Loading...
Note: File does not exist in v3.1.
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * Scheduler internal types and methods:
4 */
5#include <linux/sched.h>
6
7#include <linux/sched/autogroup.h>
8#include <linux/sched/clock.h>
9#include <linux/sched/coredump.h>
10#include <linux/sched/cpufreq.h>
11#include <linux/sched/cputime.h>
12#include <linux/sched/deadline.h>
13#include <linux/sched/debug.h>
14#include <linux/sched/hotplug.h>
15#include <linux/sched/idle.h>
16#include <linux/sched/init.h>
17#include <linux/sched/isolation.h>
18#include <linux/sched/jobctl.h>
19#include <linux/sched/loadavg.h>
20#include <linux/sched/mm.h>
21#include <linux/sched/nohz.h>
22#include <linux/sched/numa_balancing.h>
23#include <linux/sched/prio.h>
24#include <linux/sched/rt.h>
25#include <linux/sched/signal.h>
26#include <linux/sched/stat.h>
27#include <linux/sched/sysctl.h>
28#include <linux/sched/task.h>
29#include <linux/sched/task_stack.h>
30#include <linux/sched/topology.h>
31#include <linux/sched/user.h>
32#include <linux/sched/wake_q.h>
33#include <linux/sched/xacct.h>
34
35#include <uapi/linux/sched/types.h>
36
37#include <linux/binfmts.h>
38#include <linux/blkdev.h>
39#include <linux/compat.h>
40#include <linux/context_tracking.h>
41#include <linux/cpufreq.h>
42#include <linux/cpuidle.h>
43#include <linux/cpuset.h>
44#include <linux/ctype.h>
45#include <linux/debugfs.h>
46#include <linux/delayacct.h>
47#include <linux/init_task.h>
48#include <linux/kprobes.h>
49#include <linux/kthread.h>
50#include <linux/membarrier.h>
51#include <linux/migrate.h>
52#include <linux/mmu_context.h>
53#include <linux/nmi.h>
54#include <linux/proc_fs.h>
55#include <linux/prefetch.h>
56#include <linux/profile.h>
57#include <linux/rcupdate_wait.h>
58#include <linux/security.h>
59#include <linux/stackprotector.h>
60#include <linux/stop_machine.h>
61#include <linux/suspend.h>
62#include <linux/swait.h>
63#include <linux/syscalls.h>
64#include <linux/task_work.h>
65#include <linux/tsacct_kern.h>
66
67#include <asm/tlb.h>
68
69#ifdef CONFIG_PARAVIRT
70# include <asm/paravirt.h>
71#endif
72
73#include "cpupri.h"
74#include "cpudeadline.h"
75
76#ifdef CONFIG_SCHED_DEBUG
77# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
78#else
79# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
80#endif
81
82struct rq;
83struct cpuidle_state;
84
85/* task_struct::on_rq states: */
86#define TASK_ON_RQ_QUEUED 1
87#define TASK_ON_RQ_MIGRATING 2
88
89extern __read_mostly int scheduler_running;
90
91extern unsigned long calc_load_update;
92extern atomic_long_t calc_load_tasks;
93
94extern void calc_global_load_tick(struct rq *this_rq);
95extern long calc_load_fold_active(struct rq *this_rq, long adjust);
96
97#ifdef CONFIG_SMP
98extern void cpu_load_update_active(struct rq *this_rq);
99#else
100static inline void cpu_load_update_active(struct rq *this_rq) { }
101#endif
102
103/*
104 * Helpers for converting nanosecond timing to jiffy resolution
105 */
106#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107
108/*
109 * Increase resolution of nice-level calculations for 64-bit architectures.
110 * The extra resolution improves shares distribution and load balancing of
111 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
112 * hierarchies, especially on larger systems. This is not a user-visible change
113 * and does not change the user-interface for setting shares/weights.
114 *
115 * We increase resolution only if we have enough bits to allow this increased
116 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
117 * are pretty high and the returns do not justify the increased costs.
118 *
119 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
120 * increase coverage and consistency always enable it on 64-bit platforms.
121 */
122#ifdef CONFIG_64BIT
123# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
124# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
125# define scale_load_down(w) ((w) >> SCHED_FIXEDPOINT_SHIFT)
126#else
127# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
128# define scale_load(w) (w)
129# define scale_load_down(w) (w)
130#endif
131
132/*
133 * Task weight (visible to users) and its load (invisible to users) have
134 * independent resolution, but they should be well calibrated. We use
135 * scale_load() and scale_load_down(w) to convert between them. The
136 * following must be true:
137 *
138 * scale_load(sched_prio_to_weight[USER_PRIO(NICE_TO_PRIO(0))]) == NICE_0_LOAD
139 *
140 */
141#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
142
143/*
144 * Single value that decides SCHED_DEADLINE internal math precision.
145 * 10 -> just above 1us
146 * 9 -> just above 0.5us
147 */
148#define DL_SCALE 10
149
150/*
151 * Single value that denotes runtime == period, ie unlimited time.
152 */
153#define RUNTIME_INF ((u64)~0ULL)
154
155static inline int idle_policy(int policy)
156{
157 return policy == SCHED_IDLE;
158}
159static inline int fair_policy(int policy)
160{
161 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
162}
163
164static inline int rt_policy(int policy)
165{
166 return policy == SCHED_FIFO || policy == SCHED_RR;
167}
168
169static inline int dl_policy(int policy)
170{
171 return policy == SCHED_DEADLINE;
172}
173static inline bool valid_policy(int policy)
174{
175 return idle_policy(policy) || fair_policy(policy) ||
176 rt_policy(policy) || dl_policy(policy);
177}
178
179static inline int task_has_rt_policy(struct task_struct *p)
180{
181 return rt_policy(p->policy);
182}
183
184static inline int task_has_dl_policy(struct task_struct *p)
185{
186 return dl_policy(p->policy);
187}
188
189#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
190
191/*
192 * !! For sched_setattr_nocheck() (kernel) only !!
193 *
194 * This is actually gross. :(
195 *
196 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
197 * tasks, but still be able to sleep. We need this on platforms that cannot
198 * atomically change clock frequency. Remove once fast switching will be
199 * available on such platforms.
200 *
201 * SUGOV stands for SchedUtil GOVernor.
202 */
203#define SCHED_FLAG_SUGOV 0x10000000
204
205static inline bool dl_entity_is_special(struct sched_dl_entity *dl_se)
206{
207#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
208 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
209#else
210 return false;
211#endif
212}
213
214/*
215 * Tells if entity @a should preempt entity @b.
216 */
217static inline bool
218dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
219{
220 return dl_entity_is_special(a) ||
221 dl_time_before(a->deadline, b->deadline);
222}
223
224/*
225 * This is the priority-queue data structure of the RT scheduling class:
226 */
227struct rt_prio_array {
228 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
229 struct list_head queue[MAX_RT_PRIO];
230};
231
232struct rt_bandwidth {
233 /* nests inside the rq lock: */
234 raw_spinlock_t rt_runtime_lock;
235 ktime_t rt_period;
236 u64 rt_runtime;
237 struct hrtimer rt_period_timer;
238 unsigned int rt_period_active;
239};
240
241void __dl_clear_params(struct task_struct *p);
242
243/*
244 * To keep the bandwidth of -deadline tasks and groups under control
245 * we need some place where:
246 * - store the maximum -deadline bandwidth of the system (the group);
247 * - cache the fraction of that bandwidth that is currently allocated.
248 *
249 * This is all done in the data structure below. It is similar to the
250 * one used for RT-throttling (rt_bandwidth), with the main difference
251 * that, since here we are only interested in admission control, we
252 * do not decrease any runtime while the group "executes", neither we
253 * need a timer to replenish it.
254 *
255 * With respect to SMP, the bandwidth is given on a per-CPU basis,
256 * meaning that:
257 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
258 * - dl_total_bw array contains, in the i-eth element, the currently
259 * allocated bandwidth on the i-eth CPU.
260 * Moreover, groups consume bandwidth on each CPU, while tasks only
261 * consume bandwidth on the CPU they're running on.
262 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
263 * that will be shown the next time the proc or cgroup controls will
264 * be red. It on its turn can be changed by writing on its own
265 * control.
266 */
267struct dl_bandwidth {
268 raw_spinlock_t dl_runtime_lock;
269 u64 dl_runtime;
270 u64 dl_period;
271};
272
273static inline int dl_bandwidth_enabled(void)
274{
275 return sysctl_sched_rt_runtime >= 0;
276}
277
278struct dl_bw {
279 raw_spinlock_t lock;
280 u64 bw;
281 u64 total_bw;
282};
283
284static inline void __dl_update(struct dl_bw *dl_b, s64 bw);
285
286static inline
287void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
288{
289 dl_b->total_bw -= tsk_bw;
290 __dl_update(dl_b, (s32)tsk_bw / cpus);
291}
292
293static inline
294void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
295{
296 dl_b->total_bw += tsk_bw;
297 __dl_update(dl_b, -((s32)tsk_bw / cpus));
298}
299
300static inline
301bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
302{
303 return dl_b->bw != -1 &&
304 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
305}
306
307extern void dl_change_utilization(struct task_struct *p, u64 new_bw);
308extern void init_dl_bw(struct dl_bw *dl_b);
309extern int sched_dl_global_validate(void);
310extern void sched_dl_do_global(void);
311extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
312extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
313extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
314extern bool __checkparam_dl(const struct sched_attr *attr);
315extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
316extern int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed);
317extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
318extern bool dl_cpu_busy(unsigned int cpu);
319
320#ifdef CONFIG_CGROUP_SCHED
321
322#include <linux/cgroup.h>
323
324struct cfs_rq;
325struct rt_rq;
326
327extern struct list_head task_groups;
328
329struct cfs_bandwidth {
330#ifdef CONFIG_CFS_BANDWIDTH
331 raw_spinlock_t lock;
332 ktime_t period;
333 u64 quota;
334 u64 runtime;
335 s64 hierarchical_quota;
336 u64 runtime_expires;
337
338 int idle;
339 int period_active;
340 struct hrtimer period_timer;
341 struct hrtimer slack_timer;
342 struct list_head throttled_cfs_rq;
343
344 /* Statistics: */
345 int nr_periods;
346 int nr_throttled;
347 u64 throttled_time;
348#endif
349};
350
351/* Task group related information */
352struct task_group {
353 struct cgroup_subsys_state css;
354
355#ifdef CONFIG_FAIR_GROUP_SCHED
356 /* schedulable entities of this group on each CPU */
357 struct sched_entity **se;
358 /* runqueue "owned" by this group on each CPU */
359 struct cfs_rq **cfs_rq;
360 unsigned long shares;
361
362#ifdef CONFIG_SMP
363 /*
364 * load_avg can be heavily contended at clock tick time, so put
365 * it in its own cacheline separated from the fields above which
366 * will also be accessed at each tick.
367 */
368 atomic_long_t load_avg ____cacheline_aligned;
369#endif
370#endif
371
372#ifdef CONFIG_RT_GROUP_SCHED
373 struct sched_rt_entity **rt_se;
374 struct rt_rq **rt_rq;
375
376 struct rt_bandwidth rt_bandwidth;
377#endif
378
379 struct rcu_head rcu;
380 struct list_head list;
381
382 struct task_group *parent;
383 struct list_head siblings;
384 struct list_head children;
385
386#ifdef CONFIG_SCHED_AUTOGROUP
387 struct autogroup *autogroup;
388#endif
389
390 struct cfs_bandwidth cfs_bandwidth;
391};
392
393#ifdef CONFIG_FAIR_GROUP_SCHED
394#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
395
396/*
397 * A weight of 0 or 1 can cause arithmetics problems.
398 * A weight of a cfs_rq is the sum of weights of which entities
399 * are queued on this cfs_rq, so a weight of a entity should not be
400 * too large, so as the shares value of a task group.
401 * (The default weight is 1024 - so there's no practical
402 * limitation from this.)
403 */
404#define MIN_SHARES (1UL << 1)
405#define MAX_SHARES (1UL << 18)
406#endif
407
408typedef int (*tg_visitor)(struct task_group *, void *);
409
410extern int walk_tg_tree_from(struct task_group *from,
411 tg_visitor down, tg_visitor up, void *data);
412
413/*
414 * Iterate the full tree, calling @down when first entering a node and @up when
415 * leaving it for the final time.
416 *
417 * Caller must hold rcu_lock or sufficient equivalent.
418 */
419static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
420{
421 return walk_tg_tree_from(&root_task_group, down, up, data);
422}
423
424extern int tg_nop(struct task_group *tg, void *data);
425
426extern void free_fair_sched_group(struct task_group *tg);
427extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
428extern void online_fair_sched_group(struct task_group *tg);
429extern void unregister_fair_sched_group(struct task_group *tg);
430extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
431 struct sched_entity *se, int cpu,
432 struct sched_entity *parent);
433extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
434
435extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
436extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
437extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
438
439extern void free_rt_sched_group(struct task_group *tg);
440extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
441extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
442 struct sched_rt_entity *rt_se, int cpu,
443 struct sched_rt_entity *parent);
444extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
445extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
446extern long sched_group_rt_runtime(struct task_group *tg);
447extern long sched_group_rt_period(struct task_group *tg);
448extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
449
450extern struct task_group *sched_create_group(struct task_group *parent);
451extern void sched_online_group(struct task_group *tg,
452 struct task_group *parent);
453extern void sched_destroy_group(struct task_group *tg);
454extern void sched_offline_group(struct task_group *tg);
455
456extern void sched_move_task(struct task_struct *tsk);
457
458#ifdef CONFIG_FAIR_GROUP_SCHED
459extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
460
461#ifdef CONFIG_SMP
462extern void set_task_rq_fair(struct sched_entity *se,
463 struct cfs_rq *prev, struct cfs_rq *next);
464#else /* !CONFIG_SMP */
465static inline void set_task_rq_fair(struct sched_entity *se,
466 struct cfs_rq *prev, struct cfs_rq *next) { }
467#endif /* CONFIG_SMP */
468#endif /* CONFIG_FAIR_GROUP_SCHED */
469
470#else /* CONFIG_CGROUP_SCHED */
471
472struct cfs_bandwidth { };
473
474#endif /* CONFIG_CGROUP_SCHED */
475
476/* CFS-related fields in a runqueue */
477struct cfs_rq {
478 struct load_weight load;
479 unsigned long runnable_weight;
480 unsigned int nr_running;
481 unsigned int h_nr_running;
482
483 u64 exec_clock;
484 u64 min_vruntime;
485#ifndef CONFIG_64BIT
486 u64 min_vruntime_copy;
487#endif
488
489 struct rb_root_cached tasks_timeline;
490
491 /*
492 * 'curr' points to currently running entity on this cfs_rq.
493 * It is set to NULL otherwise (i.e when none are currently running).
494 */
495 struct sched_entity *curr;
496 struct sched_entity *next;
497 struct sched_entity *last;
498 struct sched_entity *skip;
499
500#ifdef CONFIG_SCHED_DEBUG
501 unsigned int nr_spread_over;
502#endif
503
504#ifdef CONFIG_SMP
505 /*
506 * CFS load tracking
507 */
508 struct sched_avg avg;
509#ifndef CONFIG_64BIT
510 u64 load_last_update_time_copy;
511#endif
512 struct {
513 raw_spinlock_t lock ____cacheline_aligned;
514 int nr;
515 unsigned long load_avg;
516 unsigned long util_avg;
517 unsigned long runnable_sum;
518 } removed;
519
520#ifdef CONFIG_FAIR_GROUP_SCHED
521 unsigned long tg_load_avg_contrib;
522 long propagate;
523 long prop_runnable_sum;
524
525 /*
526 * h_load = weight * f(tg)
527 *
528 * Where f(tg) is the recursive weight fraction assigned to
529 * this group.
530 */
531 unsigned long h_load;
532 u64 last_h_load_update;
533 struct sched_entity *h_load_next;
534#endif /* CONFIG_FAIR_GROUP_SCHED */
535#endif /* CONFIG_SMP */
536
537#ifdef CONFIG_FAIR_GROUP_SCHED
538 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
539
540 /*
541 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
542 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
543 * (like users, containers etc.)
544 *
545 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
546 * This list is used during load balance.
547 */
548 int on_list;
549 struct list_head leaf_cfs_rq_list;
550 struct task_group *tg; /* group that "owns" this runqueue */
551
552#ifdef CONFIG_CFS_BANDWIDTH
553 int runtime_enabled;
554 u64 runtime_expires;
555 s64 runtime_remaining;
556
557 u64 throttled_clock;
558 u64 throttled_clock_task;
559 u64 throttled_clock_task_time;
560 int throttled;
561 int throttle_count;
562 struct list_head throttled_list;
563#endif /* CONFIG_CFS_BANDWIDTH */
564#endif /* CONFIG_FAIR_GROUP_SCHED */
565};
566
567static inline int rt_bandwidth_enabled(void)
568{
569 return sysctl_sched_rt_runtime >= 0;
570}
571
572/* RT IPI pull logic requires IRQ_WORK */
573#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
574# define HAVE_RT_PUSH_IPI
575#endif
576
577/* Real-Time classes' related field in a runqueue: */
578struct rt_rq {
579 struct rt_prio_array active;
580 unsigned int rt_nr_running;
581 unsigned int rr_nr_running;
582#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
583 struct {
584 int curr; /* highest queued rt task prio */
585#ifdef CONFIG_SMP
586 int next; /* next highest */
587#endif
588 } highest_prio;
589#endif
590#ifdef CONFIG_SMP
591 unsigned long rt_nr_migratory;
592 unsigned long rt_nr_total;
593 int overloaded;
594 struct plist_head pushable_tasks;
595#endif /* CONFIG_SMP */
596 int rt_queued;
597
598 int rt_throttled;
599 u64 rt_time;
600 u64 rt_runtime;
601 /* Nests inside the rq lock: */
602 raw_spinlock_t rt_runtime_lock;
603
604#ifdef CONFIG_RT_GROUP_SCHED
605 unsigned long rt_nr_boosted;
606
607 struct rq *rq;
608 struct task_group *tg;
609#endif
610};
611
612/* Deadline class' related fields in a runqueue */
613struct dl_rq {
614 /* runqueue is an rbtree, ordered by deadline */
615 struct rb_root_cached root;
616
617 unsigned long dl_nr_running;
618
619#ifdef CONFIG_SMP
620 /*
621 * Deadline values of the currently executing and the
622 * earliest ready task on this rq. Caching these facilitates
623 * the decision wether or not a ready but not running task
624 * should migrate somewhere else.
625 */
626 struct {
627 u64 curr;
628 u64 next;
629 } earliest_dl;
630
631 unsigned long dl_nr_migratory;
632 int overloaded;
633
634 /*
635 * Tasks on this rq that can be pushed away. They are kept in
636 * an rb-tree, ordered by tasks' deadlines, with caching
637 * of the leftmost (earliest deadline) element.
638 */
639 struct rb_root_cached pushable_dl_tasks_root;
640#else
641 struct dl_bw dl_bw;
642#endif
643 /*
644 * "Active utilization" for this runqueue: increased when a
645 * task wakes up (becomes TASK_RUNNING) and decreased when a
646 * task blocks
647 */
648 u64 running_bw;
649
650 /*
651 * Utilization of the tasks "assigned" to this runqueue (including
652 * the tasks that are in runqueue and the tasks that executed on this
653 * CPU and blocked). Increased when a task moves to this runqueue, and
654 * decreased when the task moves away (migrates, changes scheduling
655 * policy, or terminates).
656 * This is needed to compute the "inactive utilization" for the
657 * runqueue (inactive utilization = this_bw - running_bw).
658 */
659 u64 this_bw;
660 u64 extra_bw;
661
662 /*
663 * Inverse of the fraction of CPU utilization that can be reclaimed
664 * by the GRUB algorithm.
665 */
666 u64 bw_ratio;
667};
668
669#ifdef CONFIG_SMP
670
671static inline bool sched_asym_prefer(int a, int b)
672{
673 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
674}
675
676/*
677 * We add the notion of a root-domain which will be used to define per-domain
678 * variables. Each exclusive cpuset essentially defines an island domain by
679 * fully partitioning the member CPUs from any other cpuset. Whenever a new
680 * exclusive cpuset is created, we also create and attach a new root-domain
681 * object.
682 *
683 */
684struct root_domain {
685 atomic_t refcount;
686 atomic_t rto_count;
687 struct rcu_head rcu;
688 cpumask_var_t span;
689 cpumask_var_t online;
690
691 /* Indicate more than one runnable task for any CPU */
692 bool overload;
693
694 /*
695 * The bit corresponding to a CPU gets set here if such CPU has more
696 * than one runnable -deadline task (as it is below for RT tasks).
697 */
698 cpumask_var_t dlo_mask;
699 atomic_t dlo_count;
700 struct dl_bw dl_bw;
701 struct cpudl cpudl;
702
703#ifdef HAVE_RT_PUSH_IPI
704 /*
705 * For IPI pull requests, loop across the rto_mask.
706 */
707 struct irq_work rto_push_work;
708 raw_spinlock_t rto_lock;
709 /* These are only updated and read within rto_lock */
710 int rto_loop;
711 int rto_cpu;
712 /* These atomics are updated outside of a lock */
713 atomic_t rto_loop_next;
714 atomic_t rto_loop_start;
715#endif
716 /*
717 * The "RT overload" flag: it gets set if a CPU has more than
718 * one runnable RT task.
719 */
720 cpumask_var_t rto_mask;
721 struct cpupri cpupri;
722
723 unsigned long max_cpu_capacity;
724};
725
726extern struct root_domain def_root_domain;
727extern struct mutex sched_domains_mutex;
728
729extern void init_defrootdomain(void);
730extern int sched_init_domains(const struct cpumask *cpu_map);
731extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
732extern void sched_get_rd(struct root_domain *rd);
733extern void sched_put_rd(struct root_domain *rd);
734
735#ifdef HAVE_RT_PUSH_IPI
736extern void rto_push_irq_work_func(struct irq_work *work);
737#endif
738#endif /* CONFIG_SMP */
739
740/*
741 * This is the main, per-CPU runqueue data structure.
742 *
743 * Locking rule: those places that want to lock multiple runqueues
744 * (such as the load balancing or the thread migration code), lock
745 * acquire operations must be ordered by ascending &runqueue.
746 */
747struct rq {
748 /* runqueue lock: */
749 raw_spinlock_t lock;
750
751 /*
752 * nr_running and cpu_load should be in the same cacheline because
753 * remote CPUs use both these fields when doing load calculation.
754 */
755 unsigned int nr_running;
756#ifdef CONFIG_NUMA_BALANCING
757 unsigned int nr_numa_running;
758 unsigned int nr_preferred_running;
759#endif
760 #define CPU_LOAD_IDX_MAX 5
761 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
762#ifdef CONFIG_NO_HZ_COMMON
763#ifdef CONFIG_SMP
764 unsigned long last_load_update_tick;
765 unsigned long last_blocked_load_update_tick;
766 unsigned int has_blocked_load;
767#endif /* CONFIG_SMP */
768 unsigned int nohz_tick_stopped;
769 atomic_t nohz_flags;
770#endif /* CONFIG_NO_HZ_COMMON */
771
772 /* capture load from *all* tasks on this CPU: */
773 struct load_weight load;
774 unsigned long nr_load_updates;
775 u64 nr_switches;
776
777 struct cfs_rq cfs;
778 struct rt_rq rt;
779 struct dl_rq dl;
780
781#ifdef CONFIG_FAIR_GROUP_SCHED
782 /* list of leaf cfs_rq on this CPU: */
783 struct list_head leaf_cfs_rq_list;
784 struct list_head *tmp_alone_branch;
785#endif /* CONFIG_FAIR_GROUP_SCHED */
786
787 /*
788 * This is part of a global counter where only the total sum
789 * over all CPUs matters. A task can increase this counter on
790 * one CPU and if it got migrated afterwards it may decrease
791 * it on another CPU. Always updated under the runqueue lock:
792 */
793 unsigned long nr_uninterruptible;
794
795 struct task_struct *curr;
796 struct task_struct *idle;
797 struct task_struct *stop;
798 unsigned long next_balance;
799 struct mm_struct *prev_mm;
800
801 unsigned int clock_update_flags;
802 u64 clock;
803 u64 clock_task;
804
805 atomic_t nr_iowait;
806
807#ifdef CONFIG_SMP
808 struct root_domain *rd;
809 struct sched_domain *sd;
810
811 unsigned long cpu_capacity;
812 unsigned long cpu_capacity_orig;
813
814 struct callback_head *balance_callback;
815
816 unsigned char idle_balance;
817
818 /* For active balancing */
819 int active_balance;
820 int push_cpu;
821 struct cpu_stop_work active_balance_work;
822
823 /* CPU of this runqueue: */
824 int cpu;
825 int online;
826
827 struct list_head cfs_tasks;
828
829 u64 rt_avg;
830 u64 age_stamp;
831 u64 idle_stamp;
832 u64 avg_idle;
833
834 /* This is used to determine avg_idle's max value */
835 u64 max_idle_balance_cost;
836#endif
837
838#ifdef CONFIG_IRQ_TIME_ACCOUNTING
839 u64 prev_irq_time;
840#endif
841#ifdef CONFIG_PARAVIRT
842 u64 prev_steal_time;
843#endif
844#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
845 u64 prev_steal_time_rq;
846#endif
847
848 /* calc_load related fields */
849 unsigned long calc_load_update;
850 long calc_load_active;
851
852#ifdef CONFIG_SCHED_HRTICK
853#ifdef CONFIG_SMP
854 int hrtick_csd_pending;
855 call_single_data_t hrtick_csd;
856#endif
857 struct hrtimer hrtick_timer;
858#endif
859
860#ifdef CONFIG_SCHEDSTATS
861 /* latency stats */
862 struct sched_info rq_sched_info;
863 unsigned long long rq_cpu_time;
864 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
865
866 /* sys_sched_yield() stats */
867 unsigned int yld_count;
868
869 /* schedule() stats */
870 unsigned int sched_count;
871 unsigned int sched_goidle;
872
873 /* try_to_wake_up() stats */
874 unsigned int ttwu_count;
875 unsigned int ttwu_local;
876#endif
877
878#ifdef CONFIG_SMP
879 struct llist_head wake_list;
880#endif
881
882#ifdef CONFIG_CPU_IDLE
883 /* Must be inspected within a rcu lock section */
884 struct cpuidle_state *idle_state;
885#endif
886};
887
888static inline int cpu_of(struct rq *rq)
889{
890#ifdef CONFIG_SMP
891 return rq->cpu;
892#else
893 return 0;
894#endif
895}
896
897
898#ifdef CONFIG_SCHED_SMT
899
900extern struct static_key_false sched_smt_present;
901
902extern void __update_idle_core(struct rq *rq);
903
904static inline void update_idle_core(struct rq *rq)
905{
906 if (static_branch_unlikely(&sched_smt_present))
907 __update_idle_core(rq);
908}
909
910#else
911static inline void update_idle_core(struct rq *rq) { }
912#endif
913
914DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
915
916#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
917#define this_rq() this_cpu_ptr(&runqueues)
918#define task_rq(p) cpu_rq(task_cpu(p))
919#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
920#define raw_rq() raw_cpu_ptr(&runqueues)
921
922static inline u64 __rq_clock_broken(struct rq *rq)
923{
924 return READ_ONCE(rq->clock);
925}
926
927/*
928 * rq::clock_update_flags bits
929 *
930 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
931 * call to __schedule(). This is an optimisation to avoid
932 * neighbouring rq clock updates.
933 *
934 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
935 * in effect and calls to update_rq_clock() are being ignored.
936 *
937 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
938 * made to update_rq_clock() since the last time rq::lock was pinned.
939 *
940 * If inside of __schedule(), clock_update_flags will have been
941 * shifted left (a left shift is a cheap operation for the fast path
942 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
943 *
944 * if (rq-clock_update_flags >= RQCF_UPDATED)
945 *
946 * to check if %RQCF_UPADTED is set. It'll never be shifted more than
947 * one position though, because the next rq_unpin_lock() will shift it
948 * back.
949 */
950#define RQCF_REQ_SKIP 0x01
951#define RQCF_ACT_SKIP 0x02
952#define RQCF_UPDATED 0x04
953
954static inline void assert_clock_updated(struct rq *rq)
955{
956 /*
957 * The only reason for not seeing a clock update since the
958 * last rq_pin_lock() is if we're currently skipping updates.
959 */
960 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
961}
962
963static inline u64 rq_clock(struct rq *rq)
964{
965 lockdep_assert_held(&rq->lock);
966 assert_clock_updated(rq);
967
968 return rq->clock;
969}
970
971static inline u64 rq_clock_task(struct rq *rq)
972{
973 lockdep_assert_held(&rq->lock);
974 assert_clock_updated(rq);
975
976 return rq->clock_task;
977}
978
979static inline void rq_clock_skip_update(struct rq *rq)
980{
981 lockdep_assert_held(&rq->lock);
982 rq->clock_update_flags |= RQCF_REQ_SKIP;
983}
984
985/*
986 * See rt task throttling, which is the only time a skip
987 * request is cancelled.
988 */
989static inline void rq_clock_cancel_skipupdate(struct rq *rq)
990{
991 lockdep_assert_held(&rq->lock);
992 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
993}
994
995struct rq_flags {
996 unsigned long flags;
997 struct pin_cookie cookie;
998#ifdef CONFIG_SCHED_DEBUG
999 /*
1000 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1001 * current pin context is stashed here in case it needs to be
1002 * restored in rq_repin_lock().
1003 */
1004 unsigned int clock_update_flags;
1005#endif
1006};
1007
1008static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1009{
1010 rf->cookie = lockdep_pin_lock(&rq->lock);
1011
1012#ifdef CONFIG_SCHED_DEBUG
1013 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1014 rf->clock_update_flags = 0;
1015#endif
1016}
1017
1018static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1019{
1020#ifdef CONFIG_SCHED_DEBUG
1021 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1022 rf->clock_update_flags = RQCF_UPDATED;
1023#endif
1024
1025 lockdep_unpin_lock(&rq->lock, rf->cookie);
1026}
1027
1028static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1029{
1030 lockdep_repin_lock(&rq->lock, rf->cookie);
1031
1032#ifdef CONFIG_SCHED_DEBUG
1033 /*
1034 * Restore the value we stashed in @rf for this pin context.
1035 */
1036 rq->clock_update_flags |= rf->clock_update_flags;
1037#endif
1038}
1039
1040#ifdef CONFIG_NUMA
1041enum numa_topology_type {
1042 NUMA_DIRECT,
1043 NUMA_GLUELESS_MESH,
1044 NUMA_BACKPLANE,
1045};
1046extern enum numa_topology_type sched_numa_topology_type;
1047extern int sched_max_numa_distance;
1048extern bool find_numa_distance(int distance);
1049#endif
1050
1051#ifdef CONFIG_NUMA
1052extern void sched_init_numa(void);
1053extern void sched_domains_numa_masks_set(unsigned int cpu);
1054extern void sched_domains_numa_masks_clear(unsigned int cpu);
1055#else
1056static inline void sched_init_numa(void) { }
1057static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1058static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1059#endif
1060
1061#ifdef CONFIG_NUMA_BALANCING
1062/* The regions in numa_faults array from task_struct */
1063enum numa_faults_stats {
1064 NUMA_MEM = 0,
1065 NUMA_CPU,
1066 NUMA_MEMBUF,
1067 NUMA_CPUBUF
1068};
1069extern void sched_setnuma(struct task_struct *p, int node);
1070extern int migrate_task_to(struct task_struct *p, int cpu);
1071extern int migrate_swap(struct task_struct *, struct task_struct *);
1072#endif /* CONFIG_NUMA_BALANCING */
1073
1074#ifdef CONFIG_SMP
1075
1076static inline void
1077queue_balance_callback(struct rq *rq,
1078 struct callback_head *head,
1079 void (*func)(struct rq *rq))
1080{
1081 lockdep_assert_held(&rq->lock);
1082
1083 if (unlikely(head->next))
1084 return;
1085
1086 head->func = (void (*)(struct callback_head *))func;
1087 head->next = rq->balance_callback;
1088 rq->balance_callback = head;
1089}
1090
1091extern void sched_ttwu_pending(void);
1092
1093#define rcu_dereference_check_sched_domain(p) \
1094 rcu_dereference_check((p), \
1095 lockdep_is_held(&sched_domains_mutex))
1096
1097/*
1098 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1099 * See detach_destroy_domains: synchronize_sched for details.
1100 *
1101 * The domain tree of any CPU may only be accessed from within
1102 * preempt-disabled sections.
1103 */
1104#define for_each_domain(cpu, __sd) \
1105 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1106 __sd; __sd = __sd->parent)
1107
1108#define for_each_lower_domain(sd) for (; sd; sd = sd->child)
1109
1110/**
1111 * highest_flag_domain - Return highest sched_domain containing flag.
1112 * @cpu: The CPU whose highest level of sched domain is to
1113 * be returned.
1114 * @flag: The flag to check for the highest sched_domain
1115 * for the given CPU.
1116 *
1117 * Returns the highest sched_domain of a CPU which contains the given flag.
1118 */
1119static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1120{
1121 struct sched_domain *sd, *hsd = NULL;
1122
1123 for_each_domain(cpu, sd) {
1124 if (!(sd->flags & flag))
1125 break;
1126 hsd = sd;
1127 }
1128
1129 return hsd;
1130}
1131
1132static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1133{
1134 struct sched_domain *sd;
1135
1136 for_each_domain(cpu, sd) {
1137 if (sd->flags & flag)
1138 break;
1139 }
1140
1141 return sd;
1142}
1143
1144DECLARE_PER_CPU(struct sched_domain *, sd_llc);
1145DECLARE_PER_CPU(int, sd_llc_size);
1146DECLARE_PER_CPU(int, sd_llc_id);
1147DECLARE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
1148DECLARE_PER_CPU(struct sched_domain *, sd_numa);
1149DECLARE_PER_CPU(struct sched_domain *, sd_asym);
1150
1151struct sched_group_capacity {
1152 atomic_t ref;
1153 /*
1154 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1155 * for a single CPU.
1156 */
1157 unsigned long capacity;
1158 unsigned long min_capacity; /* Min per-CPU capacity in group */
1159 unsigned long next_update;
1160 int imbalance; /* XXX unrelated to capacity but shared group state */
1161
1162#ifdef CONFIG_SCHED_DEBUG
1163 int id;
1164#endif
1165
1166 unsigned long cpumask[0]; /* Balance mask */
1167};
1168
1169struct sched_group {
1170 struct sched_group *next; /* Must be a circular list */
1171 atomic_t ref;
1172
1173 unsigned int group_weight;
1174 struct sched_group_capacity *sgc;
1175 int asym_prefer_cpu; /* CPU of highest priority in group */
1176
1177 /*
1178 * The CPUs this group covers.
1179 *
1180 * NOTE: this field is variable length. (Allocated dynamically
1181 * by attaching extra space to the end of the structure,
1182 * depending on how many CPUs the kernel has booted up with)
1183 */
1184 unsigned long cpumask[0];
1185};
1186
1187static inline struct cpumask *sched_group_span(struct sched_group *sg)
1188{
1189 return to_cpumask(sg->cpumask);
1190}
1191
1192/*
1193 * See build_balance_mask().
1194 */
1195static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1196{
1197 return to_cpumask(sg->sgc->cpumask);
1198}
1199
1200/**
1201 * group_first_cpu - Returns the first CPU in the cpumask of a sched_group.
1202 * @group: The group whose first CPU is to be returned.
1203 */
1204static inline unsigned int group_first_cpu(struct sched_group *group)
1205{
1206 return cpumask_first(sched_group_span(group));
1207}
1208
1209extern int group_balance_cpu(struct sched_group *sg);
1210
1211#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
1212void register_sched_domain_sysctl(void);
1213void dirty_sched_domain_sysctl(int cpu);
1214void unregister_sched_domain_sysctl(void);
1215#else
1216static inline void register_sched_domain_sysctl(void)
1217{
1218}
1219static inline void dirty_sched_domain_sysctl(int cpu)
1220{
1221}
1222static inline void unregister_sched_domain_sysctl(void)
1223{
1224}
1225#endif
1226
1227#else
1228
1229static inline void sched_ttwu_pending(void) { }
1230
1231#endif /* CONFIG_SMP */
1232
1233#include "stats.h"
1234#include "autogroup.h"
1235
1236#ifdef CONFIG_CGROUP_SCHED
1237
1238/*
1239 * Return the group to which this tasks belongs.
1240 *
1241 * We cannot use task_css() and friends because the cgroup subsystem
1242 * changes that value before the cgroup_subsys::attach() method is called,
1243 * therefore we cannot pin it and might observe the wrong value.
1244 *
1245 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
1246 * core changes this before calling sched_move_task().
1247 *
1248 * Instead we use a 'copy' which is updated from sched_move_task() while
1249 * holding both task_struct::pi_lock and rq::lock.
1250 */
1251static inline struct task_group *task_group(struct task_struct *p)
1252{
1253 return p->sched_task_group;
1254}
1255
1256/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
1257static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
1258{
1259#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
1260 struct task_group *tg = task_group(p);
1261#endif
1262
1263#ifdef CONFIG_FAIR_GROUP_SCHED
1264 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
1265 p->se.cfs_rq = tg->cfs_rq[cpu];
1266 p->se.parent = tg->se[cpu];
1267#endif
1268
1269#ifdef CONFIG_RT_GROUP_SCHED
1270 p->rt.rt_rq = tg->rt_rq[cpu];
1271 p->rt.parent = tg->rt_se[cpu];
1272#endif
1273}
1274
1275#else /* CONFIG_CGROUP_SCHED */
1276
1277static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
1278static inline struct task_group *task_group(struct task_struct *p)
1279{
1280 return NULL;
1281}
1282
1283#endif /* CONFIG_CGROUP_SCHED */
1284
1285static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1286{
1287 set_task_rq(p, cpu);
1288#ifdef CONFIG_SMP
1289 /*
1290 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1291 * successfuly executed on another CPU. We must ensure that updates of
1292 * per-task data have been completed by this moment.
1293 */
1294 smp_wmb();
1295#ifdef CONFIG_THREAD_INFO_IN_TASK
1296 p->cpu = cpu;
1297#else
1298 task_thread_info(p)->cpu = cpu;
1299#endif
1300 p->wake_cpu = cpu;
1301#endif
1302}
1303
1304/*
1305 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
1306 */
1307#ifdef CONFIG_SCHED_DEBUG
1308# include <linux/static_key.h>
1309# define const_debug __read_mostly
1310#else
1311# define const_debug const
1312#endif
1313
1314#define SCHED_FEAT(name, enabled) \
1315 __SCHED_FEAT_##name ,
1316
1317enum {
1318#include "features.h"
1319 __SCHED_FEAT_NR,
1320};
1321
1322#undef SCHED_FEAT
1323
1324#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
1325
1326/*
1327 * To support run-time toggling of sched features, all the translation units
1328 * (but core.c) reference the sysctl_sched_features defined in core.c.
1329 */
1330extern const_debug unsigned int sysctl_sched_features;
1331
1332#define SCHED_FEAT(name, enabled) \
1333static __always_inline bool static_branch_##name(struct static_key *key) \
1334{ \
1335 return static_key_##enabled(key); \
1336}
1337
1338#include "features.h"
1339#undef SCHED_FEAT
1340
1341extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
1342#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
1343
1344#else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
1345
1346/*
1347 * Each translation unit has its own copy of sysctl_sched_features to allow
1348 * constants propagation at compile time and compiler optimization based on
1349 * features default.
1350 */
1351#define SCHED_FEAT(name, enabled) \
1352 (1UL << __SCHED_FEAT_##name) * enabled |
1353static const_debug __maybe_unused unsigned int sysctl_sched_features =
1354#include "features.h"
1355 0;
1356#undef SCHED_FEAT
1357
1358#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
1359
1360#endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
1361
1362extern struct static_key_false sched_numa_balancing;
1363extern struct static_key_false sched_schedstats;
1364
1365static inline u64 global_rt_period(void)
1366{
1367 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
1368}
1369
1370static inline u64 global_rt_runtime(void)
1371{
1372 if (sysctl_sched_rt_runtime < 0)
1373 return RUNTIME_INF;
1374
1375 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
1376}
1377
1378static inline int task_current(struct rq *rq, struct task_struct *p)
1379{
1380 return rq->curr == p;
1381}
1382
1383static inline int task_running(struct rq *rq, struct task_struct *p)
1384{
1385#ifdef CONFIG_SMP
1386 return p->on_cpu;
1387#else
1388 return task_current(rq, p);
1389#endif
1390}
1391
1392static inline int task_on_rq_queued(struct task_struct *p)
1393{
1394 return p->on_rq == TASK_ON_RQ_QUEUED;
1395}
1396
1397static inline int task_on_rq_migrating(struct task_struct *p)
1398{
1399 return p->on_rq == TASK_ON_RQ_MIGRATING;
1400}
1401
1402/*
1403 * wake flags
1404 */
1405#define WF_SYNC 0x01 /* Waker goes to sleep after wakeup */
1406#define WF_FORK 0x02 /* Child wakeup after fork */
1407#define WF_MIGRATED 0x4 /* Internal use, task got migrated */
1408
1409/*
1410 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1411 * of tasks with abnormal "nice" values across CPUs the contribution that
1412 * each task makes to its run queue's load is weighted according to its
1413 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1414 * scaled version of the new time slice allocation that they receive on time
1415 * slice expiry etc.
1416 */
1417
1418#define WEIGHT_IDLEPRIO 3
1419#define WMULT_IDLEPRIO 1431655765
1420
1421extern const int sched_prio_to_weight[40];
1422extern const u32 sched_prio_to_wmult[40];
1423
1424/*
1425 * {de,en}queue flags:
1426 *
1427 * DEQUEUE_SLEEP - task is no longer runnable
1428 * ENQUEUE_WAKEUP - task just became runnable
1429 *
1430 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
1431 * are in a known state which allows modification. Such pairs
1432 * should preserve as much state as possible.
1433 *
1434 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
1435 * in the runqueue.
1436 *
1437 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
1438 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
1439 * ENQUEUE_MIGRATED - the task was migrated during wakeup
1440 *
1441 */
1442
1443#define DEQUEUE_SLEEP 0x01
1444#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
1445#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
1446#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
1447
1448#define ENQUEUE_WAKEUP 0x01
1449#define ENQUEUE_RESTORE 0x02
1450#define ENQUEUE_MOVE 0x04
1451#define ENQUEUE_NOCLOCK 0x08
1452
1453#define ENQUEUE_HEAD 0x10
1454#define ENQUEUE_REPLENISH 0x20
1455#ifdef CONFIG_SMP
1456#define ENQUEUE_MIGRATED 0x40
1457#else
1458#define ENQUEUE_MIGRATED 0x00
1459#endif
1460
1461#define RETRY_TASK ((void *)-1UL)
1462
1463struct sched_class {
1464 const struct sched_class *next;
1465
1466 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1467 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1468 void (*yield_task) (struct rq *rq);
1469 bool (*yield_to_task)(struct rq *rq, struct task_struct *p, bool preempt);
1470
1471 void (*check_preempt_curr)(struct rq *rq, struct task_struct *p, int flags);
1472
1473 /*
1474 * It is the responsibility of the pick_next_task() method that will
1475 * return the next task to call put_prev_task() on the @prev task or
1476 * something equivalent.
1477 *
1478 * May return RETRY_TASK when it finds a higher prio class has runnable
1479 * tasks.
1480 */
1481 struct task_struct * (*pick_next_task)(struct rq *rq,
1482 struct task_struct *prev,
1483 struct rq_flags *rf);
1484 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
1485
1486#ifdef CONFIG_SMP
1487 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1488 void (*migrate_task_rq)(struct task_struct *p);
1489
1490 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
1491
1492 void (*set_cpus_allowed)(struct task_struct *p,
1493 const struct cpumask *newmask);
1494
1495 void (*rq_online)(struct rq *rq);
1496 void (*rq_offline)(struct rq *rq);
1497#endif
1498
1499 void (*set_curr_task)(struct rq *rq);
1500 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
1501 void (*task_fork)(struct task_struct *p);
1502 void (*task_dead)(struct task_struct *p);
1503
1504 /*
1505 * The switched_from() call is allowed to drop rq->lock, therefore we
1506 * cannot assume the switched_from/switched_to pair is serliazed by
1507 * rq->lock. They are however serialized by p->pi_lock.
1508 */
1509 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
1510 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1511 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1512 int oldprio);
1513
1514 unsigned int (*get_rr_interval)(struct rq *rq,
1515 struct task_struct *task);
1516
1517 void (*update_curr)(struct rq *rq);
1518
1519#define TASK_SET_GROUP 0
1520#define TASK_MOVE_GROUP 1
1521
1522#ifdef CONFIG_FAIR_GROUP_SCHED
1523 void (*task_change_group)(struct task_struct *p, int type);
1524#endif
1525};
1526
1527static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1528{
1529 prev->sched_class->put_prev_task(rq, prev);
1530}
1531
1532static inline void set_curr_task(struct rq *rq, struct task_struct *curr)
1533{
1534 curr->sched_class->set_curr_task(rq);
1535}
1536
1537#ifdef CONFIG_SMP
1538#define sched_class_highest (&stop_sched_class)
1539#else
1540#define sched_class_highest (&dl_sched_class)
1541#endif
1542#define for_each_class(class) \
1543 for (class = sched_class_highest; class; class = class->next)
1544
1545extern const struct sched_class stop_sched_class;
1546extern const struct sched_class dl_sched_class;
1547extern const struct sched_class rt_sched_class;
1548extern const struct sched_class fair_sched_class;
1549extern const struct sched_class idle_sched_class;
1550
1551
1552#ifdef CONFIG_SMP
1553
1554extern void update_group_capacity(struct sched_domain *sd, int cpu);
1555
1556extern void trigger_load_balance(struct rq *rq);
1557
1558extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask);
1559
1560#endif
1561
1562#ifdef CONFIG_CPU_IDLE
1563static inline void idle_set_state(struct rq *rq,
1564 struct cpuidle_state *idle_state)
1565{
1566 rq->idle_state = idle_state;
1567}
1568
1569static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1570{
1571 SCHED_WARN_ON(!rcu_read_lock_held());
1572
1573 return rq->idle_state;
1574}
1575#else
1576static inline void idle_set_state(struct rq *rq,
1577 struct cpuidle_state *idle_state)
1578{
1579}
1580
1581static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1582{
1583 return NULL;
1584}
1585#endif
1586
1587extern void schedule_idle(void);
1588
1589extern void sysrq_sched_debug_show(void);
1590extern void sched_init_granularity(void);
1591extern void update_max_interval(void);
1592
1593extern void init_sched_dl_class(void);
1594extern void init_sched_rt_class(void);
1595extern void init_sched_fair_class(void);
1596
1597extern void reweight_task(struct task_struct *p, int prio);
1598
1599extern void resched_curr(struct rq *rq);
1600extern void resched_cpu(int cpu);
1601
1602extern struct rt_bandwidth def_rt_bandwidth;
1603extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1604
1605extern struct dl_bandwidth def_dl_bandwidth;
1606extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1607extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1608extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se);
1609extern void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
1610
1611#define BW_SHIFT 20
1612#define BW_UNIT (1 << BW_SHIFT)
1613#define RATIO_SHIFT 8
1614unsigned long to_ratio(u64 period, u64 runtime);
1615
1616extern void init_entity_runnable_average(struct sched_entity *se);
1617extern void post_init_entity_util_avg(struct sched_entity *se);
1618
1619#ifdef CONFIG_NO_HZ_FULL
1620extern bool sched_can_stop_tick(struct rq *rq);
1621extern int __init sched_tick_offload_init(void);
1622
1623/*
1624 * Tick may be needed by tasks in the runqueue depending on their policy and
1625 * requirements. If tick is needed, lets send the target an IPI to kick it out of
1626 * nohz mode if necessary.
1627 */
1628static inline void sched_update_tick_dependency(struct rq *rq)
1629{
1630 int cpu;
1631
1632 if (!tick_nohz_full_enabled())
1633 return;
1634
1635 cpu = cpu_of(rq);
1636
1637 if (!tick_nohz_full_cpu(cpu))
1638 return;
1639
1640 if (sched_can_stop_tick(rq))
1641 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
1642 else
1643 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
1644}
1645#else
1646static inline int sched_tick_offload_init(void) { return 0; }
1647static inline void sched_update_tick_dependency(struct rq *rq) { }
1648#endif
1649
1650static inline void add_nr_running(struct rq *rq, unsigned count)
1651{
1652 unsigned prev_nr = rq->nr_running;
1653
1654 rq->nr_running = prev_nr + count;
1655
1656 if (prev_nr < 2 && rq->nr_running >= 2) {
1657#ifdef CONFIG_SMP
1658 if (!rq->rd->overload)
1659 rq->rd->overload = true;
1660#endif
1661 }
1662
1663 sched_update_tick_dependency(rq);
1664}
1665
1666static inline void sub_nr_running(struct rq *rq, unsigned count)
1667{
1668 rq->nr_running -= count;
1669 /* Check if we still need preemption */
1670 sched_update_tick_dependency(rq);
1671}
1672
1673extern void update_rq_clock(struct rq *rq);
1674
1675extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1676extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1677
1678extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1679
1680extern const_debug unsigned int sysctl_sched_time_avg;
1681extern const_debug unsigned int sysctl_sched_nr_migrate;
1682extern const_debug unsigned int sysctl_sched_migration_cost;
1683
1684static inline u64 sched_avg_period(void)
1685{
1686 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1687}
1688
1689#ifdef CONFIG_SCHED_HRTICK
1690
1691/*
1692 * Use hrtick when:
1693 * - enabled by features
1694 * - hrtimer is actually high res
1695 */
1696static inline int hrtick_enabled(struct rq *rq)
1697{
1698 if (!sched_feat(HRTICK))
1699 return 0;
1700 if (!cpu_active(cpu_of(rq)))
1701 return 0;
1702 return hrtimer_is_hres_active(&rq->hrtick_timer);
1703}
1704
1705void hrtick_start(struct rq *rq, u64 delay);
1706
1707#else
1708
1709static inline int hrtick_enabled(struct rq *rq)
1710{
1711 return 0;
1712}
1713
1714#endif /* CONFIG_SCHED_HRTICK */
1715
1716#ifndef arch_scale_freq_capacity
1717static __always_inline
1718unsigned long arch_scale_freq_capacity(int cpu)
1719{
1720 return SCHED_CAPACITY_SCALE;
1721}
1722#endif
1723
1724#ifdef CONFIG_SMP
1725extern void sched_avg_update(struct rq *rq);
1726
1727#ifndef arch_scale_cpu_capacity
1728static __always_inline
1729unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
1730{
1731 if (sd && (sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
1732 return sd->smt_gain / sd->span_weight;
1733
1734 return SCHED_CAPACITY_SCALE;
1735}
1736#endif
1737
1738static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1739{
1740 rq->rt_avg += rt_delta * arch_scale_freq_capacity(cpu_of(rq));
1741 sched_avg_update(rq);
1742}
1743#else
1744#ifndef arch_scale_cpu_capacity
1745static __always_inline
1746unsigned long arch_scale_cpu_capacity(void __always_unused *sd, int cpu)
1747{
1748 return SCHED_CAPACITY_SCALE;
1749}
1750#endif
1751static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1752static inline void sched_avg_update(struct rq *rq) { }
1753#endif
1754
1755struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1756 __acquires(rq->lock);
1757
1758struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1759 __acquires(p->pi_lock)
1760 __acquires(rq->lock);
1761
1762static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1763 __releases(rq->lock)
1764{
1765 rq_unpin_lock(rq, rf);
1766 raw_spin_unlock(&rq->lock);
1767}
1768
1769static inline void
1770task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1771 __releases(rq->lock)
1772 __releases(p->pi_lock)
1773{
1774 rq_unpin_lock(rq, rf);
1775 raw_spin_unlock(&rq->lock);
1776 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1777}
1778
1779static inline void
1780rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1781 __acquires(rq->lock)
1782{
1783 raw_spin_lock_irqsave(&rq->lock, rf->flags);
1784 rq_pin_lock(rq, rf);
1785}
1786
1787static inline void
1788rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1789 __acquires(rq->lock)
1790{
1791 raw_spin_lock_irq(&rq->lock);
1792 rq_pin_lock(rq, rf);
1793}
1794
1795static inline void
1796rq_lock(struct rq *rq, struct rq_flags *rf)
1797 __acquires(rq->lock)
1798{
1799 raw_spin_lock(&rq->lock);
1800 rq_pin_lock(rq, rf);
1801}
1802
1803static inline void
1804rq_relock(struct rq *rq, struct rq_flags *rf)
1805 __acquires(rq->lock)
1806{
1807 raw_spin_lock(&rq->lock);
1808 rq_repin_lock(rq, rf);
1809}
1810
1811static inline void
1812rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1813 __releases(rq->lock)
1814{
1815 rq_unpin_lock(rq, rf);
1816 raw_spin_unlock_irqrestore(&rq->lock, rf->flags);
1817}
1818
1819static inline void
1820rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1821 __releases(rq->lock)
1822{
1823 rq_unpin_lock(rq, rf);
1824 raw_spin_unlock_irq(&rq->lock);
1825}
1826
1827static inline void
1828rq_unlock(struct rq *rq, struct rq_flags *rf)
1829 __releases(rq->lock)
1830{
1831 rq_unpin_lock(rq, rf);
1832 raw_spin_unlock(&rq->lock);
1833}
1834
1835#ifdef CONFIG_SMP
1836#ifdef CONFIG_PREEMPT
1837
1838static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1839
1840/*
1841 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1842 * way at the expense of forcing extra atomic operations in all
1843 * invocations. This assures that the double_lock is acquired using the
1844 * same underlying policy as the spinlock_t on this architecture, which
1845 * reduces latency compared to the unfair variant below. However, it
1846 * also adds more overhead and therefore may reduce throughput.
1847 */
1848static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1849 __releases(this_rq->lock)
1850 __acquires(busiest->lock)
1851 __acquires(this_rq->lock)
1852{
1853 raw_spin_unlock(&this_rq->lock);
1854 double_rq_lock(this_rq, busiest);
1855
1856 return 1;
1857}
1858
1859#else
1860/*
1861 * Unfair double_lock_balance: Optimizes throughput at the expense of
1862 * latency by eliminating extra atomic operations when the locks are
1863 * already in proper order on entry. This favors lower CPU-ids and will
1864 * grant the double lock to lower CPUs over higher ids under contention,
1865 * regardless of entry order into the function.
1866 */
1867static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1868 __releases(this_rq->lock)
1869 __acquires(busiest->lock)
1870 __acquires(this_rq->lock)
1871{
1872 int ret = 0;
1873
1874 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1875 if (busiest < this_rq) {
1876 raw_spin_unlock(&this_rq->lock);
1877 raw_spin_lock(&busiest->lock);
1878 raw_spin_lock_nested(&this_rq->lock,
1879 SINGLE_DEPTH_NESTING);
1880 ret = 1;
1881 } else
1882 raw_spin_lock_nested(&busiest->lock,
1883 SINGLE_DEPTH_NESTING);
1884 }
1885 return ret;
1886}
1887
1888#endif /* CONFIG_PREEMPT */
1889
1890/*
1891 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1892 */
1893static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1894{
1895 if (unlikely(!irqs_disabled())) {
1896 /* printk() doesn't work well under rq->lock */
1897 raw_spin_unlock(&this_rq->lock);
1898 BUG_ON(1);
1899 }
1900
1901 return _double_lock_balance(this_rq, busiest);
1902}
1903
1904static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1905 __releases(busiest->lock)
1906{
1907 raw_spin_unlock(&busiest->lock);
1908 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1909}
1910
1911static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1912{
1913 if (l1 > l2)
1914 swap(l1, l2);
1915
1916 spin_lock(l1);
1917 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1918}
1919
1920static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1921{
1922 if (l1 > l2)
1923 swap(l1, l2);
1924
1925 spin_lock_irq(l1);
1926 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1927}
1928
1929static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1930{
1931 if (l1 > l2)
1932 swap(l1, l2);
1933
1934 raw_spin_lock(l1);
1935 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1936}
1937
1938/*
1939 * double_rq_lock - safely lock two runqueues
1940 *
1941 * Note this does not disable interrupts like task_rq_lock,
1942 * you need to do so manually before calling.
1943 */
1944static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1945 __acquires(rq1->lock)
1946 __acquires(rq2->lock)
1947{
1948 BUG_ON(!irqs_disabled());
1949 if (rq1 == rq2) {
1950 raw_spin_lock(&rq1->lock);
1951 __acquire(rq2->lock); /* Fake it out ;) */
1952 } else {
1953 if (rq1 < rq2) {
1954 raw_spin_lock(&rq1->lock);
1955 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1956 } else {
1957 raw_spin_lock(&rq2->lock);
1958 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1959 }
1960 }
1961}
1962
1963/*
1964 * double_rq_unlock - safely unlock two runqueues
1965 *
1966 * Note this does not restore interrupts like task_rq_unlock,
1967 * you need to do so manually after calling.
1968 */
1969static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1970 __releases(rq1->lock)
1971 __releases(rq2->lock)
1972{
1973 raw_spin_unlock(&rq1->lock);
1974 if (rq1 != rq2)
1975 raw_spin_unlock(&rq2->lock);
1976 else
1977 __release(rq2->lock);
1978}
1979
1980extern void set_rq_online (struct rq *rq);
1981extern void set_rq_offline(struct rq *rq);
1982extern bool sched_smp_initialized;
1983
1984#else /* CONFIG_SMP */
1985
1986/*
1987 * double_rq_lock - safely lock two runqueues
1988 *
1989 * Note this does not disable interrupts like task_rq_lock,
1990 * you need to do so manually before calling.
1991 */
1992static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1993 __acquires(rq1->lock)
1994 __acquires(rq2->lock)
1995{
1996 BUG_ON(!irqs_disabled());
1997 BUG_ON(rq1 != rq2);
1998 raw_spin_lock(&rq1->lock);
1999 __acquire(rq2->lock); /* Fake it out ;) */
2000}
2001
2002/*
2003 * double_rq_unlock - safely unlock two runqueues
2004 *
2005 * Note this does not restore interrupts like task_rq_unlock,
2006 * you need to do so manually after calling.
2007 */
2008static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2009 __releases(rq1->lock)
2010 __releases(rq2->lock)
2011{
2012 BUG_ON(rq1 != rq2);
2013 raw_spin_unlock(&rq1->lock);
2014 __release(rq2->lock);
2015}
2016
2017#endif
2018
2019extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2020extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2021
2022#ifdef CONFIG_SCHED_DEBUG
2023extern bool sched_debug_enabled;
2024
2025extern void print_cfs_stats(struct seq_file *m, int cpu);
2026extern void print_rt_stats(struct seq_file *m, int cpu);
2027extern void print_dl_stats(struct seq_file *m, int cpu);
2028extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2029extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2030extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2031#ifdef CONFIG_NUMA_BALANCING
2032extern void
2033show_numa_stats(struct task_struct *p, struct seq_file *m);
2034extern void
2035print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2036 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2037#endif /* CONFIG_NUMA_BALANCING */
2038#endif /* CONFIG_SCHED_DEBUG */
2039
2040extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2041extern void init_rt_rq(struct rt_rq *rt_rq);
2042extern void init_dl_rq(struct dl_rq *dl_rq);
2043
2044extern void cfs_bandwidth_usage_inc(void);
2045extern void cfs_bandwidth_usage_dec(void);
2046
2047#ifdef CONFIG_NO_HZ_COMMON
2048#define NOHZ_BALANCE_KICK_BIT 0
2049#define NOHZ_STATS_KICK_BIT 1
2050
2051#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2052#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2053
2054#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK)
2055
2056#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2057
2058extern void nohz_balance_exit_idle(struct rq *rq);
2059#else
2060static inline void nohz_balance_exit_idle(struct rq *rq) { }
2061#endif
2062
2063
2064#ifdef CONFIG_SMP
2065static inline
2066void __dl_update(struct dl_bw *dl_b, s64 bw)
2067{
2068 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
2069 int i;
2070
2071 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2072 "sched RCU must be held");
2073 for_each_cpu_and(i, rd->span, cpu_active_mask) {
2074 struct rq *rq = cpu_rq(i);
2075
2076 rq->dl.extra_bw += bw;
2077 }
2078}
2079#else
2080static inline
2081void __dl_update(struct dl_bw *dl_b, s64 bw)
2082{
2083 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
2084
2085 dl->extra_bw += bw;
2086}
2087#endif
2088
2089
2090#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2091struct irqtime {
2092 u64 total;
2093 u64 tick_delta;
2094 u64 irq_start_time;
2095 struct u64_stats_sync sync;
2096};
2097
2098DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2099
2100/*
2101 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2102 * Otherwise ksoftirqd's sum_exec_runtime is substracted its own runtime
2103 * and never move forward.
2104 */
2105static inline u64 irq_time_read(int cpu)
2106{
2107 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2108 unsigned int seq;
2109 u64 total;
2110
2111 do {
2112 seq = __u64_stats_fetch_begin(&irqtime->sync);
2113 total = irqtime->total;
2114 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2115
2116 return total;
2117}
2118#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2119
2120#ifdef CONFIG_CPU_FREQ
2121DECLARE_PER_CPU(struct update_util_data *, cpufreq_update_util_data);
2122
2123/**
2124 * cpufreq_update_util - Take a note about CPU utilization changes.
2125 * @rq: Runqueue to carry out the update for.
2126 * @flags: Update reason flags.
2127 *
2128 * This function is called by the scheduler on the CPU whose utilization is
2129 * being updated.
2130 *
2131 * It can only be called from RCU-sched read-side critical sections.
2132 *
2133 * The way cpufreq is currently arranged requires it to evaluate the CPU
2134 * performance state (frequency/voltage) on a regular basis to prevent it from
2135 * being stuck in a completely inadequate performance level for too long.
2136 * That is not guaranteed to happen if the updates are only triggered from CFS
2137 * and DL, though, because they may not be coming in if only RT tasks are
2138 * active all the time (or there are RT tasks only).
2139 *
2140 * As a workaround for that issue, this function is called periodically by the
2141 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2142 * but that really is a band-aid. Going forward it should be replaced with
2143 * solutions targeted more specifically at RT tasks.
2144 */
2145static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2146{
2147 struct update_util_data *data;
2148
2149 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2150 cpu_of(rq)));
2151 if (data)
2152 data->func(data, rq_clock(rq), flags);
2153}
2154#else
2155static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2156#endif /* CONFIG_CPU_FREQ */
2157
2158#ifdef arch_scale_freq_capacity
2159# ifndef arch_scale_freq_invariant
2160# define arch_scale_freq_invariant() true
2161# endif
2162#else
2163# define arch_scale_freq_invariant() false
2164#endif
2165
2166#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
2167static inline unsigned long cpu_util_dl(struct rq *rq)
2168{
2169 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
2170}
2171
2172static inline unsigned long cpu_util_cfs(struct rq *rq)
2173{
2174 unsigned long util = READ_ONCE(rq->cfs.avg.util_avg);
2175
2176 if (sched_feat(UTIL_EST)) {
2177 util = max_t(unsigned long, util,
2178 READ_ONCE(rq->cfs.avg.util_est.enqueued));
2179 }
2180
2181 return util;
2182}
2183#endif