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 * Convert user-nice values [ -20 ... 0 ... 19 ]
28 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
31 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
32 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
33 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
36 * 'User priority' is the nice value converted to something we
37 * can work with better when scaling various scheduler parameters,
38 * it's a [ 0 ... 39 ] range.
40 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
41 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
42 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
45 * Helpers for converting nanosecond timing to jiffy resolution
47 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
50 * Increase resolution of nice-level calculations for 64-bit architectures.
51 * The extra resolution improves shares distribution and load balancing of
52 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
53 * hierarchies, especially on larger systems. This is not a user-visible change
54 * and does not change the user-interface for setting shares/weights.
56 * We increase resolution only if we have enough bits to allow this increased
57 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
58 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
61 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
62 # define SCHED_LOAD_RESOLUTION 10
63 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
64 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
66 # define SCHED_LOAD_RESOLUTION 0
67 # define scale_load(w) (w)
68 # define scale_load_down(w) (w)
71 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
72 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
74 #define NICE_0_LOAD SCHED_LOAD_SCALE
75 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
78 * Single value that decides SCHED_DEADLINE internal math precision.
79 * 10 -> just above 1us
80 * 9 -> just above 0.5us
85 * These are the 'tuning knobs' of the scheduler:
89 * single value that denotes runtime == period, ie unlimited time.
91 #define RUNTIME_INF ((u64)~0ULL)
93 static inline int fair_policy(int policy
)
95 return policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
;
98 static inline int rt_policy(int policy
)
100 return policy
== SCHED_FIFO
|| policy
== SCHED_RR
;
103 static inline int dl_policy(int policy
)
105 return policy
== SCHED_DEADLINE
;
108 static inline int task_has_rt_policy(struct task_struct
*p
)
110 return rt_policy(p
->policy
);
113 static inline int task_has_dl_policy(struct task_struct
*p
)
115 return dl_policy(p
->policy
);
118 static inline bool dl_time_before(u64 a
, u64 b
)
120 return (s64
)(a
- b
) < 0;
124 * Tells if entity @a should preempt entity @b.
127 dl_entity_preempt(struct sched_dl_entity
*a
, struct sched_dl_entity
*b
)
129 return dl_time_before(a
->deadline
, b
->deadline
);
133 * This is the priority-queue data structure of the RT scheduling class:
135 struct rt_prio_array
{
136 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
137 struct list_head queue
[MAX_RT_PRIO
];
140 struct rt_bandwidth
{
141 /* nests inside the rq lock: */
142 raw_spinlock_t rt_runtime_lock
;
145 struct hrtimer rt_period_timer
;
148 * To keep the bandwidth of -deadline tasks and groups under control
149 * we need some place where:
150 * - store the maximum -deadline bandwidth of the system (the group);
151 * - cache the fraction of that bandwidth that is currently allocated.
153 * This is all done in the data structure below. It is similar to the
154 * one used for RT-throttling (rt_bandwidth), with the main difference
155 * that, since here we are only interested in admission control, we
156 * do not decrease any runtime while the group "executes", neither we
157 * need a timer to replenish it.
159 * With respect to SMP, the bandwidth is given on a per-CPU basis,
161 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
162 * - dl_total_bw array contains, in the i-eth element, the currently
163 * allocated bandwidth on the i-eth CPU.
164 * Moreover, groups consume bandwidth on each CPU, while tasks only
165 * consume bandwidth on the CPU they're running on.
166 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
167 * that will be shown the next time the proc or cgroup controls will
168 * be red. It on its turn can be changed by writing on its own
171 struct dl_bandwidth
{
172 raw_spinlock_t dl_runtime_lock
;
177 static inline int dl_bandwidth_enabled(void)
179 return sysctl_sched_rt_runtime
>= 0;
182 extern struct dl_bw
*dl_bw_of(int i
);
189 extern struct mutex sched_domains_mutex
;
191 #ifdef CONFIG_CGROUP_SCHED
193 #include <linux/cgroup.h>
198 extern struct list_head task_groups
;
200 struct cfs_bandwidth
{
201 #ifdef CONFIG_CFS_BANDWIDTH
205 s64 hierarchal_quota
;
208 int idle
, timer_active
;
209 struct hrtimer period_timer
, slack_timer
;
210 struct list_head throttled_cfs_rq
;
213 int nr_periods
, nr_throttled
;
218 /* task group related information */
220 struct cgroup_subsys_state css
;
222 #ifdef CONFIG_FAIR_GROUP_SCHED
223 /* schedulable entities of this group on each cpu */
224 struct sched_entity
**se
;
225 /* runqueue "owned" by this group on each cpu */
226 struct cfs_rq
**cfs_rq
;
227 unsigned long shares
;
230 atomic_long_t load_avg
;
231 atomic_t runnable_avg
;
235 #ifdef CONFIG_RT_GROUP_SCHED
236 struct sched_rt_entity
**rt_se
;
237 struct rt_rq
**rt_rq
;
239 struct rt_bandwidth rt_bandwidth
;
243 struct list_head list
;
245 struct task_group
*parent
;
246 struct list_head siblings
;
247 struct list_head children
;
249 #ifdef CONFIG_SCHED_AUTOGROUP
250 struct autogroup
*autogroup
;
253 struct cfs_bandwidth cfs_bandwidth
;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
260 * A weight of 0 or 1 can cause arithmetics problems.
