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/irq_work.h>
10 #include <linux/tick.h>
11 #include <linux/slab.h>
14 #include "cpudeadline.h"
20 /* task_struct::on_rq states: */
21 #define TASK_ON_RQ_QUEUED 1
22 #define TASK_ON_RQ_MIGRATING 2
24 extern __read_mostly int scheduler_running;
26 extern unsigned long calc_load_update;
27 extern atomic_long_t calc_load_tasks;
29 extern void calc_global_load_tick(struct rq *this_rq);
30 extern long calc_load_fold_active(struct rq *this_rq);
33 extern void update_cpu_load_active(struct rq *this_rq);
35 static inline void update_cpu_load_active(struct rq *this_rq) { }
39 * Helpers for converting nanosecond timing to jiffy resolution
41 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
44 * Increase resolution of nice-level calculations for 64-bit architectures.
45 * The extra resolution improves shares distribution and load balancing of
46 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
47 * hierarchies, especially on larger systems. This is not a user-visible change
48 * and does not change the user-interface for setting shares/weights.
50 * We increase resolution only if we have enough bits to allow this increased
51 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
52 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
55 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */
56 # define SCHED_LOAD_RESOLUTION 10
57 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION)
58 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION)
60 # define SCHED_LOAD_RESOLUTION 0
61 # define scale_load(w) (w)
62 # define scale_load_down(w) (w)
65 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION)
66 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT)
68 #define NICE_0_LOAD SCHED_LOAD_SCALE
69 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
72 * Single value that decides SCHED_DEADLINE internal math precision.
73 * 10 -> just above 1us
74 * 9 -> just above 0.5us
79 * These are the 'tuning knobs' of the scheduler:
83 * single value that denotes runtime == period, ie unlimited time.
85 #define RUNTIME_INF ((u64)~0ULL)
87 static inline int fair_policy(int policy)
89 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
92 static inline int rt_policy(int policy)
94 return policy == SCHED_FIFO || policy == SCHED_RR;
97 static inline int dl_policy(int policy)
99 return policy == SCHED_DEADLINE;
102 static inline int task_has_rt_policy(struct task_struct *p)
104 return rt_policy(p->policy);
107 static inline int task_has_dl_policy(struct task_struct *p)
109 return dl_policy(p->policy);
112 static inline bool dl_time_before(u64 a, u64 b)
114 return (s64)(a - b) < 0;
118 * Tells if entity @a should preempt entity @b.
121 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
123 return dl_time_before(a->deadline, b->deadline);
127 * This is the priority-queue data structure of the RT scheduling class:
129 struct rt_prio_array {
130 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
131 struct list_head queue[MAX_RT_PRIO];
134 struct rt_bandwidth {
135 /* nests inside the rq lock: */
136 raw_spinlock_t rt_runtime_lock;
139 struct hrtimer rt_period_timer;
142 void __dl_clear_params(struct task_struct *p);
145 * To keep the bandwidth of -deadline tasks and groups under control
146 * we need some place where:
147 * - store the maximum -deadline bandwidth of the system (the group);
148 * - cache the fraction of that bandwidth that is currently allocated.
150 * This is all done in the data structure below. It is similar to the
151 * one used for RT-throttling (rt_bandwidth), with the main difference
152 * that, since here we are only interested in admission control, we
153 * do not decrease any runtime while the group "executes", neither we
154 * need a timer to replenish it.
156 * With respect to SMP, the bandwidth is given on a per-CPU basis,
158 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
159 * - dl_total_bw array contains, in the i-eth element, the currently
160 * allocated bandwidth on the i-eth CPU.
161 * Moreover, groups consume bandwidth on each CPU, while tasks only
162 * consume bandwidth on the CPU they're running on.
163 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
164 * that will be shown the next time the proc or cgroup controls will
165 * be red. It on its turn can be changed by writing on its own
168 struct dl_bandwidth {
169 raw_spinlock_t dl_runtime_lock;
174 static inline int dl_bandwidth_enabled(void)
176 return sysctl_sched_rt_runtime >= 0;
179 extern struct dl_bw *dl_bw_of(int i);
187 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
189 dl_b->total_bw -= tsk_bw;
193 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
195 dl_b->total_bw += tsk_bw;
199 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
201 return dl_b->bw != -1 &&
202 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
205 extern struct mutex sched_domains_mutex;
207 #ifdef CONFIG_CGROUP_SCHED
209 #include <linux/cgroup.h>
214 extern struct list_head task_groups;
216 struct cfs_bandwidth {
217 #ifdef CONFIG_CFS_BANDWIDTH
221 s64 hierarchical_quota;
224 int idle, timer_active;
225 struct hrtimer period_timer, slack_timer;
226 struct list_head throttled_cfs_rq;
229 int nr_periods, nr_throttled;
234 /* task group related information */
236 struct cgroup_subsys_state css;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 /* schedulable entities of this group on each cpu */
240 struct sched_entity **se;
241 /* runqueue "owned" by this group on each cpu */
242 struct cfs_rq **cfs_rq;
243 unsigned long shares;
246 atomic_long_t load_avg;
247 atomic_t runnable_avg;
251 #ifdef CONFIG_RT_GROUP_SCHED
252 struct sched_rt_entity **rt_se;
253 struct rt_rq **rt_rq;
255 struct rt_bandwidth rt_bandwidth;
259 struct list_head list;
261 struct task_group *parent;
262 struct list_head siblings;
263 struct list_head children;
265 #ifdef CONFIG_SCHED_AUTOGROUP
266 struct autogroup *autogroup;
269 struct cfs_bandwidth cfs_bandwidth;
272 #ifdef CONFIG_FAIR_GROUP_SCHED
273 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
276 * A weight of 0 or 1 can cause arithmetics problems.
277 * A weight of a cfs_rq is the sum of weights of which entities
278 * are queued on this cfs_rq, so a weight of a entity should not be
279 * too large, so as the shares value of a task group.
280 * (The default weight is 1024 - so there's no practical
281 * limitation from this.)
283 #define MIN_SHARES (1UL << 1)
284 #define MAX_SHARES (1UL << 18)
287 typedef int (*tg_visitor)(struct task_group *, void *);
289 extern int walk_tg_tree_from(struct task_group *from,
290 tg_visitor down, tg_visitor up, void *data);
293 * Iterate the full tree, calling @down when first entering a node and @up when
294 * leaving it for the final time.
296 * Caller must hold rcu_lock or sufficient equivalent.
298 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
300 return walk_tg_tree_from(&root_task_group, down, up, data);
303 extern int tg_nop(struct task_group *tg, void *data);
305 extern void free_fair_sched_group(struct task_group *tg);
306 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
307 extern void unregister_fair_sched_group(struct task_group *tg, int cpu);
308 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
309 struct sched_entity *se, int cpu,
310 struct sched_entity *parent);
311 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
312 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
314 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
315 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force);
316 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
318 extern void free_rt_sched_group(struct task_group *tg);
319 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
320 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
321 struct sched_rt_entity *rt_se, int cpu,
322 struct sched_rt_entity *parent);
324 extern struct task_group *sched_create_group(struct task_group *parent);
325 extern void sched_online_group(struct task_group *tg,
326 struct task_group *parent);
327 extern void sched_destroy_group(struct task_group *tg);
328 extern void sched_offline_group(struct task_group *tg);
330 extern void sched_move_task(struct task_struct *tsk);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
336 #else /* CONFIG_CGROUP_SCHED */
338 struct cfs_bandwidth { };
340 #endif /* CONFIG_CGROUP_SCHED */
342 /* CFS-related fields in a runqueue */
344 struct load_weight load;
345 unsigned int nr_running, h_nr_running;
350 u64 min_vruntime_copy;
353 struct rb_root tasks_timeline;
354 struct rb_node *rb_leftmost;
357 * 'curr' points to currently running entity on this cfs_rq.
358 * It is set to NULL otherwise (i.e when none are currently running).
360 struct sched_entity *curr, *next, *last, *skip;
362 #ifdef CONFIG_SCHED_DEBUG
363 unsigned int nr_spread_over;
369 * Under CFS, load is tracked on a per-entity basis and aggregated up.
370 * This allows for the description of both thread and group usage (in
371 * the FAIR_GROUP_SCHED case).
372 * runnable_load_avg is the sum of the load_avg_contrib of the
373 * sched_entities on the rq.
374 * blocked_load_avg is similar to runnable_load_avg except that its
375 * the blocked sched_entities on the rq.
376 * utilization_load_avg is the sum of the average running time of the
377 * sched_entities on the rq.
379 unsigned long runnable_load_avg, blocked_load_avg, utilization_load_avg;
380 atomic64_t decay_counter;
382 atomic_long_t removed_load;
384 #ifdef CONFIG_FAIR_GROUP_SCHED
385 /* Required to track per-cpu representation of a task_group */
386 u32 tg_runnable_contrib;
387 unsigned long tg_load_contrib;
390 * h_load = weight * f(tg)
392 * Where f(tg) is the recursive weight fraction assigned to
395 unsigned long h_load;
396 u64 last_h_load_update;
397 struct sched_entity *h_load_next;
398 #endif /* CONFIG_FAIR_GROUP_SCHED */
399 #endif /* CONFIG_SMP */
401 #ifdef CONFIG_FAIR_GROUP_SCHED
402 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
405 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
406 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
407 * (like users, containers etc.)
409 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
410 * list is used during load balance.
413 struct list_head leaf_cfs_rq_list;
414 struct task_group *tg; /* group that "owns" this runqueue */
416 #ifdef CONFIG_CFS_BANDWIDTH
419 s64 runtime_remaining;
421 u64 throttled_clock, throttled_clock_task;
422 u64 throttled_clock_task_time;
423 int throttled, throttle_count;
424 struct list_head throttled_list;
425 #endif /* CONFIG_CFS_BANDWIDTH */
426 #endif /* CONFIG_FAIR_GROUP_SCHED */
429 static inline int rt_bandwidth_enabled(void)
431 return sysctl_sched_rt_runtime >= 0;
434 /* RT IPI pull logic requires IRQ_WORK */
435 #ifdef CONFIG_IRQ_WORK
436 # define HAVE_RT_PUSH_IPI
439 /* Real-Time classes' related field in a runqueue: */
441 struct rt_prio_array active;
442 unsigned int rt_nr_running;
443 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
445 int curr; /* highest queued rt task prio */
447 int next; /* next highest */
452 unsigned long rt_nr_migratory;
453 unsigned long rt_nr_total;
455 struct plist_head pushable_tasks;
456 #ifdef HAVE_RT_PUSH_IPI
459 struct irq_work push_work;
460 raw_spinlock_t push_lock;
462 #endif /* CONFIG_SMP */
468 /* Nests inside the rq lock: */
469 raw_spinlock_t rt_runtime_lock;
471 #ifdef CONFIG_RT_GROUP_SCHED
472 unsigned long rt_nr_boosted;
475 struct task_group *tg;
479 /* Deadline class' related fields in a runqueue */
481 /* runqueue is an rbtree, ordered by deadline */
482 struct rb_root rb_root;
483 struct rb_node *rb_leftmost;
485 unsigned long dl_nr_running;
489 * Deadline values of the currently executing and the
490 * earliest ready task on this rq. Caching these facilitates
491 * the decision wether or not a ready but not running task
492 * should migrate somewhere else.
499 unsigned long dl_nr_migratory;
503 * Tasks on this rq that can be pushed away. They are kept in
504 * an rb-tree, ordered by tasks' deadlines, with caching
505 * of the leftmost (earliest deadline) element.
507 struct rb_root pushable_dl_tasks_root;
508 struct rb_node *pushable_dl_tasks_leftmost;
517 * We add the notion of a root-domain which will be used to define per-domain
518 * variables. Each exclusive cpuset essentially defines an island domain by
519 * fully partitioning the member cpus from any other cpuset. Whenever a new
520 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
531 /* Indicate more than one runnable task for any CPU */
535 * The bit corresponding to a CPU gets set here if such CPU has more
536 * than one runnable -deadline task (as it is below for RT tasks).
538 cpumask_var_t dlo_mask;
544 * The "RT overload" flag: it gets set if a CPU has more than
545 * one runnable RT task.
547 cpumask_var_t rto_mask;
548 struct cpupri cpupri;
551 extern struct root_domain def_root_domain;
553 #endif /* CONFIG_SMP */
556 * This is the main, per-CPU runqueue data structure.
558 * Locking rule: those places that want to lock multiple runqueues
559 * (such as the load balancing or the thread migration code), lock
560 * acquire operations must be ordered by ascending &runqueue.
567 * nr_running and cpu_load should be in the same cacheline because
568 * remote CPUs use both these fields when doing load calculation.
570 unsigned int nr_running;
571 #ifdef CONFIG_NUMA_BALANCING
572 unsigned int nr_numa_running;
573 unsigned int nr_preferred_running;
575 #define CPU_LOAD_IDX_MAX 5
576 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_load_update_tick;
578 #ifdef CONFIG_NO_HZ_COMMON
580 unsigned long nohz_flags;
582 #ifdef CONFIG_NO_HZ_FULL
583 unsigned long last_sched_tick;
585 /* capture load from *all* tasks on this cpu: */
586 struct load_weight load;
587 unsigned long nr_load_updates;
594 #ifdef CONFIG_FAIR_GROUP_SCHED
595 /* list of leaf cfs_rq on this cpu: */
596 struct list_head leaf_cfs_rq_list;
598 struct sched_avg avg;
599 #endif /* CONFIG_FAIR_GROUP_SCHED */
602 * This is part of a global counter where only the total sum
603 * over all CPUs matters. A task can increase this counter on
604 * one CPU and if it got migrated afterwards it may decrease
605 * it on another CPU. Always updated under the runqueue lock:
607 unsigned long nr_uninterruptible;
609 struct task_struct *curr, *idle, *stop;
610 unsigned long next_balance;
611 struct mm_struct *prev_mm;
613 unsigned int clock_skip_update;
620 struct root_domain *rd;
621 struct sched_domain *sd;
623 unsigned long cpu_capacity;
624 unsigned long cpu_capacity_orig;
626 unsigned char idle_balance;
627 /* For active balancing */
631 struct cpu_stop_work active_balance_work;
632 /* cpu of this runqueue: */
636 struct list_head cfs_tasks;
643 /* This is used to determine avg_idle's max value */
644 u64 max_idle_balance_cost;
647 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
650 #ifdef CONFIG_PARAVIRT
653 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
654 u64 prev_steal_time_rq;
657 /* calc_load related fields */
658 unsigned long calc_load_update;
659 long calc_load_active;
661 #ifdef CONFIG_SCHED_HRTICK
663 int hrtick_csd_pending;
664 struct call_single_data hrtick_csd;
666 struct hrtimer hrtick_timer;
669 #ifdef CONFIG_SCHEDSTATS
671 struct sched_info rq_sched_info;
672 unsigned long long rq_cpu_time;
673 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
675 /* sys_sched_yield() stats */
676 unsigned int yld_count;
678 /* schedule() stats */
679 unsigned int sched_count;
680 unsigned int sched_goidle;
682 /* try_to_wake_up() stats */
683 unsigned int ttwu_count;
684 unsigned int ttwu_local;
688 struct llist_head wake_list;
691 #ifdef CONFIG_CPU_IDLE
692 /* Must be inspected within a rcu lock section */
693 struct cpuidle_state *idle_state;
697 static inline int cpu_of(struct rq *rq)
706 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
708 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
709 #define this_rq() this_cpu_ptr(&runqueues)
710 #define task_rq(p) cpu_rq(task_cpu(p))
711 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
712 #define raw_rq() raw_cpu_ptr(&runqueues)
714 static inline u64 __rq_clock_broken(struct rq *rq)
716 return ACCESS_ONCE(rq->clock);
719 static inline u64 rq_clock(struct rq *rq)
721 lockdep_assert_held(&rq->lock);
725 static inline u64 rq_clock_task(struct rq *rq)
727 lockdep_assert_held(&rq->lock);
728 return rq->clock_task;
731 #define RQCF_REQ_SKIP 0x01
732 #define RQCF_ACT_SKIP 0x02
734 static inline void rq_clock_skip_update(struct rq *rq, bool skip)
736 lockdep_assert_held(&rq->lock);
738 rq->clock_skip_update |= RQCF_REQ_SKIP;
740 rq->clock_skip_update &= ~RQCF_REQ_SKIP;
744 enum numa_topology_type {
749 extern enum numa_topology_type sched_numa_topology_type;
750 extern int sched_max_numa_distance;
751 extern bool find_numa_distance(int distance);
754 #ifdef CONFIG_NUMA_BALANCING
755 /* The regions in numa_faults array from task_struct */
756 enum numa_faults_stats {
762 extern void sched_setnuma(struct task_struct *p, int node);
763 extern int migrate_task_to(struct task_struct *p, int cpu);
764 extern int migrate_swap(struct task_struct *, struct task_struct *);
765 #endif /* CONFIG_NUMA_BALANCING */
769 extern void sched_ttwu_pending(void);
771 #define rcu_dereference_check_sched_domain(p) \
772 rcu_dereference_check((p), \
773 lockdep_is_held(&sched_domains_mutex))
776 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
777 * See detach_destroy_domains: synchronize_sched for details.
779 * The domain tree of any CPU may only be accessed from within
780 * preempt-disabled sections.
782 #define for_each_domain(cpu, __sd) \
783 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
784 __sd; __sd = __sd->parent)
786 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
789 * highest_flag_domain - Return highest sched_domain containing flag.
790 * @cpu: The cpu whose highest level of sched domain is to
792 * @flag: The flag to check for the highest sched_domain
795 * Returns the highest sched_domain of a cpu which contains the given flag.
797 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
799 struct sched_domain *sd, *hsd = NULL;
801 for_each_domain(cpu, sd) {
802 if (!(sd->flags & flag))
810 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
812 struct sched_domain *sd;
814 for_each_domain(cpu, sd) {
815 if (sd->flags & flag)
822 DECLARE_PER_CPU(struct sched_domain *, sd_llc);
823 DECLARE_PER_CPU(int, sd_llc_size);
824 DECLARE_PER_CPU(int, sd_llc_id);
825 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
826 DECLARE_PER_CPU(struct sched_domain *, sd_busy);
827 DECLARE_PER_CPU(struct sched_domain *, sd_asym);
829 struct sched_group_capacity {
832 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity
835 unsigned int capacity;
836 unsigned long next_update;
837 int imbalance; /* XXX unrelated to capacity but shared group state */
839 * Number of busy cpus in this group.
841 atomic_t nr_busy_cpus;
843 unsigned long cpumask[0]; /* iteration mask */
847 struct sched_group *next; /* Must be a circular list */
850 unsigned int group_weight;
851 struct sched_group_capacity *sgc;
854 * The CPUs this group covers.
856 * NOTE: this field is variable length. (Allocated dynamically
857 * by attaching extra space to the end of the structure,
858 * depending on how many CPUs the kernel has booted up with)
860 unsigned long cpumask[0];
863 static inline struct cpumask *sched_group_cpus(struct sched_group *sg)
865 return to_cpumask(sg->cpumask);
869 * cpumask masking which cpus in the group are allowed to iterate up the domain
872 static inline struct cpumask *sched_group_mask(struct sched_group *sg)
874 return to_cpumask(sg->sgc->cpumask);
878 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
879 * @group: The group whose first cpu is to be returned.
881 static inline unsigned int group_first_cpu(struct sched_group *group)
883 return cpumask_first(sched_group_cpus(group));
886 extern int group_balance_cpu(struct sched_group *sg);
890 static inline void sched_ttwu_pending(void) { }
892 #endif /* CONFIG_SMP */
895 #include "auto_group.h"
897 #ifdef CONFIG_CGROUP_SCHED
900 * Return the group to which this tasks belongs.
902 * We cannot use task_css() and friends because the cgroup subsystem
903 * changes that value before the cgroup_subsys::attach() method is called,
904 * therefore we cannot pin it and might observe the wrong value.
906 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
907 * core changes this before calling sched_move_task().
909 * Instead we use a 'copy' which is updated from sched_move_task() while
910 * holding both task_struct::pi_lock and rq::lock.
912 static inline struct task_group *task_group(struct task_struct *p)
914 return p->sched_task_group;
917 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
918 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
920 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
921 struct task_group *tg = task_group(p);
924 #ifdef CONFIG_FAIR_GROUP_SCHED
925 p->se.cfs_rq = tg->cfs_rq[cpu];
926 p->se.parent = tg->se[cpu];
929 #ifdef CONFIG_RT_GROUP_SCHED
930 p->rt.rt_rq = tg->rt_rq[cpu];
931 p->rt.parent = tg->rt_se[cpu];
935 #else /* CONFIG_CGROUP_SCHED */
937 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
938 static inline struct task_group *task_group(struct task_struct *p)
943 #endif /* CONFIG_CGROUP_SCHED */
945 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
950 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
951 * successfuly executed on another CPU. We must ensure that updates of
952 * per-task data have been completed by this moment.
955 task_thread_info(p)->cpu = cpu;
961 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
963 #ifdef CONFIG_SCHED_DEBUG
964 # include <linux/static_key.h>
965 # define const_debug __read_mostly
967 # define const_debug const
970 extern const_debug unsigned int sysctl_sched_features;
972 #define SCHED_FEAT(name, enabled) \
973 __SCHED_FEAT_##name ,
976 #include "features.h"
982 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
983 #define SCHED_FEAT(name, enabled) \
984 static __always_inline bool static_branch_##name(struct static_key *key) \
986 return static_key_##enabled(key); \
989 #include "features.h"
993 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
994 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
995 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
996 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
997 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
999 #ifdef CONFIG_NUMA_BALANCING
1000 #define sched_feat_numa(x) sched_feat(x)
1001 #ifdef CONFIG_SCHED_DEBUG
1002 #define numabalancing_enabled sched_feat_numa(NUMA)
1004 extern bool numabalancing_enabled;
1005 #endif /* CONFIG_SCHED_DEBUG */
1007 #define sched_feat_numa(x) (0)
1008 #define numabalancing_enabled (0)
1009 #endif /* CONFIG_NUMA_BALANCING */
1011 static inline u64 global_rt_period(void)
1013 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
1016 static inline u64 global_rt_runtime(void)
1018 if (sysctl_sched_rt_runtime < 0)
1021 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
1024 static inline int task_current(struct rq *rq, struct task_struct *p)
1026 return rq->curr == p;
1029 static inline int task_running(struct rq *rq, struct task_struct *p)
1034 return task_current(rq, p);
1038 static inline int task_on_rq_queued(struct task_struct *p)
1040 return p->on_rq == TASK_ON_RQ_QUEUED;
1043 static inline int task_on_rq_migrating(struct task_struct *p)
1045 return p->on_rq == TASK_ON_RQ_MIGRATING;
1048 #ifndef prepare_arch_switch
1049 # define prepare_arch_switch(next) do { } while (0)
1051 #ifndef finish_arch_switch
1052 # define finish_arch_switch(prev) do { } while (0)
1054 #ifndef finish_arch_post_lock_switch
1055 # define finish_arch_post_lock_switch() do { } while (0)
1058 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1062 * We can optimise this out completely for !SMP, because the
1063 * SMP rebalancing from interrupt is the only thing that cares
1070 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1074 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1075 * We must ensure this doesn't happen until the switch is completely
1081 #ifdef CONFIG_DEBUG_SPINLOCK
1082 /* this is a valid case when another task releases the spinlock */
1083 rq->lock.owner = current;
1086 * If we are tracking spinlock dependencies then we have to
1087 * fix up the runqueue lock - which gets 'carried over' from
1088 * prev into current:
1090 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1092 raw_spin_unlock_irq(&rq->lock);
1098 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */
1099 #define WF_FORK 0x02 /* child wakeup after fork */
1100 #define WF_MIGRATED 0x4 /* internal use, task got migrated */
1103 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1104 * of tasks with abnormal "nice" values across CPUs the contribution that
1105 * each task makes to its run queue's load is weighted according to its
1106 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1107 * scaled version of the new time slice allocation that they receive on time
1111 #define WEIGHT_IDLEPRIO 3
1112 #define WMULT_IDLEPRIO 1431655765
1115 * Nice levels are multiplicative, with a gentle 10% change for every
1116 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1117 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1118 * that remained on nice 0.
1120 * The "10% effect" is relative and cumulative: from _any_ nice level,
1121 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1122 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1123 * If a task goes up by ~10% and another task goes down by ~10% then
1124 * the relative distance between them is ~25%.)
1126 static const int prio_to_weight[40] = {
1127 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1128 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1129 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1130 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1131 /* 0 */ 1024, 820, 655, 526, 423,
1132 /* 5 */ 335, 272, 215, 172, 137,
1133 /* 10 */ 110, 87, 70, 56, 45,
1134 /* 15 */ 36, 29, 23, 18, 15,
1138 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1140 * In cases where the weight does not change often, we can use the
1141 * precalculated inverse to speed up arithmetics by turning divisions
1142 * into multiplications:
1144 static const u32 prio_to_wmult[40] = {
1145 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1146 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1147 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1148 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1149 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1150 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1151 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1152 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1155 #define ENQUEUE_WAKEUP 1
1156 #define ENQUEUE_HEAD 2
1158 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */
1160 #define ENQUEUE_WAKING 0
1162 #define ENQUEUE_REPLENISH 8
1164 #define DEQUEUE_SLEEP 1
1166 #define RETRY_TASK ((void *)-1UL)
1168 struct sched_class {
1169 const struct sched_class *next;
1171 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
1172 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
1173 void (*yield_task) (struct rq *rq);
1174 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
1176 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
1179 * It is the responsibility of the pick_next_task() method that will
1180 * return the next task to call put_prev_task() on the @prev task or
1181 * something equivalent.
1183 * May return RETRY_TASK when it finds a higher prio class has runnable
1186 struct task_struct * (*pick_next_task) (struct rq *rq,
1187 struct task_struct *prev);
1188 void (*put_prev_task) (struct rq *rq, struct task_struct *p);
1191 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
1192 void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
1194 void (*post_schedule) (struct rq *this_rq);
1195 void (*task_waking) (struct task_struct *task);
1196 void (*task_woken) (struct rq *this_rq, struct task_struct *task);
1198 void (*set_cpus_allowed)(struct task_struct *p,
1199 const struct cpumask *newmask);
1201 void (*rq_online)(struct rq *rq);
1202 void (*rq_offline)(struct rq *rq);
1205 void (*set_curr_task) (struct rq *rq);
1206 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
1207 void (*task_fork) (struct task_struct *p);
1208 void (*task_dead) (struct task_struct *p);
1211 * The switched_from() call is allowed to drop rq->lock, therefore we
1212 * cannot assume the switched_from/switched_to pair is serliazed by
1213 * rq->lock. They are however serialized by p->pi_lock.
1215 void (*switched_from) (struct rq *this_rq, struct task_struct *task);
1216 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
1217 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
1220 unsigned int (*get_rr_interval) (struct rq *rq,
1221 struct task_struct *task);
1223 void (*update_curr) (struct rq *rq);
1225 #ifdef CONFIG_FAIR_GROUP_SCHED
1226 void (*task_move_group) (struct task_struct *p, int on_rq);
1230 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
1232 prev->sched_class->put_prev_task(rq, prev);
1235 #define sched_class_highest (&stop_sched_class)
1236 #define for_each_class(class) \
1237 for (class = sched_class_highest; class; class = class->next)
1239 extern const struct sched_class stop_sched_class;
1240 extern const struct sched_class dl_sched_class;
1241 extern const struct sched_class rt_sched_class;
1242 extern const struct sched_class fair_sched_class;
1243 extern const struct sched_class idle_sched_class;
1248 extern void update_group_capacity(struct sched_domain *sd, int cpu);
1250 extern void trigger_load_balance(struct rq *rq);
1252 extern void idle_enter_fair(struct rq *this_rq);
1253 extern void idle_exit_fair(struct rq *this_rq);
1257 static inline void idle_enter_fair(struct rq *rq) { }
1258 static inline void idle_exit_fair(struct rq *rq) { }
1262 #ifdef CONFIG_CPU_IDLE
1263 static inline void idle_set_state(struct rq *rq,
1264 struct cpuidle_state *idle_state)
1266 rq->idle_state = idle_state;
1269 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1271 WARN_ON(!rcu_read_lock_held());
1272 return rq->idle_state;
1275 static inline void idle_set_state(struct rq *rq,
1276 struct cpuidle_state *idle_state)
1280 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
1286 extern void sysrq_sched_debug_show(void);
1287 extern void sched_init_granularity(void);
1288 extern void update_max_interval(void);
1290 extern void init_sched_dl_class(void);
1291 extern void init_sched_rt_class(void);
1292 extern void init_sched_fair_class(void);
1293 extern void init_sched_dl_class(void);
1295 extern void resched_curr(struct rq *rq);
1296 extern void resched_cpu(int cpu);
1298 extern struct rt_bandwidth def_rt_bandwidth;
1299 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
1301 extern struct dl_bandwidth def_dl_bandwidth;
1302 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime);
1303 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
1305 unsigned long to_ratio(u64 period, u64 runtime);
1307 extern void init_task_runnable_average(struct task_struct *p);
1309 static inline void add_nr_running(struct rq *rq, unsigned count)
1311 unsigned prev_nr = rq->nr_running;
1313 rq->nr_running = prev_nr + count;
1315 if (prev_nr < 2 && rq->nr_running >= 2) {
1317 if (!rq->rd->overload)
1318 rq->rd->overload = true;
1321 #ifdef CONFIG_NO_HZ_FULL
1322 if (tick_nohz_full_cpu(rq->cpu)) {
1324 * Tick is needed if more than one task runs on a CPU.
1325 * Send the target an IPI to kick it out of nohz mode.
1327 * We assume that IPI implies full memory barrier and the
1328 * new value of rq->nr_running is visible on reception
1331 tick_nohz_full_kick_cpu(rq->cpu);
1337 static inline void sub_nr_running(struct rq *rq, unsigned count)
1339 rq->nr_running -= count;
1342 static inline void rq_last_tick_reset(struct rq *rq)
1344 #ifdef CONFIG_NO_HZ_FULL
1345 rq->last_sched_tick = jiffies;
1349 extern void update_rq_clock(struct rq *rq);
1351 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
1352 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
1354 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
1356 extern const_debug unsigned int sysctl_sched_time_avg;
1357 extern const_debug unsigned int sysctl_sched_nr_migrate;
1358 extern const_debug unsigned int sysctl_sched_migration_cost;
1360 static inline u64 sched_avg_period(void)
1362 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1365 #ifdef CONFIG_SCHED_HRTICK
1369 * - enabled by features
1370 * - hrtimer is actually high res
1372 static inline int hrtick_enabled(struct rq *rq)
1374 if (!sched_feat(HRTICK))
1376 if (!cpu_active(cpu_of(rq)))
1378 return hrtimer_is_hres_active(&rq->hrtick_timer);
1381 void hrtick_start(struct rq *rq, u64 delay);
1385 static inline int hrtick_enabled(struct rq *rq)
1390 #endif /* CONFIG_SCHED_HRTICK */
1393 extern void sched_avg_update(struct rq *rq);
1395 #ifndef arch_scale_freq_capacity
1396 static __always_inline
1397 unsigned long arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
1399 return SCHED_CAPACITY_SCALE;
1403 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1405 rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq));
1406 sched_avg_update(rq);
1409 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { }
1410 static inline void sched_avg_update(struct rq *rq) { }
1413 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
1416 * __task_rq_lock - lock the rq @p resides on.
1418 static inline struct rq *__task_rq_lock(struct task_struct *p)
1419 __acquires(rq->lock)
1423 lockdep_assert_held(&p->pi_lock);
1427 raw_spin_lock(&rq->lock);
1428 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
1430 raw_spin_unlock(&rq->lock);
1432 while (unlikely(task_on_rq_migrating(p)))
1438 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1440 static inline struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1441 __acquires(p->pi_lock)
1442 __acquires(rq->lock)
1447 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1449 raw_spin_lock(&rq->lock);
1451 * move_queued_task() task_rq_lock()
1453 * ACQUIRE (rq->lock)
1454 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
1455 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
1456 * [S] ->cpu = new_cpu [L] task_rq()
1458 * RELEASE (rq->lock)
1460 * If we observe the old cpu in task_rq_lock, the acquire of
1461 * the old rq->lock will fully serialize against the stores.
1463 * If we observe the new cpu in task_rq_lock, the acquire will
1464 * pair with the WMB to ensure we must then also see migrating.
1466 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
1468 raw_spin_unlock(&rq->lock);
1469 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1471 while (unlikely(task_on_rq_migrating(p)))
1476 static inline void __task_rq_unlock(struct rq *rq)
1477 __releases(rq->lock)
1479 raw_spin_unlock(&rq->lock);
1483 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1484 __releases(rq->lock)
1485 __releases(p->pi_lock)
1487 raw_spin_unlock(&rq->lock);
1488 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1492 #ifdef CONFIG_PREEMPT
1494 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2);
1497 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1498 * way at the expense of forcing extra atomic operations in all
1499 * invocations. This assures that the double_lock is acquired using the
1500 * same underlying policy as the spinlock_t on this architecture, which
1501 * reduces latency compared to the unfair variant below. However, it
1502 * also adds more overhead and therefore may reduce throughput.
1504 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1505 __releases(this_rq->lock)
1506 __acquires(busiest->lock)
1507 __acquires(this_rq->lock)
1509 raw_spin_unlock(&this_rq->lock);
1510 double_rq_lock(this_rq, busiest);
1517 * Unfair double_lock_balance: Optimizes throughput at the expense of
1518 * latency by eliminating extra atomic operations when the locks are
1519 * already in proper order on entry. This favors lower cpu-ids and will
1520 * grant the double lock to lower cpus over higher ids under contention,
1521 * regardless of entry order into the function.
1523 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1524 __releases(this_rq->lock)
1525 __acquires(busiest->lock)
1526 __acquires(this_rq->lock)
1530 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1531 if (busiest < this_rq) {
1532 raw_spin_unlock(&this_rq->lock);
1533 raw_spin_lock(&busiest->lock);
1534 raw_spin_lock_nested(&this_rq->lock,
1535 SINGLE_DEPTH_NESTING);
1538 raw_spin_lock_nested(&busiest->lock,
1539 SINGLE_DEPTH_NESTING);
1544 #endif /* CONFIG_PREEMPT */
1547 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1549 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1551 if (unlikely(!irqs_disabled())) {
1552 /* printk() doesn't work good under rq->lock */
1553 raw_spin_unlock(&this_rq->lock);
1557 return _double_lock_balance(this_rq, busiest);
1560 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1561 __releases(busiest->lock)
1563 raw_spin_unlock(&busiest->lock);
1564 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1567 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
1573 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1576 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
1582 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1585 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
1591 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1595 * double_rq_lock - safely lock two runqueues
1597 * Note this does not disable interrupts like task_rq_lock,
1598 * you need to do so manually before calling.
1600 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1601 __acquires(rq1->lock)
1602 __acquires(rq2->lock)
1604 BUG_ON(!irqs_disabled());
1606 raw_spin_lock(&rq1->lock);
1607 __acquire(rq2->lock); /* Fake it out ;) */
1610 raw_spin_lock(&rq1->lock);
1611 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1613 raw_spin_lock(&rq2->lock);
1614 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1620 * double_rq_unlock - safely unlock two runqueues
1622 * Note this does not restore interrupts like task_rq_unlock,
1623 * you need to do so manually after calling.
1625 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1626 __releases(rq1->lock)
1627 __releases(rq2->lock)
1629 raw_spin_unlock(&rq1->lock);
1631 raw_spin_unlock(&rq2->lock);
1633 __release(rq2->lock);
1636 #else /* CONFIG_SMP */
1639 * double_rq_lock - safely lock two runqueues
1641 * Note this does not disable interrupts like task_rq_lock,
1642 * you need to do so manually before calling.
1644 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
1645 __acquires(rq1->lock)
1646 __acquires(rq2->lock)
1648 BUG_ON(!irqs_disabled());
1650 raw_spin_lock(&rq1->lock);
1651 __acquire(rq2->lock); /* Fake it out ;) */
1655 * double_rq_unlock - safely unlock two runqueues
1657 * Note this does not restore interrupts like task_rq_unlock,
1658 * you need to do so manually after calling.
1660 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1661 __releases(rq1->lock)
1662 __releases(rq2->lock)
1665 raw_spin_unlock(&rq1->lock);
1666 __release(rq2->lock);
1671 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
1672 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
1673 extern void print_cfs_stats(struct seq_file *m, int cpu);
1674 extern void print_rt_stats(struct seq_file *m, int cpu);
1675 extern void print_dl_stats(struct seq_file *m, int cpu);
1677 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
1678 extern void init_rt_rq(struct rt_rq *rt_rq);
1679 extern void init_dl_rq(struct dl_rq *dl_rq);
1681 extern void cfs_bandwidth_usage_inc(void);
1682 extern void cfs_bandwidth_usage_dec(void);
1684 #ifdef CONFIG_NO_HZ_COMMON
1685 enum rq_nohz_flag_bits {
1690 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
1693 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1695 DECLARE_PER_CPU(u64, cpu_hardirq_time);
1696 DECLARE_PER_CPU(u64, cpu_softirq_time);
1698 #ifndef CONFIG_64BIT
1699 DECLARE_PER_CPU(seqcount_t, irq_time_seq);
1701 static inline void irq_time_write_begin(void)
1703 __this_cpu_inc(irq_time_seq.sequence);
1707 static inline void irq_time_write_end(void)
1710 __this_cpu_inc(irq_time_seq.sequence);
1713 static inline u64 irq_time_read(int cpu)
1719 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1720 irq_time = per_cpu(cpu_softirq_time, cpu) +
1721 per_cpu(cpu_hardirq_time, cpu);
1722 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1726 #else /* CONFIG_64BIT */
1727 static inline void irq_time_write_begin(void)
1731 static inline void irq_time_write_end(void)
1735 static inline u64 irq_time_read(int cpu)
1737 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1739 #endif /* CONFIG_64BIT */
1740 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */