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 #ifndef prepare_arch_switch
453 # define prepare_arch_switch(next) do { } while (0)
455 #ifndef finish_arch_switch
456 # define finish_arch_switch(prev) do { } while (0)
459 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
460 static inline int task_running(struct rq *rq, struct task_struct *p)
462 return rq->curr == p;
465 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
469 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
471 #ifdef CONFIG_DEBUG_SPINLOCK
472 /* this is a valid case when another task releases the spinlock */
473 rq->lock.owner = current;
476 * If we are tracking spinlock dependencies then we have to
477 * fix up the runqueue lock - which gets 'carried over' from
480 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
482 spin_unlock_irq(&rq->lock);
485 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
486 static inline int task_running(struct rq *rq, struct task_struct *p)
491 return rq->curr == p;
495 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
499 * We can optimise this out completely for !SMP, because the
500 * SMP rebalancing from interrupt is the only thing that cares
505 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
506 spin_unlock_irq(&rq->lock);
508 spin_unlock(&rq->lock);
512 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
516 * After ->oncpu is cleared, the task can be moved to a different CPU.
517 * We must ensure this doesn't happen until the switch is completely
523 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
527 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
530 * __task_rq_lock - lock the runqueue a given task resides on.
531 * Must be called interrupts disabled.
533 static inline struct rq *__task_rq_lock(struct task_struct *p)
540 spin_lock(&rq->lock);
541 if (unlikely(rq != task_rq(p))) {
542 spin_unlock(&rq->lock);
543 goto repeat_lock_task;
549 * task_rq_lock - lock the runqueue a given task resides on and disable
550 * interrupts. Note the ordering: we can safely lookup the task_rq without
551 * explicitly disabling preemption.
553 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
559 local_irq_save(*flags);
561 spin_lock(&rq->lock);
562 if (unlikely(rq != task_rq(p))) {
563 spin_unlock_irqrestore(&rq->lock, *flags);
564 goto repeat_lock_task;
569 static inline void __task_rq_unlock(struct rq *rq)
572 spin_unlock(&rq->lock);
575 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
578 spin_unlock_irqrestore(&rq->lock, *flags);
582 * this_rq_lock - lock this runqueue and disable interrupts.
584 static inline struct rq *this_rq_lock(void)
591 spin_lock(&rq->lock);
596 #include "sched_stats.h"
599 * Adding/removing a task to/from a priority array:
601 static void dequeue_task(struct task_struct *p, struct prio_array *array)
604 list_del(&p->run_list);
605 if (list_empty(array->queue + p->prio))
606 __clear_bit(p->prio, array->bitmap);
609 static void enqueue_task(struct task_struct *p, struct prio_array *array)
611 sched_info_queued(p);
612 list_add_tail(&p->run_list, array->queue + p->prio);
613 __set_bit(p->prio, array->bitmap);
619 * Put task to the end of the run list without the overhead of dequeue
620 * followed by enqueue.
622 static void requeue_task(struct task_struct *p, struct prio_array *array)
624 list_move_tail(&p->run_list, array->queue + p->prio);
628 enqueue_task_head(struct task_struct *p, struct prio_array *array)
630 list_add(&p->run_list, array->queue + p->prio);
631 __set_bit(p->prio, array->bitmap);
637 * __normal_prio - return the priority that is based on the static
638 * priority but is modified by bonuses/penalties.
640 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
641 * into the -5 ... 0 ... +5 bonus/penalty range.
643 * We use 25% of the full 0...39 priority range so that:
645 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
646 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
648 * Both properties are important to certain workloads.
651 static inline int __normal_prio(struct task_struct *p)
655 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
657 prio = p->static_prio - bonus;
658 if (prio < MAX_RT_PRIO)
660 if (prio > MAX_PRIO-1)
666 * To aid in avoiding the subversion of "niceness" due to uneven distribution
667 * of tasks with abnormal "nice" values across CPUs the contribution that
668 * each task makes to its run queue's load is weighted according to its
669 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
670 * scaled version of the new time slice allocation that they receive on time
675 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
676 * If static_prio_timeslice() is ever changed to break this assumption then
677 * this code will need modification
679 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
680 #define LOAD_WEIGHT(lp) \
681 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
682 #define PRIO_TO_LOAD_WEIGHT(prio) \
683 LOAD_WEIGHT(static_prio_timeslice(prio))
684 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
685 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
687 static void set_load_weight(struct task_struct *p)
689 if (has_rt_policy(p)) {
691 if (p == task_rq(p)->migration_thread)
693 * The migration thread does the actual balancing.
694 * Giving its load any weight will skew balancing
700 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
702 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
706 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
708 rq->raw_weighted_load += p->load_weight;
712 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
714 rq->raw_weighted_load -= p->load_weight;
717 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
720 inc_raw_weighted_load(rq, p);
723 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
726 dec_raw_weighted_load(rq, p);
730 * Calculate the expected normal priority: i.e. priority
731 * without taking RT-inheritance into account. Might be
732 * boosted by interactivity modifiers. Changes upon fork,
733 * setprio syscalls, and whenever the interactivity
734 * estimator recalculates.
736 static inline int normal_prio(struct task_struct *p)
740 if (has_rt_policy(p))
741 prio = MAX_RT_PRIO-1 - p->rt_priority;
743 prio = __normal_prio(p);
748 * Calculate the current priority, i.e. the priority
749 * taken into account by the scheduler. This value might
750 * be boosted by RT tasks, or might be boosted by
751 * interactivity modifiers. Will be RT if the task got
752 * RT-boosted. If not then it returns p->normal_prio.
754 static int effective_prio(struct task_struct *p)
756 p->normal_prio = normal_prio(p);
758 * If we are RT tasks or we were boosted to RT priority,
759 * keep the priority unchanged. Otherwise, update priority
760 * to the normal priority:
762 if (!rt_prio(p->prio))
763 return p->normal_prio;
768 * __activate_task - move a task to the runqueue.
770 static void __activate_task(struct task_struct *p, struct rq *rq)
772 struct prio_array *target = rq->active;
775 target = rq->expired;
776 enqueue_task(p, target);
777 inc_nr_running(p, rq);
781 * __activate_idle_task - move idle task to the _front_ of runqueue.
783 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
785 enqueue_task_head(p, rq->active);
786 inc_nr_running(p, rq);
790 * Recalculate p->normal_prio and p->prio after having slept,
791 * updating the sleep-average too:
793 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
795 /* Caller must always ensure 'now >= p->timestamp' */
796 unsigned long sleep_time = now - p->timestamp;
801 if (likely(sleep_time > 0)) {
803 * This ceiling is set to the lowest priority that would allow
804 * a task to be reinserted into the active array on timeslice
807 unsigned long ceiling = INTERACTIVE_SLEEP(p);
809 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
811 * Prevents user tasks from achieving best priority
812 * with one single large enough sleep.
814 p->sleep_avg = ceiling;
816 * Using INTERACTIVE_SLEEP() as a ceiling places a
817 * nice(0) task 1ms sleep away from promotion, and
818 * gives it 700ms to round-robin with no chance of
819 * being demoted. This is more than generous, so
820 * mark this sleep as non-interactive to prevent the
821 * on-runqueue bonus logic from intervening should
822 * this task not receive cpu immediately.
824 p->sleep_type = SLEEP_NONINTERACTIVE;
827 * Tasks waking from uninterruptible sleep are
828 * limited in their sleep_avg rise as they
829 * are likely to be waiting on I/O
831 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
832 if (p->sleep_avg >= ceiling)
834 else if (p->sleep_avg + sleep_time >=
836 p->sleep_avg = ceiling;
842 * This code gives a bonus to interactive tasks.
844 * The boost works by updating the 'average sleep time'
845 * value here, based on ->timestamp. The more time a
846 * task spends sleeping, the higher the average gets -
847 * and the higher the priority boost gets as well.
849 p->sleep_avg += sleep_time;
852 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
853 p->sleep_avg = NS_MAX_SLEEP_AVG;
856 return effective_prio(p);
860 * activate_task - move a task to the runqueue and do priority recalculation
862 * Update all the scheduling statistics stuff. (sleep average
863 * calculation, priority modifiers, etc.)
865 static void activate_task(struct task_struct *p, struct rq *rq, int local)
867 unsigned long long now;
875 /* Compensate for drifting sched_clock */
876 struct rq *this_rq = this_rq();
877 now = (now - this_rq->most_recent_timestamp)
878 + rq->most_recent_timestamp;
883 * Sleep time is in units of nanosecs, so shift by 20 to get a
884 * milliseconds-range estimation of the amount of time that the task
887 if (unlikely(prof_on == SLEEP_PROFILING)) {
888 if (p->state == TASK_UNINTERRUPTIBLE)
889 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
890 (now - p->timestamp) >> 20);
893 p->prio = recalc_task_prio(p, now);
896 * This checks to make sure it's not an uninterruptible task
897 * that is now waking up.
899 if (p->sleep_type == SLEEP_NORMAL) {
901 * Tasks which were woken up by interrupts (ie. hw events)
902 * are most likely of interactive nature. So we give them
903 * the credit of extending their sleep time to the period
904 * of time they spend on the runqueue, waiting for execution
905 * on a CPU, first time around:
908 p->sleep_type = SLEEP_INTERRUPTED;
911 * Normal first-time wakeups get a credit too for
912 * on-runqueue time, but it will be weighted down:
914 p->sleep_type = SLEEP_INTERACTIVE;
919 __activate_task(p, rq);
923 * deactivate_task - remove a task from the runqueue.
925 static void deactivate_task(struct task_struct *p, struct rq *rq)
927 dec_nr_running(p, rq);
928 dequeue_task(p, p->array);
933 * resched_task - mark a task 'to be rescheduled now'.
935 * On UP this means the setting of the need_resched flag, on SMP it
936 * might also involve a cross-CPU call to trigger the scheduler on
941 #ifndef tsk_is_polling
942 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
945 static void resched_task(struct task_struct *p)
949 assert_spin_locked(&task_rq(p)->lock);
951 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
954 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
957 if (cpu == smp_processor_id())
960 /* NEED_RESCHED must be visible before we test polling */
962 if (!tsk_is_polling(p))
963 smp_send_reschedule(cpu);
966 static void resched_cpu(int cpu)
968 struct rq *rq = cpu_rq(cpu);
971 if (!spin_trylock_irqsave(&rq->lock, flags))
973 resched_task(cpu_curr(cpu));
974 spin_unlock_irqrestore(&rq->lock, flags);
977 static inline void resched_task(struct task_struct *p)
979 assert_spin_locked(&task_rq(p)->lock);
980 set_tsk_need_resched(p);
985 * task_curr - is this task currently executing on a CPU?
986 * @p: the task in question.
988 inline int task_curr(const struct task_struct *p)
990 return cpu_curr(task_cpu(p)) == p;
993 /* Used instead of source_load when we know the type == 0 */
994 unsigned long weighted_cpuload(const int cpu)
996 return cpu_rq(cpu)->raw_weighted_load;
1001 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1003 task_thread_info(p)->cpu = cpu;
1006 struct migration_req {
1007 struct list_head list;
1009 struct task_struct *task;
1012 struct completion done;
1016 * The task's runqueue lock must be held.
1017 * Returns true if you have to wait for migration thread.
1020 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1022 struct rq *rq = task_rq(p);
1025 * If the task is not on a runqueue (and not running), then
1026 * it is sufficient to simply update the task's cpu field.
1028 if (!p->array && !task_running(rq, p)) {
1029 set_task_cpu(p, dest_cpu);
1033 init_completion(&req->done);
1035 req->dest_cpu = dest_cpu;
1036 list_add(&req->list, &rq->migration_queue);
1042 * wait_task_inactive - wait for a thread to unschedule.
1044 * The caller must ensure that the task *will* unschedule sometime soon,
1045 * else this function might spin for a *long* time. This function can't
1046 * be called with interrupts off, or it may introduce deadlock with
1047 * smp_call_function() if an IPI is sent by the same process we are
1048 * waiting to become inactive.
1050 void wait_task_inactive(struct task_struct *p)
1052 unsigned long flags;
1054 struct prio_array *array;
1059 * We do the initial early heuristics without holding
1060 * any task-queue locks at all. We'll only try to get
1061 * the runqueue lock when things look like they will
1067 * If the task is actively running on another CPU
1068 * still, just relax and busy-wait without holding
1071 * NOTE! Since we don't hold any locks, it's not
1072 * even sure that "rq" stays as the right runqueue!
1073 * But we don't care, since "task_running()" will
1074 * return false if the runqueue has changed and p
1075 * is actually now running somewhere else!
1077 while (task_running(rq, p))
1081 * Ok, time to look more closely! We need the rq
1082 * lock now, to be *sure*. If we're wrong, we'll
1083 * just go back and repeat.
1085 rq = task_rq_lock(p, &flags);
1086 running = task_running(rq, p);
1088 task_rq_unlock(rq, &flags);
1091 * Was it really running after all now that we
1092 * checked with the proper locks actually held?
1094 * Oops. Go back and try again..
1096 if (unlikely(running)) {
1102 * It's not enough that it's not actively running,
1103 * it must be off the runqueue _entirely_, and not
1106 * So if it wa still runnable (but just not actively
1107 * running right now), it's preempted, and we should
1108 * yield - it could be a while.
1110 if (unlikely(array)) {
1116 * Ahh, all good. It wasn't running, and it wasn't
1117 * runnable, which means that it will never become
1118 * running in the future either. We're all done!
1123 * kick_process - kick a running thread to enter/exit the kernel
1124 * @p: the to-be-kicked thread
1126 * Cause a process which is running on another CPU to enter
1127 * kernel-mode, without any delay. (to get signals handled.)
1129 * NOTE: this function doesnt have to take the runqueue lock,
1130 * because all it wants to ensure is that the remote task enters
1131 * the kernel. If the IPI races and the task has been migrated
1132 * to another CPU then no harm is done and the purpose has been
1135 void kick_process(struct task_struct *p)
1141 if ((cpu != smp_processor_id()) && task_curr(p))
1142 smp_send_reschedule(cpu);
1147 * Return a low guess at the load of a migration-source cpu weighted
1148 * according to the scheduling class and "nice" value.
1150 * We want to under-estimate the load of migration sources, to
1151 * balance conservatively.
1153 static inline unsigned long source_load(int cpu, int type)
1155 struct rq *rq = cpu_rq(cpu);
1158 return rq->raw_weighted_load;
1160 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1164 * Return a high guess at the load of a migration-target cpu weighted
1165 * according to the scheduling class and "nice" value.
1167 static inline unsigned long target_load(int cpu, int type)
1169 struct rq *rq = cpu_rq(cpu);
1172 return rq->raw_weighted_load;
1174 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1178 * Return the average load per task on the cpu's run queue
1180 static inline unsigned long cpu_avg_load_per_task(int cpu)
1182 struct rq *rq = cpu_rq(cpu);
1183 unsigned long n = rq->nr_running;
1185 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1189 * find_idlest_group finds and returns the least busy CPU group within the
1192 static struct sched_group *
1193 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1195 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1196 unsigned long min_load = ULONG_MAX, this_load = 0;
1197 int load_idx = sd->forkexec_idx;
1198 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1201 unsigned long load, avg_load;
1205 /* Skip over this group if it has no CPUs allowed */
1206 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1209 local_group = cpu_isset(this_cpu, group->cpumask);
1211 /* Tally up the load of all CPUs in the group */
1214 for_each_cpu_mask(i, group->cpumask) {
1215 /* Bias balancing toward cpus of our domain */
1217 load = source_load(i, load_idx);
1219 load = target_load(i, load_idx);
1224 /* Adjust by relative CPU power of the group */
1225 avg_load = sg_div_cpu_power(group,
1226 avg_load * SCHED_LOAD_SCALE);
1229 this_load = avg_load;
1231 } else if (avg_load < min_load) {
1232 min_load = avg_load;
1236 group = group->next;
1237 } while (group != sd->groups);
1239 if (!idlest || 100*this_load < imbalance*min_load)
1245 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1248 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1251 unsigned long load, min_load = ULONG_MAX;
1255 /* Traverse only the allowed CPUs */
1256 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1258 for_each_cpu_mask(i, tmp) {
1259 load = weighted_cpuload(i);
1261 if (load < min_load || (load == min_load && i == this_cpu)) {
1271 * sched_balance_self: balance the current task (running on cpu) in domains
1272 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1275 * Balance, ie. select the least loaded group.
1277 * Returns the target CPU number, or the same CPU if no balancing is needed.
1279 * preempt must be disabled.
1281 static int sched_balance_self(int cpu, int flag)
1283 struct task_struct *t = current;
1284 struct sched_domain *tmp, *sd = NULL;
1286 for_each_domain(cpu, tmp) {
1288 * If power savings logic is enabled for a domain, stop there.
1290 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1292 if (tmp->flags & flag)
1298 struct sched_group *group;
1299 int new_cpu, weight;
1301 if (!(sd->flags & flag)) {
1307 group = find_idlest_group(sd, t, cpu);
1313 new_cpu = find_idlest_cpu(group, t, cpu);
1314 if (new_cpu == -1 || new_cpu == cpu) {
1315 /* Now try balancing at a lower domain level of cpu */
1320 /* Now try balancing at a lower domain level of new_cpu */
1323 weight = cpus_weight(span);
1324 for_each_domain(cpu, tmp) {
1325 if (weight <= cpus_weight(tmp->span))
1327 if (tmp->flags & flag)
1330 /* while loop will break here if sd == NULL */
1336 #endif /* CONFIG_SMP */
1339 * wake_idle() will wake a task on an idle cpu if task->cpu is
1340 * not idle and an idle cpu is available. The span of cpus to
1341 * search starts with cpus closest then further out as needed,
1342 * so we always favor a closer, idle cpu.
1344 * Returns the CPU we should wake onto.
1346 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1347 static int wake_idle(int cpu, struct task_struct *p)
1350 struct sched_domain *sd;
1354 * If it is idle, then it is the best cpu to run this task.
1356 * This cpu is also the best, if it has more than one task already.
1357 * Siblings must be also busy(in most cases) as they didn't already
1358 * pickup the extra load from this cpu and hence we need not check
1359 * sibling runqueue info. This will avoid the checks and cache miss
1360 * penalities associated with that.
1362 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1365 for_each_domain(cpu, sd) {
1366 if (sd->flags & SD_WAKE_IDLE) {
1367 cpus_and(tmp, sd->span, p->cpus_allowed);
1368 for_each_cpu_mask(i, tmp) {
1379 static inline int wake_idle(int cpu, struct task_struct *p)
1386 * try_to_wake_up - wake up a thread
1387 * @p: the to-be-woken-up thread
1388 * @state: the mask of task states that can be woken
1389 * @sync: do a synchronous wakeup?
1391 * Put it on the run-queue if it's not already there. The "current"
1392 * thread is always on the run-queue (except when the actual
1393 * re-schedule is in progress), and as such you're allowed to do
1394 * the simpler "current->state = TASK_RUNNING" to mark yourself
1395 * runnable without the overhead of this.
1397 * returns failure only if the task is already active.
1399 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1401 int cpu, this_cpu, success = 0;
1402 unsigned long flags;
1406 struct sched_domain *sd, *this_sd = NULL;
1407 unsigned long load, this_load;
1411 rq = task_rq_lock(p, &flags);
1412 old_state = p->state;
1413 if (!(old_state & state))
1420 this_cpu = smp_processor_id();
1423 if (unlikely(task_running(rq, p)))
1428 schedstat_inc(rq, ttwu_cnt);
1429 if (cpu == this_cpu) {
1430 schedstat_inc(rq, ttwu_local);
1434 for_each_domain(this_cpu, sd) {
1435 if (cpu_isset(cpu, sd->span)) {
1436 schedstat_inc(sd, ttwu_wake_remote);
1442 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1446 * Check for affine wakeup and passive balancing possibilities.
1449 int idx = this_sd->wake_idx;
1450 unsigned int imbalance;
1452 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1454 load = source_load(cpu, idx);
1455 this_load = target_load(this_cpu, idx);
1457 new_cpu = this_cpu; /* Wake to this CPU if we can */
1459 if (this_sd->flags & SD_WAKE_AFFINE) {
1460 unsigned long tl = this_load;
1461 unsigned long tl_per_task;
1463 tl_per_task = cpu_avg_load_per_task(this_cpu);
1466 * If sync wakeup then subtract the (maximum possible)
1467 * effect of the currently running task from the load
1468 * of the current CPU:
1471 tl -= current->load_weight;
1474 tl + target_load(cpu, idx) <= tl_per_task) ||
1475 100*(tl + p->load_weight) <= imbalance*load) {
1477 * This domain has SD_WAKE_AFFINE and
1478 * p is cache cold in this domain, and
1479 * there is no bad imbalance.
1481 schedstat_inc(this_sd, ttwu_move_affine);
1487 * Start passive balancing when half the imbalance_pct
1490 if (this_sd->flags & SD_WAKE_BALANCE) {
1491 if (imbalance*this_load <= 100*load) {
1492 schedstat_inc(this_sd, ttwu_move_balance);
1498 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1500 new_cpu = wake_idle(new_cpu, p);
1501 if (new_cpu != cpu) {
1502 set_task_cpu(p, new_cpu);
1503 task_rq_unlock(rq, &flags);
1504 /* might preempt at this point */
1505 rq = task_rq_lock(p, &flags);
1506 old_state = p->state;
1507 if (!(old_state & state))
1512 this_cpu = smp_processor_id();
1517 #endif /* CONFIG_SMP */
1518 if (old_state == TASK_UNINTERRUPTIBLE) {
1519 rq->nr_uninterruptible--;
1521 * Tasks on involuntary sleep don't earn
1522 * sleep_avg beyond just interactive state.
1524 p->sleep_type = SLEEP_NONINTERACTIVE;
1528 * Tasks that have marked their sleep as noninteractive get
1529 * woken up with their sleep average not weighted in an
1532 if (old_state & TASK_NONINTERACTIVE)
1533 p->sleep_type = SLEEP_NONINTERACTIVE;
1536 activate_task(p, rq, cpu == this_cpu);
1538 * Sync wakeups (i.e. those types of wakeups where the waker
1539 * has indicated that it will leave the CPU in short order)
1540 * don't trigger a preemption, if the woken up task will run on
1541 * this cpu. (in this case the 'I will reschedule' promise of
1542 * the waker guarantees that the freshly woken up task is going
1543 * to be considered on this CPU.)
1545 if (!sync || cpu != this_cpu) {
1546 if (TASK_PREEMPTS_CURR(p, rq))
1547 resched_task(rq->curr);
1552 p->state = TASK_RUNNING;
1554 task_rq_unlock(rq, &flags);
1559 int fastcall wake_up_process(struct task_struct *p)
1561 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1562 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1564 EXPORT_SYMBOL(wake_up_process);
1566 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1568 return try_to_wake_up(p, state, 0);
1571 static void task_running_tick(struct rq *rq, struct task_struct *p);
1573 * Perform scheduler related setup for a newly forked process p.
1574 * p is forked by current.
1576 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1578 int cpu = get_cpu();
1581 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1583 set_task_cpu(p, cpu);
1586 * We mark the process as running here, but have not actually
1587 * inserted it onto the runqueue yet. This guarantees that
1588 * nobody will actually run it, and a signal or other external
1589 * event cannot wake it up and insert it on the runqueue either.
1591 p->state = TASK_RUNNING;
1594 * Make sure we do not leak PI boosting priority to the child:
1596 p->prio = current->normal_prio;
1598 INIT_LIST_HEAD(&p->run_list);
1600 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1601 if (unlikely(sched_info_on()))
1602 memset(&p->sched_info, 0, sizeof(p->sched_info));
1604 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1607 #ifdef CONFIG_PREEMPT
1608 /* Want to start with kernel preemption disabled. */
1609 task_thread_info(p)->preempt_count = 1;
1612 * Share the timeslice between parent and child, thus the
1613 * total amount of pending timeslices in the system doesn't change,
1614 * resulting in more scheduling fairness.
1616 local_irq_disable();
1617 p->time_slice = (current->time_slice + 1) >> 1;
1619 * The remainder of the first timeslice might be recovered by
1620 * the parent if the child exits early enough.
1622 p->first_time_slice = 1;
1623 current->time_slice >>= 1;
1624 p->timestamp = sched_clock();
1625 if (unlikely(!current->time_slice)) {
1627 * This case is rare, it happens when the parent has only
1628 * a single jiffy left from its timeslice. Taking the
1629 * runqueue lock is not a problem.
1631 current->time_slice = 1;
1632 task_running_tick(cpu_rq(cpu), current);
1639 * wake_up_new_task - wake up a newly created task for the first time.
1641 * This function will do some initial scheduler statistics housekeeping
1642 * that must be done for every newly created context, then puts the task
1643 * on the runqueue and wakes it.
1645 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1647 struct rq *rq, *this_rq;
1648 unsigned long flags;
1651 rq = task_rq_lock(p, &flags);
1652 BUG_ON(p->state != TASK_RUNNING);
1653 this_cpu = smp_processor_id();
1657 * We decrease the sleep average of forking parents
1658 * and children as well, to keep max-interactive tasks
1659 * from forking tasks that are max-interactive. The parent
1660 * (current) is done further down, under its lock.
1662 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1663 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1665 p->prio = effective_prio(p);
1667 if (likely(cpu == this_cpu)) {
1668 if (!(clone_flags & CLONE_VM)) {
1670 * The VM isn't cloned, so we're in a good position to
1671 * do child-runs-first in anticipation of an exec. This
1672 * usually avoids a lot of COW overhead.
1674 if (unlikely(!current->array))
1675 __activate_task(p, rq);
1677 p->prio = current->prio;
1678 p->normal_prio = current->normal_prio;
1679 list_add_tail(&p->run_list, ¤t->run_list);
1680 p->array = current->array;
1681 p->array->nr_active++;
1682 inc_nr_running(p, rq);
1686 /* Run child last */
1687 __activate_task(p, rq);
1689 * We skip the following code due to cpu == this_cpu
1691 * task_rq_unlock(rq, &flags);
1692 * this_rq = task_rq_lock(current, &flags);
1696 this_rq = cpu_rq(this_cpu);
1699 * Not the local CPU - must adjust timestamp. This should
1700 * get optimised away in the !CONFIG_SMP case.
1702 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1703 + rq->most_recent_timestamp;
1704 __activate_task(p, rq);
1705 if (TASK_PREEMPTS_CURR(p, rq))
1706 resched_task(rq->curr);
1709 * Parent and child are on different CPUs, now get the
1710 * parent runqueue to update the parent's ->sleep_avg:
1712 task_rq_unlock(rq, &flags);
1713 this_rq = task_rq_lock(current, &flags);
1715 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1716 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1717 task_rq_unlock(this_rq, &flags);
1721 * prepare_task_switch - prepare to switch tasks
1722 * @rq: the runqueue preparing to switch
1723 * @next: the task we are going to switch to.
1725 * This is called with the rq lock held and interrupts off. It must
1726 * be paired with a subsequent finish_task_switch after the context
1729 * prepare_task_switch sets up locking and calls architecture specific
1732 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1734 prepare_lock_switch(rq, next);
1735 prepare_arch_switch(next);
1739 * finish_task_switch - clean up after a task-switch
1740 * @rq: runqueue associated with task-switch
1741 * @prev: the thread we just switched away from.
1743 * finish_task_switch must be called after the context switch, paired
1744 * with a prepare_task_switch call before the context switch.
1745 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1746 * and do any other architecture-specific cleanup actions.
1748 * Note that we may have delayed dropping an mm in context_switch(). If
1749 * so, we finish that here outside of the runqueue lock. (Doing it
1750 * with the lock held can cause deadlocks; see schedule() for
1753 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1754 __releases(rq->lock)
1756 struct mm_struct *mm = rq->prev_mm;
1762 * A task struct has one reference for the use as "current".
1763 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1764 * schedule one last time. The schedule call will never return, and
1765 * the scheduled task must drop that reference.
1766 * The test for TASK_DEAD must occur while the runqueue locks are
1767 * still held, otherwise prev could be scheduled on another cpu, die
1768 * there before we look at prev->state, and then the reference would
1770 * Manfred Spraul <manfred@colorfullife.com>
1772 prev_state = prev->state;
1773 finish_arch_switch(prev);
1774 finish_lock_switch(rq, prev);
1777 if (unlikely(prev_state == TASK_DEAD)) {
1779 * Remove function-return probe instances associated with this
1780 * task and put them back on the free list.
1782 kprobe_flush_task(prev);
1783 put_task_struct(prev);
1788 * schedule_tail - first thing a freshly forked thread must call.
1789 * @prev: the thread we just switched away from.
1791 asmlinkage void schedule_tail(struct task_struct *prev)
1792 __releases(rq->lock)
1794 struct rq *rq = this_rq();
1796 finish_task_switch(rq, prev);
1797 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1798 /* In this case, finish_task_switch does not reenable preemption */
1801 if (current->set_child_tid)
1802 put_user(current->pid, current->set_child_tid);
1806 * context_switch - switch to the new MM and the new
1807 * thread's register state.
1809 static inline struct task_struct *
1810 context_switch(struct rq *rq, struct task_struct *prev,
1811 struct task_struct *next)
1813 struct mm_struct *mm = next->mm;
1814 struct mm_struct *oldmm = prev->active_mm;
1817 * For paravirt, this is coupled with an exit in switch_to to
1818 * combine the page table reload and the switch backend into
1821 arch_enter_lazy_cpu_mode();
1824 next->active_mm = oldmm;
1825 atomic_inc(&oldmm->mm_count);
1826 enter_lazy_tlb(oldmm, next);
1828 switch_mm(oldmm, mm, next);
1831 prev->active_mm = NULL;
1832 WARN_ON(rq->prev_mm);
1833 rq->prev_mm = oldmm;
1836 * Since the runqueue lock will be released by the next
1837 * task (which is an invalid locking op but in the case
1838 * of the scheduler it's an obvious special-case), so we
1839 * do an early lockdep release here:
1841 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1842 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1845 /* Here we just switch the register state and the stack. */
1846 switch_to(prev, next, prev);
1852 * nr_running, nr_uninterruptible and nr_context_switches:
1854 * externally visible scheduler statistics: current number of runnable
1855 * threads, current number of uninterruptible-sleeping threads, total
1856 * number of context switches performed since bootup.
1858 unsigned long nr_running(void)
1860 unsigned long i, sum = 0;
1862 for_each_online_cpu(i)
1863 sum += cpu_rq(i)->nr_running;
1868 unsigned long nr_uninterruptible(void)
1870 unsigned long i, sum = 0;
1872 for_each_possible_cpu(i)
1873 sum += cpu_rq(i)->nr_uninterruptible;
1876 * Since we read the counters lockless, it might be slightly
1877 * inaccurate. Do not allow it to go below zero though:
1879 if (unlikely((long)sum < 0))
1885 unsigned long long nr_context_switches(void)
1888 unsigned long long sum = 0;
1890 for_each_possible_cpu(i)
1891 sum += cpu_rq(i)->nr_switches;
1896 unsigned long nr_iowait(void)
1898 unsigned long i, sum = 0;
1900 for_each_possible_cpu(i)
1901 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1906 unsigned long nr_active(void)
1908 unsigned long i, running = 0, uninterruptible = 0;
1910 for_each_online_cpu(i) {
1911 running += cpu_rq(i)->nr_running;
1912 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1915 if (unlikely((long)uninterruptible < 0))
1916 uninterruptible = 0;
1918 return running + uninterruptible;
1924 * Is this task likely cache-hot:
1927 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1929 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1933 * double_rq_lock - safely lock two runqueues
1935 * Note this does not disable interrupts like task_rq_lock,
1936 * you need to do so manually before calling.
1938 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1939 __acquires(rq1->lock)
1940 __acquires(rq2->lock)
1942 BUG_ON(!irqs_disabled());
1944 spin_lock(&rq1->lock);
1945 __acquire(rq2->lock); /* Fake it out ;) */
1948 spin_lock(&rq1->lock);
1949 spin_lock(&rq2->lock);
1951 spin_lock(&rq2->lock);
1952 spin_lock(&rq1->lock);
1958 * double_rq_unlock - safely unlock two runqueues
1960 * Note this does not restore interrupts like task_rq_unlock,
1961 * you need to do so manually after calling.
1963 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1964 __releases(rq1->lock)
1965 __releases(rq2->lock)
1967 spin_unlock(&rq1->lock);
1969 spin_unlock(&rq2->lock);
1971 __release(rq2->lock);
1975 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1977 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1978 __releases(this_rq->lock)
1979 __acquires(busiest->lock)
1980 __acquires(this_rq->lock)
1982 if (unlikely(!irqs_disabled())) {
1983 /* printk() doesn't work good under rq->lock */
1984 spin_unlock(&this_rq->lock);
1987 if (unlikely(!spin_trylock(&busiest->lock))) {
1988 if (busiest < this_rq) {
1989 spin_unlock(&this_rq->lock);
1990 spin_lock(&busiest->lock);
1991 spin_lock(&this_rq->lock);
1993 spin_lock(&busiest->lock);
1998 * If dest_cpu is allowed for this process, migrate the task to it.
1999 * This is accomplished by forcing the cpu_allowed mask to only
2000 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2001 * the cpu_allowed mask is restored.
2003 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2005 struct migration_req req;
2006 unsigned long flags;
2009 rq = task_rq_lock(p, &flags);
2010 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2011 || unlikely(cpu_is_offline(dest_cpu)))
2014 /* force the process onto the specified CPU */
2015 if (migrate_task(p, dest_cpu, &req)) {
2016 /* Need to wait for migration thread (might exit: take ref). */
2017 struct task_struct *mt = rq->migration_thread;
2019 get_task_struct(mt);
2020 task_rq_unlock(rq, &flags);
2021 wake_up_process(mt);
2022 put_task_struct(mt);
2023 wait_for_completion(&req.done);
2028 task_rq_unlock(rq, &flags);
2032 * sched_exec - execve() is a valuable balancing opportunity, because at
2033 * this point the task has the smallest effective memory and cache footprint.
2035 void sched_exec(void)
2037 int new_cpu, this_cpu = get_cpu();
2038 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2040 if (new_cpu != this_cpu)
2041 sched_migrate_task(current, new_cpu);
2045 * pull_task - move a task from a remote runqueue to the local runqueue.
2046 * Both runqueues must be locked.
2048 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2049 struct task_struct *p, struct rq *this_rq,
2050 struct prio_array *this_array, int this_cpu)
2052 dequeue_task(p, src_array);
2053 dec_nr_running(p, src_rq);
2054 set_task_cpu(p, this_cpu);
2055 inc_nr_running(p, this_rq);
2056 enqueue_task(p, this_array);
2057 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2058 + this_rq->most_recent_timestamp;
2060 * Note that idle threads have a prio of MAX_PRIO, for this test
2061 * to be always true for them.
2063 if (TASK_PREEMPTS_CURR(p, this_rq))
2064 resched_task(this_rq->curr);
2068 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2071 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2072 struct sched_domain *sd, enum cpu_idle_type idle,
2076 * We do not migrate tasks that are:
2077 * 1) running (obviously), or
2078 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2079 * 3) are cache-hot on their current CPU.
2081 if (!cpu_isset(this_cpu, p->cpus_allowed))
2085 if (task_running(rq, p))
2089 * Aggressive migration if:
2090 * 1) task is cache cold, or
2091 * 2) too many balance attempts have failed.
2094 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2095 #ifdef CONFIG_SCHEDSTATS
2096 if (task_hot(p, rq->most_recent_timestamp, sd))
2097 schedstat_inc(sd, lb_hot_gained[idle]);
2102 if (task_hot(p, rq->most_recent_timestamp, sd))
2107 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2110 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2111 * load from busiest to this_rq, as part of a balancing operation within
2112 * "domain". Returns the number of tasks moved.
2114 * Called with both runqueues locked.
2116 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2117 unsigned long max_nr_move, unsigned long max_load_move,
2118 struct sched_domain *sd, enum cpu_idle_type idle,
2121 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2122 best_prio_seen, skip_for_load;
2123 struct prio_array *array, *dst_array;
2124 struct list_head *head, *curr;
2125 struct task_struct *tmp;
2128 if (max_nr_move == 0 || max_load_move == 0)
2131 rem_load_move = max_load_move;
2133 this_best_prio = rq_best_prio(this_rq);
2134 best_prio = rq_best_prio(busiest);
2136 * Enable handling of the case where there is more than one task
2137 * with the best priority. If the current running task is one
2138 * of those with prio==best_prio we know it won't be moved
2139 * and therefore it's safe to override the skip (based on load) of
2140 * any task we find with that prio.
2142 best_prio_seen = best_prio == busiest->curr->prio;
2145 * We first consider expired tasks. Those will likely not be
2146 * executed in the near future, and they are most likely to
2147 * be cache-cold, thus switching CPUs has the least effect
2150 if (busiest->expired->nr_active) {
2151 array = busiest->expired;
2152 dst_array = this_rq->expired;
2154 array = busiest->active;
2155 dst_array = this_rq->active;
2159 /* Start searching at priority 0: */
2163 idx = sched_find_first_bit(array->bitmap);
2165 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2166 if (idx >= MAX_PRIO) {
2167 if (array == busiest->expired && busiest->active->nr_active) {
2168 array = busiest->active;
2169 dst_array = this_rq->active;
2175 head = array->queue + idx;
2178 tmp = list_entry(curr, struct task_struct, run_list);
2183 * To help distribute high priority tasks accross CPUs we don't
2184 * skip a task if it will be the highest priority task (i.e. smallest
2185 * prio value) on its new queue regardless of its load weight
2187 skip_for_load = tmp->load_weight > rem_load_move;
2188 if (skip_for_load && idx < this_best_prio)
2189 skip_for_load = !best_prio_seen && idx == best_prio;
2190 if (skip_for_load ||
2191 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2193 best_prio_seen |= idx == best_prio;
2200 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2202 rem_load_move -= tmp->load_weight;
2205 * We only want to steal up to the prescribed number of tasks
2206 * and the prescribed amount of weighted load.
2208 if (pulled < max_nr_move && rem_load_move > 0) {
2209 if (idx < this_best_prio)
2210 this_best_prio = idx;
2218 * Right now, this is the only place pull_task() is called,
2219 * so we can safely collect pull_task() stats here rather than
2220 * inside pull_task().
2222 schedstat_add(sd, lb_gained[idle], pulled);
2225 *all_pinned = pinned;
2230 * find_busiest_group finds and returns the busiest CPU group within the
2231 * domain. It calculates and returns the amount of weighted load which
2232 * should be moved to restore balance via the imbalance parameter.
2234 static struct sched_group *
2235 find_busiest_group(struct sched_domain *sd, int this_cpu,
2236 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2237 cpumask_t *cpus, int *balance)
2239 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2240 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2241 unsigned long max_pull;
2242 unsigned long busiest_load_per_task, busiest_nr_running;
2243 unsigned long this_load_per_task, this_nr_running;
2245 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2246 int power_savings_balance = 1;
2247 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2248 unsigned long min_nr_running = ULONG_MAX;
2249 struct sched_group *group_min = NULL, *group_leader = NULL;
2252 max_load = this_load = total_load = total_pwr = 0;
2253 busiest_load_per_task = busiest_nr_running = 0;
2254 this_load_per_task = this_nr_running = 0;
2255 if (idle == CPU_NOT_IDLE)
2256 load_idx = sd->busy_idx;
2257 else if (idle == CPU_NEWLY_IDLE)
2258 load_idx = sd->newidle_idx;
2260 load_idx = sd->idle_idx;
2263 unsigned long load, group_capacity;
2266 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2267 unsigned long sum_nr_running, sum_weighted_load;
2269 local_group = cpu_isset(this_cpu, group->cpumask);
2272 balance_cpu = first_cpu(group->cpumask);
2274 /* Tally up the load of all CPUs in the group */
2275 sum_weighted_load = sum_nr_running = avg_load = 0;
2277 for_each_cpu_mask(i, group->cpumask) {
2280 if (!cpu_isset(i, *cpus))
2285 if (*sd_idle && !idle_cpu(i))
2288 /* Bias balancing toward cpus of our domain */
2290 if (idle_cpu(i) && !first_idle_cpu) {
2295 load = target_load(i, load_idx);
2297 load = source_load(i, load_idx);
2300 sum_nr_running += rq->nr_running;
2301 sum_weighted_load += rq->raw_weighted_load;
2305 * First idle cpu or the first cpu(busiest) in this sched group
2306 * is eligible for doing load balancing at this and above
2309 if (local_group && balance_cpu != this_cpu && balance) {
2314 total_load += avg_load;
2315 total_pwr += group->__cpu_power;
2317 /* Adjust by relative CPU power of the group */
2318 avg_load = sg_div_cpu_power(group,
2319 avg_load * SCHED_LOAD_SCALE);
2321 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2324 this_load = avg_load;
2326 this_nr_running = sum_nr_running;
2327 this_load_per_task = sum_weighted_load;
2328 } else if (avg_load > max_load &&
2329 sum_nr_running > group_capacity) {
2330 max_load = avg_load;
2332 busiest_nr_running = sum_nr_running;
2333 busiest_load_per_task = sum_weighted_load;
2336 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2338 * Busy processors will not participate in power savings
2341 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2345 * If the local group is idle or completely loaded
2346 * no need to do power savings balance at this domain
2348 if (local_group && (this_nr_running >= group_capacity ||
2350 power_savings_balance = 0;
2353 * If a group is already running at full capacity or idle,
2354 * don't include that group in power savings calculations
2356 if (!power_savings_balance || sum_nr_running >= group_capacity
2361 * Calculate the group which has the least non-idle load.
2362 * This is the group from where we need to pick up the load
2365 if ((sum_nr_running < min_nr_running) ||
2366 (sum_nr_running == min_nr_running &&
2367 first_cpu(group->cpumask) <
2368 first_cpu(group_min->cpumask))) {
2370 min_nr_running = sum_nr_running;
2371 min_load_per_task = sum_weighted_load /
2376 * Calculate the group which is almost near its
2377 * capacity but still has some space to pick up some load
2378 * from other group and save more power
2380 if (sum_nr_running <= group_capacity - 1) {
2381 if (sum_nr_running > leader_nr_running ||
2382 (sum_nr_running == leader_nr_running &&
2383 first_cpu(group->cpumask) >
2384 first_cpu(group_leader->cpumask))) {
2385 group_leader = group;
2386 leader_nr_running = sum_nr_running;
2391 group = group->next;
2392 } while (group != sd->groups);
2394 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2397 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2399 if (this_load >= avg_load ||
2400 100*max_load <= sd->imbalance_pct*this_load)
2403 busiest_load_per_task /= busiest_nr_running;
2405 * We're trying to get all the cpus to the average_load, so we don't
2406 * want to push ourselves above the average load, nor do we wish to
2407 * reduce the max loaded cpu below the average load, as either of these
2408 * actions would just result in more rebalancing later, and ping-pong
2409 * tasks around. Thus we look for the minimum possible imbalance.
2410 * Negative imbalances (*we* are more loaded than anyone else) will
2411 * be counted as no imbalance for these purposes -- we can't fix that
2412 * by pulling tasks to us. Be careful of negative numbers as they'll
2413 * appear as very large values with unsigned longs.
2415 if (max_load <= busiest_load_per_task)
2419 * In the presence of smp nice balancing, certain scenarios can have
2420 * max load less than avg load(as we skip the groups at or below
2421 * its cpu_power, while calculating max_load..)
2423 if (max_load < avg_load) {
2425 goto small_imbalance;
2428 /* Don't want to pull so many tasks that a group would go idle */
2429 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2431 /* How much load to actually move to equalise the imbalance */
2432 *imbalance = min(max_pull * busiest->__cpu_power,
2433 (avg_load - this_load) * this->__cpu_power)
2437 * if *imbalance is less than the average load per runnable task
2438 * there is no gaurantee that any tasks will be moved so we'll have
2439 * a think about bumping its value to force at least one task to be
2442 if (*imbalance < busiest_load_per_task) {
2443 unsigned long tmp, pwr_now, pwr_move;
2447 pwr_move = pwr_now = 0;
2449 if (this_nr_running) {
2450 this_load_per_task /= this_nr_running;
2451 if (busiest_load_per_task > this_load_per_task)
2454 this_load_per_task = SCHED_LOAD_SCALE;
2456 if (max_load - this_load >= busiest_load_per_task * imbn) {
2457 *imbalance = busiest_load_per_task;
2462 * OK, we don't have enough imbalance to justify moving tasks,
2463 * however we may be able to increase total CPU power used by
2467 pwr_now += busiest->__cpu_power *
2468 min(busiest_load_per_task, max_load);
2469 pwr_now += this->__cpu_power *
2470 min(this_load_per_task, this_load);
2471 pwr_now /= SCHED_LOAD_SCALE;
2473 /* Amount of load we'd subtract */
2474 tmp = sg_div_cpu_power(busiest,
2475 busiest_load_per_task * SCHED_LOAD_SCALE);
2477 pwr_move += busiest->__cpu_power *
2478 min(busiest_load_per_task, max_load - tmp);
2480 /* Amount of load we'd add */
2481 if (max_load * busiest->__cpu_power <
2482 busiest_load_per_task * SCHED_LOAD_SCALE)
2483 tmp = sg_div_cpu_power(this,
2484 max_load * busiest->__cpu_power);
2486 tmp = sg_div_cpu_power(this,
2487 busiest_load_per_task * SCHED_LOAD_SCALE);
2488 pwr_move += this->__cpu_power *
2489 min(this_load_per_task, this_load + tmp);
2490 pwr_move /= SCHED_LOAD_SCALE;
2492 /* Move if we gain throughput */
2493 if (pwr_move <= pwr_now)
2496 *imbalance = busiest_load_per_task;
2502 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2503 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2506 if (this == group_leader && group_leader != group_min) {
2507 *imbalance = min_load_per_task;
2517 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2520 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2521 unsigned long imbalance, cpumask_t *cpus)
2523 struct rq *busiest = NULL, *rq;
2524 unsigned long max_load = 0;
2527 for_each_cpu_mask(i, group->cpumask) {
2529 if (!cpu_isset(i, *cpus))
2534 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2537 if (rq->raw_weighted_load > max_load) {
2538 max_load = rq->raw_weighted_load;
2547 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2548 * so long as it is large enough.
2550 #define MAX_PINNED_INTERVAL 512
2552 static inline unsigned long minus_1_or_zero(unsigned long n)
2554 return n > 0 ? n - 1 : 0;
2558 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2559 * tasks if there is an imbalance.
2561 static int load_balance(int this_cpu, struct rq *this_rq,
2562 struct sched_domain *sd, enum cpu_idle_type idle,
2565 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2566 struct sched_group *group;
2567 unsigned long imbalance;
2569 cpumask_t cpus = CPU_MASK_ALL;
2570 unsigned long flags;
2573 * When power savings policy is enabled for the parent domain, idle
2574 * sibling can pick up load irrespective of busy siblings. In this case,
2575 * let the state of idle sibling percolate up as IDLE, instead of
2576 * portraying it as CPU_NOT_IDLE.
2578 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2579 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2582 schedstat_inc(sd, lb_cnt[idle]);
2585 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2592 schedstat_inc(sd, lb_nobusyg[idle]);
2596 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2598 schedstat_inc(sd, lb_nobusyq[idle]);
2602 BUG_ON(busiest == this_rq);
2604 schedstat_add(sd, lb_imbalance[idle], imbalance);
2607 if (busiest->nr_running > 1) {
2609 * Attempt to move tasks. If find_busiest_group has found
2610 * an imbalance but busiest->nr_running <= 1, the group is
2611 * still unbalanced. nr_moved simply stays zero, so it is
2612 * correctly treated as an imbalance.
2614 local_irq_save(flags);
2615 double_rq_lock(this_rq, busiest);
2616 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2617 minus_1_or_zero(busiest->nr_running),
2618 imbalance, sd, idle, &all_pinned);
2619 double_rq_unlock(this_rq, busiest);
2620 local_irq_restore(flags);
2623 * some other cpu did the load balance for us.
2625 if (nr_moved && this_cpu != smp_processor_id())
2626 resched_cpu(this_cpu);
2628 /* All tasks on this runqueue were pinned by CPU affinity */
2629 if (unlikely(all_pinned)) {
2630 cpu_clear(cpu_of(busiest), cpus);
2631 if (!cpus_empty(cpus))
2638 schedstat_inc(sd, lb_failed[idle]);
2639 sd->nr_balance_failed++;
2641 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2643 spin_lock_irqsave(&busiest->lock, flags);
2645 /* don't kick the migration_thread, if the curr
2646 * task on busiest cpu can't be moved to this_cpu
2648 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2649 spin_unlock_irqrestore(&busiest->lock, flags);
2651 goto out_one_pinned;
2654 if (!busiest->active_balance) {
2655 busiest->active_balance = 1;
2656 busiest->push_cpu = this_cpu;
2659 spin_unlock_irqrestore(&busiest->lock, flags);
2661 wake_up_process(busiest->migration_thread);
2664 * We've kicked active balancing, reset the failure
2667 sd->nr_balance_failed = sd->cache_nice_tries+1;
2670 sd->nr_balance_failed = 0;
2672 if (likely(!active_balance)) {
2673 /* We were unbalanced, so reset the balancing interval */
2674 sd->balance_interval = sd->min_interval;
2677 * If we've begun active balancing, start to back off. This
2678 * case may not be covered by the all_pinned logic if there
2679 * is only 1 task on the busy runqueue (because we don't call
2682 if (sd->balance_interval < sd->max_interval)
2683 sd->balance_interval *= 2;
2686 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2687 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2692 schedstat_inc(sd, lb_balanced[idle]);
2694 sd->nr_balance_failed = 0;
2697 /* tune up the balancing interval */
2698 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2699 (sd->balance_interval < sd->max_interval))
2700 sd->balance_interval *= 2;
2702 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2703 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2709 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2710 * tasks if there is an imbalance.
2712 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2713 * this_rq is locked.
2716 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2718 struct sched_group *group;
2719 struct rq *busiest = NULL;
2720 unsigned long imbalance;
2723 cpumask_t cpus = CPU_MASK_ALL;
2726 * When power savings policy is enabled for the parent domain, idle
2727 * sibling can pick up load irrespective of busy siblings. In this case,
2728 * let the state of idle sibling percolate up as IDLE, instead of
2729 * portraying it as CPU_NOT_IDLE.
2731 if (sd->flags & SD_SHARE_CPUPOWER &&
2732 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2735 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2737 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2738 &sd_idle, &cpus, NULL);
2740 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2744 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2747 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2751 BUG_ON(busiest == this_rq);
2753 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2756 if (busiest->nr_running > 1) {
2757 /* Attempt to move tasks */
2758 double_lock_balance(this_rq, busiest);
2759 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2760 minus_1_or_zero(busiest->nr_running),
2761 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2762 spin_unlock(&busiest->lock);
2765 cpu_clear(cpu_of(busiest), cpus);
2766 if (!cpus_empty(cpus))
2772 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2773 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2774 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2777 sd->nr_balance_failed = 0;
2782 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2783 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2784 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2786 sd->nr_balance_failed = 0;
2792 * idle_balance is called by schedule() if this_cpu is about to become
2793 * idle. Attempts to pull tasks from other CPUs.
2795 static void idle_balance(int this_cpu, struct rq *this_rq)
2797 struct sched_domain *sd;
2798 int pulled_task = 0;
2799 unsigned long next_balance = jiffies + 60 * HZ;
2801 for_each_domain(this_cpu, sd) {
2802 unsigned long interval;
2804 if (!(sd->flags & SD_LOAD_BALANCE))
2807 if (sd->flags & SD_BALANCE_NEWIDLE)
2808 /* If we've pulled tasks over stop searching: */
2809 pulled_task = load_balance_newidle(this_cpu,
2812 interval = msecs_to_jiffies(sd->balance_interval);
2813 if (time_after(next_balance, sd->last_balance + interval))
2814 next_balance = sd->last_balance + interval;
2820 * We are going idle. next_balance may be set based on
2821 * a busy processor. So reset next_balance.
2823 this_rq->next_balance = next_balance;
2827 * active_load_balance is run by migration threads. It pushes running tasks
2828 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2829 * running on each physical CPU where possible, and avoids physical /
2830 * logical imbalances.
2832 * Called with busiest_rq locked.
2834 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2836 int target_cpu = busiest_rq->push_cpu;
2837 struct sched_domain *sd;
2838 struct rq *target_rq;
2840 /* Is there any task to move? */
2841 if (busiest_rq->nr_running <= 1)
2844 target_rq = cpu_rq(target_cpu);
2847 * This condition is "impossible", if it occurs
2848 * we need to fix it. Originally reported by
2849 * Bjorn Helgaas on a 128-cpu setup.
2851 BUG_ON(busiest_rq == target_rq);
2853 /* move a task from busiest_rq to target_rq */
2854 double_lock_balance(busiest_rq, target_rq);
2856 /* Search for an sd spanning us and the target CPU. */
2857 for_each_domain(target_cpu, sd) {
2858 if ((sd->flags & SD_LOAD_BALANCE) &&
2859 cpu_isset(busiest_cpu, sd->span))
2864 schedstat_inc(sd, alb_cnt);
2866 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2867 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2869 schedstat_inc(sd, alb_pushed);
2871 schedstat_inc(sd, alb_failed);
2873 spin_unlock(&target_rq->lock);
2876 static void update_load(struct rq *this_rq)
2878 unsigned long this_load;
2879 unsigned int i, scale;
2881 this_load = this_rq->raw_weighted_load;
2883 /* Update our load: */
2884 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2885 unsigned long old_load, new_load;
2887 /* scale is effectively 1 << i now, and >> i divides by scale */
2889 old_load = this_rq->cpu_load[i];
2890 new_load = this_load;
2892 * Round up the averaging division if load is increasing. This
2893 * prevents us from getting stuck on 9 if the load is 10, for
2896 if (new_load > old_load)
2897 new_load += scale-1;
2898 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2904 atomic_t load_balancer;
2906 } nohz ____cacheline_aligned = {
2907 .load_balancer = ATOMIC_INIT(-1),
2908 .cpu_mask = CPU_MASK_NONE,
2912 * This routine will try to nominate the ilb (idle load balancing)
2913 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2914 * load balancing on behalf of all those cpus. If all the cpus in the system
2915 * go into this tickless mode, then there will be no ilb owner (as there is
2916 * no need for one) and all the cpus will sleep till the next wakeup event
2919 * For the ilb owner, tick is not stopped. And this tick will be used
2920 * for idle load balancing. ilb owner will still be part of
2923 * While stopping the tick, this cpu will become the ilb owner if there
2924 * is no other owner. And will be the owner till that cpu becomes busy
2925 * or if all cpus in the system stop their ticks at which point
2926 * there is no need for ilb owner.
2928 * When the ilb owner becomes busy, it nominates another owner, during the
2929 * next busy scheduler_tick()
2931 int select_nohz_load_balancer(int stop_tick)
2933 int cpu = smp_processor_id();
2936 cpu_set(cpu, nohz.cpu_mask);
2937 cpu_rq(cpu)->in_nohz_recently = 1;
2940 * If we are going offline and still the leader, give up!
2942 if (cpu_is_offline(cpu) &&
2943 atomic_read(&nohz.load_balancer) == cpu) {
2944 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2949 /* time for ilb owner also to sleep */
2950 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2951 if (atomic_read(&nohz.load_balancer) == cpu)
2952 atomic_set(&nohz.load_balancer, -1);
2956 if (atomic_read(&nohz.load_balancer) == -1) {
2957 /* make me the ilb owner */
2958 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2960 } else if (atomic_read(&nohz.load_balancer) == cpu)
2963 if (!cpu_isset(cpu, nohz.cpu_mask))
2966 cpu_clear(cpu, nohz.cpu_mask);
2968 if (atomic_read(&nohz.load_balancer) == cpu)
2969 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2976 static DEFINE_SPINLOCK(balancing);
2979 * It checks each scheduling domain to see if it is due to be balanced,
2980 * and initiates a balancing operation if so.
2982 * Balancing parameters are set up in arch_init_sched_domains.
2984 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2987 struct rq *rq = cpu_rq(cpu);
2988 unsigned long interval;
2989 struct sched_domain *sd;
2990 /* Earliest time when we have to do rebalance again */
2991 unsigned long next_balance = jiffies + 60*HZ;
2993 for_each_domain(cpu, sd) {
2994 if (!(sd->flags & SD_LOAD_BALANCE))
2997 interval = sd->balance_interval;
2998 if (idle != CPU_IDLE)
2999 interval *= sd->busy_factor;
3001 /* scale ms to jiffies */
3002 interval = msecs_to_jiffies(interval);
3003 if (unlikely(!interval))
3006 if (sd->flags & SD_SERIALIZE) {
3007 if (!spin_trylock(&balancing))
3011 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3012 if (load_balance(cpu, rq, sd, idle, &balance)) {
3014 * We've pulled tasks over so either we're no
3015 * longer idle, or one of our SMT siblings is
3018 idle = CPU_NOT_IDLE;
3020 sd->last_balance = jiffies;
3022 if (sd->flags & SD_SERIALIZE)
3023 spin_unlock(&balancing);
3025 if (time_after(next_balance, sd->last_balance + interval))
3026 next_balance = sd->last_balance + interval;
3029 * Stop the load balance at this level. There is another
3030 * CPU in our sched group which is doing load balancing more
3036 rq->next_balance = next_balance;
3040 * run_rebalance_domains is triggered when needed from the scheduler tick.
3041 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3042 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3044 static void run_rebalance_domains(struct softirq_action *h)
3046 int local_cpu = smp_processor_id();
3047 struct rq *local_rq = cpu_rq(local_cpu);
3048 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3050 rebalance_domains(local_cpu, idle);
3054 * If this cpu is the owner for idle load balancing, then do the
3055 * balancing on behalf of the other idle cpus whose ticks are
3058 if (local_rq->idle_at_tick &&
3059 atomic_read(&nohz.load_balancer) == local_cpu) {
3060 cpumask_t cpus = nohz.cpu_mask;
3064 cpu_clear(local_cpu, cpus);
3065 for_each_cpu_mask(balance_cpu, cpus) {
3067 * If this cpu gets work to do, stop the load balancing
3068 * work being done for other cpus. Next load
3069 * balancing owner will pick it up.
3074 rebalance_domains(balance_cpu, CPU_IDLE);
3076 rq = cpu_rq(balance_cpu);
3077 if (time_after(local_rq->next_balance, rq->next_balance))
3078 local_rq->next_balance = rq->next_balance;
3085 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3087 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3088 * idle load balancing owner or decide to stop the periodic load balancing,
3089 * if the whole system is idle.
3091 static inline void trigger_load_balance(int cpu)
3093 struct rq *rq = cpu_rq(cpu);
3096 * If we were in the nohz mode recently and busy at the current
3097 * scheduler tick, then check if we need to nominate new idle
3100 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3101 rq->in_nohz_recently = 0;
3103 if (atomic_read(&nohz.load_balancer) == cpu) {
3104 cpu_clear(cpu, nohz.cpu_mask);
3105 atomic_set(&nohz.load_balancer, -1);
3108 if (atomic_read(&nohz.load_balancer) == -1) {
3110 * simple selection for now: Nominate the
3111 * first cpu in the nohz list to be the next
3114 * TBD: Traverse the sched domains and nominate
3115 * the nearest cpu in the nohz.cpu_mask.
3117 int ilb = first_cpu(nohz.cpu_mask);
3125 * If this cpu is idle and doing idle load balancing for all the
3126 * cpus with ticks stopped, is it time for that to stop?
3128 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3129 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3135 * If this cpu is idle and the idle load balancing is done by
3136 * someone else, then no need raise the SCHED_SOFTIRQ
3138 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3139 cpu_isset(cpu, nohz.cpu_mask))
3142 if (time_after_eq(jiffies, rq->next_balance))
3143 raise_softirq(SCHED_SOFTIRQ);
3147 * on UP we do not need to balance between CPUs:
3149 static inline void idle_balance(int cpu, struct rq *rq)
3154 DEFINE_PER_CPU(struct kernel_stat, kstat);
3156 EXPORT_PER_CPU_SYMBOL(kstat);
3159 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3160 * that have not yet been banked in case the task is currently running.
3162 unsigned long long task_sched_runtime(struct task_struct *p)
3164 unsigned long flags;
3168 rq = task_rq_lock(p, &flags);
3169 ns = p->se.sum_exec_runtime;
3170 if (rq->curr == p) {
3171 delta_exec = rq_clock(rq) - p->se.exec_start;
3172 if ((s64)delta_exec > 0)
3175 task_rq_unlock(rq, &flags);
3181 * We place interactive tasks back into the active array, if possible.
3183 * To guarantee that this does not starve expired tasks we ignore the
3184 * interactivity of a task if the first expired task had to wait more
3185 * than a 'reasonable' amount of time. This deadline timeout is
3186 * load-dependent, as the frequency of array switched decreases with
3187 * increasing number of running tasks. We also ignore the interactivity
3188 * if a better static_prio task has expired:
3190 static inline int expired_starving(struct rq *rq)
3192 if (rq->curr->static_prio > rq->best_expired_prio)
3194 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3196 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3202 * Account user cpu time to a process.
3203 * @p: the process that the cpu time gets accounted to
3204 * @hardirq_offset: the offset to subtract from hardirq_count()
3205 * @cputime: the cpu time spent in user space since the last update
3207 void account_user_time(struct task_struct *p, cputime_t cputime)
3209 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3212 p->utime = cputime_add(p->utime, cputime);
3214 /* Add user time to cpustat. */
3215 tmp = cputime_to_cputime64(cputime);
3216 if (TASK_NICE(p) > 0)
3217 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3219 cpustat->user = cputime64_add(cpustat->user, tmp);
3223 * Account system cpu time to a process.
3224 * @p: the process that the cpu time gets accounted to
3225 * @hardirq_offset: the offset to subtract from hardirq_count()
3226 * @cputime: the cpu time spent in kernel space since the last update
3228 void account_system_time(struct task_struct *p, int hardirq_offset,
3231 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3232 struct rq *rq = this_rq();
3235 p->stime = cputime_add(p->stime, cputime);
3237 /* Add system time to cpustat. */
3238 tmp = cputime_to_cputime64(cputime);
3239 if (hardirq_count() - hardirq_offset)
3240 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3241 else if (softirq_count())
3242 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3243 else if (p != rq->idle)
3244 cpustat->system = cputime64_add(cpustat->system, tmp);
3245 else if (atomic_read(&rq->nr_iowait) > 0)
3246 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3248 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3249 /* Account for system time used */
3250 acct_update_integrals(p);
3254 * Account for involuntary wait time.
3255 * @p: the process from which the cpu time has been stolen
3256 * @steal: the cpu time spent in involuntary wait
3258 void account_steal_time(struct task_struct *p, cputime_t steal)
3260 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3261 cputime64_t tmp = cputime_to_cputime64(steal);
3262 struct rq *rq = this_rq();
3264 if (p == rq->idle) {
3265 p->stime = cputime_add(p->stime, steal);
3266 if (atomic_read(&rq->nr_iowait) > 0)
3267 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3269 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3271 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3274 static void task_running_tick(struct rq *rq, struct task_struct *p)
3276 if (p->array != rq->active) {
3277 /* Task has expired but was not scheduled yet */
3278 set_tsk_need_resched(p);
3281 spin_lock(&rq->lock);
3283 * The task was running during this tick - update the
3284 * time slice counter. Note: we do not update a thread's
3285 * priority until it either goes to sleep or uses up its
3286 * timeslice. This makes it possible for interactive tasks
3287 * to use up their timeslices at their highest priority levels.
3291 * RR tasks need a special form of timeslice management.
3292 * FIFO tasks have no timeslices.
3294 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3295 p->time_slice = task_timeslice(p);
3296 p->first_time_slice = 0;
3297 set_tsk_need_resched(p);
3299 /* put it at the end of the queue: */
3300 requeue_task(p, rq->active);
3304 if (!--p->time_slice) {
3305 dequeue_task(p, rq->active);
3306 set_tsk_need_resched(p);
3307 p->prio = effective_prio(p);
3308 p->time_slice = task_timeslice(p);
3309 p->first_time_slice = 0;
3311 if (!rq->expired_timestamp)
3312 rq->expired_timestamp = jiffies;
3313 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3314 enqueue_task(p, rq->expired);
3315 if (p->static_prio < rq->best_expired_prio)
3316 rq->best_expired_prio = p->static_prio;
3318 enqueue_task(p, rq->active);
3321 * Prevent a too long timeslice allowing a task to monopolize
3322 * the CPU. We do this by splitting up the timeslice into
3325 * Note: this does not mean the task's timeslices expire or
3326 * get lost in any way, they just might be preempted by
3327 * another task of equal priority. (one with higher
3328 * priority would have preempted this task already.) We
3329 * requeue this task to the end of the list on this priority
3330 * level, which is in essence a round-robin of tasks with
3333 * This only applies to tasks in the interactive
3334 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3336 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3337 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3338 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3339 (p->array == rq->active)) {
3341 requeue_task(p, rq->active);
3342 set_tsk_need_resched(p);
3346 spin_unlock(&rq->lock);
3350 * This function gets called by the timer code, with HZ frequency.
3351 * We call it with interrupts disabled.
3353 * It also gets called by the fork code, when changing the parent's
3356 void scheduler_tick(void)
3358 struct task_struct *p = current;
3359 int cpu = smp_processor_id();
3360 int idle_at_tick = idle_cpu(cpu);
3361 struct rq *rq = cpu_rq(cpu);
3364 task_running_tick(rq, p);
3367 rq->idle_at_tick = idle_at_tick;
3368 trigger_load_balance(cpu);
3372 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3374 void fastcall add_preempt_count(int val)
3379 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3381 preempt_count() += val;
3383 * Spinlock count overflowing soon?
3385 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3388 EXPORT_SYMBOL(add_preempt_count);
3390 void fastcall sub_preempt_count(int val)
3395 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3398 * Is the spinlock portion underflowing?
3400 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3401 !(preempt_count() & PREEMPT_MASK)))
3404 preempt_count() -= val;
3406 EXPORT_SYMBOL(sub_preempt_count);
3410 static inline int interactive_sleep(enum sleep_type sleep_type)
3412 return (sleep_type == SLEEP_INTERACTIVE ||
3413 sleep_type == SLEEP_INTERRUPTED);
3417 * schedule() is the main scheduler function.
3419 asmlinkage void __sched schedule(void)
3421 struct task_struct *prev, *next;
3422 struct prio_array *array;
3423 struct list_head *queue;
3424 unsigned long long now;
3425 unsigned long run_time;
3426 int cpu, idx, new_prio;
3431 * Test if we are atomic. Since do_exit() needs to call into
3432 * schedule() atomically, we ignore that path for now.
3433 * Otherwise, whine if we are scheduling when we should not be.
3435 if (unlikely(in_atomic() && !current->exit_state)) {
3436 printk(KERN_ERR "BUG: scheduling while atomic: "
3438 current->comm, preempt_count(), current->pid);
3439 debug_show_held_locks(current);
3440 if (irqs_disabled())
3441 print_irqtrace_events(current);
3444 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3449 release_kernel_lock(prev);
3450 need_resched_nonpreemptible:
3454 * The idle thread is not allowed to schedule!
3455 * Remove this check after it has been exercised a bit.
3457 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3458 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3462 schedstat_inc(rq, sched_cnt);
3463 now = sched_clock();
3464 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3465 run_time = now - prev->timestamp;
3466 if (unlikely((long long)(now - prev->timestamp) < 0))
3469 run_time = NS_MAX_SLEEP_AVG;
3472 * Tasks charged proportionately less run_time at high sleep_avg to
3473 * delay them losing their interactive status
3475 run_time /= (CURRENT_BONUS(prev) ? : 1);
3477 spin_lock_irq(&rq->lock);
3479 switch_count = &prev->nivcsw;
3480 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3481 switch_count = &prev->nvcsw;
3482 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3483 unlikely(signal_pending(prev))))
3484 prev->state = TASK_RUNNING;
3486 if (prev->state == TASK_UNINTERRUPTIBLE)
3487 rq->nr_uninterruptible++;
3488 deactivate_task(prev, rq);
3492 cpu = smp_processor_id();
3493 if (unlikely(!rq->nr_running)) {
3494 idle_balance(cpu, rq);
3495 if (!rq->nr_running) {
3497 rq->expired_timestamp = 0;
3503 if (unlikely(!array->nr_active)) {
3505 * Switch the active and expired arrays.
3507 schedstat_inc(rq, sched_switch);
3508 rq->active = rq->expired;
3509 rq->expired = array;
3511 rq->expired_timestamp = 0;
3512 rq->best_expired_prio = MAX_PRIO;
3515 idx = sched_find_first_bit(array->bitmap);
3516 queue = array->queue + idx;
3517 next = list_entry(queue->next, struct task_struct, run_list);
3519 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3520 unsigned long long delta = now - next->timestamp;
3521 if (unlikely((long long)(now - next->timestamp) < 0))
3524 if (next->sleep_type == SLEEP_INTERACTIVE)
3525 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3527 array = next->array;
3528 new_prio = recalc_task_prio(next, next->timestamp + delta);
3530 if (unlikely(next->prio != new_prio)) {
3531 dequeue_task(next, array);
3532 next->prio = new_prio;
3533 enqueue_task(next, array);
3536 next->sleep_type = SLEEP_NORMAL;
3538 if (next == rq->idle)
3539 schedstat_inc(rq, sched_goidle);
3541 prefetch_stack(next);
3542 clear_tsk_need_resched(prev);
3543 rcu_qsctr_inc(task_cpu(prev));
3545 prev->sleep_avg -= run_time;
3546 if ((long)prev->sleep_avg <= 0)
3547 prev->sleep_avg = 0;
3548 prev->timestamp = prev->last_ran = now;
3550 sched_info_switch(prev, next);
3551 if (likely(prev != next)) {
3552 next->timestamp = next->last_ran = now;
3557 prepare_task_switch(rq, next);
3558 prev = context_switch(rq, prev, next);
3561 * this_rq must be evaluated again because prev may have moved
3562 * CPUs since it called schedule(), thus the 'rq' on its stack
3563 * frame will be invalid.
3565 finish_task_switch(this_rq(), prev);
3567 spin_unlock_irq(&rq->lock);
3570 if (unlikely(reacquire_kernel_lock(prev) < 0))
3571 goto need_resched_nonpreemptible;
3572 preempt_enable_no_resched();
3573 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3576 EXPORT_SYMBOL(schedule);
3578 #ifdef CONFIG_PREEMPT
3580 * this is the entry point to schedule() from in-kernel preemption
3581 * off of preempt_enable. Kernel preemptions off return from interrupt
3582 * occur there and call schedule directly.
3584 asmlinkage void __sched preempt_schedule(void)
3586 struct thread_info *ti = current_thread_info();
3587 #ifdef CONFIG_PREEMPT_BKL
3588 struct task_struct *task = current;
3589 int saved_lock_depth;
3592 * If there is a non-zero preempt_count or interrupts are disabled,
3593 * we do not want to preempt the current task. Just return..
3595 if (likely(ti->preempt_count || irqs_disabled()))
3599 add_preempt_count(PREEMPT_ACTIVE);
3601 * We keep the big kernel semaphore locked, but we
3602 * clear ->lock_depth so that schedule() doesnt
3603 * auto-release the semaphore:
3605 #ifdef CONFIG_PREEMPT_BKL
3606 saved_lock_depth = task->lock_depth;
3607 task->lock_depth = -1;
3610 #ifdef CONFIG_PREEMPT_BKL
3611 task->lock_depth = saved_lock_depth;
3613 sub_preempt_count(PREEMPT_ACTIVE);
3615 /* we could miss a preemption opportunity between schedule and now */
3617 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3620 EXPORT_SYMBOL(preempt_schedule);
3623 * this is the entry point to schedule() from kernel preemption
3624 * off of irq context.
3625 * Note, that this is called and return with irqs disabled. This will
3626 * protect us against recursive calling from irq.
3628 asmlinkage void __sched preempt_schedule_irq(void)
3630 struct thread_info *ti = current_thread_info();
3631 #ifdef CONFIG_PREEMPT_BKL
3632 struct task_struct *task = current;
3633 int saved_lock_depth;
3635 /* Catch callers which need to be fixed */
3636 BUG_ON(ti->preempt_count || !irqs_disabled());
3639 add_preempt_count(PREEMPT_ACTIVE);
3641 * We keep the big kernel semaphore locked, but we
3642 * clear ->lock_depth so that schedule() doesnt
3643 * auto-release the semaphore:
3645 #ifdef CONFIG_PREEMPT_BKL
3646 saved_lock_depth = task->lock_depth;
3647 task->lock_depth = -1;
3651 local_irq_disable();
3652 #ifdef CONFIG_PREEMPT_BKL
3653 task->lock_depth = saved_lock_depth;
3655 sub_preempt_count(PREEMPT_ACTIVE);
3657 /* we could miss a preemption opportunity between schedule and now */
3659 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3663 #endif /* CONFIG_PREEMPT */
3665 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3668 return try_to_wake_up(curr->private, mode, sync);
3670 EXPORT_SYMBOL(default_wake_function);
3673 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3674 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3675 * number) then we wake all the non-exclusive tasks and one exclusive task.
3677 * There are circumstances in which we can try to wake a task which has already
3678 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3679 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3681 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3682 int nr_exclusive, int sync, void *key)
3684 struct list_head *tmp, *next;
3686 list_for_each_safe(tmp, next, &q->task_list) {
3687 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3688 unsigned flags = curr->flags;
3690 if (curr->func(curr, mode, sync, key) &&
3691 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3697 * __wake_up - wake up threads blocked on a waitqueue.
3699 * @mode: which threads
3700 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3701 * @key: is directly passed to the wakeup function
3703 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3704 int nr_exclusive, void *key)
3706 unsigned long flags;
3708 spin_lock_irqsave(&q->lock, flags);
3709 __wake_up_common(q, mode, nr_exclusive, 0, key);
3710 spin_unlock_irqrestore(&q->lock, flags);
3712 EXPORT_SYMBOL(__wake_up);
3715 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3717 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3719 __wake_up_common(q, mode, 1, 0, NULL);
3723 * __wake_up_sync - wake up threads blocked on a waitqueue.
3725 * @mode: which threads
3726 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3728 * The sync wakeup differs that the waker knows that it will schedule
3729 * away soon, so while the target thread will be woken up, it will not
3730 * be migrated to another CPU - ie. the two threads are 'synchronized'
3731 * with each other. This can prevent needless bouncing between CPUs.
3733 * On UP it can prevent extra preemption.
3736 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3738 unsigned long flags;
3744 if (unlikely(!nr_exclusive))
3747 spin_lock_irqsave(&q->lock, flags);
3748 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3749 spin_unlock_irqrestore(&q->lock, flags);
3751 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3753 void fastcall complete(struct completion *x)
3755 unsigned long flags;
3757 spin_lock_irqsave(&x->wait.lock, flags);
3759 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3761 spin_unlock_irqrestore(&x->wait.lock, flags);
3763 EXPORT_SYMBOL(complete);
3765 void fastcall complete_all(struct completion *x)
3767 unsigned long flags;
3769 spin_lock_irqsave(&x->wait.lock, flags);
3770 x->done += UINT_MAX/2;
3771 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3773 spin_unlock_irqrestore(&x->wait.lock, flags);
3775 EXPORT_SYMBOL(complete_all);
3777 void fastcall __sched wait_for_completion(struct completion *x)
3781 spin_lock_irq(&x->wait.lock);
3783 DECLARE_WAITQUEUE(wait, current);
3785 wait.flags |= WQ_FLAG_EXCLUSIVE;
3786 __add_wait_queue_tail(&x->wait, &wait);
3788 __set_current_state(TASK_UNINTERRUPTIBLE);
3789 spin_unlock_irq(&x->wait.lock);
3791 spin_lock_irq(&x->wait.lock);
3793 __remove_wait_queue(&x->wait, &wait);
3796 spin_unlock_irq(&x->wait.lock);
3798 EXPORT_SYMBOL(wait_for_completion);
3800 unsigned long fastcall __sched
3801 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3805 spin_lock_irq(&x->wait.lock);
3807 DECLARE_WAITQUEUE(wait, current);
3809 wait.flags |= WQ_FLAG_EXCLUSIVE;
3810 __add_wait_queue_tail(&x->wait, &wait);
3812 __set_current_state(TASK_UNINTERRUPTIBLE);
3813 spin_unlock_irq(&x->wait.lock);
3814 timeout = schedule_timeout(timeout);
3815 spin_lock_irq(&x->wait.lock);
3817 __remove_wait_queue(&x->wait, &wait);
3821 __remove_wait_queue(&x->wait, &wait);
3825 spin_unlock_irq(&x->wait.lock);
3828 EXPORT_SYMBOL(wait_for_completion_timeout);
3830 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3836 spin_lock_irq(&x->wait.lock);
3838 DECLARE_WAITQUEUE(wait, current);
3840 wait.flags |= WQ_FLAG_EXCLUSIVE;
3841 __add_wait_queue_tail(&x->wait, &wait);
3843 if (signal_pending(current)) {
3845 __remove_wait_queue(&x->wait, &wait);
3848 __set_current_state(TASK_INTERRUPTIBLE);
3849 spin_unlock_irq(&x->wait.lock);
3851 spin_lock_irq(&x->wait.lock);
3853 __remove_wait_queue(&x->wait, &wait);
3857 spin_unlock_irq(&x->wait.lock);
3861 EXPORT_SYMBOL(wait_for_completion_interruptible);
3863 unsigned long fastcall __sched
3864 wait_for_completion_interruptible_timeout(struct completion *x,
3865 unsigned long timeout)
3869 spin_lock_irq(&x->wait.lock);
3871 DECLARE_WAITQUEUE(wait, current);
3873 wait.flags |= WQ_FLAG_EXCLUSIVE;
3874 __add_wait_queue_tail(&x->wait, &wait);
3876 if (signal_pending(current)) {
3877 timeout = -ERESTARTSYS;
3878 __remove_wait_queue(&x->wait, &wait);
3881 __set_current_state(TASK_INTERRUPTIBLE);
3882 spin_unlock_irq(&x->wait.lock);
3883 timeout = schedule_timeout(timeout);
3884 spin_lock_irq(&x->wait.lock);
3886 __remove_wait_queue(&x->wait, &wait);
3890 __remove_wait_queue(&x->wait, &wait);
3894 spin_unlock_irq(&x->wait.lock);
3897 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3900 #define SLEEP_ON_VAR \
3901 unsigned long flags; \
3902 wait_queue_t wait; \
3903 init_waitqueue_entry(&wait, current);
3905 #define SLEEP_ON_HEAD \
3906 spin_lock_irqsave(&q->lock,flags); \
3907 __add_wait_queue(q, &wait); \
3908 spin_unlock(&q->lock);
3910 #define SLEEP_ON_TAIL \
3911 spin_lock_irq(&q->lock); \
3912 __remove_wait_queue(q, &wait); \
3913 spin_unlock_irqrestore(&q->lock, flags);
3915 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3919 current->state = TASK_INTERRUPTIBLE;
3925 EXPORT_SYMBOL(interruptible_sleep_on);
3927 long fastcall __sched
3928 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3932 current->state = TASK_INTERRUPTIBLE;
3935 timeout = schedule_timeout(timeout);
3940 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3942 void fastcall __sched sleep_on(wait_queue_head_t *q)
3946 current->state = TASK_UNINTERRUPTIBLE;
3952 EXPORT_SYMBOL(sleep_on);
3954 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3958 current->state = TASK_UNINTERRUPTIBLE;
3961 timeout = schedule_timeout(timeout);
3967 EXPORT_SYMBOL(sleep_on_timeout);
3969 #ifdef CONFIG_RT_MUTEXES
3972 * rt_mutex_setprio - set the current priority of a task
3974 * @prio: prio value (kernel-internal form)
3976 * This function changes the 'effective' priority of a task. It does
3977 * not touch ->normal_prio like __setscheduler().
3979 * Used by the rt_mutex code to implement priority inheritance logic.
3981 void rt_mutex_setprio(struct task_struct *p, int prio)
3983 struct prio_array *array;
3984 unsigned long flags;
3988 BUG_ON(prio < 0 || prio > MAX_PRIO);
3990 rq = task_rq_lock(p, &flags);
3995 dequeue_task(p, array);
4000 * If changing to an RT priority then queue it
4001 * in the active array!
4005 enqueue_task(p, array);
4007 * Reschedule if we are currently running on this runqueue and
4008 * our priority decreased, or if we are not currently running on
4009 * this runqueue and our priority is higher than the current's
4011 if (task_running(rq, p)) {
4012 if (p->prio > oldprio)
4013 resched_task(rq->curr);
4014 } else if (TASK_PREEMPTS_CURR(p, rq))
4015 resched_task(rq->curr);
4017 task_rq_unlock(rq, &flags);
4022 void set_user_nice(struct task_struct *p, long nice)
4024 struct prio_array *array;
4025 int old_prio, delta;
4026 unsigned long flags;
4029 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4032 * We have to be careful, if called from sys_setpriority(),
4033 * the task might be in the middle of scheduling on another CPU.
4035 rq = task_rq_lock(p, &flags);
4037 * The RT priorities are set via sched_setscheduler(), but we still
4038 * allow the 'normal' nice value to be set - but as expected
4039 * it wont have any effect on scheduling until the task is
4040 * not SCHED_NORMAL/SCHED_BATCH:
4042 if (has_rt_policy(p)) {
4043 p->static_prio = NICE_TO_PRIO(nice);
4048 dequeue_task(p, array);
4049 dec_raw_weighted_load(rq, p);
4052 p->static_prio = NICE_TO_PRIO(nice);
4055 p->prio = effective_prio(p);
4056 delta = p->prio - old_prio;
4059 enqueue_task(p, array);
4060 inc_raw_weighted_load(rq, p);
4062 * If the task increased its priority or is running and
4063 * lowered its priority, then reschedule its CPU:
4065 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4066 resched_task(rq->curr);
4069 task_rq_unlock(rq, &flags);
4071 EXPORT_SYMBOL(set_user_nice);
4074 * can_nice - check if a task can reduce its nice value
4078 int can_nice(const struct task_struct *p, const int nice)
4080 /* convert nice value [19,-20] to rlimit style value [1,40] */
4081 int nice_rlim = 20 - nice;
4083 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4084 capable(CAP_SYS_NICE));
4087 #ifdef __ARCH_WANT_SYS_NICE
4090 * sys_nice - change the priority of the current process.
4091 * @increment: priority increment
4093 * sys_setpriority is a more generic, but much slower function that
4094 * does similar things.
4096 asmlinkage long sys_nice(int increment)
4101 * Setpriority might change our priority at the same moment.
4102 * We don't have to worry. Conceptually one call occurs first
4103 * and we have a single winner.
4105 if (increment < -40)
4110 nice = PRIO_TO_NICE(current->static_prio) + increment;
4116 if (increment < 0 && !can_nice(current, nice))
4119 retval = security_task_setnice(current, nice);
4123 set_user_nice(current, nice);
4130 * task_prio - return the priority value of a given task.
4131 * @p: the task in question.
4133 * This is the priority value as seen by users in /proc.
4134 * RT tasks are offset by -200. Normal tasks are centered
4135 * around 0, value goes from -16 to +15.
4137 int task_prio(const struct task_struct *p)
4139 return p->prio - MAX_RT_PRIO;
4143 * task_nice - return the nice value of a given task.
4144 * @p: the task in question.
4146 int task_nice(const struct task_struct *p)
4148 return TASK_NICE(p);
4150 EXPORT_SYMBOL_GPL(task_nice);
4153 * idle_cpu - is a given cpu idle currently?
4154 * @cpu: the processor in question.
4156 int idle_cpu(int cpu)
4158 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4162 * idle_task - return the idle task for a given cpu.
4163 * @cpu: the processor in question.
4165 struct task_struct *idle_task(int cpu)
4167 return cpu_rq(cpu)->idle;
4171 * find_process_by_pid - find a process with a matching PID value.
4172 * @pid: the pid in question.
4174 static inline struct task_struct *find_process_by_pid(pid_t pid)
4176 return pid ? find_task_by_pid(pid) : current;
4179 /* Actually do priority change: must hold rq lock. */
4180 static void __setscheduler(struct task_struct *p, int policy, int prio)
4185 p->rt_priority = prio;
4186 p->normal_prio = normal_prio(p);
4187 /* we are holding p->pi_lock already */
4188 p->prio = rt_mutex_getprio(p);
4190 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4192 if (policy == SCHED_BATCH)
4198 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4199 * @p: the task in question.
4200 * @policy: new policy.
4201 * @param: structure containing the new RT priority.
4203 * NOTE that the task may be already dead.
4205 int sched_setscheduler(struct task_struct *p, int policy,
4206 struct sched_param *param)
4208 int retval, oldprio, oldpolicy = -1;
4209 struct prio_array *array;
4210 unsigned long flags;
4213 /* may grab non-irq protected spin_locks */
4214 BUG_ON(in_interrupt());
4216 /* double check policy once rq lock held */
4218 policy = oldpolicy = p->policy;
4219 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4220 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4223 * Valid priorities for SCHED_FIFO and SCHED_RR are
4224 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4227 if (param->sched_priority < 0 ||
4228 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4229 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4231 if (is_rt_policy(policy) != (param->sched_priority != 0))
4235 * Allow unprivileged RT tasks to decrease priority:
4237 if (!capable(CAP_SYS_NICE)) {
4238 if (is_rt_policy(policy)) {
4239 unsigned long rlim_rtprio;
4240 unsigned long flags;
4242 if (!lock_task_sighand(p, &flags))
4244 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4245 unlock_task_sighand(p, &flags);
4247 /* can't set/change the rt policy */
4248 if (policy != p->policy && !rlim_rtprio)
4251 /* can't increase priority */
4252 if (param->sched_priority > p->rt_priority &&
4253 param->sched_priority > rlim_rtprio)
4257 /* can't change other user's priorities */
4258 if ((current->euid != p->euid) &&
4259 (current->euid != p->uid))
4263 retval = security_task_setscheduler(p, policy, param);
4267 * make sure no PI-waiters arrive (or leave) while we are
4268 * changing the priority of the task:
4270 spin_lock_irqsave(&p->pi_lock, flags);
4272 * To be able to change p->policy safely, the apropriate
4273 * runqueue lock must be held.
4275 rq = __task_rq_lock(p);
4276 /* recheck policy now with rq lock held */
4277 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4278 policy = oldpolicy = -1;
4279 __task_rq_unlock(rq);
4280 spin_unlock_irqrestore(&p->pi_lock, flags);
4285 deactivate_task(p, rq);
4287 __setscheduler(p, policy, param->sched_priority);
4289 __activate_task(p, rq);
4291 * Reschedule if we are currently running on this runqueue and
4292 * our priority decreased, or if we are not currently running on
4293 * this runqueue and our priority is higher than the current's
4295 if (task_running(rq, p)) {
4296 if (p->prio > oldprio)
4297 resched_task(rq->curr);
4298 } else if (TASK_PREEMPTS_CURR(p, rq))
4299 resched_task(rq->curr);
4301 __task_rq_unlock(rq);
4302 spin_unlock_irqrestore(&p->pi_lock, flags);
4304 rt_mutex_adjust_pi(p);
4308 EXPORT_SYMBOL_GPL(sched_setscheduler);
4311 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4313 struct sched_param lparam;
4314 struct task_struct *p;
4317 if (!param || pid < 0)
4319 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4324 p = find_process_by_pid(pid);
4326 retval = sched_setscheduler(p, policy, &lparam);
4333 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4334 * @pid: the pid in question.
4335 * @policy: new policy.
4336 * @param: structure containing the new RT priority.
4338 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4339 struct sched_param __user *param)
4341 /* negative values for policy are not valid */
4345 return do_sched_setscheduler(pid, policy, param);
4349 * sys_sched_setparam - set/change the RT priority of a thread
4350 * @pid: the pid in question.
4351 * @param: structure containing the new RT priority.
4353 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4355 return do_sched_setscheduler(pid, -1, param);
4359 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4360 * @pid: the pid in question.
4362 asmlinkage long sys_sched_getscheduler(pid_t pid)
4364 struct task_struct *p;
4365 int retval = -EINVAL;
4371 read_lock(&tasklist_lock);
4372 p = find_process_by_pid(pid);
4374 retval = security_task_getscheduler(p);
4378 read_unlock(&tasklist_lock);
4385 * sys_sched_getscheduler - get the RT priority of a thread
4386 * @pid: the pid in question.
4387 * @param: structure containing the RT priority.
4389 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4391 struct sched_param lp;
4392 struct task_struct *p;
4393 int retval = -EINVAL;
4395 if (!param || pid < 0)
4398 read_lock(&tasklist_lock);
4399 p = find_process_by_pid(pid);
4404 retval = security_task_getscheduler(p);
4408 lp.sched_priority = p->rt_priority;
4409 read_unlock(&tasklist_lock);
4412 * This one might sleep, we cannot do it with a spinlock held ...
4414 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4420 read_unlock(&tasklist_lock);
4424 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4426 cpumask_t cpus_allowed;
4427 struct task_struct *p;
4430 mutex_lock(&sched_hotcpu_mutex);
4431 read_lock(&tasklist_lock);
4433 p = find_process_by_pid(pid);
4435 read_unlock(&tasklist_lock);
4436 mutex_unlock(&sched_hotcpu_mutex);
4441 * It is not safe to call set_cpus_allowed with the
4442 * tasklist_lock held. We will bump the task_struct's
4443 * usage count and then drop tasklist_lock.
4446 read_unlock(&tasklist_lock);
4449 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4450 !capable(CAP_SYS_NICE))
4453 retval = security_task_setscheduler(p, 0, NULL);
4457 cpus_allowed = cpuset_cpus_allowed(p);
4458 cpus_and(new_mask, new_mask, cpus_allowed);
4459 retval = set_cpus_allowed(p, new_mask);
4463 mutex_unlock(&sched_hotcpu_mutex);
4467 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4468 cpumask_t *new_mask)
4470 if (len < sizeof(cpumask_t)) {
4471 memset(new_mask, 0, sizeof(cpumask_t));
4472 } else if (len > sizeof(cpumask_t)) {
4473 len = sizeof(cpumask_t);
4475 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4479 * sys_sched_setaffinity - set the cpu affinity of a process
4480 * @pid: pid of the process
4481 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4482 * @user_mask_ptr: user-space pointer to the new cpu mask
4484 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4485 unsigned long __user *user_mask_ptr)
4490 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4494 return sched_setaffinity(pid, new_mask);
4498 * Represents all cpu's present in the system
4499 * In systems capable of hotplug, this map could dynamically grow
4500 * as new cpu's are detected in the system via any platform specific
4501 * method, such as ACPI for e.g.
4504 cpumask_t cpu_present_map __read_mostly;
4505 EXPORT_SYMBOL(cpu_present_map);
4508 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4509 EXPORT_SYMBOL(cpu_online_map);
4511 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4512 EXPORT_SYMBOL(cpu_possible_map);
4515 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4517 struct task_struct *p;
4520 mutex_lock(&sched_hotcpu_mutex);
4521 read_lock(&tasklist_lock);
4524 p = find_process_by_pid(pid);
4528 retval = security_task_getscheduler(p);
4532 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4535 read_unlock(&tasklist_lock);
4536 mutex_unlock(&sched_hotcpu_mutex);
4544 * sys_sched_getaffinity - get the cpu affinity of a process
4545 * @pid: pid of the process
4546 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4547 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4549 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4550 unsigned long __user *user_mask_ptr)
4555 if (len < sizeof(cpumask_t))
4558 ret = sched_getaffinity(pid, &mask);
4562 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4565 return sizeof(cpumask_t);
4569 * sys_sched_yield - yield the current processor to other threads.
4571 * This function yields the current CPU by moving the calling thread
4572 * to the expired array. If there are no other threads running on this
4573 * CPU then this function will return.
4575 asmlinkage long sys_sched_yield(void)
4577 struct rq *rq = this_rq_lock();
4578 struct prio_array *array = current->array, *target = rq->expired;
4580 schedstat_inc(rq, yld_cnt);
4582 * We implement yielding by moving the task into the expired
4585 * (special rule: RT tasks will just roundrobin in the active
4588 if (rt_task(current))
4589 target = rq->active;
4591 if (array->nr_active == 1) {
4592 schedstat_inc(rq, yld_act_empty);
4593 if (!rq->expired->nr_active)
4594 schedstat_inc(rq, yld_both_empty);
4595 } else if (!rq->expired->nr_active)
4596 schedstat_inc(rq, yld_exp_empty);
4598 if (array != target) {
4599 dequeue_task(current, array);
4600 enqueue_task(current, target);
4603 * requeue_task is cheaper so perform that if possible.
4605 requeue_task(current, array);
4608 * Since we are going to call schedule() anyway, there's
4609 * no need to preempt or enable interrupts:
4611 __release(rq->lock);
4612 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4613 _raw_spin_unlock(&rq->lock);
4614 preempt_enable_no_resched();
4621 static void __cond_resched(void)
4623 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4624 __might_sleep(__FILE__, __LINE__);
4627 * The BKS might be reacquired before we have dropped
4628 * PREEMPT_ACTIVE, which could trigger a second
4629 * cond_resched() call.
4632 add_preempt_count(PREEMPT_ACTIVE);
4634 sub_preempt_count(PREEMPT_ACTIVE);
4635 } while (need_resched());
4638 int __sched cond_resched(void)
4640 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4641 system_state == SYSTEM_RUNNING) {
4647 EXPORT_SYMBOL(cond_resched);
4650 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4651 * call schedule, and on return reacquire the lock.
4653 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4654 * operations here to prevent schedule() from being called twice (once via
4655 * spin_unlock(), once by hand).
4657 int cond_resched_lock(spinlock_t *lock)
4661 if (need_lockbreak(lock)) {
4667 if (need_resched() && system_state == SYSTEM_RUNNING) {
4668 spin_release(&lock->dep_map, 1, _THIS_IP_);
4669 _raw_spin_unlock(lock);
4670 preempt_enable_no_resched();
4677 EXPORT_SYMBOL(cond_resched_lock);
4679 int __sched cond_resched_softirq(void)
4681 BUG_ON(!in_softirq());
4683 if (need_resched() && system_state == SYSTEM_RUNNING) {
4691 EXPORT_SYMBOL(cond_resched_softirq);
4694 * yield - yield the current processor to other threads.
4696 * This is a shortcut for kernel-space yielding - it marks the
4697 * thread runnable and calls sys_sched_yield().
4699 void __sched yield(void)
4701 set_current_state(TASK_RUNNING);
4704 EXPORT_SYMBOL(yield);
4707 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4708 * that process accounting knows that this is a task in IO wait state.
4710 * But don't do that if it is a deliberate, throttling IO wait (this task
4711 * has set its backing_dev_info: the queue against which it should throttle)
4713 void __sched io_schedule(void)
4715 struct rq *rq = &__raw_get_cpu_var(runqueues);
4717 delayacct_blkio_start();
4718 atomic_inc(&rq->nr_iowait);
4720 atomic_dec(&rq->nr_iowait);
4721 delayacct_blkio_end();
4723 EXPORT_SYMBOL(io_schedule);
4725 long __sched io_schedule_timeout(long timeout)
4727 struct rq *rq = &__raw_get_cpu_var(runqueues);
4730 delayacct_blkio_start();
4731 atomic_inc(&rq->nr_iowait);
4732 ret = schedule_timeout(timeout);
4733 atomic_dec(&rq->nr_iowait);
4734 delayacct_blkio_end();
4739 * sys_sched_get_priority_max - return maximum RT priority.
4740 * @policy: scheduling class.
4742 * this syscall returns the maximum rt_priority that can be used
4743 * by a given scheduling class.
4745 asmlinkage long sys_sched_get_priority_max(int policy)
4752 ret = MAX_USER_RT_PRIO-1;
4763 * sys_sched_get_priority_min - return minimum RT priority.
4764 * @policy: scheduling class.
4766 * this syscall returns the minimum rt_priority that can be used
4767 * by a given scheduling class.
4769 asmlinkage long sys_sched_get_priority_min(int policy)
4786 * sys_sched_rr_get_interval - return the default timeslice of a process.
4787 * @pid: pid of the process.
4788 * @interval: userspace pointer to the timeslice value.
4790 * this syscall writes the default timeslice value of a given process
4791 * into the user-space timespec buffer. A value of '0' means infinity.
4794 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4796 struct task_struct *p;
4797 int retval = -EINVAL;
4804 read_lock(&tasklist_lock);
4805 p = find_process_by_pid(pid);
4809 retval = security_task_getscheduler(p);
4813 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4814 0 : task_timeslice(p), &t);
4815 read_unlock(&tasklist_lock);
4816 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4820 read_unlock(&tasklist_lock);
4824 static const char stat_nam[] = "RSDTtZX";
4826 static void show_task(struct task_struct *p)
4828 unsigned long free = 0;
4831 state = p->state ? __ffs(p->state) + 1 : 0;
4832 printk("%-13.13s %c", p->comm,
4833 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4834 #if (BITS_PER_LONG == 32)
4835 if (state == TASK_RUNNING)
4836 printk(" running ");
4838 printk(" %08lX ", thread_saved_pc(p));
4840 if (state == TASK_RUNNING)
4841 printk(" running task ");
4843 printk(" %016lx ", thread_saved_pc(p));
4845 #ifdef CONFIG_DEBUG_STACK_USAGE
4847 unsigned long *n = end_of_stack(p);
4850 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4853 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4855 printk(" (L-TLB)\n");
4857 printk(" (NOTLB)\n");
4859 if (state != TASK_RUNNING)
4860 show_stack(p, NULL);
4863 void show_state_filter(unsigned long state_filter)
4865 struct task_struct *g, *p;
4867 #if (BITS_PER_LONG == 32)
4870 printk(" task PC stack pid father child younger older\n");
4874 printk(" task PC stack pid father child younger older\n");
4876 read_lock(&tasklist_lock);
4877 do_each_thread(g, p) {
4879 * reset the NMI-timeout, listing all files on a slow
4880 * console might take alot of time:
4882 touch_nmi_watchdog();
4883 if (!state_filter || (p->state & state_filter))
4885 } while_each_thread(g, p);
4887 touch_all_softlockup_watchdogs();
4889 read_unlock(&tasklist_lock);
4891 * Only show locks if all tasks are dumped:
4893 if (state_filter == -1)
4894 debug_show_all_locks();
4897 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4903 * init_idle - set up an idle thread for a given CPU
4904 * @idle: task in question
4905 * @cpu: cpu the idle task belongs to
4907 * NOTE: this function does not set the idle thread's NEED_RESCHED
4908 * flag, to make booting more robust.
4910 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4912 struct rq *rq = cpu_rq(cpu);
4913 unsigned long flags;
4915 idle->timestamp = sched_clock();
4916 idle->sleep_avg = 0;
4918 idle->prio = idle->normal_prio = MAX_PRIO;
4919 idle->state = TASK_RUNNING;
4920 idle->cpus_allowed = cpumask_of_cpu(cpu);
4921 set_task_cpu(idle, cpu);
4923 spin_lock_irqsave(&rq->lock, flags);
4924 rq->curr = rq->idle = idle;
4925 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4928 spin_unlock_irqrestore(&rq->lock, flags);
4930 /* Set the preempt count _outside_ the spinlocks! */
4931 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4932 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4934 task_thread_info(idle)->preempt_count = 0;
4939 * In a system that switches off the HZ timer nohz_cpu_mask
4940 * indicates which cpus entered this state. This is used
4941 * in the rcu update to wait only for active cpus. For system
4942 * which do not switch off the HZ timer nohz_cpu_mask should
4943 * always be CPU_MASK_NONE.
4945 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4949 * This is how migration works:
4951 * 1) we queue a struct migration_req structure in the source CPU's
4952 * runqueue and wake up that CPU's migration thread.
4953 * 2) we down() the locked semaphore => thread blocks.
4954 * 3) migration thread wakes up (implicitly it forces the migrated
4955 * thread off the CPU)
4956 * 4) it gets the migration request and checks whether the migrated
4957 * task is still in the wrong runqueue.
4958 * 5) if it's in the wrong runqueue then the migration thread removes
4959 * it and puts it into the right queue.
4960 * 6) migration thread up()s the semaphore.
4961 * 7) we wake up and the migration is done.
4965 * Change a given task's CPU affinity. Migrate the thread to a
4966 * proper CPU and schedule it away if the CPU it's executing on
4967 * is removed from the allowed bitmask.
4969 * NOTE: the caller must have a valid reference to the task, the
4970 * task must not exit() & deallocate itself prematurely. The
4971 * call is not atomic; no spinlocks may be held.
4973 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4975 struct migration_req req;
4976 unsigned long flags;
4980 rq = task_rq_lock(p, &flags);
4981 if (!cpus_intersects(new_mask, cpu_online_map)) {
4986 p->cpus_allowed = new_mask;
4987 /* Can the task run on the task's current CPU? If so, we're done */
4988 if (cpu_isset(task_cpu(p), new_mask))
4991 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4992 /* Need help from migration thread: drop lock and wait. */
4993 task_rq_unlock(rq, &flags);
4994 wake_up_process(rq->migration_thread);
4995 wait_for_completion(&req.done);
4996 tlb_migrate_finish(p->mm);
5000 task_rq_unlock(rq, &flags);
5004 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5007 * Move (not current) task off this cpu, onto dest cpu. We're doing
5008 * this because either it can't run here any more (set_cpus_allowed()
5009 * away from this CPU, or CPU going down), or because we're
5010 * attempting to rebalance this task on exec (sched_exec).
5012 * So we race with normal scheduler movements, but that's OK, as long
5013 * as the task is no longer on this CPU.
5015 * Returns non-zero if task was successfully migrated.
5017 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5019 struct rq *rq_dest, *rq_src;
5022 if (unlikely(cpu_is_offline(dest_cpu)))
5025 rq_src = cpu_rq(src_cpu);
5026 rq_dest = cpu_rq(dest_cpu);
5028 double_rq_lock(rq_src, rq_dest);
5029 /* Already moved. */
5030 if (task_cpu(p) != src_cpu)
5032 /* Affinity changed (again). */
5033 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5036 set_task_cpu(p, dest_cpu);
5039 * Sync timestamp with rq_dest's before activating.
5040 * The same thing could be achieved by doing this step
5041 * afterwards, and pretending it was a local activate.
5042 * This way is cleaner and logically correct.
5044 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5045 + rq_dest->most_recent_timestamp;
5046 deactivate_task(p, rq_src);
5047 __activate_task(p, rq_dest);
5048 if (TASK_PREEMPTS_CURR(p, rq_dest))
5049 resched_task(rq_dest->curr);
5053 double_rq_unlock(rq_src, rq_dest);
5058 * migration_thread - this is a highprio system thread that performs
5059 * thread migration by bumping thread off CPU then 'pushing' onto
5062 static int migration_thread(void *data)
5064 int cpu = (long)data;
5068 BUG_ON(rq->migration_thread != current);
5070 set_current_state(TASK_INTERRUPTIBLE);
5071 while (!kthread_should_stop()) {
5072 struct migration_req *req;
5073 struct list_head *head;
5077 spin_lock_irq(&rq->lock);
5079 if (cpu_is_offline(cpu)) {
5080 spin_unlock_irq(&rq->lock);
5084 if (rq->active_balance) {
5085 active_load_balance(rq, cpu);
5086 rq->active_balance = 0;
5089 head = &rq->migration_queue;
5091 if (list_empty(head)) {
5092 spin_unlock_irq(&rq->lock);
5094 set_current_state(TASK_INTERRUPTIBLE);
5097 req = list_entry(head->next, struct migration_req, list);
5098 list_del_init(head->next);
5100 spin_unlock(&rq->lock);
5101 __migrate_task(req->task, cpu, req->dest_cpu);
5104 complete(&req->done);
5106 __set_current_state(TASK_RUNNING);
5110 /* Wait for kthread_stop */
5111 set_current_state(TASK_INTERRUPTIBLE);
5112 while (!kthread_should_stop()) {
5114 set_current_state(TASK_INTERRUPTIBLE);
5116 __set_current_state(TASK_RUNNING);
5120 #ifdef CONFIG_HOTPLUG_CPU
5122 * Figure out where task on dead CPU should go, use force if neccessary.
5123 * NOTE: interrupts should be disabled by the caller
5125 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5127 unsigned long flags;
5134 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5135 cpus_and(mask, mask, p->cpus_allowed);
5136 dest_cpu = any_online_cpu(mask);
5138 /* On any allowed CPU? */
5139 if (dest_cpu == NR_CPUS)
5140 dest_cpu = any_online_cpu(p->cpus_allowed);
5142 /* No more Mr. Nice Guy. */
5143 if (dest_cpu == NR_CPUS) {
5144 rq = task_rq_lock(p, &flags);
5145 cpus_setall(p->cpus_allowed);
5146 dest_cpu = any_online_cpu(p->cpus_allowed);
5147 task_rq_unlock(rq, &flags);
5150 * Don't tell them about moving exiting tasks or
5151 * kernel threads (both mm NULL), since they never
5154 if (p->mm && printk_ratelimit())
5155 printk(KERN_INFO "process %d (%s) no "
5156 "longer affine to cpu%d\n",
5157 p->pid, p->comm, dead_cpu);
5159 if (!__migrate_task(p, dead_cpu, dest_cpu))
5164 * While a dead CPU has no uninterruptible tasks queued at this point,
5165 * it might still have a nonzero ->nr_uninterruptible counter, because
5166 * for performance reasons the counter is not stricly tracking tasks to
5167 * their home CPUs. So we just add the counter to another CPU's counter,
5168 * to keep the global sum constant after CPU-down:
5170 static void migrate_nr_uninterruptible(struct rq *rq_src)
5172 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5173 unsigned long flags;
5175 local_irq_save(flags);
5176 double_rq_lock(rq_src, rq_dest);
5177 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5178 rq_src->nr_uninterruptible = 0;
5179 double_rq_unlock(rq_src, rq_dest);
5180 local_irq_restore(flags);
5183 /* Run through task list and migrate tasks from the dead cpu. */
5184 static void migrate_live_tasks(int src_cpu)
5186 struct task_struct *p, *t;
5188 write_lock_irq(&tasklist_lock);
5190 do_each_thread(t, p) {
5194 if (task_cpu(p) == src_cpu)
5195 move_task_off_dead_cpu(src_cpu, p);
5196 } while_each_thread(t, p);
5198 write_unlock_irq(&tasklist_lock);
5201 /* Schedules idle task to be the next runnable task on current CPU.
5202 * It does so by boosting its priority to highest possible and adding it to
5203 * the _front_ of the runqueue. Used by CPU offline code.
5205 void sched_idle_next(void)
5207 int this_cpu = smp_processor_id();
5208 struct rq *rq = cpu_rq(this_cpu);
5209 struct task_struct *p = rq->idle;
5210 unsigned long flags;
5212 /* cpu has to be offline */
5213 BUG_ON(cpu_online(this_cpu));
5216 * Strictly not necessary since rest of the CPUs are stopped by now
5217 * and interrupts disabled on the current cpu.
5219 spin_lock_irqsave(&rq->lock, flags);
5221 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5223 /* Add idle task to the _front_ of its priority queue: */
5224 __activate_idle_task(p, rq);
5226 spin_unlock_irqrestore(&rq->lock, flags);
5230 * Ensures that the idle task is using init_mm right before its cpu goes
5233 void idle_task_exit(void)
5235 struct mm_struct *mm = current->active_mm;
5237 BUG_ON(cpu_online(smp_processor_id()));
5240 switch_mm(mm, &init_mm, current);
5244 /* called under rq->lock with disabled interrupts */
5245 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5247 struct rq *rq = cpu_rq(dead_cpu);
5249 /* Must be exiting, otherwise would be on tasklist. */
5250 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5252 /* Cannot have done final schedule yet: would have vanished. */
5253 BUG_ON(p->state == TASK_DEAD);
5258 * Drop lock around migration; if someone else moves it,
5259 * that's OK. No task can be added to this CPU, so iteration is
5261 * NOTE: interrupts should be left disabled --dev@
5263 spin_unlock(&rq->lock);
5264 move_task_off_dead_cpu(dead_cpu, p);
5265 spin_lock(&rq->lock);
5270 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5271 static void migrate_dead_tasks(unsigned int dead_cpu)
5273 struct rq *rq = cpu_rq(dead_cpu);
5274 unsigned int arr, i;
5276 for (arr = 0; arr < 2; arr++) {
5277 for (i = 0; i < MAX_PRIO; i++) {
5278 struct list_head *list = &rq->arrays[arr].queue[i];
5280 while (!list_empty(list))
5281 migrate_dead(dead_cpu, list_entry(list->next,
5282 struct task_struct, run_list));
5286 #endif /* CONFIG_HOTPLUG_CPU */
5289 * migration_call - callback that gets triggered when a CPU is added.
5290 * Here we can start up the necessary migration thread for the new CPU.
5292 static int __cpuinit
5293 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5295 struct task_struct *p;
5296 int cpu = (long)hcpu;
5297 unsigned long flags;
5301 case CPU_LOCK_ACQUIRE:
5302 mutex_lock(&sched_hotcpu_mutex);
5305 case CPU_UP_PREPARE:
5306 case CPU_UP_PREPARE_FROZEN:
5307 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5310 p->flags |= PF_NOFREEZE;
5311 kthread_bind(p, cpu);
5312 /* Must be high prio: stop_machine expects to yield to it. */
5313 rq = task_rq_lock(p, &flags);
5314 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5315 task_rq_unlock(rq, &flags);
5316 cpu_rq(cpu)->migration_thread = p;
5320 case CPU_ONLINE_FROZEN:
5321 /* Strictly unneccessary, as first user will wake it. */
5322 wake_up_process(cpu_rq(cpu)->migration_thread);
5325 #ifdef CONFIG_HOTPLUG_CPU
5326 case CPU_UP_CANCELED:
5327 case CPU_UP_CANCELED_FROZEN:
5328 if (!cpu_rq(cpu)->migration_thread)
5330 /* Unbind it from offline cpu so it can run. Fall thru. */
5331 kthread_bind(cpu_rq(cpu)->migration_thread,
5332 any_online_cpu(cpu_online_map));
5333 kthread_stop(cpu_rq(cpu)->migration_thread);
5334 cpu_rq(cpu)->migration_thread = NULL;
5338 case CPU_DEAD_FROZEN:
5339 migrate_live_tasks(cpu);
5341 kthread_stop(rq->migration_thread);
5342 rq->migration_thread = NULL;
5343 /* Idle task back to normal (off runqueue, low prio) */
5344 rq = task_rq_lock(rq->idle, &flags);
5345 deactivate_task(rq->idle, rq);
5346 rq->idle->static_prio = MAX_PRIO;
5347 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5348 migrate_dead_tasks(cpu);
5349 task_rq_unlock(rq, &flags);
5350 migrate_nr_uninterruptible(rq);
5351 BUG_ON(rq->nr_running != 0);
5353 /* No need to migrate the tasks: it was best-effort if
5354 * they didn't take sched_hotcpu_mutex. Just wake up
5355 * the requestors. */
5356 spin_lock_irq(&rq->lock);
5357 while (!list_empty(&rq->migration_queue)) {
5358 struct migration_req *req;
5360 req = list_entry(rq->migration_queue.next,
5361 struct migration_req, list);
5362 list_del_init(&req->list);
5363 complete(&req->done);
5365 spin_unlock_irq(&rq->lock);
5368 case CPU_LOCK_RELEASE:
5369 mutex_unlock(&sched_hotcpu_mutex);
5375 /* Register at highest priority so that task migration (migrate_all_tasks)
5376 * happens before everything else.
5378 static struct notifier_block __cpuinitdata migration_notifier = {
5379 .notifier_call = migration_call,
5383 int __init migration_init(void)
5385 void *cpu = (void *)(long)smp_processor_id();
5388 /* Start one for the boot CPU: */
5389 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5390 BUG_ON(err == NOTIFY_BAD);
5391 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5392 register_cpu_notifier(&migration_notifier);
5400 /* Number of possible processor ids */
5401 int nr_cpu_ids __read_mostly = NR_CPUS;
5402 EXPORT_SYMBOL(nr_cpu_ids);
5404 #undef SCHED_DOMAIN_DEBUG
5405 #ifdef SCHED_DOMAIN_DEBUG
5406 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5411 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5415 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5420 struct sched_group *group = sd->groups;
5421 cpumask_t groupmask;
5423 cpumask_scnprintf(str, NR_CPUS, sd->span);
5424 cpus_clear(groupmask);
5427 for (i = 0; i < level + 1; i++)
5429 printk("domain %d: ", level);
5431 if (!(sd->flags & SD_LOAD_BALANCE)) {
5432 printk("does not load-balance\n");
5434 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5439 printk("span %s\n", str);
5441 if (!cpu_isset(cpu, sd->span))
5442 printk(KERN_ERR "ERROR: domain->span does not contain "
5444 if (!cpu_isset(cpu, group->cpumask))
5445 printk(KERN_ERR "ERROR: domain->groups does not contain"
5449 for (i = 0; i < level + 2; i++)
5455 printk(KERN_ERR "ERROR: group is NULL\n");
5459 if (!group->__cpu_power) {
5461 printk(KERN_ERR "ERROR: domain->cpu_power not "
5465 if (!cpus_weight(group->cpumask)) {
5467 printk(KERN_ERR "ERROR: empty group\n");
5470 if (cpus_intersects(groupmask, group->cpumask)) {
5472 printk(KERN_ERR "ERROR: repeated CPUs\n");
5475 cpus_or(groupmask, groupmask, group->cpumask);
5477 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5480 group = group->next;
5481 } while (group != sd->groups);
5484 if (!cpus_equal(sd->span, groupmask))
5485 printk(KERN_ERR "ERROR: groups don't span "
5493 if (!cpus_subset(groupmask, sd->span))
5494 printk(KERN_ERR "ERROR: parent span is not a superset "
5495 "of domain->span\n");
5500 # define sched_domain_debug(sd, cpu) do { } while (0)
5503 static int sd_degenerate(struct sched_domain *sd)
5505 if (cpus_weight(sd->span) == 1)
5508 /* Following flags need at least 2 groups */
5509 if (sd->flags & (SD_LOAD_BALANCE |
5510 SD_BALANCE_NEWIDLE |
5514 SD_SHARE_PKG_RESOURCES)) {
5515 if (sd->groups != sd->groups->next)
5519 /* Following flags don't use groups */
5520 if (sd->flags & (SD_WAKE_IDLE |
5529 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5531 unsigned long cflags = sd->flags, pflags = parent->flags;
5533 if (sd_degenerate(parent))
5536 if (!cpus_equal(sd->span, parent->span))
5539 /* Does parent contain flags not in child? */
5540 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5541 if (cflags & SD_WAKE_AFFINE)
5542 pflags &= ~SD_WAKE_BALANCE;
5543 /* Flags needing groups don't count if only 1 group in parent */
5544 if (parent->groups == parent->groups->next) {
5545 pflags &= ~(SD_LOAD_BALANCE |
5546 SD_BALANCE_NEWIDLE |
5550 SD_SHARE_PKG_RESOURCES);
5552 if (~cflags & pflags)
5559 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5560 * hold the hotplug lock.
5562 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5564 struct rq *rq = cpu_rq(cpu);
5565 struct sched_domain *tmp;
5567 /* Remove the sched domains which do not contribute to scheduling. */
5568 for (tmp = sd; tmp; tmp = tmp->parent) {
5569 struct sched_domain *parent = tmp->parent;
5572 if (sd_parent_degenerate(tmp, parent)) {
5573 tmp->parent = parent->parent;
5575 parent->parent->child = tmp;
5579 if (sd && sd_degenerate(sd)) {
5585 sched_domain_debug(sd, cpu);
5587 rcu_assign_pointer(rq->sd, sd);
5590 /* cpus with isolated domains */
5591 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5593 /* Setup the mask of cpus configured for isolated domains */
5594 static int __init isolated_cpu_setup(char *str)
5596 int ints[NR_CPUS], i;
5598 str = get_options(str, ARRAY_SIZE(ints), ints);
5599 cpus_clear(cpu_isolated_map);
5600 for (i = 1; i <= ints[0]; i++)
5601 if (ints[i] < NR_CPUS)
5602 cpu_set(ints[i], cpu_isolated_map);
5606 __setup ("isolcpus=", isolated_cpu_setup);
5609 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5610 * to a function which identifies what group(along with sched group) a CPU
5611 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5612 * (due to the fact that we keep track of groups covered with a cpumask_t).
5614 * init_sched_build_groups will build a circular linked list of the groups
5615 * covered by the given span, and will set each group's ->cpumask correctly,
5616 * and ->cpu_power to 0.
5619 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5620 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5621 struct sched_group **sg))
5623 struct sched_group *first = NULL, *last = NULL;
5624 cpumask_t covered = CPU_MASK_NONE;
5627 for_each_cpu_mask(i, span) {
5628 struct sched_group *sg;
5629 int group = group_fn(i, cpu_map, &sg);
5632 if (cpu_isset(i, covered))
5635 sg->cpumask = CPU_MASK_NONE;
5636 sg->__cpu_power = 0;
5638 for_each_cpu_mask(j, span) {
5639 if (group_fn(j, cpu_map, NULL) != group)
5642 cpu_set(j, covered);
5643 cpu_set(j, sg->cpumask);
5654 #define SD_NODES_PER_DOMAIN 16
5659 * find_next_best_node - find the next node to include in a sched_domain
5660 * @node: node whose sched_domain we're building
5661 * @used_nodes: nodes already in the sched_domain
5663 * Find the next node to include in a given scheduling domain. Simply
5664 * finds the closest node not already in the @used_nodes map.
5666 * Should use nodemask_t.
5668 static int find_next_best_node(int node, unsigned long *used_nodes)
5670 int i, n, val, min_val, best_node = 0;
5674 for (i = 0; i < MAX_NUMNODES; i++) {
5675 /* Start at @node */
5676 n = (node + i) % MAX_NUMNODES;
5678 if (!nr_cpus_node(n))
5681 /* Skip already used nodes */
5682 if (test_bit(n, used_nodes))
5685 /* Simple min distance search */
5686 val = node_distance(node, n);
5688 if (val < min_val) {
5694 set_bit(best_node, used_nodes);
5699 * sched_domain_node_span - get a cpumask for a node's sched_domain
5700 * @node: node whose cpumask we're constructing
5701 * @size: number of nodes to include in this span
5703 * Given a node, construct a good cpumask for its sched_domain to span. It
5704 * should be one that prevents unnecessary balancing, but also spreads tasks
5707 static cpumask_t sched_domain_node_span(int node)
5709 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5710 cpumask_t span, nodemask;
5714 bitmap_zero(used_nodes, MAX_NUMNODES);
5716 nodemask = node_to_cpumask(node);
5717 cpus_or(span, span, nodemask);
5718 set_bit(node, used_nodes);
5720 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5721 int next_node = find_next_best_node(node, used_nodes);
5723 nodemask = node_to_cpumask(next_node);
5724 cpus_or(span, span, nodemask);
5731 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5734 * SMT sched-domains:
5736 #ifdef CONFIG_SCHED_SMT
5737 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5738 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5740 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5741 struct sched_group **sg)
5744 *sg = &per_cpu(sched_group_cpus, cpu);
5750 * multi-core sched-domains:
5752 #ifdef CONFIG_SCHED_MC
5753 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5754 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5757 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5758 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5759 struct sched_group **sg)
5762 cpumask_t mask = cpu_sibling_map[cpu];
5763 cpus_and(mask, mask, *cpu_map);
5764 group = first_cpu(mask);
5766 *sg = &per_cpu(sched_group_core, group);
5769 #elif defined(CONFIG_SCHED_MC)
5770 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5771 struct sched_group **sg)
5774 *sg = &per_cpu(sched_group_core, cpu);
5779 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5780 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5782 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5783 struct sched_group **sg)
5786 #ifdef CONFIG_SCHED_MC
5787 cpumask_t mask = cpu_coregroup_map(cpu);
5788 cpus_and(mask, mask, *cpu_map);
5789 group = first_cpu(mask);
5790 #elif defined(CONFIG_SCHED_SMT)
5791 cpumask_t mask = cpu_sibling_map[cpu];
5792 cpus_and(mask, mask, *cpu_map);
5793 group = first_cpu(mask);
5798 *sg = &per_cpu(sched_group_phys, group);
5804 * The init_sched_build_groups can't handle what we want to do with node
5805 * groups, so roll our own. Now each node has its own list of groups which
5806 * gets dynamically allocated.
5808 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5809 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5811 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5812 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5814 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5815 struct sched_group **sg)
5817 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5820 cpus_and(nodemask, nodemask, *cpu_map);
5821 group = first_cpu(nodemask);
5824 *sg = &per_cpu(sched_group_allnodes, group);
5828 static void init_numa_sched_groups_power(struct sched_group *group_head)
5830 struct sched_group *sg = group_head;
5836 for_each_cpu_mask(j, sg->cpumask) {
5837 struct sched_domain *sd;
5839 sd = &per_cpu(phys_domains, j);
5840 if (j != first_cpu(sd->groups->cpumask)) {
5842 * Only add "power" once for each
5848 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5851 if (sg != group_head)
5857 /* Free memory allocated for various sched_group structures */
5858 static void free_sched_groups(const cpumask_t *cpu_map)
5862 for_each_cpu_mask(cpu, *cpu_map) {
5863 struct sched_group **sched_group_nodes
5864 = sched_group_nodes_bycpu[cpu];
5866 if (!sched_group_nodes)
5869 for (i = 0; i < MAX_NUMNODES; i++) {
5870 cpumask_t nodemask = node_to_cpumask(i);
5871 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5873 cpus_and(nodemask, nodemask, *cpu_map);
5874 if (cpus_empty(nodemask))
5884 if (oldsg != sched_group_nodes[i])
5887 kfree(sched_group_nodes);
5888 sched_group_nodes_bycpu[cpu] = NULL;
5892 static void free_sched_groups(const cpumask_t *cpu_map)
5898 * Initialize sched groups cpu_power.
5900 * cpu_power indicates the capacity of sched group, which is used while
5901 * distributing the load between different sched groups in a sched domain.
5902 * Typically cpu_power for all the groups in a sched domain will be same unless
5903 * there are asymmetries in the topology. If there are asymmetries, group
5904 * having more cpu_power will pickup more load compared to the group having
5907 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5908 * the maximum number of tasks a group can handle in the presence of other idle
5909 * or lightly loaded groups in the same sched domain.
5911 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5913 struct sched_domain *child;
5914 struct sched_group *group;
5916 WARN_ON(!sd || !sd->groups);
5918 if (cpu != first_cpu(sd->groups->cpumask))
5923 sd->groups->__cpu_power = 0;
5926 * For perf policy, if the groups in child domain share resources
5927 * (for example cores sharing some portions of the cache hierarchy
5928 * or SMT), then set this domain groups cpu_power such that each group
5929 * can handle only one task, when there are other idle groups in the
5930 * same sched domain.
5932 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5934 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5935 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5940 * add cpu_power of each child group to this groups cpu_power
5942 group = child->groups;
5944 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5945 group = group->next;
5946 } while (group != child->groups);
5950 * Build sched domains for a given set of cpus and attach the sched domains
5951 * to the individual cpus
5953 static int build_sched_domains(const cpumask_t *cpu_map)
5956 struct sched_domain *sd;
5958 struct sched_group **sched_group_nodes = NULL;
5959 int sd_allnodes = 0;
5962 * Allocate the per-node list of sched groups
5964 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5966 if (!sched_group_nodes) {
5967 printk(KERN_WARNING "Can not alloc sched group node list\n");
5970 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5974 * Set up domains for cpus specified by the cpu_map.
5976 for_each_cpu_mask(i, *cpu_map) {
5977 struct sched_domain *sd = NULL, *p;
5978 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5980 cpus_and(nodemask, nodemask, *cpu_map);
5983 if (cpus_weight(*cpu_map)
5984 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5985 sd = &per_cpu(allnodes_domains, i);
5986 *sd = SD_ALLNODES_INIT;
5987 sd->span = *cpu_map;
5988 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
5994 sd = &per_cpu(node_domains, i);
5996 sd->span = sched_domain_node_span(cpu_to_node(i));
6000 cpus_and(sd->span, sd->span, *cpu_map);
6004 sd = &per_cpu(phys_domains, i);
6006 sd->span = nodemask;
6010 cpu_to_phys_group(i, cpu_map, &sd->groups);
6012 #ifdef CONFIG_SCHED_MC
6014 sd = &per_cpu(core_domains, i);
6016 sd->span = cpu_coregroup_map(i);
6017 cpus_and(sd->span, sd->span, *cpu_map);
6020 cpu_to_core_group(i, cpu_map, &sd->groups);
6023 #ifdef CONFIG_SCHED_SMT
6025 sd = &per_cpu(cpu_domains, i);
6026 *sd = SD_SIBLING_INIT;
6027 sd->span = cpu_sibling_map[i];
6028 cpus_and(sd->span, sd->span, *cpu_map);
6031 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6035 #ifdef CONFIG_SCHED_SMT
6036 /* Set up CPU (sibling) groups */
6037 for_each_cpu_mask(i, *cpu_map) {
6038 cpumask_t this_sibling_map = cpu_sibling_map[i];
6039 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6040 if (i != first_cpu(this_sibling_map))
6043 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6047 #ifdef CONFIG_SCHED_MC
6048 /* Set up multi-core groups */
6049 for_each_cpu_mask(i, *cpu_map) {
6050 cpumask_t this_core_map = cpu_coregroup_map(i);
6051 cpus_and(this_core_map, this_core_map, *cpu_map);
6052 if (i != first_cpu(this_core_map))
6054 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6059 /* Set up physical groups */
6060 for (i = 0; i < MAX_NUMNODES; i++) {
6061 cpumask_t nodemask = node_to_cpumask(i);
6063 cpus_and(nodemask, nodemask, *cpu_map);
6064 if (cpus_empty(nodemask))
6067 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6071 /* Set up node groups */
6073 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6075 for (i = 0; i < MAX_NUMNODES; i++) {
6076 /* Set up node groups */
6077 struct sched_group *sg, *prev;
6078 cpumask_t nodemask = node_to_cpumask(i);
6079 cpumask_t domainspan;
6080 cpumask_t covered = CPU_MASK_NONE;
6083 cpus_and(nodemask, nodemask, *cpu_map);
6084 if (cpus_empty(nodemask)) {
6085 sched_group_nodes[i] = NULL;
6089 domainspan = sched_domain_node_span(i);
6090 cpus_and(domainspan, domainspan, *cpu_map);
6092 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6094 printk(KERN_WARNING "Can not alloc domain group for "
6098 sched_group_nodes[i] = sg;
6099 for_each_cpu_mask(j, nodemask) {
6100 struct sched_domain *sd;
6101 sd = &per_cpu(node_domains, j);
6104 sg->__cpu_power = 0;
6105 sg->cpumask = nodemask;
6107 cpus_or(covered, covered, nodemask);
6110 for (j = 0; j < MAX_NUMNODES; j++) {
6111 cpumask_t tmp, notcovered;
6112 int n = (i + j) % MAX_NUMNODES;
6114 cpus_complement(notcovered, covered);
6115 cpus_and(tmp, notcovered, *cpu_map);
6116 cpus_and(tmp, tmp, domainspan);
6117 if (cpus_empty(tmp))
6120 nodemask = node_to_cpumask(n);
6121 cpus_and(tmp, tmp, nodemask);
6122 if (cpus_empty(tmp))
6125 sg = kmalloc_node(sizeof(struct sched_group),
6129 "Can not alloc domain group for node %d\n", j);
6132 sg->__cpu_power = 0;
6134 sg->next = prev->next;
6135 cpus_or(covered, covered, tmp);
6142 /* Calculate CPU power for physical packages and nodes */
6143 #ifdef CONFIG_SCHED_SMT
6144 for_each_cpu_mask(i, *cpu_map) {
6145 sd = &per_cpu(cpu_domains, i);
6146 init_sched_groups_power(i, sd);
6149 #ifdef CONFIG_SCHED_MC
6150 for_each_cpu_mask(i, *cpu_map) {
6151 sd = &per_cpu(core_domains, i);
6152 init_sched_groups_power(i, sd);
6156 for_each_cpu_mask(i, *cpu_map) {
6157 sd = &per_cpu(phys_domains, i);
6158 init_sched_groups_power(i, sd);
6162 for (i = 0; i < MAX_NUMNODES; i++)
6163 init_numa_sched_groups_power(sched_group_nodes[i]);
6166 struct sched_group *sg;
6168 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6169 init_numa_sched_groups_power(sg);
6173 /* Attach the domains */
6174 for_each_cpu_mask(i, *cpu_map) {
6175 struct sched_domain *sd;
6176 #ifdef CONFIG_SCHED_SMT
6177 sd = &per_cpu(cpu_domains, i);
6178 #elif defined(CONFIG_SCHED_MC)
6179 sd = &per_cpu(core_domains, i);
6181 sd = &per_cpu(phys_domains, i);
6183 cpu_attach_domain(sd, i);
6190 free_sched_groups(cpu_map);
6195 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6197 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6199 cpumask_t cpu_default_map;
6203 * Setup mask for cpus without special case scheduling requirements.
6204 * For now this just excludes isolated cpus, but could be used to
6205 * exclude other special cases in the future.
6207 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6209 err = build_sched_domains(&cpu_default_map);
6214 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6216 free_sched_groups(cpu_map);
6220 * Detach sched domains from a group of cpus specified in cpu_map
6221 * These cpus will now be attached to the NULL domain
6223 static void detach_destroy_domains(const cpumask_t *cpu_map)
6227 for_each_cpu_mask(i, *cpu_map)
6228 cpu_attach_domain(NULL, i);
6229 synchronize_sched();
6230 arch_destroy_sched_domains(cpu_map);
6234 * Partition sched domains as specified by the cpumasks below.
6235 * This attaches all cpus from the cpumasks to the NULL domain,
6236 * waits for a RCU quiescent period, recalculates sched
6237 * domain information and then attaches them back to the
6238 * correct sched domains
6239 * Call with hotplug lock held
6241 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6243 cpumask_t change_map;
6246 cpus_and(*partition1, *partition1, cpu_online_map);
6247 cpus_and(*partition2, *partition2, cpu_online_map);
6248 cpus_or(change_map, *partition1, *partition2);
6250 /* Detach sched domains from all of the affected cpus */
6251 detach_destroy_domains(&change_map);
6252 if (!cpus_empty(*partition1))
6253 err = build_sched_domains(partition1);
6254 if (!err && !cpus_empty(*partition2))
6255 err = build_sched_domains(partition2);
6260 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6261 int arch_reinit_sched_domains(void)
6265 mutex_lock(&sched_hotcpu_mutex);
6266 detach_destroy_domains(&cpu_online_map);
6267 err = arch_init_sched_domains(&cpu_online_map);
6268 mutex_unlock(&sched_hotcpu_mutex);
6273 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6277 if (buf[0] != '0' && buf[0] != '1')
6281 sched_smt_power_savings = (buf[0] == '1');
6283 sched_mc_power_savings = (buf[0] == '1');
6285 ret = arch_reinit_sched_domains();
6287 return ret ? ret : count;
6290 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6294 #ifdef CONFIG_SCHED_SMT
6296 err = sysfs_create_file(&cls->kset.kobj,
6297 &attr_sched_smt_power_savings.attr);
6299 #ifdef CONFIG_SCHED_MC
6300 if (!err && mc_capable())
6301 err = sysfs_create_file(&cls->kset.kobj,
6302 &attr_sched_mc_power_savings.attr);
6308 #ifdef CONFIG_SCHED_MC
6309 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6311 return sprintf(page, "%u\n", sched_mc_power_savings);
6313 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6314 const char *buf, size_t count)
6316 return sched_power_savings_store(buf, count, 0);
6318 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6319 sched_mc_power_savings_store);
6322 #ifdef CONFIG_SCHED_SMT
6323 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6325 return sprintf(page, "%u\n", sched_smt_power_savings);
6327 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6328 const char *buf, size_t count)
6330 return sched_power_savings_store(buf, count, 1);
6332 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6333 sched_smt_power_savings_store);
6337 * Force a reinitialization of the sched domains hierarchy. The domains
6338 * and groups cannot be updated in place without racing with the balancing
6339 * code, so we temporarily attach all running cpus to the NULL domain
6340 * which will prevent rebalancing while the sched domains are recalculated.
6342 static int update_sched_domains(struct notifier_block *nfb,
6343 unsigned long action, void *hcpu)
6346 case CPU_UP_PREPARE:
6347 case CPU_UP_PREPARE_FROZEN:
6348 case CPU_DOWN_PREPARE:
6349 case CPU_DOWN_PREPARE_FROZEN:
6350 detach_destroy_domains(&cpu_online_map);
6353 case CPU_UP_CANCELED:
6354 case CPU_UP_CANCELED_FROZEN:
6355 case CPU_DOWN_FAILED:
6356 case CPU_DOWN_FAILED_FROZEN:
6358 case CPU_ONLINE_FROZEN:
6360 case CPU_DEAD_FROZEN:
6362 * Fall through and re-initialise the domains.
6369 /* The hotplug lock is already held by cpu_up/cpu_down */
6370 arch_init_sched_domains(&cpu_online_map);
6375 void __init sched_init_smp(void)
6377 cpumask_t non_isolated_cpus;
6379 mutex_lock(&sched_hotcpu_mutex);
6380 arch_init_sched_domains(&cpu_online_map);
6381 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6382 if (cpus_empty(non_isolated_cpus))
6383 cpu_set(smp_processor_id(), non_isolated_cpus);
6384 mutex_unlock(&sched_hotcpu_mutex);
6385 /* XXX: Theoretical race here - CPU may be hotplugged now */
6386 hotcpu_notifier(update_sched_domains, 0);
6388 /* Move init over to a non-isolated CPU */
6389 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6393 void __init sched_init_smp(void)
6396 #endif /* CONFIG_SMP */
6398 int in_sched_functions(unsigned long addr)
6400 /* Linker adds these: start and end of __sched functions */
6401 extern char __sched_text_start[], __sched_text_end[];
6403 return in_lock_functions(addr) ||
6404 (addr >= (unsigned long)__sched_text_start
6405 && addr < (unsigned long)__sched_text_end);
6408 void __init sched_init(void)
6411 int highest_cpu = 0;
6413 for_each_possible_cpu(i) {
6414 struct prio_array *array;
6418 spin_lock_init(&rq->lock);
6419 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6421 rq->active = rq->arrays;
6422 rq->expired = rq->arrays + 1;
6423 rq->best_expired_prio = MAX_PRIO;
6427 for (j = 1; j < 3; j++)
6428 rq->cpu_load[j] = 0;
6429 rq->active_balance = 0;
6432 rq->migration_thread = NULL;
6433 INIT_LIST_HEAD(&rq->migration_queue);
6435 atomic_set(&rq->nr_iowait, 0);
6437 for (j = 0; j < 2; j++) {
6438 array = rq->arrays + j;
6439 for (k = 0; k < MAX_PRIO; k++) {
6440 INIT_LIST_HEAD(array->queue + k);
6441 __clear_bit(k, array->bitmap);
6443 // delimiter for bitsearch
6444 __set_bit(MAX_PRIO, array->bitmap);
6449 set_load_weight(&init_task);
6452 nr_cpu_ids = highest_cpu + 1;
6453 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6456 #ifdef CONFIG_RT_MUTEXES
6457 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6461 * The boot idle thread does lazy MMU switching as well:
6463 atomic_inc(&init_mm.mm_count);
6464 enter_lazy_tlb(&init_mm, current);
6467 * Make us the idle thread. Technically, schedule() should not be
6468 * called from this thread, however somewhere below it might be,
6469 * but because we are the idle thread, we just pick up running again
6470 * when this runqueue becomes "idle".
6472 init_idle(current, smp_processor_id());
6475 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6476 void __might_sleep(char *file, int line)
6479 static unsigned long prev_jiffy; /* ratelimiting */
6481 if ((in_atomic() || irqs_disabled()) &&
6482 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6483 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6485 prev_jiffy = jiffies;
6486 printk(KERN_ERR "BUG: sleeping function called from invalid"
6487 " context at %s:%d\n", file, line);
6488 printk("in_atomic():%d, irqs_disabled():%d\n",
6489 in_atomic(), irqs_disabled());
6490 debug_show_held_locks(current);
6491 if (irqs_disabled())
6492 print_irqtrace_events(current);
6497 EXPORT_SYMBOL(__might_sleep);
6500 #ifdef CONFIG_MAGIC_SYSRQ
6501 void normalize_rt_tasks(void)
6503 struct prio_array *array;
6504 struct task_struct *g, *p;
6505 unsigned long flags;
6508 read_lock_irq(&tasklist_lock);
6510 do_each_thread(g, p) {
6514 spin_lock_irqsave(&p->pi_lock, flags);
6515 rq = __task_rq_lock(p);
6519 deactivate_task(p, task_rq(p));
6520 __setscheduler(p, SCHED_NORMAL, 0);
6522 __activate_task(p, task_rq(p));
6523 resched_task(rq->curr);
6526 __task_rq_unlock(rq);
6527 spin_unlock_irqrestore(&p->pi_lock, flags);
6528 } while_each_thread(g, p);
6530 read_unlock_irq(&tasklist_lock);
6533 #endif /* CONFIG_MAGIC_SYSRQ */
6537 * These functions are only useful for the IA64 MCA handling.
6539 * They can only be called when the whole system has been
6540 * stopped - every CPU needs to be quiescent, and no scheduling
6541 * activity can take place. Using them for anything else would
6542 * be a serious bug, and as a result, they aren't even visible
6543 * under any other configuration.
6547 * curr_task - return the current task for a given cpu.
6548 * @cpu: the processor in question.
6550 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6552 struct task_struct *curr_task(int cpu)
6554 return cpu_curr(cpu);
6558 * set_curr_task - set the current task for a given cpu.
6559 * @cpu: the processor in question.
6560 * @p: the task pointer to set.
6562 * Description: This function must only be used when non-maskable interrupts
6563 * are serviced on a separate stack. It allows the architecture to switch the
6564 * notion of the current task on a cpu in a non-blocking manner. This function
6565 * must be called with all CPU's synchronized, and interrupts disabled, the
6566 * and caller must save the original value of the current task (see
6567 * curr_task() above) and restore that value before reenabling interrupts and
6568 * re-starting the system.
6570 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6572 void set_curr_task(int cpu, struct task_struct *p)