261 * A weight of a cfs_rq is the sum of weights of which entities
262 * are queued on this cfs_rq, so a weight of a entity should not be
263 * too large, so as the shares value of a task group.
264 * (The default weight is 1024 - so there's no practical
265 * limitation from this.)
267 #define MIN_SHARES (1UL << 1)
268 #define MAX_SHARES (1UL << 18)
271 typedef int (*tg_visitor
)(struct task_group
*, void *);
273 extern int walk_tg_tree_from(struct task_group
*from
,
274 tg_visitor down
, tg_visitor up
, void *data
);
277 * Iterate the full tree, calling @down when first entering a node and @up when
278 * leaving it for the final time.
280 * Caller must hold rcu_lock or sufficient equivalent.
282 static inline int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
284 return walk_tg_tree_from(&root_task_group
, down
, up
, data
);
287 extern int tg_nop(struct task_group
*tg
, void *data
);
289 extern void free_fair_sched_group(struct task_group
*tg
);
290 extern int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
);
291 extern void unregister_fair_sched_group(struct task_group
*tg
, int cpu
);
292 extern void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
293 struct sched_entity
*se
, int cpu
,
294 struct sched_entity
*parent
);
295 extern void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
);
296 extern int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
);
298 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
);
299 extern void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
);
300 extern void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
);
302 extern void free_rt_sched_group(struct task_group
*tg
);
303 extern int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
);
304 extern void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
305 struct sched_rt_entity
*rt_se
, int cpu
,
306 struct sched_rt_entity
*parent
);
308 extern struct task_group
*sched_create_group(struct task_group
*parent
);
309 extern void sched_online_group(struct task_group
*tg
,
310 struct task_group
*parent
);
311 extern void sched_destroy_group(struct task_group
*tg
);
312 extern void sched_offline_group(struct task_group
*tg
);
314 extern void sched_move_task(struct task_struct
*tsk
);
316 #ifdef CONFIG_FAIR_GROUP_SCHED
317 extern int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
);
320 #else /* CONFIG_CGROUP_SCHED */
322 struct cfs_bandwidth
{ };
324 #endif /* CONFIG_CGROUP_SCHED */
326 /* CFS-related fields in a runqueue */
328 struct load_weight load
;
329 unsigned int nr_running
, h_nr_running
;
334 u64 min_vruntime_copy
;
337 struct rb_root tasks_timeline
;
338 struct rb_node
*rb_leftmost
;
341 * 'curr' points to currently running entity on this cfs_rq.
342 * It is set to NULL otherwise (i.e when none are currently running).
344 struct sched_entity
*curr
, *next
, *last
, *skip
;
346 #ifdef CONFIG_SCHED_DEBUG
347 unsigned int nr_spread_over
;
353 * Under CFS, load is tracked on a per-entity basis and aggregated up.
354 * This allows for the description of both thread and group usage (in
355 * the FAIR_GROUP_SCHED case).
357 unsigned long runnable_load_avg
, blocked_load_avg
;
358 atomic64_t decay_counter
;
360 atomic_long_t removed_load
;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 /* Required to track per-cpu representation of a task_group */
364 u32 tg_runnable_contrib
;
365 unsigned long tg_load_contrib
;
368 * h_load = weight * f(tg)
370 * Where f(tg) is the recursive weight fraction assigned to
373 unsigned long h_load
;
374 u64 last_h_load_update
;
375 struct sched_entity
*h_load_next
;
376 #endif /* CONFIG_FAIR_GROUP_SCHED */
377 #endif /* CONFIG_SMP */
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
385 * (like users, containers etc.)
387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
388 * list is used during load balance.
391 struct list_head leaf_cfs_rq_list
;
392 struct task_group
*tg
; /* group that "owns" this runqueue */
394 #ifdef CONFIG_CFS_BANDWIDTH
397 s64 runtime_remaining
;
399 u64 throttled_clock
, throttled_clock_task
;
400 u64 throttled_clock_task_time
;
401 int throttled
, throttle_count
;
402 struct list_head throttled_list
;
403 #endif /* CONFIG_CFS_BANDWIDTH */
404 #endif /* CONFIG_FAIR_GROUP_SCHED */
407 static inline int rt_bandwidth_enabled(void)
409 return sysctl_sched_rt_runtime
>= 0;
412 /* Real-Time classes' related field in a runqueue: */
414 struct rt_prio_array active
;
415 unsigned int rt_nr_running
;
416 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
418 int curr
; /* highest queued rt task prio */
420 int next
; /* next highest */
425 unsigned long rt_nr_migratory
;
426 unsigned long rt_nr_total
;
428 struct plist_head pushable_tasks
;
433 /* Nests inside the rq lock: */
434 raw_spinlock_t rt_runtime_lock
;
436 #ifdef CONFIG_RT_GROUP_SCHED
437 unsigned long rt_nr_boosted
;
440 struct task_group
*tg
;
444 /* Deadline class' related fields in a runqueue */
446 /* runqueue is an rbtree, ordered by deadline */
447 struct rb_root rb_root
;
448 struct rb_node
*rb_leftmost
;
450 unsigned long dl_nr_running
;
454 * Deadline values of the currently executing and the
455 * earliest ready task on this rq. Caching these facilitates
456 * the decision wether or not a ready but not running task
457 * should migrate somewhere else.
464 unsigned long dl_nr_migratory
;
468 * Tasks on this rq that can be pushed away. They are kept in
469 * an rb-tree, ordered by tasks' deadlines, with caching
470 * of the leftmost (earliest deadline) element.
472 struct rb_root pushable_dl_tasks_root
;
473 struct rb_node
*pushable_dl_tasks_leftmost
;
482 * We add the notion of a root-domain which will be used to define per-domain
483 * variables. Each exclusive cpuset essentially defines an island domain by
484 * fully partitioning the member cpus from any other cpuset. Whenever a new
485 * exclusive cpuset is created, we also create and attach a new root-domain
494 cpumask_var_t online
;
497 * The bit corresponding to a CPU gets set here if such CPU has more
498 * than one runnable -deadline task (as it is below for RT tasks).
500 cpumask_var_t dlo_mask
;
506 * The "RT overload" flag: it gets set if a CPU has more than
507 * one runnable RT task.
509 cpumask_var_t rto_mask
;
510 struct cpupri cpupri
;
513 extern struct root_domain def_root_domain
;
515 #endif /* CONFIG_SMP */
518 * This is the main, per-CPU runqueue data structure.
520 * Locking rule: those places that want to lock multiple runqueues
521 * (such as the load balancing or the thread migration code), lock
522 * acquire operations must be ordered by ascending &runqueue.
529 * nr_running and cpu_load should be in the same cacheline because
530 * remote CPUs use both these fields when doing load calculation.
532 unsigned int nr_running
;
533 #ifdef CONFIG_NUMA_BALANCING
534 unsigned int nr_numa_running
;
535 unsigned int nr_preferred_running
;
537 #define CPU_LOAD_IDX_MAX 5
538 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
539 unsigned long last_load_update_tick
;
540 #ifdef CONFIG_NO_HZ_COMMON
542 unsigned long nohz_flags
;
544 #ifdef CONFIG_NO_HZ_FULL
545 unsigned long last_sched_tick
;
547 int skip_clock_update
;
549 /* capture load from *all* tasks on this cpu: */
550 struct load_weight load
;
551 unsigned long nr_load_updates
;
558 #ifdef CONFIG_FAIR_GROUP_SCHED
559 /* list of leaf cfs_rq on this cpu: */
560 struct list_head leaf_cfs_rq_list
;
561 #endif /* CONFIG_FAIR_GROUP_SCHED */
563 #ifdef CONFIG_RT_GROUP_SCHED
564 struct list_head leaf_rt_rq_list
;
568 * This is part of a global counter where only the total sum
569 * over all CPUs matters. A task can increase this counter on
570 * one CPU and if it got migrated afterwards it may decrease
571 * it on another CPU. Always updated under the runqueue lock:
573 unsigned long nr_uninterruptible
;
575 struct task_struct
*curr
, *idle
, *stop
;
576 unsigned long next_balance
;
577 struct mm_struct
*prev_mm
;
585 struct root_domain
*rd
;
586 struct sched_domain
*sd
;
588 unsigned long cpu_power
;
590 unsigned char idle_balance
;
591 /* For active balancing */
595 struct cpu_stop_work active_balance_work
;
596 /* cpu of this runqueue: */
600 struct list_head cfs_tasks
;
607 /* This is used to determine avg_idle's max value */
608 u64 max_idle_balance_cost
;
611 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
614 #ifdef CONFIG_PARAVIRT
617 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
618 u64 prev_steal_time_rq
;
621 /* calc_load related fields */
622 unsigned long calc_load_update
;
623 long calc_load_active
;
625 #ifdef CONFIG_SCHED_HRTICK
627 int hrtick_csd_pending
;
628 struct call_single_data hrtick_csd
;
630 struct hrtimer hrtick_timer
;
633 #ifdef CONFIG_SCHEDSTATS
635 struct sched_info rq_sched_info
;
636 unsigned long long rq_cpu_time
;
637 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
639 /* sys_sched_yield() stats */
640 unsigned int yld_count
;
642 /* schedule() stats */
643 unsigned int sched_count
;
644 unsigned int sched_goidle
;
646 /* try_to_wake_up() stats */
647 unsigned int ttwu_count
;
648 unsigned int ttwu_local
;
652 struct llist_head wake_list
;
655 struct sched_avg avg
;
658 static inline int cpu_of(struct rq
*rq
)
667 DECLARE_PER_CPU(struct rq
, runqueues
);
669 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
670 #define this_rq() (&__get_cpu_var(runqueues))
671 #define task_rq(p) cpu_rq(task_cpu(p))
672 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
673 #define raw_rq() (&__raw_get_cpu_var(runqueues))
675 static inline u64
rq_clock(struct rq
*rq
)
680 static inline u64
rq_clock_task(struct rq
*rq
)
682 return rq
->clock_task
;
685 #ifdef CONFIG_NUMA_BALANCING
686 extern void sched_setnuma(struct task_struct
*p
, int node
);
687 extern int migrate_task_to(struct task_struct
*p
, int cpu
);
688 extern int migrate_swap(struct task_struct
*, struct task_struct
*);
689 #endif /* CONFIG_NUMA_BALANCING */
693 #define rcu_dereference_check_sched_domain(p) \
694 rcu_dereference_check((p), \
695 lockdep_is_held(&sched_domains_mutex))
698 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
699 * See detach_destroy_domains: synchronize_sched for details.
701 * The domain tree of any CPU may only be accessed from within
702 * preempt-disabled sections.
704 #define for_each_domain(cpu, __sd) \
705 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
706 __sd; __sd = __sd->parent)
708 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
711 * highest_flag_domain - Return highest sched_domain containing flag.
712 * @cpu: The cpu whose highest level of sched domain is to
714 * @flag: The flag to check for the highest sched_domain
717 * Returns the highest sched_domain of a cpu which contains the given flag.
719 static inline struct sched_domain
*highest_flag_domain(int cpu
, int flag
)
721 struct sched_domain
*sd
, *hsd
= NULL
;
723 for_each_domain(cpu
, sd
) {
724 if (!(sd
->flags
& flag
))
732 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
734 struct sched_domain
*sd
;
736 for_each_domain(cpu
, sd
) {
737 if (sd
->flags
& flag
)
744 DECLARE_PER_CPU(struct sched_domain
*, sd_llc
);
745 DECLARE_PER_CPU(int, sd_llc_size
);
746 DECLARE_PER_CPU(int, sd_llc_id
);
747 DECLARE_PER_CPU(struct sched_domain
*, sd_numa
);
748 DECLARE_PER_CPU(struct sched_domain
*, sd_busy
);
749 DECLARE_PER_CPU(struct sched_domain
*, sd_asym
);
751 struct sched_group_power
{
754 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
757 unsigned int power
, power_orig
;
758 unsigned long next_update
;
759 int imbalance
; /* XXX unrelated to power but shared group state */
761 * Number of busy cpus in this group.
763 atomic_t nr_busy_cpus
;
765 unsigned long cpumask
[0]; /* iteration mask */
769 struct sched_group
*next
; /* Must be a circular list */
772 unsigned int group_weight
;
773 struct sched_group_power
*sgp
;
776 * The CPUs this group covers.
778 * NOTE: this field is variable length. (Allocated dynamically
779 * by attaching extra space to the end of the structure,
780 * depending on how many CPUs the kernel has booted up with)
782 unsigned long cpumask
[0];
785 static inline struct cpumask
*sched_group_cpus(struct sched_group
*sg
)
787 return to_cpumask(sg
->cpumask
);
791 * cpumask masking which cpus in the group are allowed to iterate up the domain
794 static inline struct cpumask
*sched_group_mask(struct sched_group
*sg
)
796 return to_cpumask(sg
->sgp
->cpumask
);
800 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
801 * @group: The group whose first cpu is to be returned.
803 static inline unsigned int group_first_cpu(struct sched_group
*group
)
805 return cpumask_first(sched_group_cpus(group
));
808 extern int group_balance_cpu(struct sched_group
*sg
);
810 #endif /* CONFIG_SMP */
813 #include "auto_group.h"
815 #ifdef CONFIG_CGROUP_SCHED
818 * Return the group to which this tasks belongs.
820 * We cannot use task_css() and friends because the cgroup subsystem
821 * changes that value before the cgroup_subsys::attach() method is called,
822 * therefore we cannot pin it and might observe the wrong value.
824 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
825 * core changes this before calling sched_move_task().
827 * Instead we use a 'copy' which is updated from sched_move_task() while
828 * holding both task_struct::pi_lock and rq::lock.
830 static inline struct task_group
*task_group(struct task_struct
*p
)
832 return p
->sched_task_group
;
835 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
836 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
838 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
839 struct task_group
*tg
= task_group(p
);
842 #ifdef CONFIG_FAIR_GROUP_SCHED
843 p
->se
.cfs_rq
= tg
->cfs_rq
[cpu
];
844 p
->se
.parent
= tg
->se
[cpu
];
847 #ifdef CONFIG_RT_GROUP_SCHED
848 p
->rt
.rt_rq
= tg
->rt_rq
[cpu
];
849 p
->rt
.parent
= tg
->rt_se
[cpu
];
853 #else /* CONFIG_CGROUP_SCHED */
855 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
856 static inline struct task_group
*task_group(struct task_struct
*p
)
861 #endif /* CONFIG_CGROUP_SCHED */
863 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
868 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
869 * successfuly executed on another CPU. We must ensure that updates of
870 * per-task data have been completed by this moment.
873 task_thread_info(p
)->cpu
= cpu
;
879 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
881 #ifdef CONFIG_SCHED_DEBUG
882 # include <linux/static_key.h>
883 # define const_debug __read_mostly
885 # define const_debug const
888 extern const_debug
unsigned int sysctl_sched_features
;
890 #define SCHED_FEAT(name, enabled) \
891 __SCHED_FEAT_##name ,
894 #include "features.h"
900 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
901 static __always_inline
bool static_branch__true(struct static_key
*key
)
903 return static_key_true(key
); /* Not out of line branch. */
906 static __always_inline
bool static_branch__false(struct static_key
*key
)
908 return static_key_false(key
); /* Out of line branch. */
911 #define SCHED_FEAT(name, enabled) \
912 static __always_inline bool static_branch_##name(struct static_key *key) \
914 return static_branch__##enabled(key); \
917 #include "features.h"
921 extern struct static_key sched_feat_keys
[__SCHED_FEAT_NR
];
922 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
923 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
924 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
925 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
927 #ifdef CONFIG_NUMA_BALANCING
928 #define sched_feat_numa(x) sched_feat(x)
929 #ifdef CONFIG_SCHED_DEBUG
930 #define numabalancing_enabled sched_feat_numa(NUMA)
932 extern bool numabalancing_enabled
;
933 #endif /* CONFIG_SCHED_DEBUG */
935 #define sched_feat_numa(x) (0)
936 #define numabalancing_enabled (0)
937 #endif /* CONFIG_NUMA_BALANCING */
939 static inline u64
global_rt_period(void)
941 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
944 static inline u64
global_rt_runtime(void)
946 if (sysctl_sched_rt_runtime
< 0)
949 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
952 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
954 return rq
->curr
== p
;
957 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
962 return task_current(rq
, p
);
967 #ifndef prepare_arch_switch
968 # define prepare_arch_switch(next) do { } while (0)
970 #ifndef finish_arch_switch
971 # define finish_arch_switch(prev) do { } while (0)
973 #ifndef finish_arch_post_lock_switch
974 # define finish_arch_post_lock_switch() do { } while (0)
977 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
978 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
982 * We can optimise this out completely for !SMP, because the
983 * SMP rebalancing from interrupt is the only thing that cares
990 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
994 * After ->on_cpu is cleared, the task can be moved to a different CPU.
995 * We must ensure this doesn't happen until the switch is completely
1001 #ifdef CONFIG_DEBUG_SPINLOCK
1002 /* this is a valid case when another task releases the spinlock */
1003 rq
->lock
.owner
= current
;
1006 * If we are tracking spinlock dependencies then we have to
1007 * fix up the runqueue lock - which gets 'carried over' from
1008 * prev into current:
1010 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
1012 raw_spin_unlock_irq(&rq
->lock
);
1015 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1016 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1020 * We can optimise this out completely for !SMP, because the
1021 * SMP rebalancing from interrupt is the only thing that cares
1026 raw_spin_unlock(&rq
->lock
);
1029 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1033 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1034 * We must ensure this doesn't happen until the switch is completely
1042 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1047 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1048 #define WF_FORK 0x02 /* child wakeup after fork */
1049 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1052 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1053 * of tasks with abnormal "nice" values across CPUs the contribution that
1054 * each task makes to its run queue's load is weighted according to its
1055 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1056 * scaled version of the new time slice allocation that they receive on time
1060 #define WEIGHT_IDLEPRIO 3
1061 #define WMULT_IDLEPRIO 1431655765
1064 * Nice levels are multiplicative, with a gentle 10% change for every
1065 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1066 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1067 * that remained on nice 0.
1069 * The "10% effect" is relative and cumulative: from _any_ nice level,
1070 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1071 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1072 * If a task goes up by ~10% and another task goes down by ~10% then
1073 * the relative distance between them is ~25%.)
1075 static const int prio_to_weight
[40] = {
1076 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1077 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1078 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1079 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1080 /* 0 */ 1024, 820, 655, 526, 423,
1081 /* 5 */ 335, 272, 215, 172, 137,
1082 /* 10 */ 110, 87, 70, 56, 45,
1083 /* 15 */ 36, 29, 23, 18, 15,
1087 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1089 * In cases where the weight does not change often, we can use the
1090 * precalculated inverse to speed up arithmetics by turning divisions
1091 * into multiplications:
1093 static const u32 prio_to_wmult
[40] = {
1094 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1095 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1096 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1097 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1098 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1099 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1100 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1101 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1104 #define ENQUEUE_WAKEUP 1
1105 #define ENQUEUE_HEAD 2
1107 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1109 #define ENQUEUE_WAKING 0
1111 #define ENQUEUE_REPLENISH 8
1113 #define DEQUEUE_SLEEP 1
1115 struct sched_class
{
1116 const struct sched_class
*next
;
1118 void (*enqueue_task
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1119 void (*dequeue_task
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1120 void (*yield_task
) (struct rq
*rq
);
1121 bool (*yield_to_task
) (struct rq
*rq
, struct task_struct
*p
, bool preempt
);
1123 void (*check_preempt_curr
) (struct rq
*rq
, struct task_struct
*p
, int flags
);
1125 struct task_struct
* (*pick_next_task
) (struct rq
*rq
);
1126 void (*put_prev_task
) (struct rq
*rq
, struct task_struct
*p
);
1129 int (*select_task_rq
)(struct task_struct
*p
, int task_cpu
, int sd_flag
, int flags
);
1130 void (*migrate_task_rq
)(struct task_struct
*p
, int next_cpu
);
1132 void (*pre_schedule
) (struct rq
*this_rq
, struct task_struct
*task
);
1133 void (*post_schedule
) (struct rq
*this_rq
);
1134 void (*task_waking
) (struct task_struct
*task
);
1135 void (*task_woken
) (struct rq
*this_rq
, struct task_struct
*task
);
1137 void (*set_cpus_allowed
)(struct task_struct
*p
,
1138 const struct cpumask
*newmask
);
1140 void (*rq_online
)(struct rq
*rq
);
1141 void (*rq_offline
)(struct rq
*rq
);
1144 void (*set_curr_task
) (struct rq
*rq
);
1145 void (*task_tick
) (struct rq
*rq
, struct task_struct
*p
, int queued
);
1146 void (*task_fork
) (struct task_struct
*p
);
1147 void (*task_dead
) (struct task_struct
*p
);
1149 void (*switched_from
) (struct rq
*this_rq
, struct task_struct
*task
);
1150 void (*switched_to
) (struct rq
*this_rq
, struct task_struct
*task
);
1151 void (*prio_changed
) (struct rq
*this_rq
, struct task_struct
*task
,
1154 unsigned int (*get_rr_interval
) (struct rq
*rq
,
1155 struct task_struct
*task
);
1157 #ifdef CONFIG_FAIR_GROUP_SCHED
1158 void (*task_move_group
) (struct task_struct
*p
, int on_rq
);
1162 #define sched_class_highest (&stop_sched_class)
1163 #define for_each_class(class) \
1164 for (class = sched_class_highest; class; class = class->next)
1166 extern const struct sched_class stop_sched_class
;
1167 extern const struct sched_class dl_sched_class
;
1168 extern const struct sched_class rt_sched_class
;
1169 extern const struct sched_class fair_sched_class
;
1170 extern const struct sched_class idle_sched_class
;
1175 extern void update_group_power(struct sched_domain
*sd
, int cpu
);
1177 extern void trigger_load_balance(struct rq
*rq
);
1178 extern void idle_balance(int this_cpu
, struct rq
*this_rq
);
1180 extern void idle_enter_fair(struct rq
*this_rq
);
1181 extern void idle_exit_fair(struct rq
*this_rq
);
1183 #else /* CONFIG_SMP */
1185 static inline void idle_balance(int cpu
, struct rq
*rq
)
1191 extern void sysrq_sched_debug_show(void);
1192 extern void sched_init_granularity(void);
1193 extern void update_max_interval(void);
1195 extern void init_sched_dl_class(void);
1196 extern void init_sched_rt_class(void);
1197 extern void init_sched_fair_class(void);
1198 extern void init_sched_dl_class(void);
1200 extern void resched_task(struct task_struct
*p
);
1201 extern void resched_cpu(int cpu
);
1203 extern struct rt_bandwidth def_rt_bandwidth
;
1204 extern void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
);
1206 extern struct dl_bandwidth def_dl_bandwidth
;
1207 extern void init_dl_bandwidth(struct dl_bandwidth
*dl_b
, u64 period
, u64 runtime
);
1208 extern void init_dl_task_timer(struct sched_dl_entity
*dl_se
);
1210 unsigned long to_ratio(u64 period
, u64 runtime
);
1212 extern void update_idle_cpu_load(struct rq
*this_rq
);
1214 extern void init_task_runnable_average(struct task_struct
*p
);
1216 #ifdef CONFIG_PARAVIRT
1217 static inline u64
steal_ticks(u64 steal
)
1219 if (unlikely(steal
> NSEC_PER_SEC
))
1220 return div_u64(steal
, TICK_NSEC
);
1222 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
1226 static inline void inc_nr_running(struct rq
*rq
)
1230 #ifdef CONFIG_NO_HZ_FULL
1231 if (rq
->nr_running
== 2) {
1232 if (tick_nohz_full_cpu(rq
->cpu
)) {
1233 /* Order rq->nr_running write against the IPI */
1235 smp_send_reschedule(rq
->cpu
);
1241 static inline void dec_nr_running(struct rq
*rq
)
1246 static inline void rq_last_tick_reset(struct rq
*rq
)
1248 #ifdef CONFIG_NO_HZ_FULL
1249 rq
->last_sched_tick
= jiffies
;
1253 extern void update_rq_clock(struct rq
*rq
);
1255 extern void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
);
1256 extern void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
);
1258 extern void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
1260 extern const_debug
unsigned int sysctl_sched_time_avg
;
1261 extern const_debug
unsigned int sysctl_sched_nr_migrate
;
1262 extern const_debug
unsigned int sysctl_sched_migration_cost
;
1264 static inline u64
sched_avg_period(void)
1266 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1269 #ifdef CONFIG_SCHED_HRTICK
1273 * - enabled by features
1274 * - hrtimer is actually high res
1276 static inline int hrtick_enabled(struct rq
*rq
)
1278 if (!sched_feat(HRTICK
))
1280 if (!cpu_active(cpu_of(rq
)))
1282 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1285 void hrtick_start(struct rq
*rq
, u64 delay
);
1289 static inline int hrtick_enabled(struct rq
*rq
)
1294 #endif /* CONFIG_SCHED_HRTICK */
1297 extern void sched_avg_update(struct rq
*rq
);
1298 static inline void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1300 rq
->rt_avg
+= rt_delta
;
1301 sched_avg_update(rq
);
1304 static inline void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
) { }
1305 static inline void sched_avg_update(struct rq
*rq
) { }
1308 extern void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
);
1311 #ifdef CONFIG_PREEMPT
1313 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1316 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1317 * way at the expense of forcing extra atomic operations in all
1318 * invocations. This assures that the double_lock is acquired using the
1319 * same underlying policy as the spinlock_t on this architecture, which
1320 * reduces latency compared to the unfair variant below. However, it
1321 * also adds more overhead and therefore may reduce throughput.
1323 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1324 __releases(this_rq
->lock
)
1325 __acquires(busiest
->lock
)
1326 __acquires(this_rq
->lock
)
1328 raw_spin_unlock(&this_rq
->lock
);
1329 double_rq_lock(this_rq
, busiest
);
1336 * Unfair double_lock_balance: Optimizes throughput at the expense of
1337 * latency by eliminating extra atomic operations when the locks are
1338 * already in proper order on entry. This favors lower cpu-ids and will
1339 * grant the double lock to lower cpus over higher ids under contention,
1340 * regardless of entry order into the function.
