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>
13 #include "cpudeadline.h"
18 /* task_struct::on_rq states: */
19 #define TASK_ON_RQ_QUEUED 1
20 #define TASK_ON_RQ_MIGRATING 2
22 extern __read_mostly int scheduler_running;
24 extern unsigned long calc_load_update;
25 extern atomic_long_t calc_load_tasks;
27 extern long calc_load_fold_active(struct rq *this_rq);
28 extern void update_cpu_load_active(struct rq *this_rq);
31 * Helpers for converting nanosecond timing to jiffy resolution
33 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
36 * Increase resolution of nice-level calculations for 64-bit architectures.
37 * The extra resolution improves shares distribution and load balancing of
38 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
39 * hierarchies, especially on larger systems. This is not a user-visible change
40 * and does not change the user-interface for setting shares/weights.
42 * We increase resolution only if we have enough bits to allow this increased
43 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
44 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
47 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
48 # define SCHED_LOAD_RESOLUTION 10
49 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
50 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
52 # define SCHED_LOAD_RESOLUTION 0
53 # define scale_load(w) (w)
54 # define scale_load_down(w) (w)
57 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
58 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
60 #define NICE_0_LOAD SCHED_LOAD_SCALE
61 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
64 * Single value that decides SCHED_DEADLINE internal math precision.
65 * 10 -> just above 1us
66 * 9 -> just above 0.5us
71 * These are the 'tuning knobs' of the scheduler:
75 * single value that denotes runtime == period, ie unlimited time.
77 #define RUNTIME_INF ((u64)~0ULL)
79 static inline int fair_policy(int policy)
81 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
84 static inline int rt_policy(int policy)
86 return policy == SCHED_FIFO || policy == SCHED_RR;
89 static inline int dl_policy(int policy)
91 return policy == SCHED_DEADLINE;
94 static inline int task_has_rt_policy(struct task_struct *p)
96 return rt_policy(p->policy);
99 static inline int task_has_dl_policy(struct task_struct *p)
101 return dl_policy(p->policy);
104 static inline bool dl_time_before(u64 a, u64 b)
106 return (s64)(a - b) < 0;
110 * Tells if entity @a should preempt entity @b.
113 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
115 return dl_time_before(a->deadline, b->deadline);
119 * This is the priority-queue data structure of the RT scheduling class:
121 struct rt_prio_array {
122 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
123 struct list_head queue[MAX_RT_PRIO];
126 struct rt_bandwidth {
127 /* nests inside the rq lock: */
128 raw_spinlock_t rt_runtime_lock;
131 struct hrtimer rt_period_timer;
134 * To keep the bandwidth of -deadline tasks and groups under control
135 * we need some place where:
136 * - store the maximum -deadline bandwidth of the system (the group);
137 * - cache the fraction of that bandwidth that is currently allocated.
139 * This is all done in the data structure below. It is similar to the
140 * one used for RT-throttling (rt_bandwidth), with the main difference
141 * that, since here we are only interested in admission control, we
142 * do not decrease any runtime while the group "executes", neither we
143 * need a timer to replenish it.
145 * With respect to SMP, the bandwidth is given on a per-CPU basis,
147 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
148 * - dl_total_bw array contains, in the i-eth element, the currently
149 * allocated bandwidth on the i-eth CPU.
150 * Moreover, groups consume bandwidth on each CPU, while tasks only
151 * consume bandwidth on the CPU they're running on.
152 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
153 * that will be shown the next time the proc or cgroup controls will
154 * be red. It on its turn can be changed by writing on its own
157 struct dl_bandwidth {
158 raw_spinlock_t dl_runtime_lock;
163 static inline int dl_bandwidth_enabled(void)
165 return sysctl_sched_rt_runtime >= 0;
168 extern struct dl_bw *dl_bw_of(int i);
175 extern struct mutex sched_domains_mutex;
177 #ifdef CONFIG_CGROUP_SCHED
179 #include <linux/cgroup.h>
184 extern struct list_head task_groups;
186 struct cfs_bandwidth {
187 #ifdef CONFIG_CFS_BANDWIDTH
191 s64 hierarchal_quota;
194 int idle, timer_active;
195 struct hrtimer period_timer, slack_timer;
196 struct list_head throttled_cfs_rq;
199 int nr_periods, nr_throttled;
204 /* task group related information */
206 struct cgroup_subsys_state css;
208 #ifdef CONFIG_FAIR_GROUP_SCHED
209 /* schedulable entities of this group on each cpu */
210 struct sched_entity **se;
211 /* runqueue "owned" by this group on each cpu */
212 struct cfs_rq **cfs_rq;
213 unsigned long shares;
216 atomic_long_t load_avg;
217 atomic_t runnable_avg;
221 #ifdef CONFIG_RT_GROUP_SCHED
222 struct sched_rt_entity **rt_se;
223 struct rt_rq **rt_rq;
225 struct rt_bandwidth rt_bandwidth;
229 struct list_head list;
231 struct task_group *parent;
232 struct list_head siblings;
233 struct list_head children;
235 #ifdef CONFIG_SCHED_AUTOGROUP
236 struct autogroup *autogroup;
239 struct cfs_bandwidth cfs_bandwidth;
242 #ifdef CONFIG_FAIR_GROUP_SCHED
243 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
246 * A weight of 0 or 1 can cause arithmetics problems.
247 * A weight of a cfs_rq is the sum of weights of which entities
248 * are queued on this cfs_rq, so a weight of a entity should not be
249 * too large, so as the shares value of a task group.
250 * (The default weight is 1024 - so there's no practical
251 * limitation from this.)
253 #define MIN_SHARES (1UL << 1)
254 #define MAX_SHARES (1UL << 18)
257 typedef int (*tg_visitor)(struct task_group *, void *);
259 extern int walk_tg_tree_from(struct task_group *from,
260 tg_visitor down, tg_visitor up, void *data);
263 * Iterate the full tree, calling @down when first entering a node and @up when
264 * leaving it for the final time.
266 * Caller must hold rcu_lock or sufficient equivalent.
268 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
270 return walk_tg_tree_from(&root_task_group, down, up, data);
273 extern int tg_nop(struct task_group *tg, void *data);
275 extern void free_fair_sched_group(struct task_group *tg);
276 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
277 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
278 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
279 struct sched_entity *se, int cpu,
280 struct sched_entity *parent);
281 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
282 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
284 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
285 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
286 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
288 extern void free_rt_sched_group(struct task_group *tg);
289 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
290 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
291 struct sched_rt_entity *rt_se, int cpu,
292 struct sched_rt_entity *parent);
294 extern struct task_group *sched_create_group(struct task_group *parent);
295 extern void sched_online_group(struct task_group *tg,
296 struct task_group *parent);
297 extern void sched_destroy_group(struct task_group *tg);
298 extern void sched_offline_group(struct task_group *tg);
300 extern void sched_move_task(struct task_struct *tsk);
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
306 #else /* CONFIG_CGROUP_SCHED */
308 struct cfs_bandwidth { };
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned int nr_running, h_nr_running;
320 u64 min_vruntime_copy;
323 struct rb_root tasks_timeline;
324 struct rb_node *rb_leftmost;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last, *skip;
332 #ifdef CONFIG_SCHED_DEBUG
333 unsigned int nr_spread_over;
339 * Under CFS, load is tracked on a per-entity basis and aggregated up.
340 * This allows for the description of both thread and group usage (in
341 * the FAIR_GROUP_SCHED case).
343 unsigned long runnable_load_avg, blocked_load_avg;
344 atomic64_t decay_counter;
346 atomic_long_t removed_load;
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 /* Required to track per-cpu representation of a task_group */
350 u32 tg_runnable_contrib;
351 unsigned long tg_load_contrib;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load;
360 u64 last_h_load_update;
361 struct sched_entity *h_load_next;
362 #endif /* CONFIG_FAIR_GROUP_SCHED */
363 #endif /* CONFIG_SMP */
365 #ifdef CONFIG_FAIR_GROUP_SCHED
366 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
369 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
370 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
371 * (like users, containers etc.)
373 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
374 * list is used during load balance.
377 struct list_head leaf_cfs_rq_list;
378 struct task_group *tg; /* group that "owns" this runqueue */
380 #ifdef CONFIG_CFS_BANDWIDTH
383 s64 runtime_remaining;
385 u64 throttled_clock, throttled_clock_task;
386 u64 throttled_clock_task_time;
387 int throttled, throttle_count;
388 struct list_head throttled_list;
389 #endif /* CONFIG_CFS_BANDWIDTH */
390 #endif /* CONFIG_FAIR_GROUP_SCHED */
393 static inline int rt_bandwidth_enabled(void)
395 return sysctl_sched_rt_runtime >= 0;
398 /* Real-Time classes' related field in a runqueue: */
400 struct rt_prio_array active;
401 unsigned int rt_nr_running;
402 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
404 int curr; /* highest queued rt task prio */
406 int next; /* next highest */
411 unsigned long rt_nr_migratory;
412 unsigned long rt_nr_total;
414 struct plist_head pushable_tasks;
421 /* Nests inside the rq lock: */
422 raw_spinlock_t rt_runtime_lock;
424 #ifdef CONFIG_RT_GROUP_SCHED
425 unsigned long rt_nr_boosted;
428 struct task_group *tg;
432 /* Deadline class' related fields in a runqueue */
434 /* runqueue is an rbtree, ordered by deadline */
435 struct rb_root rb_root;
436 struct rb_node *rb_leftmost;
438 unsigned long dl_nr_running;
442 * Deadline values of the currently executing and the
443 * earliest ready task on this rq. Caching these facilitates
444 * the decision wether or not a ready but not running task
445 * should migrate somewhere else.
452 unsigned long dl_nr_migratory;
456 * Tasks on this rq that can be pushed away. They are kept in
457 * an rb-tree, ordered by tasks' deadlines, with caching
458 * of the leftmost (earliest deadline) element.
460 struct rb_root pushable_dl_tasks_root;
461 struct rb_node *pushable_dl_tasks_leftmost;
470 * We add the notion of a root-domain which will be used to define per-domain
471 * variables. Each exclusive cpuset essentially defines an island domain by
472 * fully partitioning the member cpus from any other cpuset. Whenever a new
473 * exclusive cpuset is created, we also create and attach a new root-domain
482 cpumask_var_t online;
484 /* Indicate more than one runnable task for any CPU */
488 * The bit corresponding to a CPU gets set here if such CPU has more
489 * than one runnable -deadline task (as it is below for RT tasks).
491 cpumask_var_t dlo_mask;
497 * The "RT overload" flag: it gets set if a CPU has more than
498 * one runnable RT task.
500 cpumask_var_t rto_mask;
501 struct cpupri cpupri;
504 extern struct root_domain def_root_domain;
506 #endif /* CONFIG_SMP */
509 * This is the main, per-CPU runqueue data structure.
511 * Locking rule: those places that want to lock multiple runqueues
512 * (such as the load balancing or the thread migration code), lock
513 * acquire operations must be ordered by ascending &runqueue.
520 * nr_running and cpu_load should be in the same cacheline because
521 * remote CPUs use both these fields when doing load calculation.
523 unsigned int nr_running;
524 #ifdef CONFIG_NUMA_BALANCING
525 unsigned int nr_numa_running;
526 unsigned int nr_preferred_running;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
530 unsigned long last_load_update_tick;
531 #ifdef CONFIG_NO_HZ_COMMON
533 unsigned long nohz_flags;
535 #ifdef CONFIG_NO_HZ_FULL
536 unsigned long last_sched_tick;
538 int skip_clock_update;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 struct sched_avg avg;
554 #endif /* CONFIG_FAIR_GROUP_SCHED */
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle, *stop;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned long cpu_capacity;
579 unsigned char idle_balance;
580 /* For active balancing */
584 struct cpu_stop_work active_balance_work;
585 /* cpu of this runqueue: */
589 struct list_head cfs_tasks;
596 /* This is used to determine avg_idle's max value */
597 u64 max_idle_balance_cost;
600 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
603 #ifdef CONFIG_PARAVIRT
606 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
607 u64 prev_steal_time_rq;
610 /* calc_load related fields */
611 unsigned long calc_load_update;
612 long calc_load_active;
614 #ifdef CONFIG_SCHED_HRTICK
616 int hrtick_csd_pending;
617 struct call_single_data hrtick_csd;
619 struct hrtimer hrtick_timer;
622 #ifdef CONFIG_SCHEDSTATS
624 struct sched_info rq_sched_info;
625 unsigned long long rq_cpu_time;
626 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
628 /* sys_sched_yield() stats */
629 unsigned int yld_count;
631 /* schedule() stats */
632 unsigned int sched_count;
633 unsigned int sched_goidle;
635 /* try_to_wake_up() stats */
636 unsigned int ttwu_count;
637 unsigned int ttwu_local;
641 struct llist_head wake_list;
645 static inline int cpu_of(struct rq *rq)
654 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
656 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
657 #define this_rq() (&__get_cpu_var(runqueues))
658 #define task_rq(p) cpu_rq(task_cpu(p))
659 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
660 #define raw_rq() (&__raw_get_cpu_var(runqueues))
662 static inline u64 rq_clock(struct rq *rq)
667 static inline u64 rq_clock_task(struct rq *rq)
669 return rq->clock_task;
672 #ifdef CONFIG_NUMA_BALANCING
673 extern void sched_setnuma(struct task_struct *p, int node);
674 extern int migrate_task_to(struct task_struct *p, int cpu);
675 extern int migrate_swap(struct task_struct *, struct task_struct *);
676 #endif /* CONFIG_NUMA_BALANCING */
680 extern void sched_ttwu_pending(void);
682 #define rcu_dereference_check_sched_domain(p) \
683 rcu_dereference_check((p), \
684 lockdep_is_held(&sched_domains_mutex))
687 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
688 * See detach_destroy_domains: synchronize_sched for details.
690 * The domain tree of any CPU may only be accessed from within
691 * preempt-disabled sections.
693 #define for_each_domain(cpu, __sd) \
694 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
695 __sd; __sd = __sd->parent)
697 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
700 * highest_flag_domain - Return highest sched_domain containing flag.
701 * @cpu: The cpu whose highest level of sched domain is to
703 * @flag: The flag to check for the highest sched_domain
706 * Returns the highest sched_domain of a cpu which contains the given flag.
708 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
710 struct sched_domain *sd, *hsd = NULL;
712 for_each_domain(cpu, sd) {
713 if (!(sd->flags & flag))
721 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
723 struct sched_domain *sd;
725 for_each_domain(cpu, sd) {
726 if (sd->flags & flag)
733 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
734 DECLARE_PER_CPU(int, sd_llc_size);
735 DECLARE_PER_CPU(int, sd_llc_id);
736 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
737 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
738 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
740 struct sched_group_capacity {
743 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
746 unsigned int capacity, capacity_orig;
747 unsigned long next_update;
748 int imbalance; /* XXX unrelated to capacity but shared group state */
750 * Number of busy cpus in this group.
752 atomic_t nr_busy_cpus;
754 unsigned long cpumask[0]; /* iteration mask */
758 struct sched_group *next; /* Must be a circular list */
761 unsigned int group_weight;
762 struct sched_group_capacity *sgc;
765 * The CPUs this group covers.
767 * NOTE: this field is variable length. (Allocated dynamically
768 * by attaching extra space to the end of the structure,
769 * depending on how many CPUs the kernel has booted up with)
771 unsigned long cpumask[0];
774 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
776 return to_cpumask(sg->cpumask);
780 * cpumask masking which cpus in the group are allowed to iterate up the domain
783 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
785 return to_cpumask(sg->sgc->cpumask);
789 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
790 * @group: The group whose first cpu is to be returned.
792 static inline unsigned int group_first_cpu(struct sched_group *group)
794 return cpumask_first(sched_group_cpus(group));
797 extern int group_balance_cpu(struct sched_group *sg);
801 static inline void sched_ttwu_pending(void) { }
803 #endif /* CONFIG_SMP */
806 #include "auto_group.h"
808 #ifdef CONFIG_CGROUP_SCHED
811 * Return the group to which this tasks belongs.
813 * We cannot use task_css() and friends because the cgroup subsystem
814 * changes that value before the cgroup_subsys::attach() method is called,
815 * therefore we cannot pin it and might observe the wrong value.
817 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
818 * core changes this before calling sched_move_task().
820 * Instead we use a 'copy' which is updated from sched_move_task() while
821 * holding both task_struct::pi_lock and rq::lock.
823 static inline struct task_group *task_group(struct task_struct *p)
825 return p->sched_task_group;
828 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
829 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
831 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
832 struct task_group *tg = task_group(p);
835 #ifdef CONFIG_FAIR_GROUP_SCHED
836 p->se.cfs_rq = tg->cfs_rq[cpu];
837 p->se.parent = tg->se[cpu];
840 #ifdef CONFIG_RT_GROUP_SCHED
841 p->rt.rt_rq = tg->rt_rq[cpu];
842 p->rt.parent = tg->rt_se[cpu];
846 #else /* CONFIG_CGROUP_SCHED */
848 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
849 static inline struct task_group *task_group(struct task_struct *p)
854 #endif /* CONFIG_CGROUP_SCHED */
856 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
861 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
862 * successfuly executed on another CPU. We must ensure that updates of
863 * per-task data have been completed by this moment.
866 task_thread_info(p)->cpu = cpu;
872 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
874 #ifdef CONFIG_SCHED_DEBUG
875 # include <linux/static_key.h>
876 # define const_debug __read_mostly
878 # define const_debug const
881 extern const_debug unsigned int sysctl_sched_features;
883 #define SCHED_FEAT(name, enabled) \
884 __SCHED_FEAT_##name ,
887 #include "features.h"
893 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
894 #define SCHED_FEAT(name, enabled) \
895 static __always_inline bool static_branch_##name(struct static_key *key) \
897 return static_key_##enabled(key); \
900 #include "features.h"
904 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
905 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
906 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
907 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
908 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
910 #ifdef CONFIG_NUMA_BALANCING
911 #define sched_feat_numa(x) sched_feat(x)
912 #ifdef CONFIG_SCHED_DEBUG
913 #define numabalancing_enabled sched_feat_numa(NUMA)
915 extern bool numabalancing_enabled;
916 #endif /* CONFIG_SCHED_DEBUG */
918 #define sched_feat_numa(x) (0)
919 #define numabalancing_enabled (0)
920 #endif /* CONFIG_NUMA_BALANCING */
922 static inline u64 global_rt_period(void)
924 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
927 static inline u64 global_rt_runtime(void)
929 if (sysctl_sched_rt_runtime < 0)
932 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
935 static inline int task_current(struct rq *rq, struct task_struct *p)
937 return rq->curr == p;
940 static inline int task_running(struct rq *rq, struct task_struct *p)
945 return task_current(rq, p);
949 static inline int task_on_rq_queued(struct task_struct *p)
951 return p->on_rq == TASK_ON_RQ_QUEUED;
954 static inline int task_on_rq_migrating(struct task_struct *p)
956 return p->on_rq == TASK_ON_RQ_MIGRATING;
959 #ifndef prepare_arch_switch
960 # define prepare_arch_switch(next) do { } while (0)
962 #ifndef finish_arch_switch
963 # define finish_arch_switch(prev) do { } while (0)
965 #ifndef finish_arch_post_lock_switch
966 # define finish_arch_post_lock_switch() do { } while (0)
969 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
970 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
974 * We can optimise this out completely for !SMP, because the
975 * SMP rebalancing from interrupt is the only thing that cares
982 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
986 * After ->on_cpu is cleared, the task can be moved to a different CPU.
987 * We must ensure this doesn't happen until the switch is completely
993 #ifdef CONFIG_DEBUG_SPINLOCK
994 /* this is a valid case when another task releases the spinlock */
995 rq->lock.owner = current;
998 * If we are tracking spinlock dependencies then we have to
999 * fix up the runqueue lock - which gets 'carried over' from
1000 * prev into current:
1002 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1004 raw_spin_unlock_irq(&rq->lock);
1007 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1008 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1012 * We can optimise this out completely for !SMP, because the
1013 * SMP rebalancing from interrupt is the only thing that cares
1018 raw_spin_unlock(&rq->lock);
1021 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1025 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1026 * We must ensure this doesn't happen until the switch is completely
1034 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1039 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1040 #define WF_FORK 0x02 /* child wakeup after fork */
1041 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1044 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1045 * of tasks with abnormal "nice" values across CPUs the contribution that
1046 * each task makes to its run queue's load is weighted according to its
1047 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1048 * scaled version of the new time slice allocation that they receive on time
1052 #define WEIGHT_IDLEPRIO 3
1053 #define WMULT_IDLEPRIO 1431655765
1056 * Nice levels are multiplicative, with a gentle 10% change for every
1057 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1058 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1059 * that remained on nice 0.
1061 * The "10% effect" is relative and cumulative: from _any_ nice level,
1062 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1063 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1064 * If a task goes up by ~10% and another task goes down by ~10% then
1065 * the relative distance between them is ~25%.)
1067 static const int prio_to_weight[40] = {
1068 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1069 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1070 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1071 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1072 /* 0 */ 1024, 820, 655, 526, 423,
1073 /* 5 */ 335, 272, 215, 172, 137,
1074 /* 10 */ 110, 87, 70, 56, 45,
1075 /* 15 */ 36, 29, 23, 18, 15,
1079 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1081 * In cases where the weight does not change often, we can use the
1082 * precalculated inverse to speed up arithmetics by turning divisions
1083 * into multiplications:
1085 static const u32 prio_to_wmult[40] = {
1086 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1087 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1088 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1089 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1090 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1091 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1092 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1093 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1096 #define ENQUEUE_WAKEUP 1
1097 #define ENQUEUE_HEAD 2
1099 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1101 #define ENQUEUE_WAKING 0
1103 #define ENQUEUE_REPLENISH 8
1105 #define DEQUEUE_SLEEP 1
1107 #define RETRY_TASK ((void *)-1UL)
1109 struct sched_class {
1110 const struct sched_class *next;
1112 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1113 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1114 void (*yield_task) (struct rq *rq);
1115 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1117 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1120 * It is the responsibility of the pick_next_task() method that will
1121 * return the next task to call put_prev_task() on the @prev task or
1122 * something equivalent.
1124 * May return RETRY_TASK when it finds a higher prio class has runnable
1127 struct task_struct * (*pick_next_task) (struct rq *rq,
1128 struct task_struct *prev);
1129 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1132 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1133 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1135 void (*post_schedule) (struct rq *this_rq);
1136 void (*task_waking) (struct task_struct *task);
1137 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1139 void (*set_cpus_allowed)(struct task_struct *p,
1140 const struct cpumask *newmask);
1142 void (*rq_online)(struct rq *rq);
1143 void (*rq_offline)(struct rq *rq);
1146 void (*set_curr_task) (struct rq *rq);
1147 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1148 void (*task_fork) (struct task_struct *p);
1149 void (*task_dead) (struct task_struct *p);
1151 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1152 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1153 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1156 unsigned int (*get_rr_interval) (struct rq *rq,
1157 struct task_struct *task);
1159 #ifdef CONFIG_FAIR_GROUP_SCHED
1160 void (*task_move_group) (struct task_struct *p, int on_rq);
1164 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1166 prev->sched_class->put_prev_task(rq, prev);
1169 #define sched_class_highest (&stop_sched_class)
1170 #define for_each_class(class) \
1171 for (class = sched_class_highest; class; class = class->next)
1173 extern const struct sched_class stop_sched_class;
1174 extern const struct sched_class dl_sched_class;
1175 extern const struct sched_class rt_sched_class;
1176 extern const struct sched_class fair_sched_class;
1177 extern const struct sched_class idle_sched_class;
1182 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1184 extern void trigger_load_balance(struct rq *rq);
1186 extern void idle_enter_fair(struct rq *this_rq);
1187 extern void idle_exit_fair(struct rq *this_rq);
1191 static inline void idle_enter_fair(struct rq *rq) { }
1192 static inline void idle_exit_fair(struct rq *rq) { }
1196 extern void sysrq_sched_debug_show(void);
1197 extern void sched_init_granularity(void);
1198 extern void update_max_interval(void);
1200 extern void init_sched_dl_class(void);
1201 extern void init_sched_rt_class(void);
1202 extern void init_sched_fair_class(void);
1203 extern void init_sched_dl_class(void);
1205 extern void resched_curr(struct rq *rq);
1206 extern void resched_cpu(int cpu);
1208 extern struct rt_bandwidth def_rt_bandwidth;
1209 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1211 extern struct dl_bandwidth def_dl_bandwidth;
1212 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1213 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1215 unsigned long to_ratio(u64 period, u64 runtime);
1217 extern void update_idle_cpu_load(struct rq *this_rq);
1219 extern void init_task_runnable_average(struct task_struct *p);
1221 static inline void add_nr_running(struct rq *rq, unsigned count)
1223 unsigned prev_nr = rq->nr_running;
1225 rq->nr_running = prev_nr + count;
1227 if (prev_nr < 2 && rq->nr_running >= 2) {
1229 if (!rq->rd->overload)
1230 rq->rd->overload = true;
1233 #ifdef CONFIG_NO_HZ_FULL
1234 if (tick_nohz_full_cpu(rq->cpu)) {
1236 * Tick is needed if more than one task runs on a CPU.
1237 * Send the target an IPI to kick it out of nohz mode.
1239 * We assume that IPI implies full memory barrier and the
1240 * new value of rq->nr_running is visible on reception
1243 tick_nohz_full_kick_cpu(rq->cpu);
1249 static inline void sub_nr_running(struct rq *rq, unsigned count)
1251 rq->nr_running -= count;
1254 static inline void rq_last_tick_reset(struct rq *rq)
1256 #ifdef CONFIG_NO_HZ_FULL
1257 rq->last_sched_tick = jiffies;
1261 extern void update_rq_clock(struct rq *rq);
1263 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1264 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1266 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1268 extern const_debug unsigned int sysctl_sched_time_avg;
1269 extern const_debug unsigned int sysctl_sched_nr_migrate;
1270 extern const_debug unsigned int sysctl_sched_migration_cost;
1272 static inline u64 sched_avg_period(void)
1274 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1277 #ifdef CONFIG_SCHED_HRTICK
1281 * - enabled by features
1282 * - hrtimer is actually high res
1284 static inline int hrtick_enabled(struct rq *rq)
1286 if (!sched_feat(HRTICK))
1288 if (!cpu_active(cpu_of(rq)))
1290 return hrtimer_is_hres_active(&rq->hrtick_timer);
1293 void hrtick_start(struct rq *rq, u64 delay);
1297 static inline int hrtick_enabled(struct rq *rq)
1302 #endif /* CONFIG_SCHED_HRTICK */
1305 extern void sched_avg_update(struct rq *rq);
1306 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1308 rq->rt_avg += rt_delta;
1309 sched_avg_update(rq);
1312 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1313 static inline void sched_avg_update(struct rq *rq) { }
1316 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1319 #ifdef CONFIG_PREEMPT
1321 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1324 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1325 * way at the expense of forcing extra atomic operations in all
1326 * invocations. This assures that the double_lock is acquired using the
1327 * same underlying policy as the spinlock_t on this architecture, which
1328 * reduces latency compared to the unfair variant below. However, it
1329 * also adds more overhead and therefore may reduce throughput.
1331 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1332 __releases(this_rq->lock)
1333 __acquires(busiest->lock)
1334 __acquires(this_rq->lock)
1336 raw_spin_unlock(&this_rq->lock);
1337 double_rq_lock(this_rq, busiest);
1344 * Unfair double_lock_balance: Optimizes throughput at the expense of
1345 * latency by eliminating extra atomic operations when the locks are
1346 * already in proper order on entry. This favors lower cpu-ids and will
1347 * grant the double lock to lower cpus over higher ids under contention,
1348 * regardless of entry order into the function.
1350 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1351 __releases(this_rq->lock)
1352 __acquires(busiest->lock)
1353 __acquires(this_rq->lock)
1357 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1358 if (busiest < this_rq) {
1359 raw_spin_unlock(&this_rq->lock);
1360 raw_spin_lock(&busiest->lock);
1361 raw_spin_lock_nested(&this_rq->lock,
1362 SINGLE_DEPTH_NESTING);
1365 raw_spin_lock_nested(&busiest->lock,
1366 SINGLE_DEPTH_NESTING);
1371 #endif /* CONFIG_PREEMPT */
1374 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1376 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1378 if (unlikely(!irqs_disabled())) {
1379 /* printk() doesn't work good under rq->lock */
1380 raw_spin_unlock(&this_rq->lock);
1384 return _double_lock_balance(this_rq, busiest);
1387 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1388 __releases(busiest->lock)
1390 raw_spin_unlock(&busiest->lock);
1391 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1394 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1400 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1403 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1409 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1412 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1418 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1422 * double_rq_lock - safely lock two runqueues
1424 * Note this does not disable interrupts like task_rq_lock,
1425 * you need to do so manually before calling.
1427 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1428 __acquires(rq1->lock)
1429 __acquires(rq2->lock)
1431 BUG_ON(!irqs_disabled());
1433 raw_spin_lock(&rq1->lock);
1434 __acquire(rq2->lock); /* Fake it out ;) */
1437 raw_spin_lock(&rq1->lock);
1438 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1440 raw_spin_lock(&rq2->lock);
1441 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1447 * double_rq_unlock - safely unlock two runqueues
1449 * Note this does not restore interrupts like task_rq_unlock,
1450 * you need to do so manually after calling.
1452 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1453 __releases(rq1->lock)
1454 __releases(rq2->lock)
1456 raw_spin_unlock(&rq1->lock);
1458 raw_spin_unlock(&rq2->lock);
1460 __release(rq2->lock);
1463 #else /* CONFIG_SMP */
1466 * double_rq_lock - safely lock two runqueues
1468 * Note this does not disable interrupts like task_rq_lock,
1469 * you need to do so manually before calling.
1471 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1472 __acquires(rq1->lock)
1473 __acquires(rq2->lock)
1475 BUG_ON(!irqs_disabled());
1477 raw_spin_lock(&rq1->lock);
1478 __acquire(rq2->lock); /* Fake it out ;) */
1482 * double_rq_unlock - safely unlock two runqueues
1484 * Note this does not restore interrupts like task_rq_unlock,
1485 * you need to do so manually after calling.
1487 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1488 __releases(rq1->lock)
1489 __releases(rq2->lock)
1492 raw_spin_unlock(&rq1->lock);
1493 __release(rq2->lock);
1498 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1499 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1500 extern void print_cfs_stats(struct seq_file *m, int cpu);
1501 extern void print_rt_stats(struct seq_file *m, int cpu);
1503 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1504 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1505 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1507 extern void cfs_bandwidth_usage_inc(void);
1508 extern void cfs_bandwidth_usage_dec(void);
1510 #ifdef CONFIG_NO_HZ_COMMON
1511 enum rq_nohz_flag_bits {
1516 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1519 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1521 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1522 DECLARE_PER_CPU(u64, cpu_softirq_time);
1524 #ifndef CONFIG_64BIT
1525 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1527 static inline void irq_time_write_begin(void)
1529 __this_cpu_inc(irq_time_seq.sequence);
1533 static inline void irq_time_write_end(void)
1536 __this_cpu_inc(irq_time_seq.sequence);
1539 static inline u64 irq_time_read(int cpu)
1545 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1546 irq_time = per_cpu(cpu_softirq_time, cpu) +
1547 per_cpu(cpu_hardirq_time, cpu);
1548 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1552 #else /* CONFIG_64BIT */
1553 static inline void irq_time_write_begin(void)
1557 static inline void irq_time_write_end(void)
1561 static inline u64 irq_time_read(int cpu)
1563 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1565 #endif /* CONFIG_64BIT */
1566 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */