2 #include <linux/sched.h>
3 #include <linux/sched/sysctl.h>
4 #include <linux/sched/rt.h>
5 #include <linux/sched/deadline.h>
6 #include <linux/mutex.h>
7 #include <linux/spinlock.h>
8 #include <linux/stop_machine.h>
9 #include <linux/tick.h>
10 #include <linux/slab.h>
13 #include "cpudeadline.h"
18 extern __read_mostly
int scheduler_running
;
20 extern unsigned long calc_load_update
;
21 extern atomic_long_t calc_load_tasks
;
23 extern long calc_load_fold_active(struct rq
*this_rq
);
24 extern void update_cpu_load_active(struct rq
*this_rq
);
27 * Helpers for converting nanosecond timing to jiffy resolution
29 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
32 * Increase resolution of nice-level calculations for 64-bit architectures.
33 * The extra resolution improves shares distribution and load balancing of
34 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
35 * hierarchies, especially on larger systems. This is not a user-visible change
36 * and does not change the user-interface for setting shares/weights.
38 * We increase resolution only if we have enough bits to allow this increased
39 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
40 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
43 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
44 # define SCHED_LOAD_RESOLUTION 10
45 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
46 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
48 # define SCHED_LOAD_RESOLUTION 0
49 # define scale_load(w) (w)
50 # define scale_load_down(w) (w)
53 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
54 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
56 #define NICE_0_LOAD SCHED_LOAD_SCALE
57 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
60 * Single value that decides SCHED_DEADLINE internal math precision.
61 * 10 -> just above 1us
62 * 9 -> just above 0.5us
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 fair_policy(int policy
)
77 return policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
;
80 static inline int rt_policy(int policy
)
82 return policy
== SCHED_FIFO
|| policy
== SCHED_RR
;
85 static inline int dl_policy(int policy
)
87 return policy
== SCHED_DEADLINE
;
90 static inline int task_has_rt_policy(struct task_struct
*p
)
92 return rt_policy(p
->policy
);
95 static inline int task_has_dl_policy(struct task_struct
*p
)
97 return dl_policy(p
->policy
);
100 static inline bool dl_time_before(u64 a
, u64 b
)
102 return (s64
)(a
- b
) < 0;
106 * Tells if entity @a should preempt entity @b.
109 dl_entity_preempt(struct sched_dl_entity
*a
, struct sched_dl_entity
*b
)
111 return dl_time_before(a
->deadline
, b
->deadline
);
115 * This is the priority-queue data structure of the RT scheduling class:
117 struct rt_prio_array
{
118 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
119 struct list_head queue
[MAX_RT_PRIO
];
122 struct rt_bandwidth
{
123 /* nests inside the rq lock: */
124 raw_spinlock_t rt_runtime_lock
;
127 struct hrtimer rt_period_timer
;
130 * To keep the bandwidth of -deadline tasks and groups under control
131 * we need some place where:
132 * - store the maximum -deadline bandwidth of the system (the group);
133 * - cache the fraction of that bandwidth that is currently allocated.
135 * This is all done in the data structure below. It is similar to the
136 * one used for RT-throttling (rt_bandwidth), with the main difference
137 * that, since here we are only interested in admission control, we
138 * do not decrease any runtime while the group "executes", neither we
139 * need a timer to replenish it.
141 * With respect to SMP, the bandwidth is given on a per-CPU basis,
143 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
144 * - dl_total_bw array contains, in the i-eth element, the currently
145 * allocated bandwidth on the i-eth CPU.
146 * Moreover, groups consume bandwidth on each CPU, while tasks only
147 * consume bandwidth on the CPU they're running on.
148 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
149 * that will be shown the next time the proc or cgroup controls will
150 * be red. It on its turn can be changed by writing on its own
153 struct dl_bandwidth
{
154 raw_spinlock_t dl_runtime_lock
;
159 static inline int dl_bandwidth_enabled(void)
161 return sysctl_sched_rt_runtime
>= 0;
164 extern struct dl_bw
*dl_bw_of(int i
);
171 extern struct mutex sched_domains_mutex
;
173 #ifdef CONFIG_CGROUP_SCHED
175 #include <linux/cgroup.h>
180 extern struct list_head task_groups
;
182 struct cfs_bandwidth
{
183 #ifdef CONFIG_CFS_BANDWIDTH
187 s64 hierarchal_quota
;
190 int idle
, timer_active
;
191 struct hrtimer period_timer
, slack_timer
;
192 struct list_head throttled_cfs_rq
;
195 int nr_periods
, nr_throttled
;
200 /* task group related information */
202 struct cgroup_subsys_state css
;
204 #ifdef CONFIG_FAIR_GROUP_SCHED
205 /* schedulable entities of this group on each cpu */
206 struct sched_entity
**se
;
207 /* runqueue "owned" by this group on each cpu */
208 struct cfs_rq
**cfs_rq
;
209 unsigned long shares
;
212 atomic_long_t load_avg
;
213 atomic_t runnable_avg
;
217 #ifdef CONFIG_RT_GROUP_SCHED
218 struct sched_rt_entity
**rt_se
;
219 struct rt_rq
**rt_rq
;
221 struct rt_bandwidth rt_bandwidth
;
225 struct list_head list
;
227 struct task_group
*parent
;
228 struct list_head siblings
;
229 struct list_head children
;
231 #ifdef CONFIG_SCHED_AUTOGROUP
232 struct autogroup
*autogroup
;
235 struct cfs_bandwidth cfs_bandwidth
;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
242 * A weight of 0 or 1 can cause arithmetics problems.
243 * A weight of a cfs_rq is the sum of weights of which entities
244 * are queued on this cfs_rq, so a weight of a entity should not be
245 * too large, so as the shares value of a task group.
246 * (The default weight is 1024 - so there's no practical
247 * limitation from this.)
249 #define MIN_SHARES (1UL << 1)
250 #define MAX_SHARES (1UL << 18)
253 typedef int (*tg_visitor
)(struct task_group
*, void *);
255 extern int walk_tg_tree_from(struct task_group
*from
,
256 tg_visitor down
, tg_visitor up
, void *data
);
259 * Iterate the full tree, calling @down when first entering a node and @up when
260 * leaving it for the final time.
262 * Caller must hold rcu_lock or sufficient equivalent.
264 static inline int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
266 return walk_tg_tree_from(&root_task_group
, down
, up
, data
);
269 extern int tg_nop(struct task_group
*tg
, void *data
);
271 extern void free_fair_sched_group(struct task_group
*tg
);
272 extern int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
);
273 extern void unregister_fair_sched_group(struct task_group
*tg
, int cpu
);
274 extern void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
275 struct sched_entity
*se
, int cpu
,
276 struct sched_entity
*parent
);
277 extern void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
);
278 extern int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
);
280 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
);
281 extern void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
);
282 extern void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
);
284 extern void free_rt_sched_group(struct task_group
*tg
);
285 extern int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
);
286 extern void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
287 struct sched_rt_entity
*rt_se
, int cpu
,
288 struct sched_rt_entity
*parent
);
290 extern struct task_group
*sched_create_group(struct task_group
*parent
);
291 extern void sched_online_group(struct task_group
*tg
,
292 struct task_group
*parent
);
293 extern void sched_destroy_group(struct task_group
*tg
);
294 extern void sched_offline_group(struct task_group
*tg
);
296 extern void sched_move_task(struct task_struct
*tsk
);
298 #ifdef CONFIG_FAIR_GROUP_SCHED
299 extern int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
);
302 #else /* CONFIG_CGROUP_SCHED */
304 struct cfs_bandwidth
{ };
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load
;
311 unsigned int nr_running
, h_nr_running
;
316 u64 min_vruntime_copy
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
323 * 'curr' points to currently running entity on this cfs_rq.
324 * It is set to NULL otherwise (i.e when none are currently running).
326 struct sched_entity
*curr
, *next
, *last
, *skip
;
328 #ifdef CONFIG_SCHED_DEBUG
329 unsigned int nr_spread_over
;
335 * Under CFS, load is tracked on a per-entity basis and aggregated up.
336 * This allows for the description of both thread and group usage (in
337 * the FAIR_GROUP_SCHED case).
339 unsigned long runnable_load_avg
, blocked_load_avg
;
340 atomic64_t decay_counter
;
342 atomic_long_t removed_load
;
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 /* Required to track per-cpu representation of a task_group */
346 u32 tg_runnable_contrib
;
347 unsigned long tg_load_contrib
;
350 * h_load = weight * f(tg)
352 * Where f(tg) is the recursive weight fraction assigned to
355 unsigned long h_load
;
356 u64 last_h_load_update
;
357 struct sched_entity
*h_load_next
;
358 #endif /* CONFIG_FAIR_GROUP_SCHED */
359 #endif /* CONFIG_SMP */
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
365 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
366 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
367 * (like users, containers etc.)
369 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
370 * list is used during load balance.
373 struct list_head leaf_cfs_rq_list
;
374 struct task_group
*tg
; /* group that "owns" this runqueue */
376 #ifdef CONFIG_CFS_BANDWIDTH
379 s64 runtime_remaining
;
381 u64 throttled_clock
, throttled_clock_task
;
382 u64 throttled_clock_task_time
;
383 int throttled
, throttle_count
;
384 struct list_head throttled_list
;
385 #endif /* CONFIG_CFS_BANDWIDTH */
386 #endif /* CONFIG_FAIR_GROUP_SCHED */
389 static inline int rt_bandwidth_enabled(void)
391 return sysctl_sched_rt_runtime
>= 0;
394 /* Real-Time classes' related field in a runqueue: */
396 struct rt_prio_array active
;
397 unsigned int rt_nr_running
;
398 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
400 int curr
; /* highest queued rt task prio */
402 int next
; /* next highest */
407 unsigned long rt_nr_migratory
;
408 unsigned long rt_nr_total
;
410 struct plist_head pushable_tasks
;
417 /* Nests inside the rq lock: */
418 raw_spinlock_t rt_runtime_lock
;
420 #ifdef CONFIG_RT_GROUP_SCHED
421 unsigned long rt_nr_boosted
;
424 struct task_group
*tg
;
428 /* Deadline class' related fields in a runqueue */
430 /* runqueue is an rbtree, ordered by deadline */
431 struct rb_root rb_root
;
432 struct rb_node
*rb_leftmost
;
434 unsigned long dl_nr_running
;
438 * Deadline values of the currently executing and the
439 * earliest ready task on this rq. Caching these facilitates
440 * the decision wether or not a ready but not running task
441 * should migrate somewhere else.
448 unsigned long dl_nr_migratory
;
452 * Tasks on this rq that can be pushed away. They are kept in
453 * an rb-tree, ordered by tasks' deadlines, with caching
454 * of the leftmost (earliest deadline) element.
456 struct rb_root pushable_dl_tasks_root
;
457 struct rb_node
*pushable_dl_tasks_leftmost
;
466 * We add the notion of a root-domain which will be used to define per-domain
467 * variables. Each exclusive cpuset essentially defines an island domain by
468 * fully partitioning the member cpus from any other cpuset. Whenever a new
469 * exclusive cpuset is created, we also create and attach a new root-domain
478 cpumask_var_t online
;
480 /* Indicate more than one runnable task for any CPU */
484 * The bit corresponding to a CPU gets set here if such CPU has more
485 * than one runnable -deadline task (as it is below for RT tasks).
487 cpumask_var_t dlo_mask
;
493 * The "RT overload" flag: it gets set if a CPU has more than
494 * one runnable RT task.
496 cpumask_var_t rto_mask
;
497 struct cpupri cpupri
;
500 extern struct root_domain def_root_domain
;
502 #endif /* CONFIG_SMP */
505 * This is the main, per-CPU runqueue data structure.
507 * Locking rule: those places that want to lock multiple runqueues
508 * (such as the load balancing or the thread migration code), lock
509 * acquire operations must be ordered by ascending &runqueue.
516 * nr_running and cpu_load should be in the same cacheline because
517 * remote CPUs use both these fields when doing load calculation.
519 unsigned int nr_running
;
520 #ifdef CONFIG_NUMA_BALANCING
521 unsigned int nr_numa_running
;
522 unsigned int nr_preferred_running
;
524 #define CPU_LOAD_IDX_MAX 5
525 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
526 unsigned long last_load_update_tick
;
527 #ifdef CONFIG_NO_HZ_COMMON
529 unsigned long nohz_flags
;
531 #ifdef CONFIG_NO_HZ_FULL
532 unsigned long last_sched_tick
;
534 int skip_clock_update
;
536 /* capture load from *all* tasks on this cpu: */
537 struct load_weight load
;
538 unsigned long nr_load_updates
;
545 #ifdef CONFIG_FAIR_GROUP_SCHED
546 /* list of leaf cfs_rq on this cpu: */
547 struct list_head leaf_cfs_rq_list
;
549 struct sched_avg avg
;
550 #endif /* CONFIG_FAIR_GROUP_SCHED */
553 * This is part of a global counter where only the total sum
554 * over all CPUs matters. A task can increase this counter on
555 * one CPU and if it got migrated afterwards it may decrease
556 * it on another CPU. Always updated under the runqueue lock:
558 unsigned long nr_uninterruptible
;
560 struct task_struct
*curr
, *idle
, *stop
;
561 unsigned long next_balance
;
562 struct mm_struct
*prev_mm
;
570 struct root_domain
*rd
;
571 struct sched_domain
*sd
;
573 unsigned long cpu_capacity
;
575 unsigned char idle_balance
;
576 /* For active balancing */
580 struct cpu_stop_work active_balance_work
;
581 /* cpu of this runqueue: */
585 struct list_head cfs_tasks
;
592 /* This is used to determine avg_idle's max value */
593 u64 max_idle_balance_cost
;
596 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
599 #ifdef CONFIG_PARAVIRT
602 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
603 u64 prev_steal_time_rq
;
606 /* calc_load related fields */
607 unsigned long calc_load_update
;
608 long calc_load_active
;
610 #ifdef CONFIG_SCHED_HRTICK
612 int hrtick_csd_pending
;
613 struct call_single_data hrtick_csd
;
615 struct hrtimer hrtick_timer
;
618 #ifdef CONFIG_SCHEDSTATS
620 struct sched_info rq_sched_info
;
621 unsigned long long rq_cpu_time
;
622 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
624 /* sys_sched_yield() stats */
625 unsigned int yld_count
;
627 /* schedule() stats */
628 unsigned int sched_count
;
629 unsigned int sched_goidle
;
631 /* try_to_wake_up() stats */
632 unsigned int ttwu_count
;
633 unsigned int ttwu_local
;
637 struct llist_head wake_list
;
641 static inline int cpu_of(struct rq
*rq
)
650 DECLARE_PER_CPU(struct rq
, runqueues
);
652 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
653 #define this_rq() (&__get_cpu_var(runqueues))
654 #define task_rq(p) cpu_rq(task_cpu(p))
655 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
656 #define raw_rq() (&__raw_get_cpu_var(runqueues))
658 static inline u64
rq_clock(struct rq
*rq
)
663 static inline u64
rq_clock_task(struct rq
*rq
)
665 return rq
->clock_task
;
668 #ifdef CONFIG_NUMA_BALANCING
669 extern void sched_setnuma(struct task_struct
*p
, int node
);
670 extern int migrate_task_to(struct task_struct
*p
, int cpu
);
671 extern int migrate_swap(struct task_struct
*, struct task_struct
*);
672 #endif /* CONFIG_NUMA_BALANCING */
676 extern void sched_ttwu_pending(void);
678 #define rcu_dereference_check_sched_domain(p) \
679 rcu_dereference_check((p), \
680 lockdep_is_held(&sched_domains_mutex))
683 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684 * See detach_destroy_domains: synchronize_sched for details.
686 * The domain tree of any CPU may only be accessed from within
687 * preempt-disabled sections.
689 #define for_each_domain(cpu, __sd) \
690 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
691 __sd; __sd = __sd->parent)
693 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
696 * highest_flag_domain - Return highest sched_domain containing flag.
697 * @cpu: The cpu whose highest level of sched domain is to
699 * @flag: The flag to check for the highest sched_domain
702 * Returns the highest sched_domain of a cpu which contains the given flag.
704 static inline struct sched_domain
*highest_flag_domain(int cpu
, int flag
)
706 struct sched_domain
*sd
, *hsd
= NULL
;
708 for_each_domain(cpu
, sd
) {
709 if (!(sd
->flags
& flag
))
717 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
719 struct sched_domain
*sd
;
721 for_each_domain(cpu
, sd
) {
722 if (sd
->flags
& flag
)
729 DECLARE_PER_CPU(struct sched_domain
*, sd_llc
);
730 DECLARE_PER_CPU(int, sd_llc_size
);
731 DECLARE_PER_CPU(int, sd_llc_id
);
732 DECLARE_PER_CPU(struct sched_domain
*, sd_numa
);
733 DECLARE_PER_CPU(struct sched_domain
*, sd_busy
);
734 DECLARE_PER_CPU(struct sched_domain
*, sd_asym
);
736 struct sched_group_capacity
{
739 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
742 unsigned int capacity
, capacity_orig
;
743 unsigned long next_update
;
744 int imbalance
; /* XXX unrelated to capacity but shared group state */
746 * Number of busy cpus in this group.
748 atomic_t nr_busy_cpus
;
750 unsigned long cpumask
[0]; /* iteration mask */
754 struct sched_group
*next
; /* Must be a circular list */
757 unsigned int group_weight
;
758 struct sched_group_capacity
*sgc
;
761 * The CPUs this group covers.
763 * NOTE: this field is variable length. (Allocated dynamically
764 * by attaching extra space to the end of the structure,
765 * depending on how many CPUs the kernel has booted up with)
767 unsigned long cpumask
[0];
770 static inline struct cpumask
*sched_group_cpus(struct sched_group
*sg
)
772 return to_cpumask(sg
->cpumask
);
776 * cpumask masking which cpus in the group are allowed to iterate up the domain
779 static inline struct cpumask
*sched_group_mask(struct sched_group
*sg
)
781 return to_cpumask(sg
->sgc
->cpumask
);
785 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
786 * @group: The group whose first cpu is to be returned.
788 static inline unsigned int group_first_cpu(struct sched_group
*group
)
790 return cpumask_first(sched_group_cpus(group
));
793 extern int group_balance_cpu(struct sched_group
*sg
);
797 static inline void sched_ttwu_pending(void) { }
799 #endif /* CONFIG_SMP */
802 #include "auto_group.h"
804 #ifdef CONFIG_CGROUP_SCHED
807 * Return the group to which this tasks belongs.
809 * We cannot use task_css() and friends because the cgroup subsystem
810 * changes that value before the cgroup_subsys::attach() method is called,
811 * therefore we cannot pin it and might observe the wrong value.
813 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
814 * core changes this before calling sched_move_task().
816 * Instead we use a 'copy' which is updated from sched_move_task() while
817 * holding both task_struct::pi_lock and rq::lock.
819 static inline struct task_group
*task_group(struct task_struct
*p
)
821 return p
->sched_task_group
;
824 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
825 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
827 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
828 struct task_group
*tg
= task_group(p
);
831 #ifdef CONFIG_FAIR_GROUP_SCHED
832 p
->se
.cfs_rq
= tg
->cfs_rq
[cpu
];
833 p
->se
.parent
= tg
->se
[cpu
];
836 #ifdef CONFIG_RT_GROUP_SCHED
837 p
->rt
.rt_rq
= tg
->rt_rq
[cpu
];
838 p
->rt
.parent
= tg
->rt_se
[cpu
];
842 #else /* CONFIG_CGROUP_SCHED */
844 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
845 static inline struct task_group
*task_group(struct task_struct
*p
)
850 #endif /* CONFIG_CGROUP_SCHED */
852 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
857 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
858 * successfuly executed on another CPU. We must ensure that updates of
859 * per-task data have been completed by this moment.
862 task_thread_info(p
)->cpu
= cpu
;
868 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
870 #ifdef CONFIG_SCHED_DEBUG
871 # include <linux/static_key.h>
872 # define const_debug __read_mostly
874 # define const_debug const
877 extern const_debug
unsigned int sysctl_sched_features
;
879 #define SCHED_FEAT(name, enabled) \
880 __SCHED_FEAT_##name ,
883 #include "features.h"
889 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
890 #define SCHED_FEAT(name, enabled) \
891 static __always_inline bool static_branch_##name(struct static_key *key) \
893 return static_key_##enabled(key); \
896 #include "features.h"
900 extern struct static_key sched_feat_keys
[__SCHED_FEAT_NR
];
901 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
902 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
903 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
904 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
906 #ifdef CONFIG_NUMA_BALANCING
907 #define sched_feat_numa(x) sched_feat(x)
908 #ifdef CONFIG_SCHED_DEBUG
909 #define numabalancing_enabled sched_feat_numa(NUMA)
911 extern bool numabalancing_enabled
;
912 #endif /* CONFIG_SCHED_DEBUG */
914 #define sched_feat_numa(x) (0)
915 #define numabalancing_enabled (0)
916 #endif /* CONFIG_NUMA_BALANCING */
918 static inline u64
global_rt_period(void)
920 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
923 static inline u64
global_rt_runtime(void)
925 if (sysctl_sched_rt_runtime
< 0)
928 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
931 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
933 return rq
->curr
== p
;
936 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
941 return task_current(rq
, p
);
946 #ifndef prepare_arch_switch
947 # define prepare_arch_switch(next) do { } while (0)
949 #ifndef finish_arch_switch
950 # define finish_arch_switch(prev) do { } while (0)
952 #ifndef finish_arch_post_lock_switch
953 # define finish_arch_post_lock_switch() do { } while (0)
956 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
957 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
961 * We can optimise this out completely for !SMP, because the
962 * SMP rebalancing from interrupt is the only thing that cares
969 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
973 * After ->on_cpu is cleared, the task can be moved to a different CPU.
974 * We must ensure this doesn't happen until the switch is completely
980 #ifdef CONFIG_DEBUG_SPINLOCK
981 /* this is a valid case when another task releases the spinlock */
982 rq
->lock
.owner
= current
;
985 * If we are tracking spinlock dependencies then we have to
986 * fix up the runqueue lock - which gets 'carried over' from
989 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
991 raw_spin_unlock_irq(&rq
->lock
);
994 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
995 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
999 * We can optimise this out completely for !SMP, because the
1000 * SMP rebalancing from interrupt is the only thing that cares
1005 raw_spin_unlock(&rq
->lock
);
1008 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1012 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1013 * We must ensure this doesn't happen until the switch is completely
1021 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1026 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1027 #define WF_FORK 0x02 /* child wakeup after fork */
1028 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1031 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1032 * of tasks with abnormal "nice" values across CPUs the contribution that
1033 * each task makes to its run queue's load is weighted according to its
1034 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1035 * scaled version of the new time slice allocation that they receive on time
1039 #define WEIGHT_IDLEPRIO 3
1040 #define WMULT_IDLEPRIO 1431655765
1043 * Nice levels are multiplicative, with a gentle 10% change for every
1044 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1045 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1046 * that remained on nice 0.
1048 * The "10% effect" is relative and cumulative: from _any_ nice level,
1049 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1050 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1051 * If a task goes up by ~10% and another task goes down by ~10% then
1052 * the relative distance between them is ~25%.)
1054 static const int prio_to_weight
[40] = {
1055 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1056 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1057 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1058 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1059 /* 0 */ 1024, 820, 655, 526, 423,
1060 /* 5 */ 335, 272, 215, 172, 137,
1061 /* 10 */ 110, 87, 70, 56, 45,
1062 /* 15 */ 36, 29, 23, 18, 15,
1066 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1068 * In cases where the weight does not change often, we can use the
1069 * precalculated inverse to speed up arithmetics by turning divisions
1070 * into multiplications:
1072 static const u32 prio_to_wmult
[40] = {
1073 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1074 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1075 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1076 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1077 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1078 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1079 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1080 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1083 #define ENQUEUE_WAKEUP 1
1084 #define ENQUEUE_HEAD 2
1086 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1088 #define ENQUEUE_WAKING 0
1090 #define ENQUEUE_REPLENISH 8
1092 #define DEQUEUE_SLEEP 1
1094 #define RETRY_TASK ((void *)-1UL)
1096 struct sched_class
{
1097 const struct sched_class
*next
;
1099 void (*enqueue_task
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1100 void (*dequeue_task
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1101 void (*yield_task
) (struct rq
*rq
);
1102 bool (*yield_to_task
) (struct rq
*rq
, struct task_struct
*p
, bool preempt
);
1104 void (*check_preempt_curr
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1107 * It is the responsibility of the pick_next_task() method that will
1108 * return the next task to call put_prev_task() on the @prev task or
1109 * something equivalent.
1111 * May return RETRY_TASK when it finds a higher prio class has runnable
1114 struct task_struct
* (*pick_next_task
) (struct rq
*rq
,
1115 struct task_struct
*prev
);
1116 void (*put_prev_task
) (struct rq
*rq
, struct task_struct
*p
);
1119 int (*select_task_rq
)(struct task_struct
*p
, int task_cpu
, int sd_flag
, int flags
);
1120 void (*migrate_task_rq
)(struct task_struct
*p
, int next_cpu
);
1122 void (*post_schedule
) (struct rq
*this_rq
);
1123 void (*task_waking
) (struct task_struct
*task
);
1124 void (*task_woken
) (struct rq
*this_rq
, struct task_struct
*task
);
1126 void (*set_cpus_allowed
)(struct task_struct
*p
,
1127 const struct cpumask
*newmask
);
1129 void (*rq_online
)(struct rq
*rq
);
1130 void (*rq_offline
)(struct rq
*rq
);
1133 void (*set_curr_task
) (struct rq
*rq
);
1134 void (*task_tick
) (struct rq
*rq
, struct task_struct
*p
, int queued
);
1135 void (*task_fork
) (struct task_struct
*p
);
1136 void (*task_dead
) (struct task_struct
*p
);
1138 void (*switched_from
) (struct rq
*this_rq
, struct task_struct
*task
);
1139 void (*switched_to
) (struct rq
*this_rq
, struct task_struct
*task
);
1140 void (*prio_changed
) (struct rq
*this_rq
, struct task_struct
*task
,
1143 unsigned int (*get_rr_interval
) (struct rq
*rq
,
1144 struct task_struct
*task
);
1146 #ifdef CONFIG_FAIR_GROUP_SCHED
1147 void (*task_move_group
) (struct task_struct
*p
, int on_rq
);
1151 static inline void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
1153 prev
->sched_class
->put_prev_task(rq
, prev
);
1156 #define sched_class_highest (&stop_sched_class)
1157 #define for_each_class(class) \
1158 for (class = sched_class_highest; class; class = class->next)
1160 extern const struct sched_class stop_sched_class
;
1161 extern const struct sched_class dl_sched_class
;
1162 extern const struct sched_class rt_sched_class
;
1163 extern const struct sched_class fair_sched_class
;
1164 extern const struct sched_class idle_sched_class
;
1169 extern void update_group_capacity(struct sched_domain
*sd
, int cpu
);
1171 extern void trigger_load_balance(struct rq
*rq
);
1173 extern void idle_enter_fair(struct rq
*this_rq
);
1174 extern void idle_exit_fair(struct rq
*this_rq
);
1178 static inline void idle_enter_fair(struct rq
*rq
) { }
1179 static inline void idle_exit_fair(struct rq
*rq
) { }
1183 extern void sysrq_sched_debug_show(void);
1184 extern void sched_init_granularity(void);
1185 extern void update_max_interval(void);
1187 extern void init_sched_dl_class(void);
1188 extern void init_sched_rt_class(void);
1189 extern void init_sched_fair_class(void);
1190 extern void init_sched_dl_class(void);
1192 extern void resched_curr(struct rq
*rq
);
1193 extern void resched_cpu(int cpu
);
1195 extern struct rt_bandwidth def_rt_bandwidth
;
1196 extern void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
);
1198 extern struct dl_bandwidth def_dl_bandwidth
;
1199 extern void init_dl_bandwidth(struct dl_bandwidth
*dl_b
, u64 period
, u64 runtime
);
1200 extern void init_dl_task_timer(struct sched_dl_entity
*dl_se
);
1202 unsigned long to_ratio(u64 period
, u64 runtime
);
1204 extern void update_idle_cpu_load(struct rq
*this_rq
);
1206 extern void init_task_runnable_average(struct task_struct
*p
);
1208 static inline void add_nr_running(struct rq
*rq
, unsigned count
)
1210 unsigned prev_nr
= rq
->nr_running
;
1212 rq
->nr_running
= prev_nr
+ count
;
1214 if (prev_nr
< 2 && rq
->nr_running
>= 2) {
1216 if (!rq
->rd
->overload
)
1217 rq
->rd
->overload
= true;
1220 #ifdef CONFIG_NO_HZ_FULL
1221 if (tick_nohz_full_cpu(rq
->cpu
)) {
1223 * Tick is needed if more than one task runs on a CPU.
1224 * Send the target an IPI to kick it out of nohz mode.
1226 * We assume that IPI implies full memory barrier and the
1227 * new value of rq->nr_running is visible on reception
1230 tick_nohz_full_kick_cpu(rq
->cpu
);
1236 static inline void sub_nr_running(struct rq
*rq
, unsigned count
)
1238 rq
->nr_running
-= count
;
1241 static inline void rq_last_tick_reset(struct rq
*rq
)
1243 #ifdef CONFIG_NO_HZ_FULL
1244 rq
->last_sched_tick
= jiffies
;
1248 extern void update_rq_clock(struct rq
*rq
);
1250 extern void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
);
1251 extern void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
);
1253 extern void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
1255 extern const_debug
unsigned int sysctl_sched_time_avg
;
1256 extern const_debug
unsigned int sysctl_sched_nr_migrate
;
1257 extern const_debug
unsigned int sysctl_sched_migration_cost
;
1259 static inline u64
sched_avg_period(void)
1261 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1264 #ifdef CONFIG_SCHED_HRTICK
1268 * - enabled by features
1269 * - hrtimer is actually high res
1271 static inline int hrtick_enabled(struct rq
*rq
)
1273 if (!sched_feat(HRTICK
))
1275 if (!cpu_active(cpu_of(rq
)))
1277 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1280 void hrtick_start(struct rq
*rq
, u64 delay
);
1284 static inline int hrtick_enabled(struct rq
*rq
)
1289 #endif /* CONFIG_SCHED_HRTICK */
1292 extern void sched_avg_update(struct rq
*rq
);
1293 static inline void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1295 rq
->rt_avg
+= rt_delta
;
1296 sched_avg_update(rq
);
1299 static inline void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
) { }
1300 static inline void sched_avg_update(struct rq
*rq
) { }
1303 extern void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
);
1306 #ifdef CONFIG_PREEMPT
1308 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1311 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1312 * way at the expense of forcing extra atomic operations in all
1313 * invocations. This assures that the double_lock is acquired using the
1314 * same underlying policy as the spinlock_t on this architecture, which
1315 * reduces latency compared to the unfair variant below. However, it
1316 * also adds more overhead and therefore may reduce throughput.
1318 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1319 __releases(this_rq
->lock
)
1320 __acquires(busiest
->lock
)
1321 __acquires(this_rq
->lock
)
1323 raw_spin_unlock(&this_rq
->lock
);
1324 double_rq_lock(this_rq
, busiest
);
1331 * Unfair double_lock_balance: Optimizes throughput at the expense of
1332 * latency by eliminating extra atomic operations when the locks are
1333 * already in proper order on entry. This favors lower cpu-ids and will
1334 * grant the double lock to lower cpus over higher ids under contention,
1335 * regardless of entry order into the function.
1337 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1338 __releases(this_rq
->lock
)
1339 __acquires(busiest
->lock
)
1340 __acquires(this_rq
->lock
)
1344 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1345 if (busiest
< this_rq
) {
1346 raw_spin_unlock(&this_rq
->lock
);
1347 raw_spin_lock(&busiest
->lock
);
1348 raw_spin_lock_nested(&this_rq
->lock
,
1349 SINGLE_DEPTH_NESTING
);
1352 raw_spin_lock_nested(&busiest
->lock
,
1353 SINGLE_DEPTH_NESTING
);
1358 #endif /* CONFIG_PREEMPT */
1361 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1363 static inline int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1365 if (unlikely(!irqs_disabled())) {
1366 /* printk() doesn't work good under rq->lock */
1367 raw_spin_unlock(&this_rq
->lock
);
1371 return _double_lock_balance(this_rq
, busiest
);
1374 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1375 __releases(busiest
->lock
)
1377 raw_spin_unlock(&busiest
->lock
);
1378 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1381 static inline void double_lock(spinlock_t
*l1
, spinlock_t
*l2
)
1387 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1390 static inline void double_lock_irq(spinlock_t
*l1
, spinlock_t
*l2
)
1396 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1399 static inline void double_raw_lock(raw_spinlock_t
*l1
, raw_spinlock_t
*l2
)
1405 raw_spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1409 * double_rq_lock - safely lock two runqueues
1411 * Note this does not disable interrupts like task_rq_lock,
1412 * you need to do so manually before calling.
1414 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1415 __acquires(rq1
->lock
)
1416 __acquires(rq2
->lock
)
1418 BUG_ON(!irqs_disabled());
1420 raw_spin_lock(&rq1
->lock
);
1421 __acquire(rq2
->lock
); /* Fake it out ;) */
1424 raw_spin_lock(&rq1
->lock
);
1425 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1427 raw_spin_lock(&rq2
->lock
);
1428 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1434 * double_rq_unlock - safely unlock two runqueues
1436 * Note this does not restore interrupts like task_rq_unlock,
1437 * you need to do so manually after calling.
1439 static inline void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1440 __releases(rq1
->lock
)
1441 __releases(rq2
->lock
)
1443 raw_spin_unlock(&rq1
->lock
);
1445 raw_spin_unlock(&rq2
->lock
);
1447 __release(rq2
->lock
);
1450 #else /* CONFIG_SMP */
1453 * double_rq_lock - safely lock two runqueues
1455 * Note this does not disable interrupts like task_rq_lock,
1456 * you need to do so manually before calling.
1458 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1459 __acquires(rq1
->lock
)
1460 __acquires(rq2
->lock
)
1462 BUG_ON(!irqs_disabled());
1464 raw_spin_lock(&rq1
->lock
);
1465 __acquire(rq2
->lock
); /* Fake it out ;) */
1469 * double_rq_unlock - safely unlock two runqueues
1471 * Note this does not restore interrupts like task_rq_unlock,
1472 * you need to do so manually after calling.
1474 static inline void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1475 __releases(rq1
->lock
)
1476 __releases(rq2
->lock
)
1479 raw_spin_unlock(&rq1
->lock
);
1480 __release(rq2
->lock
);
1485 extern struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
);
1486 extern struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
);
1487 extern void print_cfs_stats(struct seq_file
*m
, int cpu
);
1488 extern void print_rt_stats(struct seq_file
*m
, int cpu
);
1490 extern void init_cfs_rq(struct cfs_rq
*cfs_rq
);
1491 extern void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
);
1492 extern void init_dl_rq(struct dl_rq
*dl_rq
, struct rq
*rq
);
1494 extern void cfs_bandwidth_usage_inc(void);
1495 extern void cfs_bandwidth_usage_dec(void);
1497 #ifdef CONFIG_NO_HZ_COMMON
1498 enum rq_nohz_flag_bits
{
1503 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1506 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1508 DECLARE_PER_CPU(u64
, cpu_hardirq_time
);
1509 DECLARE_PER_CPU(u64
, cpu_softirq_time
);
1511 #ifndef CONFIG_64BIT
1512 DECLARE_PER_CPU(seqcount_t
, irq_time_seq
);
1514 static inline void irq_time_write_begin(void)
1516 __this_cpu_inc(irq_time_seq
.sequence
);
1520 static inline void irq_time_write_end(void)
1523 __this_cpu_inc(irq_time_seq
.sequence
);
1526 static inline u64
irq_time_read(int cpu
)
1532 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1533 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1534 per_cpu(cpu_hardirq_time
, cpu
);
1535 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1539 #else /* CONFIG_64BIT */
1540 static inline void irq_time_write_begin(void)
1544 static inline void irq_time_write_end(void)
1548 static inline u64
irq_time_read(int cpu
)
1550 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1552 #endif /* CONFIG_64BIT */
1553 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */