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
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 #define NICE_0_LOAD SCHED_LOAD_SCALE
95 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
98 * These are the 'tuning knobs' of the scheduler:
100 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
101 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
102 * Timeslices get refilled after they expire.
104 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
105 #define DEF_TIMESLICE (100 * HZ / 1000)
106 #define ON_RUNQUEUE_WEIGHT 30
107 #define CHILD_PENALTY 95
108 #define PARENT_PENALTY 100
109 #define EXIT_WEIGHT 3
110 #define PRIO_BONUS_RATIO 25
111 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
112 #define INTERACTIVE_DELTA 2
113 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
114 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
115 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
118 * If a task is 'interactive' then we reinsert it in the active
119 * array after it has expired its current timeslice. (it will not
120 * continue to run immediately, it will still roundrobin with
121 * other interactive tasks.)
123 * This part scales the interactivity limit depending on niceness.
125 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
126 * Here are a few examples of different nice levels:
128 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
129 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
130 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
131 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
132 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
134 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
135 * priority range a task can explore, a value of '1' means the
136 * task is rated interactive.)
138 * Ie. nice +19 tasks can never get 'interactive' enough to be
139 * reinserted into the active array. And only heavily CPU-hog nice -20
140 * tasks will be expired. Default nice 0 tasks are somewhere between,
141 * it takes some effort for them to get interactive, but it's not
145 #define CURRENT_BONUS(p) \
146 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
149 #define GRANULARITY (10 * HZ / 1000 ? : 1)
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
156 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
157 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
160 #define SCALE(v1,v1_max,v2_max) \
161 (v1) * (v2_max) / (v1_max)
164 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
167 #define TASK_INTERACTIVE(p) \
168 ((p)->prio <= (p)->static_prio - DELTA(p))
170 #define INTERACTIVE_SLEEP(p) \
171 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
172 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
174 #define TASK_PREEMPTS_CURR(p, rq) \
175 ((p)->prio < (rq)->curr->prio)
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio)
182 if (static_prio < NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
185 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
190 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
191 * Since cpu_power is a 'constant', we can use a reciprocal divide.
193 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
195 return reciprocal_divide(load, sg->reciprocal_cpu_power);
199 * Each time a sched group cpu_power is changed,
200 * we must compute its reciprocal value
202 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
204 sg->__cpu_power += val;
205 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
210 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
211 * to time slice values: [800ms ... 100ms ... 5ms]
213 * The higher a thread's priority, the bigger timeslices
214 * it gets during one round of execution. But even the lowest
215 * priority thread gets MIN_TIMESLICE worth of execution time.
218 static inline unsigned int task_timeslice(struct task_struct *p)
220 return static_prio_timeslice(p->static_prio);
223 static inline int rt_policy(int policy)
225 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
230 static inline int task_has_rt_policy(struct task_struct *p)
232 return rt_policy(p->policy);
236 * This is the priority-queue data structure of the RT scheduling class:
238 struct rt_prio_array {
239 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
240 struct list_head queue[MAX_RT_PRIO];
244 struct load_weight load;
245 u64 load_update_start, load_update_last;
246 unsigned long delta_fair, delta_exec, delta_stat;
249 /* CFS-related fields in a runqueue */
251 struct load_weight load;
252 unsigned long nr_running;
258 unsigned long wait_runtime_overruns, wait_runtime_underruns;
260 struct rb_root tasks_timeline;
261 struct rb_node *rb_leftmost;
262 struct rb_node *rb_load_balance_curr;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* 'curr' points to currently running entity on this cfs_rq.
265 * It is set to NULL otherwise (i.e when none are currently running).
267 struct sched_entity *curr;
268 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
270 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
271 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
272 * (like users, containers etc.)
274 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
275 * list is used during load balance.
277 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
281 /* Real-Time classes' related field in a runqueue: */
283 struct rt_prio_array active;
284 int rt_load_balance_idx;
285 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
289 * The prio-array type of the old scheduler:
292 unsigned int nr_active;
293 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
294 struct list_head queue[MAX_PRIO];
298 * This is the main, per-CPU runqueue data structure.
300 * Locking rule: those places that want to lock multiple runqueues
301 * (such as the load balancing or the thread migration code), lock
302 * acquire operations must be ordered by ascending &runqueue.
305 spinlock_t lock; /* runqueue lock */
308 * nr_running and cpu_load should be in the same cacheline because
309 * remote CPUs use both these fields when doing load calculation.
311 unsigned long nr_running;
312 unsigned long raw_weighted_load;
313 #define CPU_LOAD_IDX_MAX 5
314 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
315 unsigned char idle_at_tick;
317 unsigned char in_nohz_recently;
319 struct load_stat ls; /* capture load from *all* tasks on this cpu */
320 unsigned long nr_load_updates;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
330 * This is part of a global counter where only the total sum
331 * over all CPUs matters. A task can increase this counter on
332 * one CPU and if it got migrated afterwards it may decrease
333 * it on another CPU. Always updated under the runqueue lock:
335 unsigned long nr_uninterruptible;
337 unsigned long expired_timestamp;
338 unsigned long long most_recent_timestamp;
340 struct task_struct *curr, *idle;
341 unsigned long next_balance;
342 struct mm_struct *prev_mm;
344 struct prio_array *active, *expired, arrays[2];
345 int best_expired_prio;
347 u64 clock, prev_clock_raw;
350 unsigned int clock_warps, clock_overflows;
351 unsigned int clock_unstable_events;
353 struct sched_class *load_balance_class;
358 struct sched_domain *sd;
360 /* For active balancing */
363 int cpu; /* cpu of this runqueue */
365 struct task_struct *migration_thread;
366 struct list_head migration_queue;
369 #ifdef CONFIG_SCHEDSTATS
371 struct sched_info rq_sched_info;
373 /* sys_sched_yield() stats */
374 unsigned long yld_exp_empty;
375 unsigned long yld_act_empty;
376 unsigned long yld_both_empty;
377 unsigned long yld_cnt;
379 /* schedule() stats */
380 unsigned long sched_switch;
381 unsigned long sched_cnt;
382 unsigned long sched_goidle;
384 /* try_to_wake_up() stats */
385 unsigned long ttwu_cnt;
386 unsigned long ttwu_local;
388 struct lock_class_key rq_lock_key;
391 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
392 static DEFINE_MUTEX(sched_hotcpu_mutex);
394 static inline int cpu_of(struct rq *rq)
404 * Per-runqueue clock, as finegrained as the platform can give us:
406 static unsigned long long __rq_clock(struct rq *rq)
408 u64 prev_raw = rq->prev_clock_raw;
409 u64 now = sched_clock();
410 s64 delta = now - prev_raw;
411 u64 clock = rq->clock;
414 * Protect against sched_clock() occasionally going backwards:
416 if (unlikely(delta < 0)) {
421 * Catch too large forward jumps too:
423 if (unlikely(delta > 2*TICK_NSEC)) {
425 rq->clock_overflows++;
427 if (unlikely(delta > rq->clock_max_delta))
428 rq->clock_max_delta = delta;
433 rq->prev_clock_raw = now;
439 static inline unsigned long long rq_clock(struct rq *rq)
441 int this_cpu = smp_processor_id();
443 if (this_cpu == cpu_of(rq))
444 return __rq_clock(rq);
450 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
451 * See detach_destroy_domains: synchronize_sched for details.
453 * The domain tree of any CPU may only be accessed from within
454 * preempt-disabled sections.
456 #define for_each_domain(cpu, __sd) \
457 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
459 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
460 #define this_rq() (&__get_cpu_var(runqueues))
461 #define task_rq(p) cpu_rq(task_cpu(p))
462 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
464 #ifdef CONFIG_FAIR_GROUP_SCHED
465 /* Change a task's ->cfs_rq if it moves across CPUs */
466 static inline void set_task_cfs_rq(struct task_struct *p)
468 p->se.cfs_rq = &task_rq(p)->cfs;
471 static inline void set_task_cfs_rq(struct task_struct *p)
476 #ifndef prepare_arch_switch
477 # define prepare_arch_switch(next) do { } while (0)
479 #ifndef finish_arch_switch
480 # define finish_arch_switch(prev) do { } while (0)
483 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
484 static inline int task_running(struct rq *rq, struct task_struct *p)
486 return rq->curr == p;
489 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
493 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
495 #ifdef CONFIG_DEBUG_SPINLOCK
496 /* this is a valid case when another task releases the spinlock */
497 rq->lock.owner = current;
500 * If we are tracking spinlock dependencies then we have to
501 * fix up the runqueue lock - which gets 'carried over' from
504 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
506 spin_unlock_irq(&rq->lock);
509 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
510 static inline int task_running(struct rq *rq, struct task_struct *p)
515 return rq->curr == p;
519 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
523 * We can optimise this out completely for !SMP, because the
524 * SMP rebalancing from interrupt is the only thing that cares
529 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
530 spin_unlock_irq(&rq->lock);
532 spin_unlock(&rq->lock);
536 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
540 * After ->oncpu is cleared, the task can be moved to a different CPU.
541 * We must ensure this doesn't happen until the switch is completely
547 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
551 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
554 * __task_rq_lock - lock the runqueue a given task resides on.
555 * Must be called interrupts disabled.
557 static inline struct rq *__task_rq_lock(struct task_struct *p)
564 spin_lock(&rq->lock);
565 if (unlikely(rq != task_rq(p))) {
566 spin_unlock(&rq->lock);
567 goto repeat_lock_task;
573 * task_rq_lock - lock the runqueue a given task resides on and disable
574 * interrupts. Note the ordering: we can safely lookup the task_rq without
575 * explicitly disabling preemption.
577 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
583 local_irq_save(*flags);
585 spin_lock(&rq->lock);
586 if (unlikely(rq != task_rq(p))) {
587 spin_unlock_irqrestore(&rq->lock, *flags);
588 goto repeat_lock_task;
593 static inline void __task_rq_unlock(struct rq *rq)
596 spin_unlock(&rq->lock);
599 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
602 spin_unlock_irqrestore(&rq->lock, *flags);
606 * this_rq_lock - lock this runqueue and disable interrupts.
608 static inline struct rq *this_rq_lock(void)
615 spin_lock(&rq->lock);
621 * resched_task - mark a task 'to be rescheduled now'.
623 * On UP this means the setting of the need_resched flag, on SMP it
624 * might also involve a cross-CPU call to trigger the scheduler on
629 #ifndef tsk_is_polling
630 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
633 static void resched_task(struct task_struct *p)
637 assert_spin_locked(&task_rq(p)->lock);
639 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
642 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
645 if (cpu == smp_processor_id())
648 /* NEED_RESCHED must be visible before we test polling */
650 if (!tsk_is_polling(p))
651 smp_send_reschedule(cpu);
654 static void resched_cpu(int cpu)
656 struct rq *rq = cpu_rq(cpu);
659 if (!spin_trylock_irqsave(&rq->lock, flags))
661 resched_task(cpu_curr(cpu));
662 spin_unlock_irqrestore(&rq->lock, flags);
665 static inline void resched_task(struct task_struct *p)
667 assert_spin_locked(&task_rq(p)->lock);
668 set_tsk_need_resched(p);
672 #include "sched_stats.h"
675 * Adding/removing a task to/from a priority array:
677 static void dequeue_task(struct task_struct *p, struct prio_array *array)
680 list_del(&p->run_list);
681 if (list_empty(array->queue + p->prio))
682 __clear_bit(p->prio, array->bitmap);
685 static void enqueue_task(struct task_struct *p, struct prio_array *array)
687 sched_info_queued(p);
688 list_add_tail(&p->run_list, array->queue + p->prio);
689 __set_bit(p->prio, array->bitmap);
695 * Put task to the end of the run list without the overhead of dequeue
696 * followed by enqueue.
698 static void requeue_task(struct task_struct *p, struct prio_array *array)
700 list_move_tail(&p->run_list, array->queue + p->prio);
704 enqueue_task_head(struct task_struct *p, struct prio_array *array)
706 list_add(&p->run_list, array->queue + p->prio);
707 __set_bit(p->prio, array->bitmap);
713 * __normal_prio - return the priority that is based on the static
714 * priority but is modified by bonuses/penalties.
716 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
717 * into the -5 ... 0 ... +5 bonus/penalty range.
719 * We use 25% of the full 0...39 priority range so that:
721 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
722 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
724 * Both properties are important to certain workloads.
727 static inline int __normal_prio(struct task_struct *p)
731 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
733 prio = p->static_prio - bonus;
734 if (prio < MAX_RT_PRIO)
736 if (prio > MAX_PRIO-1)
742 * To aid in avoiding the subversion of "niceness" due to uneven distribution
743 * of tasks with abnormal "nice" values across CPUs the contribution that
744 * each task makes to its run queue's load is weighted according to its
745 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
746 * scaled version of the new time slice allocation that they receive on time
751 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
752 * If static_prio_timeslice() is ever changed to break this assumption then
753 * this code will need modification
755 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
756 #define LOAD_WEIGHT(lp) \
757 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
758 #define PRIO_TO_LOAD_WEIGHT(prio) \
759 LOAD_WEIGHT(static_prio_timeslice(prio))
760 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
761 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
763 static void set_load_weight(struct task_struct *p)
765 if (task_has_rt_policy(p)) {
767 if (p == task_rq(p)->migration_thread)
769 * The migration thread does the actual balancing.
770 * Giving its load any weight will skew balancing
776 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
778 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
782 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
784 rq->raw_weighted_load += p->load_weight;
788 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
790 rq->raw_weighted_load -= p->load_weight;
793 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
796 inc_raw_weighted_load(rq, p);
799 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
802 dec_raw_weighted_load(rq, p);
806 * Calculate the expected normal priority: i.e. priority
807 * without taking RT-inheritance into account. Might be
808 * boosted by interactivity modifiers. Changes upon fork,
809 * setprio syscalls, and whenever the interactivity
810 * estimator recalculates.
812 static inline int normal_prio(struct task_struct *p)
816 if (task_has_rt_policy(p))
817 prio = MAX_RT_PRIO-1 - p->rt_priority;
819 prio = __normal_prio(p);
824 * Calculate the current priority, i.e. the priority
825 * taken into account by the scheduler. This value might
826 * be boosted by RT tasks, or might be boosted by
827 * interactivity modifiers. Will be RT if the task got
828 * RT-boosted. If not then it returns p->normal_prio.
830 static int effective_prio(struct task_struct *p)
832 p->normal_prio = normal_prio(p);
834 * If we are RT tasks or we were boosted to RT priority,
835 * keep the priority unchanged. Otherwise, update priority
836 * to the normal priority:
838 if (!rt_prio(p->prio))
839 return p->normal_prio;
844 * __activate_task - move a task to the runqueue.
846 static void __activate_task(struct task_struct *p, struct rq *rq)
848 struct prio_array *target = rq->active;
851 target = rq->expired;
852 enqueue_task(p, target);
853 inc_nr_running(p, rq);
857 * __activate_idle_task - move idle task to the _front_ of runqueue.
859 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
861 enqueue_task_head(p, rq->active);
862 inc_nr_running(p, rq);
866 * Recalculate p->normal_prio and p->prio after having slept,
867 * updating the sleep-average too:
869 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
871 /* Caller must always ensure 'now >= p->timestamp' */
872 unsigned long sleep_time = now - p->timestamp;
877 if (likely(sleep_time > 0)) {
879 * This ceiling is set to the lowest priority that would allow
880 * a task to be reinserted into the active array on timeslice
883 unsigned long ceiling = INTERACTIVE_SLEEP(p);
885 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
887 * Prevents user tasks from achieving best priority
888 * with one single large enough sleep.
890 p->sleep_avg = ceiling;
892 * Using INTERACTIVE_SLEEP() as a ceiling places a
893 * nice(0) task 1ms sleep away from promotion, and
894 * gives it 700ms to round-robin with no chance of
895 * being demoted. This is more than generous, so
896 * mark this sleep as non-interactive to prevent the
897 * on-runqueue bonus logic from intervening should
898 * this task not receive cpu immediately.
900 p->sleep_type = SLEEP_NONINTERACTIVE;
903 * Tasks waking from uninterruptible sleep are
904 * limited in their sleep_avg rise as they
905 * are likely to be waiting on I/O
907 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
908 if (p->sleep_avg >= ceiling)
910 else if (p->sleep_avg + sleep_time >=
912 p->sleep_avg = ceiling;
918 * This code gives a bonus to interactive tasks.
920 * The boost works by updating the 'average sleep time'
921 * value here, based on ->timestamp. The more time a
922 * task spends sleeping, the higher the average gets -
923 * and the higher the priority boost gets as well.
925 p->sleep_avg += sleep_time;
928 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
929 p->sleep_avg = NS_MAX_SLEEP_AVG;
932 return effective_prio(p);
936 * activate_task - move a task to the runqueue and do priority recalculation
938 * Update all the scheduling statistics stuff. (sleep average
939 * calculation, priority modifiers, etc.)
941 static void activate_task(struct task_struct *p, struct rq *rq, int local)
943 unsigned long long now;
951 /* Compensate for drifting sched_clock */
952 struct rq *this_rq = this_rq();
953 now = (now - this_rq->most_recent_timestamp)
954 + rq->most_recent_timestamp;
959 * Sleep time is in units of nanosecs, so shift by 20 to get a
960 * milliseconds-range estimation of the amount of time that the task
963 if (unlikely(prof_on == SLEEP_PROFILING)) {
964 if (p->state == TASK_UNINTERRUPTIBLE)
965 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
966 (now - p->timestamp) >> 20);
969 p->prio = recalc_task_prio(p, now);
972 * This checks to make sure it's not an uninterruptible task
973 * that is now waking up.
975 if (p->sleep_type == SLEEP_NORMAL) {
977 * Tasks which were woken up by interrupts (ie. hw events)
978 * are most likely of interactive nature. So we give them
979 * the credit of extending their sleep time to the period
980 * of time they spend on the runqueue, waiting for execution
981 * on a CPU, first time around:
984 p->sleep_type = SLEEP_INTERRUPTED;
987 * Normal first-time wakeups get a credit too for
988 * on-runqueue time, but it will be weighted down:
990 p->sleep_type = SLEEP_INTERACTIVE;
995 __activate_task(p, rq);
999 * deactivate_task - remove a task from the runqueue.
1001 static void deactivate_task(struct task_struct *p, struct rq *rq)
1003 dec_nr_running(p, rq);
1004 dequeue_task(p, p->array);
1009 * task_curr - is this task currently executing on a CPU?
1010 * @p: the task in question.
1012 inline int task_curr(const struct task_struct *p)
1014 return cpu_curr(task_cpu(p)) == p;
1017 /* Used instead of source_load when we know the type == 0 */
1018 unsigned long weighted_cpuload(const int cpu)
1020 return cpu_rq(cpu)->raw_weighted_load;
1025 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1027 task_thread_info(p)->cpu = cpu;
1030 struct migration_req {
1031 struct list_head list;
1033 struct task_struct *task;
1036 struct completion done;
1040 * The task's runqueue lock must be held.
1041 * Returns true if you have to wait for migration thread.
1044 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1046 struct rq *rq = task_rq(p);
1049 * If the task is not on a runqueue (and not running), then
1050 * it is sufficient to simply update the task's cpu field.
1052 if (!p->array && !task_running(rq, p)) {
1053 set_task_cpu(p, dest_cpu);
1057 init_completion(&req->done);
1059 req->dest_cpu = dest_cpu;
1060 list_add(&req->list, &rq->migration_queue);
1066 * wait_task_inactive - wait for a thread to unschedule.
1068 * The caller must ensure that the task *will* unschedule sometime soon,
1069 * else this function might spin for a *long* time. This function can't
1070 * be called with interrupts off, or it may introduce deadlock with
1071 * smp_call_function() if an IPI is sent by the same process we are
1072 * waiting to become inactive.
1074 void wait_task_inactive(struct task_struct *p)
1076 unsigned long flags;
1078 struct prio_array *array;
1083 * We do the initial early heuristics without holding
1084 * any task-queue locks at all. We'll only try to get
1085 * the runqueue lock when things look like they will
1091 * If the task is actively running on another CPU
1092 * still, just relax and busy-wait without holding
1095 * NOTE! Since we don't hold any locks, it's not
1096 * even sure that "rq" stays as the right runqueue!
1097 * But we don't care, since "task_running()" will
1098 * return false if the runqueue has changed and p
1099 * is actually now running somewhere else!
1101 while (task_running(rq, p))
1105 * Ok, time to look more closely! We need the rq
1106 * lock now, to be *sure*. If we're wrong, we'll
1107 * just go back and repeat.
1109 rq = task_rq_lock(p, &flags);
1110 running = task_running(rq, p);
1112 task_rq_unlock(rq, &flags);
1115 * Was it really running after all now that we
1116 * checked with the proper locks actually held?
1118 * Oops. Go back and try again..
1120 if (unlikely(running)) {
1126 * It's not enough that it's not actively running,
1127 * it must be off the runqueue _entirely_, and not
1130 * So if it wa still runnable (but just not actively
1131 * running right now), it's preempted, and we should
1132 * yield - it could be a while.
1134 if (unlikely(array)) {
1140 * Ahh, all good. It wasn't running, and it wasn't
1141 * runnable, which means that it will never become
1142 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesnt have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct *p)
1165 if ((cpu != smp_processor_id()) && task_curr(p))
1166 smp_send_reschedule(cpu);
1171 * Return a low guess at the load of a migration-source cpu weighted
1172 * according to the scheduling class and "nice" value.
1174 * We want to under-estimate the load of migration sources, to
1175 * balance conservatively.
1177 static inline unsigned long source_load(int cpu, int type)
1179 struct rq *rq = cpu_rq(cpu);
1182 return rq->raw_weighted_load;
1184 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1188 * Return a high guess at the load of a migration-target cpu weighted
1189 * according to the scheduling class and "nice" value.
1191 static inline unsigned long target_load(int cpu, int type)
1193 struct rq *rq = cpu_rq(cpu);
1196 return rq->raw_weighted_load;
1198 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1202 * Return the average load per task on the cpu's run queue
1204 static inline unsigned long cpu_avg_load_per_task(int cpu)
1206 struct rq *rq = cpu_rq(cpu);
1207 unsigned long n = rq->nr_running;
1209 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1213 * find_idlest_group finds and returns the least busy CPU group within the
1216 static struct sched_group *
1217 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1219 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1220 unsigned long min_load = ULONG_MAX, this_load = 0;
1221 int load_idx = sd->forkexec_idx;
1222 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1225 unsigned long load, avg_load;
1229 /* Skip over this group if it has no CPUs allowed */
1230 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1233 local_group = cpu_isset(this_cpu, group->cpumask);
1235 /* Tally up the load of all CPUs in the group */
1238 for_each_cpu_mask(i, group->cpumask) {
1239 /* Bias balancing toward cpus of our domain */
1241 load = source_load(i, load_idx);
1243 load = target_load(i, load_idx);
1248 /* Adjust by relative CPU power of the group */
1249 avg_load = sg_div_cpu_power(group,
1250 avg_load * SCHED_LOAD_SCALE);
1253 this_load = avg_load;
1255 } else if (avg_load < min_load) {
1256 min_load = avg_load;
1260 group = group->next;
1261 } while (group != sd->groups);
1263 if (!idlest || 100*this_load < imbalance*min_load)
1269 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1272 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1275 unsigned long load, min_load = ULONG_MAX;
1279 /* Traverse only the allowed CPUs */
1280 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1282 for_each_cpu_mask(i, tmp) {
1283 load = weighted_cpuload(i);
1285 if (load < min_load || (load == min_load && i == this_cpu)) {
1295 * sched_balance_self: balance the current task (running on cpu) in domains
1296 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1299 * Balance, ie. select the least loaded group.
1301 * Returns the target CPU number, or the same CPU if no balancing is needed.
1303 * preempt must be disabled.
1305 static int sched_balance_self(int cpu, int flag)
1307 struct task_struct *t = current;
1308 struct sched_domain *tmp, *sd = NULL;
1310 for_each_domain(cpu, tmp) {
1312 * If power savings logic is enabled for a domain, stop there.
1314 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1316 if (tmp->flags & flag)
1322 struct sched_group *group;
1323 int new_cpu, weight;
1325 if (!(sd->flags & flag)) {
1331 group = find_idlest_group(sd, t, cpu);
1337 new_cpu = find_idlest_cpu(group, t, cpu);
1338 if (new_cpu == -1 || new_cpu == cpu) {
1339 /* Now try balancing at a lower domain level of cpu */
1344 /* Now try balancing at a lower domain level of new_cpu */
1347 weight = cpus_weight(span);
1348 for_each_domain(cpu, tmp) {
1349 if (weight <= cpus_weight(tmp->span))
1351 if (tmp->flags & flag)
1354 /* while loop will break here if sd == NULL */
1360 #endif /* CONFIG_SMP */
1363 * wake_idle() will wake a task on an idle cpu if task->cpu is
1364 * not idle and an idle cpu is available. The span of cpus to
1365 * search starts with cpus closest then further out as needed,
1366 * so we always favor a closer, idle cpu.
1368 * Returns the CPU we should wake onto.
1370 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1371 static int wake_idle(int cpu, struct task_struct *p)
1374 struct sched_domain *sd;
1378 * If it is idle, then it is the best cpu to run this task.
1380 * This cpu is also the best, if it has more than one task already.
1381 * Siblings must be also busy(in most cases) as they didn't already
1382 * pickup the extra load from this cpu and hence we need not check
1383 * sibling runqueue info. This will avoid the checks and cache miss
1384 * penalities associated with that.
1386 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1389 for_each_domain(cpu, sd) {
1390 if (sd->flags & SD_WAKE_IDLE) {
1391 cpus_and(tmp, sd->span, p->cpus_allowed);
1392 for_each_cpu_mask(i, tmp) {
1403 static inline int wake_idle(int cpu, struct task_struct *p)
1410 * try_to_wake_up - wake up a thread
1411 * @p: the to-be-woken-up thread
1412 * @state: the mask of task states that can be woken
1413 * @sync: do a synchronous wakeup?
1415 * Put it on the run-queue if it's not already there. The "current"
1416 * thread is always on the run-queue (except when the actual
1417 * re-schedule is in progress), and as such you're allowed to do
1418 * the simpler "current->state = TASK_RUNNING" to mark yourself
1419 * runnable without the overhead of this.
1421 * returns failure only if the task is already active.
1423 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1425 int cpu, this_cpu, success = 0;
1426 unsigned long flags;
1430 struct sched_domain *sd, *this_sd = NULL;
1431 unsigned long load, this_load;
1435 rq = task_rq_lock(p, &flags);
1436 old_state = p->state;
1437 if (!(old_state & state))
1444 this_cpu = smp_processor_id();
1447 if (unlikely(task_running(rq, p)))
1452 schedstat_inc(rq, ttwu_cnt);
1453 if (cpu == this_cpu) {
1454 schedstat_inc(rq, ttwu_local);
1458 for_each_domain(this_cpu, sd) {
1459 if (cpu_isset(cpu, sd->span)) {
1460 schedstat_inc(sd, ttwu_wake_remote);
1466 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1470 * Check for affine wakeup and passive balancing possibilities.
1473 int idx = this_sd->wake_idx;
1474 unsigned int imbalance;
1476 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1478 load = source_load(cpu, idx);
1479 this_load = target_load(this_cpu, idx);
1481 new_cpu = this_cpu; /* Wake to this CPU if we can */
1483 if (this_sd->flags & SD_WAKE_AFFINE) {
1484 unsigned long tl = this_load;
1485 unsigned long tl_per_task;
1487 tl_per_task = cpu_avg_load_per_task(this_cpu);
1490 * If sync wakeup then subtract the (maximum possible)
1491 * effect of the currently running task from the load
1492 * of the current CPU:
1495 tl -= current->load_weight;
1498 tl + target_load(cpu, idx) <= tl_per_task) ||
1499 100*(tl + p->load_weight) <= imbalance*load) {
1501 * This domain has SD_WAKE_AFFINE and
1502 * p is cache cold in this domain, and
1503 * there is no bad imbalance.
1505 schedstat_inc(this_sd, ttwu_move_affine);
1511 * Start passive balancing when half the imbalance_pct
1514 if (this_sd->flags & SD_WAKE_BALANCE) {
1515 if (imbalance*this_load <= 100*load) {
1516 schedstat_inc(this_sd, ttwu_move_balance);
1522 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1524 new_cpu = wake_idle(new_cpu, p);
1525 if (new_cpu != cpu) {
1526 set_task_cpu(p, new_cpu);
1527 task_rq_unlock(rq, &flags);
1528 /* might preempt at this point */
1529 rq = task_rq_lock(p, &flags);
1530 old_state = p->state;
1531 if (!(old_state & state))
1536 this_cpu = smp_processor_id();
1541 #endif /* CONFIG_SMP */
1542 if (old_state == TASK_UNINTERRUPTIBLE) {
1543 rq->nr_uninterruptible--;
1545 * Tasks on involuntary sleep don't earn
1546 * sleep_avg beyond just interactive state.
1548 p->sleep_type = SLEEP_NONINTERACTIVE;
1552 * Tasks that have marked their sleep as noninteractive get
1553 * woken up with their sleep average not weighted in an
1556 if (old_state & TASK_NONINTERACTIVE)
1557 p->sleep_type = SLEEP_NONINTERACTIVE;
1560 activate_task(p, rq, cpu == this_cpu);
1562 * Sync wakeups (i.e. those types of wakeups where the waker
1563 * has indicated that it will leave the CPU in short order)
1564 * don't trigger a preemption, if the woken up task will run on
1565 * this cpu. (in this case the 'I will reschedule' promise of
1566 * the waker guarantees that the freshly woken up task is going
1567 * to be considered on this CPU.)
1569 if (!sync || cpu != this_cpu) {
1570 if (TASK_PREEMPTS_CURR(p, rq))
1571 resched_task(rq->curr);
1576 p->state = TASK_RUNNING;
1578 task_rq_unlock(rq, &flags);
1583 int fastcall wake_up_process(struct task_struct *p)
1585 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1586 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1588 EXPORT_SYMBOL(wake_up_process);
1590 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1592 return try_to_wake_up(p, state, 0);
1595 static void task_running_tick(struct rq *rq, struct task_struct *p);
1597 * Perform scheduler related setup for a newly forked process p.
1598 * p is forked by current.
1600 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1602 int cpu = get_cpu();
1605 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1607 set_task_cpu(p, cpu);
1610 * We mark the process as running here, but have not actually
1611 * inserted it onto the runqueue yet. This guarantees that
1612 * nobody will actually run it, and a signal or other external
1613 * event cannot wake it up and insert it on the runqueue either.
1615 p->state = TASK_RUNNING;
1618 * Make sure we do not leak PI boosting priority to the child:
1620 p->prio = current->normal_prio;
1622 INIT_LIST_HEAD(&p->run_list);
1624 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1625 if (unlikely(sched_info_on()))
1626 memset(&p->sched_info, 0, sizeof(p->sched_info));
1628 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1631 #ifdef CONFIG_PREEMPT
1632 /* Want to start with kernel preemption disabled. */
1633 task_thread_info(p)->preempt_count = 1;
1636 * Share the timeslice between parent and child, thus the
1637 * total amount of pending timeslices in the system doesn't change,
1638 * resulting in more scheduling fairness.
1640 local_irq_disable();
1641 p->time_slice = (current->time_slice + 1) >> 1;
1643 * The remainder of the first timeslice might be recovered by
1644 * the parent if the child exits early enough.
1646 p->first_time_slice = 1;
1647 current->time_slice >>= 1;
1648 p->timestamp = sched_clock();
1649 if (unlikely(!current->time_slice)) {
1651 * This case is rare, it happens when the parent has only
1652 * a single jiffy left from its timeslice. Taking the
1653 * runqueue lock is not a problem.
1655 current->time_slice = 1;
1656 task_running_tick(cpu_rq(cpu), current);
1663 * wake_up_new_task - wake up a newly created task for the first time.
1665 * This function will do some initial scheduler statistics housekeeping
1666 * that must be done for every newly created context, then puts the task
1667 * on the runqueue and wakes it.
1669 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1671 struct rq *rq, *this_rq;
1672 unsigned long flags;
1675 rq = task_rq_lock(p, &flags);
1676 BUG_ON(p->state != TASK_RUNNING);
1677 this_cpu = smp_processor_id();
1681 * We decrease the sleep average of forking parents
1682 * and children as well, to keep max-interactive tasks
1683 * from forking tasks that are max-interactive. The parent
1684 * (current) is done further down, under its lock.
1686 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1687 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1689 p->prio = effective_prio(p);
1691 if (likely(cpu == this_cpu)) {
1692 if (!(clone_flags & CLONE_VM)) {
1694 * The VM isn't cloned, so we're in a good position to
1695 * do child-runs-first in anticipation of an exec. This
1696 * usually avoids a lot of COW overhead.
1698 if (unlikely(!current->array))
1699 __activate_task(p, rq);
1701 p->prio = current->prio;
1702 p->normal_prio = current->normal_prio;
1703 list_add_tail(&p->run_list, ¤t->run_list);
1704 p->array = current->array;
1705 p->array->nr_active++;
1706 inc_nr_running(p, rq);
1710 /* Run child last */
1711 __activate_task(p, rq);
1713 * We skip the following code due to cpu == this_cpu
1715 * task_rq_unlock(rq, &flags);
1716 * this_rq = task_rq_lock(current, &flags);
1720 this_rq = cpu_rq(this_cpu);
1723 * Not the local CPU - must adjust timestamp. This should
1724 * get optimised away in the !CONFIG_SMP case.
1726 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1727 + rq->most_recent_timestamp;
1728 __activate_task(p, rq);
1729 if (TASK_PREEMPTS_CURR(p, rq))
1730 resched_task(rq->curr);
1733 * Parent and child are on different CPUs, now get the
1734 * parent runqueue to update the parent's ->sleep_avg:
1736 task_rq_unlock(rq, &flags);
1737 this_rq = task_rq_lock(current, &flags);
1739 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1740 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1741 task_rq_unlock(this_rq, &flags);
1745 * prepare_task_switch - prepare to switch tasks
1746 * @rq: the runqueue preparing to switch
1747 * @next: the task we are going to switch to.
1749 * This is called with the rq lock held and interrupts off. It must
1750 * be paired with a subsequent finish_task_switch after the context
1753 * prepare_task_switch sets up locking and calls architecture specific
1756 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1758 prepare_lock_switch(rq, next);
1759 prepare_arch_switch(next);
1763 * finish_task_switch - clean up after a task-switch
1764 * @rq: runqueue associated with task-switch
1765 * @prev: the thread we just switched away from.
1767 * finish_task_switch must be called after the context switch, paired
1768 * with a prepare_task_switch call before the context switch.
1769 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1770 * and do any other architecture-specific cleanup actions.
1772 * Note that we may have delayed dropping an mm in context_switch(). If
1773 * so, we finish that here outside of the runqueue lock. (Doing it
1774 * with the lock held can cause deadlocks; see schedule() for
1777 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1778 __releases(rq->lock)
1780 struct mm_struct *mm = rq->prev_mm;
1786 * A task struct has one reference for the use as "current".
1787 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1788 * schedule one last time. The schedule call will never return, and
1789 * the scheduled task must drop that reference.
1790 * The test for TASK_DEAD must occur while the runqueue locks are
1791 * still held, otherwise prev could be scheduled on another cpu, die
1792 * there before we look at prev->state, and then the reference would
1794 * Manfred Spraul <manfred@colorfullife.com>
1796 prev_state = prev->state;
1797 finish_arch_switch(prev);
1798 finish_lock_switch(rq, prev);
1801 if (unlikely(prev_state == TASK_DEAD)) {
1803 * Remove function-return probe instances associated with this
1804 * task and put them back on the free list.
1806 kprobe_flush_task(prev);
1807 put_task_struct(prev);
1812 * schedule_tail - first thing a freshly forked thread must call.
1813 * @prev: the thread we just switched away from.
1815 asmlinkage void schedule_tail(struct task_struct *prev)
1816 __releases(rq->lock)
1818 struct rq *rq = this_rq();
1820 finish_task_switch(rq, prev);
1821 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1822 /* In this case, finish_task_switch does not reenable preemption */
1825 if (current->set_child_tid)
1826 put_user(current->pid, current->set_child_tid);
1830 * context_switch - switch to the new MM and the new
1831 * thread's register state.
1833 static inline struct task_struct *
1834 context_switch(struct rq *rq, struct task_struct *prev,
1835 struct task_struct *next)
1837 struct mm_struct *mm = next->mm;
1838 struct mm_struct *oldmm = prev->active_mm;
1841 * For paravirt, this is coupled with an exit in switch_to to
1842 * combine the page table reload and the switch backend into
1845 arch_enter_lazy_cpu_mode();
1848 next->active_mm = oldmm;
1849 atomic_inc(&oldmm->mm_count);
1850 enter_lazy_tlb(oldmm, next);
1852 switch_mm(oldmm, mm, next);
1855 prev->active_mm = NULL;
1856 WARN_ON(rq->prev_mm);
1857 rq->prev_mm = oldmm;
1860 * Since the runqueue lock will be released by the next
1861 * task (which is an invalid locking op but in the case
1862 * of the scheduler it's an obvious special-case), so we
1863 * do an early lockdep release here:
1865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1866 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1869 /* Here we just switch the register state and the stack. */
1870 switch_to(prev, next, prev);
1876 * nr_running, nr_uninterruptible and nr_context_switches:
1878 * externally visible scheduler statistics: current number of runnable
1879 * threads, current number of uninterruptible-sleeping threads, total
1880 * number of context switches performed since bootup.
1882 unsigned long nr_running(void)
1884 unsigned long i, sum = 0;
1886 for_each_online_cpu(i)
1887 sum += cpu_rq(i)->nr_running;
1892 unsigned long nr_uninterruptible(void)
1894 unsigned long i, sum = 0;
1896 for_each_possible_cpu(i)
1897 sum += cpu_rq(i)->nr_uninterruptible;
1900 * Since we read the counters lockless, it might be slightly
1901 * inaccurate. Do not allow it to go below zero though:
1903 if (unlikely((long)sum < 0))
1909 unsigned long long nr_context_switches(void)
1912 unsigned long long sum = 0;
1914 for_each_possible_cpu(i)
1915 sum += cpu_rq(i)->nr_switches;
1920 unsigned long nr_iowait(void)
1922 unsigned long i, sum = 0;
1924 for_each_possible_cpu(i)
1925 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1930 unsigned long nr_active(void)
1932 unsigned long i, running = 0, uninterruptible = 0;
1934 for_each_online_cpu(i) {
1935 running += cpu_rq(i)->nr_running;
1936 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1939 if (unlikely((long)uninterruptible < 0))
1940 uninterruptible = 0;
1942 return running + uninterruptible;
1948 * Is this task likely cache-hot:
1951 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1953 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1957 * double_rq_lock - safely lock two runqueues
1959 * Note this does not disable interrupts like task_rq_lock,
1960 * you need to do so manually before calling.
1962 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1963 __acquires(rq1->lock)
1964 __acquires(rq2->lock)
1966 BUG_ON(!irqs_disabled());
1968 spin_lock(&rq1->lock);
1969 __acquire(rq2->lock); /* Fake it out ;) */
1972 spin_lock(&rq1->lock);
1973 spin_lock(&rq2->lock);
1975 spin_lock(&rq2->lock);
1976 spin_lock(&rq1->lock);
1982 * double_rq_unlock - safely unlock two runqueues
1984 * Note this does not restore interrupts like task_rq_unlock,
1985 * you need to do so manually after calling.
1987 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1988 __releases(rq1->lock)
1989 __releases(rq2->lock)
1991 spin_unlock(&rq1->lock);
1993 spin_unlock(&rq2->lock);
1995 __release(rq2->lock);
1999 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2001 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2002 __releases(this_rq->lock)
2003 __acquires(busiest->lock)
2004 __acquires(this_rq->lock)
2006 if (unlikely(!irqs_disabled())) {
2007 /* printk() doesn't work good under rq->lock */
2008 spin_unlock(&this_rq->lock);
2011 if (unlikely(!spin_trylock(&busiest->lock))) {
2012 if (busiest < this_rq) {
2013 spin_unlock(&this_rq->lock);
2014 spin_lock(&busiest->lock);
2015 spin_lock(&this_rq->lock);
2017 spin_lock(&busiest->lock);
2022 * If dest_cpu is allowed for this process, migrate the task to it.
2023 * This is accomplished by forcing the cpu_allowed mask to only
2024 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2025 * the cpu_allowed mask is restored.
2027 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2029 struct migration_req req;
2030 unsigned long flags;
2033 rq = task_rq_lock(p, &flags);
2034 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2035 || unlikely(cpu_is_offline(dest_cpu)))
2038 /* force the process onto the specified CPU */
2039 if (migrate_task(p, dest_cpu, &req)) {
2040 /* Need to wait for migration thread (might exit: take ref). */
2041 struct task_struct *mt = rq->migration_thread;
2043 get_task_struct(mt);
2044 task_rq_unlock(rq, &flags);
2045 wake_up_process(mt);
2046 put_task_struct(mt);
2047 wait_for_completion(&req.done);
2052 task_rq_unlock(rq, &flags);
2056 * sched_exec - execve() is a valuable balancing opportunity, because at
2057 * this point the task has the smallest effective memory and cache footprint.
2059 void sched_exec(void)
2061 int new_cpu, this_cpu = get_cpu();
2062 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2064 if (new_cpu != this_cpu)
2065 sched_migrate_task(current, new_cpu);
2069 * pull_task - move a task from a remote runqueue to the local runqueue.
2070 * Both runqueues must be locked.
2072 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2073 struct task_struct *p, struct rq *this_rq,
2074 struct prio_array *this_array, int this_cpu)
2076 dequeue_task(p, src_array);
2077 dec_nr_running(p, src_rq);
2078 set_task_cpu(p, this_cpu);
2079 inc_nr_running(p, this_rq);
2080 enqueue_task(p, this_array);
2081 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2082 + this_rq->most_recent_timestamp;
2084 * Note that idle threads have a prio of MAX_PRIO, for this test
2085 * to be always true for them.
2087 if (TASK_PREEMPTS_CURR(p, this_rq))
2088 resched_task(this_rq->curr);
2092 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2095 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2096 struct sched_domain *sd, enum cpu_idle_type idle,
2100 * We do not migrate tasks that are:
2101 * 1) running (obviously), or
2102 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2103 * 3) are cache-hot on their current CPU.
2105 if (!cpu_isset(this_cpu, p->cpus_allowed))
2109 if (task_running(rq, p))
2113 * Aggressive migration if:
2114 * 1) task is cache cold, or
2115 * 2) too many balance attempts have failed.
2118 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2119 #ifdef CONFIG_SCHEDSTATS
2120 if (task_hot(p, rq->most_recent_timestamp, sd))
2121 schedstat_inc(sd, lb_hot_gained[idle]);
2126 if (task_hot(p, rq->most_recent_timestamp, sd))
2131 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2134 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2135 * load from busiest to this_rq, as part of a balancing operation within
2136 * "domain". Returns the number of tasks moved.
2138 * Called with both runqueues locked.
2140 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2141 unsigned long max_nr_move, unsigned long max_load_move,
2142 struct sched_domain *sd, enum cpu_idle_type idle,
2145 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2146 best_prio_seen, skip_for_load;
2147 struct prio_array *array, *dst_array;
2148 struct list_head *head, *curr;
2149 struct task_struct *tmp;
2152 if (max_nr_move == 0 || max_load_move == 0)
2155 rem_load_move = max_load_move;
2157 this_best_prio = rq_best_prio(this_rq);
2158 best_prio = rq_best_prio(busiest);
2160 * Enable handling of the case where there is more than one task
2161 * with the best priority. If the current running task is one
2162 * of those with prio==best_prio we know it won't be moved
2163 * and therefore it's safe to override the skip (based on load) of
2164 * any task we find with that prio.
2166 best_prio_seen = best_prio == busiest->curr->prio;
2169 * We first consider expired tasks. Those will likely not be
2170 * executed in the near future, and they are most likely to
2171 * be cache-cold, thus switching CPUs has the least effect
2174 if (busiest->expired->nr_active) {
2175 array = busiest->expired;
2176 dst_array = this_rq->expired;
2178 array = busiest->active;
2179 dst_array = this_rq->active;
2183 /* Start searching at priority 0: */
2187 idx = sched_find_first_bit(array->bitmap);
2189 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2190 if (idx >= MAX_PRIO) {
2191 if (array == busiest->expired && busiest->active->nr_active) {
2192 array = busiest->active;
2193 dst_array = this_rq->active;
2199 head = array->queue + idx;
2202 tmp = list_entry(curr, struct task_struct, run_list);
2207 * To help distribute high priority tasks accross CPUs we don't
2208 * skip a task if it will be the highest priority task (i.e. smallest
2209 * prio value) on its new queue regardless of its load weight
2211 skip_for_load = tmp->load_weight > rem_load_move;
2212 if (skip_for_load && idx < this_best_prio)
2213 skip_for_load = !best_prio_seen && idx == best_prio;
2214 if (skip_for_load ||
2215 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2217 best_prio_seen |= idx == best_prio;
2224 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2226 rem_load_move -= tmp->load_weight;
2229 * We only want to steal up to the prescribed number of tasks
2230 * and the prescribed amount of weighted load.
2232 if (pulled < max_nr_move && rem_load_move > 0) {
2233 if (idx < this_best_prio)
2234 this_best_prio = idx;
2242 * Right now, this is the only place pull_task() is called,
2243 * so we can safely collect pull_task() stats here rather than
2244 * inside pull_task().
2246 schedstat_add(sd, lb_gained[idle], pulled);
2249 *all_pinned = pinned;
2254 * find_busiest_group finds and returns the busiest CPU group within the
2255 * domain. It calculates and returns the amount of weighted load which
2256 * should be moved to restore balance via the imbalance parameter.
2258 static struct sched_group *
2259 find_busiest_group(struct sched_domain *sd, int this_cpu,
2260 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2261 cpumask_t *cpus, int *balance)
2263 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2264 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2265 unsigned long max_pull;
2266 unsigned long busiest_load_per_task, busiest_nr_running;
2267 unsigned long this_load_per_task, this_nr_running;
2269 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2270 int power_savings_balance = 1;
2271 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2272 unsigned long min_nr_running = ULONG_MAX;
2273 struct sched_group *group_min = NULL, *group_leader = NULL;
2276 max_load = this_load = total_load = total_pwr = 0;
2277 busiest_load_per_task = busiest_nr_running = 0;
2278 this_load_per_task = this_nr_running = 0;
2279 if (idle == CPU_NOT_IDLE)
2280 load_idx = sd->busy_idx;
2281 else if (idle == CPU_NEWLY_IDLE)
2282 load_idx = sd->newidle_idx;
2284 load_idx = sd->idle_idx;
2287 unsigned long load, group_capacity;
2290 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2291 unsigned long sum_nr_running, sum_weighted_load;
2293 local_group = cpu_isset(this_cpu, group->cpumask);
2296 balance_cpu = first_cpu(group->cpumask);
2298 /* Tally up the load of all CPUs in the group */
2299 sum_weighted_load = sum_nr_running = avg_load = 0;
2301 for_each_cpu_mask(i, group->cpumask) {
2304 if (!cpu_isset(i, *cpus))
2309 if (*sd_idle && !idle_cpu(i))
2312 /* Bias balancing toward cpus of our domain */
2314 if (idle_cpu(i) && !first_idle_cpu) {
2319 load = target_load(i, load_idx);
2321 load = source_load(i, load_idx);
2324 sum_nr_running += rq->nr_running;
2325 sum_weighted_load += rq->raw_weighted_load;
2329 * First idle cpu or the first cpu(busiest) in this sched group
2330 * is eligible for doing load balancing at this and above
2333 if (local_group && balance_cpu != this_cpu && balance) {
2338 total_load += avg_load;
2339 total_pwr += group->__cpu_power;
2341 /* Adjust by relative CPU power of the group */
2342 avg_load = sg_div_cpu_power(group,
2343 avg_load * SCHED_LOAD_SCALE);
2345 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2348 this_load = avg_load;
2350 this_nr_running = sum_nr_running;
2351 this_load_per_task = sum_weighted_load;
2352 } else if (avg_load > max_load &&
2353 sum_nr_running > group_capacity) {
2354 max_load = avg_load;
2356 busiest_nr_running = sum_nr_running;
2357 busiest_load_per_task = sum_weighted_load;
2360 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2362 * Busy processors will not participate in power savings
2365 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2369 * If the local group is idle or completely loaded
2370 * no need to do power savings balance at this domain
2372 if (local_group && (this_nr_running >= group_capacity ||
2374 power_savings_balance = 0;
2377 * If a group is already running at full capacity or idle,
2378 * don't include that group in power savings calculations
2380 if (!power_savings_balance || sum_nr_running >= group_capacity
2385 * Calculate the group which has the least non-idle load.
2386 * This is the group from where we need to pick up the load
2389 if ((sum_nr_running < min_nr_running) ||
2390 (sum_nr_running == min_nr_running &&
2391 first_cpu(group->cpumask) <
2392 first_cpu(group_min->cpumask))) {
2394 min_nr_running = sum_nr_running;
2395 min_load_per_task = sum_weighted_load /
2400 * Calculate the group which is almost near its
2401 * capacity but still has some space to pick up some load
2402 * from other group and save more power
2404 if (sum_nr_running <= group_capacity - 1) {
2405 if (sum_nr_running > leader_nr_running ||
2406 (sum_nr_running == leader_nr_running &&
2407 first_cpu(group->cpumask) >
2408 first_cpu(group_leader->cpumask))) {
2409 group_leader = group;
2410 leader_nr_running = sum_nr_running;
2415 group = group->next;
2416 } while (group != sd->groups);
2418 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2421 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2423 if (this_load >= avg_load ||
2424 100*max_load <= sd->imbalance_pct*this_load)
2427 busiest_load_per_task /= busiest_nr_running;
2429 * We're trying to get all the cpus to the average_load, so we don't
2430 * want to push ourselves above the average load, nor do we wish to
2431 * reduce the max loaded cpu below the average load, as either of these
2432 * actions would just result in more rebalancing later, and ping-pong
2433 * tasks around. Thus we look for the minimum possible imbalance.
2434 * Negative imbalances (*we* are more loaded than anyone else) will
2435 * be counted as no imbalance for these purposes -- we can't fix that
2436 * by pulling tasks to us. Be careful of negative numbers as they'll
2437 * appear as very large values with unsigned longs.
2439 if (max_load <= busiest_load_per_task)
2443 * In the presence of smp nice balancing, certain scenarios can have
2444 * max load less than avg load(as we skip the groups at or below
2445 * its cpu_power, while calculating max_load..)
2447 if (max_load < avg_load) {
2449 goto small_imbalance;
2452 /* Don't want to pull so many tasks that a group would go idle */
2453 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2455 /* How much load to actually move to equalise the imbalance */
2456 *imbalance = min(max_pull * busiest->__cpu_power,
2457 (avg_load - this_load) * this->__cpu_power)
2461 * if *imbalance is less than the average load per runnable task
2462 * there is no gaurantee that any tasks will be moved so we'll have
2463 * a think about bumping its value to force at least one task to be
2466 if (*imbalance < busiest_load_per_task) {
2467 unsigned long tmp, pwr_now, pwr_move;
2471 pwr_move = pwr_now = 0;
2473 if (this_nr_running) {
2474 this_load_per_task /= this_nr_running;
2475 if (busiest_load_per_task > this_load_per_task)
2478 this_load_per_task = SCHED_LOAD_SCALE;
2480 if (max_load - this_load >= busiest_load_per_task * imbn) {
2481 *imbalance = busiest_load_per_task;
2486 * OK, we don't have enough imbalance to justify moving tasks,
2487 * however we may be able to increase total CPU power used by
2491 pwr_now += busiest->__cpu_power *
2492 min(busiest_load_per_task, max_load);
2493 pwr_now += this->__cpu_power *
2494 min(this_load_per_task, this_load);
2495 pwr_now /= SCHED_LOAD_SCALE;
2497 /* Amount of load we'd subtract */
2498 tmp = sg_div_cpu_power(busiest,
2499 busiest_load_per_task * SCHED_LOAD_SCALE);
2501 pwr_move += busiest->__cpu_power *
2502 min(busiest_load_per_task, max_load - tmp);
2504 /* Amount of load we'd add */
2505 if (max_load * busiest->__cpu_power <
2506 busiest_load_per_task * SCHED_LOAD_SCALE)
2507 tmp = sg_div_cpu_power(this,
2508 max_load * busiest->__cpu_power);
2510 tmp = sg_div_cpu_power(this,
2511 busiest_load_per_task * SCHED_LOAD_SCALE);
2512 pwr_move += this->__cpu_power *
2513 min(this_load_per_task, this_load + tmp);
2514 pwr_move /= SCHED_LOAD_SCALE;
2516 /* Move if we gain throughput */
2517 if (pwr_move <= pwr_now)
2520 *imbalance = busiest_load_per_task;
2526 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2527 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2530 if (this == group_leader && group_leader != group_min) {
2531 *imbalance = min_load_per_task;
2541 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2544 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2545 unsigned long imbalance, cpumask_t *cpus)
2547 struct rq *busiest = NULL, *rq;
2548 unsigned long max_load = 0;
2551 for_each_cpu_mask(i, group->cpumask) {
2553 if (!cpu_isset(i, *cpus))
2558 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2561 if (rq->raw_weighted_load > max_load) {
2562 max_load = rq->raw_weighted_load;
2571 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2572 * so long as it is large enough.
2574 #define MAX_PINNED_INTERVAL 512
2576 static inline unsigned long minus_1_or_zero(unsigned long n)
2578 return n > 0 ? n - 1 : 0;
2582 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2583 * tasks if there is an imbalance.
2585 static int load_balance(int this_cpu, struct rq *this_rq,
2586 struct sched_domain *sd, enum cpu_idle_type idle,
2589 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2590 struct sched_group *group;
2591 unsigned long imbalance;
2593 cpumask_t cpus = CPU_MASK_ALL;
2594 unsigned long flags;
2597 * When power savings policy is enabled for the parent domain, idle
2598 * sibling can pick up load irrespective of busy siblings. In this case,
2599 * let the state of idle sibling percolate up as IDLE, instead of
2600 * portraying it as CPU_NOT_IDLE.
2602 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2603 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2606 schedstat_inc(sd, lb_cnt[idle]);
2609 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2616 schedstat_inc(sd, lb_nobusyg[idle]);
2620 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2622 schedstat_inc(sd, lb_nobusyq[idle]);
2626 BUG_ON(busiest == this_rq);
2628 schedstat_add(sd, lb_imbalance[idle], imbalance);
2631 if (busiest->nr_running > 1) {
2633 * Attempt to move tasks. If find_busiest_group has found
2634 * an imbalance but busiest->nr_running <= 1, the group is
2635 * still unbalanced. nr_moved simply stays zero, so it is
2636 * correctly treated as an imbalance.
2638 local_irq_save(flags);
2639 double_rq_lock(this_rq, busiest);
2640 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2641 minus_1_or_zero(busiest->nr_running),
2642 imbalance, sd, idle, &all_pinned);
2643 double_rq_unlock(this_rq, busiest);
2644 local_irq_restore(flags);
2647 * some other cpu did the load balance for us.
2649 if (nr_moved && this_cpu != smp_processor_id())
2650 resched_cpu(this_cpu);
2652 /* All tasks on this runqueue were pinned by CPU affinity */
2653 if (unlikely(all_pinned)) {
2654 cpu_clear(cpu_of(busiest), cpus);
2655 if (!cpus_empty(cpus))
2662 schedstat_inc(sd, lb_failed[idle]);
2663 sd->nr_balance_failed++;
2665 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2667 spin_lock_irqsave(&busiest->lock, flags);
2669 /* don't kick the migration_thread, if the curr
2670 * task on busiest cpu can't be moved to this_cpu
2672 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2673 spin_unlock_irqrestore(&busiest->lock, flags);
2675 goto out_one_pinned;
2678 if (!busiest->active_balance) {
2679 busiest->active_balance = 1;
2680 busiest->push_cpu = this_cpu;
2683 spin_unlock_irqrestore(&busiest->lock, flags);
2685 wake_up_process(busiest->migration_thread);
2688 * We've kicked active balancing, reset the failure
2691 sd->nr_balance_failed = sd->cache_nice_tries+1;
2694 sd->nr_balance_failed = 0;
2696 if (likely(!active_balance)) {
2697 /* We were unbalanced, so reset the balancing interval */
2698 sd->balance_interval = sd->min_interval;
2701 * If we've begun active balancing, start to back off. This
2702 * case may not be covered by the all_pinned logic if there
2703 * is only 1 task on the busy runqueue (because we don't call
2706 if (sd->balance_interval < sd->max_interval)
2707 sd->balance_interval *= 2;
2710 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2711 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2716 schedstat_inc(sd, lb_balanced[idle]);
2718 sd->nr_balance_failed = 0;
2721 /* tune up the balancing interval */
2722 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2723 (sd->balance_interval < sd->max_interval))
2724 sd->balance_interval *= 2;
2726 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2727 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2733 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2734 * tasks if there is an imbalance.
2736 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2737 * this_rq is locked.
2740 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2742 struct sched_group *group;
2743 struct rq *busiest = NULL;
2744 unsigned long imbalance;
2747 cpumask_t cpus = CPU_MASK_ALL;
2750 * When power savings policy is enabled for the parent domain, idle
2751 * sibling can pick up load irrespective of busy siblings. In this case,
2752 * let the state of idle sibling percolate up as IDLE, instead of
2753 * portraying it as CPU_NOT_IDLE.
2755 if (sd->flags & SD_SHARE_CPUPOWER &&
2756 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2759 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2761 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2762 &sd_idle, &cpus, NULL);
2764 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2768 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2771 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2775 BUG_ON(busiest == this_rq);
2777 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2780 if (busiest->nr_running > 1) {
2781 /* Attempt to move tasks */
2782 double_lock_balance(this_rq, busiest);
2783 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2784 minus_1_or_zero(busiest->nr_running),
2785 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2786 spin_unlock(&busiest->lock);
2789 cpu_clear(cpu_of(busiest), cpus);
2790 if (!cpus_empty(cpus))
2796 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2797 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2798 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2801 sd->nr_balance_failed = 0;
2806 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2807 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2808 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2810 sd->nr_balance_failed = 0;
2816 * idle_balance is called by schedule() if this_cpu is about to become
2817 * idle. Attempts to pull tasks from other CPUs.
2819 static void idle_balance(int this_cpu, struct rq *this_rq)
2821 struct sched_domain *sd;
2822 int pulled_task = 0;
2823 unsigned long next_balance = jiffies + 60 * HZ;
2825 for_each_domain(this_cpu, sd) {
2826 unsigned long interval;
2828 if (!(sd->flags & SD_LOAD_BALANCE))
2831 if (sd->flags & SD_BALANCE_NEWIDLE)
2832 /* If we've pulled tasks over stop searching: */
2833 pulled_task = load_balance_newidle(this_cpu,
2836 interval = msecs_to_jiffies(sd->balance_interval);
2837 if (time_after(next_balance, sd->last_balance + interval))
2838 next_balance = sd->last_balance + interval;
2844 * We are going idle. next_balance may be set based on
2845 * a busy processor. So reset next_balance.
2847 this_rq->next_balance = next_balance;
2851 * active_load_balance is run by migration threads. It pushes running tasks
2852 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2853 * running on each physical CPU where possible, and avoids physical /
2854 * logical imbalances.
2856 * Called with busiest_rq locked.
2858 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2860 int target_cpu = busiest_rq->push_cpu;
2861 struct sched_domain *sd;
2862 struct rq *target_rq;
2864 /* Is there any task to move? */
2865 if (busiest_rq->nr_running <= 1)
2868 target_rq = cpu_rq(target_cpu);
2871 * This condition is "impossible", if it occurs
2872 * we need to fix it. Originally reported by
2873 * Bjorn Helgaas on a 128-cpu setup.
2875 BUG_ON(busiest_rq == target_rq);
2877 /* move a task from busiest_rq to target_rq */
2878 double_lock_balance(busiest_rq, target_rq);
2880 /* Search for an sd spanning us and the target CPU. */
2881 for_each_domain(target_cpu, sd) {
2882 if ((sd->flags & SD_LOAD_BALANCE) &&
2883 cpu_isset(busiest_cpu, sd->span))
2888 schedstat_inc(sd, alb_cnt);
2890 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2891 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2893 schedstat_inc(sd, alb_pushed);
2895 schedstat_inc(sd, alb_failed);
2897 spin_unlock(&target_rq->lock);
2900 static void update_load(struct rq *this_rq)
2902 unsigned long this_load;
2903 unsigned int i, scale;
2905 this_load = this_rq->raw_weighted_load;
2907 /* Update our load: */
2908 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2909 unsigned long old_load, new_load;
2911 /* scale is effectively 1 << i now, and >> i divides by scale */
2913 old_load = this_rq->cpu_load[i];
2914 new_load = this_load;
2916 * Round up the averaging division if load is increasing. This
2917 * prevents us from getting stuck on 9 if the load is 10, for
2920 if (new_load > old_load)
2921 new_load += scale-1;
2922 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2928 atomic_t load_balancer;
2930 } nohz ____cacheline_aligned = {
2931 .load_balancer = ATOMIC_INIT(-1),
2932 .cpu_mask = CPU_MASK_NONE,
2936 * This routine will try to nominate the ilb (idle load balancing)
2937 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2938 * load balancing on behalf of all those cpus. If all the cpus in the system
2939 * go into this tickless mode, then there will be no ilb owner (as there is
2940 * no need for one) and all the cpus will sleep till the next wakeup event
2943 * For the ilb owner, tick is not stopped. And this tick will be used
2944 * for idle load balancing. ilb owner will still be part of
2947 * While stopping the tick, this cpu will become the ilb owner if there
2948 * is no other owner. And will be the owner till that cpu becomes busy
2949 * or if all cpus in the system stop their ticks at which point
2950 * there is no need for ilb owner.
2952 * When the ilb owner becomes busy, it nominates another owner, during the
2953 * next busy scheduler_tick()
2955 int select_nohz_load_balancer(int stop_tick)
2957 int cpu = smp_processor_id();
2960 cpu_set(cpu, nohz.cpu_mask);
2961 cpu_rq(cpu)->in_nohz_recently = 1;
2964 * If we are going offline and still the leader, give up!
2966 if (cpu_is_offline(cpu) &&
2967 atomic_read(&nohz.load_balancer) == cpu) {
2968 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2973 /* time for ilb owner also to sleep */
2974 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2975 if (atomic_read(&nohz.load_balancer) == cpu)
2976 atomic_set(&nohz.load_balancer, -1);
2980 if (atomic_read(&nohz.load_balancer) == -1) {
2981 /* make me the ilb owner */
2982 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2984 } else if (atomic_read(&nohz.load_balancer) == cpu)
2987 if (!cpu_isset(cpu, nohz.cpu_mask))
2990 cpu_clear(cpu, nohz.cpu_mask);
2992 if (atomic_read(&nohz.load_balancer) == cpu)
2993 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3000 static DEFINE_SPINLOCK(balancing);
3003 * It checks each scheduling domain to see if it is due to be balanced,
3004 * and initiates a balancing operation if so.
3006 * Balancing parameters are set up in arch_init_sched_domains.
3008 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3011 struct rq *rq = cpu_rq(cpu);
3012 unsigned long interval;
3013 struct sched_domain *sd;
3014 /* Earliest time when we have to do rebalance again */
3015 unsigned long next_balance = jiffies + 60*HZ;
3017 for_each_domain(cpu, sd) {
3018 if (!(sd->flags & SD_LOAD_BALANCE))
3021 interval = sd->balance_interval;
3022 if (idle != CPU_IDLE)
3023 interval *= sd->busy_factor;
3025 /* scale ms to jiffies */
3026 interval = msecs_to_jiffies(interval);
3027 if (unlikely(!interval))
3030 if (sd->flags & SD_SERIALIZE) {
3031 if (!spin_trylock(&balancing))
3035 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3036 if (load_balance(cpu, rq, sd, idle, &balance)) {
3038 * We've pulled tasks over so either we're no
3039 * longer idle, or one of our SMT siblings is
3042 idle = CPU_NOT_IDLE;
3044 sd->last_balance = jiffies;
3046 if (sd->flags & SD_SERIALIZE)
3047 spin_unlock(&balancing);
3049 if (time_after(next_balance, sd->last_balance + interval))
3050 next_balance = sd->last_balance + interval;
3053 * Stop the load balance at this level. There is another
3054 * CPU in our sched group which is doing load balancing more
3060 rq->next_balance = next_balance;
3064 * run_rebalance_domains is triggered when needed from the scheduler tick.
3065 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3066 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3068 static void run_rebalance_domains(struct softirq_action *h)
3070 int local_cpu = smp_processor_id();
3071 struct rq *local_rq = cpu_rq(local_cpu);
3072 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3074 rebalance_domains(local_cpu, idle);
3078 * If this cpu is the owner for idle load balancing, then do the
3079 * balancing on behalf of the other idle cpus whose ticks are
3082 if (local_rq->idle_at_tick &&
3083 atomic_read(&nohz.load_balancer) == local_cpu) {
3084 cpumask_t cpus = nohz.cpu_mask;
3088 cpu_clear(local_cpu, cpus);
3089 for_each_cpu_mask(balance_cpu, cpus) {
3091 * If this cpu gets work to do, stop the load balancing
3092 * work being done for other cpus. Next load
3093 * balancing owner will pick it up.
3098 rebalance_domains(balance_cpu, CPU_IDLE);
3100 rq = cpu_rq(balance_cpu);
3101 if (time_after(local_rq->next_balance, rq->next_balance))
3102 local_rq->next_balance = rq->next_balance;
3109 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3111 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3112 * idle load balancing owner or decide to stop the periodic load balancing,
3113 * if the whole system is idle.
3115 static inline void trigger_load_balance(int cpu)
3117 struct rq *rq = cpu_rq(cpu);
3120 * If we were in the nohz mode recently and busy at the current
3121 * scheduler tick, then check if we need to nominate new idle
3124 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3125 rq->in_nohz_recently = 0;
3127 if (atomic_read(&nohz.load_balancer) == cpu) {
3128 cpu_clear(cpu, nohz.cpu_mask);
3129 atomic_set(&nohz.load_balancer, -1);
3132 if (atomic_read(&nohz.load_balancer) == -1) {
3134 * simple selection for now: Nominate the
3135 * first cpu in the nohz list to be the next
3138 * TBD: Traverse the sched domains and nominate
3139 * the nearest cpu in the nohz.cpu_mask.
3141 int ilb = first_cpu(nohz.cpu_mask);
3149 * If this cpu is idle and doing idle load balancing for all the
3150 * cpus with ticks stopped, is it time for that to stop?
3152 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3153 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3159 * If this cpu is idle and the idle load balancing is done by
3160 * someone else, then no need raise the SCHED_SOFTIRQ
3162 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3163 cpu_isset(cpu, nohz.cpu_mask))
3166 if (time_after_eq(jiffies, rq->next_balance))
3167 raise_softirq(SCHED_SOFTIRQ);
3171 * on UP we do not need to balance between CPUs:
3173 static inline void idle_balance(int cpu, struct rq *rq)
3178 DEFINE_PER_CPU(struct kernel_stat, kstat);
3180 EXPORT_PER_CPU_SYMBOL(kstat);
3183 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3184 * that have not yet been banked in case the task is currently running.
3186 unsigned long long task_sched_runtime(struct task_struct *p)
3188 unsigned long flags;
3192 rq = task_rq_lock(p, &flags);
3193 ns = p->se.sum_exec_runtime;
3194 if (rq->curr == p) {
3195 delta_exec = rq_clock(rq) - p->se.exec_start;
3196 if ((s64)delta_exec > 0)
3199 task_rq_unlock(rq, &flags);
3205 * We place interactive tasks back into the active array, if possible.
3207 * To guarantee that this does not starve expired tasks we ignore the
3208 * interactivity of a task if the first expired task had to wait more
3209 * than a 'reasonable' amount of time. This deadline timeout is
3210 * load-dependent, as the frequency of array switched decreases with
3211 * increasing number of running tasks. We also ignore the interactivity
3212 * if a better static_prio task has expired:
3214 static inline int expired_starving(struct rq *rq)
3216 if (rq->curr->static_prio > rq->best_expired_prio)
3218 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3220 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3226 * Account user cpu time to a process.
3227 * @p: the process that the cpu time gets accounted to
3228 * @hardirq_offset: the offset to subtract from hardirq_count()
3229 * @cputime: the cpu time spent in user space since the last update
3231 void account_user_time(struct task_struct *p, cputime_t cputime)
3233 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3236 p->utime = cputime_add(p->utime, cputime);
3238 /* Add user time to cpustat. */
3239 tmp = cputime_to_cputime64(cputime);
3240 if (TASK_NICE(p) > 0)
3241 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3243 cpustat->user = cputime64_add(cpustat->user, tmp);
3247 * Account system cpu time to a process.
3248 * @p: the process that the cpu time gets accounted to
3249 * @hardirq_offset: the offset to subtract from hardirq_count()
3250 * @cputime: the cpu time spent in kernel space since the last update
3252 void account_system_time(struct task_struct *p, int hardirq_offset,
3255 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3256 struct rq *rq = this_rq();
3259 p->stime = cputime_add(p->stime, cputime);
3261 /* Add system time to cpustat. */
3262 tmp = cputime_to_cputime64(cputime);
3263 if (hardirq_count() - hardirq_offset)
3264 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3265 else if (softirq_count())
3266 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3267 else if (p != rq->idle)
3268 cpustat->system = cputime64_add(cpustat->system, tmp);
3269 else if (atomic_read(&rq->nr_iowait) > 0)
3270 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3272 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3273 /* Account for system time used */
3274 acct_update_integrals(p);
3278 * Account for involuntary wait time.
3279 * @p: the process from which the cpu time has been stolen
3280 * @steal: the cpu time spent in involuntary wait
3282 void account_steal_time(struct task_struct *p, cputime_t steal)
3284 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3285 cputime64_t tmp = cputime_to_cputime64(steal);
3286 struct rq *rq = this_rq();
3288 if (p == rq->idle) {
3289 p->stime = cputime_add(p->stime, steal);
3290 if (atomic_read(&rq->nr_iowait) > 0)
3291 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3293 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3295 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3298 static void task_running_tick(struct rq *rq, struct task_struct *p)
3300 if (p->array != rq->active) {
3301 /* Task has expired but was not scheduled yet */
3302 set_tsk_need_resched(p);
3305 spin_lock(&rq->lock);
3307 * The task was running during this tick - update the
3308 * time slice counter. Note: we do not update a thread's
3309 * priority until it either goes to sleep or uses up its
3310 * timeslice. This makes it possible for interactive tasks
3311 * to use up their timeslices at their highest priority levels.
3315 * RR tasks need a special form of timeslice management.
3316 * FIFO tasks have no timeslices.
3318 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3319 p->time_slice = task_timeslice(p);
3320 p->first_time_slice = 0;
3321 set_tsk_need_resched(p);
3323 /* put it at the end of the queue: */
3324 requeue_task(p, rq->active);
3328 if (!--p->time_slice) {
3329 dequeue_task(p, rq->active);
3330 set_tsk_need_resched(p);
3331 p->prio = effective_prio(p);
3332 p->time_slice = task_timeslice(p);
3333 p->first_time_slice = 0;
3335 if (!rq->expired_timestamp)
3336 rq->expired_timestamp = jiffies;
3337 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3338 enqueue_task(p, rq->expired);
3339 if (p->static_prio < rq->best_expired_prio)
3340 rq->best_expired_prio = p->static_prio;
3342 enqueue_task(p, rq->active);
3345 * Prevent a too long timeslice allowing a task to monopolize
3346 * the CPU. We do this by splitting up the timeslice into
3349 * Note: this does not mean the task's timeslices expire or
3350 * get lost in any way, they just might be preempted by
3351 * another task of equal priority. (one with higher
3352 * priority would have preempted this task already.) We
3353 * requeue this task to the end of the list on this priority
3354 * level, which is in essence a round-robin of tasks with
3357 * This only applies to tasks in the interactive
3358 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3360 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3361 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3362 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3363 (p->array == rq->active)) {
3365 requeue_task(p, rq->active);
3366 set_tsk_need_resched(p);
3370 spin_unlock(&rq->lock);
3374 * This function gets called by the timer code, with HZ frequency.
3375 * We call it with interrupts disabled.
3377 * It also gets called by the fork code, when changing the parent's
3380 void scheduler_tick(void)
3382 struct task_struct *p = current;
3383 int cpu = smp_processor_id();
3384 int idle_at_tick = idle_cpu(cpu);
3385 struct rq *rq = cpu_rq(cpu);
3388 task_running_tick(rq, p);
3391 rq->idle_at_tick = idle_at_tick;
3392 trigger_load_balance(cpu);
3396 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3398 void fastcall add_preempt_count(int val)
3403 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3405 preempt_count() += val;
3407 * Spinlock count overflowing soon?
3409 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3412 EXPORT_SYMBOL(add_preempt_count);
3414 void fastcall sub_preempt_count(int val)
3419 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3422 * Is the spinlock portion underflowing?
3424 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3425 !(preempt_count() & PREEMPT_MASK)))
3428 preempt_count() -= val;
3430 EXPORT_SYMBOL(sub_preempt_count);
3434 static inline int interactive_sleep(enum sleep_type sleep_type)
3436 return (sleep_type == SLEEP_INTERACTIVE ||
3437 sleep_type == SLEEP_INTERRUPTED);
3441 * schedule() is the main scheduler function.
3443 asmlinkage void __sched schedule(void)
3445 struct task_struct *prev, *next;
3446 struct prio_array *array;
3447 struct list_head *queue;
3448 unsigned long long now;
3449 unsigned long run_time;
3450 int cpu, idx, new_prio;
3455 * Test if we are atomic. Since do_exit() needs to call into
3456 * schedule() atomically, we ignore that path for now.
3457 * Otherwise, whine if we are scheduling when we should not be.
3459 if (unlikely(in_atomic() && !current->exit_state)) {
3460 printk(KERN_ERR "BUG: scheduling while atomic: "
3462 current->comm, preempt_count(), current->pid);
3463 debug_show_held_locks(current);
3464 if (irqs_disabled())
3465 print_irqtrace_events(current);
3468 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3473 release_kernel_lock(prev);
3474 need_resched_nonpreemptible:
3478 * The idle thread is not allowed to schedule!
3479 * Remove this check after it has been exercised a bit.
3481 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3482 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3486 schedstat_inc(rq, sched_cnt);
3487 now = sched_clock();
3488 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3489 run_time = now - prev->timestamp;
3490 if (unlikely((long long)(now - prev->timestamp) < 0))
3493 run_time = NS_MAX_SLEEP_AVG;
3496 * Tasks charged proportionately less run_time at high sleep_avg to
3497 * delay them losing their interactive status
3499 run_time /= (CURRENT_BONUS(prev) ? : 1);
3501 spin_lock_irq(&rq->lock);
3503 switch_count = &prev->nivcsw;
3504 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3505 switch_count = &prev->nvcsw;
3506 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3507 unlikely(signal_pending(prev))))
3508 prev->state = TASK_RUNNING;
3510 if (prev->state == TASK_UNINTERRUPTIBLE)
3511 rq->nr_uninterruptible++;
3512 deactivate_task(prev, rq);
3516 cpu = smp_processor_id();
3517 if (unlikely(!rq->nr_running)) {
3518 idle_balance(cpu, rq);
3519 if (!rq->nr_running) {
3521 rq->expired_timestamp = 0;
3527 if (unlikely(!array->nr_active)) {
3529 * Switch the active and expired arrays.
3531 schedstat_inc(rq, sched_switch);
3532 rq->active = rq->expired;
3533 rq->expired = array;
3535 rq->expired_timestamp = 0;
3536 rq->best_expired_prio = MAX_PRIO;
3539 idx = sched_find_first_bit(array->bitmap);
3540 queue = array->queue + idx;
3541 next = list_entry(queue->next, struct task_struct, run_list);
3543 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3544 unsigned long long delta = now - next->timestamp;
3545 if (unlikely((long long)(now - next->timestamp) < 0))
3548 if (next->sleep_type == SLEEP_INTERACTIVE)
3549 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3551 array = next->array;
3552 new_prio = recalc_task_prio(next, next->timestamp + delta);
3554 if (unlikely(next->prio != new_prio)) {
3555 dequeue_task(next, array);
3556 next->prio = new_prio;
3557 enqueue_task(next, array);
3560 next->sleep_type = SLEEP_NORMAL;
3562 if (next == rq->idle)
3563 schedstat_inc(rq, sched_goidle);
3565 prefetch_stack(next);
3566 clear_tsk_need_resched(prev);
3567 rcu_qsctr_inc(task_cpu(prev));
3569 prev->sleep_avg -= run_time;
3570 if ((long)prev->sleep_avg <= 0)
3571 prev->sleep_avg = 0;
3572 prev->timestamp = prev->last_ran = now;
3574 sched_info_switch(prev, next);
3575 if (likely(prev != next)) {
3576 next->timestamp = next->last_ran = now;
3581 prepare_task_switch(rq, next);
3582 prev = context_switch(rq, prev, next);
3585 * this_rq must be evaluated again because prev may have moved
3586 * CPUs since it called schedule(), thus the 'rq' on its stack
3587 * frame will be invalid.
3589 finish_task_switch(this_rq(), prev);
3591 spin_unlock_irq(&rq->lock);
3594 if (unlikely(reacquire_kernel_lock(prev) < 0))
3595 goto need_resched_nonpreemptible;
3596 preempt_enable_no_resched();
3597 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3600 EXPORT_SYMBOL(schedule);
3602 #ifdef CONFIG_PREEMPT
3604 * this is the entry point to schedule() from in-kernel preemption
3605 * off of preempt_enable. Kernel preemptions off return from interrupt
3606 * occur there and call schedule directly.
3608 asmlinkage void __sched preempt_schedule(void)
3610 struct thread_info *ti = current_thread_info();
3611 #ifdef CONFIG_PREEMPT_BKL
3612 struct task_struct *task = current;
3613 int saved_lock_depth;
3616 * If there is a non-zero preempt_count or interrupts are disabled,
3617 * we do not want to preempt the current task. Just return..
3619 if (likely(ti->preempt_count || irqs_disabled()))
3623 add_preempt_count(PREEMPT_ACTIVE);
3625 * We keep the big kernel semaphore locked, but we
3626 * clear ->lock_depth so that schedule() doesnt
3627 * auto-release the semaphore:
3629 #ifdef CONFIG_PREEMPT_BKL
3630 saved_lock_depth = task->lock_depth;
3631 task->lock_depth = -1;
3634 #ifdef CONFIG_PREEMPT_BKL
3635 task->lock_depth = saved_lock_depth;
3637 sub_preempt_count(PREEMPT_ACTIVE);
3639 /* we could miss a preemption opportunity between schedule and now */
3641 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3644 EXPORT_SYMBOL(preempt_schedule);
3647 * this is the entry point to schedule() from kernel preemption
3648 * off of irq context.
3649 * Note, that this is called and return with irqs disabled. This will
3650 * protect us against recursive calling from irq.
3652 asmlinkage void __sched preempt_schedule_irq(void)
3654 struct thread_info *ti = current_thread_info();
3655 #ifdef CONFIG_PREEMPT_BKL
3656 struct task_struct *task = current;
3657 int saved_lock_depth;
3659 /* Catch callers which need to be fixed */
3660 BUG_ON(ti->preempt_count || !irqs_disabled());
3663 add_preempt_count(PREEMPT_ACTIVE);
3665 * We keep the big kernel semaphore locked, but we
3666 * clear ->lock_depth so that schedule() doesnt
3667 * auto-release the semaphore:
3669 #ifdef CONFIG_PREEMPT_BKL
3670 saved_lock_depth = task->lock_depth;
3671 task->lock_depth = -1;
3675 local_irq_disable();
3676 #ifdef CONFIG_PREEMPT_BKL
3677 task->lock_depth = saved_lock_depth;
3679 sub_preempt_count(PREEMPT_ACTIVE);
3681 /* we could miss a preemption opportunity between schedule and now */
3683 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3687 #endif /* CONFIG_PREEMPT */
3689 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3692 return try_to_wake_up(curr->private, mode, sync);
3694 EXPORT_SYMBOL(default_wake_function);
3697 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3698 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3699 * number) then we wake all the non-exclusive tasks and one exclusive task.
3701 * There are circumstances in which we can try to wake a task which has already
3702 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3703 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3705 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3706 int nr_exclusive, int sync, void *key)
3708 struct list_head *tmp, *next;
3710 list_for_each_safe(tmp, next, &q->task_list) {
3711 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3712 unsigned flags = curr->flags;
3714 if (curr->func(curr, mode, sync, key) &&
3715 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3721 * __wake_up - wake up threads blocked on a waitqueue.
3723 * @mode: which threads
3724 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3725 * @key: is directly passed to the wakeup function
3727 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3728 int nr_exclusive, void *key)
3730 unsigned long flags;
3732 spin_lock_irqsave(&q->lock, flags);
3733 __wake_up_common(q, mode, nr_exclusive, 0, key);
3734 spin_unlock_irqrestore(&q->lock, flags);
3736 EXPORT_SYMBOL(__wake_up);
3739 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3741 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3743 __wake_up_common(q, mode, 1, 0, NULL);
3747 * __wake_up_sync - wake up threads blocked on a waitqueue.
3749 * @mode: which threads
3750 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3752 * The sync wakeup differs that the waker knows that it will schedule
3753 * away soon, so while the target thread will be woken up, it will not
3754 * be migrated to another CPU - ie. the two threads are 'synchronized'
3755 * with each other. This can prevent needless bouncing between CPUs.
3757 * On UP it can prevent extra preemption.
3760 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3762 unsigned long flags;
3768 if (unlikely(!nr_exclusive))
3771 spin_lock_irqsave(&q->lock, flags);
3772 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3773 spin_unlock_irqrestore(&q->lock, flags);
3775 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3777 void fastcall complete(struct completion *x)
3779 unsigned long flags;
3781 spin_lock_irqsave(&x->wait.lock, flags);
3783 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3785 spin_unlock_irqrestore(&x->wait.lock, flags);
3787 EXPORT_SYMBOL(complete);
3789 void fastcall complete_all(struct completion *x)
3791 unsigned long flags;
3793 spin_lock_irqsave(&x->wait.lock, flags);
3794 x->done += UINT_MAX/2;
3795 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3797 spin_unlock_irqrestore(&x->wait.lock, flags);
3799 EXPORT_SYMBOL(complete_all);
3801 void fastcall __sched wait_for_completion(struct completion *x)
3805 spin_lock_irq(&x->wait.lock);
3807 DECLARE_WAITQUEUE(wait, current);
3809 wait.flags |= WQ_FLAG_EXCLUSIVE;
3810 __add_wait_queue_tail(&x->wait, &wait);
3812 __set_current_state(TASK_UNINTERRUPTIBLE);
3813 spin_unlock_irq(&x->wait.lock);
3815 spin_lock_irq(&x->wait.lock);
3817 __remove_wait_queue(&x->wait, &wait);
3820 spin_unlock_irq(&x->wait.lock);
3822 EXPORT_SYMBOL(wait_for_completion);
3824 unsigned long fastcall __sched
3825 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3829 spin_lock_irq(&x->wait.lock);
3831 DECLARE_WAITQUEUE(wait, current);
3833 wait.flags |= WQ_FLAG_EXCLUSIVE;
3834 __add_wait_queue_tail(&x->wait, &wait);
3836 __set_current_state(TASK_UNINTERRUPTIBLE);
3837 spin_unlock_irq(&x->wait.lock);
3838 timeout = schedule_timeout(timeout);
3839 spin_lock_irq(&x->wait.lock);
3841 __remove_wait_queue(&x->wait, &wait);
3845 __remove_wait_queue(&x->wait, &wait);
3849 spin_unlock_irq(&x->wait.lock);
3852 EXPORT_SYMBOL(wait_for_completion_timeout);
3854 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3860 spin_lock_irq(&x->wait.lock);
3862 DECLARE_WAITQUEUE(wait, current);
3864 wait.flags |= WQ_FLAG_EXCLUSIVE;
3865 __add_wait_queue_tail(&x->wait, &wait);
3867 if (signal_pending(current)) {
3869 __remove_wait_queue(&x->wait, &wait);
3872 __set_current_state(TASK_INTERRUPTIBLE);
3873 spin_unlock_irq(&x->wait.lock);
3875 spin_lock_irq(&x->wait.lock);
3877 __remove_wait_queue(&x->wait, &wait);
3881 spin_unlock_irq(&x->wait.lock);
3885 EXPORT_SYMBOL(wait_for_completion_interruptible);
3887 unsigned long fastcall __sched
3888 wait_for_completion_interruptible_timeout(struct completion *x,
3889 unsigned long timeout)
3893 spin_lock_irq(&x->wait.lock);
3895 DECLARE_WAITQUEUE(wait, current);
3897 wait.flags |= WQ_FLAG_EXCLUSIVE;
3898 __add_wait_queue_tail(&x->wait, &wait);
3900 if (signal_pending(current)) {
3901 timeout = -ERESTARTSYS;
3902 __remove_wait_queue(&x->wait, &wait);
3905 __set_current_state(TASK_INTERRUPTIBLE);
3906 spin_unlock_irq(&x->wait.lock);
3907 timeout = schedule_timeout(timeout);
3908 spin_lock_irq(&x->wait.lock);
3910 __remove_wait_queue(&x->wait, &wait);
3914 __remove_wait_queue(&x->wait, &wait);
3918 spin_unlock_irq(&x->wait.lock);
3921 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3924 #define SLEEP_ON_VAR \
3925 unsigned long flags; \
3926 wait_queue_t wait; \
3927 init_waitqueue_entry(&wait, current);
3929 #define SLEEP_ON_HEAD \
3930 spin_lock_irqsave(&q->lock,flags); \
3931 __add_wait_queue(q, &wait); \
3932 spin_unlock(&q->lock);
3934 #define SLEEP_ON_TAIL \
3935 spin_lock_irq(&q->lock); \
3936 __remove_wait_queue(q, &wait); \
3937 spin_unlock_irqrestore(&q->lock, flags);
3939 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3943 current->state = TASK_INTERRUPTIBLE;
3949 EXPORT_SYMBOL(interruptible_sleep_on);
3951 long fastcall __sched
3952 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3956 current->state = TASK_INTERRUPTIBLE;
3959 timeout = schedule_timeout(timeout);
3964 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3966 void fastcall __sched sleep_on(wait_queue_head_t *q)
3970 current->state = TASK_UNINTERRUPTIBLE;
3976 EXPORT_SYMBOL(sleep_on);
3978 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3982 current->state = TASK_UNINTERRUPTIBLE;
3985 timeout = schedule_timeout(timeout);
3991 EXPORT_SYMBOL(sleep_on_timeout);
3993 #ifdef CONFIG_RT_MUTEXES
3996 * rt_mutex_setprio - set the current priority of a task
3998 * @prio: prio value (kernel-internal form)
4000 * This function changes the 'effective' priority of a task. It does
4001 * not touch ->normal_prio like __setscheduler().
4003 * Used by the rt_mutex code to implement priority inheritance logic.
4005 void rt_mutex_setprio(struct task_struct *p, int prio)
4007 struct prio_array *array;
4008 unsigned long flags;
4012 BUG_ON(prio < 0 || prio > MAX_PRIO);
4014 rq = task_rq_lock(p, &flags);
4019 dequeue_task(p, array);
4024 * If changing to an RT priority then queue it
4025 * in the active array!
4029 enqueue_task(p, array);
4031 * Reschedule if we are currently running on this runqueue and
4032 * our priority decreased, or if we are not currently running on
4033 * this runqueue and our priority is higher than the current's
4035 if (task_running(rq, p)) {
4036 if (p->prio > oldprio)
4037 resched_task(rq->curr);
4038 } else if (TASK_PREEMPTS_CURR(p, rq))
4039 resched_task(rq->curr);
4041 task_rq_unlock(rq, &flags);
4046 void set_user_nice(struct task_struct *p, long nice)
4048 struct prio_array *array;
4049 int old_prio, delta;
4050 unsigned long flags;
4053 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 * We have to be careful, if called from sys_setpriority(),
4057 * the task might be in the middle of scheduling on another CPU.
4059 rq = task_rq_lock(p, &flags);
4061 * The RT priorities are set via sched_setscheduler(), but we still
4062 * allow the 'normal' nice value to be set - but as expected
4063 * it wont have any effect on scheduling until the task is
4064 * not SCHED_NORMAL/SCHED_BATCH:
4066 if (task_has_rt_policy(p)) {
4067 p->static_prio = NICE_TO_PRIO(nice);
4072 dequeue_task(p, array);
4073 dec_raw_weighted_load(rq, p);
4076 p->static_prio = NICE_TO_PRIO(nice);
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4083 enqueue_task(p, array);
4084 inc_raw_weighted_load(rq, p);
4086 * If the task increased its priority or is running and
4087 * lowered its priority, then reschedule its CPU:
4089 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4090 resched_task(rq->curr);
4093 task_rq_unlock(rq, &flags);
4095 EXPORT_SYMBOL(set_user_nice);
4098 * can_nice - check if a task can reduce its nice value
4102 int can_nice(const struct task_struct *p, const int nice)
4104 /* convert nice value [19,-20] to rlimit style value [1,40] */
4105 int nice_rlim = 20 - nice;
4107 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4108 capable(CAP_SYS_NICE));
4111 #ifdef __ARCH_WANT_SYS_NICE
4114 * sys_nice - change the priority of the current process.
4115 * @increment: priority increment
4117 * sys_setpriority is a more generic, but much slower function that
4118 * does similar things.
4120 asmlinkage long sys_nice(int increment)
4125 * Setpriority might change our priority at the same moment.
4126 * We don't have to worry. Conceptually one call occurs first
4127 * and we have a single winner.
4129 if (increment < -40)
4134 nice = PRIO_TO_NICE(current->static_prio) + increment;
4140 if (increment < 0 && !can_nice(current, nice))
4143 retval = security_task_setnice(current, nice);
4147 set_user_nice(current, nice);
4154 * task_prio - return the priority value of a given task.
4155 * @p: the task in question.
4157 * This is the priority value as seen by users in /proc.
4158 * RT tasks are offset by -200. Normal tasks are centered
4159 * around 0, value goes from -16 to +15.
4161 int task_prio(const struct task_struct *p)
4163 return p->prio - MAX_RT_PRIO;
4167 * task_nice - return the nice value of a given task.
4168 * @p: the task in question.
4170 int task_nice(const struct task_struct *p)
4172 return TASK_NICE(p);
4174 EXPORT_SYMBOL_GPL(task_nice);
4177 * idle_cpu - is a given cpu idle currently?
4178 * @cpu: the processor in question.
4180 int idle_cpu(int cpu)
4182 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4186 * idle_task - return the idle task for a given cpu.
4187 * @cpu: the processor in question.
4189 struct task_struct *idle_task(int cpu)
4191 return cpu_rq(cpu)->idle;
4195 * find_process_by_pid - find a process with a matching PID value.
4196 * @pid: the pid in question.
4198 static inline struct task_struct *find_process_by_pid(pid_t pid)
4200 return pid ? find_task_by_pid(pid) : current;
4203 /* Actually do priority change: must hold rq lock. */
4204 static void __setscheduler(struct task_struct *p, int policy, int prio)
4209 p->rt_priority = prio;
4210 p->normal_prio = normal_prio(p);
4211 /* we are holding p->pi_lock already */
4212 p->prio = rt_mutex_getprio(p);
4214 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4216 if (policy == SCHED_BATCH)
4222 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4223 * @p: the task in question.
4224 * @policy: new policy.
4225 * @param: structure containing the new RT priority.
4227 * NOTE that the task may be already dead.
4229 int sched_setscheduler(struct task_struct *p, int policy,
4230 struct sched_param *param)
4232 int retval, oldprio, oldpolicy = -1;
4233 struct prio_array *array;
4234 unsigned long flags;
4237 /* may grab non-irq protected spin_locks */
4238 BUG_ON(in_interrupt());
4240 /* double check policy once rq lock held */
4242 policy = oldpolicy = p->policy;
4243 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4244 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4247 * Valid priorities for SCHED_FIFO and SCHED_RR are
4248 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4251 if (param->sched_priority < 0 ||
4252 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4253 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4255 if (rt_policy(policy) != (param->sched_priority != 0))
4259 * Allow unprivileged RT tasks to decrease priority:
4261 if (!capable(CAP_SYS_NICE)) {
4262 if (rt_policy(policy)) {
4263 unsigned long rlim_rtprio;
4264 unsigned long flags;
4266 if (!lock_task_sighand(p, &flags))
4268 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4269 unlock_task_sighand(p, &flags);
4271 /* can't set/change the rt policy */
4272 if (policy != p->policy && !rlim_rtprio)
4275 /* can't increase priority */
4276 if (param->sched_priority > p->rt_priority &&
4277 param->sched_priority > rlim_rtprio)
4281 /* can't change other user's priorities */
4282 if ((current->euid != p->euid) &&
4283 (current->euid != p->uid))
4287 retval = security_task_setscheduler(p, policy, param);
4291 * make sure no PI-waiters arrive (or leave) while we are
4292 * changing the priority of the task:
4294 spin_lock_irqsave(&p->pi_lock, flags);
4296 * To be able to change p->policy safely, the apropriate
4297 * runqueue lock must be held.
4299 rq = __task_rq_lock(p);
4300 /* recheck policy now with rq lock held */
4301 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4302 policy = oldpolicy = -1;
4303 __task_rq_unlock(rq);
4304 spin_unlock_irqrestore(&p->pi_lock, flags);
4309 deactivate_task(p, rq);
4311 __setscheduler(p, policy, param->sched_priority);
4313 __activate_task(p, rq);
4315 * Reschedule if we are currently running on this runqueue and
4316 * our priority decreased, or if we are not currently running on
4317 * this runqueue and our priority is higher than the current's
4319 if (task_running(rq, p)) {
4320 if (p->prio > oldprio)
4321 resched_task(rq->curr);
4322 } else if (TASK_PREEMPTS_CURR(p, rq))
4323 resched_task(rq->curr);
4325 __task_rq_unlock(rq);
4326 spin_unlock_irqrestore(&p->pi_lock, flags);
4328 rt_mutex_adjust_pi(p);
4332 EXPORT_SYMBOL_GPL(sched_setscheduler);
4335 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4337 struct sched_param lparam;
4338 struct task_struct *p;
4341 if (!param || pid < 0)
4343 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4348 p = find_process_by_pid(pid);
4350 retval = sched_setscheduler(p, policy, &lparam);
4357 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4358 * @pid: the pid in question.
4359 * @policy: new policy.
4360 * @param: structure containing the new RT priority.
4362 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4363 struct sched_param __user *param)
4365 /* negative values for policy are not valid */
4369 return do_sched_setscheduler(pid, policy, param);
4373 * sys_sched_setparam - set/change the RT priority of a thread
4374 * @pid: the pid in question.
4375 * @param: structure containing the new RT priority.
4377 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4379 return do_sched_setscheduler(pid, -1, param);
4383 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4384 * @pid: the pid in question.
4386 asmlinkage long sys_sched_getscheduler(pid_t pid)
4388 struct task_struct *p;
4389 int retval = -EINVAL;
4395 read_lock(&tasklist_lock);
4396 p = find_process_by_pid(pid);
4398 retval = security_task_getscheduler(p);
4402 read_unlock(&tasklist_lock);
4409 * sys_sched_getscheduler - get the RT priority of a thread
4410 * @pid: the pid in question.
4411 * @param: structure containing the RT priority.
4413 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4415 struct sched_param lp;
4416 struct task_struct *p;
4417 int retval = -EINVAL;
4419 if (!param || pid < 0)
4422 read_lock(&tasklist_lock);
4423 p = find_process_by_pid(pid);
4428 retval = security_task_getscheduler(p);
4432 lp.sched_priority = p->rt_priority;
4433 read_unlock(&tasklist_lock);
4436 * This one might sleep, we cannot do it with a spinlock held ...
4438 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4444 read_unlock(&tasklist_lock);
4448 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4450 cpumask_t cpus_allowed;
4451 struct task_struct *p;
4454 mutex_lock(&sched_hotcpu_mutex);
4455 read_lock(&tasklist_lock);
4457 p = find_process_by_pid(pid);
4459 read_unlock(&tasklist_lock);
4460 mutex_unlock(&sched_hotcpu_mutex);
4465 * It is not safe to call set_cpus_allowed with the
4466 * tasklist_lock held. We will bump the task_struct's
4467 * usage count and then drop tasklist_lock.
4470 read_unlock(&tasklist_lock);
4473 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4474 !capable(CAP_SYS_NICE))
4477 retval = security_task_setscheduler(p, 0, NULL);
4481 cpus_allowed = cpuset_cpus_allowed(p);
4482 cpus_and(new_mask, new_mask, cpus_allowed);
4483 retval = set_cpus_allowed(p, new_mask);
4487 mutex_unlock(&sched_hotcpu_mutex);
4491 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4492 cpumask_t *new_mask)
4494 if (len < sizeof(cpumask_t)) {
4495 memset(new_mask, 0, sizeof(cpumask_t));
4496 } else if (len > sizeof(cpumask_t)) {
4497 len = sizeof(cpumask_t);
4499 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4503 * sys_sched_setaffinity - set the cpu affinity of a process
4504 * @pid: pid of the process
4505 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4506 * @user_mask_ptr: user-space pointer to the new cpu mask
4508 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4509 unsigned long __user *user_mask_ptr)
4514 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4518 return sched_setaffinity(pid, new_mask);
4522 * Represents all cpu's present in the system
4523 * In systems capable of hotplug, this map could dynamically grow
4524 * as new cpu's are detected in the system via any platform specific
4525 * method, such as ACPI for e.g.
4528 cpumask_t cpu_present_map __read_mostly;
4529 EXPORT_SYMBOL(cpu_present_map);
4532 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4533 EXPORT_SYMBOL(cpu_online_map);
4535 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4536 EXPORT_SYMBOL(cpu_possible_map);
4539 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4541 struct task_struct *p;
4544 mutex_lock(&sched_hotcpu_mutex);
4545 read_lock(&tasklist_lock);
4548 p = find_process_by_pid(pid);
4552 retval = security_task_getscheduler(p);
4556 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4559 read_unlock(&tasklist_lock);
4560 mutex_unlock(&sched_hotcpu_mutex);
4568 * sys_sched_getaffinity - get the cpu affinity of a process
4569 * @pid: pid of the process
4570 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4571 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4573 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4574 unsigned long __user *user_mask_ptr)
4579 if (len < sizeof(cpumask_t))
4582 ret = sched_getaffinity(pid, &mask);
4586 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4589 return sizeof(cpumask_t);
4593 * sys_sched_yield - yield the current processor to other threads.
4595 * This function yields the current CPU by moving the calling thread
4596 * to the expired array. If there are no other threads running on this
4597 * CPU then this function will return.
4599 asmlinkage long sys_sched_yield(void)
4601 struct rq *rq = this_rq_lock();
4602 struct prio_array *array = current->array, *target = rq->expired;
4604 schedstat_inc(rq, yld_cnt);
4606 * We implement yielding by moving the task into the expired
4609 * (special rule: RT tasks will just roundrobin in the active
4612 if (rt_task(current))
4613 target = rq->active;
4615 if (array->nr_active == 1) {
4616 schedstat_inc(rq, yld_act_empty);
4617 if (!rq->expired->nr_active)
4618 schedstat_inc(rq, yld_both_empty);
4619 } else if (!rq->expired->nr_active)
4620 schedstat_inc(rq, yld_exp_empty);
4622 if (array != target) {
4623 dequeue_task(current, array);
4624 enqueue_task(current, target);
4627 * requeue_task is cheaper so perform that if possible.
4629 requeue_task(current, array);
4632 * Since we are going to call schedule() anyway, there's
4633 * no need to preempt or enable interrupts:
4635 __release(rq->lock);
4636 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4637 _raw_spin_unlock(&rq->lock);
4638 preempt_enable_no_resched();
4645 static void __cond_resched(void)
4647 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4648 __might_sleep(__FILE__, __LINE__);
4651 * The BKS might be reacquired before we have dropped
4652 * PREEMPT_ACTIVE, which could trigger a second
4653 * cond_resched() call.
4656 add_preempt_count(PREEMPT_ACTIVE);
4658 sub_preempt_count(PREEMPT_ACTIVE);
4659 } while (need_resched());
4662 int __sched cond_resched(void)
4664 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4665 system_state == SYSTEM_RUNNING) {
4671 EXPORT_SYMBOL(cond_resched);
4674 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4675 * call schedule, and on return reacquire the lock.
4677 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4678 * operations here to prevent schedule() from being called twice (once via
4679 * spin_unlock(), once by hand).
4681 int cond_resched_lock(spinlock_t *lock)
4685 if (need_lockbreak(lock)) {
4691 if (need_resched() && system_state == SYSTEM_RUNNING) {
4692 spin_release(&lock->dep_map, 1, _THIS_IP_);
4693 _raw_spin_unlock(lock);
4694 preempt_enable_no_resched();
4701 EXPORT_SYMBOL(cond_resched_lock);
4703 int __sched cond_resched_softirq(void)
4705 BUG_ON(!in_softirq());
4707 if (need_resched() && system_state == SYSTEM_RUNNING) {
4715 EXPORT_SYMBOL(cond_resched_softirq);
4718 * yield - yield the current processor to other threads.
4720 * This is a shortcut for kernel-space yielding - it marks the
4721 * thread runnable and calls sys_sched_yield().
4723 void __sched yield(void)
4725 set_current_state(TASK_RUNNING);
4728 EXPORT_SYMBOL(yield);
4731 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4732 * that process accounting knows that this is a task in IO wait state.
4734 * But don't do that if it is a deliberate, throttling IO wait (this task
4735 * has set its backing_dev_info: the queue against which it should throttle)
4737 void __sched io_schedule(void)
4739 struct rq *rq = &__raw_get_cpu_var(runqueues);
4741 delayacct_blkio_start();
4742 atomic_inc(&rq->nr_iowait);
4744 atomic_dec(&rq->nr_iowait);
4745 delayacct_blkio_end();
4747 EXPORT_SYMBOL(io_schedule);
4749 long __sched io_schedule_timeout(long timeout)
4751 struct rq *rq = &__raw_get_cpu_var(runqueues);
4754 delayacct_blkio_start();
4755 atomic_inc(&rq->nr_iowait);
4756 ret = schedule_timeout(timeout);
4757 atomic_dec(&rq->nr_iowait);
4758 delayacct_blkio_end();
4763 * sys_sched_get_priority_max - return maximum RT priority.
4764 * @policy: scheduling class.
4766 * this syscall returns the maximum rt_priority that can be used
4767 * by a given scheduling class.
4769 asmlinkage long sys_sched_get_priority_max(int policy)
4776 ret = MAX_USER_RT_PRIO-1;
4787 * sys_sched_get_priority_min - return minimum RT priority.
4788 * @policy: scheduling class.
4790 * this syscall returns the minimum rt_priority that can be used
4791 * by a given scheduling class.
4793 asmlinkage long sys_sched_get_priority_min(int policy)
4810 * sys_sched_rr_get_interval - return the default timeslice of a process.
4811 * @pid: pid of the process.
4812 * @interval: userspace pointer to the timeslice value.
4814 * this syscall writes the default timeslice value of a given process
4815 * into the user-space timespec buffer. A value of '0' means infinity.
4818 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4820 struct task_struct *p;
4821 int retval = -EINVAL;
4828 read_lock(&tasklist_lock);
4829 p = find_process_by_pid(pid);
4833 retval = security_task_getscheduler(p);
4837 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4838 0 : task_timeslice(p), &t);
4839 read_unlock(&tasklist_lock);
4840 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4844 read_unlock(&tasklist_lock);
4848 static const char stat_nam[] = "RSDTtZX";
4850 static void show_task(struct task_struct *p)
4852 unsigned long free = 0;
4855 state = p->state ? __ffs(p->state) + 1 : 0;
4856 printk("%-13.13s %c", p->comm,
4857 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4858 #if (BITS_PER_LONG == 32)
4859 if (state == TASK_RUNNING)
4860 printk(" running ");
4862 printk(" %08lX ", thread_saved_pc(p));
4864 if (state == TASK_RUNNING)
4865 printk(" running task ");
4867 printk(" %016lx ", thread_saved_pc(p));
4869 #ifdef CONFIG_DEBUG_STACK_USAGE
4871 unsigned long *n = end_of_stack(p);
4874 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4877 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4879 printk(" (L-TLB)\n");
4881 printk(" (NOTLB)\n");
4883 if (state != TASK_RUNNING)
4884 show_stack(p, NULL);
4887 void show_state_filter(unsigned long state_filter)
4889 struct task_struct *g, *p;
4891 #if (BITS_PER_LONG == 32)
4894 printk(" task PC stack pid father child younger older\n");
4898 printk(" task PC stack pid father child younger older\n");
4900 read_lock(&tasklist_lock);
4901 do_each_thread(g, p) {
4903 * reset the NMI-timeout, listing all files on a slow
4904 * console might take alot of time:
4906 touch_nmi_watchdog();
4907 if (!state_filter || (p->state & state_filter))
4909 } while_each_thread(g, p);
4911 touch_all_softlockup_watchdogs();
4913 read_unlock(&tasklist_lock);
4915 * Only show locks if all tasks are dumped:
4917 if (state_filter == -1)
4918 debug_show_all_locks();
4921 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4927 * init_idle - set up an idle thread for a given CPU
4928 * @idle: task in question
4929 * @cpu: cpu the idle task belongs to
4931 * NOTE: this function does not set the idle thread's NEED_RESCHED
4932 * flag, to make booting more robust.
4934 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4936 struct rq *rq = cpu_rq(cpu);
4937 unsigned long flags;
4939 idle->timestamp = sched_clock();
4940 idle->sleep_avg = 0;
4942 idle->prio = idle->normal_prio = MAX_PRIO;
4943 idle->state = TASK_RUNNING;
4944 idle->cpus_allowed = cpumask_of_cpu(cpu);
4945 set_task_cpu(idle, cpu);
4947 spin_lock_irqsave(&rq->lock, flags);
4948 rq->curr = rq->idle = idle;
4949 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4952 spin_unlock_irqrestore(&rq->lock, flags);
4954 /* Set the preempt count _outside_ the spinlocks! */
4955 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4956 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4958 task_thread_info(idle)->preempt_count = 0;
4963 * In a system that switches off the HZ timer nohz_cpu_mask
4964 * indicates which cpus entered this state. This is used
4965 * in the rcu update to wait only for active cpus. For system
4966 * which do not switch off the HZ timer nohz_cpu_mask should
4967 * always be CPU_MASK_NONE.
4969 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4973 * This is how migration works:
4975 * 1) we queue a struct migration_req structure in the source CPU's
4976 * runqueue and wake up that CPU's migration thread.
4977 * 2) we down() the locked semaphore => thread blocks.
4978 * 3) migration thread wakes up (implicitly it forces the migrated
4979 * thread off the CPU)
4980 * 4) it gets the migration request and checks whether the migrated
4981 * task is still in the wrong runqueue.
4982 * 5) if it's in the wrong runqueue then the migration thread removes
4983 * it and puts it into the right queue.
4984 * 6) migration thread up()s the semaphore.
4985 * 7) we wake up and the migration is done.
4989 * Change a given task's CPU affinity. Migrate the thread to a
4990 * proper CPU and schedule it away if the CPU it's executing on
4991 * is removed from the allowed bitmask.
4993 * NOTE: the caller must have a valid reference to the task, the
4994 * task must not exit() & deallocate itself prematurely. The
4995 * call is not atomic; no spinlocks may be held.
4997 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4999 struct migration_req req;
5000 unsigned long flags;
5004 rq = task_rq_lock(p, &flags);
5005 if (!cpus_intersects(new_mask, cpu_online_map)) {
5010 p->cpus_allowed = new_mask;
5011 /* Can the task run on the task's current CPU? If so, we're done */
5012 if (cpu_isset(task_cpu(p), new_mask))
5015 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5016 /* Need help from migration thread: drop lock and wait. */
5017 task_rq_unlock(rq, &flags);
5018 wake_up_process(rq->migration_thread);
5019 wait_for_completion(&req.done);
5020 tlb_migrate_finish(p->mm);
5024 task_rq_unlock(rq, &flags);
5028 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5031 * Move (not current) task off this cpu, onto dest cpu. We're doing
5032 * this because either it can't run here any more (set_cpus_allowed()
5033 * away from this CPU, or CPU going down), or because we're
5034 * attempting to rebalance this task on exec (sched_exec).
5036 * So we race with normal scheduler movements, but that's OK, as long
5037 * as the task is no longer on this CPU.
5039 * Returns non-zero if task was successfully migrated.
5041 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5043 struct rq *rq_dest, *rq_src;
5046 if (unlikely(cpu_is_offline(dest_cpu)))
5049 rq_src = cpu_rq(src_cpu);
5050 rq_dest = cpu_rq(dest_cpu);
5052 double_rq_lock(rq_src, rq_dest);
5053 /* Already moved. */
5054 if (task_cpu(p) != src_cpu)
5056 /* Affinity changed (again). */
5057 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5060 set_task_cpu(p, dest_cpu);
5063 * Sync timestamp with rq_dest's before activating.
5064 * The same thing could be achieved by doing this step
5065 * afterwards, and pretending it was a local activate.
5066 * This way is cleaner and logically correct.
5068 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5069 + rq_dest->most_recent_timestamp;
5070 deactivate_task(p, rq_src);
5071 __activate_task(p, rq_dest);
5072 if (TASK_PREEMPTS_CURR(p, rq_dest))
5073 resched_task(rq_dest->curr);
5077 double_rq_unlock(rq_src, rq_dest);
5082 * migration_thread - this is a highprio system thread that performs
5083 * thread migration by bumping thread off CPU then 'pushing' onto
5086 static int migration_thread(void *data)
5088 int cpu = (long)data;
5092 BUG_ON(rq->migration_thread != current);
5094 set_current_state(TASK_INTERRUPTIBLE);
5095 while (!kthread_should_stop()) {
5096 struct migration_req *req;
5097 struct list_head *head;
5101 spin_lock_irq(&rq->lock);
5103 if (cpu_is_offline(cpu)) {
5104 spin_unlock_irq(&rq->lock);
5108 if (rq->active_balance) {
5109 active_load_balance(rq, cpu);
5110 rq->active_balance = 0;
5113 head = &rq->migration_queue;
5115 if (list_empty(head)) {
5116 spin_unlock_irq(&rq->lock);
5118 set_current_state(TASK_INTERRUPTIBLE);
5121 req = list_entry(head->next, struct migration_req, list);
5122 list_del_init(head->next);
5124 spin_unlock(&rq->lock);
5125 __migrate_task(req->task, cpu, req->dest_cpu);
5128 complete(&req->done);
5130 __set_current_state(TASK_RUNNING);
5134 /* Wait for kthread_stop */
5135 set_current_state(TASK_INTERRUPTIBLE);
5136 while (!kthread_should_stop()) {
5138 set_current_state(TASK_INTERRUPTIBLE);
5140 __set_current_state(TASK_RUNNING);
5144 #ifdef CONFIG_HOTPLUG_CPU
5146 * Figure out where task on dead CPU should go, use force if neccessary.
5147 * NOTE: interrupts should be disabled by the caller
5149 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5151 unsigned long flags;
5158 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5159 cpus_and(mask, mask, p->cpus_allowed);
5160 dest_cpu = any_online_cpu(mask);
5162 /* On any allowed CPU? */
5163 if (dest_cpu == NR_CPUS)
5164 dest_cpu = any_online_cpu(p->cpus_allowed);
5166 /* No more Mr. Nice Guy. */
5167 if (dest_cpu == NR_CPUS) {
5168 rq = task_rq_lock(p, &flags);
5169 cpus_setall(p->cpus_allowed);
5170 dest_cpu = any_online_cpu(p->cpus_allowed);
5171 task_rq_unlock(rq, &flags);
5174 * Don't tell them about moving exiting tasks or
5175 * kernel threads (both mm NULL), since they never
5178 if (p->mm && printk_ratelimit())
5179 printk(KERN_INFO "process %d (%s) no "
5180 "longer affine to cpu%d\n",
5181 p->pid, p->comm, dead_cpu);
5183 if (!__migrate_task(p, dead_cpu, dest_cpu))
5188 * While a dead CPU has no uninterruptible tasks queued at this point,
5189 * it might still have a nonzero ->nr_uninterruptible counter, because
5190 * for performance reasons the counter is not stricly tracking tasks to
5191 * their home CPUs. So we just add the counter to another CPU's counter,
5192 * to keep the global sum constant after CPU-down:
5194 static void migrate_nr_uninterruptible(struct rq *rq_src)
5196 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5197 unsigned long flags;
5199 local_irq_save(flags);
5200 double_rq_lock(rq_src, rq_dest);
5201 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5202 rq_src->nr_uninterruptible = 0;
5203 double_rq_unlock(rq_src, rq_dest);
5204 local_irq_restore(flags);
5207 /* Run through task list and migrate tasks from the dead cpu. */
5208 static void migrate_live_tasks(int src_cpu)
5210 struct task_struct *p, *t;
5212 write_lock_irq(&tasklist_lock);
5214 do_each_thread(t, p) {
5218 if (task_cpu(p) == src_cpu)
5219 move_task_off_dead_cpu(src_cpu, p);
5220 } while_each_thread(t, p);
5222 write_unlock_irq(&tasklist_lock);
5225 /* Schedules idle task to be the next runnable task on current CPU.
5226 * It does so by boosting its priority to highest possible and adding it to
5227 * the _front_ of the runqueue. Used by CPU offline code.
5229 void sched_idle_next(void)
5231 int this_cpu = smp_processor_id();
5232 struct rq *rq = cpu_rq(this_cpu);
5233 struct task_struct *p = rq->idle;
5234 unsigned long flags;
5236 /* cpu has to be offline */
5237 BUG_ON(cpu_online(this_cpu));
5240 * Strictly not necessary since rest of the CPUs are stopped by now
5241 * and interrupts disabled on the current cpu.
5243 spin_lock_irqsave(&rq->lock, flags);
5245 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5247 /* Add idle task to the _front_ of its priority queue: */
5248 __activate_idle_task(p, rq);
5250 spin_unlock_irqrestore(&rq->lock, flags);
5254 * Ensures that the idle task is using init_mm right before its cpu goes
5257 void idle_task_exit(void)
5259 struct mm_struct *mm = current->active_mm;
5261 BUG_ON(cpu_online(smp_processor_id()));
5264 switch_mm(mm, &init_mm, current);
5268 /* called under rq->lock with disabled interrupts */
5269 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5271 struct rq *rq = cpu_rq(dead_cpu);
5273 /* Must be exiting, otherwise would be on tasklist. */
5274 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5276 /* Cannot have done final schedule yet: would have vanished. */
5277 BUG_ON(p->state == TASK_DEAD);
5282 * Drop lock around migration; if someone else moves it,
5283 * that's OK. No task can be added to this CPU, so iteration is
5285 * NOTE: interrupts should be left disabled --dev@
5287 spin_unlock(&rq->lock);
5288 move_task_off_dead_cpu(dead_cpu, p);
5289 spin_lock(&rq->lock);
5294 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5295 static void migrate_dead_tasks(unsigned int dead_cpu)
5297 struct rq *rq = cpu_rq(dead_cpu);
5298 unsigned int arr, i;
5300 for (arr = 0; arr < 2; arr++) {
5301 for (i = 0; i < MAX_PRIO; i++) {
5302 struct list_head *list = &rq->arrays[arr].queue[i];
5304 while (!list_empty(list))
5305 migrate_dead(dead_cpu, list_entry(list->next,
5306 struct task_struct, run_list));
5310 #endif /* CONFIG_HOTPLUG_CPU */
5313 * migration_call - callback that gets triggered when a CPU is added.
5314 * Here we can start up the necessary migration thread for the new CPU.
5316 static int __cpuinit
5317 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5319 struct task_struct *p;
5320 int cpu = (long)hcpu;
5321 unsigned long flags;
5325 case CPU_LOCK_ACQUIRE:
5326 mutex_lock(&sched_hotcpu_mutex);
5329 case CPU_UP_PREPARE:
5330 case CPU_UP_PREPARE_FROZEN:
5331 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5334 p->flags |= PF_NOFREEZE;
5335 kthread_bind(p, cpu);
5336 /* Must be high prio: stop_machine expects to yield to it. */
5337 rq = task_rq_lock(p, &flags);
5338 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5339 task_rq_unlock(rq, &flags);
5340 cpu_rq(cpu)->migration_thread = p;
5344 case CPU_ONLINE_FROZEN:
5345 /* Strictly unneccessary, as first user will wake it. */
5346 wake_up_process(cpu_rq(cpu)->migration_thread);
5349 #ifdef CONFIG_HOTPLUG_CPU
5350 case CPU_UP_CANCELED:
5351 case CPU_UP_CANCELED_FROZEN:
5352 if (!cpu_rq(cpu)->migration_thread)
5354 /* Unbind it from offline cpu so it can run. Fall thru. */
5355 kthread_bind(cpu_rq(cpu)->migration_thread,
5356 any_online_cpu(cpu_online_map));
5357 kthread_stop(cpu_rq(cpu)->migration_thread);
5358 cpu_rq(cpu)->migration_thread = NULL;
5362 case CPU_DEAD_FROZEN:
5363 migrate_live_tasks(cpu);
5365 kthread_stop(rq->migration_thread);
5366 rq->migration_thread = NULL;
5367 /* Idle task back to normal (off runqueue, low prio) */
5368 rq = task_rq_lock(rq->idle, &flags);
5369 deactivate_task(rq->idle, rq);
5370 rq->idle->static_prio = MAX_PRIO;
5371 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5372 migrate_dead_tasks(cpu);
5373 task_rq_unlock(rq, &flags);
5374 migrate_nr_uninterruptible(rq);
5375 BUG_ON(rq->nr_running != 0);
5377 /* No need to migrate the tasks: it was best-effort if
5378 * they didn't take sched_hotcpu_mutex. Just wake up
5379 * the requestors. */
5380 spin_lock_irq(&rq->lock);
5381 while (!list_empty(&rq->migration_queue)) {
5382 struct migration_req *req;
5384 req = list_entry(rq->migration_queue.next,
5385 struct migration_req, list);
5386 list_del_init(&req->list);
5387 complete(&req->done);
5389 spin_unlock_irq(&rq->lock);
5392 case CPU_LOCK_RELEASE:
5393 mutex_unlock(&sched_hotcpu_mutex);
5399 /* Register at highest priority so that task migration (migrate_all_tasks)
5400 * happens before everything else.
5402 static struct notifier_block __cpuinitdata migration_notifier = {
5403 .notifier_call = migration_call,
5407 int __init migration_init(void)
5409 void *cpu = (void *)(long)smp_processor_id();
5412 /* Start one for the boot CPU: */
5413 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5414 BUG_ON(err == NOTIFY_BAD);
5415 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5416 register_cpu_notifier(&migration_notifier);
5424 /* Number of possible processor ids */
5425 int nr_cpu_ids __read_mostly = NR_CPUS;
5426 EXPORT_SYMBOL(nr_cpu_ids);
5428 #undef SCHED_DOMAIN_DEBUG
5429 #ifdef SCHED_DOMAIN_DEBUG
5430 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5435 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5439 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5444 struct sched_group *group = sd->groups;
5445 cpumask_t groupmask;
5447 cpumask_scnprintf(str, NR_CPUS, sd->span);
5448 cpus_clear(groupmask);
5451 for (i = 0; i < level + 1; i++)
5453 printk("domain %d: ", level);
5455 if (!(sd->flags & SD_LOAD_BALANCE)) {
5456 printk("does not load-balance\n");
5458 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5463 printk("span %s\n", str);
5465 if (!cpu_isset(cpu, sd->span))
5466 printk(KERN_ERR "ERROR: domain->span does not contain "
5468 if (!cpu_isset(cpu, group->cpumask))
5469 printk(KERN_ERR "ERROR: domain->groups does not contain"
5473 for (i = 0; i < level + 2; i++)
5479 printk(KERN_ERR "ERROR: group is NULL\n");
5483 if (!group->__cpu_power) {
5485 printk(KERN_ERR "ERROR: domain->cpu_power not "
5489 if (!cpus_weight(group->cpumask)) {
5491 printk(KERN_ERR "ERROR: empty group\n");
5494 if (cpus_intersects(groupmask, group->cpumask)) {
5496 printk(KERN_ERR "ERROR: repeated CPUs\n");
5499 cpus_or(groupmask, groupmask, group->cpumask);
5501 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5504 group = group->next;
5505 } while (group != sd->groups);
5508 if (!cpus_equal(sd->span, groupmask))
5509 printk(KERN_ERR "ERROR: groups don't span "
5517 if (!cpus_subset(groupmask, sd->span))
5518 printk(KERN_ERR "ERROR: parent span is not a superset "
5519 "of domain->span\n");
5524 # define sched_domain_debug(sd, cpu) do { } while (0)
5527 static int sd_degenerate(struct sched_domain *sd)
5529 if (cpus_weight(sd->span) == 1)
5532 /* Following flags need at least 2 groups */
5533 if (sd->flags & (SD_LOAD_BALANCE |
5534 SD_BALANCE_NEWIDLE |
5538 SD_SHARE_PKG_RESOURCES)) {
5539 if (sd->groups != sd->groups->next)
5543 /* Following flags don't use groups */
5544 if (sd->flags & (SD_WAKE_IDLE |
5553 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5555 unsigned long cflags = sd->flags, pflags = parent->flags;
5557 if (sd_degenerate(parent))
5560 if (!cpus_equal(sd->span, parent->span))
5563 /* Does parent contain flags not in child? */
5564 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5565 if (cflags & SD_WAKE_AFFINE)
5566 pflags &= ~SD_WAKE_BALANCE;
5567 /* Flags needing groups don't count if only 1 group in parent */
5568 if (parent->groups == parent->groups->next) {
5569 pflags &= ~(SD_LOAD_BALANCE |
5570 SD_BALANCE_NEWIDLE |
5574 SD_SHARE_PKG_RESOURCES);
5576 if (~cflags & pflags)
5583 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5584 * hold the hotplug lock.
5586 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5588 struct rq *rq = cpu_rq(cpu);
5589 struct sched_domain *tmp;
5591 /* Remove the sched domains which do not contribute to scheduling. */
5592 for (tmp = sd; tmp; tmp = tmp->parent) {
5593 struct sched_domain *parent = tmp->parent;
5596 if (sd_parent_degenerate(tmp, parent)) {
5597 tmp->parent = parent->parent;
5599 parent->parent->child = tmp;
5603 if (sd && sd_degenerate(sd)) {
5609 sched_domain_debug(sd, cpu);
5611 rcu_assign_pointer(rq->sd, sd);
5614 /* cpus with isolated domains */
5615 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5617 /* Setup the mask of cpus configured for isolated domains */
5618 static int __init isolated_cpu_setup(char *str)
5620 int ints[NR_CPUS], i;
5622 str = get_options(str, ARRAY_SIZE(ints), ints);
5623 cpus_clear(cpu_isolated_map);
5624 for (i = 1; i <= ints[0]; i++)
5625 if (ints[i] < NR_CPUS)
5626 cpu_set(ints[i], cpu_isolated_map);
5630 __setup ("isolcpus=", isolated_cpu_setup);
5633 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5634 * to a function which identifies what group(along with sched group) a CPU
5635 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5636 * (due to the fact that we keep track of groups covered with a cpumask_t).
5638 * init_sched_build_groups will build a circular linked list of the groups
5639 * covered by the given span, and will set each group's ->cpumask correctly,
5640 * and ->cpu_power to 0.
5643 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5644 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5645 struct sched_group **sg))
5647 struct sched_group *first = NULL, *last = NULL;
5648 cpumask_t covered = CPU_MASK_NONE;
5651 for_each_cpu_mask(i, span) {
5652 struct sched_group *sg;
5653 int group = group_fn(i, cpu_map, &sg);
5656 if (cpu_isset(i, covered))
5659 sg->cpumask = CPU_MASK_NONE;
5660 sg->__cpu_power = 0;
5662 for_each_cpu_mask(j, span) {
5663 if (group_fn(j, cpu_map, NULL) != group)
5666 cpu_set(j, covered);
5667 cpu_set(j, sg->cpumask);
5678 #define SD_NODES_PER_DOMAIN 16
5683 * find_next_best_node - find the next node to include in a sched_domain
5684 * @node: node whose sched_domain we're building
5685 * @used_nodes: nodes already in the sched_domain
5687 * Find the next node to include in a given scheduling domain. Simply
5688 * finds the closest node not already in the @used_nodes map.
5690 * Should use nodemask_t.
5692 static int find_next_best_node(int node, unsigned long *used_nodes)
5694 int i, n, val, min_val, best_node = 0;
5698 for (i = 0; i < MAX_NUMNODES; i++) {
5699 /* Start at @node */
5700 n = (node + i) % MAX_NUMNODES;
5702 if (!nr_cpus_node(n))
5705 /* Skip already used nodes */
5706 if (test_bit(n, used_nodes))
5709 /* Simple min distance search */
5710 val = node_distance(node, n);
5712 if (val < min_val) {
5718 set_bit(best_node, used_nodes);
5723 * sched_domain_node_span - get a cpumask for a node's sched_domain
5724 * @node: node whose cpumask we're constructing
5725 * @size: number of nodes to include in this span
5727 * Given a node, construct a good cpumask for its sched_domain to span. It
5728 * should be one that prevents unnecessary balancing, but also spreads tasks
5731 static cpumask_t sched_domain_node_span(int node)
5733 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5734 cpumask_t span, nodemask;
5738 bitmap_zero(used_nodes, MAX_NUMNODES);
5740 nodemask = node_to_cpumask(node);
5741 cpus_or(span, span, nodemask);
5742 set_bit(node, used_nodes);
5744 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5745 int next_node = find_next_best_node(node, used_nodes);
5747 nodemask = node_to_cpumask(next_node);
5748 cpus_or(span, span, nodemask);
5755 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5758 * SMT sched-domains:
5760 #ifdef CONFIG_SCHED_SMT
5761 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5762 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5764 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5765 struct sched_group **sg)
5768 *sg = &per_cpu(sched_group_cpus, cpu);
5774 * multi-core sched-domains:
5776 #ifdef CONFIG_SCHED_MC
5777 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5778 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5781 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5782 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5783 struct sched_group **sg)
5786 cpumask_t mask = cpu_sibling_map[cpu];
5787 cpus_and(mask, mask, *cpu_map);
5788 group = first_cpu(mask);
5790 *sg = &per_cpu(sched_group_core, group);
5793 #elif defined(CONFIG_SCHED_MC)
5794 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5795 struct sched_group **sg)
5798 *sg = &per_cpu(sched_group_core, cpu);
5803 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5804 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5806 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5807 struct sched_group **sg)
5810 #ifdef CONFIG_SCHED_MC
5811 cpumask_t mask = cpu_coregroup_map(cpu);
5812 cpus_and(mask, mask, *cpu_map);
5813 group = first_cpu(mask);
5814 #elif defined(CONFIG_SCHED_SMT)
5815 cpumask_t mask = cpu_sibling_map[cpu];
5816 cpus_and(mask, mask, *cpu_map);
5817 group = first_cpu(mask);
5822 *sg = &per_cpu(sched_group_phys, group);
5828 * The init_sched_build_groups can't handle what we want to do with node
5829 * groups, so roll our own. Now each node has its own list of groups which
5830 * gets dynamically allocated.
5832 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5833 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5835 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5836 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5838 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5839 struct sched_group **sg)
5841 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5844 cpus_and(nodemask, nodemask, *cpu_map);
5845 group = first_cpu(nodemask);
5848 *sg = &per_cpu(sched_group_allnodes, group);
5852 static void init_numa_sched_groups_power(struct sched_group *group_head)
5854 struct sched_group *sg = group_head;
5860 for_each_cpu_mask(j, sg->cpumask) {
5861 struct sched_domain *sd;
5863 sd = &per_cpu(phys_domains, j);
5864 if (j != first_cpu(sd->groups->cpumask)) {
5866 * Only add "power" once for each
5872 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5875 if (sg != group_head)
5881 /* Free memory allocated for various sched_group structures */
5882 static void free_sched_groups(const cpumask_t *cpu_map)
5886 for_each_cpu_mask(cpu, *cpu_map) {
5887 struct sched_group **sched_group_nodes
5888 = sched_group_nodes_bycpu[cpu];
5890 if (!sched_group_nodes)
5893 for (i = 0; i < MAX_NUMNODES; i++) {
5894 cpumask_t nodemask = node_to_cpumask(i);
5895 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5897 cpus_and(nodemask, nodemask, *cpu_map);
5898 if (cpus_empty(nodemask))
5908 if (oldsg != sched_group_nodes[i])
5911 kfree(sched_group_nodes);
5912 sched_group_nodes_bycpu[cpu] = NULL;
5916 static void free_sched_groups(const cpumask_t *cpu_map)
5922 * Initialize sched groups cpu_power.
5924 * cpu_power indicates the capacity of sched group, which is used while
5925 * distributing the load between different sched groups in a sched domain.
5926 * Typically cpu_power for all the groups in a sched domain will be same unless
5927 * there are asymmetries in the topology. If there are asymmetries, group
5928 * having more cpu_power will pickup more load compared to the group having
5931 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5932 * the maximum number of tasks a group can handle in the presence of other idle
5933 * or lightly loaded groups in the same sched domain.
5935 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5937 struct sched_domain *child;
5938 struct sched_group *group;
5940 WARN_ON(!sd || !sd->groups);
5942 if (cpu != first_cpu(sd->groups->cpumask))
5947 sd->groups->__cpu_power = 0;
5950 * For perf policy, if the groups in child domain share resources
5951 * (for example cores sharing some portions of the cache hierarchy
5952 * or SMT), then set this domain groups cpu_power such that each group
5953 * can handle only one task, when there are other idle groups in the
5954 * same sched domain.
5956 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5958 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5959 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5964 * add cpu_power of each child group to this groups cpu_power
5966 group = child->groups;
5968 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5969 group = group->next;
5970 } while (group != child->groups);
5974 * Build sched domains for a given set of cpus and attach the sched domains
5975 * to the individual cpus
5977 static int build_sched_domains(const cpumask_t *cpu_map)
5980 struct sched_domain *sd;
5982 struct sched_group **sched_group_nodes = NULL;
5983 int sd_allnodes = 0;
5986 * Allocate the per-node list of sched groups
5988 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5990 if (!sched_group_nodes) {
5991 printk(KERN_WARNING "Can not alloc sched group node list\n");
5994 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5998 * Set up domains for cpus specified by the cpu_map.
6000 for_each_cpu_mask(i, *cpu_map) {
6001 struct sched_domain *sd = NULL, *p;
6002 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6004 cpus_and(nodemask, nodemask, *cpu_map);
6007 if (cpus_weight(*cpu_map)
6008 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6009 sd = &per_cpu(allnodes_domains, i);
6010 *sd = SD_ALLNODES_INIT;
6011 sd->span = *cpu_map;
6012 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6018 sd = &per_cpu(node_domains, i);
6020 sd->span = sched_domain_node_span(cpu_to_node(i));
6024 cpus_and(sd->span, sd->span, *cpu_map);
6028 sd = &per_cpu(phys_domains, i);
6030 sd->span = nodemask;
6034 cpu_to_phys_group(i, cpu_map, &sd->groups);
6036 #ifdef CONFIG_SCHED_MC
6038 sd = &per_cpu(core_domains, i);
6040 sd->span = cpu_coregroup_map(i);
6041 cpus_and(sd->span, sd->span, *cpu_map);
6044 cpu_to_core_group(i, cpu_map, &sd->groups);
6047 #ifdef CONFIG_SCHED_SMT
6049 sd = &per_cpu(cpu_domains, i);
6050 *sd = SD_SIBLING_INIT;
6051 sd->span = cpu_sibling_map[i];
6052 cpus_and(sd->span, sd->span, *cpu_map);
6055 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6059 #ifdef CONFIG_SCHED_SMT
6060 /* Set up CPU (sibling) groups */
6061 for_each_cpu_mask(i, *cpu_map) {
6062 cpumask_t this_sibling_map = cpu_sibling_map[i];
6063 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6064 if (i != first_cpu(this_sibling_map))
6067 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6071 #ifdef CONFIG_SCHED_MC
6072 /* Set up multi-core groups */
6073 for_each_cpu_mask(i, *cpu_map) {
6074 cpumask_t this_core_map = cpu_coregroup_map(i);
6075 cpus_and(this_core_map, this_core_map, *cpu_map);
6076 if (i != first_cpu(this_core_map))
6078 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6083 /* Set up physical groups */
6084 for (i = 0; i < MAX_NUMNODES; i++) {
6085 cpumask_t nodemask = node_to_cpumask(i);
6087 cpus_and(nodemask, nodemask, *cpu_map);
6088 if (cpus_empty(nodemask))
6091 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6095 /* Set up node groups */
6097 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6099 for (i = 0; i < MAX_NUMNODES; i++) {
6100 /* Set up node groups */
6101 struct sched_group *sg, *prev;
6102 cpumask_t nodemask = node_to_cpumask(i);
6103 cpumask_t domainspan;
6104 cpumask_t covered = CPU_MASK_NONE;
6107 cpus_and(nodemask, nodemask, *cpu_map);
6108 if (cpus_empty(nodemask)) {
6109 sched_group_nodes[i] = NULL;
6113 domainspan = sched_domain_node_span(i);
6114 cpus_and(domainspan, domainspan, *cpu_map);
6116 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6118 printk(KERN_WARNING "Can not alloc domain group for "
6122 sched_group_nodes[i] = sg;
6123 for_each_cpu_mask(j, nodemask) {
6124 struct sched_domain *sd;
6125 sd = &per_cpu(node_domains, j);
6128 sg->__cpu_power = 0;
6129 sg->cpumask = nodemask;
6131 cpus_or(covered, covered, nodemask);
6134 for (j = 0; j < MAX_NUMNODES; j++) {
6135 cpumask_t tmp, notcovered;
6136 int n = (i + j) % MAX_NUMNODES;
6138 cpus_complement(notcovered, covered);
6139 cpus_and(tmp, notcovered, *cpu_map);
6140 cpus_and(tmp, tmp, domainspan);
6141 if (cpus_empty(tmp))
6144 nodemask = node_to_cpumask(n);
6145 cpus_and(tmp, tmp, nodemask);
6146 if (cpus_empty(tmp))
6149 sg = kmalloc_node(sizeof(struct sched_group),
6153 "Can not alloc domain group for node %d\n", j);
6156 sg->__cpu_power = 0;
6158 sg->next = prev->next;
6159 cpus_or(covered, covered, tmp);
6166 /* Calculate CPU power for physical packages and nodes */
6167 #ifdef CONFIG_SCHED_SMT
6168 for_each_cpu_mask(i, *cpu_map) {
6169 sd = &per_cpu(cpu_domains, i);
6170 init_sched_groups_power(i, sd);
6173 #ifdef CONFIG_SCHED_MC
6174 for_each_cpu_mask(i, *cpu_map) {
6175 sd = &per_cpu(core_domains, i);
6176 init_sched_groups_power(i, sd);
6180 for_each_cpu_mask(i, *cpu_map) {
6181 sd = &per_cpu(phys_domains, i);
6182 init_sched_groups_power(i, sd);
6186 for (i = 0; i < MAX_NUMNODES; i++)
6187 init_numa_sched_groups_power(sched_group_nodes[i]);
6190 struct sched_group *sg;
6192 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6193 init_numa_sched_groups_power(sg);
6197 /* Attach the domains */
6198 for_each_cpu_mask(i, *cpu_map) {
6199 struct sched_domain *sd;
6200 #ifdef CONFIG_SCHED_SMT
6201 sd = &per_cpu(cpu_domains, i);
6202 #elif defined(CONFIG_SCHED_MC)
6203 sd = &per_cpu(core_domains, i);
6205 sd = &per_cpu(phys_domains, i);
6207 cpu_attach_domain(sd, i);
6214 free_sched_groups(cpu_map);
6219 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6221 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6223 cpumask_t cpu_default_map;
6227 * Setup mask for cpus without special case scheduling requirements.
6228 * For now this just excludes isolated cpus, but could be used to
6229 * exclude other special cases in the future.
6231 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6233 err = build_sched_domains(&cpu_default_map);
6238 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6240 free_sched_groups(cpu_map);
6244 * Detach sched domains from a group of cpus specified in cpu_map
6245 * These cpus will now be attached to the NULL domain
6247 static void detach_destroy_domains(const cpumask_t *cpu_map)
6251 for_each_cpu_mask(i, *cpu_map)
6252 cpu_attach_domain(NULL, i);
6253 synchronize_sched();
6254 arch_destroy_sched_domains(cpu_map);
6258 * Partition sched domains as specified by the cpumasks below.
6259 * This attaches all cpus from the cpumasks to the NULL domain,
6260 * waits for a RCU quiescent period, recalculates sched
6261 * domain information and then attaches them back to the
6262 * correct sched domains
6263 * Call with hotplug lock held
6265 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6267 cpumask_t change_map;
6270 cpus_and(*partition1, *partition1, cpu_online_map);
6271 cpus_and(*partition2, *partition2, cpu_online_map);
6272 cpus_or(change_map, *partition1, *partition2);
6274 /* Detach sched domains from all of the affected cpus */
6275 detach_destroy_domains(&change_map);
6276 if (!cpus_empty(*partition1))
6277 err = build_sched_domains(partition1);
6278 if (!err && !cpus_empty(*partition2))
6279 err = build_sched_domains(partition2);
6284 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6285 int arch_reinit_sched_domains(void)
6289 mutex_lock(&sched_hotcpu_mutex);
6290 detach_destroy_domains(&cpu_online_map);
6291 err = arch_init_sched_domains(&cpu_online_map);
6292 mutex_unlock(&sched_hotcpu_mutex);
6297 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6301 if (buf[0] != '0' && buf[0] != '1')
6305 sched_smt_power_savings = (buf[0] == '1');
6307 sched_mc_power_savings = (buf[0] == '1');
6309 ret = arch_reinit_sched_domains();
6311 return ret ? ret : count;
6314 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6318 #ifdef CONFIG_SCHED_SMT
6320 err = sysfs_create_file(&cls->kset.kobj,
6321 &attr_sched_smt_power_savings.attr);
6323 #ifdef CONFIG_SCHED_MC
6324 if (!err && mc_capable())
6325 err = sysfs_create_file(&cls->kset.kobj,
6326 &attr_sched_mc_power_savings.attr);
6332 #ifdef CONFIG_SCHED_MC
6333 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6335 return sprintf(page, "%u\n", sched_mc_power_savings);
6337 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6338 const char *buf, size_t count)
6340 return sched_power_savings_store(buf, count, 0);
6342 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6343 sched_mc_power_savings_store);
6346 #ifdef CONFIG_SCHED_SMT
6347 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6349 return sprintf(page, "%u\n", sched_smt_power_savings);
6351 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6352 const char *buf, size_t count)
6354 return sched_power_savings_store(buf, count, 1);
6356 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6357 sched_smt_power_savings_store);
6361 * Force a reinitialization of the sched domains hierarchy. The domains
6362 * and groups cannot be updated in place without racing with the balancing
6363 * code, so we temporarily attach all running cpus to the NULL domain
6364 * which will prevent rebalancing while the sched domains are recalculated.
6366 static int update_sched_domains(struct notifier_block *nfb,
6367 unsigned long action, void *hcpu)
6370 case CPU_UP_PREPARE:
6371 case CPU_UP_PREPARE_FROZEN:
6372 case CPU_DOWN_PREPARE:
6373 case CPU_DOWN_PREPARE_FROZEN:
6374 detach_destroy_domains(&cpu_online_map);
6377 case CPU_UP_CANCELED:
6378 case CPU_UP_CANCELED_FROZEN:
6379 case CPU_DOWN_FAILED:
6380 case CPU_DOWN_FAILED_FROZEN:
6382 case CPU_ONLINE_FROZEN:
6384 case CPU_DEAD_FROZEN:
6386 * Fall through and re-initialise the domains.
6393 /* The hotplug lock is already held by cpu_up/cpu_down */
6394 arch_init_sched_domains(&cpu_online_map);
6399 void __init sched_init_smp(void)
6401 cpumask_t non_isolated_cpus;
6403 mutex_lock(&sched_hotcpu_mutex);
6404 arch_init_sched_domains(&cpu_online_map);
6405 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6406 if (cpus_empty(non_isolated_cpus))
6407 cpu_set(smp_processor_id(), non_isolated_cpus);
6408 mutex_unlock(&sched_hotcpu_mutex);
6409 /* XXX: Theoretical race here - CPU may be hotplugged now */
6410 hotcpu_notifier(update_sched_domains, 0);
6412 /* Move init over to a non-isolated CPU */
6413 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6417 void __init sched_init_smp(void)
6420 #endif /* CONFIG_SMP */
6422 int in_sched_functions(unsigned long addr)
6424 /* Linker adds these: start and end of __sched functions */
6425 extern char __sched_text_start[], __sched_text_end[];
6427 return in_lock_functions(addr) ||
6428 (addr >= (unsigned long)__sched_text_start
6429 && addr < (unsigned long)__sched_text_end);
6432 void __init sched_init(void)
6435 int highest_cpu = 0;
6437 for_each_possible_cpu(i) {
6438 struct prio_array *array;
6442 spin_lock_init(&rq->lock);
6443 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6445 rq->active = rq->arrays;
6446 rq->expired = rq->arrays + 1;
6447 rq->best_expired_prio = MAX_PRIO;
6451 for (j = 1; j < 3; j++)
6452 rq->cpu_load[j] = 0;
6453 rq->active_balance = 0;
6456 rq->migration_thread = NULL;
6457 INIT_LIST_HEAD(&rq->migration_queue);
6459 atomic_set(&rq->nr_iowait, 0);
6461 for (j = 0; j < 2; j++) {
6462 array = rq->arrays + j;
6463 for (k = 0; k < MAX_PRIO; k++) {
6464 INIT_LIST_HEAD(array->queue + k);
6465 __clear_bit(k, array->bitmap);
6467 // delimiter for bitsearch
6468 __set_bit(MAX_PRIO, array->bitmap);
6473 set_load_weight(&init_task);
6476 nr_cpu_ids = highest_cpu + 1;
6477 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6480 #ifdef CONFIG_RT_MUTEXES
6481 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6485 * The boot idle thread does lazy MMU switching as well:
6487 atomic_inc(&init_mm.mm_count);
6488 enter_lazy_tlb(&init_mm, current);
6491 * Make us the idle thread. Technically, schedule() should not be
6492 * called from this thread, however somewhere below it might be,
6493 * but because we are the idle thread, we just pick up running again
6494 * when this runqueue becomes "idle".
6496 init_idle(current, smp_processor_id());
6499 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6500 void __might_sleep(char *file, int line)
6503 static unsigned long prev_jiffy; /* ratelimiting */
6505 if ((in_atomic() || irqs_disabled()) &&
6506 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6507 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6509 prev_jiffy = jiffies;
6510 printk(KERN_ERR "BUG: sleeping function called from invalid"
6511 " context at %s:%d\n", file, line);
6512 printk("in_atomic():%d, irqs_disabled():%d\n",
6513 in_atomic(), irqs_disabled());
6514 debug_show_held_locks(current);
6515 if (irqs_disabled())
6516 print_irqtrace_events(current);
6521 EXPORT_SYMBOL(__might_sleep);
6524 #ifdef CONFIG_MAGIC_SYSRQ
6525 void normalize_rt_tasks(void)
6527 struct prio_array *array;
6528 struct task_struct *g, *p;
6529 unsigned long flags;
6532 read_lock_irq(&tasklist_lock);
6534 do_each_thread(g, p) {
6538 spin_lock_irqsave(&p->pi_lock, flags);
6539 rq = __task_rq_lock(p);
6543 deactivate_task(p, task_rq(p));
6544 __setscheduler(p, SCHED_NORMAL, 0);
6546 __activate_task(p, task_rq(p));
6547 resched_task(rq->curr);
6550 __task_rq_unlock(rq);
6551 spin_unlock_irqrestore(&p->pi_lock, flags);
6552 } while_each_thread(g, p);
6554 read_unlock_irq(&tasklist_lock);
6557 #endif /* CONFIG_MAGIC_SYSRQ */
6561 * These functions are only useful for the IA64 MCA handling.
6563 * They can only be called when the whole system has been
6564 * stopped - every CPU needs to be quiescent, and no scheduling
6565 * activity can take place. Using them for anything else would
6566 * be a serious bug, and as a result, they aren't even visible
6567 * under any other configuration.
6571 * curr_task - return the current task for a given cpu.
6572 * @cpu: the processor in question.
6574 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6576 struct task_struct *curr_task(int cpu)
6578 return cpu_curr(cpu);
6582 * set_curr_task - set the current task for a given cpu.
6583 * @cpu: the processor in question.
6584 * @p: the task pointer to set.
6586 * Description: This function must only be used when non-maskable interrupts
6587 * are serviced on a separate stack. It allows the architecture to switch the
6588 * notion of the current task on a cpu in a non-blocking manner. This function
6589 * must be called with all CPU's synchronized, and interrupts disabled, the
6590 * and caller must save the original value of the current task (see
6591 * curr_task() above) and restore that value before reenabling interrupts and
6592 * re-starting the system.
6594 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6596 void set_curr_task(int cpu, struct task_struct *p)