1342 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1343 __releases(this_rq
->lock
)
1344 __acquires(busiest
->lock
)
1345 __acquires(this_rq
->lock
)
1349 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1350 if (busiest
< this_rq
) {
1351 raw_spin_unlock(&this_rq
->lock
);
1352 raw_spin_lock(&busiest
->lock
);
1353 raw_spin_lock_nested(&this_rq
->lock
,
1354 SINGLE_DEPTH_NESTING
);
1357 raw_spin_lock_nested(&busiest
->lock
,
1358 SINGLE_DEPTH_NESTING
);
1363 #endif /* CONFIG_PREEMPT */
1366 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1368 static inline int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1370 if (unlikely(!irqs_disabled())) {
1371 /* printk() doesn't work good under rq->lock */
1372 raw_spin_unlock(&this_rq
->lock
);
1376 return _double_lock_balance(this_rq
, busiest
);
1379 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1380 __releases(busiest
->lock
)
1382 raw_spin_unlock(&busiest
->lock
);
1383 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1386 static inline void double_lock(spinlock_t
*l1
, spinlock_t
*l2
)
1392 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1395 static inline void double_raw_lock(raw_spinlock_t
*l1
, raw_spinlock_t
*l2
)
1401 raw_spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1405 * double_rq_lock - safely lock two runqueues
1407 * Note this does not disable interrupts like task_rq_lock,
1408 * you need to do so manually before calling.
1410 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1411 __acquires(rq1
->lock
)
1412 __acquires(rq2
->lock
)
1414 BUG_ON(!irqs_disabled());
1416 raw_spin_lock(&rq1
->lock
);
1417 __acquire(rq2
->lock
); /* Fake it out ;) */
1420 raw_spin_lock(&rq1
->lock
);
1421 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1423 raw_spin_lock(&rq2
->lock
);
1424 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1430 * double_rq_unlock - safely unlock two runqueues
1432 * Note this does not restore interrupts like task_rq_unlock,
1433 * you need to do so manually after calling.
1435 static inline void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1436 __releases(rq1
->lock
)
1437 __releases(rq2
->lock
)
1439 raw_spin_unlock(&rq1
->lock
);
1441 raw_spin_unlock(&rq2
->lock
);
1443 __release(rq2
->lock
);
1446 #else /* CONFIG_SMP */
1449 * double_rq_lock - safely lock two runqueues
1451 * Note this does not disable interrupts like task_rq_lock,
1452 * you need to do so manually before calling.
1454 static inline void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1455 __acquires(rq1
->lock
)
1456 __acquires(rq2
->lock
)
1458 BUG_ON(!irqs_disabled());
1460 raw_spin_lock(&rq1
->lock
);
1461 __acquire(rq2
->lock
); /* Fake it out ;) */
1465 * double_rq_unlock - safely unlock two runqueues
1467 * Note this does not restore interrupts like task_rq_unlock,
1468 * you need to do so manually after calling.
1470 static inline void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1471 __releases(rq1
->lock
)
1472 __releases(rq2
->lock
)
1475 raw_spin_unlock(&rq1
->lock
);
1476 __release(rq2
->lock
);
1481 extern struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
);
1482 extern struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
);
1483 extern void print_cfs_stats(struct seq_file
*m
, int cpu
);
1484 extern void print_rt_stats(struct seq_file
*m
, int cpu
);
1486 extern void init_cfs_rq(struct cfs_rq
*cfs_rq
);
1487 extern void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
);
1488 extern void init_dl_rq(struct dl_rq
*dl_rq
, struct rq
*rq
);
1490 extern void cfs_bandwidth_usage_inc(void);
1491 extern void cfs_bandwidth_usage_dec(void);
1493 #ifdef CONFIG_NO_HZ_COMMON
1494 enum rq_nohz_flag_bits
{
1499 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1502 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1504 DECLARE_PER_CPU(u64
, cpu_hardirq_time
);
1505 DECLARE_PER_CPU(u64
, cpu_softirq_time
);
1507 #ifndef CONFIG_64BIT
1508 DECLARE_PER_CPU(seqcount_t
, irq_time_seq
);
1510 static inline void irq_time_write_begin(void)
1512 __this_cpu_inc(irq_time_seq
.sequence
);
1516 static inline void irq_time_write_end(void)
1519 __this_cpu_inc(irq_time_seq
.sequence
);
1522 static inline u64
irq_time_read(int cpu
)
1528 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1529 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1530 per_cpu(cpu_hardirq_time
, cpu
);
1531 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1535 #else /* CONFIG_64BIT */
1536 static inline void irq_time_write_begin(void)
1540 static inline void irq_time_write_end(void)
1544 static inline u64
irq_time_read(int cpu
)
1546 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1548 #endif /* CONFIG_64BIT */
1549 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */