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))
95 * These are the 'tuning knobs' of the scheduler:
97 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
98 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
99 * Timeslices get refilled after they expire.
101 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
102 #define DEF_TIMESLICE (100 * HZ / 1000)
103 #define ON_RUNQUEUE_WEIGHT 30
104 #define CHILD_PENALTY 95
105 #define PARENT_PENALTY 100
106 #define EXIT_WEIGHT 3
107 #define PRIO_BONUS_RATIO 25
108 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
109 #define INTERACTIVE_DELTA 2
110 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
111 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
112 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
115 * If a task is 'interactive' then we reinsert it in the active
116 * array after it has expired its current timeslice. (it will not
117 * continue to run immediately, it will still roundrobin with
118 * other interactive tasks.)
120 * This part scales the interactivity limit depending on niceness.
122 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
123 * Here are a few examples of different nice levels:
125 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
126 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
127 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
129 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
131 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
132 * priority range a task can explore, a value of '1' means the
133 * task is rated interactive.)
135 * Ie. nice +19 tasks can never get 'interactive' enough to be
136 * reinserted into the active array. And only heavily CPU-hog nice -20
137 * tasks will be expired. Default nice 0 tasks are somewhere between,
138 * it takes some effort for them to get interactive, but it's not
142 #define CURRENT_BONUS(p) \
143 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
146 #define GRANULARITY (10 * HZ / 1000 ? : 1)
149 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
150 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
153 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
154 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
157 #define SCALE(v1,v1_max,v2_max) \
158 (v1) * (v2_max) / (v1_max)
161 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
164 #define TASK_INTERACTIVE(p) \
165 ((p)->prio <= (p)->static_prio - DELTA(p))
167 #define INTERACTIVE_SLEEP(p) \
168 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
169 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
171 #define TASK_PREEMPTS_CURR(p, rq) \
172 ((p)->prio < (rq)->curr->prio)
174 #define SCALE_PRIO(x, prio) \
175 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
177 static unsigned int static_prio_timeslice(int static_prio)
179 if (static_prio < NICE_TO_PRIO(0))
180 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
182 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
187 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
188 * Since cpu_power is a 'constant', we can use a reciprocal divide.
190 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
192 return reciprocal_divide(load, sg->reciprocal_cpu_power);
196 * Each time a sched group cpu_power is changed,
197 * we must compute its reciprocal value
199 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
201 sg->__cpu_power += val;
202 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
207 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
208 * to time slice values: [800ms ... 100ms ... 5ms]
210 * The higher a thread's priority, the bigger timeslices
211 * it gets during one round of execution. But even the lowest
212 * priority thread gets MIN_TIMESLICE worth of execution time.
215 static inline unsigned int task_timeslice(struct task_struct *p)
217 return static_prio_timeslice(p->static_prio);
221 * These are the runqueue data structures:
225 unsigned int nr_active;
226 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
227 struct list_head queue[MAX_PRIO];
231 * This is the main, per-CPU runqueue data structure.
233 * Locking rule: those places that want to lock multiple runqueues
234 * (such as the load balancing or the thread migration code), lock
235 * acquire operations must be ordered by ascending &runqueue.
241 * nr_running and cpu_load should be in the same cacheline because
242 * remote CPUs use both these fields when doing load calculation.
244 unsigned long nr_running;
245 unsigned long raw_weighted_load;
247 unsigned long cpu_load[3];
248 unsigned char idle_at_tick;
250 unsigned char in_nohz_recently;
253 unsigned long long nr_switches;
256 * This is part of a global counter where only the total sum
257 * over all CPUs matters. A task can increase this counter on
258 * one CPU and if it got migrated afterwards it may decrease
259 * it on another CPU. Always updated under the runqueue lock:
261 unsigned long nr_uninterruptible;
263 unsigned long expired_timestamp;
264 /* Cached timestamp set by update_cpu_clock() */
265 unsigned long long most_recent_timestamp;
266 struct task_struct *curr, *idle;
267 unsigned long next_balance;
268 struct mm_struct *prev_mm;
269 struct prio_array *active, *expired, arrays[2];
270 int best_expired_prio;
274 struct sched_domain *sd;
276 /* For active balancing */
279 int cpu; /* cpu of this runqueue */
281 struct task_struct *migration_thread;
282 struct list_head migration_queue;
285 #ifdef CONFIG_SCHEDSTATS
287 struct sched_info rq_sched_info;
289 /* sys_sched_yield() stats */
290 unsigned long yld_exp_empty;
291 unsigned long yld_act_empty;
292 unsigned long yld_both_empty;
293 unsigned long yld_cnt;
295 /* schedule() stats */
296 unsigned long sched_switch;
297 unsigned long sched_cnt;
298 unsigned long sched_goidle;
300 /* try_to_wake_up() stats */
301 unsigned long ttwu_cnt;
302 unsigned long ttwu_local;
304 struct lock_class_key rq_lock_key;
307 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
308 static DEFINE_MUTEX(sched_hotcpu_mutex);
310 static inline int cpu_of(struct rq *rq)
320 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
321 * See detach_destroy_domains: synchronize_sched for details.
323 * The domain tree of any CPU may only be accessed from within
324 * preempt-disabled sections.
326 #define for_each_domain(cpu, __sd) \
327 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
329 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
330 #define this_rq() (&__get_cpu_var(runqueues))
331 #define task_rq(p) cpu_rq(task_cpu(p))
332 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
334 #ifndef prepare_arch_switch
335 # define prepare_arch_switch(next) do { } while (0)
337 #ifndef finish_arch_switch
338 # define finish_arch_switch(prev) do { } while (0)
341 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
342 static inline int task_running(struct rq *rq, struct task_struct *p)
344 return rq->curr == p;
347 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
351 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
353 #ifdef CONFIG_DEBUG_SPINLOCK
354 /* this is a valid case when another task releases the spinlock */
355 rq->lock.owner = current;
358 * If we are tracking spinlock dependencies then we have to
359 * fix up the runqueue lock - which gets 'carried over' from
362 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
364 spin_unlock_irq(&rq->lock);
367 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
368 static inline int task_running(struct rq *rq, struct task_struct *p)
373 return rq->curr == p;
377 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
381 * We can optimise this out completely for !SMP, because the
382 * SMP rebalancing from interrupt is the only thing that cares
387 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
388 spin_unlock_irq(&rq->lock);
390 spin_unlock(&rq->lock);
394 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
398 * After ->oncpu is cleared, the task can be moved to a different CPU.
399 * We must ensure this doesn't happen until the switch is completely
405 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
409 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
412 * __task_rq_lock - lock the runqueue a given task resides on.
413 * Must be called interrupts disabled.
415 static inline struct rq *__task_rq_lock(struct task_struct *p)
422 spin_lock(&rq->lock);
423 if (unlikely(rq != task_rq(p))) {
424 spin_unlock(&rq->lock);
425 goto repeat_lock_task;
431 * task_rq_lock - lock the runqueue a given task resides on and disable
432 * interrupts. Note the ordering: we can safely lookup the task_rq without
433 * explicitly disabling preemption.
435 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
441 local_irq_save(*flags);
443 spin_lock(&rq->lock);
444 if (unlikely(rq != task_rq(p))) {
445 spin_unlock_irqrestore(&rq->lock, *flags);
446 goto repeat_lock_task;
451 static inline void __task_rq_unlock(struct rq *rq)
454 spin_unlock(&rq->lock);
457 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
460 spin_unlock_irqrestore(&rq->lock, *flags);
464 * this_rq_lock - lock this runqueue and disable interrupts.
466 static inline struct rq *this_rq_lock(void)
473 spin_lock(&rq->lock);
478 #include "sched_stats.h"
481 * Adding/removing a task to/from a priority array:
483 static void dequeue_task(struct task_struct *p, struct prio_array *array)
486 list_del(&p->run_list);
487 if (list_empty(array->queue + p->prio))
488 __clear_bit(p->prio, array->bitmap);
491 static void enqueue_task(struct task_struct *p, struct prio_array *array)
493 sched_info_queued(p);
494 list_add_tail(&p->run_list, array->queue + p->prio);
495 __set_bit(p->prio, array->bitmap);
501 * Put task to the end of the run list without the overhead of dequeue
502 * followed by enqueue.
504 static void requeue_task(struct task_struct *p, struct prio_array *array)
506 list_move_tail(&p->run_list, array->queue + p->prio);
510 enqueue_task_head(struct task_struct *p, struct prio_array *array)
512 list_add(&p->run_list, array->queue + p->prio);
513 __set_bit(p->prio, array->bitmap);
519 * __normal_prio - return the priority that is based on the static
520 * priority but is modified by bonuses/penalties.
522 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
523 * into the -5 ... 0 ... +5 bonus/penalty range.
525 * We use 25% of the full 0...39 priority range so that:
527 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
528 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
530 * Both properties are important to certain workloads.
533 static inline int __normal_prio(struct task_struct *p)
537 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
539 prio = p->static_prio - bonus;
540 if (prio < MAX_RT_PRIO)
542 if (prio > MAX_PRIO-1)
548 * To aid in avoiding the subversion of "niceness" due to uneven distribution
549 * of tasks with abnormal "nice" values across CPUs the contribution that
550 * each task makes to its run queue's load is weighted according to its
551 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
552 * scaled version of the new time slice allocation that they receive on time
557 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
558 * If static_prio_timeslice() is ever changed to break this assumption then
559 * this code will need modification
561 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
562 #define LOAD_WEIGHT(lp) \
563 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
564 #define PRIO_TO_LOAD_WEIGHT(prio) \
565 LOAD_WEIGHT(static_prio_timeslice(prio))
566 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
567 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
569 static void set_load_weight(struct task_struct *p)
571 if (has_rt_policy(p)) {
573 if (p == task_rq(p)->migration_thread)
575 * The migration thread does the actual balancing.
576 * Giving its load any weight will skew balancing
582 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
584 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
588 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
590 rq->raw_weighted_load += p->load_weight;
594 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
596 rq->raw_weighted_load -= p->load_weight;
599 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
602 inc_raw_weighted_load(rq, p);
605 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
608 dec_raw_weighted_load(rq, p);
612 * Calculate the expected normal priority: i.e. priority
613 * without taking RT-inheritance into account. Might be
614 * boosted by interactivity modifiers. Changes upon fork,
615 * setprio syscalls, and whenever the interactivity
616 * estimator recalculates.
618 static inline int normal_prio(struct task_struct *p)
622 if (has_rt_policy(p))
623 prio = MAX_RT_PRIO-1 - p->rt_priority;
625 prio = __normal_prio(p);
630 * Calculate the current priority, i.e. the priority
631 * taken into account by the scheduler. This value might
632 * be boosted by RT tasks, or might be boosted by
633 * interactivity modifiers. Will be RT if the task got
634 * RT-boosted. If not then it returns p->normal_prio.
636 static int effective_prio(struct task_struct *p)
638 p->normal_prio = normal_prio(p);
640 * If we are RT tasks or we were boosted to RT priority,
641 * keep the priority unchanged. Otherwise, update priority
642 * to the normal priority:
644 if (!rt_prio(p->prio))
645 return p->normal_prio;
650 * __activate_task - move a task to the runqueue.
652 static void __activate_task(struct task_struct *p, struct rq *rq)
654 struct prio_array *target = rq->active;
657 target = rq->expired;
658 enqueue_task(p, target);
659 inc_nr_running(p, rq);
663 * __activate_idle_task - move idle task to the _front_ of runqueue.
665 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
667 enqueue_task_head(p, rq->active);
668 inc_nr_running(p, rq);
672 * Recalculate p->normal_prio and p->prio after having slept,
673 * updating the sleep-average too:
675 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
677 /* Caller must always ensure 'now >= p->timestamp' */
678 unsigned long sleep_time = now - p->timestamp;
683 if (likely(sleep_time > 0)) {
685 * This ceiling is set to the lowest priority that would allow
686 * a task to be reinserted into the active array on timeslice
689 unsigned long ceiling = INTERACTIVE_SLEEP(p);
691 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
693 * Prevents user tasks from achieving best priority
694 * with one single large enough sleep.
696 p->sleep_avg = ceiling;
698 * Using INTERACTIVE_SLEEP() as a ceiling places a
699 * nice(0) task 1ms sleep away from promotion, and
700 * gives it 700ms to round-robin with no chance of
701 * being demoted. This is more than generous, so
702 * mark this sleep as non-interactive to prevent the
703 * on-runqueue bonus logic from intervening should
704 * this task not receive cpu immediately.
706 p->sleep_type = SLEEP_NONINTERACTIVE;
709 * Tasks waking from uninterruptible sleep are
710 * limited in their sleep_avg rise as they
711 * are likely to be waiting on I/O
713 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
714 if (p->sleep_avg >= ceiling)
716 else if (p->sleep_avg + sleep_time >=
718 p->sleep_avg = ceiling;
724 * This code gives a bonus to interactive tasks.
726 * The boost works by updating the 'average sleep time'
727 * value here, based on ->timestamp. The more time a
728 * task spends sleeping, the higher the average gets -
729 * and the higher the priority boost gets as well.
731 p->sleep_avg += sleep_time;
734 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
735 p->sleep_avg = NS_MAX_SLEEP_AVG;
738 return effective_prio(p);
742 * activate_task - move a task to the runqueue and do priority recalculation
744 * Update all the scheduling statistics stuff. (sleep average
745 * calculation, priority modifiers, etc.)
747 static void activate_task(struct task_struct *p, struct rq *rq, int local)
749 unsigned long long now;
757 /* Compensate for drifting sched_clock */
758 struct rq *this_rq = this_rq();
759 now = (now - this_rq->most_recent_timestamp)
760 + rq->most_recent_timestamp;
765 * Sleep time is in units of nanosecs, so shift by 20 to get a
766 * milliseconds-range estimation of the amount of time that the task
769 if (unlikely(prof_on == SLEEP_PROFILING)) {
770 if (p->state == TASK_UNINTERRUPTIBLE)
771 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
772 (now - p->timestamp) >> 20);
775 p->prio = recalc_task_prio(p, now);
778 * This checks to make sure it's not an uninterruptible task
779 * that is now waking up.
781 if (p->sleep_type == SLEEP_NORMAL) {
783 * Tasks which were woken up by interrupts (ie. hw events)
784 * are most likely of interactive nature. So we give them
785 * the credit of extending their sleep time to the period
786 * of time they spend on the runqueue, waiting for execution
787 * on a CPU, first time around:
790 p->sleep_type = SLEEP_INTERRUPTED;
793 * Normal first-time wakeups get a credit too for
794 * on-runqueue time, but it will be weighted down:
796 p->sleep_type = SLEEP_INTERACTIVE;
801 __activate_task(p, rq);
805 * deactivate_task - remove a task from the runqueue.
807 static void deactivate_task(struct task_struct *p, struct rq *rq)
809 dec_nr_running(p, rq);
810 dequeue_task(p, p->array);
815 * resched_task - mark a task 'to be rescheduled now'.
817 * On UP this means the setting of the need_resched flag, on SMP it
818 * might also involve a cross-CPU call to trigger the scheduler on
823 #ifndef tsk_is_polling
824 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
827 static void resched_task(struct task_struct *p)
831 assert_spin_locked(&task_rq(p)->lock);
833 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
836 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
839 if (cpu == smp_processor_id())
842 /* NEED_RESCHED must be visible before we test polling */
844 if (!tsk_is_polling(p))
845 smp_send_reschedule(cpu);
848 static void resched_cpu(int cpu)
850 struct rq *rq = cpu_rq(cpu);
853 if (!spin_trylock_irqsave(&rq->lock, flags))
855 resched_task(cpu_curr(cpu));
856 spin_unlock_irqrestore(&rq->lock, flags);
859 static inline void resched_task(struct task_struct *p)
861 assert_spin_locked(&task_rq(p)->lock);
862 set_tsk_need_resched(p);
867 * task_curr - is this task currently executing on a CPU?
868 * @p: the task in question.
870 inline int task_curr(const struct task_struct *p)
872 return cpu_curr(task_cpu(p)) == p;
875 /* Used instead of source_load when we know the type == 0 */
876 unsigned long weighted_cpuload(const int cpu)
878 return cpu_rq(cpu)->raw_weighted_load;
883 void set_task_cpu(struct task_struct *p, unsigned int cpu)
885 task_thread_info(p)->cpu = cpu;
888 struct migration_req {
889 struct list_head list;
891 struct task_struct *task;
894 struct completion done;
898 * The task's runqueue lock must be held.
899 * Returns true if you have to wait for migration thread.
902 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
904 struct rq *rq = task_rq(p);
907 * If the task is not on a runqueue (and not running), then
908 * it is sufficient to simply update the task's cpu field.
910 if (!p->array && !task_running(rq, p)) {
911 set_task_cpu(p, dest_cpu);
915 init_completion(&req->done);
917 req->dest_cpu = dest_cpu;
918 list_add(&req->list, &rq->migration_queue);
924 * wait_task_inactive - wait for a thread to unschedule.
926 * The caller must ensure that the task *will* unschedule sometime soon,
927 * else this function might spin for a *long* time. This function can't
928 * be called with interrupts off, or it may introduce deadlock with
929 * smp_call_function() if an IPI is sent by the same process we are
930 * waiting to become inactive.
932 void wait_task_inactive(struct task_struct *p)
936 struct prio_array *array;
941 * We do the initial early heuristics without holding
942 * any task-queue locks at all. We'll only try to get
943 * the runqueue lock when things look like they will
949 * If the task is actively running on another CPU
950 * still, just relax and busy-wait without holding
953 * NOTE! Since we don't hold any locks, it's not
954 * even sure that "rq" stays as the right runqueue!
955 * But we don't care, since "task_running()" will
956 * return false if the runqueue has changed and p
957 * is actually now running somewhere else!
959 while (task_running(rq, p))
963 * Ok, time to look more closely! We need the rq
964 * lock now, to be *sure*. If we're wrong, we'll
965 * just go back and repeat.
967 rq = task_rq_lock(p, &flags);
968 running = task_running(rq, p);
970 task_rq_unlock(rq, &flags);
973 * Was it really running after all now that we
974 * checked with the proper locks actually held?
976 * Oops. Go back and try again..
978 if (unlikely(running)) {
984 * It's not enough that it's not actively running,
985 * it must be off the runqueue _entirely_, and not
988 * So if it wa still runnable (but just not actively
989 * running right now), it's preempted, and we should
990 * yield - it could be a while.
992 if (unlikely(array)) {
998 * Ahh, all good. It wasn't running, and it wasn't
999 * runnable, which means that it will never become
1000 * running in the future either. We're all done!
1005 * kick_process - kick a running thread to enter/exit the kernel
1006 * @p: the to-be-kicked thread
1008 * Cause a process which is running on another CPU to enter
1009 * kernel-mode, without any delay. (to get signals handled.)
1011 * NOTE: this function doesnt have to take the runqueue lock,
1012 * because all it wants to ensure is that the remote task enters
1013 * the kernel. If the IPI races and the task has been migrated
1014 * to another CPU then no harm is done and the purpose has been
1017 void kick_process(struct task_struct *p)
1023 if ((cpu != smp_processor_id()) && task_curr(p))
1024 smp_send_reschedule(cpu);
1029 * Return a low guess at the load of a migration-source cpu weighted
1030 * according to the scheduling class and "nice" value.
1032 * We want to under-estimate the load of migration sources, to
1033 * balance conservatively.
1035 static inline unsigned long source_load(int cpu, int type)
1037 struct rq *rq = cpu_rq(cpu);
1040 return rq->raw_weighted_load;
1042 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1046 * Return a high guess at the load of a migration-target cpu weighted
1047 * according to the scheduling class and "nice" value.
1049 static inline unsigned long target_load(int cpu, int type)
1051 struct rq *rq = cpu_rq(cpu);
1054 return rq->raw_weighted_load;
1056 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1060 * Return the average load per task on the cpu's run queue
1062 static inline unsigned long cpu_avg_load_per_task(int cpu)
1064 struct rq *rq = cpu_rq(cpu);
1065 unsigned long n = rq->nr_running;
1067 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1071 * find_idlest_group finds and returns the least busy CPU group within the
1074 static struct sched_group *
1075 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1077 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1078 unsigned long min_load = ULONG_MAX, this_load = 0;
1079 int load_idx = sd->forkexec_idx;
1080 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1083 unsigned long load, avg_load;
1087 /* Skip over this group if it has no CPUs allowed */
1088 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1091 local_group = cpu_isset(this_cpu, group->cpumask);
1093 /* Tally up the load of all CPUs in the group */
1096 for_each_cpu_mask(i, group->cpumask) {
1097 /* Bias balancing toward cpus of our domain */
1099 load = source_load(i, load_idx);
1101 load = target_load(i, load_idx);
1106 /* Adjust by relative CPU power of the group */
1107 avg_load = sg_div_cpu_power(group,
1108 avg_load * SCHED_LOAD_SCALE);
1111 this_load = avg_load;
1113 } else if (avg_load < min_load) {
1114 min_load = avg_load;
1118 group = group->next;
1119 } while (group != sd->groups);
1121 if (!idlest || 100*this_load < imbalance*min_load)
1127 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1130 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1133 unsigned long load, min_load = ULONG_MAX;
1137 /* Traverse only the allowed CPUs */
1138 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1140 for_each_cpu_mask(i, tmp) {
1141 load = weighted_cpuload(i);
1143 if (load < min_load || (load == min_load && i == this_cpu)) {
1153 * sched_balance_self: balance the current task (running on cpu) in domains
1154 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1157 * Balance, ie. select the least loaded group.
1159 * Returns the target CPU number, or the same CPU if no balancing is needed.
1161 * preempt must be disabled.
1163 static int sched_balance_self(int cpu, int flag)
1165 struct task_struct *t = current;
1166 struct sched_domain *tmp, *sd = NULL;
1168 for_each_domain(cpu, tmp) {
1170 * If power savings logic is enabled for a domain, stop there.
1172 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1174 if (tmp->flags & flag)
1180 struct sched_group *group;
1181 int new_cpu, weight;
1183 if (!(sd->flags & flag)) {
1189 group = find_idlest_group(sd, t, cpu);
1195 new_cpu = find_idlest_cpu(group, t, cpu);
1196 if (new_cpu == -1 || new_cpu == cpu) {
1197 /* Now try balancing at a lower domain level of cpu */
1202 /* Now try balancing at a lower domain level of new_cpu */
1205 weight = cpus_weight(span);
1206 for_each_domain(cpu, tmp) {
1207 if (weight <= cpus_weight(tmp->span))
1209 if (tmp->flags & flag)
1212 /* while loop will break here if sd == NULL */
1218 #endif /* CONFIG_SMP */
1221 * wake_idle() will wake a task on an idle cpu if task->cpu is
1222 * not idle and an idle cpu is available. The span of cpus to
1223 * search starts with cpus closest then further out as needed,
1224 * so we always favor a closer, idle cpu.
1226 * Returns the CPU we should wake onto.
1228 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1229 static int wake_idle(int cpu, struct task_struct *p)
1232 struct sched_domain *sd;
1236 * If it is idle, then it is the best cpu to run this task.
1238 * This cpu is also the best, if it has more than one task already.
1239 * Siblings must be also busy(in most cases) as they didn't already
1240 * pickup the extra load from this cpu and hence we need not check
1241 * sibling runqueue info. This will avoid the checks and cache miss
1242 * penalities associated with that.
1244 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1247 for_each_domain(cpu, sd) {
1248 if (sd->flags & SD_WAKE_IDLE) {
1249 cpus_and(tmp, sd->span, p->cpus_allowed);
1250 for_each_cpu_mask(i, tmp) {
1261 static inline int wake_idle(int cpu, struct task_struct *p)
1268 * try_to_wake_up - wake up a thread
1269 * @p: the to-be-woken-up thread
1270 * @state: the mask of task states that can be woken
1271 * @sync: do a synchronous wakeup?
1273 * Put it on the run-queue if it's not already there. The "current"
1274 * thread is always on the run-queue (except when the actual
1275 * re-schedule is in progress), and as such you're allowed to do
1276 * the simpler "current->state = TASK_RUNNING" to mark yourself
1277 * runnable without the overhead of this.
1279 * returns failure only if the task is already active.
1281 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1283 int cpu, this_cpu, success = 0;
1284 unsigned long flags;
1288 struct sched_domain *sd, *this_sd = NULL;
1289 unsigned long load, this_load;
1293 rq = task_rq_lock(p, &flags);
1294 old_state = p->state;
1295 if (!(old_state & state))
1302 this_cpu = smp_processor_id();
1305 if (unlikely(task_running(rq, p)))
1310 schedstat_inc(rq, ttwu_cnt);
1311 if (cpu == this_cpu) {
1312 schedstat_inc(rq, ttwu_local);
1316 for_each_domain(this_cpu, sd) {
1317 if (cpu_isset(cpu, sd->span)) {
1318 schedstat_inc(sd, ttwu_wake_remote);
1324 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1328 * Check for affine wakeup and passive balancing possibilities.
1331 int idx = this_sd->wake_idx;
1332 unsigned int imbalance;
1334 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1336 load = source_load(cpu, idx);
1337 this_load = target_load(this_cpu, idx);
1339 new_cpu = this_cpu; /* Wake to this CPU if we can */
1341 if (this_sd->flags & SD_WAKE_AFFINE) {
1342 unsigned long tl = this_load;
1343 unsigned long tl_per_task;
1345 tl_per_task = cpu_avg_load_per_task(this_cpu);
1348 * If sync wakeup then subtract the (maximum possible)
1349 * effect of the currently running task from the load
1350 * of the current CPU:
1353 tl -= current->load_weight;
1356 tl + target_load(cpu, idx) <= tl_per_task) ||
1357 100*(tl + p->load_weight) <= imbalance*load) {
1359 * This domain has SD_WAKE_AFFINE and
1360 * p is cache cold in this domain, and
1361 * there is no bad imbalance.
1363 schedstat_inc(this_sd, ttwu_move_affine);
1369 * Start passive balancing when half the imbalance_pct
1372 if (this_sd->flags & SD_WAKE_BALANCE) {
1373 if (imbalance*this_load <= 100*load) {
1374 schedstat_inc(this_sd, ttwu_move_balance);
1380 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1382 new_cpu = wake_idle(new_cpu, p);
1383 if (new_cpu != cpu) {
1384 set_task_cpu(p, new_cpu);
1385 task_rq_unlock(rq, &flags);
1386 /* might preempt at this point */
1387 rq = task_rq_lock(p, &flags);
1388 old_state = p->state;
1389 if (!(old_state & state))
1394 this_cpu = smp_processor_id();
1399 #endif /* CONFIG_SMP */
1400 if (old_state == TASK_UNINTERRUPTIBLE) {
1401 rq->nr_uninterruptible--;
1403 * Tasks on involuntary sleep don't earn
1404 * sleep_avg beyond just interactive state.
1406 p->sleep_type = SLEEP_NONINTERACTIVE;
1410 * Tasks that have marked their sleep as noninteractive get
1411 * woken up with their sleep average not weighted in an
1414 if (old_state & TASK_NONINTERACTIVE)
1415 p->sleep_type = SLEEP_NONINTERACTIVE;
1418 activate_task(p, rq, cpu == this_cpu);
1420 * Sync wakeups (i.e. those types of wakeups where the waker
1421 * has indicated that it will leave the CPU in short order)
1422 * don't trigger a preemption, if the woken up task will run on
1423 * this cpu. (in this case the 'I will reschedule' promise of
1424 * the waker guarantees that the freshly woken up task is going
1425 * to be considered on this CPU.)
1427 if (!sync || cpu != this_cpu) {
1428 if (TASK_PREEMPTS_CURR(p, rq))
1429 resched_task(rq->curr);
1434 p->state = TASK_RUNNING;
1436 task_rq_unlock(rq, &flags);
1441 int fastcall wake_up_process(struct task_struct *p)
1443 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1444 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1446 EXPORT_SYMBOL(wake_up_process);
1448 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1450 return try_to_wake_up(p, state, 0);
1453 static void task_running_tick(struct rq *rq, struct task_struct *p);
1455 * Perform scheduler related setup for a newly forked process p.
1456 * p is forked by current.
1458 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1460 int cpu = get_cpu();
1463 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1465 set_task_cpu(p, cpu);
1468 * We mark the process as running here, but have not actually
1469 * inserted it onto the runqueue yet. This guarantees that
1470 * nobody will actually run it, and a signal or other external
1471 * event cannot wake it up and insert it on the runqueue either.
1473 p->state = TASK_RUNNING;
1476 * Make sure we do not leak PI boosting priority to the child:
1478 p->prio = current->normal_prio;
1480 INIT_LIST_HEAD(&p->run_list);
1482 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1483 if (unlikely(sched_info_on()))
1484 memset(&p->sched_info, 0, sizeof(p->sched_info));
1486 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1489 #ifdef CONFIG_PREEMPT
1490 /* Want to start with kernel preemption disabled. */
1491 task_thread_info(p)->preempt_count = 1;
1494 * Share the timeslice between parent and child, thus the
1495 * total amount of pending timeslices in the system doesn't change,
1496 * resulting in more scheduling fairness.
1498 local_irq_disable();
1499 p->time_slice = (current->time_slice + 1) >> 1;
1501 * The remainder of the first timeslice might be recovered by
1502 * the parent if the child exits early enough.
1504 p->first_time_slice = 1;
1505 current->time_slice >>= 1;
1506 p->timestamp = sched_clock();
1507 if (unlikely(!current->time_slice)) {
1509 * This case is rare, it happens when the parent has only
1510 * a single jiffy left from its timeslice. Taking the
1511 * runqueue lock is not a problem.
1513 current->time_slice = 1;
1514 task_running_tick(cpu_rq(cpu), current);
1521 * wake_up_new_task - wake up a newly created task for the first time.
1523 * This function will do some initial scheduler statistics housekeeping
1524 * that must be done for every newly created context, then puts the task
1525 * on the runqueue and wakes it.
1527 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1529 struct rq *rq, *this_rq;
1530 unsigned long flags;
1533 rq = task_rq_lock(p, &flags);
1534 BUG_ON(p->state != TASK_RUNNING);
1535 this_cpu = smp_processor_id();
1539 * We decrease the sleep average of forking parents
1540 * and children as well, to keep max-interactive tasks
1541 * from forking tasks that are max-interactive. The parent
1542 * (current) is done further down, under its lock.
1544 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1545 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1547 p->prio = effective_prio(p);
1549 if (likely(cpu == this_cpu)) {
1550 if (!(clone_flags & CLONE_VM)) {
1552 * The VM isn't cloned, so we're in a good position to
1553 * do child-runs-first in anticipation of an exec. This
1554 * usually avoids a lot of COW overhead.
1556 if (unlikely(!current->array))
1557 __activate_task(p, rq);
1559 p->prio = current->prio;
1560 p->normal_prio = current->normal_prio;
1561 list_add_tail(&p->run_list, ¤t->run_list);
1562 p->array = current->array;
1563 p->array->nr_active++;
1564 inc_nr_running(p, rq);
1568 /* Run child last */
1569 __activate_task(p, rq);
1571 * We skip the following code due to cpu == this_cpu
1573 * task_rq_unlock(rq, &flags);
1574 * this_rq = task_rq_lock(current, &flags);
1578 this_rq = cpu_rq(this_cpu);
1581 * Not the local CPU - must adjust timestamp. This should
1582 * get optimised away in the !CONFIG_SMP case.
1584 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1585 + rq->most_recent_timestamp;
1586 __activate_task(p, rq);
1587 if (TASK_PREEMPTS_CURR(p, rq))
1588 resched_task(rq->curr);
1591 * Parent and child are on different CPUs, now get the
1592 * parent runqueue to update the parent's ->sleep_avg:
1594 task_rq_unlock(rq, &flags);
1595 this_rq = task_rq_lock(current, &flags);
1597 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1598 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1599 task_rq_unlock(this_rq, &flags);
1603 * prepare_task_switch - prepare to switch tasks
1604 * @rq: the runqueue preparing to switch
1605 * @next: the task we are going to switch to.
1607 * This is called with the rq lock held and interrupts off. It must
1608 * be paired with a subsequent finish_task_switch after the context
1611 * prepare_task_switch sets up locking and calls architecture specific
1614 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1616 prepare_lock_switch(rq, next);
1617 prepare_arch_switch(next);
1621 * finish_task_switch - clean up after a task-switch
1622 * @rq: runqueue associated with task-switch
1623 * @prev: the thread we just switched away from.
1625 * finish_task_switch must be called after the context switch, paired
1626 * with a prepare_task_switch call before the context switch.
1627 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1628 * and do any other architecture-specific cleanup actions.
1630 * Note that we may have delayed dropping an mm in context_switch(). If
1631 * so, we finish that here outside of the runqueue lock. (Doing it
1632 * with the lock held can cause deadlocks; see schedule() for
1635 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1636 __releases(rq->lock)
1638 struct mm_struct *mm = rq->prev_mm;
1644 * A task struct has one reference for the use as "current".
1645 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1646 * schedule one last time. The schedule call will never return, and
1647 * the scheduled task must drop that reference.
1648 * The test for TASK_DEAD must occur while the runqueue locks are
1649 * still held, otherwise prev could be scheduled on another cpu, die
1650 * there before we look at prev->state, and then the reference would
1652 * Manfred Spraul <manfred@colorfullife.com>
1654 prev_state = prev->state;
1655 finish_arch_switch(prev);
1656 finish_lock_switch(rq, prev);
1659 if (unlikely(prev_state == TASK_DEAD)) {
1661 * Remove function-return probe instances associated with this
1662 * task and put them back on the free list.
1664 kprobe_flush_task(prev);
1665 put_task_struct(prev);
1670 * schedule_tail - first thing a freshly forked thread must call.
1671 * @prev: the thread we just switched away from.
1673 asmlinkage void schedule_tail(struct task_struct *prev)
1674 __releases(rq->lock)
1676 struct rq *rq = this_rq();
1678 finish_task_switch(rq, prev);
1679 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1680 /* In this case, finish_task_switch does not reenable preemption */
1683 if (current->set_child_tid)
1684 put_user(current->pid, current->set_child_tid);
1688 * context_switch - switch to the new MM and the new
1689 * thread's register state.
1691 static inline struct task_struct *
1692 context_switch(struct rq *rq, struct task_struct *prev,
1693 struct task_struct *next)
1695 struct mm_struct *mm = next->mm;
1696 struct mm_struct *oldmm = prev->active_mm;
1699 * For paravirt, this is coupled with an exit in switch_to to
1700 * combine the page table reload and the switch backend into
1703 arch_enter_lazy_cpu_mode();
1706 next->active_mm = oldmm;
1707 atomic_inc(&oldmm->mm_count);
1708 enter_lazy_tlb(oldmm, next);
1710 switch_mm(oldmm, mm, next);
1713 prev->active_mm = NULL;
1714 WARN_ON(rq->prev_mm);
1715 rq->prev_mm = oldmm;
1718 * Since the runqueue lock will be released by the next
1719 * task (which is an invalid locking op but in the case
1720 * of the scheduler it's an obvious special-case), so we
1721 * do an early lockdep release here:
1723 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1724 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1727 /* Here we just switch the register state and the stack. */
1728 switch_to(prev, next, prev);
1734 * nr_running, nr_uninterruptible and nr_context_switches:
1736 * externally visible scheduler statistics: current number of runnable
1737 * threads, current number of uninterruptible-sleeping threads, total
1738 * number of context switches performed since bootup.
1740 unsigned long nr_running(void)
1742 unsigned long i, sum = 0;
1744 for_each_online_cpu(i)
1745 sum += cpu_rq(i)->nr_running;
1750 unsigned long nr_uninterruptible(void)
1752 unsigned long i, sum = 0;
1754 for_each_possible_cpu(i)
1755 sum += cpu_rq(i)->nr_uninterruptible;
1758 * Since we read the counters lockless, it might be slightly
1759 * inaccurate. Do not allow it to go below zero though:
1761 if (unlikely((long)sum < 0))
1767 unsigned long long nr_context_switches(void)
1770 unsigned long long sum = 0;
1772 for_each_possible_cpu(i)
1773 sum += cpu_rq(i)->nr_switches;
1778 unsigned long nr_iowait(void)
1780 unsigned long i, sum = 0;
1782 for_each_possible_cpu(i)
1783 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1788 unsigned long nr_active(void)
1790 unsigned long i, running = 0, uninterruptible = 0;
1792 for_each_online_cpu(i) {
1793 running += cpu_rq(i)->nr_running;
1794 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1797 if (unlikely((long)uninterruptible < 0))
1798 uninterruptible = 0;
1800 return running + uninterruptible;
1806 * Is this task likely cache-hot:
1809 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1811 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1815 * double_rq_lock - safely lock two runqueues
1817 * Note this does not disable interrupts like task_rq_lock,
1818 * you need to do so manually before calling.
1820 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1821 __acquires(rq1->lock)
1822 __acquires(rq2->lock)
1824 BUG_ON(!irqs_disabled());
1826 spin_lock(&rq1->lock);
1827 __acquire(rq2->lock); /* Fake it out ;) */
1830 spin_lock(&rq1->lock);
1831 spin_lock(&rq2->lock);
1833 spin_lock(&rq2->lock);
1834 spin_lock(&rq1->lock);
1840 * double_rq_unlock - safely unlock two runqueues
1842 * Note this does not restore interrupts like task_rq_unlock,
1843 * you need to do so manually after calling.
1845 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1846 __releases(rq1->lock)
1847 __releases(rq2->lock)
1849 spin_unlock(&rq1->lock);
1851 spin_unlock(&rq2->lock);
1853 __release(rq2->lock);
1857 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1859 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1860 __releases(this_rq->lock)
1861 __acquires(busiest->lock)
1862 __acquires(this_rq->lock)
1864 if (unlikely(!irqs_disabled())) {
1865 /* printk() doesn't work good under rq->lock */
1866 spin_unlock(&this_rq->lock);
1869 if (unlikely(!spin_trylock(&busiest->lock))) {
1870 if (busiest < this_rq) {
1871 spin_unlock(&this_rq->lock);
1872 spin_lock(&busiest->lock);
1873 spin_lock(&this_rq->lock);
1875 spin_lock(&busiest->lock);
1880 * If dest_cpu is allowed for this process, migrate the task to it.
1881 * This is accomplished by forcing the cpu_allowed mask to only
1882 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1883 * the cpu_allowed mask is restored.
1885 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
1887 struct migration_req req;
1888 unsigned long flags;
1891 rq = task_rq_lock(p, &flags);
1892 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1893 || unlikely(cpu_is_offline(dest_cpu)))
1896 /* force the process onto the specified CPU */
1897 if (migrate_task(p, dest_cpu, &req)) {
1898 /* Need to wait for migration thread (might exit: take ref). */
1899 struct task_struct *mt = rq->migration_thread;
1901 get_task_struct(mt);
1902 task_rq_unlock(rq, &flags);
1903 wake_up_process(mt);
1904 put_task_struct(mt);
1905 wait_for_completion(&req.done);
1910 task_rq_unlock(rq, &flags);
1914 * sched_exec - execve() is a valuable balancing opportunity, because at
1915 * this point the task has the smallest effective memory and cache footprint.
1917 void sched_exec(void)
1919 int new_cpu, this_cpu = get_cpu();
1920 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1922 if (new_cpu != this_cpu)
1923 sched_migrate_task(current, new_cpu);
1927 * pull_task - move a task from a remote runqueue to the local runqueue.
1928 * Both runqueues must be locked.
1930 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
1931 struct task_struct *p, struct rq *this_rq,
1932 struct prio_array *this_array, int this_cpu)
1934 dequeue_task(p, src_array);
1935 dec_nr_running(p, src_rq);
1936 set_task_cpu(p, this_cpu);
1937 inc_nr_running(p, this_rq);
1938 enqueue_task(p, this_array);
1939 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
1940 + this_rq->most_recent_timestamp;
1942 * Note that idle threads have a prio of MAX_PRIO, for this test
1943 * to be always true for them.
1945 if (TASK_PREEMPTS_CURR(p, this_rq))
1946 resched_task(this_rq->curr);
1950 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1953 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1954 struct sched_domain *sd, enum cpu_idle_type idle,
1958 * We do not migrate tasks that are:
1959 * 1) running (obviously), or
1960 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1961 * 3) are cache-hot on their current CPU.
1963 if (!cpu_isset(this_cpu, p->cpus_allowed))
1967 if (task_running(rq, p))
1971 * Aggressive migration if:
1972 * 1) task is cache cold, or
1973 * 2) too many balance attempts have failed.
1976 if (sd->nr_balance_failed > sd->cache_nice_tries) {
1977 #ifdef CONFIG_SCHEDSTATS
1978 if (task_hot(p, rq->most_recent_timestamp, sd))
1979 schedstat_inc(sd, lb_hot_gained[idle]);
1984 if (task_hot(p, rq->most_recent_timestamp, sd))
1989 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
1992 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
1993 * load from busiest to this_rq, as part of a balancing operation within
1994 * "domain". Returns the number of tasks moved.
1996 * Called with both runqueues locked.
1998 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1999 unsigned long max_nr_move, unsigned long max_load_move,
2000 struct sched_domain *sd, enum cpu_idle_type idle,
2003 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2004 best_prio_seen, skip_for_load;
2005 struct prio_array *array, *dst_array;
2006 struct list_head *head, *curr;
2007 struct task_struct *tmp;
2010 if (max_nr_move == 0 || max_load_move == 0)
2013 rem_load_move = max_load_move;
2015 this_best_prio = rq_best_prio(this_rq);
2016 best_prio = rq_best_prio(busiest);
2018 * Enable handling of the case where there is more than one task
2019 * with the best priority. If the current running task is one
2020 * of those with prio==best_prio we know it won't be moved
2021 * and therefore it's safe to override the skip (based on load) of
2022 * any task we find with that prio.
2024 best_prio_seen = best_prio == busiest->curr->prio;
2027 * We first consider expired tasks. Those will likely not be
2028 * executed in the near future, and they are most likely to
2029 * be cache-cold, thus switching CPUs has the least effect
2032 if (busiest->expired->nr_active) {
2033 array = busiest->expired;
2034 dst_array = this_rq->expired;
2036 array = busiest->active;
2037 dst_array = this_rq->active;
2041 /* Start searching at priority 0: */
2045 idx = sched_find_first_bit(array->bitmap);
2047 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2048 if (idx >= MAX_PRIO) {
2049 if (array == busiest->expired && busiest->active->nr_active) {
2050 array = busiest->active;
2051 dst_array = this_rq->active;
2057 head = array->queue + idx;
2060 tmp = list_entry(curr, struct task_struct, run_list);
2065 * To help distribute high priority tasks accross CPUs we don't
2066 * skip a task if it will be the highest priority task (i.e. smallest
2067 * prio value) on its new queue regardless of its load weight
2069 skip_for_load = tmp->load_weight > rem_load_move;
2070 if (skip_for_load && idx < this_best_prio)
2071 skip_for_load = !best_prio_seen && idx == best_prio;
2072 if (skip_for_load ||
2073 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2075 best_prio_seen |= idx == best_prio;
2082 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2084 rem_load_move -= tmp->load_weight;
2087 * We only want to steal up to the prescribed number of tasks
2088 * and the prescribed amount of weighted load.
2090 if (pulled < max_nr_move && rem_load_move > 0) {
2091 if (idx < this_best_prio)
2092 this_best_prio = idx;
2100 * Right now, this is the only place pull_task() is called,
2101 * so we can safely collect pull_task() stats here rather than
2102 * inside pull_task().
2104 schedstat_add(sd, lb_gained[idle], pulled);
2107 *all_pinned = pinned;
2112 * find_busiest_group finds and returns the busiest CPU group within the
2113 * domain. It calculates and returns the amount of weighted load which
2114 * should be moved to restore balance via the imbalance parameter.
2116 static struct sched_group *
2117 find_busiest_group(struct sched_domain *sd, int this_cpu,
2118 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2119 cpumask_t *cpus, int *balance)
2121 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2122 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2123 unsigned long max_pull;
2124 unsigned long busiest_load_per_task, busiest_nr_running;
2125 unsigned long this_load_per_task, this_nr_running;
2127 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2128 int power_savings_balance = 1;
2129 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2130 unsigned long min_nr_running = ULONG_MAX;
2131 struct sched_group *group_min = NULL, *group_leader = NULL;
2134 max_load = this_load = total_load = total_pwr = 0;
2135 busiest_load_per_task = busiest_nr_running = 0;
2136 this_load_per_task = this_nr_running = 0;
2137 if (idle == CPU_NOT_IDLE)
2138 load_idx = sd->busy_idx;
2139 else if (idle == CPU_NEWLY_IDLE)
2140 load_idx = sd->newidle_idx;
2142 load_idx = sd->idle_idx;
2145 unsigned long load, group_capacity;
2148 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2149 unsigned long sum_nr_running, sum_weighted_load;
2151 local_group = cpu_isset(this_cpu, group->cpumask);
2154 balance_cpu = first_cpu(group->cpumask);
2156 /* Tally up the load of all CPUs in the group */
2157 sum_weighted_load = sum_nr_running = avg_load = 0;
2159 for_each_cpu_mask(i, group->cpumask) {
2162 if (!cpu_isset(i, *cpus))
2167 if (*sd_idle && !idle_cpu(i))
2170 /* Bias balancing toward cpus of our domain */
2172 if (idle_cpu(i) && !first_idle_cpu) {
2177 load = target_load(i, load_idx);
2179 load = source_load(i, load_idx);
2182 sum_nr_running += rq->nr_running;
2183 sum_weighted_load += rq->raw_weighted_load;
2187 * First idle cpu or the first cpu(busiest) in this sched group
2188 * is eligible for doing load balancing at this and above
2191 if (local_group && balance_cpu != this_cpu && balance) {
2196 total_load += avg_load;
2197 total_pwr += group->__cpu_power;
2199 /* Adjust by relative CPU power of the group */
2200 avg_load = sg_div_cpu_power(group,
2201 avg_load * SCHED_LOAD_SCALE);
2203 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2206 this_load = avg_load;
2208 this_nr_running = sum_nr_running;
2209 this_load_per_task = sum_weighted_load;
2210 } else if (avg_load > max_load &&
2211 sum_nr_running > group_capacity) {
2212 max_load = avg_load;
2214 busiest_nr_running = sum_nr_running;
2215 busiest_load_per_task = sum_weighted_load;
2218 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2220 * Busy processors will not participate in power savings
2223 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2227 * If the local group is idle or completely loaded
2228 * no need to do power savings balance at this domain
2230 if (local_group && (this_nr_running >= group_capacity ||
2232 power_savings_balance = 0;
2235 * If a group is already running at full capacity or idle,
2236 * don't include that group in power savings calculations
2238 if (!power_savings_balance || sum_nr_running >= group_capacity
2243 * Calculate the group which has the least non-idle load.
2244 * This is the group from where we need to pick up the load
2247 if ((sum_nr_running < min_nr_running) ||
2248 (sum_nr_running == min_nr_running &&
2249 first_cpu(group->cpumask) <
2250 first_cpu(group_min->cpumask))) {
2252 min_nr_running = sum_nr_running;
2253 min_load_per_task = sum_weighted_load /
2258 * Calculate the group which is almost near its
2259 * capacity but still has some space to pick up some load
2260 * from other group and save more power
2262 if (sum_nr_running <= group_capacity - 1) {
2263 if (sum_nr_running > leader_nr_running ||
2264 (sum_nr_running == leader_nr_running &&
2265 first_cpu(group->cpumask) >
2266 first_cpu(group_leader->cpumask))) {
2267 group_leader = group;
2268 leader_nr_running = sum_nr_running;
2273 group = group->next;
2274 } while (group != sd->groups);
2276 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2279 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2281 if (this_load >= avg_load ||
2282 100*max_load <= sd->imbalance_pct*this_load)
2285 busiest_load_per_task /= busiest_nr_running;
2287 * We're trying to get all the cpus to the average_load, so we don't
2288 * want to push ourselves above the average load, nor do we wish to
2289 * reduce the max loaded cpu below the average load, as either of these
2290 * actions would just result in more rebalancing later, and ping-pong
2291 * tasks around. Thus we look for the minimum possible imbalance.
2292 * Negative imbalances (*we* are more loaded than anyone else) will
2293 * be counted as no imbalance for these purposes -- we can't fix that
2294 * by pulling tasks to us. Be careful of negative numbers as they'll
2295 * appear as very large values with unsigned longs.
2297 if (max_load <= busiest_load_per_task)
2301 * In the presence of smp nice balancing, certain scenarios can have
2302 * max load less than avg load(as we skip the groups at or below
2303 * its cpu_power, while calculating max_load..)
2305 if (max_load < avg_load) {
2307 goto small_imbalance;
2310 /* Don't want to pull so many tasks that a group would go idle */
2311 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2313 /* How much load to actually move to equalise the imbalance */
2314 *imbalance = min(max_pull * busiest->__cpu_power,
2315 (avg_load - this_load) * this->__cpu_power)
2319 * if *imbalance is less than the average load per runnable task
2320 * there is no gaurantee that any tasks will be moved so we'll have
2321 * a think about bumping its value to force at least one task to be
2324 if (*imbalance < busiest_load_per_task) {
2325 unsigned long tmp, pwr_now, pwr_move;
2329 pwr_move = pwr_now = 0;
2331 if (this_nr_running) {
2332 this_load_per_task /= this_nr_running;
2333 if (busiest_load_per_task > this_load_per_task)
2336 this_load_per_task = SCHED_LOAD_SCALE;
2338 if (max_load - this_load >= busiest_load_per_task * imbn) {
2339 *imbalance = busiest_load_per_task;
2344 * OK, we don't have enough imbalance to justify moving tasks,
2345 * however we may be able to increase total CPU power used by
2349 pwr_now += busiest->__cpu_power *
2350 min(busiest_load_per_task, max_load);
2351 pwr_now += this->__cpu_power *
2352 min(this_load_per_task, this_load);
2353 pwr_now /= SCHED_LOAD_SCALE;
2355 /* Amount of load we'd subtract */
2356 tmp = sg_div_cpu_power(busiest,
2357 busiest_load_per_task * SCHED_LOAD_SCALE);
2359 pwr_move += busiest->__cpu_power *
2360 min(busiest_load_per_task, max_load - tmp);
2362 /* Amount of load we'd add */
2363 if (max_load * busiest->__cpu_power <
2364 busiest_load_per_task * SCHED_LOAD_SCALE)
2365 tmp = sg_div_cpu_power(this,
2366 max_load * busiest->__cpu_power);
2368 tmp = sg_div_cpu_power(this,
2369 busiest_load_per_task * SCHED_LOAD_SCALE);
2370 pwr_move += this->__cpu_power *
2371 min(this_load_per_task, this_load + tmp);
2372 pwr_move /= SCHED_LOAD_SCALE;
2374 /* Move if we gain throughput */
2375 if (pwr_move <= pwr_now)
2378 *imbalance = busiest_load_per_task;
2384 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2385 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2388 if (this == group_leader && group_leader != group_min) {
2389 *imbalance = min_load_per_task;
2399 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2402 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2403 unsigned long imbalance, cpumask_t *cpus)
2405 struct rq *busiest = NULL, *rq;
2406 unsigned long max_load = 0;
2409 for_each_cpu_mask(i, group->cpumask) {
2411 if (!cpu_isset(i, *cpus))
2416 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2419 if (rq->raw_weighted_load > max_load) {
2420 max_load = rq->raw_weighted_load;
2429 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2430 * so long as it is large enough.
2432 #define MAX_PINNED_INTERVAL 512
2434 static inline unsigned long minus_1_or_zero(unsigned long n)
2436 return n > 0 ? n - 1 : 0;
2440 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2441 * tasks if there is an imbalance.
2443 static int load_balance(int this_cpu, struct rq *this_rq,
2444 struct sched_domain *sd, enum cpu_idle_type idle,
2447 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2448 struct sched_group *group;
2449 unsigned long imbalance;
2451 cpumask_t cpus = CPU_MASK_ALL;
2452 unsigned long flags;
2455 * When power savings policy is enabled for the parent domain, idle
2456 * sibling can pick up load irrespective of busy siblings. In this case,
2457 * let the state of idle sibling percolate up as IDLE, instead of
2458 * portraying it as CPU_NOT_IDLE.
2460 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2461 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2464 schedstat_inc(sd, lb_cnt[idle]);
2467 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2474 schedstat_inc(sd, lb_nobusyg[idle]);
2478 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2480 schedstat_inc(sd, lb_nobusyq[idle]);
2484 BUG_ON(busiest == this_rq);
2486 schedstat_add(sd, lb_imbalance[idle], imbalance);
2489 if (busiest->nr_running > 1) {
2491 * Attempt to move tasks. If find_busiest_group has found
2492 * an imbalance but busiest->nr_running <= 1, the group is
2493 * still unbalanced. nr_moved simply stays zero, so it is
2494 * correctly treated as an imbalance.
2496 local_irq_save(flags);
2497 double_rq_lock(this_rq, busiest);
2498 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2499 minus_1_or_zero(busiest->nr_running),
2500 imbalance, sd, idle, &all_pinned);
2501 double_rq_unlock(this_rq, busiest);
2502 local_irq_restore(flags);
2505 * some other cpu did the load balance for us.
2507 if (nr_moved && this_cpu != smp_processor_id())
2508 resched_cpu(this_cpu);
2510 /* All tasks on this runqueue were pinned by CPU affinity */
2511 if (unlikely(all_pinned)) {
2512 cpu_clear(cpu_of(busiest), cpus);
2513 if (!cpus_empty(cpus))
2520 schedstat_inc(sd, lb_failed[idle]);
2521 sd->nr_balance_failed++;
2523 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2525 spin_lock_irqsave(&busiest->lock, flags);
2527 /* don't kick the migration_thread, if the curr
2528 * task on busiest cpu can't be moved to this_cpu
2530 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2531 spin_unlock_irqrestore(&busiest->lock, flags);
2533 goto out_one_pinned;
2536 if (!busiest->active_balance) {
2537 busiest->active_balance = 1;
2538 busiest->push_cpu = this_cpu;
2541 spin_unlock_irqrestore(&busiest->lock, flags);
2543 wake_up_process(busiest->migration_thread);
2546 * We've kicked active balancing, reset the failure
2549 sd->nr_balance_failed = sd->cache_nice_tries+1;
2552 sd->nr_balance_failed = 0;
2554 if (likely(!active_balance)) {
2555 /* We were unbalanced, so reset the balancing interval */
2556 sd->balance_interval = sd->min_interval;
2559 * If we've begun active balancing, start to back off. This
2560 * case may not be covered by the all_pinned logic if there
2561 * is only 1 task on the busy runqueue (because we don't call
2564 if (sd->balance_interval < sd->max_interval)
2565 sd->balance_interval *= 2;
2568 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2569 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2574 schedstat_inc(sd, lb_balanced[idle]);
2576 sd->nr_balance_failed = 0;
2579 /* tune up the balancing interval */
2580 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2581 (sd->balance_interval < sd->max_interval))
2582 sd->balance_interval *= 2;
2584 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2585 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2591 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2592 * tasks if there is an imbalance.
2594 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2595 * this_rq is locked.
2598 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2600 struct sched_group *group;
2601 struct rq *busiest = NULL;
2602 unsigned long imbalance;
2605 cpumask_t cpus = CPU_MASK_ALL;
2608 * When power savings policy is enabled for the parent domain, idle
2609 * sibling can pick up load irrespective of busy siblings. In this case,
2610 * let the state of idle sibling percolate up as IDLE, instead of
2611 * portraying it as CPU_NOT_IDLE.
2613 if (sd->flags & SD_SHARE_CPUPOWER &&
2614 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2617 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2619 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2620 &sd_idle, &cpus, NULL);
2622 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2626 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2629 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2633 BUG_ON(busiest == this_rq);
2635 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2638 if (busiest->nr_running > 1) {
2639 /* Attempt to move tasks */
2640 double_lock_balance(this_rq, busiest);
2641 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2642 minus_1_or_zero(busiest->nr_running),
2643 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2644 spin_unlock(&busiest->lock);
2647 cpu_clear(cpu_of(busiest), cpus);
2648 if (!cpus_empty(cpus))
2654 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2655 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2656 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2659 sd->nr_balance_failed = 0;
2664 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2665 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2666 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2668 sd->nr_balance_failed = 0;
2674 * idle_balance is called by schedule() if this_cpu is about to become
2675 * idle. Attempts to pull tasks from other CPUs.
2677 static void idle_balance(int this_cpu, struct rq *this_rq)
2679 struct sched_domain *sd;
2680 int pulled_task = 0;
2681 unsigned long next_balance = jiffies + 60 * HZ;
2683 for_each_domain(this_cpu, sd) {
2684 unsigned long interval;
2686 if (!(sd->flags & SD_LOAD_BALANCE))
2689 if (sd->flags & SD_BALANCE_NEWIDLE)
2690 /* If we've pulled tasks over stop searching: */
2691 pulled_task = load_balance_newidle(this_cpu,
2694 interval = msecs_to_jiffies(sd->balance_interval);
2695 if (time_after(next_balance, sd->last_balance + interval))
2696 next_balance = sd->last_balance + interval;
2702 * We are going idle. next_balance may be set based on
2703 * a busy processor. So reset next_balance.
2705 this_rq->next_balance = next_balance;
2709 * active_load_balance is run by migration threads. It pushes running tasks
2710 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2711 * running on each physical CPU where possible, and avoids physical /
2712 * logical imbalances.
2714 * Called with busiest_rq locked.
2716 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2718 int target_cpu = busiest_rq->push_cpu;
2719 struct sched_domain *sd;
2720 struct rq *target_rq;
2722 /* Is there any task to move? */
2723 if (busiest_rq->nr_running <= 1)
2726 target_rq = cpu_rq(target_cpu);
2729 * This condition is "impossible", if it occurs
2730 * we need to fix it. Originally reported by
2731 * Bjorn Helgaas on a 128-cpu setup.
2733 BUG_ON(busiest_rq == target_rq);
2735 /* move a task from busiest_rq to target_rq */
2736 double_lock_balance(busiest_rq, target_rq);
2738 /* Search for an sd spanning us and the target CPU. */
2739 for_each_domain(target_cpu, sd) {
2740 if ((sd->flags & SD_LOAD_BALANCE) &&
2741 cpu_isset(busiest_cpu, sd->span))
2746 schedstat_inc(sd, alb_cnt);
2748 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2749 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2751 schedstat_inc(sd, alb_pushed);
2753 schedstat_inc(sd, alb_failed);
2755 spin_unlock(&target_rq->lock);
2758 static void update_load(struct rq *this_rq)
2760 unsigned long this_load;
2761 unsigned int i, scale;
2763 this_load = this_rq->raw_weighted_load;
2765 /* Update our load: */
2766 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2767 unsigned long old_load, new_load;
2769 /* scale is effectively 1 << i now, and >> i divides by scale */
2771 old_load = this_rq->cpu_load[i];
2772 new_load = this_load;
2774 * Round up the averaging division if load is increasing. This
2775 * prevents us from getting stuck on 9 if the load is 10, for
2778 if (new_load > old_load)
2779 new_load += scale-1;
2780 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2786 atomic_t load_balancer;
2788 } nohz ____cacheline_aligned = {
2789 .load_balancer = ATOMIC_INIT(-1),
2790 .cpu_mask = CPU_MASK_NONE,
2794 * This routine will try to nominate the ilb (idle load balancing)
2795 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2796 * load balancing on behalf of all those cpus. If all the cpus in the system
2797 * go into this tickless mode, then there will be no ilb owner (as there is
2798 * no need for one) and all the cpus will sleep till the next wakeup event
2801 * For the ilb owner, tick is not stopped. And this tick will be used
2802 * for idle load balancing. ilb owner will still be part of
2805 * While stopping the tick, this cpu will become the ilb owner if there
2806 * is no other owner. And will be the owner till that cpu becomes busy
2807 * or if all cpus in the system stop their ticks at which point
2808 * there is no need for ilb owner.
2810 * When the ilb owner becomes busy, it nominates another owner, during the
2811 * next busy scheduler_tick()
2813 int select_nohz_load_balancer(int stop_tick)
2815 int cpu = smp_processor_id();
2818 cpu_set(cpu, nohz.cpu_mask);
2819 cpu_rq(cpu)->in_nohz_recently = 1;
2822 * If we are going offline and still the leader, give up!
2824 if (cpu_is_offline(cpu) &&
2825 atomic_read(&nohz.load_balancer) == cpu) {
2826 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2831 /* time for ilb owner also to sleep */
2832 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2833 if (atomic_read(&nohz.load_balancer) == cpu)
2834 atomic_set(&nohz.load_balancer, -1);
2838 if (atomic_read(&nohz.load_balancer) == -1) {
2839 /* make me the ilb owner */
2840 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2842 } else if (atomic_read(&nohz.load_balancer) == cpu)
2845 if (!cpu_isset(cpu, nohz.cpu_mask))
2848 cpu_clear(cpu, nohz.cpu_mask);
2850 if (atomic_read(&nohz.load_balancer) == cpu)
2851 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2858 static DEFINE_SPINLOCK(balancing);
2861 * It checks each scheduling domain to see if it is due to be balanced,
2862 * and initiates a balancing operation if so.
2864 * Balancing parameters are set up in arch_init_sched_domains.
2866 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2869 struct rq *rq = cpu_rq(cpu);
2870 unsigned long interval;
2871 struct sched_domain *sd;
2872 /* Earliest time when we have to do rebalance again */
2873 unsigned long next_balance = jiffies + 60*HZ;
2875 for_each_domain(cpu, sd) {
2876 if (!(sd->flags & SD_LOAD_BALANCE))
2879 interval = sd->balance_interval;
2880 if (idle != CPU_IDLE)
2881 interval *= sd->busy_factor;
2883 /* scale ms to jiffies */
2884 interval = msecs_to_jiffies(interval);
2885 if (unlikely(!interval))
2888 if (sd->flags & SD_SERIALIZE) {
2889 if (!spin_trylock(&balancing))
2893 if (time_after_eq(jiffies, sd->last_balance + interval)) {
2894 if (load_balance(cpu, rq, sd, idle, &balance)) {
2896 * We've pulled tasks over so either we're no
2897 * longer idle, or one of our SMT siblings is
2900 idle = CPU_NOT_IDLE;
2902 sd->last_balance = jiffies;
2904 if (sd->flags & SD_SERIALIZE)
2905 spin_unlock(&balancing);
2907 if (time_after(next_balance, sd->last_balance + interval))
2908 next_balance = sd->last_balance + interval;
2911 * Stop the load balance at this level. There is another
2912 * CPU in our sched group which is doing load balancing more
2918 rq->next_balance = next_balance;
2922 * run_rebalance_domains is triggered when needed from the scheduler tick.
2923 * In CONFIG_NO_HZ case, the idle load balance owner will do the
2924 * rebalancing for all the cpus for whom scheduler ticks are stopped.
2926 static void run_rebalance_domains(struct softirq_action *h)
2928 int local_cpu = smp_processor_id();
2929 struct rq *local_rq = cpu_rq(local_cpu);
2930 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
2932 rebalance_domains(local_cpu, idle);
2936 * If this cpu is the owner for idle load balancing, then do the
2937 * balancing on behalf of the other idle cpus whose ticks are
2940 if (local_rq->idle_at_tick &&
2941 atomic_read(&nohz.load_balancer) == local_cpu) {
2942 cpumask_t cpus = nohz.cpu_mask;
2946 cpu_clear(local_cpu, cpus);
2947 for_each_cpu_mask(balance_cpu, cpus) {
2949 * If this cpu gets work to do, stop the load balancing
2950 * work being done for other cpus. Next load
2951 * balancing owner will pick it up.
2956 rebalance_domains(balance_cpu, CPU_IDLE);
2958 rq = cpu_rq(balance_cpu);
2959 if (time_after(local_rq->next_balance, rq->next_balance))
2960 local_rq->next_balance = rq->next_balance;
2967 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
2969 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
2970 * idle load balancing owner or decide to stop the periodic load balancing,
2971 * if the whole system is idle.
2973 static inline void trigger_load_balance(int cpu)
2975 struct rq *rq = cpu_rq(cpu);
2978 * If we were in the nohz mode recently and busy at the current
2979 * scheduler tick, then check if we need to nominate new idle
2982 if (rq->in_nohz_recently && !rq->idle_at_tick) {
2983 rq->in_nohz_recently = 0;
2985 if (atomic_read(&nohz.load_balancer) == cpu) {
2986 cpu_clear(cpu, nohz.cpu_mask);
2987 atomic_set(&nohz.load_balancer, -1);
2990 if (atomic_read(&nohz.load_balancer) == -1) {
2992 * simple selection for now: Nominate the
2993 * first cpu in the nohz list to be the next
2996 * TBD: Traverse the sched domains and nominate
2997 * the nearest cpu in the nohz.cpu_mask.
2999 int ilb = first_cpu(nohz.cpu_mask);
3007 * If this cpu is idle and doing idle load balancing for all the
3008 * cpus with ticks stopped, is it time for that to stop?
3010 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3011 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3017 * If this cpu is idle and the idle load balancing is done by
3018 * someone else, then no need raise the SCHED_SOFTIRQ
3020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3021 cpu_isset(cpu, nohz.cpu_mask))
3024 if (time_after_eq(jiffies, rq->next_balance))
3025 raise_softirq(SCHED_SOFTIRQ);
3029 * on UP we do not need to balance between CPUs:
3031 static inline void idle_balance(int cpu, struct rq *rq)
3036 DEFINE_PER_CPU(struct kernel_stat, kstat);
3038 EXPORT_PER_CPU_SYMBOL(kstat);
3041 * This is called on clock ticks and on context switches.
3042 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3045 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3047 p->sched_time += now - p->last_ran;
3048 p->last_ran = rq->most_recent_timestamp = now;
3052 * Return current->sched_time plus any more ns on the sched_clock
3053 * that have not yet been banked.
3055 unsigned long long current_sched_time(const struct task_struct *p)
3057 unsigned long long ns;
3058 unsigned long flags;
3060 local_irq_save(flags);
3061 ns = p->sched_time + sched_clock() - p->last_ran;
3062 local_irq_restore(flags);
3068 * We place interactive tasks back into the active array, if possible.
3070 * To guarantee that this does not starve expired tasks we ignore the
3071 * interactivity of a task if the first expired task had to wait more
3072 * than a 'reasonable' amount of time. This deadline timeout is
3073 * load-dependent, as the frequency of array switched decreases with
3074 * increasing number of running tasks. We also ignore the interactivity
3075 * if a better static_prio task has expired:
3077 static inline int expired_starving(struct rq *rq)
3079 if (rq->curr->static_prio > rq->best_expired_prio)
3081 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3083 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3089 * Account user cpu time to a process.
3090 * @p: the process that the cpu time gets accounted to
3091 * @hardirq_offset: the offset to subtract from hardirq_count()
3092 * @cputime: the cpu time spent in user space since the last update
3094 void account_user_time(struct task_struct *p, cputime_t cputime)
3096 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3099 p->utime = cputime_add(p->utime, cputime);
3101 /* Add user time to cpustat. */
3102 tmp = cputime_to_cputime64(cputime);
3103 if (TASK_NICE(p) > 0)
3104 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3106 cpustat->user = cputime64_add(cpustat->user, tmp);
3110 * Account system cpu time to a process.
3111 * @p: the process that the cpu time gets accounted to
3112 * @hardirq_offset: the offset to subtract from hardirq_count()
3113 * @cputime: the cpu time spent in kernel space since the last update
3115 void account_system_time(struct task_struct *p, int hardirq_offset,
3118 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3119 struct rq *rq = this_rq();
3122 p->stime = cputime_add(p->stime, cputime);
3124 /* Add system time to cpustat. */
3125 tmp = cputime_to_cputime64(cputime);
3126 if (hardirq_count() - hardirq_offset)
3127 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3128 else if (softirq_count())
3129 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3130 else if (p != rq->idle)
3131 cpustat->system = cputime64_add(cpustat->system, tmp);
3132 else if (atomic_read(&rq->nr_iowait) > 0)
3133 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3135 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3136 /* Account for system time used */
3137 acct_update_integrals(p);
3141 * Account for involuntary wait time.
3142 * @p: the process from which the cpu time has been stolen
3143 * @steal: the cpu time spent in involuntary wait
3145 void account_steal_time(struct task_struct *p, cputime_t steal)
3147 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3148 cputime64_t tmp = cputime_to_cputime64(steal);
3149 struct rq *rq = this_rq();
3151 if (p == rq->idle) {
3152 p->stime = cputime_add(p->stime, steal);
3153 if (atomic_read(&rq->nr_iowait) > 0)
3154 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3156 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3158 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3161 static void task_running_tick(struct rq *rq, struct task_struct *p)
3163 if (p->array != rq->active) {
3164 /* Task has expired but was not scheduled yet */
3165 set_tsk_need_resched(p);
3168 spin_lock(&rq->lock);
3170 * The task was running during this tick - update the
3171 * time slice counter. Note: we do not update a thread's
3172 * priority until it either goes to sleep or uses up its
3173 * timeslice. This makes it possible for interactive tasks
3174 * to use up their timeslices at their highest priority levels.
3178 * RR tasks need a special form of timeslice management.
3179 * FIFO tasks have no timeslices.
3181 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3182 p->time_slice = task_timeslice(p);
3183 p->first_time_slice = 0;
3184 set_tsk_need_resched(p);
3186 /* put it at the end of the queue: */
3187 requeue_task(p, rq->active);
3191 if (!--p->time_slice) {
3192 dequeue_task(p, rq->active);
3193 set_tsk_need_resched(p);
3194 p->prio = effective_prio(p);
3195 p->time_slice = task_timeslice(p);
3196 p->first_time_slice = 0;
3198 if (!rq->expired_timestamp)
3199 rq->expired_timestamp = jiffies;
3200 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3201 enqueue_task(p, rq->expired);
3202 if (p->static_prio < rq->best_expired_prio)
3203 rq->best_expired_prio = p->static_prio;
3205 enqueue_task(p, rq->active);
3208 * Prevent a too long timeslice allowing a task to monopolize
3209 * the CPU. We do this by splitting up the timeslice into
3212 * Note: this does not mean the task's timeslices expire or
3213 * get lost in any way, they just might be preempted by
3214 * another task of equal priority. (one with higher
3215 * priority would have preempted this task already.) We
3216 * requeue this task to the end of the list on this priority
3217 * level, which is in essence a round-robin of tasks with
3220 * This only applies to tasks in the interactive
3221 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3223 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3224 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3225 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3226 (p->array == rq->active)) {
3228 requeue_task(p, rq->active);
3229 set_tsk_need_resched(p);
3233 spin_unlock(&rq->lock);
3237 * This function gets called by the timer code, with HZ frequency.
3238 * We call it with interrupts disabled.
3240 * It also gets called by the fork code, when changing the parent's
3243 void scheduler_tick(void)
3245 unsigned long long now = sched_clock();
3246 struct task_struct *p = current;
3247 int cpu = smp_processor_id();
3248 int idle_at_tick = idle_cpu(cpu);
3249 struct rq *rq = cpu_rq(cpu);
3251 update_cpu_clock(p, rq, now);
3254 task_running_tick(rq, p);
3257 rq->idle_at_tick = idle_at_tick;
3258 trigger_load_balance(cpu);
3262 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3264 void fastcall add_preempt_count(int val)
3269 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3271 preempt_count() += val;
3273 * Spinlock count overflowing soon?
3275 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3278 EXPORT_SYMBOL(add_preempt_count);
3280 void fastcall sub_preempt_count(int val)
3285 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3288 * Is the spinlock portion underflowing?
3290 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3291 !(preempt_count() & PREEMPT_MASK)))
3294 preempt_count() -= val;
3296 EXPORT_SYMBOL(sub_preempt_count);
3300 static inline int interactive_sleep(enum sleep_type sleep_type)
3302 return (sleep_type == SLEEP_INTERACTIVE ||
3303 sleep_type == SLEEP_INTERRUPTED);
3307 * schedule() is the main scheduler function.
3309 asmlinkage void __sched schedule(void)
3311 struct task_struct *prev, *next;
3312 struct prio_array *array;
3313 struct list_head *queue;
3314 unsigned long long now;
3315 unsigned long run_time;
3316 int cpu, idx, new_prio;
3321 * Test if we are atomic. Since do_exit() needs to call into
3322 * schedule() atomically, we ignore that path for now.
3323 * Otherwise, whine if we are scheduling when we should not be.
3325 if (unlikely(in_atomic() && !current->exit_state)) {
3326 printk(KERN_ERR "BUG: scheduling while atomic: "
3328 current->comm, preempt_count(), current->pid);
3329 debug_show_held_locks(current);
3330 if (irqs_disabled())
3331 print_irqtrace_events(current);
3334 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3339 release_kernel_lock(prev);
3340 need_resched_nonpreemptible:
3344 * The idle thread is not allowed to schedule!
3345 * Remove this check after it has been exercised a bit.
3347 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3348 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3352 schedstat_inc(rq, sched_cnt);
3353 now = sched_clock();
3354 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3355 run_time = now - prev->timestamp;
3356 if (unlikely((long long)(now - prev->timestamp) < 0))
3359 run_time = NS_MAX_SLEEP_AVG;
3362 * Tasks charged proportionately less run_time at high sleep_avg to
3363 * delay them losing their interactive status
3365 run_time /= (CURRENT_BONUS(prev) ? : 1);
3367 spin_lock_irq(&rq->lock);
3369 switch_count = &prev->nivcsw;
3370 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3371 switch_count = &prev->nvcsw;
3372 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3373 unlikely(signal_pending(prev))))
3374 prev->state = TASK_RUNNING;
3376 if (prev->state == TASK_UNINTERRUPTIBLE)
3377 rq->nr_uninterruptible++;
3378 deactivate_task(prev, rq);
3382 cpu = smp_processor_id();
3383 if (unlikely(!rq->nr_running)) {
3384 idle_balance(cpu, rq);
3385 if (!rq->nr_running) {
3387 rq->expired_timestamp = 0;
3393 if (unlikely(!array->nr_active)) {
3395 * Switch the active and expired arrays.
3397 schedstat_inc(rq, sched_switch);
3398 rq->active = rq->expired;
3399 rq->expired = array;
3401 rq->expired_timestamp = 0;
3402 rq->best_expired_prio = MAX_PRIO;
3405 idx = sched_find_first_bit(array->bitmap);
3406 queue = array->queue + idx;
3407 next = list_entry(queue->next, struct task_struct, run_list);
3409 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3410 unsigned long long delta = now - next->timestamp;
3411 if (unlikely((long long)(now - next->timestamp) < 0))
3414 if (next->sleep_type == SLEEP_INTERACTIVE)
3415 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3417 array = next->array;
3418 new_prio = recalc_task_prio(next, next->timestamp + delta);
3420 if (unlikely(next->prio != new_prio)) {
3421 dequeue_task(next, array);
3422 next->prio = new_prio;
3423 enqueue_task(next, array);
3426 next->sleep_type = SLEEP_NORMAL;
3428 if (next == rq->idle)
3429 schedstat_inc(rq, sched_goidle);
3431 prefetch_stack(next);
3432 clear_tsk_need_resched(prev);
3433 rcu_qsctr_inc(task_cpu(prev));
3435 update_cpu_clock(prev, rq, now);
3437 prev->sleep_avg -= run_time;
3438 if ((long)prev->sleep_avg <= 0)
3439 prev->sleep_avg = 0;
3440 prev->timestamp = prev->last_ran = now;
3442 sched_info_switch(prev, next);
3443 if (likely(prev != next)) {
3444 next->timestamp = next->last_ran = now;
3449 prepare_task_switch(rq, next);
3450 prev = context_switch(rq, prev, next);
3453 * this_rq must be evaluated again because prev may have moved
3454 * CPUs since it called schedule(), thus the 'rq' on its stack
3455 * frame will be invalid.
3457 finish_task_switch(this_rq(), prev);
3459 spin_unlock_irq(&rq->lock);
3462 if (unlikely(reacquire_kernel_lock(prev) < 0))
3463 goto need_resched_nonpreemptible;
3464 preempt_enable_no_resched();
3465 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3468 EXPORT_SYMBOL(schedule);
3470 #ifdef CONFIG_PREEMPT
3472 * this is the entry point to schedule() from in-kernel preemption
3473 * off of preempt_enable. Kernel preemptions off return from interrupt
3474 * occur there and call schedule directly.
3476 asmlinkage void __sched preempt_schedule(void)
3478 struct thread_info *ti = current_thread_info();
3479 #ifdef CONFIG_PREEMPT_BKL
3480 struct task_struct *task = current;
3481 int saved_lock_depth;
3484 * If there is a non-zero preempt_count or interrupts are disabled,
3485 * we do not want to preempt the current task. Just return..
3487 if (likely(ti->preempt_count || irqs_disabled()))
3491 add_preempt_count(PREEMPT_ACTIVE);
3493 * We keep the big kernel semaphore locked, but we
3494 * clear ->lock_depth so that schedule() doesnt
3495 * auto-release the semaphore:
3497 #ifdef CONFIG_PREEMPT_BKL
3498 saved_lock_depth = task->lock_depth;
3499 task->lock_depth = -1;
3502 #ifdef CONFIG_PREEMPT_BKL
3503 task->lock_depth = saved_lock_depth;
3505 sub_preempt_count(PREEMPT_ACTIVE);
3507 /* we could miss a preemption opportunity between schedule and now */
3509 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3512 EXPORT_SYMBOL(preempt_schedule);
3515 * this is the entry point to schedule() from kernel preemption
3516 * off of irq context.
3517 * Note, that this is called and return with irqs disabled. This will
3518 * protect us against recursive calling from irq.
3520 asmlinkage void __sched preempt_schedule_irq(void)
3522 struct thread_info *ti = current_thread_info();
3523 #ifdef CONFIG_PREEMPT_BKL
3524 struct task_struct *task = current;
3525 int saved_lock_depth;
3527 /* Catch callers which need to be fixed */
3528 BUG_ON(ti->preempt_count || !irqs_disabled());
3531 add_preempt_count(PREEMPT_ACTIVE);
3533 * We keep the big kernel semaphore locked, but we
3534 * clear ->lock_depth so that schedule() doesnt
3535 * auto-release the semaphore:
3537 #ifdef CONFIG_PREEMPT_BKL
3538 saved_lock_depth = task->lock_depth;
3539 task->lock_depth = -1;
3543 local_irq_disable();
3544 #ifdef CONFIG_PREEMPT_BKL
3545 task->lock_depth = saved_lock_depth;
3547 sub_preempt_count(PREEMPT_ACTIVE);
3549 /* we could miss a preemption opportunity between schedule and now */
3551 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3555 #endif /* CONFIG_PREEMPT */
3557 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3560 return try_to_wake_up(curr->private, mode, sync);
3562 EXPORT_SYMBOL(default_wake_function);
3565 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3566 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3567 * number) then we wake all the non-exclusive tasks and one exclusive task.
3569 * There are circumstances in which we can try to wake a task which has already
3570 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3571 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3573 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3574 int nr_exclusive, int sync, void *key)
3576 struct list_head *tmp, *next;
3578 list_for_each_safe(tmp, next, &q->task_list) {
3579 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3580 unsigned flags = curr->flags;
3582 if (curr->func(curr, mode, sync, key) &&
3583 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3589 * __wake_up - wake up threads blocked on a waitqueue.
3591 * @mode: which threads
3592 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3593 * @key: is directly passed to the wakeup function
3595 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3596 int nr_exclusive, void *key)
3598 unsigned long flags;
3600 spin_lock_irqsave(&q->lock, flags);
3601 __wake_up_common(q, mode, nr_exclusive, 0, key);
3602 spin_unlock_irqrestore(&q->lock, flags);
3604 EXPORT_SYMBOL(__wake_up);
3607 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3609 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3611 __wake_up_common(q, mode, 1, 0, NULL);
3615 * __wake_up_sync - wake up threads blocked on a waitqueue.
3617 * @mode: which threads
3618 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3620 * The sync wakeup differs that the waker knows that it will schedule
3621 * away soon, so while the target thread will be woken up, it will not
3622 * be migrated to another CPU - ie. the two threads are 'synchronized'
3623 * with each other. This can prevent needless bouncing between CPUs.
3625 * On UP it can prevent extra preemption.
3628 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3630 unsigned long flags;
3636 if (unlikely(!nr_exclusive))
3639 spin_lock_irqsave(&q->lock, flags);
3640 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3641 spin_unlock_irqrestore(&q->lock, flags);
3643 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3645 void fastcall complete(struct completion *x)
3647 unsigned long flags;
3649 spin_lock_irqsave(&x->wait.lock, flags);
3651 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3653 spin_unlock_irqrestore(&x->wait.lock, flags);
3655 EXPORT_SYMBOL(complete);
3657 void fastcall complete_all(struct completion *x)
3659 unsigned long flags;
3661 spin_lock_irqsave(&x->wait.lock, flags);
3662 x->done += UINT_MAX/2;
3663 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3665 spin_unlock_irqrestore(&x->wait.lock, flags);
3667 EXPORT_SYMBOL(complete_all);
3669 void fastcall __sched wait_for_completion(struct completion *x)
3673 spin_lock_irq(&x->wait.lock);
3675 DECLARE_WAITQUEUE(wait, current);
3677 wait.flags |= WQ_FLAG_EXCLUSIVE;
3678 __add_wait_queue_tail(&x->wait, &wait);
3680 __set_current_state(TASK_UNINTERRUPTIBLE);
3681 spin_unlock_irq(&x->wait.lock);
3683 spin_lock_irq(&x->wait.lock);
3685 __remove_wait_queue(&x->wait, &wait);
3688 spin_unlock_irq(&x->wait.lock);
3690 EXPORT_SYMBOL(wait_for_completion);
3692 unsigned long fastcall __sched
3693 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3697 spin_lock_irq(&x->wait.lock);
3699 DECLARE_WAITQUEUE(wait, current);
3701 wait.flags |= WQ_FLAG_EXCLUSIVE;
3702 __add_wait_queue_tail(&x->wait, &wait);
3704 __set_current_state(TASK_UNINTERRUPTIBLE);
3705 spin_unlock_irq(&x->wait.lock);
3706 timeout = schedule_timeout(timeout);
3707 spin_lock_irq(&x->wait.lock);
3709 __remove_wait_queue(&x->wait, &wait);
3713 __remove_wait_queue(&x->wait, &wait);
3717 spin_unlock_irq(&x->wait.lock);
3720 EXPORT_SYMBOL(wait_for_completion_timeout);
3722 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3728 spin_lock_irq(&x->wait.lock);
3730 DECLARE_WAITQUEUE(wait, current);
3732 wait.flags |= WQ_FLAG_EXCLUSIVE;
3733 __add_wait_queue_tail(&x->wait, &wait);
3735 if (signal_pending(current)) {
3737 __remove_wait_queue(&x->wait, &wait);
3740 __set_current_state(TASK_INTERRUPTIBLE);
3741 spin_unlock_irq(&x->wait.lock);
3743 spin_lock_irq(&x->wait.lock);
3745 __remove_wait_queue(&x->wait, &wait);
3749 spin_unlock_irq(&x->wait.lock);
3753 EXPORT_SYMBOL(wait_for_completion_interruptible);
3755 unsigned long fastcall __sched
3756 wait_for_completion_interruptible_timeout(struct completion *x,
3757 unsigned long timeout)
3761 spin_lock_irq(&x->wait.lock);
3763 DECLARE_WAITQUEUE(wait, current);
3765 wait.flags |= WQ_FLAG_EXCLUSIVE;
3766 __add_wait_queue_tail(&x->wait, &wait);
3768 if (signal_pending(current)) {
3769 timeout = -ERESTARTSYS;
3770 __remove_wait_queue(&x->wait, &wait);
3773 __set_current_state(TASK_INTERRUPTIBLE);
3774 spin_unlock_irq(&x->wait.lock);
3775 timeout = schedule_timeout(timeout);
3776 spin_lock_irq(&x->wait.lock);
3778 __remove_wait_queue(&x->wait, &wait);
3782 __remove_wait_queue(&x->wait, &wait);
3786 spin_unlock_irq(&x->wait.lock);
3789 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3792 #define SLEEP_ON_VAR \
3793 unsigned long flags; \
3794 wait_queue_t wait; \
3795 init_waitqueue_entry(&wait, current);
3797 #define SLEEP_ON_HEAD \
3798 spin_lock_irqsave(&q->lock,flags); \
3799 __add_wait_queue(q, &wait); \
3800 spin_unlock(&q->lock);
3802 #define SLEEP_ON_TAIL \
3803 spin_lock_irq(&q->lock); \
3804 __remove_wait_queue(q, &wait); \
3805 spin_unlock_irqrestore(&q->lock, flags);
3807 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3811 current->state = TASK_INTERRUPTIBLE;
3817 EXPORT_SYMBOL(interruptible_sleep_on);
3819 long fastcall __sched
3820 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3824 current->state = TASK_INTERRUPTIBLE;
3827 timeout = schedule_timeout(timeout);
3832 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3834 void fastcall __sched sleep_on(wait_queue_head_t *q)
3838 current->state = TASK_UNINTERRUPTIBLE;
3844 EXPORT_SYMBOL(sleep_on);
3846 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3850 current->state = TASK_UNINTERRUPTIBLE;
3853 timeout = schedule_timeout(timeout);
3859 EXPORT_SYMBOL(sleep_on_timeout);
3861 #ifdef CONFIG_RT_MUTEXES
3864 * rt_mutex_setprio - set the current priority of a task
3866 * @prio: prio value (kernel-internal form)
3868 * This function changes the 'effective' priority of a task. It does
3869 * not touch ->normal_prio like __setscheduler().
3871 * Used by the rt_mutex code to implement priority inheritance logic.
3873 void rt_mutex_setprio(struct task_struct *p, int prio)
3875 struct prio_array *array;
3876 unsigned long flags;
3880 BUG_ON(prio < 0 || prio > MAX_PRIO);
3882 rq = task_rq_lock(p, &flags);
3887 dequeue_task(p, array);
3892 * If changing to an RT priority then queue it
3893 * in the active array!
3897 enqueue_task(p, array);
3899 * Reschedule if we are currently running on this runqueue and
3900 * our priority decreased, or if we are not currently running on
3901 * this runqueue and our priority is higher than the current's
3903 if (task_running(rq, p)) {
3904 if (p->prio > oldprio)
3905 resched_task(rq->curr);
3906 } else if (TASK_PREEMPTS_CURR(p, rq))
3907 resched_task(rq->curr);
3909 task_rq_unlock(rq, &flags);
3914 void set_user_nice(struct task_struct *p, long nice)
3916 struct prio_array *array;
3917 int old_prio, delta;
3918 unsigned long flags;
3921 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3924 * We have to be careful, if called from sys_setpriority(),
3925 * the task might be in the middle of scheduling on another CPU.
3927 rq = task_rq_lock(p, &flags);
3929 * The RT priorities are set via sched_setscheduler(), but we still
3930 * allow the 'normal' nice value to be set - but as expected
3931 * it wont have any effect on scheduling until the task is
3932 * not SCHED_NORMAL/SCHED_BATCH:
3934 if (has_rt_policy(p)) {
3935 p->static_prio = NICE_TO_PRIO(nice);
3940 dequeue_task(p, array);
3941 dec_raw_weighted_load(rq, p);
3944 p->static_prio = NICE_TO_PRIO(nice);
3947 p->prio = effective_prio(p);
3948 delta = p->prio - old_prio;
3951 enqueue_task(p, array);
3952 inc_raw_weighted_load(rq, p);
3954 * If the task increased its priority or is running and
3955 * lowered its priority, then reschedule its CPU:
3957 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3958 resched_task(rq->curr);
3961 task_rq_unlock(rq, &flags);
3963 EXPORT_SYMBOL(set_user_nice);
3966 * can_nice - check if a task can reduce its nice value
3970 int can_nice(const struct task_struct *p, const int nice)
3972 /* convert nice value [19,-20] to rlimit style value [1,40] */
3973 int nice_rlim = 20 - nice;
3975 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3976 capable(CAP_SYS_NICE));
3979 #ifdef __ARCH_WANT_SYS_NICE
3982 * sys_nice - change the priority of the current process.
3983 * @increment: priority increment
3985 * sys_setpriority is a more generic, but much slower function that
3986 * does similar things.
3988 asmlinkage long sys_nice(int increment)
3993 * Setpriority might change our priority at the same moment.
3994 * We don't have to worry. Conceptually one call occurs first
3995 * and we have a single winner.
3997 if (increment < -40)
4002 nice = PRIO_TO_NICE(current->static_prio) + increment;
4008 if (increment < 0 && !can_nice(current, nice))
4011 retval = security_task_setnice(current, nice);
4015 set_user_nice(current, nice);
4022 * task_prio - return the priority value of a given task.
4023 * @p: the task in question.
4025 * This is the priority value as seen by users in /proc.
4026 * RT tasks are offset by -200. Normal tasks are centered
4027 * around 0, value goes from -16 to +15.
4029 int task_prio(const struct task_struct *p)
4031 return p->prio - MAX_RT_PRIO;
4035 * task_nice - return the nice value of a given task.
4036 * @p: the task in question.
4038 int task_nice(const struct task_struct *p)
4040 return TASK_NICE(p);
4042 EXPORT_SYMBOL_GPL(task_nice);
4045 * idle_cpu - is a given cpu idle currently?
4046 * @cpu: the processor in question.
4048 int idle_cpu(int cpu)
4050 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4054 * idle_task - return the idle task for a given cpu.
4055 * @cpu: the processor in question.
4057 struct task_struct *idle_task(int cpu)
4059 return cpu_rq(cpu)->idle;
4063 * find_process_by_pid - find a process with a matching PID value.
4064 * @pid: the pid in question.
4066 static inline struct task_struct *find_process_by_pid(pid_t pid)
4068 return pid ? find_task_by_pid(pid) : current;
4071 /* Actually do priority change: must hold rq lock. */
4072 static void __setscheduler(struct task_struct *p, int policy, int prio)
4077 p->rt_priority = prio;
4078 p->normal_prio = normal_prio(p);
4079 /* we are holding p->pi_lock already */
4080 p->prio = rt_mutex_getprio(p);
4082 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4084 if (policy == SCHED_BATCH)
4090 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4091 * @p: the task in question.
4092 * @policy: new policy.
4093 * @param: structure containing the new RT priority.
4095 * NOTE that the task may be already dead.
4097 int sched_setscheduler(struct task_struct *p, int policy,
4098 struct sched_param *param)
4100 int retval, oldprio, oldpolicy = -1;
4101 struct prio_array *array;
4102 unsigned long flags;
4105 /* may grab non-irq protected spin_locks */
4106 BUG_ON(in_interrupt());
4108 /* double check policy once rq lock held */
4110 policy = oldpolicy = p->policy;
4111 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4112 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4115 * Valid priorities for SCHED_FIFO and SCHED_RR are
4116 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4119 if (param->sched_priority < 0 ||
4120 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4121 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4123 if (is_rt_policy(policy) != (param->sched_priority != 0))
4127 * Allow unprivileged RT tasks to decrease priority:
4129 if (!capable(CAP_SYS_NICE)) {
4130 if (is_rt_policy(policy)) {
4131 unsigned long rlim_rtprio;
4132 unsigned long flags;
4134 if (!lock_task_sighand(p, &flags))
4136 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4137 unlock_task_sighand(p, &flags);
4139 /* can't set/change the rt policy */
4140 if (policy != p->policy && !rlim_rtprio)
4143 /* can't increase priority */
4144 if (param->sched_priority > p->rt_priority &&
4145 param->sched_priority > rlim_rtprio)
4149 /* can't change other user's priorities */
4150 if ((current->euid != p->euid) &&
4151 (current->euid != p->uid))
4155 retval = security_task_setscheduler(p, policy, param);
4159 * make sure no PI-waiters arrive (or leave) while we are
4160 * changing the priority of the task:
4162 spin_lock_irqsave(&p->pi_lock, flags);
4164 * To be able to change p->policy safely, the apropriate
4165 * runqueue lock must be held.
4167 rq = __task_rq_lock(p);
4168 /* recheck policy now with rq lock held */
4169 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4170 policy = oldpolicy = -1;
4171 __task_rq_unlock(rq);
4172 spin_unlock_irqrestore(&p->pi_lock, flags);
4177 deactivate_task(p, rq);
4179 __setscheduler(p, policy, param->sched_priority);
4181 __activate_task(p, rq);
4183 * Reschedule if we are currently running on this runqueue and
4184 * our priority decreased, or if we are not currently running on
4185 * this runqueue and our priority is higher than the current's
4187 if (task_running(rq, p)) {
4188 if (p->prio > oldprio)
4189 resched_task(rq->curr);
4190 } else if (TASK_PREEMPTS_CURR(p, rq))
4191 resched_task(rq->curr);
4193 __task_rq_unlock(rq);
4194 spin_unlock_irqrestore(&p->pi_lock, flags);
4196 rt_mutex_adjust_pi(p);
4200 EXPORT_SYMBOL_GPL(sched_setscheduler);
4203 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4205 struct sched_param lparam;
4206 struct task_struct *p;
4209 if (!param || pid < 0)
4211 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4216 p = find_process_by_pid(pid);
4218 retval = sched_setscheduler(p, policy, &lparam);
4225 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4226 * @pid: the pid in question.
4227 * @policy: new policy.
4228 * @param: structure containing the new RT priority.
4230 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4231 struct sched_param __user *param)
4233 /* negative values for policy are not valid */
4237 return do_sched_setscheduler(pid, policy, param);
4241 * sys_sched_setparam - set/change the RT priority of a thread
4242 * @pid: the pid in question.
4243 * @param: structure containing the new RT priority.
4245 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4247 return do_sched_setscheduler(pid, -1, param);
4251 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4252 * @pid: the pid in question.
4254 asmlinkage long sys_sched_getscheduler(pid_t pid)
4256 struct task_struct *p;
4257 int retval = -EINVAL;
4263 read_lock(&tasklist_lock);
4264 p = find_process_by_pid(pid);
4266 retval = security_task_getscheduler(p);
4270 read_unlock(&tasklist_lock);
4277 * sys_sched_getscheduler - get the RT priority of a thread
4278 * @pid: the pid in question.
4279 * @param: structure containing the RT priority.
4281 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4283 struct sched_param lp;
4284 struct task_struct *p;
4285 int retval = -EINVAL;
4287 if (!param || pid < 0)
4290 read_lock(&tasklist_lock);
4291 p = find_process_by_pid(pid);
4296 retval = security_task_getscheduler(p);
4300 lp.sched_priority = p->rt_priority;
4301 read_unlock(&tasklist_lock);
4304 * This one might sleep, we cannot do it with a spinlock held ...
4306 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4312 read_unlock(&tasklist_lock);
4316 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4318 cpumask_t cpus_allowed;
4319 struct task_struct *p;
4322 mutex_lock(&sched_hotcpu_mutex);
4323 read_lock(&tasklist_lock);
4325 p = find_process_by_pid(pid);
4327 read_unlock(&tasklist_lock);
4328 mutex_unlock(&sched_hotcpu_mutex);
4333 * It is not safe to call set_cpus_allowed with the
4334 * tasklist_lock held. We will bump the task_struct's
4335 * usage count and then drop tasklist_lock.
4338 read_unlock(&tasklist_lock);
4341 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4342 !capable(CAP_SYS_NICE))
4345 retval = security_task_setscheduler(p, 0, NULL);
4349 cpus_allowed = cpuset_cpus_allowed(p);
4350 cpus_and(new_mask, new_mask, cpus_allowed);
4351 retval = set_cpus_allowed(p, new_mask);
4355 mutex_unlock(&sched_hotcpu_mutex);
4359 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4360 cpumask_t *new_mask)
4362 if (len < sizeof(cpumask_t)) {
4363 memset(new_mask, 0, sizeof(cpumask_t));
4364 } else if (len > sizeof(cpumask_t)) {
4365 len = sizeof(cpumask_t);
4367 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4371 * sys_sched_setaffinity - set the cpu affinity of a process
4372 * @pid: pid of the process
4373 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4374 * @user_mask_ptr: user-space pointer to the new cpu mask
4376 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4377 unsigned long __user *user_mask_ptr)
4382 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4386 return sched_setaffinity(pid, new_mask);
4390 * Represents all cpu's present in the system
4391 * In systems capable of hotplug, this map could dynamically grow
4392 * as new cpu's are detected in the system via any platform specific
4393 * method, such as ACPI for e.g.
4396 cpumask_t cpu_present_map __read_mostly;
4397 EXPORT_SYMBOL(cpu_present_map);
4400 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4401 EXPORT_SYMBOL(cpu_online_map);
4403 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4404 EXPORT_SYMBOL(cpu_possible_map);
4407 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4409 struct task_struct *p;
4412 mutex_lock(&sched_hotcpu_mutex);
4413 read_lock(&tasklist_lock);
4416 p = find_process_by_pid(pid);
4420 retval = security_task_getscheduler(p);
4424 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4427 read_unlock(&tasklist_lock);
4428 mutex_unlock(&sched_hotcpu_mutex);
4436 * sys_sched_getaffinity - get the cpu affinity of a process
4437 * @pid: pid of the process
4438 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4439 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4441 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4442 unsigned long __user *user_mask_ptr)
4447 if (len < sizeof(cpumask_t))
4450 ret = sched_getaffinity(pid, &mask);
4454 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4457 return sizeof(cpumask_t);
4461 * sys_sched_yield - yield the current processor to other threads.
4463 * This function yields the current CPU by moving the calling thread
4464 * to the expired array. If there are no other threads running on this
4465 * CPU then this function will return.
4467 asmlinkage long sys_sched_yield(void)
4469 struct rq *rq = this_rq_lock();
4470 struct prio_array *array = current->array, *target = rq->expired;
4472 schedstat_inc(rq, yld_cnt);
4474 * We implement yielding by moving the task into the expired
4477 * (special rule: RT tasks will just roundrobin in the active
4480 if (rt_task(current))
4481 target = rq->active;
4483 if (array->nr_active == 1) {
4484 schedstat_inc(rq, yld_act_empty);
4485 if (!rq->expired->nr_active)
4486 schedstat_inc(rq, yld_both_empty);
4487 } else if (!rq->expired->nr_active)
4488 schedstat_inc(rq, yld_exp_empty);
4490 if (array != target) {
4491 dequeue_task(current, array);
4492 enqueue_task(current, target);
4495 * requeue_task is cheaper so perform that if possible.
4497 requeue_task(current, array);
4500 * Since we are going to call schedule() anyway, there's
4501 * no need to preempt or enable interrupts:
4503 __release(rq->lock);
4504 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4505 _raw_spin_unlock(&rq->lock);
4506 preempt_enable_no_resched();
4513 static void __cond_resched(void)
4515 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4516 __might_sleep(__FILE__, __LINE__);
4519 * The BKS might be reacquired before we have dropped
4520 * PREEMPT_ACTIVE, which could trigger a second
4521 * cond_resched() call.
4524 add_preempt_count(PREEMPT_ACTIVE);
4526 sub_preempt_count(PREEMPT_ACTIVE);
4527 } while (need_resched());
4530 int __sched cond_resched(void)
4532 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4533 system_state == SYSTEM_RUNNING) {
4539 EXPORT_SYMBOL(cond_resched);
4542 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4543 * call schedule, and on return reacquire the lock.
4545 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4546 * operations here to prevent schedule() from being called twice (once via
4547 * spin_unlock(), once by hand).
4549 int cond_resched_lock(spinlock_t *lock)
4553 if (need_lockbreak(lock)) {
4559 if (need_resched() && system_state == SYSTEM_RUNNING) {
4560 spin_release(&lock->dep_map, 1, _THIS_IP_);
4561 _raw_spin_unlock(lock);
4562 preempt_enable_no_resched();
4569 EXPORT_SYMBOL(cond_resched_lock);
4571 int __sched cond_resched_softirq(void)
4573 BUG_ON(!in_softirq());
4575 if (need_resched() && system_state == SYSTEM_RUNNING) {
4583 EXPORT_SYMBOL(cond_resched_softirq);
4586 * yield - yield the current processor to other threads.
4588 * This is a shortcut for kernel-space yielding - it marks the
4589 * thread runnable and calls sys_sched_yield().
4591 void __sched yield(void)
4593 set_current_state(TASK_RUNNING);
4596 EXPORT_SYMBOL(yield);
4599 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4600 * that process accounting knows that this is a task in IO wait state.
4602 * But don't do that if it is a deliberate, throttling IO wait (this task
4603 * has set its backing_dev_info: the queue against which it should throttle)
4605 void __sched io_schedule(void)
4607 struct rq *rq = &__raw_get_cpu_var(runqueues);
4609 delayacct_blkio_start();
4610 atomic_inc(&rq->nr_iowait);
4612 atomic_dec(&rq->nr_iowait);
4613 delayacct_blkio_end();
4615 EXPORT_SYMBOL(io_schedule);
4617 long __sched io_schedule_timeout(long timeout)
4619 struct rq *rq = &__raw_get_cpu_var(runqueues);
4622 delayacct_blkio_start();
4623 atomic_inc(&rq->nr_iowait);
4624 ret = schedule_timeout(timeout);
4625 atomic_dec(&rq->nr_iowait);
4626 delayacct_blkio_end();
4631 * sys_sched_get_priority_max - return maximum RT priority.
4632 * @policy: scheduling class.
4634 * this syscall returns the maximum rt_priority that can be used
4635 * by a given scheduling class.
4637 asmlinkage long sys_sched_get_priority_max(int policy)
4644 ret = MAX_USER_RT_PRIO-1;
4655 * sys_sched_get_priority_min - return minimum RT priority.
4656 * @policy: scheduling class.
4658 * this syscall returns the minimum rt_priority that can be used
4659 * by a given scheduling class.
4661 asmlinkage long sys_sched_get_priority_min(int policy)
4678 * sys_sched_rr_get_interval - return the default timeslice of a process.
4679 * @pid: pid of the process.
4680 * @interval: userspace pointer to the timeslice value.
4682 * this syscall writes the default timeslice value of a given process
4683 * into the user-space timespec buffer. A value of '0' means infinity.
4686 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4688 struct task_struct *p;
4689 int retval = -EINVAL;
4696 read_lock(&tasklist_lock);
4697 p = find_process_by_pid(pid);
4701 retval = security_task_getscheduler(p);
4705 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4706 0 : task_timeslice(p), &t);
4707 read_unlock(&tasklist_lock);
4708 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4712 read_unlock(&tasklist_lock);
4716 static const char stat_nam[] = "RSDTtZX";
4718 static void show_task(struct task_struct *p)
4720 unsigned long free = 0;
4723 state = p->state ? __ffs(p->state) + 1 : 0;
4724 printk("%-13.13s %c", p->comm,
4725 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4726 #if (BITS_PER_LONG == 32)
4727 if (state == TASK_RUNNING)
4728 printk(" running ");
4730 printk(" %08lX ", thread_saved_pc(p));
4732 if (state == TASK_RUNNING)
4733 printk(" running task ");
4735 printk(" %016lx ", thread_saved_pc(p));
4737 #ifdef CONFIG_DEBUG_STACK_USAGE
4739 unsigned long *n = end_of_stack(p);
4742 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4745 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4747 printk(" (L-TLB)\n");
4749 printk(" (NOTLB)\n");
4751 if (state != TASK_RUNNING)
4752 show_stack(p, NULL);
4755 void show_state_filter(unsigned long state_filter)
4757 struct task_struct *g, *p;
4759 #if (BITS_PER_LONG == 32)
4762 printk(" task PC stack pid father child younger older\n");
4766 printk(" task PC stack pid father child younger older\n");
4768 read_lock(&tasklist_lock);
4769 do_each_thread(g, p) {
4771 * reset the NMI-timeout, listing all files on a slow
4772 * console might take alot of time:
4774 touch_nmi_watchdog();
4775 if (!state_filter || (p->state & state_filter))
4777 } while_each_thread(g, p);
4779 touch_all_softlockup_watchdogs();
4781 read_unlock(&tasklist_lock);
4783 * Only show locks if all tasks are dumped:
4785 if (state_filter == -1)
4786 debug_show_all_locks();
4789 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4795 * init_idle - set up an idle thread for a given CPU
4796 * @idle: task in question
4797 * @cpu: cpu the idle task belongs to
4799 * NOTE: this function does not set the idle thread's NEED_RESCHED
4800 * flag, to make booting more robust.
4802 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4804 struct rq *rq = cpu_rq(cpu);
4805 unsigned long flags;
4807 idle->timestamp = sched_clock();
4808 idle->sleep_avg = 0;
4810 idle->prio = idle->normal_prio = MAX_PRIO;
4811 idle->state = TASK_RUNNING;
4812 idle->cpus_allowed = cpumask_of_cpu(cpu);
4813 set_task_cpu(idle, cpu);
4815 spin_lock_irqsave(&rq->lock, flags);
4816 rq->curr = rq->idle = idle;
4817 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4820 spin_unlock_irqrestore(&rq->lock, flags);
4822 /* Set the preempt count _outside_ the spinlocks! */
4823 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4824 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4826 task_thread_info(idle)->preempt_count = 0;
4831 * In a system that switches off the HZ timer nohz_cpu_mask
4832 * indicates which cpus entered this state. This is used
4833 * in the rcu update to wait only for active cpus. For system
4834 * which do not switch off the HZ timer nohz_cpu_mask should
4835 * always be CPU_MASK_NONE.
4837 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4841 * This is how migration works:
4843 * 1) we queue a struct migration_req structure in the source CPU's
4844 * runqueue and wake up that CPU's migration thread.
4845 * 2) we down() the locked semaphore => thread blocks.
4846 * 3) migration thread wakes up (implicitly it forces the migrated
4847 * thread off the CPU)
4848 * 4) it gets the migration request and checks whether the migrated
4849 * task is still in the wrong runqueue.
4850 * 5) if it's in the wrong runqueue then the migration thread removes
4851 * it and puts it into the right queue.
4852 * 6) migration thread up()s the semaphore.
4853 * 7) we wake up and the migration is done.
4857 * Change a given task's CPU affinity. Migrate the thread to a
4858 * proper CPU and schedule it away if the CPU it's executing on
4859 * is removed from the allowed bitmask.
4861 * NOTE: the caller must have a valid reference to the task, the
4862 * task must not exit() & deallocate itself prematurely. The
4863 * call is not atomic; no spinlocks may be held.
4865 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4867 struct migration_req req;
4868 unsigned long flags;
4872 rq = task_rq_lock(p, &flags);
4873 if (!cpus_intersects(new_mask, cpu_online_map)) {
4878 p->cpus_allowed = new_mask;
4879 /* Can the task run on the task's current CPU? If so, we're done */
4880 if (cpu_isset(task_cpu(p), new_mask))
4883 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4884 /* Need help from migration thread: drop lock and wait. */
4885 task_rq_unlock(rq, &flags);
4886 wake_up_process(rq->migration_thread);
4887 wait_for_completion(&req.done);
4888 tlb_migrate_finish(p->mm);
4892 task_rq_unlock(rq, &flags);
4896 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4899 * Move (not current) task off this cpu, onto dest cpu. We're doing
4900 * this because either it can't run here any more (set_cpus_allowed()
4901 * away from this CPU, or CPU going down), or because we're
4902 * attempting to rebalance this task on exec (sched_exec).
4904 * So we race with normal scheduler movements, but that's OK, as long
4905 * as the task is no longer on this CPU.
4907 * Returns non-zero if task was successfully migrated.
4909 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4911 struct rq *rq_dest, *rq_src;
4914 if (unlikely(cpu_is_offline(dest_cpu)))
4917 rq_src = cpu_rq(src_cpu);
4918 rq_dest = cpu_rq(dest_cpu);
4920 double_rq_lock(rq_src, rq_dest);
4921 /* Already moved. */
4922 if (task_cpu(p) != src_cpu)
4924 /* Affinity changed (again). */
4925 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4928 set_task_cpu(p, dest_cpu);
4931 * Sync timestamp with rq_dest's before activating.
4932 * The same thing could be achieved by doing this step
4933 * afterwards, and pretending it was a local activate.
4934 * This way is cleaner and logically correct.
4936 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
4937 + rq_dest->most_recent_timestamp;
4938 deactivate_task(p, rq_src);
4939 __activate_task(p, rq_dest);
4940 if (TASK_PREEMPTS_CURR(p, rq_dest))
4941 resched_task(rq_dest->curr);
4945 double_rq_unlock(rq_src, rq_dest);
4950 * migration_thread - this is a highprio system thread that performs
4951 * thread migration by bumping thread off CPU then 'pushing' onto
4954 static int migration_thread(void *data)
4956 int cpu = (long)data;
4960 BUG_ON(rq->migration_thread != current);
4962 set_current_state(TASK_INTERRUPTIBLE);
4963 while (!kthread_should_stop()) {
4964 struct migration_req *req;
4965 struct list_head *head;
4969 spin_lock_irq(&rq->lock);
4971 if (cpu_is_offline(cpu)) {
4972 spin_unlock_irq(&rq->lock);
4976 if (rq->active_balance) {
4977 active_load_balance(rq, cpu);
4978 rq->active_balance = 0;
4981 head = &rq->migration_queue;
4983 if (list_empty(head)) {
4984 spin_unlock_irq(&rq->lock);
4986 set_current_state(TASK_INTERRUPTIBLE);
4989 req = list_entry(head->next, struct migration_req, list);
4990 list_del_init(head->next);
4992 spin_unlock(&rq->lock);
4993 __migrate_task(req->task, cpu, req->dest_cpu);
4996 complete(&req->done);
4998 __set_current_state(TASK_RUNNING);
5002 /* Wait for kthread_stop */
5003 set_current_state(TASK_INTERRUPTIBLE);
5004 while (!kthread_should_stop()) {
5006 set_current_state(TASK_INTERRUPTIBLE);
5008 __set_current_state(TASK_RUNNING);
5012 #ifdef CONFIG_HOTPLUG_CPU
5014 * Figure out where task on dead CPU should go, use force if neccessary.
5015 * NOTE: interrupts should be disabled by the caller
5017 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5019 unsigned long flags;
5026 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5027 cpus_and(mask, mask, p->cpus_allowed);
5028 dest_cpu = any_online_cpu(mask);
5030 /* On any allowed CPU? */
5031 if (dest_cpu == NR_CPUS)
5032 dest_cpu = any_online_cpu(p->cpus_allowed);
5034 /* No more Mr. Nice Guy. */
5035 if (dest_cpu == NR_CPUS) {
5036 rq = task_rq_lock(p, &flags);
5037 cpus_setall(p->cpus_allowed);
5038 dest_cpu = any_online_cpu(p->cpus_allowed);
5039 task_rq_unlock(rq, &flags);
5042 * Don't tell them about moving exiting tasks or
5043 * kernel threads (both mm NULL), since they never
5046 if (p->mm && printk_ratelimit())
5047 printk(KERN_INFO "process %d (%s) no "
5048 "longer affine to cpu%d\n",
5049 p->pid, p->comm, dead_cpu);
5051 if (!__migrate_task(p, dead_cpu, dest_cpu))
5056 * While a dead CPU has no uninterruptible tasks queued at this point,
5057 * it might still have a nonzero ->nr_uninterruptible counter, because
5058 * for performance reasons the counter is not stricly tracking tasks to
5059 * their home CPUs. So we just add the counter to another CPU's counter,
5060 * to keep the global sum constant after CPU-down:
5062 static void migrate_nr_uninterruptible(struct rq *rq_src)
5064 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5065 unsigned long flags;
5067 local_irq_save(flags);
5068 double_rq_lock(rq_src, rq_dest);
5069 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5070 rq_src->nr_uninterruptible = 0;
5071 double_rq_unlock(rq_src, rq_dest);
5072 local_irq_restore(flags);
5075 /* Run through task list and migrate tasks from the dead cpu. */
5076 static void migrate_live_tasks(int src_cpu)
5078 struct task_struct *p, *t;
5080 write_lock_irq(&tasklist_lock);
5082 do_each_thread(t, p) {
5086 if (task_cpu(p) == src_cpu)
5087 move_task_off_dead_cpu(src_cpu, p);
5088 } while_each_thread(t, p);
5090 write_unlock_irq(&tasklist_lock);
5093 /* Schedules idle task to be the next runnable task on current CPU.
5094 * It does so by boosting its priority to highest possible and adding it to
5095 * the _front_ of the runqueue. Used by CPU offline code.
5097 void sched_idle_next(void)
5099 int this_cpu = smp_processor_id();
5100 struct rq *rq = cpu_rq(this_cpu);
5101 struct task_struct *p = rq->idle;
5102 unsigned long flags;
5104 /* cpu has to be offline */
5105 BUG_ON(cpu_online(this_cpu));
5108 * Strictly not necessary since rest of the CPUs are stopped by now
5109 * and interrupts disabled on the current cpu.
5111 spin_lock_irqsave(&rq->lock, flags);
5113 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5115 /* Add idle task to the _front_ of its priority queue: */
5116 __activate_idle_task(p, rq);
5118 spin_unlock_irqrestore(&rq->lock, flags);
5122 * Ensures that the idle task is using init_mm right before its cpu goes
5125 void idle_task_exit(void)
5127 struct mm_struct *mm = current->active_mm;
5129 BUG_ON(cpu_online(smp_processor_id()));
5132 switch_mm(mm, &init_mm, current);
5136 /* called under rq->lock with disabled interrupts */
5137 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5139 struct rq *rq = cpu_rq(dead_cpu);
5141 /* Must be exiting, otherwise would be on tasklist. */
5142 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5144 /* Cannot have done final schedule yet: would have vanished. */
5145 BUG_ON(p->state == TASK_DEAD);
5150 * Drop lock around migration; if someone else moves it,
5151 * that's OK. No task can be added to this CPU, so iteration is
5153 * NOTE: interrupts should be left disabled --dev@
5155 spin_unlock(&rq->lock);
5156 move_task_off_dead_cpu(dead_cpu, p);
5157 spin_lock(&rq->lock);
5162 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5163 static void migrate_dead_tasks(unsigned int dead_cpu)
5165 struct rq *rq = cpu_rq(dead_cpu);
5166 unsigned int arr, i;
5168 for (arr = 0; arr < 2; arr++) {
5169 for (i = 0; i < MAX_PRIO; i++) {
5170 struct list_head *list = &rq->arrays[arr].queue[i];
5172 while (!list_empty(list))
5173 migrate_dead(dead_cpu, list_entry(list->next,
5174 struct task_struct, run_list));
5178 #endif /* CONFIG_HOTPLUG_CPU */
5181 * migration_call - callback that gets triggered when a CPU is added.
5182 * Here we can start up the necessary migration thread for the new CPU.
5184 static int __cpuinit
5185 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5187 struct task_struct *p;
5188 int cpu = (long)hcpu;
5189 unsigned long flags;
5193 case CPU_LOCK_ACQUIRE:
5194 mutex_lock(&sched_hotcpu_mutex);
5197 case CPU_UP_PREPARE:
5198 case CPU_UP_PREPARE_FROZEN:
5199 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5202 p->flags |= PF_NOFREEZE;
5203 kthread_bind(p, cpu);
5204 /* Must be high prio: stop_machine expects to yield to it. */
5205 rq = task_rq_lock(p, &flags);
5206 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5207 task_rq_unlock(rq, &flags);
5208 cpu_rq(cpu)->migration_thread = p;
5212 case CPU_ONLINE_FROZEN:
5213 /* Strictly unneccessary, as first user will wake it. */
5214 wake_up_process(cpu_rq(cpu)->migration_thread);
5217 #ifdef CONFIG_HOTPLUG_CPU
5218 case CPU_UP_CANCELED:
5219 case CPU_UP_CANCELED_FROZEN:
5220 if (!cpu_rq(cpu)->migration_thread)
5222 /* Unbind it from offline cpu so it can run. Fall thru. */
5223 kthread_bind(cpu_rq(cpu)->migration_thread,
5224 any_online_cpu(cpu_online_map));
5225 kthread_stop(cpu_rq(cpu)->migration_thread);
5226 cpu_rq(cpu)->migration_thread = NULL;
5230 case CPU_DEAD_FROZEN:
5231 migrate_live_tasks(cpu);
5233 kthread_stop(rq->migration_thread);
5234 rq->migration_thread = NULL;
5235 /* Idle task back to normal (off runqueue, low prio) */
5236 rq = task_rq_lock(rq->idle, &flags);
5237 deactivate_task(rq->idle, rq);
5238 rq->idle->static_prio = MAX_PRIO;
5239 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5240 migrate_dead_tasks(cpu);
5241 task_rq_unlock(rq, &flags);
5242 migrate_nr_uninterruptible(rq);
5243 BUG_ON(rq->nr_running != 0);
5245 /* No need to migrate the tasks: it was best-effort if
5246 * they didn't take sched_hotcpu_mutex. Just wake up
5247 * the requestors. */
5248 spin_lock_irq(&rq->lock);
5249 while (!list_empty(&rq->migration_queue)) {
5250 struct migration_req *req;
5252 req = list_entry(rq->migration_queue.next,
5253 struct migration_req, list);
5254 list_del_init(&req->list);
5255 complete(&req->done);
5257 spin_unlock_irq(&rq->lock);
5260 case CPU_LOCK_RELEASE:
5261 mutex_unlock(&sched_hotcpu_mutex);
5267 /* Register at highest priority so that task migration (migrate_all_tasks)
5268 * happens before everything else.
5270 static struct notifier_block __cpuinitdata migration_notifier = {
5271 .notifier_call = migration_call,
5275 int __init migration_init(void)
5277 void *cpu = (void *)(long)smp_processor_id();
5280 /* Start one for the boot CPU: */
5281 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5282 BUG_ON(err == NOTIFY_BAD);
5283 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5284 register_cpu_notifier(&migration_notifier);
5292 /* Number of possible processor ids */
5293 int nr_cpu_ids __read_mostly = NR_CPUS;
5294 EXPORT_SYMBOL(nr_cpu_ids);
5296 #undef SCHED_DOMAIN_DEBUG
5297 #ifdef SCHED_DOMAIN_DEBUG
5298 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5303 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5307 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5312 struct sched_group *group = sd->groups;
5313 cpumask_t groupmask;
5315 cpumask_scnprintf(str, NR_CPUS, sd->span);
5316 cpus_clear(groupmask);
5319 for (i = 0; i < level + 1; i++)
5321 printk("domain %d: ", level);
5323 if (!(sd->flags & SD_LOAD_BALANCE)) {
5324 printk("does not load-balance\n");
5326 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5331 printk("span %s\n", str);
5333 if (!cpu_isset(cpu, sd->span))
5334 printk(KERN_ERR "ERROR: domain->span does not contain "
5336 if (!cpu_isset(cpu, group->cpumask))
5337 printk(KERN_ERR "ERROR: domain->groups does not contain"
5341 for (i = 0; i < level + 2; i++)
5347 printk(KERN_ERR "ERROR: group is NULL\n");
5351 if (!group->__cpu_power) {
5353 printk(KERN_ERR "ERROR: domain->cpu_power not "
5357 if (!cpus_weight(group->cpumask)) {
5359 printk(KERN_ERR "ERROR: empty group\n");
5362 if (cpus_intersects(groupmask, group->cpumask)) {
5364 printk(KERN_ERR "ERROR: repeated CPUs\n");
5367 cpus_or(groupmask, groupmask, group->cpumask);
5369 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5372 group = group->next;
5373 } while (group != sd->groups);
5376 if (!cpus_equal(sd->span, groupmask))
5377 printk(KERN_ERR "ERROR: groups don't span "
5385 if (!cpus_subset(groupmask, sd->span))
5386 printk(KERN_ERR "ERROR: parent span is not a superset "
5387 "of domain->span\n");
5392 # define sched_domain_debug(sd, cpu) do { } while (0)
5395 static int sd_degenerate(struct sched_domain *sd)
5397 if (cpus_weight(sd->span) == 1)
5400 /* Following flags need at least 2 groups */
5401 if (sd->flags & (SD_LOAD_BALANCE |
5402 SD_BALANCE_NEWIDLE |
5406 SD_SHARE_PKG_RESOURCES)) {
5407 if (sd->groups != sd->groups->next)
5411 /* Following flags don't use groups */
5412 if (sd->flags & (SD_WAKE_IDLE |
5421 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5423 unsigned long cflags = sd->flags, pflags = parent->flags;
5425 if (sd_degenerate(parent))
5428 if (!cpus_equal(sd->span, parent->span))
5431 /* Does parent contain flags not in child? */
5432 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5433 if (cflags & SD_WAKE_AFFINE)
5434 pflags &= ~SD_WAKE_BALANCE;
5435 /* Flags needing groups don't count if only 1 group in parent */
5436 if (parent->groups == parent->groups->next) {
5437 pflags &= ~(SD_LOAD_BALANCE |
5438 SD_BALANCE_NEWIDLE |
5442 SD_SHARE_PKG_RESOURCES);
5444 if (~cflags & pflags)
5451 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5452 * hold the hotplug lock.
5454 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5456 struct rq *rq = cpu_rq(cpu);
5457 struct sched_domain *tmp;
5459 /* Remove the sched domains which do not contribute to scheduling. */
5460 for (tmp = sd; tmp; tmp = tmp->parent) {
5461 struct sched_domain *parent = tmp->parent;
5464 if (sd_parent_degenerate(tmp, parent)) {
5465 tmp->parent = parent->parent;
5467 parent->parent->child = tmp;
5471 if (sd && sd_degenerate(sd)) {
5477 sched_domain_debug(sd, cpu);
5479 rcu_assign_pointer(rq->sd, sd);
5482 /* cpus with isolated domains */
5483 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5485 /* Setup the mask of cpus configured for isolated domains */
5486 static int __init isolated_cpu_setup(char *str)
5488 int ints[NR_CPUS], i;
5490 str = get_options(str, ARRAY_SIZE(ints), ints);
5491 cpus_clear(cpu_isolated_map);
5492 for (i = 1; i <= ints[0]; i++)
5493 if (ints[i] < NR_CPUS)
5494 cpu_set(ints[i], cpu_isolated_map);
5498 __setup ("isolcpus=", isolated_cpu_setup);
5501 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5502 * to a function which identifies what group(along with sched group) a CPU
5503 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5504 * (due to the fact that we keep track of groups covered with a cpumask_t).
5506 * init_sched_build_groups will build a circular linked list of the groups
5507 * covered by the given span, and will set each group's ->cpumask correctly,
5508 * and ->cpu_power to 0.
5511 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5512 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5513 struct sched_group **sg))
5515 struct sched_group *first = NULL, *last = NULL;
5516 cpumask_t covered = CPU_MASK_NONE;
5519 for_each_cpu_mask(i, span) {
5520 struct sched_group *sg;
5521 int group = group_fn(i, cpu_map, &sg);
5524 if (cpu_isset(i, covered))
5527 sg->cpumask = CPU_MASK_NONE;
5528 sg->__cpu_power = 0;
5530 for_each_cpu_mask(j, span) {
5531 if (group_fn(j, cpu_map, NULL) != group)
5534 cpu_set(j, covered);
5535 cpu_set(j, sg->cpumask);
5546 #define SD_NODES_PER_DOMAIN 16
5551 * find_next_best_node - find the next node to include in a sched_domain
5552 * @node: node whose sched_domain we're building
5553 * @used_nodes: nodes already in the sched_domain
5555 * Find the next node to include in a given scheduling domain. Simply
5556 * finds the closest node not already in the @used_nodes map.
5558 * Should use nodemask_t.
5560 static int find_next_best_node(int node, unsigned long *used_nodes)
5562 int i, n, val, min_val, best_node = 0;
5566 for (i = 0; i < MAX_NUMNODES; i++) {
5567 /* Start at @node */
5568 n = (node + i) % MAX_NUMNODES;
5570 if (!nr_cpus_node(n))
5573 /* Skip already used nodes */
5574 if (test_bit(n, used_nodes))
5577 /* Simple min distance search */
5578 val = node_distance(node, n);
5580 if (val < min_val) {
5586 set_bit(best_node, used_nodes);
5591 * sched_domain_node_span - get a cpumask for a node's sched_domain
5592 * @node: node whose cpumask we're constructing
5593 * @size: number of nodes to include in this span
5595 * Given a node, construct a good cpumask for its sched_domain to span. It
5596 * should be one that prevents unnecessary balancing, but also spreads tasks
5599 static cpumask_t sched_domain_node_span(int node)
5601 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5602 cpumask_t span, nodemask;
5606 bitmap_zero(used_nodes, MAX_NUMNODES);
5608 nodemask = node_to_cpumask(node);
5609 cpus_or(span, span, nodemask);
5610 set_bit(node, used_nodes);
5612 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5613 int next_node = find_next_best_node(node, used_nodes);
5615 nodemask = node_to_cpumask(next_node);
5616 cpus_or(span, span, nodemask);
5623 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5626 * SMT sched-domains:
5628 #ifdef CONFIG_SCHED_SMT
5629 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5630 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5632 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5633 struct sched_group **sg)
5636 *sg = &per_cpu(sched_group_cpus, cpu);
5642 * multi-core sched-domains:
5644 #ifdef CONFIG_SCHED_MC
5645 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5646 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5649 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5650 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5651 struct sched_group **sg)
5654 cpumask_t mask = cpu_sibling_map[cpu];
5655 cpus_and(mask, mask, *cpu_map);
5656 group = first_cpu(mask);
5658 *sg = &per_cpu(sched_group_core, group);
5661 #elif defined(CONFIG_SCHED_MC)
5662 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5663 struct sched_group **sg)
5666 *sg = &per_cpu(sched_group_core, cpu);
5671 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5672 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5674 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5675 struct sched_group **sg)
5678 #ifdef CONFIG_SCHED_MC
5679 cpumask_t mask = cpu_coregroup_map(cpu);
5680 cpus_and(mask, mask, *cpu_map);
5681 group = first_cpu(mask);
5682 #elif defined(CONFIG_SCHED_SMT)
5683 cpumask_t mask = cpu_sibling_map[cpu];
5684 cpus_and(mask, mask, *cpu_map);
5685 group = first_cpu(mask);
5690 *sg = &per_cpu(sched_group_phys, group);
5696 * The init_sched_build_groups can't handle what we want to do with node
5697 * groups, so roll our own. Now each node has its own list of groups which
5698 * gets dynamically allocated.
5700 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5701 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5703 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5704 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5706 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5707 struct sched_group **sg)
5709 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5712 cpus_and(nodemask, nodemask, *cpu_map);
5713 group = first_cpu(nodemask);
5716 *sg = &per_cpu(sched_group_allnodes, group);
5720 static void init_numa_sched_groups_power(struct sched_group *group_head)
5722 struct sched_group *sg = group_head;
5728 for_each_cpu_mask(j, sg->cpumask) {
5729 struct sched_domain *sd;
5731 sd = &per_cpu(phys_domains, j);
5732 if (j != first_cpu(sd->groups->cpumask)) {
5734 * Only add "power" once for each
5740 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5743 if (sg != group_head)
5749 /* Free memory allocated for various sched_group structures */
5750 static void free_sched_groups(const cpumask_t *cpu_map)
5754 for_each_cpu_mask(cpu, *cpu_map) {
5755 struct sched_group **sched_group_nodes
5756 = sched_group_nodes_bycpu[cpu];
5758 if (!sched_group_nodes)
5761 for (i = 0; i < MAX_NUMNODES; i++) {
5762 cpumask_t nodemask = node_to_cpumask(i);
5763 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5765 cpus_and(nodemask, nodemask, *cpu_map);
5766 if (cpus_empty(nodemask))
5776 if (oldsg != sched_group_nodes[i])
5779 kfree(sched_group_nodes);
5780 sched_group_nodes_bycpu[cpu] = NULL;
5784 static void free_sched_groups(const cpumask_t *cpu_map)
5790 * Initialize sched groups cpu_power.
5792 * cpu_power indicates the capacity of sched group, which is used while
5793 * distributing the load between different sched groups in a sched domain.
5794 * Typically cpu_power for all the groups in a sched domain will be same unless
5795 * there are asymmetries in the topology. If there are asymmetries, group
5796 * having more cpu_power will pickup more load compared to the group having
5799 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5800 * the maximum number of tasks a group can handle in the presence of other idle
5801 * or lightly loaded groups in the same sched domain.
5803 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5805 struct sched_domain *child;
5806 struct sched_group *group;
5808 WARN_ON(!sd || !sd->groups);
5810 if (cpu != first_cpu(sd->groups->cpumask))
5815 sd->groups->__cpu_power = 0;
5818 * For perf policy, if the groups in child domain share resources
5819 * (for example cores sharing some portions of the cache hierarchy
5820 * or SMT), then set this domain groups cpu_power such that each group
5821 * can handle only one task, when there are other idle groups in the
5822 * same sched domain.
5824 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5826 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5827 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5832 * add cpu_power of each child group to this groups cpu_power
5834 group = child->groups;
5836 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5837 group = group->next;
5838 } while (group != child->groups);
5842 * Build sched domains for a given set of cpus and attach the sched domains
5843 * to the individual cpus
5845 static int build_sched_domains(const cpumask_t *cpu_map)
5848 struct sched_domain *sd;
5850 struct sched_group **sched_group_nodes = NULL;
5851 int sd_allnodes = 0;
5854 * Allocate the per-node list of sched groups
5856 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5858 if (!sched_group_nodes) {
5859 printk(KERN_WARNING "Can not alloc sched group node list\n");
5862 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5866 * Set up domains for cpus specified by the cpu_map.
5868 for_each_cpu_mask(i, *cpu_map) {
5869 struct sched_domain *sd = NULL, *p;
5870 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5872 cpus_and(nodemask, nodemask, *cpu_map);
5875 if (cpus_weight(*cpu_map)
5876 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5877 sd = &per_cpu(allnodes_domains, i);
5878 *sd = SD_ALLNODES_INIT;
5879 sd->span = *cpu_map;
5880 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
5886 sd = &per_cpu(node_domains, i);
5888 sd->span = sched_domain_node_span(cpu_to_node(i));
5892 cpus_and(sd->span, sd->span, *cpu_map);
5896 sd = &per_cpu(phys_domains, i);
5898 sd->span = nodemask;
5902 cpu_to_phys_group(i, cpu_map, &sd->groups);
5904 #ifdef CONFIG_SCHED_MC
5906 sd = &per_cpu(core_domains, i);
5908 sd->span = cpu_coregroup_map(i);
5909 cpus_and(sd->span, sd->span, *cpu_map);
5912 cpu_to_core_group(i, cpu_map, &sd->groups);
5915 #ifdef CONFIG_SCHED_SMT
5917 sd = &per_cpu(cpu_domains, i);
5918 *sd = SD_SIBLING_INIT;
5919 sd->span = cpu_sibling_map[i];
5920 cpus_and(sd->span, sd->span, *cpu_map);
5923 cpu_to_cpu_group(i, cpu_map, &sd->groups);
5927 #ifdef CONFIG_SCHED_SMT
5928 /* Set up CPU (sibling) groups */
5929 for_each_cpu_mask(i, *cpu_map) {
5930 cpumask_t this_sibling_map = cpu_sibling_map[i];
5931 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
5932 if (i != first_cpu(this_sibling_map))
5935 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
5939 #ifdef CONFIG_SCHED_MC
5940 /* Set up multi-core groups */
5941 for_each_cpu_mask(i, *cpu_map) {
5942 cpumask_t this_core_map = cpu_coregroup_map(i);
5943 cpus_and(this_core_map, this_core_map, *cpu_map);
5944 if (i != first_cpu(this_core_map))
5946 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
5951 /* Set up physical groups */
5952 for (i = 0; i < MAX_NUMNODES; i++) {
5953 cpumask_t nodemask = node_to_cpumask(i);
5955 cpus_and(nodemask, nodemask, *cpu_map);
5956 if (cpus_empty(nodemask))
5959 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
5963 /* Set up node groups */
5965 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
5967 for (i = 0; i < MAX_NUMNODES; i++) {
5968 /* Set up node groups */
5969 struct sched_group *sg, *prev;
5970 cpumask_t nodemask = node_to_cpumask(i);
5971 cpumask_t domainspan;
5972 cpumask_t covered = CPU_MASK_NONE;
5975 cpus_and(nodemask, nodemask, *cpu_map);
5976 if (cpus_empty(nodemask)) {
5977 sched_group_nodes[i] = NULL;
5981 domainspan = sched_domain_node_span(i);
5982 cpus_and(domainspan, domainspan, *cpu_map);
5984 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
5986 printk(KERN_WARNING "Can not alloc domain group for "
5990 sched_group_nodes[i] = sg;
5991 for_each_cpu_mask(j, nodemask) {
5992 struct sched_domain *sd;
5993 sd = &per_cpu(node_domains, j);
5996 sg->__cpu_power = 0;
5997 sg->cpumask = nodemask;
5999 cpus_or(covered, covered, nodemask);
6002 for (j = 0; j < MAX_NUMNODES; j++) {
6003 cpumask_t tmp, notcovered;
6004 int n = (i + j) % MAX_NUMNODES;
6006 cpus_complement(notcovered, covered);
6007 cpus_and(tmp, notcovered, *cpu_map);
6008 cpus_and(tmp, tmp, domainspan);
6009 if (cpus_empty(tmp))
6012 nodemask = node_to_cpumask(n);
6013 cpus_and(tmp, tmp, nodemask);
6014 if (cpus_empty(tmp))
6017 sg = kmalloc_node(sizeof(struct sched_group),
6021 "Can not alloc domain group for node %d\n", j);
6024 sg->__cpu_power = 0;
6026 sg->next = prev->next;
6027 cpus_or(covered, covered, tmp);
6034 /* Calculate CPU power for physical packages and nodes */
6035 #ifdef CONFIG_SCHED_SMT
6036 for_each_cpu_mask(i, *cpu_map) {
6037 sd = &per_cpu(cpu_domains, i);
6038 init_sched_groups_power(i, sd);
6041 #ifdef CONFIG_SCHED_MC
6042 for_each_cpu_mask(i, *cpu_map) {
6043 sd = &per_cpu(core_domains, i);
6044 init_sched_groups_power(i, sd);
6048 for_each_cpu_mask(i, *cpu_map) {
6049 sd = &per_cpu(phys_domains, i);
6050 init_sched_groups_power(i, sd);
6054 for (i = 0; i < MAX_NUMNODES; i++)
6055 init_numa_sched_groups_power(sched_group_nodes[i]);
6058 struct sched_group *sg;
6060 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6061 init_numa_sched_groups_power(sg);
6065 /* Attach the domains */
6066 for_each_cpu_mask(i, *cpu_map) {
6067 struct sched_domain *sd;
6068 #ifdef CONFIG_SCHED_SMT
6069 sd = &per_cpu(cpu_domains, i);
6070 #elif defined(CONFIG_SCHED_MC)
6071 sd = &per_cpu(core_domains, i);
6073 sd = &per_cpu(phys_domains, i);
6075 cpu_attach_domain(sd, i);
6082 free_sched_groups(cpu_map);
6087 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6089 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6091 cpumask_t cpu_default_map;
6095 * Setup mask for cpus without special case scheduling requirements.
6096 * For now this just excludes isolated cpus, but could be used to
6097 * exclude other special cases in the future.
6099 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6101 err = build_sched_domains(&cpu_default_map);
6106 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6108 free_sched_groups(cpu_map);
6112 * Detach sched domains from a group of cpus specified in cpu_map
6113 * These cpus will now be attached to the NULL domain
6115 static void detach_destroy_domains(const cpumask_t *cpu_map)
6119 for_each_cpu_mask(i, *cpu_map)
6120 cpu_attach_domain(NULL, i);
6121 synchronize_sched();
6122 arch_destroy_sched_domains(cpu_map);
6126 * Partition sched domains as specified by the cpumasks below.
6127 * This attaches all cpus from the cpumasks to the NULL domain,
6128 * waits for a RCU quiescent period, recalculates sched
6129 * domain information and then attaches them back to the
6130 * correct sched domains
6131 * Call with hotplug lock held
6133 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6135 cpumask_t change_map;
6138 cpus_and(*partition1, *partition1, cpu_online_map);
6139 cpus_and(*partition2, *partition2, cpu_online_map);
6140 cpus_or(change_map, *partition1, *partition2);
6142 /* Detach sched domains from all of the affected cpus */
6143 detach_destroy_domains(&change_map);
6144 if (!cpus_empty(*partition1))
6145 err = build_sched_domains(partition1);
6146 if (!err && !cpus_empty(*partition2))
6147 err = build_sched_domains(partition2);
6152 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6153 int arch_reinit_sched_domains(void)
6157 mutex_lock(&sched_hotcpu_mutex);
6158 detach_destroy_domains(&cpu_online_map);
6159 err = arch_init_sched_domains(&cpu_online_map);
6160 mutex_unlock(&sched_hotcpu_mutex);
6165 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6169 if (buf[0] != '0' && buf[0] != '1')
6173 sched_smt_power_savings = (buf[0] == '1');
6175 sched_mc_power_savings = (buf[0] == '1');
6177 ret = arch_reinit_sched_domains();
6179 return ret ? ret : count;
6182 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6186 #ifdef CONFIG_SCHED_SMT
6188 err = sysfs_create_file(&cls->kset.kobj,
6189 &attr_sched_smt_power_savings.attr);
6191 #ifdef CONFIG_SCHED_MC
6192 if (!err && mc_capable())
6193 err = sysfs_create_file(&cls->kset.kobj,
6194 &attr_sched_mc_power_savings.attr);
6200 #ifdef CONFIG_SCHED_MC
6201 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6203 return sprintf(page, "%u\n", sched_mc_power_savings);
6205 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6206 const char *buf, size_t count)
6208 return sched_power_savings_store(buf, count, 0);
6210 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6211 sched_mc_power_savings_store);
6214 #ifdef CONFIG_SCHED_SMT
6215 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6217 return sprintf(page, "%u\n", sched_smt_power_savings);
6219 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6220 const char *buf, size_t count)
6222 return sched_power_savings_store(buf, count, 1);
6224 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6225 sched_smt_power_savings_store);
6229 * Force a reinitialization of the sched domains hierarchy. The domains
6230 * and groups cannot be updated in place without racing with the balancing
6231 * code, so we temporarily attach all running cpus to the NULL domain
6232 * which will prevent rebalancing while the sched domains are recalculated.
6234 static int update_sched_domains(struct notifier_block *nfb,
6235 unsigned long action, void *hcpu)
6238 case CPU_UP_PREPARE:
6239 case CPU_UP_PREPARE_FROZEN:
6240 case CPU_DOWN_PREPARE:
6241 case CPU_DOWN_PREPARE_FROZEN:
6242 detach_destroy_domains(&cpu_online_map);
6245 case CPU_UP_CANCELED:
6246 case CPU_UP_CANCELED_FROZEN:
6247 case CPU_DOWN_FAILED:
6248 case CPU_DOWN_FAILED_FROZEN:
6250 case CPU_ONLINE_FROZEN:
6252 case CPU_DEAD_FROZEN:
6254 * Fall through and re-initialise the domains.
6261 /* The hotplug lock is already held by cpu_up/cpu_down */
6262 arch_init_sched_domains(&cpu_online_map);
6267 void __init sched_init_smp(void)
6269 cpumask_t non_isolated_cpus;
6271 mutex_lock(&sched_hotcpu_mutex);
6272 arch_init_sched_domains(&cpu_online_map);
6273 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6274 if (cpus_empty(non_isolated_cpus))
6275 cpu_set(smp_processor_id(), non_isolated_cpus);
6276 mutex_unlock(&sched_hotcpu_mutex);
6277 /* XXX: Theoretical race here - CPU may be hotplugged now */
6278 hotcpu_notifier(update_sched_domains, 0);
6280 /* Move init over to a non-isolated CPU */
6281 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6285 void __init sched_init_smp(void)
6288 #endif /* CONFIG_SMP */
6290 int in_sched_functions(unsigned long addr)
6292 /* Linker adds these: start and end of __sched functions */
6293 extern char __sched_text_start[], __sched_text_end[];
6295 return in_lock_functions(addr) ||
6296 (addr >= (unsigned long)__sched_text_start
6297 && addr < (unsigned long)__sched_text_end);
6300 void __init sched_init(void)
6303 int highest_cpu = 0;
6305 for_each_possible_cpu(i) {
6306 struct prio_array *array;
6310 spin_lock_init(&rq->lock);
6311 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6313 rq->active = rq->arrays;
6314 rq->expired = rq->arrays + 1;
6315 rq->best_expired_prio = MAX_PRIO;
6319 for (j = 1; j < 3; j++)
6320 rq->cpu_load[j] = 0;
6321 rq->active_balance = 0;
6324 rq->migration_thread = NULL;
6325 INIT_LIST_HEAD(&rq->migration_queue);
6327 atomic_set(&rq->nr_iowait, 0);
6329 for (j = 0; j < 2; j++) {
6330 array = rq->arrays + j;
6331 for (k = 0; k < MAX_PRIO; k++) {
6332 INIT_LIST_HEAD(array->queue + k);
6333 __clear_bit(k, array->bitmap);
6335 // delimiter for bitsearch
6336 __set_bit(MAX_PRIO, array->bitmap);
6341 set_load_weight(&init_task);
6344 nr_cpu_ids = highest_cpu + 1;
6345 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6348 #ifdef CONFIG_RT_MUTEXES
6349 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6353 * The boot idle thread does lazy MMU switching as well:
6355 atomic_inc(&init_mm.mm_count);
6356 enter_lazy_tlb(&init_mm, current);
6359 * Make us the idle thread. Technically, schedule() should not be
6360 * called from this thread, however somewhere below it might be,
6361 * but because we are the idle thread, we just pick up running again
6362 * when this runqueue becomes "idle".
6364 init_idle(current, smp_processor_id());
6367 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6368 void __might_sleep(char *file, int line)
6371 static unsigned long prev_jiffy; /* ratelimiting */
6373 if ((in_atomic() || irqs_disabled()) &&
6374 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6375 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6377 prev_jiffy = jiffies;
6378 printk(KERN_ERR "BUG: sleeping function called from invalid"
6379 " context at %s:%d\n", file, line);
6380 printk("in_atomic():%d, irqs_disabled():%d\n",
6381 in_atomic(), irqs_disabled());
6382 debug_show_held_locks(current);
6383 if (irqs_disabled())
6384 print_irqtrace_events(current);
6389 EXPORT_SYMBOL(__might_sleep);
6392 #ifdef CONFIG_MAGIC_SYSRQ
6393 void normalize_rt_tasks(void)
6395 struct prio_array *array;
6396 struct task_struct *g, *p;
6397 unsigned long flags;
6400 read_lock_irq(&tasklist_lock);
6402 do_each_thread(g, p) {
6406 spin_lock_irqsave(&p->pi_lock, flags);
6407 rq = __task_rq_lock(p);
6411 deactivate_task(p, task_rq(p));
6412 __setscheduler(p, SCHED_NORMAL, 0);
6414 __activate_task(p, task_rq(p));
6415 resched_task(rq->curr);
6418 __task_rq_unlock(rq);
6419 spin_unlock_irqrestore(&p->pi_lock, flags);
6420 } while_each_thread(g, p);
6422 read_unlock_irq(&tasklist_lock);
6425 #endif /* CONFIG_MAGIC_SYSRQ */
6429 * These functions are only useful for the IA64 MCA handling.
6431 * They can only be called when the whole system has been
6432 * stopped - every CPU needs to be quiescent, and no scheduling
6433 * activity can take place. Using them for anything else would
6434 * be a serious bug, and as a result, they aren't even visible
6435 * under any other configuration.
6439 * curr_task - return the current task for a given cpu.
6440 * @cpu: the processor in question.
6442 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6444 struct task_struct *curr_task(int cpu)
6446 return cpu_curr(cpu);
6450 * set_curr_task - set the current task for a given cpu.
6451 * @cpu: the processor in question.
6452 * @p: the task pointer to set.
6454 * Description: This function must only be used when non-maskable interrupts
6455 * are serviced on a separate stack. It allows the architecture to switch the
6456 * notion of the current task on a cpu in a non-blocking manner. This function
6457 * must be called with all CPU's synchronized, and interrupts disabled, the
6458 * and caller must save the original value of the current task (see
6459 * curr_task() above) and restore that value before reenabling interrupts and
6460 * re-starting the system.
6462 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6464 void set_curr_task(int cpu, struct task_struct *p)