2 #include <linux/sched.h>
3 #include <linux/sched/sysctl.h>
4 #include <linux/sched/rt.h>
5 #include <linux/sched/deadline.h>
6 #include <linux/mutex.h>
7 #include <linux/spinlock.h>
8 #include <linux/stop_machine.h>
9 #include <linux/tick.h>
10 #include <linux/slab.h>
13 #include "cpudeadline.h"
19 /* task_struct::on_rq states: */
20 #define TASK_ON_RQ_QUEUED 1
21 #define TASK_ON_RQ_MIGRATING 2
23 extern __read_mostly int scheduler_running;
25 extern unsigned long calc_load_update;
26 extern atomic_long_t calc_load_tasks;
28 extern long calc_load_fold_active(struct rq *this_rq);
29 extern void update_cpu_load_active(struct rq *this_rq);
32 * Helpers for converting nanosecond timing to jiffy resolution
34 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
37 * Increase resolution of nice-level calculations for 64-bit architectures.
38 * The extra resolution improves shares distribution and load balancing of
39 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
40 * hierarchies, especially on larger systems. This is not a user-visible change
41 * and does not change the user-interface for setting shares/weights.
43 * We increase resolution only if we have enough bits to allow this increased
44 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
45 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
48 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
49 # define SCHED_LOAD_RESOLUTION 10
50 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
51 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
53 # define SCHED_LOAD_RESOLUTION 0
54 # define scale_load(w) (w)
55 # define scale_load_down(w) (w)
58 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
59 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
61 #define NICE_0_LOAD SCHED_LOAD_SCALE
62 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
65 * Single value that decides SCHED_DEADLINE internal math precision.
66 * 10 -> just above 1us
67 * 9 -> just above 0.5us
72 * These are the 'tuning knobs' of the scheduler:
76 * single value that denotes runtime == period, ie unlimited time.
78 #define RUNTIME_INF ((u64)~0ULL)
80 static inline int fair_policy(int policy)
82 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
85 static inline int rt_policy(int policy)
87 return policy == SCHED_FIFO || policy == SCHED_RR;
90 static inline int dl_policy(int policy)
92 return policy == SCHED_DEADLINE;
95 static inline int task_has_rt_policy(struct task_struct *p)
97 return rt_policy(p->policy);
100 static inline int task_has_dl_policy(struct task_struct *p)
102 return dl_policy(p->policy);
105 static inline bool dl_time_before(u64 a, u64 b)
107 return (s64)(a - b) < 0;
111 * Tells if entity @a should preempt entity @b.
114 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
116 return dl_time_before(a->deadline, b->deadline);
120 * This is the priority-queue data structure of the RT scheduling class:
122 struct rt_prio_array {
123 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
124 struct list_head queue[MAX_RT_PRIO];
127 struct rt_bandwidth {
128 /* nests inside the rq lock: */
129 raw_spinlock_t rt_runtime_lock;
132 struct hrtimer rt_period_timer;
135 void __dl_clear_params(struct task_struct *p);
138 * To keep the bandwidth of -deadline tasks and groups under control
139 * we need some place where:
140 * - store the maximum -deadline bandwidth of the system (the group);
141 * - cache the fraction of that bandwidth that is currently allocated.
143 * This is all done in the data structure below. It is similar to the
144 * one used for RT-throttling (rt_bandwidth), with the main difference
145 * that, since here we are only interested in admission control, we
146 * do not decrease any runtime while the group "executes", neither we
147 * need a timer to replenish it.
149 * With respect to SMP, the bandwidth is given on a per-CPU basis,
151 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
152 * - dl_total_bw array contains, in the i-eth element, the currently
153 * allocated bandwidth on the i-eth CPU.
154 * Moreover, groups consume bandwidth on each CPU, while tasks only
155 * consume bandwidth on the CPU they're running on.
156 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
157 * that will be shown the next time the proc or cgroup controls will
158 * be red. It on its turn can be changed by writing on its own
161 struct dl_bandwidth {
162 raw_spinlock_t dl_runtime_lock;
167 static inline int dl_bandwidth_enabled(void)
169 return sysctl_sched_rt_runtime >= 0;
172 extern struct dl_bw *dl_bw_of(int i);
180 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
182 dl_b->total_bw -= tsk_bw;
186 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
188 dl_b->total_bw += tsk_bw;
192 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
194 return dl_b->bw != -1 &&
195 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
198 extern struct mutex sched_domains_mutex;
200 #ifdef CONFIG_CGROUP_SCHED
202 #include <linux/cgroup.h>
207 extern struct list_head task_groups;
209 struct cfs_bandwidth {
210 #ifdef CONFIG_CFS_BANDWIDTH
214 s64 hierarchical_quota;
217 int idle, timer_active;
218 struct hrtimer period_timer, slack_timer;
219 struct list_head throttled_cfs_rq;
222 int nr_periods, nr_throttled;
227 /* task group related information */
229 struct cgroup_subsys_state css;
231 #ifdef CONFIG_FAIR_GROUP_SCHED
232 /* schedulable entities of this group on each cpu */
233 struct sched_entity **se;
234 /* runqueue "owned" by this group on each cpu */
235 struct cfs_rq **cfs_rq;
236 unsigned long shares;
239 atomic_long_t load_avg;
240 atomic_t runnable_avg;
244 #ifdef CONFIG_RT_GROUP_SCHED
245 struct sched_rt_entity **rt_se;
246 struct rt_rq **rt_rq;
248 struct rt_bandwidth rt_bandwidth;
252 struct list_head list;
254 struct task_group *parent;
255 struct list_head siblings;
256 struct list_head children;
258 #ifdef CONFIG_SCHED_AUTOGROUP
259 struct autogroup *autogroup;
262 struct cfs_bandwidth cfs_bandwidth;
265 #ifdef CONFIG_FAIR_GROUP_SCHED
266 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
269 * A weight of 0 or 1 can cause arithmetics problems.
270 * A weight of a cfs_rq is the sum of weights of which entities
271 * are queued on this cfs_rq, so a weight of a entity should not be
272 * too large, so as the shares value of a task group.
273 * (The default weight is 1024 - so there's no practical
274 * limitation from this.)
276 #define MIN_SHARES (1UL << 1)
277 #define MAX_SHARES (1UL << 18)
280 typedef int (*tg_visitor)(struct task_group *, void *);
282 extern int walk_tg_tree_from(struct task_group *from,
283 tg_visitor down, tg_visitor up, void *data);
286 * Iterate the full tree, calling @down when first entering a node and @up when
287 * leaving it for the final time.
289 * Caller must hold rcu_lock or sufficient equivalent.
291 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
293 return walk_tg_tree_from(&root_task_group, down, up, data);
296 extern int tg_nop(struct task_group *tg, void *data);
298 extern void free_fair_sched_group(struct task_group *tg);
299 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
300 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
301 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
302 struct sched_entity *se, int cpu,
303 struct sched_entity *parent);
304 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
305 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
307 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
308 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
309 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
311 extern void free_rt_sched_group(struct task_group *tg);
312 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
313 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
314 struct sched_rt_entity *rt_se, int cpu,
315 struct sched_rt_entity *parent);
317 extern struct task_group *sched_create_group(struct task_group *parent);
318 extern void sched_online_group(struct task_group *tg,
319 struct task_group *parent);
320 extern void sched_destroy_group(struct task_group *tg);
321 extern void sched_offline_group(struct task_group *tg);
323 extern void sched_move_task(struct task_struct *tsk);
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
329 #else /* CONFIG_CGROUP_SCHED */
331 struct cfs_bandwidth { };
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load;
338 unsigned int nr_running, h_nr_running;
343 u64 min_vruntime_copy;
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
350 * 'curr' points to currently running entity on this cfs_rq.
351 * It is set to NULL otherwise (i.e when none are currently running).
353 struct sched_entity *curr, *next, *last, *skip;
355 #ifdef CONFIG_SCHED_DEBUG
356 unsigned int nr_spread_over;
362 * Under CFS, load is tracked on a per-entity basis and aggregated up.
363 * This allows for the description of both thread and group usage (in
364 * the FAIR_GROUP_SCHED case).
366 unsigned long runnable_load_avg, blocked_load_avg;
367 atomic64_t decay_counter;
369 atomic_long_t removed_load;
371 #ifdef CONFIG_FAIR_GROUP_SCHED
372 /* Required to track per-cpu representation of a task_group */
373 u32 tg_runnable_contrib;
374 unsigned long tg_load_contrib;
377 * h_load = weight * f(tg)
379 * Where f(tg) is the recursive weight fraction assigned to
382 unsigned long h_load;
383 u64 last_h_load_update;
384 struct sched_entity *h_load_next;
385 #endif /* CONFIG_FAIR_GROUP_SCHED */
386 #endif /* CONFIG_SMP */
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
392 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
393 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
394 * (like users, containers etc.)
396 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
397 * list is used during load balance.
400 struct list_head leaf_cfs_rq_list;
401 struct task_group *tg; /* group that "owns" this runqueue */
403 #ifdef CONFIG_CFS_BANDWIDTH
406 s64 runtime_remaining;
408 u64 throttled_clock, throttled_clock_task;
409 u64 throttled_clock_task_time;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
412 #endif /* CONFIG_CFS_BANDWIDTH */
413 #endif /* CONFIG_FAIR_GROUP_SCHED */
416 static inline int rt_bandwidth_enabled(void)
418 return sysctl_sched_rt_runtime >= 0;
421 /* Real-Time classes' related field in a runqueue: */
423 struct rt_prio_array active;
424 unsigned int rt_nr_running;
425 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
427 int curr; /* highest queued rt task prio */
429 int next; /* next highest */
434 unsigned long rt_nr_migratory;
435 unsigned long rt_nr_total;
437 struct plist_head pushable_tasks;
444 /* Nests inside the rq lock: */
445 raw_spinlock_t rt_runtime_lock;
447 #ifdef CONFIG_RT_GROUP_SCHED
448 unsigned long rt_nr_boosted;
451 struct task_group *tg;
455 /* Deadline class' related fields in a runqueue */
457 /* runqueue is an rbtree, ordered by deadline */
458 struct rb_root rb_root;
459 struct rb_node *rb_leftmost;
461 unsigned long dl_nr_running;
465 * Deadline values of the currently executing and the
466 * earliest ready task on this rq. Caching these facilitates
467 * the decision wether or not a ready but not running task
468 * should migrate somewhere else.
475 unsigned long dl_nr_migratory;
479 * Tasks on this rq that can be pushed away. They are kept in
480 * an rb-tree, ordered by tasks' deadlines, with caching
481 * of the leftmost (earliest deadline) element.
483 struct rb_root pushable_dl_tasks_root;
484 struct rb_node *pushable_dl_tasks_leftmost;
493 * We add the notion of a root-domain which will be used to define per-domain
494 * variables. Each exclusive cpuset essentially defines an island domain by
495 * fully partitioning the member cpus from any other cpuset. Whenever a new
496 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online;
507 /* Indicate more than one runnable task for any CPU */
511 * The bit corresponding to a CPU gets set here if such CPU has more
512 * than one runnable -deadline task (as it is below for RT tasks).
514 cpumask_var_t dlo_mask;
520 * The "RT overload" flag: it gets set if a CPU has more than
521 * one runnable RT task.
523 cpumask_var_t rto_mask;
524 struct cpupri cpupri;
527 extern struct root_domain def_root_domain;
529 #endif /* CONFIG_SMP */
532 * This is the main, per-CPU runqueue data structure.
534 * Locking rule: those places that want to lock multiple runqueues
535 * (such as the load balancing or the thread migration code), lock
536 * acquire operations must be ordered by ascending &runqueue.
543 * nr_running and cpu_load should be in the same cacheline because
544 * remote CPUs use both these fields when doing load calculation.
546 unsigned int nr_running;
547 #ifdef CONFIG_NUMA_BALANCING
548 unsigned int nr_numa_running;
549 unsigned int nr_preferred_running;
551 #define CPU_LOAD_IDX_MAX 5
552 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
553 unsigned long last_load_update_tick;
554 #ifdef CONFIG_NO_HZ_COMMON
556 unsigned long nohz_flags;
558 #ifdef CONFIG_NO_HZ_FULL
559 unsigned long last_sched_tick;
561 /* capture load from *all* tasks on this cpu: */
562 struct load_weight load;
563 unsigned long nr_load_updates;
570 #ifdef CONFIG_FAIR_GROUP_SCHED
571 /* list of leaf cfs_rq on this cpu: */
572 struct list_head leaf_cfs_rq_list;
574 struct sched_avg avg;
575 #endif /* CONFIG_FAIR_GROUP_SCHED */
578 * This is part of a global counter where only the total sum
579 * over all CPUs matters. A task can increase this counter on
580 * one CPU and if it got migrated afterwards it may decrease
581 * it on another CPU. Always updated under the runqueue lock:
583 unsigned long nr_uninterruptible;
585 struct task_struct *curr, *idle, *stop;
586 unsigned long next_balance;
587 struct mm_struct *prev_mm;
589 unsigned int clock_skip_update;
596 struct root_domain *rd;
597 struct sched_domain *sd;
599 unsigned long cpu_capacity;
601 unsigned char idle_balance;
602 /* For active balancing */
606 struct cpu_stop_work active_balance_work;
607 /* cpu of this runqueue: */
611 struct list_head cfs_tasks;
618 /* This is used to determine avg_idle's max value */
619 u64 max_idle_balance_cost;
622 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
625 #ifdef CONFIG_PARAVIRT
628 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
629 u64 prev_steal_time_rq;
632 /* calc_load related fields */
633 unsigned long calc_load_update;
634 long calc_load_active;
636 #ifdef CONFIG_SCHED_HRTICK
638 int hrtick_csd_pending;
639 struct call_single_data hrtick_csd;
641 struct hrtimer hrtick_timer;
644 #ifdef CONFIG_SCHEDSTATS
646 struct sched_info rq_sched_info;
647 unsigned long long rq_cpu_time;
648 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
650 /* sys_sched_yield() stats */
651 unsigned int yld_count;
653 /* schedule() stats */
654 unsigned int sched_count;
655 unsigned int sched_goidle;
657 /* try_to_wake_up() stats */
658 unsigned int ttwu_count;
659 unsigned int ttwu_local;
663 struct llist_head wake_list;
666 #ifdef CONFIG_CPU_IDLE
667 /* Must be inspected within a rcu lock section */
668 struct cpuidle_state *idle_state;
672 static inline int cpu_of(struct rq *rq)
681 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
683 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
684 #define this_rq() this_cpu_ptr(&runqueues)
685 #define task_rq(p) cpu_rq(task_cpu(p))
686 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
687 #define raw_rq() raw_cpu_ptr(&runqueues)
689 static inline u64 __rq_clock_broken(struct rq *rq)
691 return ACCESS_ONCE(rq->clock);
694 static inline u64 rq_clock(struct rq *rq)
696 lockdep_assert_held(&rq->lock);
700 static inline u64 rq_clock_task(struct rq *rq)
702 lockdep_assert_held(&rq->lock);
703 return rq->clock_task;
706 #define RQCF_REQ_SKIP 0x01
707 #define RQCF_ACT_SKIP 0x02
709 static inline void rq_clock_skip_update(struct rq *rq, bool skip)
711 lockdep_assert_held(&rq->lock);
713 rq->clock_skip_update |= RQCF_REQ_SKIP;
715 rq->clock_skip_update &= ~RQCF_REQ_SKIP;
719 enum numa_topology_type {
724 extern enum numa_topology_type sched_numa_topology_type;
725 extern int sched_max_numa_distance;
726 extern bool find_numa_distance(int distance);
729 #ifdef CONFIG_NUMA_BALANCING
730 /* The regions in numa_faults array from task_struct */
731 enum numa_faults_stats {
737 extern void sched_setnuma(struct task_struct *p, int node);
738 extern int migrate_task_to(struct task_struct *p, int cpu);
739 extern int migrate_swap(struct task_struct *, struct task_struct *);
740 #endif /* CONFIG_NUMA_BALANCING */
744 extern void sched_ttwu_pending(void);
746 #define rcu_dereference_check_sched_domain(p) \
747 rcu_dereference_check((p), \
748 lockdep_is_held(&sched_domains_mutex))
751 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
752 * See detach_destroy_domains: synchronize_sched for details.
754 * The domain tree of any CPU may only be accessed from within
755 * preempt-disabled sections.
757 #define for_each_domain(cpu, __sd) \
758 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
759 __sd; __sd = __sd->parent)
761 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
764 * highest_flag_domain - Return highest sched_domain containing flag.
765 * @cpu: The cpu whose highest level of sched domain is to
767 * @flag: The flag to check for the highest sched_domain
770 * Returns the highest sched_domain of a cpu which contains the given flag.
772 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
774 struct sched_domain *sd, *hsd = NULL;
776 for_each_domain(cpu, sd) {
777 if (!(sd->flags & flag))
785 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
787 struct sched_domain *sd;
789 for_each_domain(cpu, sd) {
790 if (sd->flags & flag)
797 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
798 DECLARE_PER_CPU(int, sd_llc_size);
799 DECLARE_PER_CPU(int, sd_llc_id);
800 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
801 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
802 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
804 struct sched_group_capacity {
807 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
810 unsigned int capacity, capacity_orig;
811 unsigned long next_update;
812 int imbalance; /* XXX unrelated to capacity but shared group state */
814 * Number of busy cpus in this group.
816 atomic_t nr_busy_cpus;
818 unsigned long cpumask[0]; /* iteration mask */
822 struct sched_group *next; /* Must be a circular list */
825 unsigned int group_weight;
826 struct sched_group_capacity *sgc;
829 * The CPUs this group covers.
831 * NOTE: this field is variable length. (Allocated dynamically
832 * by attaching extra space to the end of the structure,
833 * depending on how many CPUs the kernel has booted up with)
835 unsigned long cpumask[0];
838 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
840 return to_cpumask(sg->cpumask);
844 * cpumask masking which cpus in the group are allowed to iterate up the domain
847 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
849 return to_cpumask(sg->sgc->cpumask);
853 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
854 * @group: The group whose first cpu is to be returned.
856 static inline unsigned int group_first_cpu(struct sched_group *group)
858 return cpumask_first(sched_group_cpus(group));
861 extern int group_balance_cpu(struct sched_group *sg);
865 static inline void sched_ttwu_pending(void) { }
867 #endif /* CONFIG_SMP */
870 #include "auto_group.h"
872 #ifdef CONFIG_CGROUP_SCHED
875 * Return the group to which this tasks belongs.
877 * We cannot use task_css() and friends because the cgroup subsystem
878 * changes that value before the cgroup_subsys::attach() method is called,
879 * therefore we cannot pin it and might observe the wrong value.
881 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
882 * core changes this before calling sched_move_task().
884 * Instead we use a 'copy' which is updated from sched_move_task() while
885 * holding both task_struct::pi_lock and rq::lock.
887 static inline struct task_group *task_group(struct task_struct *p)
889 return p->sched_task_group;
892 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
893 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
895 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
896 struct task_group *tg = task_group(p);
899 #ifdef CONFIG_FAIR_GROUP_SCHED
900 p->se.cfs_rq = tg->cfs_rq[cpu];
901 p->se.parent = tg->se[cpu];
904 #ifdef CONFIG_RT_GROUP_SCHED
905 p->rt.rt_rq = tg->rt_rq[cpu];
906 p->rt.parent = tg->rt_se[cpu];
910 #else /* CONFIG_CGROUP_SCHED */
912 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
913 static inline struct task_group *task_group(struct task_struct *p)
918 #endif /* CONFIG_CGROUP_SCHED */
920 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
925 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
926 * successfuly executed on another CPU. We must ensure that updates of
927 * per-task data have been completed by this moment.
930 task_thread_info(p)->cpu = cpu;
936 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
938 #ifdef CONFIG_SCHED_DEBUG
939 # include <linux/static_key.h>
940 # define const_debug __read_mostly
942 # define const_debug const
945 extern const_debug unsigned int sysctl_sched_features;
947 #define SCHED_FEAT(name, enabled) \
948 __SCHED_FEAT_##name ,
951 #include "features.h"
957 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
958 #define SCHED_FEAT(name, enabled) \
959 static __always_inline bool static_branch_##name(struct static_key *key) \
961 return static_key_##enabled(key); \
964 #include "features.h"
968 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
969 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
970 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
971 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
972 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
974 #ifdef CONFIG_NUMA_BALANCING
975 #define sched_feat_numa(x) sched_feat(x)
976 #ifdef CONFIG_SCHED_DEBUG
977 #define numabalancing_enabled sched_feat_numa(NUMA)
979 extern bool numabalancing_enabled;
980 #endif /* CONFIG_SCHED_DEBUG */
982 #define sched_feat_numa(x) (0)
983 #define numabalancing_enabled (0)
984 #endif /* CONFIG_NUMA_BALANCING */
986 static inline u64 global_rt_period(void)
988 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
991 static inline u64 global_rt_runtime(void)
993 if (sysctl_sched_rt_runtime < 0)
996 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
999 static inline int task_current(struct rq *rq, struct task_struct *p)
1001 return rq->curr == p;
1004 static inline int task_running(struct rq *rq, struct task_struct *p)
1009 return task_current(rq, p);
1013 static inline int task_on_rq_queued(struct task_struct *p)
1015 return p->on_rq == TASK_ON_RQ_QUEUED;
1018 static inline int task_on_rq_migrating(struct task_struct *p)
1020 return p->on_rq == TASK_ON_RQ_MIGRATING;
1023 #ifndef prepare_arch_switch
1024 # define prepare_arch_switch(next) do { } while (0)
1026 #ifndef finish_arch_switch
1027 # define finish_arch_switch(prev) do { } while (0)
1029 #ifndef finish_arch_post_lock_switch
1030 # define finish_arch_post_lock_switch() do { } while (0)
1033 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1037 * We can optimise this out completely for !SMP, because the
1038 * SMP rebalancing from interrupt is the only thing that cares
1045 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1049 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1050 * We must ensure this doesn't happen until the switch is completely
1056 #ifdef CONFIG_DEBUG_SPINLOCK
1057 /* this is a valid case when another task releases the spinlock */
1058 rq->lock.owner = current;
1061 * If we are tracking spinlock dependencies then we have to
1062 * fix up the runqueue lock - which gets 'carried over' from
1063 * prev into current:
1065 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1067 raw_spin_unlock_irq(&rq->lock);
1073 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1074 #define WF_FORK 0x02 /* child wakeup after fork */
1075 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1078 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1079 * of tasks with abnormal "nice" values across CPUs the contribution that
1080 * each task makes to its run queue's load is weighted according to its
1081 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1082 * scaled version of the new time slice allocation that they receive on time
1086 #define WEIGHT_IDLEPRIO 3
1087 #define WMULT_IDLEPRIO 1431655765
1090 * Nice levels are multiplicative, with a gentle 10% change for every
1091 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1092 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1093 * that remained on nice 0.
1095 * The "10% effect" is relative and cumulative: from _any_ nice level,
1096 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1097 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1098 * If a task goes up by ~10% and another task goes down by ~10% then
1099 * the relative distance between them is ~25%.)
1101 static const int prio_to_weight[40] = {
1102 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1103 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1104 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1105 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1106 /* 0 */ 1024, 820, 655, 526, 423,
1107 /* 5 */ 335, 272, 215, 172, 137,
1108 /* 10 */ 110, 87, 70, 56, 45,
1109 /* 15 */ 36, 29, 23, 18, 15,
1113 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1115 * In cases where the weight does not change often, we can use the
1116 * precalculated inverse to speed up arithmetics by turning divisions
1117 * into multiplications:
1119 static const u32 prio_to_wmult[40] = {
1120 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1121 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1122 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1123 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1124 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1125 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1126 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1127 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1130 #define ENQUEUE_WAKEUP 1
1131 #define ENQUEUE_HEAD 2
1133 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1135 #define ENQUEUE_WAKING 0
1137 #define ENQUEUE_REPLENISH 8
1139 #define DEQUEUE_SLEEP 1
1141 #define RETRY_TASK ((void *)-1UL)
1143 struct sched_class {
1144 const struct sched_class *next;
1146 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1147 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1148 void (*yield_task) (struct rq *rq);
1149 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1151 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1154 * It is the responsibility of the pick_next_task() method that will
1155 * return the next task to call put_prev_task() on the @prev task or
1156 * something equivalent.
1158 * May return RETRY_TASK when it finds a higher prio class has runnable
1161 struct task_struct * (*pick_next_task) (struct rq *rq,
1162 struct task_struct *prev);
1163 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1166 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1167 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1169 void (*post_schedule) (struct rq *this_rq);
1170 void (*task_waking) (struct task_struct *task);
1171 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1173 void (*set_cpus_allowed)(struct task_struct *p,
1174 const struct cpumask *newmask);
1176 void (*rq_online)(struct rq *rq);
1177 void (*rq_offline)(struct rq *rq);
1180 void (*set_curr_task) (struct rq *rq);
1181 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1182 void (*task_fork) (struct task_struct *p);
1183 void (*task_dead) (struct task_struct *p);
1186 * The switched_from() call is allowed to drop rq->lock, therefore we
1187 * cannot assume the switched_from/switched_to pair is serliazed by
1188 * rq->lock. They are however serialized by p->pi_lock.
1190 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1191 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1192 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1195 unsigned int (*get_rr_interval) (struct rq *rq,
1196 struct task_struct *task);
1198 void (*update_curr) (struct rq *rq);
1200 #ifdef CONFIG_FAIR_GROUP_SCHED
1201 void (*task_move_group) (struct task_struct *p, int on_rq);
1205 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1207 prev->sched_class->put_prev_task(rq, prev);
1210 #define sched_class_highest (&stop_sched_class)
1211 #define for_each_class(class) \
1212 for (class = sched_class_highest; class; class = class->next)
1214 extern const struct sched_class stop_sched_class;
1215 extern const struct sched_class dl_sched_class;
1216 extern const struct sched_class rt_sched_class;
1217 extern const struct sched_class fair_sched_class;
1218 extern const struct sched_class idle_sched_class;
1223 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1225 extern void trigger_load_balance(struct rq *rq);
1227 extern void idle_enter_fair(struct rq *this_rq);
1228 extern void idle_exit_fair(struct rq *this_rq);
1232 static inline void idle_enter_fair(struct rq *rq) { }
1233 static inline void idle_exit_fair(struct rq *rq) { }
1237 #ifdef CONFIG_CPU_IDLE
1238 static inline void idle_set_state(struct rq *rq,
1239 struct cpuidle_state *idle_state)
1241 rq->idle_state = idle_state;
1244 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1246 WARN_ON(!rcu_read_lock_held());
1247 return rq->idle_state;
1250 static inline void idle_set_state(struct rq *rq,
1251 struct cpuidle_state *idle_state)
1255 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1261 extern void sysrq_sched_debug_show(void);
1262 extern void sched_init_granularity(void);
1263 extern void update_max_interval(void);
1265 extern void init_sched_dl_class(void);
1266 extern void init_sched_rt_class(void);
1267 extern void init_sched_fair_class(void);
1268 extern void init_sched_dl_class(void);
1270 extern void resched_curr(struct rq *rq);
1271 extern void resched_cpu(int cpu);
1273 extern struct rt_bandwidth def_rt_bandwidth;
1274 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1276 extern struct dl_bandwidth def_dl_bandwidth;
1277 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1278 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1280 unsigned long to_ratio(u64 period, u64 runtime);
1282 extern void update_idle_cpu_load(struct rq *this_rq);
1284 extern void init_task_runnable_average(struct task_struct *p);
1286 static inline void add_nr_running(struct rq *rq, unsigned count)
1288 unsigned prev_nr = rq->nr_running;
1290 rq->nr_running = prev_nr + count;
1292 if (prev_nr < 2 && rq->nr_running >= 2) {
1294 if (!rq->rd->overload)
1295 rq->rd->overload = true;
1298 #ifdef CONFIG_NO_HZ_FULL
1299 if (tick_nohz_full_cpu(rq->cpu)) {
1301 * Tick is needed if more than one task runs on a CPU.
1302 * Send the target an IPI to kick it out of nohz mode.
1304 * We assume that IPI implies full memory barrier and the
1305 * new value of rq->nr_running is visible on reception
1308 tick_nohz_full_kick_cpu(rq->cpu);
1314 static inline void sub_nr_running(struct rq *rq, unsigned count)
1316 rq->nr_running -= count;
1319 static inline void rq_last_tick_reset(struct rq *rq)
1321 #ifdef CONFIG_NO_HZ_FULL
1322 rq->last_sched_tick = jiffies;
1326 extern void update_rq_clock(struct rq *rq);
1328 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1329 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1331 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1333 extern const_debug unsigned int sysctl_sched_time_avg;
1334 extern const_debug unsigned int sysctl_sched_nr_migrate;
1335 extern const_debug unsigned int sysctl_sched_migration_cost;
1337 static inline u64 sched_avg_period(void)
1339 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1342 #ifdef CONFIG_SCHED_HRTICK
1346 * - enabled by features
1347 * - hrtimer is actually high res
1349 static inline int hrtick_enabled(struct rq *rq)
1351 if (!sched_feat(HRTICK))
1353 if (!cpu_active(cpu_of(rq)))
1355 return hrtimer_is_hres_active(&rq->hrtick_timer);
1358 void hrtick_start(struct rq *rq, u64 delay);
1362 static inline int hrtick_enabled(struct rq *rq)
1367 #endif /* CONFIG_SCHED_HRTICK */
1370 extern void sched_avg_update(struct rq *rq);
1371 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1373 rq->rt_avg += rt_delta;
1374 sched_avg_update(rq);
1377 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1378 static inline void sched_avg_update(struct rq *rq) { }
1381 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1384 #ifdef CONFIG_PREEMPT
1386 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1389 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1390 * way at the expense of forcing extra atomic operations in all
1391 * invocations. This assures that the double_lock is acquired using the
1392 * same underlying policy as the spinlock_t on this architecture, which
1393 * reduces latency compared to the unfair variant below. However, it
1394 * also adds more overhead and therefore may reduce throughput.
1396 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1397 __releases(this_rq->lock)
1398 __acquires(busiest->lock)
1399 __acquires(this_rq->lock)
1401 raw_spin_unlock(&this_rq->lock);
1402 double_rq_lock(this_rq, busiest);
1409 * Unfair double_lock_balance: Optimizes throughput at the expense of
1410 * latency by eliminating extra atomic operations when the locks are
1411 * already in proper order on entry. This favors lower cpu-ids and will
1412 * grant the double lock to lower cpus over higher ids under contention,
1413 * regardless of entry order into the function.
1415 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1416 __releases(this_rq->lock)
1417 __acquires(busiest->lock)
1418 __acquires(this_rq->lock)
1422 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1423 if (busiest < this_rq) {
1424 raw_spin_unlock(&this_rq->lock);
1425 raw_spin_lock(&busiest->lock);
1426 raw_spin_lock_nested(&this_rq->lock,
1427 SINGLE_DEPTH_NESTING);
1430 raw_spin_lock_nested(&busiest->lock,
1431 SINGLE_DEPTH_NESTING);
1436 #endif /* CONFIG_PREEMPT */
1439 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1441 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1443 if (unlikely(!irqs_disabled())) {
1444 /* printk() doesn't work good under rq->lock */
1445 raw_spin_unlock(&this_rq->lock);
1449 return _double_lock_balance(this_rq, busiest);
1452 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1453 __releases(busiest->lock)
1455 raw_spin_unlock(&busiest->lock);
1456 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1459 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1465 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1468 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1474 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1477 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1483 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1487 * double_rq_lock - safely lock two runqueues
1489 * Note this does not disable interrupts like task_rq_lock,
1490 * you need to do so manually before calling.
1492 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1493 __acquires(rq1->lock)
1494 __acquires(rq2->lock)
1496 BUG_ON(!irqs_disabled());
1498 raw_spin_lock(&rq1->lock);
1499 __acquire(rq2->lock); /* Fake it out ;) */
1502 raw_spin_lock(&rq1->lock);
1503 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1505 raw_spin_lock(&rq2->lock);
1506 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1512 * double_rq_unlock - safely unlock two runqueues
1514 * Note this does not restore interrupts like task_rq_unlock,
1515 * you need to do so manually after calling.
1517 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1518 __releases(rq1->lock)
1519 __releases(rq2->lock)
1521 raw_spin_unlock(&rq1->lock);
1523 raw_spin_unlock(&rq2->lock);
1525 __release(rq2->lock);
1528 #else /* CONFIG_SMP */
1531 * double_rq_lock - safely lock two runqueues
1533 * Note this does not disable interrupts like task_rq_lock,
1534 * you need to do so manually before calling.
1536 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1537 __acquires(rq1->lock)
1538 __acquires(rq2->lock)
1540 BUG_ON(!irqs_disabled());
1542 raw_spin_lock(&rq1->lock);
1543 __acquire(rq2->lock); /* Fake it out ;) */
1547 * double_rq_unlock - safely unlock two runqueues
1549 * Note this does not restore interrupts like task_rq_unlock,
1550 * you need to do so manually after calling.
1552 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1553 __releases(rq1->lock)
1554 __releases(rq2->lock)
1557 raw_spin_unlock(&rq1->lock);
1558 __release(rq2->lock);
1563 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1564 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1565 extern void print_cfs_stats(struct seq_file *m, int cpu);
1566 extern void print_rt_stats(struct seq_file *m, int cpu);
1567 extern void print_dl_stats(struct seq_file *m, int cpu);
1569 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1570 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
1571 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq);
1573 extern void cfs_bandwidth_usage_inc(void);
1574 extern void cfs_bandwidth_usage_dec(void);
1576 #ifdef CONFIG_NO_HZ_COMMON
1577 enum rq_nohz_flag_bits {
1582 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1585 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1587 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1588 DECLARE_PER_CPU(u64, cpu_softirq_time);
1590 #ifndef CONFIG_64BIT
1591 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1593 static inline void irq_time_write_begin(void)
1595 __this_cpu_inc(irq_time_seq.sequence);
1599 static inline void irq_time_write_end(void)
1602 __this_cpu_inc(irq_time_seq.sequence);
1605 static inline u64 irq_time_read(int cpu)
1611 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1612 irq_time = per_cpu(cpu_softirq_time, cpu) +
1613 per_cpu(cpu_hardirq_time, cpu);
1614 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1618 #else /* CONFIG_64BIT */
1619 static inline void irq_time_write_begin(void)
1623 static inline void irq_time_write_end(void)
1627 static inline u64 irq_time_read(int cpu)
1629 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1631 #endif /* CONFIG_64BIT */
1632 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */