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>
70 #include <linux/bootmem.h>
73 #include <asm/irq_regs.h>
76 * Scheduler clock - returns current time in nanosec units.
77 * This is default implementation.
78 * Architectures and sub-architectures can override this.
80 unsigned long long __attribute__((weak)) sched_clock(void)
82 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
126 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
127 * Since cpu_power is a 'constant', we can use a reciprocal divide.
129 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
131 return reciprocal_divide(load, sg->reciprocal_cpu_power);
135 * Each time a sched group cpu_power is changed,
136 * we must compute its reciprocal value
138 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
140 sg->__cpu_power += val;
141 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
145 static inline int rt_policy(int policy)
147 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
152 static inline int task_has_rt_policy(struct task_struct *p)
154 return rt_policy(p->policy);
158 * This is the priority-queue data structure of the RT scheduling class:
160 struct rt_prio_array {
161 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
162 struct list_head queue[MAX_RT_PRIO];
165 struct rt_bandwidth {
166 /* nests inside the rq lock: */
167 spinlock_t rt_runtime_lock;
170 struct hrtimer rt_period_timer;
173 static struct rt_bandwidth def_rt_bandwidth;
175 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
177 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
179 struct rt_bandwidth *rt_b =
180 container_of(timer, struct rt_bandwidth, rt_period_timer);
186 now = hrtimer_cb_get_time(timer);
187 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
192 idle = do_sched_rt_period_timer(rt_b, overrun);
195 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
199 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
201 rt_b->rt_period = ns_to_ktime(period);
202 rt_b->rt_runtime = runtime;
204 spin_lock_init(&rt_b->rt_runtime_lock);
206 hrtimer_init(&rt_b->rt_period_timer,
207 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
208 rt_b->rt_period_timer.function = sched_rt_period_timer;
209 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
212 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
216 if (rt_b->rt_runtime == RUNTIME_INF)
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 spin_lock(&rt_b->rt_runtime_lock);
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
228 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
229 hrtimer_start(&rt_b->rt_period_timer,
230 rt_b->rt_period_timer.expires,
233 spin_unlock(&rt_b->rt_runtime_lock);
236 #ifdef CONFIG_RT_GROUP_SCHED
237 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
239 hrtimer_cancel(&rt_b->rt_period_timer);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
276 #ifdef CONFIG_USER_SCHED
280 * Every UID task group (including init_task_group aka UID-0) will
281 * be a child to this group.
283 struct task_group root_task_group;
285 #ifdef CONFIG_FAIR_GROUP_SCHED
286 /* Default task group's sched entity on each cpu */
287 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
288 /* Default task group's cfs_rq on each cpu */
289 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
292 #ifdef CONFIG_RT_GROUP_SCHED
293 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
294 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
297 #define root_task_group init_task_group
300 /* task_group_lock serializes add/remove of task groups and also changes to
301 * a task group's cpu shares.
303 static DEFINE_SPINLOCK(task_group_lock);
305 /* doms_cur_mutex serializes access to doms_cur[] array */
306 static DEFINE_MUTEX(doms_cur_mutex);
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 #ifdef CONFIG_USER_SCHED
310 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
318 /* Default task group.
319 * Every task in system belong to this group at bootup.
321 struct task_group init_task_group;
323 /* return group to which a task belongs */
324 static inline struct task_group *task_group(struct task_struct *p)
326 struct task_group *tg;
328 #ifdef CONFIG_USER_SCHED
330 #elif defined(CONFIG_CGROUP_SCHED)
331 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
332 struct task_group, css);
334 tg = &init_task_group;
339 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
340 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
342 #ifdef CONFIG_FAIR_GROUP_SCHED
343 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
344 p->se.parent = task_group(p)->se[cpu];
347 #ifdef CONFIG_RT_GROUP_SCHED
348 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
349 p->rt.parent = task_group(p)->rt_se[cpu];
353 static inline void lock_doms_cur(void)
355 mutex_lock(&doms_cur_mutex);
358 static inline void unlock_doms_cur(void)
360 mutex_unlock(&doms_cur_mutex);
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
366 static inline void lock_doms_cur(void) { }
367 static inline void unlock_doms_cur(void) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
379 struct rb_root tasks_timeline;
380 struct rb_node *rb_leftmost;
381 struct rb_node *rb_load_balance_curr;
382 /* 'curr' points to currently running entity on this cfs_rq.
383 * It is set to NULL otherwise (i.e when none are currently running).
385 struct sched_entity *curr, *next;
387 unsigned long nr_spread_over;
389 #ifdef CONFIG_FAIR_GROUP_SCHED
390 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
393 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
394 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
395 * (like users, containers etc.)
397 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
398 * list is used during load balance.
400 struct list_head leaf_cfs_rq_list;
401 struct task_group *tg; /* group that "owns" this runqueue */
405 /* Real-Time classes' related field in a runqueue: */
407 struct rt_prio_array active;
408 unsigned long rt_nr_running;
409 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
410 int highest_prio; /* highest queued rt task prio */
413 unsigned long rt_nr_migratory;
419 /* Nests inside the rq lock: */
420 spinlock_t rt_runtime_lock;
422 #ifdef CONFIG_RT_GROUP_SCHED
423 unsigned long rt_nr_boosted;
426 struct list_head leaf_rt_rq_list;
427 struct task_group *tg;
428 struct sched_rt_entity *rt_se;
435 * We add the notion of a root-domain which will be used to define per-domain
436 * variables. Each exclusive cpuset essentially defines an island domain by
437 * fully partitioning the member cpus from any other cpuset. Whenever a new
438 * exclusive cpuset is created, we also create and attach a new root-domain
448 * The "RT overload" flag: it gets set if a CPU has more than
449 * one runnable RT task.
456 * By default the system creates a single root-domain with all cpus as
457 * members (mimicking the global state we have today).
459 static struct root_domain def_root_domain;
464 * This is the main, per-CPU runqueue data structure.
466 * Locking rule: those places that want to lock multiple runqueues
467 * (such as the load balancing or the thread migration code), lock
468 * acquire operations must be ordered by ascending &runqueue.
475 * nr_running and cpu_load should be in the same cacheline because
476 * remote CPUs use both these fields when doing load calculation.
478 unsigned long nr_running;
479 #define CPU_LOAD_IDX_MAX 5
480 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
481 unsigned char idle_at_tick;
483 unsigned long last_tick_seen;
484 unsigned char in_nohz_recently;
486 /* capture load from *all* tasks on this cpu: */
487 struct load_weight load;
488 unsigned long nr_load_updates;
494 #ifdef CONFIG_FAIR_GROUP_SCHED
495 /* list of leaf cfs_rq on this cpu: */
496 struct list_head leaf_cfs_rq_list;
498 #ifdef CONFIG_RT_GROUP_SCHED
499 struct list_head leaf_rt_rq_list;
503 * This is part of a global counter where only the total sum
504 * over all CPUs matters. A task can increase this counter on
505 * one CPU and if it got migrated afterwards it may decrease
506 * it on another CPU. Always updated under the runqueue lock:
508 unsigned long nr_uninterruptible;
510 struct task_struct *curr, *idle;
511 unsigned long next_balance;
512 struct mm_struct *prev_mm;
514 u64 clock, prev_clock_raw;
517 unsigned int clock_warps, clock_overflows, clock_underflows;
519 unsigned int clock_deep_idle_events;
525 struct root_domain *rd;
526 struct sched_domain *sd;
528 /* For active balancing */
531 /* cpu of this runqueue: */
534 struct task_struct *migration_thread;
535 struct list_head migration_queue;
538 #ifdef CONFIG_SCHED_HRTICK
539 unsigned long hrtick_flags;
540 ktime_t hrtick_expire;
541 struct hrtimer hrtick_timer;
544 #ifdef CONFIG_SCHEDSTATS
546 struct sched_info rq_sched_info;
548 /* sys_sched_yield() stats */
549 unsigned int yld_exp_empty;
550 unsigned int yld_act_empty;
551 unsigned int yld_both_empty;
552 unsigned int yld_count;
554 /* schedule() stats */
555 unsigned int sched_switch;
556 unsigned int sched_count;
557 unsigned int sched_goidle;
559 /* try_to_wake_up() stats */
560 unsigned int ttwu_count;
561 unsigned int ttwu_local;
564 unsigned int bkl_count;
566 struct lock_class_key rq_lock_key;
569 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
571 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
573 rq->curr->sched_class->check_preempt_curr(rq, p);
576 static inline int cpu_of(struct rq *rq)
586 static inline bool nohz_on(int cpu)
588 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
591 static inline u64 max_skipped_ticks(struct rq *rq)
593 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
596 static inline void update_last_tick_seen(struct rq *rq)
598 rq->last_tick_seen = jiffies;
601 static inline u64 max_skipped_ticks(struct rq *rq)
606 static inline void update_last_tick_seen(struct rq *rq)
612 * Update the per-runqueue clock, as finegrained as the platform can give
613 * us, but without assuming monotonicity, etc.:
615 static void __update_rq_clock(struct rq *rq)
617 u64 prev_raw = rq->prev_clock_raw;
618 u64 now = sched_clock();
619 s64 delta = now - prev_raw;
620 u64 clock = rq->clock;
622 #ifdef CONFIG_SCHED_DEBUG
623 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
626 * Protect against sched_clock() occasionally going backwards:
628 if (unlikely(delta < 0)) {
633 * Catch too large forward jumps too:
635 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
636 u64 max_time = rq->tick_timestamp + max_jump;
638 if (unlikely(clock + delta > max_time)) {
639 if (clock < max_time)
643 rq->clock_overflows++;
645 if (unlikely(delta > rq->clock_max_delta))
646 rq->clock_max_delta = delta;
651 rq->prev_clock_raw = now;
655 static void update_rq_clock(struct rq *rq)
657 if (likely(smp_processor_id() == cpu_of(rq)))
658 __update_rq_clock(rq);
662 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
663 * See detach_destroy_domains: synchronize_sched for details.
665 * The domain tree of any CPU may only be accessed from within
666 * preempt-disabled sections.
668 #define for_each_domain(cpu, __sd) \
669 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
671 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
672 #define this_rq() (&__get_cpu_var(runqueues))
673 #define task_rq(p) cpu_rq(task_cpu(p))
674 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
677 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
679 #ifdef CONFIG_SCHED_DEBUG
680 # define const_debug __read_mostly
682 # define const_debug static const
686 * Debugging: various feature bits
689 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
690 SCHED_FEAT_WAKEUP_PREEMPT = 2,
691 SCHED_FEAT_START_DEBIT = 4,
692 SCHED_FEAT_AFFINE_WAKEUPS = 8,
693 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
694 SCHED_FEAT_SYNC_WAKEUPS = 32,
695 SCHED_FEAT_HRTICK = 64,
696 SCHED_FEAT_DOUBLE_TICK = 128,
697 SCHED_FEAT_NORMALIZED_SLEEPER = 256,
700 const_debug unsigned int sysctl_sched_features =
701 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
702 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
703 SCHED_FEAT_START_DEBIT * 1 |
704 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
705 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
706 SCHED_FEAT_SYNC_WAKEUPS * 1 |
707 SCHED_FEAT_HRTICK * 1 |
708 SCHED_FEAT_DOUBLE_TICK * 0 |
709 SCHED_FEAT_NORMALIZED_SLEEPER * 1;
711 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
714 * Number of tasks to iterate in a single balance run.
715 * Limited because this is done with IRQs disabled.
717 const_debug unsigned int sysctl_sched_nr_migrate = 32;
720 * period over which we measure -rt task cpu usage in us.
723 unsigned int sysctl_sched_rt_period = 1000000;
725 static __read_mostly int scheduler_running;
728 * part of the period that we allow rt tasks to run in us.
731 int sysctl_sched_rt_runtime = 950000;
733 static inline u64 global_rt_period(void)
735 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
738 static inline u64 global_rt_runtime(void)
740 if (sysctl_sched_rt_period < 0)
743 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
746 static const unsigned long long time_sync_thresh = 100000;
748 static DEFINE_PER_CPU(unsigned long long, time_offset);
749 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
752 * Global lock which we take every now and then to synchronize
753 * the CPUs time. This method is not warp-safe, but it's good
754 * enough to synchronize slowly diverging time sources and thus
755 * it's good enough for tracing:
757 static DEFINE_SPINLOCK(time_sync_lock);
758 static unsigned long long prev_global_time;
760 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
764 spin_lock_irqsave(&time_sync_lock, flags);
766 if (time < prev_global_time) {
767 per_cpu(time_offset, cpu) += prev_global_time - time;
768 time = prev_global_time;
770 prev_global_time = time;
773 spin_unlock_irqrestore(&time_sync_lock, flags);
778 static unsigned long long __cpu_clock(int cpu)
780 unsigned long long now;
785 * Only call sched_clock() if the scheduler has already been
786 * initialized (some code might call cpu_clock() very early):
788 if (unlikely(!scheduler_running))
791 local_irq_save(flags);
795 local_irq_restore(flags);
801 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
802 * clock constructed from sched_clock():
804 unsigned long long cpu_clock(int cpu)
806 unsigned long long prev_cpu_time, time, delta_time;
808 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
809 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
810 delta_time = time-prev_cpu_time;
812 if (unlikely(delta_time > time_sync_thresh))
813 time = __sync_cpu_clock(time, cpu);
817 EXPORT_SYMBOL_GPL(cpu_clock);
819 #ifndef prepare_arch_switch
820 # define prepare_arch_switch(next) do { } while (0)
822 #ifndef finish_arch_switch
823 # define finish_arch_switch(prev) do { } while (0)
826 static inline int task_current(struct rq *rq, struct task_struct *p)
828 return rq->curr == p;
831 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
832 static inline int task_running(struct rq *rq, struct task_struct *p)
834 return task_current(rq, p);
837 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
841 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
843 #ifdef CONFIG_DEBUG_SPINLOCK
844 /* this is a valid case when another task releases the spinlock */
845 rq->lock.owner = current;
848 * If we are tracking spinlock dependencies then we have to
849 * fix up the runqueue lock - which gets 'carried over' from
852 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
854 spin_unlock_irq(&rq->lock);
857 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
858 static inline int task_running(struct rq *rq, struct task_struct *p)
863 return task_current(rq, p);
867 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871 * We can optimise this out completely for !SMP, because the
872 * SMP rebalancing from interrupt is the only thing that cares
877 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
878 spin_unlock_irq(&rq->lock);
880 spin_unlock(&rq->lock);
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
888 * After ->oncpu is cleared, the task can be moved to a different CPU.
889 * We must ensure this doesn't happen until the switch is completely
895 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
902 * __task_rq_lock - lock the runqueue a given task resides on.
903 * Must be called interrupts disabled.
905 static inline struct rq *__task_rq_lock(struct task_struct *p)
909 struct rq *rq = task_rq(p);
910 spin_lock(&rq->lock);
911 if (likely(rq == task_rq(p)))
913 spin_unlock(&rq->lock);
918 * task_rq_lock - lock the runqueue a given task resides on and disable
919 * interrupts. Note the ordering: we can safely lookup the task_rq without
920 * explicitly disabling preemption.
922 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
928 local_irq_save(*flags);
930 spin_lock(&rq->lock);
931 if (likely(rq == task_rq(p)))
933 spin_unlock_irqrestore(&rq->lock, *flags);
937 static void __task_rq_unlock(struct rq *rq)
940 spin_unlock(&rq->lock);
943 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
946 spin_unlock_irqrestore(&rq->lock, *flags);
950 * this_rq_lock - lock this runqueue and disable interrupts.
952 static struct rq *this_rq_lock(void)
959 spin_lock(&rq->lock);
965 * We are going deep-idle (irqs are disabled):
967 void sched_clock_idle_sleep_event(void)
969 struct rq *rq = cpu_rq(smp_processor_id());
971 spin_lock(&rq->lock);
972 __update_rq_clock(rq);
973 spin_unlock(&rq->lock);
974 rq->clock_deep_idle_events++;
976 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
979 * We just idled delta nanoseconds (called with irqs disabled):
981 void sched_clock_idle_wakeup_event(u64 delta_ns)
983 struct rq *rq = cpu_rq(smp_processor_id());
984 u64 now = sched_clock();
986 rq->idle_clock += delta_ns;
988 * Override the previous timestamp and ignore all
989 * sched_clock() deltas that occured while we idled,
990 * and use the PM-provided delta_ns to advance the
993 spin_lock(&rq->lock);
994 rq->prev_clock_raw = now;
995 rq->clock += delta_ns;
996 spin_unlock(&rq->lock);
997 touch_softlockup_watchdog();
999 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
1001 static void __resched_task(struct task_struct *p, int tif_bit);
1003 static inline void resched_task(struct task_struct *p)
1005 __resched_task(p, TIF_NEED_RESCHED);
1008 #ifdef CONFIG_SCHED_HRTICK
1010 * Use HR-timers to deliver accurate preemption points.
1012 * Its all a bit involved since we cannot program an hrt while holding the
1013 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1016 * When we get rescheduled we reprogram the hrtick_timer outside of the
1019 static inline void resched_hrt(struct task_struct *p)
1021 __resched_task(p, TIF_HRTICK_RESCHED);
1024 static inline void resched_rq(struct rq *rq)
1026 unsigned long flags;
1028 spin_lock_irqsave(&rq->lock, flags);
1029 resched_task(rq->curr);
1030 spin_unlock_irqrestore(&rq->lock, flags);
1034 HRTICK_SET, /* re-programm hrtick_timer */
1035 HRTICK_RESET, /* not a new slice */
1040 * - enabled by features
1041 * - hrtimer is actually high res
1043 static inline int hrtick_enabled(struct rq *rq)
1045 if (!sched_feat(HRTICK))
1047 return hrtimer_is_hres_active(&rq->hrtick_timer);
1051 * Called to set the hrtick timer state.
1053 * called with rq->lock held and irqs disabled
1055 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1057 assert_spin_locked(&rq->lock);
1060 * preempt at: now + delay
1063 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1065 * indicate we need to program the timer
1067 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1069 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1072 * New slices are called from the schedule path and don't need a
1073 * forced reschedule.
1076 resched_hrt(rq->curr);
1079 static void hrtick_clear(struct rq *rq)
1081 if (hrtimer_active(&rq->hrtick_timer))
1082 hrtimer_cancel(&rq->hrtick_timer);
1086 * Update the timer from the possible pending state.
1088 static void hrtick_set(struct rq *rq)
1092 unsigned long flags;
1094 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1096 spin_lock_irqsave(&rq->lock, flags);
1097 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1098 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1099 time = rq->hrtick_expire;
1100 clear_thread_flag(TIF_HRTICK_RESCHED);
1101 spin_unlock_irqrestore(&rq->lock, flags);
1104 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1105 if (reset && !hrtimer_active(&rq->hrtick_timer))
1112 * High-resolution timer tick.
1113 * Runs from hardirq context with interrupts disabled.
1115 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1117 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1119 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1121 spin_lock(&rq->lock);
1122 __update_rq_clock(rq);
1123 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1124 spin_unlock(&rq->lock);
1126 return HRTIMER_NORESTART;
1129 static inline void init_rq_hrtick(struct rq *rq)
1131 rq->hrtick_flags = 0;
1132 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1133 rq->hrtick_timer.function = hrtick;
1134 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1137 void hrtick_resched(void)
1140 unsigned long flags;
1142 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1145 local_irq_save(flags);
1146 rq = cpu_rq(smp_processor_id());
1148 local_irq_restore(flags);
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void hrtick_set(struct rq *rq)
1159 static inline void init_rq_hrtick(struct rq *rq)
1163 void hrtick_resched(void)
1169 * resched_task - mark a task 'to be rescheduled now'.
1171 * On UP this means the setting of the need_resched flag, on SMP it
1172 * might also involve a cross-CPU call to trigger the scheduler on
1177 #ifndef tsk_is_polling
1178 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1181 static void __resched_task(struct task_struct *p, int tif_bit)
1185 assert_spin_locked(&task_rq(p)->lock);
1187 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1190 set_tsk_thread_flag(p, tif_bit);
1193 if (cpu == smp_processor_id())
1196 /* NEED_RESCHED must be visible before we test polling */
1198 if (!tsk_is_polling(p))
1199 smp_send_reschedule(cpu);
1202 static void resched_cpu(int cpu)
1204 struct rq *rq = cpu_rq(cpu);
1205 unsigned long flags;
1207 if (!spin_trylock_irqsave(&rq->lock, flags))
1209 resched_task(cpu_curr(cpu));
1210 spin_unlock_irqrestore(&rq->lock, flags);
1215 * When add_timer_on() enqueues a timer into the timer wheel of an
1216 * idle CPU then this timer might expire before the next timer event
1217 * which is scheduled to wake up that CPU. In case of a completely
1218 * idle system the next event might even be infinite time into the
1219 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1220 * leaves the inner idle loop so the newly added timer is taken into
1221 * account when the CPU goes back to idle and evaluates the timer
1222 * wheel for the next timer event.
1224 void wake_up_idle_cpu(int cpu)
1226 struct rq *rq = cpu_rq(cpu);
1228 if (cpu == smp_processor_id())
1232 * This is safe, as this function is called with the timer
1233 * wheel base lock of (cpu) held. When the CPU is on the way
1234 * to idle and has not yet set rq->curr to idle then it will
1235 * be serialized on the timer wheel base lock and take the new
1236 * timer into account automatically.
1238 if (rq->curr != rq->idle)
1242 * We can set TIF_RESCHED on the idle task of the other CPU
1243 * lockless. The worst case is that the other CPU runs the
1244 * idle task through an additional NOOP schedule()
1246 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1248 /* NEED_RESCHED must be visible before we test polling */
1250 if (!tsk_is_polling(rq->idle))
1251 smp_send_reschedule(cpu);
1256 static void __resched_task(struct task_struct *p, int tif_bit)
1258 assert_spin_locked(&task_rq(p)->lock);
1259 set_tsk_thread_flag(p, tif_bit);
1263 #if BITS_PER_LONG == 32
1264 # define WMULT_CONST (~0UL)
1266 # define WMULT_CONST (1UL << 32)
1269 #define WMULT_SHIFT 32
1272 * Shift right and round:
1274 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1276 static unsigned long
1277 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1278 struct load_weight *lw)
1282 if (unlikely(!lw->inv_weight))
1283 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1285 tmp = (u64)delta_exec * weight;
1287 * Check whether we'd overflow the 64-bit multiplication:
1289 if (unlikely(tmp > WMULT_CONST))
1290 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1293 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1295 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1298 static inline unsigned long
1299 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1301 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1304 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1310 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1317 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1318 * of tasks with abnormal "nice" values across CPUs the contribution that
1319 * each task makes to its run queue's load is weighted according to its
1320 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1321 * scaled version of the new time slice allocation that they receive on time
1325 #define WEIGHT_IDLEPRIO 2
1326 #define WMULT_IDLEPRIO (1 << 31)
1329 * Nice levels are multiplicative, with a gentle 10% change for every
1330 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1331 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1332 * that remained on nice 0.
1334 * The "10% effect" is relative and cumulative: from _any_ nice level,
1335 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1336 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1337 * If a task goes up by ~10% and another task goes down by ~10% then
1338 * the relative distance between them is ~25%.)
1340 static const int prio_to_weight[40] = {
1341 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1342 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1343 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1344 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1345 /* 0 */ 1024, 820, 655, 526, 423,
1346 /* 5 */ 335, 272, 215, 172, 137,
1347 /* 10 */ 110, 87, 70, 56, 45,
1348 /* 15 */ 36, 29, 23, 18, 15,
1352 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1354 * In cases where the weight does not change often, we can use the
1355 * precalculated inverse to speed up arithmetics by turning divisions
1356 * into multiplications:
1358 static const u32 prio_to_wmult[40] = {
1359 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1360 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1361 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1362 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1363 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1364 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1365 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1366 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1369 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1372 * runqueue iterator, to support SMP load-balancing between different
1373 * scheduling classes, without having to expose their internal data
1374 * structures to the load-balancing proper:
1376 struct rq_iterator {
1378 struct task_struct *(*start)(void *);
1379 struct task_struct *(*next)(void *);
1383 static unsigned long
1384 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1385 unsigned long max_load_move, struct sched_domain *sd,
1386 enum cpu_idle_type idle, int *all_pinned,
1387 int *this_best_prio, struct rq_iterator *iterator);
1390 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1391 struct sched_domain *sd, enum cpu_idle_type idle,
1392 struct rq_iterator *iterator);
1395 #ifdef CONFIG_CGROUP_CPUACCT
1396 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1402 static unsigned long source_load(int cpu, int type);
1403 static unsigned long target_load(int cpu, int type);
1404 static unsigned long cpu_avg_load_per_task(int cpu);
1405 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1406 #endif /* CONFIG_SMP */
1408 #include "sched_stats.h"
1409 #include "sched_idletask.c"
1410 #include "sched_fair.c"
1411 #include "sched_rt.c"
1412 #ifdef CONFIG_SCHED_DEBUG
1413 # include "sched_debug.c"
1416 #define sched_class_highest (&rt_sched_class)
1418 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1420 update_load_add(&rq->load, p->se.load.weight);
1423 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1425 update_load_sub(&rq->load, p->se.load.weight);
1428 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1434 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1440 static void set_load_weight(struct task_struct *p)
1442 if (task_has_rt_policy(p)) {
1443 p->se.load.weight = prio_to_weight[0] * 2;
1444 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1449 * SCHED_IDLE tasks get minimal weight:
1451 if (p->policy == SCHED_IDLE) {
1452 p->se.load.weight = WEIGHT_IDLEPRIO;
1453 p->se.load.inv_weight = WMULT_IDLEPRIO;
1457 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1458 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1461 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1463 sched_info_queued(p);
1464 p->sched_class->enqueue_task(rq, p, wakeup);
1468 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1470 p->sched_class->dequeue_task(rq, p, sleep);
1475 * __normal_prio - return the priority that is based on the static prio
1477 static inline int __normal_prio(struct task_struct *p)
1479 return p->static_prio;
1483 * Calculate the expected normal priority: i.e. priority
1484 * without taking RT-inheritance into account. Might be
1485 * boosted by interactivity modifiers. Changes upon fork,
1486 * setprio syscalls, and whenever the interactivity
1487 * estimator recalculates.
1489 static inline int normal_prio(struct task_struct *p)
1493 if (task_has_rt_policy(p))
1494 prio = MAX_RT_PRIO-1 - p->rt_priority;
1496 prio = __normal_prio(p);
1501 * Calculate the current priority, i.e. the priority
1502 * taken into account by the scheduler. This value might
1503 * be boosted by RT tasks, or might be boosted by
1504 * interactivity modifiers. Will be RT if the task got
1505 * RT-boosted. If not then it returns p->normal_prio.
1507 static int effective_prio(struct task_struct *p)
1509 p->normal_prio = normal_prio(p);
1511 * If we are RT tasks or we were boosted to RT priority,
1512 * keep the priority unchanged. Otherwise, update priority
1513 * to the normal priority:
1515 if (!rt_prio(p->prio))
1516 return p->normal_prio;
1521 * activate_task - move a task to the runqueue.
1523 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1525 if (task_contributes_to_load(p))
1526 rq->nr_uninterruptible--;
1528 enqueue_task(rq, p, wakeup);
1529 inc_nr_running(p, rq);
1533 * deactivate_task - remove a task from the runqueue.
1535 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1537 if (task_contributes_to_load(p))
1538 rq->nr_uninterruptible++;
1540 dequeue_task(rq, p, sleep);
1541 dec_nr_running(p, rq);
1545 * task_curr - is this task currently executing on a CPU?
1546 * @p: the task in question.
1548 inline int task_curr(const struct task_struct *p)
1550 return cpu_curr(task_cpu(p)) == p;
1553 /* Used instead of source_load when we know the type == 0 */
1554 unsigned long weighted_cpuload(const int cpu)
1556 return cpu_rq(cpu)->load.weight;
1559 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1561 set_task_rq(p, cpu);
1564 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1565 * successfuly executed on another CPU. We must ensure that updates of
1566 * per-task data have been completed by this moment.
1569 task_thread_info(p)->cpu = cpu;
1573 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1574 const struct sched_class *prev_class,
1575 int oldprio, int running)
1577 if (prev_class != p->sched_class) {
1578 if (prev_class->switched_from)
1579 prev_class->switched_from(rq, p, running);
1580 p->sched_class->switched_to(rq, p, running);
1582 p->sched_class->prio_changed(rq, p, oldprio, running);
1588 * Is this task likely cache-hot:
1591 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1596 * Buddy candidates are cache hot:
1598 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1601 if (p->sched_class != &fair_sched_class)
1604 if (sysctl_sched_migration_cost == -1)
1606 if (sysctl_sched_migration_cost == 0)
1609 delta = now - p->se.exec_start;
1611 return delta < (s64)sysctl_sched_migration_cost;
1615 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1617 int old_cpu = task_cpu(p);
1618 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1619 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1620 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1623 clock_offset = old_rq->clock - new_rq->clock;
1625 #ifdef CONFIG_SCHEDSTATS
1626 if (p->se.wait_start)
1627 p->se.wait_start -= clock_offset;
1628 if (p->se.sleep_start)
1629 p->se.sleep_start -= clock_offset;
1630 if (p->se.block_start)
1631 p->se.block_start -= clock_offset;
1632 if (old_cpu != new_cpu) {
1633 schedstat_inc(p, se.nr_migrations);
1634 if (task_hot(p, old_rq->clock, NULL))
1635 schedstat_inc(p, se.nr_forced2_migrations);
1638 p->se.vruntime -= old_cfsrq->min_vruntime -
1639 new_cfsrq->min_vruntime;
1641 __set_task_cpu(p, new_cpu);
1644 struct migration_req {
1645 struct list_head list;
1647 struct task_struct *task;
1650 struct completion done;
1654 * The task's runqueue lock must be held.
1655 * Returns true if you have to wait for migration thread.
1658 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1660 struct rq *rq = task_rq(p);
1663 * If the task is not on a runqueue (and not running), then
1664 * it is sufficient to simply update the task's cpu field.
1666 if (!p->se.on_rq && !task_running(rq, p)) {
1667 set_task_cpu(p, dest_cpu);
1671 init_completion(&req->done);
1673 req->dest_cpu = dest_cpu;
1674 list_add(&req->list, &rq->migration_queue);
1680 * wait_task_inactive - wait for a thread to unschedule.
1682 * The caller must ensure that the task *will* unschedule sometime soon,
1683 * else this function might spin for a *long* time. This function can't
1684 * be called with interrupts off, or it may introduce deadlock with
1685 * smp_call_function() if an IPI is sent by the same process we are
1686 * waiting to become inactive.
1688 void wait_task_inactive(struct task_struct *p)
1690 unsigned long flags;
1696 * We do the initial early heuristics without holding
1697 * any task-queue locks at all. We'll only try to get
1698 * the runqueue lock when things look like they will
1704 * If the task is actively running on another CPU
1705 * still, just relax and busy-wait without holding
1708 * NOTE! Since we don't hold any locks, it's not
1709 * even sure that "rq" stays as the right runqueue!
1710 * But we don't care, since "task_running()" will
1711 * return false if the runqueue has changed and p
1712 * is actually now running somewhere else!
1714 while (task_running(rq, p))
1718 * Ok, time to look more closely! We need the rq
1719 * lock now, to be *sure*. If we're wrong, we'll
1720 * just go back and repeat.
1722 rq = task_rq_lock(p, &flags);
1723 running = task_running(rq, p);
1724 on_rq = p->se.on_rq;
1725 task_rq_unlock(rq, &flags);
1728 * Was it really running after all now that we
1729 * checked with the proper locks actually held?
1731 * Oops. Go back and try again..
1733 if (unlikely(running)) {
1739 * It's not enough that it's not actively running,
1740 * it must be off the runqueue _entirely_, and not
1743 * So if it wa still runnable (but just not actively
1744 * running right now), it's preempted, and we should
1745 * yield - it could be a while.
1747 if (unlikely(on_rq)) {
1748 schedule_timeout_uninterruptible(1);
1753 * Ahh, all good. It wasn't running, and it wasn't
1754 * runnable, which means that it will never become
1755 * running in the future either. We're all done!
1762 * kick_process - kick a running thread to enter/exit the kernel
1763 * @p: the to-be-kicked thread
1765 * Cause a process which is running on another CPU to enter
1766 * kernel-mode, without any delay. (to get signals handled.)
1768 * NOTE: this function doesnt have to take the runqueue lock,
1769 * because all it wants to ensure is that the remote task enters
1770 * the kernel. If the IPI races and the task has been migrated
1771 * to another CPU then no harm is done and the purpose has been
1774 void kick_process(struct task_struct *p)
1780 if ((cpu != smp_processor_id()) && task_curr(p))
1781 smp_send_reschedule(cpu);
1786 * Return a low guess at the load of a migration-source cpu weighted
1787 * according to the scheduling class and "nice" value.
1789 * We want to under-estimate the load of migration sources, to
1790 * balance conservatively.
1792 static unsigned long source_load(int cpu, int type)
1794 struct rq *rq = cpu_rq(cpu);
1795 unsigned long total = weighted_cpuload(cpu);
1800 return min(rq->cpu_load[type-1], total);
1804 * Return a high guess at the load of a migration-target cpu weighted
1805 * according to the scheduling class and "nice" value.
1807 static unsigned long target_load(int cpu, int type)
1809 struct rq *rq = cpu_rq(cpu);
1810 unsigned long total = weighted_cpuload(cpu);
1815 return max(rq->cpu_load[type-1], total);
1819 * Return the average load per task on the cpu's run queue
1821 static unsigned long cpu_avg_load_per_task(int cpu)
1823 struct rq *rq = cpu_rq(cpu);
1824 unsigned long total = weighted_cpuload(cpu);
1825 unsigned long n = rq->nr_running;
1827 return n ? total / n : SCHED_LOAD_SCALE;
1831 * find_idlest_group finds and returns the least busy CPU group within the
1834 static struct sched_group *
1835 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1837 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1838 unsigned long min_load = ULONG_MAX, this_load = 0;
1839 int load_idx = sd->forkexec_idx;
1840 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1843 unsigned long load, avg_load;
1847 /* Skip over this group if it has no CPUs allowed */
1848 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1851 local_group = cpu_isset(this_cpu, group->cpumask);
1853 /* Tally up the load of all CPUs in the group */
1856 for_each_cpu_mask(i, group->cpumask) {
1857 /* Bias balancing toward cpus of our domain */
1859 load = source_load(i, load_idx);
1861 load = target_load(i, load_idx);
1866 /* Adjust by relative CPU power of the group */
1867 avg_load = sg_div_cpu_power(group,
1868 avg_load * SCHED_LOAD_SCALE);
1871 this_load = avg_load;
1873 } else if (avg_load < min_load) {
1874 min_load = avg_load;
1877 } while (group = group->next, group != sd->groups);
1879 if (!idlest || 100*this_load < imbalance*min_load)
1885 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1888 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1891 unsigned long load, min_load = ULONG_MAX;
1895 /* Traverse only the allowed CPUs */
1896 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1898 for_each_cpu_mask(i, *tmp) {
1899 load = weighted_cpuload(i);
1901 if (load < min_load || (load == min_load && i == this_cpu)) {
1911 * sched_balance_self: balance the current task (running on cpu) in domains
1912 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1915 * Balance, ie. select the least loaded group.
1917 * Returns the target CPU number, or the same CPU if no balancing is needed.
1919 * preempt must be disabled.
1921 static int sched_balance_self(int cpu, int flag)
1923 struct task_struct *t = current;
1924 struct sched_domain *tmp, *sd = NULL;
1926 for_each_domain(cpu, tmp) {
1928 * If power savings logic is enabled for a domain, stop there.
1930 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1932 if (tmp->flags & flag)
1937 cpumask_t span, tmpmask;
1938 struct sched_group *group;
1939 int new_cpu, weight;
1941 if (!(sd->flags & flag)) {
1947 group = find_idlest_group(sd, t, cpu);
1953 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1954 if (new_cpu == -1 || new_cpu == cpu) {
1955 /* Now try balancing at a lower domain level of cpu */
1960 /* Now try balancing at a lower domain level of new_cpu */
1963 weight = cpus_weight(span);
1964 for_each_domain(cpu, tmp) {
1965 if (weight <= cpus_weight(tmp->span))
1967 if (tmp->flags & flag)
1970 /* while loop will break here if sd == NULL */
1976 #endif /* CONFIG_SMP */
1979 * try_to_wake_up - wake up a thread
1980 * @p: the to-be-woken-up thread
1981 * @state: the mask of task states that can be woken
1982 * @sync: do a synchronous wakeup?
1984 * Put it on the run-queue if it's not already there. The "current"
1985 * thread is always on the run-queue (except when the actual
1986 * re-schedule is in progress), and as such you're allowed to do
1987 * the simpler "current->state = TASK_RUNNING" to mark yourself
1988 * runnable without the overhead of this.
1990 * returns failure only if the task is already active.
1992 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1994 int cpu, orig_cpu, this_cpu, success = 0;
1995 unsigned long flags;
1999 if (!sched_feat(SYNC_WAKEUPS))
2003 rq = task_rq_lock(p, &flags);
2004 old_state = p->state;
2005 if (!(old_state & state))
2013 this_cpu = smp_processor_id();
2016 if (unlikely(task_running(rq, p)))
2019 cpu = p->sched_class->select_task_rq(p, sync);
2020 if (cpu != orig_cpu) {
2021 set_task_cpu(p, cpu);
2022 task_rq_unlock(rq, &flags);
2023 /* might preempt at this point */
2024 rq = task_rq_lock(p, &flags);
2025 old_state = p->state;
2026 if (!(old_state & state))
2031 this_cpu = smp_processor_id();
2035 #ifdef CONFIG_SCHEDSTATS
2036 schedstat_inc(rq, ttwu_count);
2037 if (cpu == this_cpu)
2038 schedstat_inc(rq, ttwu_local);
2040 struct sched_domain *sd;
2041 for_each_domain(this_cpu, sd) {
2042 if (cpu_isset(cpu, sd->span)) {
2043 schedstat_inc(sd, ttwu_wake_remote);
2051 #endif /* CONFIG_SMP */
2052 schedstat_inc(p, se.nr_wakeups);
2054 schedstat_inc(p, se.nr_wakeups_sync);
2055 if (orig_cpu != cpu)
2056 schedstat_inc(p, se.nr_wakeups_migrate);
2057 if (cpu == this_cpu)
2058 schedstat_inc(p, se.nr_wakeups_local);
2060 schedstat_inc(p, se.nr_wakeups_remote);
2061 update_rq_clock(rq);
2062 activate_task(rq, p, 1);
2066 check_preempt_curr(rq, p);
2068 p->state = TASK_RUNNING;
2070 if (p->sched_class->task_wake_up)
2071 p->sched_class->task_wake_up(rq, p);
2074 task_rq_unlock(rq, &flags);
2079 int wake_up_process(struct task_struct *p)
2081 return try_to_wake_up(p, TASK_ALL, 0);
2083 EXPORT_SYMBOL(wake_up_process);
2085 int wake_up_state(struct task_struct *p, unsigned int state)
2087 return try_to_wake_up(p, state, 0);
2091 * Perform scheduler related setup for a newly forked process p.
2092 * p is forked by current.
2094 * __sched_fork() is basic setup used by init_idle() too:
2096 static void __sched_fork(struct task_struct *p)
2098 p->se.exec_start = 0;
2099 p->se.sum_exec_runtime = 0;
2100 p->se.prev_sum_exec_runtime = 0;
2101 p->se.last_wakeup = 0;
2102 p->se.avg_overlap = 0;
2104 #ifdef CONFIG_SCHEDSTATS
2105 p->se.wait_start = 0;
2106 p->se.sum_sleep_runtime = 0;
2107 p->se.sleep_start = 0;
2108 p->se.block_start = 0;
2109 p->se.sleep_max = 0;
2110 p->se.block_max = 0;
2112 p->se.slice_max = 0;
2116 INIT_LIST_HEAD(&p->rt.run_list);
2119 #ifdef CONFIG_PREEMPT_NOTIFIERS
2120 INIT_HLIST_HEAD(&p->preempt_notifiers);
2124 * We mark the process as running here, but have not actually
2125 * inserted it onto the runqueue yet. This guarantees that
2126 * nobody will actually run it, and a signal or other external
2127 * event cannot wake it up and insert it on the runqueue either.
2129 p->state = TASK_RUNNING;
2133 * fork()/clone()-time setup:
2135 void sched_fork(struct task_struct *p, int clone_flags)
2137 int cpu = get_cpu();
2142 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2144 set_task_cpu(p, cpu);
2147 * Make sure we do not leak PI boosting priority to the child:
2149 p->prio = current->normal_prio;
2150 if (!rt_prio(p->prio))
2151 p->sched_class = &fair_sched_class;
2153 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2154 if (likely(sched_info_on()))
2155 memset(&p->sched_info, 0, sizeof(p->sched_info));
2157 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2160 #ifdef CONFIG_PREEMPT
2161 /* Want to start with kernel preemption disabled. */
2162 task_thread_info(p)->preempt_count = 1;
2168 * wake_up_new_task - wake up a newly created task for the first time.
2170 * This function will do some initial scheduler statistics housekeeping
2171 * that must be done for every newly created context, then puts the task
2172 * on the runqueue and wakes it.
2174 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2176 unsigned long flags;
2179 rq = task_rq_lock(p, &flags);
2180 BUG_ON(p->state != TASK_RUNNING);
2181 update_rq_clock(rq);
2183 p->prio = effective_prio(p);
2185 if (!p->sched_class->task_new || !current->se.on_rq) {
2186 activate_task(rq, p, 0);
2189 * Let the scheduling class do new task startup
2190 * management (if any):
2192 p->sched_class->task_new(rq, p);
2193 inc_nr_running(p, rq);
2195 check_preempt_curr(rq, p);
2197 if (p->sched_class->task_wake_up)
2198 p->sched_class->task_wake_up(rq, p);
2200 task_rq_unlock(rq, &flags);
2203 #ifdef CONFIG_PREEMPT_NOTIFIERS
2206 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2207 * @notifier: notifier struct to register
2209 void preempt_notifier_register(struct preempt_notifier *notifier)
2211 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2213 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2216 * preempt_notifier_unregister - no longer interested in preemption notifications
2217 * @notifier: notifier struct to unregister
2219 * This is safe to call from within a preemption notifier.
2221 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2223 hlist_del(¬ifier->link);
2225 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2227 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2229 struct preempt_notifier *notifier;
2230 struct hlist_node *node;
2232 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2233 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2237 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2238 struct task_struct *next)
2240 struct preempt_notifier *notifier;
2241 struct hlist_node *node;
2243 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2244 notifier->ops->sched_out(notifier, next);
2249 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2254 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2255 struct task_struct *next)
2262 * prepare_task_switch - prepare to switch tasks
2263 * @rq: the runqueue preparing to switch
2264 * @prev: the current task that is being switched out
2265 * @next: the task we are going to switch to.
2267 * This is called with the rq lock held and interrupts off. It must
2268 * be paired with a subsequent finish_task_switch after the context
2271 * prepare_task_switch sets up locking and calls architecture specific
2275 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2276 struct task_struct *next)
2278 fire_sched_out_preempt_notifiers(prev, next);
2279 prepare_lock_switch(rq, next);
2280 prepare_arch_switch(next);
2284 * finish_task_switch - clean up after a task-switch
2285 * @rq: runqueue associated with task-switch
2286 * @prev: the thread we just switched away from.
2288 * finish_task_switch must be called after the context switch, paired
2289 * with a prepare_task_switch call before the context switch.
2290 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2291 * and do any other architecture-specific cleanup actions.
2293 * Note that we may have delayed dropping an mm in context_switch(). If
2294 * so, we finish that here outside of the runqueue lock. (Doing it
2295 * with the lock held can cause deadlocks; see schedule() for
2298 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2299 __releases(rq->lock)
2301 struct mm_struct *mm = rq->prev_mm;
2307 * A task struct has one reference for the use as "current".
2308 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2309 * schedule one last time. The schedule call will never return, and
2310 * the scheduled task must drop that reference.
2311 * The test for TASK_DEAD must occur while the runqueue locks are
2312 * still held, otherwise prev could be scheduled on another cpu, die
2313 * there before we look at prev->state, and then the reference would
2315 * Manfred Spraul <manfred@colorfullife.com>
2317 prev_state = prev->state;
2318 finish_arch_switch(prev);
2319 finish_lock_switch(rq, prev);
2321 if (current->sched_class->post_schedule)
2322 current->sched_class->post_schedule(rq);
2325 fire_sched_in_preempt_notifiers(current);
2328 if (unlikely(prev_state == TASK_DEAD)) {
2330 * Remove function-return probe instances associated with this
2331 * task and put them back on the free list.
2333 kprobe_flush_task(prev);
2334 put_task_struct(prev);
2339 * schedule_tail - first thing a freshly forked thread must call.
2340 * @prev: the thread we just switched away from.
2342 asmlinkage void schedule_tail(struct task_struct *prev)
2343 __releases(rq->lock)
2345 struct rq *rq = this_rq();
2347 finish_task_switch(rq, prev);
2348 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2349 /* In this case, finish_task_switch does not reenable preemption */
2352 if (current->set_child_tid)
2353 put_user(task_pid_vnr(current), current->set_child_tid);
2357 * context_switch - switch to the new MM and the new
2358 * thread's register state.
2361 context_switch(struct rq *rq, struct task_struct *prev,
2362 struct task_struct *next)
2364 struct mm_struct *mm, *oldmm;
2366 prepare_task_switch(rq, prev, next);
2368 oldmm = prev->active_mm;
2370 * For paravirt, this is coupled with an exit in switch_to to
2371 * combine the page table reload and the switch backend into
2374 arch_enter_lazy_cpu_mode();
2376 if (unlikely(!mm)) {
2377 next->active_mm = oldmm;
2378 atomic_inc(&oldmm->mm_count);
2379 enter_lazy_tlb(oldmm, next);
2381 switch_mm(oldmm, mm, next);
2383 if (unlikely(!prev->mm)) {
2384 prev->active_mm = NULL;
2385 rq->prev_mm = oldmm;
2388 * Since the runqueue lock will be released by the next
2389 * task (which is an invalid locking op but in the case
2390 * of the scheduler it's an obvious special-case), so we
2391 * do an early lockdep release here:
2393 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2394 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2397 /* Here we just switch the register state and the stack. */
2398 switch_to(prev, next, prev);
2402 * this_rq must be evaluated again because prev may have moved
2403 * CPUs since it called schedule(), thus the 'rq' on its stack
2404 * frame will be invalid.
2406 finish_task_switch(this_rq(), prev);
2410 * nr_running, nr_uninterruptible and nr_context_switches:
2412 * externally visible scheduler statistics: current number of runnable
2413 * threads, current number of uninterruptible-sleeping threads, total
2414 * number of context switches performed since bootup.
2416 unsigned long nr_running(void)
2418 unsigned long i, sum = 0;
2420 for_each_online_cpu(i)
2421 sum += cpu_rq(i)->nr_running;
2426 unsigned long nr_uninterruptible(void)
2428 unsigned long i, sum = 0;
2430 for_each_possible_cpu(i)
2431 sum += cpu_rq(i)->nr_uninterruptible;
2434 * Since we read the counters lockless, it might be slightly
2435 * inaccurate. Do not allow it to go below zero though:
2437 if (unlikely((long)sum < 0))
2443 unsigned long long nr_context_switches(void)
2446 unsigned long long sum = 0;
2448 for_each_possible_cpu(i)
2449 sum += cpu_rq(i)->nr_switches;
2454 unsigned long nr_iowait(void)
2456 unsigned long i, sum = 0;
2458 for_each_possible_cpu(i)
2459 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2464 unsigned long nr_active(void)
2466 unsigned long i, running = 0, uninterruptible = 0;
2468 for_each_online_cpu(i) {
2469 running += cpu_rq(i)->nr_running;
2470 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2473 if (unlikely((long)uninterruptible < 0))
2474 uninterruptible = 0;
2476 return running + uninterruptible;
2480 * Update rq->cpu_load[] statistics. This function is usually called every
2481 * scheduler tick (TICK_NSEC).
2483 static void update_cpu_load(struct rq *this_rq)
2485 unsigned long this_load = this_rq->load.weight;
2488 this_rq->nr_load_updates++;
2490 /* Update our load: */
2491 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2492 unsigned long old_load, new_load;
2494 /* scale is effectively 1 << i now, and >> i divides by scale */
2496 old_load = this_rq->cpu_load[i];
2497 new_load = this_load;
2499 * Round up the averaging division if load is increasing. This
2500 * prevents us from getting stuck on 9 if the load is 10, for
2503 if (new_load > old_load)
2504 new_load += scale-1;
2505 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2512 * double_rq_lock - safely lock two runqueues
2514 * Note this does not disable interrupts like task_rq_lock,
2515 * you need to do so manually before calling.
2517 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2518 __acquires(rq1->lock)
2519 __acquires(rq2->lock)
2521 BUG_ON(!irqs_disabled());
2523 spin_lock(&rq1->lock);
2524 __acquire(rq2->lock); /* Fake it out ;) */
2527 spin_lock(&rq1->lock);
2528 spin_lock(&rq2->lock);
2530 spin_lock(&rq2->lock);
2531 spin_lock(&rq1->lock);
2534 update_rq_clock(rq1);
2535 update_rq_clock(rq2);
2539 * double_rq_unlock - safely unlock two runqueues
2541 * Note this does not restore interrupts like task_rq_unlock,
2542 * you need to do so manually after calling.
2544 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2545 __releases(rq1->lock)
2546 __releases(rq2->lock)
2548 spin_unlock(&rq1->lock);
2550 spin_unlock(&rq2->lock);
2552 __release(rq2->lock);
2556 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2558 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2559 __releases(this_rq->lock)
2560 __acquires(busiest->lock)
2561 __acquires(this_rq->lock)
2565 if (unlikely(!irqs_disabled())) {
2566 /* printk() doesn't work good under rq->lock */
2567 spin_unlock(&this_rq->lock);
2570 if (unlikely(!spin_trylock(&busiest->lock))) {
2571 if (busiest < this_rq) {
2572 spin_unlock(&this_rq->lock);
2573 spin_lock(&busiest->lock);
2574 spin_lock(&this_rq->lock);
2577 spin_lock(&busiest->lock);
2583 * If dest_cpu is allowed for this process, migrate the task to it.
2584 * This is accomplished by forcing the cpu_allowed mask to only
2585 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2586 * the cpu_allowed mask is restored.
2588 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2590 struct migration_req req;
2591 unsigned long flags;
2594 rq = task_rq_lock(p, &flags);
2595 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2596 || unlikely(cpu_is_offline(dest_cpu)))
2599 /* force the process onto the specified CPU */
2600 if (migrate_task(p, dest_cpu, &req)) {
2601 /* Need to wait for migration thread (might exit: take ref). */
2602 struct task_struct *mt = rq->migration_thread;
2604 get_task_struct(mt);
2605 task_rq_unlock(rq, &flags);
2606 wake_up_process(mt);
2607 put_task_struct(mt);
2608 wait_for_completion(&req.done);
2613 task_rq_unlock(rq, &flags);
2617 * sched_exec - execve() is a valuable balancing opportunity, because at
2618 * this point the task has the smallest effective memory and cache footprint.
2620 void sched_exec(void)
2622 int new_cpu, this_cpu = get_cpu();
2623 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2625 if (new_cpu != this_cpu)
2626 sched_migrate_task(current, new_cpu);
2630 * pull_task - move a task from a remote runqueue to the local runqueue.
2631 * Both runqueues must be locked.
2633 static void pull_task(struct rq *src_rq, struct task_struct *p,
2634 struct rq *this_rq, int this_cpu)
2636 deactivate_task(src_rq, p, 0);
2637 set_task_cpu(p, this_cpu);
2638 activate_task(this_rq, p, 0);
2640 * Note that idle threads have a prio of MAX_PRIO, for this test
2641 * to be always true for them.
2643 check_preempt_curr(this_rq, p);
2647 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2650 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2651 struct sched_domain *sd, enum cpu_idle_type idle,
2655 * We do not migrate tasks that are:
2656 * 1) running (obviously), or
2657 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2658 * 3) are cache-hot on their current CPU.
2660 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2661 schedstat_inc(p, se.nr_failed_migrations_affine);
2666 if (task_running(rq, p)) {
2667 schedstat_inc(p, se.nr_failed_migrations_running);
2672 * Aggressive migration if:
2673 * 1) task is cache cold, or
2674 * 2) too many balance attempts have failed.
2677 if (!task_hot(p, rq->clock, sd) ||
2678 sd->nr_balance_failed > sd->cache_nice_tries) {
2679 #ifdef CONFIG_SCHEDSTATS
2680 if (task_hot(p, rq->clock, sd)) {
2681 schedstat_inc(sd, lb_hot_gained[idle]);
2682 schedstat_inc(p, se.nr_forced_migrations);
2688 if (task_hot(p, rq->clock, sd)) {
2689 schedstat_inc(p, se.nr_failed_migrations_hot);
2695 static unsigned long
2696 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2697 unsigned long max_load_move, struct sched_domain *sd,
2698 enum cpu_idle_type idle, int *all_pinned,
2699 int *this_best_prio, struct rq_iterator *iterator)
2701 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2702 struct task_struct *p;
2703 long rem_load_move = max_load_move;
2705 if (max_load_move == 0)
2711 * Start the load-balancing iterator:
2713 p = iterator->start(iterator->arg);
2715 if (!p || loops++ > sysctl_sched_nr_migrate)
2718 * To help distribute high priority tasks across CPUs we don't
2719 * skip a task if it will be the highest priority task (i.e. smallest
2720 * prio value) on its new queue regardless of its load weight
2722 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2723 SCHED_LOAD_SCALE_FUZZ;
2724 if ((skip_for_load && p->prio >= *this_best_prio) ||
2725 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2726 p = iterator->next(iterator->arg);
2730 pull_task(busiest, p, this_rq, this_cpu);
2732 rem_load_move -= p->se.load.weight;
2735 * We only want to steal up to the prescribed amount of weighted load.
2737 if (rem_load_move > 0) {
2738 if (p->prio < *this_best_prio)
2739 *this_best_prio = p->prio;
2740 p = iterator->next(iterator->arg);
2745 * Right now, this is one of only two places pull_task() is called,
2746 * so we can safely collect pull_task() stats here rather than
2747 * inside pull_task().
2749 schedstat_add(sd, lb_gained[idle], pulled);
2752 *all_pinned = pinned;
2754 return max_load_move - rem_load_move;
2758 * move_tasks tries to move up to max_load_move weighted load from busiest to
2759 * this_rq, as part of a balancing operation within domain "sd".
2760 * Returns 1 if successful and 0 otherwise.
2762 * Called with both runqueues locked.
2764 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2765 unsigned long max_load_move,
2766 struct sched_domain *sd, enum cpu_idle_type idle,
2769 const struct sched_class *class = sched_class_highest;
2770 unsigned long total_load_moved = 0;
2771 int this_best_prio = this_rq->curr->prio;
2775 class->load_balance(this_rq, this_cpu, busiest,
2776 max_load_move - total_load_moved,
2777 sd, idle, all_pinned, &this_best_prio);
2778 class = class->next;
2779 } while (class && max_load_move > total_load_moved);
2781 return total_load_moved > 0;
2785 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2786 struct sched_domain *sd, enum cpu_idle_type idle,
2787 struct rq_iterator *iterator)
2789 struct task_struct *p = iterator->start(iterator->arg);
2793 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2794 pull_task(busiest, p, this_rq, this_cpu);
2796 * Right now, this is only the second place pull_task()
2797 * is called, so we can safely collect pull_task()
2798 * stats here rather than inside pull_task().
2800 schedstat_inc(sd, lb_gained[idle]);
2804 p = iterator->next(iterator->arg);
2811 * move_one_task tries to move exactly one task from busiest to this_rq, as
2812 * part of active balancing operations within "domain".
2813 * Returns 1 if successful and 0 otherwise.
2815 * Called with both runqueues locked.
2817 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2818 struct sched_domain *sd, enum cpu_idle_type idle)
2820 const struct sched_class *class;
2822 for (class = sched_class_highest; class; class = class->next)
2823 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2830 * find_busiest_group finds and returns the busiest CPU group within the
2831 * domain. It calculates and returns the amount of weighted load which
2832 * should be moved to restore balance via the imbalance parameter.
2834 static struct sched_group *
2835 find_busiest_group(struct sched_domain *sd, int this_cpu,
2836 unsigned long *imbalance, enum cpu_idle_type idle,
2837 int *sd_idle, const cpumask_t *cpus, int *balance)
2839 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2840 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2841 unsigned long max_pull;
2842 unsigned long busiest_load_per_task, busiest_nr_running;
2843 unsigned long this_load_per_task, this_nr_running;
2844 int load_idx, group_imb = 0;
2845 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2846 int power_savings_balance = 1;
2847 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2848 unsigned long min_nr_running = ULONG_MAX;
2849 struct sched_group *group_min = NULL, *group_leader = NULL;
2852 max_load = this_load = total_load = total_pwr = 0;
2853 busiest_load_per_task = busiest_nr_running = 0;
2854 this_load_per_task = this_nr_running = 0;
2855 if (idle == CPU_NOT_IDLE)
2856 load_idx = sd->busy_idx;
2857 else if (idle == CPU_NEWLY_IDLE)
2858 load_idx = sd->newidle_idx;
2860 load_idx = sd->idle_idx;
2863 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2866 int __group_imb = 0;
2867 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2868 unsigned long sum_nr_running, sum_weighted_load;
2870 local_group = cpu_isset(this_cpu, group->cpumask);
2873 balance_cpu = first_cpu(group->cpumask);
2875 /* Tally up the load of all CPUs in the group */
2876 sum_weighted_load = sum_nr_running = avg_load = 0;
2878 min_cpu_load = ~0UL;
2880 for_each_cpu_mask(i, group->cpumask) {
2883 if (!cpu_isset(i, *cpus))
2888 if (*sd_idle && rq->nr_running)
2891 /* Bias balancing toward cpus of our domain */
2893 if (idle_cpu(i) && !first_idle_cpu) {
2898 load = target_load(i, load_idx);
2900 load = source_load(i, load_idx);
2901 if (load > max_cpu_load)
2902 max_cpu_load = load;
2903 if (min_cpu_load > load)
2904 min_cpu_load = load;
2908 sum_nr_running += rq->nr_running;
2909 sum_weighted_load += weighted_cpuload(i);
2913 * First idle cpu or the first cpu(busiest) in this sched group
2914 * is eligible for doing load balancing at this and above
2915 * domains. In the newly idle case, we will allow all the cpu's
2916 * to do the newly idle load balance.
2918 if (idle != CPU_NEWLY_IDLE && local_group &&
2919 balance_cpu != this_cpu && balance) {
2924 total_load += avg_load;
2925 total_pwr += group->__cpu_power;
2927 /* Adjust by relative CPU power of the group */
2928 avg_load = sg_div_cpu_power(group,
2929 avg_load * SCHED_LOAD_SCALE);
2931 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2934 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2937 this_load = avg_load;
2939 this_nr_running = sum_nr_running;
2940 this_load_per_task = sum_weighted_load;
2941 } else if (avg_load > max_load &&
2942 (sum_nr_running > group_capacity || __group_imb)) {
2943 max_load = avg_load;
2945 busiest_nr_running = sum_nr_running;
2946 busiest_load_per_task = sum_weighted_load;
2947 group_imb = __group_imb;
2950 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2952 * Busy processors will not participate in power savings
2955 if (idle == CPU_NOT_IDLE ||
2956 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2960 * If the local group is idle or completely loaded
2961 * no need to do power savings balance at this domain
2963 if (local_group && (this_nr_running >= group_capacity ||
2965 power_savings_balance = 0;
2968 * If a group is already running at full capacity or idle,
2969 * don't include that group in power savings calculations
2971 if (!power_savings_balance || sum_nr_running >= group_capacity
2976 * Calculate the group which has the least non-idle load.
2977 * This is the group from where we need to pick up the load
2980 if ((sum_nr_running < min_nr_running) ||
2981 (sum_nr_running == min_nr_running &&
2982 first_cpu(group->cpumask) <
2983 first_cpu(group_min->cpumask))) {
2985 min_nr_running = sum_nr_running;
2986 min_load_per_task = sum_weighted_load /
2991 * Calculate the group which is almost near its
2992 * capacity but still has some space to pick up some load
2993 * from other group and save more power
2995 if (sum_nr_running <= group_capacity - 1) {
2996 if (sum_nr_running > leader_nr_running ||
2997 (sum_nr_running == leader_nr_running &&
2998 first_cpu(group->cpumask) >
2999 first_cpu(group_leader->cpumask))) {
3000 group_leader = group;
3001 leader_nr_running = sum_nr_running;
3006 group = group->next;
3007 } while (group != sd->groups);
3009 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3012 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3014 if (this_load >= avg_load ||
3015 100*max_load <= sd->imbalance_pct*this_load)
3018 busiest_load_per_task /= busiest_nr_running;
3020 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3023 * We're trying to get all the cpus to the average_load, so we don't
3024 * want to push ourselves above the average load, nor do we wish to
3025 * reduce the max loaded cpu below the average load, as either of these
3026 * actions would just result in more rebalancing later, and ping-pong
3027 * tasks around. Thus we look for the minimum possible imbalance.
3028 * Negative imbalances (*we* are more loaded than anyone else) will
3029 * be counted as no imbalance for these purposes -- we can't fix that
3030 * by pulling tasks to us. Be careful of negative numbers as they'll
3031 * appear as very large values with unsigned longs.
3033 if (max_load <= busiest_load_per_task)
3037 * In the presence of smp nice balancing, certain scenarios can have
3038 * max load less than avg load(as we skip the groups at or below
3039 * its cpu_power, while calculating max_load..)
3041 if (max_load < avg_load) {
3043 goto small_imbalance;
3046 /* Don't want to pull so many tasks that a group would go idle */
3047 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3049 /* How much load to actually move to equalise the imbalance */
3050 *imbalance = min(max_pull * busiest->__cpu_power,
3051 (avg_load - this_load) * this->__cpu_power)
3055 * if *imbalance is less than the average load per runnable task
3056 * there is no gaurantee that any tasks will be moved so we'll have
3057 * a think about bumping its value to force at least one task to be
3060 if (*imbalance < busiest_load_per_task) {
3061 unsigned long tmp, pwr_now, pwr_move;
3065 pwr_move = pwr_now = 0;
3067 if (this_nr_running) {
3068 this_load_per_task /= this_nr_running;
3069 if (busiest_load_per_task > this_load_per_task)
3072 this_load_per_task = SCHED_LOAD_SCALE;
3074 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3075 busiest_load_per_task * imbn) {
3076 *imbalance = busiest_load_per_task;
3081 * OK, we don't have enough imbalance to justify moving tasks,
3082 * however we may be able to increase total CPU power used by
3086 pwr_now += busiest->__cpu_power *
3087 min(busiest_load_per_task, max_load);
3088 pwr_now += this->__cpu_power *
3089 min(this_load_per_task, this_load);
3090 pwr_now /= SCHED_LOAD_SCALE;
3092 /* Amount of load we'd subtract */
3093 tmp = sg_div_cpu_power(busiest,
3094 busiest_load_per_task * SCHED_LOAD_SCALE);
3096 pwr_move += busiest->__cpu_power *
3097 min(busiest_load_per_task, max_load - tmp);
3099 /* Amount of load we'd add */
3100 if (max_load * busiest->__cpu_power <
3101 busiest_load_per_task * SCHED_LOAD_SCALE)
3102 tmp = sg_div_cpu_power(this,
3103 max_load * busiest->__cpu_power);
3105 tmp = sg_div_cpu_power(this,
3106 busiest_load_per_task * SCHED_LOAD_SCALE);
3107 pwr_move += this->__cpu_power *
3108 min(this_load_per_task, this_load + tmp);
3109 pwr_move /= SCHED_LOAD_SCALE;
3111 /* Move if we gain throughput */
3112 if (pwr_move > pwr_now)
3113 *imbalance = busiest_load_per_task;
3119 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3120 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3123 if (this == group_leader && group_leader != group_min) {
3124 *imbalance = min_load_per_task;
3134 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3137 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3138 unsigned long imbalance, const cpumask_t *cpus)
3140 struct rq *busiest = NULL, *rq;
3141 unsigned long max_load = 0;
3144 for_each_cpu_mask(i, group->cpumask) {
3147 if (!cpu_isset(i, *cpus))
3151 wl = weighted_cpuload(i);
3153 if (rq->nr_running == 1 && wl > imbalance)
3156 if (wl > max_load) {
3166 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3167 * so long as it is large enough.
3169 #define MAX_PINNED_INTERVAL 512
3172 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3173 * tasks if there is an imbalance.
3175 static int load_balance(int this_cpu, struct rq *this_rq,
3176 struct sched_domain *sd, enum cpu_idle_type idle,
3177 int *balance, cpumask_t *cpus)
3179 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3180 struct sched_group *group;
3181 unsigned long imbalance;
3183 unsigned long flags;
3188 * When power savings policy is enabled for the parent domain, idle
3189 * sibling can pick up load irrespective of busy siblings. In this case,
3190 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3191 * portraying it as CPU_NOT_IDLE.
3193 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3194 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3197 schedstat_inc(sd, lb_count[idle]);
3200 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3207 schedstat_inc(sd, lb_nobusyg[idle]);
3211 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3213 schedstat_inc(sd, lb_nobusyq[idle]);
3217 BUG_ON(busiest == this_rq);
3219 schedstat_add(sd, lb_imbalance[idle], imbalance);
3222 if (busiest->nr_running > 1) {
3224 * Attempt to move tasks. If find_busiest_group has found
3225 * an imbalance but busiest->nr_running <= 1, the group is
3226 * still unbalanced. ld_moved simply stays zero, so it is
3227 * correctly treated as an imbalance.
3229 local_irq_save(flags);
3230 double_rq_lock(this_rq, busiest);
3231 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3232 imbalance, sd, idle, &all_pinned);
3233 double_rq_unlock(this_rq, busiest);
3234 local_irq_restore(flags);
3237 * some other cpu did the load balance for us.
3239 if (ld_moved && this_cpu != smp_processor_id())
3240 resched_cpu(this_cpu);
3242 /* All tasks on this runqueue were pinned by CPU affinity */
3243 if (unlikely(all_pinned)) {
3244 cpu_clear(cpu_of(busiest), *cpus);
3245 if (!cpus_empty(*cpus))
3252 schedstat_inc(sd, lb_failed[idle]);
3253 sd->nr_balance_failed++;
3255 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3257 spin_lock_irqsave(&busiest->lock, flags);
3259 /* don't kick the migration_thread, if the curr
3260 * task on busiest cpu can't be moved to this_cpu
3262 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3263 spin_unlock_irqrestore(&busiest->lock, flags);
3265 goto out_one_pinned;
3268 if (!busiest->active_balance) {
3269 busiest->active_balance = 1;
3270 busiest->push_cpu = this_cpu;
3273 spin_unlock_irqrestore(&busiest->lock, flags);
3275 wake_up_process(busiest->migration_thread);
3278 * We've kicked active balancing, reset the failure
3281 sd->nr_balance_failed = sd->cache_nice_tries+1;
3284 sd->nr_balance_failed = 0;
3286 if (likely(!active_balance)) {
3287 /* We were unbalanced, so reset the balancing interval */
3288 sd->balance_interval = sd->min_interval;
3291 * If we've begun active balancing, start to back off. This
3292 * case may not be covered by the all_pinned logic if there
3293 * is only 1 task on the busy runqueue (because we don't call
3296 if (sd->balance_interval < sd->max_interval)
3297 sd->balance_interval *= 2;
3300 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3301 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3306 schedstat_inc(sd, lb_balanced[idle]);
3308 sd->nr_balance_failed = 0;
3311 /* tune up the balancing interval */
3312 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3313 (sd->balance_interval < sd->max_interval))
3314 sd->balance_interval *= 2;
3316 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3317 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3323 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3324 * tasks if there is an imbalance.
3326 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3327 * this_rq is locked.
3330 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3333 struct sched_group *group;
3334 struct rq *busiest = NULL;
3335 unsigned long imbalance;
3343 * When power savings policy is enabled for the parent domain, idle
3344 * sibling can pick up load irrespective of busy siblings. In this case,
3345 * let the state of idle sibling percolate up as IDLE, instead of
3346 * portraying it as CPU_NOT_IDLE.
3348 if (sd->flags & SD_SHARE_CPUPOWER &&
3349 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3352 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3354 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3355 &sd_idle, cpus, NULL);
3357 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3361 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3363 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3367 BUG_ON(busiest == this_rq);
3369 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3372 if (busiest->nr_running > 1) {
3373 /* Attempt to move tasks */
3374 double_lock_balance(this_rq, busiest);
3375 /* this_rq->clock is already updated */
3376 update_rq_clock(busiest);
3377 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3378 imbalance, sd, CPU_NEWLY_IDLE,
3380 spin_unlock(&busiest->lock);
3382 if (unlikely(all_pinned)) {
3383 cpu_clear(cpu_of(busiest), *cpus);
3384 if (!cpus_empty(*cpus))
3390 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3391 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3392 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3395 sd->nr_balance_failed = 0;
3400 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3401 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3402 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3404 sd->nr_balance_failed = 0;
3410 * idle_balance is called by schedule() if this_cpu is about to become
3411 * idle. Attempts to pull tasks from other CPUs.
3413 static void idle_balance(int this_cpu, struct rq *this_rq)
3415 struct sched_domain *sd;
3416 int pulled_task = -1;
3417 unsigned long next_balance = jiffies + HZ;
3420 for_each_domain(this_cpu, sd) {
3421 unsigned long interval;
3423 if (!(sd->flags & SD_LOAD_BALANCE))
3426 if (sd->flags & SD_BALANCE_NEWIDLE)
3427 /* If we've pulled tasks over stop searching: */
3428 pulled_task = load_balance_newidle(this_cpu, this_rq,
3431 interval = msecs_to_jiffies(sd->balance_interval);
3432 if (time_after(next_balance, sd->last_balance + interval))
3433 next_balance = sd->last_balance + interval;
3437 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3439 * We are going idle. next_balance may be set based on
3440 * a busy processor. So reset next_balance.
3442 this_rq->next_balance = next_balance;
3447 * active_load_balance is run by migration threads. It pushes running tasks
3448 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3449 * running on each physical CPU where possible, and avoids physical /
3450 * logical imbalances.
3452 * Called with busiest_rq locked.
3454 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3456 int target_cpu = busiest_rq->push_cpu;
3457 struct sched_domain *sd;
3458 struct rq *target_rq;
3460 /* Is there any task to move? */
3461 if (busiest_rq->nr_running <= 1)
3464 target_rq = cpu_rq(target_cpu);
3467 * This condition is "impossible", if it occurs
3468 * we need to fix it. Originally reported by
3469 * Bjorn Helgaas on a 128-cpu setup.
3471 BUG_ON(busiest_rq == target_rq);
3473 /* move a task from busiest_rq to target_rq */
3474 double_lock_balance(busiest_rq, target_rq);
3475 update_rq_clock(busiest_rq);
3476 update_rq_clock(target_rq);
3478 /* Search for an sd spanning us and the target CPU. */
3479 for_each_domain(target_cpu, sd) {
3480 if ((sd->flags & SD_LOAD_BALANCE) &&
3481 cpu_isset(busiest_cpu, sd->span))
3486 schedstat_inc(sd, alb_count);
3488 if (move_one_task(target_rq, target_cpu, busiest_rq,
3490 schedstat_inc(sd, alb_pushed);
3492 schedstat_inc(sd, alb_failed);
3494 spin_unlock(&target_rq->lock);
3499 atomic_t load_balancer;
3501 } nohz ____cacheline_aligned = {
3502 .load_balancer = ATOMIC_INIT(-1),
3503 .cpu_mask = CPU_MASK_NONE,
3507 * This routine will try to nominate the ilb (idle load balancing)
3508 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3509 * load balancing on behalf of all those cpus. If all the cpus in the system
3510 * go into this tickless mode, then there will be no ilb owner (as there is
3511 * no need for one) and all the cpus will sleep till the next wakeup event
3514 * For the ilb owner, tick is not stopped. And this tick will be used
3515 * for idle load balancing. ilb owner will still be part of
3518 * While stopping the tick, this cpu will become the ilb owner if there
3519 * is no other owner. And will be the owner till that cpu becomes busy
3520 * or if all cpus in the system stop their ticks at which point
3521 * there is no need for ilb owner.
3523 * When the ilb owner becomes busy, it nominates another owner, during the
3524 * next busy scheduler_tick()
3526 int select_nohz_load_balancer(int stop_tick)
3528 int cpu = smp_processor_id();
3531 cpu_set(cpu, nohz.cpu_mask);
3532 cpu_rq(cpu)->in_nohz_recently = 1;
3535 * If we are going offline and still the leader, give up!
3537 if (cpu_is_offline(cpu) &&
3538 atomic_read(&nohz.load_balancer) == cpu) {
3539 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3544 /* time for ilb owner also to sleep */
3545 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3546 if (atomic_read(&nohz.load_balancer) == cpu)
3547 atomic_set(&nohz.load_balancer, -1);
3551 if (atomic_read(&nohz.load_balancer) == -1) {
3552 /* make me the ilb owner */
3553 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3555 } else if (atomic_read(&nohz.load_balancer) == cpu)
3558 if (!cpu_isset(cpu, nohz.cpu_mask))
3561 cpu_clear(cpu, nohz.cpu_mask);
3563 if (atomic_read(&nohz.load_balancer) == cpu)
3564 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3571 static DEFINE_SPINLOCK(balancing);
3574 * It checks each scheduling domain to see if it is due to be balanced,
3575 * and initiates a balancing operation if so.
3577 * Balancing parameters are set up in arch_init_sched_domains.
3579 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3582 struct rq *rq = cpu_rq(cpu);
3583 unsigned long interval;
3584 struct sched_domain *sd;
3585 /* Earliest time when we have to do rebalance again */
3586 unsigned long next_balance = jiffies + 60*HZ;
3587 int update_next_balance = 0;
3590 for_each_domain(cpu, sd) {
3591 if (!(sd->flags & SD_LOAD_BALANCE))
3594 interval = sd->balance_interval;
3595 if (idle != CPU_IDLE)
3596 interval *= sd->busy_factor;
3598 /* scale ms to jiffies */
3599 interval = msecs_to_jiffies(interval);
3600 if (unlikely(!interval))
3602 if (interval > HZ*NR_CPUS/10)
3603 interval = HZ*NR_CPUS/10;
3606 if (sd->flags & SD_SERIALIZE) {
3607 if (!spin_trylock(&balancing))
3611 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3612 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3614 * We've pulled tasks over so either we're no
3615 * longer idle, or one of our SMT siblings is
3618 idle = CPU_NOT_IDLE;
3620 sd->last_balance = jiffies;
3622 if (sd->flags & SD_SERIALIZE)
3623 spin_unlock(&balancing);
3625 if (time_after(next_balance, sd->last_balance + interval)) {
3626 next_balance = sd->last_balance + interval;
3627 update_next_balance = 1;
3631 * Stop the load balance at this level. There is another
3632 * CPU in our sched group which is doing load balancing more
3640 * next_balance will be updated only when there is a need.
3641 * When the cpu is attached to null domain for ex, it will not be
3644 if (likely(update_next_balance))
3645 rq->next_balance = next_balance;
3649 * run_rebalance_domains is triggered when needed from the scheduler tick.
3650 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3651 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3653 static void run_rebalance_domains(struct softirq_action *h)
3655 int this_cpu = smp_processor_id();
3656 struct rq *this_rq = cpu_rq(this_cpu);
3657 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3658 CPU_IDLE : CPU_NOT_IDLE;
3660 rebalance_domains(this_cpu, idle);
3664 * If this cpu is the owner for idle load balancing, then do the
3665 * balancing on behalf of the other idle cpus whose ticks are
3668 if (this_rq->idle_at_tick &&
3669 atomic_read(&nohz.load_balancer) == this_cpu) {
3670 cpumask_t cpus = nohz.cpu_mask;
3674 cpu_clear(this_cpu, cpus);
3675 for_each_cpu_mask(balance_cpu, cpus) {
3677 * If this cpu gets work to do, stop the load balancing
3678 * work being done for other cpus. Next load
3679 * balancing owner will pick it up.
3684 rebalance_domains(balance_cpu, CPU_IDLE);
3686 rq = cpu_rq(balance_cpu);
3687 if (time_after(this_rq->next_balance, rq->next_balance))
3688 this_rq->next_balance = rq->next_balance;
3695 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3697 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3698 * idle load balancing owner or decide to stop the periodic load balancing,
3699 * if the whole system is idle.
3701 static inline void trigger_load_balance(struct rq *rq, int cpu)
3705 * If we were in the nohz mode recently and busy at the current
3706 * scheduler tick, then check if we need to nominate new idle
3709 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3710 rq->in_nohz_recently = 0;
3712 if (atomic_read(&nohz.load_balancer) == cpu) {
3713 cpu_clear(cpu, nohz.cpu_mask);
3714 atomic_set(&nohz.load_balancer, -1);
3717 if (atomic_read(&nohz.load_balancer) == -1) {
3719 * simple selection for now: Nominate the
3720 * first cpu in the nohz list to be the next
3723 * TBD: Traverse the sched domains and nominate
3724 * the nearest cpu in the nohz.cpu_mask.
3726 int ilb = first_cpu(nohz.cpu_mask);
3728 if (ilb < nr_cpu_ids)
3734 * If this cpu is idle and doing idle load balancing for all the
3735 * cpus with ticks stopped, is it time for that to stop?
3737 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3738 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3744 * If this cpu is idle and the idle load balancing is done by
3745 * someone else, then no need raise the SCHED_SOFTIRQ
3747 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3748 cpu_isset(cpu, nohz.cpu_mask))
3751 if (time_after_eq(jiffies, rq->next_balance))
3752 raise_softirq(SCHED_SOFTIRQ);
3755 #else /* CONFIG_SMP */
3758 * on UP we do not need to balance between CPUs:
3760 static inline void idle_balance(int cpu, struct rq *rq)
3766 DEFINE_PER_CPU(struct kernel_stat, kstat);
3768 EXPORT_PER_CPU_SYMBOL(kstat);
3771 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3772 * that have not yet been banked in case the task is currently running.
3774 unsigned long long task_sched_runtime(struct task_struct *p)
3776 unsigned long flags;
3780 rq = task_rq_lock(p, &flags);
3781 ns = p->se.sum_exec_runtime;
3782 if (task_current(rq, p)) {
3783 update_rq_clock(rq);
3784 delta_exec = rq->clock - p->se.exec_start;
3785 if ((s64)delta_exec > 0)
3788 task_rq_unlock(rq, &flags);
3794 * Account user cpu time to a process.
3795 * @p: the process that the cpu time gets accounted to
3796 * @cputime: the cpu time spent in user space since the last update
3798 void account_user_time(struct task_struct *p, cputime_t cputime)
3800 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3803 p->utime = cputime_add(p->utime, cputime);
3805 /* Add user time to cpustat. */
3806 tmp = cputime_to_cputime64(cputime);
3807 if (TASK_NICE(p) > 0)
3808 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3810 cpustat->user = cputime64_add(cpustat->user, tmp);
3814 * Account guest cpu time to a process.
3815 * @p: the process that the cpu time gets accounted to
3816 * @cputime: the cpu time spent in virtual machine since the last update
3818 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3821 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3823 tmp = cputime_to_cputime64(cputime);
3825 p->utime = cputime_add(p->utime, cputime);
3826 p->gtime = cputime_add(p->gtime, cputime);
3828 cpustat->user = cputime64_add(cpustat->user, tmp);
3829 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3833 * Account scaled user cpu time to a process.
3834 * @p: the process that the cpu time gets accounted to
3835 * @cputime: the cpu time spent in user space since the last update
3837 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3839 p->utimescaled = cputime_add(p->utimescaled, cputime);
3843 * Account system cpu time to a process.
3844 * @p: the process that the cpu time gets accounted to
3845 * @hardirq_offset: the offset to subtract from hardirq_count()
3846 * @cputime: the cpu time spent in kernel space since the last update
3848 void account_system_time(struct task_struct *p, int hardirq_offset,
3851 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3852 struct rq *rq = this_rq();
3855 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3856 return account_guest_time(p, cputime);
3858 p->stime = cputime_add(p->stime, cputime);
3860 /* Add system time to cpustat. */
3861 tmp = cputime_to_cputime64(cputime);
3862 if (hardirq_count() - hardirq_offset)
3863 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3864 else if (softirq_count())
3865 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3866 else if (p != rq->idle)
3867 cpustat->system = cputime64_add(cpustat->system, tmp);
3868 else if (atomic_read(&rq->nr_iowait) > 0)
3869 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3871 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3872 /* Account for system time used */
3873 acct_update_integrals(p);
3877 * Account scaled system cpu time to a process.
3878 * @p: the process that the cpu time gets accounted to
3879 * @hardirq_offset: the offset to subtract from hardirq_count()
3880 * @cputime: the cpu time spent in kernel space since the last update
3882 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3884 p->stimescaled = cputime_add(p->stimescaled, cputime);
3888 * Account for involuntary wait time.
3889 * @p: the process from which the cpu time has been stolen
3890 * @steal: the cpu time spent in involuntary wait
3892 void account_steal_time(struct task_struct *p, cputime_t steal)
3894 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3895 cputime64_t tmp = cputime_to_cputime64(steal);
3896 struct rq *rq = this_rq();
3898 if (p == rq->idle) {
3899 p->stime = cputime_add(p->stime, steal);
3900 if (atomic_read(&rq->nr_iowait) > 0)
3901 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3903 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3905 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3909 * This function gets called by the timer code, with HZ frequency.
3910 * We call it with interrupts disabled.
3912 * It also gets called by the fork code, when changing the parent's
3915 void scheduler_tick(void)
3917 int cpu = smp_processor_id();
3918 struct rq *rq = cpu_rq(cpu);
3919 struct task_struct *curr = rq->curr;
3920 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3922 spin_lock(&rq->lock);
3923 __update_rq_clock(rq);
3925 * Let rq->clock advance by at least TICK_NSEC:
3927 if (unlikely(rq->clock < next_tick)) {
3928 rq->clock = next_tick;
3929 rq->clock_underflows++;
3931 rq->tick_timestamp = rq->clock;
3932 update_last_tick_seen(rq);
3933 update_cpu_load(rq);
3934 curr->sched_class->task_tick(rq, curr, 0);
3935 spin_unlock(&rq->lock);
3938 rq->idle_at_tick = idle_cpu(cpu);
3939 trigger_load_balance(rq, cpu);
3943 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3945 void __kprobes add_preempt_count(int val)
3950 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3952 preempt_count() += val;
3954 * Spinlock count overflowing soon?
3956 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3959 EXPORT_SYMBOL(add_preempt_count);
3961 void __kprobes sub_preempt_count(int val)
3966 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3969 * Is the spinlock portion underflowing?
3971 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3972 !(preempt_count() & PREEMPT_MASK)))
3975 preempt_count() -= val;
3977 EXPORT_SYMBOL(sub_preempt_count);
3982 * Print scheduling while atomic bug:
3984 static noinline void __schedule_bug(struct task_struct *prev)
3986 struct pt_regs *regs = get_irq_regs();
3988 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3989 prev->comm, prev->pid, preempt_count());
3991 debug_show_held_locks(prev);
3992 if (irqs_disabled())
3993 print_irqtrace_events(prev);
4002 * Various schedule()-time debugging checks and statistics:
4004 static inline void schedule_debug(struct task_struct *prev)
4007 * Test if we are atomic. Since do_exit() needs to call into
4008 * schedule() atomically, we ignore that path for now.
4009 * Otherwise, whine if we are scheduling when we should not be.
4011 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4012 __schedule_bug(prev);
4014 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4016 schedstat_inc(this_rq(), sched_count);
4017 #ifdef CONFIG_SCHEDSTATS
4018 if (unlikely(prev->lock_depth >= 0)) {
4019 schedstat_inc(this_rq(), bkl_count);
4020 schedstat_inc(prev, sched_info.bkl_count);
4026 * Pick up the highest-prio task:
4028 static inline struct task_struct *
4029 pick_next_task(struct rq *rq, struct task_struct *prev)
4031 const struct sched_class *class;
4032 struct task_struct *p;
4035 * Optimization: we know that if all tasks are in
4036 * the fair class we can call that function directly:
4038 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4039 p = fair_sched_class.pick_next_task(rq);
4044 class = sched_class_highest;
4046 p = class->pick_next_task(rq);
4050 * Will never be NULL as the idle class always
4051 * returns a non-NULL p:
4053 class = class->next;
4058 * schedule() is the main scheduler function.
4060 asmlinkage void __sched schedule(void)
4062 struct task_struct *prev, *next;
4063 unsigned long *switch_count;
4069 cpu = smp_processor_id();
4073 switch_count = &prev->nivcsw;
4075 release_kernel_lock(prev);
4076 need_resched_nonpreemptible:
4078 schedule_debug(prev);
4083 * Do the rq-clock update outside the rq lock:
4085 local_irq_disable();
4086 __update_rq_clock(rq);
4087 spin_lock(&rq->lock);
4088 clear_tsk_need_resched(prev);
4090 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4091 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4092 signal_pending(prev))) {
4093 prev->state = TASK_RUNNING;
4095 deactivate_task(rq, prev, 1);
4097 switch_count = &prev->nvcsw;
4101 if (prev->sched_class->pre_schedule)
4102 prev->sched_class->pre_schedule(rq, prev);
4105 if (unlikely(!rq->nr_running))
4106 idle_balance(cpu, rq);
4108 prev->sched_class->put_prev_task(rq, prev);
4109 next = pick_next_task(rq, prev);
4111 sched_info_switch(prev, next);
4113 if (likely(prev != next)) {
4118 context_switch(rq, prev, next); /* unlocks the rq */
4120 * the context switch might have flipped the stack from under
4121 * us, hence refresh the local variables.
4123 cpu = smp_processor_id();
4126 spin_unlock_irq(&rq->lock);
4130 if (unlikely(reacquire_kernel_lock(current) < 0))
4131 goto need_resched_nonpreemptible;
4133 preempt_enable_no_resched();
4134 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4137 EXPORT_SYMBOL(schedule);
4139 #ifdef CONFIG_PREEMPT
4141 * this is the entry point to schedule() from in-kernel preemption
4142 * off of preempt_enable. Kernel preemptions off return from interrupt
4143 * occur there and call schedule directly.
4145 asmlinkage void __sched preempt_schedule(void)
4147 struct thread_info *ti = current_thread_info();
4148 struct task_struct *task = current;
4149 int saved_lock_depth;
4152 * If there is a non-zero preempt_count or interrupts are disabled,
4153 * we do not want to preempt the current task. Just return..
4155 if (likely(ti->preempt_count || irqs_disabled()))
4159 add_preempt_count(PREEMPT_ACTIVE);
4162 * We keep the big kernel semaphore locked, but we
4163 * clear ->lock_depth so that schedule() doesnt
4164 * auto-release the semaphore:
4166 saved_lock_depth = task->lock_depth;
4167 task->lock_depth = -1;
4169 task->lock_depth = saved_lock_depth;
4170 sub_preempt_count(PREEMPT_ACTIVE);
4173 * Check again in case we missed a preemption opportunity
4174 * between schedule and now.
4177 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4179 EXPORT_SYMBOL(preempt_schedule);
4182 * this is the entry point to schedule() from kernel preemption
4183 * off of irq context.
4184 * Note, that this is called and return with irqs disabled. This will
4185 * protect us against recursive calling from irq.
4187 asmlinkage void __sched preempt_schedule_irq(void)
4189 struct thread_info *ti = current_thread_info();
4190 struct task_struct *task = current;
4191 int saved_lock_depth;
4193 /* Catch callers which need to be fixed */
4194 BUG_ON(ti->preempt_count || !irqs_disabled());
4197 add_preempt_count(PREEMPT_ACTIVE);
4200 * We keep the big kernel semaphore locked, but we
4201 * clear ->lock_depth so that schedule() doesnt
4202 * auto-release the semaphore:
4204 saved_lock_depth = task->lock_depth;
4205 task->lock_depth = -1;
4208 local_irq_disable();
4209 task->lock_depth = saved_lock_depth;
4210 sub_preempt_count(PREEMPT_ACTIVE);
4213 * Check again in case we missed a preemption opportunity
4214 * between schedule and now.
4217 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4220 #endif /* CONFIG_PREEMPT */
4222 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4225 return try_to_wake_up(curr->private, mode, sync);
4227 EXPORT_SYMBOL(default_wake_function);
4230 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4231 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4232 * number) then we wake all the non-exclusive tasks and one exclusive task.
4234 * There are circumstances in which we can try to wake a task which has already
4235 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4236 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4238 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4239 int nr_exclusive, int sync, void *key)
4241 wait_queue_t *curr, *next;
4243 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4244 unsigned flags = curr->flags;
4246 if (curr->func(curr, mode, sync, key) &&
4247 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4253 * __wake_up - wake up threads blocked on a waitqueue.
4255 * @mode: which threads
4256 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4257 * @key: is directly passed to the wakeup function
4259 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4260 int nr_exclusive, void *key)
4262 unsigned long flags;
4264 spin_lock_irqsave(&q->lock, flags);
4265 __wake_up_common(q, mode, nr_exclusive, 0, key);
4266 spin_unlock_irqrestore(&q->lock, flags);
4268 EXPORT_SYMBOL(__wake_up);
4271 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4273 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4275 __wake_up_common(q, mode, 1, 0, NULL);
4279 * __wake_up_sync - wake up threads blocked on a waitqueue.
4281 * @mode: which threads
4282 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4284 * The sync wakeup differs that the waker knows that it will schedule
4285 * away soon, so while the target thread will be woken up, it will not
4286 * be migrated to another CPU - ie. the two threads are 'synchronized'
4287 * with each other. This can prevent needless bouncing between CPUs.
4289 * On UP it can prevent extra preemption.
4292 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4294 unsigned long flags;
4300 if (unlikely(!nr_exclusive))
4303 spin_lock_irqsave(&q->lock, flags);
4304 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4305 spin_unlock_irqrestore(&q->lock, flags);
4307 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4309 void complete(struct completion *x)
4311 unsigned long flags;
4313 spin_lock_irqsave(&x->wait.lock, flags);
4315 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4316 spin_unlock_irqrestore(&x->wait.lock, flags);
4318 EXPORT_SYMBOL(complete);
4320 void complete_all(struct completion *x)
4322 unsigned long flags;
4324 spin_lock_irqsave(&x->wait.lock, flags);
4325 x->done += UINT_MAX/2;
4326 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4327 spin_unlock_irqrestore(&x->wait.lock, flags);
4329 EXPORT_SYMBOL(complete_all);
4331 static inline long __sched
4332 do_wait_for_common(struct completion *x, long timeout, int state)
4335 DECLARE_WAITQUEUE(wait, current);
4337 wait.flags |= WQ_FLAG_EXCLUSIVE;
4338 __add_wait_queue_tail(&x->wait, &wait);
4340 if ((state == TASK_INTERRUPTIBLE &&
4341 signal_pending(current)) ||
4342 (state == TASK_KILLABLE &&
4343 fatal_signal_pending(current))) {
4344 __remove_wait_queue(&x->wait, &wait);
4345 return -ERESTARTSYS;
4347 __set_current_state(state);
4348 spin_unlock_irq(&x->wait.lock);
4349 timeout = schedule_timeout(timeout);
4350 spin_lock_irq(&x->wait.lock);
4352 __remove_wait_queue(&x->wait, &wait);
4356 __remove_wait_queue(&x->wait, &wait);
4363 wait_for_common(struct completion *x, long timeout, int state)
4367 spin_lock_irq(&x->wait.lock);
4368 timeout = do_wait_for_common(x, timeout, state);
4369 spin_unlock_irq(&x->wait.lock);
4373 void __sched wait_for_completion(struct completion *x)
4375 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4377 EXPORT_SYMBOL(wait_for_completion);
4379 unsigned long __sched
4380 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4382 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4384 EXPORT_SYMBOL(wait_for_completion_timeout);
4386 int __sched wait_for_completion_interruptible(struct completion *x)
4388 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4389 if (t == -ERESTARTSYS)
4393 EXPORT_SYMBOL(wait_for_completion_interruptible);
4395 unsigned long __sched
4396 wait_for_completion_interruptible_timeout(struct completion *x,
4397 unsigned long timeout)
4399 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4401 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4403 int __sched wait_for_completion_killable(struct completion *x)
4405 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4406 if (t == -ERESTARTSYS)
4410 EXPORT_SYMBOL(wait_for_completion_killable);
4413 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4415 unsigned long flags;
4418 init_waitqueue_entry(&wait, current);
4420 __set_current_state(state);
4422 spin_lock_irqsave(&q->lock, flags);
4423 __add_wait_queue(q, &wait);
4424 spin_unlock(&q->lock);
4425 timeout = schedule_timeout(timeout);
4426 spin_lock_irq(&q->lock);
4427 __remove_wait_queue(q, &wait);
4428 spin_unlock_irqrestore(&q->lock, flags);
4433 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4435 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4437 EXPORT_SYMBOL(interruptible_sleep_on);
4440 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4442 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4444 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4446 void __sched sleep_on(wait_queue_head_t *q)
4448 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4450 EXPORT_SYMBOL(sleep_on);
4452 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4454 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4456 EXPORT_SYMBOL(sleep_on_timeout);
4458 #ifdef CONFIG_RT_MUTEXES
4461 * rt_mutex_setprio - set the current priority of a task
4463 * @prio: prio value (kernel-internal form)
4465 * This function changes the 'effective' priority of a task. It does
4466 * not touch ->normal_prio like __setscheduler().
4468 * Used by the rt_mutex code to implement priority inheritance logic.
4470 void rt_mutex_setprio(struct task_struct *p, int prio)
4472 unsigned long flags;
4473 int oldprio, on_rq, running;
4475 const struct sched_class *prev_class = p->sched_class;
4477 BUG_ON(prio < 0 || prio > MAX_PRIO);
4479 rq = task_rq_lock(p, &flags);
4480 update_rq_clock(rq);
4483 on_rq = p->se.on_rq;
4484 running = task_current(rq, p);
4486 dequeue_task(rq, p, 0);
4488 p->sched_class->put_prev_task(rq, p);
4491 p->sched_class = &rt_sched_class;
4493 p->sched_class = &fair_sched_class;
4498 p->sched_class->set_curr_task(rq);
4500 enqueue_task(rq, p, 0);
4502 check_class_changed(rq, p, prev_class, oldprio, running);
4504 task_rq_unlock(rq, &flags);
4509 void set_user_nice(struct task_struct *p, long nice)
4511 int old_prio, delta, on_rq;
4512 unsigned long flags;
4515 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4518 * We have to be careful, if called from sys_setpriority(),
4519 * the task might be in the middle of scheduling on another CPU.
4521 rq = task_rq_lock(p, &flags);
4522 update_rq_clock(rq);
4524 * The RT priorities are set via sched_setscheduler(), but we still
4525 * allow the 'normal' nice value to be set - but as expected
4526 * it wont have any effect on scheduling until the task is
4527 * SCHED_FIFO/SCHED_RR:
4529 if (task_has_rt_policy(p)) {
4530 p->static_prio = NICE_TO_PRIO(nice);
4533 on_rq = p->se.on_rq;
4535 dequeue_task(rq, p, 0);
4539 p->static_prio = NICE_TO_PRIO(nice);
4542 p->prio = effective_prio(p);
4543 delta = p->prio - old_prio;
4546 enqueue_task(rq, p, 0);
4549 * If the task increased its priority or is running and
4550 * lowered its priority, then reschedule its CPU:
4552 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4553 resched_task(rq->curr);
4556 task_rq_unlock(rq, &flags);
4558 EXPORT_SYMBOL(set_user_nice);
4561 * can_nice - check if a task can reduce its nice value
4565 int can_nice(const struct task_struct *p, const int nice)
4567 /* convert nice value [19,-20] to rlimit style value [1,40] */
4568 int nice_rlim = 20 - nice;
4570 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4571 capable(CAP_SYS_NICE));
4574 #ifdef __ARCH_WANT_SYS_NICE
4577 * sys_nice - change the priority of the current process.
4578 * @increment: priority increment
4580 * sys_setpriority is a more generic, but much slower function that
4581 * does similar things.
4583 asmlinkage long sys_nice(int increment)
4588 * Setpriority might change our priority at the same moment.
4589 * We don't have to worry. Conceptually one call occurs first
4590 * and we have a single winner.
4592 if (increment < -40)
4597 nice = PRIO_TO_NICE(current->static_prio) + increment;
4603 if (increment < 0 && !can_nice(current, nice))
4606 retval = security_task_setnice(current, nice);
4610 set_user_nice(current, nice);
4617 * task_prio - return the priority value of a given task.
4618 * @p: the task in question.
4620 * This is the priority value as seen by users in /proc.
4621 * RT tasks are offset by -200. Normal tasks are centered
4622 * around 0, value goes from -16 to +15.
4624 int task_prio(const struct task_struct *p)
4626 return p->prio - MAX_RT_PRIO;
4630 * task_nice - return the nice value of a given task.
4631 * @p: the task in question.
4633 int task_nice(const struct task_struct *p)
4635 return TASK_NICE(p);
4637 EXPORT_SYMBOL(task_nice);
4640 * idle_cpu - is a given cpu idle currently?
4641 * @cpu: the processor in question.
4643 int idle_cpu(int cpu)
4645 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4649 * idle_task - return the idle task for a given cpu.
4650 * @cpu: the processor in question.
4652 struct task_struct *idle_task(int cpu)
4654 return cpu_rq(cpu)->idle;
4658 * find_process_by_pid - find a process with a matching PID value.
4659 * @pid: the pid in question.
4661 static struct task_struct *find_process_by_pid(pid_t pid)
4663 return pid ? find_task_by_vpid(pid) : current;
4666 /* Actually do priority change: must hold rq lock. */
4668 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4670 BUG_ON(p->se.on_rq);
4673 switch (p->policy) {
4677 p->sched_class = &fair_sched_class;
4681 p->sched_class = &rt_sched_class;
4685 p->rt_priority = prio;
4686 p->normal_prio = normal_prio(p);
4687 /* we are holding p->pi_lock already */
4688 p->prio = rt_mutex_getprio(p);
4693 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4694 * @p: the task in question.
4695 * @policy: new policy.
4696 * @param: structure containing the new RT priority.
4698 * NOTE that the task may be already dead.
4700 int sched_setscheduler(struct task_struct *p, int policy,
4701 struct sched_param *param)
4703 int retval, oldprio, oldpolicy = -1, on_rq, running;
4704 unsigned long flags;
4705 const struct sched_class *prev_class = p->sched_class;
4708 /* may grab non-irq protected spin_locks */
4709 BUG_ON(in_interrupt());
4711 /* double check policy once rq lock held */
4713 policy = oldpolicy = p->policy;
4714 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4715 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4716 policy != SCHED_IDLE)
4719 * Valid priorities for SCHED_FIFO and SCHED_RR are
4720 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4721 * SCHED_BATCH and SCHED_IDLE is 0.
4723 if (param->sched_priority < 0 ||
4724 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4725 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4727 if (rt_policy(policy) != (param->sched_priority != 0))
4731 * Allow unprivileged RT tasks to decrease priority:
4733 if (!capable(CAP_SYS_NICE)) {
4734 if (rt_policy(policy)) {
4735 unsigned long rlim_rtprio;
4737 if (!lock_task_sighand(p, &flags))
4739 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4740 unlock_task_sighand(p, &flags);
4742 /* can't set/change the rt policy */
4743 if (policy != p->policy && !rlim_rtprio)
4746 /* can't increase priority */
4747 if (param->sched_priority > p->rt_priority &&
4748 param->sched_priority > rlim_rtprio)
4752 * Like positive nice levels, dont allow tasks to
4753 * move out of SCHED_IDLE either:
4755 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4758 /* can't change other user's priorities */
4759 if ((current->euid != p->euid) &&
4760 (current->euid != p->uid))
4764 #ifdef CONFIG_RT_GROUP_SCHED
4766 * Do not allow realtime tasks into groups that have no runtime
4769 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4773 retval = security_task_setscheduler(p, policy, param);
4777 * make sure no PI-waiters arrive (or leave) while we are
4778 * changing the priority of the task:
4780 spin_lock_irqsave(&p->pi_lock, flags);
4782 * To be able to change p->policy safely, the apropriate
4783 * runqueue lock must be held.
4785 rq = __task_rq_lock(p);
4786 /* recheck policy now with rq lock held */
4787 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4788 policy = oldpolicy = -1;
4789 __task_rq_unlock(rq);
4790 spin_unlock_irqrestore(&p->pi_lock, flags);
4793 update_rq_clock(rq);
4794 on_rq = p->se.on_rq;
4795 running = task_current(rq, p);
4797 deactivate_task(rq, p, 0);
4799 p->sched_class->put_prev_task(rq, p);
4802 __setscheduler(rq, p, policy, param->sched_priority);
4805 p->sched_class->set_curr_task(rq);
4807 activate_task(rq, p, 0);
4809 check_class_changed(rq, p, prev_class, oldprio, running);
4811 __task_rq_unlock(rq);
4812 spin_unlock_irqrestore(&p->pi_lock, flags);
4814 rt_mutex_adjust_pi(p);
4818 EXPORT_SYMBOL_GPL(sched_setscheduler);
4821 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4823 struct sched_param lparam;
4824 struct task_struct *p;
4827 if (!param || pid < 0)
4829 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4834 p = find_process_by_pid(pid);
4836 retval = sched_setscheduler(p, policy, &lparam);
4843 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4844 * @pid: the pid in question.
4845 * @policy: new policy.
4846 * @param: structure containing the new RT priority.
4849 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4851 /* negative values for policy are not valid */
4855 return do_sched_setscheduler(pid, policy, param);
4859 * sys_sched_setparam - set/change the RT priority of a thread
4860 * @pid: the pid in question.
4861 * @param: structure containing the new RT priority.
4863 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4865 return do_sched_setscheduler(pid, -1, param);
4869 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4870 * @pid: the pid in question.
4872 asmlinkage long sys_sched_getscheduler(pid_t pid)
4874 struct task_struct *p;
4881 read_lock(&tasklist_lock);
4882 p = find_process_by_pid(pid);
4884 retval = security_task_getscheduler(p);
4888 read_unlock(&tasklist_lock);
4893 * sys_sched_getscheduler - get the RT priority of a thread
4894 * @pid: the pid in question.
4895 * @param: structure containing the RT priority.
4897 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4899 struct sched_param lp;
4900 struct task_struct *p;
4903 if (!param || pid < 0)
4906 read_lock(&tasklist_lock);
4907 p = find_process_by_pid(pid);
4912 retval = security_task_getscheduler(p);
4916 lp.sched_priority = p->rt_priority;
4917 read_unlock(&tasklist_lock);
4920 * This one might sleep, we cannot do it with a spinlock held ...
4922 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4927 read_unlock(&tasklist_lock);
4931 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4933 cpumask_t cpus_allowed;
4934 cpumask_t new_mask = *in_mask;
4935 struct task_struct *p;
4939 read_lock(&tasklist_lock);
4941 p = find_process_by_pid(pid);
4943 read_unlock(&tasklist_lock);
4949 * It is not safe to call set_cpus_allowed with the
4950 * tasklist_lock held. We will bump the task_struct's
4951 * usage count and then drop tasklist_lock.
4954 read_unlock(&tasklist_lock);
4957 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4958 !capable(CAP_SYS_NICE))
4961 retval = security_task_setscheduler(p, 0, NULL);
4965 cpuset_cpus_allowed(p, &cpus_allowed);
4966 cpus_and(new_mask, new_mask, cpus_allowed);
4968 retval = set_cpus_allowed_ptr(p, &new_mask);
4971 cpuset_cpus_allowed(p, &cpus_allowed);
4972 if (!cpus_subset(new_mask, cpus_allowed)) {
4974 * We must have raced with a concurrent cpuset
4975 * update. Just reset the cpus_allowed to the
4976 * cpuset's cpus_allowed
4978 new_mask = cpus_allowed;
4988 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4989 cpumask_t *new_mask)
4991 if (len < sizeof(cpumask_t)) {
4992 memset(new_mask, 0, sizeof(cpumask_t));
4993 } else if (len > sizeof(cpumask_t)) {
4994 len = sizeof(cpumask_t);
4996 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5000 * sys_sched_setaffinity - set the cpu affinity of a process
5001 * @pid: pid of the process
5002 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5003 * @user_mask_ptr: user-space pointer to the new cpu mask
5005 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5006 unsigned long __user *user_mask_ptr)
5011 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5015 return sched_setaffinity(pid, &new_mask);
5019 * Represents all cpu's present in the system
5020 * In systems capable of hotplug, this map could dynamically grow
5021 * as new cpu's are detected in the system via any platform specific
5022 * method, such as ACPI for e.g.
5025 cpumask_t cpu_present_map __read_mostly;
5026 EXPORT_SYMBOL(cpu_present_map);
5029 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5030 EXPORT_SYMBOL(cpu_online_map);
5032 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5033 EXPORT_SYMBOL(cpu_possible_map);
5036 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5038 struct task_struct *p;
5042 read_lock(&tasklist_lock);
5045 p = find_process_by_pid(pid);
5049 retval = security_task_getscheduler(p);
5053 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5056 read_unlock(&tasklist_lock);
5063 * sys_sched_getaffinity - get the cpu affinity of a process
5064 * @pid: pid of the process
5065 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5066 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5068 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5069 unsigned long __user *user_mask_ptr)
5074 if (len < sizeof(cpumask_t))
5077 ret = sched_getaffinity(pid, &mask);
5081 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5084 return sizeof(cpumask_t);
5088 * sys_sched_yield - yield the current processor to other threads.
5090 * This function yields the current CPU to other tasks. If there are no
5091 * other threads running on this CPU then this function will return.
5093 asmlinkage long sys_sched_yield(void)
5095 struct rq *rq = this_rq_lock();
5097 schedstat_inc(rq, yld_count);
5098 current->sched_class->yield_task(rq);
5101 * Since we are going to call schedule() anyway, there's
5102 * no need to preempt or enable interrupts:
5104 __release(rq->lock);
5105 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5106 _raw_spin_unlock(&rq->lock);
5107 preempt_enable_no_resched();
5114 static void __cond_resched(void)
5116 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5117 __might_sleep(__FILE__, __LINE__);
5120 * The BKS might be reacquired before we have dropped
5121 * PREEMPT_ACTIVE, which could trigger a second
5122 * cond_resched() call.
5125 add_preempt_count(PREEMPT_ACTIVE);
5127 sub_preempt_count(PREEMPT_ACTIVE);
5128 } while (need_resched());
5131 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5132 int __sched _cond_resched(void)
5134 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5135 system_state == SYSTEM_RUNNING) {
5141 EXPORT_SYMBOL(_cond_resched);
5145 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5146 * call schedule, and on return reacquire the lock.
5148 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5149 * operations here to prevent schedule() from being called twice (once via
5150 * spin_unlock(), once by hand).
5152 int cond_resched_lock(spinlock_t *lock)
5154 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5157 if (spin_needbreak(lock) || resched) {
5159 if (resched && need_resched())
5168 EXPORT_SYMBOL(cond_resched_lock);
5170 int __sched cond_resched_softirq(void)
5172 BUG_ON(!in_softirq());
5174 if (need_resched() && system_state == SYSTEM_RUNNING) {
5182 EXPORT_SYMBOL(cond_resched_softirq);
5185 * yield - yield the current processor to other threads.
5187 * This is a shortcut for kernel-space yielding - it marks the
5188 * thread runnable and calls sys_sched_yield().
5190 void __sched yield(void)
5192 set_current_state(TASK_RUNNING);
5195 EXPORT_SYMBOL(yield);
5198 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5199 * that process accounting knows that this is a task in IO wait state.
5201 * But don't do that if it is a deliberate, throttling IO wait (this task
5202 * has set its backing_dev_info: the queue against which it should throttle)
5204 void __sched io_schedule(void)
5206 struct rq *rq = &__raw_get_cpu_var(runqueues);
5208 delayacct_blkio_start();
5209 atomic_inc(&rq->nr_iowait);
5211 atomic_dec(&rq->nr_iowait);
5212 delayacct_blkio_end();
5214 EXPORT_SYMBOL(io_schedule);
5216 long __sched io_schedule_timeout(long timeout)
5218 struct rq *rq = &__raw_get_cpu_var(runqueues);
5221 delayacct_blkio_start();
5222 atomic_inc(&rq->nr_iowait);
5223 ret = schedule_timeout(timeout);
5224 atomic_dec(&rq->nr_iowait);
5225 delayacct_blkio_end();
5230 * sys_sched_get_priority_max - return maximum RT priority.
5231 * @policy: scheduling class.
5233 * this syscall returns the maximum rt_priority that can be used
5234 * by a given scheduling class.
5236 asmlinkage long sys_sched_get_priority_max(int policy)
5243 ret = MAX_USER_RT_PRIO-1;
5255 * sys_sched_get_priority_min - return minimum RT priority.
5256 * @policy: scheduling class.
5258 * this syscall returns the minimum rt_priority that can be used
5259 * by a given scheduling class.
5261 asmlinkage long sys_sched_get_priority_min(int policy)
5279 * sys_sched_rr_get_interval - return the default timeslice of a process.
5280 * @pid: pid of the process.
5281 * @interval: userspace pointer to the timeslice value.
5283 * this syscall writes the default timeslice value of a given process
5284 * into the user-space timespec buffer. A value of '0' means infinity.
5287 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5289 struct task_struct *p;
5290 unsigned int time_slice;
5298 read_lock(&tasklist_lock);
5299 p = find_process_by_pid(pid);
5303 retval = security_task_getscheduler(p);
5308 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5309 * tasks that are on an otherwise idle runqueue:
5312 if (p->policy == SCHED_RR) {
5313 time_slice = DEF_TIMESLICE;
5314 } else if (p->policy != SCHED_FIFO) {
5315 struct sched_entity *se = &p->se;
5316 unsigned long flags;
5319 rq = task_rq_lock(p, &flags);
5320 if (rq->cfs.load.weight)
5321 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5322 task_rq_unlock(rq, &flags);
5324 read_unlock(&tasklist_lock);
5325 jiffies_to_timespec(time_slice, &t);
5326 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5330 read_unlock(&tasklist_lock);
5334 static const char stat_nam[] = "RSDTtZX";
5336 void sched_show_task(struct task_struct *p)
5338 unsigned long free = 0;
5341 state = p->state ? __ffs(p->state) + 1 : 0;
5342 printk(KERN_INFO "%-13.13s %c", p->comm,
5343 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5344 #if BITS_PER_LONG == 32
5345 if (state == TASK_RUNNING)
5346 printk(KERN_CONT " running ");
5348 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5350 if (state == TASK_RUNNING)
5351 printk(KERN_CONT " running task ");
5353 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5355 #ifdef CONFIG_DEBUG_STACK_USAGE
5357 unsigned long *n = end_of_stack(p);
5360 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5363 printk(KERN_CONT "%5lu %5d %6d\n", free,
5364 task_pid_nr(p), task_pid_nr(p->real_parent));
5366 show_stack(p, NULL);
5369 void show_state_filter(unsigned long state_filter)
5371 struct task_struct *g, *p;
5373 #if BITS_PER_LONG == 32
5375 " task PC stack pid father\n");
5378 " task PC stack pid father\n");
5380 read_lock(&tasklist_lock);
5381 do_each_thread(g, p) {
5383 * reset the NMI-timeout, listing all files on a slow
5384 * console might take alot of time:
5386 touch_nmi_watchdog();
5387 if (!state_filter || (p->state & state_filter))
5389 } while_each_thread(g, p);
5391 touch_all_softlockup_watchdogs();
5393 #ifdef CONFIG_SCHED_DEBUG
5394 sysrq_sched_debug_show();
5396 read_unlock(&tasklist_lock);
5398 * Only show locks if all tasks are dumped:
5400 if (state_filter == -1)
5401 debug_show_all_locks();
5404 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5406 idle->sched_class = &idle_sched_class;
5410 * init_idle - set up an idle thread for a given CPU
5411 * @idle: task in question
5412 * @cpu: cpu the idle task belongs to
5414 * NOTE: this function does not set the idle thread's NEED_RESCHED
5415 * flag, to make booting more robust.
5417 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5419 struct rq *rq = cpu_rq(cpu);
5420 unsigned long flags;
5423 idle->se.exec_start = sched_clock();
5425 idle->prio = idle->normal_prio = MAX_PRIO;
5426 idle->cpus_allowed = cpumask_of_cpu(cpu);
5427 __set_task_cpu(idle, cpu);
5429 spin_lock_irqsave(&rq->lock, flags);
5430 rq->curr = rq->idle = idle;
5431 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5434 spin_unlock_irqrestore(&rq->lock, flags);
5436 /* Set the preempt count _outside_ the spinlocks! */
5437 task_thread_info(idle)->preempt_count = 0;
5440 * The idle tasks have their own, simple scheduling class:
5442 idle->sched_class = &idle_sched_class;
5446 * In a system that switches off the HZ timer nohz_cpu_mask
5447 * indicates which cpus entered this state. This is used
5448 * in the rcu update to wait only for active cpus. For system
5449 * which do not switch off the HZ timer nohz_cpu_mask should
5450 * always be CPU_MASK_NONE.
5452 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5455 * Increase the granularity value when there are more CPUs,
5456 * because with more CPUs the 'effective latency' as visible
5457 * to users decreases. But the relationship is not linear,
5458 * so pick a second-best guess by going with the log2 of the
5461 * This idea comes from the SD scheduler of Con Kolivas:
5463 static inline void sched_init_granularity(void)
5465 unsigned int factor = 1 + ilog2(num_online_cpus());
5466 const unsigned long limit = 200000000;
5468 sysctl_sched_min_granularity *= factor;
5469 if (sysctl_sched_min_granularity > limit)
5470 sysctl_sched_min_granularity = limit;
5472 sysctl_sched_latency *= factor;
5473 if (sysctl_sched_latency > limit)
5474 sysctl_sched_latency = limit;
5476 sysctl_sched_wakeup_granularity *= factor;
5481 * This is how migration works:
5483 * 1) we queue a struct migration_req structure in the source CPU's
5484 * runqueue and wake up that CPU's migration thread.
5485 * 2) we down() the locked semaphore => thread blocks.
5486 * 3) migration thread wakes up (implicitly it forces the migrated
5487 * thread off the CPU)
5488 * 4) it gets the migration request and checks whether the migrated
5489 * task is still in the wrong runqueue.
5490 * 5) if it's in the wrong runqueue then the migration thread removes
5491 * it and puts it into the right queue.
5492 * 6) migration thread up()s the semaphore.
5493 * 7) we wake up and the migration is done.
5497 * Change a given task's CPU affinity. Migrate the thread to a
5498 * proper CPU and schedule it away if the CPU it's executing on
5499 * is removed from the allowed bitmask.
5501 * NOTE: the caller must have a valid reference to the task, the
5502 * task must not exit() & deallocate itself prematurely. The
5503 * call is not atomic; no spinlocks may be held.
5505 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5507 struct migration_req req;
5508 unsigned long flags;
5512 rq = task_rq_lock(p, &flags);
5513 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5518 if (p->sched_class->set_cpus_allowed)
5519 p->sched_class->set_cpus_allowed(p, new_mask);
5521 p->cpus_allowed = *new_mask;
5522 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5525 /* Can the task run on the task's current CPU? If so, we're done */
5526 if (cpu_isset(task_cpu(p), *new_mask))
5529 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5530 /* Need help from migration thread: drop lock and wait. */
5531 task_rq_unlock(rq, &flags);
5532 wake_up_process(rq->migration_thread);
5533 wait_for_completion(&req.done);
5534 tlb_migrate_finish(p->mm);
5538 task_rq_unlock(rq, &flags);
5542 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5545 * Move (not current) task off this cpu, onto dest cpu. We're doing
5546 * this because either it can't run here any more (set_cpus_allowed()
5547 * away from this CPU, or CPU going down), or because we're
5548 * attempting to rebalance this task on exec (sched_exec).
5550 * So we race with normal scheduler movements, but that's OK, as long
5551 * as the task is no longer on this CPU.
5553 * Returns non-zero if task was successfully migrated.
5555 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5557 struct rq *rq_dest, *rq_src;
5560 if (unlikely(cpu_is_offline(dest_cpu)))
5563 rq_src = cpu_rq(src_cpu);
5564 rq_dest = cpu_rq(dest_cpu);
5566 double_rq_lock(rq_src, rq_dest);
5567 /* Already moved. */
5568 if (task_cpu(p) != src_cpu)
5570 /* Affinity changed (again). */
5571 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5574 on_rq = p->se.on_rq;
5576 deactivate_task(rq_src, p, 0);
5578 set_task_cpu(p, dest_cpu);
5580 activate_task(rq_dest, p, 0);
5581 check_preempt_curr(rq_dest, p);
5585 double_rq_unlock(rq_src, rq_dest);
5590 * migration_thread - this is a highprio system thread that performs
5591 * thread migration by bumping thread off CPU then 'pushing' onto
5594 static int migration_thread(void *data)
5596 int cpu = (long)data;
5600 BUG_ON(rq->migration_thread != current);
5602 set_current_state(TASK_INTERRUPTIBLE);
5603 while (!kthread_should_stop()) {
5604 struct migration_req *req;
5605 struct list_head *head;
5607 spin_lock_irq(&rq->lock);
5609 if (cpu_is_offline(cpu)) {
5610 spin_unlock_irq(&rq->lock);
5614 if (rq->active_balance) {
5615 active_load_balance(rq, cpu);
5616 rq->active_balance = 0;
5619 head = &rq->migration_queue;
5621 if (list_empty(head)) {
5622 spin_unlock_irq(&rq->lock);
5624 set_current_state(TASK_INTERRUPTIBLE);
5627 req = list_entry(head->next, struct migration_req, list);
5628 list_del_init(head->next);
5630 spin_unlock(&rq->lock);
5631 __migrate_task(req->task, cpu, req->dest_cpu);
5634 complete(&req->done);
5636 __set_current_state(TASK_RUNNING);
5640 /* Wait for kthread_stop */
5641 set_current_state(TASK_INTERRUPTIBLE);
5642 while (!kthread_should_stop()) {
5644 set_current_state(TASK_INTERRUPTIBLE);
5646 __set_current_state(TASK_RUNNING);
5650 #ifdef CONFIG_HOTPLUG_CPU
5652 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5656 local_irq_disable();
5657 ret = __migrate_task(p, src_cpu, dest_cpu);
5663 * Figure out where task on dead CPU should go, use force if necessary.
5664 * NOTE: interrupts should be disabled by the caller
5666 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5668 unsigned long flags;
5675 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5676 cpus_and(mask, mask, p->cpus_allowed);
5677 dest_cpu = any_online_cpu(mask);
5679 /* On any allowed CPU? */
5680 if (dest_cpu >= nr_cpu_ids)
5681 dest_cpu = any_online_cpu(p->cpus_allowed);
5683 /* No more Mr. Nice Guy. */
5684 if (dest_cpu >= nr_cpu_ids) {
5685 cpumask_t cpus_allowed;
5687 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5689 * Try to stay on the same cpuset, where the
5690 * current cpuset may be a subset of all cpus.
5691 * The cpuset_cpus_allowed_locked() variant of
5692 * cpuset_cpus_allowed() will not block. It must be
5693 * called within calls to cpuset_lock/cpuset_unlock.
5695 rq = task_rq_lock(p, &flags);
5696 p->cpus_allowed = cpus_allowed;
5697 dest_cpu = any_online_cpu(p->cpus_allowed);
5698 task_rq_unlock(rq, &flags);
5701 * Don't tell them about moving exiting tasks or
5702 * kernel threads (both mm NULL), since they never
5705 if (p->mm && printk_ratelimit()) {
5706 printk(KERN_INFO "process %d (%s) no "
5707 "longer affine to cpu%d\n",
5708 task_pid_nr(p), p->comm, dead_cpu);
5711 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5715 * While a dead CPU has no uninterruptible tasks queued at this point,
5716 * it might still have a nonzero ->nr_uninterruptible counter, because
5717 * for performance reasons the counter is not stricly tracking tasks to
5718 * their home CPUs. So we just add the counter to another CPU's counter,
5719 * to keep the global sum constant after CPU-down:
5721 static void migrate_nr_uninterruptible(struct rq *rq_src)
5723 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5724 unsigned long flags;
5726 local_irq_save(flags);
5727 double_rq_lock(rq_src, rq_dest);
5728 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5729 rq_src->nr_uninterruptible = 0;
5730 double_rq_unlock(rq_src, rq_dest);
5731 local_irq_restore(flags);
5734 /* Run through task list and migrate tasks from the dead cpu. */
5735 static void migrate_live_tasks(int src_cpu)
5737 struct task_struct *p, *t;
5739 read_lock(&tasklist_lock);
5741 do_each_thread(t, p) {
5745 if (task_cpu(p) == src_cpu)
5746 move_task_off_dead_cpu(src_cpu, p);
5747 } while_each_thread(t, p);
5749 read_unlock(&tasklist_lock);
5753 * Schedules idle task to be the next runnable task on current CPU.
5754 * It does so by boosting its priority to highest possible.
5755 * Used by CPU offline code.
5757 void sched_idle_next(void)
5759 int this_cpu = smp_processor_id();
5760 struct rq *rq = cpu_rq(this_cpu);
5761 struct task_struct *p = rq->idle;
5762 unsigned long flags;
5764 /* cpu has to be offline */
5765 BUG_ON(cpu_online(this_cpu));
5768 * Strictly not necessary since rest of the CPUs are stopped by now
5769 * and interrupts disabled on the current cpu.
5771 spin_lock_irqsave(&rq->lock, flags);
5773 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5775 update_rq_clock(rq);
5776 activate_task(rq, p, 0);
5778 spin_unlock_irqrestore(&rq->lock, flags);
5782 * Ensures that the idle task is using init_mm right before its cpu goes
5785 void idle_task_exit(void)
5787 struct mm_struct *mm = current->active_mm;
5789 BUG_ON(cpu_online(smp_processor_id()));
5792 switch_mm(mm, &init_mm, current);
5796 /* called under rq->lock with disabled interrupts */
5797 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5799 struct rq *rq = cpu_rq(dead_cpu);
5801 /* Must be exiting, otherwise would be on tasklist. */
5802 BUG_ON(!p->exit_state);
5804 /* Cannot have done final schedule yet: would have vanished. */
5805 BUG_ON(p->state == TASK_DEAD);
5810 * Drop lock around migration; if someone else moves it,
5811 * that's OK. No task can be added to this CPU, so iteration is
5814 spin_unlock_irq(&rq->lock);
5815 move_task_off_dead_cpu(dead_cpu, p);
5816 spin_lock_irq(&rq->lock);
5821 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5822 static void migrate_dead_tasks(unsigned int dead_cpu)
5824 struct rq *rq = cpu_rq(dead_cpu);
5825 struct task_struct *next;
5828 if (!rq->nr_running)
5830 update_rq_clock(rq);
5831 next = pick_next_task(rq, rq->curr);
5834 migrate_dead(dead_cpu, next);
5838 #endif /* CONFIG_HOTPLUG_CPU */
5840 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5842 static struct ctl_table sd_ctl_dir[] = {
5844 .procname = "sched_domain",
5850 static struct ctl_table sd_ctl_root[] = {
5852 .ctl_name = CTL_KERN,
5853 .procname = "kernel",
5855 .child = sd_ctl_dir,
5860 static struct ctl_table *sd_alloc_ctl_entry(int n)
5862 struct ctl_table *entry =
5863 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5868 static void sd_free_ctl_entry(struct ctl_table **tablep)
5870 struct ctl_table *entry;
5873 * In the intermediate directories, both the child directory and
5874 * procname are dynamically allocated and could fail but the mode
5875 * will always be set. In the lowest directory the names are
5876 * static strings and all have proc handlers.
5878 for (entry = *tablep; entry->mode; entry++) {
5880 sd_free_ctl_entry(&entry->child);
5881 if (entry->proc_handler == NULL)
5882 kfree(entry->procname);
5890 set_table_entry(struct ctl_table *entry,
5891 const char *procname, void *data, int maxlen,
5892 mode_t mode, proc_handler *proc_handler)
5894 entry->procname = procname;
5896 entry->maxlen = maxlen;
5898 entry->proc_handler = proc_handler;
5901 static struct ctl_table *
5902 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5904 struct ctl_table *table = sd_alloc_ctl_entry(12);
5909 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5910 sizeof(long), 0644, proc_doulongvec_minmax);
5911 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5912 sizeof(long), 0644, proc_doulongvec_minmax);
5913 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5914 sizeof(int), 0644, proc_dointvec_minmax);
5915 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5916 sizeof(int), 0644, proc_dointvec_minmax);
5917 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5918 sizeof(int), 0644, proc_dointvec_minmax);
5919 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5920 sizeof(int), 0644, proc_dointvec_minmax);
5921 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5922 sizeof(int), 0644, proc_dointvec_minmax);
5923 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5924 sizeof(int), 0644, proc_dointvec_minmax);
5925 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5926 sizeof(int), 0644, proc_dointvec_minmax);
5927 set_table_entry(&table[9], "cache_nice_tries",
5928 &sd->cache_nice_tries,
5929 sizeof(int), 0644, proc_dointvec_minmax);
5930 set_table_entry(&table[10], "flags", &sd->flags,
5931 sizeof(int), 0644, proc_dointvec_minmax);
5932 /* &table[11] is terminator */
5937 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5939 struct ctl_table *entry, *table;
5940 struct sched_domain *sd;
5941 int domain_num = 0, i;
5944 for_each_domain(cpu, sd)
5946 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5951 for_each_domain(cpu, sd) {
5952 snprintf(buf, 32, "domain%d", i);
5953 entry->procname = kstrdup(buf, GFP_KERNEL);
5955 entry->child = sd_alloc_ctl_domain_table(sd);
5962 static struct ctl_table_header *sd_sysctl_header;
5963 static void register_sched_domain_sysctl(void)
5965 int i, cpu_num = num_online_cpus();
5966 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5969 WARN_ON(sd_ctl_dir[0].child);
5970 sd_ctl_dir[0].child = entry;
5975 for_each_online_cpu(i) {
5976 snprintf(buf, 32, "cpu%d", i);
5977 entry->procname = kstrdup(buf, GFP_KERNEL);
5979 entry->child = sd_alloc_ctl_cpu_table(i);
5983 WARN_ON(sd_sysctl_header);
5984 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5987 /* may be called multiple times per register */
5988 static void unregister_sched_domain_sysctl(void)
5990 if (sd_sysctl_header)
5991 unregister_sysctl_table(sd_sysctl_header);
5992 sd_sysctl_header = NULL;
5993 if (sd_ctl_dir[0].child)
5994 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5997 static void register_sched_domain_sysctl(void)
6000 static void unregister_sched_domain_sysctl(void)
6006 * migration_call - callback that gets triggered when a CPU is added.
6007 * Here we can start up the necessary migration thread for the new CPU.
6009 static int __cpuinit
6010 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6012 struct task_struct *p;
6013 int cpu = (long)hcpu;
6014 unsigned long flags;
6019 case CPU_UP_PREPARE:
6020 case CPU_UP_PREPARE_FROZEN:
6021 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6024 kthread_bind(p, cpu);
6025 /* Must be high prio: stop_machine expects to yield to it. */
6026 rq = task_rq_lock(p, &flags);
6027 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6028 task_rq_unlock(rq, &flags);
6029 cpu_rq(cpu)->migration_thread = p;
6033 case CPU_ONLINE_FROZEN:
6034 /* Strictly unnecessary, as first user will wake it. */
6035 wake_up_process(cpu_rq(cpu)->migration_thread);
6037 /* Update our root-domain */
6039 spin_lock_irqsave(&rq->lock, flags);
6041 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6042 cpu_set(cpu, rq->rd->online);
6044 spin_unlock_irqrestore(&rq->lock, flags);
6047 #ifdef CONFIG_HOTPLUG_CPU
6048 case CPU_UP_CANCELED:
6049 case CPU_UP_CANCELED_FROZEN:
6050 if (!cpu_rq(cpu)->migration_thread)
6052 /* Unbind it from offline cpu so it can run. Fall thru. */
6053 kthread_bind(cpu_rq(cpu)->migration_thread,
6054 any_online_cpu(cpu_online_map));
6055 kthread_stop(cpu_rq(cpu)->migration_thread);
6056 cpu_rq(cpu)->migration_thread = NULL;
6060 case CPU_DEAD_FROZEN:
6061 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6062 migrate_live_tasks(cpu);
6064 kthread_stop(rq->migration_thread);
6065 rq->migration_thread = NULL;
6066 /* Idle task back to normal (off runqueue, low prio) */
6067 spin_lock_irq(&rq->lock);
6068 update_rq_clock(rq);
6069 deactivate_task(rq, rq->idle, 0);
6070 rq->idle->static_prio = MAX_PRIO;
6071 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6072 rq->idle->sched_class = &idle_sched_class;
6073 migrate_dead_tasks(cpu);
6074 spin_unlock_irq(&rq->lock);
6076 migrate_nr_uninterruptible(rq);
6077 BUG_ON(rq->nr_running != 0);
6080 * No need to migrate the tasks: it was best-effort if
6081 * they didn't take sched_hotcpu_mutex. Just wake up
6084 spin_lock_irq(&rq->lock);
6085 while (!list_empty(&rq->migration_queue)) {
6086 struct migration_req *req;
6088 req = list_entry(rq->migration_queue.next,
6089 struct migration_req, list);
6090 list_del_init(&req->list);
6091 complete(&req->done);
6093 spin_unlock_irq(&rq->lock);
6097 case CPU_DYING_FROZEN:
6098 /* Update our root-domain */
6100 spin_lock_irqsave(&rq->lock, flags);
6102 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6103 cpu_clear(cpu, rq->rd->online);
6105 spin_unlock_irqrestore(&rq->lock, flags);
6112 /* Register at highest priority so that task migration (migrate_all_tasks)
6113 * happens before everything else.
6115 static struct notifier_block __cpuinitdata migration_notifier = {
6116 .notifier_call = migration_call,
6120 void __init migration_init(void)
6122 void *cpu = (void *)(long)smp_processor_id();
6125 /* Start one for the boot CPU: */
6126 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6127 BUG_ON(err == NOTIFY_BAD);
6128 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6129 register_cpu_notifier(&migration_notifier);
6135 #ifdef CONFIG_SCHED_DEBUG
6137 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6138 cpumask_t *groupmask)
6140 struct sched_group *group = sd->groups;
6143 cpulist_scnprintf(str, sizeof(str), sd->span);
6144 cpus_clear(*groupmask);
6146 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6148 if (!(sd->flags & SD_LOAD_BALANCE)) {
6149 printk("does not load-balance\n");
6151 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6156 printk(KERN_CONT "span %s\n", str);
6158 if (!cpu_isset(cpu, sd->span)) {
6159 printk(KERN_ERR "ERROR: domain->span does not contain "
6162 if (!cpu_isset(cpu, group->cpumask)) {
6163 printk(KERN_ERR "ERROR: domain->groups does not contain"
6167 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6171 printk(KERN_ERR "ERROR: group is NULL\n");
6175 if (!group->__cpu_power) {
6176 printk(KERN_CONT "\n");
6177 printk(KERN_ERR "ERROR: domain->cpu_power not "
6182 if (!cpus_weight(group->cpumask)) {
6183 printk(KERN_CONT "\n");
6184 printk(KERN_ERR "ERROR: empty group\n");
6188 if (cpus_intersects(*groupmask, group->cpumask)) {
6189 printk(KERN_CONT "\n");
6190 printk(KERN_ERR "ERROR: repeated CPUs\n");
6194 cpus_or(*groupmask, *groupmask, group->cpumask);
6196 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6197 printk(KERN_CONT " %s", str);
6199 group = group->next;
6200 } while (group != sd->groups);
6201 printk(KERN_CONT "\n");
6203 if (!cpus_equal(sd->span, *groupmask))
6204 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6206 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6207 printk(KERN_ERR "ERROR: parent span is not a superset "
6208 "of domain->span\n");
6212 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6214 cpumask_t *groupmask;
6218 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6222 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6224 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6226 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6231 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6241 # define sched_domain_debug(sd, cpu) do { } while (0)
6244 static int sd_degenerate(struct sched_domain *sd)
6246 if (cpus_weight(sd->span) == 1)
6249 /* Following flags need at least 2 groups */
6250 if (sd->flags & (SD_LOAD_BALANCE |
6251 SD_BALANCE_NEWIDLE |
6255 SD_SHARE_PKG_RESOURCES)) {
6256 if (sd->groups != sd->groups->next)
6260 /* Following flags don't use groups */
6261 if (sd->flags & (SD_WAKE_IDLE |
6270 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6272 unsigned long cflags = sd->flags, pflags = parent->flags;
6274 if (sd_degenerate(parent))
6277 if (!cpus_equal(sd->span, parent->span))
6280 /* Does parent contain flags not in child? */
6281 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6282 if (cflags & SD_WAKE_AFFINE)
6283 pflags &= ~SD_WAKE_BALANCE;
6284 /* Flags needing groups don't count if only 1 group in parent */
6285 if (parent->groups == parent->groups->next) {
6286 pflags &= ~(SD_LOAD_BALANCE |
6287 SD_BALANCE_NEWIDLE |
6291 SD_SHARE_PKG_RESOURCES);
6293 if (~cflags & pflags)
6299 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6301 unsigned long flags;
6302 const struct sched_class *class;
6304 spin_lock_irqsave(&rq->lock, flags);
6307 struct root_domain *old_rd = rq->rd;
6309 for (class = sched_class_highest; class; class = class->next) {
6310 if (class->leave_domain)
6311 class->leave_domain(rq);
6314 cpu_clear(rq->cpu, old_rd->span);
6315 cpu_clear(rq->cpu, old_rd->online);
6317 if (atomic_dec_and_test(&old_rd->refcount))
6321 atomic_inc(&rd->refcount);
6324 cpu_set(rq->cpu, rd->span);
6325 if (cpu_isset(rq->cpu, cpu_online_map))
6326 cpu_set(rq->cpu, rd->online);
6328 for (class = sched_class_highest; class; class = class->next) {
6329 if (class->join_domain)
6330 class->join_domain(rq);
6333 spin_unlock_irqrestore(&rq->lock, flags);
6336 static void init_rootdomain(struct root_domain *rd)
6338 memset(rd, 0, sizeof(*rd));
6340 cpus_clear(rd->span);
6341 cpus_clear(rd->online);
6344 static void init_defrootdomain(void)
6346 init_rootdomain(&def_root_domain);
6347 atomic_set(&def_root_domain.refcount, 1);
6350 static struct root_domain *alloc_rootdomain(void)
6352 struct root_domain *rd;
6354 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6358 init_rootdomain(rd);
6364 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6365 * hold the hotplug lock.
6368 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6370 struct rq *rq = cpu_rq(cpu);
6371 struct sched_domain *tmp;
6373 /* Remove the sched domains which do not contribute to scheduling. */
6374 for (tmp = sd; tmp; tmp = tmp->parent) {
6375 struct sched_domain *parent = tmp->parent;
6378 if (sd_parent_degenerate(tmp, parent)) {
6379 tmp->parent = parent->parent;
6381 parent->parent->child = tmp;
6385 if (sd && sd_degenerate(sd)) {
6391 sched_domain_debug(sd, cpu);
6393 rq_attach_root(rq, rd);
6394 rcu_assign_pointer(rq->sd, sd);
6397 /* cpus with isolated domains */
6398 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6400 /* Setup the mask of cpus configured for isolated domains */
6401 static int __init isolated_cpu_setup(char *str)
6403 int ints[NR_CPUS], i;
6405 str = get_options(str, ARRAY_SIZE(ints), ints);
6406 cpus_clear(cpu_isolated_map);
6407 for (i = 1; i <= ints[0]; i++)
6408 if (ints[i] < NR_CPUS)
6409 cpu_set(ints[i], cpu_isolated_map);
6413 __setup("isolcpus=", isolated_cpu_setup);
6416 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6417 * to a function which identifies what group(along with sched group) a CPU
6418 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6419 * (due to the fact that we keep track of groups covered with a cpumask_t).
6421 * init_sched_build_groups will build a circular linked list of the groups
6422 * covered by the given span, and will set each group's ->cpumask correctly,
6423 * and ->cpu_power to 0.
6426 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6427 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6428 struct sched_group **sg,
6429 cpumask_t *tmpmask),
6430 cpumask_t *covered, cpumask_t *tmpmask)
6432 struct sched_group *first = NULL, *last = NULL;
6435 cpus_clear(*covered);
6437 for_each_cpu_mask(i, *span) {
6438 struct sched_group *sg;
6439 int group = group_fn(i, cpu_map, &sg, tmpmask);
6442 if (cpu_isset(i, *covered))
6445 cpus_clear(sg->cpumask);
6446 sg->__cpu_power = 0;
6448 for_each_cpu_mask(j, *span) {
6449 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6452 cpu_set(j, *covered);
6453 cpu_set(j, sg->cpumask);
6464 #define SD_NODES_PER_DOMAIN 16
6469 * find_next_best_node - find the next node to include in a sched_domain
6470 * @node: node whose sched_domain we're building
6471 * @used_nodes: nodes already in the sched_domain
6473 * Find the next node to include in a given scheduling domain. Simply
6474 * finds the closest node not already in the @used_nodes map.
6476 * Should use nodemask_t.
6478 static int find_next_best_node(int node, nodemask_t *used_nodes)
6480 int i, n, val, min_val, best_node = 0;
6484 for (i = 0; i < MAX_NUMNODES; i++) {
6485 /* Start at @node */
6486 n = (node + i) % MAX_NUMNODES;
6488 if (!nr_cpus_node(n))
6491 /* Skip already used nodes */
6492 if (node_isset(n, *used_nodes))
6495 /* Simple min distance search */
6496 val = node_distance(node, n);
6498 if (val < min_val) {
6504 node_set(best_node, *used_nodes);
6509 * sched_domain_node_span - get a cpumask for a node's sched_domain
6510 * @node: node whose cpumask we're constructing
6512 * Given a node, construct a good cpumask for its sched_domain to span. It
6513 * should be one that prevents unnecessary balancing, but also spreads tasks
6516 static void sched_domain_node_span(int node, cpumask_t *span)
6518 nodemask_t used_nodes;
6519 node_to_cpumask_ptr(nodemask, node);
6523 nodes_clear(used_nodes);
6525 cpus_or(*span, *span, *nodemask);
6526 node_set(node, used_nodes);
6528 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6529 int next_node = find_next_best_node(node, &used_nodes);
6531 node_to_cpumask_ptr_next(nodemask, next_node);
6532 cpus_or(*span, *span, *nodemask);
6537 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6540 * SMT sched-domains:
6542 #ifdef CONFIG_SCHED_SMT
6543 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6544 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6547 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6551 *sg = &per_cpu(sched_group_cpus, cpu);
6557 * multi-core sched-domains:
6559 #ifdef CONFIG_SCHED_MC
6560 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6561 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6564 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6566 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6571 *mask = per_cpu(cpu_sibling_map, cpu);
6572 cpus_and(*mask, *mask, *cpu_map);
6573 group = first_cpu(*mask);
6575 *sg = &per_cpu(sched_group_core, group);
6578 #elif defined(CONFIG_SCHED_MC)
6580 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6584 *sg = &per_cpu(sched_group_core, cpu);
6589 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6590 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6593 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6597 #ifdef CONFIG_SCHED_MC
6598 *mask = cpu_coregroup_map(cpu);
6599 cpus_and(*mask, *mask, *cpu_map);
6600 group = first_cpu(*mask);
6601 #elif defined(CONFIG_SCHED_SMT)
6602 *mask = per_cpu(cpu_sibling_map, cpu);
6603 cpus_and(*mask, *mask, *cpu_map);
6604 group = first_cpu(*mask);
6609 *sg = &per_cpu(sched_group_phys, group);
6615 * The init_sched_build_groups can't handle what we want to do with node
6616 * groups, so roll our own. Now each node has its own list of groups which
6617 * gets dynamically allocated.
6619 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6620 static struct sched_group ***sched_group_nodes_bycpu;
6622 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6623 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6625 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6626 struct sched_group **sg, cpumask_t *nodemask)
6630 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6631 cpus_and(*nodemask, *nodemask, *cpu_map);
6632 group = first_cpu(*nodemask);
6635 *sg = &per_cpu(sched_group_allnodes, group);
6639 static void init_numa_sched_groups_power(struct sched_group *group_head)
6641 struct sched_group *sg = group_head;
6647 for_each_cpu_mask(j, sg->cpumask) {
6648 struct sched_domain *sd;
6650 sd = &per_cpu(phys_domains, j);
6651 if (j != first_cpu(sd->groups->cpumask)) {
6653 * Only add "power" once for each
6659 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6662 } while (sg != group_head);
6667 /* Free memory allocated for various sched_group structures */
6668 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6672 for_each_cpu_mask(cpu, *cpu_map) {
6673 struct sched_group **sched_group_nodes
6674 = sched_group_nodes_bycpu[cpu];
6676 if (!sched_group_nodes)
6679 for (i = 0; i < MAX_NUMNODES; i++) {
6680 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6682 *nodemask = node_to_cpumask(i);
6683 cpus_and(*nodemask, *nodemask, *cpu_map);
6684 if (cpus_empty(*nodemask))
6694 if (oldsg != sched_group_nodes[i])
6697 kfree(sched_group_nodes);
6698 sched_group_nodes_bycpu[cpu] = NULL;
6702 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6708 * Initialize sched groups cpu_power.
6710 * cpu_power indicates the capacity of sched group, which is used while
6711 * distributing the load between different sched groups in a sched domain.
6712 * Typically cpu_power for all the groups in a sched domain will be same unless
6713 * there are asymmetries in the topology. If there are asymmetries, group
6714 * having more cpu_power will pickup more load compared to the group having
6717 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6718 * the maximum number of tasks a group can handle in the presence of other idle
6719 * or lightly loaded groups in the same sched domain.
6721 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6723 struct sched_domain *child;
6724 struct sched_group *group;
6726 WARN_ON(!sd || !sd->groups);
6728 if (cpu != first_cpu(sd->groups->cpumask))
6733 sd->groups->__cpu_power = 0;
6736 * For perf policy, if the groups in child domain share resources
6737 * (for example cores sharing some portions of the cache hierarchy
6738 * or SMT), then set this domain groups cpu_power such that each group
6739 * can handle only one task, when there are other idle groups in the
6740 * same sched domain.
6742 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6744 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6745 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6750 * add cpu_power of each child group to this groups cpu_power
6752 group = child->groups;
6754 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6755 group = group->next;
6756 } while (group != child->groups);
6760 * Initializers for schedule domains
6761 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6764 #define SD_INIT(sd, type) sd_init_##type(sd)
6765 #define SD_INIT_FUNC(type) \
6766 static noinline void sd_init_##type(struct sched_domain *sd) \
6768 memset(sd, 0, sizeof(*sd)); \
6769 *sd = SD_##type##_INIT; \
6774 SD_INIT_FUNC(ALLNODES)
6777 #ifdef CONFIG_SCHED_SMT
6778 SD_INIT_FUNC(SIBLING)
6780 #ifdef CONFIG_SCHED_MC
6785 * To minimize stack usage kmalloc room for cpumasks and share the
6786 * space as the usage in build_sched_domains() dictates. Used only
6787 * if the amount of space is significant.
6790 cpumask_t tmpmask; /* make this one first */
6793 cpumask_t this_sibling_map;
6794 cpumask_t this_core_map;
6796 cpumask_t send_covered;
6799 cpumask_t domainspan;
6801 cpumask_t notcovered;
6806 #define SCHED_CPUMASK_ALLOC 1
6807 #define SCHED_CPUMASK_FREE(v) kfree(v)
6808 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6810 #define SCHED_CPUMASK_ALLOC 0
6811 #define SCHED_CPUMASK_FREE(v)
6812 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6815 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6816 ((unsigned long)(a) + offsetof(struct allmasks, v))
6819 * Build sched domains for a given set of cpus and attach the sched domains
6820 * to the individual cpus
6822 static int build_sched_domains(const cpumask_t *cpu_map)
6825 struct root_domain *rd;
6826 SCHED_CPUMASK_DECLARE(allmasks);
6829 struct sched_group **sched_group_nodes = NULL;
6830 int sd_allnodes = 0;
6833 * Allocate the per-node list of sched groups
6835 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6837 if (!sched_group_nodes) {
6838 printk(KERN_WARNING "Can not alloc sched group node list\n");
6843 rd = alloc_rootdomain();
6845 printk(KERN_WARNING "Cannot alloc root domain\n");
6847 kfree(sched_group_nodes);
6852 #if SCHED_CPUMASK_ALLOC
6853 /* get space for all scratch cpumask variables */
6854 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6856 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6859 kfree(sched_group_nodes);
6864 tmpmask = (cpumask_t *)allmasks;
6868 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6872 * Set up domains for cpus specified by the cpu_map.
6874 for_each_cpu_mask(i, *cpu_map) {
6875 struct sched_domain *sd = NULL, *p;
6876 SCHED_CPUMASK_VAR(nodemask, allmasks);
6878 *nodemask = node_to_cpumask(cpu_to_node(i));
6879 cpus_and(*nodemask, *nodemask, *cpu_map);
6882 if (cpus_weight(*cpu_map) >
6883 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6884 sd = &per_cpu(allnodes_domains, i);
6885 SD_INIT(sd, ALLNODES);
6886 sd->span = *cpu_map;
6887 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6893 sd = &per_cpu(node_domains, i);
6895 sched_domain_node_span(cpu_to_node(i), &sd->span);
6899 cpus_and(sd->span, sd->span, *cpu_map);
6903 sd = &per_cpu(phys_domains, i);
6905 sd->span = *nodemask;
6909 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6911 #ifdef CONFIG_SCHED_MC
6913 sd = &per_cpu(core_domains, i);
6915 sd->span = cpu_coregroup_map(i);
6916 cpus_and(sd->span, sd->span, *cpu_map);
6919 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
6922 #ifdef CONFIG_SCHED_SMT
6924 sd = &per_cpu(cpu_domains, i);
6925 SD_INIT(sd, SIBLING);
6926 sd->span = per_cpu(cpu_sibling_map, i);
6927 cpus_and(sd->span, sd->span, *cpu_map);
6930 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
6934 #ifdef CONFIG_SCHED_SMT
6935 /* Set up CPU (sibling) groups */
6936 for_each_cpu_mask(i, *cpu_map) {
6937 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
6938 SCHED_CPUMASK_VAR(send_covered, allmasks);
6940 *this_sibling_map = per_cpu(cpu_sibling_map, i);
6941 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
6942 if (i != first_cpu(*this_sibling_map))
6945 init_sched_build_groups(this_sibling_map, cpu_map,
6947 send_covered, tmpmask);
6951 #ifdef CONFIG_SCHED_MC
6952 /* Set up multi-core groups */
6953 for_each_cpu_mask(i, *cpu_map) {
6954 SCHED_CPUMASK_VAR(this_core_map, allmasks);
6955 SCHED_CPUMASK_VAR(send_covered, allmasks);
6957 *this_core_map = cpu_coregroup_map(i);
6958 cpus_and(*this_core_map, *this_core_map, *cpu_map);
6959 if (i != first_cpu(*this_core_map))
6962 init_sched_build_groups(this_core_map, cpu_map,
6964 send_covered, tmpmask);
6968 /* Set up physical groups */
6969 for (i = 0; i < MAX_NUMNODES; i++) {
6970 SCHED_CPUMASK_VAR(nodemask, allmasks);
6971 SCHED_CPUMASK_VAR(send_covered, allmasks);
6973 *nodemask = node_to_cpumask(i);
6974 cpus_and(*nodemask, *nodemask, *cpu_map);
6975 if (cpus_empty(*nodemask))
6978 init_sched_build_groups(nodemask, cpu_map,
6980 send_covered, tmpmask);
6984 /* Set up node groups */
6986 SCHED_CPUMASK_VAR(send_covered, allmasks);
6988 init_sched_build_groups(cpu_map, cpu_map,
6989 &cpu_to_allnodes_group,
6990 send_covered, tmpmask);
6993 for (i = 0; i < MAX_NUMNODES; i++) {
6994 /* Set up node groups */
6995 struct sched_group *sg, *prev;
6996 SCHED_CPUMASK_VAR(nodemask, allmasks);
6997 SCHED_CPUMASK_VAR(domainspan, allmasks);
6998 SCHED_CPUMASK_VAR(covered, allmasks);
7001 *nodemask = node_to_cpumask(i);
7002 cpus_clear(*covered);
7004 cpus_and(*nodemask, *nodemask, *cpu_map);
7005 if (cpus_empty(*nodemask)) {
7006 sched_group_nodes[i] = NULL;
7010 sched_domain_node_span(i, domainspan);
7011 cpus_and(*domainspan, *domainspan, *cpu_map);
7013 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7015 printk(KERN_WARNING "Can not alloc domain group for "
7019 sched_group_nodes[i] = sg;
7020 for_each_cpu_mask(j, *nodemask) {
7021 struct sched_domain *sd;
7023 sd = &per_cpu(node_domains, j);
7026 sg->__cpu_power = 0;
7027 sg->cpumask = *nodemask;
7029 cpus_or(*covered, *covered, *nodemask);
7032 for (j = 0; j < MAX_NUMNODES; j++) {
7033 SCHED_CPUMASK_VAR(notcovered, allmasks);
7034 int n = (i + j) % MAX_NUMNODES;
7035 node_to_cpumask_ptr(pnodemask, n);
7037 cpus_complement(*notcovered, *covered);
7038 cpus_and(*tmpmask, *notcovered, *cpu_map);
7039 cpus_and(*tmpmask, *tmpmask, *domainspan);
7040 if (cpus_empty(*tmpmask))
7043 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7044 if (cpus_empty(*tmpmask))
7047 sg = kmalloc_node(sizeof(struct sched_group),
7051 "Can not alloc domain group for node %d\n", j);
7054 sg->__cpu_power = 0;
7055 sg->cpumask = *tmpmask;
7056 sg->next = prev->next;
7057 cpus_or(*covered, *covered, *tmpmask);
7064 /* Calculate CPU power for physical packages and nodes */
7065 #ifdef CONFIG_SCHED_SMT
7066 for_each_cpu_mask(i, *cpu_map) {
7067 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7069 init_sched_groups_power(i, sd);
7072 #ifdef CONFIG_SCHED_MC
7073 for_each_cpu_mask(i, *cpu_map) {
7074 struct sched_domain *sd = &per_cpu(core_domains, i);
7076 init_sched_groups_power(i, sd);
7080 for_each_cpu_mask(i, *cpu_map) {
7081 struct sched_domain *sd = &per_cpu(phys_domains, i);
7083 init_sched_groups_power(i, sd);
7087 for (i = 0; i < MAX_NUMNODES; i++)
7088 init_numa_sched_groups_power(sched_group_nodes[i]);
7091 struct sched_group *sg;
7093 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7095 init_numa_sched_groups_power(sg);
7099 /* Attach the domains */
7100 for_each_cpu_mask(i, *cpu_map) {
7101 struct sched_domain *sd;
7102 #ifdef CONFIG_SCHED_SMT
7103 sd = &per_cpu(cpu_domains, i);
7104 #elif defined(CONFIG_SCHED_MC)
7105 sd = &per_cpu(core_domains, i);
7107 sd = &per_cpu(phys_domains, i);
7109 cpu_attach_domain(sd, rd, i);
7112 SCHED_CPUMASK_FREE((void *)allmasks);
7117 free_sched_groups(cpu_map, tmpmask);
7118 SCHED_CPUMASK_FREE((void *)allmasks);
7123 static cpumask_t *doms_cur; /* current sched domains */
7124 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7127 * Special case: If a kmalloc of a doms_cur partition (array of
7128 * cpumask_t) fails, then fallback to a single sched domain,
7129 * as determined by the single cpumask_t fallback_doms.
7131 static cpumask_t fallback_doms;
7133 void __attribute__((weak)) arch_update_cpu_topology(void)
7138 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7139 * For now this just excludes isolated cpus, but could be used to
7140 * exclude other special cases in the future.
7142 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7146 arch_update_cpu_topology();
7148 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7150 doms_cur = &fallback_doms;
7151 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7152 err = build_sched_domains(doms_cur);
7153 register_sched_domain_sysctl();
7158 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7161 free_sched_groups(cpu_map, tmpmask);
7165 * Detach sched domains from a group of cpus specified in cpu_map
7166 * These cpus will now be attached to the NULL domain
7168 static void detach_destroy_domains(const cpumask_t *cpu_map)
7173 unregister_sched_domain_sysctl();
7175 for_each_cpu_mask(i, *cpu_map)
7176 cpu_attach_domain(NULL, &def_root_domain, i);
7177 synchronize_sched();
7178 arch_destroy_sched_domains(cpu_map, &tmpmask);
7182 * Partition sched domains as specified by the 'ndoms_new'
7183 * cpumasks in the array doms_new[] of cpumasks. This compares
7184 * doms_new[] to the current sched domain partitioning, doms_cur[].
7185 * It destroys each deleted domain and builds each new domain.
7187 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7188 * The masks don't intersect (don't overlap.) We should setup one
7189 * sched domain for each mask. CPUs not in any of the cpumasks will
7190 * not be load balanced. If the same cpumask appears both in the
7191 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7194 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7195 * ownership of it and will kfree it when done with it. If the caller
7196 * failed the kmalloc call, then it can pass in doms_new == NULL,
7197 * and partition_sched_domains() will fallback to the single partition
7200 * Call with hotplug lock held
7202 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7208 /* always unregister in case we don't destroy any domains */
7209 unregister_sched_domain_sysctl();
7211 if (doms_new == NULL) {
7213 doms_new = &fallback_doms;
7214 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7217 /* Destroy deleted domains */
7218 for (i = 0; i < ndoms_cur; i++) {
7219 for (j = 0; j < ndoms_new; j++) {
7220 if (cpus_equal(doms_cur[i], doms_new[j]))
7223 /* no match - a current sched domain not in new doms_new[] */
7224 detach_destroy_domains(doms_cur + i);
7229 /* Build new domains */
7230 for (i = 0; i < ndoms_new; i++) {
7231 for (j = 0; j < ndoms_cur; j++) {
7232 if (cpus_equal(doms_new[i], doms_cur[j]))
7235 /* no match - add a new doms_new */
7236 build_sched_domains(doms_new + i);
7241 /* Remember the new sched domains */
7242 if (doms_cur != &fallback_doms)
7244 doms_cur = doms_new;
7245 ndoms_cur = ndoms_new;
7247 register_sched_domain_sysctl();
7252 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7253 int arch_reinit_sched_domains(void)
7258 detach_destroy_domains(&cpu_online_map);
7259 err = arch_init_sched_domains(&cpu_online_map);
7265 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7269 if (buf[0] != '0' && buf[0] != '1')
7273 sched_smt_power_savings = (buf[0] == '1');
7275 sched_mc_power_savings = (buf[0] == '1');
7277 ret = arch_reinit_sched_domains();
7279 return ret ? ret : count;
7282 #ifdef CONFIG_SCHED_MC
7283 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7285 return sprintf(page, "%u\n", sched_mc_power_savings);
7287 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7288 const char *buf, size_t count)
7290 return sched_power_savings_store(buf, count, 0);
7292 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7293 sched_mc_power_savings_store);
7296 #ifdef CONFIG_SCHED_SMT
7297 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7299 return sprintf(page, "%u\n", sched_smt_power_savings);
7301 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7302 const char *buf, size_t count)
7304 return sched_power_savings_store(buf, count, 1);
7306 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7307 sched_smt_power_savings_store);
7310 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7314 #ifdef CONFIG_SCHED_SMT
7316 err = sysfs_create_file(&cls->kset.kobj,
7317 &attr_sched_smt_power_savings.attr);
7319 #ifdef CONFIG_SCHED_MC
7320 if (!err && mc_capable())
7321 err = sysfs_create_file(&cls->kset.kobj,
7322 &attr_sched_mc_power_savings.attr);
7329 * Force a reinitialization of the sched domains hierarchy. The domains
7330 * and groups cannot be updated in place without racing with the balancing
7331 * code, so we temporarily attach all running cpus to the NULL domain
7332 * which will prevent rebalancing while the sched domains are recalculated.
7334 static int update_sched_domains(struct notifier_block *nfb,
7335 unsigned long action, void *hcpu)
7338 case CPU_UP_PREPARE:
7339 case CPU_UP_PREPARE_FROZEN:
7340 case CPU_DOWN_PREPARE:
7341 case CPU_DOWN_PREPARE_FROZEN:
7342 detach_destroy_domains(&cpu_online_map);
7345 case CPU_UP_CANCELED:
7346 case CPU_UP_CANCELED_FROZEN:
7347 case CPU_DOWN_FAILED:
7348 case CPU_DOWN_FAILED_FROZEN:
7350 case CPU_ONLINE_FROZEN:
7352 case CPU_DEAD_FROZEN:
7354 * Fall through and re-initialise the domains.
7361 /* The hotplug lock is already held by cpu_up/cpu_down */
7362 arch_init_sched_domains(&cpu_online_map);
7367 void __init sched_init_smp(void)
7369 cpumask_t non_isolated_cpus;
7371 #if defined(CONFIG_NUMA)
7372 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7374 BUG_ON(sched_group_nodes_bycpu == NULL);
7377 arch_init_sched_domains(&cpu_online_map);
7378 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7379 if (cpus_empty(non_isolated_cpus))
7380 cpu_set(smp_processor_id(), non_isolated_cpus);
7382 /* XXX: Theoretical race here - CPU may be hotplugged now */
7383 hotcpu_notifier(update_sched_domains, 0);
7385 /* Move init over to a non-isolated CPU */
7386 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7388 sched_init_granularity();
7391 void __init sched_init_smp(void)
7393 #if defined(CONFIG_NUMA)
7394 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7396 BUG_ON(sched_group_nodes_bycpu == NULL);
7398 sched_init_granularity();
7400 #endif /* CONFIG_SMP */
7402 int in_sched_functions(unsigned long addr)
7404 return in_lock_functions(addr) ||
7405 (addr >= (unsigned long)__sched_text_start
7406 && addr < (unsigned long)__sched_text_end);
7409 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7411 cfs_rq->tasks_timeline = RB_ROOT;
7412 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7418 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7420 struct rt_prio_array *array;
7423 array = &rt_rq->active;
7424 for (i = 0; i < MAX_RT_PRIO; i++) {
7425 INIT_LIST_HEAD(array->queue + i);
7426 __clear_bit(i, array->bitmap);
7428 /* delimiter for bitsearch: */
7429 __set_bit(MAX_RT_PRIO, array->bitmap);
7431 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7432 rt_rq->highest_prio = MAX_RT_PRIO;
7435 rt_rq->rt_nr_migratory = 0;
7436 rt_rq->overloaded = 0;
7440 rt_rq->rt_throttled = 0;
7441 rt_rq->rt_runtime = 0;
7442 spin_lock_init(&rt_rq->rt_runtime_lock);
7444 #ifdef CONFIG_RT_GROUP_SCHED
7445 rt_rq->rt_nr_boosted = 0;
7450 #ifdef CONFIG_FAIR_GROUP_SCHED
7451 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7452 struct sched_entity *se, int cpu, int add,
7453 struct sched_entity *parent)
7455 struct rq *rq = cpu_rq(cpu);
7456 tg->cfs_rq[cpu] = cfs_rq;
7457 init_cfs_rq(cfs_rq, rq);
7460 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7463 /* se could be NULL for init_task_group */
7468 se->cfs_rq = &rq->cfs;
7470 se->cfs_rq = parent->my_q;
7473 se->load.weight = tg->shares;
7474 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7475 se->parent = parent;
7479 #ifdef CONFIG_RT_GROUP_SCHED
7480 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7481 struct sched_rt_entity *rt_se, int cpu, int add,
7482 struct sched_rt_entity *parent)
7484 struct rq *rq = cpu_rq(cpu);
7486 tg->rt_rq[cpu] = rt_rq;
7487 init_rt_rq(rt_rq, rq);
7489 rt_rq->rt_se = rt_se;
7490 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7492 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7494 tg->rt_se[cpu] = rt_se;
7499 rt_se->rt_rq = &rq->rt;
7501 rt_se->rt_rq = parent->my_q;
7503 rt_se->rt_rq = &rq->rt;
7504 rt_se->my_q = rt_rq;
7505 rt_se->parent = parent;
7506 INIT_LIST_HEAD(&rt_se->run_list);
7510 void __init sched_init(void)
7513 unsigned long alloc_size = 0, ptr;
7515 #ifdef CONFIG_FAIR_GROUP_SCHED
7516 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7518 #ifdef CONFIG_RT_GROUP_SCHED
7519 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7521 #ifdef CONFIG_USER_SCHED
7525 * As sched_init() is called before page_alloc is setup,
7526 * we use alloc_bootmem().
7529 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7531 #ifdef CONFIG_FAIR_GROUP_SCHED
7532 init_task_group.se = (struct sched_entity **)ptr;
7533 ptr += nr_cpu_ids * sizeof(void **);
7535 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7536 ptr += nr_cpu_ids * sizeof(void **);
7538 #ifdef CONFIG_USER_SCHED
7539 root_task_group.se = (struct sched_entity **)ptr;
7540 ptr += nr_cpu_ids * sizeof(void **);
7542 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7543 ptr += nr_cpu_ids * sizeof(void **);
7546 #ifdef CONFIG_RT_GROUP_SCHED
7547 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7548 ptr += nr_cpu_ids * sizeof(void **);
7550 init_task_group.rt_rq = (struct rt_rq **)ptr;
7551 ptr += nr_cpu_ids * sizeof(void **);
7553 #ifdef CONFIG_USER_SCHED
7554 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7555 ptr += nr_cpu_ids * sizeof(void **);
7557 root_task_group.rt_rq = (struct rt_rq **)ptr;
7558 ptr += nr_cpu_ids * sizeof(void **);
7564 init_defrootdomain();
7567 init_rt_bandwidth(&def_rt_bandwidth,
7568 global_rt_period(), global_rt_runtime());
7570 #ifdef CONFIG_RT_GROUP_SCHED
7571 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7572 global_rt_period(), global_rt_runtime());
7573 #ifdef CONFIG_USER_SCHED
7574 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7575 global_rt_period(), RUNTIME_INF);
7579 #ifdef CONFIG_GROUP_SCHED
7580 list_add(&init_task_group.list, &task_groups);
7583 for_each_possible_cpu(i) {
7587 spin_lock_init(&rq->lock);
7588 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7591 update_last_tick_seen(rq);
7592 init_cfs_rq(&rq->cfs, rq);
7593 init_rt_rq(&rq->rt, rq);
7594 #ifdef CONFIG_FAIR_GROUP_SCHED
7595 init_task_group.shares = init_task_group_load;
7596 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7597 #ifdef CONFIG_CGROUP_SCHED
7599 * How much cpu bandwidth does init_task_group get?
7601 * In case of task-groups formed thr' the cgroup filesystem, it
7602 * gets 100% of the cpu resources in the system. This overall
7603 * system cpu resource is divided among the tasks of
7604 * init_task_group and its child task-groups in a fair manner,
7605 * based on each entity's (task or task-group's) weight
7606 * (se->load.weight).
7608 * In other words, if init_task_group has 10 tasks of weight
7609 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7610 * then A0's share of the cpu resource is:
7612 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7614 * We achieve this by letting init_task_group's tasks sit
7615 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7617 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7618 #elif defined CONFIG_USER_SCHED
7619 root_task_group.shares = NICE_0_LOAD;
7620 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7622 * In case of task-groups formed thr' the user id of tasks,
7623 * init_task_group represents tasks belonging to root user.
7624 * Hence it forms a sibling of all subsequent groups formed.
7625 * In this case, init_task_group gets only a fraction of overall
7626 * system cpu resource, based on the weight assigned to root
7627 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7628 * by letting tasks of init_task_group sit in a separate cfs_rq
7629 * (init_cfs_rq) and having one entity represent this group of
7630 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7632 init_tg_cfs_entry(&init_task_group,
7633 &per_cpu(init_cfs_rq, i),
7634 &per_cpu(init_sched_entity, i), i, 1,
7635 root_task_group.se[i]);
7638 #endif /* CONFIG_FAIR_GROUP_SCHED */
7640 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7641 #ifdef CONFIG_RT_GROUP_SCHED
7642 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7643 #ifdef CONFIG_CGROUP_SCHED
7644 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7645 #elif defined CONFIG_USER_SCHED
7646 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7647 init_tg_rt_entry(&init_task_group,
7648 &per_cpu(init_rt_rq, i),
7649 &per_cpu(init_sched_rt_entity, i), i, 1,
7650 root_task_group.rt_se[i]);
7654 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7655 rq->cpu_load[j] = 0;
7659 rq->active_balance = 0;
7660 rq->next_balance = jiffies;
7663 rq->migration_thread = NULL;
7664 INIT_LIST_HEAD(&rq->migration_queue);
7665 rq_attach_root(rq, &def_root_domain);
7668 atomic_set(&rq->nr_iowait, 0);
7671 set_load_weight(&init_task);
7673 #ifdef CONFIG_PREEMPT_NOTIFIERS
7674 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7678 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7681 #ifdef CONFIG_RT_MUTEXES
7682 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7686 * The boot idle thread does lazy MMU switching as well:
7688 atomic_inc(&init_mm.mm_count);
7689 enter_lazy_tlb(&init_mm, current);
7692 * Make us the idle thread. Technically, schedule() should not be
7693 * called from this thread, however somewhere below it might be,
7694 * but because we are the idle thread, we just pick up running again
7695 * when this runqueue becomes "idle".
7697 init_idle(current, smp_processor_id());
7699 * During early bootup we pretend to be a normal task:
7701 current->sched_class = &fair_sched_class;
7703 scheduler_running = 1;
7706 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7707 void __might_sleep(char *file, int line)
7710 static unsigned long prev_jiffy; /* ratelimiting */
7712 if ((in_atomic() || irqs_disabled()) &&
7713 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7714 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7716 prev_jiffy = jiffies;
7717 printk(KERN_ERR "BUG: sleeping function called from invalid"
7718 " context at %s:%d\n", file, line);
7719 printk("in_atomic():%d, irqs_disabled():%d\n",
7720 in_atomic(), irqs_disabled());
7721 debug_show_held_locks(current);
7722 if (irqs_disabled())
7723 print_irqtrace_events(current);
7728 EXPORT_SYMBOL(__might_sleep);
7731 #ifdef CONFIG_MAGIC_SYSRQ
7732 static void normalize_task(struct rq *rq, struct task_struct *p)
7735 update_rq_clock(rq);
7736 on_rq = p->se.on_rq;
7738 deactivate_task(rq, p, 0);
7739 __setscheduler(rq, p, SCHED_NORMAL, 0);
7741 activate_task(rq, p, 0);
7742 resched_task(rq->curr);
7746 void normalize_rt_tasks(void)
7748 struct task_struct *g, *p;
7749 unsigned long flags;
7752 read_lock_irqsave(&tasklist_lock, flags);
7753 do_each_thread(g, p) {
7755 * Only normalize user tasks:
7760 p->se.exec_start = 0;
7761 #ifdef CONFIG_SCHEDSTATS
7762 p->se.wait_start = 0;
7763 p->se.sleep_start = 0;
7764 p->se.block_start = 0;
7766 task_rq(p)->clock = 0;
7770 * Renice negative nice level userspace
7773 if (TASK_NICE(p) < 0 && p->mm)
7774 set_user_nice(p, 0);
7778 spin_lock(&p->pi_lock);
7779 rq = __task_rq_lock(p);
7781 normalize_task(rq, p);
7783 __task_rq_unlock(rq);
7784 spin_unlock(&p->pi_lock);
7785 } while_each_thread(g, p);
7787 read_unlock_irqrestore(&tasklist_lock, flags);
7790 #endif /* CONFIG_MAGIC_SYSRQ */
7794 * These functions are only useful for the IA64 MCA handling.
7796 * They can only be called when the whole system has been
7797 * stopped - every CPU needs to be quiescent, and no scheduling
7798 * activity can take place. Using them for anything else would
7799 * be a serious bug, and as a result, they aren't even visible
7800 * under any other configuration.
7804 * curr_task - return the current task for a given cpu.
7805 * @cpu: the processor in question.
7807 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7809 struct task_struct *curr_task(int cpu)
7811 return cpu_curr(cpu);
7815 * set_curr_task - set the current task for a given cpu.
7816 * @cpu: the processor in question.
7817 * @p: the task pointer to set.
7819 * Description: This function must only be used when non-maskable interrupts
7820 * are serviced on a separate stack. It allows the architecture to switch the
7821 * notion of the current task on a cpu in a non-blocking manner. This function
7822 * must be called with all CPU's synchronized, and interrupts disabled, the
7823 * and caller must save the original value of the current task (see
7824 * curr_task() above) and restore that value before reenabling interrupts and
7825 * re-starting the system.
7827 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7829 void set_curr_task(int cpu, struct task_struct *p)
7836 #ifdef CONFIG_FAIR_GROUP_SCHED
7837 static void free_fair_sched_group(struct task_group *tg)
7841 for_each_possible_cpu(i) {
7843 kfree(tg->cfs_rq[i]);
7853 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7855 struct cfs_rq *cfs_rq;
7856 struct sched_entity *se, *parent_se;
7860 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7863 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7867 tg->shares = NICE_0_LOAD;
7869 for_each_possible_cpu(i) {
7872 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7873 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7877 se = kmalloc_node(sizeof(struct sched_entity),
7878 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7882 parent_se = parent ? parent->se[i] : NULL;
7883 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
7892 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7894 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7895 &cpu_rq(cpu)->leaf_cfs_rq_list);
7898 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7900 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7903 static inline void free_fair_sched_group(struct task_group *tg)
7908 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7913 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7917 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7922 #ifdef CONFIG_RT_GROUP_SCHED
7923 static void free_rt_sched_group(struct task_group *tg)
7927 destroy_rt_bandwidth(&tg->rt_bandwidth);
7929 for_each_possible_cpu(i) {
7931 kfree(tg->rt_rq[i]);
7933 kfree(tg->rt_se[i]);
7941 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7943 struct rt_rq *rt_rq;
7944 struct sched_rt_entity *rt_se, *parent_se;
7948 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7951 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7955 init_rt_bandwidth(&tg->rt_bandwidth,
7956 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7958 for_each_possible_cpu(i) {
7961 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7962 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7966 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7967 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7971 parent_se = parent ? parent->rt_se[i] : NULL;
7972 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
7981 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7983 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7984 &cpu_rq(cpu)->leaf_rt_rq_list);
7987 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7989 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7992 static inline void free_rt_sched_group(struct task_group *tg)
7997 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8002 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8006 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8011 #ifdef CONFIG_GROUP_SCHED
8012 static void free_sched_group(struct task_group *tg)
8014 free_fair_sched_group(tg);
8015 free_rt_sched_group(tg);
8019 /* allocate runqueue etc for a new task group */
8020 struct task_group *sched_create_group(struct task_group *parent)
8022 struct task_group *tg;
8023 unsigned long flags;
8026 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8028 return ERR_PTR(-ENOMEM);
8030 if (!alloc_fair_sched_group(tg, parent))
8033 if (!alloc_rt_sched_group(tg, parent))
8036 spin_lock_irqsave(&task_group_lock, flags);
8037 for_each_possible_cpu(i) {
8038 register_fair_sched_group(tg, i);
8039 register_rt_sched_group(tg, i);
8041 list_add_rcu(&tg->list, &task_groups);
8042 spin_unlock_irqrestore(&task_group_lock, flags);
8047 free_sched_group(tg);
8048 return ERR_PTR(-ENOMEM);
8051 /* rcu callback to free various structures associated with a task group */
8052 static void free_sched_group_rcu(struct rcu_head *rhp)
8054 /* now it should be safe to free those cfs_rqs */
8055 free_sched_group(container_of(rhp, struct task_group, rcu));
8058 /* Destroy runqueue etc associated with a task group */
8059 void sched_destroy_group(struct task_group *tg)
8061 unsigned long flags;
8064 spin_lock_irqsave(&task_group_lock, flags);
8065 for_each_possible_cpu(i) {
8066 unregister_fair_sched_group(tg, i);
8067 unregister_rt_sched_group(tg, i);
8069 list_del_rcu(&tg->list);
8070 spin_unlock_irqrestore(&task_group_lock, flags);
8072 /* wait for possible concurrent references to cfs_rqs complete */
8073 call_rcu(&tg->rcu, free_sched_group_rcu);
8076 /* change task's runqueue when it moves between groups.
8077 * The caller of this function should have put the task in its new group
8078 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8079 * reflect its new group.
8081 void sched_move_task(struct task_struct *tsk)
8084 unsigned long flags;
8087 rq = task_rq_lock(tsk, &flags);
8089 update_rq_clock(rq);
8091 running = task_current(rq, tsk);
8092 on_rq = tsk->se.on_rq;
8095 dequeue_task(rq, tsk, 0);
8096 if (unlikely(running))
8097 tsk->sched_class->put_prev_task(rq, tsk);
8099 set_task_rq(tsk, task_cpu(tsk));
8101 #ifdef CONFIG_FAIR_GROUP_SCHED
8102 if (tsk->sched_class->moved_group)
8103 tsk->sched_class->moved_group(tsk);
8106 if (unlikely(running))
8107 tsk->sched_class->set_curr_task(rq);
8109 enqueue_task(rq, tsk, 0);
8111 task_rq_unlock(rq, &flags);
8115 #ifdef CONFIG_FAIR_GROUP_SCHED
8116 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8118 struct cfs_rq *cfs_rq = se->cfs_rq;
8119 struct rq *rq = cfs_rq->rq;
8122 spin_lock_irq(&rq->lock);
8126 dequeue_entity(cfs_rq, se, 0);
8128 se->load.weight = shares;
8129 se->load.inv_weight = div64_64((1ULL<<32), shares);
8132 enqueue_entity(cfs_rq, se, 0);
8134 spin_unlock_irq(&rq->lock);
8137 static DEFINE_MUTEX(shares_mutex);
8139 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8142 unsigned long flags;
8145 * We can't change the weight of the root cgroup.
8151 * A weight of 0 or 1 can cause arithmetics problems.
8152 * (The default weight is 1024 - so there's no practical
8153 * limitation from this.)
8158 mutex_lock(&shares_mutex);
8159 if (tg->shares == shares)
8162 spin_lock_irqsave(&task_group_lock, flags);
8163 for_each_possible_cpu(i)
8164 unregister_fair_sched_group(tg, i);
8165 spin_unlock_irqrestore(&task_group_lock, flags);
8167 /* wait for any ongoing reference to this group to finish */
8168 synchronize_sched();
8171 * Now we are free to modify the group's share on each cpu
8172 * w/o tripping rebalance_share or load_balance_fair.
8174 tg->shares = shares;
8175 for_each_possible_cpu(i)
8176 set_se_shares(tg->se[i], shares);
8179 * Enable load balance activity on this group, by inserting it back on
8180 * each cpu's rq->leaf_cfs_rq_list.
8182 spin_lock_irqsave(&task_group_lock, flags);
8183 for_each_possible_cpu(i)
8184 register_fair_sched_group(tg, i);
8185 spin_unlock_irqrestore(&task_group_lock, flags);
8187 mutex_unlock(&shares_mutex);
8191 unsigned long sched_group_shares(struct task_group *tg)
8197 #ifdef CONFIG_RT_GROUP_SCHED
8199 * Ensure that the real time constraints are schedulable.
8201 static DEFINE_MUTEX(rt_constraints_mutex);
8203 static unsigned long to_ratio(u64 period, u64 runtime)
8205 if (runtime == RUNTIME_INF)
8208 return div64_64(runtime << 16, period);
8211 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8213 struct task_group *tgi;
8214 unsigned long total = 0;
8215 unsigned long global_ratio =
8216 to_ratio(global_rt_period(), global_rt_runtime());
8219 list_for_each_entry_rcu(tgi, &task_groups, list) {
8223 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8224 tgi->rt_bandwidth.rt_runtime);
8228 return total + to_ratio(period, runtime) < global_ratio;
8231 /* Must be called with tasklist_lock held */
8232 static inline int tg_has_rt_tasks(struct task_group *tg)
8234 struct task_struct *g, *p;
8235 do_each_thread(g, p) {
8236 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8238 } while_each_thread(g, p);
8242 static int tg_set_bandwidth(struct task_group *tg,
8243 u64 rt_period, u64 rt_runtime)
8247 mutex_lock(&rt_constraints_mutex);
8248 read_lock(&tasklist_lock);
8249 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8253 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8258 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8259 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8260 tg->rt_bandwidth.rt_runtime = rt_runtime;
8262 for_each_possible_cpu(i) {
8263 struct rt_rq *rt_rq = tg->rt_rq[i];
8265 spin_lock(&rt_rq->rt_runtime_lock);
8266 rt_rq->rt_runtime = rt_runtime;
8267 spin_unlock(&rt_rq->rt_runtime_lock);
8269 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8271 read_unlock(&tasklist_lock);
8272 mutex_unlock(&rt_constraints_mutex);
8277 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8279 u64 rt_runtime, rt_period;
8281 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8282 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8283 if (rt_runtime_us < 0)
8284 rt_runtime = RUNTIME_INF;
8286 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8289 long sched_group_rt_runtime(struct task_group *tg)
8293 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8296 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8297 do_div(rt_runtime_us, NSEC_PER_USEC);
8298 return rt_runtime_us;
8301 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8303 u64 rt_runtime, rt_period;
8305 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8306 rt_runtime = tg->rt_bandwidth.rt_runtime;
8308 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8311 long sched_group_rt_period(struct task_group *tg)
8315 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8316 do_div(rt_period_us, NSEC_PER_USEC);
8317 return rt_period_us;
8320 static int sched_rt_global_constraints(void)
8324 mutex_lock(&rt_constraints_mutex);
8325 if (!__rt_schedulable(NULL, 1, 0))
8327 mutex_unlock(&rt_constraints_mutex);
8332 static int sched_rt_global_constraints(void)
8334 unsigned long flags;
8337 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8338 for_each_possible_cpu(i) {
8339 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8341 spin_lock(&rt_rq->rt_runtime_lock);
8342 rt_rq->rt_runtime = global_rt_runtime();
8343 spin_unlock(&rt_rq->rt_runtime_lock);
8345 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8351 int sched_rt_handler(struct ctl_table *table, int write,
8352 struct file *filp, void __user *buffer, size_t *lenp,
8356 int old_period, old_runtime;
8357 static DEFINE_MUTEX(mutex);
8360 old_period = sysctl_sched_rt_period;
8361 old_runtime = sysctl_sched_rt_runtime;
8363 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8365 if (!ret && write) {
8366 ret = sched_rt_global_constraints();
8368 sysctl_sched_rt_period = old_period;
8369 sysctl_sched_rt_runtime = old_runtime;
8371 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8372 def_rt_bandwidth.rt_period =
8373 ns_to_ktime(global_rt_period());
8376 mutex_unlock(&mutex);
8381 #ifdef CONFIG_CGROUP_SCHED
8383 /* return corresponding task_group object of a cgroup */
8384 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8386 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8387 struct task_group, css);
8390 static struct cgroup_subsys_state *
8391 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8393 struct task_group *tg, *parent;
8395 if (!cgrp->parent) {
8396 /* This is early initialization for the top cgroup */
8397 init_task_group.css.cgroup = cgrp;
8398 return &init_task_group.css;
8401 parent = cgroup_tg(cgrp->parent);
8402 tg = sched_create_group(parent);
8404 return ERR_PTR(-ENOMEM);
8406 /* Bind the cgroup to task_group object we just created */
8407 tg->css.cgroup = cgrp;
8413 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8415 struct task_group *tg = cgroup_tg(cgrp);
8417 sched_destroy_group(tg);
8421 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8422 struct task_struct *tsk)
8424 #ifdef CONFIG_RT_GROUP_SCHED
8425 /* Don't accept realtime tasks when there is no way for them to run */
8426 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8429 /* We don't support RT-tasks being in separate groups */
8430 if (tsk->sched_class != &fair_sched_class)
8438 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8439 struct cgroup *old_cont, struct task_struct *tsk)
8441 sched_move_task(tsk);
8444 #ifdef CONFIG_FAIR_GROUP_SCHED
8445 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8448 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8451 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8453 struct task_group *tg = cgroup_tg(cgrp);
8455 return (u64) tg->shares;
8459 #ifdef CONFIG_RT_GROUP_SCHED
8460 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8462 const char __user *userbuf,
8463 size_t nbytes, loff_t *unused_ppos)
8472 if (nbytes >= sizeof(buffer))
8474 if (copy_from_user(buffer, userbuf, nbytes))
8477 buffer[nbytes] = 0; /* nul-terminate */
8479 /* strip newline if necessary */
8480 if (nbytes && (buffer[nbytes-1] == '\n'))
8481 buffer[nbytes-1] = 0;
8482 val = simple_strtoll(buffer, &end, 0);
8486 /* Pass to subsystem */
8487 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8493 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8495 char __user *buf, size_t nbytes,
8499 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8500 int len = sprintf(tmp, "%ld\n", val);
8502 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8505 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8508 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8511 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8513 return sched_group_rt_period(cgroup_tg(cgrp));
8517 static struct cftype cpu_files[] = {
8518 #ifdef CONFIG_FAIR_GROUP_SCHED
8521 .read_uint = cpu_shares_read_uint,
8522 .write_uint = cpu_shares_write_uint,
8525 #ifdef CONFIG_RT_GROUP_SCHED
8527 .name = "rt_runtime_us",
8528 .read = cpu_rt_runtime_read,
8529 .write = cpu_rt_runtime_write,
8532 .name = "rt_period_us",
8533 .read_uint = cpu_rt_period_read_uint,
8534 .write_uint = cpu_rt_period_write_uint,
8539 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8541 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8544 struct cgroup_subsys cpu_cgroup_subsys = {
8546 .create = cpu_cgroup_create,
8547 .destroy = cpu_cgroup_destroy,
8548 .can_attach = cpu_cgroup_can_attach,
8549 .attach = cpu_cgroup_attach,
8550 .populate = cpu_cgroup_populate,
8551 .subsys_id = cpu_cgroup_subsys_id,
8555 #endif /* CONFIG_CGROUP_SCHED */
8557 #ifdef CONFIG_CGROUP_CPUACCT
8560 * CPU accounting code for task groups.
8562 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8563 * (balbir@in.ibm.com).
8566 /* track cpu usage of a group of tasks */
8568 struct cgroup_subsys_state css;
8569 /* cpuusage holds pointer to a u64-type object on every cpu */
8573 struct cgroup_subsys cpuacct_subsys;
8575 /* return cpu accounting group corresponding to this container */
8576 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8578 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8579 struct cpuacct, css);
8582 /* return cpu accounting group to which this task belongs */
8583 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8585 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8586 struct cpuacct, css);
8589 /* create a new cpu accounting group */
8590 static struct cgroup_subsys_state *cpuacct_create(
8591 struct cgroup_subsys *ss, struct cgroup *cgrp)
8593 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8596 return ERR_PTR(-ENOMEM);
8598 ca->cpuusage = alloc_percpu(u64);
8599 if (!ca->cpuusage) {
8601 return ERR_PTR(-ENOMEM);
8607 /* destroy an existing cpu accounting group */
8609 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8611 struct cpuacct *ca = cgroup_ca(cgrp);
8613 free_percpu(ca->cpuusage);
8617 /* return total cpu usage (in nanoseconds) of a group */
8618 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8620 struct cpuacct *ca = cgroup_ca(cgrp);
8621 u64 totalcpuusage = 0;
8624 for_each_possible_cpu(i) {
8625 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8628 * Take rq->lock to make 64-bit addition safe on 32-bit
8631 spin_lock_irq(&cpu_rq(i)->lock);
8632 totalcpuusage += *cpuusage;
8633 spin_unlock_irq(&cpu_rq(i)->lock);
8636 return totalcpuusage;
8639 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8642 struct cpuacct *ca = cgroup_ca(cgrp);
8651 for_each_possible_cpu(i) {
8652 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8654 spin_lock_irq(&cpu_rq(i)->lock);
8656 spin_unlock_irq(&cpu_rq(i)->lock);
8662 static struct cftype files[] = {
8665 .read_uint = cpuusage_read,
8666 .write_uint = cpuusage_write,
8670 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8672 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8676 * charge this task's execution time to its accounting group.
8678 * called with rq->lock held.
8680 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8684 if (!cpuacct_subsys.active)
8689 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8691 *cpuusage += cputime;
8695 struct cgroup_subsys cpuacct_subsys = {
8697 .create = cpuacct_create,
8698 .destroy = cpuacct_destroy,
8699 .populate = cpuacct_populate,
8700 .subsys_id = cpuacct_subsys_id,
8702 #endif /* CONFIG_CGROUP_CPUACCT */