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 void __dl_clear_params(struct task_struct *p);
137 * To keep the bandwidth of -deadline tasks and groups under control
138 * we need some place where:
139 * - store the maximum -deadline bandwidth of the system (the group);
140 * - cache the fraction of that bandwidth that is currently allocated.
142 * This is all done in the data structure below. It is similar to the
143 * one used for RT-throttling (rt_bandwidth), with the main difference
144 * that, since here we are only interested in admission control, we
145 * do not decrease any runtime while the group "executes", neither we
146 * need a timer to replenish it.
148 * With respect to SMP, the bandwidth is given on a per-CPU basis,
150 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
151 * - dl_total_bw array contains, in the i-eth element, the currently
152 * allocated bandwidth on the i-eth CPU.
153 * Moreover, groups consume bandwidth on each CPU, while tasks only
154 * consume bandwidth on the CPU they're running on.
155 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
156 * that will be shown the next time the proc or cgroup controls will
157 * be red. It on its turn can be changed by writing on its own
160 struct dl_bandwidth {
161 raw_spinlock_t dl_runtime_lock;
166 static inline int dl_bandwidth_enabled(void)
168 return sysctl_sched_rt_runtime >= 0;
171 extern struct dl_bw *dl_bw_of(int i);
178 extern struct mutex sched_domains_mutex;
180 #ifdef CONFIG_CGROUP_SCHED
182 #include <linux/cgroup.h>
187 extern struct list_head task_groups;
189 struct cfs_bandwidth {
190 #ifdef CONFIG_CFS_BANDWIDTH
194 s64 hierarchical_quota;
197 int idle, timer_active;
198 struct hrtimer period_timer, slack_timer;
199 struct list_head throttled_cfs_rq;
202 int nr_periods, nr_throttled;
207 /* task group related information */
209 struct cgroup_subsys_state css;
211 #ifdef CONFIG_FAIR_GROUP_SCHED
212 /* schedulable entities of this group on each cpu */
213 struct sched_entity **se;
214 /* runqueue "owned" by this group on each cpu */
215 struct cfs_rq **cfs_rq;
216 unsigned long shares;
219 atomic_long_t load_avg;
220 atomic_t runnable_avg;
224 #ifdef CONFIG_RT_GROUP_SCHED
225 struct sched_rt_entity **rt_se;
226 struct rt_rq **rt_rq;
228 struct rt_bandwidth rt_bandwidth;
232 struct list_head list;
234 struct task_group *parent;
235 struct list_head siblings;
236 struct list_head children;
238 #ifdef CONFIG_SCHED_AUTOGROUP
239 struct autogroup *autogroup;
242 struct cfs_bandwidth cfs_bandwidth;
245 #ifdef CONFIG_FAIR_GROUP_SCHED
246 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
249 * A weight of 0 or 1 can cause arithmetics problems.
250 * A weight of a cfs_rq is the sum of weights of which entities
251 * are queued on this cfs_rq, so a weight of a entity should not be
252 * too large, so as the shares value of a task group.
253 * (The default weight is 1024 - so there's no practical
254 * limitation from this.)
256 #define MIN_SHARES (1UL << 1)
257 #define MAX_SHARES (1UL << 18)
260 typedef int (*tg_visitor)(struct task_group *, void *);
262 extern int walk_tg_tree_from(struct task_group *from,
263 tg_visitor down, tg_visitor up, void *data);
266 * Iterate the full tree, calling @down when first entering a node and @up when
267 * leaving it for the final time.
269 * Caller must hold rcu_lock or sufficient equivalent.
271 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
273 return walk_tg_tree_from(&root_task_group, down, up, data);
276 extern int tg_nop(struct task_group *tg, void *data);
278 extern void free_fair_sched_group(struct task_group *tg);
279 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
280 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
281 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
282 struct sched_entity *se, int cpu,
283 struct sched_entity *parent);
284 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
285 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
287 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
288 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
289 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
291 extern void free_rt_sched_group(struct task_group *tg);
292 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
293 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
294 struct sched_rt_entity *rt_se, int cpu,
295 struct sched_rt_entity *parent);
297 extern struct task_group *sched_create_group(struct task_group *parent);
298 extern void sched_online_group(struct task_group *tg,
299 struct task_group *parent);
300 extern void sched_destroy_group(struct task_group *tg);
301 extern void sched_offline_group(struct task_group *tg);
303 extern void sched_move_task(struct task_struct *tsk);
305 #ifdef CONFIG_FAIR_GROUP_SCHED
306 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
309 #else /* CONFIG_CGROUP_SCHED */
311 struct cfs_bandwidth { };
313 #endif /* CONFIG_CGROUP_SCHED */
315 /* CFS-related fields in a runqueue */
317 struct load_weight load;
318 unsigned int nr_running, h_nr_running;
323 u64 min_vruntime_copy;
326 struct rb_root tasks_timeline;
327 struct rb_node *rb_leftmost;
330 * 'curr' points to currently running entity on this cfs_rq.
331 * It is set to NULL otherwise (i.e when none are currently running).
333 struct sched_entity *curr, *next, *last, *skip;
335 #ifdef CONFIG_SCHED_DEBUG
336 unsigned int nr_spread_over;
342 * Under CFS, load is tracked on a per-entity basis and aggregated up.
343 * This allows for the description of both thread and group usage (in
344 * the FAIR_GROUP_SCHED case).
346 unsigned long runnable_load_avg, blocked_load_avg;
347 atomic64_t decay_counter;
349 atomic_long_t removed_load;
351 #ifdef CONFIG_FAIR_GROUP_SCHED
352 /* Required to track per-cpu representation of a task_group */
353 u32 tg_runnable_contrib;
354 unsigned long tg_load_contrib;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
362 unsigned long h_load;
363 u64 last_h_load_update;
364 struct sched_entity *h_load_next;
365 #endif /* CONFIG_FAIR_GROUP_SCHED */
366 #endif /* CONFIG_SMP */
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
380 struct list_head leaf_cfs_rq_list;
381 struct task_group *tg; /* group that "owns" this runqueue */
383 #ifdef CONFIG_CFS_BANDWIDTH
386 s64 runtime_remaining;
388 u64 throttled_clock, throttled_clock_task;
389 u64 throttled_clock_task_time;
390 int throttled, throttle_count;
391 struct list_head throttled_list;
392 #endif /* CONFIG_CFS_BANDWIDTH */
393 #endif /* CONFIG_FAIR_GROUP_SCHED */
396 static inline int rt_bandwidth_enabled(void)
398 return sysctl_sched_rt_runtime >= 0;
401 /* Real-Time classes' related field in a runqueue: */
403 struct rt_prio_array active;
404 unsigned int rt_nr_running;
405 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
407 int curr; /* highest queued rt task prio */
409 int next; /* next highest */
414 unsigned long rt_nr_migratory;
415 unsigned long rt_nr_total;
417 struct plist_head pushable_tasks;
424 /* Nests inside the rq lock: */
425 raw_spinlock_t rt_runtime_lock;
427 #ifdef CONFIG_RT_GROUP_SCHED
428 unsigned long rt_nr_boosted;
431 struct task_group *tg;
435 /* Deadline class' related fields in a runqueue */
437 /* runqueue is an rbtree, ordered by deadline */
438 struct rb_root rb_root;
439 struct rb_node *rb_leftmost;
441 unsigned long dl_nr_running;
445 * Deadline values of the currently executing and the
446 * earliest ready task on this rq. Caching these facilitates
447 * the decision wether or not a ready but not running task
448 * should migrate somewhere else.
455 unsigned long dl_nr_migratory;
459 * Tasks on this rq that can be pushed away. They are kept in
460 * an rb-tree, ordered by tasks' deadlines, with caching
461 * of the leftmost (earliest deadline) element.
463 struct rb_root pushable_dl_tasks_root;
464 struct rb_node *pushable_dl_tasks_leftmost;
473 * We add the notion of a root-domain which will be used to define per-domain
474 * variables. Each exclusive cpuset essentially defines an island domain by
475 * fully partitioning the member cpus from any other cpuset. Whenever a new
476 * exclusive cpuset is created, we also create and attach a new root-domain
485 cpumask_var_t online;
487 /* Indicate more than one runnable task for any CPU */
491 * The bit corresponding to a CPU gets set here if such CPU has more
492 * than one runnable -deadline task (as it is below for RT tasks).
494 cpumask_var_t dlo_mask;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask;
504 struct cpupri cpupri;
507 extern struct root_domain def_root_domain;
509 #endif /* CONFIG_SMP */
512 * This is the main, per-CPU runqueue data structure.
514 * Locking rule: those places that want to lock multiple runqueues
515 * (such as the load balancing or the thread migration code), lock
516 * acquire operations must be ordered by ascending &runqueue.
523 * nr_running and cpu_load should be in the same cacheline because
524 * remote CPUs use both these fields when doing load calculation.
526 unsigned int nr_running;
527 #ifdef CONFIG_NUMA_BALANCING
528 unsigned int nr_numa_running;
529 unsigned int nr_preferred_running;
531 #define CPU_LOAD_IDX_MAX 5
532 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
533 unsigned long last_load_update_tick;
534 #ifdef CONFIG_NO_HZ_COMMON
536 unsigned long nohz_flags;
538 #ifdef CONFIG_NO_HZ_FULL
539 unsigned long last_sched_tick;
541 int skip_clock_update;
543 /* capture load from *all* tasks on this cpu: */
544 struct load_weight load;
545 unsigned long nr_load_updates;
552 #ifdef CONFIG_FAIR_GROUP_SCHED
553 /* list of leaf cfs_rq on this cpu: */
554 struct list_head leaf_cfs_rq_list;
556 struct sched_avg avg;
557 #endif /* CONFIG_FAIR_GROUP_SCHED */
560 * This is part of a global counter where only the total sum
561 * over all CPUs matters. A task can increase this counter on
562 * one CPU and if it got migrated afterwards it may decrease
563 * it on another CPU. Always updated under the runqueue lock:
565 unsigned long nr_uninterruptible;
567 struct task_struct *curr, *idle, *stop;
568 unsigned long next_balance;
569 struct mm_struct *prev_mm;
577 struct root_domain *rd;
578 struct sched_domain *sd;
580 unsigned long cpu_capacity;
582 unsigned char idle_balance;
583 /* For active balancing */
587 struct cpu_stop_work active_balance_work;
588 /* cpu of this runqueue: */
592 struct list_head cfs_tasks;
599 /* This is used to determine avg_idle's max value */
600 u64 max_idle_balance_cost;
603 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
606 #ifdef CONFIG_PARAVIRT
609 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
610 u64 prev_steal_time_rq;
613 /* calc_load related fields */
614 unsigned long calc_load_update;
615 long calc_load_active;
617 #ifdef CONFIG_SCHED_HRTICK
619 int hrtick_csd_pending;
620 struct call_single_data hrtick_csd;
622 struct hrtimer hrtick_timer;
625 #ifdef CONFIG_SCHEDSTATS
627 struct sched_info rq_sched_info;
628 unsigned long long rq_cpu_time;
629 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
631 /* sys_sched_yield() stats */
632 unsigned int yld_count;
634 /* schedule() stats */
635 unsigned int sched_count;
636 unsigned int sched_goidle;
638 /* try_to_wake_up() stats */
639 unsigned int ttwu_count;
640 unsigned int ttwu_local;
644 struct llist_head wake_list;
648 static inline int cpu_of(struct rq *rq)
657 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
659 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660 #define this_rq() (&__get_cpu_var(runqueues))
661 #define task_rq(p) cpu_rq(task_cpu(p))
662 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663 #define raw_rq() (&__raw_get_cpu_var(runqueues))
665 static inline u64 rq_clock(struct rq *rq)
670 static inline u64 rq_clock_task(struct rq *rq)
672 return rq->clock_task;
675 #ifdef CONFIG_NUMA_BALANCING
676 extern void sched_setnuma(struct task_struct *p, int node);
677 extern int migrate_task_to(struct task_struct *p, int cpu);
678 extern int migrate_swap(struct task_struct *, struct task_struct *);
679 #endif /* CONFIG_NUMA_BALANCING */
683 extern void sched_ttwu_pending(void);
685 #define rcu_dereference_check_sched_domain(p) \
686 rcu_dereference_check((p), \
687 lockdep_is_held(&sched_domains_mutex))
690 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
691 * See detach_destroy_domains: synchronize_sched for details.
693 * The domain tree of any CPU may only be accessed from within
694 * preempt-disabled sections.
696 #define for_each_domain(cpu, __sd) \
697 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
698 __sd; __sd = __sd->parent)
700 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
703 * highest_flag_domain - Return highest sched_domain containing flag.
704 * @cpu: The cpu whose highest level of sched domain is to
706 * @flag: The flag to check for the highest sched_domain
709 * Returns the highest sched_domain of a cpu which contains the given flag.
711 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
713 struct sched_domain *sd, *hsd = NULL;
715 for_each_domain(cpu, sd) {
716 if (!(sd->flags & flag))
724 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
726 struct sched_domain *sd;
728 for_each_domain(cpu, sd) {
729 if (sd->flags & flag)
736 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
737 DECLARE_PER_CPU(int, sd_llc_size);
738 DECLARE_PER_CPU(int, sd_llc_id);
739 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
740 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
741 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
743 struct sched_group_capacity {
746 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
749 unsigned int capacity, capacity_orig;
750 unsigned long next_update;
751 int imbalance; /* XXX unrelated to capacity but shared group state */
753 * Number of busy cpus in this group.
755 atomic_t nr_busy_cpus;
757 unsigned long cpumask[0]; /* iteration mask */
761 struct sched_group *next; /* Must be a circular list */
764 unsigned int group_weight;
765 struct sched_group_capacity *sgc;
768 * The CPUs this group covers.
770 * NOTE: this field is variable length. (Allocated dynamically
771 * by attaching extra space to the end of the structure,
772 * depending on how many CPUs the kernel has booted up with)
774 unsigned long cpumask[0];
777 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
779 return to_cpumask(sg->cpumask);
783 * cpumask masking which cpus in the group are allowed to iterate up the domain
786 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
788 return to_cpumask(sg->sgc->cpumask);
792 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
793 * @group: The group whose first cpu is to be returned.
795 static inline unsigned int group_first_cpu(struct sched_group *group)
797 return cpumask_first(sched_group_cpus(group));
800 extern int group_balance_cpu(struct sched_group *sg);
804 static inline void sched_ttwu_pending(void) { }
806 #endif /* CONFIG_SMP */
809 #include "auto_group.h"
811 #ifdef CONFIG_CGROUP_SCHED
814 * Return the group to which this tasks belongs.
816 * We cannot use task_css() and friends because the cgroup subsystem
817 * changes that value before the cgroup_subsys::attach() method is called,
818 * therefore we cannot pin it and might observe the wrong value.
820 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
821 * core changes this before calling sched_move_task().
823 * Instead we use a 'copy' which is updated from sched_move_task() while
824 * holding both task_struct::pi_lock and rq::lock.
