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
277 #ifdef CONFIG_FAIR_GROUP_SCHED
278 /* Default task group's sched entity on each cpu */
279 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
280 /* Default task group's cfs_rq on each cpu */
281 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
286 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
290 /* task_group_lock serializes add/remove of task groups and also changes to
291 * a task group's cpu shares.
293 static DEFINE_SPINLOCK(task_group_lock);
295 /* doms_cur_mutex serializes access to doms_cur[] array */
296 static DEFINE_MUTEX(doms_cur_mutex);
298 #ifdef CONFIG_FAIR_GROUP_SCHED
299 #ifdef CONFIG_USER_SCHED
300 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
302 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
305 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
308 /* Default task group.
309 * Every task in system belong to this group at bootup.
311 struct task_group init_task_group;
313 /* return group to which a task belongs */
314 static inline struct task_group *task_group(struct task_struct *p)
316 struct task_group *tg;
318 #ifdef CONFIG_USER_SCHED
320 #elif defined(CONFIG_CGROUP_SCHED)
321 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
322 struct task_group, css);
324 tg = &init_task_group;
329 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
330 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
334 p->se.parent = task_group(p)->se[cpu];
337 #ifdef CONFIG_RT_GROUP_SCHED
338 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
339 p->rt.parent = task_group(p)->rt_se[cpu];
343 static inline void lock_doms_cur(void)
345 mutex_lock(&doms_cur_mutex);
348 static inline void unlock_doms_cur(void)
350 mutex_unlock(&doms_cur_mutex);
355 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
356 static inline void lock_doms_cur(void) { }
357 static inline void unlock_doms_cur(void) { }
359 #endif /* CONFIG_GROUP_SCHED */
361 /* CFS-related fields in a runqueue */
363 struct load_weight load;
364 unsigned long nr_running;
369 struct rb_root tasks_timeline;
370 struct rb_node *rb_leftmost;
371 struct rb_node *rb_load_balance_curr;
372 /* 'curr' points to currently running entity on this cfs_rq.
373 * It is set to NULL otherwise (i.e when none are currently running).
375 struct sched_entity *curr, *next;
377 unsigned long nr_spread_over;
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
385 * (like users, containers etc.)
387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
388 * list is used during load balance.
390 struct list_head leaf_cfs_rq_list;
391 struct task_group *tg; /* group that "owns" this runqueue */
395 /* Real-Time classes' related field in a runqueue: */
397 struct rt_prio_array active;
398 unsigned long rt_nr_running;
399 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
400 int highest_prio; /* highest queued rt task prio */
403 unsigned long rt_nr_migratory;
409 /* Nests inside the rq lock: */
410 spinlock_t rt_runtime_lock;
412 #ifdef CONFIG_RT_GROUP_SCHED
413 unsigned long rt_nr_boosted;
416 struct list_head leaf_rt_rq_list;
417 struct task_group *tg;
418 struct sched_rt_entity *rt_se;
425 * We add the notion of a root-domain which will be used to define per-domain
426 * variables. Each exclusive cpuset essentially defines an island domain by
427 * fully partitioning the member cpus from any other cpuset. Whenever a new
428 * exclusive cpuset is created, we also create and attach a new root-domain
438 * The "RT overload" flag: it gets set if a CPU has more than
439 * one runnable RT task.
446 * By default the system creates a single root-domain with all cpus as
447 * members (mimicking the global state we have today).
449 static struct root_domain def_root_domain;
454 * This is the main, per-CPU runqueue data structure.
456 * Locking rule: those places that want to lock multiple runqueues
457 * (such as the load balancing or the thread migration code), lock
458 * acquire operations must be ordered by ascending &runqueue.
465 * nr_running and cpu_load should be in the same cacheline because
466 * remote CPUs use both these fields when doing load calculation.
468 unsigned long nr_running;
469 #define CPU_LOAD_IDX_MAX 5
470 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
471 unsigned char idle_at_tick;
473 unsigned long last_tick_seen;
474 unsigned char in_nohz_recently;
476 /* capture load from *all* tasks on this cpu: */
477 struct load_weight load;
478 unsigned long nr_load_updates;
484 #ifdef CONFIG_FAIR_GROUP_SCHED
485 /* list of leaf cfs_rq on this cpu: */
486 struct list_head leaf_cfs_rq_list;
488 #ifdef CONFIG_RT_GROUP_SCHED
489 struct list_head leaf_rt_rq_list;
493 * This is part of a global counter where only the total sum
494 * over all CPUs matters. A task can increase this counter on
495 * one CPU and if it got migrated afterwards it may decrease
496 * it on another CPU. Always updated under the runqueue lock:
498 unsigned long nr_uninterruptible;
500 struct task_struct *curr, *idle;
501 unsigned long next_balance;
502 struct mm_struct *prev_mm;
504 u64 clock, prev_clock_raw;
507 unsigned int clock_warps, clock_overflows, clock_underflows;
509 unsigned int clock_deep_idle_events;
515 struct root_domain *rd;
516 struct sched_domain *sd;
518 /* For active balancing */
521 /* cpu of this runqueue: */
524 struct task_struct *migration_thread;
525 struct list_head migration_queue;
528 #ifdef CONFIG_SCHED_HRTICK
529 unsigned long hrtick_flags;
530 ktime_t hrtick_expire;
531 struct hrtimer hrtick_timer;
534 #ifdef CONFIG_SCHEDSTATS
536 struct sched_info rq_sched_info;
538 /* sys_sched_yield() stats */
539 unsigned int yld_exp_empty;
540 unsigned int yld_act_empty;
541 unsigned int yld_both_empty;
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
556 struct lock_class_key rq_lock_key;
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
563 rq->curr->sched_class->check_preempt_curr(rq, p);
566 static inline int cpu_of(struct rq *rq)
576 static inline bool nohz_on(int cpu)
578 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
581 static inline u64 max_skipped_ticks(struct rq *rq)
583 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
586 static inline void update_last_tick_seen(struct rq *rq)
588 rq->last_tick_seen = jiffies;
591 static inline u64 max_skipped_ticks(struct rq *rq)
596 static inline void update_last_tick_seen(struct rq *rq)
602 * Update the per-runqueue clock, as finegrained as the platform can give
603 * us, but without assuming monotonicity, etc.:
605 static void __update_rq_clock(struct rq *rq)
607 u64 prev_raw = rq->prev_clock_raw;
608 u64 now = sched_clock();
609 s64 delta = now - prev_raw;
610 u64 clock = rq->clock;
612 #ifdef CONFIG_SCHED_DEBUG
613 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
616 * Protect against sched_clock() occasionally going backwards:
618 if (unlikely(delta < 0)) {
623 * Catch too large forward jumps too:
625 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
626 u64 max_time = rq->tick_timestamp + max_jump;
628 if (unlikely(clock + delta > max_time)) {
629 if (clock < max_time)
633 rq->clock_overflows++;
635 if (unlikely(delta > rq->clock_max_delta))
636 rq->clock_max_delta = delta;
641 rq->prev_clock_raw = now;
645 static void update_rq_clock(struct rq *rq)
647 if (likely(smp_processor_id() == cpu_of(rq)))
648 __update_rq_clock(rq);
652 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
653 * See detach_destroy_domains: synchronize_sched for details.
655 * The domain tree of any CPU may only be accessed from within
656 * preempt-disabled sections.
658 #define for_each_domain(cpu, __sd) \
659 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
661 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
662 #define this_rq() (&__get_cpu_var(runqueues))
663 #define task_rq(p) cpu_rq(task_cpu(p))
664 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
667 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
669 #ifdef CONFIG_SCHED_DEBUG
670 # define const_debug __read_mostly
672 # define const_debug static const
676 * Debugging: various feature bits
679 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
680 SCHED_FEAT_WAKEUP_PREEMPT = 2,
681 SCHED_FEAT_START_DEBIT = 4,
682 SCHED_FEAT_AFFINE_WAKEUPS = 8,
683 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
684 SCHED_FEAT_SYNC_WAKEUPS = 32,
685 SCHED_FEAT_HRTICK = 64,
686 SCHED_FEAT_DOUBLE_TICK = 128,
687 SCHED_FEAT_NORMALIZED_SLEEPER = 256,
690 const_debug unsigned int sysctl_sched_features =
691 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
692 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
693 SCHED_FEAT_START_DEBIT * 1 |
694 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
695 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
696 SCHED_FEAT_SYNC_WAKEUPS * 1 |
697 SCHED_FEAT_HRTICK * 1 |
698 SCHED_FEAT_DOUBLE_TICK * 0 |
699 SCHED_FEAT_NORMALIZED_SLEEPER * 1;
701 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
704 * Number of tasks to iterate in a single balance run.
705 * Limited because this is done with IRQs disabled.
707 const_debug unsigned int sysctl_sched_nr_migrate = 32;
710 * period over which we measure -rt task cpu usage in us.
713 unsigned int sysctl_sched_rt_period = 1000000;
715 static __read_mostly int scheduler_running;
718 * part of the period that we allow rt tasks to run in us.
721 int sysctl_sched_rt_runtime = 950000;
723 static inline u64 global_rt_period(void)
725 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
728 static inline u64 global_rt_runtime(void)
730 if (sysctl_sched_rt_period < 0)
733 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
736 static const unsigned long long time_sync_thresh = 100000;
738 static DEFINE_PER_CPU(unsigned long long, time_offset);
739 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
742 * Global lock which we take every now and then to synchronize
743 * the CPUs time. This method is not warp-safe, but it's good
744 * enough to synchronize slowly diverging time sources and thus
745 * it's good enough for tracing:
747 static DEFINE_SPINLOCK(time_sync_lock);
748 static unsigned long long prev_global_time;
750 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
754 spin_lock_irqsave(&time_sync_lock, flags);
756 if (time < prev_global_time) {
757 per_cpu(time_offset, cpu) += prev_global_time - time;
758 time = prev_global_time;
760 prev_global_time = time;
763 spin_unlock_irqrestore(&time_sync_lock, flags);
768 static unsigned long long __cpu_clock(int cpu)
770 unsigned long long now;
775 * Only call sched_clock() if the scheduler has already been
776 * initialized (some code might call cpu_clock() very early):
778 if (unlikely(!scheduler_running))
781 local_irq_save(flags);
785 local_irq_restore(flags);
791 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
792 * clock constructed from sched_clock():
794 unsigned long long cpu_clock(int cpu)
796 unsigned long long prev_cpu_time, time, delta_time;
798 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
799 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
800 delta_time = time-prev_cpu_time;
802 if (unlikely(delta_time > time_sync_thresh))
803 time = __sync_cpu_clock(time, cpu);
807 EXPORT_SYMBOL_GPL(cpu_clock);
809 #ifndef prepare_arch_switch
810 # define prepare_arch_switch(next) do { } while (0)
812 #ifndef finish_arch_switch
813 # define finish_arch_switch(prev) do { } while (0)
816 static inline int task_current(struct rq *rq, struct task_struct *p)
818 return rq->curr == p;
821 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
822 static inline int task_running(struct rq *rq, struct task_struct *p)
824 return task_current(rq, p);
827 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
831 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
833 #ifdef CONFIG_DEBUG_SPINLOCK
834 /* this is a valid case when another task releases the spinlock */
835 rq->lock.owner = current;
838 * If we are tracking spinlock dependencies then we have to
839 * fix up the runqueue lock - which gets 'carried over' from
842 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
844 spin_unlock_irq(&rq->lock);
847 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
848 static inline int task_running(struct rq *rq, struct task_struct *p)
853 return task_current(rq, p);
857 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
861 * We can optimise this out completely for !SMP, because the
862 * SMP rebalancing from interrupt is the only thing that cares
867 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
868 spin_unlock_irq(&rq->lock);
870 spin_unlock(&rq->lock);
874 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
878 * After ->oncpu is cleared, the task can be moved to a different CPU.
879 * We must ensure this doesn't happen until the switch is completely
885 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
889 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
892 * __task_rq_lock - lock the runqueue a given task resides on.
893 * Must be called interrupts disabled.
895 static inline struct rq *__task_rq_lock(struct task_struct *p)
899 struct rq *rq = task_rq(p);
900 spin_lock(&rq->lock);
901 if (likely(rq == task_rq(p)))
903 spin_unlock(&rq->lock);
908 * task_rq_lock - lock the runqueue a given task resides on and disable
909 * interrupts. Note the ordering: we can safely lookup the task_rq without
910 * explicitly disabling preemption.
912 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
918 local_irq_save(*flags);
920 spin_lock(&rq->lock);
921 if (likely(rq == task_rq(p)))
923 spin_unlock_irqrestore(&rq->lock, *flags);
927 static void __task_rq_unlock(struct rq *rq)
930 spin_unlock(&rq->lock);
933 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
936 spin_unlock_irqrestore(&rq->lock, *flags);
940 * this_rq_lock - lock this runqueue and disable interrupts.
942 static struct rq *this_rq_lock(void)
949 spin_lock(&rq->lock);
955 * We are going deep-idle (irqs are disabled):
957 void sched_clock_idle_sleep_event(void)
959 struct rq *rq = cpu_rq(smp_processor_id());
961 spin_lock(&rq->lock);
962 __update_rq_clock(rq);
963 spin_unlock(&rq->lock);
964 rq->clock_deep_idle_events++;
966 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
969 * We just idled delta nanoseconds (called with irqs disabled):
971 void sched_clock_idle_wakeup_event(u64 delta_ns)
973 struct rq *rq = cpu_rq(smp_processor_id());
974 u64 now = sched_clock();
976 rq->idle_clock += delta_ns;
978 * Override the previous timestamp and ignore all
979 * sched_clock() deltas that occured while we idled,
980 * and use the PM-provided delta_ns to advance the
983 spin_lock(&rq->lock);
984 rq->prev_clock_raw = now;
985 rq->clock += delta_ns;
986 spin_unlock(&rq->lock);
987 touch_softlockup_watchdog();
989 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
991 static void __resched_task(struct task_struct *p, int tif_bit);
993 static inline void resched_task(struct task_struct *p)
995 __resched_task(p, TIF_NEED_RESCHED);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1009 static inline void resched_hrt(struct task_struct *p)
1011 __resched_task(p, TIF_HRTICK_RESCHED);
1014 static inline void resched_rq(struct rq *rq)
1016 unsigned long flags;
1018 spin_lock_irqsave(&rq->lock, flags);
1019 resched_task(rq->curr);
1020 spin_unlock_irqrestore(&rq->lock, flags);
1024 HRTICK_SET, /* re-programm hrtick_timer */
1025 HRTICK_RESET, /* not a new slice */
1030 * - enabled by features
1031 * - hrtimer is actually high res
1033 static inline int hrtick_enabled(struct rq *rq)
1035 if (!sched_feat(HRTICK))
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1041 * Called to set the hrtick timer state.
1043 * called with rq->lock held and irqs disabled
1045 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1047 assert_spin_locked(&rq->lock);
1050 * preempt at: now + delay
1053 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1055 * indicate we need to program the timer
1057 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1059 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1062 * New slices are called from the schedule path and don't need a
1063 * forced reschedule.
1066 resched_hrt(rq->curr);
1069 static void hrtick_clear(struct rq *rq)
1071 if (hrtimer_active(&rq->hrtick_timer))
1072 hrtimer_cancel(&rq->hrtick_timer);
1076 * Update the timer from the possible pending state.
1078 static void hrtick_set(struct rq *rq)
1082 unsigned long flags;
1084 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1086 spin_lock_irqsave(&rq->lock, flags);
1087 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1088 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1089 time = rq->hrtick_expire;
1090 clear_thread_flag(TIF_HRTICK_RESCHED);
1091 spin_unlock_irqrestore(&rq->lock, flags);
1094 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1095 if (reset && !hrtimer_active(&rq->hrtick_timer))
1102 * High-resolution timer tick.
1103 * Runs from hardirq context with interrupts disabled.
1105 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1107 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1109 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1111 spin_lock(&rq->lock);
1112 __update_rq_clock(rq);
1113 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1114 spin_unlock(&rq->lock);
1116 return HRTIMER_NORESTART;
1119 static inline void init_rq_hrtick(struct rq *rq)
1121 rq->hrtick_flags = 0;
1122 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1123 rq->hrtick_timer.function = hrtick;
1124 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1127 void hrtick_resched(void)
1130 unsigned long flags;
1132 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1135 local_irq_save(flags);
1136 rq = cpu_rq(smp_processor_id());
1138 local_irq_restore(flags);
1141 static inline void hrtick_clear(struct rq *rq)
1145 static inline void hrtick_set(struct rq *rq)
1149 static inline void init_rq_hrtick(struct rq *rq)
1153 void hrtick_resched(void)
1159 * resched_task - mark a task 'to be rescheduled now'.
1161 * On UP this means the setting of the need_resched flag, on SMP it
1162 * might also involve a cross-CPU call to trigger the scheduler on
1167 #ifndef tsk_is_polling
1168 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1171 static void __resched_task(struct task_struct *p, int tif_bit)
1175 assert_spin_locked(&task_rq(p)->lock);
1177 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1180 set_tsk_thread_flag(p, tif_bit);
1183 if (cpu == smp_processor_id())
1186 /* NEED_RESCHED must be visible before we test polling */
1188 if (!tsk_is_polling(p))
1189 smp_send_reschedule(cpu);
1192 static void resched_cpu(int cpu)
1194 struct rq *rq = cpu_rq(cpu);
1195 unsigned long flags;
1197 if (!spin_trylock_irqsave(&rq->lock, flags))
1199 resched_task(cpu_curr(cpu));
1200 spin_unlock_irqrestore(&rq->lock, flags);
1205 * When add_timer_on() enqueues a timer into the timer wheel of an
1206 * idle CPU then this timer might expire before the next timer event
1207 * which is scheduled to wake up that CPU. In case of a completely
1208 * idle system the next event might even be infinite time into the
1209 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1210 * leaves the inner idle loop so the newly added timer is taken into
1211 * account when the CPU goes back to idle and evaluates the timer
1212 * wheel for the next timer event.
1214 void wake_up_idle_cpu(int cpu)
1216 struct rq *rq = cpu_rq(cpu);
1218 if (cpu == smp_processor_id())
1222 * This is safe, as this function is called with the timer
1223 * wheel base lock of (cpu) held. When the CPU is on the way
1224 * to idle and has not yet set rq->curr to idle then it will
1225 * be serialized on the timer wheel base lock and take the new
1226 * timer into account automatically.
1228 if (rq->curr != rq->idle)
1232 * We can set TIF_RESCHED on the idle task of the other CPU
1233 * lockless. The worst case is that the other CPU runs the
1234 * idle task through an additional NOOP schedule()
1236 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1238 /* NEED_RESCHED must be visible before we test polling */
1240 if (!tsk_is_polling(rq->idle))
1241 smp_send_reschedule(cpu);
1246 static void __resched_task(struct task_struct *p, int tif_bit)
1248 assert_spin_locked(&task_rq(p)->lock);
1249 set_tsk_thread_flag(p, tif_bit);
1253 #if BITS_PER_LONG == 32
1254 # define WMULT_CONST (~0UL)
1256 # define WMULT_CONST (1UL << 32)
1259 #define WMULT_SHIFT 32
1262 * Shift right and round:
1264 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1266 static unsigned long
1267 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1268 struct load_weight *lw)
1272 if (unlikely(!lw->inv_weight))
1273 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1275 tmp = (u64)delta_exec * weight;
1277 * Check whether we'd overflow the 64-bit multiplication:
1279 if (unlikely(tmp > WMULT_CONST))
1280 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1283 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1285 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1288 static inline unsigned long
1289 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1291 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1294 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1300 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1307 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1308 * of tasks with abnormal "nice" values across CPUs the contribution that
1309 * each task makes to its run queue's load is weighted according to its
1310 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1311 * scaled version of the new time slice allocation that they receive on time
1315 #define WEIGHT_IDLEPRIO 2
1316 #define WMULT_IDLEPRIO (1 << 31)
1319 * Nice levels are multiplicative, with a gentle 10% change for every
1320 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1321 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1322 * that remained on nice 0.
1324 * The "10% effect" is relative and cumulative: from _any_ nice level,
1325 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1326 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1327 * If a task goes up by ~10% and another task goes down by ~10% then
1328 * the relative distance between them is ~25%.)
1330 static const int prio_to_weight[40] = {
1331 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1332 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1333 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1334 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1335 /* 0 */ 1024, 820, 655, 526, 423,
1336 /* 5 */ 335, 272, 215, 172, 137,
1337 /* 10 */ 110, 87, 70, 56, 45,
1338 /* 15 */ 36, 29, 23, 18, 15,
1342 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1344 * In cases where the weight does not change often, we can use the
1345 * precalculated inverse to speed up arithmetics by turning divisions
1346 * into multiplications:
1348 static const u32 prio_to_wmult[40] = {
1349 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1350 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1351 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1352 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1353 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1354 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1355 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1356 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1359 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1362 * runqueue iterator, to support SMP load-balancing between different
1363 * scheduling classes, without having to expose their internal data
1364 * structures to the load-balancing proper:
1366 struct rq_iterator {
1368 struct task_struct *(*start)(void *);
1369 struct task_struct *(*next)(void *);
1373 static unsigned long
1374 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1375 unsigned long max_load_move, struct sched_domain *sd,
1376 enum cpu_idle_type idle, int *all_pinned,
1377 int *this_best_prio, struct rq_iterator *iterator);
1380 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1381 struct sched_domain *sd, enum cpu_idle_type idle,
1382 struct rq_iterator *iterator);
1385 #ifdef CONFIG_CGROUP_CPUACCT
1386 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1388 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1392 static unsigned long source_load(int cpu, int type);
1393 static unsigned long target_load(int cpu, int type);
1394 static unsigned long cpu_avg_load_per_task(int cpu);
1395 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1396 #endif /* CONFIG_SMP */
1398 #include "sched_stats.h"
1399 #include "sched_idletask.c"
1400 #include "sched_fair.c"
1401 #include "sched_rt.c"
1402 #ifdef CONFIG_SCHED_DEBUG
1403 # include "sched_debug.c"
1406 #define sched_class_highest (&rt_sched_class)
1408 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1410 update_load_add(&rq->load, p->se.load.weight);
1413 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1415 update_load_sub(&rq->load, p->se.load.weight);
1418 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1424 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1430 static void set_load_weight(struct task_struct *p)
1432 if (task_has_rt_policy(p)) {
1433 p->se.load.weight = prio_to_weight[0] * 2;
1434 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1439 * SCHED_IDLE tasks get minimal weight:
1441 if (p->policy == SCHED_IDLE) {
1442 p->se.load.weight = WEIGHT_IDLEPRIO;
1443 p->se.load.inv_weight = WMULT_IDLEPRIO;
1447 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1448 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1451 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1453 sched_info_queued(p);
1454 p->sched_class->enqueue_task(rq, p, wakeup);
1458 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1460 p->sched_class->dequeue_task(rq, p, sleep);
1465 * __normal_prio - return the priority that is based on the static prio
1467 static inline int __normal_prio(struct task_struct *p)
1469 return p->static_prio;
1473 * Calculate the expected normal priority: i.e. priority
1474 * without taking RT-inheritance into account. Might be
1475 * boosted by interactivity modifiers. Changes upon fork,
1476 * setprio syscalls, and whenever the interactivity
1477 * estimator recalculates.
1479 static inline int normal_prio(struct task_struct *p)
1483 if (task_has_rt_policy(p))
1484 prio = MAX_RT_PRIO-1 - p->rt_priority;
1486 prio = __normal_prio(p);
1491 * Calculate the current priority, i.e. the priority
1492 * taken into account by the scheduler. This value might
1493 * be boosted by RT tasks, or might be boosted by
1494 * interactivity modifiers. Will be RT if the task got
1495 * RT-boosted. If not then it returns p->normal_prio.
1497 static int effective_prio(struct task_struct *p)
1499 p->normal_prio = normal_prio(p);
1501 * If we are RT tasks or we were boosted to RT priority,
1502 * keep the priority unchanged. Otherwise, update priority
1503 * to the normal priority:
1505 if (!rt_prio(p->prio))
1506 return p->normal_prio;
1511 * activate_task - move a task to the runqueue.
1513 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1515 if (task_contributes_to_load(p))
1516 rq->nr_uninterruptible--;
1518 enqueue_task(rq, p, wakeup);
1519 inc_nr_running(p, rq);
1523 * deactivate_task - remove a task from the runqueue.
1525 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1527 if (task_contributes_to_load(p))
1528 rq->nr_uninterruptible++;
1530 dequeue_task(rq, p, sleep);
1531 dec_nr_running(p, rq);
1535 * task_curr - is this task currently executing on a CPU?
1536 * @p: the task in question.
1538 inline int task_curr(const struct task_struct *p)
1540 return cpu_curr(task_cpu(p)) == p;
1543 /* Used instead of source_load when we know the type == 0 */
1544 unsigned long weighted_cpuload(const int cpu)
1546 return cpu_rq(cpu)->load.weight;
1549 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1551 set_task_rq(p, cpu);
1554 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1555 * successfuly executed on another CPU. We must ensure that updates of
1556 * per-task data have been completed by this moment.
1559 task_thread_info(p)->cpu = cpu;
1563 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1564 const struct sched_class *prev_class,
1565 int oldprio, int running)
1567 if (prev_class != p->sched_class) {
1568 if (prev_class->switched_from)
1569 prev_class->switched_from(rq, p, running);
1570 p->sched_class->switched_to(rq, p, running);
1572 p->sched_class->prio_changed(rq, p, oldprio, running);
1578 * Is this task likely cache-hot:
1581 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1586 * Buddy candidates are cache hot:
1588 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1591 if (p->sched_class != &fair_sched_class)
1594 if (sysctl_sched_migration_cost == -1)
1596 if (sysctl_sched_migration_cost == 0)
1599 delta = now - p->se.exec_start;
1601 return delta < (s64)sysctl_sched_migration_cost;
1605 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1607 int old_cpu = task_cpu(p);
1608 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1609 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1610 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1613 clock_offset = old_rq->clock - new_rq->clock;
1615 #ifdef CONFIG_SCHEDSTATS
1616 if (p->se.wait_start)
1617 p->se.wait_start -= clock_offset;
1618 if (p->se.sleep_start)
1619 p->se.sleep_start -= clock_offset;
1620 if (p->se.block_start)
1621 p->se.block_start -= clock_offset;
1622 if (old_cpu != new_cpu) {
1623 schedstat_inc(p, se.nr_migrations);
1624 if (task_hot(p, old_rq->clock, NULL))
1625 schedstat_inc(p, se.nr_forced2_migrations);
1628 p->se.vruntime -= old_cfsrq->min_vruntime -
1629 new_cfsrq->min_vruntime;
1631 __set_task_cpu(p, new_cpu);
1634 struct migration_req {
1635 struct list_head list;
1637 struct task_struct *task;
1640 struct completion done;
1644 * The task's runqueue lock must be held.
1645 * Returns true if you have to wait for migration thread.
1648 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1650 struct rq *rq = task_rq(p);
1653 * If the task is not on a runqueue (and not running), then
1654 * it is sufficient to simply update the task's cpu field.
1656 if (!p->se.on_rq && !task_running(rq, p)) {
1657 set_task_cpu(p, dest_cpu);
1661 init_completion(&req->done);
1663 req->dest_cpu = dest_cpu;
1664 list_add(&req->list, &rq->migration_queue);
1670 * wait_task_inactive - wait for a thread to unschedule.
1672 * The caller must ensure that the task *will* unschedule sometime soon,
1673 * else this function might spin for a *long* time. This function can't
1674 * be called with interrupts off, or it may introduce deadlock with
1675 * smp_call_function() if an IPI is sent by the same process we are
1676 * waiting to become inactive.
1678 void wait_task_inactive(struct task_struct *p)
1680 unsigned long flags;
1686 * We do the initial early heuristics without holding
1687 * any task-queue locks at all. We'll only try to get
1688 * the runqueue lock when things look like they will
1694 * If the task is actively running on another CPU
1695 * still, just relax and busy-wait without holding
1698 * NOTE! Since we don't hold any locks, it's not
1699 * even sure that "rq" stays as the right runqueue!
1700 * But we don't care, since "task_running()" will
1701 * return false if the runqueue has changed and p
1702 * is actually now running somewhere else!
1704 while (task_running(rq, p))
1708 * Ok, time to look more closely! We need the rq
1709 * lock now, to be *sure*. If we're wrong, we'll
1710 * just go back and repeat.
1712 rq = task_rq_lock(p, &flags);
1713 running = task_running(rq, p);
1714 on_rq = p->se.on_rq;
1715 task_rq_unlock(rq, &flags);
1718 * Was it really running after all now that we
1719 * checked with the proper locks actually held?
1721 * Oops. Go back and try again..
1723 if (unlikely(running)) {
1729 * It's not enough that it's not actively running,
1730 * it must be off the runqueue _entirely_, and not
1733 * So if it wa still runnable (but just not actively
1734 * running right now), it's preempted, and we should
1735 * yield - it could be a while.
1737 if (unlikely(on_rq)) {
1738 schedule_timeout_uninterruptible(1);
1743 * Ahh, all good. It wasn't running, and it wasn't
1744 * runnable, which means that it will never become
1745 * running in the future either. We're all done!
1752 * kick_process - kick a running thread to enter/exit the kernel
1753 * @p: the to-be-kicked thread
1755 * Cause a process which is running on another CPU to enter
1756 * kernel-mode, without any delay. (to get signals handled.)
1758 * NOTE: this function doesnt have to take the runqueue lock,
1759 * because all it wants to ensure is that the remote task enters
1760 * the kernel. If the IPI races and the task has been migrated
1761 * to another CPU then no harm is done and the purpose has been
1764 void kick_process(struct task_struct *p)
1770 if ((cpu != smp_processor_id()) && task_curr(p))
1771 smp_send_reschedule(cpu);
1776 * Return a low guess at the load of a migration-source cpu weighted
1777 * according to the scheduling class and "nice" value.
1779 * We want to under-estimate the load of migration sources, to
1780 * balance conservatively.
1782 static unsigned long source_load(int cpu, int type)
1784 struct rq *rq = cpu_rq(cpu);
1785 unsigned long total = weighted_cpuload(cpu);
1790 return min(rq->cpu_load[type-1], total);
1794 * Return a high guess at the load of a migration-target cpu weighted
1795 * according to the scheduling class and "nice" value.
1797 static unsigned long target_load(int cpu, int type)
1799 struct rq *rq = cpu_rq(cpu);
1800 unsigned long total = weighted_cpuload(cpu);
1805 return max(rq->cpu_load[type-1], total);
1809 * Return the average load per task on the cpu's run queue
1811 static unsigned long cpu_avg_load_per_task(int cpu)
1813 struct rq *rq = cpu_rq(cpu);
1814 unsigned long total = weighted_cpuload(cpu);
1815 unsigned long n = rq->nr_running;
1817 return n ? total / n : SCHED_LOAD_SCALE;
1821 * find_idlest_group finds and returns the least busy CPU group within the
1824 static struct sched_group *
1825 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1827 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1828 unsigned long min_load = ULONG_MAX, this_load = 0;
1829 int load_idx = sd->forkexec_idx;
1830 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1833 unsigned long load, avg_load;
1837 /* Skip over this group if it has no CPUs allowed */
1838 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1841 local_group = cpu_isset(this_cpu, group->cpumask);
1843 /* Tally up the load of all CPUs in the group */
1846 for_each_cpu_mask(i, group->cpumask) {
1847 /* Bias balancing toward cpus of our domain */
1849 load = source_load(i, load_idx);
1851 load = target_load(i, load_idx);
1856 /* Adjust by relative CPU power of the group */
1857 avg_load = sg_div_cpu_power(group,
1858 avg_load * SCHED_LOAD_SCALE);
1861 this_load = avg_load;
1863 } else if (avg_load < min_load) {
1864 min_load = avg_load;
1867 } while (group = group->next, group != sd->groups);
1869 if (!idlest || 100*this_load < imbalance*min_load)
1875 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1878 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1881 unsigned long load, min_load = ULONG_MAX;
1885 /* Traverse only the allowed CPUs */
1886 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1888 for_each_cpu_mask(i, *tmp) {
1889 load = weighted_cpuload(i);
1891 if (load < min_load || (load == min_load && i == this_cpu)) {
1901 * sched_balance_self: balance the current task (running on cpu) in domains
1902 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1905 * Balance, ie. select the least loaded group.
1907 * Returns the target CPU number, or the same CPU if no balancing is needed.
1909 * preempt must be disabled.
1911 static int sched_balance_self(int cpu, int flag)
1913 struct task_struct *t = current;
1914 struct sched_domain *tmp, *sd = NULL;
1916 for_each_domain(cpu, tmp) {
1918 * If power savings logic is enabled for a domain, stop there.
1920 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1922 if (tmp->flags & flag)
1927 cpumask_t span, tmpmask;
1928 struct sched_group *group;
1929 int new_cpu, weight;
1931 if (!(sd->flags & flag)) {
1937 group = find_idlest_group(sd, t, cpu);
1943 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1944 if (new_cpu == -1 || new_cpu == cpu) {
1945 /* Now try balancing at a lower domain level of cpu */
1950 /* Now try balancing at a lower domain level of new_cpu */
1953 weight = cpus_weight(span);
1954 for_each_domain(cpu, tmp) {
1955 if (weight <= cpus_weight(tmp->span))
1957 if (tmp->flags & flag)
1960 /* while loop will break here if sd == NULL */
1966 #endif /* CONFIG_SMP */
1969 * try_to_wake_up - wake up a thread
1970 * @p: the to-be-woken-up thread
1971 * @state: the mask of task states that can be woken
1972 * @sync: do a synchronous wakeup?
1974 * Put it on the run-queue if it's not already there. The "current"
1975 * thread is always on the run-queue (except when the actual
1976 * re-schedule is in progress), and as such you're allowed to do
1977 * the simpler "current->state = TASK_RUNNING" to mark yourself
1978 * runnable without the overhead of this.
1980 * returns failure only if the task is already active.
1982 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1984 int cpu, orig_cpu, this_cpu, success = 0;
1985 unsigned long flags;
1989 if (!sched_feat(SYNC_WAKEUPS))
1993 rq = task_rq_lock(p, &flags);
1994 old_state = p->state;
1995 if (!(old_state & state))
2003 this_cpu = smp_processor_id();
2006 if (unlikely(task_running(rq, p)))
2009 cpu = p->sched_class->select_task_rq(p, sync);
2010 if (cpu != orig_cpu) {
2011 set_task_cpu(p, cpu);
2012 task_rq_unlock(rq, &flags);
2013 /* might preempt at this point */
2014 rq = task_rq_lock(p, &flags);
2015 old_state = p->state;
2016 if (!(old_state & state))
2021 this_cpu = smp_processor_id();
2025 #ifdef CONFIG_SCHEDSTATS
2026 schedstat_inc(rq, ttwu_count);
2027 if (cpu == this_cpu)
2028 schedstat_inc(rq, ttwu_local);
2030 struct sched_domain *sd;
2031 for_each_domain(this_cpu, sd) {
2032 if (cpu_isset(cpu, sd->span)) {
2033 schedstat_inc(sd, ttwu_wake_remote);
2041 #endif /* CONFIG_SMP */
2042 schedstat_inc(p, se.nr_wakeups);
2044 schedstat_inc(p, se.nr_wakeups_sync);
2045 if (orig_cpu != cpu)
2046 schedstat_inc(p, se.nr_wakeups_migrate);
2047 if (cpu == this_cpu)
2048 schedstat_inc(p, se.nr_wakeups_local);
2050 schedstat_inc(p, se.nr_wakeups_remote);
2051 update_rq_clock(rq);
2052 activate_task(rq, p, 1);
2056 check_preempt_curr(rq, p);
2058 p->state = TASK_RUNNING;
2060 if (p->sched_class->task_wake_up)
2061 p->sched_class->task_wake_up(rq, p);
2064 task_rq_unlock(rq, &flags);
2069 int wake_up_process(struct task_struct *p)
2071 return try_to_wake_up(p, TASK_ALL, 0);
2073 EXPORT_SYMBOL(wake_up_process);
2075 int wake_up_state(struct task_struct *p, unsigned int state)
2077 return try_to_wake_up(p, state, 0);
2081 * Perform scheduler related setup for a newly forked process p.
2082 * p is forked by current.
2084 * __sched_fork() is basic setup used by init_idle() too:
2086 static void __sched_fork(struct task_struct *p)
2088 p->se.exec_start = 0;
2089 p->se.sum_exec_runtime = 0;
2090 p->se.prev_sum_exec_runtime = 0;
2091 p->se.last_wakeup = 0;
2092 p->se.avg_overlap = 0;
2094 #ifdef CONFIG_SCHEDSTATS
2095 p->se.wait_start = 0;
2096 p->se.sum_sleep_runtime = 0;
2097 p->se.sleep_start = 0;
2098 p->se.block_start = 0;
2099 p->se.sleep_max = 0;
2100 p->se.block_max = 0;
2102 p->se.slice_max = 0;
2106 INIT_LIST_HEAD(&p->rt.run_list);
2109 #ifdef CONFIG_PREEMPT_NOTIFIERS
2110 INIT_HLIST_HEAD(&p->preempt_notifiers);
2114 * We mark the process as running here, but have not actually
2115 * inserted it onto the runqueue yet. This guarantees that
2116 * nobody will actually run it, and a signal or other external
2117 * event cannot wake it up and insert it on the runqueue either.
2119 p->state = TASK_RUNNING;
2123 * fork()/clone()-time setup:
2125 void sched_fork(struct task_struct *p, int clone_flags)
2127 int cpu = get_cpu();
2132 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2134 set_task_cpu(p, cpu);
2137 * Make sure we do not leak PI boosting priority to the child:
2139 p->prio = current->normal_prio;
2140 if (!rt_prio(p->prio))
2141 p->sched_class = &fair_sched_class;
2143 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2144 if (likely(sched_info_on()))
2145 memset(&p->sched_info, 0, sizeof(p->sched_info));
2147 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2150 #ifdef CONFIG_PREEMPT
2151 /* Want to start with kernel preemption disabled. */
2152 task_thread_info(p)->preempt_count = 1;
2158 * wake_up_new_task - wake up a newly created task for the first time.
2160 * This function will do some initial scheduler statistics housekeeping
2161 * that must be done for every newly created context, then puts the task
2162 * on the runqueue and wakes it.
2164 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2166 unsigned long flags;
2169 rq = task_rq_lock(p, &flags);
2170 BUG_ON(p->state != TASK_RUNNING);
2171 update_rq_clock(rq);
2173 p->prio = effective_prio(p);
2175 if (!p->sched_class->task_new || !current->se.on_rq) {
2176 activate_task(rq, p, 0);
2179 * Let the scheduling class do new task startup
2180 * management (if any):
2182 p->sched_class->task_new(rq, p);
2183 inc_nr_running(p, rq);
2185 check_preempt_curr(rq, p);
2187 if (p->sched_class->task_wake_up)
2188 p->sched_class->task_wake_up(rq, p);
2190 task_rq_unlock(rq, &flags);
2193 #ifdef CONFIG_PREEMPT_NOTIFIERS
2196 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2197 * @notifier: notifier struct to register
2199 void preempt_notifier_register(struct preempt_notifier *notifier)
2201 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2203 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2206 * preempt_notifier_unregister - no longer interested in preemption notifications
2207 * @notifier: notifier struct to unregister
2209 * This is safe to call from within a preemption notifier.
2211 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2213 hlist_del(¬ifier->link);
2215 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2217 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2219 struct preempt_notifier *notifier;
2220 struct hlist_node *node;
2222 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2223 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2227 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2228 struct task_struct *next)
2230 struct preempt_notifier *notifier;
2231 struct hlist_node *node;
2233 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2234 notifier->ops->sched_out(notifier, next);
2239 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2244 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2245 struct task_struct *next)
2252 * prepare_task_switch - prepare to switch tasks
2253 * @rq: the runqueue preparing to switch
2254 * @prev: the current task that is being switched out
2255 * @next: the task we are going to switch to.
2257 * This is called with the rq lock held and interrupts off. It must
2258 * be paired with a subsequent finish_task_switch after the context
2261 * prepare_task_switch sets up locking and calls architecture specific
2265 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2266 struct task_struct *next)
2268 fire_sched_out_preempt_notifiers(prev, next);
2269 prepare_lock_switch(rq, next);
2270 prepare_arch_switch(next);
2274 * finish_task_switch - clean up after a task-switch
2275 * @rq: runqueue associated with task-switch
2276 * @prev: the thread we just switched away from.
2278 * finish_task_switch must be called after the context switch, paired
2279 * with a prepare_task_switch call before the context switch.
2280 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2281 * and do any other architecture-specific cleanup actions.
2283 * Note that we may have delayed dropping an mm in context_switch(). If
2284 * so, we finish that here outside of the runqueue lock. (Doing it
2285 * with the lock held can cause deadlocks; see schedule() for
2288 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2289 __releases(rq->lock)
2291 struct mm_struct *mm = rq->prev_mm;
2297 * A task struct has one reference for the use as "current".
2298 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2299 * schedule one last time. The schedule call will never return, and
2300 * the scheduled task must drop that reference.
2301 * The test for TASK_DEAD must occur while the runqueue locks are
2302 * still held, otherwise prev could be scheduled on another cpu, die
2303 * there before we look at prev->state, and then the reference would
2305 * Manfred Spraul <manfred@colorfullife.com>
2307 prev_state = prev->state;
2308 finish_arch_switch(prev);
2309 finish_lock_switch(rq, prev);
2311 if (current->sched_class->post_schedule)
2312 current->sched_class->post_schedule(rq);
2315 fire_sched_in_preempt_notifiers(current);
2318 if (unlikely(prev_state == TASK_DEAD)) {
2320 * Remove function-return probe instances associated with this
2321 * task and put them back on the free list.
2323 kprobe_flush_task(prev);
2324 put_task_struct(prev);
2329 * schedule_tail - first thing a freshly forked thread must call.
2330 * @prev: the thread we just switched away from.
2332 asmlinkage void schedule_tail(struct task_struct *prev)
2333 __releases(rq->lock)
2335 struct rq *rq = this_rq();
2337 finish_task_switch(rq, prev);
2338 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2339 /* In this case, finish_task_switch does not reenable preemption */
2342 if (current->set_child_tid)
2343 put_user(task_pid_vnr(current), current->set_child_tid);
2347 * context_switch - switch to the new MM and the new
2348 * thread's register state.
2351 context_switch(struct rq *rq, struct task_struct *prev,
2352 struct task_struct *next)
2354 struct mm_struct *mm, *oldmm;
2356 prepare_task_switch(rq, prev, next);
2358 oldmm = prev->active_mm;
2360 * For paravirt, this is coupled with an exit in switch_to to
2361 * combine the page table reload and the switch backend into
2364 arch_enter_lazy_cpu_mode();
2366 if (unlikely(!mm)) {
2367 next->active_mm = oldmm;
2368 atomic_inc(&oldmm->mm_count);
2369 enter_lazy_tlb(oldmm, next);
2371 switch_mm(oldmm, mm, next);
2373 if (unlikely(!prev->mm)) {
2374 prev->active_mm = NULL;
2375 rq->prev_mm = oldmm;
2378 * Since the runqueue lock will be released by the next
2379 * task (which is an invalid locking op but in the case
2380 * of the scheduler it's an obvious special-case), so we
2381 * do an early lockdep release here:
2383 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2384 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2387 /* Here we just switch the register state and the stack. */
2388 switch_to(prev, next, prev);
2392 * this_rq must be evaluated again because prev may have moved
2393 * CPUs since it called schedule(), thus the 'rq' on its stack
2394 * frame will be invalid.
2396 finish_task_switch(this_rq(), prev);
2400 * nr_running, nr_uninterruptible and nr_context_switches:
2402 * externally visible scheduler statistics: current number of runnable
2403 * threads, current number of uninterruptible-sleeping threads, total
2404 * number of context switches performed since bootup.
2406 unsigned long nr_running(void)
2408 unsigned long i, sum = 0;
2410 for_each_online_cpu(i)
2411 sum += cpu_rq(i)->nr_running;
2416 unsigned long nr_uninterruptible(void)
2418 unsigned long i, sum = 0;
2420 for_each_possible_cpu(i)
2421 sum += cpu_rq(i)->nr_uninterruptible;
2424 * Since we read the counters lockless, it might be slightly
2425 * inaccurate. Do not allow it to go below zero though:
2427 if (unlikely((long)sum < 0))
2433 unsigned long long nr_context_switches(void)
2436 unsigned long long sum = 0;
2438 for_each_possible_cpu(i)
2439 sum += cpu_rq(i)->nr_switches;
2444 unsigned long nr_iowait(void)
2446 unsigned long i, sum = 0;
2448 for_each_possible_cpu(i)
2449 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2454 unsigned long nr_active(void)
2456 unsigned long i, running = 0, uninterruptible = 0;
2458 for_each_online_cpu(i) {
2459 running += cpu_rq(i)->nr_running;
2460 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2463 if (unlikely((long)uninterruptible < 0))
2464 uninterruptible = 0;
2466 return running + uninterruptible;
2470 * Update rq->cpu_load[] statistics. This function is usually called every
2471 * scheduler tick (TICK_NSEC).
2473 static void update_cpu_load(struct rq *this_rq)
2475 unsigned long this_load = this_rq->load.weight;
2478 this_rq->nr_load_updates++;
2480 /* Update our load: */
2481 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2482 unsigned long old_load, new_load;
2484 /* scale is effectively 1 << i now, and >> i divides by scale */
2486 old_load = this_rq->cpu_load[i];
2487 new_load = this_load;
2489 * Round up the averaging division if load is increasing. This
2490 * prevents us from getting stuck on 9 if the load is 10, for
2493 if (new_load > old_load)
2494 new_load += scale-1;
2495 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2502 * double_rq_lock - safely lock two runqueues
2504 * Note this does not disable interrupts like task_rq_lock,
2505 * you need to do so manually before calling.
2507 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2508 __acquires(rq1->lock)
2509 __acquires(rq2->lock)
2511 BUG_ON(!irqs_disabled());
2513 spin_lock(&rq1->lock);
2514 __acquire(rq2->lock); /* Fake it out ;) */
2517 spin_lock(&rq1->lock);
2518 spin_lock(&rq2->lock);
2520 spin_lock(&rq2->lock);
2521 spin_lock(&rq1->lock);
2524 update_rq_clock(rq1);
2525 update_rq_clock(rq2);
2529 * double_rq_unlock - safely unlock two runqueues
2531 * Note this does not restore interrupts like task_rq_unlock,
2532 * you need to do so manually after calling.
2534 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2535 __releases(rq1->lock)
2536 __releases(rq2->lock)
2538 spin_unlock(&rq1->lock);
2540 spin_unlock(&rq2->lock);
2542 __release(rq2->lock);
2546 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2548 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2549 __releases(this_rq->lock)
2550 __acquires(busiest->lock)
2551 __acquires(this_rq->lock)
2555 if (unlikely(!irqs_disabled())) {
2556 /* printk() doesn't work good under rq->lock */
2557 spin_unlock(&this_rq->lock);
2560 if (unlikely(!spin_trylock(&busiest->lock))) {
2561 if (busiest < this_rq) {
2562 spin_unlock(&this_rq->lock);
2563 spin_lock(&busiest->lock);
2564 spin_lock(&this_rq->lock);
2567 spin_lock(&busiest->lock);
2573 * If dest_cpu is allowed for this process, migrate the task to it.
2574 * This is accomplished by forcing the cpu_allowed mask to only
2575 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2576 * the cpu_allowed mask is restored.
2578 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2580 struct migration_req req;
2581 unsigned long flags;
2584 rq = task_rq_lock(p, &flags);
2585 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2586 || unlikely(cpu_is_offline(dest_cpu)))
2589 /* force the process onto the specified CPU */
2590 if (migrate_task(p, dest_cpu, &req)) {
2591 /* Need to wait for migration thread (might exit: take ref). */
2592 struct task_struct *mt = rq->migration_thread;
2594 get_task_struct(mt);
2595 task_rq_unlock(rq, &flags);
2596 wake_up_process(mt);
2597 put_task_struct(mt);
2598 wait_for_completion(&req.done);
2603 task_rq_unlock(rq, &flags);
2607 * sched_exec - execve() is a valuable balancing opportunity, because at
2608 * this point the task has the smallest effective memory and cache footprint.
2610 void sched_exec(void)
2612 int new_cpu, this_cpu = get_cpu();
2613 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2615 if (new_cpu != this_cpu)
2616 sched_migrate_task(current, new_cpu);
2620 * pull_task - move a task from a remote runqueue to the local runqueue.
2621 * Both runqueues must be locked.
2623 static void pull_task(struct rq *src_rq, struct task_struct *p,
2624 struct rq *this_rq, int this_cpu)
2626 deactivate_task(src_rq, p, 0);
2627 set_task_cpu(p, this_cpu);
2628 activate_task(this_rq, p, 0);
2630 * Note that idle threads have a prio of MAX_PRIO, for this test
2631 * to be always true for them.
2633 check_preempt_curr(this_rq, p);
2637 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2640 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2641 struct sched_domain *sd, enum cpu_idle_type idle,
2645 * We do not migrate tasks that are:
2646 * 1) running (obviously), or
2647 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2648 * 3) are cache-hot on their current CPU.
2650 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2651 schedstat_inc(p, se.nr_failed_migrations_affine);
2656 if (task_running(rq, p)) {
2657 schedstat_inc(p, se.nr_failed_migrations_running);
2662 * Aggressive migration if:
2663 * 1) task is cache cold, or
2664 * 2) too many balance attempts have failed.
2667 if (!task_hot(p, rq->clock, sd) ||
2668 sd->nr_balance_failed > sd->cache_nice_tries) {
2669 #ifdef CONFIG_SCHEDSTATS
2670 if (task_hot(p, rq->clock, sd)) {
2671 schedstat_inc(sd, lb_hot_gained[idle]);
2672 schedstat_inc(p, se.nr_forced_migrations);
2678 if (task_hot(p, rq->clock, sd)) {
2679 schedstat_inc(p, se.nr_failed_migrations_hot);
2685 static unsigned long
2686 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2687 unsigned long max_load_move, struct sched_domain *sd,
2688 enum cpu_idle_type idle, int *all_pinned,
2689 int *this_best_prio, struct rq_iterator *iterator)
2691 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2692 struct task_struct *p;
2693 long rem_load_move = max_load_move;
2695 if (max_load_move == 0)
2701 * Start the load-balancing iterator:
2703 p = iterator->start(iterator->arg);
2705 if (!p || loops++ > sysctl_sched_nr_migrate)
2708 * To help distribute high priority tasks across CPUs we don't
2709 * skip a task if it will be the highest priority task (i.e. smallest
2710 * prio value) on its new queue regardless of its load weight
2712 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2713 SCHED_LOAD_SCALE_FUZZ;
2714 if ((skip_for_load && p->prio >= *this_best_prio) ||
2715 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2716 p = iterator->next(iterator->arg);
2720 pull_task(busiest, p, this_rq, this_cpu);
2722 rem_load_move -= p->se.load.weight;
2725 * We only want to steal up to the prescribed amount of weighted load.
2727 if (rem_load_move > 0) {
2728 if (p->prio < *this_best_prio)
2729 *this_best_prio = p->prio;
2730 p = iterator->next(iterator->arg);
2735 * Right now, this is one of only two places pull_task() is called,
2736 * so we can safely collect pull_task() stats here rather than
2737 * inside pull_task().
2739 schedstat_add(sd, lb_gained[idle], pulled);
2742 *all_pinned = pinned;
2744 return max_load_move - rem_load_move;
2748 * move_tasks tries to move up to max_load_move weighted load from busiest to
2749 * this_rq, as part of a balancing operation within domain "sd".
2750 * Returns 1 if successful and 0 otherwise.
2752 * Called with both runqueues locked.
2754 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2755 unsigned long max_load_move,
2756 struct sched_domain *sd, enum cpu_idle_type idle,
2759 const struct sched_class *class = sched_class_highest;
2760 unsigned long total_load_moved = 0;
2761 int this_best_prio = this_rq->curr->prio;
2765 class->load_balance(this_rq, this_cpu, busiest,
2766 max_load_move - total_load_moved,
2767 sd, idle, all_pinned, &this_best_prio);
2768 class = class->next;
2769 } while (class && max_load_move > total_load_moved);
2771 return total_load_moved > 0;
2775 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2776 struct sched_domain *sd, enum cpu_idle_type idle,
2777 struct rq_iterator *iterator)
2779 struct task_struct *p = iterator->start(iterator->arg);
2783 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2784 pull_task(busiest, p, this_rq, this_cpu);
2786 * Right now, this is only the second place pull_task()
2787 * is called, so we can safely collect pull_task()
2788 * stats here rather than inside pull_task().
2790 schedstat_inc(sd, lb_gained[idle]);
2794 p = iterator->next(iterator->arg);
2801 * move_one_task tries to move exactly one task from busiest to this_rq, as
2802 * part of active balancing operations within "domain".
2803 * Returns 1 if successful and 0 otherwise.
2805 * Called with both runqueues locked.
2807 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2808 struct sched_domain *sd, enum cpu_idle_type idle)
2810 const struct sched_class *class;
2812 for (class = sched_class_highest; class; class = class->next)
2813 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2820 * find_busiest_group finds and returns the busiest CPU group within the
2821 * domain. It calculates and returns the amount of weighted load which
2822 * should be moved to restore balance via the imbalance parameter.
2824 static struct sched_group *
2825 find_busiest_group(struct sched_domain *sd, int this_cpu,
2826 unsigned long *imbalance, enum cpu_idle_type idle,
2827 int *sd_idle, const cpumask_t *cpus, int *balance)
2829 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2830 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2831 unsigned long max_pull;
2832 unsigned long busiest_load_per_task, busiest_nr_running;
2833 unsigned long this_load_per_task, this_nr_running;
2834 int load_idx, group_imb = 0;
2835 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2836 int power_savings_balance = 1;
2837 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2838 unsigned long min_nr_running = ULONG_MAX;
2839 struct sched_group *group_min = NULL, *group_leader = NULL;
2842 max_load = this_load = total_load = total_pwr = 0;
2843 busiest_load_per_task = busiest_nr_running = 0;
2844 this_load_per_task = this_nr_running = 0;
2845 if (idle == CPU_NOT_IDLE)
2846 load_idx = sd->busy_idx;
2847 else if (idle == CPU_NEWLY_IDLE)
2848 load_idx = sd->newidle_idx;
2850 load_idx = sd->idle_idx;
2853 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2856 int __group_imb = 0;
2857 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2858 unsigned long sum_nr_running, sum_weighted_load;
2860 local_group = cpu_isset(this_cpu, group->cpumask);
2863 balance_cpu = first_cpu(group->cpumask);
2865 /* Tally up the load of all CPUs in the group */
2866 sum_weighted_load = sum_nr_running = avg_load = 0;
2868 min_cpu_load = ~0UL;
2870 for_each_cpu_mask(i, group->cpumask) {
2873 if (!cpu_isset(i, *cpus))
2878 if (*sd_idle && rq->nr_running)
2881 /* Bias balancing toward cpus of our domain */
2883 if (idle_cpu(i) && !first_idle_cpu) {
2888 load = target_load(i, load_idx);
2890 load = source_load(i, load_idx);
2891 if (load > max_cpu_load)
2892 max_cpu_load = load;
2893 if (min_cpu_load > load)
2894 min_cpu_load = load;
2898 sum_nr_running += rq->nr_running;
2899 sum_weighted_load += weighted_cpuload(i);
2903 * First idle cpu or the first cpu(busiest) in this sched group
2904 * is eligible for doing load balancing at this and above
2905 * domains. In the newly idle case, we will allow all the cpu's
2906 * to do the newly idle load balance.
2908 if (idle != CPU_NEWLY_IDLE && local_group &&
2909 balance_cpu != this_cpu && balance) {
2914 total_load += avg_load;
2915 total_pwr += group->__cpu_power;
2917 /* Adjust by relative CPU power of the group */
2918 avg_load = sg_div_cpu_power(group,
2919 avg_load * SCHED_LOAD_SCALE);
2921 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2924 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2927 this_load = avg_load;
2929 this_nr_running = sum_nr_running;
2930 this_load_per_task = sum_weighted_load;
2931 } else if (avg_load > max_load &&
2932 (sum_nr_running > group_capacity || __group_imb)) {
2933 max_load = avg_load;
2935 busiest_nr_running = sum_nr_running;
2936 busiest_load_per_task = sum_weighted_load;
2937 group_imb = __group_imb;
2940 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2942 * Busy processors will not participate in power savings
2945 if (idle == CPU_NOT_IDLE ||
2946 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2950 * If the local group is idle or completely loaded
2951 * no need to do power savings balance at this domain
2953 if (local_group && (this_nr_running >= group_capacity ||
2955 power_savings_balance = 0;
2958 * If a group is already running at full capacity or idle,
2959 * don't include that group in power savings calculations
2961 if (!power_savings_balance || sum_nr_running >= group_capacity
2966 * Calculate the group which has the least non-idle load.
2967 * This is the group from where we need to pick up the load
2970 if ((sum_nr_running < min_nr_running) ||
2971 (sum_nr_running == min_nr_running &&
2972 first_cpu(group->cpumask) <
2973 first_cpu(group_min->cpumask))) {
2975 min_nr_running = sum_nr_running;
2976 min_load_per_task = sum_weighted_load /
2981 * Calculate the group which is almost near its
2982 * capacity but still has some space to pick up some load
2983 * from other group and save more power
2985 if (sum_nr_running <= group_capacity - 1) {
2986 if (sum_nr_running > leader_nr_running ||
2987 (sum_nr_running == leader_nr_running &&
2988 first_cpu(group->cpumask) >
2989 first_cpu(group_leader->cpumask))) {
2990 group_leader = group;
2991 leader_nr_running = sum_nr_running;
2996 group = group->next;
2997 } while (group != sd->groups);
2999 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3002 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3004 if (this_load >= avg_load ||
3005 100*max_load <= sd->imbalance_pct*this_load)
3008 busiest_load_per_task /= busiest_nr_running;
3010 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3013 * We're trying to get all the cpus to the average_load, so we don't
3014 * want to push ourselves above the average load, nor do we wish to
3015 * reduce the max loaded cpu below the average load, as either of these
3016 * actions would just result in more rebalancing later, and ping-pong
3017 * tasks around. Thus we look for the minimum possible imbalance.
3018 * Negative imbalances (*we* are more loaded than anyone else) will
3019 * be counted as no imbalance for these purposes -- we can't fix that
3020 * by pulling tasks to us. Be careful of negative numbers as they'll
3021 * appear as very large values with unsigned longs.
3023 if (max_load <= busiest_load_per_task)
3027 * In the presence of smp nice balancing, certain scenarios can have
3028 * max load less than avg load(as we skip the groups at or below
3029 * its cpu_power, while calculating max_load..)
3031 if (max_load < avg_load) {
3033 goto small_imbalance;
3036 /* Don't want to pull so many tasks that a group would go idle */
3037 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3039 /* How much load to actually move to equalise the imbalance */
3040 *imbalance = min(max_pull * busiest->__cpu_power,
3041 (avg_load - this_load) * this->__cpu_power)
3045 * if *imbalance is less than the average load per runnable task
3046 * there is no gaurantee that any tasks will be moved so we'll have
3047 * a think about bumping its value to force at least one task to be
3050 if (*imbalance < busiest_load_per_task) {
3051 unsigned long tmp, pwr_now, pwr_move;
3055 pwr_move = pwr_now = 0;
3057 if (this_nr_running) {
3058 this_load_per_task /= this_nr_running;
3059 if (busiest_load_per_task > this_load_per_task)
3062 this_load_per_task = SCHED_LOAD_SCALE;
3064 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3065 busiest_load_per_task * imbn) {
3066 *imbalance = busiest_load_per_task;
3071 * OK, we don't have enough imbalance to justify moving tasks,
3072 * however we may be able to increase total CPU power used by
3076 pwr_now += busiest->__cpu_power *
3077 min(busiest_load_per_task, max_load);
3078 pwr_now += this->__cpu_power *
3079 min(this_load_per_task, this_load);
3080 pwr_now /= SCHED_LOAD_SCALE;
3082 /* Amount of load we'd subtract */
3083 tmp = sg_div_cpu_power(busiest,
3084 busiest_load_per_task * SCHED_LOAD_SCALE);
3086 pwr_move += busiest->__cpu_power *
3087 min(busiest_load_per_task, max_load - tmp);
3089 /* Amount of load we'd add */
3090 if (max_load * busiest->__cpu_power <
3091 busiest_load_per_task * SCHED_LOAD_SCALE)
3092 tmp = sg_div_cpu_power(this,
3093 max_load * busiest->__cpu_power);
3095 tmp = sg_div_cpu_power(this,
3096 busiest_load_per_task * SCHED_LOAD_SCALE);
3097 pwr_move += this->__cpu_power *
3098 min(this_load_per_task, this_load + tmp);
3099 pwr_move /= SCHED_LOAD_SCALE;
3101 /* Move if we gain throughput */
3102 if (pwr_move > pwr_now)
3103 *imbalance = busiest_load_per_task;
3109 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3110 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3113 if (this == group_leader && group_leader != group_min) {
3114 *imbalance = min_load_per_task;
3124 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3127 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3128 unsigned long imbalance, const cpumask_t *cpus)
3130 struct rq *busiest = NULL, *rq;
3131 unsigned long max_load = 0;
3134 for_each_cpu_mask(i, group->cpumask) {
3137 if (!cpu_isset(i, *cpus))
3141 wl = weighted_cpuload(i);
3143 if (rq->nr_running == 1 && wl > imbalance)
3146 if (wl > max_load) {
3156 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3157 * so long as it is large enough.
3159 #define MAX_PINNED_INTERVAL 512
3162 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3163 * tasks if there is an imbalance.
3165 static int load_balance(int this_cpu, struct rq *this_rq,
3166 struct sched_domain *sd, enum cpu_idle_type idle,
3167 int *balance, cpumask_t *cpus)
3169 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3170 struct sched_group *group;
3171 unsigned long imbalance;
3173 unsigned long flags;
3178 * When power savings policy is enabled for the parent domain, idle
3179 * sibling can pick up load irrespective of busy siblings. In this case,
3180 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3181 * portraying it as CPU_NOT_IDLE.
3183 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3184 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3187 schedstat_inc(sd, lb_count[idle]);
3190 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3197 schedstat_inc(sd, lb_nobusyg[idle]);
3201 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3203 schedstat_inc(sd, lb_nobusyq[idle]);
3207 BUG_ON(busiest == this_rq);
3209 schedstat_add(sd, lb_imbalance[idle], imbalance);
3212 if (busiest->nr_running > 1) {
3214 * Attempt to move tasks. If find_busiest_group has found
3215 * an imbalance but busiest->nr_running <= 1, the group is
3216 * still unbalanced. ld_moved simply stays zero, so it is
3217 * correctly treated as an imbalance.
3219 local_irq_save(flags);
3220 double_rq_lock(this_rq, busiest);
3221 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3222 imbalance, sd, idle, &all_pinned);
3223 double_rq_unlock(this_rq, busiest);
3224 local_irq_restore(flags);
3227 * some other cpu did the load balance for us.
3229 if (ld_moved && this_cpu != smp_processor_id())
3230 resched_cpu(this_cpu);
3232 /* All tasks on this runqueue were pinned by CPU affinity */
3233 if (unlikely(all_pinned)) {
3234 cpu_clear(cpu_of(busiest), *cpus);
3235 if (!cpus_empty(*cpus))
3242 schedstat_inc(sd, lb_failed[idle]);
3243 sd->nr_balance_failed++;
3245 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3247 spin_lock_irqsave(&busiest->lock, flags);
3249 /* don't kick the migration_thread, if the curr
3250 * task on busiest cpu can't be moved to this_cpu
3252 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3253 spin_unlock_irqrestore(&busiest->lock, flags);
3255 goto out_one_pinned;
3258 if (!busiest->active_balance) {
3259 busiest->active_balance = 1;
3260 busiest->push_cpu = this_cpu;
3263 spin_unlock_irqrestore(&busiest->lock, flags);
3265 wake_up_process(busiest->migration_thread);
3268 * We've kicked active balancing, reset the failure
3271 sd->nr_balance_failed = sd->cache_nice_tries+1;
3274 sd->nr_balance_failed = 0;
3276 if (likely(!active_balance)) {
3277 /* We were unbalanced, so reset the balancing interval */
3278 sd->balance_interval = sd->min_interval;
3281 * If we've begun active balancing, start to back off. This
3282 * case may not be covered by the all_pinned logic if there
3283 * is only 1 task on the busy runqueue (because we don't call
3286 if (sd->balance_interval < sd->max_interval)
3287 sd->balance_interval *= 2;
3290 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3291 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3296 schedstat_inc(sd, lb_balanced[idle]);
3298 sd->nr_balance_failed = 0;
3301 /* tune up the balancing interval */
3302 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3303 (sd->balance_interval < sd->max_interval))
3304 sd->balance_interval *= 2;
3306 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3307 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3313 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3314 * tasks if there is an imbalance.
3316 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3317 * this_rq is locked.
3320 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3323 struct sched_group *group;
3324 struct rq *busiest = NULL;
3325 unsigned long imbalance;
3333 * When power savings policy is enabled for the parent domain, idle
3334 * sibling can pick up load irrespective of busy siblings. In this case,
3335 * let the state of idle sibling percolate up as IDLE, instead of
3336 * portraying it as CPU_NOT_IDLE.
3338 if (sd->flags & SD_SHARE_CPUPOWER &&
3339 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3342 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3344 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3345 &sd_idle, cpus, NULL);
3347 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3351 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3353 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3357 BUG_ON(busiest == this_rq);
3359 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3362 if (busiest->nr_running > 1) {
3363 /* Attempt to move tasks */
3364 double_lock_balance(this_rq, busiest);
3365 /* this_rq->clock is already updated */
3366 update_rq_clock(busiest);
3367 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3368 imbalance, sd, CPU_NEWLY_IDLE,
3370 spin_unlock(&busiest->lock);
3372 if (unlikely(all_pinned)) {
3373 cpu_clear(cpu_of(busiest), *cpus);
3374 if (!cpus_empty(*cpus))
3380 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3381 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3382 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3385 sd->nr_balance_failed = 0;
3390 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3391 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3392 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3394 sd->nr_balance_failed = 0;
3400 * idle_balance is called by schedule() if this_cpu is about to become
3401 * idle. Attempts to pull tasks from other CPUs.
3403 static void idle_balance(int this_cpu, struct rq *this_rq)
3405 struct sched_domain *sd;
3406 int pulled_task = -1;
3407 unsigned long next_balance = jiffies + HZ;
3410 for_each_domain(this_cpu, sd) {
3411 unsigned long interval;
3413 if (!(sd->flags & SD_LOAD_BALANCE))
3416 if (sd->flags & SD_BALANCE_NEWIDLE)
3417 /* If we've pulled tasks over stop searching: */
3418 pulled_task = load_balance_newidle(this_cpu, this_rq,
3421 interval = msecs_to_jiffies(sd->balance_interval);
3422 if (time_after(next_balance, sd->last_balance + interval))
3423 next_balance = sd->last_balance + interval;
3427 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3429 * We are going idle. next_balance may be set based on
3430 * a busy processor. So reset next_balance.
3432 this_rq->next_balance = next_balance;
3437 * active_load_balance is run by migration threads. It pushes running tasks
3438 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3439 * running on each physical CPU where possible, and avoids physical /
3440 * logical imbalances.
3442 * Called with busiest_rq locked.
3444 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3446 int target_cpu = busiest_rq->push_cpu;
3447 struct sched_domain *sd;
3448 struct rq *target_rq;
3450 /* Is there any task to move? */
3451 if (busiest_rq->nr_running <= 1)
3454 target_rq = cpu_rq(target_cpu);
3457 * This condition is "impossible", if it occurs
3458 * we need to fix it. Originally reported by
3459 * Bjorn Helgaas on a 128-cpu setup.
3461 BUG_ON(busiest_rq == target_rq);
3463 /* move a task from busiest_rq to target_rq */
3464 double_lock_balance(busiest_rq, target_rq);
3465 update_rq_clock(busiest_rq);
3466 update_rq_clock(target_rq);
3468 /* Search for an sd spanning us and the target CPU. */
3469 for_each_domain(target_cpu, sd) {
3470 if ((sd->flags & SD_LOAD_BALANCE) &&
3471 cpu_isset(busiest_cpu, sd->span))
3476 schedstat_inc(sd, alb_count);
3478 if (move_one_task(target_rq, target_cpu, busiest_rq,
3480 schedstat_inc(sd, alb_pushed);
3482 schedstat_inc(sd, alb_failed);
3484 spin_unlock(&target_rq->lock);
3489 atomic_t load_balancer;
3491 } nohz ____cacheline_aligned = {
3492 .load_balancer = ATOMIC_INIT(-1),
3493 .cpu_mask = CPU_MASK_NONE,
3497 * This routine will try to nominate the ilb (idle load balancing)
3498 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3499 * load balancing on behalf of all those cpus. If all the cpus in the system
3500 * go into this tickless mode, then there will be no ilb owner (as there is
3501 * no need for one) and all the cpus will sleep till the next wakeup event
3504 * For the ilb owner, tick is not stopped. And this tick will be used
3505 * for idle load balancing. ilb owner will still be part of
3508 * While stopping the tick, this cpu will become the ilb owner if there
3509 * is no other owner. And will be the owner till that cpu becomes busy
3510 * or if all cpus in the system stop their ticks at which point
3511 * there is no need for ilb owner.
3513 * When the ilb owner becomes busy, it nominates another owner, during the
3514 * next busy scheduler_tick()
3516 int select_nohz_load_balancer(int stop_tick)
3518 int cpu = smp_processor_id();
3521 cpu_set(cpu, nohz.cpu_mask);
3522 cpu_rq(cpu)->in_nohz_recently = 1;
3525 * If we are going offline and still the leader, give up!
3527 if (cpu_is_offline(cpu) &&
3528 atomic_read(&nohz.load_balancer) == cpu) {
3529 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3534 /* time for ilb owner also to sleep */
3535 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3536 if (atomic_read(&nohz.load_balancer) == cpu)
3537 atomic_set(&nohz.load_balancer, -1);
3541 if (atomic_read(&nohz.load_balancer) == -1) {
3542 /* make me the ilb owner */
3543 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3545 } else if (atomic_read(&nohz.load_balancer) == cpu)
3548 if (!cpu_isset(cpu, nohz.cpu_mask))
3551 cpu_clear(cpu, nohz.cpu_mask);
3553 if (atomic_read(&nohz.load_balancer) == cpu)
3554 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3561 static DEFINE_SPINLOCK(balancing);
3564 * It checks each scheduling domain to see if it is due to be balanced,
3565 * and initiates a balancing operation if so.
3567 * Balancing parameters are set up in arch_init_sched_domains.
3569 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3572 struct rq *rq = cpu_rq(cpu);
3573 unsigned long interval;
3574 struct sched_domain *sd;
3575 /* Earliest time when we have to do rebalance again */
3576 unsigned long next_balance = jiffies + 60*HZ;
3577 int update_next_balance = 0;
3580 for_each_domain(cpu, sd) {
3581 if (!(sd->flags & SD_LOAD_BALANCE))
3584 interval = sd->balance_interval;
3585 if (idle != CPU_IDLE)
3586 interval *= sd->busy_factor;
3588 /* scale ms to jiffies */
3589 interval = msecs_to_jiffies(interval);
3590 if (unlikely(!interval))
3592 if (interval > HZ*NR_CPUS/10)
3593 interval = HZ*NR_CPUS/10;
3596 if (sd->flags & SD_SERIALIZE) {
3597 if (!spin_trylock(&balancing))
3601 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3602 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3604 * We've pulled tasks over so either we're no
3605 * longer idle, or one of our SMT siblings is
3608 idle = CPU_NOT_IDLE;
3610 sd->last_balance = jiffies;
3612 if (sd->flags & SD_SERIALIZE)
3613 spin_unlock(&balancing);
3615 if (time_after(next_balance, sd->last_balance + interval)) {
3616 next_balance = sd->last_balance + interval;
3617 update_next_balance = 1;
3621 * Stop the load balance at this level. There is another
3622 * CPU in our sched group which is doing load balancing more
3630 * next_balance will be updated only when there is a need.
3631 * When the cpu is attached to null domain for ex, it will not be
3634 if (likely(update_next_balance))
3635 rq->next_balance = next_balance;
3639 * run_rebalance_domains is triggered when needed from the scheduler tick.
3640 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3641 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3643 static void run_rebalance_domains(struct softirq_action *h)
3645 int this_cpu = smp_processor_id();
3646 struct rq *this_rq = cpu_rq(this_cpu);
3647 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3648 CPU_IDLE : CPU_NOT_IDLE;
3650 rebalance_domains(this_cpu, idle);
3654 * If this cpu is the owner for idle load balancing, then do the
3655 * balancing on behalf of the other idle cpus whose ticks are
3658 if (this_rq->idle_at_tick &&
3659 atomic_read(&nohz.load_balancer) == this_cpu) {
3660 cpumask_t cpus = nohz.cpu_mask;
3664 cpu_clear(this_cpu, cpus);
3665 for_each_cpu_mask(balance_cpu, cpus) {
3667 * If this cpu gets work to do, stop the load balancing
3668 * work being done for other cpus. Next load
3669 * balancing owner will pick it up.
3674 rebalance_domains(balance_cpu, CPU_IDLE);
3676 rq = cpu_rq(balance_cpu);
3677 if (time_after(this_rq->next_balance, rq->next_balance))
3678 this_rq->next_balance = rq->next_balance;
3685 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3687 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3688 * idle load balancing owner or decide to stop the periodic load balancing,
3689 * if the whole system is idle.
3691 static inline void trigger_load_balance(struct rq *rq, int cpu)
3695 * If we were in the nohz mode recently and busy at the current
3696 * scheduler tick, then check if we need to nominate new idle
3699 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3700 rq->in_nohz_recently = 0;
3702 if (atomic_read(&nohz.load_balancer) == cpu) {
3703 cpu_clear(cpu, nohz.cpu_mask);
3704 atomic_set(&nohz.load_balancer, -1);
3707 if (atomic_read(&nohz.load_balancer) == -1) {
3709 * simple selection for now: Nominate the
3710 * first cpu in the nohz list to be the next
3713 * TBD: Traverse the sched domains and nominate
3714 * the nearest cpu in the nohz.cpu_mask.
3716 int ilb = first_cpu(nohz.cpu_mask);
3718 if (ilb < nr_cpu_ids)
3724 * If this cpu is idle and doing idle load balancing for all the
3725 * cpus with ticks stopped, is it time for that to stop?
3727 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3728 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3734 * If this cpu is idle and the idle load balancing is done by
3735 * someone else, then no need raise the SCHED_SOFTIRQ
3737 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3738 cpu_isset(cpu, nohz.cpu_mask))
3741 if (time_after_eq(jiffies, rq->next_balance))
3742 raise_softirq(SCHED_SOFTIRQ);
3745 #else /* CONFIG_SMP */
3748 * on UP we do not need to balance between CPUs:
3750 static inline void idle_balance(int cpu, struct rq *rq)
3756 DEFINE_PER_CPU(struct kernel_stat, kstat);
3758 EXPORT_PER_CPU_SYMBOL(kstat);
3761 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3762 * that have not yet been banked in case the task is currently running.
3764 unsigned long long task_sched_runtime(struct task_struct *p)
3766 unsigned long flags;
3770 rq = task_rq_lock(p, &flags);
3771 ns = p->se.sum_exec_runtime;
3772 if (task_current(rq, p)) {
3773 update_rq_clock(rq);
3774 delta_exec = rq->clock - p->se.exec_start;
3775 if ((s64)delta_exec > 0)
3778 task_rq_unlock(rq, &flags);
3784 * Account user cpu time to a process.
3785 * @p: the process that the cpu time gets accounted to
3786 * @cputime: the cpu time spent in user space since the last update
3788 void account_user_time(struct task_struct *p, cputime_t cputime)
3790 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3793 p->utime = cputime_add(p->utime, cputime);
3795 /* Add user time to cpustat. */
3796 tmp = cputime_to_cputime64(cputime);
3797 if (TASK_NICE(p) > 0)
3798 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3800 cpustat->user = cputime64_add(cpustat->user, tmp);
3804 * Account guest cpu time to a process.
3805 * @p: the process that the cpu time gets accounted to
3806 * @cputime: the cpu time spent in virtual machine since the last update
3808 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3811 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3813 tmp = cputime_to_cputime64(cputime);
3815 p->utime = cputime_add(p->utime, cputime);
3816 p->gtime = cputime_add(p->gtime, cputime);
3818 cpustat->user = cputime64_add(cpustat->user, tmp);
3819 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3823 * Account scaled user cpu time to a process.
3824 * @p: the process that the cpu time gets accounted to
3825 * @cputime: the cpu time spent in user space since the last update
3827 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3829 p->utimescaled = cputime_add(p->utimescaled, cputime);
3833 * Account system cpu time to a process.
3834 * @p: the process that the cpu time gets accounted to
3835 * @hardirq_offset: the offset to subtract from hardirq_count()
3836 * @cputime: the cpu time spent in kernel space since the last update
3838 void account_system_time(struct task_struct *p, int hardirq_offset,
3841 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3842 struct rq *rq = this_rq();
3845 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3846 return account_guest_time(p, cputime);
3848 p->stime = cputime_add(p->stime, cputime);
3850 /* Add system time to cpustat. */
3851 tmp = cputime_to_cputime64(cputime);
3852 if (hardirq_count() - hardirq_offset)
3853 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3854 else if (softirq_count())
3855 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3856 else if (p != rq->idle)
3857 cpustat->system = cputime64_add(cpustat->system, tmp);
3858 else if (atomic_read(&rq->nr_iowait) > 0)
3859 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3861 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3862 /* Account for system time used */
3863 acct_update_integrals(p);
3867 * Account scaled system cpu time to a process.
3868 * @p: the process that the cpu time gets accounted to
3869 * @hardirq_offset: the offset to subtract from hardirq_count()
3870 * @cputime: the cpu time spent in kernel space since the last update
3872 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3874 p->stimescaled = cputime_add(p->stimescaled, cputime);
3878 * Account for involuntary wait time.
3879 * @p: the process from which the cpu time has been stolen
3880 * @steal: the cpu time spent in involuntary wait
3882 void account_steal_time(struct task_struct *p, cputime_t steal)
3884 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3885 cputime64_t tmp = cputime_to_cputime64(steal);
3886 struct rq *rq = this_rq();
3888 if (p == rq->idle) {
3889 p->stime = cputime_add(p->stime, steal);
3890 if (atomic_read(&rq->nr_iowait) > 0)
3891 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3893 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3895 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3899 * This function gets called by the timer code, with HZ frequency.
3900 * We call it with interrupts disabled.
3902 * It also gets called by the fork code, when changing the parent's
3905 void scheduler_tick(void)
3907 int cpu = smp_processor_id();
3908 struct rq *rq = cpu_rq(cpu);
3909 struct task_struct *curr = rq->curr;
3910 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3912 spin_lock(&rq->lock);
3913 __update_rq_clock(rq);
3915 * Let rq->clock advance by at least TICK_NSEC:
3917 if (unlikely(rq->clock < next_tick)) {
3918 rq->clock = next_tick;
3919 rq->clock_underflows++;
3921 rq->tick_timestamp = rq->clock;
3922 update_last_tick_seen(rq);
3923 update_cpu_load(rq);
3924 curr->sched_class->task_tick(rq, curr, 0);
3925 spin_unlock(&rq->lock);
3928 rq->idle_at_tick = idle_cpu(cpu);
3929 trigger_load_balance(rq, cpu);
3933 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3935 void __kprobes add_preempt_count(int val)
3940 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3942 preempt_count() += val;
3944 * Spinlock count overflowing soon?
3946 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3949 EXPORT_SYMBOL(add_preempt_count);
3951 void __kprobes sub_preempt_count(int val)
3956 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3959 * Is the spinlock portion underflowing?
3961 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3962 !(preempt_count() & PREEMPT_MASK)))
3965 preempt_count() -= val;
3967 EXPORT_SYMBOL(sub_preempt_count);
3972 * Print scheduling while atomic bug:
3974 static noinline void __schedule_bug(struct task_struct *prev)
3976 struct pt_regs *regs = get_irq_regs();
3978 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3979 prev->comm, prev->pid, preempt_count());
3981 debug_show_held_locks(prev);
3982 if (irqs_disabled())
3983 print_irqtrace_events(prev);
3992 * Various schedule()-time debugging checks and statistics:
3994 static inline void schedule_debug(struct task_struct *prev)
3997 * Test if we are atomic. Since do_exit() needs to call into
3998 * schedule() atomically, we ignore that path for now.
3999 * Otherwise, whine if we are scheduling when we should not be.
4001 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4002 __schedule_bug(prev);
4004 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4006 schedstat_inc(this_rq(), sched_count);
4007 #ifdef CONFIG_SCHEDSTATS
4008 if (unlikely(prev->lock_depth >= 0)) {
4009 schedstat_inc(this_rq(), bkl_count);
4010 schedstat_inc(prev, sched_info.bkl_count);
4016 * Pick up the highest-prio task:
4018 static inline struct task_struct *
4019 pick_next_task(struct rq *rq, struct task_struct *prev)
4021 const struct sched_class *class;
4022 struct task_struct *p;
4025 * Optimization: we know that if all tasks are in
4026 * the fair class we can call that function directly:
4028 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4029 p = fair_sched_class.pick_next_task(rq);
4034 class = sched_class_highest;
4036 p = class->pick_next_task(rq);
4040 * Will never be NULL as the idle class always
4041 * returns a non-NULL p:
4043 class = class->next;
4048 * schedule() is the main scheduler function.
4050 asmlinkage void __sched schedule(void)
4052 struct task_struct *prev, *next;
4053 unsigned long *switch_count;
4059 cpu = smp_processor_id();
4063 switch_count = &prev->nivcsw;
4065 release_kernel_lock(prev);
4066 need_resched_nonpreemptible:
4068 schedule_debug(prev);
4073 * Do the rq-clock update outside the rq lock:
4075 local_irq_disable();
4076 __update_rq_clock(rq);
4077 spin_lock(&rq->lock);
4078 clear_tsk_need_resched(prev);
4080 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4081 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4082 signal_pending(prev))) {
4083 prev->state = TASK_RUNNING;
4085 deactivate_task(rq, prev, 1);
4087 switch_count = &prev->nvcsw;
4091 if (prev->sched_class->pre_schedule)
4092 prev->sched_class->pre_schedule(rq, prev);
4095 if (unlikely(!rq->nr_running))
4096 idle_balance(cpu, rq);
4098 prev->sched_class->put_prev_task(rq, prev);
4099 next = pick_next_task(rq, prev);
4101 sched_info_switch(prev, next);
4103 if (likely(prev != next)) {
4108 context_switch(rq, prev, next); /* unlocks the rq */
4110 * the context switch might have flipped the stack from under
4111 * us, hence refresh the local variables.
4113 cpu = smp_processor_id();
4116 spin_unlock_irq(&rq->lock);
4120 if (unlikely(reacquire_kernel_lock(current) < 0))
4121 goto need_resched_nonpreemptible;
4123 preempt_enable_no_resched();
4124 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4127 EXPORT_SYMBOL(schedule);
4129 #ifdef CONFIG_PREEMPT
4131 * this is the entry point to schedule() from in-kernel preemption
4132 * off of preempt_enable. Kernel preemptions off return from interrupt
4133 * occur there and call schedule directly.
4135 asmlinkage void __sched preempt_schedule(void)
4137 struct thread_info *ti = current_thread_info();
4138 struct task_struct *task = current;
4139 int saved_lock_depth;
4142 * If there is a non-zero preempt_count or interrupts are disabled,
4143 * we do not want to preempt the current task. Just return..
4145 if (likely(ti->preempt_count || irqs_disabled()))
4149 add_preempt_count(PREEMPT_ACTIVE);
4152 * We keep the big kernel semaphore locked, but we
4153 * clear ->lock_depth so that schedule() doesnt
4154 * auto-release the semaphore:
4156 saved_lock_depth = task->lock_depth;
4157 task->lock_depth = -1;
4159 task->lock_depth = saved_lock_depth;
4160 sub_preempt_count(PREEMPT_ACTIVE);
4163 * Check again in case we missed a preemption opportunity
4164 * between schedule and now.
4167 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4169 EXPORT_SYMBOL(preempt_schedule);
4172 * this is the entry point to schedule() from kernel preemption
4173 * off of irq context.
4174 * Note, that this is called and return with irqs disabled. This will
4175 * protect us against recursive calling from irq.
4177 asmlinkage void __sched preempt_schedule_irq(void)
4179 struct thread_info *ti = current_thread_info();
4180 struct task_struct *task = current;
4181 int saved_lock_depth;
4183 /* Catch callers which need to be fixed */
4184 BUG_ON(ti->preempt_count || !irqs_disabled());
4187 add_preempt_count(PREEMPT_ACTIVE);
4190 * We keep the big kernel semaphore locked, but we
4191 * clear ->lock_depth so that schedule() doesnt
4192 * auto-release the semaphore:
4194 saved_lock_depth = task->lock_depth;
4195 task->lock_depth = -1;
4198 local_irq_disable();
4199 task->lock_depth = saved_lock_depth;
4200 sub_preempt_count(PREEMPT_ACTIVE);
4203 * Check again in case we missed a preemption opportunity
4204 * between schedule and now.
4207 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4210 #endif /* CONFIG_PREEMPT */
4212 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4215 return try_to_wake_up(curr->private, mode, sync);
4217 EXPORT_SYMBOL(default_wake_function);
4220 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4221 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4222 * number) then we wake all the non-exclusive tasks and one exclusive task.
4224 * There are circumstances in which we can try to wake a task which has already
4225 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4226 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4228 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4229 int nr_exclusive, int sync, void *key)
4231 wait_queue_t *curr, *next;
4233 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4234 unsigned flags = curr->flags;
4236 if (curr->func(curr, mode, sync, key) &&
4237 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4243 * __wake_up - wake up threads blocked on a waitqueue.
4245 * @mode: which threads
4246 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4247 * @key: is directly passed to the wakeup function
4249 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4250 int nr_exclusive, void *key)
4252 unsigned long flags;
4254 spin_lock_irqsave(&q->lock, flags);
4255 __wake_up_common(q, mode, nr_exclusive, 0, key);
4256 spin_unlock_irqrestore(&q->lock, flags);
4258 EXPORT_SYMBOL(__wake_up);
4261 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4263 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4265 __wake_up_common(q, mode, 1, 0, NULL);
4269 * __wake_up_sync - wake up threads blocked on a waitqueue.
4271 * @mode: which threads
4272 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4274 * The sync wakeup differs that the waker knows that it will schedule
4275 * away soon, so while the target thread will be woken up, it will not
4276 * be migrated to another CPU - ie. the two threads are 'synchronized'
4277 * with each other. This can prevent needless bouncing between CPUs.
4279 * On UP it can prevent extra preemption.
4282 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4284 unsigned long flags;
4290 if (unlikely(!nr_exclusive))
4293 spin_lock_irqsave(&q->lock, flags);
4294 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4295 spin_unlock_irqrestore(&q->lock, flags);
4297 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4299 void complete(struct completion *x)
4301 unsigned long flags;
4303 spin_lock_irqsave(&x->wait.lock, flags);
4305 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4306 spin_unlock_irqrestore(&x->wait.lock, flags);
4308 EXPORT_SYMBOL(complete);
4310 void complete_all(struct completion *x)
4312 unsigned long flags;
4314 spin_lock_irqsave(&x->wait.lock, flags);
4315 x->done += UINT_MAX/2;
4316 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4317 spin_unlock_irqrestore(&x->wait.lock, flags);
4319 EXPORT_SYMBOL(complete_all);
4321 static inline long __sched
4322 do_wait_for_common(struct completion *x, long timeout, int state)
4325 DECLARE_WAITQUEUE(wait, current);
4327 wait.flags |= WQ_FLAG_EXCLUSIVE;
4328 __add_wait_queue_tail(&x->wait, &wait);
4330 if ((state == TASK_INTERRUPTIBLE &&
4331 signal_pending(current)) ||
4332 (state == TASK_KILLABLE &&
4333 fatal_signal_pending(current))) {
4334 __remove_wait_queue(&x->wait, &wait);
4335 return -ERESTARTSYS;
4337 __set_current_state(state);
4338 spin_unlock_irq(&x->wait.lock);
4339 timeout = schedule_timeout(timeout);
4340 spin_lock_irq(&x->wait.lock);
4342 __remove_wait_queue(&x->wait, &wait);
4346 __remove_wait_queue(&x->wait, &wait);
4353 wait_for_common(struct completion *x, long timeout, int state)
4357 spin_lock_irq(&x->wait.lock);
4358 timeout = do_wait_for_common(x, timeout, state);
4359 spin_unlock_irq(&x->wait.lock);
4363 void __sched wait_for_completion(struct completion *x)
4365 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4367 EXPORT_SYMBOL(wait_for_completion);
4369 unsigned long __sched
4370 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4372 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4374 EXPORT_SYMBOL(wait_for_completion_timeout);
4376 int __sched wait_for_completion_interruptible(struct completion *x)
4378 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4379 if (t == -ERESTARTSYS)
4383 EXPORT_SYMBOL(wait_for_completion_interruptible);
4385 unsigned long __sched
4386 wait_for_completion_interruptible_timeout(struct completion *x,
4387 unsigned long timeout)
4389 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4391 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4393 int __sched wait_for_completion_killable(struct completion *x)
4395 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4396 if (t == -ERESTARTSYS)
4400 EXPORT_SYMBOL(wait_for_completion_killable);
4403 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4405 unsigned long flags;
4408 init_waitqueue_entry(&wait, current);
4410 __set_current_state(state);
4412 spin_lock_irqsave(&q->lock, flags);
4413 __add_wait_queue(q, &wait);
4414 spin_unlock(&q->lock);
4415 timeout = schedule_timeout(timeout);
4416 spin_lock_irq(&q->lock);
4417 __remove_wait_queue(q, &wait);
4418 spin_unlock_irqrestore(&q->lock, flags);
4423 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4425 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4427 EXPORT_SYMBOL(interruptible_sleep_on);
4430 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4432 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4434 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4436 void __sched sleep_on(wait_queue_head_t *q)
4438 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4440 EXPORT_SYMBOL(sleep_on);
4442 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4444 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4446 EXPORT_SYMBOL(sleep_on_timeout);
4448 #ifdef CONFIG_RT_MUTEXES
4451 * rt_mutex_setprio - set the current priority of a task
4453 * @prio: prio value (kernel-internal form)
4455 * This function changes the 'effective' priority of a task. It does
4456 * not touch ->normal_prio like __setscheduler().
4458 * Used by the rt_mutex code to implement priority inheritance logic.
4460 void rt_mutex_setprio(struct task_struct *p, int prio)
4462 unsigned long flags;
4463 int oldprio, on_rq, running;
4465 const struct sched_class *prev_class = p->sched_class;
4467 BUG_ON(prio < 0 || prio > MAX_PRIO);
4469 rq = task_rq_lock(p, &flags);
4470 update_rq_clock(rq);
4473 on_rq = p->se.on_rq;
4474 running = task_current(rq, p);
4476 dequeue_task(rq, p, 0);
4478 p->sched_class->put_prev_task(rq, p);
4481 p->sched_class = &rt_sched_class;
4483 p->sched_class = &fair_sched_class;
4488 p->sched_class->set_curr_task(rq);
4490 enqueue_task(rq, p, 0);
4492 check_class_changed(rq, p, prev_class, oldprio, running);
4494 task_rq_unlock(rq, &flags);
4499 void set_user_nice(struct task_struct *p, long nice)
4501 int old_prio, delta, on_rq;
4502 unsigned long flags;
4505 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4508 * We have to be careful, if called from sys_setpriority(),
4509 * the task might be in the middle of scheduling on another CPU.
4511 rq = task_rq_lock(p, &flags);
4512 update_rq_clock(rq);
4514 * The RT priorities are set via sched_setscheduler(), but we still
4515 * allow the 'normal' nice value to be set - but as expected
4516 * it wont have any effect on scheduling until the task is
4517 * SCHED_FIFO/SCHED_RR:
4519 if (task_has_rt_policy(p)) {
4520 p->static_prio = NICE_TO_PRIO(nice);
4523 on_rq = p->se.on_rq;
4525 dequeue_task(rq, p, 0);
4529 p->static_prio = NICE_TO_PRIO(nice);
4532 p->prio = effective_prio(p);
4533 delta = p->prio - old_prio;
4536 enqueue_task(rq, p, 0);
4539 * If the task increased its priority or is running and
4540 * lowered its priority, then reschedule its CPU:
4542 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4543 resched_task(rq->curr);
4546 task_rq_unlock(rq, &flags);
4548 EXPORT_SYMBOL(set_user_nice);
4551 * can_nice - check if a task can reduce its nice value
4555 int can_nice(const struct task_struct *p, const int nice)
4557 /* convert nice value [19,-20] to rlimit style value [1,40] */
4558 int nice_rlim = 20 - nice;
4560 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4561 capable(CAP_SYS_NICE));
4564 #ifdef __ARCH_WANT_SYS_NICE
4567 * sys_nice - change the priority of the current process.
4568 * @increment: priority increment
4570 * sys_setpriority is a more generic, but much slower function that
4571 * does similar things.
4573 asmlinkage long sys_nice(int increment)
4578 * Setpriority might change our priority at the same moment.
4579 * We don't have to worry. Conceptually one call occurs first
4580 * and we have a single winner.
4582 if (increment < -40)
4587 nice = PRIO_TO_NICE(current->static_prio) + increment;
4593 if (increment < 0 && !can_nice(current, nice))
4596 retval = security_task_setnice(current, nice);
4600 set_user_nice(current, nice);
4607 * task_prio - return the priority value of a given task.
4608 * @p: the task in question.
4610 * This is the priority value as seen by users in /proc.
4611 * RT tasks are offset by -200. Normal tasks are centered
4612 * around 0, value goes from -16 to +15.
4614 int task_prio(const struct task_struct *p)
4616 return p->prio - MAX_RT_PRIO;
4620 * task_nice - return the nice value of a given task.
4621 * @p: the task in question.
4623 int task_nice(const struct task_struct *p)
4625 return TASK_NICE(p);
4627 EXPORT_SYMBOL(task_nice);
4630 * idle_cpu - is a given cpu idle currently?
4631 * @cpu: the processor in question.
4633 int idle_cpu(int cpu)
4635 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4639 * idle_task - return the idle task for a given cpu.
4640 * @cpu: the processor in question.
4642 struct task_struct *idle_task(int cpu)
4644 return cpu_rq(cpu)->idle;
4648 * find_process_by_pid - find a process with a matching PID value.
4649 * @pid: the pid in question.
4651 static struct task_struct *find_process_by_pid(pid_t pid)
4653 return pid ? find_task_by_vpid(pid) : current;
4656 /* Actually do priority change: must hold rq lock. */
4658 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4660 BUG_ON(p->se.on_rq);
4663 switch (p->policy) {
4667 p->sched_class = &fair_sched_class;
4671 p->sched_class = &rt_sched_class;
4675 p->rt_priority = prio;
4676 p->normal_prio = normal_prio(p);
4677 /* we are holding p->pi_lock already */
4678 p->prio = rt_mutex_getprio(p);
4683 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4684 * @p: the task in question.
4685 * @policy: new policy.
4686 * @param: structure containing the new RT priority.
4688 * NOTE that the task may be already dead.
4690 int sched_setscheduler(struct task_struct *p, int policy,
4691 struct sched_param *param)
4693 int retval, oldprio, oldpolicy = -1, on_rq, running;
4694 unsigned long flags;
4695 const struct sched_class *prev_class = p->sched_class;
4698 /* may grab non-irq protected spin_locks */
4699 BUG_ON(in_interrupt());
4701 /* double check policy once rq lock held */
4703 policy = oldpolicy = p->policy;
4704 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4705 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4706 policy != SCHED_IDLE)
4709 * Valid priorities for SCHED_FIFO and SCHED_RR are
4710 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4711 * SCHED_BATCH and SCHED_IDLE is 0.
4713 if (param->sched_priority < 0 ||
4714 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4715 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4717 if (rt_policy(policy) != (param->sched_priority != 0))
4721 * Allow unprivileged RT tasks to decrease priority:
4723 if (!capable(CAP_SYS_NICE)) {
4724 if (rt_policy(policy)) {
4725 unsigned long rlim_rtprio;
4727 if (!lock_task_sighand(p, &flags))
4729 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4730 unlock_task_sighand(p, &flags);
4732 /* can't set/change the rt policy */
4733 if (policy != p->policy && !rlim_rtprio)
4736 /* can't increase priority */
4737 if (param->sched_priority > p->rt_priority &&
4738 param->sched_priority > rlim_rtprio)
4742 * Like positive nice levels, dont allow tasks to
4743 * move out of SCHED_IDLE either:
4745 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4748 /* can't change other user's priorities */
4749 if ((current->euid != p->euid) &&
4750 (current->euid != p->uid))
4754 #ifdef CONFIG_RT_GROUP_SCHED
4756 * Do not allow realtime tasks into groups that have no runtime
4759 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4763 retval = security_task_setscheduler(p, policy, param);
4767 * make sure no PI-waiters arrive (or leave) while we are
4768 * changing the priority of the task:
4770 spin_lock_irqsave(&p->pi_lock, flags);
4772 * To be able to change p->policy safely, the apropriate
4773 * runqueue lock must be held.
4775 rq = __task_rq_lock(p);
4776 /* recheck policy now with rq lock held */
4777 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4778 policy = oldpolicy = -1;
4779 __task_rq_unlock(rq);
4780 spin_unlock_irqrestore(&p->pi_lock, flags);
4783 update_rq_clock(rq);
4784 on_rq = p->se.on_rq;
4785 running = task_current(rq, p);
4787 deactivate_task(rq, p, 0);
4789 p->sched_class->put_prev_task(rq, p);
4792 __setscheduler(rq, p, policy, param->sched_priority);
4795 p->sched_class->set_curr_task(rq);
4797 activate_task(rq, p, 0);
4799 check_class_changed(rq, p, prev_class, oldprio, running);
4801 __task_rq_unlock(rq);
4802 spin_unlock_irqrestore(&p->pi_lock, flags);
4804 rt_mutex_adjust_pi(p);
4808 EXPORT_SYMBOL_GPL(sched_setscheduler);
4811 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4813 struct sched_param lparam;
4814 struct task_struct *p;
4817 if (!param || pid < 0)
4819 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4824 p = find_process_by_pid(pid);
4826 retval = sched_setscheduler(p, policy, &lparam);
4833 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4834 * @pid: the pid in question.
4835 * @policy: new policy.
4836 * @param: structure containing the new RT priority.
4839 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4841 /* negative values for policy are not valid */
4845 return do_sched_setscheduler(pid, policy, param);
4849 * sys_sched_setparam - set/change the RT priority of a thread
4850 * @pid: the pid in question.
4851 * @param: structure containing the new RT priority.
4853 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4855 return do_sched_setscheduler(pid, -1, param);
4859 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4860 * @pid: the pid in question.
4862 asmlinkage long sys_sched_getscheduler(pid_t pid)
4864 struct task_struct *p;
4871 read_lock(&tasklist_lock);
4872 p = find_process_by_pid(pid);
4874 retval = security_task_getscheduler(p);
4878 read_unlock(&tasklist_lock);
4883 * sys_sched_getscheduler - get the RT priority of a thread
4884 * @pid: the pid in question.
4885 * @param: structure containing the RT priority.
4887 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4889 struct sched_param lp;
4890 struct task_struct *p;
4893 if (!param || pid < 0)
4896 read_lock(&tasklist_lock);
4897 p = find_process_by_pid(pid);
4902 retval = security_task_getscheduler(p);
4906 lp.sched_priority = p->rt_priority;
4907 read_unlock(&tasklist_lock);
4910 * This one might sleep, we cannot do it with a spinlock held ...
4912 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4917 read_unlock(&tasklist_lock);
4921 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4923 cpumask_t cpus_allowed;
4924 cpumask_t new_mask = *in_mask;
4925 struct task_struct *p;
4929 read_lock(&tasklist_lock);
4931 p = find_process_by_pid(pid);
4933 read_unlock(&tasklist_lock);
4939 * It is not safe to call set_cpus_allowed with the
4940 * tasklist_lock held. We will bump the task_struct's
4941 * usage count and then drop tasklist_lock.
4944 read_unlock(&tasklist_lock);
4947 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4948 !capable(CAP_SYS_NICE))
4951 retval = security_task_setscheduler(p, 0, NULL);
4955 cpuset_cpus_allowed(p, &cpus_allowed);
4956 cpus_and(new_mask, new_mask, cpus_allowed);
4958 retval = set_cpus_allowed_ptr(p, &new_mask);
4961 cpuset_cpus_allowed(p, &cpus_allowed);
4962 if (!cpus_subset(new_mask, cpus_allowed)) {
4964 * We must have raced with a concurrent cpuset
4965 * update. Just reset the cpus_allowed to the
4966 * cpuset's cpus_allowed
4968 new_mask = cpus_allowed;
4978 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4979 cpumask_t *new_mask)
4981 if (len < sizeof(cpumask_t)) {
4982 memset(new_mask, 0, sizeof(cpumask_t));
4983 } else if (len > sizeof(cpumask_t)) {
4984 len = sizeof(cpumask_t);
4986 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4990 * sys_sched_setaffinity - set the cpu affinity of a process
4991 * @pid: pid of the process
4992 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4993 * @user_mask_ptr: user-space pointer to the new cpu mask
4995 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4996 unsigned long __user *user_mask_ptr)
5001 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5005 return sched_setaffinity(pid, &new_mask);
5009 * Represents all cpu's present in the system
5010 * In systems capable of hotplug, this map could dynamically grow
5011 * as new cpu's are detected in the system via any platform specific
5012 * method, such as ACPI for e.g.
5015 cpumask_t cpu_present_map __read_mostly;
5016 EXPORT_SYMBOL(cpu_present_map);
5019 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5020 EXPORT_SYMBOL(cpu_online_map);
5022 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5023 EXPORT_SYMBOL(cpu_possible_map);
5026 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5028 struct task_struct *p;
5032 read_lock(&tasklist_lock);
5035 p = find_process_by_pid(pid);
5039 retval = security_task_getscheduler(p);
5043 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5046 read_unlock(&tasklist_lock);
5053 * sys_sched_getaffinity - get the cpu affinity of a process
5054 * @pid: pid of the process
5055 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5056 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5058 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5059 unsigned long __user *user_mask_ptr)
5064 if (len < sizeof(cpumask_t))
5067 ret = sched_getaffinity(pid, &mask);
5071 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5074 return sizeof(cpumask_t);
5078 * sys_sched_yield - yield the current processor to other threads.
5080 * This function yields the current CPU to other tasks. If there are no
5081 * other threads running on this CPU then this function will return.
5083 asmlinkage long sys_sched_yield(void)
5085 struct rq *rq = this_rq_lock();
5087 schedstat_inc(rq, yld_count);
5088 current->sched_class->yield_task(rq);
5091 * Since we are going to call schedule() anyway, there's
5092 * no need to preempt or enable interrupts:
5094 __release(rq->lock);
5095 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5096 _raw_spin_unlock(&rq->lock);
5097 preempt_enable_no_resched();
5104 static void __cond_resched(void)
5106 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5107 __might_sleep(__FILE__, __LINE__);
5110 * The BKS might be reacquired before we have dropped
5111 * PREEMPT_ACTIVE, which could trigger a second
5112 * cond_resched() call.
5115 add_preempt_count(PREEMPT_ACTIVE);
5117 sub_preempt_count(PREEMPT_ACTIVE);
5118 } while (need_resched());
5121 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5122 int __sched _cond_resched(void)
5124 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5125 system_state == SYSTEM_RUNNING) {
5131 EXPORT_SYMBOL(_cond_resched);
5135 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5136 * call schedule, and on return reacquire the lock.
5138 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5139 * operations here to prevent schedule() from being called twice (once via
5140 * spin_unlock(), once by hand).
5142 int cond_resched_lock(spinlock_t *lock)
5144 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5147 if (spin_needbreak(lock) || resched) {
5149 if (resched && need_resched())
5158 EXPORT_SYMBOL(cond_resched_lock);
5160 int __sched cond_resched_softirq(void)
5162 BUG_ON(!in_softirq());
5164 if (need_resched() && system_state == SYSTEM_RUNNING) {
5172 EXPORT_SYMBOL(cond_resched_softirq);
5175 * yield - yield the current processor to other threads.
5177 * This is a shortcut for kernel-space yielding - it marks the
5178 * thread runnable and calls sys_sched_yield().
5180 void __sched yield(void)
5182 set_current_state(TASK_RUNNING);
5185 EXPORT_SYMBOL(yield);
5188 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5189 * that process accounting knows that this is a task in IO wait state.
5191 * But don't do that if it is a deliberate, throttling IO wait (this task
5192 * has set its backing_dev_info: the queue against which it should throttle)
5194 void __sched io_schedule(void)
5196 struct rq *rq = &__raw_get_cpu_var(runqueues);
5198 delayacct_blkio_start();
5199 atomic_inc(&rq->nr_iowait);
5201 atomic_dec(&rq->nr_iowait);
5202 delayacct_blkio_end();
5204 EXPORT_SYMBOL(io_schedule);
5206 long __sched io_schedule_timeout(long timeout)
5208 struct rq *rq = &__raw_get_cpu_var(runqueues);
5211 delayacct_blkio_start();
5212 atomic_inc(&rq->nr_iowait);
5213 ret = schedule_timeout(timeout);
5214 atomic_dec(&rq->nr_iowait);
5215 delayacct_blkio_end();
5220 * sys_sched_get_priority_max - return maximum RT priority.
5221 * @policy: scheduling class.
5223 * this syscall returns the maximum rt_priority that can be used
5224 * by a given scheduling class.
5226 asmlinkage long sys_sched_get_priority_max(int policy)
5233 ret = MAX_USER_RT_PRIO-1;
5245 * sys_sched_get_priority_min - return minimum RT priority.
5246 * @policy: scheduling class.
5248 * this syscall returns the minimum rt_priority that can be used
5249 * by a given scheduling class.
5251 asmlinkage long sys_sched_get_priority_min(int policy)
5269 * sys_sched_rr_get_interval - return the default timeslice of a process.
5270 * @pid: pid of the process.
5271 * @interval: userspace pointer to the timeslice value.
5273 * this syscall writes the default timeslice value of a given process
5274 * into the user-space timespec buffer. A value of '0' means infinity.
5277 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5279 struct task_struct *p;
5280 unsigned int time_slice;
5288 read_lock(&tasklist_lock);
5289 p = find_process_by_pid(pid);
5293 retval = security_task_getscheduler(p);
5298 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5299 * tasks that are on an otherwise idle runqueue:
5302 if (p->policy == SCHED_RR) {
5303 time_slice = DEF_TIMESLICE;
5304 } else if (p->policy != SCHED_FIFO) {
5305 struct sched_entity *se = &p->se;
5306 unsigned long flags;
5309 rq = task_rq_lock(p, &flags);
5310 if (rq->cfs.load.weight)
5311 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5312 task_rq_unlock(rq, &flags);
5314 read_unlock(&tasklist_lock);
5315 jiffies_to_timespec(time_slice, &t);
5316 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5320 read_unlock(&tasklist_lock);
5324 static const char stat_nam[] = "RSDTtZX";
5326 void sched_show_task(struct task_struct *p)
5328 unsigned long free = 0;
5331 state = p->state ? __ffs(p->state) + 1 : 0;
5332 printk(KERN_INFO "%-13.13s %c", p->comm,
5333 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5334 #if BITS_PER_LONG == 32
5335 if (state == TASK_RUNNING)
5336 printk(KERN_CONT " running ");
5338 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5340 if (state == TASK_RUNNING)
5341 printk(KERN_CONT " running task ");
5343 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5345 #ifdef CONFIG_DEBUG_STACK_USAGE
5347 unsigned long *n = end_of_stack(p);
5350 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5353 printk(KERN_CONT "%5lu %5d %6d\n", free,
5354 task_pid_nr(p), task_pid_nr(p->real_parent));
5356 show_stack(p, NULL);
5359 void show_state_filter(unsigned long state_filter)
5361 struct task_struct *g, *p;
5363 #if BITS_PER_LONG == 32
5365 " task PC stack pid father\n");
5368 " task PC stack pid father\n");
5370 read_lock(&tasklist_lock);
5371 do_each_thread(g, p) {
5373 * reset the NMI-timeout, listing all files on a slow
5374 * console might take alot of time:
5376 touch_nmi_watchdog();
5377 if (!state_filter || (p->state & state_filter))
5379 } while_each_thread(g, p);
5381 touch_all_softlockup_watchdogs();
5383 #ifdef CONFIG_SCHED_DEBUG
5384 sysrq_sched_debug_show();
5386 read_unlock(&tasklist_lock);
5388 * Only show locks if all tasks are dumped:
5390 if (state_filter == -1)
5391 debug_show_all_locks();
5394 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5396 idle->sched_class = &idle_sched_class;
5400 * init_idle - set up an idle thread for a given CPU
5401 * @idle: task in question
5402 * @cpu: cpu the idle task belongs to
5404 * NOTE: this function does not set the idle thread's NEED_RESCHED
5405 * flag, to make booting more robust.
5407 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5409 struct rq *rq = cpu_rq(cpu);
5410 unsigned long flags;
5413 idle->se.exec_start = sched_clock();
5415 idle->prio = idle->normal_prio = MAX_PRIO;
5416 idle->cpus_allowed = cpumask_of_cpu(cpu);
5417 __set_task_cpu(idle, cpu);
5419 spin_lock_irqsave(&rq->lock, flags);
5420 rq->curr = rq->idle = idle;
5421 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5424 spin_unlock_irqrestore(&rq->lock, flags);
5426 /* Set the preempt count _outside_ the spinlocks! */
5427 task_thread_info(idle)->preempt_count = 0;
5430 * The idle tasks have their own, simple scheduling class:
5432 idle->sched_class = &idle_sched_class;
5436 * In a system that switches off the HZ timer nohz_cpu_mask
5437 * indicates which cpus entered this state. This is used
5438 * in the rcu update to wait only for active cpus. For system
5439 * which do not switch off the HZ timer nohz_cpu_mask should
5440 * always be CPU_MASK_NONE.
5442 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5445 * Increase the granularity value when there are more CPUs,
5446 * because with more CPUs the 'effective latency' as visible
5447 * to users decreases. But the relationship is not linear,
5448 * so pick a second-best guess by going with the log2 of the
5451 * This idea comes from the SD scheduler of Con Kolivas:
5453 static inline void sched_init_granularity(void)
5455 unsigned int factor = 1 + ilog2(num_online_cpus());
5456 const unsigned long limit = 200000000;
5458 sysctl_sched_min_granularity *= factor;
5459 if (sysctl_sched_min_granularity > limit)
5460 sysctl_sched_min_granularity = limit;
5462 sysctl_sched_latency *= factor;
5463 if (sysctl_sched_latency > limit)
5464 sysctl_sched_latency = limit;
5466 sysctl_sched_wakeup_granularity *= factor;
5471 * This is how migration works:
5473 * 1) we queue a struct migration_req structure in the source CPU's
5474 * runqueue and wake up that CPU's migration thread.
5475 * 2) we down() the locked semaphore => thread blocks.
5476 * 3) migration thread wakes up (implicitly it forces the migrated
5477 * thread off the CPU)
5478 * 4) it gets the migration request and checks whether the migrated
5479 * task is still in the wrong runqueue.
5480 * 5) if it's in the wrong runqueue then the migration thread removes
5481 * it and puts it into the right queue.
5482 * 6) migration thread up()s the semaphore.
5483 * 7) we wake up and the migration is done.
5487 * Change a given task's CPU affinity. Migrate the thread to a
5488 * proper CPU and schedule it away if the CPU it's executing on
5489 * is removed from the allowed bitmask.
5491 * NOTE: the caller must have a valid reference to the task, the
5492 * task must not exit() & deallocate itself prematurely. The
5493 * call is not atomic; no spinlocks may be held.
5495 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5497 struct migration_req req;
5498 unsigned long flags;
5502 rq = task_rq_lock(p, &flags);
5503 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5508 if (p->sched_class->set_cpus_allowed)
5509 p->sched_class->set_cpus_allowed(p, new_mask);
5511 p->cpus_allowed = *new_mask;
5512 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5515 /* Can the task run on the task's current CPU? If so, we're done */
5516 if (cpu_isset(task_cpu(p), *new_mask))
5519 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5520 /* Need help from migration thread: drop lock and wait. */
5521 task_rq_unlock(rq, &flags);
5522 wake_up_process(rq->migration_thread);
5523 wait_for_completion(&req.done);
5524 tlb_migrate_finish(p->mm);
5528 task_rq_unlock(rq, &flags);
5532 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5535 * Move (not current) task off this cpu, onto dest cpu. We're doing
5536 * this because either it can't run here any more (set_cpus_allowed()
5537 * away from this CPU, or CPU going down), or because we're
5538 * attempting to rebalance this task on exec (sched_exec).
5540 * So we race with normal scheduler movements, but that's OK, as long
5541 * as the task is no longer on this CPU.
5543 * Returns non-zero if task was successfully migrated.
5545 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5547 struct rq *rq_dest, *rq_src;
5550 if (unlikely(cpu_is_offline(dest_cpu)))
5553 rq_src = cpu_rq(src_cpu);
5554 rq_dest = cpu_rq(dest_cpu);
5556 double_rq_lock(rq_src, rq_dest);
5557 /* Already moved. */
5558 if (task_cpu(p) != src_cpu)
5560 /* Affinity changed (again). */
5561 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5564 on_rq = p->se.on_rq;
5566 deactivate_task(rq_src, p, 0);
5568 set_task_cpu(p, dest_cpu);
5570 activate_task(rq_dest, p, 0);
5571 check_preempt_curr(rq_dest, p);
5575 double_rq_unlock(rq_src, rq_dest);
5580 * migration_thread - this is a highprio system thread that performs
5581 * thread migration by bumping thread off CPU then 'pushing' onto
5584 static int migration_thread(void *data)
5586 int cpu = (long)data;
5590 BUG_ON(rq->migration_thread != current);
5592 set_current_state(TASK_INTERRUPTIBLE);
5593 while (!kthread_should_stop()) {
5594 struct migration_req *req;
5595 struct list_head *head;
5597 spin_lock_irq(&rq->lock);
5599 if (cpu_is_offline(cpu)) {
5600 spin_unlock_irq(&rq->lock);
5604 if (rq->active_balance) {
5605 active_load_balance(rq, cpu);
5606 rq->active_balance = 0;
5609 head = &rq->migration_queue;
5611 if (list_empty(head)) {
5612 spin_unlock_irq(&rq->lock);
5614 set_current_state(TASK_INTERRUPTIBLE);
5617 req = list_entry(head->next, struct migration_req, list);
5618 list_del_init(head->next);
5620 spin_unlock(&rq->lock);
5621 __migrate_task(req->task, cpu, req->dest_cpu);
5624 complete(&req->done);
5626 __set_current_state(TASK_RUNNING);
5630 /* Wait for kthread_stop */
5631 set_current_state(TASK_INTERRUPTIBLE);
5632 while (!kthread_should_stop()) {
5634 set_current_state(TASK_INTERRUPTIBLE);
5636 __set_current_state(TASK_RUNNING);
5640 #ifdef CONFIG_HOTPLUG_CPU
5642 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5646 local_irq_disable();
5647 ret = __migrate_task(p, src_cpu, dest_cpu);
5653 * Figure out where task on dead CPU should go, use force if necessary.
5654 * NOTE: interrupts should be disabled by the caller
5656 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5658 unsigned long flags;
5665 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5666 cpus_and(mask, mask, p->cpus_allowed);
5667 dest_cpu = any_online_cpu(mask);
5669 /* On any allowed CPU? */
5670 if (dest_cpu >= nr_cpu_ids)
5671 dest_cpu = any_online_cpu(p->cpus_allowed);
5673 /* No more Mr. Nice Guy. */
5674 if (dest_cpu >= nr_cpu_ids) {
5675 cpumask_t cpus_allowed;
5677 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5679 * Try to stay on the same cpuset, where the
5680 * current cpuset may be a subset of all cpus.
5681 * The cpuset_cpus_allowed_locked() variant of
5682 * cpuset_cpus_allowed() will not block. It must be
5683 * called within calls to cpuset_lock/cpuset_unlock.
5685 rq = task_rq_lock(p, &flags);
5686 p->cpus_allowed = cpus_allowed;
5687 dest_cpu = any_online_cpu(p->cpus_allowed);
5688 task_rq_unlock(rq, &flags);
5691 * Don't tell them about moving exiting tasks or
5692 * kernel threads (both mm NULL), since they never
5695 if (p->mm && printk_ratelimit()) {
5696 printk(KERN_INFO "process %d (%s) no "
5697 "longer affine to cpu%d\n",
5698 task_pid_nr(p), p->comm, dead_cpu);
5701 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5705 * While a dead CPU has no uninterruptible tasks queued at this point,
5706 * it might still have a nonzero ->nr_uninterruptible counter, because
5707 * for performance reasons the counter is not stricly tracking tasks to
5708 * their home CPUs. So we just add the counter to another CPU's counter,
5709 * to keep the global sum constant after CPU-down:
5711 static void migrate_nr_uninterruptible(struct rq *rq_src)
5713 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5714 unsigned long flags;
5716 local_irq_save(flags);
5717 double_rq_lock(rq_src, rq_dest);
5718 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5719 rq_src->nr_uninterruptible = 0;
5720 double_rq_unlock(rq_src, rq_dest);
5721 local_irq_restore(flags);
5724 /* Run through task list and migrate tasks from the dead cpu. */
5725 static void migrate_live_tasks(int src_cpu)
5727 struct task_struct *p, *t;
5729 read_lock(&tasklist_lock);
5731 do_each_thread(t, p) {
5735 if (task_cpu(p) == src_cpu)
5736 move_task_off_dead_cpu(src_cpu, p);
5737 } while_each_thread(t, p);
5739 read_unlock(&tasklist_lock);
5743 * Schedules idle task to be the next runnable task on current CPU.
5744 * It does so by boosting its priority to highest possible.
5745 * Used by CPU offline code.
5747 void sched_idle_next(void)
5749 int this_cpu = smp_processor_id();
5750 struct rq *rq = cpu_rq(this_cpu);
5751 struct task_struct *p = rq->idle;
5752 unsigned long flags;
5754 /* cpu has to be offline */
5755 BUG_ON(cpu_online(this_cpu));
5758 * Strictly not necessary since rest of the CPUs are stopped by now
5759 * and interrupts disabled on the current cpu.
5761 spin_lock_irqsave(&rq->lock, flags);
5763 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5765 update_rq_clock(rq);
5766 activate_task(rq, p, 0);
5768 spin_unlock_irqrestore(&rq->lock, flags);
5772 * Ensures that the idle task is using init_mm right before its cpu goes
5775 void idle_task_exit(void)
5777 struct mm_struct *mm = current->active_mm;
5779 BUG_ON(cpu_online(smp_processor_id()));
5782 switch_mm(mm, &init_mm, current);
5786 /* called under rq->lock with disabled interrupts */
5787 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5789 struct rq *rq = cpu_rq(dead_cpu);
5791 /* Must be exiting, otherwise would be on tasklist. */
5792 BUG_ON(!p->exit_state);
5794 /* Cannot have done final schedule yet: would have vanished. */
5795 BUG_ON(p->state == TASK_DEAD);
5800 * Drop lock around migration; if someone else moves it,
5801 * that's OK. No task can be added to this CPU, so iteration is
5804 spin_unlock_irq(&rq->lock);
5805 move_task_off_dead_cpu(dead_cpu, p);
5806 spin_lock_irq(&rq->lock);
5811 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5812 static void migrate_dead_tasks(unsigned int dead_cpu)
5814 struct rq *rq = cpu_rq(dead_cpu);
5815 struct task_struct *next;
5818 if (!rq->nr_running)
5820 update_rq_clock(rq);
5821 next = pick_next_task(rq, rq->curr);
5824 migrate_dead(dead_cpu, next);
5828 #endif /* CONFIG_HOTPLUG_CPU */
5830 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5832 static struct ctl_table sd_ctl_dir[] = {
5834 .procname = "sched_domain",
5840 static struct ctl_table sd_ctl_root[] = {
5842 .ctl_name = CTL_KERN,
5843 .procname = "kernel",
5845 .child = sd_ctl_dir,
5850 static struct ctl_table *sd_alloc_ctl_entry(int n)
5852 struct ctl_table *entry =
5853 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5858 static void sd_free_ctl_entry(struct ctl_table **tablep)
5860 struct ctl_table *entry;
5863 * In the intermediate directories, both the child directory and
5864 * procname are dynamically allocated and could fail but the mode
5865 * will always be set. In the lowest directory the names are
5866 * static strings and all have proc handlers.
5868 for (entry = *tablep; entry->mode; entry++) {
5870 sd_free_ctl_entry(&entry->child);
5871 if (entry->proc_handler == NULL)
5872 kfree(entry->procname);
5880 set_table_entry(struct ctl_table *entry,
5881 const char *procname, void *data, int maxlen,
5882 mode_t mode, proc_handler *proc_handler)
5884 entry->procname = procname;
5886 entry->maxlen = maxlen;
5888 entry->proc_handler = proc_handler;
5891 static struct ctl_table *
5892 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5894 struct ctl_table *table = sd_alloc_ctl_entry(12);
5899 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5900 sizeof(long), 0644, proc_doulongvec_minmax);
5901 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5902 sizeof(long), 0644, proc_doulongvec_minmax);
5903 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5904 sizeof(int), 0644, proc_dointvec_minmax);
5905 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5906 sizeof(int), 0644, proc_dointvec_minmax);
5907 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5908 sizeof(int), 0644, proc_dointvec_minmax);
5909 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5910 sizeof(int), 0644, proc_dointvec_minmax);
5911 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5912 sizeof(int), 0644, proc_dointvec_minmax);
5913 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5914 sizeof(int), 0644, proc_dointvec_minmax);
5915 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5916 sizeof(int), 0644, proc_dointvec_minmax);
5917 set_table_entry(&table[9], "cache_nice_tries",
5918 &sd->cache_nice_tries,
5919 sizeof(int), 0644, proc_dointvec_minmax);
5920 set_table_entry(&table[10], "flags", &sd->flags,
5921 sizeof(int), 0644, proc_dointvec_minmax);
5922 /* &table[11] is terminator */
5927 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5929 struct ctl_table *entry, *table;
5930 struct sched_domain *sd;
5931 int domain_num = 0, i;
5934 for_each_domain(cpu, sd)
5936 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5941 for_each_domain(cpu, sd) {
5942 snprintf(buf, 32, "domain%d", i);
5943 entry->procname = kstrdup(buf, GFP_KERNEL);
5945 entry->child = sd_alloc_ctl_domain_table(sd);
5952 static struct ctl_table_header *sd_sysctl_header;
5953 static void register_sched_domain_sysctl(void)
5955 int i, cpu_num = num_online_cpus();
5956 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5959 WARN_ON(sd_ctl_dir[0].child);
5960 sd_ctl_dir[0].child = entry;
5965 for_each_online_cpu(i) {
5966 snprintf(buf, 32, "cpu%d", i);
5967 entry->procname = kstrdup(buf, GFP_KERNEL);
5969 entry->child = sd_alloc_ctl_cpu_table(i);
5973 WARN_ON(sd_sysctl_header);
5974 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5977 /* may be called multiple times per register */
5978 static void unregister_sched_domain_sysctl(void)
5980 if (sd_sysctl_header)
5981 unregister_sysctl_table(sd_sysctl_header);
5982 sd_sysctl_header = NULL;
5983 if (sd_ctl_dir[0].child)
5984 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5987 static void register_sched_domain_sysctl(void)
5990 static void unregister_sched_domain_sysctl(void)
5996 * migration_call - callback that gets triggered when a CPU is added.
5997 * Here we can start up the necessary migration thread for the new CPU.
5999 static int __cpuinit
6000 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6002 struct task_struct *p;
6003 int cpu = (long)hcpu;
6004 unsigned long flags;
6009 case CPU_UP_PREPARE:
6010 case CPU_UP_PREPARE_FROZEN:
6011 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6014 kthread_bind(p, cpu);
6015 /* Must be high prio: stop_machine expects to yield to it. */
6016 rq = task_rq_lock(p, &flags);
6017 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6018 task_rq_unlock(rq, &flags);
6019 cpu_rq(cpu)->migration_thread = p;
6023 case CPU_ONLINE_FROZEN:
6024 /* Strictly unnecessary, as first user will wake it. */
6025 wake_up_process(cpu_rq(cpu)->migration_thread);
6027 /* Update our root-domain */
6029 spin_lock_irqsave(&rq->lock, flags);
6031 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6032 cpu_set(cpu, rq->rd->online);
6034 spin_unlock_irqrestore(&rq->lock, flags);
6037 #ifdef CONFIG_HOTPLUG_CPU
6038 case CPU_UP_CANCELED:
6039 case CPU_UP_CANCELED_FROZEN:
6040 if (!cpu_rq(cpu)->migration_thread)
6042 /* Unbind it from offline cpu so it can run. Fall thru. */
6043 kthread_bind(cpu_rq(cpu)->migration_thread,
6044 any_online_cpu(cpu_online_map));
6045 kthread_stop(cpu_rq(cpu)->migration_thread);
6046 cpu_rq(cpu)->migration_thread = NULL;
6050 case CPU_DEAD_FROZEN:
6051 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6052 migrate_live_tasks(cpu);
6054 kthread_stop(rq->migration_thread);
6055 rq->migration_thread = NULL;
6056 /* Idle task back to normal (off runqueue, low prio) */
6057 spin_lock_irq(&rq->lock);
6058 update_rq_clock(rq);
6059 deactivate_task(rq, rq->idle, 0);
6060 rq->idle->static_prio = MAX_PRIO;
6061 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6062 rq->idle->sched_class = &idle_sched_class;
6063 migrate_dead_tasks(cpu);
6064 spin_unlock_irq(&rq->lock);
6066 migrate_nr_uninterruptible(rq);
6067 BUG_ON(rq->nr_running != 0);
6070 * No need to migrate the tasks: it was best-effort if
6071 * they didn't take sched_hotcpu_mutex. Just wake up
6074 spin_lock_irq(&rq->lock);
6075 while (!list_empty(&rq->migration_queue)) {
6076 struct migration_req *req;
6078 req = list_entry(rq->migration_queue.next,
6079 struct migration_req, list);
6080 list_del_init(&req->list);
6081 complete(&req->done);
6083 spin_unlock_irq(&rq->lock);
6087 case CPU_DYING_FROZEN:
6088 /* Update our root-domain */
6090 spin_lock_irqsave(&rq->lock, flags);
6092 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6093 cpu_clear(cpu, rq->rd->online);
6095 spin_unlock_irqrestore(&rq->lock, flags);
6102 /* Register at highest priority so that task migration (migrate_all_tasks)
6103 * happens before everything else.
6105 static struct notifier_block __cpuinitdata migration_notifier = {
6106 .notifier_call = migration_call,
6110 void __init migration_init(void)
6112 void *cpu = (void *)(long)smp_processor_id();
6115 /* Start one for the boot CPU: */
6116 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6117 BUG_ON(err == NOTIFY_BAD);
6118 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6119 register_cpu_notifier(&migration_notifier);
6125 #ifdef CONFIG_SCHED_DEBUG
6127 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6128 cpumask_t *groupmask)
6130 struct sched_group *group = sd->groups;
6133 cpulist_scnprintf(str, sizeof(str), sd->span);
6134 cpus_clear(*groupmask);
6136 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6138 if (!(sd->flags & SD_LOAD_BALANCE)) {
6139 printk("does not load-balance\n");
6141 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6146 printk(KERN_CONT "span %s\n", str);
6148 if (!cpu_isset(cpu, sd->span)) {
6149 printk(KERN_ERR "ERROR: domain->span does not contain "
6152 if (!cpu_isset(cpu, group->cpumask)) {
6153 printk(KERN_ERR "ERROR: domain->groups does not contain"
6157 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6161 printk(KERN_ERR "ERROR: group is NULL\n");
6165 if (!group->__cpu_power) {
6166 printk(KERN_CONT "\n");
6167 printk(KERN_ERR "ERROR: domain->cpu_power not "
6172 if (!cpus_weight(group->cpumask)) {
6173 printk(KERN_CONT "\n");
6174 printk(KERN_ERR "ERROR: empty group\n");
6178 if (cpus_intersects(*groupmask, group->cpumask)) {
6179 printk(KERN_CONT "\n");
6180 printk(KERN_ERR "ERROR: repeated CPUs\n");
6184 cpus_or(*groupmask, *groupmask, group->cpumask);
6186 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6187 printk(KERN_CONT " %s", str);
6189 group = group->next;
6190 } while (group != sd->groups);
6191 printk(KERN_CONT "\n");
6193 if (!cpus_equal(sd->span, *groupmask))
6194 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6196 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6197 printk(KERN_ERR "ERROR: parent span is not a superset "
6198 "of domain->span\n");
6202 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6204 cpumask_t *groupmask;
6208 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6212 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6214 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6216 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6221 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6231 # define sched_domain_debug(sd, cpu) do { } while (0)
6234 static int sd_degenerate(struct sched_domain *sd)
6236 if (cpus_weight(sd->span) == 1)
6239 /* Following flags need at least 2 groups */
6240 if (sd->flags & (SD_LOAD_BALANCE |
6241 SD_BALANCE_NEWIDLE |
6245 SD_SHARE_PKG_RESOURCES)) {
6246 if (sd->groups != sd->groups->next)
6250 /* Following flags don't use groups */
6251 if (sd->flags & (SD_WAKE_IDLE |
6260 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6262 unsigned long cflags = sd->flags, pflags = parent->flags;
6264 if (sd_degenerate(parent))
6267 if (!cpus_equal(sd->span, parent->span))
6270 /* Does parent contain flags not in child? */
6271 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6272 if (cflags & SD_WAKE_AFFINE)
6273 pflags &= ~SD_WAKE_BALANCE;
6274 /* Flags needing groups don't count if only 1 group in parent */
6275 if (parent->groups == parent->groups->next) {
6276 pflags &= ~(SD_LOAD_BALANCE |
6277 SD_BALANCE_NEWIDLE |
6281 SD_SHARE_PKG_RESOURCES);
6283 if (~cflags & pflags)
6289 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6291 unsigned long flags;
6292 const struct sched_class *class;
6294 spin_lock_irqsave(&rq->lock, flags);
6297 struct root_domain *old_rd = rq->rd;
6299 for (class = sched_class_highest; class; class = class->next) {
6300 if (class->leave_domain)
6301 class->leave_domain(rq);
6304 cpu_clear(rq->cpu, old_rd->span);
6305 cpu_clear(rq->cpu, old_rd->online);
6307 if (atomic_dec_and_test(&old_rd->refcount))
6311 atomic_inc(&rd->refcount);
6314 cpu_set(rq->cpu, rd->span);
6315 if (cpu_isset(rq->cpu, cpu_online_map))
6316 cpu_set(rq->cpu, rd->online);
6318 for (class = sched_class_highest; class; class = class->next) {
6319 if (class->join_domain)
6320 class->join_domain(rq);
6323 spin_unlock_irqrestore(&rq->lock, flags);
6326 static void init_rootdomain(struct root_domain *rd)
6328 memset(rd, 0, sizeof(*rd));
6330 cpus_clear(rd->span);
6331 cpus_clear(rd->online);
6334 static void init_defrootdomain(void)
6336 init_rootdomain(&def_root_domain);
6337 atomic_set(&def_root_domain.refcount, 1);
6340 static struct root_domain *alloc_rootdomain(void)
6342 struct root_domain *rd;
6344 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6348 init_rootdomain(rd);
6354 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6355 * hold the hotplug lock.
6358 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6360 struct rq *rq = cpu_rq(cpu);
6361 struct sched_domain *tmp;
6363 /* Remove the sched domains which do not contribute to scheduling. */
6364 for (tmp = sd; tmp; tmp = tmp->parent) {
6365 struct sched_domain *parent = tmp->parent;
6368 if (sd_parent_degenerate(tmp, parent)) {
6369 tmp->parent = parent->parent;
6371 parent->parent->child = tmp;
6375 if (sd && sd_degenerate(sd)) {
6381 sched_domain_debug(sd, cpu);
6383 rq_attach_root(rq, rd);
6384 rcu_assign_pointer(rq->sd, sd);
6387 /* cpus with isolated domains */
6388 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6390 /* Setup the mask of cpus configured for isolated domains */
6391 static int __init isolated_cpu_setup(char *str)
6393 int ints[NR_CPUS], i;
6395 str = get_options(str, ARRAY_SIZE(ints), ints);
6396 cpus_clear(cpu_isolated_map);
6397 for (i = 1; i <= ints[0]; i++)
6398 if (ints[i] < NR_CPUS)
6399 cpu_set(ints[i], cpu_isolated_map);
6403 __setup("isolcpus=", isolated_cpu_setup);
6406 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6407 * to a function which identifies what group(along with sched group) a CPU
6408 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6409 * (due to the fact that we keep track of groups covered with a cpumask_t).
6411 * init_sched_build_groups will build a circular linked list of the groups
6412 * covered by the given span, and will set each group's ->cpumask correctly,
6413 * and ->cpu_power to 0.
6416 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6417 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6418 struct sched_group **sg,
6419 cpumask_t *tmpmask),
6420 cpumask_t *covered, cpumask_t *tmpmask)
6422 struct sched_group *first = NULL, *last = NULL;
6425 cpus_clear(*covered);
6427 for_each_cpu_mask(i, *span) {
6428 struct sched_group *sg;
6429 int group = group_fn(i, cpu_map, &sg, tmpmask);
6432 if (cpu_isset(i, *covered))
6435 cpus_clear(sg->cpumask);
6436 sg->__cpu_power = 0;
6438 for_each_cpu_mask(j, *span) {
6439 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6442 cpu_set(j, *covered);
6443 cpu_set(j, sg->cpumask);
6454 #define SD_NODES_PER_DOMAIN 16
6459 * find_next_best_node - find the next node to include in a sched_domain
6460 * @node: node whose sched_domain we're building
6461 * @used_nodes: nodes already in the sched_domain
6463 * Find the next node to include in a given scheduling domain. Simply
6464 * finds the closest node not already in the @used_nodes map.
6466 * Should use nodemask_t.
6468 static int find_next_best_node(int node, nodemask_t *used_nodes)
6470 int i, n, val, min_val, best_node = 0;
6474 for (i = 0; i < MAX_NUMNODES; i++) {
6475 /* Start at @node */
6476 n = (node + i) % MAX_NUMNODES;
6478 if (!nr_cpus_node(n))
6481 /* Skip already used nodes */
6482 if (node_isset(n, *used_nodes))
6485 /* Simple min distance search */
6486 val = node_distance(node, n);
6488 if (val < min_val) {
6494 node_set(best_node, *used_nodes);
6499 * sched_domain_node_span - get a cpumask for a node's sched_domain
6500 * @node: node whose cpumask we're constructing
6502 * Given a node, construct a good cpumask for its sched_domain to span. It
6503 * should be one that prevents unnecessary balancing, but also spreads tasks
6506 static void sched_domain_node_span(int node, cpumask_t *span)
6508 nodemask_t used_nodes;
6509 node_to_cpumask_ptr(nodemask, node);
6513 nodes_clear(used_nodes);
6515 cpus_or(*span, *span, *nodemask);
6516 node_set(node, used_nodes);
6518 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6519 int next_node = find_next_best_node(node, &used_nodes);
6521 node_to_cpumask_ptr_next(nodemask, next_node);
6522 cpus_or(*span, *span, *nodemask);
6527 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6530 * SMT sched-domains:
6532 #ifdef CONFIG_SCHED_SMT
6533 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6534 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6537 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6541 *sg = &per_cpu(sched_group_cpus, cpu);
6547 * multi-core sched-domains:
6549 #ifdef CONFIG_SCHED_MC
6550 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6551 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6554 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6556 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6561 *mask = per_cpu(cpu_sibling_map, cpu);
6562 cpus_and(*mask, *mask, *cpu_map);
6563 group = first_cpu(*mask);
6565 *sg = &per_cpu(sched_group_core, group);
6568 #elif defined(CONFIG_SCHED_MC)
6570 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6574 *sg = &per_cpu(sched_group_core, cpu);
6579 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6580 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6583 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6587 #ifdef CONFIG_SCHED_MC
6588 *mask = cpu_coregroup_map(cpu);
6589 cpus_and(*mask, *mask, *cpu_map);
6590 group = first_cpu(*mask);
6591 #elif defined(CONFIG_SCHED_SMT)
6592 *mask = per_cpu(cpu_sibling_map, cpu);
6593 cpus_and(*mask, *mask, *cpu_map);
6594 group = first_cpu(*mask);
6599 *sg = &per_cpu(sched_group_phys, group);
6605 * The init_sched_build_groups can't handle what we want to do with node
6606 * groups, so roll our own. Now each node has its own list of groups which
6607 * gets dynamically allocated.
6609 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6610 static struct sched_group ***sched_group_nodes_bycpu;
6612 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6613 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6615 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6616 struct sched_group **sg, cpumask_t *nodemask)
6620 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6621 cpus_and(*nodemask, *nodemask, *cpu_map);
6622 group = first_cpu(*nodemask);
6625 *sg = &per_cpu(sched_group_allnodes, group);
6629 static void init_numa_sched_groups_power(struct sched_group *group_head)
6631 struct sched_group *sg = group_head;
6637 for_each_cpu_mask(j, sg->cpumask) {
6638 struct sched_domain *sd;
6640 sd = &per_cpu(phys_domains, j);
6641 if (j != first_cpu(sd->groups->cpumask)) {
6643 * Only add "power" once for each
6649 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6652 } while (sg != group_head);
6657 /* Free memory allocated for various sched_group structures */
6658 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6662 for_each_cpu_mask(cpu, *cpu_map) {
6663 struct sched_group **sched_group_nodes
6664 = sched_group_nodes_bycpu[cpu];
6666 if (!sched_group_nodes)
6669 for (i = 0; i < MAX_NUMNODES; i++) {
6670 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6672 *nodemask = node_to_cpumask(i);
6673 cpus_and(*nodemask, *nodemask, *cpu_map);
6674 if (cpus_empty(*nodemask))
6684 if (oldsg != sched_group_nodes[i])
6687 kfree(sched_group_nodes);
6688 sched_group_nodes_bycpu[cpu] = NULL;
6692 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6698 * Initialize sched groups cpu_power.
6700 * cpu_power indicates the capacity of sched group, which is used while
6701 * distributing the load between different sched groups in a sched domain.
6702 * Typically cpu_power for all the groups in a sched domain will be same unless
6703 * there are asymmetries in the topology. If there are asymmetries, group
6704 * having more cpu_power will pickup more load compared to the group having
6707 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6708 * the maximum number of tasks a group can handle in the presence of other idle
6709 * or lightly loaded groups in the same sched domain.
6711 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6713 struct sched_domain *child;
6714 struct sched_group *group;
6716 WARN_ON(!sd || !sd->groups);
6718 if (cpu != first_cpu(sd->groups->cpumask))
6723 sd->groups->__cpu_power = 0;
6726 * For perf policy, if the groups in child domain share resources
6727 * (for example cores sharing some portions of the cache hierarchy
6728 * or SMT), then set this domain groups cpu_power such that each group
6729 * can handle only one task, when there are other idle groups in the
6730 * same sched domain.
6732 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6734 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6735 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6740 * add cpu_power of each child group to this groups cpu_power
6742 group = child->groups;
6744 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6745 group = group->next;
6746 } while (group != child->groups);
6750 * Initializers for schedule domains
6751 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6754 #define SD_INIT(sd, type) sd_init_##type(sd)
6755 #define SD_INIT_FUNC(type) \
6756 static noinline void sd_init_##type(struct sched_domain *sd) \
6758 memset(sd, 0, sizeof(*sd)); \
6759 *sd = SD_##type##_INIT; \
6764 SD_INIT_FUNC(ALLNODES)
6767 #ifdef CONFIG_SCHED_SMT
6768 SD_INIT_FUNC(SIBLING)
6770 #ifdef CONFIG_SCHED_MC
6775 * To minimize stack usage kmalloc room for cpumasks and share the
6776 * space as the usage in build_sched_domains() dictates. Used only
6777 * if the amount of space is significant.
6780 cpumask_t tmpmask; /* make this one first */
6783 cpumask_t this_sibling_map;
6784 cpumask_t this_core_map;
6786 cpumask_t send_covered;
6789 cpumask_t domainspan;
6791 cpumask_t notcovered;
6796 #define SCHED_CPUMASK_ALLOC 1
6797 #define SCHED_CPUMASK_FREE(v) kfree(v)
6798 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6800 #define SCHED_CPUMASK_ALLOC 0
6801 #define SCHED_CPUMASK_FREE(v)
6802 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6805 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6806 ((unsigned long)(a) + offsetof(struct allmasks, v))
6809 * Build sched domains for a given set of cpus and attach the sched domains
6810 * to the individual cpus
6812 static int build_sched_domains(const cpumask_t *cpu_map)
6815 struct root_domain *rd;
6816 SCHED_CPUMASK_DECLARE(allmasks);
6819 struct sched_group **sched_group_nodes = NULL;
6820 int sd_allnodes = 0;
6823 * Allocate the per-node list of sched groups
6825 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6827 if (!sched_group_nodes) {
6828 printk(KERN_WARNING "Can not alloc sched group node list\n");
6833 rd = alloc_rootdomain();
6835 printk(KERN_WARNING "Cannot alloc root domain\n");
6837 kfree(sched_group_nodes);
6842 #if SCHED_CPUMASK_ALLOC
6843 /* get space for all scratch cpumask variables */
6844 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6846 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6849 kfree(sched_group_nodes);
6854 tmpmask = (cpumask_t *)allmasks;
6858 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6862 * Set up domains for cpus specified by the cpu_map.
6864 for_each_cpu_mask(i, *cpu_map) {
6865 struct sched_domain *sd = NULL, *p;
6866 SCHED_CPUMASK_VAR(nodemask, allmasks);
6868 *nodemask = node_to_cpumask(cpu_to_node(i));
6869 cpus_and(*nodemask, *nodemask, *cpu_map);
6872 if (cpus_weight(*cpu_map) >
6873 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6874 sd = &per_cpu(allnodes_domains, i);
6875 SD_INIT(sd, ALLNODES);
6876 sd->span = *cpu_map;
6877 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6883 sd = &per_cpu(node_domains, i);
6885 sched_domain_node_span(cpu_to_node(i), &sd->span);
6889 cpus_and(sd->span, sd->span, *cpu_map);
6893 sd = &per_cpu(phys_domains, i);
6895 sd->span = *nodemask;
6899 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6901 #ifdef CONFIG_SCHED_MC
6903 sd = &per_cpu(core_domains, i);
6905 sd->span = cpu_coregroup_map(i);
6906 cpus_and(sd->span, sd->span, *cpu_map);
6909 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
6912 #ifdef CONFIG_SCHED_SMT
6914 sd = &per_cpu(cpu_domains, i);
6915 SD_INIT(sd, SIBLING);
6916 sd->span = per_cpu(cpu_sibling_map, i);
6917 cpus_and(sd->span, sd->span, *cpu_map);
6920 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
6924 #ifdef CONFIG_SCHED_SMT
6925 /* Set up CPU (sibling) groups */
6926 for_each_cpu_mask(i, *cpu_map) {
6927 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
6928 SCHED_CPUMASK_VAR(send_covered, allmasks);
6930 *this_sibling_map = per_cpu(cpu_sibling_map, i);
6931 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
6932 if (i != first_cpu(*this_sibling_map))
6935 init_sched_build_groups(this_sibling_map, cpu_map,
6937 send_covered, tmpmask);
6941 #ifdef CONFIG_SCHED_MC
6942 /* Set up multi-core groups */
6943 for_each_cpu_mask(i, *cpu_map) {
6944 SCHED_CPUMASK_VAR(this_core_map, allmasks);
6945 SCHED_CPUMASK_VAR(send_covered, allmasks);
6947 *this_core_map = cpu_coregroup_map(i);
6948 cpus_and(*this_core_map, *this_core_map, *cpu_map);
6949 if (i != first_cpu(*this_core_map))
6952 init_sched_build_groups(this_core_map, cpu_map,
6954 send_covered, tmpmask);
6958 /* Set up physical groups */
6959 for (i = 0; i < MAX_NUMNODES; i++) {
6960 SCHED_CPUMASK_VAR(nodemask, allmasks);
6961 SCHED_CPUMASK_VAR(send_covered, allmasks);
6963 *nodemask = node_to_cpumask(i);
6964 cpus_and(*nodemask, *nodemask, *cpu_map);
6965 if (cpus_empty(*nodemask))
6968 init_sched_build_groups(nodemask, cpu_map,
6970 send_covered, tmpmask);
6974 /* Set up node groups */
6976 SCHED_CPUMASK_VAR(send_covered, allmasks);
6978 init_sched_build_groups(cpu_map, cpu_map,
6979 &cpu_to_allnodes_group,
6980 send_covered, tmpmask);
6983 for (i = 0; i < MAX_NUMNODES; i++) {
6984 /* Set up node groups */
6985 struct sched_group *sg, *prev;
6986 SCHED_CPUMASK_VAR(nodemask, allmasks);
6987 SCHED_CPUMASK_VAR(domainspan, allmasks);
6988 SCHED_CPUMASK_VAR(covered, allmasks);
6991 *nodemask = node_to_cpumask(i);
6992 cpus_clear(*covered);
6994 cpus_and(*nodemask, *nodemask, *cpu_map);
6995 if (cpus_empty(*nodemask)) {
6996 sched_group_nodes[i] = NULL;
7000 sched_domain_node_span(i, domainspan);
7001 cpus_and(*domainspan, *domainspan, *cpu_map);
7003 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7005 printk(KERN_WARNING "Can not alloc domain group for "
7009 sched_group_nodes[i] = sg;
7010 for_each_cpu_mask(j, *nodemask) {
7011 struct sched_domain *sd;
7013 sd = &per_cpu(node_domains, j);
7016 sg->__cpu_power = 0;
7017 sg->cpumask = *nodemask;
7019 cpus_or(*covered, *covered, *nodemask);
7022 for (j = 0; j < MAX_NUMNODES; j++) {
7023 SCHED_CPUMASK_VAR(notcovered, allmasks);
7024 int n = (i + j) % MAX_NUMNODES;
7025 node_to_cpumask_ptr(pnodemask, n);
7027 cpus_complement(*notcovered, *covered);
7028 cpus_and(*tmpmask, *notcovered, *cpu_map);
7029 cpus_and(*tmpmask, *tmpmask, *domainspan);
7030 if (cpus_empty(*tmpmask))
7033 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7034 if (cpus_empty(*tmpmask))
7037 sg = kmalloc_node(sizeof(struct sched_group),
7041 "Can not alloc domain group for node %d\n", j);
7044 sg->__cpu_power = 0;
7045 sg->cpumask = *tmpmask;
7046 sg->next = prev->next;
7047 cpus_or(*covered, *covered, *tmpmask);
7054 /* Calculate CPU power for physical packages and nodes */
7055 #ifdef CONFIG_SCHED_SMT
7056 for_each_cpu_mask(i, *cpu_map) {
7057 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7059 init_sched_groups_power(i, sd);
7062 #ifdef CONFIG_SCHED_MC
7063 for_each_cpu_mask(i, *cpu_map) {
7064 struct sched_domain *sd = &per_cpu(core_domains, i);
7066 init_sched_groups_power(i, sd);
7070 for_each_cpu_mask(i, *cpu_map) {
7071 struct sched_domain *sd = &per_cpu(phys_domains, i);
7073 init_sched_groups_power(i, sd);
7077 for (i = 0; i < MAX_NUMNODES; i++)
7078 init_numa_sched_groups_power(sched_group_nodes[i]);
7081 struct sched_group *sg;
7083 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7085 init_numa_sched_groups_power(sg);
7089 /* Attach the domains */
7090 for_each_cpu_mask(i, *cpu_map) {
7091 struct sched_domain *sd;
7092 #ifdef CONFIG_SCHED_SMT
7093 sd = &per_cpu(cpu_domains, i);
7094 #elif defined(CONFIG_SCHED_MC)
7095 sd = &per_cpu(core_domains, i);
7097 sd = &per_cpu(phys_domains, i);
7099 cpu_attach_domain(sd, rd, i);
7102 SCHED_CPUMASK_FREE((void *)allmasks);
7107 free_sched_groups(cpu_map, tmpmask);
7108 SCHED_CPUMASK_FREE((void *)allmasks);
7113 static cpumask_t *doms_cur; /* current sched domains */
7114 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7117 * Special case: If a kmalloc of a doms_cur partition (array of
7118 * cpumask_t) fails, then fallback to a single sched domain,
7119 * as determined by the single cpumask_t fallback_doms.
7121 static cpumask_t fallback_doms;
7123 void __attribute__((weak)) arch_update_cpu_topology(void)
7128 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7129 * For now this just excludes isolated cpus, but could be used to
7130 * exclude other special cases in the future.
7132 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7136 arch_update_cpu_topology();
7138 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7140 doms_cur = &fallback_doms;
7141 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7142 err = build_sched_domains(doms_cur);
7143 register_sched_domain_sysctl();
7148 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7151 free_sched_groups(cpu_map, tmpmask);
7155 * Detach sched domains from a group of cpus specified in cpu_map
7156 * These cpus will now be attached to the NULL domain
7158 static void detach_destroy_domains(const cpumask_t *cpu_map)
7163 unregister_sched_domain_sysctl();
7165 for_each_cpu_mask(i, *cpu_map)
7166 cpu_attach_domain(NULL, &def_root_domain, i);
7167 synchronize_sched();
7168 arch_destroy_sched_domains(cpu_map, &tmpmask);
7172 * Partition sched domains as specified by the 'ndoms_new'
7173 * cpumasks in the array doms_new[] of cpumasks. This compares
7174 * doms_new[] to the current sched domain partitioning, doms_cur[].
7175 * It destroys each deleted domain and builds each new domain.
7177 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7178 * The masks don't intersect (don't overlap.) We should setup one
7179 * sched domain for each mask. CPUs not in any of the cpumasks will
7180 * not be load balanced. If the same cpumask appears both in the
7181 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7184 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7185 * ownership of it and will kfree it when done with it. If the caller
7186 * failed the kmalloc call, then it can pass in doms_new == NULL,
7187 * and partition_sched_domains() will fallback to the single partition
7190 * Call with hotplug lock held
7192 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7198 /* always unregister in case we don't destroy any domains */
7199 unregister_sched_domain_sysctl();
7201 if (doms_new == NULL) {
7203 doms_new = &fallback_doms;
7204 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7207 /* Destroy deleted domains */
7208 for (i = 0; i < ndoms_cur; i++) {
7209 for (j = 0; j < ndoms_new; j++) {
7210 if (cpus_equal(doms_cur[i], doms_new[j]))
7213 /* no match - a current sched domain not in new doms_new[] */
7214 detach_destroy_domains(doms_cur + i);
7219 /* Build new domains */
7220 for (i = 0; i < ndoms_new; i++) {
7221 for (j = 0; j < ndoms_cur; j++) {
7222 if (cpus_equal(doms_new[i], doms_cur[j]))
7225 /* no match - add a new doms_new */
7226 build_sched_domains(doms_new + i);
7231 /* Remember the new sched domains */
7232 if (doms_cur != &fallback_doms)
7234 doms_cur = doms_new;
7235 ndoms_cur = ndoms_new;
7237 register_sched_domain_sysctl();
7242 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7243 int arch_reinit_sched_domains(void)
7248 detach_destroy_domains(&cpu_online_map);
7249 err = arch_init_sched_domains(&cpu_online_map);
7255 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7259 if (buf[0] != '0' && buf[0] != '1')
7263 sched_smt_power_savings = (buf[0] == '1');
7265 sched_mc_power_savings = (buf[0] == '1');
7267 ret = arch_reinit_sched_domains();
7269 return ret ? ret : count;
7272 #ifdef CONFIG_SCHED_MC
7273 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7275 return sprintf(page, "%u\n", sched_mc_power_savings);
7277 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7278 const char *buf, size_t count)
7280 return sched_power_savings_store(buf, count, 0);
7282 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7283 sched_mc_power_savings_store);
7286 #ifdef CONFIG_SCHED_SMT
7287 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7289 return sprintf(page, "%u\n", sched_smt_power_savings);
7291 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7292 const char *buf, size_t count)
7294 return sched_power_savings_store(buf, count, 1);
7296 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7297 sched_smt_power_savings_store);
7300 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7304 #ifdef CONFIG_SCHED_SMT
7306 err = sysfs_create_file(&cls->kset.kobj,
7307 &attr_sched_smt_power_savings.attr);
7309 #ifdef CONFIG_SCHED_MC
7310 if (!err && mc_capable())
7311 err = sysfs_create_file(&cls->kset.kobj,
7312 &attr_sched_mc_power_savings.attr);
7319 * Force a reinitialization of the sched domains hierarchy. The domains
7320 * and groups cannot be updated in place without racing with the balancing
7321 * code, so we temporarily attach all running cpus to the NULL domain
7322 * which will prevent rebalancing while the sched domains are recalculated.
7324 static int update_sched_domains(struct notifier_block *nfb,
7325 unsigned long action, void *hcpu)
7328 case CPU_UP_PREPARE:
7329 case CPU_UP_PREPARE_FROZEN:
7330 case CPU_DOWN_PREPARE:
7331 case CPU_DOWN_PREPARE_FROZEN:
7332 detach_destroy_domains(&cpu_online_map);
7335 case CPU_UP_CANCELED:
7336 case CPU_UP_CANCELED_FROZEN:
7337 case CPU_DOWN_FAILED:
7338 case CPU_DOWN_FAILED_FROZEN:
7340 case CPU_ONLINE_FROZEN:
7342 case CPU_DEAD_FROZEN:
7344 * Fall through and re-initialise the domains.
7351 /* The hotplug lock is already held by cpu_up/cpu_down */
7352 arch_init_sched_domains(&cpu_online_map);
7357 void __init sched_init_smp(void)
7359 cpumask_t non_isolated_cpus;
7361 #if defined(CONFIG_NUMA)
7362 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7364 BUG_ON(sched_group_nodes_bycpu == NULL);
7367 arch_init_sched_domains(&cpu_online_map);
7368 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7369 if (cpus_empty(non_isolated_cpus))
7370 cpu_set(smp_processor_id(), non_isolated_cpus);
7372 /* XXX: Theoretical race here - CPU may be hotplugged now */
7373 hotcpu_notifier(update_sched_domains, 0);
7375 /* Move init over to a non-isolated CPU */
7376 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7378 sched_init_granularity();
7381 void __init sched_init_smp(void)
7383 #if defined(CONFIG_NUMA)
7384 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7386 BUG_ON(sched_group_nodes_bycpu == NULL);
7388 sched_init_granularity();
7390 #endif /* CONFIG_SMP */
7392 int in_sched_functions(unsigned long addr)
7394 return in_lock_functions(addr) ||
7395 (addr >= (unsigned long)__sched_text_start
7396 && addr < (unsigned long)__sched_text_end);
7399 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7401 cfs_rq->tasks_timeline = RB_ROOT;
7402 #ifdef CONFIG_FAIR_GROUP_SCHED
7405 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7408 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7410 struct rt_prio_array *array;
7413 array = &rt_rq->active;
7414 for (i = 0; i < MAX_RT_PRIO; i++) {
7415 INIT_LIST_HEAD(array->queue + i);
7416 __clear_bit(i, array->bitmap);
7418 /* delimiter for bitsearch: */
7419 __set_bit(MAX_RT_PRIO, array->bitmap);
7421 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7422 rt_rq->highest_prio = MAX_RT_PRIO;
7425 rt_rq->rt_nr_migratory = 0;
7426 rt_rq->overloaded = 0;
7430 rt_rq->rt_throttled = 0;
7431 rt_rq->rt_runtime = 0;
7432 spin_lock_init(&rt_rq->rt_runtime_lock);
7434 #ifdef CONFIG_RT_GROUP_SCHED
7435 rt_rq->rt_nr_boosted = 0;
7440 #ifdef CONFIG_FAIR_GROUP_SCHED
7441 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7442 struct cfs_rq *cfs_rq, struct sched_entity *se,
7445 tg->cfs_rq[cpu] = cfs_rq;
7446 init_cfs_rq(cfs_rq, rq);
7449 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7452 /* se could be NULL for init_task_group */
7456 se->cfs_rq = &rq->cfs;
7458 se->load.weight = tg->shares;
7459 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7464 #ifdef CONFIG_RT_GROUP_SCHED
7465 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7466 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7469 tg->rt_rq[cpu] = rt_rq;
7470 init_rt_rq(rt_rq, rq);
7472 rt_rq->rt_se = rt_se;
7473 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7475 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7477 tg->rt_se[cpu] = rt_se;
7481 rt_se->rt_rq = &rq->rt;
7482 rt_se->my_q = rt_rq;
7483 rt_se->parent = NULL;
7484 INIT_LIST_HEAD(&rt_se->run_list);
7488 void __init sched_init(void)
7491 unsigned long alloc_size = 0, ptr;
7493 #ifdef CONFIG_FAIR_GROUP_SCHED
7494 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7496 #ifdef CONFIG_RT_GROUP_SCHED
7497 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7500 * As sched_init() is called before page_alloc is setup,
7501 * we use alloc_bootmem().
7504 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7506 #ifdef CONFIG_FAIR_GROUP_SCHED
7507 init_task_group.se = (struct sched_entity **)ptr;
7508 ptr += nr_cpu_ids * sizeof(void **);
7510 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7511 ptr += nr_cpu_ids * sizeof(void **);
7513 #ifdef CONFIG_RT_GROUP_SCHED
7514 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7515 ptr += nr_cpu_ids * sizeof(void **);
7517 init_task_group.rt_rq = (struct rt_rq **)ptr;
7522 init_defrootdomain();
7525 init_rt_bandwidth(&def_rt_bandwidth,
7526 global_rt_period(), global_rt_runtime());
7528 #ifdef CONFIG_RT_GROUP_SCHED
7529 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7530 global_rt_period(), global_rt_runtime());
7533 #ifdef CONFIG_GROUP_SCHED
7534 list_add(&init_task_group.list, &task_groups);
7537 for_each_possible_cpu(i) {
7541 spin_lock_init(&rq->lock);
7542 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7545 update_last_tick_seen(rq);
7546 init_cfs_rq(&rq->cfs, rq);
7547 init_rt_rq(&rq->rt, rq);
7548 #ifdef CONFIG_FAIR_GROUP_SCHED
7549 init_task_group.shares = init_task_group_load;
7550 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7551 #ifdef CONFIG_CGROUP_SCHED
7553 * How much cpu bandwidth does init_task_group get?
7555 * In case of task-groups formed thr' the cgroup filesystem, it
7556 * gets 100% of the cpu resources in the system. This overall
7557 * system cpu resource is divided among the tasks of
7558 * init_task_group and its child task-groups in a fair manner,
7559 * based on each entity's (task or task-group's) weight
7560 * (se->load.weight).
7562 * In other words, if init_task_group has 10 tasks of weight
7563 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7564 * then A0's share of the cpu resource is:
7566 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7568 * We achieve this by letting init_task_group's tasks sit
7569 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7571 init_tg_cfs_entry(rq, &init_task_group, &rq->cfs, NULL, i, 1);
7572 #elif defined CONFIG_USER_SCHED
7574 * In case of task-groups formed thr' the user id of tasks,
7575 * init_task_group represents tasks belonging to root user.
7576 * Hence it forms a sibling of all subsequent groups formed.
7577 * In this case, init_task_group gets only a fraction of overall
7578 * system cpu resource, based on the weight assigned to root
7579 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7580 * by letting tasks of init_task_group sit in a separate cfs_rq
7581 * (init_cfs_rq) and having one entity represent this group of
7582 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7584 init_tg_cfs_entry(rq, &init_task_group,
7585 &per_cpu(init_cfs_rq, i),
7586 &per_cpu(init_sched_entity, i), i, 1);
7589 #endif /* CONFIG_FAIR_GROUP_SCHED */
7591 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7592 #ifdef CONFIG_RT_GROUP_SCHED
7593 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7594 #ifdef CONFIG_CGROUP_SCHED
7595 init_tg_rt_entry(rq, &init_task_group, &rq->rt, NULL, i, 1);
7596 #elif defined CONFIG_USER_SCHED
7597 init_tg_rt_entry(rq, &init_task_group,
7598 &per_cpu(init_rt_rq, i),
7599 &per_cpu(init_sched_rt_entity, i), i, 1);
7603 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7604 rq->cpu_load[j] = 0;
7608 rq->active_balance = 0;
7609 rq->next_balance = jiffies;
7612 rq->migration_thread = NULL;
7613 INIT_LIST_HEAD(&rq->migration_queue);
7614 rq_attach_root(rq, &def_root_domain);
7617 atomic_set(&rq->nr_iowait, 0);
7620 set_load_weight(&init_task);
7622 #ifdef CONFIG_PREEMPT_NOTIFIERS
7623 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7627 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7630 #ifdef CONFIG_RT_MUTEXES
7631 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7635 * The boot idle thread does lazy MMU switching as well:
7637 atomic_inc(&init_mm.mm_count);
7638 enter_lazy_tlb(&init_mm, current);
7641 * Make us the idle thread. Technically, schedule() should not be
7642 * called from this thread, however somewhere below it might be,
7643 * but because we are the idle thread, we just pick up running again
7644 * when this runqueue becomes "idle".
7646 init_idle(current, smp_processor_id());
7648 * During early bootup we pretend to be a normal task:
7650 current->sched_class = &fair_sched_class;
7652 scheduler_running = 1;
7655 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7656 void __might_sleep(char *file, int line)
7659 static unsigned long prev_jiffy; /* ratelimiting */
7661 if ((in_atomic() || irqs_disabled()) &&
7662 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7663 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7665 prev_jiffy = jiffies;
7666 printk(KERN_ERR "BUG: sleeping function called from invalid"
7667 " context at %s:%d\n", file, line);
7668 printk("in_atomic():%d, irqs_disabled():%d\n",
7669 in_atomic(), irqs_disabled());
7670 debug_show_held_locks(current);
7671 if (irqs_disabled())
7672 print_irqtrace_events(current);
7677 EXPORT_SYMBOL(__might_sleep);
7680 #ifdef CONFIG_MAGIC_SYSRQ
7681 static void normalize_task(struct rq *rq, struct task_struct *p)
7684 update_rq_clock(rq);
7685 on_rq = p->se.on_rq;
7687 deactivate_task(rq, p, 0);
7688 __setscheduler(rq, p, SCHED_NORMAL, 0);
7690 activate_task(rq, p, 0);
7691 resched_task(rq->curr);
7695 void normalize_rt_tasks(void)
7697 struct task_struct *g, *p;
7698 unsigned long flags;
7701 read_lock_irqsave(&tasklist_lock, flags);
7702 do_each_thread(g, p) {
7704 * Only normalize user tasks:
7709 p->se.exec_start = 0;
7710 #ifdef CONFIG_SCHEDSTATS
7711 p->se.wait_start = 0;
7712 p->se.sleep_start = 0;
7713 p->se.block_start = 0;
7715 task_rq(p)->clock = 0;
7719 * Renice negative nice level userspace
7722 if (TASK_NICE(p) < 0 && p->mm)
7723 set_user_nice(p, 0);
7727 spin_lock(&p->pi_lock);
7728 rq = __task_rq_lock(p);
7730 normalize_task(rq, p);
7732 __task_rq_unlock(rq);
7733 spin_unlock(&p->pi_lock);
7734 } while_each_thread(g, p);
7736 read_unlock_irqrestore(&tasklist_lock, flags);
7739 #endif /* CONFIG_MAGIC_SYSRQ */
7743 * These functions are only useful for the IA64 MCA handling.
7745 * They can only be called when the whole system has been
7746 * stopped - every CPU needs to be quiescent, and no scheduling
7747 * activity can take place. Using them for anything else would
7748 * be a serious bug, and as a result, they aren't even visible
7749 * under any other configuration.
7753 * curr_task - return the current task for a given cpu.
7754 * @cpu: the processor in question.
7756 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7758 struct task_struct *curr_task(int cpu)
7760 return cpu_curr(cpu);
7764 * set_curr_task - set the current task for a given cpu.
7765 * @cpu: the processor in question.
7766 * @p: the task pointer to set.
7768 * Description: This function must only be used when non-maskable interrupts
7769 * are serviced on a separate stack. It allows the architecture to switch the
7770 * notion of the current task on a cpu in a non-blocking manner. This function
7771 * must be called with all CPU's synchronized, and interrupts disabled, the
7772 * and caller must save the original value of the current task (see
7773 * curr_task() above) and restore that value before reenabling interrupts and
7774 * re-starting the system.
7776 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7778 void set_curr_task(int cpu, struct task_struct *p)
7785 #ifdef CONFIG_FAIR_GROUP_SCHED
7786 static void free_fair_sched_group(struct task_group *tg)
7790 for_each_possible_cpu(i) {
7792 kfree(tg->cfs_rq[i]);
7801 static int alloc_fair_sched_group(struct task_group *tg)
7803 struct cfs_rq *cfs_rq;
7804 struct sched_entity *se;
7808 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7811 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7815 tg->shares = NICE_0_LOAD;
7817 for_each_possible_cpu(i) {
7820 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7821 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7825 se = kmalloc_node(sizeof(struct sched_entity),
7826 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7830 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7839 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7841 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7842 &cpu_rq(cpu)->leaf_cfs_rq_list);
7845 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7847 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7850 static inline void free_fair_sched_group(struct task_group *tg)
7854 static inline int alloc_fair_sched_group(struct task_group *tg)
7859 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7863 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7868 #ifdef CONFIG_RT_GROUP_SCHED
7869 static void free_rt_sched_group(struct task_group *tg)
7873 destroy_rt_bandwidth(&tg->rt_bandwidth);
7875 for_each_possible_cpu(i) {
7877 kfree(tg->rt_rq[i]);
7879 kfree(tg->rt_se[i]);
7886 static int alloc_rt_sched_group(struct task_group *tg)
7888 struct rt_rq *rt_rq;
7889 struct sched_rt_entity *rt_se;
7893 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7896 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7900 init_rt_bandwidth(&tg->rt_bandwidth,
7901 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7903 for_each_possible_cpu(i) {
7906 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7907 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7911 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7912 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7916 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7925 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7927 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7928 &cpu_rq(cpu)->leaf_rt_rq_list);
7931 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7933 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7936 static inline void free_rt_sched_group(struct task_group *tg)
7940 static inline int alloc_rt_sched_group(struct task_group *tg)
7945 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7949 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7954 #ifdef CONFIG_GROUP_SCHED
7955 static void free_sched_group(struct task_group *tg)
7957 free_fair_sched_group(tg);
7958 free_rt_sched_group(tg);
7962 /* allocate runqueue etc for a new task group */
7963 struct task_group *sched_create_group(void)
7965 struct task_group *tg;
7966 unsigned long flags;
7969 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7971 return ERR_PTR(-ENOMEM);
7973 if (!alloc_fair_sched_group(tg))
7976 if (!alloc_rt_sched_group(tg))
7979 spin_lock_irqsave(&task_group_lock, flags);
7980 for_each_possible_cpu(i) {
7981 register_fair_sched_group(tg, i);
7982 register_rt_sched_group(tg, i);
7984 list_add_rcu(&tg->list, &task_groups);
7985 spin_unlock_irqrestore(&task_group_lock, flags);
7990 free_sched_group(tg);
7991 return ERR_PTR(-ENOMEM);
7994 /* rcu callback to free various structures associated with a task group */
7995 static void free_sched_group_rcu(struct rcu_head *rhp)
7997 /* now it should be safe to free those cfs_rqs */
7998 free_sched_group(container_of(rhp, struct task_group, rcu));
8001 /* Destroy runqueue etc associated with a task group */
8002 void sched_destroy_group(struct task_group *tg)
8004 unsigned long flags;
8007 spin_lock_irqsave(&task_group_lock, flags);
8008 for_each_possible_cpu(i) {
8009 unregister_fair_sched_group(tg, i);
8010 unregister_rt_sched_group(tg, i);
8012 list_del_rcu(&tg->list);
8013 spin_unlock_irqrestore(&task_group_lock, flags);
8015 /* wait for possible concurrent references to cfs_rqs complete */
8016 call_rcu(&tg->rcu, free_sched_group_rcu);
8019 /* change task's runqueue when it moves between groups.
8020 * The caller of this function should have put the task in its new group
8021 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8022 * reflect its new group.
8024 void sched_move_task(struct task_struct *tsk)
8027 unsigned long flags;
8030 rq = task_rq_lock(tsk, &flags);
8032 update_rq_clock(rq);
8034 running = task_current(rq, tsk);
8035 on_rq = tsk->se.on_rq;
8038 dequeue_task(rq, tsk, 0);
8039 if (unlikely(running))
8040 tsk->sched_class->put_prev_task(rq, tsk);
8042 set_task_rq(tsk, task_cpu(tsk));
8044 #ifdef CONFIG_FAIR_GROUP_SCHED
8045 if (tsk->sched_class->moved_group)
8046 tsk->sched_class->moved_group(tsk);
8049 if (unlikely(running))
8050 tsk->sched_class->set_curr_task(rq);
8052 enqueue_task(rq, tsk, 0);
8054 task_rq_unlock(rq, &flags);
8058 #ifdef CONFIG_FAIR_GROUP_SCHED
8059 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8061 struct cfs_rq *cfs_rq = se->cfs_rq;
8062 struct rq *rq = cfs_rq->rq;
8065 spin_lock_irq(&rq->lock);
8069 dequeue_entity(cfs_rq, se, 0);
8071 se->load.weight = shares;
8072 se->load.inv_weight = div64_64((1ULL<<32), shares);
8075 enqueue_entity(cfs_rq, se, 0);
8077 spin_unlock_irq(&rq->lock);
8080 static DEFINE_MUTEX(shares_mutex);
8082 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8085 unsigned long flags;
8088 * A weight of 0 or 1 can cause arithmetics problems.
8089 * (The default weight is 1024 - so there's no practical
8090 * limitation from this.)
8095 mutex_lock(&shares_mutex);
8096 if (tg->shares == shares)
8099 spin_lock_irqsave(&task_group_lock, flags);
8100 for_each_possible_cpu(i)
8101 unregister_fair_sched_group(tg, i);
8102 spin_unlock_irqrestore(&task_group_lock, flags);
8104 /* wait for any ongoing reference to this group to finish */
8105 synchronize_sched();
8108 * Now we are free to modify the group's share on each cpu
8109 * w/o tripping rebalance_share or load_balance_fair.
8111 tg->shares = shares;
8112 for_each_possible_cpu(i)
8113 set_se_shares(tg->se[i], shares);
8116 * Enable load balance activity on this group, by inserting it back on
8117 * each cpu's rq->leaf_cfs_rq_list.
8119 spin_lock_irqsave(&task_group_lock, flags);
8120 for_each_possible_cpu(i)
8121 register_fair_sched_group(tg, i);
8122 spin_unlock_irqrestore(&task_group_lock, flags);
8124 mutex_unlock(&shares_mutex);
8128 unsigned long sched_group_shares(struct task_group *tg)
8134 #ifdef CONFIG_RT_GROUP_SCHED
8136 * Ensure that the real time constraints are schedulable.
8138 static DEFINE_MUTEX(rt_constraints_mutex);
8140 static unsigned long to_ratio(u64 period, u64 runtime)
8142 if (runtime == RUNTIME_INF)
8145 return div64_64(runtime << 16, period);
8148 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8150 struct task_group *tgi;
8151 unsigned long total = 0;
8152 unsigned long global_ratio =
8153 to_ratio(global_rt_period(), global_rt_runtime());
8156 list_for_each_entry_rcu(tgi, &task_groups, list) {
8160 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8161 tgi->rt_bandwidth.rt_runtime);
8165 return total + to_ratio(period, runtime) < global_ratio;
8168 /* Must be called with tasklist_lock held */
8169 static inline int tg_has_rt_tasks(struct task_group *tg)
8171 struct task_struct *g, *p;
8172 do_each_thread(g, p) {
8173 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8175 } while_each_thread(g, p);
8179 static int tg_set_bandwidth(struct task_group *tg,
8180 u64 rt_period, u64 rt_runtime)
8184 mutex_lock(&rt_constraints_mutex);
8185 read_lock(&tasklist_lock);
8186 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8190 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8195 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8196 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8197 tg->rt_bandwidth.rt_runtime = rt_runtime;
8199 for_each_possible_cpu(i) {
8200 struct rt_rq *rt_rq = tg->rt_rq[i];
8202 spin_lock(&rt_rq->rt_runtime_lock);
8203 rt_rq->rt_runtime = rt_runtime;
8204 spin_unlock(&rt_rq->rt_runtime_lock);
8206 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8208 read_unlock(&tasklist_lock);
8209 mutex_unlock(&rt_constraints_mutex);
8214 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8216 u64 rt_runtime, rt_period;
8218 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8219 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8220 if (rt_runtime_us < 0)
8221 rt_runtime = RUNTIME_INF;
8223 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8226 long sched_group_rt_runtime(struct task_group *tg)
8230 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8233 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8234 do_div(rt_runtime_us, NSEC_PER_USEC);
8235 return rt_runtime_us;
8238 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8240 u64 rt_runtime, rt_period;
8242 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8243 rt_runtime = tg->rt_bandwidth.rt_runtime;
8245 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8248 long sched_group_rt_period(struct task_group *tg)
8252 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8253 do_div(rt_period_us, NSEC_PER_USEC);
8254 return rt_period_us;
8257 static int sched_rt_global_constraints(void)
8261 mutex_lock(&rt_constraints_mutex);
8262 if (!__rt_schedulable(NULL, 1, 0))
8264 mutex_unlock(&rt_constraints_mutex);
8269 static int sched_rt_global_constraints(void)
8271 unsigned long flags;
8274 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8275 for_each_possible_cpu(i) {
8276 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8278 spin_lock(&rt_rq->rt_runtime_lock);
8279 rt_rq->rt_runtime = global_rt_runtime();
8280 spin_unlock(&rt_rq->rt_runtime_lock);
8282 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8288 int sched_rt_handler(struct ctl_table *table, int write,
8289 struct file *filp, void __user *buffer, size_t *lenp,
8293 int old_period, old_runtime;
8294 static DEFINE_MUTEX(mutex);
8297 old_period = sysctl_sched_rt_period;
8298 old_runtime = sysctl_sched_rt_runtime;
8300 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8302 if (!ret && write) {
8303 ret = sched_rt_global_constraints();
8305 sysctl_sched_rt_period = old_period;
8306 sysctl_sched_rt_runtime = old_runtime;
8308 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8309 def_rt_bandwidth.rt_period =
8310 ns_to_ktime(global_rt_period());
8313 mutex_unlock(&mutex);
8318 #ifdef CONFIG_CGROUP_SCHED
8320 /* return corresponding task_group object of a cgroup */
8321 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8323 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8324 struct task_group, css);
8327 static struct cgroup_subsys_state *
8328 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8330 struct task_group *tg;
8332 if (!cgrp->parent) {
8333 /* This is early initialization for the top cgroup */
8334 init_task_group.css.cgroup = cgrp;
8335 return &init_task_group.css;
8338 /* we support only 1-level deep hierarchical scheduler atm */
8339 if (cgrp->parent->parent)
8340 return ERR_PTR(-EINVAL);
8342 tg = sched_create_group();
8344 return ERR_PTR(-ENOMEM);
8346 /* Bind the cgroup to task_group object we just created */
8347 tg->css.cgroup = cgrp;
8353 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8355 struct task_group *tg = cgroup_tg(cgrp);
8357 sched_destroy_group(tg);
8361 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8362 struct task_struct *tsk)
8364 #ifdef CONFIG_RT_GROUP_SCHED
8365 /* Don't accept realtime tasks when there is no way for them to run */
8366 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8369 /* We don't support RT-tasks being in separate groups */
8370 if (tsk->sched_class != &fair_sched_class)
8378 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8379 struct cgroup *old_cont, struct task_struct *tsk)
8381 sched_move_task(tsk);
8384 #ifdef CONFIG_FAIR_GROUP_SCHED
8385 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8388 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8391 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8393 struct task_group *tg = cgroup_tg(cgrp);
8395 return (u64) tg->shares;
8399 #ifdef CONFIG_RT_GROUP_SCHED
8400 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8402 const char __user *userbuf,
8403 size_t nbytes, loff_t *unused_ppos)
8412 if (nbytes >= sizeof(buffer))
8414 if (copy_from_user(buffer, userbuf, nbytes))
8417 buffer[nbytes] = 0; /* nul-terminate */
8419 /* strip newline if necessary */
8420 if (nbytes && (buffer[nbytes-1] == '\n'))
8421 buffer[nbytes-1] = 0;
8422 val = simple_strtoll(buffer, &end, 0);
8426 /* Pass to subsystem */
8427 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8433 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8435 char __user *buf, size_t nbytes,
8439 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8440 int len = sprintf(tmp, "%ld\n", val);
8442 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8445 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8448 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8451 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8453 return sched_group_rt_period(cgroup_tg(cgrp));
8457 static struct cftype cpu_files[] = {
8458 #ifdef CONFIG_FAIR_GROUP_SCHED
8461 .read_uint = cpu_shares_read_uint,
8462 .write_uint = cpu_shares_write_uint,
8465 #ifdef CONFIG_RT_GROUP_SCHED
8467 .name = "rt_runtime_us",
8468 .read = cpu_rt_runtime_read,
8469 .write = cpu_rt_runtime_write,
8472 .name = "rt_period_us",
8473 .read_uint = cpu_rt_period_read_uint,
8474 .write_uint = cpu_rt_period_write_uint,
8479 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8481 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8484 struct cgroup_subsys cpu_cgroup_subsys = {
8486 .create = cpu_cgroup_create,
8487 .destroy = cpu_cgroup_destroy,
8488 .can_attach = cpu_cgroup_can_attach,
8489 .attach = cpu_cgroup_attach,
8490 .populate = cpu_cgroup_populate,
8491 .subsys_id = cpu_cgroup_subsys_id,
8495 #endif /* CONFIG_CGROUP_SCHED */
8497 #ifdef CONFIG_CGROUP_CPUACCT
8500 * CPU accounting code for task groups.
8502 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8503 * (balbir@in.ibm.com).
8506 /* track cpu usage of a group of tasks */
8508 struct cgroup_subsys_state css;
8509 /* cpuusage holds pointer to a u64-type object on every cpu */
8513 struct cgroup_subsys cpuacct_subsys;
8515 /* return cpu accounting group corresponding to this container */
8516 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8518 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8519 struct cpuacct, css);
8522 /* return cpu accounting group to which this task belongs */
8523 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8525 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8526 struct cpuacct, css);
8529 /* create a new cpu accounting group */
8530 static struct cgroup_subsys_state *cpuacct_create(
8531 struct cgroup_subsys *ss, struct cgroup *cgrp)
8533 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8536 return ERR_PTR(-ENOMEM);
8538 ca->cpuusage = alloc_percpu(u64);
8539 if (!ca->cpuusage) {
8541 return ERR_PTR(-ENOMEM);
8547 /* destroy an existing cpu accounting group */
8549 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8551 struct cpuacct *ca = cgroup_ca(cgrp);
8553 free_percpu(ca->cpuusage);
8557 /* return total cpu usage (in nanoseconds) of a group */
8558 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8560 struct cpuacct *ca = cgroup_ca(cgrp);
8561 u64 totalcpuusage = 0;
8564 for_each_possible_cpu(i) {
8565 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8568 * Take rq->lock to make 64-bit addition safe on 32-bit
8571 spin_lock_irq(&cpu_rq(i)->lock);
8572 totalcpuusage += *cpuusage;
8573 spin_unlock_irq(&cpu_rq(i)->lock);
8576 return totalcpuusage;
8579 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8582 struct cpuacct *ca = cgroup_ca(cgrp);
8591 for_each_possible_cpu(i) {
8592 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8594 spin_lock_irq(&cpu_rq(i)->lock);
8596 spin_unlock_irq(&cpu_rq(i)->lock);
8602 static struct cftype files[] = {
8605 .read_uint = cpuusage_read,
8606 .write_uint = cpuusage_write,
8610 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8612 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8616 * charge this task's execution time to its accounting group.
8618 * called with rq->lock held.
8620 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8624 if (!cpuacct_subsys.active)
8629 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8631 *cpuusage += cputime;
8635 struct cgroup_subsys cpuacct_subsys = {
8637 .create = cpuacct_create,
8638 .destroy = cpuacct_destroy,
8639 .populate = cpuacct_populate,
8640 .subsys_id = cpuacct_subsys_id,
8642 #endif /* CONFIG_CGROUP_CPUACCT */