4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
72 #include <asm/irq_regs.h>
75 * Scheduler clock - returns current time in nanosec units.
76 * This is default implementation.
77 * Architectures and sub-architectures can override this.
79 unsigned long long __attribute__((weak)) sched_clock(void)
81 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 #ifdef CONFIG_GROUP_SCHED
161 #include <linux/cgroup.h>
165 static LIST_HEAD(task_groups);
167 /* task group related information */
169 #ifdef CONFIG_CGROUP_SCHED
170 struct cgroup_subsys_state css;
173 #ifdef CONFIG_FAIR_GROUP_SCHED
174 /* schedulable entities of this group on each cpu */
175 struct sched_entity **se;
176 /* runqueue "owned" by this group on each cpu */
177 struct cfs_rq **cfs_rq;
178 unsigned long shares;
181 #ifdef CONFIG_RT_GROUP_SCHED
182 struct sched_rt_entity **rt_se;
183 struct rt_rq **rt_rq;
189 struct list_head list;
192 #ifdef CONFIG_FAIR_GROUP_SCHED
193 /* Default task group's sched entity on each cpu */
194 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
195 /* Default task group's cfs_rq on each cpu */
196 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
198 static struct sched_entity *init_sched_entity_p[NR_CPUS];
199 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
202 #ifdef CONFIG_RT_GROUP_SCHED
203 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
204 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
206 static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
207 static struct rt_rq *init_rt_rq_p[NR_CPUS];
210 /* task_group_lock serializes add/remove of task groups and also changes to
211 * a task group's cpu shares.
213 static DEFINE_SPINLOCK(task_group_lock);
215 /* doms_cur_mutex serializes access to doms_cur[] array */
216 static DEFINE_MUTEX(doms_cur_mutex);
218 #ifdef CONFIG_FAIR_GROUP_SCHED
219 #ifdef CONFIG_USER_SCHED
220 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
222 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
225 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
228 /* Default task group.
229 * Every task in system belong to this group at bootup.
231 struct task_group init_task_group = {
232 #ifdef CONFIG_FAIR_GROUP_SCHED
233 .se = init_sched_entity_p,
234 .cfs_rq = init_cfs_rq_p,
237 #ifdef CONFIG_RT_GROUP_SCHED
238 .rt_se = init_sched_rt_entity_p,
239 .rt_rq = init_rt_rq_p,
243 /* return group to which a task belongs */
244 static inline struct task_group *task_group(struct task_struct *p)
246 struct task_group *tg;
248 #ifdef CONFIG_USER_SCHED
250 #elif defined(CONFIG_CGROUP_SCHED)
251 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
252 struct task_group, css);
254 tg = &init_task_group;
259 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
260 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
262 #ifdef CONFIG_FAIR_GROUP_SCHED
263 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
264 p->se.parent = task_group(p)->se[cpu];
267 #ifdef CONFIG_RT_GROUP_SCHED
268 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
269 p->rt.parent = task_group(p)->rt_se[cpu];
273 static inline void lock_doms_cur(void)
275 mutex_lock(&doms_cur_mutex);
278 static inline void unlock_doms_cur(void)
280 mutex_unlock(&doms_cur_mutex);
285 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
286 static inline void lock_doms_cur(void) { }
287 static inline void unlock_doms_cur(void) { }
289 #endif /* CONFIG_GROUP_SCHED */
291 /* CFS-related fields in a runqueue */
293 struct load_weight load;
294 unsigned long nr_running;
299 struct rb_root tasks_timeline;
300 struct rb_node *rb_leftmost;
301 struct rb_node *rb_load_balance_curr;
302 /* 'curr' points to currently running entity on this cfs_rq.
303 * It is set to NULL otherwise (i.e when none are currently running).
305 struct sched_entity *curr, *next;
307 unsigned long nr_spread_over;
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
313 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
314 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
315 * (like users, containers etc.)
317 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
318 * list is used during load balance.
320 struct list_head leaf_cfs_rq_list;
321 struct task_group *tg; /* group that "owns" this runqueue */
325 /* Real-Time classes' related field in a runqueue: */
327 struct rt_prio_array active;
328 unsigned long rt_nr_running;
329 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
330 int highest_prio; /* highest queued rt task prio */
333 unsigned long rt_nr_migratory;
339 #ifdef CONFIG_RT_GROUP_SCHED
340 unsigned long rt_nr_boosted;
343 struct list_head leaf_rt_rq_list;
344 struct task_group *tg;
345 struct sched_rt_entity *rt_se;
352 * We add the notion of a root-domain which will be used to define per-domain
353 * variables. Each exclusive cpuset essentially defines an island domain by
354 * fully partitioning the member cpus from any other cpuset. Whenever a new
355 * exclusive cpuset is created, we also create and attach a new root-domain
365 * The "RT overload" flag: it gets set if a CPU has more than
366 * one runnable RT task.
373 * By default the system creates a single root-domain with all cpus as
374 * members (mimicking the global state we have today).
376 static struct root_domain def_root_domain;
381 * This is the main, per-CPU runqueue data structure.
383 * Locking rule: those places that want to lock multiple runqueues
384 * (such as the load balancing or the thread migration code), lock
385 * acquire operations must be ordered by ascending &runqueue.
392 * nr_running and cpu_load should be in the same cacheline because
393 * remote CPUs use both these fields when doing load calculation.
395 unsigned long nr_running;
396 #define CPU_LOAD_IDX_MAX 5
397 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
398 unsigned char idle_at_tick;
400 unsigned long last_tick_seen;
401 unsigned char in_nohz_recently;
403 /* capture load from *all* tasks on this cpu: */
404 struct load_weight load;
405 unsigned long nr_load_updates;
410 u64 rt_period_expire;
413 #ifdef CONFIG_FAIR_GROUP_SCHED
414 /* list of leaf cfs_rq on this cpu: */
415 struct list_head leaf_cfs_rq_list;
417 #ifdef CONFIG_RT_GROUP_SCHED
418 struct list_head leaf_rt_rq_list;
422 * This is part of a global counter where only the total sum
423 * over all CPUs matters. A task can increase this counter on
424 * one CPU and if it got migrated afterwards it may decrease
425 * it on another CPU. Always updated under the runqueue lock:
427 unsigned long nr_uninterruptible;
429 struct task_struct *curr, *idle;
430 unsigned long next_balance;
431 struct mm_struct *prev_mm;
433 u64 clock, prev_clock_raw;
436 unsigned int clock_warps, clock_overflows, clock_underflows;
438 unsigned int clock_deep_idle_events;
444 struct root_domain *rd;
445 struct sched_domain *sd;
447 /* For active balancing */
450 /* cpu of this runqueue: */
453 struct task_struct *migration_thread;
454 struct list_head migration_queue;
457 #ifdef CONFIG_SCHED_HRTICK
458 unsigned long hrtick_flags;
459 ktime_t hrtick_expire;
460 struct hrtimer hrtick_timer;
463 #ifdef CONFIG_SCHEDSTATS
465 struct sched_info rq_sched_info;
467 /* sys_sched_yield() stats */
468 unsigned int yld_exp_empty;
469 unsigned int yld_act_empty;
470 unsigned int yld_both_empty;
471 unsigned int yld_count;
473 /* schedule() stats */
474 unsigned int sched_switch;
475 unsigned int sched_count;
476 unsigned int sched_goidle;
478 /* try_to_wake_up() stats */
479 unsigned int ttwu_count;
480 unsigned int ttwu_local;
483 unsigned int bkl_count;
485 struct lock_class_key rq_lock_key;
488 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
490 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
492 rq->curr->sched_class->check_preempt_curr(rq, p);
495 static inline int cpu_of(struct rq *rq)
505 static inline bool nohz_on(int cpu)
507 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
510 static inline u64 max_skipped_ticks(struct rq *rq)
512 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
515 static inline void update_last_tick_seen(struct rq *rq)
517 rq->last_tick_seen = jiffies;
520 static inline u64 max_skipped_ticks(struct rq *rq)
525 static inline void update_last_tick_seen(struct rq *rq)
531 * Update the per-runqueue clock, as finegrained as the platform can give
532 * us, but without assuming monotonicity, etc.:
534 static void __update_rq_clock(struct rq *rq)
536 u64 prev_raw = rq->prev_clock_raw;
537 u64 now = sched_clock();
538 s64 delta = now - prev_raw;
539 u64 clock = rq->clock;
541 #ifdef CONFIG_SCHED_DEBUG
542 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
545 * Protect against sched_clock() occasionally going backwards:
547 if (unlikely(delta < 0)) {
552 * Catch too large forward jumps too:
554 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
555 u64 max_time = rq->tick_timestamp + max_jump;
557 if (unlikely(clock + delta > max_time)) {
558 if (clock < max_time)
562 rq->clock_overflows++;
564 if (unlikely(delta > rq->clock_max_delta))
565 rq->clock_max_delta = delta;
570 rq->prev_clock_raw = now;
574 static void update_rq_clock(struct rq *rq)
576 if (likely(smp_processor_id() == cpu_of(rq)))
577 __update_rq_clock(rq);
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
595 unsigned long rt_needs_cpu(int cpu)
597 struct rq *rq = cpu_rq(cpu);
600 if (!rq->rt_throttled)
603 if (rq->clock > rq->rt_period_expire)
606 delta = rq->rt_period_expire - rq->clock;
607 do_div(delta, NSEC_PER_SEC / HZ);
609 return (unsigned long)delta;
613 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
615 #ifdef CONFIG_SCHED_DEBUG
616 # define const_debug __read_mostly
618 # define const_debug static const
622 * Debugging: various feature bits
625 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
626 SCHED_FEAT_WAKEUP_PREEMPT = 2,
627 SCHED_FEAT_START_DEBIT = 4,
628 SCHED_FEAT_AFFINE_WAKEUPS = 8,
629 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
630 SCHED_FEAT_SYNC_WAKEUPS = 32,
631 SCHED_FEAT_HRTICK = 64,
632 SCHED_FEAT_DOUBLE_TICK = 128,
635 const_debug unsigned int sysctl_sched_features =
636 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
637 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
638 SCHED_FEAT_START_DEBIT * 1 |
639 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
640 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
641 SCHED_FEAT_SYNC_WAKEUPS * 1 |
642 SCHED_FEAT_HRTICK * 1 |
643 SCHED_FEAT_DOUBLE_TICK * 0;
645 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
648 * Number of tasks to iterate in a single balance run.
649 * Limited because this is done with IRQs disabled.
651 const_debug unsigned int sysctl_sched_nr_migrate = 32;
654 * period over which we measure -rt task cpu usage in us.
657 unsigned int sysctl_sched_rt_period = 1000000;
659 static __read_mostly int scheduler_running;
662 * part of the period that we allow rt tasks to run in us.
665 int sysctl_sched_rt_runtime = 950000;
668 * single value that denotes runtime == period, ie unlimited time.
670 #define RUNTIME_INF ((u64)~0ULL)
672 static const unsigned long long time_sync_thresh = 100000;
674 static DEFINE_PER_CPU(unsigned long long, time_offset);
675 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
678 * Global lock which we take every now and then to synchronize
679 * the CPUs time. This method is not warp-safe, but it's good
680 * enough to synchronize slowly diverging time sources and thus
681 * it's good enough for tracing:
683 static DEFINE_SPINLOCK(time_sync_lock);
684 static unsigned long long prev_global_time;
686 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
690 spin_lock_irqsave(&time_sync_lock, flags);
692 if (time < prev_global_time) {
693 per_cpu(time_offset, cpu) += prev_global_time - time;
694 time = prev_global_time;
696 prev_global_time = time;
699 spin_unlock_irqrestore(&time_sync_lock, flags);
704 static unsigned long long __cpu_clock(int cpu)
706 unsigned long long now;
711 * Only call sched_clock() if the scheduler has already been
712 * initialized (some code might call cpu_clock() very early):
714 if (unlikely(!scheduler_running))
717 local_irq_save(flags);
721 local_irq_restore(flags);
727 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
728 * clock constructed from sched_clock():
730 unsigned long long cpu_clock(int cpu)
732 unsigned long long prev_cpu_time, time, delta_time;
734 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
735 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
736 delta_time = time-prev_cpu_time;
738 if (unlikely(delta_time > time_sync_thresh))
739 time = __sync_cpu_clock(time, cpu);
743 EXPORT_SYMBOL_GPL(cpu_clock);
745 #ifndef prepare_arch_switch
746 # define prepare_arch_switch(next) do { } while (0)
748 #ifndef finish_arch_switch
749 # define finish_arch_switch(prev) do { } while (0)
752 static inline int task_current(struct rq *rq, struct task_struct *p)
754 return rq->curr == p;
757 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
758 static inline int task_running(struct rq *rq, struct task_struct *p)
760 return task_current(rq, p);
763 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
767 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
769 #ifdef CONFIG_DEBUG_SPINLOCK
770 /* this is a valid case when another task releases the spinlock */
771 rq->lock.owner = current;
774 * If we are tracking spinlock dependencies then we have to
775 * fix up the runqueue lock - which gets 'carried over' from
778 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
780 spin_unlock_irq(&rq->lock);
783 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
784 static inline int task_running(struct rq *rq, struct task_struct *p)
789 return task_current(rq, p);
793 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
797 * We can optimise this out completely for !SMP, because the
798 * SMP rebalancing from interrupt is the only thing that cares
803 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
804 spin_unlock_irq(&rq->lock);
806 spin_unlock(&rq->lock);
810 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
814 * After ->oncpu is cleared, the task can be moved to a different CPU.
815 * We must ensure this doesn't happen until the switch is completely
821 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
825 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
828 * __task_rq_lock - lock the runqueue a given task resides on.
829 * Must be called interrupts disabled.
831 static inline struct rq *__task_rq_lock(struct task_struct *p)
835 struct rq *rq = task_rq(p);
836 spin_lock(&rq->lock);
837 if (likely(rq == task_rq(p)))
839 spin_unlock(&rq->lock);
844 * task_rq_lock - lock the runqueue a given task resides on and disable
845 * interrupts. Note the ordering: we can safely lookup the task_rq without
846 * explicitly disabling preemption.
848 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
854 local_irq_save(*flags);
856 spin_lock(&rq->lock);
857 if (likely(rq == task_rq(p)))
859 spin_unlock_irqrestore(&rq->lock, *flags);
863 static void __task_rq_unlock(struct rq *rq)
866 spin_unlock(&rq->lock);
869 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
872 spin_unlock_irqrestore(&rq->lock, *flags);
876 * this_rq_lock - lock this runqueue and disable interrupts.
878 static struct rq *this_rq_lock(void)
885 spin_lock(&rq->lock);
891 * We are going deep-idle (irqs are disabled):
893 void sched_clock_idle_sleep_event(void)
895 struct rq *rq = cpu_rq(smp_processor_id());
897 spin_lock(&rq->lock);
898 __update_rq_clock(rq);
899 spin_unlock(&rq->lock);
900 rq->clock_deep_idle_events++;
902 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
905 * We just idled delta nanoseconds (called with irqs disabled):
907 void sched_clock_idle_wakeup_event(u64 delta_ns)
909 struct rq *rq = cpu_rq(smp_processor_id());
910 u64 now = sched_clock();
912 rq->idle_clock += delta_ns;
914 * Override the previous timestamp and ignore all
915 * sched_clock() deltas that occured while we idled,
916 * and use the PM-provided delta_ns to advance the
919 spin_lock(&rq->lock);
920 rq->prev_clock_raw = now;
921 rq->clock += delta_ns;
922 spin_unlock(&rq->lock);
923 touch_softlockup_watchdog();
925 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
927 static void __resched_task(struct task_struct *p, int tif_bit);
929 static inline void resched_task(struct task_struct *p)
931 __resched_task(p, TIF_NEED_RESCHED);
934 #ifdef CONFIG_SCHED_HRTICK
936 * Use HR-timers to deliver accurate preemption points.
938 * Its all a bit involved since we cannot program an hrt while holding the
939 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
942 * When we get rescheduled we reprogram the hrtick_timer outside of the
945 static inline void resched_hrt(struct task_struct *p)
947 __resched_task(p, TIF_HRTICK_RESCHED);
950 static inline void resched_rq(struct rq *rq)
954 spin_lock_irqsave(&rq->lock, flags);
955 resched_task(rq->curr);
956 spin_unlock_irqrestore(&rq->lock, flags);
960 HRTICK_SET, /* re-programm hrtick_timer */
961 HRTICK_RESET, /* not a new slice */
966 * - enabled by features
967 * - hrtimer is actually high res
969 static inline int hrtick_enabled(struct rq *rq)
971 if (!sched_feat(HRTICK))
973 return hrtimer_is_hres_active(&rq->hrtick_timer);
977 * Called to set the hrtick timer state.
979 * called with rq->lock held and irqs disabled
981 static void hrtick_start(struct rq *rq, u64 delay, int reset)
983 assert_spin_locked(&rq->lock);
986 * preempt at: now + delay
989 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
991 * indicate we need to program the timer
993 __set_bit(HRTICK_SET, &rq->hrtick_flags);
995 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
998 * New slices are called from the schedule path and don't need a
1002 resched_hrt(rq->curr);
1005 static void hrtick_clear(struct rq *rq)
1007 if (hrtimer_active(&rq->hrtick_timer))
1008 hrtimer_cancel(&rq->hrtick_timer);
1012 * Update the timer from the possible pending state.
1014 static void hrtick_set(struct rq *rq)
1018 unsigned long flags;
1020 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1022 spin_lock_irqsave(&rq->lock, flags);
1023 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1024 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1025 time = rq->hrtick_expire;
1026 clear_thread_flag(TIF_HRTICK_RESCHED);
1027 spin_unlock_irqrestore(&rq->lock, flags);
1030 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1031 if (reset && !hrtimer_active(&rq->hrtick_timer))
1038 * High-resolution timer tick.
1039 * Runs from hardirq context with interrupts disabled.
1041 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1043 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1045 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1047 spin_lock(&rq->lock);
1048 __update_rq_clock(rq);
1049 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1050 spin_unlock(&rq->lock);
1052 return HRTIMER_NORESTART;
1055 static inline void init_rq_hrtick(struct rq *rq)
1057 rq->hrtick_flags = 0;
1058 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1059 rq->hrtick_timer.function = hrtick;
1060 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1063 void hrtick_resched(void)
1066 unsigned long flags;
1068 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1071 local_irq_save(flags);
1072 rq = cpu_rq(smp_processor_id());
1074 local_irq_restore(flags);
1077 static inline void hrtick_clear(struct rq *rq)
1081 static inline void hrtick_set(struct rq *rq)
1085 static inline void init_rq_hrtick(struct rq *rq)
1089 void hrtick_resched(void)
1095 * resched_task - mark a task 'to be rescheduled now'.
1097 * On UP this means the setting of the need_resched flag, on SMP it
1098 * might also involve a cross-CPU call to trigger the scheduler on
1103 #ifndef tsk_is_polling
1104 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1107 static void __resched_task(struct task_struct *p, int tif_bit)
1111 assert_spin_locked(&task_rq(p)->lock);
1113 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1116 set_tsk_thread_flag(p, tif_bit);
1119 if (cpu == smp_processor_id())
1122 /* NEED_RESCHED must be visible before we test polling */
1124 if (!tsk_is_polling(p))
1125 smp_send_reschedule(cpu);
1128 static void resched_cpu(int cpu)
1130 struct rq *rq = cpu_rq(cpu);
1131 unsigned long flags;
1133 if (!spin_trylock_irqsave(&rq->lock, flags))
1135 resched_task(cpu_curr(cpu));
1136 spin_unlock_irqrestore(&rq->lock, flags);
1141 * When add_timer_on() enqueues a timer into the timer wheel of an
1142 * idle CPU then this timer might expire before the next timer event
1143 * which is scheduled to wake up that CPU. In case of a completely
1144 * idle system the next event might even be infinite time into the
1145 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1146 * leaves the inner idle loop so the newly added timer is taken into
1147 * account when the CPU goes back to idle and evaluates the timer
1148 * wheel for the next timer event.
1150 void wake_up_idle_cpu(int cpu)
1152 struct rq *rq = cpu_rq(cpu);
1154 if (cpu == smp_processor_id())
1158 * This is safe, as this function is called with the timer
1159 * wheel base lock of (cpu) held. When the CPU is on the way
1160 * to idle and has not yet set rq->curr to idle then it will
1161 * be serialized on the timer wheel base lock and take the new
1162 * timer into account automatically.
1164 if (rq->curr != rq->idle)
1168 * We can set TIF_RESCHED on the idle task of the other CPU
1169 * lockless. The worst case is that the other CPU runs the
1170 * idle task through an additional NOOP schedule()
1172 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1174 /* NEED_RESCHED must be visible before we test polling */
1176 if (!tsk_is_polling(rq->idle))
1177 smp_send_reschedule(cpu);
1182 static void __resched_task(struct task_struct *p, int tif_bit)
1184 assert_spin_locked(&task_rq(p)->lock);
1185 set_tsk_thread_flag(p, tif_bit);
1189 #if BITS_PER_LONG == 32
1190 # define WMULT_CONST (~0UL)
1192 # define WMULT_CONST (1UL << 32)
1195 #define WMULT_SHIFT 32
1198 * Shift right and round:
1200 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1202 static unsigned long
1203 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1204 struct load_weight *lw)
1208 if (unlikely(!lw->inv_weight))
1209 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1211 tmp = (u64)delta_exec * weight;
1213 * Check whether we'd overflow the 64-bit multiplication:
1215 if (unlikely(tmp > WMULT_CONST))
1216 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1219 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1221 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1224 static inline unsigned long
1225 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1227 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1230 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1236 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1243 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1244 * of tasks with abnormal "nice" values across CPUs the contribution that
1245 * each task makes to its run queue's load is weighted according to its
1246 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1247 * scaled version of the new time slice allocation that they receive on time
1251 #define WEIGHT_IDLEPRIO 2
1252 #define WMULT_IDLEPRIO (1 << 31)
1255 * Nice levels are multiplicative, with a gentle 10% change for every
1256 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1257 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1258 * that remained on nice 0.
1260 * The "10% effect" is relative and cumulative: from _any_ nice level,
1261 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1262 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1263 * If a task goes up by ~10% and another task goes down by ~10% then
1264 * the relative distance between them is ~25%.)
1266 static const int prio_to_weight[40] = {
1267 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1268 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1269 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1270 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1271 /* 0 */ 1024, 820, 655, 526, 423,
1272 /* 5 */ 335, 272, 215, 172, 137,
1273 /* 10 */ 110, 87, 70, 56, 45,
1274 /* 15 */ 36, 29, 23, 18, 15,
1278 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1280 * In cases where the weight does not change often, we can use the
1281 * precalculated inverse to speed up arithmetics by turning divisions
1282 * into multiplications:
1284 static const u32 prio_to_wmult[40] = {
1285 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1286 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1287 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1288 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1289 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1290 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1291 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1292 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1295 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1298 * runqueue iterator, to support SMP load-balancing between different
1299 * scheduling classes, without having to expose their internal data
1300 * structures to the load-balancing proper:
1302 struct rq_iterator {
1304 struct task_struct *(*start)(void *);
1305 struct task_struct *(*next)(void *);
1309 static unsigned long
1310 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1311 unsigned long max_load_move, struct sched_domain *sd,
1312 enum cpu_idle_type idle, int *all_pinned,
1313 int *this_best_prio, struct rq_iterator *iterator);
1316 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1317 struct sched_domain *sd, enum cpu_idle_type idle,
1318 struct rq_iterator *iterator);
1321 #ifdef CONFIG_CGROUP_CPUACCT
1322 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1324 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1328 static unsigned long source_load(int cpu, int type);
1329 static unsigned long target_load(int cpu, int type);
1330 static unsigned long cpu_avg_load_per_task(int cpu);
1331 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1332 #endif /* CONFIG_SMP */
1334 #include "sched_stats.h"
1335 #include "sched_idletask.c"
1336 #include "sched_fair.c"
1337 #include "sched_rt.c"
1338 #ifdef CONFIG_SCHED_DEBUG
1339 # include "sched_debug.c"
1342 #define sched_class_highest (&rt_sched_class)
1344 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1346 update_load_add(&rq->load, p->se.load.weight);
1349 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1351 update_load_sub(&rq->load, p->se.load.weight);
1354 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1360 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1366 static void set_load_weight(struct task_struct *p)
1368 if (task_has_rt_policy(p)) {
1369 p->se.load.weight = prio_to_weight[0] * 2;
1370 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1375 * SCHED_IDLE tasks get minimal weight:
1377 if (p->policy == SCHED_IDLE) {
1378 p->se.load.weight = WEIGHT_IDLEPRIO;
1379 p->se.load.inv_weight = WMULT_IDLEPRIO;
1383 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1384 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1387 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1389 sched_info_queued(p);
1390 p->sched_class->enqueue_task(rq, p, wakeup);
1394 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1396 p->sched_class->dequeue_task(rq, p, sleep);
1401 * __normal_prio - return the priority that is based on the static prio
1403 static inline int __normal_prio(struct task_struct *p)
1405 return p->static_prio;
1409 * Calculate the expected normal priority: i.e. priority
1410 * without taking RT-inheritance into account. Might be
1411 * boosted by interactivity modifiers. Changes upon fork,
1412 * setprio syscalls, and whenever the interactivity
1413 * estimator recalculates.
1415 static inline int normal_prio(struct task_struct *p)
1419 if (task_has_rt_policy(p))
1420 prio = MAX_RT_PRIO-1 - p->rt_priority;
1422 prio = __normal_prio(p);
1427 * Calculate the current priority, i.e. the priority
1428 * taken into account by the scheduler. This value might
1429 * be boosted by RT tasks, or might be boosted by
1430 * interactivity modifiers. Will be RT if the task got
1431 * RT-boosted. If not then it returns p->normal_prio.
1433 static int effective_prio(struct task_struct *p)
1435 p->normal_prio = normal_prio(p);
1437 * If we are RT tasks or we were boosted to RT priority,
1438 * keep the priority unchanged. Otherwise, update priority
1439 * to the normal priority:
1441 if (!rt_prio(p->prio))
1442 return p->normal_prio;
1447 * activate_task - move a task to the runqueue.
1449 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1451 if (task_contributes_to_load(p))
1452 rq->nr_uninterruptible--;
1454 enqueue_task(rq, p, wakeup);
1455 inc_nr_running(p, rq);
1459 * deactivate_task - remove a task from the runqueue.
1461 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1463 if (task_contributes_to_load(p))
1464 rq->nr_uninterruptible++;
1466 dequeue_task(rq, p, sleep);
1467 dec_nr_running(p, rq);
1471 * task_curr - is this task currently executing on a CPU?
1472 * @p: the task in question.
1474 inline int task_curr(const struct task_struct *p)
1476 return cpu_curr(task_cpu(p)) == p;
1479 /* Used instead of source_load when we know the type == 0 */
1480 unsigned long weighted_cpuload(const int cpu)
1482 return cpu_rq(cpu)->load.weight;
1485 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1487 set_task_rq(p, cpu);
1490 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1491 * successfuly executed on another CPU. We must ensure that updates of
1492 * per-task data have been completed by this moment.
1495 task_thread_info(p)->cpu = cpu;
1499 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1500 const struct sched_class *prev_class,
1501 int oldprio, int running)
1503 if (prev_class != p->sched_class) {
1504 if (prev_class->switched_from)
1505 prev_class->switched_from(rq, p, running);
1506 p->sched_class->switched_to(rq, p, running);
1508 p->sched_class->prio_changed(rq, p, oldprio, running);
1514 * Is this task likely cache-hot:
1517 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1522 * Buddy candidates are cache hot:
1524 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1527 if (p->sched_class != &fair_sched_class)
1530 if (sysctl_sched_migration_cost == -1)
1532 if (sysctl_sched_migration_cost == 0)
1535 delta = now - p->se.exec_start;
1537 return delta < (s64)sysctl_sched_migration_cost;
1541 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1543 int old_cpu = task_cpu(p);
1544 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1545 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1546 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1549 clock_offset = old_rq->clock - new_rq->clock;
1551 #ifdef CONFIG_SCHEDSTATS
1552 if (p->se.wait_start)
1553 p->se.wait_start -= clock_offset;
1554 if (p->se.sleep_start)
1555 p->se.sleep_start -= clock_offset;
1556 if (p->se.block_start)
1557 p->se.block_start -= clock_offset;
1558 if (old_cpu != new_cpu) {
1559 schedstat_inc(p, se.nr_migrations);
1560 if (task_hot(p, old_rq->clock, NULL))
1561 schedstat_inc(p, se.nr_forced2_migrations);
1564 p->se.vruntime -= old_cfsrq->min_vruntime -
1565 new_cfsrq->min_vruntime;
1567 __set_task_cpu(p, new_cpu);
1570 struct migration_req {
1571 struct list_head list;
1573 struct task_struct *task;
1576 struct completion done;
1580 * The task's runqueue lock must be held.
1581 * Returns true if you have to wait for migration thread.
1584 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1586 struct rq *rq = task_rq(p);
1589 * If the task is not on a runqueue (and not running), then
1590 * it is sufficient to simply update the task's cpu field.
1592 if (!p->se.on_rq && !task_running(rq, p)) {
1593 set_task_cpu(p, dest_cpu);
1597 init_completion(&req->done);
1599 req->dest_cpu = dest_cpu;
1600 list_add(&req->list, &rq->migration_queue);
1606 * wait_task_inactive - wait for a thread to unschedule.
1608 * The caller must ensure that the task *will* unschedule sometime soon,
1609 * else this function might spin for a *long* time. This function can't
1610 * be called with interrupts off, or it may introduce deadlock with
1611 * smp_call_function() if an IPI is sent by the same process we are
1612 * waiting to become inactive.
1614 void wait_task_inactive(struct task_struct *p)
1616 unsigned long flags;
1622 * We do the initial early heuristics without holding
1623 * any task-queue locks at all. We'll only try to get
1624 * the runqueue lock when things look like they will
1630 * If the task is actively running on another CPU
1631 * still, just relax and busy-wait without holding
1634 * NOTE! Since we don't hold any locks, it's not
1635 * even sure that "rq" stays as the right runqueue!
1636 * But we don't care, since "task_running()" will
1637 * return false if the runqueue has changed and p
1638 * is actually now running somewhere else!
1640 while (task_running(rq, p))
1644 * Ok, time to look more closely! We need the rq
1645 * lock now, to be *sure*. If we're wrong, we'll
1646 * just go back and repeat.
1648 rq = task_rq_lock(p, &flags);
1649 running = task_running(rq, p);
1650 on_rq = p->se.on_rq;
1651 task_rq_unlock(rq, &flags);
1654 * Was it really running after all now that we
1655 * checked with the proper locks actually held?
1657 * Oops. Go back and try again..
1659 if (unlikely(running)) {
1665 * It's not enough that it's not actively running,
1666 * it must be off the runqueue _entirely_, and not
1669 * So if it wa still runnable (but just not actively
1670 * running right now), it's preempted, and we should
1671 * yield - it could be a while.
1673 if (unlikely(on_rq)) {
1674 schedule_timeout_uninterruptible(1);
1679 * Ahh, all good. It wasn't running, and it wasn't
1680 * runnable, which means that it will never become
1681 * running in the future either. We're all done!
1688 * kick_process - kick a running thread to enter/exit the kernel
1689 * @p: the to-be-kicked thread
1691 * Cause a process which is running on another CPU to enter
1692 * kernel-mode, without any delay. (to get signals handled.)
1694 * NOTE: this function doesnt have to take the runqueue lock,
1695 * because all it wants to ensure is that the remote task enters
1696 * the kernel. If the IPI races and the task has been migrated
1697 * to another CPU then no harm is done and the purpose has been
1700 void kick_process(struct task_struct *p)
1706 if ((cpu != smp_processor_id()) && task_curr(p))
1707 smp_send_reschedule(cpu);
1712 * Return a low guess at the load of a migration-source cpu weighted
1713 * according to the scheduling class and "nice" value.
1715 * We want to under-estimate the load of migration sources, to
1716 * balance conservatively.
1718 static unsigned long source_load(int cpu, int type)
1720 struct rq *rq = cpu_rq(cpu);
1721 unsigned long total = weighted_cpuload(cpu);
1726 return min(rq->cpu_load[type-1], total);
1730 * Return a high guess at the load of a migration-target cpu weighted
1731 * according to the scheduling class and "nice" value.
1733 static unsigned long target_load(int cpu, int type)
1735 struct rq *rq = cpu_rq(cpu);
1736 unsigned long total = weighted_cpuload(cpu);
1741 return max(rq->cpu_load[type-1], total);
1745 * Return the average load per task on the cpu's run queue
1747 static unsigned long cpu_avg_load_per_task(int cpu)
1749 struct rq *rq = cpu_rq(cpu);
1750 unsigned long total = weighted_cpuload(cpu);
1751 unsigned long n = rq->nr_running;
1753 return n ? total / n : SCHED_LOAD_SCALE;
1757 * find_idlest_group finds and returns the least busy CPU group within the
1760 static struct sched_group *
1761 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1763 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1764 unsigned long min_load = ULONG_MAX, this_load = 0;
1765 int load_idx = sd->forkexec_idx;
1766 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1769 unsigned long load, avg_load;
1773 /* Skip over this group if it has no CPUs allowed */
1774 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1777 local_group = cpu_isset(this_cpu, group->cpumask);
1779 /* Tally up the load of all CPUs in the group */
1782 for_each_cpu_mask(i, group->cpumask) {
1783 /* Bias balancing toward cpus of our domain */
1785 load = source_load(i, load_idx);
1787 load = target_load(i, load_idx);
1792 /* Adjust by relative CPU power of the group */
1793 avg_load = sg_div_cpu_power(group,
1794 avg_load * SCHED_LOAD_SCALE);
1797 this_load = avg_load;
1799 } else if (avg_load < min_load) {
1800 min_load = avg_load;
1803 } while (group = group->next, group != sd->groups);
1805 if (!idlest || 100*this_load < imbalance*min_load)
1811 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1814 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1817 unsigned long load, min_load = ULONG_MAX;
1821 /* Traverse only the allowed CPUs */
1822 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1824 for_each_cpu_mask(i, tmp) {
1825 load = weighted_cpuload(i);
1827 if (load < min_load || (load == min_load && i == this_cpu)) {
1837 * sched_balance_self: balance the current task (running on cpu) in domains
1838 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1841 * Balance, ie. select the least loaded group.
1843 * Returns the target CPU number, or the same CPU if no balancing is needed.
1845 * preempt must be disabled.
1847 static int sched_balance_self(int cpu, int flag)
1849 struct task_struct *t = current;
1850 struct sched_domain *tmp, *sd = NULL;
1852 for_each_domain(cpu, tmp) {
1854 * If power savings logic is enabled for a domain, stop there.
1856 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1858 if (tmp->flags & flag)
1864 struct sched_group *group;
1865 int new_cpu, weight;
1867 if (!(sd->flags & flag)) {
1873 group = find_idlest_group(sd, t, cpu);
1879 new_cpu = find_idlest_cpu(group, t, cpu);
1880 if (new_cpu == -1 || new_cpu == cpu) {
1881 /* Now try balancing at a lower domain level of cpu */
1886 /* Now try balancing at a lower domain level of new_cpu */
1889 weight = cpus_weight(span);
1890 for_each_domain(cpu, tmp) {
1891 if (weight <= cpus_weight(tmp->span))
1893 if (tmp->flags & flag)
1896 /* while loop will break here if sd == NULL */
1902 #endif /* CONFIG_SMP */
1905 * try_to_wake_up - wake up a thread
1906 * @p: the to-be-woken-up thread
1907 * @state: the mask of task states that can be woken
1908 * @sync: do a synchronous wakeup?
1910 * Put it on the run-queue if it's not already there. The "current"
1911 * thread is always on the run-queue (except when the actual
1912 * re-schedule is in progress), and as such you're allowed to do
1913 * the simpler "current->state = TASK_RUNNING" to mark yourself
1914 * runnable without the overhead of this.
1916 * returns failure only if the task is already active.
1918 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1920 int cpu, orig_cpu, this_cpu, success = 0;
1921 unsigned long flags;
1925 if (!sched_feat(SYNC_WAKEUPS))
1929 rq = task_rq_lock(p, &flags);
1930 old_state = p->state;
1931 if (!(old_state & state))
1939 this_cpu = smp_processor_id();
1942 if (unlikely(task_running(rq, p)))
1945 cpu = p->sched_class->select_task_rq(p, sync);
1946 if (cpu != orig_cpu) {
1947 set_task_cpu(p, cpu);
1948 task_rq_unlock(rq, &flags);
1949 /* might preempt at this point */
1950 rq = task_rq_lock(p, &flags);
1951 old_state = p->state;
1952 if (!(old_state & state))
1957 this_cpu = smp_processor_id();
1961 #ifdef CONFIG_SCHEDSTATS
1962 schedstat_inc(rq, ttwu_count);
1963 if (cpu == this_cpu)
1964 schedstat_inc(rq, ttwu_local);
1966 struct sched_domain *sd;
1967 for_each_domain(this_cpu, sd) {
1968 if (cpu_isset(cpu, sd->span)) {
1969 schedstat_inc(sd, ttwu_wake_remote);
1977 #endif /* CONFIG_SMP */
1978 schedstat_inc(p, se.nr_wakeups);
1980 schedstat_inc(p, se.nr_wakeups_sync);
1981 if (orig_cpu != cpu)
1982 schedstat_inc(p, se.nr_wakeups_migrate);
1983 if (cpu == this_cpu)
1984 schedstat_inc(p, se.nr_wakeups_local);
1986 schedstat_inc(p, se.nr_wakeups_remote);
1987 update_rq_clock(rq);
1988 activate_task(rq, p, 1);
1992 check_preempt_curr(rq, p);
1994 p->state = TASK_RUNNING;
1996 if (p->sched_class->task_wake_up)
1997 p->sched_class->task_wake_up(rq, p);
2000 task_rq_unlock(rq, &flags);
2005 int wake_up_process(struct task_struct *p)
2007 return try_to_wake_up(p, TASK_ALL, 0);
2009 EXPORT_SYMBOL(wake_up_process);
2011 int wake_up_state(struct task_struct *p, unsigned int state)
2013 return try_to_wake_up(p, state, 0);
2017 * Perform scheduler related setup for a newly forked process p.
2018 * p is forked by current.
2020 * __sched_fork() is basic setup used by init_idle() too:
2022 static void __sched_fork(struct task_struct *p)
2024 p->se.exec_start = 0;
2025 p->se.sum_exec_runtime = 0;
2026 p->se.prev_sum_exec_runtime = 0;
2027 p->se.last_wakeup = 0;
2028 p->se.avg_overlap = 0;
2030 #ifdef CONFIG_SCHEDSTATS
2031 p->se.wait_start = 0;
2032 p->se.sum_sleep_runtime = 0;
2033 p->se.sleep_start = 0;
2034 p->se.block_start = 0;
2035 p->se.sleep_max = 0;
2036 p->se.block_max = 0;
2038 p->se.slice_max = 0;
2042 INIT_LIST_HEAD(&p->rt.run_list);
2045 #ifdef CONFIG_PREEMPT_NOTIFIERS
2046 INIT_HLIST_HEAD(&p->preempt_notifiers);
2050 * We mark the process as running here, but have not actually
2051 * inserted it onto the runqueue yet. This guarantees that
2052 * nobody will actually run it, and a signal or other external
2053 * event cannot wake it up and insert it on the runqueue either.
2055 p->state = TASK_RUNNING;
2059 * fork()/clone()-time setup:
2061 void sched_fork(struct task_struct *p, int clone_flags)
2063 int cpu = get_cpu();
2068 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2070 set_task_cpu(p, cpu);
2073 * Make sure we do not leak PI boosting priority to the child:
2075 p->prio = current->normal_prio;
2076 if (!rt_prio(p->prio))
2077 p->sched_class = &fair_sched_class;
2079 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2080 if (likely(sched_info_on()))
2081 memset(&p->sched_info, 0, sizeof(p->sched_info));
2083 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2086 #ifdef CONFIG_PREEMPT
2087 /* Want to start with kernel preemption disabled. */
2088 task_thread_info(p)->preempt_count = 1;
2094 * wake_up_new_task - wake up a newly created task for the first time.
2096 * This function will do some initial scheduler statistics housekeeping
2097 * that must be done for every newly created context, then puts the task
2098 * on the runqueue and wakes it.
2100 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2102 unsigned long flags;
2105 rq = task_rq_lock(p, &flags);
2106 BUG_ON(p->state != TASK_RUNNING);
2107 update_rq_clock(rq);
2109 p->prio = effective_prio(p);
2111 if (!p->sched_class->task_new || !current->se.on_rq) {
2112 activate_task(rq, p, 0);
2115 * Let the scheduling class do new task startup
2116 * management (if any):
2118 p->sched_class->task_new(rq, p);
2119 inc_nr_running(p, rq);
2121 check_preempt_curr(rq, p);
2123 if (p->sched_class->task_wake_up)
2124 p->sched_class->task_wake_up(rq, p);
2126 task_rq_unlock(rq, &flags);
2129 #ifdef CONFIG_PREEMPT_NOTIFIERS
2132 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2133 * @notifier: notifier struct to register
2135 void preempt_notifier_register(struct preempt_notifier *notifier)
2137 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2139 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2142 * preempt_notifier_unregister - no longer interested in preemption notifications
2143 * @notifier: notifier struct to unregister
2145 * This is safe to call from within a preemption notifier.
2147 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2149 hlist_del(¬ifier->link);
2151 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2153 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2155 struct preempt_notifier *notifier;
2156 struct hlist_node *node;
2158 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2159 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2163 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2164 struct task_struct *next)
2166 struct preempt_notifier *notifier;
2167 struct hlist_node *node;
2169 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2170 notifier->ops->sched_out(notifier, next);
2175 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2180 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2181 struct task_struct *next)
2188 * prepare_task_switch - prepare to switch tasks
2189 * @rq: the runqueue preparing to switch
2190 * @prev: the current task that is being switched out
2191 * @next: the task we are going to switch to.
2193 * This is called with the rq lock held and interrupts off. It must
2194 * be paired with a subsequent finish_task_switch after the context
2197 * prepare_task_switch sets up locking and calls architecture specific
2201 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2202 struct task_struct *next)
2204 fire_sched_out_preempt_notifiers(prev, next);
2205 prepare_lock_switch(rq, next);
2206 prepare_arch_switch(next);
2210 * finish_task_switch - clean up after a task-switch
2211 * @rq: runqueue associated with task-switch
2212 * @prev: the thread we just switched away from.
2214 * finish_task_switch must be called after the context switch, paired
2215 * with a prepare_task_switch call before the context switch.
2216 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2217 * and do any other architecture-specific cleanup actions.
2219 * Note that we may have delayed dropping an mm in context_switch(). If
2220 * so, we finish that here outside of the runqueue lock. (Doing it
2221 * with the lock held can cause deadlocks; see schedule() for
2224 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2225 __releases(rq->lock)
2227 struct mm_struct *mm = rq->prev_mm;
2233 * A task struct has one reference for the use as "current".
2234 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2235 * schedule one last time. The schedule call will never return, and
2236 * the scheduled task must drop that reference.
2237 * The test for TASK_DEAD must occur while the runqueue locks are
2238 * still held, otherwise prev could be scheduled on another cpu, die
2239 * there before we look at prev->state, and then the reference would
2241 * Manfred Spraul <manfred@colorfullife.com>
2243 prev_state = prev->state;
2244 finish_arch_switch(prev);
2245 finish_lock_switch(rq, prev);
2247 if (current->sched_class->post_schedule)
2248 current->sched_class->post_schedule(rq);
2251 fire_sched_in_preempt_notifiers(current);
2254 if (unlikely(prev_state == TASK_DEAD)) {
2256 * Remove function-return probe instances associated with this
2257 * task and put them back on the free list.
2259 kprobe_flush_task(prev);
2260 put_task_struct(prev);
2265 * schedule_tail - first thing a freshly forked thread must call.
2266 * @prev: the thread we just switched away from.
2268 asmlinkage void schedule_tail(struct task_struct *prev)
2269 __releases(rq->lock)
2271 struct rq *rq = this_rq();
2273 finish_task_switch(rq, prev);
2274 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2275 /* In this case, finish_task_switch does not reenable preemption */
2278 if (current->set_child_tid)
2279 put_user(task_pid_vnr(current), current->set_child_tid);
2283 * context_switch - switch to the new MM and the new
2284 * thread's register state.
2287 context_switch(struct rq *rq, struct task_struct *prev,
2288 struct task_struct *next)
2290 struct mm_struct *mm, *oldmm;
2292 prepare_task_switch(rq, prev, next);
2294 oldmm = prev->active_mm;
2296 * For paravirt, this is coupled with an exit in switch_to to
2297 * combine the page table reload and the switch backend into
2300 arch_enter_lazy_cpu_mode();
2302 if (unlikely(!mm)) {
2303 next->active_mm = oldmm;
2304 atomic_inc(&oldmm->mm_count);
2305 enter_lazy_tlb(oldmm, next);
2307 switch_mm(oldmm, mm, next);
2309 if (unlikely(!prev->mm)) {
2310 prev->active_mm = NULL;
2311 rq->prev_mm = oldmm;
2314 * Since the runqueue lock will be released by the next
2315 * task (which is an invalid locking op but in the case
2316 * of the scheduler it's an obvious special-case), so we
2317 * do an early lockdep release here:
2319 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2320 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2323 /* Here we just switch the register state and the stack. */
2324 switch_to(prev, next, prev);
2328 * this_rq must be evaluated again because prev may have moved
2329 * CPUs since it called schedule(), thus the 'rq' on its stack
2330 * frame will be invalid.
2332 finish_task_switch(this_rq(), prev);
2336 * nr_running, nr_uninterruptible and nr_context_switches:
2338 * externally visible scheduler statistics: current number of runnable
2339 * threads, current number of uninterruptible-sleeping threads, total
2340 * number of context switches performed since bootup.
2342 unsigned long nr_running(void)
2344 unsigned long i, sum = 0;
2346 for_each_online_cpu(i)
2347 sum += cpu_rq(i)->nr_running;
2352 unsigned long nr_uninterruptible(void)
2354 unsigned long i, sum = 0;
2356 for_each_possible_cpu(i)
2357 sum += cpu_rq(i)->nr_uninterruptible;
2360 * Since we read the counters lockless, it might be slightly
2361 * inaccurate. Do not allow it to go below zero though:
2363 if (unlikely((long)sum < 0))
2369 unsigned long long nr_context_switches(void)
2372 unsigned long long sum = 0;
2374 for_each_possible_cpu(i)
2375 sum += cpu_rq(i)->nr_switches;
2380 unsigned long nr_iowait(void)
2382 unsigned long i, sum = 0;
2384 for_each_possible_cpu(i)
2385 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2390 unsigned long nr_active(void)
2392 unsigned long i, running = 0, uninterruptible = 0;
2394 for_each_online_cpu(i) {
2395 running += cpu_rq(i)->nr_running;
2396 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2399 if (unlikely((long)uninterruptible < 0))
2400 uninterruptible = 0;
2402 return running + uninterruptible;
2406 * Update rq->cpu_load[] statistics. This function is usually called every
2407 * scheduler tick (TICK_NSEC).
2409 static void update_cpu_load(struct rq *this_rq)
2411 unsigned long this_load = this_rq->load.weight;
2414 this_rq->nr_load_updates++;
2416 /* Update our load: */
2417 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2418 unsigned long old_load, new_load;
2420 /* scale is effectively 1 << i now, and >> i divides by scale */
2422 old_load = this_rq->cpu_load[i];
2423 new_load = this_load;
2425 * Round up the averaging division if load is increasing. This
2426 * prevents us from getting stuck on 9 if the load is 10, for
2429 if (new_load > old_load)
2430 new_load += scale-1;
2431 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2438 * double_rq_lock - safely lock two runqueues
2440 * Note this does not disable interrupts like task_rq_lock,
2441 * you need to do so manually before calling.
2443 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2444 __acquires(rq1->lock)
2445 __acquires(rq2->lock)
2447 BUG_ON(!irqs_disabled());
2449 spin_lock(&rq1->lock);
2450 __acquire(rq2->lock); /* Fake it out ;) */
2453 spin_lock(&rq1->lock);
2454 spin_lock(&rq2->lock);
2456 spin_lock(&rq2->lock);
2457 spin_lock(&rq1->lock);
2460 update_rq_clock(rq1);
2461 update_rq_clock(rq2);
2465 * double_rq_unlock - safely unlock two runqueues
2467 * Note this does not restore interrupts like task_rq_unlock,
2468 * you need to do so manually after calling.
2470 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2471 __releases(rq1->lock)
2472 __releases(rq2->lock)
2474 spin_unlock(&rq1->lock);
2476 spin_unlock(&rq2->lock);
2478 __release(rq2->lock);
2482 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2484 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2485 __releases(this_rq->lock)
2486 __acquires(busiest->lock)
2487 __acquires(this_rq->lock)
2491 if (unlikely(!irqs_disabled())) {
2492 /* printk() doesn't work good under rq->lock */
2493 spin_unlock(&this_rq->lock);
2496 if (unlikely(!spin_trylock(&busiest->lock))) {
2497 if (busiest < this_rq) {
2498 spin_unlock(&this_rq->lock);
2499 spin_lock(&busiest->lock);
2500 spin_lock(&this_rq->lock);
2503 spin_lock(&busiest->lock);
2509 * If dest_cpu is allowed for this process, migrate the task to it.
2510 * This is accomplished by forcing the cpu_allowed mask to only
2511 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2512 * the cpu_allowed mask is restored.
2514 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2516 struct migration_req req;
2517 unsigned long flags;
2520 rq = task_rq_lock(p, &flags);
2521 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2522 || unlikely(cpu_is_offline(dest_cpu)))
2525 /* force the process onto the specified CPU */
2526 if (migrate_task(p, dest_cpu, &req)) {
2527 /* Need to wait for migration thread (might exit: take ref). */
2528 struct task_struct *mt = rq->migration_thread;
2530 get_task_struct(mt);
2531 task_rq_unlock(rq, &flags);
2532 wake_up_process(mt);
2533 put_task_struct(mt);
2534 wait_for_completion(&req.done);
2539 task_rq_unlock(rq, &flags);
2543 * sched_exec - execve() is a valuable balancing opportunity, because at
2544 * this point the task has the smallest effective memory and cache footprint.
2546 void sched_exec(void)
2548 int new_cpu, this_cpu = get_cpu();
2549 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2551 if (new_cpu != this_cpu)
2552 sched_migrate_task(current, new_cpu);
2556 * pull_task - move a task from a remote runqueue to the local runqueue.
2557 * Both runqueues must be locked.
2559 static void pull_task(struct rq *src_rq, struct task_struct *p,
2560 struct rq *this_rq, int this_cpu)
2562 deactivate_task(src_rq, p, 0);
2563 set_task_cpu(p, this_cpu);
2564 activate_task(this_rq, p, 0);
2566 * Note that idle threads have a prio of MAX_PRIO, for this test
2567 * to be always true for them.
2569 check_preempt_curr(this_rq, p);
2573 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2576 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2577 struct sched_domain *sd, enum cpu_idle_type idle,
2581 * We do not migrate tasks that are:
2582 * 1) running (obviously), or
2583 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2584 * 3) are cache-hot on their current CPU.
2586 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2587 schedstat_inc(p, se.nr_failed_migrations_affine);
2592 if (task_running(rq, p)) {
2593 schedstat_inc(p, se.nr_failed_migrations_running);
2598 * Aggressive migration if:
2599 * 1) task is cache cold, or
2600 * 2) too many balance attempts have failed.
2603 if (!task_hot(p, rq->clock, sd) ||
2604 sd->nr_balance_failed > sd->cache_nice_tries) {
2605 #ifdef CONFIG_SCHEDSTATS
2606 if (task_hot(p, rq->clock, sd)) {
2607 schedstat_inc(sd, lb_hot_gained[idle]);
2608 schedstat_inc(p, se.nr_forced_migrations);
2614 if (task_hot(p, rq->clock, sd)) {
2615 schedstat_inc(p, se.nr_failed_migrations_hot);
2621 static unsigned long
2622 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2623 unsigned long max_load_move, struct sched_domain *sd,
2624 enum cpu_idle_type idle, int *all_pinned,
2625 int *this_best_prio, struct rq_iterator *iterator)
2627 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2628 struct task_struct *p;
2629 long rem_load_move = max_load_move;
2631 if (max_load_move == 0)
2637 * Start the load-balancing iterator:
2639 p = iterator->start(iterator->arg);
2641 if (!p || loops++ > sysctl_sched_nr_migrate)
2644 * To help distribute high priority tasks across CPUs we don't
2645 * skip a task if it will be the highest priority task (i.e. smallest
2646 * prio value) on its new queue regardless of its load weight
2648 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2649 SCHED_LOAD_SCALE_FUZZ;
2650 if ((skip_for_load && p->prio >= *this_best_prio) ||
2651 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2652 p = iterator->next(iterator->arg);
2656 pull_task(busiest, p, this_rq, this_cpu);
2658 rem_load_move -= p->se.load.weight;
2661 * We only want to steal up to the prescribed amount of weighted load.
2663 if (rem_load_move > 0) {
2664 if (p->prio < *this_best_prio)
2665 *this_best_prio = p->prio;
2666 p = iterator->next(iterator->arg);
2671 * Right now, this is one of only two places pull_task() is called,
2672 * so we can safely collect pull_task() stats here rather than
2673 * inside pull_task().
2675 schedstat_add(sd, lb_gained[idle], pulled);
2678 *all_pinned = pinned;
2680 return max_load_move - rem_load_move;
2684 * move_tasks tries to move up to max_load_move weighted load from busiest to
2685 * this_rq, as part of a balancing operation within domain "sd".
2686 * Returns 1 if successful and 0 otherwise.
2688 * Called with both runqueues locked.
2690 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2691 unsigned long max_load_move,
2692 struct sched_domain *sd, enum cpu_idle_type idle,
2695 const struct sched_class *class = sched_class_highest;
2696 unsigned long total_load_moved = 0;
2697 int this_best_prio = this_rq->curr->prio;
2701 class->load_balance(this_rq, this_cpu, busiest,
2702 max_load_move - total_load_moved,
2703 sd, idle, all_pinned, &this_best_prio);
2704 class = class->next;
2705 } while (class && max_load_move > total_load_moved);
2707 return total_load_moved > 0;
2711 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2712 struct sched_domain *sd, enum cpu_idle_type idle,
2713 struct rq_iterator *iterator)
2715 struct task_struct *p = iterator->start(iterator->arg);
2719 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2720 pull_task(busiest, p, this_rq, this_cpu);
2722 * Right now, this is only the second place pull_task()
2723 * is called, so we can safely collect pull_task()
2724 * stats here rather than inside pull_task().
2726 schedstat_inc(sd, lb_gained[idle]);
2730 p = iterator->next(iterator->arg);
2737 * move_one_task tries to move exactly one task from busiest to this_rq, as
2738 * part of active balancing operations within "domain".
2739 * Returns 1 if successful and 0 otherwise.
2741 * Called with both runqueues locked.
2743 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2744 struct sched_domain *sd, enum cpu_idle_type idle)
2746 const struct sched_class *class;
2748 for (class = sched_class_highest; class; class = class->next)
2749 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2756 * find_busiest_group finds and returns the busiest CPU group within the
2757 * domain. It calculates and returns the amount of weighted load which
2758 * should be moved to restore balance via the imbalance parameter.
2760 static struct sched_group *
2761 find_busiest_group(struct sched_domain *sd, int this_cpu,
2762 unsigned long *imbalance, enum cpu_idle_type idle,
2763 int *sd_idle, cpumask_t *cpus, int *balance)
2765 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2766 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2767 unsigned long max_pull;
2768 unsigned long busiest_load_per_task, busiest_nr_running;
2769 unsigned long this_load_per_task, this_nr_running;
2770 int load_idx, group_imb = 0;
2771 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2772 int power_savings_balance = 1;
2773 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2774 unsigned long min_nr_running = ULONG_MAX;
2775 struct sched_group *group_min = NULL, *group_leader = NULL;
2778 max_load = this_load = total_load = total_pwr = 0;
2779 busiest_load_per_task = busiest_nr_running = 0;
2780 this_load_per_task = this_nr_running = 0;
2781 if (idle == CPU_NOT_IDLE)
2782 load_idx = sd->busy_idx;
2783 else if (idle == CPU_NEWLY_IDLE)
2784 load_idx = sd->newidle_idx;
2786 load_idx = sd->idle_idx;
2789 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2792 int __group_imb = 0;
2793 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2794 unsigned long sum_nr_running, sum_weighted_load;
2796 local_group = cpu_isset(this_cpu, group->cpumask);
2799 balance_cpu = first_cpu(group->cpumask);
2801 /* Tally up the load of all CPUs in the group */
2802 sum_weighted_load = sum_nr_running = avg_load = 0;
2804 min_cpu_load = ~0UL;
2806 for_each_cpu_mask(i, group->cpumask) {
2809 if (!cpu_isset(i, *cpus))
2814 if (*sd_idle && rq->nr_running)
2817 /* Bias balancing toward cpus of our domain */
2819 if (idle_cpu(i) && !first_idle_cpu) {
2824 load = target_load(i, load_idx);
2826 load = source_load(i, load_idx);
2827 if (load > max_cpu_load)
2828 max_cpu_load = load;
2829 if (min_cpu_load > load)
2830 min_cpu_load = load;
2834 sum_nr_running += rq->nr_running;
2835 sum_weighted_load += weighted_cpuload(i);
2839 * First idle cpu or the first cpu(busiest) in this sched group
2840 * is eligible for doing load balancing at this and above
2841 * domains. In the newly idle case, we will allow all the cpu's
2842 * to do the newly idle load balance.
2844 if (idle != CPU_NEWLY_IDLE && local_group &&
2845 balance_cpu != this_cpu && balance) {
2850 total_load += avg_load;
2851 total_pwr += group->__cpu_power;
2853 /* Adjust by relative CPU power of the group */
2854 avg_load = sg_div_cpu_power(group,
2855 avg_load * SCHED_LOAD_SCALE);
2857 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2860 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2863 this_load = avg_load;
2865 this_nr_running = sum_nr_running;
2866 this_load_per_task = sum_weighted_load;
2867 } else if (avg_load > max_load &&
2868 (sum_nr_running > group_capacity || __group_imb)) {
2869 max_load = avg_load;
2871 busiest_nr_running = sum_nr_running;
2872 busiest_load_per_task = sum_weighted_load;
2873 group_imb = __group_imb;
2876 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2878 * Busy processors will not participate in power savings
2881 if (idle == CPU_NOT_IDLE ||
2882 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2886 * If the local group is idle or completely loaded
2887 * no need to do power savings balance at this domain
2889 if (local_group && (this_nr_running >= group_capacity ||
2891 power_savings_balance = 0;
2894 * If a group is already running at full capacity or idle,
2895 * don't include that group in power savings calculations
2897 if (!power_savings_balance || sum_nr_running >= group_capacity
2902 * Calculate the group which has the least non-idle load.
2903 * This is the group from where we need to pick up the load
2906 if ((sum_nr_running < min_nr_running) ||
2907 (sum_nr_running == min_nr_running &&
2908 first_cpu(group->cpumask) <
2909 first_cpu(group_min->cpumask))) {
2911 min_nr_running = sum_nr_running;
2912 min_load_per_task = sum_weighted_load /
2917 * Calculate the group which is almost near its
2918 * capacity but still has some space to pick up some load
2919 * from other group and save more power
2921 if (sum_nr_running <= group_capacity - 1) {
2922 if (sum_nr_running > leader_nr_running ||
2923 (sum_nr_running == leader_nr_running &&
2924 first_cpu(group->cpumask) >
2925 first_cpu(group_leader->cpumask))) {
2926 group_leader = group;
2927 leader_nr_running = sum_nr_running;
2932 group = group->next;
2933 } while (group != sd->groups);
2935 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2938 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2940 if (this_load >= avg_load ||
2941 100*max_load <= sd->imbalance_pct*this_load)
2944 busiest_load_per_task /= busiest_nr_running;
2946 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2949 * We're trying to get all the cpus to the average_load, so we don't
2950 * want to push ourselves above the average load, nor do we wish to
2951 * reduce the max loaded cpu below the average load, as either of these
2952 * actions would just result in more rebalancing later, and ping-pong
2953 * tasks around. Thus we look for the minimum possible imbalance.
2954 * Negative imbalances (*we* are more loaded than anyone else) will
2955 * be counted as no imbalance for these purposes -- we can't fix that
2956 * by pulling tasks to us. Be careful of negative numbers as they'll
2957 * appear as very large values with unsigned longs.
2959 if (max_load <= busiest_load_per_task)
2963 * In the presence of smp nice balancing, certain scenarios can have
2964 * max load less than avg load(as we skip the groups at or below
2965 * its cpu_power, while calculating max_load..)
2967 if (max_load < avg_load) {
2969 goto small_imbalance;
2972 /* Don't want to pull so many tasks that a group would go idle */
2973 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2975 /* How much load to actually move to equalise the imbalance */
2976 *imbalance = min(max_pull * busiest->__cpu_power,
2977 (avg_load - this_load) * this->__cpu_power)
2981 * if *imbalance is less than the average load per runnable task
2982 * there is no gaurantee that any tasks will be moved so we'll have
2983 * a think about bumping its value to force at least one task to be
2986 if (*imbalance < busiest_load_per_task) {
2987 unsigned long tmp, pwr_now, pwr_move;
2991 pwr_move = pwr_now = 0;
2993 if (this_nr_running) {
2994 this_load_per_task /= this_nr_running;
2995 if (busiest_load_per_task > this_load_per_task)
2998 this_load_per_task = SCHED_LOAD_SCALE;
3000 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3001 busiest_load_per_task * imbn) {
3002 *imbalance = busiest_load_per_task;
3007 * OK, we don't have enough imbalance to justify moving tasks,
3008 * however we may be able to increase total CPU power used by
3012 pwr_now += busiest->__cpu_power *
3013 min(busiest_load_per_task, max_load);
3014 pwr_now += this->__cpu_power *
3015 min(this_load_per_task, this_load);
3016 pwr_now /= SCHED_LOAD_SCALE;
3018 /* Amount of load we'd subtract */
3019 tmp = sg_div_cpu_power(busiest,
3020 busiest_load_per_task * SCHED_LOAD_SCALE);
3022 pwr_move += busiest->__cpu_power *
3023 min(busiest_load_per_task, max_load - tmp);
3025 /* Amount of load we'd add */
3026 if (max_load * busiest->__cpu_power <
3027 busiest_load_per_task * SCHED_LOAD_SCALE)
3028 tmp = sg_div_cpu_power(this,
3029 max_load * busiest->__cpu_power);
3031 tmp = sg_div_cpu_power(this,
3032 busiest_load_per_task * SCHED_LOAD_SCALE);
3033 pwr_move += this->__cpu_power *
3034 min(this_load_per_task, this_load + tmp);
3035 pwr_move /= SCHED_LOAD_SCALE;
3037 /* Move if we gain throughput */
3038 if (pwr_move > pwr_now)
3039 *imbalance = busiest_load_per_task;
3045 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3046 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3049 if (this == group_leader && group_leader != group_min) {
3050 *imbalance = min_load_per_task;
3060 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3063 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3064 unsigned long imbalance, cpumask_t *cpus)
3066 struct rq *busiest = NULL, *rq;
3067 unsigned long max_load = 0;
3070 for_each_cpu_mask(i, group->cpumask) {
3073 if (!cpu_isset(i, *cpus))
3077 wl = weighted_cpuload(i);
3079 if (rq->nr_running == 1 && wl > imbalance)
3082 if (wl > max_load) {
3092 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3093 * so long as it is large enough.
3095 #define MAX_PINNED_INTERVAL 512
3098 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3099 * tasks if there is an imbalance.
3101 static int load_balance(int this_cpu, struct rq *this_rq,
3102 struct sched_domain *sd, enum cpu_idle_type idle,
3105 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3106 struct sched_group *group;
3107 unsigned long imbalance;
3109 cpumask_t cpus = CPU_MASK_ALL;
3110 unsigned long flags;
3113 * When power savings policy is enabled for the parent domain, idle
3114 * sibling can pick up load irrespective of busy siblings. In this case,
3115 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3116 * portraying it as CPU_NOT_IDLE.
3118 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3119 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3122 schedstat_inc(sd, lb_count[idle]);
3125 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3132 schedstat_inc(sd, lb_nobusyg[idle]);
3136 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3138 schedstat_inc(sd, lb_nobusyq[idle]);
3142 BUG_ON(busiest == this_rq);
3144 schedstat_add(sd, lb_imbalance[idle], imbalance);
3147 if (busiest->nr_running > 1) {
3149 * Attempt to move tasks. If find_busiest_group has found
3150 * an imbalance but busiest->nr_running <= 1, the group is
3151 * still unbalanced. ld_moved simply stays zero, so it is
3152 * correctly treated as an imbalance.
3154 local_irq_save(flags);
3155 double_rq_lock(this_rq, busiest);
3156 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3157 imbalance, sd, idle, &all_pinned);
3158 double_rq_unlock(this_rq, busiest);
3159 local_irq_restore(flags);
3162 * some other cpu did the load balance for us.
3164 if (ld_moved && this_cpu != smp_processor_id())
3165 resched_cpu(this_cpu);
3167 /* All tasks on this runqueue were pinned by CPU affinity */
3168 if (unlikely(all_pinned)) {
3169 cpu_clear(cpu_of(busiest), cpus);
3170 if (!cpus_empty(cpus))
3177 schedstat_inc(sd, lb_failed[idle]);
3178 sd->nr_balance_failed++;
3180 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3182 spin_lock_irqsave(&busiest->lock, flags);
3184 /* don't kick the migration_thread, if the curr
3185 * task on busiest cpu can't be moved to this_cpu
3187 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3188 spin_unlock_irqrestore(&busiest->lock, flags);
3190 goto out_one_pinned;
3193 if (!busiest->active_balance) {
3194 busiest->active_balance = 1;
3195 busiest->push_cpu = this_cpu;
3198 spin_unlock_irqrestore(&busiest->lock, flags);
3200 wake_up_process(busiest->migration_thread);
3203 * We've kicked active balancing, reset the failure
3206 sd->nr_balance_failed = sd->cache_nice_tries+1;
3209 sd->nr_balance_failed = 0;
3211 if (likely(!active_balance)) {
3212 /* We were unbalanced, so reset the balancing interval */
3213 sd->balance_interval = sd->min_interval;
3216 * If we've begun active balancing, start to back off. This
3217 * case may not be covered by the all_pinned logic if there
3218 * is only 1 task on the busy runqueue (because we don't call
3221 if (sd->balance_interval < sd->max_interval)
3222 sd->balance_interval *= 2;
3225 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3226 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3231 schedstat_inc(sd, lb_balanced[idle]);
3233 sd->nr_balance_failed = 0;
3236 /* tune up the balancing interval */
3237 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3238 (sd->balance_interval < sd->max_interval))
3239 sd->balance_interval *= 2;
3241 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3242 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3248 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3249 * tasks if there is an imbalance.
3251 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3252 * this_rq is locked.
3255 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3257 struct sched_group *group;
3258 struct rq *busiest = NULL;
3259 unsigned long imbalance;
3263 cpumask_t cpus = CPU_MASK_ALL;
3266 * When power savings policy is enabled for the parent domain, idle
3267 * sibling can pick up load irrespective of busy siblings. In this case,
3268 * let the state of idle sibling percolate up as IDLE, instead of
3269 * portraying it as CPU_NOT_IDLE.
3271 if (sd->flags & SD_SHARE_CPUPOWER &&
3272 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3275 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3277 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3278 &sd_idle, &cpus, NULL);
3280 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3284 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3287 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3291 BUG_ON(busiest == this_rq);
3293 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3296 if (busiest->nr_running > 1) {
3297 /* Attempt to move tasks */
3298 double_lock_balance(this_rq, busiest);
3299 /* this_rq->clock is already updated */
3300 update_rq_clock(busiest);
3301 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3302 imbalance, sd, CPU_NEWLY_IDLE,
3304 spin_unlock(&busiest->lock);
3306 if (unlikely(all_pinned)) {
3307 cpu_clear(cpu_of(busiest), cpus);
3308 if (!cpus_empty(cpus))
3314 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3315 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3316 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3319 sd->nr_balance_failed = 0;
3324 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3325 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3326 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3328 sd->nr_balance_failed = 0;
3334 * idle_balance is called by schedule() if this_cpu is about to become
3335 * idle. Attempts to pull tasks from other CPUs.
3337 static void idle_balance(int this_cpu, struct rq *this_rq)
3339 struct sched_domain *sd;
3340 int pulled_task = -1;
3341 unsigned long next_balance = jiffies + HZ;
3343 for_each_domain(this_cpu, sd) {
3344 unsigned long interval;
3346 if (!(sd->flags & SD_LOAD_BALANCE))
3349 if (sd->flags & SD_BALANCE_NEWIDLE)
3350 /* If we've pulled tasks over stop searching: */
3351 pulled_task = load_balance_newidle(this_cpu,
3354 interval = msecs_to_jiffies(sd->balance_interval);
3355 if (time_after(next_balance, sd->last_balance + interval))
3356 next_balance = sd->last_balance + interval;
3360 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3362 * We are going idle. next_balance may be set based on
3363 * a busy processor. So reset next_balance.
3365 this_rq->next_balance = next_balance;
3370 * active_load_balance is run by migration threads. It pushes running tasks
3371 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3372 * running on each physical CPU where possible, and avoids physical /
3373 * logical imbalances.
3375 * Called with busiest_rq locked.
3377 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3379 int target_cpu = busiest_rq->push_cpu;
3380 struct sched_domain *sd;
3381 struct rq *target_rq;
3383 /* Is there any task to move? */
3384 if (busiest_rq->nr_running <= 1)
3387 target_rq = cpu_rq(target_cpu);
3390 * This condition is "impossible", if it occurs
3391 * we need to fix it. Originally reported by
3392 * Bjorn Helgaas on a 128-cpu setup.
3394 BUG_ON(busiest_rq == target_rq);
3396 /* move a task from busiest_rq to target_rq */
3397 double_lock_balance(busiest_rq, target_rq);
3398 update_rq_clock(busiest_rq);
3399 update_rq_clock(target_rq);
3401 /* Search for an sd spanning us and the target CPU. */
3402 for_each_domain(target_cpu, sd) {
3403 if ((sd->flags & SD_LOAD_BALANCE) &&
3404 cpu_isset(busiest_cpu, sd->span))
3409 schedstat_inc(sd, alb_count);
3411 if (move_one_task(target_rq, target_cpu, busiest_rq,
3413 schedstat_inc(sd, alb_pushed);
3415 schedstat_inc(sd, alb_failed);
3417 spin_unlock(&target_rq->lock);
3422 atomic_t load_balancer;
3424 } nohz ____cacheline_aligned = {
3425 .load_balancer = ATOMIC_INIT(-1),
3426 .cpu_mask = CPU_MASK_NONE,
3430 * This routine will try to nominate the ilb (idle load balancing)
3431 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3432 * load balancing on behalf of all those cpus. If all the cpus in the system
3433 * go into this tickless mode, then there will be no ilb owner (as there is
3434 * no need for one) and all the cpus will sleep till the next wakeup event
3437 * For the ilb owner, tick is not stopped. And this tick will be used
3438 * for idle load balancing. ilb owner will still be part of
3441 * While stopping the tick, this cpu will become the ilb owner if there
3442 * is no other owner. And will be the owner till that cpu becomes busy
3443 * or if all cpus in the system stop their ticks at which point
3444 * there is no need for ilb owner.
3446 * When the ilb owner becomes busy, it nominates another owner, during the
3447 * next busy scheduler_tick()
3449 int select_nohz_load_balancer(int stop_tick)
3451 int cpu = smp_processor_id();
3454 cpu_set(cpu, nohz.cpu_mask);
3455 cpu_rq(cpu)->in_nohz_recently = 1;
3458 * If we are going offline and still the leader, give up!
3460 if (cpu_is_offline(cpu) &&
3461 atomic_read(&nohz.load_balancer) == cpu) {
3462 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3467 /* time for ilb owner also to sleep */
3468 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3469 if (atomic_read(&nohz.load_balancer) == cpu)
3470 atomic_set(&nohz.load_balancer, -1);
3474 if (atomic_read(&nohz.load_balancer) == -1) {
3475 /* make me the ilb owner */
3476 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3478 } else if (atomic_read(&nohz.load_balancer) == cpu)
3481 if (!cpu_isset(cpu, nohz.cpu_mask))
3484 cpu_clear(cpu, nohz.cpu_mask);
3486 if (atomic_read(&nohz.load_balancer) == cpu)
3487 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3494 static DEFINE_SPINLOCK(balancing);
3497 * It checks each scheduling domain to see if it is due to be balanced,
3498 * and initiates a balancing operation if so.
3500 * Balancing parameters are set up in arch_init_sched_domains.
3502 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3505 struct rq *rq = cpu_rq(cpu);
3506 unsigned long interval;
3507 struct sched_domain *sd;
3508 /* Earliest time when we have to do rebalance again */
3509 unsigned long next_balance = jiffies + 60*HZ;
3510 int update_next_balance = 0;
3512 for_each_domain(cpu, sd) {
3513 if (!(sd->flags & SD_LOAD_BALANCE))
3516 interval = sd->balance_interval;
3517 if (idle != CPU_IDLE)
3518 interval *= sd->busy_factor;
3520 /* scale ms to jiffies */
3521 interval = msecs_to_jiffies(interval);
3522 if (unlikely(!interval))
3524 if (interval > HZ*NR_CPUS/10)
3525 interval = HZ*NR_CPUS/10;
3528 if (sd->flags & SD_SERIALIZE) {
3529 if (!spin_trylock(&balancing))
3533 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3534 if (load_balance(cpu, rq, sd, idle, &balance)) {
3536 * We've pulled tasks over so either we're no
3537 * longer idle, or one of our SMT siblings is
3540 idle = CPU_NOT_IDLE;
3542 sd->last_balance = jiffies;
3544 if (sd->flags & SD_SERIALIZE)
3545 spin_unlock(&balancing);
3547 if (time_after(next_balance, sd->last_balance + interval)) {
3548 next_balance = sd->last_balance + interval;
3549 update_next_balance = 1;
3553 * Stop the load balance at this level. There is another
3554 * CPU in our sched group which is doing load balancing more
3562 * next_balance will be updated only when there is a need.
3563 * When the cpu is attached to null domain for ex, it will not be
3566 if (likely(update_next_balance))
3567 rq->next_balance = next_balance;
3571 * run_rebalance_domains is triggered when needed from the scheduler tick.
3572 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3573 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3575 static void run_rebalance_domains(struct softirq_action *h)
3577 int this_cpu = smp_processor_id();
3578 struct rq *this_rq = cpu_rq(this_cpu);
3579 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3580 CPU_IDLE : CPU_NOT_IDLE;
3582 rebalance_domains(this_cpu, idle);
3586 * If this cpu is the owner for idle load balancing, then do the
3587 * balancing on behalf of the other idle cpus whose ticks are
3590 if (this_rq->idle_at_tick &&
3591 atomic_read(&nohz.load_balancer) == this_cpu) {
3592 cpumask_t cpus = nohz.cpu_mask;
3596 cpu_clear(this_cpu, cpus);
3597 for_each_cpu_mask(balance_cpu, cpus) {
3599 * If this cpu gets work to do, stop the load balancing
3600 * work being done for other cpus. Next load
3601 * balancing owner will pick it up.
3606 rebalance_domains(balance_cpu, CPU_IDLE);
3608 rq = cpu_rq(balance_cpu);
3609 if (time_after(this_rq->next_balance, rq->next_balance))
3610 this_rq->next_balance = rq->next_balance;
3617 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3619 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3620 * idle load balancing owner or decide to stop the periodic load balancing,
3621 * if the whole system is idle.
3623 static inline void trigger_load_balance(struct rq *rq, int cpu)
3627 * If we were in the nohz mode recently and busy at the current
3628 * scheduler tick, then check if we need to nominate new idle
3631 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3632 rq->in_nohz_recently = 0;
3634 if (atomic_read(&nohz.load_balancer) == cpu) {
3635 cpu_clear(cpu, nohz.cpu_mask);
3636 atomic_set(&nohz.load_balancer, -1);
3639 if (atomic_read(&nohz.load_balancer) == -1) {
3641 * simple selection for now: Nominate the
3642 * first cpu in the nohz list to be the next
3645 * TBD: Traverse the sched domains and nominate
3646 * the nearest cpu in the nohz.cpu_mask.
3648 int ilb = first_cpu(nohz.cpu_mask);
3656 * If this cpu is idle and doing idle load balancing for all the
3657 * cpus with ticks stopped, is it time for that to stop?
3659 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3660 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3666 * If this cpu is idle and the idle load balancing is done by
3667 * someone else, then no need raise the SCHED_SOFTIRQ
3669 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3670 cpu_isset(cpu, nohz.cpu_mask))
3673 if (time_after_eq(jiffies, rq->next_balance))
3674 raise_softirq(SCHED_SOFTIRQ);
3677 #else /* CONFIG_SMP */
3680 * on UP we do not need to balance between CPUs:
3682 static inline void idle_balance(int cpu, struct rq *rq)
3688 DEFINE_PER_CPU(struct kernel_stat, kstat);
3690 EXPORT_PER_CPU_SYMBOL(kstat);
3693 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3694 * that have not yet been banked in case the task is currently running.
3696 unsigned long long task_sched_runtime(struct task_struct *p)
3698 unsigned long flags;
3702 rq = task_rq_lock(p, &flags);
3703 ns = p->se.sum_exec_runtime;
3704 if (task_current(rq, p)) {
3705 update_rq_clock(rq);
3706 delta_exec = rq->clock - p->se.exec_start;
3707 if ((s64)delta_exec > 0)
3710 task_rq_unlock(rq, &flags);
3716 * Account user cpu time to a process.
3717 * @p: the process that the cpu time gets accounted to
3718 * @cputime: the cpu time spent in user space since the last update
3720 void account_user_time(struct task_struct *p, cputime_t cputime)
3722 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3725 p->utime = cputime_add(p->utime, cputime);
3727 /* Add user time to cpustat. */
3728 tmp = cputime_to_cputime64(cputime);
3729 if (TASK_NICE(p) > 0)
3730 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3732 cpustat->user = cputime64_add(cpustat->user, tmp);
3736 * Account guest cpu time to a process.
3737 * @p: the process that the cpu time gets accounted to
3738 * @cputime: the cpu time spent in virtual machine since the last update
3740 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3743 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3745 tmp = cputime_to_cputime64(cputime);
3747 p->utime = cputime_add(p->utime, cputime);
3748 p->gtime = cputime_add(p->gtime, cputime);
3750 cpustat->user = cputime64_add(cpustat->user, tmp);
3751 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3755 * Account scaled user cpu time to a process.
3756 * @p: the process that the cpu time gets accounted to
3757 * @cputime: the cpu time spent in user space since the last update
3759 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3761 p->utimescaled = cputime_add(p->utimescaled, cputime);
3765 * Account system cpu time to a process.
3766 * @p: the process that the cpu time gets accounted to
3767 * @hardirq_offset: the offset to subtract from hardirq_count()
3768 * @cputime: the cpu time spent in kernel space since the last update
3770 void account_system_time(struct task_struct *p, int hardirq_offset,
3773 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3774 struct rq *rq = this_rq();
3777 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3778 return account_guest_time(p, cputime);
3780 p->stime = cputime_add(p->stime, cputime);
3782 /* Add system time to cpustat. */
3783 tmp = cputime_to_cputime64(cputime);
3784 if (hardirq_count() - hardirq_offset)
3785 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3786 else if (softirq_count())
3787 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3788 else if (p != rq->idle)
3789 cpustat->system = cputime64_add(cpustat->system, tmp);
3790 else if (atomic_read(&rq->nr_iowait) > 0)
3791 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3793 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3794 /* Account for system time used */
3795 acct_update_integrals(p);
3799 * Account scaled system cpu time to a process.
3800 * @p: the process that the cpu time gets accounted to
3801 * @hardirq_offset: the offset to subtract from hardirq_count()
3802 * @cputime: the cpu time spent in kernel space since the last update
3804 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3806 p->stimescaled = cputime_add(p->stimescaled, cputime);
3810 * Account for involuntary wait time.
3811 * @p: the process from which the cpu time has been stolen
3812 * @steal: the cpu time spent in involuntary wait
3814 void account_steal_time(struct task_struct *p, cputime_t steal)
3816 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3817 cputime64_t tmp = cputime_to_cputime64(steal);
3818 struct rq *rq = this_rq();
3820 if (p == rq->idle) {
3821 p->stime = cputime_add(p->stime, steal);
3822 if (atomic_read(&rq->nr_iowait) > 0)
3823 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3825 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3827 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3831 * This function gets called by the timer code, with HZ frequency.
3832 * We call it with interrupts disabled.
3834 * It also gets called by the fork code, when changing the parent's
3837 void scheduler_tick(void)
3839 int cpu = smp_processor_id();
3840 struct rq *rq = cpu_rq(cpu);
3841 struct task_struct *curr = rq->curr;
3842 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3844 spin_lock(&rq->lock);
3845 __update_rq_clock(rq);
3847 * Let rq->clock advance by at least TICK_NSEC:
3849 if (unlikely(rq->clock < next_tick)) {
3850 rq->clock = next_tick;
3851 rq->clock_underflows++;
3853 rq->tick_timestamp = rq->clock;
3854 update_last_tick_seen(rq);
3855 update_cpu_load(rq);
3856 curr->sched_class->task_tick(rq, curr, 0);
3857 update_sched_rt_period(rq);
3858 spin_unlock(&rq->lock);
3861 rq->idle_at_tick = idle_cpu(cpu);
3862 trigger_load_balance(rq, cpu);
3866 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3868 void __kprobes add_preempt_count(int val)
3873 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3875 preempt_count() += val;
3877 * Spinlock count overflowing soon?
3879 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3882 EXPORT_SYMBOL(add_preempt_count);
3884 void __kprobes sub_preempt_count(int val)
3889 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3892 * Is the spinlock portion underflowing?
3894 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3895 !(preempt_count() & PREEMPT_MASK)))
3898 preempt_count() -= val;
3900 EXPORT_SYMBOL(sub_preempt_count);
3905 * Print scheduling while atomic bug:
3907 static noinline void __schedule_bug(struct task_struct *prev)
3909 struct pt_regs *regs = get_irq_regs();
3911 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3912 prev->comm, prev->pid, preempt_count());
3914 debug_show_held_locks(prev);
3915 if (irqs_disabled())
3916 print_irqtrace_events(prev);
3925 * Various schedule()-time debugging checks and statistics:
3927 static inline void schedule_debug(struct task_struct *prev)
3930 * Test if we are atomic. Since do_exit() needs to call into
3931 * schedule() atomically, we ignore that path for now.
3932 * Otherwise, whine if we are scheduling when we should not be.
3934 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3935 __schedule_bug(prev);
3937 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3939 schedstat_inc(this_rq(), sched_count);
3940 #ifdef CONFIG_SCHEDSTATS
3941 if (unlikely(prev->lock_depth >= 0)) {
3942 schedstat_inc(this_rq(), bkl_count);
3943 schedstat_inc(prev, sched_info.bkl_count);
3949 * Pick up the highest-prio task:
3951 static inline struct task_struct *
3952 pick_next_task(struct rq *rq, struct task_struct *prev)
3954 const struct sched_class *class;
3955 struct task_struct *p;
3958 * Optimization: we know that if all tasks are in
3959 * the fair class we can call that function directly:
3961 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3962 p = fair_sched_class.pick_next_task(rq);
3967 class = sched_class_highest;
3969 p = class->pick_next_task(rq);
3973 * Will never be NULL as the idle class always
3974 * returns a non-NULL p:
3976 class = class->next;
3981 * schedule() is the main scheduler function.
3983 asmlinkage void __sched schedule(void)
3985 struct task_struct *prev, *next;
3986 unsigned long *switch_count;
3992 cpu = smp_processor_id();
3996 switch_count = &prev->nivcsw;
3998 release_kernel_lock(prev);
3999 need_resched_nonpreemptible:
4001 schedule_debug(prev);
4006 * Do the rq-clock update outside the rq lock:
4008 local_irq_disable();
4009 __update_rq_clock(rq);
4010 spin_lock(&rq->lock);
4011 clear_tsk_need_resched(prev);
4013 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4014 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4015 signal_pending(prev))) {
4016 prev->state = TASK_RUNNING;
4018 deactivate_task(rq, prev, 1);
4020 switch_count = &prev->nvcsw;
4024 if (prev->sched_class->pre_schedule)
4025 prev->sched_class->pre_schedule(rq, prev);
4028 if (unlikely(!rq->nr_running))
4029 idle_balance(cpu, rq);
4031 prev->sched_class->put_prev_task(rq, prev);
4032 next = pick_next_task(rq, prev);
4034 sched_info_switch(prev, next);
4036 if (likely(prev != next)) {
4041 context_switch(rq, prev, next); /* unlocks the rq */
4043 * the context switch might have flipped the stack from under
4044 * us, hence refresh the local variables.
4046 cpu = smp_processor_id();
4049 spin_unlock_irq(&rq->lock);
4053 if (unlikely(reacquire_kernel_lock(current) < 0))
4054 goto need_resched_nonpreemptible;
4056 preempt_enable_no_resched();
4057 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4060 EXPORT_SYMBOL(schedule);
4062 #ifdef CONFIG_PREEMPT
4064 * this is the entry point to schedule() from in-kernel preemption
4065 * off of preempt_enable. Kernel preemptions off return from interrupt
4066 * occur there and call schedule directly.
4068 asmlinkage void __sched preempt_schedule(void)
4070 struct thread_info *ti = current_thread_info();
4071 struct task_struct *task = current;
4072 int saved_lock_depth;
4075 * If there is a non-zero preempt_count or interrupts are disabled,
4076 * we do not want to preempt the current task. Just return..
4078 if (likely(ti->preempt_count || irqs_disabled()))
4082 add_preempt_count(PREEMPT_ACTIVE);
4085 * We keep the big kernel semaphore locked, but we
4086 * clear ->lock_depth so that schedule() doesnt
4087 * auto-release the semaphore:
4089 saved_lock_depth = task->lock_depth;
4090 task->lock_depth = -1;
4092 task->lock_depth = saved_lock_depth;
4093 sub_preempt_count(PREEMPT_ACTIVE);
4096 * Check again in case we missed a preemption opportunity
4097 * between schedule and now.
4100 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4102 EXPORT_SYMBOL(preempt_schedule);
4105 * this is the entry point to schedule() from kernel preemption
4106 * off of irq context.
4107 * Note, that this is called and return with irqs disabled. This will
4108 * protect us against recursive calling from irq.
4110 asmlinkage void __sched preempt_schedule_irq(void)
4112 struct thread_info *ti = current_thread_info();
4113 struct task_struct *task = current;
4114 int saved_lock_depth;
4116 /* Catch callers which need to be fixed */
4117 BUG_ON(ti->preempt_count || !irqs_disabled());
4120 add_preempt_count(PREEMPT_ACTIVE);
4123 * We keep the big kernel semaphore locked, but we
4124 * clear ->lock_depth so that schedule() doesnt
4125 * auto-release the semaphore:
4127 saved_lock_depth = task->lock_depth;
4128 task->lock_depth = -1;
4131 local_irq_disable();
4132 task->lock_depth = saved_lock_depth;
4133 sub_preempt_count(PREEMPT_ACTIVE);
4136 * Check again in case we missed a preemption opportunity
4137 * between schedule and now.
4140 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4143 #endif /* CONFIG_PREEMPT */
4145 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4148 return try_to_wake_up(curr->private, mode, sync);
4150 EXPORT_SYMBOL(default_wake_function);
4153 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4154 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4155 * number) then we wake all the non-exclusive tasks and one exclusive task.
4157 * There are circumstances in which we can try to wake a task which has already
4158 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4159 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4161 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4162 int nr_exclusive, int sync, void *key)
4164 wait_queue_t *curr, *next;
4166 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4167 unsigned flags = curr->flags;
4169 if (curr->func(curr, mode, sync, key) &&
4170 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4176 * __wake_up - wake up threads blocked on a waitqueue.
4178 * @mode: which threads
4179 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4180 * @key: is directly passed to the wakeup function
4182 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4183 int nr_exclusive, void *key)
4185 unsigned long flags;
4187 spin_lock_irqsave(&q->lock, flags);
4188 __wake_up_common(q, mode, nr_exclusive, 0, key);
4189 spin_unlock_irqrestore(&q->lock, flags);
4191 EXPORT_SYMBOL(__wake_up);
4194 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4196 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4198 __wake_up_common(q, mode, 1, 0, NULL);
4202 * __wake_up_sync - wake up threads blocked on a waitqueue.
4204 * @mode: which threads
4205 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4207 * The sync wakeup differs that the waker knows that it will schedule
4208 * away soon, so while the target thread will be woken up, it will not
4209 * be migrated to another CPU - ie. the two threads are 'synchronized'
4210 * with each other. This can prevent needless bouncing between CPUs.
4212 * On UP it can prevent extra preemption.
4215 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4217 unsigned long flags;
4223 if (unlikely(!nr_exclusive))
4226 spin_lock_irqsave(&q->lock, flags);
4227 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4228 spin_unlock_irqrestore(&q->lock, flags);
4230 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4232 void complete(struct completion *x)
4234 unsigned long flags;
4236 spin_lock_irqsave(&x->wait.lock, flags);
4238 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4239 spin_unlock_irqrestore(&x->wait.lock, flags);
4241 EXPORT_SYMBOL(complete);
4243 void complete_all(struct completion *x)
4245 unsigned long flags;
4247 spin_lock_irqsave(&x->wait.lock, flags);
4248 x->done += UINT_MAX/2;
4249 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4250 spin_unlock_irqrestore(&x->wait.lock, flags);
4252 EXPORT_SYMBOL(complete_all);
4254 static inline long __sched
4255 do_wait_for_common(struct completion *x, long timeout, int state)
4258 DECLARE_WAITQUEUE(wait, current);
4260 wait.flags |= WQ_FLAG_EXCLUSIVE;
4261 __add_wait_queue_tail(&x->wait, &wait);
4263 if ((state == TASK_INTERRUPTIBLE &&
4264 signal_pending(current)) ||
4265 (state == TASK_KILLABLE &&
4266 fatal_signal_pending(current))) {
4267 __remove_wait_queue(&x->wait, &wait);
4268 return -ERESTARTSYS;
4270 __set_current_state(state);
4271 spin_unlock_irq(&x->wait.lock);
4272 timeout = schedule_timeout(timeout);
4273 spin_lock_irq(&x->wait.lock);
4275 __remove_wait_queue(&x->wait, &wait);
4279 __remove_wait_queue(&x->wait, &wait);
4286 wait_for_common(struct completion *x, long timeout, int state)
4290 spin_lock_irq(&x->wait.lock);
4291 timeout = do_wait_for_common(x, timeout, state);
4292 spin_unlock_irq(&x->wait.lock);
4296 void __sched wait_for_completion(struct completion *x)
4298 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4300 EXPORT_SYMBOL(wait_for_completion);
4302 unsigned long __sched
4303 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4305 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4307 EXPORT_SYMBOL(wait_for_completion_timeout);
4309 int __sched wait_for_completion_interruptible(struct completion *x)
4311 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4312 if (t == -ERESTARTSYS)
4316 EXPORT_SYMBOL(wait_for_completion_interruptible);
4318 unsigned long __sched
4319 wait_for_completion_interruptible_timeout(struct completion *x,
4320 unsigned long timeout)
4322 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4324 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4326 int __sched wait_for_completion_killable(struct completion *x)
4328 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4329 if (t == -ERESTARTSYS)
4333 EXPORT_SYMBOL(wait_for_completion_killable);
4336 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4338 unsigned long flags;
4341 init_waitqueue_entry(&wait, current);
4343 __set_current_state(state);
4345 spin_lock_irqsave(&q->lock, flags);
4346 __add_wait_queue(q, &wait);
4347 spin_unlock(&q->lock);
4348 timeout = schedule_timeout(timeout);
4349 spin_lock_irq(&q->lock);
4350 __remove_wait_queue(q, &wait);
4351 spin_unlock_irqrestore(&q->lock, flags);
4356 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4358 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4360 EXPORT_SYMBOL(interruptible_sleep_on);
4363 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4365 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4367 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4369 void __sched sleep_on(wait_queue_head_t *q)
4371 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4373 EXPORT_SYMBOL(sleep_on);
4375 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4377 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4379 EXPORT_SYMBOL(sleep_on_timeout);
4381 #ifdef CONFIG_RT_MUTEXES
4384 * rt_mutex_setprio - set the current priority of a task
4386 * @prio: prio value (kernel-internal form)
4388 * This function changes the 'effective' priority of a task. It does
4389 * not touch ->normal_prio like __setscheduler().
4391 * Used by the rt_mutex code to implement priority inheritance logic.
4393 void rt_mutex_setprio(struct task_struct *p, int prio)
4395 unsigned long flags;
4396 int oldprio, on_rq, running;
4398 const struct sched_class *prev_class = p->sched_class;
4400 BUG_ON(prio < 0 || prio > MAX_PRIO);
4402 rq = task_rq_lock(p, &flags);
4403 update_rq_clock(rq);
4406 on_rq = p->se.on_rq;
4407 running = task_current(rq, p);
4409 dequeue_task(rq, p, 0);
4411 p->sched_class->put_prev_task(rq, p);
4414 p->sched_class = &rt_sched_class;
4416 p->sched_class = &fair_sched_class;
4421 p->sched_class->set_curr_task(rq);
4423 enqueue_task(rq, p, 0);
4425 check_class_changed(rq, p, prev_class, oldprio, running);
4427 task_rq_unlock(rq, &flags);
4432 void set_user_nice(struct task_struct *p, long nice)
4434 int old_prio, delta, on_rq;
4435 unsigned long flags;
4438 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4441 * We have to be careful, if called from sys_setpriority(),
4442 * the task might be in the middle of scheduling on another CPU.
4444 rq = task_rq_lock(p, &flags);
4445 update_rq_clock(rq);
4447 * The RT priorities are set via sched_setscheduler(), but we still
4448 * allow the 'normal' nice value to be set - but as expected
4449 * it wont have any effect on scheduling until the task is
4450 * SCHED_FIFO/SCHED_RR:
4452 if (task_has_rt_policy(p)) {
4453 p->static_prio = NICE_TO_PRIO(nice);
4456 on_rq = p->se.on_rq;
4458 dequeue_task(rq, p, 0);
4462 p->static_prio = NICE_TO_PRIO(nice);
4465 p->prio = effective_prio(p);
4466 delta = p->prio - old_prio;
4469 enqueue_task(rq, p, 0);
4472 * If the task increased its priority or is running and
4473 * lowered its priority, then reschedule its CPU:
4475 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4476 resched_task(rq->curr);
4479 task_rq_unlock(rq, &flags);
4481 EXPORT_SYMBOL(set_user_nice);
4484 * can_nice - check if a task can reduce its nice value
4488 int can_nice(const struct task_struct *p, const int nice)
4490 /* convert nice value [19,-20] to rlimit style value [1,40] */
4491 int nice_rlim = 20 - nice;
4493 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4494 capable(CAP_SYS_NICE));
4497 #ifdef __ARCH_WANT_SYS_NICE
4500 * sys_nice - change the priority of the current process.
4501 * @increment: priority increment
4503 * sys_setpriority is a more generic, but much slower function that
4504 * does similar things.
4506 asmlinkage long sys_nice(int increment)
4511 * Setpriority might change our priority at the same moment.
4512 * We don't have to worry. Conceptually one call occurs first
4513 * and we have a single winner.
4515 if (increment < -40)
4520 nice = PRIO_TO_NICE(current->static_prio) + increment;
4526 if (increment < 0 && !can_nice(current, nice))
4529 retval = security_task_setnice(current, nice);
4533 set_user_nice(current, nice);
4540 * task_prio - return the priority value of a given task.
4541 * @p: the task in question.
4543 * This is the priority value as seen by users in /proc.
4544 * RT tasks are offset by -200. Normal tasks are centered
4545 * around 0, value goes from -16 to +15.
4547 int task_prio(const struct task_struct *p)
4549 return p->prio - MAX_RT_PRIO;
4553 * task_nice - return the nice value of a given task.
4554 * @p: the task in question.
4556 int task_nice(const struct task_struct *p)
4558 return TASK_NICE(p);
4560 EXPORT_SYMBOL(task_nice);
4563 * idle_cpu - is a given cpu idle currently?
4564 * @cpu: the processor in question.
4566 int idle_cpu(int cpu)
4568 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4572 * idle_task - return the idle task for a given cpu.
4573 * @cpu: the processor in question.
4575 struct task_struct *idle_task(int cpu)
4577 return cpu_rq(cpu)->idle;
4581 * find_process_by_pid - find a process with a matching PID value.
4582 * @pid: the pid in question.
4584 static struct task_struct *find_process_by_pid(pid_t pid)
4586 return pid ? find_task_by_vpid(pid) : current;
4589 /* Actually do priority change: must hold rq lock. */
4591 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4593 BUG_ON(p->se.on_rq);
4596 switch (p->policy) {
4600 p->sched_class = &fair_sched_class;
4604 p->sched_class = &rt_sched_class;
4608 p->rt_priority = prio;
4609 p->normal_prio = normal_prio(p);
4610 /* we are holding p->pi_lock already */
4611 p->prio = rt_mutex_getprio(p);
4616 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4617 * @p: the task in question.
4618 * @policy: new policy.
4619 * @param: structure containing the new RT priority.
4621 * NOTE that the task may be already dead.
4623 int sched_setscheduler(struct task_struct *p, int policy,
4624 struct sched_param *param)
4626 int retval, oldprio, oldpolicy = -1, on_rq, running;
4627 unsigned long flags;
4628 const struct sched_class *prev_class = p->sched_class;
4631 /* may grab non-irq protected spin_locks */
4632 BUG_ON(in_interrupt());
4634 /* double check policy once rq lock held */
4636 policy = oldpolicy = p->policy;
4637 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4638 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4639 policy != SCHED_IDLE)
4642 * Valid priorities for SCHED_FIFO and SCHED_RR are
4643 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4644 * SCHED_BATCH and SCHED_IDLE is 0.
4646 if (param->sched_priority < 0 ||
4647 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4648 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4650 if (rt_policy(policy) != (param->sched_priority != 0))
4654 * Allow unprivileged RT tasks to decrease priority:
4656 if (!capable(CAP_SYS_NICE)) {
4657 if (rt_policy(policy)) {
4658 unsigned long rlim_rtprio;
4660 if (!lock_task_sighand(p, &flags))
4662 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4663 unlock_task_sighand(p, &flags);
4665 /* can't set/change the rt policy */
4666 if (policy != p->policy && !rlim_rtprio)
4669 /* can't increase priority */
4670 if (param->sched_priority > p->rt_priority &&
4671 param->sched_priority > rlim_rtprio)
4675 * Like positive nice levels, dont allow tasks to
4676 * move out of SCHED_IDLE either:
4678 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4681 /* can't change other user's priorities */
4682 if ((current->euid != p->euid) &&
4683 (current->euid != p->uid))
4687 #ifdef CONFIG_RT_GROUP_SCHED
4689 * Do not allow realtime tasks into groups that have no runtime
4692 if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4696 retval = security_task_setscheduler(p, policy, param);
4700 * make sure no PI-waiters arrive (or leave) while we are
4701 * changing the priority of the task:
4703 spin_lock_irqsave(&p->pi_lock, flags);
4705 * To be able to change p->policy safely, the apropriate
4706 * runqueue lock must be held.
4708 rq = __task_rq_lock(p);
4709 /* recheck policy now with rq lock held */
4710 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4711 policy = oldpolicy = -1;
4712 __task_rq_unlock(rq);
4713 spin_unlock_irqrestore(&p->pi_lock, flags);
4716 update_rq_clock(rq);
4717 on_rq = p->se.on_rq;
4718 running = task_current(rq, p);
4720 deactivate_task(rq, p, 0);
4722 p->sched_class->put_prev_task(rq, p);
4725 __setscheduler(rq, p, policy, param->sched_priority);
4728 p->sched_class->set_curr_task(rq);
4730 activate_task(rq, p, 0);
4732 check_class_changed(rq, p, prev_class, oldprio, running);
4734 __task_rq_unlock(rq);
4735 spin_unlock_irqrestore(&p->pi_lock, flags);
4737 rt_mutex_adjust_pi(p);
4741 EXPORT_SYMBOL_GPL(sched_setscheduler);
4744 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4746 struct sched_param lparam;
4747 struct task_struct *p;
4750 if (!param || pid < 0)
4752 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4757 p = find_process_by_pid(pid);
4759 retval = sched_setscheduler(p, policy, &lparam);
4766 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4767 * @pid: the pid in question.
4768 * @policy: new policy.
4769 * @param: structure containing the new RT priority.
4772 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4774 /* negative values for policy are not valid */
4778 return do_sched_setscheduler(pid, policy, param);
4782 * sys_sched_setparam - set/change the RT priority of a thread
4783 * @pid: the pid in question.
4784 * @param: structure containing the new RT priority.
4786 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4788 return do_sched_setscheduler(pid, -1, param);
4792 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4793 * @pid: the pid in question.
4795 asmlinkage long sys_sched_getscheduler(pid_t pid)
4797 struct task_struct *p;
4804 read_lock(&tasklist_lock);
4805 p = find_process_by_pid(pid);
4807 retval = security_task_getscheduler(p);
4811 read_unlock(&tasklist_lock);
4816 * sys_sched_getscheduler - get the RT priority of a thread
4817 * @pid: the pid in question.
4818 * @param: structure containing the RT priority.
4820 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4822 struct sched_param lp;
4823 struct task_struct *p;
4826 if (!param || pid < 0)
4829 read_lock(&tasklist_lock);
4830 p = find_process_by_pid(pid);
4835 retval = security_task_getscheduler(p);
4839 lp.sched_priority = p->rt_priority;
4840 read_unlock(&tasklist_lock);
4843 * This one might sleep, we cannot do it with a spinlock held ...
4845 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4850 read_unlock(&tasklist_lock);
4854 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4856 cpumask_t cpus_allowed;
4857 struct task_struct *p;
4861 read_lock(&tasklist_lock);
4863 p = find_process_by_pid(pid);
4865 read_unlock(&tasklist_lock);
4871 * It is not safe to call set_cpus_allowed with the
4872 * tasklist_lock held. We will bump the task_struct's
4873 * usage count and then drop tasklist_lock.
4876 read_unlock(&tasklist_lock);
4879 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4880 !capable(CAP_SYS_NICE))
4883 retval = security_task_setscheduler(p, 0, NULL);
4887 cpus_allowed = cpuset_cpus_allowed(p);
4888 cpus_and(new_mask, new_mask, cpus_allowed);
4890 retval = set_cpus_allowed(p, new_mask);
4893 cpus_allowed = cpuset_cpus_allowed(p);
4894 if (!cpus_subset(new_mask, cpus_allowed)) {
4896 * We must have raced with a concurrent cpuset
4897 * update. Just reset the cpus_allowed to the
4898 * cpuset's cpus_allowed
4900 new_mask = cpus_allowed;
4910 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4911 cpumask_t *new_mask)
4913 if (len < sizeof(cpumask_t)) {
4914 memset(new_mask, 0, sizeof(cpumask_t));
4915 } else if (len > sizeof(cpumask_t)) {
4916 len = sizeof(cpumask_t);
4918 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4922 * sys_sched_setaffinity - set the cpu affinity of a process
4923 * @pid: pid of the process
4924 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4925 * @user_mask_ptr: user-space pointer to the new cpu mask
4927 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4928 unsigned long __user *user_mask_ptr)
4933 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4937 return sched_setaffinity(pid, new_mask);
4941 * Represents all cpu's present in the system
4942 * In systems capable of hotplug, this map could dynamically grow
4943 * as new cpu's are detected in the system via any platform specific
4944 * method, such as ACPI for e.g.
4947 cpumask_t cpu_present_map __read_mostly;
4948 EXPORT_SYMBOL(cpu_present_map);
4951 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4952 EXPORT_SYMBOL(cpu_online_map);
4954 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4955 EXPORT_SYMBOL(cpu_possible_map);
4958 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4960 struct task_struct *p;
4964 read_lock(&tasklist_lock);
4967 p = find_process_by_pid(pid);
4971 retval = security_task_getscheduler(p);
4975 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4978 read_unlock(&tasklist_lock);
4985 * sys_sched_getaffinity - get the cpu affinity of a process
4986 * @pid: pid of the process
4987 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4988 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4990 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4991 unsigned long __user *user_mask_ptr)
4996 if (len < sizeof(cpumask_t))
4999 ret = sched_getaffinity(pid, &mask);
5003 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5006 return sizeof(cpumask_t);
5010 * sys_sched_yield - yield the current processor to other threads.
5012 * This function yields the current CPU to other tasks. If there are no
5013 * other threads running on this CPU then this function will return.
5015 asmlinkage long sys_sched_yield(void)
5017 struct rq *rq = this_rq_lock();
5019 schedstat_inc(rq, yld_count);
5020 current->sched_class->yield_task(rq);
5023 * Since we are going to call schedule() anyway, there's
5024 * no need to preempt or enable interrupts:
5026 __release(rq->lock);
5027 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5028 _raw_spin_unlock(&rq->lock);
5029 preempt_enable_no_resched();
5036 static void __cond_resched(void)
5038 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5039 __might_sleep(__FILE__, __LINE__);
5042 * The BKS might be reacquired before we have dropped
5043 * PREEMPT_ACTIVE, which could trigger a second
5044 * cond_resched() call.
5047 add_preempt_count(PREEMPT_ACTIVE);
5049 sub_preempt_count(PREEMPT_ACTIVE);
5050 } while (need_resched());
5053 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5054 int __sched _cond_resched(void)
5056 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5057 system_state == SYSTEM_RUNNING) {
5063 EXPORT_SYMBOL(_cond_resched);
5067 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5068 * call schedule, and on return reacquire the lock.
5070 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5071 * operations here to prevent schedule() from being called twice (once via
5072 * spin_unlock(), once by hand).
5074 int cond_resched_lock(spinlock_t *lock)
5076 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5079 if (spin_needbreak(lock) || resched) {
5081 if (resched && need_resched())
5090 EXPORT_SYMBOL(cond_resched_lock);
5092 int __sched cond_resched_softirq(void)
5094 BUG_ON(!in_softirq());
5096 if (need_resched() && system_state == SYSTEM_RUNNING) {
5104 EXPORT_SYMBOL(cond_resched_softirq);
5107 * yield - yield the current processor to other threads.
5109 * This is a shortcut for kernel-space yielding - it marks the
5110 * thread runnable and calls sys_sched_yield().
5112 void __sched yield(void)
5114 set_current_state(TASK_RUNNING);
5117 EXPORT_SYMBOL(yield);
5120 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5121 * that process accounting knows that this is a task in IO wait state.
5123 * But don't do that if it is a deliberate, throttling IO wait (this task
5124 * has set its backing_dev_info: the queue against which it should throttle)
5126 void __sched io_schedule(void)
5128 struct rq *rq = &__raw_get_cpu_var(runqueues);
5130 delayacct_blkio_start();
5131 atomic_inc(&rq->nr_iowait);
5133 atomic_dec(&rq->nr_iowait);
5134 delayacct_blkio_end();
5136 EXPORT_SYMBOL(io_schedule);
5138 long __sched io_schedule_timeout(long timeout)
5140 struct rq *rq = &__raw_get_cpu_var(runqueues);
5143 delayacct_blkio_start();
5144 atomic_inc(&rq->nr_iowait);
5145 ret = schedule_timeout(timeout);
5146 atomic_dec(&rq->nr_iowait);
5147 delayacct_blkio_end();
5152 * sys_sched_get_priority_max - return maximum RT priority.
5153 * @policy: scheduling class.
5155 * this syscall returns the maximum rt_priority that can be used
5156 * by a given scheduling class.
5158 asmlinkage long sys_sched_get_priority_max(int policy)
5165 ret = MAX_USER_RT_PRIO-1;
5177 * sys_sched_get_priority_min - return minimum RT priority.
5178 * @policy: scheduling class.
5180 * this syscall returns the minimum rt_priority that can be used
5181 * by a given scheduling class.
5183 asmlinkage long sys_sched_get_priority_min(int policy)
5201 * sys_sched_rr_get_interval - return the default timeslice of a process.
5202 * @pid: pid of the process.
5203 * @interval: userspace pointer to the timeslice value.
5205 * this syscall writes the default timeslice value of a given process
5206 * into the user-space timespec buffer. A value of '0' means infinity.
5209 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5211 struct task_struct *p;
5212 unsigned int time_slice;
5220 read_lock(&tasklist_lock);
5221 p = find_process_by_pid(pid);
5225 retval = security_task_getscheduler(p);
5230 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5231 * tasks that are on an otherwise idle runqueue:
5234 if (p->policy == SCHED_RR) {
5235 time_slice = DEF_TIMESLICE;
5236 } else if (p->policy != SCHED_FIFO) {
5237 struct sched_entity *se = &p->se;
5238 unsigned long flags;
5241 rq = task_rq_lock(p, &flags);
5242 if (rq->cfs.load.weight)
5243 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5244 task_rq_unlock(rq, &flags);
5246 read_unlock(&tasklist_lock);
5247 jiffies_to_timespec(time_slice, &t);
5248 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5252 read_unlock(&tasklist_lock);
5256 static const char stat_nam[] = "RSDTtZX";
5258 void sched_show_task(struct task_struct *p)
5260 unsigned long free = 0;
5263 state = p->state ? __ffs(p->state) + 1 : 0;
5264 printk(KERN_INFO "%-13.13s %c", p->comm,
5265 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5266 #if BITS_PER_LONG == 32
5267 if (state == TASK_RUNNING)
5268 printk(KERN_CONT " running ");
5270 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5272 if (state == TASK_RUNNING)
5273 printk(KERN_CONT " running task ");
5275 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5277 #ifdef CONFIG_DEBUG_STACK_USAGE
5279 unsigned long *n = end_of_stack(p);
5282 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5285 printk(KERN_CONT "%5lu %5d %6d\n", free,
5286 task_pid_nr(p), task_pid_nr(p->real_parent));
5288 show_stack(p, NULL);
5291 void show_state_filter(unsigned long state_filter)
5293 struct task_struct *g, *p;
5295 #if BITS_PER_LONG == 32
5297 " task PC stack pid father\n");
5300 " task PC stack pid father\n");
5302 read_lock(&tasklist_lock);
5303 do_each_thread(g, p) {
5305 * reset the NMI-timeout, listing all files on a slow
5306 * console might take alot of time:
5308 touch_nmi_watchdog();
5309 if (!state_filter || (p->state & state_filter))
5311 } while_each_thread(g, p);
5313 touch_all_softlockup_watchdogs();
5315 #ifdef CONFIG_SCHED_DEBUG
5316 sysrq_sched_debug_show();
5318 read_unlock(&tasklist_lock);
5320 * Only show locks if all tasks are dumped:
5322 if (state_filter == -1)
5323 debug_show_all_locks();
5326 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5328 idle->sched_class = &idle_sched_class;
5332 * init_idle - set up an idle thread for a given CPU
5333 * @idle: task in question
5334 * @cpu: cpu the idle task belongs to
5336 * NOTE: this function does not set the idle thread's NEED_RESCHED
5337 * flag, to make booting more robust.
5339 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5341 struct rq *rq = cpu_rq(cpu);
5342 unsigned long flags;
5345 idle->se.exec_start = sched_clock();
5347 idle->prio = idle->normal_prio = MAX_PRIO;
5348 idle->cpus_allowed = cpumask_of_cpu(cpu);
5349 __set_task_cpu(idle, cpu);
5351 spin_lock_irqsave(&rq->lock, flags);
5352 rq->curr = rq->idle = idle;
5353 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5356 spin_unlock_irqrestore(&rq->lock, flags);
5358 /* Set the preempt count _outside_ the spinlocks! */
5359 task_thread_info(idle)->preempt_count = 0;
5362 * The idle tasks have their own, simple scheduling class:
5364 idle->sched_class = &idle_sched_class;
5368 * In a system that switches off the HZ timer nohz_cpu_mask
5369 * indicates which cpus entered this state. This is used
5370 * in the rcu update to wait only for active cpus. For system
5371 * which do not switch off the HZ timer nohz_cpu_mask should
5372 * always be CPU_MASK_NONE.
5374 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5377 * Increase the granularity value when there are more CPUs,
5378 * because with more CPUs the 'effective latency' as visible
5379 * to users decreases. But the relationship is not linear,
5380 * so pick a second-best guess by going with the log2 of the
5383 * This idea comes from the SD scheduler of Con Kolivas:
5385 static inline void sched_init_granularity(void)
5387 unsigned int factor = 1 + ilog2(num_online_cpus());
5388 const unsigned long limit = 200000000;
5390 sysctl_sched_min_granularity *= factor;
5391 if (sysctl_sched_min_granularity > limit)
5392 sysctl_sched_min_granularity = limit;
5394 sysctl_sched_latency *= factor;
5395 if (sysctl_sched_latency > limit)
5396 sysctl_sched_latency = limit;
5398 sysctl_sched_wakeup_granularity *= factor;
5403 * This is how migration works:
5405 * 1) we queue a struct migration_req structure in the source CPU's
5406 * runqueue and wake up that CPU's migration thread.
5407 * 2) we down() the locked semaphore => thread blocks.
5408 * 3) migration thread wakes up (implicitly it forces the migrated
5409 * thread off the CPU)
5410 * 4) it gets the migration request and checks whether the migrated
5411 * task is still in the wrong runqueue.
5412 * 5) if it's in the wrong runqueue then the migration thread removes
5413 * it and puts it into the right queue.
5414 * 6) migration thread up()s the semaphore.
5415 * 7) we wake up and the migration is done.
5419 * Change a given task's CPU affinity. Migrate the thread to a
5420 * proper CPU and schedule it away if the CPU it's executing on
5421 * is removed from the allowed bitmask.
5423 * NOTE: the caller must have a valid reference to the task, the
5424 * task must not exit() & deallocate itself prematurely. The
5425 * call is not atomic; no spinlocks may be held.
5427 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5429 struct migration_req req;
5430 unsigned long flags;
5434 rq = task_rq_lock(p, &flags);
5435 if (!cpus_intersects(new_mask, cpu_online_map)) {
5440 if (p->sched_class->set_cpus_allowed)
5441 p->sched_class->set_cpus_allowed(p, &new_mask);
5443 p->cpus_allowed = new_mask;
5444 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5447 /* Can the task run on the task's current CPU? If so, we're done */
5448 if (cpu_isset(task_cpu(p), new_mask))
5451 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5452 /* Need help from migration thread: drop lock and wait. */
5453 task_rq_unlock(rq, &flags);
5454 wake_up_process(rq->migration_thread);
5455 wait_for_completion(&req.done);
5456 tlb_migrate_finish(p->mm);
5460 task_rq_unlock(rq, &flags);
5464 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5467 * Move (not current) task off this cpu, onto dest cpu. We're doing
5468 * this because either it can't run here any more (set_cpus_allowed()
5469 * away from this CPU, or CPU going down), or because we're
5470 * attempting to rebalance this task on exec (sched_exec).
5472 * So we race with normal scheduler movements, but that's OK, as long
5473 * as the task is no longer on this CPU.
5475 * Returns non-zero if task was successfully migrated.
5477 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5479 struct rq *rq_dest, *rq_src;
5482 if (unlikely(cpu_is_offline(dest_cpu)))
5485 rq_src = cpu_rq(src_cpu);
5486 rq_dest = cpu_rq(dest_cpu);
5488 double_rq_lock(rq_src, rq_dest);
5489 /* Already moved. */
5490 if (task_cpu(p) != src_cpu)
5492 /* Affinity changed (again). */
5493 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5496 on_rq = p->se.on_rq;
5498 deactivate_task(rq_src, p, 0);
5500 set_task_cpu(p, dest_cpu);
5502 activate_task(rq_dest, p, 0);
5503 check_preempt_curr(rq_dest, p);
5507 double_rq_unlock(rq_src, rq_dest);
5512 * migration_thread - this is a highprio system thread that performs
5513 * thread migration by bumping thread off CPU then 'pushing' onto
5516 static int migration_thread(void *data)
5518 int cpu = (long)data;
5522 BUG_ON(rq->migration_thread != current);
5524 set_current_state(TASK_INTERRUPTIBLE);
5525 while (!kthread_should_stop()) {
5526 struct migration_req *req;
5527 struct list_head *head;
5529 spin_lock_irq(&rq->lock);
5531 if (cpu_is_offline(cpu)) {
5532 spin_unlock_irq(&rq->lock);
5536 if (rq->active_balance) {
5537 active_load_balance(rq, cpu);
5538 rq->active_balance = 0;
5541 head = &rq->migration_queue;
5543 if (list_empty(head)) {
5544 spin_unlock_irq(&rq->lock);
5546 set_current_state(TASK_INTERRUPTIBLE);
5549 req = list_entry(head->next, struct migration_req, list);
5550 list_del_init(head->next);
5552 spin_unlock(&rq->lock);
5553 __migrate_task(req->task, cpu, req->dest_cpu);
5556 complete(&req->done);
5558 __set_current_state(TASK_RUNNING);
5562 /* Wait for kthread_stop */
5563 set_current_state(TASK_INTERRUPTIBLE);
5564 while (!kthread_should_stop()) {
5566 set_current_state(TASK_INTERRUPTIBLE);
5568 __set_current_state(TASK_RUNNING);
5572 #ifdef CONFIG_HOTPLUG_CPU
5574 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5578 local_irq_disable();
5579 ret = __migrate_task(p, src_cpu, dest_cpu);
5585 * Figure out where task on dead CPU should go, use force if necessary.
5586 * NOTE: interrupts should be disabled by the caller
5588 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5590 unsigned long flags;
5597 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5598 cpus_and(mask, mask, p->cpus_allowed);
5599 dest_cpu = any_online_cpu(mask);
5601 /* On any allowed CPU? */
5602 if (dest_cpu == NR_CPUS)
5603 dest_cpu = any_online_cpu(p->cpus_allowed);
5605 /* No more Mr. Nice Guy. */
5606 if (dest_cpu == NR_CPUS) {
5607 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5609 * Try to stay on the same cpuset, where the
5610 * current cpuset may be a subset of all cpus.
5611 * The cpuset_cpus_allowed_locked() variant of
5612 * cpuset_cpus_allowed() will not block. It must be
5613 * called within calls to cpuset_lock/cpuset_unlock.
5615 rq = task_rq_lock(p, &flags);
5616 p->cpus_allowed = cpus_allowed;
5617 dest_cpu = any_online_cpu(p->cpus_allowed);
5618 task_rq_unlock(rq, &flags);
5621 * Don't tell them about moving exiting tasks or
5622 * kernel threads (both mm NULL), since they never
5625 if (p->mm && printk_ratelimit()) {
5626 printk(KERN_INFO "process %d (%s) no "
5627 "longer affine to cpu%d\n",
5628 task_pid_nr(p), p->comm, dead_cpu);
5631 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5635 * While a dead CPU has no uninterruptible tasks queued at this point,
5636 * it might still have a nonzero ->nr_uninterruptible counter, because
5637 * for performance reasons the counter is not stricly tracking tasks to
5638 * their home CPUs. So we just add the counter to another CPU's counter,
5639 * to keep the global sum constant after CPU-down:
5641 static void migrate_nr_uninterruptible(struct rq *rq_src)
5643 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5644 unsigned long flags;
5646 local_irq_save(flags);
5647 double_rq_lock(rq_src, rq_dest);
5648 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5649 rq_src->nr_uninterruptible = 0;
5650 double_rq_unlock(rq_src, rq_dest);
5651 local_irq_restore(flags);
5654 /* Run through task list and migrate tasks from the dead cpu. */
5655 static void migrate_live_tasks(int src_cpu)
5657 struct task_struct *p, *t;
5659 read_lock(&tasklist_lock);
5661 do_each_thread(t, p) {
5665 if (task_cpu(p) == src_cpu)
5666 move_task_off_dead_cpu(src_cpu, p);
5667 } while_each_thread(t, p);
5669 read_unlock(&tasklist_lock);
5673 * Schedules idle task to be the next runnable task on current CPU.
5674 * It does so by boosting its priority to highest possible.
5675 * Used by CPU offline code.
5677 void sched_idle_next(void)
5679 int this_cpu = smp_processor_id();
5680 struct rq *rq = cpu_rq(this_cpu);
5681 struct task_struct *p = rq->idle;
5682 unsigned long flags;
5684 /* cpu has to be offline */
5685 BUG_ON(cpu_online(this_cpu));
5688 * Strictly not necessary since rest of the CPUs are stopped by now
5689 * and interrupts disabled on the current cpu.
5691 spin_lock_irqsave(&rq->lock, flags);
5693 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5695 update_rq_clock(rq);
5696 activate_task(rq, p, 0);
5698 spin_unlock_irqrestore(&rq->lock, flags);
5702 * Ensures that the idle task is using init_mm right before its cpu goes
5705 void idle_task_exit(void)
5707 struct mm_struct *mm = current->active_mm;
5709 BUG_ON(cpu_online(smp_processor_id()));
5712 switch_mm(mm, &init_mm, current);
5716 /* called under rq->lock with disabled interrupts */
5717 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5719 struct rq *rq = cpu_rq(dead_cpu);
5721 /* Must be exiting, otherwise would be on tasklist. */
5722 BUG_ON(!p->exit_state);
5724 /* Cannot have done final schedule yet: would have vanished. */
5725 BUG_ON(p->state == TASK_DEAD);
5730 * Drop lock around migration; if someone else moves it,
5731 * that's OK. No task can be added to this CPU, so iteration is
5734 spin_unlock_irq(&rq->lock);
5735 move_task_off_dead_cpu(dead_cpu, p);
5736 spin_lock_irq(&rq->lock);
5741 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5742 static void migrate_dead_tasks(unsigned int dead_cpu)
5744 struct rq *rq = cpu_rq(dead_cpu);
5745 struct task_struct *next;
5748 if (!rq->nr_running)
5750 update_rq_clock(rq);
5751 next = pick_next_task(rq, rq->curr);
5754 migrate_dead(dead_cpu, next);
5758 #endif /* CONFIG_HOTPLUG_CPU */
5760 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5762 static struct ctl_table sd_ctl_dir[] = {
5764 .procname = "sched_domain",
5770 static struct ctl_table sd_ctl_root[] = {
5772 .ctl_name = CTL_KERN,
5773 .procname = "kernel",
5775 .child = sd_ctl_dir,
5780 static struct ctl_table *sd_alloc_ctl_entry(int n)
5782 struct ctl_table *entry =
5783 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5788 static void sd_free_ctl_entry(struct ctl_table **tablep)
5790 struct ctl_table *entry;
5793 * In the intermediate directories, both the child directory and
5794 * procname are dynamically allocated and could fail but the mode
5795 * will always be set. In the lowest directory the names are
5796 * static strings and all have proc handlers.
5798 for (entry = *tablep; entry->mode; entry++) {
5800 sd_free_ctl_entry(&entry->child);
5801 if (entry->proc_handler == NULL)
5802 kfree(entry->procname);
5810 set_table_entry(struct ctl_table *entry,
5811 const char *procname, void *data, int maxlen,
5812 mode_t mode, proc_handler *proc_handler)
5814 entry->procname = procname;
5816 entry->maxlen = maxlen;
5818 entry->proc_handler = proc_handler;
5821 static struct ctl_table *
5822 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5824 struct ctl_table *table = sd_alloc_ctl_entry(12);
5829 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5830 sizeof(long), 0644, proc_doulongvec_minmax);
5831 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5832 sizeof(long), 0644, proc_doulongvec_minmax);
5833 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5834 sizeof(int), 0644, proc_dointvec_minmax);
5835 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5836 sizeof(int), 0644, proc_dointvec_minmax);
5837 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5838 sizeof(int), 0644, proc_dointvec_minmax);
5839 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5840 sizeof(int), 0644, proc_dointvec_minmax);
5841 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5842 sizeof(int), 0644, proc_dointvec_minmax);
5843 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5844 sizeof(int), 0644, proc_dointvec_minmax);
5845 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5846 sizeof(int), 0644, proc_dointvec_minmax);
5847 set_table_entry(&table[9], "cache_nice_tries",
5848 &sd->cache_nice_tries,
5849 sizeof(int), 0644, proc_dointvec_minmax);
5850 set_table_entry(&table[10], "flags", &sd->flags,
5851 sizeof(int), 0644, proc_dointvec_minmax);
5852 /* &table[11] is terminator */
5857 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5859 struct ctl_table *entry, *table;
5860 struct sched_domain *sd;
5861 int domain_num = 0, i;
5864 for_each_domain(cpu, sd)
5866 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5871 for_each_domain(cpu, sd) {
5872 snprintf(buf, 32, "domain%d", i);
5873 entry->procname = kstrdup(buf, GFP_KERNEL);
5875 entry->child = sd_alloc_ctl_domain_table(sd);
5882 static struct ctl_table_header *sd_sysctl_header;
5883 static void register_sched_domain_sysctl(void)
5885 int i, cpu_num = num_online_cpus();
5886 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5889 WARN_ON(sd_ctl_dir[0].child);
5890 sd_ctl_dir[0].child = entry;
5895 for_each_online_cpu(i) {
5896 snprintf(buf, 32, "cpu%d", i);
5897 entry->procname = kstrdup(buf, GFP_KERNEL);
5899 entry->child = sd_alloc_ctl_cpu_table(i);
5903 WARN_ON(sd_sysctl_header);
5904 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5907 /* may be called multiple times per register */
5908 static void unregister_sched_domain_sysctl(void)
5910 if (sd_sysctl_header)
5911 unregister_sysctl_table(sd_sysctl_header);
5912 sd_sysctl_header = NULL;
5913 if (sd_ctl_dir[0].child)
5914 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5917 static void register_sched_domain_sysctl(void)
5920 static void unregister_sched_domain_sysctl(void)
5926 * migration_call - callback that gets triggered when a CPU is added.
5927 * Here we can start up the necessary migration thread for the new CPU.
5929 static int __cpuinit
5930 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5932 struct task_struct *p;
5933 int cpu = (long)hcpu;
5934 unsigned long flags;
5939 case CPU_UP_PREPARE:
5940 case CPU_UP_PREPARE_FROZEN:
5941 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5944 kthread_bind(p, cpu);
5945 /* Must be high prio: stop_machine expects to yield to it. */
5946 rq = task_rq_lock(p, &flags);
5947 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5948 task_rq_unlock(rq, &flags);
5949 cpu_rq(cpu)->migration_thread = p;
5953 case CPU_ONLINE_FROZEN:
5954 /* Strictly unnecessary, as first user will wake it. */
5955 wake_up_process(cpu_rq(cpu)->migration_thread);
5957 /* Update our root-domain */
5959 spin_lock_irqsave(&rq->lock, flags);
5961 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5962 cpu_set(cpu, rq->rd->online);
5964 spin_unlock_irqrestore(&rq->lock, flags);
5967 #ifdef CONFIG_HOTPLUG_CPU
5968 case CPU_UP_CANCELED:
5969 case CPU_UP_CANCELED_FROZEN:
5970 if (!cpu_rq(cpu)->migration_thread)
5972 /* Unbind it from offline cpu so it can run. Fall thru. */
5973 kthread_bind(cpu_rq(cpu)->migration_thread,
5974 any_online_cpu(cpu_online_map));
5975 kthread_stop(cpu_rq(cpu)->migration_thread);
5976 cpu_rq(cpu)->migration_thread = NULL;
5980 case CPU_DEAD_FROZEN:
5981 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5982 migrate_live_tasks(cpu);
5984 kthread_stop(rq->migration_thread);
5985 rq->migration_thread = NULL;
5986 /* Idle task back to normal (off runqueue, low prio) */
5987 spin_lock_irq(&rq->lock);
5988 update_rq_clock(rq);
5989 deactivate_task(rq, rq->idle, 0);
5990 rq->idle->static_prio = MAX_PRIO;
5991 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5992 rq->idle->sched_class = &idle_sched_class;
5993 migrate_dead_tasks(cpu);
5994 spin_unlock_irq(&rq->lock);
5996 migrate_nr_uninterruptible(rq);
5997 BUG_ON(rq->nr_running != 0);
6000 * No need to migrate the tasks: it was best-effort if
6001 * they didn't take sched_hotcpu_mutex. Just wake up
6004 spin_lock_irq(&rq->lock);
6005 while (!list_empty(&rq->migration_queue)) {
6006 struct migration_req *req;
6008 req = list_entry(rq->migration_queue.next,
6009 struct migration_req, list);
6010 list_del_init(&req->list);
6011 complete(&req->done);
6013 spin_unlock_irq(&rq->lock);
6017 case CPU_DYING_FROZEN:
6018 /* Update our root-domain */
6020 spin_lock_irqsave(&rq->lock, flags);
6022 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6023 cpu_clear(cpu, rq->rd->online);
6025 spin_unlock_irqrestore(&rq->lock, flags);
6032 /* Register at highest priority so that task migration (migrate_all_tasks)
6033 * happens before everything else.
6035 static struct notifier_block __cpuinitdata migration_notifier = {
6036 .notifier_call = migration_call,
6040 void __init migration_init(void)
6042 void *cpu = (void *)(long)smp_processor_id();
6045 /* Start one for the boot CPU: */
6046 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6047 BUG_ON(err == NOTIFY_BAD);
6048 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6049 register_cpu_notifier(&migration_notifier);
6055 /* Number of possible processor ids */
6056 int nr_cpu_ids __read_mostly = NR_CPUS;
6057 EXPORT_SYMBOL(nr_cpu_ids);
6059 #ifdef CONFIG_SCHED_DEBUG
6061 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
6063 struct sched_group *group = sd->groups;
6064 cpumask_t groupmask;
6067 cpumask_scnprintf(str, NR_CPUS, sd->span);
6068 cpus_clear(groupmask);
6070 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6072 if (!(sd->flags & SD_LOAD_BALANCE)) {
6073 printk("does not load-balance\n");
6075 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6080 printk(KERN_CONT "span %s\n", str);
6082 if (!cpu_isset(cpu, sd->span)) {
6083 printk(KERN_ERR "ERROR: domain->span does not contain "
6086 if (!cpu_isset(cpu, group->cpumask)) {
6087 printk(KERN_ERR "ERROR: domain->groups does not contain"
6091 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6095 printk(KERN_ERR "ERROR: group is NULL\n");
6099 if (!group->__cpu_power) {
6100 printk(KERN_CONT "\n");
6101 printk(KERN_ERR "ERROR: domain->cpu_power not "
6106 if (!cpus_weight(group->cpumask)) {
6107 printk(KERN_CONT "\n");
6108 printk(KERN_ERR "ERROR: empty group\n");
6112 if (cpus_intersects(groupmask, group->cpumask)) {
6113 printk(KERN_CONT "\n");
6114 printk(KERN_ERR "ERROR: repeated CPUs\n");
6118 cpus_or(groupmask, groupmask, group->cpumask);
6120 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6121 printk(KERN_CONT " %s", str);
6123 group = group->next;
6124 } while (group != sd->groups);
6125 printk(KERN_CONT "\n");
6127 if (!cpus_equal(sd->span, groupmask))
6128 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6130 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6131 printk(KERN_ERR "ERROR: parent span is not a superset "
6132 "of domain->span\n");
6136 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6141 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6145 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6148 if (sched_domain_debug_one(sd, cpu, level))
6157 # define sched_domain_debug(sd, cpu) do { } while (0)
6160 static int sd_degenerate(struct sched_domain *sd)
6162 if (cpus_weight(sd->span) == 1)
6165 /* Following flags need at least 2 groups */
6166 if (sd->flags & (SD_LOAD_BALANCE |
6167 SD_BALANCE_NEWIDLE |
6171 SD_SHARE_PKG_RESOURCES)) {
6172 if (sd->groups != sd->groups->next)
6176 /* Following flags don't use groups */
6177 if (sd->flags & (SD_WAKE_IDLE |
6186 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6188 unsigned long cflags = sd->flags, pflags = parent->flags;
6190 if (sd_degenerate(parent))
6193 if (!cpus_equal(sd->span, parent->span))
6196 /* Does parent contain flags not in child? */
6197 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6198 if (cflags & SD_WAKE_AFFINE)
6199 pflags &= ~SD_WAKE_BALANCE;
6200 /* Flags needing groups don't count if only 1 group in parent */
6201 if (parent->groups == parent->groups->next) {
6202 pflags &= ~(SD_LOAD_BALANCE |
6203 SD_BALANCE_NEWIDLE |
6207 SD_SHARE_PKG_RESOURCES);
6209 if (~cflags & pflags)
6215 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6217 unsigned long flags;
6218 const struct sched_class *class;
6220 spin_lock_irqsave(&rq->lock, flags);
6223 struct root_domain *old_rd = rq->rd;
6225 for (class = sched_class_highest; class; class = class->next) {
6226 if (class->leave_domain)
6227 class->leave_domain(rq);
6230 cpu_clear(rq->cpu, old_rd->span);
6231 cpu_clear(rq->cpu, old_rd->online);
6233 if (atomic_dec_and_test(&old_rd->refcount))
6237 atomic_inc(&rd->refcount);
6240 cpu_set(rq->cpu, rd->span);
6241 if (cpu_isset(rq->cpu, cpu_online_map))
6242 cpu_set(rq->cpu, rd->online);
6244 for (class = sched_class_highest; class; class = class->next) {
6245 if (class->join_domain)
6246 class->join_domain(rq);
6249 spin_unlock_irqrestore(&rq->lock, flags);
6252 static void init_rootdomain(struct root_domain *rd)
6254 memset(rd, 0, sizeof(*rd));
6256 cpus_clear(rd->span);
6257 cpus_clear(rd->online);
6260 static void init_defrootdomain(void)
6262 init_rootdomain(&def_root_domain);
6263 atomic_set(&def_root_domain.refcount, 1);
6266 static struct root_domain *alloc_rootdomain(void)
6268 struct root_domain *rd;
6270 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6274 init_rootdomain(rd);
6280 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6281 * hold the hotplug lock.
6284 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6286 struct rq *rq = cpu_rq(cpu);
6287 struct sched_domain *tmp;
6289 /* Remove the sched domains which do not contribute to scheduling. */
6290 for (tmp = sd; tmp; tmp = tmp->parent) {
6291 struct sched_domain *parent = tmp->parent;
6294 if (sd_parent_degenerate(tmp, parent)) {
6295 tmp->parent = parent->parent;
6297 parent->parent->child = tmp;
6301 if (sd && sd_degenerate(sd)) {
6307 sched_domain_debug(sd, cpu);
6309 rq_attach_root(rq, rd);
6310 rcu_assign_pointer(rq->sd, sd);
6313 /* cpus with isolated domains */
6314 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6316 /* Setup the mask of cpus configured for isolated domains */
6317 static int __init isolated_cpu_setup(char *str)
6319 int ints[NR_CPUS], i;
6321 str = get_options(str, ARRAY_SIZE(ints), ints);
6322 cpus_clear(cpu_isolated_map);
6323 for (i = 1; i <= ints[0]; i++)
6324 if (ints[i] < NR_CPUS)
6325 cpu_set(ints[i], cpu_isolated_map);
6329 __setup("isolcpus=", isolated_cpu_setup);
6332 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6333 * to a function which identifies what group(along with sched group) a CPU
6334 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6335 * (due to the fact that we keep track of groups covered with a cpumask_t).
6337 * init_sched_build_groups will build a circular linked list of the groups
6338 * covered by the given span, and will set each group's ->cpumask correctly,
6339 * and ->cpu_power to 0.
6342 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6343 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6344 struct sched_group **sg))
6346 struct sched_group *first = NULL, *last = NULL;
6347 cpumask_t covered = CPU_MASK_NONE;
6350 for_each_cpu_mask(i, span) {
6351 struct sched_group *sg;
6352 int group = group_fn(i, cpu_map, &sg);
6355 if (cpu_isset(i, covered))
6358 sg->cpumask = CPU_MASK_NONE;
6359 sg->__cpu_power = 0;
6361 for_each_cpu_mask(j, span) {
6362 if (group_fn(j, cpu_map, NULL) != group)
6365 cpu_set(j, covered);
6366 cpu_set(j, sg->cpumask);
6377 #define SD_NODES_PER_DOMAIN 16
6382 * find_next_best_node - find the next node to include in a sched_domain
6383 * @node: node whose sched_domain we're building
6384 * @used_nodes: nodes already in the sched_domain
6386 * Find the next node to include in a given scheduling domain. Simply
6387 * finds the closest node not already in the @used_nodes map.
6389 * Should use nodemask_t.
6391 static int find_next_best_node(int node, unsigned long *used_nodes)
6393 int i, n, val, min_val, best_node = 0;
6397 for (i = 0; i < MAX_NUMNODES; i++) {
6398 /* Start at @node */
6399 n = (node + i) % MAX_NUMNODES;
6401 if (!nr_cpus_node(n))
6404 /* Skip already used nodes */
6405 if (test_bit(n, used_nodes))
6408 /* Simple min distance search */
6409 val = node_distance(node, n);
6411 if (val < min_val) {
6417 set_bit(best_node, used_nodes);
6422 * sched_domain_node_span - get a cpumask for a node's sched_domain
6423 * @node: node whose cpumask we're constructing
6424 * @size: number of nodes to include in this span
6426 * Given a node, construct a good cpumask for its sched_domain to span. It
6427 * should be one that prevents unnecessary balancing, but also spreads tasks
6430 static cpumask_t sched_domain_node_span(int node)
6432 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6433 cpumask_t span, nodemask;
6437 bitmap_zero(used_nodes, MAX_NUMNODES);
6439 nodemask = node_to_cpumask(node);
6440 cpus_or(span, span, nodemask);
6441 set_bit(node, used_nodes);
6443 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6444 int next_node = find_next_best_node(node, used_nodes);
6446 nodemask = node_to_cpumask(next_node);
6447 cpus_or(span, span, nodemask);
6454 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6457 * SMT sched-domains:
6459 #ifdef CONFIG_SCHED_SMT
6460 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6461 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6464 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6467 *sg = &per_cpu(sched_group_cpus, cpu);
6473 * multi-core sched-domains:
6475 #ifdef CONFIG_SCHED_MC
6476 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6477 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6480 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6482 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6485 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6486 cpus_and(mask, mask, *cpu_map);
6487 group = first_cpu(mask);
6489 *sg = &per_cpu(sched_group_core, group);
6492 #elif defined(CONFIG_SCHED_MC)
6494 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6497 *sg = &per_cpu(sched_group_core, cpu);
6502 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6503 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6506 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6509 #ifdef CONFIG_SCHED_MC
6510 cpumask_t mask = cpu_coregroup_map(cpu);
6511 cpus_and(mask, mask, *cpu_map);
6512 group = first_cpu(mask);
6513 #elif defined(CONFIG_SCHED_SMT)
6514 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6515 cpus_and(mask, mask, *cpu_map);
6516 group = first_cpu(mask);
6521 *sg = &per_cpu(sched_group_phys, group);
6527 * The init_sched_build_groups can't handle what we want to do with node
6528 * groups, so roll our own. Now each node has its own list of groups which
6529 * gets dynamically allocated.
6531 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6532 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6534 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6535 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6537 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6538 struct sched_group **sg)
6540 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6543 cpus_and(nodemask, nodemask, *cpu_map);
6544 group = first_cpu(nodemask);
6547 *sg = &per_cpu(sched_group_allnodes, group);
6551 static void init_numa_sched_groups_power(struct sched_group *group_head)
6553 struct sched_group *sg = group_head;
6559 for_each_cpu_mask(j, sg->cpumask) {
6560 struct sched_domain *sd;
6562 sd = &per_cpu(phys_domains, j);
6563 if (j != first_cpu(sd->groups->cpumask)) {
6565 * Only add "power" once for each
6571 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6574 } while (sg != group_head);
6579 /* Free memory allocated for various sched_group structures */
6580 static void free_sched_groups(const cpumask_t *cpu_map)
6584 for_each_cpu_mask(cpu, *cpu_map) {
6585 struct sched_group **sched_group_nodes
6586 = sched_group_nodes_bycpu[cpu];
6588 if (!sched_group_nodes)
6591 for (i = 0; i < MAX_NUMNODES; i++) {
6592 cpumask_t nodemask = node_to_cpumask(i);
6593 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6595 cpus_and(nodemask, nodemask, *cpu_map);
6596 if (cpus_empty(nodemask))
6606 if (oldsg != sched_group_nodes[i])
6609 kfree(sched_group_nodes);
6610 sched_group_nodes_bycpu[cpu] = NULL;
6614 static void free_sched_groups(const cpumask_t *cpu_map)
6620 * Initialize sched groups cpu_power.
6622 * cpu_power indicates the capacity of sched group, which is used while
6623 * distributing the load between different sched groups in a sched domain.
6624 * Typically cpu_power for all the groups in a sched domain will be same unless
6625 * there are asymmetries in the topology. If there are asymmetries, group
6626 * having more cpu_power will pickup more load compared to the group having
6629 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6630 * the maximum number of tasks a group can handle in the presence of other idle
6631 * or lightly loaded groups in the same sched domain.
6633 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6635 struct sched_domain *child;
6636 struct sched_group *group;
6638 WARN_ON(!sd || !sd->groups);
6640 if (cpu != first_cpu(sd->groups->cpumask))
6645 sd->groups->__cpu_power = 0;
6648 * For perf policy, if the groups in child domain share resources
6649 * (for example cores sharing some portions of the cache hierarchy
6650 * or SMT), then set this domain groups cpu_power such that each group
6651 * can handle only one task, when there are other idle groups in the
6652 * same sched domain.
6654 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6656 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6657 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6662 * add cpu_power of each child group to this groups cpu_power
6664 group = child->groups;
6666 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6667 group = group->next;
6668 } while (group != child->groups);
6672 * Build sched domains for a given set of cpus and attach the sched domains
6673 * to the individual cpus
6675 static int build_sched_domains(const cpumask_t *cpu_map)
6678 struct root_domain *rd;
6680 struct sched_group **sched_group_nodes = NULL;
6681 int sd_allnodes = 0;
6684 * Allocate the per-node list of sched groups
6686 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6688 if (!sched_group_nodes) {
6689 printk(KERN_WARNING "Can not alloc sched group node list\n");
6692 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6695 rd = alloc_rootdomain();
6697 printk(KERN_WARNING "Cannot alloc root domain\n");
6702 * Set up domains for cpus specified by the cpu_map.
6704 for_each_cpu_mask(i, *cpu_map) {
6705 struct sched_domain *sd = NULL, *p;
6706 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6708 cpus_and(nodemask, nodemask, *cpu_map);
6711 if (cpus_weight(*cpu_map) >
6712 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6713 sd = &per_cpu(allnodes_domains, i);
6714 *sd = SD_ALLNODES_INIT;
6715 sd->span = *cpu_map;
6716 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6722 sd = &per_cpu(node_domains, i);
6724 sd->span = sched_domain_node_span(cpu_to_node(i));
6728 cpus_and(sd->span, sd->span, *cpu_map);
6732 sd = &per_cpu(phys_domains, i);
6734 sd->span = nodemask;
6738 cpu_to_phys_group(i, cpu_map, &sd->groups);
6740 #ifdef CONFIG_SCHED_MC
6742 sd = &per_cpu(core_domains, i);
6744 sd->span = cpu_coregroup_map(i);
6745 cpus_and(sd->span, sd->span, *cpu_map);
6748 cpu_to_core_group(i, cpu_map, &sd->groups);
6751 #ifdef CONFIG_SCHED_SMT
6753 sd = &per_cpu(cpu_domains, i);
6754 *sd = SD_SIBLING_INIT;
6755 sd->span = per_cpu(cpu_sibling_map, i);
6756 cpus_and(sd->span, sd->span, *cpu_map);
6759 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6763 #ifdef CONFIG_SCHED_SMT
6764 /* Set up CPU (sibling) groups */
6765 for_each_cpu_mask(i, *cpu_map) {
6766 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6767 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6768 if (i != first_cpu(this_sibling_map))
6771 init_sched_build_groups(this_sibling_map, cpu_map,
6776 #ifdef CONFIG_SCHED_MC
6777 /* Set up multi-core groups */
6778 for_each_cpu_mask(i, *cpu_map) {
6779 cpumask_t this_core_map = cpu_coregroup_map(i);
6780 cpus_and(this_core_map, this_core_map, *cpu_map);
6781 if (i != first_cpu(this_core_map))
6783 init_sched_build_groups(this_core_map, cpu_map,
6784 &cpu_to_core_group);
6788 /* Set up physical groups */
6789 for (i = 0; i < MAX_NUMNODES; i++) {
6790 cpumask_t nodemask = node_to_cpumask(i);
6792 cpus_and(nodemask, nodemask, *cpu_map);
6793 if (cpus_empty(nodemask))
6796 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6800 /* Set up node groups */
6802 init_sched_build_groups(*cpu_map, cpu_map,
6803 &cpu_to_allnodes_group);
6805 for (i = 0; i < MAX_NUMNODES; i++) {
6806 /* Set up node groups */
6807 struct sched_group *sg, *prev;
6808 cpumask_t nodemask = node_to_cpumask(i);
6809 cpumask_t domainspan;
6810 cpumask_t covered = CPU_MASK_NONE;
6813 cpus_and(nodemask, nodemask, *cpu_map);
6814 if (cpus_empty(nodemask)) {
6815 sched_group_nodes[i] = NULL;
6819 domainspan = sched_domain_node_span(i);
6820 cpus_and(domainspan, domainspan, *cpu_map);
6822 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6824 printk(KERN_WARNING "Can not alloc domain group for "
6828 sched_group_nodes[i] = sg;
6829 for_each_cpu_mask(j, nodemask) {
6830 struct sched_domain *sd;
6832 sd = &per_cpu(node_domains, j);
6835 sg->__cpu_power = 0;
6836 sg->cpumask = nodemask;
6838 cpus_or(covered, covered, nodemask);
6841 for (j = 0; j < MAX_NUMNODES; j++) {
6842 cpumask_t tmp, notcovered;
6843 int n = (i + j) % MAX_NUMNODES;
6845 cpus_complement(notcovered, covered);
6846 cpus_and(tmp, notcovered, *cpu_map);
6847 cpus_and(tmp, tmp, domainspan);
6848 if (cpus_empty(tmp))
6851 nodemask = node_to_cpumask(n);
6852 cpus_and(tmp, tmp, nodemask);
6853 if (cpus_empty(tmp))
6856 sg = kmalloc_node(sizeof(struct sched_group),
6860 "Can not alloc domain group for node %d\n", j);
6863 sg->__cpu_power = 0;
6865 sg->next = prev->next;
6866 cpus_or(covered, covered, tmp);
6873 /* Calculate CPU power for physical packages and nodes */
6874 #ifdef CONFIG_SCHED_SMT
6875 for_each_cpu_mask(i, *cpu_map) {
6876 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6878 init_sched_groups_power(i, sd);
6881 #ifdef CONFIG_SCHED_MC
6882 for_each_cpu_mask(i, *cpu_map) {
6883 struct sched_domain *sd = &per_cpu(core_domains, i);
6885 init_sched_groups_power(i, sd);
6889 for_each_cpu_mask(i, *cpu_map) {
6890 struct sched_domain *sd = &per_cpu(phys_domains, i);
6892 init_sched_groups_power(i, sd);
6896 for (i = 0; i < MAX_NUMNODES; i++)
6897 init_numa_sched_groups_power(sched_group_nodes[i]);
6900 struct sched_group *sg;
6902 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6903 init_numa_sched_groups_power(sg);
6907 /* Attach the domains */
6908 for_each_cpu_mask(i, *cpu_map) {
6909 struct sched_domain *sd;
6910 #ifdef CONFIG_SCHED_SMT
6911 sd = &per_cpu(cpu_domains, i);
6912 #elif defined(CONFIG_SCHED_MC)
6913 sd = &per_cpu(core_domains, i);
6915 sd = &per_cpu(phys_domains, i);
6917 cpu_attach_domain(sd, rd, i);
6924 free_sched_groups(cpu_map);
6929 static cpumask_t *doms_cur; /* current sched domains */
6930 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6933 * Special case: If a kmalloc of a doms_cur partition (array of
6934 * cpumask_t) fails, then fallback to a single sched domain,
6935 * as determined by the single cpumask_t fallback_doms.
6937 static cpumask_t fallback_doms;
6939 void __attribute__((weak)) arch_update_cpu_topology(void)
6944 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6945 * For now this just excludes isolated cpus, but could be used to
6946 * exclude other special cases in the future.
6948 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6952 arch_update_cpu_topology();
6954 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6956 doms_cur = &fallback_doms;
6957 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6958 err = build_sched_domains(doms_cur);
6959 register_sched_domain_sysctl();
6964 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6966 free_sched_groups(cpu_map);
6970 * Detach sched domains from a group of cpus specified in cpu_map
6971 * These cpus will now be attached to the NULL domain
6973 static void detach_destroy_domains(const cpumask_t *cpu_map)
6977 unregister_sched_domain_sysctl();
6979 for_each_cpu_mask(i, *cpu_map)
6980 cpu_attach_domain(NULL, &def_root_domain, i);
6981 synchronize_sched();
6982 arch_destroy_sched_domains(cpu_map);
6986 * Partition sched domains as specified by the 'ndoms_new'
6987 * cpumasks in the array doms_new[] of cpumasks. This compares
6988 * doms_new[] to the current sched domain partitioning, doms_cur[].
6989 * It destroys each deleted domain and builds each new domain.
6991 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6992 * The masks don't intersect (don't overlap.) We should setup one
6993 * sched domain for each mask. CPUs not in any of the cpumasks will
6994 * not be load balanced. If the same cpumask appears both in the
6995 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6998 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6999 * ownership of it and will kfree it when done with it. If the caller
7000 * failed the kmalloc call, then it can pass in doms_new == NULL,
7001 * and partition_sched_domains() will fallback to the single partition
7004 * Call with hotplug lock held
7006 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7012 /* always unregister in case we don't destroy any domains */
7013 unregister_sched_domain_sysctl();
7015 if (doms_new == NULL) {
7017 doms_new = &fallback_doms;
7018 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7021 /* Destroy deleted domains */
7022 for (i = 0; i < ndoms_cur; i++) {
7023 for (j = 0; j < ndoms_new; j++) {
7024 if (cpus_equal(doms_cur[i], doms_new[j]))
7027 /* no match - a current sched domain not in new doms_new[] */
7028 detach_destroy_domains(doms_cur + i);
7033 /* Build new domains */
7034 for (i = 0; i < ndoms_new; i++) {
7035 for (j = 0; j < ndoms_cur; j++) {
7036 if (cpus_equal(doms_new[i], doms_cur[j]))
7039 /* no match - add a new doms_new */
7040 build_sched_domains(doms_new + i);
7045 /* Remember the new sched domains */
7046 if (doms_cur != &fallback_doms)
7048 doms_cur = doms_new;
7049 ndoms_cur = ndoms_new;
7051 register_sched_domain_sysctl();
7056 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7057 int arch_reinit_sched_domains(void)
7062 detach_destroy_domains(&cpu_online_map);
7063 err = arch_init_sched_domains(&cpu_online_map);
7069 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7073 if (buf[0] != '0' && buf[0] != '1')
7077 sched_smt_power_savings = (buf[0] == '1');
7079 sched_mc_power_savings = (buf[0] == '1');
7081 ret = arch_reinit_sched_domains();
7083 return ret ? ret : count;
7086 #ifdef CONFIG_SCHED_MC
7087 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7089 return sprintf(page, "%u\n", sched_mc_power_savings);
7091 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7092 const char *buf, size_t count)
7094 return sched_power_savings_store(buf, count, 0);
7096 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7097 sched_mc_power_savings_store);
7100 #ifdef CONFIG_SCHED_SMT
7101 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7103 return sprintf(page, "%u\n", sched_smt_power_savings);
7105 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7106 const char *buf, size_t count)
7108 return sched_power_savings_store(buf, count, 1);
7110 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7111 sched_smt_power_savings_store);
7114 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7118 #ifdef CONFIG_SCHED_SMT
7120 err = sysfs_create_file(&cls->kset.kobj,
7121 &attr_sched_smt_power_savings.attr);
7123 #ifdef CONFIG_SCHED_MC
7124 if (!err && mc_capable())
7125 err = sysfs_create_file(&cls->kset.kobj,
7126 &attr_sched_mc_power_savings.attr);
7133 * Force a reinitialization of the sched domains hierarchy. The domains
7134 * and groups cannot be updated in place without racing with the balancing
7135 * code, so we temporarily attach all running cpus to the NULL domain
7136 * which will prevent rebalancing while the sched domains are recalculated.
7138 static int update_sched_domains(struct notifier_block *nfb,
7139 unsigned long action, void *hcpu)
7142 case CPU_UP_PREPARE:
7143 case CPU_UP_PREPARE_FROZEN:
7144 case CPU_DOWN_PREPARE:
7145 case CPU_DOWN_PREPARE_FROZEN:
7146 detach_destroy_domains(&cpu_online_map);
7149 case CPU_UP_CANCELED:
7150 case CPU_UP_CANCELED_FROZEN:
7151 case CPU_DOWN_FAILED:
7152 case CPU_DOWN_FAILED_FROZEN:
7154 case CPU_ONLINE_FROZEN:
7156 case CPU_DEAD_FROZEN:
7158 * Fall through and re-initialise the domains.
7165 /* The hotplug lock is already held by cpu_up/cpu_down */
7166 arch_init_sched_domains(&cpu_online_map);
7171 void __init sched_init_smp(void)
7173 cpumask_t non_isolated_cpus;
7176 arch_init_sched_domains(&cpu_online_map);
7177 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7178 if (cpus_empty(non_isolated_cpus))
7179 cpu_set(smp_processor_id(), non_isolated_cpus);
7181 /* XXX: Theoretical race here - CPU may be hotplugged now */
7182 hotcpu_notifier(update_sched_domains, 0);
7184 /* Move init over to a non-isolated CPU */
7185 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7187 sched_init_granularity();
7190 void __init sched_init_smp(void)
7192 sched_init_granularity();
7194 #endif /* CONFIG_SMP */
7196 int in_sched_functions(unsigned long addr)
7198 return in_lock_functions(addr) ||
7199 (addr >= (unsigned long)__sched_text_start
7200 && addr < (unsigned long)__sched_text_end);
7203 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7205 cfs_rq->tasks_timeline = RB_ROOT;
7206 #ifdef CONFIG_FAIR_GROUP_SCHED
7209 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7212 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7214 struct rt_prio_array *array;
7217 array = &rt_rq->active;
7218 for (i = 0; i < MAX_RT_PRIO; i++) {
7219 INIT_LIST_HEAD(array->queue + i);
7220 __clear_bit(i, array->bitmap);
7222 /* delimiter for bitsearch: */
7223 __set_bit(MAX_RT_PRIO, array->bitmap);
7225 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7226 rt_rq->highest_prio = MAX_RT_PRIO;
7229 rt_rq->rt_nr_migratory = 0;
7230 rt_rq->overloaded = 0;
7234 rt_rq->rt_throttled = 0;
7236 #ifdef CONFIG_RT_GROUP_SCHED
7237 rt_rq->rt_nr_boosted = 0;
7242 #ifdef CONFIG_FAIR_GROUP_SCHED
7243 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7244 struct cfs_rq *cfs_rq, struct sched_entity *se,
7247 tg->cfs_rq[cpu] = cfs_rq;
7248 init_cfs_rq(cfs_rq, rq);
7251 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7254 se->cfs_rq = &rq->cfs;
7256 se->load.weight = tg->shares;
7257 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7262 #ifdef CONFIG_RT_GROUP_SCHED
7263 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7264 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7267 tg->rt_rq[cpu] = rt_rq;
7268 init_rt_rq(rt_rq, rq);
7270 rt_rq->rt_se = rt_se;
7272 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7274 tg->rt_se[cpu] = rt_se;
7275 rt_se->rt_rq = &rq->rt;
7276 rt_se->my_q = rt_rq;
7277 rt_se->parent = NULL;
7278 INIT_LIST_HEAD(&rt_se->run_list);
7282 void __init sched_init(void)
7284 int highest_cpu = 0;
7288 init_defrootdomain();
7291 #ifdef CONFIG_GROUP_SCHED
7292 list_add(&init_task_group.list, &task_groups);
7295 for_each_possible_cpu(i) {
7299 spin_lock_init(&rq->lock);
7300 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7303 update_last_tick_seen(rq);
7304 init_cfs_rq(&rq->cfs, rq);
7305 init_rt_rq(&rq->rt, rq);
7306 #ifdef CONFIG_FAIR_GROUP_SCHED
7307 init_task_group.shares = init_task_group_load;
7308 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7309 init_tg_cfs_entry(rq, &init_task_group,
7310 &per_cpu(init_cfs_rq, i),
7311 &per_cpu(init_sched_entity, i), i, 1);
7314 #ifdef CONFIG_RT_GROUP_SCHED
7315 init_task_group.rt_runtime =
7316 sysctl_sched_rt_runtime * NSEC_PER_USEC;
7317 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7318 init_tg_rt_entry(rq, &init_task_group,
7319 &per_cpu(init_rt_rq, i),
7320 &per_cpu(init_sched_rt_entity, i), i, 1);
7322 rq->rt_period_expire = 0;
7323 rq->rt_throttled = 0;
7325 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7326 rq->cpu_load[j] = 0;
7330 rq->active_balance = 0;
7331 rq->next_balance = jiffies;
7334 rq->migration_thread = NULL;
7335 INIT_LIST_HEAD(&rq->migration_queue);
7336 rq_attach_root(rq, &def_root_domain);
7339 atomic_set(&rq->nr_iowait, 0);
7343 set_load_weight(&init_task);
7345 #ifdef CONFIG_PREEMPT_NOTIFIERS
7346 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7350 nr_cpu_ids = highest_cpu + 1;
7351 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7354 #ifdef CONFIG_RT_MUTEXES
7355 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7359 * The boot idle thread does lazy MMU switching as well:
7361 atomic_inc(&init_mm.mm_count);
7362 enter_lazy_tlb(&init_mm, current);
7365 * Make us the idle thread. Technically, schedule() should not be
7366 * called from this thread, however somewhere below it might be,
7367 * but because we are the idle thread, we just pick up running again
7368 * when this runqueue becomes "idle".
7370 init_idle(current, smp_processor_id());
7372 * During early bootup we pretend to be a normal task:
7374 current->sched_class = &fair_sched_class;
7376 scheduler_running = 1;
7379 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7380 void __might_sleep(char *file, int line)
7383 static unsigned long prev_jiffy; /* ratelimiting */
7385 if ((in_atomic() || irqs_disabled()) &&
7386 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7387 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7389 prev_jiffy = jiffies;
7390 printk(KERN_ERR "BUG: sleeping function called from invalid"
7391 " context at %s:%d\n", file, line);
7392 printk("in_atomic():%d, irqs_disabled():%d\n",
7393 in_atomic(), irqs_disabled());
7394 debug_show_held_locks(current);
7395 if (irqs_disabled())
7396 print_irqtrace_events(current);
7401 EXPORT_SYMBOL(__might_sleep);
7404 #ifdef CONFIG_MAGIC_SYSRQ
7405 static void normalize_task(struct rq *rq, struct task_struct *p)
7408 update_rq_clock(rq);
7409 on_rq = p->se.on_rq;
7411 deactivate_task(rq, p, 0);
7412 __setscheduler(rq, p, SCHED_NORMAL, 0);
7414 activate_task(rq, p, 0);
7415 resched_task(rq->curr);
7419 void normalize_rt_tasks(void)
7421 struct task_struct *g, *p;
7422 unsigned long flags;
7425 read_lock_irqsave(&tasklist_lock, flags);
7426 do_each_thread(g, p) {
7428 * Only normalize user tasks:
7433 p->se.exec_start = 0;
7434 #ifdef CONFIG_SCHEDSTATS
7435 p->se.wait_start = 0;
7436 p->se.sleep_start = 0;
7437 p->se.block_start = 0;
7439 task_rq(p)->clock = 0;
7443 * Renice negative nice level userspace
7446 if (TASK_NICE(p) < 0 && p->mm)
7447 set_user_nice(p, 0);
7451 spin_lock(&p->pi_lock);
7452 rq = __task_rq_lock(p);
7454 normalize_task(rq, p);
7456 __task_rq_unlock(rq);
7457 spin_unlock(&p->pi_lock);
7458 } while_each_thread(g, p);
7460 read_unlock_irqrestore(&tasklist_lock, flags);
7463 #endif /* CONFIG_MAGIC_SYSRQ */
7467 * These functions are only useful for the IA64 MCA handling.
7469 * They can only be called when the whole system has been
7470 * stopped - every CPU needs to be quiescent, and no scheduling
7471 * activity can take place. Using them for anything else would
7472 * be a serious bug, and as a result, they aren't even visible
7473 * under any other configuration.
7477 * curr_task - return the current task for a given cpu.
7478 * @cpu: the processor in question.
7480 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7482 struct task_struct *curr_task(int cpu)
7484 return cpu_curr(cpu);
7488 * set_curr_task - set the current task for a given cpu.
7489 * @cpu: the processor in question.
7490 * @p: the task pointer to set.
7492 * Description: This function must only be used when non-maskable interrupts
7493 * are serviced on a separate stack. It allows the architecture to switch the
7494 * notion of the current task on a cpu in a non-blocking manner. This function
7495 * must be called with all CPU's synchronized, and interrupts disabled, the
7496 * and caller must save the original value of the current task (see
7497 * curr_task() above) and restore that value before reenabling interrupts and
7498 * re-starting the system.
7500 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7502 void set_curr_task(int cpu, struct task_struct *p)
7509 #ifdef CONFIG_GROUP_SCHED
7511 #ifdef CONFIG_FAIR_GROUP_SCHED
7512 static void free_fair_sched_group(struct task_group *tg)
7516 for_each_possible_cpu(i) {
7518 kfree(tg->cfs_rq[i]);
7527 static int alloc_fair_sched_group(struct task_group *tg)
7529 struct cfs_rq *cfs_rq;
7530 struct sched_entity *se;
7534 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7537 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7541 tg->shares = NICE_0_LOAD;
7543 for_each_possible_cpu(i) {
7546 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7547 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7551 se = kmalloc_node(sizeof(struct sched_entity),
7552 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7556 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7565 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7567 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7568 &cpu_rq(cpu)->leaf_cfs_rq_list);
7571 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7573 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7576 static inline void free_fair_sched_group(struct task_group *tg)
7580 static inline int alloc_fair_sched_group(struct task_group *tg)
7585 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7589 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7594 #ifdef CONFIG_RT_GROUP_SCHED
7595 static void free_rt_sched_group(struct task_group *tg)
7599 for_each_possible_cpu(i) {
7601 kfree(tg->rt_rq[i]);
7603 kfree(tg->rt_se[i]);
7610 static int alloc_rt_sched_group(struct task_group *tg)
7612 struct rt_rq *rt_rq;
7613 struct sched_rt_entity *rt_se;
7617 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7620 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7626 for_each_possible_cpu(i) {
7629 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7630 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7634 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7635 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7639 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7648 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7650 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7651 &cpu_rq(cpu)->leaf_rt_rq_list);
7654 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7656 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7659 static inline void free_rt_sched_group(struct task_group *tg)
7663 static inline int alloc_rt_sched_group(struct task_group *tg)
7668 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7672 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7677 static void free_sched_group(struct task_group *tg)
7679 free_fair_sched_group(tg);
7680 free_rt_sched_group(tg);
7684 /* allocate runqueue etc for a new task group */
7685 struct task_group *sched_create_group(void)
7687 struct task_group *tg;
7688 unsigned long flags;
7691 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7693 return ERR_PTR(-ENOMEM);
7695 if (!alloc_fair_sched_group(tg))
7698 if (!alloc_rt_sched_group(tg))
7701 spin_lock_irqsave(&task_group_lock, flags);
7702 for_each_possible_cpu(i) {
7703 register_fair_sched_group(tg, i);
7704 register_rt_sched_group(tg, i);
7706 list_add_rcu(&tg->list, &task_groups);
7707 spin_unlock_irqrestore(&task_group_lock, flags);
7712 free_sched_group(tg);
7713 return ERR_PTR(-ENOMEM);
7716 /* rcu callback to free various structures associated with a task group */
7717 static void free_sched_group_rcu(struct rcu_head *rhp)
7719 /* now it should be safe to free those cfs_rqs */
7720 free_sched_group(container_of(rhp, struct task_group, rcu));
7723 /* Destroy runqueue etc associated with a task group */
7724 void sched_destroy_group(struct task_group *tg)
7726 unsigned long flags;
7729 spin_lock_irqsave(&task_group_lock, flags);
7730 for_each_possible_cpu(i) {
7731 unregister_fair_sched_group(tg, i);
7732 unregister_rt_sched_group(tg, i);
7734 list_del_rcu(&tg->list);
7735 spin_unlock_irqrestore(&task_group_lock, flags);
7737 /* wait for possible concurrent references to cfs_rqs complete */
7738 call_rcu(&tg->rcu, free_sched_group_rcu);
7741 /* change task's runqueue when it moves between groups.
7742 * The caller of this function should have put the task in its new group
7743 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7744 * reflect its new group.
7746 void sched_move_task(struct task_struct *tsk)
7749 unsigned long flags;
7752 rq = task_rq_lock(tsk, &flags);
7754 update_rq_clock(rq);
7756 running = task_current(rq, tsk);
7757 on_rq = tsk->se.on_rq;
7760 dequeue_task(rq, tsk, 0);
7761 if (unlikely(running))
7762 tsk->sched_class->put_prev_task(rq, tsk);
7764 set_task_rq(tsk, task_cpu(tsk));
7766 #ifdef CONFIG_FAIR_GROUP_SCHED
7767 if (tsk->sched_class->moved_group)
7768 tsk->sched_class->moved_group(tsk);
7771 if (unlikely(running))
7772 tsk->sched_class->set_curr_task(rq);
7774 enqueue_task(rq, tsk, 0);
7776 task_rq_unlock(rq, &flags);
7779 #ifdef CONFIG_FAIR_GROUP_SCHED
7780 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7782 struct cfs_rq *cfs_rq = se->cfs_rq;
7783 struct rq *rq = cfs_rq->rq;
7786 spin_lock_irq(&rq->lock);
7790 dequeue_entity(cfs_rq, se, 0);
7792 se->load.weight = shares;
7793 se->load.inv_weight = div64_64((1ULL<<32), shares);
7796 enqueue_entity(cfs_rq, se, 0);
7798 spin_unlock_irq(&rq->lock);
7801 static DEFINE_MUTEX(shares_mutex);
7803 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7806 unsigned long flags;
7809 * A weight of 0 or 1 can cause arithmetics problems.
7810 * (The default weight is 1024 - so there's no practical
7811 * limitation from this.)
7816 mutex_lock(&shares_mutex);
7817 if (tg->shares == shares)
7820 spin_lock_irqsave(&task_group_lock, flags);
7821 for_each_possible_cpu(i)
7822 unregister_fair_sched_group(tg, i);
7823 spin_unlock_irqrestore(&task_group_lock, flags);
7825 /* wait for any ongoing reference to this group to finish */
7826 synchronize_sched();
7829 * Now we are free to modify the group's share on each cpu
7830 * w/o tripping rebalance_share or load_balance_fair.
7832 tg->shares = shares;
7833 for_each_possible_cpu(i)
7834 set_se_shares(tg->se[i], shares);
7837 * Enable load balance activity on this group, by inserting it back on
7838 * each cpu's rq->leaf_cfs_rq_list.
7840 spin_lock_irqsave(&task_group_lock, flags);
7841 for_each_possible_cpu(i)
7842 register_fair_sched_group(tg, i);
7843 spin_unlock_irqrestore(&task_group_lock, flags);
7845 mutex_unlock(&shares_mutex);
7849 unsigned long sched_group_shares(struct task_group *tg)
7855 #ifdef CONFIG_RT_GROUP_SCHED
7857 * Ensure that the real time constraints are schedulable.
7859 static DEFINE_MUTEX(rt_constraints_mutex);
7861 static unsigned long to_ratio(u64 period, u64 runtime)
7863 if (runtime == RUNTIME_INF)
7866 return div64_64(runtime << 16, period);
7869 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7871 struct task_group *tgi;
7872 unsigned long total = 0;
7873 unsigned long global_ratio =
7874 to_ratio(sysctl_sched_rt_period,
7875 sysctl_sched_rt_runtime < 0 ?
7876 RUNTIME_INF : sysctl_sched_rt_runtime);
7879 list_for_each_entry_rcu(tgi, &task_groups, list) {
7883 total += to_ratio(period, tgi->rt_runtime);
7887 return total + to_ratio(period, runtime) < global_ratio;
7890 /* Must be called with tasklist_lock held */
7891 static inline int tg_has_rt_tasks(struct task_group *tg)
7893 struct task_struct *g, *p;
7894 do_each_thread(g, p) {
7895 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
7897 } while_each_thread(g, p);
7901 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7903 u64 rt_runtime, rt_period;
7906 rt_period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
7907 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7908 if (rt_runtime_us == -1)
7909 rt_runtime = RUNTIME_INF;
7911 mutex_lock(&rt_constraints_mutex);
7912 read_lock(&tasklist_lock);
7913 if (rt_runtime_us == 0 && tg_has_rt_tasks(tg)) {
7917 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7921 tg->rt_runtime = rt_runtime;
7923 read_unlock(&tasklist_lock);
7924 mutex_unlock(&rt_constraints_mutex);
7929 long sched_group_rt_runtime(struct task_group *tg)
7933 if (tg->rt_runtime == RUNTIME_INF)
7936 rt_runtime_us = tg->rt_runtime;
7937 do_div(rt_runtime_us, NSEC_PER_USEC);
7938 return rt_runtime_us;
7941 #endif /* CONFIG_GROUP_SCHED */
7943 #ifdef CONFIG_CGROUP_SCHED
7945 /* return corresponding task_group object of a cgroup */
7946 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7948 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7949 struct task_group, css);
7952 static struct cgroup_subsys_state *
7953 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7955 struct task_group *tg;
7957 if (!cgrp->parent) {
7958 /* This is early initialization for the top cgroup */
7959 init_task_group.css.cgroup = cgrp;
7960 return &init_task_group.css;
7963 /* we support only 1-level deep hierarchical scheduler atm */
7964 if (cgrp->parent->parent)
7965 return ERR_PTR(-EINVAL);
7967 tg = sched_create_group();
7969 return ERR_PTR(-ENOMEM);
7971 /* Bind the cgroup to task_group object we just created */
7972 tg->css.cgroup = cgrp;
7978 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7980 struct task_group *tg = cgroup_tg(cgrp);
7982 sched_destroy_group(tg);
7986 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7987 struct task_struct *tsk)
7989 #ifdef CONFIG_RT_GROUP_SCHED
7990 /* Don't accept realtime tasks when there is no way for them to run */
7991 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7994 /* We don't support RT-tasks being in separate groups */
7995 if (tsk->sched_class != &fair_sched_class)
8003 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8004 struct cgroup *old_cont, struct task_struct *tsk)
8006 sched_move_task(tsk);
8009 #ifdef CONFIG_FAIR_GROUP_SCHED
8010 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8013 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8016 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8018 struct task_group *tg = cgroup_tg(cgrp);
8020 return (u64) tg->shares;
8024 #ifdef CONFIG_RT_GROUP_SCHED
8025 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8027 const char __user *userbuf,
8028 size_t nbytes, loff_t *unused_ppos)
8037 if (nbytes >= sizeof(buffer))
8039 if (copy_from_user(buffer, userbuf, nbytes))
8042 buffer[nbytes] = 0; /* nul-terminate */
8044 /* strip newline if necessary */
8045 if (nbytes && (buffer[nbytes-1] == '\n'))
8046 buffer[nbytes-1] = 0;
8047 val = simple_strtoll(buffer, &end, 0);
8051 /* Pass to subsystem */
8052 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8058 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8060 char __user *buf, size_t nbytes,
8064 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8065 int len = sprintf(tmp, "%ld\n", val);
8067 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8071 static struct cftype cpu_files[] = {
8072 #ifdef CONFIG_FAIR_GROUP_SCHED
8075 .read_uint = cpu_shares_read_uint,
8076 .write_uint = cpu_shares_write_uint,
8079 #ifdef CONFIG_RT_GROUP_SCHED
8081 .name = "rt_runtime_us",
8082 .read = cpu_rt_runtime_read,
8083 .write = cpu_rt_runtime_write,
8088 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8090 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8093 struct cgroup_subsys cpu_cgroup_subsys = {
8095 .create = cpu_cgroup_create,
8096 .destroy = cpu_cgroup_destroy,
8097 .can_attach = cpu_cgroup_can_attach,
8098 .attach = cpu_cgroup_attach,
8099 .populate = cpu_cgroup_populate,
8100 .subsys_id = cpu_cgroup_subsys_id,
8104 #endif /* CONFIG_CGROUP_SCHED */
8106 #ifdef CONFIG_CGROUP_CPUACCT
8109 * CPU accounting code for task groups.
8111 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8112 * (balbir@in.ibm.com).
8115 /* track cpu usage of a group of tasks */
8117 struct cgroup_subsys_state css;
8118 /* cpuusage holds pointer to a u64-type object on every cpu */
8122 struct cgroup_subsys cpuacct_subsys;
8124 /* return cpu accounting group corresponding to this container */
8125 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
8127 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
8128 struct cpuacct, css);
8131 /* return cpu accounting group to which this task belongs */
8132 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8134 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8135 struct cpuacct, css);
8138 /* create a new cpu accounting group */
8139 static struct cgroup_subsys_state *cpuacct_create(
8140 struct cgroup_subsys *ss, struct cgroup *cont)
8142 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8145 return ERR_PTR(-ENOMEM);
8147 ca->cpuusage = alloc_percpu(u64);
8148 if (!ca->cpuusage) {
8150 return ERR_PTR(-ENOMEM);
8156 /* destroy an existing cpu accounting group */
8158 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8160 struct cpuacct *ca = cgroup_ca(cont);
8162 free_percpu(ca->cpuusage);
8166 /* return total cpu usage (in nanoseconds) of a group */
8167 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8169 struct cpuacct *ca = cgroup_ca(cont);
8170 u64 totalcpuusage = 0;
8173 for_each_possible_cpu(i) {
8174 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8177 * Take rq->lock to make 64-bit addition safe on 32-bit
8180 spin_lock_irq(&cpu_rq(i)->lock);
8181 totalcpuusage += *cpuusage;
8182 spin_unlock_irq(&cpu_rq(i)->lock);
8185 return totalcpuusage;
8188 static struct cftype files[] = {
8191 .read_uint = cpuusage_read,
8195 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8197 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8201 * charge this task's execution time to its accounting group.
8203 * called with rq->lock held.
8205 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8209 if (!cpuacct_subsys.active)
8214 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8216 *cpuusage += cputime;
8220 struct cgroup_subsys cpuacct_subsys = {
8222 .create = cpuacct_create,
8223 .destroy = cpuacct_destroy,
8224 .populate = cpuacct_populate,
8225 .subsys_id = cpuacct_subsys_id,
8227 #endif /* CONFIG_CGROUP_CPUACCT */