826 static inline struct task_group *task_group(struct task_struct *p)
828 return p->sched_task_group;
831 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
832 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
834 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
835 struct task_group *tg = task_group(p);
838 #ifdef CONFIG_FAIR_GROUP_SCHED
839 p->se.cfs_rq = tg->cfs_rq[cpu];
840 p->se.parent = tg->se[cpu];
843 #ifdef CONFIG_RT_GROUP_SCHED
844 p->rt.rt_rq = tg->rt_rq[cpu];
845 p->rt.parent = tg->rt_se[cpu];
849 #else /* CONFIG_CGROUP_SCHED */
851 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
852 static inline struct task_group *task_group(struct task_struct *p)
857 #endif /* CONFIG_CGROUP_SCHED */
859 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
864 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
865 * successfuly executed on another CPU. We must ensure that updates of
866 * per-task data have been completed by this moment.
869 task_thread_info(p)->cpu = cpu;
875 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
877 #ifdef CONFIG_SCHED_DEBUG
878 # include <linux/static_key.h>
879 # define const_debug __read_mostly
881 # define const_debug const
884 extern const_debug unsigned int sysctl_sched_features;
886 #define SCHED_FEAT(name, enabled) \
887 __SCHED_FEAT_##name ,
890 #include "features.h"
896 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
897 #define SCHED_FEAT(name, enabled) \
898 static __always_inline bool static_branch_##name(struct static_key *key) \
900 return static_key_##enabled(key); \
903 #include "features.h"
907 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
908 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
909 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
910 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
911 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
913 #ifdef CONFIG_NUMA_BALANCING
914 #define sched_feat_numa(x) sched_feat(x)
915 #ifdef CONFIG_SCHED_DEBUG
916 #define numabalancing_enabled sched_feat_numa(NUMA)
918 extern bool numabalancing_enabled;
919 #endif /* CONFIG_SCHED_DEBUG */
921 #define sched_feat_numa(x) (0)
922 #define numabalancing_enabled (0)
923 #endif /* CONFIG_NUMA_BALANCING */
925 static inline u64 global_rt_period(void)
927 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
930 static inline u64 global_rt_runtime(void)
932 if (sysctl_sched_rt_runtime < 0)
935 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
938 static inline int task_current(struct rq *rq, struct task_struct *p)
940 return rq->curr == p;
943 static inline int task_running(struct rq *rq, struct task_struct *p)
948 return task_current(rq, p);
952 static inline int task_on_rq_queued(struct task_struct *p)
954 return p->on_rq == TASK_ON_RQ_QUEUED;
957 static inline int task_on_rq_migrating(struct task_struct *p)
959 return p->on_rq == TASK_ON_RQ_MIGRATING;
962 #ifndef prepare_arch_switch
963 # define prepare_arch_switch(next) do { } while (0)
965 #ifndef finish_arch_switch
966 # define finish_arch_switch(prev) do { } while (0)
968 #ifndef finish_arch_post_lock_switch
969 # define finish_arch_post_lock_switch() do { } while (0)
972 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
973 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
977 * We can optimise this out completely for !SMP, because the
978 * SMP rebalancing from interrupt is the only thing that cares
985 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
989 * After ->on_cpu is cleared, the task can be moved to a different CPU.
990 * We must ensure this doesn't happen until the switch is completely
996 #ifdef CONFIG_DEBUG_SPINLOCK
997 /* this is a valid case when another task releases the spinlock */
998 rq->lock.owner = current;
1001 * If we are tracking spinlock dependencies then we have to
1002 * fix up the runqueue lock - which gets 'carried over' from
1003 * prev into current:
1005 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1007 raw_spin_unlock_irq(&rq->lock);
1010 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1011 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1015 * We can optimise this out completely for !SMP, because the
1016 * SMP rebalancing from interrupt is the only thing that cares
1021 raw_spin_unlock(&rq->lock);
1024 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1028 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1029 * We must ensure this doesn't happen until the switch is completely
1037 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1042 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1043 #define WF_FORK 0x02 /* child wakeup after fork */
1044 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1047 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1048 * of tasks with abnormal "nice" values across CPUs the contribution that
1049 * each task makes to its run queue's load is weighted according to its
1050 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1051 * scaled version of the new time slice allocation that they receive on time
1055 #define WEIGHT_IDLEPRIO 3
1056 #define WMULT_IDLEPRIO 1431655765
1059 * Nice levels are multiplicative, with a gentle 10% change for every
1060 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1061 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1062 * that remained on nice 0.
1064 * The "10% effect" is relative and cumulative: from _any_ nice level,
1065 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1066 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1067 * If a task goes up by ~10% and another task goes down by ~10% then
1068 * the relative distance between them is ~25%.)
1070 static const int prio_to_weight[40] = {
1071 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1072 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1073 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1074 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1075 /* 0 */ 1024, 820, 655, 526, 423,
1076 /* 5 */ 335, 272, 215, 172, 137,
1077 /* 10 */ 110, 87, 70, 56, 45,
1078 /* 15 */ 36, 29, 23, 18, 15,
1082 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1084 * In cases where the weight does not change often, we can use the
1085 * precalculated inverse to speed up arithmetics by turning divisions
1086 * into multiplications:
1088 static const u32 prio_to_wmult[40] = {
1089 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1090 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1091 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1092 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1093 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1094 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1095 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1096 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1099 #define ENQUEUE_WAKEUP 1
1100 #define ENQUEUE_HEAD 2
1102 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1104 #define ENQUEUE_WAKING 0
1106 #define ENQUEUE_REPLENISH 8
1108 #define DEQUEUE_SLEEP 1
1110 #define RETRY_TASK ((void *)-1UL)
1112 struct sched_class {
1113 const struct sched_class *next;
1115 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1116 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1117 void (*yield_task) (struct rq *rq);
1118 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1120 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1123 * It is the responsibility of the pick_next_task() method that will
1124 * return the next task to call put_prev_task() on the @prev task or
1125 * something equivalent.
1127 * May return RETRY_TASK when it finds a higher prio class has runnable
1130 struct task_struct * (*pick_next_task) (struct rq *rq,
1131 struct task_struct *prev);
1132 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1135 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1136 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1138 void (*post_schedule) (struct rq *this_rq);
1139 void (*task_waking) (struct task_struct *task);
1140 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1142 void (*set_cpus_allowed)(struct task_struct *p,
1143 const struct cpumask *newmask);
1145 void (*rq_online)(struct rq *rq);
1146 void (*rq_offline)(struct rq *rq);
1149 void (*set_curr_task) (struct rq *rq);
1150 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1151 void (*task_fork) (struct task_struct *p);
1152 void (*task_dead) (struct task_struct *p);
1154 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1155 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1156 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1159 unsigned int (*get_rr_interval) (struct rq *rq,
1160 struct task_struct *task);
1162 #ifdef CONFIG_FAIR_GROUP_SCHED
1163 void (*task_move_group) (struct task_struct *p, int on_rq);
1167 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1169 prev->sched_class->put_prev_task(rq, prev);
1172 #define sched_class_highest (&stop_sched_class)
1173 #define for_each_class(class) \
1174 for (class = sched_class_highest; class; class = class->next)
1176 extern const struct sched_class stop_sched_class;
1177 extern const struct sched_class dl_sched_class;
1178 extern const struct sched_class rt_sched_class;
1179 extern const struct sched_class fair_sched_class;
1180 extern const struct sched_class idle_sched_class;
1185 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1187 extern void trigger_load_balance(struct rq *rq);
1189 extern void idle_enter_fair(struct rq *this_rq);
1190 extern void idle_exit_fair(struct rq *this_rq);
1194 static inline void idle_enter_fair(struct rq *rq) { }
1195 static inline void idle_exit_fair(struct rq *rq) { }
1199 extern void sysrq_sched_debug_show(void);
1200 extern void sched_init_granularity(void);
1201 extern void update_max_interval(void);
1203 extern void init_sched_dl_class(void);
1204 extern void init_sched_rt_class(void);
1205 extern void init_sched_fair_class(void);
1206 extern void init_sched_dl_class(void);
1208 extern void resched_curr(struct rq *rq);
1209 extern void resched_cpu(int cpu);
1211 extern struct rt_bandwidth def_rt_bandwidth;
1212 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1214 extern struct dl_bandwidth def_dl_bandwidth;
1215 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1216 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1218 unsigned long to_ratio(u64 period, u64 runtime);
1220 extern void update_idle_cpu_load(struct rq *this_rq);
1222 extern void init_task_runnable_average(struct task_struct *p);
1224 static inline void add_nr_running(struct rq *rq, unsigned count)
1226 unsigned prev_nr = rq->nr_running;
1228 rq->nr_running = prev_nr + count;
1230 if (prev_nr < 2 && rq->nr_running >= 2) {
1232 if (!rq->rd->overload)
1233 rq->rd->overload = true;
1236 #ifdef CONFIG_NO_HZ_FULL
1237 if (tick_nohz_full_cpu(rq->cpu)) {
1239 * Tick is needed if more than one task runs on a CPU.
1240 * Send the target an IPI to kick it out of nohz mode.
1242 * We assume that IPI implies full memory barrier and the
1243 * new value of rq->nr_running is visible on reception
1246 tick_nohz_full_kick_cpu(rq->cpu);
1252 static inline void sub_nr_running(struct rq *rq, unsigned count)
1254 rq->nr_running -= count;
1257 static inline void rq_last_tick_reset(struct rq *rq)
1259 #ifdef CONFIG_NO_HZ_FULL
1260 rq->last_sched_tick = jiffies;
1264 extern void update_rq_clock(struct rq *rq);
1266 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1267 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1269 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1271 extern const_debug unsigned int sysctl_sched_time_avg;
1272 extern const_debug unsigned int sysctl_sched_nr_migrate;
1273 extern const_debug unsigned int sysctl_sched_migration_cost;
1275 static inline u64 sched_avg_period(void)
1277 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1280 #ifdef CONFIG_SCHED_HRTICK
1284 * - enabled by features
1285 * - hrtimer is actually high res
1287 static inline int hrtick_enabled(struct rq *rq)
1289 if (!sched_feat(HRTICK))
1291 if (!cpu_active(cpu_of(rq)))
1293 return hrtimer_is_hres_active(&rq->hrtick_timer);
1296 void hrtick_start(struct rq *rq, u64 delay);
1300 static inline int hrtick_enabled(struct rq *rq)
1305 #endif /* CONFIG_SCHED_HRTICK */
1308 extern void sched_avg_update(struct rq *rq);
1309 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1311 rq->rt_avg += rt_delta;
1312 sched_avg_update(rq);
1315 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1316 static inline void sched_avg_update(struct rq *rq) { }
1319 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1322 #ifdef CONFIG_PREEMPT
1324 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1327 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1328 * way at the expense of forcing extra atomic operations in all
1329 * invocations. This assures that the double_lock is acquired using the
1330 * same underlying policy as the spinlock_t on this architecture, which
1331 * reduces latency compared to the unfair variant below. However, it
1332 * also adds more overhead and therefore may reduce throughput.
1334 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1335 __releases(this_rq->lock)
1336 __acquires(busiest->lock)
1337 __acquires(this_rq->lock)
1339 raw_spin_unlock(&this_rq->lock);
1340 double_rq_lock(this_rq, busiest);
1347 * Unfair double_lock_balance: Optimizes throughput at the expense of
1348 * latency by eliminating extra atomic operations when the locks are
1349 * already in proper order on entry. This favors lower cpu-ids and will
1350 * grant the double lock to lower cpus over higher ids under contention,
1351 * regardless of entry order into the function.
1353 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1354 __releases(this_rq->lock)
1355 __acquires(busiest->lock)
1356 __acquires(this_rq->lock)
1360 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1361 if (busiest < this_rq) {
1362 raw_spin_unlock(&this_rq->lock);
1363 raw_spin_lock(&busiest->lock);
1364 raw_spin_lock_nested(&this_rq->lock,
1365 SINGLE_DEPTH_NESTING);
1368 raw_spin_lock_nested(&busiest->lock,
1369 SINGLE_DEPTH_NESTING);
1374 #endif /* CONFIG_PREEMPT */
1377 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1379 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1381 if (unlikely(!irqs_disabled())) {
1382 /* printk() doesn't work good under rq->lock */
1383 raw_spin_unlock(&this_rq->lock);
1387 return _double_lock_balance(this_rq, busiest);
1390 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1391 __releases(busiest->lock)
1393 raw_spin_unlock(&busiest->lock);
1394 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1397 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1403 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1406 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1412 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1415 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1421 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1425 * double_rq_lock - safely lock two runqueues
1427 * Note this does not disable interrupts like task_rq_lock,
1428 * you need to do so manually before calling.
1430 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1431 __acquires(rq1->lock)
1432 __acquires(rq2->lock)
1434 BUG_ON(!irqs_disabled());
1436 raw_spin_lock(&rq1->lock);
1437 __acquire(rq2->lock); /* Fake it out ;) */
1440 raw_spin_lock(&rq1->lock);
1441 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1443 raw_spin_lock(&rq2->lock);
1444 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1450 * double_rq_unlock - safely unlock two runqueues
1452 * Note this does not restore interrupts like task_rq_unlock,
1453 * you need to do so manually after calling.
1455 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1456 __releases(rq1->lock)
1457 __releases(rq2->lock)
1459 raw_spin_unlock(&rq1->lock);
1461 raw_spin_unlock(&rq2->lock);
1463 __release(rq2->lock);
1466 #else /* CONFIG_SMP */
1469 * double_rq_lock - safely lock two runqueues
1471 * Note this does not disable interrupts like task_rq_lock,
1472 * you need to do so manually before calling.
1474 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1475 __acquires(rq1->lock)
1476 __acquires(rq2->lock)
1478 BUG_ON(!irqs_disabled());
1480 raw_spin_lock(&rq1->lock);
1481 __acquire(rq2->lock); /* Fake it out ;) */
1485 * double_rq_unlock - safely unlock two runqueues
1487 * Note this does not restore interrupts like task_rq_unlock,
1488 * you need to do so manually after calling.
1490 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1491 __releases(rq1->lock)
1492 __releases(rq2->lock)
1495 raw_spin_unlock(&rq1->lock);
1496 __release(rq2->lock);
1501 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1502 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1503 extern void print_cfs_stats(struct seq_file *m, int cpu);
1504 extern void print_rt_stats(struct seq_file *m, int cpu);
1506 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1507 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1508 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1510 extern void cfs_bandwidth_usage_inc(void);
1511 extern void cfs_bandwidth_usage_dec(void);
1513 #ifdef CONFIG_NO_HZ_COMMON
1514 enum rq_nohz_flag_bits {
1519 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1524 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1525 DECLARE_PER_CPU(u64, cpu_softirq_time);
1527 #ifndef CONFIG_64BIT
1528 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1530 static inline void irq_time_write_begin(void)
1532 __this_cpu_inc(irq_time_seq.sequence);
1536 static inline void irq_time_write_end(void)
1539 __this_cpu_inc(irq_time_seq.sequence);
1542 static inline u64 irq_time_read(int cpu)
1548 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1549 irq_time = per_cpu(cpu_softirq_time, cpu) +
1550 per_cpu(cpu_hardirq_time, cpu);
1551 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1555 #else /* CONFIG_64BIT */
1556 static inline void irq_time_write_begin(void)
1560 static inline void irq_time_write_end(void)
1564 static inline u64 irq_time_read(int cpu)
1566 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1568 #endif /* CONFIG_64BIT */
1569 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */