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>
17 extern __read_mostly int scheduler_running;
19 extern unsigned long calc_load_update;
20 extern atomic_long_t calc_load_tasks;
22 extern long calc_load_fold_active(struct rq *this_rq);
23 extern void update_cpu_load_active(struct rq *this_rq);
26 * Convert user-nice values [ -20 ... 0 ... 19 ]
27 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
30 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
31 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
32 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
35 * 'User priority' is the nice value converted to something we
36 * can work with better when scaling various scheduler parameters,
37 * it's a [ 0 ... 39 ] range.
39 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
40 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
41 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
44 * Helpers for converting nanosecond timing to jiffy resolution
46 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
49 * Increase resolution of nice-level calculations for 64-bit architectures.
50 * The extra resolution improves shares distribution and load balancing of
51 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
52 * hierarchies, especially on larger systems. This is not a user-visible change
53 * and does not change the user-interface for setting shares/weights.
55 * We increase resolution only if we have enough bits to allow this increased
56 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
57 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
60 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
61 # define SCHED_LOAD_RESOLUTION 10
62 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
63 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
65 # define SCHED_LOAD_RESOLUTION 0
66 # define scale_load(w) (w)
67 # define scale_load_down(w) (w)
70 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
71 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
73 #define NICE_0_LOAD SCHED_LOAD_SCALE
74 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
77 * Single value that decides SCHED_DEADLINE internal math precision.
78 * 10 -> just above 1us
79 * 9 -> just above 0.5us
84 * These are the 'tuning knobs' of the scheduler:
88 * single value that denotes runtime == period, ie unlimited time.
90 #define RUNTIME_INF ((u64)~0ULL)
92 static inline int fair_policy(int policy)
94 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
97 static inline int rt_policy(int policy)
99 return policy == SCHED_FIFO || policy == SCHED_RR;
102 static inline int dl_policy(int policy)
104 return policy == SCHED_DEADLINE;
107 static inline int task_has_rt_policy(struct task_struct *p)
109 return rt_policy(p->policy);
112 static inline int task_has_dl_policy(struct task_struct *p)
114 return dl_policy(p->policy);
117 static inline bool dl_time_before(u64 a, u64 b)
119 return (s64)(a - b) < 0;
123 * Tells if entity @a should preempt entity @b.
126 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
128 return dl_time_before(a->deadline, b->deadline);
132 * This is the priority-queue data structure of the RT scheduling class:
134 struct rt_prio_array {
135 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
136 struct list_head queue[MAX_RT_PRIO];
139 struct rt_bandwidth {
140 /* nests inside the rq lock: */
141 raw_spinlock_t rt_runtime_lock;
144 struct hrtimer rt_period_timer;
147 * To keep the bandwidth of -deadline tasks and groups under control
148 * we need some place where:
149 * - store the maximum -deadline bandwidth of the system (the group);
150 * - cache the fraction of that bandwidth that is currently allocated.
152 * This is all done in the data structure below. It is similar to the
153 * one used for RT-throttling (rt_bandwidth), with the main difference
154 * that, since here we are only interested in admission control, we
155 * do not decrease any runtime while the group "executes", neither we
156 * need a timer to replenish it.
158 * With respect to SMP, the bandwidth is given on a per-CPU basis,
160 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
161 * - dl_total_bw array contains, in the i-eth element, the currently
162 * allocated bandwidth on the i-eth CPU.
163 * Moreover, groups consume bandwidth on each CPU, while tasks only
164 * consume bandwidth on the CPU they're running on.
165 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
166 * that will be shown the next time the proc or cgroup controls will
167 * be red. It on its turn can be changed by writing on its own
170 struct dl_bandwidth {
171 raw_spinlock_t dl_runtime_lock;
176 static inline int dl_bandwidth_enabled(void)
178 return sysctl_sched_dl_runtime >= 0;
181 extern struct dl_bw *dl_bw_of(int i);
188 static inline u64 global_dl_period(void);
189 static inline u64 global_dl_runtime(void);
191 extern struct mutex sched_domains_mutex;
193 #ifdef CONFIG_CGROUP_SCHED
195 #include <linux/cgroup.h>
200 extern struct list_head task_groups;
202 struct cfs_bandwidth {
203 #ifdef CONFIG_CFS_BANDWIDTH
207 s64 hierarchal_quota;
210 int idle, timer_active;
211 struct hrtimer period_timer, slack_timer;
212 struct list_head throttled_cfs_rq;
215 int nr_periods, nr_throttled;
220 /* task group related information */
222 struct cgroup_subsys_state css;
224 #ifdef CONFIG_FAIR_GROUP_SCHED
225 /* schedulable entities of this group on each cpu */
226 struct sched_entity **se;
227 /* runqueue "owned" by this group on each cpu */
228 struct cfs_rq **cfs_rq;
229 unsigned long shares;
232 atomic_long_t load_avg;
233 atomic_t runnable_avg;
237 #ifdef CONFIG_RT_GROUP_SCHED
238 struct sched_rt_entity **rt_se;
239 struct rt_rq **rt_rq;
241 struct rt_bandwidth rt_bandwidth;
245 struct list_head list;
247 struct task_group *parent;
248 struct list_head siblings;
249 struct list_head children;
251 #ifdef CONFIG_SCHED_AUTOGROUP
252 struct autogroup *autogroup;
255 struct cfs_bandwidth cfs_bandwidth;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
262 * A weight of 0 or 1 can cause arithmetics problems.
263 * A weight of a cfs_rq is the sum of weights of which entities
264 * are queued on this cfs_rq, so a weight of a entity should not be
265 * too large, so as the shares value of a task group.
266 * (The default weight is 1024 - so there's no practical
267 * limitation from this.)
269 #define MIN_SHARES (1UL << 1)
270 #define MAX_SHARES (1UL << 18)
273 typedef int (*tg_visitor)(struct task_group *, void *);
275 extern int walk_tg_tree_from(struct task_group *from,
276 tg_visitor down, tg_visitor up, void *data);
279 * Iterate the full tree, calling @down when first entering a node and @up when
280 * leaving it for the final time.
282 * Caller must hold rcu_lock or sufficient equivalent.
284 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
286 return walk_tg_tree_from(&root_task_group, down, up, data);
289 extern int tg_nop(struct task_group *tg, void *data);
291 extern void free_fair_sched_group(struct task_group *tg);
292 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
293 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
294 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
295 struct sched_entity *se, int cpu,
296 struct sched_entity *parent);
297 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
298 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
300 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
301 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
302 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
304 extern void free_rt_sched_group(struct task_group *tg);
305 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
306 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
307 struct sched_rt_entity *rt_se, int cpu,
308 struct sched_rt_entity *parent);
310 extern struct task_group *sched_create_group(struct task_group *parent);
311 extern void sched_online_group(struct task_group *tg,
312 struct task_group *parent);
313 extern void sched_destroy_group(struct task_group *tg);
314 extern void sched_offline_group(struct task_group *tg);
316 extern void sched_move_task(struct task_struct *tsk);
318 #ifdef CONFIG_FAIR_GROUP_SCHED
319 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
322 #else /* CONFIG_CGROUP_SCHED */
324 struct cfs_bandwidth { };
326 #endif /* CONFIG_CGROUP_SCHED */
328 /* CFS-related fields in a runqueue */
330 struct load_weight load;
331 unsigned int nr_running, h_nr_running;
336 u64 min_vruntime_copy;
339 struct rb_root tasks_timeline;
340 struct rb_node *rb_leftmost;
343 * 'curr' points to currently running entity on this cfs_rq.
344 * It is set to NULL otherwise (i.e when none are currently running).
346 struct sched_entity *curr, *next, *last, *skip;
348 #ifdef CONFIG_SCHED_DEBUG
349 unsigned int nr_spread_over;
355 * Under CFS, load is tracked on a per-entity basis and aggregated up.
356 * This allows for the description of both thread and group usage (in
357 * the FAIR_GROUP_SCHED case).
359 unsigned long runnable_load_avg, blocked_load_avg;
360 atomic64_t decay_counter;
362 atomic_long_t removed_load;
364 #ifdef CONFIG_FAIR_GROUP_SCHED
365 /* Required to track per-cpu representation of a task_group */
366 u32 tg_runnable_contrib;
367 unsigned long tg_load_contrib;
370 * h_load = weight * f(tg)
372 * Where f(tg) is the recursive weight fraction assigned to
375 unsigned long h_load;
376 u64 last_h_load_update;
377 struct sched_entity *h_load_next;
378 #endif /* CONFIG_FAIR_GROUP_SCHED */
379 #endif /* CONFIG_SMP */
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
385 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
386 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
387 * (like users, containers etc.)
389 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
390 * list is used during load balance.
393 struct list_head leaf_cfs_rq_list;
394 struct task_group *tg; /* group that "owns" this runqueue */
396 #ifdef CONFIG_CFS_BANDWIDTH
399 s64 runtime_remaining;
401 u64 throttled_clock, throttled_clock_task;
402 u64 throttled_clock_task_time;
403 int throttled, throttle_count;
404 struct list_head throttled_list;
405 #endif /* CONFIG_CFS_BANDWIDTH */
406 #endif /* CONFIG_FAIR_GROUP_SCHED */
409 static inline int rt_bandwidth_enabled(void)
411 return sysctl_sched_rt_runtime >= 0;
414 /* Real-Time classes' related field in a runqueue: */
416 struct rt_prio_array active;
417 unsigned int rt_nr_running;
418 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
420 int curr; /* highest queued rt task prio */
422 int next; /* next highest */
427 unsigned long rt_nr_migratory;
428 unsigned long rt_nr_total;
430 struct plist_head pushable_tasks;
435 /* Nests inside the rq lock: */
436 raw_spinlock_t rt_runtime_lock;
438 #ifdef CONFIG_RT_GROUP_SCHED
439 unsigned long rt_nr_boosted;
442 struct task_group *tg;
446 /* Deadline class' related fields in a runqueue */
448 /* runqueue is an rbtree, ordered by deadline */
449 struct rb_root rb_root;
450 struct rb_node *rb_leftmost;
452 unsigned long dl_nr_running;
456 * Deadline values of the currently executing and the
457 * earliest ready task on this rq. Caching these facilitates
458 * the decision wether or not a ready but not running task
459 * should migrate somewhere else.
466 unsigned long dl_nr_migratory;
467 unsigned long dl_nr_total;
471 * Tasks on this rq that can be pushed away. They are kept in
472 * an rb-tree, ordered by tasks' deadlines, with caching
473 * of the leftmost (earliest deadline) element.
475 struct rb_root pushable_dl_tasks_root;
476 struct rb_node *pushable_dl_tasks_leftmost;
485 * We add the notion of a root-domain which will be used to define per-domain
486 * variables. Each exclusive cpuset essentially defines an island domain by
487 * fully partitioning the member cpus from any other cpuset. Whenever a new
488 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online;
500 * The bit corresponding to a CPU gets set here if such CPU has more
501 * than one runnable -deadline task (as it is below for RT tasks).
503 cpumask_var_t dlo_mask;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask;
512 struct cpupri cpupri;
515 extern struct root_domain def_root_domain;
517 #endif /* CONFIG_SMP */
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned int nr_running;
535 #ifdef CONFIG_NUMA_BALANCING
536 unsigned int nr_numa_running;
537 unsigned int nr_preferred_running;
539 #define CPU_LOAD_IDX_MAX 5
540 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
541 unsigned long last_load_update_tick;
542 #ifdef CONFIG_NO_HZ_COMMON
544 unsigned long nohz_flags;
546 #ifdef CONFIG_NO_HZ_FULL
547 unsigned long last_sched_tick;
549 int skip_clock_update;
551 /* capture load from *all* tasks on this cpu: */
552 struct load_weight load;
553 unsigned long nr_load_updates;
560 #ifdef CONFIG_FAIR_GROUP_SCHED
561 /* list of leaf cfs_rq on this cpu: */
562 struct list_head leaf_cfs_rq_list;
563 #endif /* CONFIG_FAIR_GROUP_SCHED */
565 #ifdef CONFIG_RT_GROUP_SCHED
566 struct list_head leaf_rt_rq_list;
570 * This is part of a global counter where only the total sum
571 * over all CPUs matters. A task can increase this counter on
572 * one CPU and if it got migrated afterwards it may decrease
573 * it on another CPU. Always updated under the runqueue lock:
575 unsigned long nr_uninterruptible;
577 struct task_struct *curr, *idle, *stop;
578 unsigned long next_balance;
579 struct mm_struct *prev_mm;
587 struct root_domain *rd;
588 struct sched_domain *sd;
590 unsigned long cpu_power;
592 unsigned char idle_balance;
593 /* For active balancing */
597 struct cpu_stop_work active_balance_work;
598 /* cpu of this runqueue: */
602 struct list_head cfs_tasks;
609 /* This is used to determine avg_idle's max value */
610 u64 max_idle_balance_cost;
613 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
616 #ifdef CONFIG_PARAVIRT
619 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
620 u64 prev_steal_time_rq;
623 /* calc_load related fields */
624 unsigned long calc_load_update;
625 long calc_load_active;
627 #ifdef CONFIG_SCHED_HRTICK
629 int hrtick_csd_pending;
630 struct call_single_data hrtick_csd;
632 struct hrtimer hrtick_timer;
635 #ifdef CONFIG_SCHEDSTATS
637 struct sched_info rq_sched_info;
638 unsigned long long rq_cpu_time;
639 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
641 /* sys_sched_yield() stats */
642 unsigned int yld_count;
644 /* schedule() stats */
645 unsigned int sched_count;
646 unsigned int sched_goidle;
648 /* try_to_wake_up() stats */
649 unsigned int ttwu_count;
650 unsigned int ttwu_local;
654 struct llist_head wake_list;
657 struct sched_avg avg;
660 static inline int cpu_of(struct rq *rq)
669 DECLARE_PER_CPU(struct rq, runqueues);
671 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
672 #define this_rq() (&__get_cpu_var(runqueues))
673 #define task_rq(p) cpu_rq(task_cpu(p))
674 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
675 #define raw_rq() (&__raw_get_cpu_var(runqueues))
677 static inline u64 rq_clock(struct rq *rq)
682 static inline u64 rq_clock_task(struct rq *rq)
684 return rq->clock_task;
687 #ifdef CONFIG_NUMA_BALANCING
688 extern void sched_setnuma(struct task_struct *p, int node);
689 extern int migrate_task_to(struct task_struct *p, int cpu);
690 extern int migrate_swap(struct task_struct *, struct task_struct *);
691 #endif /* CONFIG_NUMA_BALANCING */
695 #define rcu_dereference_check_sched_domain(p) \
696 rcu_dereference_check((p), \
697 lockdep_is_held(&sched_domains_mutex))
700 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
701 * See detach_destroy_domains: synchronize_sched for details.
703 * The domain tree of any CPU may only be accessed from within
704 * preempt-disabled sections.
706 #define for_each_domain(cpu, __sd) \
707 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
708 __sd; __sd = __sd->parent)
710 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
713 * highest_flag_domain - Return highest sched_domain containing flag.
714 * @cpu: The cpu whose highest level of sched domain is to
716 * @flag: The flag to check for the highest sched_domain
719 * Returns the highest sched_domain of a cpu which contains the given flag.
721 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
723 struct sched_domain *sd, *hsd = NULL;
725 for_each_domain(cpu, sd) {
726 if (!(sd->flags & flag))
734 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
736 struct sched_domain *sd;
738 for_each_domain(cpu, sd) {
739 if (sd->flags & flag)
746 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
747 DECLARE_PER_CPU(int, sd_llc_size);
748 DECLARE_PER_CPU(int, sd_llc_id);
749 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
750 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
751 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
753 struct sched_group_power {
756 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
759 unsigned int power, power_orig;
760 unsigned long next_update;
761 int imbalance; /* XXX unrelated to power but shared group state */
763 * Number of busy cpus in this group.
765 atomic_t nr_busy_cpus;
767 unsigned long cpumask[0]; /* iteration mask */
771 struct sched_group *next; /* Must be a circular list */
774 unsigned int group_weight;
775 struct sched_group_power *sgp;
778 * The CPUs this group covers.
780 * NOTE: this field is variable length. (Allocated dynamically
781 * by attaching extra space to the end of the structure,
782 * depending on how many CPUs the kernel has booted up with)
784 unsigned long cpumask[0];
787 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
789 return to_cpumask(sg->cpumask);
793 * cpumask masking which cpus in the group are allowed to iterate up the domain
796 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
798 return to_cpumask(sg->sgp->cpumask);
802 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
803 * @group: The group whose first cpu is to be returned.
805 static inline unsigned int group_first_cpu(struct sched_group *group)
807 return cpumask_first(sched_group_cpus(group));
810 extern int group_balance_cpu(struct sched_group *sg);
812 #endif /* CONFIG_SMP */
815 #include "auto_group.h"
817 #ifdef CONFIG_CGROUP_SCHED
820 * Return the group to which this tasks belongs.
822 * We cannot use task_css() and friends because the cgroup subsystem
823 * changes that value before the cgroup_subsys::attach() method is called,
824 * therefore we cannot pin it and might observe the wrong value.
826 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
827 * core changes this before calling sched_move_task().
829 * Instead we use a 'copy' which is updated from sched_move_task() while
830 * holding both task_struct::pi_lock and rq::lock.
832 static inline struct task_group *task_group(struct task_struct *p)
834 return p->sched_task_group;
837 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
838 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
840 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
841 struct task_group *tg = task_group(p);
844 #ifdef CONFIG_FAIR_GROUP_SCHED
845 p->se.cfs_rq = tg->cfs_rq[cpu];
846 p->se.parent = tg->se[cpu];
849 #ifdef CONFIG_RT_GROUP_SCHED
850 p->rt.rt_rq = tg->rt_rq[cpu];
851 p->rt.parent = tg->rt_se[cpu];
855 #else /* CONFIG_CGROUP_SCHED */
857 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
858 static inline struct task_group *task_group(struct task_struct *p)
863 #endif /* CONFIG_CGROUP_SCHED */
865 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
870 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
871 * successfuly executed on another CPU. We must ensure that updates of
872 * per-task data have been completed by this moment.
875 task_thread_info(p)->cpu = cpu;
881 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
883 #ifdef CONFIG_SCHED_DEBUG
884 # include <linux/static_key.h>
885 # define const_debug __read_mostly
887 # define const_debug const
890 extern const_debug unsigned int sysctl_sched_features;
892 #define SCHED_FEAT(name, enabled) \
893 __SCHED_FEAT_##name ,
896 #include "features.h"
902 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
903 static __always_inline bool static_branch__true(struct static_key *key)
905 return static_key_true(key); /* Not out of line branch. */
908 static __always_inline bool static_branch__false(struct static_key *key)
910 return static_key_false(key); /* Out of line branch. */
913 #define SCHED_FEAT(name, enabled) \
914 static __always_inline bool static_branch_##name(struct static_key *key) \
916 return static_branch__##enabled(key); \
919 #include "features.h"
923 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
924 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
925 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
926 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
927 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
929 #ifdef CONFIG_NUMA_BALANCING
930 #define sched_feat_numa(x) sched_feat(x)
931 #ifdef CONFIG_SCHED_DEBUG
932 #define numabalancing_enabled sched_feat_numa(NUMA)
934 extern bool numabalancing_enabled;
935 #endif /* CONFIG_SCHED_DEBUG */
937 #define sched_feat_numa(x) (0)
938 #define numabalancing_enabled (0)
939 #endif /* CONFIG_NUMA_BALANCING */
941 static inline u64 global_rt_period(void)
943 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
946 static inline u64 global_rt_runtime(void)
948 if (sysctl_sched_rt_runtime < 0)
951 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
954 static inline u64 global_dl_period(void)
956 return (u64)sysctl_sched_dl_period * NSEC_PER_USEC;
959 static inline u64 global_dl_runtime(void)
961 if (sysctl_sched_dl_runtime < 0)
964 return (u64)sysctl_sched_dl_runtime * NSEC_PER_USEC;
967 static inline int task_current(struct rq *rq, struct task_struct *p)
969 return rq->curr == p;
972 static inline int task_running(struct rq *rq, struct task_struct *p)
977 return task_current(rq, p);
982 #ifndef prepare_arch_switch
983 # define prepare_arch_switch(next) do { } while (0)
985 #ifndef finish_arch_switch
986 # define finish_arch_switch(prev) do { } while (0)
988 #ifndef finish_arch_post_lock_switch
989 # define finish_arch_post_lock_switch() do { } while (0)
992 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
993 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
997 * We can optimise this out completely for !SMP, because the
998 * SMP rebalancing from interrupt is the only thing that cares
1005 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1009 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1010 * We must ensure this doesn't happen until the switch is completely
1016 #ifdef CONFIG_DEBUG_SPINLOCK
1017 /* this is a valid case when another task releases the spinlock */
1018 rq->lock.owner = current;
1021 * If we are tracking spinlock dependencies then we have to
1022 * fix up the runqueue lock - which gets 'carried over' from
1023 * prev into current:
1025 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1027 raw_spin_unlock_irq(&rq->lock);
1030 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1031 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1035 * We can optimise this out completely for !SMP, because the
1036 * SMP rebalancing from interrupt is the only thing that cares
1041 raw_spin_unlock(&rq->lock);
1044 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1048 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1049 * We must ensure this doesn't happen until the switch is completely
1057 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1062 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1063 #define WF_FORK 0x02 /* child wakeup after fork */
1064 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1067 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1068 * of tasks with abnormal "nice" values across CPUs the contribution that
1069 * each task makes to its run queue's load is weighted according to its
1070 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1071 * scaled version of the new time slice allocation that they receive on time
1075 #define WEIGHT_IDLEPRIO 3
1076 #define WMULT_IDLEPRIO 1431655765
1079 * Nice levels are multiplicative, with a gentle 10% change for every
1080 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1081 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1082 * that remained on nice 0.
1084 * The "10% effect" is relative and cumulative: from _any_ nice level,
1085 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1086 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1087 * If a task goes up by ~10% and another task goes down by ~10% then
1088 * the relative distance between them is ~25%.)
1090 static const int prio_to_weight[40] = {
1091 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1092 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1093 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1094 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1095 /* 0 */ 1024, 820, 655, 526, 423,
1096 /* 5 */ 335, 272, 215, 172, 137,
1097 /* 10 */ 110, 87, 70, 56, 45,
1098 /* 15 */ 36, 29, 23, 18, 15,
1102 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1104 * In cases where the weight does not change often, we can use the
1105 * precalculated inverse to speed up arithmetics by turning divisions
1106 * into multiplications:
1108 static const u32 prio_to_wmult[40] = {
1109 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1110 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1111 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1112 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1113 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1114 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1115 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1116 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1119 #define ENQUEUE_WAKEUP 1
1120 #define ENQUEUE_HEAD 2
1122 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1124 #define ENQUEUE_WAKING 0
1126 #define ENQUEUE_REPLENISH 8
1128 #define DEQUEUE_SLEEP 1
1130 struct sched_class {
1131 const struct sched_class *next;
1133 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1134 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1135 void (*yield_task) (struct rq *rq);
1136 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1138 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1140 struct task_struct * (*pick_next_task) (struct rq *rq);
1141 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1144 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1145 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1147 void (*pre_schedule) (struct rq *this_rq, struct task_struct *task);
1148 void (*post_schedule) (struct rq *this_rq);
1149 void (*task_waking) (struct task_struct *task);
1150 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1152 void (*set_cpus_allowed)(struct task_struct *p,
1153 const struct cpumask *newmask);
1155 void (*rq_online)(struct rq *rq);
1156 void (*rq_offline)(struct rq *rq);
1159 void (*set_curr_task) (struct rq *rq);
1160 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1161 void (*task_fork) (struct task_struct *p);
1162 void (*task_dead) (struct task_struct *p);
1164 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1165 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1166 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1169 unsigned int (*get_rr_interval) (struct rq *rq,
1170 struct task_struct *task);
1172 #ifdef CONFIG_FAIR_GROUP_SCHED
1173 void (*task_move_group) (struct task_struct *p, int on_rq);
1177 #define sched_class_highest (&stop_sched_class)
1178 #define for_each_class(class) \
1179 for (class = sched_class_highest; class; class = class->next)
1181 extern const struct sched_class stop_sched_class;
1182 extern const struct sched_class dl_sched_class;
1183 extern const struct sched_class rt_sched_class;
1184 extern const struct sched_class fair_sched_class;
1185 extern const struct sched_class idle_sched_class;
1190 extern void update_group_power(struct sched_domain *sd, int cpu);
1192 extern void trigger_load_balance(struct rq *rq, int cpu);
1193 extern void idle_balance(int this_cpu, struct rq *this_rq);
1195 extern void idle_enter_fair(struct rq *this_rq);
1196 extern void idle_exit_fair(struct rq *this_rq);
1198 #else /* CONFIG_SMP */
1200 static inline void idle_balance(int cpu, struct rq *rq)
1206 extern void sysrq_sched_debug_show(void);
1207 extern void sched_init_granularity(void);
1208 extern void update_max_interval(void);
1210 extern void init_sched_dl_class(void);
1211 extern void init_sched_rt_class(void);
1212 extern void init_sched_fair_class(void);
1213 extern void init_sched_dl_class(void);
1215 extern void resched_task(struct task_struct *p);
1216 extern void resched_cpu(int cpu);
1218 extern struct rt_bandwidth def_rt_bandwidth;
1219 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1221 extern struct dl_bandwidth def_dl_bandwidth;
1222 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1223 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1225 unsigned long to_ratio(u64 period, u64 runtime);
1227 extern void update_idle_cpu_load(struct rq *this_rq);
1229 extern void init_task_runnable_average(struct task_struct *p);
1231 #ifdef CONFIG_PARAVIRT
1232 static inline u64 steal_ticks(u64 steal)
1234 if (unlikely(steal > NSEC_PER_SEC))
1235 return div_u64(steal, TICK_NSEC);
1237 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1241 static inline void inc_nr_running(struct rq *rq)
1245 #ifdef CONFIG_NO_HZ_FULL
1246 if (rq->nr_running == 2) {
1247 if (tick_nohz_full_cpu(rq->cpu)) {
1248 /* Order rq->nr_running write against the IPI */
1250 smp_send_reschedule(rq->cpu);
1256 static inline void dec_nr_running(struct rq *rq)
1261 static inline void rq_last_tick_reset(struct rq *rq)
1263 #ifdef CONFIG_NO_HZ_FULL
1264 rq->last_sched_tick = jiffies;
1268 extern void update_rq_clock(struct rq *rq);
1270 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1271 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1273 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1275 extern const_debug unsigned int sysctl_sched_time_avg;
1276 extern const_debug unsigned int sysctl_sched_nr_migrate;
1277 extern const_debug unsigned int sysctl_sched_migration_cost;
1279 static inline u64 sched_avg_period(void)
1281 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1284 #ifdef CONFIG_SCHED_HRTICK
1288 * - enabled by features
1289 * - hrtimer is actually high res
1291 static inline int hrtick_enabled(struct rq *rq)
1293 if (!sched_feat(HRTICK))
1295 if (!cpu_active(cpu_of(rq)))
1297 return hrtimer_is_hres_active(&rq->hrtick_timer);
1300 void hrtick_start(struct rq *rq, u64 delay);
1304 static inline int hrtick_enabled(struct rq *rq)
1309 #endif /* CONFIG_SCHED_HRTICK */
1312 extern void sched_avg_update(struct rq *rq);
1313 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1315 rq->rt_avg += rt_delta;
1316 sched_avg_update(rq);
1319 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1320 static inline void sched_avg_update(struct rq *rq) { }
1323 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1326 #ifdef CONFIG_PREEMPT
1328 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1331 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1332 * way at the expense of forcing extra atomic operations in all
1333 * invocations. This assures that the double_lock is acquired using the
1334 * same underlying policy as the spinlock_t on this architecture, which
1335 * reduces latency compared to the unfair variant below. However, it
1336 * also adds more overhead and therefore may reduce throughput.
1338 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1339 __releases(this_rq->lock)
1340 __acquires(busiest->lock)
1341 __acquires(this_rq->lock)
1343 raw_spin_unlock(&this_rq->lock);
1344 double_rq_lock(this_rq, busiest);
1351 * Unfair double_lock_balance: Optimizes throughput at the expense of
1352 * latency by eliminating extra atomic operations when the locks are
1353 * already in proper order on entry. This favors lower cpu-ids and will
1354 * grant the double lock to lower cpus over higher ids under contention,
1355 * regardless of entry order into the function.
1357 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1358 __releases(this_rq->lock)
1359 __acquires(busiest->lock)
1360 __acquires(this_rq->lock)
1364 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1365 if (busiest < this_rq) {
1366 raw_spin_unlock(&this_rq->lock);
1367 raw_spin_lock(&busiest->lock);
1368 raw_spin_lock_nested(&this_rq->lock,
1369 SINGLE_DEPTH_NESTING);
1372 raw_spin_lock_nested(&busiest->lock,
1373 SINGLE_DEPTH_NESTING);
1378 #endif /* CONFIG_PREEMPT */
1381 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1383 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1385 if (unlikely(!irqs_disabled())) {
1386 /* printk() doesn't work good under rq->lock */
1387 raw_spin_unlock(&this_rq->lock);
1391 return _double_lock_balance(this_rq, busiest);
1394 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1395 __releases(busiest->lock)
1397 raw_spin_unlock(&busiest->lock);
1398 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1401 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1407 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1410 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1416 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1420 * double_rq_lock - safely lock two runqueues
1422 * Note this does not disable interrupts like task_rq_lock,
1423 * you need to do so manually before calling.
1425 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1426 __acquires(rq1->lock)
1427 __acquires(rq2->lock)
1429 BUG_ON(!irqs_disabled());
1431 raw_spin_lock(&rq1->lock);
1432 __acquire(rq2->lock); /* Fake it out ;) */
1435 raw_spin_lock(&rq1->lock);
1436 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1438 raw_spin_lock(&rq2->lock);
1439 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1445 * double_rq_unlock - safely unlock two runqueues
1447 * Note this does not restore interrupts like task_rq_unlock,
1448 * you need to do so manually after calling.
1450 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1451 __releases(rq1->lock)
1452 __releases(rq2->lock)
1454 raw_spin_unlock(&rq1->lock);
1456 raw_spin_unlock(&rq2->lock);
1458 __release(rq2->lock);
1461 #else /* CONFIG_SMP */
1464 * double_rq_lock - safely lock two runqueues
1466 * Note this does not disable interrupts like task_rq_lock,
1467 * you need to do so manually before calling.
1469 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1470 __acquires(rq1->lock)
1471 __acquires(rq2->lock)
1473 BUG_ON(!irqs_disabled());
1475 raw_spin_lock(&rq1->lock);
1476 __acquire(rq2->lock); /* Fake it out ;) */
1480 * double_rq_unlock - safely unlock two runqueues
1482 * Note this does not restore interrupts like task_rq_unlock,
1483 * you need to do so manually after calling.
1485 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1486 __releases(rq1->lock)
1487 __releases(rq2->lock)
1490 raw_spin_unlock(&rq1->lock);
1491 __release(rq2->lock);
1496 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1497 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1498 extern void print_cfs_stats(struct seq_file *m, int cpu);
1499 extern void print_rt_stats(struct seq_file *m, int cpu);
1501 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1502 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1503 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1505 extern void cfs_bandwidth_usage_inc(void);
1506 extern void cfs_bandwidth_usage_dec(void);
1508 #ifdef CONFIG_NO_HZ_COMMON
1509 enum rq_nohz_flag_bits {
1514 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1517 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1519 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1520 DECLARE_PER_CPU(u64, cpu_softirq_time);
1522 #ifndef CONFIG_64BIT
1523 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1525 static inline void irq_time_write_begin(void)
1527 __this_cpu_inc(irq_time_seq.sequence);
1531 static inline void irq_time_write_end(void)
1534 __this_cpu_inc(irq_time_seq.sequence);
1537 static inline u64 irq_time_read(int cpu)
1543 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1544 irq_time = per_cpu(cpu_softirq_time, cpu) +
1545 per_cpu(cpu_hardirq_time, cpu);
1546 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1550 #else /* CONFIG_64BIT */
1551 static inline void irq_time_write_begin(void)
1555 static inline void irq_time_write_end(void)
1559 static inline u64 irq_time_read(int cpu)
1561 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1563 #endif /* CONFIG_64BIT */
1564 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */