4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 #define NICE_0_LOAD SCHED_LOAD_SCALE
95 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
98 * These are the 'tuning knobs' of the scheduler:
100 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
101 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
102 * Timeslices get refilled after they expire.
104 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
105 #define DEF_TIMESLICE (100 * HZ / 1000)
106 #define ON_RUNQUEUE_WEIGHT 30
107 #define CHILD_PENALTY 95
108 #define PARENT_PENALTY 100
109 #define EXIT_WEIGHT 3
110 #define PRIO_BONUS_RATIO 25
111 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
112 #define INTERACTIVE_DELTA 2
113 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
114 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
115 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
118 * If a task is 'interactive' then we reinsert it in the active
119 * array after it has expired its current timeslice. (it will not
120 * continue to run immediately, it will still roundrobin with
121 * other interactive tasks.)
123 * This part scales the interactivity limit depending on niceness.
125 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
126 * Here are a few examples of different nice levels:
128 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
129 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
130 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
131 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
132 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
134 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
135 * priority range a task can explore, a value of '1' means the
136 * task is rated interactive.)
138 * Ie. nice +19 tasks can never get 'interactive' enough to be
139 * reinserted into the active array. And only heavily CPU-hog nice -20
140 * tasks will be expired. Default nice 0 tasks are somewhere between,
141 * it takes some effort for them to get interactive, but it's not
145 #define CURRENT_BONUS(p) \
146 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
149 #define GRANULARITY (10 * HZ / 1000 ? : 1)
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
156 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
157 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
160 #define SCALE(v1,v1_max,v2_max) \
161 (v1) * (v2_max) / (v1_max)
164 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
167 #define TASK_INTERACTIVE(p) \
168 ((p)->prio <= (p)->static_prio - DELTA(p))
170 #define INTERACTIVE_SLEEP(p) \
171 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
172 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
174 #define TASK_PREEMPTS_CURR(p, rq) \
175 ((p)->prio < (rq)->curr->prio)
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio)
182 if (static_prio < NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
185 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
190 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
191 * Since cpu_power is a 'constant', we can use a reciprocal divide.
193 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
195 return reciprocal_divide(load, sg->reciprocal_cpu_power);
199 * Each time a sched group cpu_power is changed,
200 * we must compute its reciprocal value
202 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
204 sg->__cpu_power += val;
205 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
210 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
211 * to time slice values: [800ms ... 100ms ... 5ms]
213 * The higher a thread's priority, the bigger timeslices
214 * it gets during one round of execution. But even the lowest
215 * priority thread gets MIN_TIMESLICE worth of execution time.
218 static inline unsigned int task_timeslice(struct task_struct *p)
220 return static_prio_timeslice(p->static_prio);
224 * This is the priority-queue data structure of the RT scheduling class:
226 struct rt_prio_array {
227 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
228 struct list_head queue[MAX_RT_PRIO];
232 struct load_weight load;
233 u64 load_update_start, load_update_last;
234 unsigned long delta_fair, delta_exec, delta_stat;
237 /* CFS-related fields in a runqueue */
239 struct load_weight load;
240 unsigned long nr_running;
246 unsigned long wait_runtime_overruns, wait_runtime_underruns;
248 struct rb_root tasks_timeline;
249 struct rb_node *rb_leftmost;
250 struct rb_node *rb_load_balance_curr;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* 'curr' points to currently running entity on this cfs_rq.
253 * It is set to NULL otherwise (i.e when none are currently running).
255 struct sched_entity *curr;
256 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
258 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
259 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
260 * (like users, containers etc.)
262 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
263 * list is used during load balance.
265 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
269 /* Real-Time classes' related field in a runqueue: */
271 struct rt_prio_array active;
272 int rt_load_balance_idx;
273 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
277 * The prio-array type of the old scheduler:
280 unsigned int nr_active;
281 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
282 struct list_head queue[MAX_PRIO];
286 * This is the main, per-CPU runqueue data structure.
288 * Locking rule: those places that want to lock multiple runqueues
289 * (such as the load balancing or the thread migration code), lock
290 * acquire operations must be ordered by ascending &runqueue.
293 spinlock_t lock; /* runqueue lock */
296 * nr_running and cpu_load should be in the same cacheline because
297 * remote CPUs use both these fields when doing load calculation.
299 unsigned long nr_running;
300 unsigned long raw_weighted_load;
301 #define CPU_LOAD_IDX_MAX 5
302 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
303 unsigned char idle_at_tick;
305 unsigned char in_nohz_recently;
307 struct load_stat ls; /* capture load from *all* tasks on this cpu */
308 unsigned long nr_load_updates;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
318 * This is part of a global counter where only the total sum
319 * over all CPUs matters. A task can increase this counter on
320 * one CPU and if it got migrated afterwards it may decrease
321 * it on another CPU. Always updated under the runqueue lock:
323 unsigned long nr_uninterruptible;
325 unsigned long expired_timestamp;
326 unsigned long long most_recent_timestamp;
328 struct task_struct *curr, *idle;
329 unsigned long next_balance;
330 struct mm_struct *prev_mm;
332 struct prio_array *active, *expired, arrays[2];
333 int best_expired_prio;
335 u64 clock, prev_clock_raw;
338 unsigned int clock_warps, clock_overflows;
339 unsigned int clock_unstable_events;
341 struct sched_class *load_balance_class;
346 struct sched_domain *sd;
348 /* For active balancing */
351 int cpu; /* cpu of this runqueue */
353 struct task_struct *migration_thread;
354 struct list_head migration_queue;
357 #ifdef CONFIG_SCHEDSTATS
359 struct sched_info rq_sched_info;
361 /* sys_sched_yield() stats */
362 unsigned long yld_exp_empty;
363 unsigned long yld_act_empty;
364 unsigned long yld_both_empty;
365 unsigned long yld_cnt;
367 /* schedule() stats */
368 unsigned long sched_switch;
369 unsigned long sched_cnt;
370 unsigned long sched_goidle;
372 /* try_to_wake_up() stats */
373 unsigned long ttwu_cnt;
374 unsigned long ttwu_local;
376 struct lock_class_key rq_lock_key;
379 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
380 static DEFINE_MUTEX(sched_hotcpu_mutex);
382 static inline int cpu_of(struct rq *rq)
392 * Per-runqueue clock, as finegrained as the platform can give us:
394 static unsigned long long __rq_clock(struct rq *rq)
396 u64 prev_raw = rq->prev_clock_raw;
397 u64 now = sched_clock();
398 s64 delta = now - prev_raw;
399 u64 clock = rq->clock;
402 * Protect against sched_clock() occasionally going backwards:
404 if (unlikely(delta < 0)) {
409 * Catch too large forward jumps too:
411 if (unlikely(delta > 2*TICK_NSEC)) {
413 rq->clock_overflows++;
415 if (unlikely(delta > rq->clock_max_delta))
416 rq->clock_max_delta = delta;
421 rq->prev_clock_raw = now;
427 static inline unsigned long long rq_clock(struct rq *rq)
429 int this_cpu = smp_processor_id();
431 if (this_cpu == cpu_of(rq))
432 return __rq_clock(rq);
438 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
439 * See detach_destroy_domains: synchronize_sched for details.
441 * The domain tree of any CPU may only be accessed from within
442 * preempt-disabled sections.
444 #define for_each_domain(cpu, __sd) \
445 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
447 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
448 #define this_rq() (&__get_cpu_var(runqueues))
449 #define task_rq(p) cpu_rq(task_cpu(p))
450 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
452 #ifdef CONFIG_FAIR_GROUP_SCHED
453 /* Change a task's ->cfs_rq if it moves across CPUs */
454 static inline void set_task_cfs_rq(struct task_struct *p)
456 p->se.cfs_rq = &task_rq(p)->cfs;
459 static inline void set_task_cfs_rq(struct task_struct *p)
464 #ifndef prepare_arch_switch
465 # define prepare_arch_switch(next) do { } while (0)
467 #ifndef finish_arch_switch
468 # define finish_arch_switch(prev) do { } while (0)
471 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
472 static inline int task_running(struct rq *rq, struct task_struct *p)
474 return rq->curr == p;
477 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
481 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
483 #ifdef CONFIG_DEBUG_SPINLOCK
484 /* this is a valid case when another task releases the spinlock */
485 rq->lock.owner = current;
488 * If we are tracking spinlock dependencies then we have to
489 * fix up the runqueue lock - which gets 'carried over' from
492 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
494 spin_unlock_irq(&rq->lock);
497 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
498 static inline int task_running(struct rq *rq, struct task_struct *p)
503 return rq->curr == p;
507 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
511 * We can optimise this out completely for !SMP, because the
512 * SMP rebalancing from interrupt is the only thing that cares
517 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
518 spin_unlock_irq(&rq->lock);
520 spin_unlock(&rq->lock);
524 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
528 * After ->oncpu is cleared, the task can be moved to a different CPU.
529 * We must ensure this doesn't happen until the switch is completely
535 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
539 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
542 * __task_rq_lock - lock the runqueue a given task resides on.
543 * Must be called interrupts disabled.
545 static inline struct rq *__task_rq_lock(struct task_struct *p)
552 spin_lock(&rq->lock);
553 if (unlikely(rq != task_rq(p))) {
554 spin_unlock(&rq->lock);
555 goto repeat_lock_task;
561 * task_rq_lock - lock the runqueue a given task resides on and disable
562 * interrupts. Note the ordering: we can safely lookup the task_rq without
563 * explicitly disabling preemption.
565 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
571 local_irq_save(*flags);
573 spin_lock(&rq->lock);
574 if (unlikely(rq != task_rq(p))) {
575 spin_unlock_irqrestore(&rq->lock, *flags);
576 goto repeat_lock_task;
581 static inline void __task_rq_unlock(struct rq *rq)
584 spin_unlock(&rq->lock);
587 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
590 spin_unlock_irqrestore(&rq->lock, *flags);
594 * this_rq_lock - lock this runqueue and disable interrupts.
596 static inline struct rq *this_rq_lock(void)
603 spin_lock(&rq->lock);
608 #include "sched_stats.h"
611 * Adding/removing a task to/from a priority array:
613 static void dequeue_task(struct task_struct *p, struct prio_array *array)
616 list_del(&p->run_list);
617 if (list_empty(array->queue + p->prio))
618 __clear_bit(p->prio, array->bitmap);
621 static void enqueue_task(struct task_struct *p, struct prio_array *array)
623 sched_info_queued(p);
624 list_add_tail(&p->run_list, array->queue + p->prio);
625 __set_bit(p->prio, array->bitmap);
631 * Put task to the end of the run list without the overhead of dequeue
632 * followed by enqueue.
634 static void requeue_task(struct task_struct *p, struct prio_array *array)
636 list_move_tail(&p->run_list, array->queue + p->prio);
640 enqueue_task_head(struct task_struct *p, struct prio_array *array)
642 list_add(&p->run_list, array->queue + p->prio);
643 __set_bit(p->prio, array->bitmap);
649 * __normal_prio - return the priority that is based on the static
650 * priority but is modified by bonuses/penalties.
652 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
653 * into the -5 ... 0 ... +5 bonus/penalty range.
655 * We use 25% of the full 0...39 priority range so that:
657 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
658 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
660 * Both properties are important to certain workloads.
663 static inline int __normal_prio(struct task_struct *p)
667 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
669 prio = p->static_prio - bonus;
670 if (prio < MAX_RT_PRIO)
672 if (prio > MAX_PRIO-1)
678 * To aid in avoiding the subversion of "niceness" due to uneven distribution
679 * of tasks with abnormal "nice" values across CPUs the contribution that
680 * each task makes to its run queue's load is weighted according to its
681 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
682 * scaled version of the new time slice allocation that they receive on time
687 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
688 * If static_prio_timeslice() is ever changed to break this assumption then
689 * this code will need modification
691 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
692 #define LOAD_WEIGHT(lp) \
693 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
694 #define PRIO_TO_LOAD_WEIGHT(prio) \
695 LOAD_WEIGHT(static_prio_timeslice(prio))
696 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
697 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
699 static void set_load_weight(struct task_struct *p)
701 if (has_rt_policy(p)) {
703 if (p == task_rq(p)->migration_thread)
705 * The migration thread does the actual balancing.
706 * Giving its load any weight will skew balancing
712 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
714 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
718 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
720 rq->raw_weighted_load += p->load_weight;
724 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
726 rq->raw_weighted_load -= p->load_weight;
729 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
732 inc_raw_weighted_load(rq, p);
735 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
738 dec_raw_weighted_load(rq, p);
742 * Calculate the expected normal priority: i.e. priority
743 * without taking RT-inheritance into account. Might be
744 * boosted by interactivity modifiers. Changes upon fork,
745 * setprio syscalls, and whenever the interactivity
746 * estimator recalculates.
748 static inline int normal_prio(struct task_struct *p)
752 if (has_rt_policy(p))
753 prio = MAX_RT_PRIO-1 - p->rt_priority;
755 prio = __normal_prio(p);
760 * Calculate the current priority, i.e. the priority
761 * taken into account by the scheduler. This value might
762 * be boosted by RT tasks, or might be boosted by
763 * interactivity modifiers. Will be RT if the task got
764 * RT-boosted. If not then it returns p->normal_prio.
766 static int effective_prio(struct task_struct *p)
768 p->normal_prio = normal_prio(p);
770 * If we are RT tasks or we were boosted to RT priority,
771 * keep the priority unchanged. Otherwise, update priority
772 * to the normal priority:
774 if (!rt_prio(p->prio))
775 return p->normal_prio;
780 * __activate_task - move a task to the runqueue.
782 static void __activate_task(struct task_struct *p, struct rq *rq)
784 struct prio_array *target = rq->active;
787 target = rq->expired;
788 enqueue_task(p, target);
789 inc_nr_running(p, rq);
793 * __activate_idle_task - move idle task to the _front_ of runqueue.
795 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
797 enqueue_task_head(p, rq->active);
798 inc_nr_running(p, rq);
802 * Recalculate p->normal_prio and p->prio after having slept,
803 * updating the sleep-average too:
805 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
807 /* Caller must always ensure 'now >= p->timestamp' */
808 unsigned long sleep_time = now - p->timestamp;
813 if (likely(sleep_time > 0)) {
815 * This ceiling is set to the lowest priority that would allow
816 * a task to be reinserted into the active array on timeslice
819 unsigned long ceiling = INTERACTIVE_SLEEP(p);
821 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
823 * Prevents user tasks from achieving best priority
824 * with one single large enough sleep.
826 p->sleep_avg = ceiling;
828 * Using INTERACTIVE_SLEEP() as a ceiling places a
829 * nice(0) task 1ms sleep away from promotion, and
830 * gives it 700ms to round-robin with no chance of
831 * being demoted. This is more than generous, so
832 * mark this sleep as non-interactive to prevent the
833 * on-runqueue bonus logic from intervening should
834 * this task not receive cpu immediately.
836 p->sleep_type = SLEEP_NONINTERACTIVE;
839 * Tasks waking from uninterruptible sleep are
840 * limited in their sleep_avg rise as they
841 * are likely to be waiting on I/O
843 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
844 if (p->sleep_avg >= ceiling)
846 else if (p->sleep_avg + sleep_time >=
848 p->sleep_avg = ceiling;
854 * This code gives a bonus to interactive tasks.
856 * The boost works by updating the 'average sleep time'
857 * value here, based on ->timestamp. The more time a
858 * task spends sleeping, the higher the average gets -
859 * and the higher the priority boost gets as well.
861 p->sleep_avg += sleep_time;
864 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
865 p->sleep_avg = NS_MAX_SLEEP_AVG;
868 return effective_prio(p);
872 * activate_task - move a task to the runqueue and do priority recalculation
874 * Update all the scheduling statistics stuff. (sleep average
875 * calculation, priority modifiers, etc.)
877 static void activate_task(struct task_struct *p, struct rq *rq, int local)
879 unsigned long long now;
887 /* Compensate for drifting sched_clock */
888 struct rq *this_rq = this_rq();
889 now = (now - this_rq->most_recent_timestamp)
890 + rq->most_recent_timestamp;
895 * Sleep time is in units of nanosecs, so shift by 20 to get a
896 * milliseconds-range estimation of the amount of time that the task
899 if (unlikely(prof_on == SLEEP_PROFILING)) {
900 if (p->state == TASK_UNINTERRUPTIBLE)
901 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
902 (now - p->timestamp) >> 20);
905 p->prio = recalc_task_prio(p, now);
908 * This checks to make sure it's not an uninterruptible task
909 * that is now waking up.
911 if (p->sleep_type == SLEEP_NORMAL) {
913 * Tasks which were woken up by interrupts (ie. hw events)
914 * are most likely of interactive nature. So we give them
915 * the credit of extending their sleep time to the period
916 * of time they spend on the runqueue, waiting for execution
917 * on a CPU, first time around:
920 p->sleep_type = SLEEP_INTERRUPTED;
923 * Normal first-time wakeups get a credit too for
924 * on-runqueue time, but it will be weighted down:
926 p->sleep_type = SLEEP_INTERACTIVE;
931 __activate_task(p, rq);
935 * deactivate_task - remove a task from the runqueue.
937 static void deactivate_task(struct task_struct *p, struct rq *rq)
939 dec_nr_running(p, rq);
940 dequeue_task(p, p->array);
945 * resched_task - mark a task 'to be rescheduled now'.
947 * On UP this means the setting of the need_resched flag, on SMP it
948 * might also involve a cross-CPU call to trigger the scheduler on
953 #ifndef tsk_is_polling
954 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
957 static void resched_task(struct task_struct *p)
961 assert_spin_locked(&task_rq(p)->lock);
963 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
966 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
969 if (cpu == smp_processor_id())
972 /* NEED_RESCHED must be visible before we test polling */
974 if (!tsk_is_polling(p))
975 smp_send_reschedule(cpu);
978 static void resched_cpu(int cpu)
980 struct rq *rq = cpu_rq(cpu);
983 if (!spin_trylock_irqsave(&rq->lock, flags))
985 resched_task(cpu_curr(cpu));
986 spin_unlock_irqrestore(&rq->lock, flags);
989 static inline void resched_task(struct task_struct *p)
991 assert_spin_locked(&task_rq(p)->lock);
992 set_tsk_need_resched(p);
997 * task_curr - is this task currently executing on a CPU?
998 * @p: the task in question.
1000 inline int task_curr(const struct task_struct *p)
1002 return cpu_curr(task_cpu(p)) == p;
1005 /* Used instead of source_load when we know the type == 0 */
1006 unsigned long weighted_cpuload(const int cpu)
1008 return cpu_rq(cpu)->raw_weighted_load;
1013 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1015 task_thread_info(p)->cpu = cpu;
1018 struct migration_req {
1019 struct list_head list;
1021 struct task_struct *task;
1024 struct completion done;
1028 * The task's runqueue lock must be held.
1029 * Returns true if you have to wait for migration thread.
1032 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1034 struct rq *rq = task_rq(p);
1037 * If the task is not on a runqueue (and not running), then
1038 * it is sufficient to simply update the task's cpu field.
1040 if (!p->array && !task_running(rq, p)) {
1041 set_task_cpu(p, dest_cpu);
1045 init_completion(&req->done);
1047 req->dest_cpu = dest_cpu;
1048 list_add(&req->list, &rq->migration_queue);
1054 * wait_task_inactive - wait for a thread to unschedule.
1056 * The caller must ensure that the task *will* unschedule sometime soon,
1057 * else this function might spin for a *long* time. This function can't
1058 * be called with interrupts off, or it may introduce deadlock with
1059 * smp_call_function() if an IPI is sent by the same process we are
1060 * waiting to become inactive.
1062 void wait_task_inactive(struct task_struct *p)
1064 unsigned long flags;
1066 struct prio_array *array;
1071 * We do the initial early heuristics without holding
1072 * any task-queue locks at all. We'll only try to get
1073 * the runqueue lock when things look like they will
1079 * If the task is actively running on another CPU
1080 * still, just relax and busy-wait without holding
1083 * NOTE! Since we don't hold any locks, it's not
1084 * even sure that "rq" stays as the right runqueue!
1085 * But we don't care, since "task_running()" will
1086 * return false if the runqueue has changed and p
1087 * is actually now running somewhere else!
1089 while (task_running(rq, p))
1093 * Ok, time to look more closely! We need the rq
1094 * lock now, to be *sure*. If we're wrong, we'll
1095 * just go back and repeat.
1097 rq = task_rq_lock(p, &flags);
1098 running = task_running(rq, p);
1100 task_rq_unlock(rq, &flags);
1103 * Was it really running after all now that we
1104 * checked with the proper locks actually held?
1106 * Oops. Go back and try again..
1108 if (unlikely(running)) {
1114 * It's not enough that it's not actively running,
1115 * it must be off the runqueue _entirely_, and not
1118 * So if it wa still runnable (but just not actively
1119 * running right now), it's preempted, and we should
1120 * yield - it could be a while.
1122 if (unlikely(array)) {
1128 * Ahh, all good. It wasn't running, and it wasn't
1129 * runnable, which means that it will never become
1130 * running in the future either. We're all done!
1135 * kick_process - kick a running thread to enter/exit the kernel
1136 * @p: the to-be-kicked thread
1138 * Cause a process which is running on another CPU to enter
1139 * kernel-mode, without any delay. (to get signals handled.)
1141 * NOTE: this function doesnt have to take the runqueue lock,
1142 * because all it wants to ensure is that the remote task enters
1143 * the kernel. If the IPI races and the task has been migrated
1144 * to another CPU then no harm is done and the purpose has been
1147 void kick_process(struct task_struct *p)
1153 if ((cpu != smp_processor_id()) && task_curr(p))
1154 smp_send_reschedule(cpu);
1159 * Return a low guess at the load of a migration-source cpu weighted
1160 * according to the scheduling class and "nice" value.
1162 * We want to under-estimate the load of migration sources, to
1163 * balance conservatively.
1165 static inline unsigned long source_load(int cpu, int type)
1167 struct rq *rq = cpu_rq(cpu);
1170 return rq->raw_weighted_load;
1172 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1176 * Return a high guess at the load of a migration-target cpu weighted
1177 * according to the scheduling class and "nice" value.
1179 static inline unsigned long target_load(int cpu, int type)
1181 struct rq *rq = cpu_rq(cpu);
1184 return rq->raw_weighted_load;
1186 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1190 * Return the average load per task on the cpu's run queue
1192 static inline unsigned long cpu_avg_load_per_task(int cpu)
1194 struct rq *rq = cpu_rq(cpu);
1195 unsigned long n = rq->nr_running;
1197 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1201 * find_idlest_group finds and returns the least busy CPU group within the
1204 static struct sched_group *
1205 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1207 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1208 unsigned long min_load = ULONG_MAX, this_load = 0;
1209 int load_idx = sd->forkexec_idx;
1210 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1213 unsigned long load, avg_load;
1217 /* Skip over this group if it has no CPUs allowed */
1218 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1221 local_group = cpu_isset(this_cpu, group->cpumask);
1223 /* Tally up the load of all CPUs in the group */
1226 for_each_cpu_mask(i, group->cpumask) {
1227 /* Bias balancing toward cpus of our domain */
1229 load = source_load(i, load_idx);
1231 load = target_load(i, load_idx);
1236 /* Adjust by relative CPU power of the group */
1237 avg_load = sg_div_cpu_power(group,
1238 avg_load * SCHED_LOAD_SCALE);
1241 this_load = avg_load;
1243 } else if (avg_load < min_load) {
1244 min_load = avg_load;
1248 group = group->next;
1249 } while (group != sd->groups);
1251 if (!idlest || 100*this_load < imbalance*min_load)
1257 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1260 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1263 unsigned long load, min_load = ULONG_MAX;
1267 /* Traverse only the allowed CPUs */
1268 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1270 for_each_cpu_mask(i, tmp) {
1271 load = weighted_cpuload(i);
1273 if (load < min_load || (load == min_load && i == this_cpu)) {
1283 * sched_balance_self: balance the current task (running on cpu) in domains
1284 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1287 * Balance, ie. select the least loaded group.
1289 * Returns the target CPU number, or the same CPU if no balancing is needed.
1291 * preempt must be disabled.
1293 static int sched_balance_self(int cpu, int flag)
1295 struct task_struct *t = current;
1296 struct sched_domain *tmp, *sd = NULL;
1298 for_each_domain(cpu, tmp) {
1300 * If power savings logic is enabled for a domain, stop there.
1302 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1304 if (tmp->flags & flag)
1310 struct sched_group *group;
1311 int new_cpu, weight;
1313 if (!(sd->flags & flag)) {
1319 group = find_idlest_group(sd, t, cpu);
1325 new_cpu = find_idlest_cpu(group, t, cpu);
1326 if (new_cpu == -1 || new_cpu == cpu) {
1327 /* Now try balancing at a lower domain level of cpu */
1332 /* Now try balancing at a lower domain level of new_cpu */
1335 weight = cpus_weight(span);
1336 for_each_domain(cpu, tmp) {
1337 if (weight <= cpus_weight(tmp->span))
1339 if (tmp->flags & flag)
1342 /* while loop will break here if sd == NULL */
1348 #endif /* CONFIG_SMP */
1351 * wake_idle() will wake a task on an idle cpu if task->cpu is
1352 * not idle and an idle cpu is available. The span of cpus to
1353 * search starts with cpus closest then further out as needed,
1354 * so we always favor a closer, idle cpu.
1356 * Returns the CPU we should wake onto.
1358 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1359 static int wake_idle(int cpu, struct task_struct *p)
1362 struct sched_domain *sd;
1366 * If it is idle, then it is the best cpu to run this task.
1368 * This cpu is also the best, if it has more than one task already.
1369 * Siblings must be also busy(in most cases) as they didn't already
1370 * pickup the extra load from this cpu and hence we need not check
1371 * sibling runqueue info. This will avoid the checks and cache miss
1372 * penalities associated with that.
1374 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1377 for_each_domain(cpu, sd) {
1378 if (sd->flags & SD_WAKE_IDLE) {
1379 cpus_and(tmp, sd->span, p->cpus_allowed);
1380 for_each_cpu_mask(i, tmp) {
1391 static inline int wake_idle(int cpu, struct task_struct *p)
1398 * try_to_wake_up - wake up a thread
1399 * @p: the to-be-woken-up thread
1400 * @state: the mask of task states that can be woken
1401 * @sync: do a synchronous wakeup?
1403 * Put it on the run-queue if it's not already there. The "current"
1404 * thread is always on the run-queue (except when the actual
1405 * re-schedule is in progress), and as such you're allowed to do
1406 * the simpler "current->state = TASK_RUNNING" to mark yourself
1407 * runnable without the overhead of this.
1409 * returns failure only if the task is already active.
1411 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1413 int cpu, this_cpu, success = 0;
1414 unsigned long flags;
1418 struct sched_domain *sd, *this_sd = NULL;
1419 unsigned long load, this_load;
1423 rq = task_rq_lock(p, &flags);
1424 old_state = p->state;
1425 if (!(old_state & state))
1432 this_cpu = smp_processor_id();
1435 if (unlikely(task_running(rq, p)))
1440 schedstat_inc(rq, ttwu_cnt);
1441 if (cpu == this_cpu) {
1442 schedstat_inc(rq, ttwu_local);
1446 for_each_domain(this_cpu, sd) {
1447 if (cpu_isset(cpu, sd->span)) {
1448 schedstat_inc(sd, ttwu_wake_remote);
1454 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1458 * Check for affine wakeup and passive balancing possibilities.
1461 int idx = this_sd->wake_idx;
1462 unsigned int imbalance;
1464 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1466 load = source_load(cpu, idx);
1467 this_load = target_load(this_cpu, idx);
1469 new_cpu = this_cpu; /* Wake to this CPU if we can */
1471 if (this_sd->flags & SD_WAKE_AFFINE) {
1472 unsigned long tl = this_load;
1473 unsigned long tl_per_task;
1475 tl_per_task = cpu_avg_load_per_task(this_cpu);
1478 * If sync wakeup then subtract the (maximum possible)
1479 * effect of the currently running task from the load
1480 * of the current CPU:
1483 tl -= current->load_weight;
1486 tl + target_load(cpu, idx) <= tl_per_task) ||
1487 100*(tl + p->load_weight) <= imbalance*load) {
1489 * This domain has SD_WAKE_AFFINE and
1490 * p is cache cold in this domain, and
1491 * there is no bad imbalance.
1493 schedstat_inc(this_sd, ttwu_move_affine);
1499 * Start passive balancing when half the imbalance_pct
1502 if (this_sd->flags & SD_WAKE_BALANCE) {
1503 if (imbalance*this_load <= 100*load) {
1504 schedstat_inc(this_sd, ttwu_move_balance);
1510 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1512 new_cpu = wake_idle(new_cpu, p);
1513 if (new_cpu != cpu) {
1514 set_task_cpu(p, new_cpu);
1515 task_rq_unlock(rq, &flags);
1516 /* might preempt at this point */
1517 rq = task_rq_lock(p, &flags);
1518 old_state = p->state;
1519 if (!(old_state & state))
1524 this_cpu = smp_processor_id();
1529 #endif /* CONFIG_SMP */
1530 if (old_state == TASK_UNINTERRUPTIBLE) {
1531 rq->nr_uninterruptible--;
1533 * Tasks on involuntary sleep don't earn
1534 * sleep_avg beyond just interactive state.
1536 p->sleep_type = SLEEP_NONINTERACTIVE;
1540 * Tasks that have marked their sleep as noninteractive get
1541 * woken up with their sleep average not weighted in an
1544 if (old_state & TASK_NONINTERACTIVE)
1545 p->sleep_type = SLEEP_NONINTERACTIVE;
1548 activate_task(p, rq, cpu == this_cpu);
1550 * Sync wakeups (i.e. those types of wakeups where the waker
1551 * has indicated that it will leave the CPU in short order)
1552 * don't trigger a preemption, if the woken up task will run on
1553 * this cpu. (in this case the 'I will reschedule' promise of
1554 * the waker guarantees that the freshly woken up task is going
1555 * to be considered on this CPU.)
1557 if (!sync || cpu != this_cpu) {
1558 if (TASK_PREEMPTS_CURR(p, rq))
1559 resched_task(rq->curr);
1564 p->state = TASK_RUNNING;
1566 task_rq_unlock(rq, &flags);
1571 int fastcall wake_up_process(struct task_struct *p)
1573 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1574 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1576 EXPORT_SYMBOL(wake_up_process);
1578 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1580 return try_to_wake_up(p, state, 0);
1583 static void task_running_tick(struct rq *rq, struct task_struct *p);
1585 * Perform scheduler related setup for a newly forked process p.
1586 * p is forked by current.
1588 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1590 int cpu = get_cpu();
1593 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1595 set_task_cpu(p, cpu);
1598 * We mark the process as running here, but have not actually
1599 * inserted it onto the runqueue yet. This guarantees that
1600 * nobody will actually run it, and a signal or other external
1601 * event cannot wake it up and insert it on the runqueue either.
1603 p->state = TASK_RUNNING;
1606 * Make sure we do not leak PI boosting priority to the child:
1608 p->prio = current->normal_prio;
1610 INIT_LIST_HEAD(&p->run_list);
1612 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1613 if (unlikely(sched_info_on()))
1614 memset(&p->sched_info, 0, sizeof(p->sched_info));
1616 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1619 #ifdef CONFIG_PREEMPT
1620 /* Want to start with kernel preemption disabled. */
1621 task_thread_info(p)->preempt_count = 1;
1624 * Share the timeslice between parent and child, thus the
1625 * total amount of pending timeslices in the system doesn't change,
1626 * resulting in more scheduling fairness.
1628 local_irq_disable();
1629 p->time_slice = (current->time_slice + 1) >> 1;
1631 * The remainder of the first timeslice might be recovered by
1632 * the parent if the child exits early enough.
1634 p->first_time_slice = 1;
1635 current->time_slice >>= 1;
1636 p->timestamp = sched_clock();
1637 if (unlikely(!current->time_slice)) {
1639 * This case is rare, it happens when the parent has only
1640 * a single jiffy left from its timeslice. Taking the
1641 * runqueue lock is not a problem.
1643 current->time_slice = 1;
1644 task_running_tick(cpu_rq(cpu), current);
1651 * wake_up_new_task - wake up a newly created task for the first time.
1653 * This function will do some initial scheduler statistics housekeeping
1654 * that must be done for every newly created context, then puts the task
1655 * on the runqueue and wakes it.
1657 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1659 struct rq *rq, *this_rq;
1660 unsigned long flags;
1663 rq = task_rq_lock(p, &flags);
1664 BUG_ON(p->state != TASK_RUNNING);
1665 this_cpu = smp_processor_id();
1669 * We decrease the sleep average of forking parents
1670 * and children as well, to keep max-interactive tasks
1671 * from forking tasks that are max-interactive. The parent
1672 * (current) is done further down, under its lock.
1674 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1675 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1677 p->prio = effective_prio(p);
1679 if (likely(cpu == this_cpu)) {
1680 if (!(clone_flags & CLONE_VM)) {
1682 * The VM isn't cloned, so we're in a good position to
1683 * do child-runs-first in anticipation of an exec. This
1684 * usually avoids a lot of COW overhead.
1686 if (unlikely(!current->array))
1687 __activate_task(p, rq);
1689 p->prio = current->prio;
1690 p->normal_prio = current->normal_prio;
1691 list_add_tail(&p->run_list, ¤t->run_list);
1692 p->array = current->array;
1693 p->array->nr_active++;
1694 inc_nr_running(p, rq);
1698 /* Run child last */
1699 __activate_task(p, rq);
1701 * We skip the following code due to cpu == this_cpu
1703 * task_rq_unlock(rq, &flags);
1704 * this_rq = task_rq_lock(current, &flags);
1708 this_rq = cpu_rq(this_cpu);
1711 * Not the local CPU - must adjust timestamp. This should
1712 * get optimised away in the !CONFIG_SMP case.
1714 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1715 + rq->most_recent_timestamp;
1716 __activate_task(p, rq);
1717 if (TASK_PREEMPTS_CURR(p, rq))
1718 resched_task(rq->curr);
1721 * Parent and child are on different CPUs, now get the
1722 * parent runqueue to update the parent's ->sleep_avg:
1724 task_rq_unlock(rq, &flags);
1725 this_rq = task_rq_lock(current, &flags);
1727 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1728 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1729 task_rq_unlock(this_rq, &flags);
1733 * prepare_task_switch - prepare to switch tasks
1734 * @rq: the runqueue preparing to switch
1735 * @next: the task we are going to switch to.
1737 * This is called with the rq lock held and interrupts off. It must
1738 * be paired with a subsequent finish_task_switch after the context
1741 * prepare_task_switch sets up locking and calls architecture specific
1744 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1746 prepare_lock_switch(rq, next);
1747 prepare_arch_switch(next);
1751 * finish_task_switch - clean up after a task-switch
1752 * @rq: runqueue associated with task-switch
1753 * @prev: the thread we just switched away from.
1755 * finish_task_switch must be called after the context switch, paired
1756 * with a prepare_task_switch call before the context switch.
1757 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1758 * and do any other architecture-specific cleanup actions.
1760 * Note that we may have delayed dropping an mm in context_switch(). If
1761 * so, we finish that here outside of the runqueue lock. (Doing it
1762 * with the lock held can cause deadlocks; see schedule() for
1765 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1766 __releases(rq->lock)
1768 struct mm_struct *mm = rq->prev_mm;
1774 * A task struct has one reference for the use as "current".
1775 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1776 * schedule one last time. The schedule call will never return, and
1777 * the scheduled task must drop that reference.
1778 * The test for TASK_DEAD must occur while the runqueue locks are
1779 * still held, otherwise prev could be scheduled on another cpu, die
1780 * there before we look at prev->state, and then the reference would
1782 * Manfred Spraul <manfred@colorfullife.com>
1784 prev_state = prev->state;
1785 finish_arch_switch(prev);
1786 finish_lock_switch(rq, prev);
1789 if (unlikely(prev_state == TASK_DEAD)) {
1791 * Remove function-return probe instances associated with this
1792 * task and put them back on the free list.
1794 kprobe_flush_task(prev);
1795 put_task_struct(prev);
1800 * schedule_tail - first thing a freshly forked thread must call.
1801 * @prev: the thread we just switched away from.
1803 asmlinkage void schedule_tail(struct task_struct *prev)
1804 __releases(rq->lock)
1806 struct rq *rq = this_rq();
1808 finish_task_switch(rq, prev);
1809 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1810 /* In this case, finish_task_switch does not reenable preemption */
1813 if (current->set_child_tid)
1814 put_user(current->pid, current->set_child_tid);
1818 * context_switch - switch to the new MM and the new
1819 * thread's register state.
1821 static inline struct task_struct *
1822 context_switch(struct rq *rq, struct task_struct *prev,
1823 struct task_struct *next)
1825 struct mm_struct *mm = next->mm;
1826 struct mm_struct *oldmm = prev->active_mm;
1829 * For paravirt, this is coupled with an exit in switch_to to
1830 * combine the page table reload and the switch backend into
1833 arch_enter_lazy_cpu_mode();
1836 next->active_mm = oldmm;
1837 atomic_inc(&oldmm->mm_count);
1838 enter_lazy_tlb(oldmm, next);
1840 switch_mm(oldmm, mm, next);
1843 prev->active_mm = NULL;
1844 WARN_ON(rq->prev_mm);
1845 rq->prev_mm = oldmm;
1848 * Since the runqueue lock will be released by the next
1849 * task (which is an invalid locking op but in the case
1850 * of the scheduler it's an obvious special-case), so we
1851 * do an early lockdep release here:
1853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1854 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1857 /* Here we just switch the register state and the stack. */
1858 switch_to(prev, next, prev);
1864 * nr_running, nr_uninterruptible and nr_context_switches:
1866 * externally visible scheduler statistics: current number of runnable
1867 * threads, current number of uninterruptible-sleeping threads, total
1868 * number of context switches performed since bootup.
1870 unsigned long nr_running(void)
1872 unsigned long i, sum = 0;
1874 for_each_online_cpu(i)
1875 sum += cpu_rq(i)->nr_running;
1880 unsigned long nr_uninterruptible(void)
1882 unsigned long i, sum = 0;
1884 for_each_possible_cpu(i)
1885 sum += cpu_rq(i)->nr_uninterruptible;
1888 * Since we read the counters lockless, it might be slightly
1889 * inaccurate. Do not allow it to go below zero though:
1891 if (unlikely((long)sum < 0))
1897 unsigned long long nr_context_switches(void)
1900 unsigned long long sum = 0;
1902 for_each_possible_cpu(i)
1903 sum += cpu_rq(i)->nr_switches;
1908 unsigned long nr_iowait(void)
1910 unsigned long i, sum = 0;
1912 for_each_possible_cpu(i)
1913 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1918 unsigned long nr_active(void)
1920 unsigned long i, running = 0, uninterruptible = 0;
1922 for_each_online_cpu(i) {
1923 running += cpu_rq(i)->nr_running;
1924 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1927 if (unlikely((long)uninterruptible < 0))
1928 uninterruptible = 0;
1930 return running + uninterruptible;
1936 * Is this task likely cache-hot:
1939 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1941 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1945 * double_rq_lock - safely lock two runqueues
1947 * Note this does not disable interrupts like task_rq_lock,
1948 * you need to do so manually before calling.
1950 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1951 __acquires(rq1->lock)
1952 __acquires(rq2->lock)
1954 BUG_ON(!irqs_disabled());
1956 spin_lock(&rq1->lock);
1957 __acquire(rq2->lock); /* Fake it out ;) */
1960 spin_lock(&rq1->lock);
1961 spin_lock(&rq2->lock);
1963 spin_lock(&rq2->lock);
1964 spin_lock(&rq1->lock);
1970 * double_rq_unlock - safely unlock two runqueues
1972 * Note this does not restore interrupts like task_rq_unlock,
1973 * you need to do so manually after calling.
1975 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1976 __releases(rq1->lock)
1977 __releases(rq2->lock)
1979 spin_unlock(&rq1->lock);
1981 spin_unlock(&rq2->lock);
1983 __release(rq2->lock);
1987 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1989 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1990 __releases(this_rq->lock)
1991 __acquires(busiest->lock)
1992 __acquires(this_rq->lock)
1994 if (unlikely(!irqs_disabled())) {
1995 /* printk() doesn't work good under rq->lock */
1996 spin_unlock(&this_rq->lock);
1999 if (unlikely(!spin_trylock(&busiest->lock))) {
2000 if (busiest < this_rq) {
2001 spin_unlock(&this_rq->lock);
2002 spin_lock(&busiest->lock);
2003 spin_lock(&this_rq->lock);
2005 spin_lock(&busiest->lock);
2010 * If dest_cpu is allowed for this process, migrate the task to it.
2011 * This is accomplished by forcing the cpu_allowed mask to only
2012 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2013 * the cpu_allowed mask is restored.
2015 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2017 struct migration_req req;
2018 unsigned long flags;
2021 rq = task_rq_lock(p, &flags);
2022 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2023 || unlikely(cpu_is_offline(dest_cpu)))
2026 /* force the process onto the specified CPU */
2027 if (migrate_task(p, dest_cpu, &req)) {
2028 /* Need to wait for migration thread (might exit: take ref). */
2029 struct task_struct *mt = rq->migration_thread;
2031 get_task_struct(mt);
2032 task_rq_unlock(rq, &flags);
2033 wake_up_process(mt);
2034 put_task_struct(mt);
2035 wait_for_completion(&req.done);
2040 task_rq_unlock(rq, &flags);
2044 * sched_exec - execve() is a valuable balancing opportunity, because at
2045 * this point the task has the smallest effective memory and cache footprint.
2047 void sched_exec(void)
2049 int new_cpu, this_cpu = get_cpu();
2050 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2052 if (new_cpu != this_cpu)
2053 sched_migrate_task(current, new_cpu);
2057 * pull_task - move a task from a remote runqueue to the local runqueue.
2058 * Both runqueues must be locked.
2060 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2061 struct task_struct *p, struct rq *this_rq,
2062 struct prio_array *this_array, int this_cpu)
2064 dequeue_task(p, src_array);
2065 dec_nr_running(p, src_rq);
2066 set_task_cpu(p, this_cpu);
2067 inc_nr_running(p, this_rq);
2068 enqueue_task(p, this_array);
2069 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2070 + this_rq->most_recent_timestamp;
2072 * Note that idle threads have a prio of MAX_PRIO, for this test
2073 * to be always true for them.
2075 if (TASK_PREEMPTS_CURR(p, this_rq))
2076 resched_task(this_rq->curr);
2080 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2083 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2084 struct sched_domain *sd, enum cpu_idle_type idle,
2088 * We do not migrate tasks that are:
2089 * 1) running (obviously), or
2090 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2091 * 3) are cache-hot on their current CPU.
2093 if (!cpu_isset(this_cpu, p->cpus_allowed))
2097 if (task_running(rq, p))
2101 * Aggressive migration if:
2102 * 1) task is cache cold, or
2103 * 2) too many balance attempts have failed.
2106 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2107 #ifdef CONFIG_SCHEDSTATS
2108 if (task_hot(p, rq->most_recent_timestamp, sd))
2109 schedstat_inc(sd, lb_hot_gained[idle]);
2114 if (task_hot(p, rq->most_recent_timestamp, sd))
2119 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2122 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2123 * load from busiest to this_rq, as part of a balancing operation within
2124 * "domain". Returns the number of tasks moved.
2126 * Called with both runqueues locked.
2128 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2129 unsigned long max_nr_move, unsigned long max_load_move,
2130 struct sched_domain *sd, enum cpu_idle_type idle,
2133 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2134 best_prio_seen, skip_for_load;
2135 struct prio_array *array, *dst_array;
2136 struct list_head *head, *curr;
2137 struct task_struct *tmp;
2140 if (max_nr_move == 0 || max_load_move == 0)
2143 rem_load_move = max_load_move;
2145 this_best_prio = rq_best_prio(this_rq);
2146 best_prio = rq_best_prio(busiest);
2148 * Enable handling of the case where there is more than one task
2149 * with the best priority. If the current running task is one
2150 * of those with prio==best_prio we know it won't be moved
2151 * and therefore it's safe to override the skip (based on load) of
2152 * any task we find with that prio.
2154 best_prio_seen = best_prio == busiest->curr->prio;
2157 * We first consider expired tasks. Those will likely not be
2158 * executed in the near future, and they are most likely to
2159 * be cache-cold, thus switching CPUs has the least effect
2162 if (busiest->expired->nr_active) {
2163 array = busiest->expired;
2164 dst_array = this_rq->expired;
2166 array = busiest->active;
2167 dst_array = this_rq->active;
2171 /* Start searching at priority 0: */
2175 idx = sched_find_first_bit(array->bitmap);
2177 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2178 if (idx >= MAX_PRIO) {
2179 if (array == busiest->expired && busiest->active->nr_active) {
2180 array = busiest->active;
2181 dst_array = this_rq->active;
2187 head = array->queue + idx;
2190 tmp = list_entry(curr, struct task_struct, run_list);
2195 * To help distribute high priority tasks accross CPUs we don't
2196 * skip a task if it will be the highest priority task (i.e. smallest
2197 * prio value) on its new queue regardless of its load weight
2199 skip_for_load = tmp->load_weight > rem_load_move;
2200 if (skip_for_load && idx < this_best_prio)
2201 skip_for_load = !best_prio_seen && idx == best_prio;
2202 if (skip_for_load ||
2203 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2205 best_prio_seen |= idx == best_prio;
2212 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2214 rem_load_move -= tmp->load_weight;
2217 * We only want to steal up to the prescribed number of tasks
2218 * and the prescribed amount of weighted load.
2220 if (pulled < max_nr_move && rem_load_move > 0) {
2221 if (idx < this_best_prio)
2222 this_best_prio = idx;
2230 * Right now, this is the only place pull_task() is called,
2231 * so we can safely collect pull_task() stats here rather than
2232 * inside pull_task().
2234 schedstat_add(sd, lb_gained[idle], pulled);
2237 *all_pinned = pinned;
2242 * find_busiest_group finds and returns the busiest CPU group within the
2243 * domain. It calculates and returns the amount of weighted load which
2244 * should be moved to restore balance via the imbalance parameter.
2246 static struct sched_group *
2247 find_busiest_group(struct sched_domain *sd, int this_cpu,
2248 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2249 cpumask_t *cpus, int *balance)
2251 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2252 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2253 unsigned long max_pull;
2254 unsigned long busiest_load_per_task, busiest_nr_running;
2255 unsigned long this_load_per_task, this_nr_running;
2257 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2258 int power_savings_balance = 1;
2259 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2260 unsigned long min_nr_running = ULONG_MAX;
2261 struct sched_group *group_min = NULL, *group_leader = NULL;
2264 max_load = this_load = total_load = total_pwr = 0;
2265 busiest_load_per_task = busiest_nr_running = 0;
2266 this_load_per_task = this_nr_running = 0;
2267 if (idle == CPU_NOT_IDLE)
2268 load_idx = sd->busy_idx;
2269 else if (idle == CPU_NEWLY_IDLE)
2270 load_idx = sd->newidle_idx;
2272 load_idx = sd->idle_idx;
2275 unsigned long load, group_capacity;
2278 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2279 unsigned long sum_nr_running, sum_weighted_load;
2281 local_group = cpu_isset(this_cpu, group->cpumask);
2284 balance_cpu = first_cpu(group->cpumask);
2286 /* Tally up the load of all CPUs in the group */
2287 sum_weighted_load = sum_nr_running = avg_load = 0;
2289 for_each_cpu_mask(i, group->cpumask) {
2292 if (!cpu_isset(i, *cpus))
2297 if (*sd_idle && !idle_cpu(i))
2300 /* Bias balancing toward cpus of our domain */
2302 if (idle_cpu(i) && !first_idle_cpu) {
2307 load = target_load(i, load_idx);
2309 load = source_load(i, load_idx);
2312 sum_nr_running += rq->nr_running;
2313 sum_weighted_load += rq->raw_weighted_load;
2317 * First idle cpu or the first cpu(busiest) in this sched group
2318 * is eligible for doing load balancing at this and above
2321 if (local_group && balance_cpu != this_cpu && balance) {
2326 total_load += avg_load;
2327 total_pwr += group->__cpu_power;
2329 /* Adjust by relative CPU power of the group */
2330 avg_load = sg_div_cpu_power(group,
2331 avg_load * SCHED_LOAD_SCALE);
2333 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2336 this_load = avg_load;
2338 this_nr_running = sum_nr_running;
2339 this_load_per_task = sum_weighted_load;
2340 } else if (avg_load > max_load &&
2341 sum_nr_running > group_capacity) {
2342 max_load = avg_load;
2344 busiest_nr_running = sum_nr_running;
2345 busiest_load_per_task = sum_weighted_load;
2348 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2350 * Busy processors will not participate in power savings
2353 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2357 * If the local group is idle or completely loaded
2358 * no need to do power savings balance at this domain
2360 if (local_group && (this_nr_running >= group_capacity ||
2362 power_savings_balance = 0;
2365 * If a group is already running at full capacity or idle,
2366 * don't include that group in power savings calculations
2368 if (!power_savings_balance || sum_nr_running >= group_capacity
2373 * Calculate the group which has the least non-idle load.
2374 * This is the group from where we need to pick up the load
2377 if ((sum_nr_running < min_nr_running) ||
2378 (sum_nr_running == min_nr_running &&
2379 first_cpu(group->cpumask) <
2380 first_cpu(group_min->cpumask))) {
2382 min_nr_running = sum_nr_running;
2383 min_load_per_task = sum_weighted_load /
2388 * Calculate the group which is almost near its
2389 * capacity but still has some space to pick up some load
2390 * from other group and save more power
2392 if (sum_nr_running <= group_capacity - 1) {
2393 if (sum_nr_running > leader_nr_running ||
2394 (sum_nr_running == leader_nr_running &&
2395 first_cpu(group->cpumask) >
2396 first_cpu(group_leader->cpumask))) {
2397 group_leader = group;
2398 leader_nr_running = sum_nr_running;
2403 group = group->next;
2404 } while (group != sd->groups);
2406 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2409 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2411 if (this_load >= avg_load ||
2412 100*max_load <= sd->imbalance_pct*this_load)
2415 busiest_load_per_task /= busiest_nr_running;
2417 * We're trying to get all the cpus to the average_load, so we don't
2418 * want to push ourselves above the average load, nor do we wish to
2419 * reduce the max loaded cpu below the average load, as either of these
2420 * actions would just result in more rebalancing later, and ping-pong
2421 * tasks around. Thus we look for the minimum possible imbalance.
2422 * Negative imbalances (*we* are more loaded than anyone else) will
2423 * be counted as no imbalance for these purposes -- we can't fix that
2424 * by pulling tasks to us. Be careful of negative numbers as they'll
2425 * appear as very large values with unsigned longs.
2427 if (max_load <= busiest_load_per_task)
2431 * In the presence of smp nice balancing, certain scenarios can have
2432 * max load less than avg load(as we skip the groups at or below
2433 * its cpu_power, while calculating max_load..)
2435 if (max_load < avg_load) {
2437 goto small_imbalance;
2440 /* Don't want to pull so many tasks that a group would go idle */
2441 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2443 /* How much load to actually move to equalise the imbalance */
2444 *imbalance = min(max_pull * busiest->__cpu_power,
2445 (avg_load - this_load) * this->__cpu_power)
2449 * if *imbalance is less than the average load per runnable task
2450 * there is no gaurantee that any tasks will be moved so we'll have
2451 * a think about bumping its value to force at least one task to be
2454 if (*imbalance < busiest_load_per_task) {
2455 unsigned long tmp, pwr_now, pwr_move;
2459 pwr_move = pwr_now = 0;
2461 if (this_nr_running) {
2462 this_load_per_task /= this_nr_running;
2463 if (busiest_load_per_task > this_load_per_task)
2466 this_load_per_task = SCHED_LOAD_SCALE;
2468 if (max_load - this_load >= busiest_load_per_task * imbn) {
2469 *imbalance = busiest_load_per_task;
2474 * OK, we don't have enough imbalance to justify moving tasks,
2475 * however we may be able to increase total CPU power used by
2479 pwr_now += busiest->__cpu_power *
2480 min(busiest_load_per_task, max_load);
2481 pwr_now += this->__cpu_power *
2482 min(this_load_per_task, this_load);
2483 pwr_now /= SCHED_LOAD_SCALE;
2485 /* Amount of load we'd subtract */
2486 tmp = sg_div_cpu_power(busiest,
2487 busiest_load_per_task * SCHED_LOAD_SCALE);
2489 pwr_move += busiest->__cpu_power *
2490 min(busiest_load_per_task, max_load - tmp);
2492 /* Amount of load we'd add */
2493 if (max_load * busiest->__cpu_power <
2494 busiest_load_per_task * SCHED_LOAD_SCALE)
2495 tmp = sg_div_cpu_power(this,
2496 max_load * busiest->__cpu_power);
2498 tmp = sg_div_cpu_power(this,
2499 busiest_load_per_task * SCHED_LOAD_SCALE);
2500 pwr_move += this->__cpu_power *
2501 min(this_load_per_task, this_load + tmp);
2502 pwr_move /= SCHED_LOAD_SCALE;
2504 /* Move if we gain throughput */
2505 if (pwr_move <= pwr_now)
2508 *imbalance = busiest_load_per_task;
2514 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2515 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2518 if (this == group_leader && group_leader != group_min) {
2519 *imbalance = min_load_per_task;
2529 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2532 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2533 unsigned long imbalance, cpumask_t *cpus)
2535 struct rq *busiest = NULL, *rq;
2536 unsigned long max_load = 0;
2539 for_each_cpu_mask(i, group->cpumask) {
2541 if (!cpu_isset(i, *cpus))
2546 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2549 if (rq->raw_weighted_load > max_load) {
2550 max_load = rq->raw_weighted_load;
2559 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2560 * so long as it is large enough.
2562 #define MAX_PINNED_INTERVAL 512
2564 static inline unsigned long minus_1_or_zero(unsigned long n)
2566 return n > 0 ? n - 1 : 0;
2570 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2571 * tasks if there is an imbalance.
2573 static int load_balance(int this_cpu, struct rq *this_rq,
2574 struct sched_domain *sd, enum cpu_idle_type idle,
2577 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2578 struct sched_group *group;
2579 unsigned long imbalance;
2581 cpumask_t cpus = CPU_MASK_ALL;
2582 unsigned long flags;
2585 * When power savings policy is enabled for the parent domain, idle
2586 * sibling can pick up load irrespective of busy siblings. In this case,
2587 * let the state of idle sibling percolate up as IDLE, instead of
2588 * portraying it as CPU_NOT_IDLE.
2590 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2591 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2594 schedstat_inc(sd, lb_cnt[idle]);
2597 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2604 schedstat_inc(sd, lb_nobusyg[idle]);
2608 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2610 schedstat_inc(sd, lb_nobusyq[idle]);
2614 BUG_ON(busiest == this_rq);
2616 schedstat_add(sd, lb_imbalance[idle], imbalance);
2619 if (busiest->nr_running > 1) {
2621 * Attempt to move tasks. If find_busiest_group has found
2622 * an imbalance but busiest->nr_running <= 1, the group is
2623 * still unbalanced. nr_moved simply stays zero, so it is
2624 * correctly treated as an imbalance.
2626 local_irq_save(flags);
2627 double_rq_lock(this_rq, busiest);
2628 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2629 minus_1_or_zero(busiest->nr_running),
2630 imbalance, sd, idle, &all_pinned);
2631 double_rq_unlock(this_rq, busiest);
2632 local_irq_restore(flags);
2635 * some other cpu did the load balance for us.
2637 if (nr_moved && this_cpu != smp_processor_id())
2638 resched_cpu(this_cpu);
2640 /* All tasks on this runqueue were pinned by CPU affinity */
2641 if (unlikely(all_pinned)) {
2642 cpu_clear(cpu_of(busiest), cpus);
2643 if (!cpus_empty(cpus))
2650 schedstat_inc(sd, lb_failed[idle]);
2651 sd->nr_balance_failed++;
2653 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2655 spin_lock_irqsave(&busiest->lock, flags);
2657 /* don't kick the migration_thread, if the curr
2658 * task on busiest cpu can't be moved to this_cpu
2660 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2661 spin_unlock_irqrestore(&busiest->lock, flags);
2663 goto out_one_pinned;
2666 if (!busiest->active_balance) {
2667 busiest->active_balance = 1;
2668 busiest->push_cpu = this_cpu;
2671 spin_unlock_irqrestore(&busiest->lock, flags);
2673 wake_up_process(busiest->migration_thread);
2676 * We've kicked active balancing, reset the failure
2679 sd->nr_balance_failed = sd->cache_nice_tries+1;
2682 sd->nr_balance_failed = 0;
2684 if (likely(!active_balance)) {
2685 /* We were unbalanced, so reset the balancing interval */
2686 sd->balance_interval = sd->min_interval;
2689 * If we've begun active balancing, start to back off. This
2690 * case may not be covered by the all_pinned logic if there
2691 * is only 1 task on the busy runqueue (because we don't call
2694 if (sd->balance_interval < sd->max_interval)
2695 sd->balance_interval *= 2;
2698 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2699 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2704 schedstat_inc(sd, lb_balanced[idle]);
2706 sd->nr_balance_failed = 0;
2709 /* tune up the balancing interval */
2710 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2711 (sd->balance_interval < sd->max_interval))
2712 sd->balance_interval *= 2;
2714 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2715 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2721 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2722 * tasks if there is an imbalance.
2724 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2725 * this_rq is locked.
2728 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2730 struct sched_group *group;
2731 struct rq *busiest = NULL;
2732 unsigned long imbalance;
2735 cpumask_t cpus = CPU_MASK_ALL;
2738 * When power savings policy is enabled for the parent domain, idle
2739 * sibling can pick up load irrespective of busy siblings. In this case,
2740 * let the state of idle sibling percolate up as IDLE, instead of
2741 * portraying it as CPU_NOT_IDLE.
2743 if (sd->flags & SD_SHARE_CPUPOWER &&
2744 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2747 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2749 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2750 &sd_idle, &cpus, NULL);
2752 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2756 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2759 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2763 BUG_ON(busiest == this_rq);
2765 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2768 if (busiest->nr_running > 1) {
2769 /* Attempt to move tasks */
2770 double_lock_balance(this_rq, busiest);
2771 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2772 minus_1_or_zero(busiest->nr_running),
2773 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2774 spin_unlock(&busiest->lock);
2777 cpu_clear(cpu_of(busiest), cpus);
2778 if (!cpus_empty(cpus))
2784 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2785 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2786 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2789 sd->nr_balance_failed = 0;
2794 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2795 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2796 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2798 sd->nr_balance_failed = 0;
2804 * idle_balance is called by schedule() if this_cpu is about to become
2805 * idle. Attempts to pull tasks from other CPUs.
2807 static void idle_balance(int this_cpu, struct rq *this_rq)
2809 struct sched_domain *sd;
2810 int pulled_task = 0;
2811 unsigned long next_balance = jiffies + 60 * HZ;
2813 for_each_domain(this_cpu, sd) {
2814 unsigned long interval;
2816 if (!(sd->flags & SD_LOAD_BALANCE))
2819 if (sd->flags & SD_BALANCE_NEWIDLE)
2820 /* If we've pulled tasks over stop searching: */
2821 pulled_task = load_balance_newidle(this_cpu,
2824 interval = msecs_to_jiffies(sd->balance_interval);
2825 if (time_after(next_balance, sd->last_balance + interval))
2826 next_balance = sd->last_balance + interval;
2832 * We are going idle. next_balance may be set based on
2833 * a busy processor. So reset next_balance.
2835 this_rq->next_balance = next_balance;
2839 * active_load_balance is run by migration threads. It pushes running tasks
2840 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2841 * running on each physical CPU where possible, and avoids physical /
2842 * logical imbalances.
2844 * Called with busiest_rq locked.
2846 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2848 int target_cpu = busiest_rq->push_cpu;
2849 struct sched_domain *sd;
2850 struct rq *target_rq;
2852 /* Is there any task to move? */
2853 if (busiest_rq->nr_running <= 1)
2856 target_rq = cpu_rq(target_cpu);
2859 * This condition is "impossible", if it occurs
2860 * we need to fix it. Originally reported by
2861 * Bjorn Helgaas on a 128-cpu setup.
2863 BUG_ON(busiest_rq == target_rq);
2865 /* move a task from busiest_rq to target_rq */
2866 double_lock_balance(busiest_rq, target_rq);
2868 /* Search for an sd spanning us and the target CPU. */
2869 for_each_domain(target_cpu, sd) {
2870 if ((sd->flags & SD_LOAD_BALANCE) &&
2871 cpu_isset(busiest_cpu, sd->span))
2876 schedstat_inc(sd, alb_cnt);
2878 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2879 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2881 schedstat_inc(sd, alb_pushed);
2883 schedstat_inc(sd, alb_failed);
2885 spin_unlock(&target_rq->lock);
2888 static void update_load(struct rq *this_rq)
2890 unsigned long this_load;
2891 unsigned int i, scale;
2893 this_load = this_rq->raw_weighted_load;
2895 /* Update our load: */
2896 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2897 unsigned long old_load, new_load;
2899 /* scale is effectively 1 << i now, and >> i divides by scale */
2901 old_load = this_rq->cpu_load[i];
2902 new_load = this_load;
2904 * Round up the averaging division if load is increasing. This
2905 * prevents us from getting stuck on 9 if the load is 10, for
2908 if (new_load > old_load)
2909 new_load += scale-1;
2910 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2916 atomic_t load_balancer;
2918 } nohz ____cacheline_aligned = {
2919 .load_balancer = ATOMIC_INIT(-1),
2920 .cpu_mask = CPU_MASK_NONE,
2924 * This routine will try to nominate the ilb (idle load balancing)
2925 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2926 * load balancing on behalf of all those cpus. If all the cpus in the system
2927 * go into this tickless mode, then there will be no ilb owner (as there is
2928 * no need for one) and all the cpus will sleep till the next wakeup event
2931 * For the ilb owner, tick is not stopped. And this tick will be used
2932 * for idle load balancing. ilb owner will still be part of
2935 * While stopping the tick, this cpu will become the ilb owner if there
2936 * is no other owner. And will be the owner till that cpu becomes busy
2937 * or if all cpus in the system stop their ticks at which point
2938 * there is no need for ilb owner.
2940 * When the ilb owner becomes busy, it nominates another owner, during the
2941 * next busy scheduler_tick()
2943 int select_nohz_load_balancer(int stop_tick)
2945 int cpu = smp_processor_id();
2948 cpu_set(cpu, nohz.cpu_mask);
2949 cpu_rq(cpu)->in_nohz_recently = 1;
2952 * If we are going offline and still the leader, give up!
2954 if (cpu_is_offline(cpu) &&
2955 atomic_read(&nohz.load_balancer) == cpu) {
2956 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2961 /* time for ilb owner also to sleep */
2962 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2963 if (atomic_read(&nohz.load_balancer) == cpu)
2964 atomic_set(&nohz.load_balancer, -1);
2968 if (atomic_read(&nohz.load_balancer) == -1) {
2969 /* make me the ilb owner */
2970 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2972 } else if (atomic_read(&nohz.load_balancer) == cpu)
2975 if (!cpu_isset(cpu, nohz.cpu_mask))
2978 cpu_clear(cpu, nohz.cpu_mask);
2980 if (atomic_read(&nohz.load_balancer) == cpu)
2981 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2988 static DEFINE_SPINLOCK(balancing);
2991 * It checks each scheduling domain to see if it is due to be balanced,
2992 * and initiates a balancing operation if so.
2994 * Balancing parameters are set up in arch_init_sched_domains.
2996 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2999 struct rq *rq = cpu_rq(cpu);
3000 unsigned long interval;
3001 struct sched_domain *sd;
3002 /* Earliest time when we have to do rebalance again */
3003 unsigned long next_balance = jiffies + 60*HZ;
3005 for_each_domain(cpu, sd) {
3006 if (!(sd->flags & SD_LOAD_BALANCE))
3009 interval = sd->balance_interval;
3010 if (idle != CPU_IDLE)
3011 interval *= sd->busy_factor;
3013 /* scale ms to jiffies */
3014 interval = msecs_to_jiffies(interval);
3015 if (unlikely(!interval))
3018 if (sd->flags & SD_SERIALIZE) {
3019 if (!spin_trylock(&balancing))
3023 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3024 if (load_balance(cpu, rq, sd, idle, &balance)) {
3026 * We've pulled tasks over so either we're no
3027 * longer idle, or one of our SMT siblings is
3030 idle = CPU_NOT_IDLE;
3032 sd->last_balance = jiffies;
3034 if (sd->flags & SD_SERIALIZE)
3035 spin_unlock(&balancing);
3037 if (time_after(next_balance, sd->last_balance + interval))
3038 next_balance = sd->last_balance + interval;
3041 * Stop the load balance at this level. There is another
3042 * CPU in our sched group which is doing load balancing more
3048 rq->next_balance = next_balance;
3052 * run_rebalance_domains is triggered when needed from the scheduler tick.
3053 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3054 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3056 static void run_rebalance_domains(struct softirq_action *h)
3058 int local_cpu = smp_processor_id();
3059 struct rq *local_rq = cpu_rq(local_cpu);
3060 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3062 rebalance_domains(local_cpu, idle);
3066 * If this cpu is the owner for idle load balancing, then do the
3067 * balancing on behalf of the other idle cpus whose ticks are
3070 if (local_rq->idle_at_tick &&
3071 atomic_read(&nohz.load_balancer) == local_cpu) {
3072 cpumask_t cpus = nohz.cpu_mask;
3076 cpu_clear(local_cpu, cpus);
3077 for_each_cpu_mask(balance_cpu, cpus) {
3079 * If this cpu gets work to do, stop the load balancing
3080 * work being done for other cpus. Next load
3081 * balancing owner will pick it up.
3086 rebalance_domains(balance_cpu, CPU_IDLE);
3088 rq = cpu_rq(balance_cpu);
3089 if (time_after(local_rq->next_balance, rq->next_balance))
3090 local_rq->next_balance = rq->next_balance;
3097 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3099 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3100 * idle load balancing owner or decide to stop the periodic load balancing,
3101 * if the whole system is idle.
3103 static inline void trigger_load_balance(int cpu)
3105 struct rq *rq = cpu_rq(cpu);
3108 * If we were in the nohz mode recently and busy at the current
3109 * scheduler tick, then check if we need to nominate new idle
3112 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3113 rq->in_nohz_recently = 0;
3115 if (atomic_read(&nohz.load_balancer) == cpu) {
3116 cpu_clear(cpu, nohz.cpu_mask);
3117 atomic_set(&nohz.load_balancer, -1);
3120 if (atomic_read(&nohz.load_balancer) == -1) {
3122 * simple selection for now: Nominate the
3123 * first cpu in the nohz list to be the next
3126 * TBD: Traverse the sched domains and nominate
3127 * the nearest cpu in the nohz.cpu_mask.
3129 int ilb = first_cpu(nohz.cpu_mask);
3137 * If this cpu is idle and doing idle load balancing for all the
3138 * cpus with ticks stopped, is it time for that to stop?
3140 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3141 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3147 * If this cpu is idle and the idle load balancing is done by
3148 * someone else, then no need raise the SCHED_SOFTIRQ
3150 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3151 cpu_isset(cpu, nohz.cpu_mask))
3154 if (time_after_eq(jiffies, rq->next_balance))
3155 raise_softirq(SCHED_SOFTIRQ);
3159 * on UP we do not need to balance between CPUs:
3161 static inline void idle_balance(int cpu, struct rq *rq)
3166 DEFINE_PER_CPU(struct kernel_stat, kstat);
3168 EXPORT_PER_CPU_SYMBOL(kstat);
3171 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3172 * that have not yet been banked in case the task is currently running.
3174 unsigned long long task_sched_runtime(struct task_struct *p)
3176 unsigned long flags;
3180 rq = task_rq_lock(p, &flags);
3181 ns = p->se.sum_exec_runtime;
3182 if (rq->curr == p) {
3183 delta_exec = rq_clock(rq) - p->se.exec_start;
3184 if ((s64)delta_exec > 0)
3187 task_rq_unlock(rq, &flags);
3193 * We place interactive tasks back into the active array, if possible.
3195 * To guarantee that this does not starve expired tasks we ignore the
3196 * interactivity of a task if the first expired task had to wait more
3197 * than a 'reasonable' amount of time. This deadline timeout is
3198 * load-dependent, as the frequency of array switched decreases with
3199 * increasing number of running tasks. We also ignore the interactivity
3200 * if a better static_prio task has expired:
3202 static inline int expired_starving(struct rq *rq)
3204 if (rq->curr->static_prio > rq->best_expired_prio)
3206 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3208 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3214 * Account user cpu time to a process.
3215 * @p: the process that the cpu time gets accounted to
3216 * @hardirq_offset: the offset to subtract from hardirq_count()
3217 * @cputime: the cpu time spent in user space since the last update
3219 void account_user_time(struct task_struct *p, cputime_t cputime)
3221 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3224 p->utime = cputime_add(p->utime, cputime);
3226 /* Add user time to cpustat. */
3227 tmp = cputime_to_cputime64(cputime);
3228 if (TASK_NICE(p) > 0)
3229 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3231 cpustat->user = cputime64_add(cpustat->user, tmp);
3235 * Account system cpu time to a process.
3236 * @p: the process that the cpu time gets accounted to
3237 * @hardirq_offset: the offset to subtract from hardirq_count()
3238 * @cputime: the cpu time spent in kernel space since the last update
3240 void account_system_time(struct task_struct *p, int hardirq_offset,
3243 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3244 struct rq *rq = this_rq();
3247 p->stime = cputime_add(p->stime, cputime);
3249 /* Add system time to cpustat. */
3250 tmp = cputime_to_cputime64(cputime);
3251 if (hardirq_count() - hardirq_offset)
3252 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3253 else if (softirq_count())
3254 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3255 else if (p != rq->idle)
3256 cpustat->system = cputime64_add(cpustat->system, tmp);
3257 else if (atomic_read(&rq->nr_iowait) > 0)
3258 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3260 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3261 /* Account for system time used */
3262 acct_update_integrals(p);
3266 * Account for involuntary wait time.
3267 * @p: the process from which the cpu time has been stolen
3268 * @steal: the cpu time spent in involuntary wait
3270 void account_steal_time(struct task_struct *p, cputime_t steal)
3272 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3273 cputime64_t tmp = cputime_to_cputime64(steal);
3274 struct rq *rq = this_rq();
3276 if (p == rq->idle) {
3277 p->stime = cputime_add(p->stime, steal);
3278 if (atomic_read(&rq->nr_iowait) > 0)
3279 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3281 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3283 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3286 static void task_running_tick(struct rq *rq, struct task_struct *p)
3288 if (p->array != rq->active) {
3289 /* Task has expired but was not scheduled yet */
3290 set_tsk_need_resched(p);
3293 spin_lock(&rq->lock);
3295 * The task was running during this tick - update the
3296 * time slice counter. Note: we do not update a thread's
3297 * priority until it either goes to sleep or uses up its
3298 * timeslice. This makes it possible for interactive tasks
3299 * to use up their timeslices at their highest priority levels.
3303 * RR tasks need a special form of timeslice management.
3304 * FIFO tasks have no timeslices.
3306 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3307 p->time_slice = task_timeslice(p);
3308 p->first_time_slice = 0;
3309 set_tsk_need_resched(p);
3311 /* put it at the end of the queue: */
3312 requeue_task(p, rq->active);
3316 if (!--p->time_slice) {
3317 dequeue_task(p, rq->active);
3318 set_tsk_need_resched(p);
3319 p->prio = effective_prio(p);
3320 p->time_slice = task_timeslice(p);
3321 p->first_time_slice = 0;
3323 if (!rq->expired_timestamp)
3324 rq->expired_timestamp = jiffies;
3325 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3326 enqueue_task(p, rq->expired);
3327 if (p->static_prio < rq->best_expired_prio)
3328 rq->best_expired_prio = p->static_prio;
3330 enqueue_task(p, rq->active);
3333 * Prevent a too long timeslice allowing a task to monopolize
3334 * the CPU. We do this by splitting up the timeslice into
3337 * Note: this does not mean the task's timeslices expire or
3338 * get lost in any way, they just might be preempted by
3339 * another task of equal priority. (one with higher
3340 * priority would have preempted this task already.) We
3341 * requeue this task to the end of the list on this priority
3342 * level, which is in essence a round-robin of tasks with
3345 * This only applies to tasks in the interactive
3346 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3348 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3349 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3350 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3351 (p->array == rq->active)) {
3353 requeue_task(p, rq->active);
3354 set_tsk_need_resched(p);
3358 spin_unlock(&rq->lock);
3362 * This function gets called by the timer code, with HZ frequency.
3363 * We call it with interrupts disabled.
3365 * It also gets called by the fork code, when changing the parent's
3368 void scheduler_tick(void)
3370 struct task_struct *p = current;
3371 int cpu = smp_processor_id();
3372 int idle_at_tick = idle_cpu(cpu);
3373 struct rq *rq = cpu_rq(cpu);
3376 task_running_tick(rq, p);
3379 rq->idle_at_tick = idle_at_tick;
3380 trigger_load_balance(cpu);
3384 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3386 void fastcall add_preempt_count(int val)
3391 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3393 preempt_count() += val;
3395 * Spinlock count overflowing soon?
3397 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3400 EXPORT_SYMBOL(add_preempt_count);
3402 void fastcall sub_preempt_count(int val)
3407 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3410 * Is the spinlock portion underflowing?
3412 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3413 !(preempt_count() & PREEMPT_MASK)))
3416 preempt_count() -= val;
3418 EXPORT_SYMBOL(sub_preempt_count);
3422 static inline int interactive_sleep(enum sleep_type sleep_type)
3424 return (sleep_type == SLEEP_INTERACTIVE ||
3425 sleep_type == SLEEP_INTERRUPTED);
3429 * schedule() is the main scheduler function.
3431 asmlinkage void __sched schedule(void)
3433 struct task_struct *prev, *next;
3434 struct prio_array *array;
3435 struct list_head *queue;
3436 unsigned long long now;
3437 unsigned long run_time;
3438 int cpu, idx, new_prio;
3443 * Test if we are atomic. Since do_exit() needs to call into
3444 * schedule() atomically, we ignore that path for now.
3445 * Otherwise, whine if we are scheduling when we should not be.
3447 if (unlikely(in_atomic() && !current->exit_state)) {
3448 printk(KERN_ERR "BUG: scheduling while atomic: "
3450 current->comm, preempt_count(), current->pid);
3451 debug_show_held_locks(current);
3452 if (irqs_disabled())
3453 print_irqtrace_events(current);
3456 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3461 release_kernel_lock(prev);
3462 need_resched_nonpreemptible:
3466 * The idle thread is not allowed to schedule!
3467 * Remove this check after it has been exercised a bit.
3469 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3470 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3474 schedstat_inc(rq, sched_cnt);
3475 now = sched_clock();
3476 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3477 run_time = now - prev->timestamp;
3478 if (unlikely((long long)(now - prev->timestamp) < 0))
3481 run_time = NS_MAX_SLEEP_AVG;
3484 * Tasks charged proportionately less run_time at high sleep_avg to
3485 * delay them losing their interactive status
3487 run_time /= (CURRENT_BONUS(prev) ? : 1);
3489 spin_lock_irq(&rq->lock);
3491 switch_count = &prev->nivcsw;
3492 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3493 switch_count = &prev->nvcsw;
3494 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3495 unlikely(signal_pending(prev))))
3496 prev->state = TASK_RUNNING;
3498 if (prev->state == TASK_UNINTERRUPTIBLE)
3499 rq->nr_uninterruptible++;
3500 deactivate_task(prev, rq);
3504 cpu = smp_processor_id();
3505 if (unlikely(!rq->nr_running)) {
3506 idle_balance(cpu, rq);
3507 if (!rq->nr_running) {
3509 rq->expired_timestamp = 0;
3515 if (unlikely(!array->nr_active)) {
3517 * Switch the active and expired arrays.
3519 schedstat_inc(rq, sched_switch);
3520 rq->active = rq->expired;
3521 rq->expired = array;
3523 rq->expired_timestamp = 0;
3524 rq->best_expired_prio = MAX_PRIO;
3527 idx = sched_find_first_bit(array->bitmap);
3528 queue = array->queue + idx;
3529 next = list_entry(queue->next, struct task_struct, run_list);
3531 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3532 unsigned long long delta = now - next->timestamp;
3533 if (unlikely((long long)(now - next->timestamp) < 0))
3536 if (next->sleep_type == SLEEP_INTERACTIVE)
3537 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3539 array = next->array;
3540 new_prio = recalc_task_prio(next, next->timestamp + delta);
3542 if (unlikely(next->prio != new_prio)) {
3543 dequeue_task(next, array);
3544 next->prio = new_prio;
3545 enqueue_task(next, array);
3548 next->sleep_type = SLEEP_NORMAL;
3550 if (next == rq->idle)
3551 schedstat_inc(rq, sched_goidle);
3553 prefetch_stack(next);
3554 clear_tsk_need_resched(prev);
3555 rcu_qsctr_inc(task_cpu(prev));
3557 prev->sleep_avg -= run_time;
3558 if ((long)prev->sleep_avg <= 0)
3559 prev->sleep_avg = 0;
3560 prev->timestamp = prev->last_ran = now;
3562 sched_info_switch(prev, next);
3563 if (likely(prev != next)) {
3564 next->timestamp = next->last_ran = now;
3569 prepare_task_switch(rq, next);
3570 prev = context_switch(rq, prev, next);
3573 * this_rq must be evaluated again because prev may have moved
3574 * CPUs since it called schedule(), thus the 'rq' on its stack
3575 * frame will be invalid.
3577 finish_task_switch(this_rq(), prev);
3579 spin_unlock_irq(&rq->lock);
3582 if (unlikely(reacquire_kernel_lock(prev) < 0))
3583 goto need_resched_nonpreemptible;
3584 preempt_enable_no_resched();
3585 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3588 EXPORT_SYMBOL(schedule);
3590 #ifdef CONFIG_PREEMPT
3592 * this is the entry point to schedule() from in-kernel preemption
3593 * off of preempt_enable. Kernel preemptions off return from interrupt
3594 * occur there and call schedule directly.
3596 asmlinkage void __sched preempt_schedule(void)
3598 struct thread_info *ti = current_thread_info();
3599 #ifdef CONFIG_PREEMPT_BKL
3600 struct task_struct *task = current;
3601 int saved_lock_depth;
3604 * If there is a non-zero preempt_count or interrupts are disabled,
3605 * we do not want to preempt the current task. Just return..
3607 if (likely(ti->preempt_count || irqs_disabled()))
3611 add_preempt_count(PREEMPT_ACTIVE);
3613 * We keep the big kernel semaphore locked, but we
3614 * clear ->lock_depth so that schedule() doesnt
3615 * auto-release the semaphore:
3617 #ifdef CONFIG_PREEMPT_BKL
3618 saved_lock_depth = task->lock_depth;
3619 task->lock_depth = -1;
3622 #ifdef CONFIG_PREEMPT_BKL
3623 task->lock_depth = saved_lock_depth;
3625 sub_preempt_count(PREEMPT_ACTIVE);
3627 /* we could miss a preemption opportunity between schedule and now */
3629 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3632 EXPORT_SYMBOL(preempt_schedule);
3635 * this is the entry point to schedule() from kernel preemption
3636 * off of irq context.
3637 * Note, that this is called and return with irqs disabled. This will
3638 * protect us against recursive calling from irq.
3640 asmlinkage void __sched preempt_schedule_irq(void)
3642 struct thread_info *ti = current_thread_info();
3643 #ifdef CONFIG_PREEMPT_BKL
3644 struct task_struct *task = current;
3645 int saved_lock_depth;
3647 /* Catch callers which need to be fixed */
3648 BUG_ON(ti->preempt_count || !irqs_disabled());
3651 add_preempt_count(PREEMPT_ACTIVE);
3653 * We keep the big kernel semaphore locked, but we
3654 * clear ->lock_depth so that schedule() doesnt
3655 * auto-release the semaphore:
3657 #ifdef CONFIG_PREEMPT_BKL
3658 saved_lock_depth = task->lock_depth;
3659 task->lock_depth = -1;
3663 local_irq_disable();
3664 #ifdef CONFIG_PREEMPT_BKL
3665 task->lock_depth = saved_lock_depth;
3667 sub_preempt_count(PREEMPT_ACTIVE);
3669 /* we could miss a preemption opportunity between schedule and now */
3671 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3675 #endif /* CONFIG_PREEMPT */
3677 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3680 return try_to_wake_up(curr->private, mode, sync);
3682 EXPORT_SYMBOL(default_wake_function);
3685 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3686 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3687 * number) then we wake all the non-exclusive tasks and one exclusive task.
3689 * There are circumstances in which we can try to wake a task which has already
3690 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3691 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3693 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3694 int nr_exclusive, int sync, void *key)
3696 struct list_head *tmp, *next;
3698 list_for_each_safe(tmp, next, &q->task_list) {
3699 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3700 unsigned flags = curr->flags;
3702 if (curr->func(curr, mode, sync, key) &&
3703 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3709 * __wake_up - wake up threads blocked on a waitqueue.
3711 * @mode: which threads
3712 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3713 * @key: is directly passed to the wakeup function
3715 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3716 int nr_exclusive, void *key)
3718 unsigned long flags;
3720 spin_lock_irqsave(&q->lock, flags);
3721 __wake_up_common(q, mode, nr_exclusive, 0, key);
3722 spin_unlock_irqrestore(&q->lock, flags);
3724 EXPORT_SYMBOL(__wake_up);
3727 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3729 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3731 __wake_up_common(q, mode, 1, 0, NULL);
3735 * __wake_up_sync - wake up threads blocked on a waitqueue.
3737 * @mode: which threads
3738 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3740 * The sync wakeup differs that the waker knows that it will schedule
3741 * away soon, so while the target thread will be woken up, it will not
3742 * be migrated to another CPU - ie. the two threads are 'synchronized'
3743 * with each other. This can prevent needless bouncing between CPUs.
3745 * On UP it can prevent extra preemption.
3748 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3750 unsigned long flags;
3756 if (unlikely(!nr_exclusive))
3759 spin_lock_irqsave(&q->lock, flags);
3760 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3761 spin_unlock_irqrestore(&q->lock, flags);
3763 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3765 void fastcall complete(struct completion *x)
3767 unsigned long flags;
3769 spin_lock_irqsave(&x->wait.lock, flags);
3771 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3773 spin_unlock_irqrestore(&x->wait.lock, flags);
3775 EXPORT_SYMBOL(complete);
3777 void fastcall complete_all(struct completion *x)
3779 unsigned long flags;
3781 spin_lock_irqsave(&x->wait.lock, flags);
3782 x->done += UINT_MAX/2;
3783 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3785 spin_unlock_irqrestore(&x->wait.lock, flags);
3787 EXPORT_SYMBOL(complete_all);
3789 void fastcall __sched wait_for_completion(struct completion *x)
3793 spin_lock_irq(&x->wait.lock);
3795 DECLARE_WAITQUEUE(wait, current);
3797 wait.flags |= WQ_FLAG_EXCLUSIVE;
3798 __add_wait_queue_tail(&x->wait, &wait);
3800 __set_current_state(TASK_UNINTERRUPTIBLE);
3801 spin_unlock_irq(&x->wait.lock);
3803 spin_lock_irq(&x->wait.lock);
3805 __remove_wait_queue(&x->wait, &wait);
3808 spin_unlock_irq(&x->wait.lock);
3810 EXPORT_SYMBOL(wait_for_completion);
3812 unsigned long fastcall __sched
3813 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3817 spin_lock_irq(&x->wait.lock);
3819 DECLARE_WAITQUEUE(wait, current);
3821 wait.flags |= WQ_FLAG_EXCLUSIVE;
3822 __add_wait_queue_tail(&x->wait, &wait);
3824 __set_current_state(TASK_UNINTERRUPTIBLE);
3825 spin_unlock_irq(&x->wait.lock);
3826 timeout = schedule_timeout(timeout);
3827 spin_lock_irq(&x->wait.lock);
3829 __remove_wait_queue(&x->wait, &wait);
3833 __remove_wait_queue(&x->wait, &wait);
3837 spin_unlock_irq(&x->wait.lock);
3840 EXPORT_SYMBOL(wait_for_completion_timeout);
3842 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3848 spin_lock_irq(&x->wait.lock);
3850 DECLARE_WAITQUEUE(wait, current);
3852 wait.flags |= WQ_FLAG_EXCLUSIVE;
3853 __add_wait_queue_tail(&x->wait, &wait);
3855 if (signal_pending(current)) {
3857 __remove_wait_queue(&x->wait, &wait);
3860 __set_current_state(TASK_INTERRUPTIBLE);
3861 spin_unlock_irq(&x->wait.lock);
3863 spin_lock_irq(&x->wait.lock);
3865 __remove_wait_queue(&x->wait, &wait);
3869 spin_unlock_irq(&x->wait.lock);
3873 EXPORT_SYMBOL(wait_for_completion_interruptible);
3875 unsigned long fastcall __sched
3876 wait_for_completion_interruptible_timeout(struct completion *x,
3877 unsigned long timeout)
3881 spin_lock_irq(&x->wait.lock);
3883 DECLARE_WAITQUEUE(wait, current);
3885 wait.flags |= WQ_FLAG_EXCLUSIVE;
3886 __add_wait_queue_tail(&x->wait, &wait);
3888 if (signal_pending(current)) {
3889 timeout = -ERESTARTSYS;
3890 __remove_wait_queue(&x->wait, &wait);
3893 __set_current_state(TASK_INTERRUPTIBLE);
3894 spin_unlock_irq(&x->wait.lock);
3895 timeout = schedule_timeout(timeout);
3896 spin_lock_irq(&x->wait.lock);
3898 __remove_wait_queue(&x->wait, &wait);
3902 __remove_wait_queue(&x->wait, &wait);
3906 spin_unlock_irq(&x->wait.lock);
3909 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3912 #define SLEEP_ON_VAR \
3913 unsigned long flags; \
3914 wait_queue_t wait; \
3915 init_waitqueue_entry(&wait, current);
3917 #define SLEEP_ON_HEAD \
3918 spin_lock_irqsave(&q->lock,flags); \
3919 __add_wait_queue(q, &wait); \
3920 spin_unlock(&q->lock);
3922 #define SLEEP_ON_TAIL \
3923 spin_lock_irq(&q->lock); \
3924 __remove_wait_queue(q, &wait); \
3925 spin_unlock_irqrestore(&q->lock, flags);
3927 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3931 current->state = TASK_INTERRUPTIBLE;
3937 EXPORT_SYMBOL(interruptible_sleep_on);
3939 long fastcall __sched
3940 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3944 current->state = TASK_INTERRUPTIBLE;
3947 timeout = schedule_timeout(timeout);
3952 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3954 void fastcall __sched sleep_on(wait_queue_head_t *q)
3958 current->state = TASK_UNINTERRUPTIBLE;
3964 EXPORT_SYMBOL(sleep_on);
3966 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3970 current->state = TASK_UNINTERRUPTIBLE;
3973 timeout = schedule_timeout(timeout);
3979 EXPORT_SYMBOL(sleep_on_timeout);
3981 #ifdef CONFIG_RT_MUTEXES
3984 * rt_mutex_setprio - set the current priority of a task
3986 * @prio: prio value (kernel-internal form)
3988 * This function changes the 'effective' priority of a task. It does
3989 * not touch ->normal_prio like __setscheduler().
3991 * Used by the rt_mutex code to implement priority inheritance logic.
3993 void rt_mutex_setprio(struct task_struct *p, int prio)
3995 struct prio_array *array;
3996 unsigned long flags;
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4002 rq = task_rq_lock(p, &flags);
4007 dequeue_task(p, array);
4012 * If changing to an RT priority then queue it
4013 * in the active array!
4017 enqueue_task(p, array);
4019 * Reschedule if we are currently running on this runqueue and
4020 * our priority decreased, or if we are not currently running on
4021 * this runqueue and our priority is higher than the current's
4023 if (task_running(rq, p)) {
4024 if (p->prio > oldprio)
4025 resched_task(rq->curr);
4026 } else if (TASK_PREEMPTS_CURR(p, rq))
4027 resched_task(rq->curr);
4029 task_rq_unlock(rq, &flags);
4034 void set_user_nice(struct task_struct *p, long nice)
4036 struct prio_array *array;
4037 int old_prio, delta;
4038 unsigned long flags;
4041 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4044 * We have to be careful, if called from sys_setpriority(),
4045 * the task might be in the middle of scheduling on another CPU.
4047 rq = task_rq_lock(p, &flags);
4049 * The RT priorities are set via sched_setscheduler(), but we still
4050 * allow the 'normal' nice value to be set - but as expected
4051 * it wont have any effect on scheduling until the task is
4052 * not SCHED_NORMAL/SCHED_BATCH:
4054 if (has_rt_policy(p)) {
4055 p->static_prio = NICE_TO_PRIO(nice);
4060 dequeue_task(p, array);
4061 dec_raw_weighted_load(rq, p);
4064 p->static_prio = NICE_TO_PRIO(nice);
4067 p->prio = effective_prio(p);
4068 delta = p->prio - old_prio;
4071 enqueue_task(p, array);
4072 inc_raw_weighted_load(rq, p);
4074 * If the task increased its priority or is running and
4075 * lowered its priority, then reschedule its CPU:
4077 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4078 resched_task(rq->curr);
4081 task_rq_unlock(rq, &flags);
4083 EXPORT_SYMBOL(set_user_nice);
4086 * can_nice - check if a task can reduce its nice value
4090 int can_nice(const struct task_struct *p, const int nice)
4092 /* convert nice value [19,-20] to rlimit style value [1,40] */
4093 int nice_rlim = 20 - nice;
4095 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4096 capable(CAP_SYS_NICE));
4099 #ifdef __ARCH_WANT_SYS_NICE
4102 * sys_nice - change the priority of the current process.
4103 * @increment: priority increment
4105 * sys_setpriority is a more generic, but much slower function that
4106 * does similar things.
4108 asmlinkage long sys_nice(int increment)
4113 * Setpriority might change our priority at the same moment.
4114 * We don't have to worry. Conceptually one call occurs first
4115 * and we have a single winner.
4117 if (increment < -40)
4122 nice = PRIO_TO_NICE(current->static_prio) + increment;
4128 if (increment < 0 && !can_nice(current, nice))
4131 retval = security_task_setnice(current, nice);
4135 set_user_nice(current, nice);
4142 * task_prio - return the priority value of a given task.
4143 * @p: the task in question.
4145 * This is the priority value as seen by users in /proc.
4146 * RT tasks are offset by -200. Normal tasks are centered
4147 * around 0, value goes from -16 to +15.
4149 int task_prio(const struct task_struct *p)
4151 return p->prio - MAX_RT_PRIO;
4155 * task_nice - return the nice value of a given task.
4156 * @p: the task in question.
4158 int task_nice(const struct task_struct *p)
4160 return TASK_NICE(p);
4162 EXPORT_SYMBOL_GPL(task_nice);
4165 * idle_cpu - is a given cpu idle currently?
4166 * @cpu: the processor in question.
4168 int idle_cpu(int cpu)
4170 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4174 * idle_task - return the idle task for a given cpu.
4175 * @cpu: the processor in question.
4177 struct task_struct *idle_task(int cpu)
4179 return cpu_rq(cpu)->idle;
4183 * find_process_by_pid - find a process with a matching PID value.
4184 * @pid: the pid in question.
4186 static inline struct task_struct *find_process_by_pid(pid_t pid)
4188 return pid ? find_task_by_pid(pid) : current;
4191 /* Actually do priority change: must hold rq lock. */
4192 static void __setscheduler(struct task_struct *p, int policy, int prio)
4197 p->rt_priority = prio;
4198 p->normal_prio = normal_prio(p);
4199 /* we are holding p->pi_lock already */
4200 p->prio = rt_mutex_getprio(p);
4202 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4204 if (policy == SCHED_BATCH)
4210 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4211 * @p: the task in question.
4212 * @policy: new policy.
4213 * @param: structure containing the new RT priority.
4215 * NOTE that the task may be already dead.
4217 int sched_setscheduler(struct task_struct *p, int policy,
4218 struct sched_param *param)
4220 int retval, oldprio, oldpolicy = -1;
4221 struct prio_array *array;
4222 unsigned long flags;
4225 /* may grab non-irq protected spin_locks */
4226 BUG_ON(in_interrupt());
4228 /* double check policy once rq lock held */
4230 policy = oldpolicy = p->policy;
4231 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4232 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4235 * Valid priorities for SCHED_FIFO and SCHED_RR are
4236 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4239 if (param->sched_priority < 0 ||
4240 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4241 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4243 if (is_rt_policy(policy) != (param->sched_priority != 0))
4247 * Allow unprivileged RT tasks to decrease priority:
4249 if (!capable(CAP_SYS_NICE)) {
4250 if (is_rt_policy(policy)) {
4251 unsigned long rlim_rtprio;
4252 unsigned long flags;
4254 if (!lock_task_sighand(p, &flags))
4256 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4257 unlock_task_sighand(p, &flags);
4259 /* can't set/change the rt policy */
4260 if (policy != p->policy && !rlim_rtprio)
4263 /* can't increase priority */
4264 if (param->sched_priority > p->rt_priority &&
4265 param->sched_priority > rlim_rtprio)
4269 /* can't change other user's priorities */
4270 if ((current->euid != p->euid) &&
4271 (current->euid != p->uid))
4275 retval = security_task_setscheduler(p, policy, param);
4279 * make sure no PI-waiters arrive (or leave) while we are
4280 * changing the priority of the task:
4282 spin_lock_irqsave(&p->pi_lock, flags);
4284 * To be able to change p->policy safely, the apropriate
4285 * runqueue lock must be held.
4287 rq = __task_rq_lock(p);
4288 /* recheck policy now with rq lock held */
4289 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4290 policy = oldpolicy = -1;
4291 __task_rq_unlock(rq);
4292 spin_unlock_irqrestore(&p->pi_lock, flags);
4297 deactivate_task(p, rq);
4299 __setscheduler(p, policy, param->sched_priority);
4301 __activate_task(p, rq);
4303 * Reschedule if we are currently running on this runqueue and
4304 * our priority decreased, or if we are not currently running on
4305 * this runqueue and our priority is higher than the current's
4307 if (task_running(rq, p)) {
4308 if (p->prio > oldprio)
4309 resched_task(rq->curr);
4310 } else if (TASK_PREEMPTS_CURR(p, rq))
4311 resched_task(rq->curr);
4313 __task_rq_unlock(rq);
4314 spin_unlock_irqrestore(&p->pi_lock, flags);
4316 rt_mutex_adjust_pi(p);
4320 EXPORT_SYMBOL_GPL(sched_setscheduler);
4323 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4325 struct sched_param lparam;
4326 struct task_struct *p;
4329 if (!param || pid < 0)
4331 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4336 p = find_process_by_pid(pid);
4338 retval = sched_setscheduler(p, policy, &lparam);
4345 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4346 * @pid: the pid in question.
4347 * @policy: new policy.
4348 * @param: structure containing the new RT priority.
4350 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4351 struct sched_param __user *param)
4353 /* negative values for policy are not valid */
4357 return do_sched_setscheduler(pid, policy, param);
4361 * sys_sched_setparam - set/change the RT priority of a thread
4362 * @pid: the pid in question.
4363 * @param: structure containing the new RT priority.
4365 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4367 return do_sched_setscheduler(pid, -1, param);
4371 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4372 * @pid: the pid in question.
4374 asmlinkage long sys_sched_getscheduler(pid_t pid)
4376 struct task_struct *p;
4377 int retval = -EINVAL;
4383 read_lock(&tasklist_lock);
4384 p = find_process_by_pid(pid);
4386 retval = security_task_getscheduler(p);
4390 read_unlock(&tasklist_lock);
4397 * sys_sched_getscheduler - get the RT priority of a thread
4398 * @pid: the pid in question.
4399 * @param: structure containing the RT priority.
4401 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4403 struct sched_param lp;
4404 struct task_struct *p;
4405 int retval = -EINVAL;
4407 if (!param || pid < 0)
4410 read_lock(&tasklist_lock);
4411 p = find_process_by_pid(pid);
4416 retval = security_task_getscheduler(p);
4420 lp.sched_priority = p->rt_priority;
4421 read_unlock(&tasklist_lock);
4424 * This one might sleep, we cannot do it with a spinlock held ...
4426 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4432 read_unlock(&tasklist_lock);
4436 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4438 cpumask_t cpus_allowed;
4439 struct task_struct *p;
4442 mutex_lock(&sched_hotcpu_mutex);
4443 read_lock(&tasklist_lock);
4445 p = find_process_by_pid(pid);
4447 read_unlock(&tasklist_lock);
4448 mutex_unlock(&sched_hotcpu_mutex);
4453 * It is not safe to call set_cpus_allowed with the
4454 * tasklist_lock held. We will bump the task_struct's
4455 * usage count and then drop tasklist_lock.
4458 read_unlock(&tasklist_lock);
4461 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4462 !capable(CAP_SYS_NICE))
4465 retval = security_task_setscheduler(p, 0, NULL);
4469 cpus_allowed = cpuset_cpus_allowed(p);
4470 cpus_and(new_mask, new_mask, cpus_allowed);
4471 retval = set_cpus_allowed(p, new_mask);
4475 mutex_unlock(&sched_hotcpu_mutex);
4479 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4480 cpumask_t *new_mask)
4482 if (len < sizeof(cpumask_t)) {
4483 memset(new_mask, 0, sizeof(cpumask_t));
4484 } else if (len > sizeof(cpumask_t)) {
4485 len = sizeof(cpumask_t);
4487 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4491 * sys_sched_setaffinity - set the cpu affinity of a process
4492 * @pid: pid of the process
4493 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4494 * @user_mask_ptr: user-space pointer to the new cpu mask
4496 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4497 unsigned long __user *user_mask_ptr)
4502 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4506 return sched_setaffinity(pid, new_mask);
4510 * Represents all cpu's present in the system
4511 * In systems capable of hotplug, this map could dynamically grow
4512 * as new cpu's are detected in the system via any platform specific
4513 * method, such as ACPI for e.g.
4516 cpumask_t cpu_present_map __read_mostly;
4517 EXPORT_SYMBOL(cpu_present_map);
4520 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4521 EXPORT_SYMBOL(cpu_online_map);
4523 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4524 EXPORT_SYMBOL(cpu_possible_map);
4527 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4529 struct task_struct *p;
4532 mutex_lock(&sched_hotcpu_mutex);
4533 read_lock(&tasklist_lock);
4536 p = find_process_by_pid(pid);
4540 retval = security_task_getscheduler(p);
4544 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4547 read_unlock(&tasklist_lock);
4548 mutex_unlock(&sched_hotcpu_mutex);
4556 * sys_sched_getaffinity - get the cpu affinity of a process
4557 * @pid: pid of the process
4558 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4559 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4561 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4562 unsigned long __user *user_mask_ptr)
4567 if (len < sizeof(cpumask_t))
4570 ret = sched_getaffinity(pid, &mask);
4574 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4577 return sizeof(cpumask_t);
4581 * sys_sched_yield - yield the current processor to other threads.
4583 * This function yields the current CPU by moving the calling thread
4584 * to the expired array. If there are no other threads running on this
4585 * CPU then this function will return.
4587 asmlinkage long sys_sched_yield(void)
4589 struct rq *rq = this_rq_lock();
4590 struct prio_array *array = current->array, *target = rq->expired;
4592 schedstat_inc(rq, yld_cnt);
4594 * We implement yielding by moving the task into the expired
4597 * (special rule: RT tasks will just roundrobin in the active
4600 if (rt_task(current))
4601 target = rq->active;
4603 if (array->nr_active == 1) {
4604 schedstat_inc(rq, yld_act_empty);
4605 if (!rq->expired->nr_active)
4606 schedstat_inc(rq, yld_both_empty);
4607 } else if (!rq->expired->nr_active)
4608 schedstat_inc(rq, yld_exp_empty);
4610 if (array != target) {
4611 dequeue_task(current, array);
4612 enqueue_task(current, target);
4615 * requeue_task is cheaper so perform that if possible.
4617 requeue_task(current, array);
4620 * Since we are going to call schedule() anyway, there's
4621 * no need to preempt or enable interrupts:
4623 __release(rq->lock);
4624 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4625 _raw_spin_unlock(&rq->lock);
4626 preempt_enable_no_resched();
4633 static void __cond_resched(void)
4635 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4636 __might_sleep(__FILE__, __LINE__);
4639 * The BKS might be reacquired before we have dropped
4640 * PREEMPT_ACTIVE, which could trigger a second
4641 * cond_resched() call.
4644 add_preempt_count(PREEMPT_ACTIVE);
4646 sub_preempt_count(PREEMPT_ACTIVE);
4647 } while (need_resched());
4650 int __sched cond_resched(void)
4652 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4653 system_state == SYSTEM_RUNNING) {
4659 EXPORT_SYMBOL(cond_resched);
4662 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4663 * call schedule, and on return reacquire the lock.
4665 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4666 * operations here to prevent schedule() from being called twice (once via
4667 * spin_unlock(), once by hand).
4669 int cond_resched_lock(spinlock_t *lock)
4673 if (need_lockbreak(lock)) {
4679 if (need_resched() && system_state == SYSTEM_RUNNING) {
4680 spin_release(&lock->dep_map, 1, _THIS_IP_);
4681 _raw_spin_unlock(lock);
4682 preempt_enable_no_resched();
4689 EXPORT_SYMBOL(cond_resched_lock);
4691 int __sched cond_resched_softirq(void)
4693 BUG_ON(!in_softirq());
4695 if (need_resched() && system_state == SYSTEM_RUNNING) {
4703 EXPORT_SYMBOL(cond_resched_softirq);
4706 * yield - yield the current processor to other threads.
4708 * This is a shortcut for kernel-space yielding - it marks the
4709 * thread runnable and calls sys_sched_yield().
4711 void __sched yield(void)
4713 set_current_state(TASK_RUNNING);
4716 EXPORT_SYMBOL(yield);
4719 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4720 * that process accounting knows that this is a task in IO wait state.
4722 * But don't do that if it is a deliberate, throttling IO wait (this task
4723 * has set its backing_dev_info: the queue against which it should throttle)
4725 void __sched io_schedule(void)
4727 struct rq *rq = &__raw_get_cpu_var(runqueues);
4729 delayacct_blkio_start();
4730 atomic_inc(&rq->nr_iowait);
4732 atomic_dec(&rq->nr_iowait);
4733 delayacct_blkio_end();
4735 EXPORT_SYMBOL(io_schedule);
4737 long __sched io_schedule_timeout(long timeout)
4739 struct rq *rq = &__raw_get_cpu_var(runqueues);
4742 delayacct_blkio_start();
4743 atomic_inc(&rq->nr_iowait);
4744 ret = schedule_timeout(timeout);
4745 atomic_dec(&rq->nr_iowait);
4746 delayacct_blkio_end();
4751 * sys_sched_get_priority_max - return maximum RT priority.
4752 * @policy: scheduling class.
4754 * this syscall returns the maximum rt_priority that can be used
4755 * by a given scheduling class.
4757 asmlinkage long sys_sched_get_priority_max(int policy)
4764 ret = MAX_USER_RT_PRIO-1;
4775 * sys_sched_get_priority_min - return minimum RT priority.
4776 * @policy: scheduling class.
4778 * this syscall returns the minimum rt_priority that can be used
4779 * by a given scheduling class.
4781 asmlinkage long sys_sched_get_priority_min(int policy)
4798 * sys_sched_rr_get_interval - return the default timeslice of a process.
4799 * @pid: pid of the process.
4800 * @interval: userspace pointer to the timeslice value.
4802 * this syscall writes the default timeslice value of a given process
4803 * into the user-space timespec buffer. A value of '0' means infinity.
4806 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4808 struct task_struct *p;
4809 int retval = -EINVAL;
4816 read_lock(&tasklist_lock);
4817 p = find_process_by_pid(pid);
4821 retval = security_task_getscheduler(p);
4825 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4826 0 : task_timeslice(p), &t);
4827 read_unlock(&tasklist_lock);
4828 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4832 read_unlock(&tasklist_lock);
4836 static const char stat_nam[] = "RSDTtZX";
4838 static void show_task(struct task_struct *p)
4840 unsigned long free = 0;
4843 state = p->state ? __ffs(p->state) + 1 : 0;
4844 printk("%-13.13s %c", p->comm,
4845 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4846 #if (BITS_PER_LONG == 32)
4847 if (state == TASK_RUNNING)
4848 printk(" running ");
4850 printk(" %08lX ", thread_saved_pc(p));
4852 if (state == TASK_RUNNING)
4853 printk(" running task ");
4855 printk(" %016lx ", thread_saved_pc(p));
4857 #ifdef CONFIG_DEBUG_STACK_USAGE
4859 unsigned long *n = end_of_stack(p);
4862 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4865 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4867 printk(" (L-TLB)\n");
4869 printk(" (NOTLB)\n");
4871 if (state != TASK_RUNNING)
4872 show_stack(p, NULL);
4875 void show_state_filter(unsigned long state_filter)
4877 struct task_struct *g, *p;
4879 #if (BITS_PER_LONG == 32)
4882 printk(" task PC stack pid father child younger older\n");
4886 printk(" task PC stack pid father child younger older\n");
4888 read_lock(&tasklist_lock);
4889 do_each_thread(g, p) {
4891 * reset the NMI-timeout, listing all files on a slow
4892 * console might take alot of time:
4894 touch_nmi_watchdog();
4895 if (!state_filter || (p->state & state_filter))
4897 } while_each_thread(g, p);
4899 touch_all_softlockup_watchdogs();
4901 read_unlock(&tasklist_lock);
4903 * Only show locks if all tasks are dumped:
4905 if (state_filter == -1)
4906 debug_show_all_locks();
4909 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4915 * init_idle - set up an idle thread for a given CPU
4916 * @idle: task in question
4917 * @cpu: cpu the idle task belongs to
4919 * NOTE: this function does not set the idle thread's NEED_RESCHED
4920 * flag, to make booting more robust.
4922 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4924 struct rq *rq = cpu_rq(cpu);
4925 unsigned long flags;
4927 idle->timestamp = sched_clock();
4928 idle->sleep_avg = 0;
4930 idle->prio = idle->normal_prio = MAX_PRIO;
4931 idle->state = TASK_RUNNING;
4932 idle->cpus_allowed = cpumask_of_cpu(cpu);
4933 set_task_cpu(idle, cpu);
4935 spin_lock_irqsave(&rq->lock, flags);
4936 rq->curr = rq->idle = idle;
4937 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4940 spin_unlock_irqrestore(&rq->lock, flags);
4942 /* Set the preempt count _outside_ the spinlocks! */
4943 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4944 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4946 task_thread_info(idle)->preempt_count = 0;
4951 * In a system that switches off the HZ timer nohz_cpu_mask
4952 * indicates which cpus entered this state. This is used
4953 * in the rcu update to wait only for active cpus. For system
4954 * which do not switch off the HZ timer nohz_cpu_mask should
4955 * always be CPU_MASK_NONE.
4957 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4961 * This is how migration works:
4963 * 1) we queue a struct migration_req structure in the source CPU's
4964 * runqueue and wake up that CPU's migration thread.
4965 * 2) we down() the locked semaphore => thread blocks.
4966 * 3) migration thread wakes up (implicitly it forces the migrated
4967 * thread off the CPU)
4968 * 4) it gets the migration request and checks whether the migrated
4969 * task is still in the wrong runqueue.
4970 * 5) if it's in the wrong runqueue then the migration thread removes
4971 * it and puts it into the right queue.
4972 * 6) migration thread up()s the semaphore.
4973 * 7) we wake up and the migration is done.
4977 * Change a given task's CPU affinity. Migrate the thread to a
4978 * proper CPU and schedule it away if the CPU it's executing on
4979 * is removed from the allowed bitmask.
4981 * NOTE: the caller must have a valid reference to the task, the
4982 * task must not exit() & deallocate itself prematurely. The
4983 * call is not atomic; no spinlocks may be held.
4985 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4987 struct migration_req req;
4988 unsigned long flags;
4992 rq = task_rq_lock(p, &flags);
4993 if (!cpus_intersects(new_mask, cpu_online_map)) {
4998 p->cpus_allowed = new_mask;
4999 /* Can the task run on the task's current CPU? If so, we're done */
5000 if (cpu_isset(task_cpu(p), new_mask))
5003 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5004 /* Need help from migration thread: drop lock and wait. */
5005 task_rq_unlock(rq, &flags);
5006 wake_up_process(rq->migration_thread);
5007 wait_for_completion(&req.done);
5008 tlb_migrate_finish(p->mm);
5012 task_rq_unlock(rq, &flags);
5016 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5019 * Move (not current) task off this cpu, onto dest cpu. We're doing
5020 * this because either it can't run here any more (set_cpus_allowed()
5021 * away from this CPU, or CPU going down), or because we're
5022 * attempting to rebalance this task on exec (sched_exec).
5024 * So we race with normal scheduler movements, but that's OK, as long
5025 * as the task is no longer on this CPU.
5027 * Returns non-zero if task was successfully migrated.
5029 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5031 struct rq *rq_dest, *rq_src;
5034 if (unlikely(cpu_is_offline(dest_cpu)))
5037 rq_src = cpu_rq(src_cpu);
5038 rq_dest = cpu_rq(dest_cpu);
5040 double_rq_lock(rq_src, rq_dest);
5041 /* Already moved. */
5042 if (task_cpu(p) != src_cpu)
5044 /* Affinity changed (again). */
5045 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5048 set_task_cpu(p, dest_cpu);
5051 * Sync timestamp with rq_dest's before activating.
5052 * The same thing could be achieved by doing this step
5053 * afterwards, and pretending it was a local activate.
5054 * This way is cleaner and logically correct.
5056 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5057 + rq_dest->most_recent_timestamp;
5058 deactivate_task(p, rq_src);
5059 __activate_task(p, rq_dest);
5060 if (TASK_PREEMPTS_CURR(p, rq_dest))
5061 resched_task(rq_dest->curr);
5065 double_rq_unlock(rq_src, rq_dest);
5070 * migration_thread - this is a highprio system thread that performs
5071 * thread migration by bumping thread off CPU then 'pushing' onto
5074 static int migration_thread(void *data)
5076 int cpu = (long)data;
5080 BUG_ON(rq->migration_thread != current);
5082 set_current_state(TASK_INTERRUPTIBLE);
5083 while (!kthread_should_stop()) {
5084 struct migration_req *req;
5085 struct list_head *head;
5089 spin_lock_irq(&rq->lock);
5091 if (cpu_is_offline(cpu)) {
5092 spin_unlock_irq(&rq->lock);
5096 if (rq->active_balance) {
5097 active_load_balance(rq, cpu);
5098 rq->active_balance = 0;
5101 head = &rq->migration_queue;
5103 if (list_empty(head)) {
5104 spin_unlock_irq(&rq->lock);
5106 set_current_state(TASK_INTERRUPTIBLE);
5109 req = list_entry(head->next, struct migration_req, list);
5110 list_del_init(head->next);
5112 spin_unlock(&rq->lock);
5113 __migrate_task(req->task, cpu, req->dest_cpu);
5116 complete(&req->done);
5118 __set_current_state(TASK_RUNNING);
5122 /* Wait for kthread_stop */
5123 set_current_state(TASK_INTERRUPTIBLE);
5124 while (!kthread_should_stop()) {
5126 set_current_state(TASK_INTERRUPTIBLE);
5128 __set_current_state(TASK_RUNNING);
5132 #ifdef CONFIG_HOTPLUG_CPU
5134 * Figure out where task on dead CPU should go, use force if neccessary.
5135 * NOTE: interrupts should be disabled by the caller
5137 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5139 unsigned long flags;
5146 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5147 cpus_and(mask, mask, p->cpus_allowed);
5148 dest_cpu = any_online_cpu(mask);
5150 /* On any allowed CPU? */
5151 if (dest_cpu == NR_CPUS)
5152 dest_cpu = any_online_cpu(p->cpus_allowed);
5154 /* No more Mr. Nice Guy. */
5155 if (dest_cpu == NR_CPUS) {
5156 rq = task_rq_lock(p, &flags);
5157 cpus_setall(p->cpus_allowed);
5158 dest_cpu = any_online_cpu(p->cpus_allowed);
5159 task_rq_unlock(rq, &flags);
5162 * Don't tell them about moving exiting tasks or
5163 * kernel threads (both mm NULL), since they never
5166 if (p->mm && printk_ratelimit())
5167 printk(KERN_INFO "process %d (%s) no "
5168 "longer affine to cpu%d\n",
5169 p->pid, p->comm, dead_cpu);
5171 if (!__migrate_task(p, dead_cpu, dest_cpu))
5176 * While a dead CPU has no uninterruptible tasks queued at this point,
5177 * it might still have a nonzero ->nr_uninterruptible counter, because
5178 * for performance reasons the counter is not stricly tracking tasks to
5179 * their home CPUs. So we just add the counter to another CPU's counter,
5180 * to keep the global sum constant after CPU-down:
5182 static void migrate_nr_uninterruptible(struct rq *rq_src)
5184 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5185 unsigned long flags;
5187 local_irq_save(flags);
5188 double_rq_lock(rq_src, rq_dest);
5189 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5190 rq_src->nr_uninterruptible = 0;
5191 double_rq_unlock(rq_src, rq_dest);
5192 local_irq_restore(flags);
5195 /* Run through task list and migrate tasks from the dead cpu. */
5196 static void migrate_live_tasks(int src_cpu)
5198 struct task_struct *p, *t;
5200 write_lock_irq(&tasklist_lock);
5202 do_each_thread(t, p) {
5206 if (task_cpu(p) == src_cpu)
5207 move_task_off_dead_cpu(src_cpu, p);
5208 } while_each_thread(t, p);
5210 write_unlock_irq(&tasklist_lock);
5213 /* Schedules idle task to be the next runnable task on current CPU.
5214 * It does so by boosting its priority to highest possible and adding it to
5215 * the _front_ of the runqueue. Used by CPU offline code.
5217 void sched_idle_next(void)
5219 int this_cpu = smp_processor_id();
5220 struct rq *rq = cpu_rq(this_cpu);
5221 struct task_struct *p = rq->idle;
5222 unsigned long flags;
5224 /* cpu has to be offline */
5225 BUG_ON(cpu_online(this_cpu));
5228 * Strictly not necessary since rest of the CPUs are stopped by now
5229 * and interrupts disabled on the current cpu.
5231 spin_lock_irqsave(&rq->lock, flags);
5233 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5235 /* Add idle task to the _front_ of its priority queue: */
5236 __activate_idle_task(p, rq);
5238 spin_unlock_irqrestore(&rq->lock, flags);
5242 * Ensures that the idle task is using init_mm right before its cpu goes
5245 void idle_task_exit(void)
5247 struct mm_struct *mm = current->active_mm;
5249 BUG_ON(cpu_online(smp_processor_id()));
5252 switch_mm(mm, &init_mm, current);
5256 /* called under rq->lock with disabled interrupts */
5257 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5259 struct rq *rq = cpu_rq(dead_cpu);
5261 /* Must be exiting, otherwise would be on tasklist. */
5262 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5264 /* Cannot have done final schedule yet: would have vanished. */
5265 BUG_ON(p->state == TASK_DEAD);
5270 * Drop lock around migration; if someone else moves it,
5271 * that's OK. No task can be added to this CPU, so iteration is
5273 * NOTE: interrupts should be left disabled --dev@
5275 spin_unlock(&rq->lock);
5276 move_task_off_dead_cpu(dead_cpu, p);
5277 spin_lock(&rq->lock);
5282 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5283 static void migrate_dead_tasks(unsigned int dead_cpu)
5285 struct rq *rq = cpu_rq(dead_cpu);
5286 unsigned int arr, i;
5288 for (arr = 0; arr < 2; arr++) {
5289 for (i = 0; i < MAX_PRIO; i++) {
5290 struct list_head *list = &rq->arrays[arr].queue[i];
5292 while (!list_empty(list))
5293 migrate_dead(dead_cpu, list_entry(list->next,
5294 struct task_struct, run_list));
5298 #endif /* CONFIG_HOTPLUG_CPU */
5301 * migration_call - callback that gets triggered when a CPU is added.
5302 * Here we can start up the necessary migration thread for the new CPU.
5304 static int __cpuinit
5305 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5307 struct task_struct *p;
5308 int cpu = (long)hcpu;
5309 unsigned long flags;
5313 case CPU_LOCK_ACQUIRE:
5314 mutex_lock(&sched_hotcpu_mutex);
5317 case CPU_UP_PREPARE:
5318 case CPU_UP_PREPARE_FROZEN:
5319 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5322 p->flags |= PF_NOFREEZE;
5323 kthread_bind(p, cpu);
5324 /* Must be high prio: stop_machine expects to yield to it. */
5325 rq = task_rq_lock(p, &flags);
5326 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5327 task_rq_unlock(rq, &flags);
5328 cpu_rq(cpu)->migration_thread = p;
5332 case CPU_ONLINE_FROZEN:
5333 /* Strictly unneccessary, as first user will wake it. */
5334 wake_up_process(cpu_rq(cpu)->migration_thread);
5337 #ifdef CONFIG_HOTPLUG_CPU
5338 case CPU_UP_CANCELED:
5339 case CPU_UP_CANCELED_FROZEN:
5340 if (!cpu_rq(cpu)->migration_thread)
5342 /* Unbind it from offline cpu so it can run. Fall thru. */
5343 kthread_bind(cpu_rq(cpu)->migration_thread,
5344 any_online_cpu(cpu_online_map));
5345 kthread_stop(cpu_rq(cpu)->migration_thread);
5346 cpu_rq(cpu)->migration_thread = NULL;
5350 case CPU_DEAD_FROZEN:
5351 migrate_live_tasks(cpu);
5353 kthread_stop(rq->migration_thread);
5354 rq->migration_thread = NULL;
5355 /* Idle task back to normal (off runqueue, low prio) */
5356 rq = task_rq_lock(rq->idle, &flags);
5357 deactivate_task(rq->idle, rq);
5358 rq->idle->static_prio = MAX_PRIO;
5359 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5360 migrate_dead_tasks(cpu);
5361 task_rq_unlock(rq, &flags);
5362 migrate_nr_uninterruptible(rq);
5363 BUG_ON(rq->nr_running != 0);
5365 /* No need to migrate the tasks: it was best-effort if
5366 * they didn't take sched_hotcpu_mutex. Just wake up
5367 * the requestors. */
5368 spin_lock_irq(&rq->lock);
5369 while (!list_empty(&rq->migration_queue)) {
5370 struct migration_req *req;
5372 req = list_entry(rq->migration_queue.next,
5373 struct migration_req, list);
5374 list_del_init(&req->list);
5375 complete(&req->done);
5377 spin_unlock_irq(&rq->lock);
5380 case CPU_LOCK_RELEASE:
5381 mutex_unlock(&sched_hotcpu_mutex);
5387 /* Register at highest priority so that task migration (migrate_all_tasks)
5388 * happens before everything else.
5390 static struct notifier_block __cpuinitdata migration_notifier = {
5391 .notifier_call = migration_call,
5395 int __init migration_init(void)
5397 void *cpu = (void *)(long)smp_processor_id();
5400 /* Start one for the boot CPU: */
5401 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5402 BUG_ON(err == NOTIFY_BAD);
5403 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5404 register_cpu_notifier(&migration_notifier);
5412 /* Number of possible processor ids */
5413 int nr_cpu_ids __read_mostly = NR_CPUS;
5414 EXPORT_SYMBOL(nr_cpu_ids);
5416 #undef SCHED_DOMAIN_DEBUG
5417 #ifdef SCHED_DOMAIN_DEBUG
5418 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5423 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5427 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5432 struct sched_group *group = sd->groups;
5433 cpumask_t groupmask;
5435 cpumask_scnprintf(str, NR_CPUS, sd->span);
5436 cpus_clear(groupmask);
5439 for (i = 0; i < level + 1; i++)
5441 printk("domain %d: ", level);
5443 if (!(sd->flags & SD_LOAD_BALANCE)) {
5444 printk("does not load-balance\n");
5446 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5451 printk("span %s\n", str);
5453 if (!cpu_isset(cpu, sd->span))
5454 printk(KERN_ERR "ERROR: domain->span does not contain "
5456 if (!cpu_isset(cpu, group->cpumask))
5457 printk(KERN_ERR "ERROR: domain->groups does not contain"
5461 for (i = 0; i < level + 2; i++)
5467 printk(KERN_ERR "ERROR: group is NULL\n");
5471 if (!group->__cpu_power) {
5473 printk(KERN_ERR "ERROR: domain->cpu_power not "
5477 if (!cpus_weight(group->cpumask)) {
5479 printk(KERN_ERR "ERROR: empty group\n");
5482 if (cpus_intersects(groupmask, group->cpumask)) {
5484 printk(KERN_ERR "ERROR: repeated CPUs\n");
5487 cpus_or(groupmask, groupmask, group->cpumask);
5489 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5492 group = group->next;
5493 } while (group != sd->groups);
5496 if (!cpus_equal(sd->span, groupmask))
5497 printk(KERN_ERR "ERROR: groups don't span "
5505 if (!cpus_subset(groupmask, sd->span))
5506 printk(KERN_ERR "ERROR: parent span is not a superset "
5507 "of domain->span\n");
5512 # define sched_domain_debug(sd, cpu) do { } while (0)
5515 static int sd_degenerate(struct sched_domain *sd)
5517 if (cpus_weight(sd->span) == 1)
5520 /* Following flags need at least 2 groups */
5521 if (sd->flags & (SD_LOAD_BALANCE |
5522 SD_BALANCE_NEWIDLE |
5526 SD_SHARE_PKG_RESOURCES)) {
5527 if (sd->groups != sd->groups->next)
5531 /* Following flags don't use groups */
5532 if (sd->flags & (SD_WAKE_IDLE |
5541 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5543 unsigned long cflags = sd->flags, pflags = parent->flags;
5545 if (sd_degenerate(parent))
5548 if (!cpus_equal(sd->span, parent->span))
5551 /* Does parent contain flags not in child? */
5552 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5553 if (cflags & SD_WAKE_AFFINE)
5554 pflags &= ~SD_WAKE_BALANCE;
5555 /* Flags needing groups don't count if only 1 group in parent */
5556 if (parent->groups == parent->groups->next) {
5557 pflags &= ~(SD_LOAD_BALANCE |
5558 SD_BALANCE_NEWIDLE |
5562 SD_SHARE_PKG_RESOURCES);
5564 if (~cflags & pflags)
5571 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5572 * hold the hotplug lock.
5574 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5576 struct rq *rq = cpu_rq(cpu);
5577 struct sched_domain *tmp;
5579 /* Remove the sched domains which do not contribute to scheduling. */
5580 for (tmp = sd; tmp; tmp = tmp->parent) {
5581 struct sched_domain *parent = tmp->parent;
5584 if (sd_parent_degenerate(tmp, parent)) {
5585 tmp->parent = parent->parent;
5587 parent->parent->child = tmp;
5591 if (sd && sd_degenerate(sd)) {
5597 sched_domain_debug(sd, cpu);
5599 rcu_assign_pointer(rq->sd, sd);
5602 /* cpus with isolated domains */
5603 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5605 /* Setup the mask of cpus configured for isolated domains */
5606 static int __init isolated_cpu_setup(char *str)
5608 int ints[NR_CPUS], i;
5610 str = get_options(str, ARRAY_SIZE(ints), ints);
5611 cpus_clear(cpu_isolated_map);
5612 for (i = 1; i <= ints[0]; i++)
5613 if (ints[i] < NR_CPUS)
5614 cpu_set(ints[i], cpu_isolated_map);
5618 __setup ("isolcpus=", isolated_cpu_setup);
5621 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5622 * to a function which identifies what group(along with sched group) a CPU
5623 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5624 * (due to the fact that we keep track of groups covered with a cpumask_t).
5626 * init_sched_build_groups will build a circular linked list of the groups
5627 * covered by the given span, and will set each group's ->cpumask correctly,
5628 * and ->cpu_power to 0.
5631 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5632 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5633 struct sched_group **sg))
5635 struct sched_group *first = NULL, *last = NULL;
5636 cpumask_t covered = CPU_MASK_NONE;
5639 for_each_cpu_mask(i, span) {
5640 struct sched_group *sg;
5641 int group = group_fn(i, cpu_map, &sg);
5644 if (cpu_isset(i, covered))
5647 sg->cpumask = CPU_MASK_NONE;
5648 sg->__cpu_power = 0;
5650 for_each_cpu_mask(j, span) {
5651 if (group_fn(j, cpu_map, NULL) != group)
5654 cpu_set(j, covered);
5655 cpu_set(j, sg->cpumask);
5666 #define SD_NODES_PER_DOMAIN 16
5671 * find_next_best_node - find the next node to include in a sched_domain
5672 * @node: node whose sched_domain we're building
5673 * @used_nodes: nodes already in the sched_domain
5675 * Find the next node to include in a given scheduling domain. Simply
5676 * finds the closest node not already in the @used_nodes map.
5678 * Should use nodemask_t.
5680 static int find_next_best_node(int node, unsigned long *used_nodes)
5682 int i, n, val, min_val, best_node = 0;
5686 for (i = 0; i < MAX_NUMNODES; i++) {
5687 /* Start at @node */
5688 n = (node + i) % MAX_NUMNODES;
5690 if (!nr_cpus_node(n))
5693 /* Skip already used nodes */
5694 if (test_bit(n, used_nodes))
5697 /* Simple min distance search */
5698 val = node_distance(node, n);
5700 if (val < min_val) {
5706 set_bit(best_node, used_nodes);
5711 * sched_domain_node_span - get a cpumask for a node's sched_domain
5712 * @node: node whose cpumask we're constructing
5713 * @size: number of nodes to include in this span
5715 * Given a node, construct a good cpumask for its sched_domain to span. It
5716 * should be one that prevents unnecessary balancing, but also spreads tasks
5719 static cpumask_t sched_domain_node_span(int node)
5721 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5722 cpumask_t span, nodemask;
5726 bitmap_zero(used_nodes, MAX_NUMNODES);
5728 nodemask = node_to_cpumask(node);
5729 cpus_or(span, span, nodemask);
5730 set_bit(node, used_nodes);
5732 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5733 int next_node = find_next_best_node(node, used_nodes);
5735 nodemask = node_to_cpumask(next_node);
5736 cpus_or(span, span, nodemask);
5743 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5746 * SMT sched-domains:
5748 #ifdef CONFIG_SCHED_SMT
5749 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5750 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5752 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5753 struct sched_group **sg)
5756 *sg = &per_cpu(sched_group_cpus, cpu);
5762 * multi-core sched-domains:
5764 #ifdef CONFIG_SCHED_MC
5765 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5766 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5769 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5770 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5771 struct sched_group **sg)
5774 cpumask_t mask = cpu_sibling_map[cpu];
5775 cpus_and(mask, mask, *cpu_map);
5776 group = first_cpu(mask);
5778 *sg = &per_cpu(sched_group_core, group);
5781 #elif defined(CONFIG_SCHED_MC)
5782 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5783 struct sched_group **sg)
5786 *sg = &per_cpu(sched_group_core, cpu);
5791 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5792 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5794 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5795 struct sched_group **sg)
5798 #ifdef CONFIG_SCHED_MC
5799 cpumask_t mask = cpu_coregroup_map(cpu);
5800 cpus_and(mask, mask, *cpu_map);
5801 group = first_cpu(mask);
5802 #elif defined(CONFIG_SCHED_SMT)
5803 cpumask_t mask = cpu_sibling_map[cpu];
5804 cpus_and(mask, mask, *cpu_map);
5805 group = first_cpu(mask);
5810 *sg = &per_cpu(sched_group_phys, group);
5816 * The init_sched_build_groups can't handle what we want to do with node
5817 * groups, so roll our own. Now each node has its own list of groups which
5818 * gets dynamically allocated.
5820 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5821 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5823 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5824 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5826 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5827 struct sched_group **sg)
5829 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5832 cpus_and(nodemask, nodemask, *cpu_map);
5833 group = first_cpu(nodemask);
5836 *sg = &per_cpu(sched_group_allnodes, group);
5840 static void init_numa_sched_groups_power(struct sched_group *group_head)
5842 struct sched_group *sg = group_head;
5848 for_each_cpu_mask(j, sg->cpumask) {
5849 struct sched_domain *sd;
5851 sd = &per_cpu(phys_domains, j);
5852 if (j != first_cpu(sd->groups->cpumask)) {
5854 * Only add "power" once for each
5860 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5863 if (sg != group_head)
5869 /* Free memory allocated for various sched_group structures */
5870 static void free_sched_groups(const cpumask_t *cpu_map)
5874 for_each_cpu_mask(cpu, *cpu_map) {
5875 struct sched_group **sched_group_nodes
5876 = sched_group_nodes_bycpu[cpu];
5878 if (!sched_group_nodes)
5881 for (i = 0; i < MAX_NUMNODES; i++) {
5882 cpumask_t nodemask = node_to_cpumask(i);
5883 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5885 cpus_and(nodemask, nodemask, *cpu_map);
5886 if (cpus_empty(nodemask))
5896 if (oldsg != sched_group_nodes[i])
5899 kfree(sched_group_nodes);
5900 sched_group_nodes_bycpu[cpu] = NULL;
5904 static void free_sched_groups(const cpumask_t *cpu_map)
5910 * Initialize sched groups cpu_power.
5912 * cpu_power indicates the capacity of sched group, which is used while
5913 * distributing the load between different sched groups in a sched domain.
5914 * Typically cpu_power for all the groups in a sched domain will be same unless
5915 * there are asymmetries in the topology. If there are asymmetries, group
5916 * having more cpu_power will pickup more load compared to the group having
5919 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5920 * the maximum number of tasks a group can handle in the presence of other idle
5921 * or lightly loaded groups in the same sched domain.
5923 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5925 struct sched_domain *child;
5926 struct sched_group *group;
5928 WARN_ON(!sd || !sd->groups);
5930 if (cpu != first_cpu(sd->groups->cpumask))
5935 sd->groups->__cpu_power = 0;
5938 * For perf policy, if the groups in child domain share resources
5939 * (for example cores sharing some portions of the cache hierarchy
5940 * or SMT), then set this domain groups cpu_power such that each group
5941 * can handle only one task, when there are other idle groups in the
5942 * same sched domain.
5944 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5946 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5947 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5952 * add cpu_power of each child group to this groups cpu_power
5954 group = child->groups;
5956 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5957 group = group->next;
5958 } while (group != child->groups);
5962 * Build sched domains for a given set of cpus and attach the sched domains
5963 * to the individual cpus
5965 static int build_sched_domains(const cpumask_t *cpu_map)
5968 struct sched_domain *sd;
5970 struct sched_group **sched_group_nodes = NULL;
5971 int sd_allnodes = 0;
5974 * Allocate the per-node list of sched groups
5976 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5978 if (!sched_group_nodes) {
5979 printk(KERN_WARNING "Can not alloc sched group node list\n");
5982 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5986 * Set up domains for cpus specified by the cpu_map.
5988 for_each_cpu_mask(i, *cpu_map) {
5989 struct sched_domain *sd = NULL, *p;
5990 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5992 cpus_and(nodemask, nodemask, *cpu_map);
5995 if (cpus_weight(*cpu_map)
5996 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5997 sd = &per_cpu(allnodes_domains, i);
5998 *sd = SD_ALLNODES_INIT;
5999 sd->span = *cpu_map;
6000 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6006 sd = &per_cpu(node_domains, i);
6008 sd->span = sched_domain_node_span(cpu_to_node(i));
6012 cpus_and(sd->span, sd->span, *cpu_map);
6016 sd = &per_cpu(phys_domains, i);
6018 sd->span = nodemask;
6022 cpu_to_phys_group(i, cpu_map, &sd->groups);
6024 #ifdef CONFIG_SCHED_MC
6026 sd = &per_cpu(core_domains, i);
6028 sd->span = cpu_coregroup_map(i);
6029 cpus_and(sd->span, sd->span, *cpu_map);
6032 cpu_to_core_group(i, cpu_map, &sd->groups);
6035 #ifdef CONFIG_SCHED_SMT
6037 sd = &per_cpu(cpu_domains, i);
6038 *sd = SD_SIBLING_INIT;
6039 sd->span = cpu_sibling_map[i];
6040 cpus_and(sd->span, sd->span, *cpu_map);
6043 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6047 #ifdef CONFIG_SCHED_SMT
6048 /* Set up CPU (sibling) groups */
6049 for_each_cpu_mask(i, *cpu_map) {
6050 cpumask_t this_sibling_map = cpu_sibling_map[i];
6051 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6052 if (i != first_cpu(this_sibling_map))
6055 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6059 #ifdef CONFIG_SCHED_MC
6060 /* Set up multi-core groups */
6061 for_each_cpu_mask(i, *cpu_map) {
6062 cpumask_t this_core_map = cpu_coregroup_map(i);
6063 cpus_and(this_core_map, this_core_map, *cpu_map);
6064 if (i != first_cpu(this_core_map))
6066 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6071 /* Set up physical groups */
6072 for (i = 0; i < MAX_NUMNODES; i++) {
6073 cpumask_t nodemask = node_to_cpumask(i);
6075 cpus_and(nodemask, nodemask, *cpu_map);
6076 if (cpus_empty(nodemask))
6079 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6083 /* Set up node groups */
6085 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6087 for (i = 0; i < MAX_NUMNODES; i++) {
6088 /* Set up node groups */
6089 struct sched_group *sg, *prev;
6090 cpumask_t nodemask = node_to_cpumask(i);
6091 cpumask_t domainspan;
6092 cpumask_t covered = CPU_MASK_NONE;
6095 cpus_and(nodemask, nodemask, *cpu_map);
6096 if (cpus_empty(nodemask)) {
6097 sched_group_nodes[i] = NULL;
6101 domainspan = sched_domain_node_span(i);
6102 cpus_and(domainspan, domainspan, *cpu_map);
6104 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6106 printk(KERN_WARNING "Can not alloc domain group for "
6110 sched_group_nodes[i] = sg;
6111 for_each_cpu_mask(j, nodemask) {
6112 struct sched_domain *sd;
6113 sd = &per_cpu(node_domains, j);
6116 sg->__cpu_power = 0;
6117 sg->cpumask = nodemask;
6119 cpus_or(covered, covered, nodemask);
6122 for (j = 0; j < MAX_NUMNODES; j++) {
6123 cpumask_t tmp, notcovered;
6124 int n = (i + j) % MAX_NUMNODES;
6126 cpus_complement(notcovered, covered);
6127 cpus_and(tmp, notcovered, *cpu_map);
6128 cpus_and(tmp, tmp, domainspan);
6129 if (cpus_empty(tmp))
6132 nodemask = node_to_cpumask(n);
6133 cpus_and(tmp, tmp, nodemask);
6134 if (cpus_empty(tmp))
6137 sg = kmalloc_node(sizeof(struct sched_group),
6141 "Can not alloc domain group for node %d\n", j);
6144 sg->__cpu_power = 0;
6146 sg->next = prev->next;
6147 cpus_or(covered, covered, tmp);
6154 /* Calculate CPU power for physical packages and nodes */
6155 #ifdef CONFIG_SCHED_SMT
6156 for_each_cpu_mask(i, *cpu_map) {
6157 sd = &per_cpu(cpu_domains, i);
6158 init_sched_groups_power(i, sd);
6161 #ifdef CONFIG_SCHED_MC
6162 for_each_cpu_mask(i, *cpu_map) {
6163 sd = &per_cpu(core_domains, i);
6164 init_sched_groups_power(i, sd);
6168 for_each_cpu_mask(i, *cpu_map) {
6169 sd = &per_cpu(phys_domains, i);
6170 init_sched_groups_power(i, sd);
6174 for (i = 0; i < MAX_NUMNODES; i++)
6175 init_numa_sched_groups_power(sched_group_nodes[i]);
6178 struct sched_group *sg;
6180 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6181 init_numa_sched_groups_power(sg);
6185 /* Attach the domains */
6186 for_each_cpu_mask(i, *cpu_map) {
6187 struct sched_domain *sd;
6188 #ifdef CONFIG_SCHED_SMT
6189 sd = &per_cpu(cpu_domains, i);
6190 #elif defined(CONFIG_SCHED_MC)
6191 sd = &per_cpu(core_domains, i);
6193 sd = &per_cpu(phys_domains, i);
6195 cpu_attach_domain(sd, i);
6202 free_sched_groups(cpu_map);
6207 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6209 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6211 cpumask_t cpu_default_map;
6215 * Setup mask for cpus without special case scheduling requirements.
6216 * For now this just excludes isolated cpus, but could be used to
6217 * exclude other special cases in the future.
6219 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6221 err = build_sched_domains(&cpu_default_map);
6226 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6228 free_sched_groups(cpu_map);
6232 * Detach sched domains from a group of cpus specified in cpu_map
6233 * These cpus will now be attached to the NULL domain
6235 static void detach_destroy_domains(const cpumask_t *cpu_map)
6239 for_each_cpu_mask(i, *cpu_map)
6240 cpu_attach_domain(NULL, i);
6241 synchronize_sched();
6242 arch_destroy_sched_domains(cpu_map);
6246 * Partition sched domains as specified by the cpumasks below.
6247 * This attaches all cpus from the cpumasks to the NULL domain,
6248 * waits for a RCU quiescent period, recalculates sched
6249 * domain information and then attaches them back to the
6250 * correct sched domains
6251 * Call with hotplug lock held
6253 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6255 cpumask_t change_map;
6258 cpus_and(*partition1, *partition1, cpu_online_map);
6259 cpus_and(*partition2, *partition2, cpu_online_map);
6260 cpus_or(change_map, *partition1, *partition2);
6262 /* Detach sched domains from all of the affected cpus */
6263 detach_destroy_domains(&change_map);
6264 if (!cpus_empty(*partition1))
6265 err = build_sched_domains(partition1);
6266 if (!err && !cpus_empty(*partition2))
6267 err = build_sched_domains(partition2);
6272 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6273 int arch_reinit_sched_domains(void)
6277 mutex_lock(&sched_hotcpu_mutex);
6278 detach_destroy_domains(&cpu_online_map);
6279 err = arch_init_sched_domains(&cpu_online_map);
6280 mutex_unlock(&sched_hotcpu_mutex);
6285 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6289 if (buf[0] != '0' && buf[0] != '1')
6293 sched_smt_power_savings = (buf[0] == '1');
6295 sched_mc_power_savings = (buf[0] == '1');
6297 ret = arch_reinit_sched_domains();
6299 return ret ? ret : count;
6302 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6306 #ifdef CONFIG_SCHED_SMT
6308 err = sysfs_create_file(&cls->kset.kobj,
6309 &attr_sched_smt_power_savings.attr);
6311 #ifdef CONFIG_SCHED_MC
6312 if (!err && mc_capable())
6313 err = sysfs_create_file(&cls->kset.kobj,
6314 &attr_sched_mc_power_savings.attr);
6320 #ifdef CONFIG_SCHED_MC
6321 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6323 return sprintf(page, "%u\n", sched_mc_power_savings);
6325 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6326 const char *buf, size_t count)
6328 return sched_power_savings_store(buf, count, 0);
6330 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6331 sched_mc_power_savings_store);
6334 #ifdef CONFIG_SCHED_SMT
6335 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6337 return sprintf(page, "%u\n", sched_smt_power_savings);
6339 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6340 const char *buf, size_t count)
6342 return sched_power_savings_store(buf, count, 1);
6344 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6345 sched_smt_power_savings_store);
6349 * Force a reinitialization of the sched domains hierarchy. The domains
6350 * and groups cannot be updated in place without racing with the balancing
6351 * code, so we temporarily attach all running cpus to the NULL domain
6352 * which will prevent rebalancing while the sched domains are recalculated.
6354 static int update_sched_domains(struct notifier_block *nfb,
6355 unsigned long action, void *hcpu)
6358 case CPU_UP_PREPARE:
6359 case CPU_UP_PREPARE_FROZEN:
6360 case CPU_DOWN_PREPARE:
6361 case CPU_DOWN_PREPARE_FROZEN:
6362 detach_destroy_domains(&cpu_online_map);
6365 case CPU_UP_CANCELED:
6366 case CPU_UP_CANCELED_FROZEN:
6367 case CPU_DOWN_FAILED:
6368 case CPU_DOWN_FAILED_FROZEN:
6370 case CPU_ONLINE_FROZEN:
6372 case CPU_DEAD_FROZEN:
6374 * Fall through and re-initialise the domains.
6381 /* The hotplug lock is already held by cpu_up/cpu_down */
6382 arch_init_sched_domains(&cpu_online_map);
6387 void __init sched_init_smp(void)
6389 cpumask_t non_isolated_cpus;
6391 mutex_lock(&sched_hotcpu_mutex);
6392 arch_init_sched_domains(&cpu_online_map);
6393 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6394 if (cpus_empty(non_isolated_cpus))
6395 cpu_set(smp_processor_id(), non_isolated_cpus);
6396 mutex_unlock(&sched_hotcpu_mutex);
6397 /* XXX: Theoretical race here - CPU may be hotplugged now */
6398 hotcpu_notifier(update_sched_domains, 0);
6400 /* Move init over to a non-isolated CPU */
6401 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6405 void __init sched_init_smp(void)
6408 #endif /* CONFIG_SMP */
6410 int in_sched_functions(unsigned long addr)
6412 /* Linker adds these: start and end of __sched functions */
6413 extern char __sched_text_start[], __sched_text_end[];
6415 return in_lock_functions(addr) ||
6416 (addr >= (unsigned long)__sched_text_start
6417 && addr < (unsigned long)__sched_text_end);
6420 void __init sched_init(void)
6423 int highest_cpu = 0;
6425 for_each_possible_cpu(i) {
6426 struct prio_array *array;
6430 spin_lock_init(&rq->lock);
6431 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6433 rq->active = rq->arrays;
6434 rq->expired = rq->arrays + 1;
6435 rq->best_expired_prio = MAX_PRIO;
6439 for (j = 1; j < 3; j++)
6440 rq->cpu_load[j] = 0;
6441 rq->active_balance = 0;
6444 rq->migration_thread = NULL;
6445 INIT_LIST_HEAD(&rq->migration_queue);
6447 atomic_set(&rq->nr_iowait, 0);
6449 for (j = 0; j < 2; j++) {
6450 array = rq->arrays + j;
6451 for (k = 0; k < MAX_PRIO; k++) {
6452 INIT_LIST_HEAD(array->queue + k);
6453 __clear_bit(k, array->bitmap);
6455 // delimiter for bitsearch
6456 __set_bit(MAX_PRIO, array->bitmap);
6461 set_load_weight(&init_task);
6464 nr_cpu_ids = highest_cpu + 1;
6465 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6468 #ifdef CONFIG_RT_MUTEXES
6469 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6473 * The boot idle thread does lazy MMU switching as well:
6475 atomic_inc(&init_mm.mm_count);
6476 enter_lazy_tlb(&init_mm, current);
6479 * Make us the idle thread. Technically, schedule() should not be
6480 * called from this thread, however somewhere below it might be,
6481 * but because we are the idle thread, we just pick up running again
6482 * when this runqueue becomes "idle".
6484 init_idle(current, smp_processor_id());
6487 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6488 void __might_sleep(char *file, int line)
6491 static unsigned long prev_jiffy; /* ratelimiting */
6493 if ((in_atomic() || irqs_disabled()) &&
6494 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6495 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6497 prev_jiffy = jiffies;
6498 printk(KERN_ERR "BUG: sleeping function called from invalid"
6499 " context at %s:%d\n", file, line);
6500 printk("in_atomic():%d, irqs_disabled():%d\n",
6501 in_atomic(), irqs_disabled());
6502 debug_show_held_locks(current);
6503 if (irqs_disabled())
6504 print_irqtrace_events(current);
6509 EXPORT_SYMBOL(__might_sleep);
6512 #ifdef CONFIG_MAGIC_SYSRQ
6513 void normalize_rt_tasks(void)
6515 struct prio_array *array;
6516 struct task_struct *g, *p;
6517 unsigned long flags;
6520 read_lock_irq(&tasklist_lock);
6522 do_each_thread(g, p) {
6526 spin_lock_irqsave(&p->pi_lock, flags);
6527 rq = __task_rq_lock(p);
6531 deactivate_task(p, task_rq(p));
6532 __setscheduler(p, SCHED_NORMAL, 0);
6534 __activate_task(p, task_rq(p));
6535 resched_task(rq->curr);
6538 __task_rq_unlock(rq);
6539 spin_unlock_irqrestore(&p->pi_lock, flags);
6540 } while_each_thread(g, p);
6542 read_unlock_irq(&tasklist_lock);
6545 #endif /* CONFIG_MAGIC_SYSRQ */
6549 * These functions are only useful for the IA64 MCA handling.
6551 * They can only be called when the whole system has been
6552 * stopped - every CPU needs to be quiescent, and no scheduling
6553 * activity can take place. Using them for anything else would
6554 * be a serious bug, and as a result, they aren't even visible
6555 * under any other configuration.
6559 * curr_task - return the current task for a given cpu.
6560 * @cpu: the processor in question.
6562 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6564 struct task_struct *curr_task(int cpu)
6566 return cpu_curr(cpu);
6570 * set_curr_task - set the current task for a given cpu.
6571 * @cpu: the processor in question.
6572 * @p: the task pointer to set.
6574 * Description: This function must only be used when non-maskable interrupts
6575 * are serviced on a separate stack. It allows the architecture to switch the
6576 * notion of the current task on a cpu in a non-blocking manner. This function
6577 * must be called with all CPU's synchronized, and interrupts disabled, the
6578 * and caller must save the original value of the current task (see
6579 * curr_task() above) and restore that value before reenabling interrupts and
6580 * re-starting the system.
6582 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6584 void set_curr_task(int cpu, struct task_struct *p)