2 #include <linux/sched.h>
3 #include <linux/sched/sysctl.h>
4 #include <linux/sched/rt.h>
5 #include <linux/mutex.h>
6 #include <linux/spinlock.h>
7 #include <linux/stop_machine.h>
8 #include <linux/tick.h>
9 #include <linux/slab.h>
16 extern __read_mostly int scheduler_running;
18 extern unsigned long calc_load_update;
19 extern atomic_long_t calc_load_tasks;
21 extern long calc_load_fold_active(struct rq *this_rq);
22 extern void update_cpu_load_active(struct rq *this_rq);
25 * Convert user-nice values [ -20 ... 0 ... 19 ]
26 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
29 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
30 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
31 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
34 * 'User priority' is the nice value converted to something we
35 * can work with better when scaling various scheduler parameters,
36 * it's a [ 0 ... 39 ] range.
38 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
39 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
40 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
43 * Helpers for converting nanosecond timing to jiffy resolution
45 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
48 * Increase resolution of nice-level calculations for 64-bit architectures.
49 * The extra resolution improves shares distribution and load balancing of
50 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
51 * hierarchies, especially on larger systems. This is not a user-visible change
52 * and does not change the user-interface for setting shares/weights.
54 * We increase resolution only if we have enough bits to allow this increased
55 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
56 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
59 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
60 # define SCHED_LOAD_RESOLUTION 10
61 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
62 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
64 # define SCHED_LOAD_RESOLUTION 0
65 # define scale_load(w) (w)
66 # define scale_load_down(w) (w)
69 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
70 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
72 #define NICE_0_LOAD SCHED_LOAD_SCALE
73 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
76 * These are the 'tuning knobs' of the scheduler:
80 * single value that denotes runtime == period, ie unlimited time.
82 #define RUNTIME_INF ((u64)~0ULL)
84 static inline int rt_policy(int policy)
86 if (policy == SCHED_FIFO || policy == SCHED_RR)
91 static inline int task_has_rt_policy(struct task_struct *p)
93 return rt_policy(p->policy);
97 * This is the priority-queue data structure of the RT scheduling class:
99 struct rt_prio_array {
100 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
101 struct list_head queue[MAX_RT_PRIO];
104 struct rt_bandwidth {
105 /* nests inside the rq lock: */
106 raw_spinlock_t rt_runtime_lock;
109 struct hrtimer rt_period_timer;
112 extern struct mutex sched_domains_mutex;
114 #ifdef CONFIG_CGROUP_SCHED
116 #include <linux/cgroup.h>
121 extern struct list_head task_groups;
123 struct cfs_bandwidth {
124 #ifdef CONFIG_CFS_BANDWIDTH
128 s64 hierarchal_quota;
131 int idle, timer_active;
132 struct hrtimer period_timer, slack_timer;
133 struct list_head throttled_cfs_rq;
136 int nr_periods, nr_throttled;
141 /* task group related information */
143 struct cgroup_subsys_state css;
145 #ifdef CONFIG_FAIR_GROUP_SCHED
146 /* schedulable entities of this group on each cpu */
147 struct sched_entity **se;
148 /* runqueue "owned" by this group on each cpu */
149 struct cfs_rq **cfs_rq;
150 unsigned long shares;
153 atomic_long_t load_avg;
154 atomic_t runnable_avg;
158 #ifdef CONFIG_RT_GROUP_SCHED
159 struct sched_rt_entity **rt_se;
160 struct rt_rq **rt_rq;
162 struct rt_bandwidth rt_bandwidth;
166 struct list_head list;
168 struct task_group *parent;
169 struct list_head siblings;
170 struct list_head children;
172 #ifdef CONFIG_SCHED_AUTOGROUP
173 struct autogroup *autogroup;
176 struct cfs_bandwidth cfs_bandwidth;
179 #ifdef CONFIG_FAIR_GROUP_SCHED
180 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
183 * A weight of 0 or 1 can cause arithmetics problems.
184 * A weight of a cfs_rq is the sum of weights of which entities
185 * are queued on this cfs_rq, so a weight of a entity should not be
186 * too large, so as the shares value of a task group.
187 * (The default weight is 1024 - so there's no practical
188 * limitation from this.)
190 #define MIN_SHARES (1UL << 1)
191 #define MAX_SHARES (1UL << 18)
194 typedef int (*tg_visitor)(struct task_group *, void *);
196 extern int walk_tg_tree_from(struct task_group *from,
197 tg_visitor down, tg_visitor up, void *data);
200 * Iterate the full tree, calling @down when first entering a node and @up when
201 * leaving it for the final time.
203 * Caller must hold rcu_lock or sufficient equivalent.
205 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
207 return walk_tg_tree_from(&root_task_group, down, up, data);
210 extern int tg_nop(struct task_group *tg, void *data);
212 extern void free_fair_sched_group(struct task_group *tg);
213 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
214 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
215 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
216 struct sched_entity *se, int cpu,
217 struct sched_entity *parent);
218 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
219 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
221 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
222 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
223 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
225 extern void free_rt_sched_group(struct task_group *tg);
226 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
227 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
228 struct sched_rt_entity *rt_se, int cpu,
229 struct sched_rt_entity *parent);
231 extern struct task_group *sched_create_group(struct task_group *parent);
232 extern void sched_online_group(struct task_group *tg,
233 struct task_group *parent);
234 extern void sched_destroy_group(struct task_group *tg);
235 extern void sched_offline_group(struct task_group *tg);
237 extern void sched_move_task(struct task_struct *tsk);
239 #ifdef CONFIG_FAIR_GROUP_SCHED
240 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
243 #else /* CONFIG_CGROUP_SCHED */
245 struct cfs_bandwidth { };
247 #endif /* CONFIG_CGROUP_SCHED */
249 /* CFS-related fields in a runqueue */
251 struct load_weight load;
252 unsigned int nr_running, h_nr_running;
257 u64 min_vruntime_copy;
260 struct rb_root tasks_timeline;
261 struct rb_node *rb_leftmost;
264 * 'curr' points to currently running entity on this cfs_rq.
265 * It is set to NULL otherwise (i.e when none are currently running).
267 struct sched_entity *curr, *next, *last, *skip;
269 #ifdef CONFIG_SCHED_DEBUG
270 unsigned int nr_spread_over;
276 * Under CFS, load is tracked on a per-entity basis and aggregated up.
277 * This allows for the description of both thread and group usage (in
278 * the FAIR_GROUP_SCHED case).
280 unsigned long runnable_load_avg, blocked_load_avg;
281 atomic64_t decay_counter;
283 atomic_long_t removed_load;
285 #ifdef CONFIG_FAIR_GROUP_SCHED
286 /* Required to track per-cpu representation of a task_group */
287 u32 tg_runnable_contrib;
288 unsigned long tg_load_contrib;
291 * h_load = weight * f(tg)
293 * Where f(tg) is the recursive weight fraction assigned to
296 unsigned long h_load;
297 u64 last_h_load_update;
298 struct sched_entity *h_load_next;
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
300 #endif /* CONFIG_SMP */
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
306 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
307 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
308 * (like users, containers etc.)
310 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
311 * list is used during load balance.
314 struct list_head leaf_cfs_rq_list;
315 struct task_group *tg; /* group that "owns" this runqueue */
317 #ifdef CONFIG_CFS_BANDWIDTH
320 s64 runtime_remaining;
322 u64 throttled_clock, throttled_clock_task;
323 u64 throttled_clock_task_time;
324 int throttled, throttle_count;
325 struct list_head throttled_list;
326 #endif /* CONFIG_CFS_BANDWIDTH */
327 #endif /* CONFIG_FAIR_GROUP_SCHED */
330 static inline int rt_bandwidth_enabled(void)
332 return sysctl_sched_rt_runtime >= 0;
335 /* Real-Time classes' related field in a runqueue: */
337 struct rt_prio_array active;
338 unsigned int rt_nr_running;
339 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
341 int curr; /* highest queued rt task prio */
343 int next; /* next highest */
348 unsigned long rt_nr_migratory;
349 unsigned long rt_nr_total;
351 struct plist_head pushable_tasks;
356 /* Nests inside the rq lock: */
357 raw_spinlock_t rt_runtime_lock;
359 #ifdef CONFIG_RT_GROUP_SCHED
360 unsigned long rt_nr_boosted;
363 struct task_group *tg;
370 * We add the notion of a root-domain which will be used to define per-domain
371 * variables. Each exclusive cpuset essentially defines an island domain by
372 * fully partitioning the member cpus from any other cpuset. Whenever a new
373 * exclusive cpuset is created, we also create and attach a new root-domain
382 cpumask_var_t online;
385 * The "RT overload" flag: it gets set if a CPU has more than
386 * one runnable RT task.
388 cpumask_var_t rto_mask;
389 struct cpupri cpupri;
392 extern struct root_domain def_root_domain;
394 #endif /* CONFIG_SMP */
397 * This is the main, per-CPU runqueue data structure.
399 * Locking rule: those places that want to lock multiple runqueues
400 * (such as the load balancing or the thread migration code), lock
401 * acquire operations must be ordered by ascending &runqueue.
408 * nr_running and cpu_load should be in the same cacheline because
409 * remote CPUs use both these fields when doing load calculation.
411 unsigned int nr_running;
412 #ifdef CONFIG_NUMA_BALANCING
413 unsigned int nr_numa_running;
414 unsigned int nr_preferred_running;
416 #define CPU_LOAD_IDX_MAX 5
417 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
418 unsigned long last_load_update_tick;
419 #ifdef CONFIG_NO_HZ_COMMON
421 unsigned long nohz_flags;
423 #ifdef CONFIG_NO_HZ_FULL
424 unsigned long last_sched_tick;
426 int skip_clock_update;
428 /* capture load from *all* tasks on this cpu: */
429 struct load_weight load;
430 unsigned long nr_load_updates;
436 #ifdef CONFIG_FAIR_GROUP_SCHED
437 /* list of leaf cfs_rq on this cpu: */
438 struct list_head leaf_cfs_rq_list;
439 #endif /* CONFIG_FAIR_GROUP_SCHED */
441 #ifdef CONFIG_RT_GROUP_SCHED
442 struct list_head leaf_rt_rq_list;
446 * This is part of a global counter where only the total sum
447 * over all CPUs matters. A task can increase this counter on
448 * one CPU and if it got migrated afterwards it may decrease
449 * it on another CPU. Always updated under the runqueue lock:
451 unsigned long nr_uninterruptible;
453 struct task_struct *curr, *idle, *stop;
454 unsigned long next_balance;
455 struct mm_struct *prev_mm;
463 struct root_domain *rd;
464 struct sched_domain *sd;
466 unsigned long cpu_power;
468 unsigned char idle_balance;
469 /* For active balancing */
473 struct cpu_stop_work active_balance_work;
474 /* cpu of this runqueue: */
478 struct list_head cfs_tasks;
485 /* This is used to determine avg_idle's max value */
486 u64 max_idle_balance_cost;
489 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
492 #ifdef CONFIG_PARAVIRT
495 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
496 u64 prev_steal_time_rq;
499 /* calc_load related fields */
500 unsigned long calc_load_update;
501 long calc_load_active;
503 #ifdef CONFIG_SCHED_HRTICK
505 int hrtick_csd_pending;
506 struct call_single_data hrtick_csd;
508 struct hrtimer hrtick_timer;
511 #ifdef CONFIG_SCHEDSTATS
513 struct sched_info rq_sched_info;
514 unsigned long long rq_cpu_time;
515 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
517 /* sys_sched_yield() stats */
518 unsigned int yld_count;
520 /* schedule() stats */
521 unsigned int sched_count;
522 unsigned int sched_goidle;
524 /* try_to_wake_up() stats */
525 unsigned int ttwu_count;
526 unsigned int ttwu_local;
530 struct llist_head wake_list;
533 struct sched_avg avg;
536 static inline int cpu_of(struct rq *rq)
545 DECLARE_PER_CPU(struct rq, runqueues);
547 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
548 #define this_rq() (&__get_cpu_var(runqueues))
549 #define task_rq(p) cpu_rq(task_cpu(p))
550 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
551 #define raw_rq() (&__raw_get_cpu_var(runqueues))
553 static inline u64 rq_clock(struct rq *rq)
558 static inline u64 rq_clock_task(struct rq *rq)
560 return rq->clock_task;
563 #ifdef CONFIG_NUMA_BALANCING
564 extern void sched_setnuma(struct task_struct *p, int node);
565 extern int migrate_task_to(struct task_struct *p, int cpu);
566 extern int migrate_swap(struct task_struct *, struct task_struct *);
567 #endif /* CONFIG_NUMA_BALANCING */
571 #define rcu_dereference_check_sched_domain(p) \
572 rcu_dereference_check((p), \
573 lockdep_is_held(&sched_domains_mutex))
576 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
577 * See detach_destroy_domains: synchronize_sched for details.
579 * The domain tree of any CPU may only be accessed from within
580 * preempt-disabled sections.
582 #define for_each_domain(cpu, __sd) \
583 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
584 __sd; __sd = __sd->parent)
586 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
589 * highest_flag_domain - Return highest sched_domain containing flag.
590 * @cpu: The cpu whose highest level of sched domain is to
592 * @flag: The flag to check for the highest sched_domain
595 * Returns the highest sched_domain of a cpu which contains the given flag.
597 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
599 struct sched_domain *sd, *hsd = NULL;
601 for_each_domain(cpu, sd) {
602 if (!(sd->flags & flag))
610 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
612 struct sched_domain *sd;
614 for_each_domain(cpu, sd) {
615 if (sd->flags & flag)
622 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
623 DECLARE_PER_CPU(int, sd_llc_size);
624 DECLARE_PER_CPU(int, sd_llc_id);
625 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
627 struct sched_group_power {
630 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
633 unsigned int power, power_orig;
634 unsigned long next_update;
635 int imbalance; /* XXX unrelated to power but shared group state */
637 * Number of busy cpus in this group.
639 atomic_t nr_busy_cpus;
641 unsigned long cpumask[0]; /* iteration mask */
645 struct sched_group *next; /* Must be a circular list */
648 unsigned int group_weight;
649 struct sched_group_power *sgp;
652 * The CPUs this group covers.
654 * NOTE: this field is variable length. (Allocated dynamically
655 * by attaching extra space to the end of the structure,
656 * depending on how many CPUs the kernel has booted up with)
658 unsigned long cpumask[0];
661 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
663 return to_cpumask(sg->cpumask);
667 * cpumask masking which cpus in the group are allowed to iterate up the domain
670 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
672 return to_cpumask(sg->sgp->cpumask);
676 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
677 * @group: The group whose first cpu is to be returned.
679 static inline unsigned int group_first_cpu(struct sched_group *group)
681 return cpumask_first(sched_group_cpus(group));
684 extern int group_balance_cpu(struct sched_group *sg);
686 #endif /* CONFIG_SMP */
689 #include "auto_group.h"
691 #ifdef CONFIG_CGROUP_SCHED
694 * Return the group to which this tasks belongs.
696 * We cannot use task_css() and friends because the cgroup subsystem
697 * changes that value before the cgroup_subsys::attach() method is called,
698 * therefore we cannot pin it and might observe the wrong value.
700 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
701 * core changes this before calling sched_move_task().
703 * Instead we use a 'copy' which is updated from sched_move_task() while
704 * holding both task_struct::pi_lock and rq::lock.
706 static inline struct task_group *task_group(struct task_struct *p)
708 return p->sched_task_group;
711 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
712 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
714 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
715 struct task_group *tg = task_group(p);
718 #ifdef CONFIG_FAIR_GROUP_SCHED
719 p->se.cfs_rq = tg->cfs_rq[cpu];
720 p->se.parent = tg->se[cpu];
723 #ifdef CONFIG_RT_GROUP_SCHED
724 p->rt.rt_rq = tg->rt_rq[cpu];
725 p->rt.parent = tg->rt_se[cpu];
729 #else /* CONFIG_CGROUP_SCHED */
731 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
732 static inline struct task_group *task_group(struct task_struct *p)
737 #endif /* CONFIG_CGROUP_SCHED */
739 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
744 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
745 * successfuly executed on another CPU. We must ensure that updates of
746 * per-task data have been completed by this moment.
749 task_thread_info(p)->cpu = cpu;
755 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
757 #ifdef CONFIG_SCHED_DEBUG
758 # include <linux/static_key.h>
759 # define const_debug __read_mostly
761 # define const_debug const
764 extern const_debug unsigned int sysctl_sched_features;
766 #define SCHED_FEAT(name, enabled) \
767 __SCHED_FEAT_##name ,
770 #include "features.h"
776 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
777 static __always_inline bool static_branch__true(struct static_key *key)
779 return static_key_true(key); /* Not out of line branch. */
782 static __always_inline bool static_branch__false(struct static_key *key)
784 return static_key_false(key); /* Out of line branch. */
787 #define SCHED_FEAT(name, enabled) \
788 static __always_inline bool static_branch_##name(struct static_key *key) \
790 return static_branch__##enabled(key); \
793 #include "features.h"
797 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
798 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
799 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
800 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
801 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
803 #ifdef CONFIG_NUMA_BALANCING
804 #define sched_feat_numa(x) sched_feat(x)
805 #ifdef CONFIG_SCHED_DEBUG
806 #define numabalancing_enabled sched_feat_numa(NUMA)
808 extern bool numabalancing_enabled;
809 #endif /* CONFIG_SCHED_DEBUG */
811 #define sched_feat_numa(x) (0)
812 #define numabalancing_enabled (0)
813 #endif /* CONFIG_NUMA_BALANCING */
815 static inline u64 global_rt_period(void)
817 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
820 static inline u64 global_rt_runtime(void)
822 if (sysctl_sched_rt_runtime < 0)
825 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
830 static inline int task_current(struct rq *rq, struct task_struct *p)
832 return rq->curr == p;
835 static inline int task_running(struct rq *rq, struct task_struct *p)
840 return task_current(rq, p);
845 #ifndef prepare_arch_switch
846 # define prepare_arch_switch(next) do { } while (0)
848 #ifndef finish_arch_switch
849 # define finish_arch_switch(prev) do { } while (0)
851 #ifndef finish_arch_post_lock_switch
852 # define finish_arch_post_lock_switch() do { } while (0)
855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
856 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
860 * We can optimise this out completely for !SMP, because the
861 * SMP rebalancing from interrupt is the only thing that cares
868 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
872 * After ->on_cpu is cleared, the task can be moved to a different CPU.
873 * We must ensure this doesn't happen until the switch is completely
879 #ifdef CONFIG_DEBUG_SPINLOCK
880 /* this is a valid case when another task releases the spinlock */
881 rq->lock.owner = current;
884 * If we are tracking spinlock dependencies then we have to
885 * fix up the runqueue lock - which gets 'carried over' from
888 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
890 raw_spin_unlock_irq(&rq->lock);
893 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
894 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
898 * We can optimise this out completely for !SMP, because the
899 * SMP rebalancing from interrupt is the only thing that cares
904 raw_spin_unlock(&rq->lock);
907 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
911 * After ->on_cpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
920 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
926 #define WF_FORK 0x02 /* child wakeup after fork */
927 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
930 * To aid in avoiding the subversion of "niceness" due to uneven distribution
931 * of tasks with abnormal "nice" values across CPUs the contribution that
932 * each task makes to its run queue's load is weighted according to its
933 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
934 * scaled version of the new time slice allocation that they receive on time
938 #define WEIGHT_IDLEPRIO 3
939 #define WMULT_IDLEPRIO 1431655765
942 * Nice levels are multiplicative, with a gentle 10% change for every
943 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
944 * nice 1, it will get ~10% less CPU time than another CPU-bound task
945 * that remained on nice 0.
947 * The "10% effect" is relative and cumulative: from _any_ nice level,
948 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
949 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
950 * If a task goes up by ~10% and another task goes down by ~10% then
951 * the relative distance between them is ~25%.)
953 static const int prio_to_weight[40] = {
954 /* -20 */ 88761, 71755, 56483, 46273, 36291,
955 /* -15 */ 29154, 23254, 18705, 14949, 11916,
956 /* -10 */ 9548, 7620, 6100, 4904, 3906,
957 /* -5 */ 3121, 2501, 1991, 1586, 1277,
958 /* 0 */ 1024, 820, 655, 526, 423,
959 /* 5 */ 335, 272, 215, 172, 137,
960 /* 10 */ 110, 87, 70, 56, 45,
961 /* 15 */ 36, 29, 23, 18, 15,
965 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
967 * In cases where the weight does not change often, we can use the
968 * precalculated inverse to speed up arithmetics by turning divisions
969 * into multiplications:
971 static const u32 prio_to_wmult[40] = {
972 /* -20 */ 48388, 59856, 76040, 92818, 118348,
973 /* -15 */ 147320, 184698, 229616, 287308, 360437,
974 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
975 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
976 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
977 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
978 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
979 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
982 #define ENQUEUE_WAKEUP 1
983 #define ENQUEUE_HEAD 2
985 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
987 #define ENQUEUE_WAKING 0
990 #define DEQUEUE_SLEEP 1
993 const struct sched_class *next;
995 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
996 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
997 void (*yield_task) (struct rq *rq);
998 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1000 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1002 struct task_struct * (*pick_next_task) (struct rq *rq);
1003 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1006 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1007 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1009 void (*pre_schedule) (struct rq *this_rq, struct task_struct *task);
1010 void (*post_schedule) (struct rq *this_rq);
1011 void (*task_waking) (struct task_struct *task);
1012 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1014 void (*set_cpus_allowed)(struct task_struct *p,
1015 const struct cpumask *newmask);
1017 void (*rq_online)(struct rq *rq);
1018 void (*rq_offline)(struct rq *rq);
1021 void (*set_curr_task) (struct rq *rq);
1022 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1023 void (*task_fork) (struct task_struct *p);
1025 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1026 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1027 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1030 unsigned int (*get_rr_interval) (struct rq *rq,
1031 struct task_struct *task);
1033 #ifdef CONFIG_FAIR_GROUP_SCHED
1034 void (*task_move_group) (struct task_struct *p, int on_rq);
1038 #define sched_class_highest (&stop_sched_class)
1039 #define for_each_class(class) \
1040 for (class = sched_class_highest; class; class = class->next)
1042 extern const struct sched_class stop_sched_class;
1043 extern const struct sched_class rt_sched_class;
1044 extern const struct sched_class fair_sched_class;
1045 extern const struct sched_class idle_sched_class;
1050 extern void update_group_power(struct sched_domain *sd, int cpu);
1052 extern void trigger_load_balance(struct rq *rq, int cpu);
1053 extern void idle_balance(int this_cpu, struct rq *this_rq);
1055 extern void idle_enter_fair(struct rq *this_rq);
1056 extern void idle_exit_fair(struct rq *this_rq);
1058 #else /* CONFIG_SMP */
1060 static inline void idle_balance(int cpu, struct rq *rq)
1066 extern void sysrq_sched_debug_show(void);
1067 extern void sched_init_granularity(void);
1068 extern void update_max_interval(void);
1069 extern void init_sched_rt_class(void);
1070 extern void init_sched_fair_class(void);
1072 extern void resched_task(struct task_struct *p);
1073 extern void resched_cpu(int cpu);
1075 extern struct rt_bandwidth def_rt_bandwidth;
1076 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1078 extern void update_idle_cpu_load(struct rq *this_rq);
1080 extern void init_task_runnable_average(struct task_struct *p);
1082 #ifdef CONFIG_PARAVIRT
1083 static inline u64 steal_ticks(u64 steal)
1085 if (unlikely(steal > NSEC_PER_SEC))
1086 return div_u64(steal, TICK_NSEC);
1088 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1092 static inline void inc_nr_running(struct rq *rq)
1096 #ifdef CONFIG_NO_HZ_FULL
1097 if (rq->nr_running == 2) {
1098 if (tick_nohz_full_cpu(rq->cpu)) {
1099 /* Order rq->nr_running write against the IPI */
1101 smp_send_reschedule(rq->cpu);
1107 static inline void dec_nr_running(struct rq *rq)
1112 static inline void rq_last_tick_reset(struct rq *rq)
1114 #ifdef CONFIG_NO_HZ_FULL
1115 rq->last_sched_tick = jiffies;
1119 extern void update_rq_clock(struct rq *rq);
1121 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1122 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1124 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1126 extern const_debug unsigned int sysctl_sched_time_avg;
1127 extern const_debug unsigned int sysctl_sched_nr_migrate;
1128 extern const_debug unsigned int sysctl_sched_migration_cost;
1130 static inline u64 sched_avg_period(void)
1132 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1135 #ifdef CONFIG_SCHED_HRTICK
1139 * - enabled by features
1140 * - hrtimer is actually high res
1142 static inline int hrtick_enabled(struct rq *rq)
1144 if (!sched_feat(HRTICK))
1146 if (!cpu_active(cpu_of(rq)))
1148 return hrtimer_is_hres_active(&rq->hrtick_timer);
1151 void hrtick_start(struct rq *rq, u64 delay);
1155 static inline int hrtick_enabled(struct rq *rq)
1160 #endif /* CONFIG_SCHED_HRTICK */
1163 extern void sched_avg_update(struct rq *rq);
1164 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1166 rq->rt_avg += rt_delta;
1167 sched_avg_update(rq);
1170 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1171 static inline void sched_avg_update(struct rq *rq) { }
1174 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1177 #ifdef CONFIG_PREEMPT
1179 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1182 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1183 * way at the expense of forcing extra atomic operations in all
1184 * invocations. This assures that the double_lock is acquired using the
1185 * same underlying policy as the spinlock_t on this architecture, which
1186 * reduces latency compared to the unfair variant below. However, it
1187 * also adds more overhead and therefore may reduce throughput.
1189 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1190 __releases(this_rq->lock)
1191 __acquires(busiest->lock)
1192 __acquires(this_rq->lock)
1194 raw_spin_unlock(&this_rq->lock);
1195 double_rq_lock(this_rq, busiest);
1202 * Unfair double_lock_balance: Optimizes throughput at the expense of
1203 * latency by eliminating extra atomic operations when the locks are
1204 * already in proper order on entry. This favors lower cpu-ids and will
1205 * grant the double lock to lower cpus over higher ids under contention,
1206 * regardless of entry order into the function.
1208 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1209 __releases(this_rq->lock)
1210 __acquires(busiest->lock)
1211 __acquires(this_rq->lock)
1215 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1216 if (busiest < this_rq) {
1217 raw_spin_unlock(&this_rq->lock);
1218 raw_spin_lock(&busiest->lock);
1219 raw_spin_lock_nested(&this_rq->lock,
1220 SINGLE_DEPTH_NESTING);
1223 raw_spin_lock_nested(&busiest->lock,
1224 SINGLE_DEPTH_NESTING);
1229 #endif /* CONFIG_PREEMPT */
1232 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1234 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1236 if (unlikely(!irqs_disabled())) {
1237 /* printk() doesn't work good under rq->lock */
1238 raw_spin_unlock(&this_rq->lock);
1242 return _double_lock_balance(this_rq, busiest);
1245 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1246 __releases(busiest->lock)
1248 raw_spin_unlock(&busiest->lock);
1249 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1252 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1258 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1261 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1267 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1271 * double_rq_lock - safely lock two runqueues
1273 * Note this does not disable interrupts like task_rq_lock,
1274 * you need to do so manually before calling.
1276 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1277 __acquires(rq1->lock)
1278 __acquires(rq2->lock)
1280 BUG_ON(!irqs_disabled());
1282 raw_spin_lock(&rq1->lock);
1283 __acquire(rq2->lock); /* Fake it out ;) */
1286 raw_spin_lock(&rq1->lock);
1287 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1289 raw_spin_lock(&rq2->lock);
1290 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1296 * double_rq_unlock - safely unlock two runqueues
1298 * Note this does not restore interrupts like task_rq_unlock,
1299 * you need to do so manually after calling.
1301 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1302 __releases(rq1->lock)
1303 __releases(rq2->lock)
1305 raw_spin_unlock(&rq1->lock);
1307 raw_spin_unlock(&rq2->lock);
1309 __release(rq2->lock);
1312 #else /* CONFIG_SMP */
1315 * double_rq_lock - safely lock two runqueues
1317 * Note this does not disable interrupts like task_rq_lock,
1318 * you need to do so manually before calling.
1320 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1321 __acquires(rq1->lock)
1322 __acquires(rq2->lock)
1324 BUG_ON(!irqs_disabled());
1326 raw_spin_lock(&rq1->lock);
1327 __acquire(rq2->lock); /* Fake it out ;) */
1331 * double_rq_unlock - safely unlock two runqueues
1333 * Note this does not restore interrupts like task_rq_unlock,
1334 * you need to do so manually after calling.
1336 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1337 __releases(rq1->lock)
1338 __releases(rq2->lock)
1341 raw_spin_unlock(&rq1->lock);
1342 __release(rq2->lock);
1347 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1348 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1349 extern void print_cfs_stats(struct seq_file *m, int cpu);
1350 extern void print_rt_stats(struct seq_file *m, int cpu);
1352 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1353 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1355 extern void account_cfs_bandwidth_used(int enabled, int was_enabled);
1357 #ifdef CONFIG_NO_HZ_COMMON
1358 enum rq_nohz_flag_bits {
1363 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1366 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1368 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1369 DECLARE_PER_CPU(u64, cpu_softirq_time);
1371 #ifndef CONFIG_64BIT
1372 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1374 static inline void irq_time_write_begin(void)
1376 __this_cpu_inc(irq_time_seq.sequence);
1380 static inline void irq_time_write_end(void)
1383 __this_cpu_inc(irq_time_seq.sequence);
1386 static inline u64 irq_time_read(int cpu)
1392 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1393 irq_time = per_cpu(cpu_softirq_time, cpu) +
1394 per_cpu(cpu_hardirq_time, cpu);
1395 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1399 #else /* CONFIG_64BIT */
1400 static inline void irq_time_write_begin(void)
1404 static inline void irq_time_write_end(void)
1408 static inline u64 irq_time_read(int cpu)
1410 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1412 #endif /* CONFIG_64BIT */
1413 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */