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>
13 extern __read_mostly int scheduler_running;
16 * Convert user-nice values [ -20 ... 0 ... 19 ]
17 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
20 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
21 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
22 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
25 * 'User priority' is the nice value converted to something we
26 * can work with better when scaling various scheduler parameters,
27 * it's a [ 0 ... 39 ] range.
29 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
30 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
31 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
34 * Helpers for converting nanosecond timing to jiffy resolution
36 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
39 * Increase resolution of nice-level calculations for 64-bit architectures.
40 * The extra resolution improves shares distribution and load balancing of
41 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
42 * hierarchies, especially on larger systems. This is not a user-visible change
43 * and does not change the user-interface for setting shares/weights.
45 * We increase resolution only if we have enough bits to allow this increased
46 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
47 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
50 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
51 # define SCHED_LOAD_RESOLUTION 10
52 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
53 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
55 # define SCHED_LOAD_RESOLUTION 0
56 # define scale_load(w) (w)
57 # define scale_load_down(w) (w)
60 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
61 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
63 #define NICE_0_LOAD SCHED_LOAD_SCALE
64 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
67 * These are the 'tuning knobs' of the scheduler:
71 * single value that denotes runtime == period, ie unlimited time.
73 #define RUNTIME_INF ((u64)~0ULL)
75 static inline int rt_policy(int policy)
77 if (policy == SCHED_FIFO || policy == SCHED_RR)
82 static inline int task_has_rt_policy(struct task_struct *p)
84 return rt_policy(p->policy);
88 * This is the priority-queue data structure of the RT scheduling class:
90 struct rt_prio_array {
91 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
92 struct list_head queue[MAX_RT_PRIO];
96 /* nests inside the rq lock: */
97 raw_spinlock_t rt_runtime_lock;
100 struct hrtimer rt_period_timer;
103 extern struct mutex sched_domains_mutex;
105 #ifdef CONFIG_CGROUP_SCHED
107 #include <linux/cgroup.h>
112 extern struct list_head task_groups;
114 struct cfs_bandwidth {
115 #ifdef CONFIG_CFS_BANDWIDTH
119 s64 hierarchal_quota;
122 int idle, timer_active;
123 struct hrtimer period_timer, slack_timer;
124 struct list_head throttled_cfs_rq;
127 int nr_periods, nr_throttled;
132 /* task group related information */
134 struct cgroup_subsys_state css;
136 #ifdef CONFIG_FAIR_GROUP_SCHED
137 /* schedulable entities of this group on each cpu */
138 struct sched_entity **se;
139 /* runqueue "owned" by this group on each cpu */
140 struct cfs_rq **cfs_rq;
141 unsigned long shares;
143 atomic_t load_weight;
145 atomic_t runnable_avg, usage_avg;
148 #ifdef CONFIG_RT_GROUP_SCHED
149 struct sched_rt_entity **rt_se;
150 struct rt_rq **rt_rq;
152 struct rt_bandwidth rt_bandwidth;
156 struct list_head list;
158 struct task_group *parent;
159 struct list_head siblings;
160 struct list_head children;
162 #ifdef CONFIG_SCHED_AUTOGROUP
163 struct autogroup *autogroup;
166 struct cfs_bandwidth cfs_bandwidth;
169 #ifdef CONFIG_FAIR_GROUP_SCHED
170 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
173 * A weight of 0 or 1 can cause arithmetics problems.
174 * A weight of a cfs_rq is the sum of weights of which entities
175 * are queued on this cfs_rq, so a weight of a entity should not be
176 * too large, so as the shares value of a task group.
177 * (The default weight is 1024 - so there's no practical
178 * limitation from this.)
180 #define MIN_SHARES (1UL << 1)
181 #define MAX_SHARES (1UL << 18)
184 typedef int (*tg_visitor)(struct task_group *, void *);
186 extern int walk_tg_tree_from(struct task_group *from,
187 tg_visitor down, tg_visitor up, void *data);
190 * Iterate the full tree, calling @down when first entering a node and @up when
191 * leaving it for the final time.
193 * Caller must hold rcu_lock or sufficient equivalent.
195 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
197 return walk_tg_tree_from(&root_task_group, down, up, data);
200 extern int tg_nop(struct task_group *tg, void *data);
202 extern void free_fair_sched_group(struct task_group *tg);
203 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
204 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
205 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
206 struct sched_entity *se, int cpu,
207 struct sched_entity *parent);
208 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
209 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
211 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
212 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
213 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
215 extern void free_rt_sched_group(struct task_group *tg);
216 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
217 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
218 struct sched_rt_entity *rt_se, int cpu,
219 struct sched_rt_entity *parent);
221 extern struct task_group *sched_create_group(struct task_group *parent);
222 extern void sched_online_group(struct task_group *tg,
223 struct task_group *parent);
224 extern void sched_destroy_group(struct task_group *tg);
225 extern void sched_offline_group(struct task_group *tg);
227 extern void sched_move_task(struct task_struct *tsk);
229 #ifdef CONFIG_FAIR_GROUP_SCHED
230 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
233 #else /* CONFIG_CGROUP_SCHED */
235 struct cfs_bandwidth { };
237 #endif /* CONFIG_CGROUP_SCHED */
239 /* CFS-related fields in a runqueue */
241 struct load_weight load;
242 unsigned int nr_running, h_nr_running;
247 u64 min_vruntime_copy;
250 struct rb_root tasks_timeline;
251 struct rb_node *rb_leftmost;
254 * 'curr' points to currently running entity on this cfs_rq.
255 * It is set to NULL otherwise (i.e when none are currently running).
257 struct sched_entity *curr, *next, *last, *skip;
259 #ifdef CONFIG_SCHED_DEBUG
260 unsigned int nr_spread_over;
265 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
266 * removed when useful for applications beyond shares distribution (e.g.
269 #ifdef CONFIG_FAIR_GROUP_SCHED
272 * Under CFS, load is tracked on a per-entity basis and aggregated up.
273 * This allows for the description of both thread and group usage (in
274 * the FAIR_GROUP_SCHED case).
276 u64 runnable_load_avg, blocked_load_avg;
277 atomic64_t decay_counter, removed_load;
279 #endif /* CONFIG_FAIR_GROUP_SCHED */
280 /* These always depend on CONFIG_FAIR_GROUP_SCHED */
281 #ifdef CONFIG_FAIR_GROUP_SCHED
282 u32 tg_runnable_contrib, tg_usage_contrib;
284 #endif /* CONFIG_FAIR_GROUP_SCHED */
287 * h_load = weight * f(tg)
289 * Where f(tg) is the recursive weight fraction assigned to
292 unsigned long h_load;
293 #endif /* CONFIG_SMP */
295 #ifdef CONFIG_FAIR_GROUP_SCHED
296 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
299 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
300 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
301 * (like users, containers etc.)
303 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
304 * list is used during load balance.
307 struct list_head leaf_cfs_rq_list;
308 struct task_group *tg; /* group that "owns" this runqueue */
310 #ifdef CONFIG_CFS_BANDWIDTH
313 s64 runtime_remaining;
315 u64 throttled_clock, throttled_clock_task;
316 u64 throttled_clock_task_time;
317 int throttled, throttle_count;
318 struct list_head throttled_list;
319 #endif /* CONFIG_CFS_BANDWIDTH */
320 #endif /* CONFIG_FAIR_GROUP_SCHED */
323 static inline int rt_bandwidth_enabled(void)
325 return sysctl_sched_rt_runtime >= 0;
328 /* Real-Time classes' related field in a runqueue: */
330 struct rt_prio_array active;
331 unsigned int rt_nr_running;
332 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
334 int curr; /* highest queued rt task prio */
336 int next; /* next highest */
341 unsigned long rt_nr_migratory;
342 unsigned long rt_nr_total;
344 struct plist_head pushable_tasks;
349 /* Nests inside the rq lock: */
350 raw_spinlock_t rt_runtime_lock;
352 #ifdef CONFIG_RT_GROUP_SCHED
353 unsigned long rt_nr_boosted;
356 struct list_head leaf_rt_rq_list;
357 struct task_group *tg;
364 * We add the notion of a root-domain which will be used to define per-domain
365 * variables. Each exclusive cpuset essentially defines an island domain by
366 * fully partitioning the member cpus from any other cpuset. Whenever a new
367 * exclusive cpuset is created, we also create and attach a new root-domain
376 cpumask_var_t online;
379 * The "RT overload" flag: it gets set if a CPU has more than
380 * one runnable RT task.
382 cpumask_var_t rto_mask;
383 struct cpupri cpupri;
386 extern struct root_domain def_root_domain;
388 #endif /* CONFIG_SMP */
391 * This is the main, per-CPU runqueue data structure.
393 * Locking rule: those places that want to lock multiple runqueues
394 * (such as the load balancing or the thread migration code), lock
395 * acquire operations must be ordered by ascending &runqueue.
402 * nr_running and cpu_load should be in the same cacheline because
403 * remote CPUs use both these fields when doing load calculation.
405 unsigned int nr_running;
406 #define CPU_LOAD_IDX_MAX 5
407 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
408 unsigned long last_load_update_tick;
409 #ifdef CONFIG_NO_HZ_COMMON
411 unsigned long nohz_flags;
413 #ifdef CONFIG_NO_HZ_FULL
414 unsigned long last_sched_tick;
416 int skip_clock_update;
418 /* capture load from *all* tasks on this cpu: */
419 struct load_weight load;
420 unsigned long nr_load_updates;
426 #ifdef CONFIG_FAIR_GROUP_SCHED
427 /* list of leaf cfs_rq on this cpu: */
428 struct list_head leaf_cfs_rq_list;
430 unsigned long h_load_throttle;
431 #endif /* CONFIG_SMP */
432 #endif /* CONFIG_FAIR_GROUP_SCHED */
434 #ifdef CONFIG_RT_GROUP_SCHED
435 struct list_head leaf_rt_rq_list;
439 * This is part of a global counter where only the total sum
440 * over all CPUs matters. A task can increase this counter on
441 * one CPU and if it got migrated afterwards it may decrease
442 * it on another CPU. Always updated under the runqueue lock:
444 unsigned long nr_uninterruptible;
446 struct task_struct *curr, *idle, *stop;
447 unsigned long next_balance;
448 struct mm_struct *prev_mm;
456 struct root_domain *rd;
457 struct sched_domain *sd;
459 unsigned long cpu_power;
461 unsigned char idle_balance;
462 /* For active balancing */
466 struct cpu_stop_work active_balance_work;
467 #ifdef CONFIG_SCHED_HMP
468 struct task_struct *migrate_task;
470 /* cpu of this runqueue: */
474 struct list_head cfs_tasks;
482 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
485 #ifdef CONFIG_PARAVIRT
488 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
489 u64 prev_steal_time_rq;
492 /* calc_load related fields */
493 unsigned long calc_load_update;
494 long calc_load_active;
496 #ifdef CONFIG_SCHED_HRTICK
498 int hrtick_csd_pending;
499 struct call_single_data hrtick_csd;
501 struct hrtimer hrtick_timer;
504 #ifdef CONFIG_SCHEDSTATS
506 struct sched_info rq_sched_info;
507 unsigned long long rq_cpu_time;
508 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
510 /* sys_sched_yield() stats */
511 unsigned int yld_count;
513 /* schedule() stats */
514 unsigned int sched_count;
515 unsigned int sched_goidle;
517 /* try_to_wake_up() stats */
518 unsigned int ttwu_count;
519 unsigned int ttwu_local;
523 struct llist_head wake_list;
526 struct sched_avg avg;
529 static inline int cpu_of(struct rq *rq)
538 DECLARE_PER_CPU(struct rq, runqueues);
540 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
541 #define this_rq() (&__get_cpu_var(runqueues))
542 #define task_rq(p) cpu_rq(task_cpu(p))
543 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
544 #define raw_rq() (&__raw_get_cpu_var(runqueues))
548 #define rcu_dereference_check_sched_domain(p) \
549 rcu_dereference_check((p), \
550 lockdep_is_held(&sched_domains_mutex))
553 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
554 * See detach_destroy_domains: synchronize_sched for details.
556 * The domain tree of any CPU may only be accessed from within
557 * preempt-disabled sections.
559 #define for_each_domain(cpu, __sd) \
560 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
561 __sd; __sd = __sd->parent)
563 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
566 * highest_flag_domain - Return highest sched_domain containing flag.
567 * @cpu: The cpu whose highest level of sched domain is to
569 * @flag: The flag to check for the highest sched_domain
572 * Returns the highest sched_domain of a cpu which contains the given flag.
574 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
576 struct sched_domain *sd, *hsd = NULL;
578 for_each_domain(cpu, sd) {
579 if (!(sd->flags & flag))
587 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
588 DECLARE_PER_CPU(int, sd_llc_id);
590 struct sched_group_power {
593 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
596 unsigned int power, power_orig;
597 unsigned long next_update;
599 * Number of busy cpus in this group.
601 atomic_t nr_busy_cpus;
603 unsigned long cpumask[0]; /* iteration mask */
607 struct sched_group *next; /* Must be a circular list */
610 unsigned int group_weight;
611 struct sched_group_power *sgp;
614 * The CPUs this group covers.
616 * NOTE: this field is variable length. (Allocated dynamically
617 * by attaching extra space to the end of the structure,
618 * depending on how many CPUs the kernel has booted up with)
620 unsigned long cpumask[0];
623 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
625 return to_cpumask(sg->cpumask);
629 * cpumask masking which cpus in the group are allowed to iterate up the domain
632 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
634 return to_cpumask(sg->sgp->cpumask);
638 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
639 * @group: The group whose first cpu is to be returned.
641 static inline unsigned int group_first_cpu(struct sched_group *group)
643 return cpumask_first(sched_group_cpus(group));
646 extern int group_balance_cpu(struct sched_group *sg);
648 #ifdef CONFIG_SCHED_HMP
649 static LIST_HEAD(hmp_domains);
650 DECLARE_PER_CPU(struct hmp_domain *, hmp_cpu_domain);
651 #define hmp_cpu_domain(cpu) (per_cpu(hmp_cpu_domain, (cpu)))
652 #endif /* CONFIG_SCHED_HMP */
654 #endif /* CONFIG_SMP */
657 #include "auto_group.h"
659 #ifdef CONFIG_CGROUP_SCHED
662 * Return the group to which this tasks belongs.
664 * We cannot use task_subsys_state() and friends because the cgroup
665 * subsystem changes that value before the cgroup_subsys::attach() method
666 * is called, therefore we cannot pin it and might observe the wrong value.
668 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
669 * core changes this before calling sched_move_task().
671 * Instead we use a 'copy' which is updated from sched_move_task() while
672 * holding both task_struct::pi_lock and rq::lock.
674 static inline struct task_group *task_group(struct task_struct *p)
676 return p->sched_task_group;
679 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
680 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
682 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
683 struct task_group *tg = task_group(p);
686 #ifdef CONFIG_FAIR_GROUP_SCHED
687 p->se.cfs_rq = tg->cfs_rq[cpu];
688 p->se.parent = tg->se[cpu];
691 #ifdef CONFIG_RT_GROUP_SCHED
692 p->rt.rt_rq = tg->rt_rq[cpu];
693 p->rt.parent = tg->rt_se[cpu];
697 #else /* CONFIG_CGROUP_SCHED */
699 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
700 static inline struct task_group *task_group(struct task_struct *p)
705 #endif /* CONFIG_CGROUP_SCHED */
707 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
712 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
713 * successfuly executed on another CPU. We must ensure that updates of
714 * per-task data have been completed by this moment.
717 task_thread_info(p)->cpu = cpu;
722 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
724 #ifdef CONFIG_SCHED_DEBUG
725 # include <linux/static_key.h>
726 # define const_debug __read_mostly
728 # define const_debug const
731 extern const_debug unsigned int sysctl_sched_features;
733 #define SCHED_FEAT(name, enabled) \
734 __SCHED_FEAT_##name ,
737 #include "features.h"
743 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
744 static __always_inline bool static_branch__true(struct static_key *key)
746 return static_key_true(key); /* Not out of line branch. */
749 static __always_inline bool static_branch__false(struct static_key *key)
751 return static_key_false(key); /* Out of line branch. */
754 #define SCHED_FEAT(name, enabled) \
755 static __always_inline bool static_branch_##name(struct static_key *key) \
757 return static_branch__##enabled(key); \
760 #include "features.h"
764 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
765 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
766 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
767 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
768 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
770 #ifdef CONFIG_NUMA_BALANCING
771 #define sched_feat_numa(x) sched_feat(x)
772 #ifdef CONFIG_SCHED_DEBUG
773 #define numabalancing_enabled sched_feat_numa(NUMA)
775 extern bool numabalancing_enabled;
776 #endif /* CONFIG_SCHED_DEBUG */
778 #define sched_feat_numa(x) (0)
779 #define numabalancing_enabled (0)
780 #endif /* CONFIG_NUMA_BALANCING */
782 static inline u64 global_rt_period(void)
784 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
787 static inline u64 global_rt_runtime(void)
789 if (sysctl_sched_rt_runtime < 0)
792 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
797 static inline int task_current(struct rq *rq, struct task_struct *p)
799 return rq->curr == p;
802 static inline int task_running(struct rq *rq, struct task_struct *p)
807 return task_current(rq, p);
812 #ifndef prepare_arch_switch
813 # define prepare_arch_switch(next) do { } while (0)
815 #ifndef finish_arch_switch
816 # define finish_arch_switch(prev) do { } while (0)
818 #ifndef finish_arch_post_lock_switch
819 # define finish_arch_post_lock_switch() do { } while (0)
822 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
823 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
827 * We can optimise this out completely for !SMP, because the
828 * SMP rebalancing from interrupt is the only thing that cares
835 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
839 * After ->on_cpu is cleared, the task can be moved to a different CPU.
840 * We must ensure this doesn't happen until the switch is completely
846 #ifdef CONFIG_DEBUG_SPINLOCK
847 /* this is a valid case when another task releases the spinlock */
848 rq->lock.owner = current;
851 * If we are tracking spinlock dependencies then we have to
852 * fix up the runqueue lock - which gets 'carried over' from
855 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
857 raw_spin_unlock_irq(&rq->lock);
860 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
865 * We can optimise this out completely for !SMP, because the
866 * SMP rebalancing from interrupt is the only thing that cares
871 raw_spin_unlock(&rq->lock);
874 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
878 * After ->on_cpu is cleared, the task can be moved to a different CPU.
879 * We must ensure this doesn't happen until the switch is completely
887 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
892 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
893 #define WF_FORK 0x02 /* child wakeup after fork */
894 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
896 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
902 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
908 static inline void update_load_set(struct load_weight *lw, unsigned long w)
915 * To aid in avoiding the subversion of "niceness" due to uneven distribution
916 * of tasks with abnormal "nice" values across CPUs the contribution that
917 * each task makes to its run queue's load is weighted according to its
918 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
919 * scaled version of the new time slice allocation that they receive on time
923 #define WEIGHT_IDLEPRIO 3
924 #define WMULT_IDLEPRIO 1431655765
927 * Nice levels are multiplicative, with a gentle 10% change for every
928 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
929 * nice 1, it will get ~10% less CPU time than another CPU-bound task
930 * that remained on nice 0.
932 * The "10% effect" is relative and cumulative: from _any_ nice level,
933 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
934 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
935 * If a task goes up by ~10% and another task goes down by ~10% then
936 * the relative distance between them is ~25%.)
938 static const int prio_to_weight[40] = {
939 /* -20 */ 88761, 71755, 56483, 46273, 36291,
940 /* -15 */ 29154, 23254, 18705, 14949, 11916,
941 /* -10 */ 9548, 7620, 6100, 4904, 3906,
942 /* -5 */ 3121, 2501, 1991, 1586, 1277,
943 /* 0 */ 1024, 820, 655, 526, 423,
944 /* 5 */ 335, 272, 215, 172, 137,
945 /* 10 */ 110, 87, 70, 56, 45,
946 /* 15 */ 36, 29, 23, 18, 15,
950 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
952 * In cases where the weight does not change often, we can use the
953 * precalculated inverse to speed up arithmetics by turning divisions
954 * into multiplications:
956 static const u32 prio_to_wmult[40] = {
957 /* -20 */ 48388, 59856, 76040, 92818, 118348,
958 /* -15 */ 147320, 184698, 229616, 287308, 360437,
959 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
960 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
961 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
962 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
963 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
964 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
967 #define ENQUEUE_WAKEUP 1
968 #define ENQUEUE_HEAD 2
970 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
972 #define ENQUEUE_WAKING 0
975 #define DEQUEUE_SLEEP 1
978 const struct sched_class *next;
980 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
981 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
982 void (*yield_task) (struct rq *rq);
983 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
985 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
987 struct task_struct * (*pick_next_task) (struct rq *rq);
988 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
991 int (*select_task_rq)(struct task_struct *p, int sd_flag, int flags);
992 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
994 void (*pre_schedule) (struct rq *this_rq, struct task_struct *task);
995 void (*post_schedule) (struct rq *this_rq);
996 void (*task_waking) (struct task_struct *task);
997 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
999 void (*set_cpus_allowed)(struct task_struct *p,
1000 const struct cpumask *newmask);
1002 void (*rq_online)(struct rq *rq);
1003 void (*rq_offline)(struct rq *rq);
1006 void (*set_curr_task) (struct rq *rq);
1007 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1008 void (*task_fork) (struct task_struct *p);
1010 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1011 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1012 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1015 unsigned int (*get_rr_interval) (struct rq *rq,
1016 struct task_struct *task);
1018 #ifdef CONFIG_FAIR_GROUP_SCHED
1019 void (*task_move_group) (struct task_struct *p, int on_rq);
1023 #define sched_class_highest (&stop_sched_class)
1024 #define for_each_class(class) \
1025 for (class = sched_class_highest; class; class = class->next)
1027 extern const struct sched_class stop_sched_class;
1028 extern const struct sched_class rt_sched_class;
1029 extern const struct sched_class fair_sched_class;
1030 extern const struct sched_class idle_sched_class;
1035 extern void update_group_power(struct sched_domain *sd, int cpu);
1037 extern void trigger_load_balance(struct rq *rq, int cpu);
1038 extern void idle_balance(int this_cpu, struct rq *this_rq);
1041 * Only depends on SMP, FAIR_GROUP_SCHED may be removed when runnable_avg
1042 * becomes useful in lb
1044 #if defined(CONFIG_FAIR_GROUP_SCHED)
1045 extern void idle_enter_fair(struct rq *this_rq);
1046 extern void idle_exit_fair(struct rq *this_rq);
1048 static inline void idle_enter_fair(struct rq *this_rq) {}
1049 static inline void idle_exit_fair(struct rq *this_rq) {}
1052 #else /* CONFIG_SMP */
1054 static inline void idle_balance(int cpu, struct rq *rq)
1060 extern void sysrq_sched_debug_show(void);
1061 extern void sched_init_granularity(void);
1062 extern void update_max_interval(void);
1063 extern int update_runtime(struct notifier_block *nfb, unsigned long action, void *hcpu);
1064 extern void init_sched_rt_class(void);
1065 extern void init_sched_fair_class(void);
1067 extern void resched_task(struct task_struct *p);
1068 extern void resched_cpu(int cpu);
1070 extern struct rt_bandwidth def_rt_bandwidth;
1071 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1073 extern void update_idle_cpu_load(struct rq *this_rq);
1075 #ifdef CONFIG_PARAVIRT
1076 static inline u64 steal_ticks(u64 steal)
1078 if (unlikely(steal > NSEC_PER_SEC))
1079 return div_u64(steal, TICK_NSEC);
1081 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1085 static inline void inc_nr_running(struct rq *rq)
1089 #ifdef CONFIG_NO_HZ_FULL
1090 if (rq->nr_running == 2) {
1091 if (tick_nohz_full_cpu(rq->cpu)) {
1092 /* Order rq->nr_running write against the IPI */
1094 smp_send_reschedule(rq->cpu);
1100 static inline void dec_nr_running(struct rq *rq)
1105 static inline void rq_last_tick_reset(struct rq *rq)
1107 #ifdef CONFIG_NO_HZ_FULL
1108 rq->last_sched_tick = jiffies;
1112 extern void update_rq_clock(struct rq *rq);
1114 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1115 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1117 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1119 extern const_debug unsigned int sysctl_sched_time_avg;
1120 extern const_debug unsigned int sysctl_sched_nr_migrate;
1121 extern const_debug unsigned int sysctl_sched_migration_cost;
1123 static inline u64 sched_avg_period(void)
1125 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1128 #ifdef CONFIG_SCHED_HRTICK
1132 * - enabled by features
1133 * - hrtimer is actually high res
1135 static inline int hrtick_enabled(struct rq *rq)
1137 if (!sched_feat(HRTICK))
1139 if (!cpu_active(cpu_of(rq)))
1141 return hrtimer_is_hres_active(&rq->hrtick_timer);
1144 void hrtick_start(struct rq *rq, u64 delay);
1148 static inline int hrtick_enabled(struct rq *rq)
1153 #endif /* CONFIG_SCHED_HRTICK */
1156 extern void sched_avg_update(struct rq *rq);
1157 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1159 rq->rt_avg += rt_delta;
1160 sched_avg_update(rq);
1163 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1164 static inline void sched_avg_update(struct rq *rq) { }
1167 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1170 #ifdef CONFIG_PREEMPT
1172 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1175 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1176 * way at the expense of forcing extra atomic operations in all
1177 * invocations. This assures that the double_lock is acquired using the
1178 * same underlying policy as the spinlock_t on this architecture, which
1179 * reduces latency compared to the unfair variant below. However, it
1180 * also adds more overhead and therefore may reduce throughput.
1182 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1183 __releases(this_rq->lock)
1184 __acquires(busiest->lock)
1185 __acquires(this_rq->lock)
1187 raw_spin_unlock(&this_rq->lock);
1188 double_rq_lock(this_rq, busiest);
1195 * Unfair double_lock_balance: Optimizes throughput at the expense of
1196 * latency by eliminating extra atomic operations when the locks are
1197 * already in proper order on entry. This favors lower cpu-ids and will
1198 * grant the double lock to lower cpus over higher ids under contention,
1199 * regardless of entry order into the function.
1201 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1202 __releases(this_rq->lock)
1203 __acquires(busiest->lock)
1204 __acquires(this_rq->lock)
1208 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1209 if (busiest < this_rq) {
1210 raw_spin_unlock(&this_rq->lock);
1211 raw_spin_lock(&busiest->lock);
1212 raw_spin_lock_nested(&this_rq->lock,
1213 SINGLE_DEPTH_NESTING);
1216 raw_spin_lock_nested(&busiest->lock,
1217 SINGLE_DEPTH_NESTING);
1222 #endif /* CONFIG_PREEMPT */
1225 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1227 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1229 if (unlikely(!irqs_disabled())) {
1230 /* printk() doesn't work good under rq->lock */
1231 raw_spin_unlock(&this_rq->lock);
1235 return _double_lock_balance(this_rq, busiest);
1238 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1239 __releases(busiest->lock)
1241 raw_spin_unlock(&busiest->lock);
1242 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1246 * double_rq_lock - safely lock two runqueues
1248 * Note this does not disable interrupts like task_rq_lock,
1249 * you need to do so manually before calling.
1251 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1252 __acquires(rq1->lock)
1253 __acquires(rq2->lock)
1255 BUG_ON(!irqs_disabled());
1257 raw_spin_lock(&rq1->lock);
1258 __acquire(rq2->lock); /* Fake it out ;) */
1261 raw_spin_lock(&rq1->lock);
1262 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1264 raw_spin_lock(&rq2->lock);
1265 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1271 * double_rq_unlock - safely unlock two runqueues
1273 * Note this does not restore interrupts like task_rq_unlock,
1274 * you need to do so manually after calling.
1276 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1277 __releases(rq1->lock)
1278 __releases(rq2->lock)
1280 raw_spin_unlock(&rq1->lock);
1282 raw_spin_unlock(&rq2->lock);
1284 __release(rq2->lock);
1287 #else /* CONFIG_SMP */
1290 * double_rq_lock - safely lock two runqueues
1292 * Note this does not disable interrupts like task_rq_lock,
1293 * you need to do so manually before calling.
1295 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1296 __acquires(rq1->lock)
1297 __acquires(rq2->lock)
1299 BUG_ON(!irqs_disabled());
1301 raw_spin_lock(&rq1->lock);
1302 __acquire(rq2->lock); /* Fake it out ;) */
1306 * double_rq_unlock - safely unlock two runqueues
1308 * Note this does not restore interrupts like task_rq_unlock,
1309 * you need to do so manually after calling.
1311 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1312 __releases(rq1->lock)
1313 __releases(rq2->lock)
1316 raw_spin_unlock(&rq1->lock);
1317 __release(rq2->lock);
1322 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1323 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1324 extern void print_cfs_stats(struct seq_file *m, int cpu);
1325 extern void print_rt_stats(struct seq_file *m, int cpu);
1327 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1328 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1330 extern void cfs_bandwidth_usage_inc(void);
1331 extern void cfs_bandwidth_usage_dec(void);
1333 #ifdef CONFIG_NO_HZ_COMMON
1334 enum rq_nohz_flag_bits {
1339 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1342 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1344 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1345 DECLARE_PER_CPU(u64, cpu_softirq_time);
1347 #ifndef CONFIG_64BIT
1348 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1350 static inline void irq_time_write_begin(void)
1352 __this_cpu_inc(irq_time_seq.sequence);
1356 static inline void irq_time_write_end(void)
1359 __this_cpu_inc(irq_time_seq.sequence);
1362 static inline u64 irq_time_read(int cpu)
1368 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1369 irq_time = per_cpu(cpu_softirq_time, cpu) +
1370 per_cpu(cpu_hardirq_time, cpu);
1371 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1375 #else /* CONFIG_64BIT */
1376 static inline void irq_time_write_begin(void)
1380 static inline void irq_time_write_end(void)
1384 static inline u64 irq_time_read(int cpu)
1386 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1388 #endif /* CONFIG_64BIT */
1389 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */