and cumulative: from _any_ nice level,
1060 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1061 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1062 * If a task goes up by ~10% and another task goes down by ~10% then
1063 * the relative distance between them is ~25%.)
1064 */
1065static const int prio_to_weight[40] = {
1066 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1067 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1068 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1069 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1070 /* 0 */ 1024, 820, 655, 526, 423,
1071 /* 5 */ 335, 272, 215, 172, 137,
1072 /* 10 */ 110, 87, 70, 56, 45,
1073 /* 15 */ 36, 29, 23, 18, 15,
1074};
1075
1076/*
1077 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1078 *
1079 * In cases where the weight does not change often, we can use the
1080 * precalculated inverse to speed up arithmetics by turning divisions
1081 * into multiplications:
1082 */
1083static const u32 prio_to_wmult[40] = {
1084 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1085 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1086 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1087 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1088 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1089 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1090 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1091 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1092};
1093
1094#define ENQUEUE_WAKEUP 1
1095#define ENQUEUE_HEAD 2
1096#ifdef CONFIG_SMP
1097#define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1098#else
1099#define ENQUEUE_WAKING 0
1100#endif
1101#define ENQUEUE_REPLENISH 8
1102
1103#define DEQUEUE_SLEEP 1
1104
1105#define RETRY_TASK ((void *)-1UL)
1106
1107struct sched_class {
1108 const struct sched_class *next;
1109
1110 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1111 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1112 void (*yield_task) (struct rq *rq);
1113 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1114
1115 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1116
1117 /*
1118 * It is the responsibility of the pick_next_task() method that will
1119 * return the next task to call put_prev_task() on the @prev task or
1120 * something equivalent.
1121 *
1122 * May return RETRY_TASK when it finds a higher prio class has runnable
1123 * tasks.
1124 */
1125 struct task_struct * (*pick_next_task) (struct rq *rq,
1126 struct task_struct *prev);
1127 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1128
1129#ifdef CONFIG_SMP
1130 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1131 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1132
1133 void (*post_schedule) (struct rq *this_rq);
1134 void (*task_waking) (struct task_struct *task);
1135 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1136
1137 void (*set_cpus_allowed)(struct task_struct *p,
1138 const struct cpumask *newmask);
1139
1140 void (*rq_online)(struct rq *rq);
1141 void (*rq_offline)(struct rq *rq);
1142#endif
1143
1144 void (*set_curr_task) (struct rq *rq);
1145 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1146 void (*task_fork) (struct task_struct *p);
1147 void (*task_dead) (struct task_struct *p);
1148
1149 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1150 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1151 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1152 int oldprio);
1153
1154 unsigned int (*get_rr_interval) (struct rq *rq,
1155 struct task_struct *task);
1156
1157#ifdef CONFIG_FAIR_GROUP_SCHED
1158 void (*task_move_group) (struct task_struct *p, int on_rq);
1159#endif
1160};
1161
1162static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1163{
1164 prev->sched_class->put_prev_task(rq, prev);
1165}
1166
1167#define sched_class_highest (&stop_sched_class)
1168#define for_each_class(class) \
1169 for (class = sched_class_highest; class; class = class->next)
1170
1171extern const struct sched_class stop_sched_class;
1172extern const struct sched_class dl_sched_class;
1173extern const struct sched_class rt_sched_class;
1174extern const struct sched_class fair_sched_class;
1175extern const struct sched_class idle_sched_class;
1176
1177
1178#ifdef CONFIG_SMP
1179
1180extern void update_group_power(struct sched_domain *sd, int cpu);
1181
1182extern void trigger_load_balance(struct rq *rq);
1183
1184extern void idle_enter_fair(struct rq *this_rq);
1185extern void idle_exit_fair(struct rq *this_rq);
1186
1187#else
1188
1189static inline void idle_enter_fair(struct rq *rq) { }
1190static inline void idle_exit_fair(struct rq *rq) { }
1191
1192#endif
1193
1194extern void sysrq_sched_debug_show(void);
1195extern void sched_init_granularity(void);
1196extern void update_max_interval(void);
1197
1198extern void init_sched_dl_class(void);
1199extern void init_sched_rt_class(void);
1200extern void init_sched_fair_class(void);
1201extern void init_sched_dl_class(void);
1202
1203extern void resched_task(struct task_struct *p);
1204extern void resched_cpu(int cpu);
1205
1206extern struct rt_bandwidth def_rt_bandwidth;
1207extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1208
1209extern struct dl_bandwidth def_dl_bandwidth;
1210extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1211extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1212
1213unsigned long to_ratio(u64 period, u64 runtime);
1214
1215extern void update_idle_cpu_load(struct rq *this_rq);
1216
1217extern void init_task_runnable_average(struct task_struct *p);
1218
1219static inline void inc_nr_running(struct rq *rq)
1220{
1221 rq->nr_running++;
1222
1223#ifdef CONFIG_NO_HZ_FULL
1224 if (rq->nr_running == 2) {
1225 if (tick_nohz_full_cpu(rq->cpu)) {
1226 /* Order rq->nr_running write against the IPI */
1227 smp_wmb();
1228 smp_send_reschedule(rq->cpu);
1229 }
1230 }
1231#endif
1232}
1233
1234static inline void dec_nr_running(struct rq *rq)
1235{
1236 rq->nr_running--;
1237}
1238
1239static inline void rq_last_tick_reset(struct rq *rq)
1240{
1241#ifdef CONFIG_NO_HZ_FULL
1242 rq->last_sched_tick = jiffies;
1243#endif
1244}
1245
1246extern void update_rq_clock(struct rq *rq);
1247
1248extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1249extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1250
1251extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1252
1253extern const_debug unsigned int sysctl_sched_time_avg;
1254extern const_debug unsigned int sysctl_sched_nr_migrate;
1255extern const_debug unsigned int sysctl_sched_migration_cost;
1256
1257static inline u64 sched_avg_period(void)
1258{
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1260}
1261
1262#ifdef CONFIG_SCHED_HRTICK
1263
1264/*
1265 * Use hrtick when:
1266 * - enabled by features
1267 * - hrtimer is actually high res
1268 */
1269static inline int hrtick_enabled(struct rq *rq)
1270{
1271 if (!sched_feat(HRTICK))
1272 return 0;
1273 if (!cpu_active(cpu_of(rq)))
1274 return 0;
1275 return hrtimer_is_hres_active(&rq->hrtick_timer);
1276}
1277
1278void hrtick_start(struct rq *rq, u64 delay);
1279
1280#else
1281
1282static inline int hrtick_enabled(struct rq *rq)
1283{
1284 return 0;
1285}
1286
1287#endif /* CONFIG_SCHED_HRTICK */
1288
1289#ifdef CONFIG_SMP
1290extern void sched_avg_update(struct rq *rq);
1291static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1292{
1293 rq->rt_avg += rt_delta;
1294 sched_avg_update(rq);
1295}
1296#else
1297static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1298static inline void sched_avg_update(struct rq *rq) { }
1299#endif
1300
1301extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1302
1303#ifdef CONFIG_SMP
1304#ifdef CONFIG_PREEMPT
1305
1306static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1307
1308/*
1309 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1310 * way at the expense of forcing extra atomic operations in all
1311 * invocations. This assures that the double_lock is acquired using the
1312 * same underlying policy as the spinlock_t on this architecture, which
1313 * reduces latency compared to the unfair variant below. However, it
1314 * also adds more overhead and therefore may reduce throughput.
1315 */
1316static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1317 __releases(this_rq->lock)
1318 __acquires(busiest->lock)
1319 __acquires(this_rq->lock)
1320{
1321 raw_spin_unlock(&this_rq->lock);
1322 double_rq_lock(this_rq, busiest);
1323
1324 return 1;
1325}
1326
1327#else
1328/*
1329 * Unfair double_lock_balance: Optimizes throughput at the expense of
1330 * latency by eliminating extra atomic operations when the locks are
1331 * already in proper order on entry. This favors lower cpu-ids and will
1332 * grant the double lock to lower cpus over higher ids under contention,
1333 * regardless of entry order into the function.
1334 */
1335static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1336 __releases(this_rq->lock)
1337 __acquires(busiest->lock)
1338 __acquires(this_rq->lock)
1339{
1340 int ret = 0;
1341
1342 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1343 if (busiest < this_rq) {
1344 raw_spin_unlock(&this_rq->lock);
1345 raw_spin_lock(&busiest->lock);
1346 raw_spin_lock_nested(&this_rq->lock,
1347 SINGLE_DEPTH_NESTING);
1348 ret = 1;
1349 } else
1350 raw_spin_lock_nested(&busiest->lock,
1351 SINGLE_DEPTH_NESTING);
1352 }
1353 return ret;
1354}
1355
1356#endif /* CONFIG_PREEMPT */
1357
1358/*
1359 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1360 */
1361static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1362{
1363 if (unlikely(!irqs_disabled())) {
1364 /* printk() doesn't work good under rq->lock */
1365 raw_spin_unlock(&this_rq->lock);
1366 BUG_ON(1);
1367 }
1368
1369 return _double_lock_balance(this_rq, busiest);
1370}
1371
1372static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1373 __releases(busiest->lock)
1374{
1375 raw_spin_unlock(&busiest->lock);
1376 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1377}
1378
1379static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1380{
1381 if (l1 > l2)
1382 swap(l1, l2);
1383
1384 spin_lock(l1);
1385 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1386}
1387
1388static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1389{
1390 if (l1 > l2)
1391 swap(l1, l2);
1392
1393 spin_lock_irq(l1);
1394 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1395}
1396
1397static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1398{
1399 if (l1 > l2)
1400 swap(l1, l2);
1401
1402 raw_spin_lock(l1);
1403 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1404}
1405
1406/*
1407 * double_rq_lock - safely lock two runqueues
1408 *
1409 * Note this does not disable interrupts like task_rq_lock,
1410 * you need to do so manually before calling.
1411 */
1412static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1413 __acquires(rq1->lock)
1414 __acquires(rq2->lock)
1415{
1416 BUG_ON(!irqs_disabled());
1417 if (rq1 == rq2) {
1418 raw_spin_lock(&rq1->lock);
1419 __acquire(rq2->lock); /* Fake it out ;) */
1420 } else {
1421 if (rq1 < rq2) {
1422 raw_spin_lock(&rq1->lock);
1423 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1424 } else {
1425 raw_spin_lock(&rq2->lock);
1426 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1427 }
1428 }
1429}
1430
1431/*
1432 * double_rq_unlock - safely unlock two runqueues
1433 *
1434 * Note this does not restore interrupts like task_rq_unlock,
1435 * you need to do so manually after calling.
1436 */
1437static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1438 __releases(rq1->lock)
1439 __releases(rq2->lock)
1440{
1441 raw_spin_unlock(&rq1->lock);
1442 if (rq1 != rq2)
1443 raw_spin_unlock(&rq2->lock);
1444 else
1445 __release(rq2->lock);
1446}
1447
1448#else /* CONFIG_SMP */
1449
1450/*
1451 * double_rq_lock - safely lock two runqueues
1452 *
1453 * Note this does not disable interrupts like task_rq_lock,
1454 * you need to do so manually before calling.
1455 */
1456static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1457 __acquires(rq1->lock)
1458 __acquires(rq2->lock)
1459{
1460 BUG_ON(!irqs_disabled());
1461 BUG_ON(rq1 != rq2);
1462 raw_spin_lock(&rq1->lock);
1463 __acquire(rq2->lock); /* Fake it out ;) */
1464}
1465
1466/*
1467 * double_rq_unlock - safely unlock two runqueues
1468 *
1469 * Note this does not restore interrupts like task_rq_unlock,
1470 * you need to do so manually after calling.
1471 */
1472static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1473 __releases(rq1->lock)
1474 __releases(rq2->lock)
1475{
1476 BUG_ON(rq1 != rq2);
1477 raw_spin_unlock(&rq1->lock);
1478 __release(rq2->lock);
1479}
1480
1481#endif
1482
1483extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1484extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1485extern void print_cfs_stats(struct seq_file *m, int cpu);
1486extern void print_rt_stats(struct seq_file *m, int cpu);
1487
1488extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1489extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1490extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1491
1492extern void cfs_bandwidth_usage_inc(void);
1493extern void cfs_bandwidth_usage_dec(void);
1494
1495#ifdef CONFIG_NO_HZ_COMMON
1496enum rq_nohz_flag_bits {
1497 NOHZ_TICK_STOPPED,
1498 NOHZ_BALANCE_KICK,
1499};
1500
1501#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1502#endif
1503
1504#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1505
1506DECLARE_PER_CPU(u64, cpu_hardirq_time);
1507DECLARE_PER_CPU(u64, cpu_softirq_time);
1508
1509#ifndef CONFIG_64BIT
1510DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1511
1512static inline void irq_time_write_begin(void)
1513{
1514 __this_cpu_inc(irq_time_seq.sequence);
1515 smp_wmb();
1516}
1517
1518static inline void irq_time_write_end(void)
1519{
1520 smp_wmb();
1521 __this_cpu_inc(irq_time_seq.sequence);
1522}
1523
1524static inline u64 irq_time_read(int cpu)
1525{
1526 u64 irq_time;
1527 unsigned seq;
1528
1529 do {
1530 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1531 irq_time = per_cpu(cpu_softirq_time, cpu) +
1532 per_cpu(cpu_hardirq_time, cpu);
1533 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1534
1535 return irq_time;
1536}
1537#else /* CONFIG_64BIT */
1538static inline void irq_time_write_begin(void)
1539{
1540}
1541
1542static inline void irq_time_write_end(void)
1543{
1544}
1545
1546static inline u64 irq_time_read(int cpu)
1547{
1548 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1549}
1550#endif /* CONFIG_64BIT */
1551#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1/* 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
72#include <trace/events/power.h>
73#include <trace/events/sched.h>
74
75#include "../workqueue_internal.h"
76
77#ifdef CONFIG_PARAVIRT
78# include <asm/paravirt.h>
79# include <asm/paravirt_api_clock.h>
80#endif
81
82#include <asm/barrier.h>
83
84#include "cpupri.h"
85#include "cpudeadline.h"
86
87#ifdef CONFIG_SCHED_DEBUG
88# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
89#else
90# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
91#endif
92
93struct rq;
94struct cpuidle_state;
95
96/* task_struct::on_rq states: */
97#define TASK_ON_RQ_QUEUED 1
98#define TASK_ON_RQ_MIGRATING 2
99
100extern __read_mostly int scheduler_running;
101
102extern unsigned long calc_load_update;
103extern atomic_long_t calc_load_tasks;
104
105extern void calc_global_load_tick(struct rq *this_rq);
106extern long calc_load_fold_active(struct rq *this_rq, long adjust);
107
108extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
109
110extern int sysctl_sched_rt_period;
111extern int sysctl_sched_rt_runtime;
112extern int sched_rr_timeslice;
113
114/*
115 * Helpers for converting nanosecond timing to jiffy resolution
116 */
117#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
118
119/*
120 * Increase resolution of nice-level calculations for 64-bit architectures.
121 * The extra resolution improves shares distribution and load balancing of
122 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
123 * hierarchies, especially on larger systems. This is not a user-visible change
124 * and does not change the user-interface for setting shares/weights.
125 *
126 * We increase resolution only if we have enough bits to allow this increased
127 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
128 * are pretty high and the returns do not justify the increased costs.
129 *
130 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
131 * increase coverage and consistency always enable it on 64-bit platforms.
132 */
133#ifdef CONFIG_64BIT
134# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
135# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
136# define scale_load_down(w) \
137({ \
138 unsigned long __w = (w); \
139 if (__w) \
140 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
141 __w; \
142})
143#else
144# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
145# define scale_load(w) (w)
146# define scale_load_down(w) (w)
147#endif
148
149/*
150 * Task weight (visible to users) and its load (invisible to users) have
151 * independent resolution, but they should be well calibrated. We use
152 * scale_load() and scale_load_down(w) to convert between them. The
153 * following must be true:
154 *
155 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
156 *
157 */
158#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
159
160/*
161 * Single value that decides SCHED_DEADLINE internal math precision.
162 * 10 -> just above 1us
163 * 9 -> just above 0.5us
164 */
165#define DL_SCALE 10
166
167/*
168 * Single value that denotes runtime == period, ie unlimited time.
169 */
170#define RUNTIME_INF ((u64)~0ULL)
171
172static inline int idle_policy(int policy)
173{
174 return policy == SCHED_IDLE;
175}
176static inline int fair_policy(int policy)
177{
178 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
179}
180
181static inline int rt_policy(int policy)
182{
183 return policy == SCHED_FIFO || policy == SCHED_RR;
184}
185
186static inline int dl_policy(int policy)
187{
188 return policy == SCHED_DEADLINE;
189}
190static inline bool valid_policy(int policy)
191{
192 return idle_policy(policy) || fair_policy(policy) ||
193 rt_policy(policy) || dl_policy(policy);
194}
195
196static inline int task_has_idle_policy(struct task_struct *p)
197{
198 return idle_policy(p->policy);
199}
200
201static inline int task_has_rt_policy(struct task_struct *p)
202{
203 return rt_policy(p->policy);
204}
205
206static inline int task_has_dl_policy(struct task_struct *p)
207{
208 return dl_policy(p->policy);
209}
210
211#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
212
213static inline void update_avg(u64 *avg, u64 sample)
214{
215 s64 diff = sample - *avg;
216 *avg += diff / 8;
217}
218
219/*
220 * Shifting a value by an exponent greater *or equal* to the size of said value
221 * is UB; cap at size-1.
222 */
223#define shr_bound(val, shift) \
224 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
225
226/*
227 * !! For sched_setattr_nocheck() (kernel) only !!
228 *
229 * This is actually gross. :(
230 *
231 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
232 * tasks, but still be able to sleep. We need this on platforms that cannot
233 * atomically change clock frequency. Remove once fast switching will be
234 * available on such platforms.
235 *
236 * SUGOV stands for SchedUtil GOVernor.
237 */
238#define SCHED_FLAG_SUGOV 0x10000000
239
240#define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
241
242static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
243{
244#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
245 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
246#else
247 return false;
248#endif
249}
250
251/*
252 * Tells if entity @a should preempt entity @b.
253 */
254static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
255 const struct sched_dl_entity *b)
256{
257 return dl_entity_is_special(a) ||
258 dl_time_before(a->deadline, b->deadline);
259}
260
261/*
262 * This is the priority-queue data structure of the RT scheduling class:
263 */
264struct rt_prio_array {
265 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
266 struct list_head queue[MAX_RT_PRIO];
267};
268
269struct rt_bandwidth {
270 /* nests inside the rq lock: */
271 raw_spinlock_t rt_runtime_lock;
272 ktime_t rt_period;
273 u64 rt_runtime;
274 struct hrtimer rt_period_timer;
275 unsigned int rt_period_active;
276};
277
278static inline int dl_bandwidth_enabled(void)
279{
280 return sysctl_sched_rt_runtime >= 0;
281}
282
283/*
284 * To keep the bandwidth of -deadline tasks under control
285 * we need some place where:
286 * - store the maximum -deadline bandwidth of each cpu;
287 * - cache the fraction of bandwidth that is currently allocated in
288 * each root domain;
289 *
290 * This is all done in the data structure below. It is similar to the
291 * one used for RT-throttling (rt_bandwidth), with the main difference
292 * that, since here we are only interested in admission control, we
293 * do not decrease any runtime while the group "executes", neither we
294 * need a timer to replenish it.
295 *
296 * With respect to SMP, bandwidth is given on a per root domain basis,
297 * meaning that:
298 * - bw (< 100%) is the deadline bandwidth of each CPU;
299 * - total_bw is the currently allocated bandwidth in each root domain;
300 */
301struct dl_bw {
302 raw_spinlock_t lock;
303 u64 bw;
304 u64 total_bw;
305};
306
307extern void init_dl_bw(struct dl_bw *dl_b);
308extern int sched_dl_global_validate(void);
309extern void sched_dl_do_global(void);
310extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
311extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
312extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
313extern bool __checkparam_dl(const struct sched_attr *attr);
314extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
315extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
316extern int dl_bw_check_overflow(int cpu);
317
318/*
319 * SCHED_DEADLINE supports servers (nested scheduling) with the following
320 * interface:
321 *
322 * dl_se::rq -- runqueue we belong to.
323 *
324 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
325 * server when it runs out of tasks to run.
326 *
327 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
328 * returns NULL.
329 *
330 * dl_server_update() -- called from update_curr_common(), propagates runtime
331 * to the server.
332 *
333 * dl_server_start()
334 * dl_server_stop() -- start/stop the server when it has (no) tasks.
335 *
336 * dl_server_init() -- initializes the server.
337 */
338extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
339extern void dl_server_start(struct sched_dl_entity *dl_se);
340extern void dl_server_stop(struct sched_dl_entity *dl_se);
341extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
342 dl_server_has_tasks_f has_tasks,
343 dl_server_pick_f pick);
344
345#ifdef CONFIG_CGROUP_SCHED
346
347struct cfs_rq;
348struct rt_rq;
349
350extern struct list_head task_groups;
351
352struct cfs_bandwidth {
353#ifdef CONFIG_CFS_BANDWIDTH
354 raw_spinlock_t lock;
355 ktime_t period;
356 u64 quota;
357 u64 runtime;
358 u64 burst;
359 u64 runtime_snap;
360 s64 hierarchical_quota;
361
362 u8 idle;
363 u8 period_active;
364 u8 slack_started;
365 struct hrtimer period_timer;
366 struct hrtimer slack_timer;
367 struct list_head throttled_cfs_rq;
368
369 /* Statistics: */
370 int nr_periods;
371 int nr_throttled;
372 int nr_burst;
373 u64 throttled_time;
374 u64 burst_time;
375#endif
376};
377
378/* Task group related information */
379struct task_group {
380 struct cgroup_subsys_state css;
381
382#ifdef CONFIG_FAIR_GROUP_SCHED
383 /* schedulable entities of this group on each CPU */
384 struct sched_entity **se;
385 /* runqueue "owned" by this group on each CPU */
386 struct cfs_rq **cfs_rq;
387 unsigned long shares;
388
389 /* A positive value indicates that this is a SCHED_IDLE group. */
390 int idle;
391
392#ifdef CONFIG_SMP
393 /*
394 * load_avg can be heavily contended at clock tick time, so put
395 * it in its own cacheline separated from the fields above which
396 * will also be accessed at each tick.
397 */
398 atomic_long_t load_avg ____cacheline_aligned;
399#endif
400#endif
401
402#ifdef CONFIG_RT_GROUP_SCHED
403 struct sched_rt_entity **rt_se;
404 struct rt_rq **rt_rq;
405
406 struct rt_bandwidth rt_bandwidth;
407#endif
408
409 struct rcu_head rcu;
410 struct list_head list;
411
412 struct task_group *parent;
413 struct list_head siblings;
414 struct list_head children;
415
416#ifdef CONFIG_SCHED_AUTOGROUP
417 struct autogroup *autogroup;
418#endif
419
420 struct cfs_bandwidth cfs_bandwidth;
421
422#ifdef CONFIG_UCLAMP_TASK_GROUP
423 /* The two decimal precision [%] value requested from user-space */
424 unsigned int uclamp_pct[UCLAMP_CNT];
425 /* Clamp values requested for a task group */
426 struct uclamp_se uclamp_req[UCLAMP_CNT];
427 /* Effective clamp values used for a task group */
428 struct uclamp_se uclamp[UCLAMP_CNT];
429#endif
430
431};
432
433#ifdef CONFIG_FAIR_GROUP_SCHED
434#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
435
436/*
437 * A weight of 0 or 1 can cause arithmetics problems.
438 * A weight of a cfs_rq is the sum of weights of which entities
439 * are queued on this cfs_rq, so a weight of a entity should not be
440 * too large, so as the shares value of a task group.
441 * (The default weight is 1024 - so there's no practical
442 * limitation from this.)
443 */
444#define MIN_SHARES (1UL << 1)
445#define MAX_SHARES (1UL << 18)
446#endif
447
448typedef int (*tg_visitor)(struct task_group *, void *);
449
450extern int walk_tg_tree_from(struct task_group *from,
451 tg_visitor down, tg_visitor up, void *data);
452
453/*
454 * Iterate the full tree, calling @down when first entering a node and @up when
455 * leaving it for the final time.
456 *
457 * Caller must hold rcu_lock or sufficient equivalent.
458 */
459static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
460{
461 return walk_tg_tree_from(&root_task_group, down, up, data);
462}
463
464extern int tg_nop(struct task_group *tg, void *data);
465
466#ifdef CONFIG_FAIR_GROUP_SCHED
467extern void free_fair_sched_group(struct task_group *tg);
468extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
469extern void online_fair_sched_group(struct task_group *tg);
470extern void unregister_fair_sched_group(struct task_group *tg);
471#else
472static inline void free_fair_sched_group(struct task_group *tg) { }
473static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
474{
475 return 1;
476}
477static inline void online_fair_sched_group(struct task_group *tg) { }
478static inline void unregister_fair_sched_group(struct task_group *tg) { }
479#endif
480
481extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
482 struct sched_entity *se, int cpu,
483 struct sched_entity *parent);
484extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
485
486extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
487extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
488extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
489extern bool cfs_task_bw_constrained(struct task_struct *p);
490
491extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
492 struct sched_rt_entity *rt_se, int cpu,
493 struct sched_rt_entity *parent);
494extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
495extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
496extern long sched_group_rt_runtime(struct task_group *tg);
497extern long sched_group_rt_period(struct task_group *tg);
498extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
499
500extern struct task_group *sched_create_group(struct task_group *parent);
501extern void sched_online_group(struct task_group *tg,
502 struct task_group *parent);
503extern void sched_destroy_group(struct task_group *tg);
504extern void sched_release_group(struct task_group *tg);
505
506extern void sched_move_task(struct task_struct *tsk);
507
508#ifdef CONFIG_FAIR_GROUP_SCHED
509extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
510
511extern int sched_group_set_idle(struct task_group *tg, long idle);
512
513#ifdef CONFIG_SMP
514extern void set_task_rq_fair(struct sched_entity *se,
515 struct cfs_rq *prev, struct cfs_rq *next);
516#else /* !CONFIG_SMP */
517static inline void set_task_rq_fair(struct sched_entity *se,
518 struct cfs_rq *prev, struct cfs_rq *next) { }
519#endif /* CONFIG_SMP */
520#endif /* CONFIG_FAIR_GROUP_SCHED */
521
522#else /* CONFIG_CGROUP_SCHED */
523
524struct cfs_bandwidth { };
525static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
526
527#endif /* CONFIG_CGROUP_SCHED */
528
529extern void unregister_rt_sched_group(struct task_group *tg);
530extern void free_rt_sched_group(struct task_group *tg);
531extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
532
533/*
534 * u64_u32_load/u64_u32_store
535 *
536 * Use a copy of a u64 value to protect against data race. This is only
537 * applicable for 32-bits architectures.
538 */
539#ifdef CONFIG_64BIT
540# define u64_u32_load_copy(var, copy) var
541# define u64_u32_store_copy(var, copy, val) (var = val)
542#else
543# define u64_u32_load_copy(var, copy) \
544({ \
545 u64 __val, __val_copy; \
546 do { \
547 __val_copy = copy; \
548 /* \
549 * paired with u64_u32_store_copy(), ordering access \
550 * to var and copy. \
551 */ \
552 smp_rmb(); \
553 __val = var; \
554 } while (__val != __val_copy); \
555 __val; \
556})
557# define u64_u32_store_copy(var, copy, val) \
558do { \
559 typeof(val) __val = (val); \
560 var = __val; \
561 /* \
562 * paired with u64_u32_load_copy(), ordering access to var and \
563 * copy. \
564 */ \
565 smp_wmb(); \
566 copy = __val; \
567} while (0)
568#endif
569# define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
570# define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
571
572/* CFS-related fields in a runqueue */
573struct cfs_rq {
574 struct load_weight load;
575 unsigned int nr_running;
576 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
577 unsigned int idle_nr_running; /* SCHED_IDLE */
578 unsigned int idle_h_nr_running; /* SCHED_IDLE */
579
580 s64 avg_vruntime;
581 u64 avg_load;
582
583 u64 exec_clock;
584 u64 min_vruntime;
585#ifdef CONFIG_SCHED_CORE
586 unsigned int forceidle_seq;
587 u64 min_vruntime_fi;
588#endif
589
590#ifndef CONFIG_64BIT
591 u64 min_vruntime_copy;
592#endif
593
594 struct rb_root_cached tasks_timeline;
595
596 /*
597 * 'curr' points to currently running entity on this cfs_rq.
598 * It is set to NULL otherwise (i.e when none are currently running).
599 */
600 struct sched_entity *curr;
601 struct sched_entity *next;
602
603#ifdef CONFIG_SCHED_DEBUG
604 unsigned int nr_spread_over;
605#endif
606
607#ifdef CONFIG_SMP
608 /*
609 * CFS load tracking
610 */
611 struct sched_avg avg;
612#ifndef CONFIG_64BIT
613 u64 last_update_time_copy;
614#endif
615 struct {
616 raw_spinlock_t lock ____cacheline_aligned;
617 int nr;
618 unsigned long load_avg;
619 unsigned long util_avg;
620 unsigned long runnable_avg;
621 } removed;
622
623#ifdef CONFIG_FAIR_GROUP_SCHED
624 u64 last_update_tg_load_avg;
625 unsigned long tg_load_avg_contrib;
626 long propagate;
627 long prop_runnable_sum;
628
629 /*
630 * h_load = weight * f(tg)
631 *
632 * Where f(tg) is the recursive weight fraction assigned to
633 * this group.
634 */
635 unsigned long h_load;
636 u64 last_h_load_update;
637 struct sched_entity *h_load_next;
638#endif /* CONFIG_FAIR_GROUP_SCHED */
639#endif /* CONFIG_SMP */
640
641#ifdef CONFIG_FAIR_GROUP_SCHED
642 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
643
644 /*
645 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
646 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
647 * (like users, containers etc.)
648 *
649 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
650 * This list is used during load balance.
651 */
652 int on_list;
653 struct list_head leaf_cfs_rq_list;
654 struct task_group *tg; /* group that "owns" this runqueue */
655
656 /* Locally cached copy of our task_group's idle value */
657 int idle;
658
659#ifdef CONFIG_CFS_BANDWIDTH
660 int runtime_enabled;
661 s64 runtime_remaining;
662
663 u64 throttled_pelt_idle;
664#ifndef CONFIG_64BIT
665 u64 throttled_pelt_idle_copy;
666#endif
667 u64 throttled_clock;
668 u64 throttled_clock_pelt;
669 u64 throttled_clock_pelt_time;
670 u64 throttled_clock_self;
671 u64 throttled_clock_self_time;
672 int throttled;
673 int throttle_count;
674 struct list_head throttled_list;
675 struct list_head throttled_csd_list;
676#endif /* CONFIG_CFS_BANDWIDTH */
677#endif /* CONFIG_FAIR_GROUP_SCHED */
678};
679
680static inline int rt_bandwidth_enabled(void)
681{
682 return sysctl_sched_rt_runtime >= 0;
683}
684
685/* RT IPI pull logic requires IRQ_WORK */
686#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
687# define HAVE_RT_PUSH_IPI
688#endif
689
690/* Real-Time classes' related field in a runqueue: */
691struct rt_rq {
692 struct rt_prio_array active;
693 unsigned int rt_nr_running;
694 unsigned int rr_nr_running;
695#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
696 struct {
697 int curr; /* highest queued rt task prio */
698#ifdef CONFIG_SMP
699 int next; /* next highest */
700#endif
701 } highest_prio;
702#endif
703#ifdef CONFIG_SMP
704 int overloaded;
705 struct plist_head pushable_tasks;
706
707#endif /* CONFIG_SMP */
708 int rt_queued;
709
710 int rt_throttled;
711 u64 rt_time;
712 u64 rt_runtime;
713 /* Nests inside the rq lock: */
714 raw_spinlock_t rt_runtime_lock;
715
716#ifdef CONFIG_RT_GROUP_SCHED
717 unsigned int rt_nr_boosted;
718
719 struct rq *rq;
720 struct task_group *tg;
721#endif
722};
723
724static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
725{
726 return rt_rq->rt_queued && rt_rq->rt_nr_running;
727}
728
729/* Deadline class' related fields in a runqueue */
730struct dl_rq {
731 /* runqueue is an rbtree, ordered by deadline */
732 struct rb_root_cached root;
733
734 unsigned int dl_nr_running;
735
736#ifdef CONFIG_SMP
737 /*
738 * Deadline values of the currently executing and the
739 * earliest ready task on this rq. Caching these facilitates
740 * the decision whether or not a ready but not running task
741 * should migrate somewhere else.
742 */
743 struct {
744 u64 curr;
745 u64 next;
746 } earliest_dl;
747
748 int overloaded;
749
750 /*
751 * Tasks on this rq that can be pushed away. They are kept in
752 * an rb-tree, ordered by tasks' deadlines, with caching
753 * of the leftmost (earliest deadline) element.
754 */
755 struct rb_root_cached pushable_dl_tasks_root;
756#else
757 struct dl_bw dl_bw;
758#endif
759 /*
760 * "Active utilization" for this runqueue: increased when a
761 * task wakes up (becomes TASK_RUNNING) and decreased when a
762 * task blocks
763 */
764 u64 running_bw;
765
766 /*
767 * Utilization of the tasks "assigned" to this runqueue (including
768 * the tasks that are in runqueue and the tasks that executed on this
769 * CPU and blocked). Increased when a task moves to this runqueue, and
770 * decreased when the task moves away (migrates, changes scheduling
771 * policy, or terminates).
772 * This is needed to compute the "inactive utilization" for the
773 * runqueue (inactive utilization = this_bw - running_bw).
774 */
775 u64 this_bw;
776 u64 extra_bw;
777
778 /*
779 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
780 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
781 */
782 u64 max_bw;
783
784 /*
785 * Inverse of the fraction of CPU utilization that can be reclaimed
786 * by the GRUB algorithm.
787 */
788 u64 bw_ratio;
789};
790
791#ifdef CONFIG_FAIR_GROUP_SCHED
792/* An entity is a task if it doesn't "own" a runqueue */
793#define entity_is_task(se) (!se->my_q)
794
795static inline void se_update_runnable(struct sched_entity *se)
796{
797 if (!entity_is_task(se))
798 se->runnable_weight = se->my_q->h_nr_running;
799}
800
801static inline long se_runnable(struct sched_entity *se)
802{
803 if (entity_is_task(se))
804 return !!se->on_rq;
805 else
806 return se->runnable_weight;
807}
808
809#else
810#define entity_is_task(se) 1
811
812static inline void se_update_runnable(struct sched_entity *se) {}
813
814static inline long se_runnable(struct sched_entity *se)
815{
816 return !!se->on_rq;
817}
818#endif
819
820#ifdef CONFIG_SMP
821/*
822 * XXX we want to get rid of these helpers and use the full load resolution.
823 */
824static inline long se_weight(struct sched_entity *se)
825{
826 return scale_load_down(se->load.weight);
827}
828
829
830static inline bool sched_asym_prefer(int a, int b)
831{
832 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
833}
834
835struct perf_domain {
836 struct em_perf_domain *em_pd;
837 struct perf_domain *next;
838 struct rcu_head rcu;
839};
840
841/* Scheduling group status flags */
842#define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
843#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
844
845/*
846 * We add the notion of a root-domain which will be used to define per-domain
847 * variables. Each exclusive cpuset essentially defines an island domain by
848 * fully partitioning the member CPUs from any other cpuset. Whenever a new
849 * exclusive cpuset is created, we also create and attach a new root-domain
850 * object.
851 *
852 */
853struct root_domain {
854 atomic_t refcount;
855 atomic_t rto_count;
856 struct rcu_head rcu;
857 cpumask_var_t span;
858 cpumask_var_t online;
859
860 /*
861 * Indicate pullable load on at least one CPU, e.g:
862 * - More than one runnable task
863 * - Running task is misfit
864 */
865 int overload;
866
867 /* Indicate one or more cpus over-utilized (tipping point) */
868 int overutilized;
869
870 /*
871 * The bit corresponding to a CPU gets set here if such CPU has more
872 * than one runnable -deadline task (as it is below for RT tasks).
873 */
874 cpumask_var_t dlo_mask;
875 atomic_t dlo_count;
876 struct dl_bw dl_bw;
877 struct cpudl cpudl;
878
879 /*
880 * Indicate whether a root_domain's dl_bw has been checked or
881 * updated. It's monotonously increasing value.
882 *
883 * Also, some corner cases, like 'wrap around' is dangerous, but given
884 * that u64 is 'big enough'. So that shouldn't be a concern.
885 */
886 u64 visit_gen;
887
888#ifdef HAVE_RT_PUSH_IPI
889 /*
890 * For IPI pull requests, loop across the rto_mask.
891 */
892 struct irq_work rto_push_work;
893 raw_spinlock_t rto_lock;
894 /* These are only updated and read within rto_lock */
895 int rto_loop;
896 int rto_cpu;
897 /* These atomics are updated outside of a lock */
898 atomic_t rto_loop_next;
899 atomic_t rto_loop_start;
900#endif
901 /*
902 * The "RT overload" flag: it gets set if a CPU has more than
903 * one runnable RT task.
904 */
905 cpumask_var_t rto_mask;
906 struct cpupri cpupri;
907
908 unsigned long max_cpu_capacity;
909
910 /*
911 * NULL-terminated list of performance domains intersecting with the
912 * CPUs of the rd. Protected by RCU.
913 */
914 struct perf_domain __rcu *pd;
915};
916
917extern void init_defrootdomain(void);
918extern int sched_init_domains(const struct cpumask *cpu_map);
919extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
920extern void sched_get_rd(struct root_domain *rd);
921extern void sched_put_rd(struct root_domain *rd);
922
923#ifdef HAVE_RT_PUSH_IPI
924extern void rto_push_irq_work_func(struct irq_work *work);
925#endif
926#endif /* CONFIG_SMP */
927
928#ifdef CONFIG_UCLAMP_TASK
929/*
930 * struct uclamp_bucket - Utilization clamp bucket
931 * @value: utilization clamp value for tasks on this clamp bucket
932 * @tasks: number of RUNNABLE tasks on this clamp bucket
933 *
934 * Keep track of how many tasks are RUNNABLE for a given utilization
935 * clamp value.
936 */
937struct uclamp_bucket {
938 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
939 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
940};
941
942/*
943 * struct uclamp_rq - rq's utilization clamp
944 * @value: currently active clamp values for a rq
945 * @bucket: utilization clamp buckets affecting a rq
946 *
947 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
948 * A clamp value is affecting a rq when there is at least one task RUNNABLE
949 * (or actually running) with that value.
950 *
951 * There are up to UCLAMP_CNT possible different clamp values, currently there
952 * are only two: minimum utilization and maximum utilization.
953 *
954 * All utilization clamping values are MAX aggregated, since:
955 * - for util_min: we want to run the CPU at least at the max of the minimum
956 * utilization required by its currently RUNNABLE tasks.
957 * - for util_max: we want to allow the CPU to run up to the max of the
958 * maximum utilization allowed by its currently RUNNABLE tasks.
959 *
960 * Since on each system we expect only a limited number of different
961 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
962 * the metrics required to compute all the per-rq utilization clamp values.
963 */
964struct uclamp_rq {
965 unsigned int value;
966 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
967};
968
969DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
970#endif /* CONFIG_UCLAMP_TASK */
971
972struct rq;
973struct balance_callback {
974 struct balance_callback *next;
975 void (*func)(struct rq *rq);
976};
977
978/*
979 * This is the main, per-CPU runqueue data structure.
980 *
981 * Locking rule: those places that want to lock multiple runqueues
982 * (such as the load balancing or the thread migration code), lock
983 * acquire operations must be ordered by ascending &runqueue.
984 */
985struct rq {
986 /* runqueue lock: */
987 raw_spinlock_t __lock;
988
989 unsigned int nr_running;
990#ifdef CONFIG_NUMA_BALANCING
991 unsigned int nr_numa_running;
992 unsigned int nr_preferred_running;
993 unsigned int numa_migrate_on;
994#endif
995#ifdef CONFIG_NO_HZ_COMMON
996#ifdef CONFIG_SMP
997 unsigned long last_blocked_load_update_tick;
998 unsigned int has_blocked_load;
999 call_single_data_t nohz_csd;
1000#endif /* CONFIG_SMP */
1001 unsigned int nohz_tick_stopped;
1002 atomic_t nohz_flags;
1003#endif /* CONFIG_NO_HZ_COMMON */
1004
1005#ifdef CONFIG_SMP
1006 unsigned int ttwu_pending;
1007#endif
1008 u64 nr_switches;
1009
1010#ifdef CONFIG_UCLAMP_TASK
1011 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1012 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1013 unsigned int uclamp_flags;
1014#define UCLAMP_FLAG_IDLE 0x01
1015#endif
1016
1017 struct cfs_rq cfs;
1018 struct rt_rq rt;
1019 struct dl_rq dl;
1020
1021#ifdef CONFIG_FAIR_GROUP_SCHED
1022 /* list of leaf cfs_rq on this CPU: */
1023 struct list_head leaf_cfs_rq_list;
1024 struct list_head *tmp_alone_branch;
1025#endif /* CONFIG_FAIR_GROUP_SCHED */
1026
1027 /*
1028 * This is part of a global counter where only the total sum
1029 * over all CPUs matters. A task can increase this counter on
1030 * one CPU and if it got migrated afterwards it may decrease
1031 * it on another CPU. Always updated under the runqueue lock:
1032 */
1033 unsigned int nr_uninterruptible;
1034
1035 struct task_struct __rcu *curr;
1036 struct task_struct *idle;
1037 struct task_struct *stop;
1038 unsigned long next_balance;
1039 struct mm_struct *prev_mm;
1040
1041 unsigned int clock_update_flags;
1042 u64 clock;
1043 /* Ensure that all clocks are in the same cache line */
1044 u64 clock_task ____cacheline_aligned;
1045 u64 clock_pelt;
1046 unsigned long lost_idle_time;
1047 u64 clock_pelt_idle;
1048 u64 clock_idle;
1049#ifndef CONFIG_64BIT
1050 u64 clock_pelt_idle_copy;
1051 u64 clock_idle_copy;
1052#endif
1053
1054 atomic_t nr_iowait;
1055
1056#ifdef CONFIG_SCHED_DEBUG
1057 u64 last_seen_need_resched_ns;
1058 int ticks_without_resched;
1059#endif
1060
1061#ifdef CONFIG_MEMBARRIER
1062 int membarrier_state;
1063#endif
1064
1065#ifdef CONFIG_SMP
1066 struct root_domain *rd;
1067 struct sched_domain __rcu *sd;
1068
1069 unsigned long cpu_capacity;
1070
1071 struct balance_callback *balance_callback;
1072
1073 unsigned char nohz_idle_balance;
1074 unsigned char idle_balance;
1075
1076 unsigned long misfit_task_load;
1077
1078 /* For active balancing */
1079 int active_balance;
1080 int push_cpu;
1081 struct cpu_stop_work active_balance_work;
1082
1083 /* CPU of this runqueue: */
1084 int cpu;
1085 int online;
1086
1087 struct list_head cfs_tasks;
1088
1089 struct sched_avg avg_rt;
1090 struct sched_avg avg_dl;
1091#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1092 struct sched_avg avg_irq;
1093#endif
1094#ifdef CONFIG_SCHED_THERMAL_PRESSURE
1095 struct sched_avg avg_thermal;
1096#endif
1097 u64 idle_stamp;
1098 u64 avg_idle;
1099
1100 /* This is used to determine avg_idle's max value */
1101 u64 max_idle_balance_cost;
1102
1103#ifdef CONFIG_HOTPLUG_CPU
1104 struct rcuwait hotplug_wait;
1105#endif
1106#endif /* CONFIG_SMP */
1107
1108#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1109 u64 prev_irq_time;
1110#endif
1111#ifdef CONFIG_PARAVIRT
1112 u64 prev_steal_time;
1113#endif
1114#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1115 u64 prev_steal_time_rq;
1116#endif
1117
1118 /* calc_load related fields */
1119 unsigned long calc_load_update;
1120 long calc_load_active;
1121
1122#ifdef CONFIG_SCHED_HRTICK
1123#ifdef CONFIG_SMP
1124 call_single_data_t hrtick_csd;
1125#endif
1126 struct hrtimer hrtick_timer;
1127 ktime_t hrtick_time;
1128#endif
1129
1130#ifdef CONFIG_SCHEDSTATS
1131 /* latency stats */
1132 struct sched_info rq_sched_info;
1133 unsigned long long rq_cpu_time;
1134 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1135
1136 /* sys_sched_yield() stats */
1137 unsigned int yld_count;
1138
1139 /* schedule() stats */
1140 unsigned int sched_count;
1141 unsigned int sched_goidle;
1142
1143 /* try_to_wake_up() stats */
1144 unsigned int ttwu_count;
1145 unsigned int ttwu_local;
1146#endif
1147
1148#ifdef CONFIG_CPU_IDLE
1149 /* Must be inspected within a rcu lock section */
1150 struct cpuidle_state *idle_state;
1151#endif
1152
1153#ifdef CONFIG_SMP
1154 unsigned int nr_pinned;
1155#endif
1156 unsigned int push_busy;
1157 struct cpu_stop_work push_work;
1158
1159#ifdef CONFIG_SCHED_CORE
1160 /* per rq */
1161 struct rq *core;
1162 struct task_struct *core_pick;
1163 unsigned int core_enabled;
1164 unsigned int core_sched_seq;
1165 struct rb_root core_tree;
1166
1167 /* shared state -- careful with sched_core_cpu_deactivate() */
1168 unsigned int core_task_seq;
1169 unsigned int core_pick_seq;
1170 unsigned long core_cookie;
1171 unsigned int core_forceidle_count;
1172 unsigned int core_forceidle_seq;
1173 unsigned int core_forceidle_occupation;
1174 u64 core_forceidle_start;
1175#endif
1176
1177 /* Scratch cpumask to be temporarily used under rq_lock */
1178 cpumask_var_t scratch_mask;
1179
1180#if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1181 call_single_data_t cfsb_csd;
1182 struct list_head cfsb_csd_list;
1183#endif
1184};
1185
1186#ifdef CONFIG_FAIR_GROUP_SCHED
1187
1188/* CPU runqueue to which this cfs_rq is attached */
1189static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1190{
1191 return cfs_rq->rq;
1192}
1193
1194#else
1195
1196static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1197{
1198 return container_of(cfs_rq, struct rq, cfs);
1199}
1200#endif
1201
1202static inline int cpu_of(struct rq *rq)
1203{
1204#ifdef CONFIG_SMP
1205 return rq->cpu;
1206#else
1207 return 0;
1208#endif
1209}
1210
1211#define MDF_PUSH 0x01
1212
1213static inline bool is_migration_disabled(struct task_struct *p)
1214{
1215#ifdef CONFIG_SMP
1216 return p->migration_disabled;
1217#else
1218 return false;
1219#endif
1220}
1221
1222DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1223
1224#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1225#define this_rq() this_cpu_ptr(&runqueues)
1226#define task_rq(p) cpu_rq(task_cpu(p))
1227#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1228#define raw_rq() raw_cpu_ptr(&runqueues)
1229
1230struct sched_group;
1231#ifdef CONFIG_SCHED_CORE
1232static inline struct cpumask *sched_group_span(struct sched_group *sg);
1233
1234DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1235
1236static inline bool sched_core_enabled(struct rq *rq)
1237{
1238 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1239}
1240
1241static inline bool sched_core_disabled(void)
1242{
1243 return !static_branch_unlikely(&__sched_core_enabled);
1244}
1245
1246/*
1247 * Be careful with this function; not for general use. The return value isn't
1248 * stable unless you actually hold a relevant rq->__lock.
1249 */
1250static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1251{
1252 if (sched_core_enabled(rq))
1253 return &rq->core->__lock;
1254
1255 return &rq->__lock;
1256}
1257
1258static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1259{
1260 if (rq->core_enabled)
1261 return &rq->core->__lock;
1262
1263 return &rq->__lock;
1264}
1265
1266bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1267 bool fi);
1268void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1269
1270/*
1271 * Helpers to check if the CPU's core cookie matches with the task's cookie
1272 * when core scheduling is enabled.
1273 * A special case is that the task's cookie always matches with CPU's core
1274 * cookie if the CPU is in an idle core.
1275 */
1276static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1277{
1278 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1279 if (!sched_core_enabled(rq))
1280 return true;
1281
1282 return rq->core->core_cookie == p->core_cookie;
1283}
1284
1285static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1286{
1287 bool idle_core = true;
1288 int cpu;
1289
1290 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1291 if (!sched_core_enabled(rq))
1292 return true;
1293
1294 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1295 if (!available_idle_cpu(cpu)) {
1296 idle_core = false;
1297 break;
1298 }
1299 }
1300
1301 /*
1302 * A CPU in an idle core is always the best choice for tasks with
1303 * cookies.
1304 */
1305 return idle_core || rq->core->core_cookie == p->core_cookie;
1306}
1307
1308static inline bool sched_group_cookie_match(struct rq *rq,
1309 struct task_struct *p,
1310 struct sched_group *group)
1311{
1312 int cpu;
1313
1314 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1315 if (!sched_core_enabled(rq))
1316 return true;
1317
1318 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1319 if (sched_core_cookie_match(cpu_rq(cpu), p))
1320 return true;
1321 }
1322 return false;
1323}
1324
1325static inline bool sched_core_enqueued(struct task_struct *p)
1326{
1327 return !RB_EMPTY_NODE(&p->core_node);
1328}
1329
1330extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1331extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1332
1333extern void sched_core_get(void);
1334extern void sched_core_put(void);
1335
1336#else /* !CONFIG_SCHED_CORE */
1337
1338static inline bool sched_core_enabled(struct rq *rq)
1339{
1340 return false;
1341}
1342
1343static inline bool sched_core_disabled(void)
1344{
1345 return true;
1346}
1347
1348static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1349{
1350 return &rq->__lock;
1351}
1352
1353static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1354{
1355 return &rq->__lock;
1356}
1357
1358static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1359{
1360 return true;
1361}
1362
1363static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1364{
1365 return true;
1366}
1367
1368static inline bool sched_group_cookie_match(struct rq *rq,
1369 struct task_struct *p,
1370 struct sched_group *group)
1371{
1372 return true;
1373}
1374#endif /* CONFIG_SCHED_CORE */
1375
1376static inline void lockdep_assert_rq_held(struct rq *rq)
1377{
1378 lockdep_assert_held(__rq_lockp(rq));
1379}
1380
1381extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1382extern bool raw_spin_rq_trylock(struct rq *rq);
1383extern void raw_spin_rq_unlock(struct rq *rq);
1384
1385static inline void raw_spin_rq_lock(struct rq *rq)
1386{
1387 raw_spin_rq_lock_nested(rq, 0);
1388}
1389
1390static inline void raw_spin_rq_lock_irq(struct rq *rq)
1391{
1392 local_irq_disable();
1393 raw_spin_rq_lock(rq);
1394}
1395
1396static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1397{
1398 raw_spin_rq_unlock(rq);
1399 local_irq_enable();
1400}
1401
1402static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1403{
1404 unsigned long flags;
1405 local_irq_save(flags);
1406 raw_spin_rq_lock(rq);
1407 return flags;
1408}
1409
1410static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1411{
1412 raw_spin_rq_unlock(rq);
1413 local_irq_restore(flags);
1414}
1415
1416#define raw_spin_rq_lock_irqsave(rq, flags) \
1417do { \
1418 flags = _raw_spin_rq_lock_irqsave(rq); \
1419} while (0)
1420
1421#ifdef CONFIG_SCHED_SMT
1422extern void __update_idle_core(struct rq *rq);
1423
1424static inline void update_idle_core(struct rq *rq)
1425{
1426 if (static_branch_unlikely(&sched_smt_present))
1427 __update_idle_core(rq);
1428}
1429
1430#else
1431static inline void update_idle_core(struct rq *rq) { }
1432#endif
1433
1434#ifdef CONFIG_FAIR_GROUP_SCHED
1435static inline struct task_struct *task_of(struct sched_entity *se)
1436{
1437 SCHED_WARN_ON(!entity_is_task(se));
1438 return container_of(se, struct task_struct, se);
1439}
1440
1441static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1442{
1443 return p->se.cfs_rq;
1444}
1445
1446/* runqueue on which this entity is (to be) queued */
1447static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1448{
1449 return se->cfs_rq;
1450}
1451
1452/* runqueue "owned" by this group */
1453static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1454{
1455 return grp->my_q;
1456}
1457
1458#else
1459
1460#define task_of(_se) container_of(_se, struct task_struct, se)
1461
1462static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1463{
1464 return &task_rq(p)->cfs;
1465}
1466
1467static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1468{
1469 const struct task_struct *p = task_of(se);
1470 struct rq *rq = task_rq(p);
1471
1472 return &rq->cfs;
1473}
1474
1475/* runqueue "owned" by this group */
1476static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1477{
1478 return NULL;
1479}
1480#endif
1481
1482extern void update_rq_clock(struct rq *rq);
1483
1484/*
1485 * rq::clock_update_flags bits
1486 *
1487 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1488 * call to __schedule(). This is an optimisation to avoid
1489 * neighbouring rq clock updates.
1490 *
1491 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1492 * in effect and calls to update_rq_clock() are being ignored.
1493 *
1494 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1495 * made to update_rq_clock() since the last time rq::lock was pinned.
1496 *
1497 * If inside of __schedule(), clock_update_flags will have been
1498 * shifted left (a left shift is a cheap operation for the fast path
1499 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1500 *
1501 * if (rq-clock_update_flags >= RQCF_UPDATED)
1502 *
1503 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1504 * one position though, because the next rq_unpin_lock() will shift it
1505 * back.
1506 */
1507#define RQCF_REQ_SKIP 0x01
1508#define RQCF_ACT_SKIP 0x02
1509#define RQCF_UPDATED 0x04
1510
1511static inline void assert_clock_updated(struct rq *rq)
1512{
1513 /*
1514 * The only reason for not seeing a clock update since the
1515 * last rq_pin_lock() is if we're currently skipping updates.
1516 */
1517 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1518}
1519
1520static inline u64 rq_clock(struct rq *rq)
1521{
1522 lockdep_assert_rq_held(rq);
1523 assert_clock_updated(rq);
1524
1525 return rq->clock;
1526}
1527
1528static inline u64 rq_clock_task(struct rq *rq)
1529{
1530 lockdep_assert_rq_held(rq);
1531 assert_clock_updated(rq);
1532
1533 return rq->clock_task;
1534}
1535
1536/**
1537 * By default the decay is the default pelt decay period.
1538 * The decay shift can change the decay period in
1539 * multiples of 32.
1540 * Decay shift Decay period(ms)
1541 * 0 32
1542 * 1 64
1543 * 2 128
1544 * 3 256
1545 * 4 512
1546 */
1547extern int sched_thermal_decay_shift;
1548
1549static inline u64 rq_clock_thermal(struct rq *rq)
1550{
1551 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1552}
1553
1554static inline void rq_clock_skip_update(struct rq *rq)
1555{
1556 lockdep_assert_rq_held(rq);
1557 rq->clock_update_flags |= RQCF_REQ_SKIP;
1558}
1559
1560/*
1561 * See rt task throttling, which is the only time a skip
1562 * request is canceled.
1563 */
1564static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1565{
1566 lockdep_assert_rq_held(rq);
1567 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1568}
1569
1570/*
1571 * During cpu offlining and rq wide unthrottling, we can trigger
1572 * an update_rq_clock() for several cfs and rt runqueues (Typically
1573 * when using list_for_each_entry_*)
1574 * rq_clock_start_loop_update() can be called after updating the clock
1575 * once and before iterating over the list to prevent multiple update.
1576 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1577 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1578 */
1579static inline void rq_clock_start_loop_update(struct rq *rq)
1580{
1581 lockdep_assert_rq_held(rq);
1582 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1583 rq->clock_update_flags |= RQCF_ACT_SKIP;
1584}
1585
1586static inline void rq_clock_stop_loop_update(struct rq *rq)
1587{
1588 lockdep_assert_rq_held(rq);
1589 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1590}
1591
1592struct rq_flags {
1593 unsigned long flags;
1594 struct pin_cookie cookie;
1595#ifdef CONFIG_SCHED_DEBUG
1596 /*
1597 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1598 * current pin context is stashed here in case it needs to be
1599 * restored in rq_repin_lock().
1600 */
1601 unsigned int clock_update_flags;
1602#endif
1603};
1604
1605extern struct balance_callback balance_push_callback;
1606
1607/*
1608 * Lockdep annotation that avoids accidental unlocks; it's like a
1609 * sticky/continuous lockdep_assert_held().
1610 *
1611 * This avoids code that has access to 'struct rq *rq' (basically everything in
1612 * the scheduler) from accidentally unlocking the rq if they do not also have a
1613 * copy of the (on-stack) 'struct rq_flags rf'.
1614 *
1615 * Also see Documentation/locking/lockdep-design.rst.
1616 */
1617static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1618{
1619 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1620
1621#ifdef CONFIG_SCHED_DEBUG
1622 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1623 rf->clock_update_flags = 0;
1624#ifdef CONFIG_SMP
1625 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1626#endif
1627#endif
1628}
1629
1630static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1631{
1632#ifdef CONFIG_SCHED_DEBUG
1633 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1634 rf->clock_update_flags = RQCF_UPDATED;
1635#endif
1636
1637 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1638}
1639
1640static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1641{
1642 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1643
1644#ifdef CONFIG_SCHED_DEBUG
1645 /*
1646 * Restore the value we stashed in @rf for this pin context.
1647 */
1648 rq->clock_update_flags |= rf->clock_update_flags;
1649#endif
1650}
1651
1652struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1653 __acquires(rq->lock);
1654
1655struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1656 __acquires(p->pi_lock)
1657 __acquires(rq->lock);
1658
1659static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1660 __releases(rq->lock)
1661{
1662 rq_unpin_lock(rq, rf);
1663 raw_spin_rq_unlock(rq);
1664}
1665
1666static inline void
1667task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1668 __releases(rq->lock)
1669 __releases(p->pi_lock)
1670{
1671 rq_unpin_lock(rq, rf);
1672 raw_spin_rq_unlock(rq);
1673 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1674}
1675
1676DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1677 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1678 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1679 struct rq *rq; struct rq_flags rf)
1680
1681static inline void
1682rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1683 __acquires(rq->lock)
1684{
1685 raw_spin_rq_lock_irqsave(rq, rf->flags);
1686 rq_pin_lock(rq, rf);
1687}
1688
1689static inline void
1690rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1691 __acquires(rq->lock)
1692{
1693 raw_spin_rq_lock_irq(rq);
1694 rq_pin_lock(rq, rf);
1695}
1696
1697static inline void
1698rq_lock(struct rq *rq, struct rq_flags *rf)
1699 __acquires(rq->lock)
1700{
1701 raw_spin_rq_lock(rq);
1702 rq_pin_lock(rq, rf);
1703}
1704
1705static inline void
1706rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1707 __releases(rq->lock)
1708{
1709 rq_unpin_lock(rq, rf);
1710 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1711}
1712
1713static inline void
1714rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1715 __releases(rq->lock)
1716{
1717 rq_unpin_lock(rq, rf);
1718 raw_spin_rq_unlock_irq(rq);
1719}
1720
1721static inline void
1722rq_unlock(struct rq *rq, struct rq_flags *rf)
1723 __releases(rq->lock)
1724{
1725 rq_unpin_lock(rq, rf);
1726 raw_spin_rq_unlock(rq);
1727}
1728
1729DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1730 rq_lock(_T->lock, &_T->rf),
1731 rq_unlock(_T->lock, &_T->rf),
1732 struct rq_flags rf)
1733
1734DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1735 rq_lock_irq(_T->lock, &_T->rf),
1736 rq_unlock_irq(_T->lock, &_T->rf),
1737 struct rq_flags rf)
1738
1739DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1740 rq_lock_irqsave(_T->lock, &_T->rf),
1741 rq_unlock_irqrestore(_T->lock, &_T->rf),
1742 struct rq_flags rf)
1743
1744static inline struct rq *
1745this_rq_lock_irq(struct rq_flags *rf)
1746 __acquires(rq->lock)
1747{
1748 struct rq *rq;
1749
1750 local_irq_disable();
1751 rq = this_rq();
1752 rq_lock(rq, rf);
1753 return rq;
1754}
1755
1756#ifdef CONFIG_NUMA
1757enum numa_topology_type {
1758 NUMA_DIRECT,
1759 NUMA_GLUELESS_MESH,
1760 NUMA_BACKPLANE,
1761};
1762extern enum numa_topology_type sched_numa_topology_type;
1763extern int sched_max_numa_distance;
1764extern bool find_numa_distance(int distance);
1765extern void sched_init_numa(int offline_node);
1766extern void sched_update_numa(int cpu, bool online);
1767extern void sched_domains_numa_masks_set(unsigned int cpu);
1768extern void sched_domains_numa_masks_clear(unsigned int cpu);
1769extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1770#else
1771static inline void sched_init_numa(int offline_node) { }
1772static inline void sched_update_numa(int cpu, bool online) { }
1773static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1774static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1775static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1776{
1777 return nr_cpu_ids;
1778}
1779#endif
1780
1781#ifdef CONFIG_NUMA_BALANCING
1782/* The regions in numa_faults array from task_struct */
1783enum numa_faults_stats {
1784 NUMA_MEM = 0,
1785 NUMA_CPU,
1786 NUMA_MEMBUF,
1787 NUMA_CPUBUF
1788};
1789extern void sched_setnuma(struct task_struct *p, int node);
1790extern int migrate_task_to(struct task_struct *p, int cpu);
1791extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1792 int cpu, int scpu);
1793extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1794#else
1795static inline void
1796init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1797{
1798}
1799#endif /* CONFIG_NUMA_BALANCING */
1800
1801#ifdef CONFIG_SMP
1802
1803static inline void
1804queue_balance_callback(struct rq *rq,
1805 struct balance_callback *head,
1806 void (*func)(struct rq *rq))
1807{
1808 lockdep_assert_rq_held(rq);
1809
1810 /*
1811 * Don't (re)queue an already queued item; nor queue anything when
1812 * balance_push() is active, see the comment with
1813 * balance_push_callback.
1814 */
1815 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1816 return;
1817
1818 head->func = func;
1819 head->next = rq->balance_callback;
1820 rq->balance_callback = head;
1821}
1822
1823#define rcu_dereference_check_sched_domain(p) \
1824 rcu_dereference_check((p), \
1825 lockdep_is_held(&sched_domains_mutex))
1826
1827/*
1828 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1829 * See destroy_sched_domains: call_rcu for details.
1830 *
1831 * The domain tree of any CPU may only be accessed from within
1832 * preempt-disabled sections.
1833 */
1834#define for_each_domain(cpu, __sd) \
1835 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1836 __sd; __sd = __sd->parent)
1837
1838/* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1839#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1840static const unsigned int SD_SHARED_CHILD_MASK =
1841#include <linux/sched/sd_flags.h>
18420;
1843#undef SD_FLAG
1844
1845/**
1846 * highest_flag_domain - Return highest sched_domain containing flag.
1847 * @cpu: The CPU whose highest level of sched domain is to
1848 * be returned.
1849 * @flag: The flag to check for the highest sched_domain
1850 * for the given CPU.
1851 *
1852 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1853 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1854 */
1855static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1856{
1857 struct sched_domain *sd, *hsd = NULL;
1858
1859 for_each_domain(cpu, sd) {
1860 if (sd->flags & flag) {
1861 hsd = sd;
1862 continue;
1863 }
1864
1865 /*
1866 * Stop the search if @flag is known to be shared at lower
1867 * levels. It will not be found further up.
1868 */
1869 if (flag & SD_SHARED_CHILD_MASK)
1870 break;
1871 }
1872
1873 return hsd;
1874}
1875
1876static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1877{
1878 struct sched_domain *sd;
1879
1880 for_each_domain(cpu, sd) {
1881 if (sd->flags & flag)
1882 break;
1883 }
1884
1885 return sd;
1886}
1887
1888DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1889DECLARE_PER_CPU(int, sd_llc_size);
1890DECLARE_PER_CPU(int, sd_llc_id);
1891DECLARE_PER_CPU(int, sd_share_id);
1892DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1893DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1894DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1895DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1896extern struct static_key_false sched_asym_cpucapacity;
1897extern struct static_key_false sched_cluster_active;
1898
1899static __always_inline bool sched_asym_cpucap_active(void)
1900{
1901 return static_branch_unlikely(&sched_asym_cpucapacity);
1902}
1903
1904struct sched_group_capacity {
1905 atomic_t ref;
1906 /*
1907 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1908 * for a single CPU.
1909 */
1910 unsigned long capacity;
1911 unsigned long min_capacity; /* Min per-CPU capacity in group */
1912 unsigned long max_capacity; /* Max per-CPU capacity in group */
1913 unsigned long next_update;
1914 int imbalance; /* XXX unrelated to capacity but shared group state */
1915
1916#ifdef CONFIG_SCHED_DEBUG
1917 int id;
1918#endif
1919
1920 unsigned long cpumask[]; /* Balance mask */
1921};
1922
1923struct sched_group {
1924 struct sched_group *next; /* Must be a circular list */
1925 atomic_t ref;
1926
1927 unsigned int group_weight;
1928 unsigned int cores;
1929 struct sched_group_capacity *sgc;
1930 int asym_prefer_cpu; /* CPU of highest priority in group */
1931 int flags;
1932
1933 /*
1934 * The CPUs this group covers.
1935 *
1936 * NOTE: this field is variable length. (Allocated dynamically
1937 * by attaching extra space to the end of the structure,
1938 * depending on how many CPUs the kernel has booted up with)
1939 */
1940 unsigned long cpumask[];
1941};
1942
1943static inline struct cpumask *sched_group_span(struct sched_group *sg)
1944{
1945 return to_cpumask(sg->cpumask);
1946}
1947
1948/*
1949 * See build_balance_mask().
1950 */
1951static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1952{
1953 return to_cpumask(sg->sgc->cpumask);
1954}
1955
1956extern int group_balance_cpu(struct sched_group *sg);
1957
1958#ifdef CONFIG_SCHED_DEBUG
1959void update_sched_domain_debugfs(void);
1960void dirty_sched_domain_sysctl(int cpu);
1961#else
1962static inline void update_sched_domain_debugfs(void)
1963{
1964}
1965static inline void dirty_sched_domain_sysctl(int cpu)
1966{
1967}
1968#endif
1969
1970extern int sched_update_scaling(void);
1971
1972static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1973{
1974 if (!p->user_cpus_ptr)
1975 return cpu_possible_mask; /* &init_task.cpus_mask */
1976 return p->user_cpus_ptr;
1977}
1978#endif /* CONFIG_SMP */
1979
1980#include "stats.h"
1981
1982#if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1983
1984extern void __sched_core_account_forceidle(struct rq *rq);
1985
1986static inline void sched_core_account_forceidle(struct rq *rq)
1987{
1988 if (schedstat_enabled())
1989 __sched_core_account_forceidle(rq);
1990}
1991
1992extern void __sched_core_tick(struct rq *rq);
1993
1994static inline void sched_core_tick(struct rq *rq)
1995{
1996 if (sched_core_enabled(rq) && schedstat_enabled())
1997 __sched_core_tick(rq);
1998}
1999
2000#else
2001
2002static inline void sched_core_account_forceidle(struct rq *rq) {}
2003
2004static inline void sched_core_tick(struct rq *rq) {}
2005
2006#endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
2007
2008#ifdef CONFIG_CGROUP_SCHED
2009
2010/*
2011 * Return the group to which this tasks belongs.
2012 *
2013 * We cannot use task_css() and friends because the cgroup subsystem
2014 * changes that value before the cgroup_subsys::attach() method is called,
2015 * therefore we cannot pin it and might observe the wrong value.
2016 *
2017 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2018 * core changes this before calling sched_move_task().
2019 *
2020 * Instead we use a 'copy' which is updated from sched_move_task() while
2021 * holding both task_struct::pi_lock and rq::lock.
2022 */
2023static inline struct task_group *task_group(struct task_struct *p)
2024{
2025 return p->sched_task_group;
2026}
2027
2028/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2029static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2030{
2031#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2032 struct task_group *tg = task_group(p);
2033#endif
2034
2035#ifdef CONFIG_FAIR_GROUP_SCHED
2036 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2037 p->se.cfs_rq = tg->cfs_rq[cpu];
2038 p->se.parent = tg->se[cpu];
2039 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2040#endif
2041
2042#ifdef CONFIG_RT_GROUP_SCHED
2043 p->rt.rt_rq = tg->rt_rq[cpu];
2044 p->rt.parent = tg->rt_se[cpu];
2045#endif
2046}
2047
2048#else /* CONFIG_CGROUP_SCHED */
2049
2050static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2051static inline struct task_group *task_group(struct task_struct *p)
2052{
2053 return NULL;
2054}
2055
2056#endif /* CONFIG_CGROUP_SCHED */
2057
2058static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2059{
2060 set_task_rq(p, cpu);
2061#ifdef CONFIG_SMP
2062 /*
2063 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2064 * successfully executed on another CPU. We must ensure that updates of
2065 * per-task data have been completed by this moment.
2066 */
2067 smp_wmb();
2068 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2069 p->wake_cpu = cpu;
2070#endif
2071}
2072
2073/*
2074 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2075 */
2076#ifdef CONFIG_SCHED_DEBUG
2077# define const_debug __read_mostly
2078#else
2079# define const_debug const
2080#endif
2081
2082#define SCHED_FEAT(name, enabled) \
2083 __SCHED_FEAT_##name ,
2084
2085enum {
2086#include "features.h"
2087 __SCHED_FEAT_NR,
2088};
2089
2090#undef SCHED_FEAT
2091
2092#ifdef CONFIG_SCHED_DEBUG
2093
2094/*
2095 * To support run-time toggling of sched features, all the translation units
2096 * (but core.c) reference the sysctl_sched_features defined in core.c.
2097 */
2098extern const_debug unsigned int sysctl_sched_features;
2099
2100#ifdef CONFIG_JUMP_LABEL
2101#define SCHED_FEAT(name, enabled) \
2102static __always_inline bool static_branch_##name(struct static_key *key) \
2103{ \
2104 return static_key_##enabled(key); \
2105}
2106
2107#include "features.h"
2108#undef SCHED_FEAT
2109
2110extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2111#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2112
2113#else /* !CONFIG_JUMP_LABEL */
2114
2115#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2116
2117#endif /* CONFIG_JUMP_LABEL */
2118
2119#else /* !SCHED_DEBUG */
2120
2121/*
2122 * Each translation unit has its own copy of sysctl_sched_features to allow
2123 * constants propagation at compile time and compiler optimization based on
2124 * features default.
2125 */
2126#define SCHED_FEAT(name, enabled) \
2127 (1UL << __SCHED_FEAT_##name) * enabled |
2128static const_debug __maybe_unused unsigned int sysctl_sched_features =
2129#include "features.h"
2130 0;
2131#undef SCHED_FEAT
2132
2133#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2134
2135#endif /* SCHED_DEBUG */
2136
2137extern struct static_key_false sched_numa_balancing;
2138extern struct static_key_false sched_schedstats;
2139
2140static inline u64 global_rt_period(void)
2141{
2142 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2143}
2144
2145static inline u64 global_rt_runtime(void)
2146{
2147 if (sysctl_sched_rt_runtime < 0)
2148 return RUNTIME_INF;
2149
2150 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2151}
2152
2153static inline int task_current(struct rq *rq, struct task_struct *p)
2154{
2155 return rq->curr == p;
2156}
2157
2158static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2159{
2160#ifdef CONFIG_SMP
2161 return p->on_cpu;
2162#else
2163 return task_current(rq, p);
2164#endif
2165}
2166
2167static inline int task_on_rq_queued(struct task_struct *p)
2168{
2169 return p->on_rq == TASK_ON_RQ_QUEUED;
2170}
2171
2172static inline int task_on_rq_migrating(struct task_struct *p)
2173{
2174 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2175}
2176
2177/* Wake flags. The first three directly map to some SD flag value */
2178#define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2179#define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2180#define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2181
2182#define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2183#define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2184#define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2185
2186#ifdef CONFIG_SMP
2187static_assert(WF_EXEC == SD_BALANCE_EXEC);
2188static_assert(WF_FORK == SD_BALANCE_FORK);
2189static_assert(WF_TTWU == SD_BALANCE_WAKE);
2190#endif
2191
2192/*
2193 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2194 * of tasks with abnormal "nice" values across CPUs the contribution that
2195 * each task makes to its run queue's load is weighted according to its
2196 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2197 * scaled version of the new time slice allocation that they receive on time
2198 * slice expiry etc.
2199 */
2200
2201#define WEIGHT_IDLEPRIO 3
2202#define WMULT_IDLEPRIO 1431655765
2203
2204extern const int sched_prio_to_weight[40];
2205extern const u32 sched_prio_to_wmult[40];
2206
2207/*
2208 * {de,en}queue flags:
2209 *
2210 * DEQUEUE_SLEEP - task is no longer runnable
2211 * ENQUEUE_WAKEUP - task just became runnable
2212 *
2213 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2214 * are in a known state which allows modification. Such pairs
2215 * should preserve as much state as possible.
2216 *
2217 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2218 * in the runqueue.
2219 *
2220 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2221 *
2222 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2223 *
2224 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2225 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2226 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2227 *
2228 */
2229
2230#define DEQUEUE_SLEEP 0x01
2231#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2232#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2233#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2234#define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2235
2236#define ENQUEUE_WAKEUP 0x01
2237#define ENQUEUE_RESTORE 0x02
2238#define ENQUEUE_MOVE 0x04
2239#define ENQUEUE_NOCLOCK 0x08
2240
2241#define ENQUEUE_HEAD 0x10
2242#define ENQUEUE_REPLENISH 0x20
2243#ifdef CONFIG_SMP
2244#define ENQUEUE_MIGRATED 0x40
2245#else
2246#define ENQUEUE_MIGRATED 0x00
2247#endif
2248#define ENQUEUE_INITIAL 0x80
2249#define ENQUEUE_MIGRATING 0x100
2250
2251#define RETRY_TASK ((void *)-1UL)
2252
2253struct affinity_context {
2254 const struct cpumask *new_mask;
2255 struct cpumask *user_mask;
2256 unsigned int flags;
2257};
2258
2259extern s64 update_curr_common(struct rq *rq);
2260
2261struct sched_class {
2262
2263#ifdef CONFIG_UCLAMP_TASK
2264 int uclamp_enabled;
2265#endif
2266
2267 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2268 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2269 void (*yield_task) (struct rq *rq);
2270 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2271
2272 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2273
2274 struct task_struct *(*pick_next_task)(struct rq *rq);
2275
2276 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2277 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2278
2279#ifdef CONFIG_SMP
2280 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2281 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2282
2283 struct task_struct * (*pick_task)(struct rq *rq);
2284
2285 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2286
2287 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2288
2289 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2290
2291 void (*rq_online)(struct rq *rq);
2292 void (*rq_offline)(struct rq *rq);
2293
2294 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2295#endif
2296
2297 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2298 void (*task_fork)(struct task_struct *p);
2299 void (*task_dead)(struct task_struct *p);
2300
2301 /*
2302 * The switched_from() call is allowed to drop rq->lock, therefore we
2303 * cannot assume the switched_from/switched_to pair is serialized by
2304 * rq->lock. They are however serialized by p->pi_lock.
2305 */
2306 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2307 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2308 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2309 int oldprio);
2310
2311 unsigned int (*get_rr_interval)(struct rq *rq,
2312 struct task_struct *task);
2313
2314 void (*update_curr)(struct rq *rq);
2315
2316#ifdef CONFIG_FAIR_GROUP_SCHED
2317 void (*task_change_group)(struct task_struct *p);
2318#endif
2319
2320#ifdef CONFIG_SCHED_CORE
2321 int (*task_is_throttled)(struct task_struct *p, int cpu);
2322#endif
2323};
2324
2325static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2326{
2327 WARN_ON_ONCE(rq->curr != prev);
2328 prev->sched_class->put_prev_task(rq, prev);
2329}
2330
2331static inline void set_next_task(struct rq *rq, struct task_struct *next)
2332{
2333 next->sched_class->set_next_task(rq, next, false);
2334}
2335
2336
2337/*
2338 * Helper to define a sched_class instance; each one is placed in a separate
2339 * section which is ordered by the linker script:
2340 *
2341 * include/asm-generic/vmlinux.lds.h
2342 *
2343 * *CAREFUL* they are laid out in *REVERSE* order!!!
2344 *
2345 * Also enforce alignment on the instance, not the type, to guarantee layout.
2346 */
2347#define DEFINE_SCHED_CLASS(name) \
2348const struct sched_class name##_sched_class \
2349 __aligned(__alignof__(struct sched_class)) \
2350 __section("__" #name "_sched_class")
2351
2352/* Defined in include/asm-generic/vmlinux.lds.h */
2353extern struct sched_class __sched_class_highest[];
2354extern struct sched_class __sched_class_lowest[];
2355
2356#define for_class_range(class, _from, _to) \
2357 for (class = (_from); class < (_to); class++)
2358
2359#define for_each_class(class) \
2360 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2361
2362#define sched_class_above(_a, _b) ((_a) < (_b))
2363
2364extern const struct sched_class stop_sched_class;
2365extern const struct sched_class dl_sched_class;
2366extern const struct sched_class rt_sched_class;
2367extern const struct sched_class fair_sched_class;
2368extern const struct sched_class idle_sched_class;
2369
2370static inline bool sched_stop_runnable(struct rq *rq)
2371{
2372 return rq->stop && task_on_rq_queued(rq->stop);
2373}
2374
2375static inline bool sched_dl_runnable(struct rq *rq)
2376{
2377 return rq->dl.dl_nr_running > 0;
2378}
2379
2380static inline bool sched_rt_runnable(struct rq *rq)
2381{
2382 return rq->rt.rt_queued > 0;
2383}
2384
2385static inline bool sched_fair_runnable(struct rq *rq)
2386{
2387 return rq->cfs.nr_running > 0;
2388}
2389
2390extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2391extern struct task_struct *pick_next_task_idle(struct rq *rq);
2392
2393#define SCA_CHECK 0x01
2394#define SCA_MIGRATE_DISABLE 0x02
2395#define SCA_MIGRATE_ENABLE 0x04
2396#define SCA_USER 0x08
2397
2398#ifdef CONFIG_SMP
2399
2400extern void update_group_capacity(struct sched_domain *sd, int cpu);
2401
2402extern void trigger_load_balance(struct rq *rq);
2403
2404extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2405
2406static inline struct task_struct *get_push_task(struct rq *rq)
2407{
2408 struct task_struct *p = rq->curr;
2409
2410 lockdep_assert_rq_held(rq);
2411
2412 if (rq->push_busy)
2413 return NULL;
2414
2415 if (p->nr_cpus_allowed == 1)
2416 return NULL;
2417
2418 if (p->migration_disabled)
2419 return NULL;
2420
2421 rq->push_busy = true;
2422 return get_task_struct(p);
2423}
2424
2425extern int push_cpu_stop(void *arg);
2426
2427#endif
2428
2429#ifdef CONFIG_CPU_IDLE
2430static inline void idle_set_state(struct rq *rq,
2431 struct cpuidle_state *idle_state)
2432{
2433 rq->idle_state = idle_state;
2434}
2435
2436static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2437{
2438 SCHED_WARN_ON(!rcu_read_lock_held());
2439
2440 return rq->idle_state;
2441}
2442#else
2443static inline void idle_set_state(struct rq *rq,
2444 struct cpuidle_state *idle_state)
2445{
2446}
2447
2448static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2449{
2450 return NULL;
2451}
2452#endif
2453
2454extern void schedule_idle(void);
2455asmlinkage void schedule_user(void);
2456
2457extern void sysrq_sched_debug_show(void);
2458extern void sched_init_granularity(void);
2459extern void update_max_interval(void);
2460
2461extern void init_sched_dl_class(void);
2462extern void init_sched_rt_class(void);
2463extern void init_sched_fair_class(void);
2464
2465extern void reweight_task(struct task_struct *p, int prio);
2466
2467extern void resched_curr(struct rq *rq);
2468extern void resched_cpu(int cpu);
2469
2470extern struct rt_bandwidth def_rt_bandwidth;
2471extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2472extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2473
2474extern void init_dl_entity(struct sched_dl_entity *dl_se);
2475
2476#define BW_SHIFT 20
2477#define BW_UNIT (1 << BW_SHIFT)
2478#define RATIO_SHIFT 8
2479#define MAX_BW_BITS (64 - BW_SHIFT)
2480#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2481unsigned long to_ratio(u64 period, u64 runtime);
2482
2483extern void init_entity_runnable_average(struct sched_entity *se);
2484extern void post_init_entity_util_avg(struct task_struct *p);
2485
2486#ifdef CONFIG_NO_HZ_FULL
2487extern bool sched_can_stop_tick(struct rq *rq);
2488extern int __init sched_tick_offload_init(void);
2489
2490/*
2491 * Tick may be needed by tasks in the runqueue depending on their policy and
2492 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2493 * nohz mode if necessary.
2494 */
2495static inline void sched_update_tick_dependency(struct rq *rq)
2496{
2497 int cpu = cpu_of(rq);
2498
2499 if (!tick_nohz_full_cpu(cpu))
2500 return;
2501
2502 if (sched_can_stop_tick(rq))
2503 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2504 else
2505 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2506}
2507#else
2508static inline int sched_tick_offload_init(void) { return 0; }
2509static inline void sched_update_tick_dependency(struct rq *rq) { }
2510#endif
2511
2512static inline void add_nr_running(struct rq *rq, unsigned count)
2513{
2514 unsigned prev_nr = rq->nr_running;
2515
2516 rq->nr_running = prev_nr + count;
2517 if (trace_sched_update_nr_running_tp_enabled()) {
2518 call_trace_sched_update_nr_running(rq, count);
2519 }
2520
2521#ifdef CONFIG_SMP
2522 if (prev_nr < 2 && rq->nr_running >= 2) {
2523 if (!READ_ONCE(rq->rd->overload))
2524 WRITE_ONCE(rq->rd->overload, 1);
2525 }
2526#endif
2527
2528 sched_update_tick_dependency(rq);
2529}
2530
2531static inline void sub_nr_running(struct rq *rq, unsigned count)
2532{
2533 rq->nr_running -= count;
2534 if (trace_sched_update_nr_running_tp_enabled()) {
2535 call_trace_sched_update_nr_running(rq, -count);
2536 }
2537
2538 /* Check if we still need preemption */
2539 sched_update_tick_dependency(rq);
2540}
2541
2542extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2543extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2544
2545extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2546
2547#ifdef CONFIG_PREEMPT_RT
2548#define SCHED_NR_MIGRATE_BREAK 8
2549#else
2550#define SCHED_NR_MIGRATE_BREAK 32
2551#endif
2552
2553extern const_debug unsigned int sysctl_sched_nr_migrate;
2554extern const_debug unsigned int sysctl_sched_migration_cost;
2555
2556extern unsigned int sysctl_sched_base_slice;
2557
2558#ifdef CONFIG_SCHED_DEBUG
2559extern int sysctl_resched_latency_warn_ms;
2560extern int sysctl_resched_latency_warn_once;
2561
2562extern unsigned int sysctl_sched_tunable_scaling;
2563
2564extern unsigned int sysctl_numa_balancing_scan_delay;
2565extern unsigned int sysctl_numa_balancing_scan_period_min;
2566extern unsigned int sysctl_numa_balancing_scan_period_max;
2567extern unsigned int sysctl_numa_balancing_scan_size;
2568extern unsigned int sysctl_numa_balancing_hot_threshold;
2569#endif
2570
2571#ifdef CONFIG_SCHED_HRTICK
2572
2573/*
2574 * Use hrtick when:
2575 * - enabled by features
2576 * - hrtimer is actually high res
2577 */
2578static inline int hrtick_enabled(struct rq *rq)
2579{
2580 if (!cpu_active(cpu_of(rq)))
2581 return 0;
2582 return hrtimer_is_hres_active(&rq->hrtick_timer);
2583}
2584
2585static inline int hrtick_enabled_fair(struct rq *rq)
2586{
2587 if (!sched_feat(HRTICK))
2588 return 0;
2589 return hrtick_enabled(rq);
2590}
2591
2592static inline int hrtick_enabled_dl(struct rq *rq)
2593{
2594 if (!sched_feat(HRTICK_DL))
2595 return 0;
2596 return hrtick_enabled(rq);
2597}
2598
2599void hrtick_start(struct rq *rq, u64 delay);
2600
2601#else
2602
2603static inline int hrtick_enabled_fair(struct rq *rq)
2604{
2605 return 0;
2606}
2607
2608static inline int hrtick_enabled_dl(struct rq *rq)
2609{
2610 return 0;
2611}
2612
2613static inline int hrtick_enabled(struct rq *rq)
2614{
2615 return 0;
2616}
2617
2618#endif /* CONFIG_SCHED_HRTICK */
2619
2620#ifndef arch_scale_freq_tick
2621static __always_inline
2622void arch_scale_freq_tick(void)
2623{
2624}
2625#endif
2626
2627#ifndef arch_scale_freq_capacity
2628/**
2629 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2630 * @cpu: the CPU in question.
2631 *
2632 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2633 *
2634 * f_curr
2635 * ------ * SCHED_CAPACITY_SCALE
2636 * f_max
2637 */
2638static __always_inline
2639unsigned long arch_scale_freq_capacity(int cpu)
2640{
2641 return SCHED_CAPACITY_SCALE;
2642}
2643#endif
2644
2645#ifdef CONFIG_SCHED_DEBUG
2646/*
2647 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2648 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2649 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2650 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2651 */
2652static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2653{
2654 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2655 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2656#ifdef CONFIG_SMP
2657 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2658#endif
2659}
2660#else
2661static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2662#endif
2663
2664#define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2665__DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2666static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2667{ class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2668 _lock; return _t; }
2669
2670#ifdef CONFIG_SMP
2671
2672static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2673{
2674#ifdef CONFIG_SCHED_CORE
2675 /*
2676 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2677 * order by core-id first and cpu-id second.
2678 *
2679 * Notably:
2680 *
2681 * double_rq_lock(0,3); will take core-0, core-1 lock
2682 * double_rq_lock(1,2); will take core-1, core-0 lock
2683 *
2684 * when only cpu-id is considered.
2685 */
2686 if (rq1->core->cpu < rq2->core->cpu)
2687 return true;
2688 if (rq1->core->cpu > rq2->core->cpu)
2689 return false;
2690
2691 /*
2692 * __sched_core_flip() relies on SMT having cpu-id lock order.
2693 */
2694#endif
2695 return rq1->cpu < rq2->cpu;
2696}
2697
2698extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2699
2700#ifdef CONFIG_PREEMPTION
2701
2702/*
2703 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2704 * way at the expense of forcing extra atomic operations in all
2705 * invocations. This assures that the double_lock is acquired using the
2706 * same underlying policy as the spinlock_t on this architecture, which
2707 * reduces latency compared to the unfair variant below. However, it
2708 * also adds more overhead and therefore may reduce throughput.
2709 */
2710static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2711 __releases(this_rq->lock)
2712 __acquires(busiest->lock)
2713 __acquires(this_rq->lock)
2714{
2715 raw_spin_rq_unlock(this_rq);
2716 double_rq_lock(this_rq, busiest);
2717
2718 return 1;
2719}
2720
2721#else
2722/*
2723 * Unfair double_lock_balance: Optimizes throughput at the expense of
2724 * latency by eliminating extra atomic operations when the locks are
2725 * already in proper order on entry. This favors lower CPU-ids and will
2726 * grant the double lock to lower CPUs over higher ids under contention,
2727 * regardless of entry order into the function.
2728 */
2729static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2730 __releases(this_rq->lock)
2731 __acquires(busiest->lock)
2732 __acquires(this_rq->lock)
2733{
2734 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2735 likely(raw_spin_rq_trylock(busiest))) {
2736 double_rq_clock_clear_update(this_rq, busiest);
2737 return 0;
2738 }
2739
2740 if (rq_order_less(this_rq, busiest)) {
2741 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2742 double_rq_clock_clear_update(this_rq, busiest);
2743 return 0;
2744 }
2745
2746 raw_spin_rq_unlock(this_rq);
2747 double_rq_lock(this_rq, busiest);
2748
2749 return 1;
2750}
2751
2752#endif /* CONFIG_PREEMPTION */
2753
2754/*
2755 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2756 */
2757static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2758{
2759 lockdep_assert_irqs_disabled();
2760
2761 return _double_lock_balance(this_rq, busiest);
2762}
2763
2764static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2765 __releases(busiest->lock)
2766{
2767 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2768 raw_spin_rq_unlock(busiest);
2769 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2770}
2771
2772static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2773{
2774 if (l1 > l2)
2775 swap(l1, l2);
2776
2777 spin_lock(l1);
2778 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2779}
2780
2781static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2782{
2783 if (l1 > l2)
2784 swap(l1, l2);
2785
2786 spin_lock_irq(l1);
2787 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2788}
2789
2790static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2791{
2792 if (l1 > l2)
2793 swap(l1, l2);
2794
2795 raw_spin_lock(l1);
2796 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2797}
2798
2799static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2800{
2801 raw_spin_unlock(l1);
2802 raw_spin_unlock(l2);
2803}
2804
2805DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
2806 double_raw_lock(_T->lock, _T->lock2),
2807 double_raw_unlock(_T->lock, _T->lock2))
2808
2809/*
2810 * double_rq_unlock - safely unlock two runqueues
2811 *
2812 * Note this does not restore interrupts like task_rq_unlock,
2813 * you need to do so manually after calling.
2814 */
2815static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2816 __releases(rq1->lock)
2817 __releases(rq2->lock)
2818{
2819 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2820 raw_spin_rq_unlock(rq2);
2821 else
2822 __release(rq2->lock);
2823 raw_spin_rq_unlock(rq1);
2824}
2825
2826extern void set_rq_online (struct rq *rq);
2827extern void set_rq_offline(struct rq *rq);
2828extern bool sched_smp_initialized;
2829
2830#else /* CONFIG_SMP */
2831
2832/*
2833 * double_rq_lock - safely lock two runqueues
2834 *
2835 * Note this does not disable interrupts like task_rq_lock,
2836 * you need to do so manually before calling.
2837 */
2838static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2839 __acquires(rq1->lock)
2840 __acquires(rq2->lock)
2841{
2842 WARN_ON_ONCE(!irqs_disabled());
2843 WARN_ON_ONCE(rq1 != rq2);
2844 raw_spin_rq_lock(rq1);
2845 __acquire(rq2->lock); /* Fake it out ;) */
2846 double_rq_clock_clear_update(rq1, rq2);
2847}
2848
2849/*
2850 * double_rq_unlock - safely unlock two runqueues
2851 *
2852 * Note this does not restore interrupts like task_rq_unlock,
2853 * you need to do so manually after calling.
2854 */
2855static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2856 __releases(rq1->lock)
2857 __releases(rq2->lock)
2858{
2859 WARN_ON_ONCE(rq1 != rq2);
2860 raw_spin_rq_unlock(rq1);
2861 __release(rq2->lock);
2862}
2863
2864#endif
2865
2866DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
2867 double_rq_lock(_T->lock, _T->lock2),
2868 double_rq_unlock(_T->lock, _T->lock2))
2869
2870extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
2871extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2872extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2873
2874#ifdef CONFIG_SCHED_DEBUG
2875extern bool sched_debug_verbose;
2876
2877extern void print_cfs_stats(struct seq_file *m, int cpu);
2878extern void print_rt_stats(struct seq_file *m, int cpu);
2879extern void print_dl_stats(struct seq_file *m, int cpu);
2880extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2881extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2882extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2883
2884extern void resched_latency_warn(int cpu, u64 latency);
2885#ifdef CONFIG_NUMA_BALANCING
2886extern void
2887show_numa_stats(struct task_struct *p, struct seq_file *m);
2888extern void
2889print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2890 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2891#endif /* CONFIG_NUMA_BALANCING */
2892#else
2893static inline void resched_latency_warn(int cpu, u64 latency) {}
2894#endif /* CONFIG_SCHED_DEBUG */
2895
2896extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2897extern void init_rt_rq(struct rt_rq *rt_rq);
2898extern void init_dl_rq(struct dl_rq *dl_rq);
2899
2900extern void cfs_bandwidth_usage_inc(void);
2901extern void cfs_bandwidth_usage_dec(void);
2902
2903#ifdef CONFIG_NO_HZ_COMMON
2904#define NOHZ_BALANCE_KICK_BIT 0
2905#define NOHZ_STATS_KICK_BIT 1
2906#define NOHZ_NEWILB_KICK_BIT 2
2907#define NOHZ_NEXT_KICK_BIT 3
2908
2909/* Run rebalance_domains() */
2910#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2911/* Update blocked load */
2912#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2913/* Update blocked load when entering idle */
2914#define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2915/* Update nohz.next_balance */
2916#define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2917
2918#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2919
2920#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2921
2922extern void nohz_balance_exit_idle(struct rq *rq);
2923#else
2924static inline void nohz_balance_exit_idle(struct rq *rq) { }
2925#endif
2926
2927#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2928extern void nohz_run_idle_balance(int cpu);
2929#else
2930static inline void nohz_run_idle_balance(int cpu) { }
2931#endif
2932
2933#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2934struct irqtime {
2935 u64 total;
2936 u64 tick_delta;
2937 u64 irq_start_time;
2938 struct u64_stats_sync sync;
2939};
2940
2941DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2942
2943/*
2944 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2945 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2946 * and never move forward.
2947 */
2948static inline u64 irq_time_read(int cpu)
2949{
2950 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2951 unsigned int seq;
2952 u64 total;
2953
2954 do {
2955 seq = __u64_stats_fetch_begin(&irqtime->sync);
2956 total = irqtime->total;
2957 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2958
2959 return total;
2960}
2961#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2962
2963#ifdef CONFIG_CPU_FREQ
2964DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2965
2966/**
2967 * cpufreq_update_util - Take a note about CPU utilization changes.
2968 * @rq: Runqueue to carry out the update for.
2969 * @flags: Update reason flags.
2970 *
2971 * This function is called by the scheduler on the CPU whose utilization is
2972 * being updated.
2973 *
2974 * It can only be called from RCU-sched read-side critical sections.
2975 *
2976 * The way cpufreq is currently arranged requires it to evaluate the CPU
2977 * performance state (frequency/voltage) on a regular basis to prevent it from
2978 * being stuck in a completely inadequate performance level for too long.
2979 * That is not guaranteed to happen if the updates are only triggered from CFS
2980 * and DL, though, because they may not be coming in if only RT tasks are
2981 * active all the time (or there are RT tasks only).
2982 *
2983 * As a workaround for that issue, this function is called periodically by the
2984 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2985 * but that really is a band-aid. Going forward it should be replaced with
2986 * solutions targeted more specifically at RT tasks.
2987 */
2988static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2989{
2990 struct update_util_data *data;
2991
2992 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2993 cpu_of(rq)));
2994 if (data)
2995 data->func(data, rq_clock(rq), flags);
2996}
2997#else
2998static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2999#endif /* CONFIG_CPU_FREQ */
3000
3001#ifdef arch_scale_freq_capacity
3002# ifndef arch_scale_freq_invariant
3003# define arch_scale_freq_invariant() true
3004# endif
3005#else
3006# define arch_scale_freq_invariant() false
3007#endif
3008
3009#ifdef CONFIG_SMP
3010unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3011 unsigned long *min,
3012 unsigned long *max);
3013
3014unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3015 unsigned long min,
3016 unsigned long max);
3017
3018
3019/*
3020 * Verify the fitness of task @p to run on @cpu taking into account the
3021 * CPU original capacity and the runtime/deadline ratio of the task.
3022 *
3023 * The function will return true if the original capacity of @cpu is
3024 * greater than or equal to task's deadline density right shifted by
3025 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3026 */
3027static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3028{
3029 unsigned long cap = arch_scale_cpu_capacity(cpu);
3030
3031 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3032}
3033
3034static inline unsigned long cpu_bw_dl(struct rq *rq)
3035{
3036 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3037}
3038
3039static inline unsigned long cpu_util_dl(struct rq *rq)
3040{
3041 return READ_ONCE(rq->avg_dl.util_avg);
3042}
3043
3044
3045extern unsigned long cpu_util_cfs(int cpu);
3046extern unsigned long cpu_util_cfs_boost(int cpu);
3047
3048static inline unsigned long cpu_util_rt(struct rq *rq)
3049{
3050 return READ_ONCE(rq->avg_rt.util_avg);
3051}
3052#endif
3053
3054#ifdef CONFIG_UCLAMP_TASK
3055unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3056
3057static inline unsigned long uclamp_rq_get(struct rq *rq,
3058 enum uclamp_id clamp_id)
3059{
3060 return READ_ONCE(rq->uclamp[clamp_id].value);
3061}
3062
3063static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3064 unsigned int value)
3065{
3066 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3067}
3068
3069static inline bool uclamp_rq_is_idle(struct rq *rq)
3070{
3071 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3072}
3073
3074/* Is the rq being capped/throttled by uclamp_max? */
3075static inline bool uclamp_rq_is_capped(struct rq *rq)
3076{
3077 unsigned long rq_util;
3078 unsigned long max_util;
3079
3080 if (!static_branch_likely(&sched_uclamp_used))
3081 return false;
3082
3083 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3084 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3085
3086 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3087}
3088
3089/*
3090 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3091 * by default in the fast path and only gets turned on once userspace performs
3092 * an operation that requires it.
3093 *
3094 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3095 * hence is active.
3096 */
3097static inline bool uclamp_is_used(void)
3098{
3099 return static_branch_likely(&sched_uclamp_used);
3100}
3101#else /* CONFIG_UCLAMP_TASK */
3102static inline unsigned long uclamp_eff_value(struct task_struct *p,
3103 enum uclamp_id clamp_id)
3104{
3105 if (clamp_id == UCLAMP_MIN)
3106 return 0;
3107
3108 return SCHED_CAPACITY_SCALE;
3109}
3110
3111static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3112
3113static inline bool uclamp_is_used(void)
3114{
3115 return false;
3116}
3117
3118static inline unsigned long uclamp_rq_get(struct rq *rq,
3119 enum uclamp_id clamp_id)
3120{
3121 if (clamp_id == UCLAMP_MIN)
3122 return 0;
3123
3124 return SCHED_CAPACITY_SCALE;
3125}
3126
3127static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3128 unsigned int value)
3129{
3130}
3131
3132static inline bool uclamp_rq_is_idle(struct rq *rq)
3133{
3134 return false;
3135}
3136#endif /* CONFIG_UCLAMP_TASK */
3137
3138#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3139static inline unsigned long cpu_util_irq(struct rq *rq)
3140{
3141 return READ_ONCE(rq->avg_irq.util_avg);
3142}
3143
3144static inline
3145unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3146{
3147 util *= (max - irq);
3148 util /= max;
3149
3150 return util;
3151
3152}
3153#else
3154static inline unsigned long cpu_util_irq(struct rq *rq)
3155{
3156 return 0;
3157}
3158
3159static inline
3160unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3161{
3162 return util;
3163}
3164#endif
3165
3166#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3167
3168#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3169
3170DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3171
3172static inline bool sched_energy_enabled(void)
3173{
3174 return static_branch_unlikely(&sched_energy_present);
3175}
3176
3177extern struct cpufreq_governor schedutil_gov;
3178
3179#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3180
3181#define perf_domain_span(pd) NULL
3182static inline bool sched_energy_enabled(void) { return false; }
3183
3184#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3185
3186#ifdef CONFIG_MEMBARRIER
3187/*
3188 * The scheduler provides memory barriers required by membarrier between:
3189 * - prior user-space memory accesses and store to rq->membarrier_state,
3190 * - store to rq->membarrier_state and following user-space memory accesses.
3191 * In the same way it provides those guarantees around store to rq->curr.
3192 */
3193static inline void membarrier_switch_mm(struct rq *rq,
3194 struct mm_struct *prev_mm,
3195 struct mm_struct *next_mm)
3196{
3197 int membarrier_state;
3198
3199 if (prev_mm == next_mm)
3200 return;
3201
3202 membarrier_state = atomic_read(&next_mm->membarrier_state);
3203 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3204 return;
3205
3206 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3207}
3208#else
3209static inline void membarrier_switch_mm(struct rq *rq,
3210 struct mm_struct *prev_mm,
3211 struct mm_struct *next_mm)
3212{
3213}
3214#endif
3215
3216#ifdef CONFIG_SMP
3217static inline bool is_per_cpu_kthread(struct task_struct *p)
3218{
3219 if (!(p->flags & PF_KTHREAD))
3220 return false;
3221
3222 if (p->nr_cpus_allowed != 1)
3223 return false;
3224
3225 return true;
3226}
3227#endif
3228
3229extern void swake_up_all_locked(struct swait_queue_head *q);
3230extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3231
3232extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3233
3234#ifdef CONFIG_PREEMPT_DYNAMIC
3235extern int preempt_dynamic_mode;
3236extern int sched_dynamic_mode(const char *str);
3237extern void sched_dynamic_update(int mode);
3238#endif
3239
3240#ifdef CONFIG_SCHED_MM_CID
3241
3242#define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3243#define MM_CID_SCAN_DELAY 100 /* 100ms */
3244
3245extern raw_spinlock_t cid_lock;
3246extern int use_cid_lock;
3247
3248extern void sched_mm_cid_migrate_from(struct task_struct *t);
3249extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3250extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3251extern void init_sched_mm_cid(struct task_struct *t);
3252
3253static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3254{
3255 if (cid < 0)
3256 return;
3257 cpumask_clear_cpu(cid, mm_cidmask(mm));
3258}
3259
3260/*
3261 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3262 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3263 * be held to transition to other states.
3264 *
3265 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3266 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3267 */
3268static inline void mm_cid_put_lazy(struct task_struct *t)
3269{
3270 struct mm_struct *mm = t->mm;
3271 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3272 int cid;
3273
3274 lockdep_assert_irqs_disabled();
3275 cid = __this_cpu_read(pcpu_cid->cid);
3276 if (!mm_cid_is_lazy_put(cid) ||
3277 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3278 return;
3279 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3280}
3281
3282static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3283{
3284 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3285 int cid, res;
3286
3287 lockdep_assert_irqs_disabled();
3288 cid = __this_cpu_read(pcpu_cid->cid);
3289 for (;;) {
3290 if (mm_cid_is_unset(cid))
3291 return MM_CID_UNSET;
3292 /*
3293 * Attempt transition from valid or lazy-put to unset.
3294 */
3295 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3296 if (res == cid)
3297 break;
3298 cid = res;
3299 }
3300 return cid;
3301}
3302
3303static inline void mm_cid_put(struct mm_struct *mm)
3304{
3305 int cid;
3306
3307 lockdep_assert_irqs_disabled();
3308 cid = mm_cid_pcpu_unset(mm);
3309 if (cid == MM_CID_UNSET)
3310 return;
3311 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3312}
3313
3314static inline int __mm_cid_try_get(struct mm_struct *mm)
3315{
3316 struct cpumask *cpumask;
3317 int cid;
3318
3319 cpumask = mm_cidmask(mm);
3320 /*
3321 * Retry finding first zero bit if the mask is temporarily
3322 * filled. This only happens during concurrent remote-clear
3323 * which owns a cid without holding a rq lock.
3324 */
3325 for (;;) {
3326 cid = cpumask_first_zero(cpumask);
3327 if (cid < nr_cpu_ids)
3328 break;
3329 cpu_relax();
3330 }
3331 if (cpumask_test_and_set_cpu(cid, cpumask))
3332 return -1;
3333 return cid;
3334}
3335
3336/*
3337 * Save a snapshot of the current runqueue time of this cpu
3338 * with the per-cpu cid value, allowing to estimate how recently it was used.
3339 */
3340static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3341{
3342 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3343
3344 lockdep_assert_rq_held(rq);
3345 WRITE_ONCE(pcpu_cid->time, rq->clock);
3346}
3347
3348static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3349{
3350 int cid;
3351
3352 /*
3353 * All allocations (even those using the cid_lock) are lock-free. If
3354 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3355 * guarantee forward progress.
3356 */
3357 if (!READ_ONCE(use_cid_lock)) {
3358 cid = __mm_cid_try_get(mm);
3359 if (cid >= 0)
3360 goto end;
3361 raw_spin_lock(&cid_lock);
3362 } else {
3363 raw_spin_lock(&cid_lock);
3364 cid = __mm_cid_try_get(mm);
3365 if (cid >= 0)
3366 goto unlock;
3367 }
3368
3369 /*
3370 * cid concurrently allocated. Retry while forcing following
3371 * allocations to use the cid_lock to ensure forward progress.
3372 */
3373 WRITE_ONCE(use_cid_lock, 1);
3374 /*
3375 * Set use_cid_lock before allocation. Only care about program order
3376 * because this is only required for forward progress.
3377 */
3378 barrier();
3379 /*
3380 * Retry until it succeeds. It is guaranteed to eventually succeed once
3381 * all newcoming allocations observe the use_cid_lock flag set.
3382 */
3383 do {
3384 cid = __mm_cid_try_get(mm);
3385 cpu_relax();
3386 } while (cid < 0);
3387 /*
3388 * Allocate before clearing use_cid_lock. Only care about
3389 * program order because this is for forward progress.
3390 */
3391 barrier();
3392 WRITE_ONCE(use_cid_lock, 0);
3393unlock:
3394 raw_spin_unlock(&cid_lock);
3395end:
3396 mm_cid_snapshot_time(rq, mm);
3397 return cid;
3398}
3399
3400static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3401{
3402 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3403 struct cpumask *cpumask;
3404 int cid;
3405
3406 lockdep_assert_rq_held(rq);
3407 cpumask = mm_cidmask(mm);
3408 cid = __this_cpu_read(pcpu_cid->cid);
3409 if (mm_cid_is_valid(cid)) {
3410 mm_cid_snapshot_time(rq, mm);
3411 return cid;
3412 }
3413 if (mm_cid_is_lazy_put(cid)) {
3414 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3415 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3416 }
3417 cid = __mm_cid_get(rq, mm);
3418 __this_cpu_write(pcpu_cid->cid, cid);
3419 return cid;
3420}
3421
3422static inline void switch_mm_cid(struct rq *rq,
3423 struct task_struct *prev,
3424 struct task_struct *next)
3425{
3426 /*
3427 * Provide a memory barrier between rq->curr store and load of
3428 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3429 *
3430 * Should be adapted if context_switch() is modified.
3431 */
3432 if (!next->mm) { // to kernel
3433 /*
3434 * user -> kernel transition does not guarantee a barrier, but
3435 * we can use the fact that it performs an atomic operation in
3436 * mmgrab().
3437 */
3438 if (prev->mm) // from user
3439 smp_mb__after_mmgrab();
3440 /*
3441 * kernel -> kernel transition does not change rq->curr->mm
3442 * state. It stays NULL.
3443 */
3444 } else { // to user
3445 /*
3446 * kernel -> user transition does not provide a barrier
3447 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3448 * Provide it here.
3449 */
3450 if (!prev->mm) { // from kernel
3451 smp_mb();
3452 } else { // from user
3453 /*
3454 * user->user transition relies on an implicit
3455 * memory barrier in switch_mm() when
3456 * current->mm changes. If the architecture
3457 * switch_mm() does not have an implicit memory
3458 * barrier, it is emitted here. If current->mm
3459 * is unchanged, no barrier is needed.
3460 */
3461 smp_mb__after_switch_mm();
3462 }
3463 }
3464 if (prev->mm_cid_active) {
3465 mm_cid_snapshot_time(rq, prev->mm);
3466 mm_cid_put_lazy(prev);
3467 prev->mm_cid = -1;
3468 }
3469 if (next->mm_cid_active)
3470 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3471}
3472
3473#else
3474static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3475static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3476static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3477static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3478static inline void init_sched_mm_cid(struct task_struct *t) { }
3479#endif
3480
3481extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3482extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3483
3484#endif /* _KERNEL_SCHED_SCHED_H */