4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
95 * These are the 'tuning knobs' of the scheduler:
97 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
98 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
99 * Timeslices get refilled after they expire.
101 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
102 #define DEF_TIMESLICE (100 * HZ / 1000)
103 #define ON_RUNQUEUE_WEIGHT 30
104 #define CHILD_PENALTY 95
105 #define PARENT_PENALTY 100
106 #define EXIT_WEIGHT 3
107 #define PRIO_BONUS_RATIO 25
108 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
109 #define INTERACTIVE_DELTA 2
110 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
111 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
112 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
115 * If a task is 'interactive' then we reinsert it in the active
116 * array after it has expired its current timeslice. (it will not
117 * continue to run immediately, it will still roundrobin with
118 * other interactive tasks.)
120 * This part scales the interactivity limit depending on niceness.
122 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
123 * Here are a few examples of different nice levels:
125 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
126 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
127 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
129 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
131 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
132 * priority range a task can explore, a value of '1' means the
133 * task is rated interactive.)
135 * Ie. nice +19 tasks can never get 'interactive' enough to be
136 * reinserted into the active array. And only heavily CPU-hog nice -20
137 * tasks will be expired. Default nice 0 tasks are somewhere between,
138 * it takes some effort for them to get interactive, but it's not
142 #define CURRENT_BONUS(p) \
143 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
146 #define GRANULARITY (10 * HZ / 1000 ? : 1)
149 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
150 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
153 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
154 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
157 #define SCALE(v1,v1_max,v2_max) \
158 (v1) * (v2_max) / (v1_max)
161 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
164 #define TASK_INTERACTIVE(p) \
165 ((p)->prio <= (p)->static_prio - DELTA(p))
167 #define INTERACTIVE_SLEEP(p) \
168 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
169 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
171 #define TASK_PREEMPTS_CURR(p, rq) \
172 ((p)->prio < (rq)->curr->prio)
174 #define SCALE_PRIO(x, prio) \
175 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
177 static unsigned int static_prio_timeslice(int static_prio)
179 if (static_prio < NICE_TO_PRIO(0))
180 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
182 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
187 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
188 * Since cpu_power is a 'constant', we can use a reciprocal divide.
190 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
192 return reciprocal_divide(load, sg->reciprocal_cpu_power);
196 * Each time a sched group cpu_power is changed,
197 * we must compute its reciprocal value
199 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
201 sg->__cpu_power += val;
202 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
207 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
208 * to time slice values: [800ms ... 100ms ... 5ms]
210 * The higher a thread's priority, the bigger timeslices
211 * it gets during one round of execution. But even the lowest
212 * priority thread gets MIN_TIMESLICE worth of execution time.
215 static inline unsigned int task_timeslice(struct task_struct *p)
217 return static_prio_timeslice(p->static_prio);
221 * These are the runqueue data structures:
225 unsigned int nr_active;
226 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
227 struct list_head queue[MAX_PRIO];
231 * This is the main, per-CPU runqueue data structure.
233 * Locking rule: those places that want to lock multiple runqueues
234 * (such as the load balancing or the thread migration code), lock
235 * acquire operations must be ordered by ascending &runqueue.
241 * nr_running and cpu_load should be in the same cacheline because
242 * remote CPUs use both these fields when doing load calculation.
244 unsigned long nr_running;
245 unsigned long raw_weighted_load;
247 unsigned long cpu_load[3];
248 unsigned char idle_at_tick;
250 unsigned char in_nohz_recently;
253 unsigned long long nr_switches;
256 * This is part of a global counter where only the total sum
257 * over all CPUs matters. A task can increase this counter on
258 * one CPU and if it got migrated afterwards it may decrease
259 * it on another CPU. Always updated under the runqueue lock:
261 unsigned long nr_uninterruptible;
263 unsigned long expired_timestamp;
264 /* Cached timestamp set by update_cpu_clock() */
265 unsigned long long most_recent_timestamp;
266 struct task_struct *curr, *idle;
267 unsigned long next_balance;
268 struct mm_struct *prev_mm;
269 struct prio_array *active, *expired, arrays[2];
270 int best_expired_prio;
274 struct sched_domain *sd;
276 /* For active balancing */
279 int cpu; /* cpu of this runqueue */
281 struct task_struct *migration_thread;
282 struct list_head migration_queue;
285 #ifdef CONFIG_SCHEDSTATS
287 struct sched_info rq_sched_info;
289 /* sys_sched_yield() stats */
290 unsigned long yld_exp_empty;
291 unsigned long yld_act_empty;
292 unsigned long yld_both_empty;
293 unsigned long yld_cnt;
295 /* schedule() stats */
296 unsigned long sched_switch;
297 unsigned long sched_cnt;
298 unsigned long sched_goidle;
300 /* try_to_wake_up() stats */
301 unsigned long ttwu_cnt;
302 unsigned long ttwu_local;
304 struct lock_class_key rq_lock_key;
307 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
308 static DEFINE_MUTEX(sched_hotcpu_mutex);
310 static inline int cpu_of(struct rq *rq)
320 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
321 * See detach_destroy_domains: synchronize_sched for details.
323 * The domain tree of any CPU may only be accessed from within
324 * preempt-disabled sections.
326 #define for_each_domain(cpu, __sd) \
327 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
329 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
330 #define this_rq() (&__get_cpu_var(runqueues))
331 #define task_rq(p) cpu_rq(task_cpu(p))
332 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
334 #ifndef prepare_arch_switch
335 # define prepare_arch_switch(next) do { } while (0)
337 #ifndef finish_arch_switch
338 # define finish_arch_switch(prev) do { } while (0)
341 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
342 static inline int task_running(struct rq *rq, struct task_struct *p)
344 return rq->curr == p;
347 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
351 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
353 #ifdef CONFIG_DEBUG_SPINLOCK
354 /* this is a valid case when another task releases the spinlock */
355 rq->lock.owner = current;
358 * If we are tracking spinlock dependencies then we have to
359 * fix up the runqueue lock - which gets 'carried over' from
362 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
364 spin_unlock_irq(&rq->lock);
367 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
368 static inline int task_running(struct rq *rq, struct task_struct *p)
373 return rq->curr == p;
377 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
381 * We can optimise this out completely for !SMP, because the
382 * SMP rebalancing from interrupt is the only thing that cares
387 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
388 spin_unlock_irq(&rq->lock);
390 spin_unlock(&rq->lock);
394 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
398 * After ->oncpu is cleared, the task can be moved to a different CPU.
399 * We must ensure this doesn't happen until the switch is completely
405 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
409 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
412 * __task_rq_lock - lock the runqueue a given task resides on.
413 * Must be called interrupts disabled.
415 static inline struct rq *__task_rq_lock(struct task_struct *p)
422 spin_lock(&rq->lock);
423 if (unlikely(rq != task_rq(p))) {
424 spin_unlock(&rq->lock);
425 goto repeat_lock_task;
431 * task_rq_lock - lock the runqueue a given task resides on and disable
432 * interrupts. Note the ordering: we can safely lookup the task_rq without
433 * explicitly disabling preemption.
435 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
441 local_irq_save(*flags);
443 spin_lock(&rq->lock);
444 if (unlikely(rq != task_rq(p))) {
445 spin_unlock_irqrestore(&rq->lock, *flags);
446 goto repeat_lock_task;
451 static inline void __task_rq_unlock(struct rq *rq)
454 spin_unlock(&rq->lock);
457 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
460 spin_unlock_irqrestore(&rq->lock, *flags);
463 #ifdef CONFIG_SCHEDSTATS
465 * bump this up when changing the output format or the meaning of an existing
466 * format, so that tools can adapt (or abort)
468 #define SCHEDSTAT_VERSION 14
470 static int show_schedstat(struct seq_file *seq, void *v)
474 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
475 seq_printf(seq, "timestamp %lu\n", jiffies);
476 for_each_online_cpu(cpu) {
477 struct rq *rq = cpu_rq(cpu);
479 struct sched_domain *sd;
483 /* runqueue-specific stats */
485 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
486 cpu, rq->yld_both_empty,
487 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
488 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
489 rq->ttwu_cnt, rq->ttwu_local,
490 rq->rq_sched_info.cpu_time,
491 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
493 seq_printf(seq, "\n");
496 /* domain-specific stats */
498 for_each_domain(cpu, sd) {
499 enum cpu_idle_type itype;
500 char mask_str[NR_CPUS];
502 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
503 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
504 for (itype = CPU_IDLE; itype < CPU_MAX_IDLE_TYPES;
506 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu "
509 sd->lb_balanced[itype],
510 sd->lb_failed[itype],
511 sd->lb_imbalance[itype],
512 sd->lb_gained[itype],
513 sd->lb_hot_gained[itype],
514 sd->lb_nobusyq[itype],
515 sd->lb_nobusyg[itype]);
517 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
519 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
520 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
521 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
522 sd->ttwu_wake_remote, sd->ttwu_move_affine,
523 sd->ttwu_move_balance);
531 static int schedstat_open(struct inode *inode, struct file *file)
533 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
534 char *buf = kmalloc(size, GFP_KERNEL);
540 res = single_open(file, show_schedstat, NULL);
542 m = file->private_data;
550 const struct file_operations proc_schedstat_operations = {
551 .open = schedstat_open,
554 .release = single_release,
558 * Expects runqueue lock to be held for atomicity of update
561 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
564 rq->rq_sched_info.run_delay += delta_jiffies;
565 rq->rq_sched_info.pcnt++;
570 * Expects runqueue lock to be held for atomicity of update
573 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
576 rq->rq_sched_info.cpu_time += delta_jiffies;
578 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
579 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
580 #else /* !CONFIG_SCHEDSTATS */
582 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
585 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
587 # define schedstat_inc(rq, field) do { } while (0)
588 # define schedstat_add(rq, field, amt) do { } while (0)
592 * this_rq_lock - lock this runqueue and disable interrupts.
594 static inline struct rq *this_rq_lock(void)
601 spin_lock(&rq->lock);
606 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
608 * Called when a process is dequeued from the active array and given
609 * the cpu. We should note that with the exception of interactive
610 * tasks, the expired queue will become the active queue after the active
611 * queue is empty, without explicitly dequeuing and requeuing tasks in the
612 * expired queue. (Interactive tasks may be requeued directly to the
613 * active queue, thus delaying tasks in the expired queue from running;
614 * see scheduler_tick()).
616 * This function is only called from sched_info_arrive(), rather than
617 * dequeue_task(). Even though a task may be queued and dequeued multiple
618 * times as it is shuffled about, we're really interested in knowing how
619 * long it was from the *first* time it was queued to the time that it
622 static inline void sched_info_dequeued(struct task_struct *t)
624 t->sched_info.last_queued = 0;
628 * Called when a task finally hits the cpu. We can now calculate how
629 * long it was waiting to run. We also note when it began so that we
630 * can keep stats on how long its timeslice is.
632 static void sched_info_arrive(struct task_struct *t)
634 unsigned long now = jiffies, delta_jiffies = 0;
636 if (t->sched_info.last_queued)
637 delta_jiffies = now - t->sched_info.last_queued;
638 sched_info_dequeued(t);
639 t->sched_info.run_delay += delta_jiffies;
640 t->sched_info.last_arrival = now;
641 t->sched_info.pcnt++;
643 rq_sched_info_arrive(task_rq(t), delta_jiffies);
647 * Called when a process is queued into either the active or expired
648 * array. The time is noted and later used to determine how long we
649 * had to wait for us to reach the cpu. Since the expired queue will
650 * become the active queue after active queue is empty, without dequeuing
651 * and requeuing any tasks, we are interested in queuing to either. It
652 * is unusual but not impossible for tasks to be dequeued and immediately
653 * requeued in the same or another array: this can happen in sched_yield(),
654 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
657 * This function is only called from enqueue_task(), but also only updates
658 * the timestamp if it is already not set. It's assumed that
659 * sched_info_dequeued() will clear that stamp when appropriate.
661 static inline void sched_info_queued(struct task_struct *t)
663 if (unlikely(sched_info_on()))
664 if (!t->sched_info.last_queued)
665 t->sched_info.last_queued = jiffies;
669 * Called when a process ceases being the active-running process, either
670 * voluntarily or involuntarily. Now we can calculate how long we ran.
672 static inline void sched_info_depart(struct task_struct *t)
674 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
676 t->sched_info.cpu_time += delta_jiffies;
677 rq_sched_info_depart(task_rq(t), delta_jiffies);
681 * Called when tasks are switched involuntarily due, typically, to expiring
682 * their time slice. (This may also be called when switching to or from
683 * the idle task.) We are only called when prev != next.
686 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
688 struct rq *rq = task_rq(prev);
691 * prev now departs the cpu. It's not interesting to record
692 * stats about how efficient we were at scheduling the idle
695 if (prev != rq->idle)
696 sched_info_depart(prev);
698 if (next != rq->idle)
699 sched_info_arrive(next);
702 sched_info_switch(struct task_struct *prev, struct task_struct *next)
704 if (unlikely(sched_info_on()))
705 __sched_info_switch(prev, next);
708 #define sched_info_queued(t) do { } while (0)
709 #define sched_info_switch(t, next) do { } while (0)
710 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
713 * Adding/removing a task to/from a priority array:
715 static void dequeue_task(struct task_struct *p, struct prio_array *array)
718 list_del(&p->run_list);
719 if (list_empty(array->queue + p->prio))
720 __clear_bit(p->prio, array->bitmap);
723 static void enqueue_task(struct task_struct *p, struct prio_array *array)
725 sched_info_queued(p);
726 list_add_tail(&p->run_list, array->queue + p->prio);
727 __set_bit(p->prio, array->bitmap);
733 * Put task to the end of the run list without the overhead of dequeue
734 * followed by enqueue.
736 static void requeue_task(struct task_struct *p, struct prio_array *array)
738 list_move_tail(&p->run_list, array->queue + p->prio);
742 enqueue_task_head(struct task_struct *p, struct prio_array *array)
744 list_add(&p->run_list, array->queue + p->prio);
745 __set_bit(p->prio, array->bitmap);
751 * __normal_prio - return the priority that is based on the static
752 * priority but is modified by bonuses/penalties.
754 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
755 * into the -5 ... 0 ... +5 bonus/penalty range.
757 * We use 25% of the full 0...39 priority range so that:
759 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
760 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
762 * Both properties are important to certain workloads.
765 static inline int __normal_prio(struct task_struct *p)
769 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
771 prio = p->static_prio - bonus;
772 if (prio < MAX_RT_PRIO)
774 if (prio > MAX_PRIO-1)
780 * To aid in avoiding the subversion of "niceness" due to uneven distribution
781 * of tasks with abnormal "nice" values across CPUs the contribution that
782 * each task makes to its run queue's load is weighted according to its
783 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
784 * scaled version of the new time slice allocation that they receive on time
789 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
790 * If static_prio_timeslice() is ever changed to break this assumption then
791 * this code will need modification
793 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
794 #define LOAD_WEIGHT(lp) \
795 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
796 #define PRIO_TO_LOAD_WEIGHT(prio) \
797 LOAD_WEIGHT(static_prio_timeslice(prio))
798 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
799 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
801 static void set_load_weight(struct task_struct *p)
803 if (has_rt_policy(p)) {
805 if (p == task_rq(p)->migration_thread)
807 * The migration thread does the actual balancing.
808 * Giving its load any weight will skew balancing
814 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
816 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
820 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
822 rq->raw_weighted_load += p->load_weight;
826 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
828 rq->raw_weighted_load -= p->load_weight;
831 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
834 inc_raw_weighted_load(rq, p);
837 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
840 dec_raw_weighted_load(rq, p);
844 * Calculate the expected normal priority: i.e. priority
845 * without taking RT-inheritance into account. Might be
846 * boosted by interactivity modifiers. Changes upon fork,
847 * setprio syscalls, and whenever the interactivity
848 * estimator recalculates.
850 static inline int normal_prio(struct task_struct *p)
854 if (has_rt_policy(p))
855 prio = MAX_RT_PRIO-1 - p->rt_priority;
857 prio = __normal_prio(p);
862 * Calculate the current priority, i.e. the priority
863 * taken into account by the scheduler. This value might
864 * be boosted by RT tasks, or might be boosted by
865 * interactivity modifiers. Will be RT if the task got
866 * RT-boosted. If not then it returns p->normal_prio.
868 static int effective_prio(struct task_struct *p)
870 p->normal_prio = normal_prio(p);
872 * If we are RT tasks or we were boosted to RT priority,
873 * keep the priority unchanged. Otherwise, update priority
874 * to the normal priority:
876 if (!rt_prio(p->prio))
877 return p->normal_prio;
882 * __activate_task - move a task to the runqueue.
884 static void __activate_task(struct task_struct *p, struct rq *rq)
886 struct prio_array *target = rq->active;
889 target = rq->expired;
890 enqueue_task(p, target);
891 inc_nr_running(p, rq);
895 * __activate_idle_task - move idle task to the _front_ of runqueue.
897 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
899 enqueue_task_head(p, rq->active);
900 inc_nr_running(p, rq);
904 * Recalculate p->normal_prio and p->prio after having slept,
905 * updating the sleep-average too:
907 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
909 /* Caller must always ensure 'now >= p->timestamp' */
910 unsigned long sleep_time = now - p->timestamp;
915 if (likely(sleep_time > 0)) {
917 * This ceiling is set to the lowest priority that would allow
918 * a task to be reinserted into the active array on timeslice
921 unsigned long ceiling = INTERACTIVE_SLEEP(p);
923 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
925 * Prevents user tasks from achieving best priority
926 * with one single large enough sleep.
928 p->sleep_avg = ceiling;
930 * Using INTERACTIVE_SLEEP() as a ceiling places a
931 * nice(0) task 1ms sleep away from promotion, and
932 * gives it 700ms to round-robin with no chance of
933 * being demoted. This is more than generous, so
934 * mark this sleep as non-interactive to prevent the
935 * on-runqueue bonus logic from intervening should
936 * this task not receive cpu immediately.
938 p->sleep_type = SLEEP_NONINTERACTIVE;
941 * Tasks waking from uninterruptible sleep are
942 * limited in their sleep_avg rise as they
943 * are likely to be waiting on I/O
945 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
946 if (p->sleep_avg >= ceiling)
948 else if (p->sleep_avg + sleep_time >=
950 p->sleep_avg = ceiling;
956 * This code gives a bonus to interactive tasks.
958 * The boost works by updating the 'average sleep time'
959 * value here, based on ->timestamp. The more time a
960 * task spends sleeping, the higher the average gets -
961 * and the higher the priority boost gets as well.
963 p->sleep_avg += sleep_time;
966 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
967 p->sleep_avg = NS_MAX_SLEEP_AVG;
970 return effective_prio(p);
974 * activate_task - move a task to the runqueue and do priority recalculation
976 * Update all the scheduling statistics stuff. (sleep average
977 * calculation, priority modifiers, etc.)
979 static void activate_task(struct task_struct *p, struct rq *rq, int local)
981 unsigned long long now;
989 /* Compensate for drifting sched_clock */
990 struct rq *this_rq = this_rq();
991 now = (now - this_rq->most_recent_timestamp)
992 + rq->most_recent_timestamp;
997 * Sleep time is in units of nanosecs, so shift by 20 to get a
998 * milliseconds-range estimation of the amount of time that the task
1001 if (unlikely(prof_on == SLEEP_PROFILING)) {
1002 if (p->state == TASK_UNINTERRUPTIBLE)
1003 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
1004 (now - p->timestamp) >> 20);
1007 p->prio = recalc_task_prio(p, now);
1010 * This checks to make sure it's not an uninterruptible task
1011 * that is now waking up.
1013 if (p->sleep_type == SLEEP_NORMAL) {
1015 * Tasks which were woken up by interrupts (ie. hw events)
1016 * are most likely of interactive nature. So we give them
1017 * the credit of extending their sleep time to the period
1018 * of time they spend on the runqueue, waiting for execution
1019 * on a CPU, first time around:
1022 p->sleep_type = SLEEP_INTERRUPTED;
1025 * Normal first-time wakeups get a credit too for
1026 * on-runqueue time, but it will be weighted down:
1028 p->sleep_type = SLEEP_INTERACTIVE;
1033 __activate_task(p, rq);
1037 * deactivate_task - remove a task from the runqueue.
1039 static void deactivate_task(struct task_struct *p, struct rq *rq)
1041 dec_nr_running(p, rq);
1042 dequeue_task(p, p->array);
1047 * resched_task - mark a task 'to be rescheduled now'.
1049 * On UP this means the setting of the need_resched flag, on SMP it
1050 * might also involve a cross-CPU call to trigger the scheduler on
1055 #ifndef tsk_is_polling
1056 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1059 static void resched_task(struct task_struct *p)
1063 assert_spin_locked(&task_rq(p)->lock);
1065 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1068 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1071 if (cpu == smp_processor_id())
1074 /* NEED_RESCHED must be visible before we test polling */
1076 if (!tsk_is_polling(p))
1077 smp_send_reschedule(cpu);
1080 static void resched_cpu(int cpu)
1082 struct rq *rq = cpu_rq(cpu);
1083 unsigned long flags;
1085 if (!spin_trylock_irqsave(&rq->lock, flags))
1087 resched_task(cpu_curr(cpu));
1088 spin_unlock_irqrestore(&rq->lock, flags);
1091 static inline void resched_task(struct task_struct *p)
1093 assert_spin_locked(&task_rq(p)->lock);
1094 set_tsk_need_resched(p);
1099 * task_curr - is this task currently executing on a CPU?
1100 * @p: the task in question.
1102 inline int task_curr(const struct task_struct *p)
1104 return cpu_curr(task_cpu(p)) == p;
1107 /* Used instead of source_load when we know the type == 0 */
1108 unsigned long weighted_cpuload(const int cpu)
1110 return cpu_rq(cpu)->raw_weighted_load;
1115 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1117 task_thread_info(p)->cpu = cpu;
1120 struct migration_req {
1121 struct list_head list;
1123 struct task_struct *task;
1126 struct completion done;
1130 * The task's runqueue lock must be held.
1131 * Returns true if you have to wait for migration thread.
1134 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1136 struct rq *rq = task_rq(p);
1139 * If the task is not on a runqueue (and not running), then
1140 * it is sufficient to simply update the task's cpu field.
1142 if (!p->array && !task_running(rq, p)) {
1143 set_task_cpu(p, dest_cpu);
1147 init_completion(&req->done);
1149 req->dest_cpu = dest_cpu;
1150 list_add(&req->list, &rq->migration_queue);
1156 * wait_task_inactive - wait for a thread to unschedule.
1158 * The caller must ensure that the task *will* unschedule sometime soon,
1159 * else this function might spin for a *long* time. This function can't
1160 * be called with interrupts off, or it may introduce deadlock with
1161 * smp_call_function() if an IPI is sent by the same process we are
1162 * waiting to become inactive.
1164 void wait_task_inactive(struct task_struct *p)
1166 unsigned long flags;
1168 struct prio_array *array;
1173 * We do the initial early heuristics without holding
1174 * any task-queue locks at all. We'll only try to get
1175 * the runqueue lock when things look like they will
1181 * If the task is actively running on another CPU
1182 * still, just relax and busy-wait without holding
1185 * NOTE! Since we don't hold any locks, it's not
1186 * even sure that "rq" stays as the right runqueue!
1187 * But we don't care, since "task_running()" will
1188 * return false if the runqueue has changed and p
1189 * is actually now running somewhere else!
1191 while (task_running(rq, p))
1195 * Ok, time to look more closely! We need the rq
1196 * lock now, to be *sure*. If we're wrong, we'll
1197 * just go back and repeat.
1199 rq = task_rq_lock(p, &flags);
1200 running = task_running(rq, p);
1202 task_rq_unlock(rq, &flags);
1205 * Was it really running after all now that we
1206 * checked with the proper locks actually held?
1208 * Oops. Go back and try again..
1210 if (unlikely(running)) {
1216 * It's not enough that it's not actively running,
1217 * it must be off the runqueue _entirely_, and not
1220 * So if it wa still runnable (but just not actively
1221 * running right now), it's preempted, and we should
1222 * yield - it could be a while.
1224 if (unlikely(array)) {
1230 * Ahh, all good. It wasn't running, and it wasn't
1231 * runnable, which means that it will never become
1232 * running in the future either. We're all done!
1237 * kick_process - kick a running thread to enter/exit the kernel
1238 * @p: the to-be-kicked thread
1240 * Cause a process which is running on another CPU to enter
1241 * kernel-mode, without any delay. (to get signals handled.)
1243 * NOTE: this function doesnt have to take the runqueue lock,
1244 * because all it wants to ensure is that the remote task enters
1245 * the kernel. If the IPI races and the task has been migrated
1246 * to another CPU then no harm is done and the purpose has been
1249 void kick_process(struct task_struct *p)
1255 if ((cpu != smp_processor_id()) && task_curr(p))
1256 smp_send_reschedule(cpu);
1261 * Return a low guess at the load of a migration-source cpu weighted
1262 * according to the scheduling class and "nice" value.
1264 * We want to under-estimate the load of migration sources, to
1265 * balance conservatively.
1267 static inline unsigned long source_load(int cpu, int type)
1269 struct rq *rq = cpu_rq(cpu);
1272 return rq->raw_weighted_load;
1274 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1278 * Return a high guess at the load of a migration-target cpu weighted
1279 * according to the scheduling class and "nice" value.
1281 static inline unsigned long target_load(int cpu, int type)
1283 struct rq *rq = cpu_rq(cpu);
1286 return rq->raw_weighted_load;
1288 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1292 * Return the average load per task on the cpu's run queue
1294 static inline unsigned long cpu_avg_load_per_task(int cpu)
1296 struct rq *rq = cpu_rq(cpu);
1297 unsigned long n = rq->nr_running;
1299 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1303 * find_idlest_group finds and returns the least busy CPU group within the
1306 static struct sched_group *
1307 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1309 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1310 unsigned long min_load = ULONG_MAX, this_load = 0;
1311 int load_idx = sd->forkexec_idx;
1312 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1315 unsigned long load, avg_load;
1319 /* Skip over this group if it has no CPUs allowed */
1320 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1323 local_group = cpu_isset(this_cpu, group->cpumask);
1325 /* Tally up the load of all CPUs in the group */
1328 for_each_cpu_mask(i, group->cpumask) {
1329 /* Bias balancing toward cpus of our domain */
1331 load = source_load(i, load_idx);
1333 load = target_load(i, load_idx);
1338 /* Adjust by relative CPU power of the group */
1339 avg_load = sg_div_cpu_power(group,
1340 avg_load * SCHED_LOAD_SCALE);
1343 this_load = avg_load;
1345 } else if (avg_load < min_load) {
1346 min_load = avg_load;
1350 group = group->next;
1351 } while (group != sd->groups);
1353 if (!idlest || 100*this_load < imbalance*min_load)
1359 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1362 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1365 unsigned long load, min_load = ULONG_MAX;
1369 /* Traverse only the allowed CPUs */
1370 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1372 for_each_cpu_mask(i, tmp) {
1373 load = weighted_cpuload(i);
1375 if (load < min_load || (load == min_load && i == this_cpu)) {
1385 * sched_balance_self: balance the current task (running on cpu) in domains
1386 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1389 * Balance, ie. select the least loaded group.
1391 * Returns the target CPU number, or the same CPU if no balancing is needed.
1393 * preempt must be disabled.
1395 static int sched_balance_self(int cpu, int flag)
1397 struct task_struct *t = current;
1398 struct sched_domain *tmp, *sd = NULL;
1400 for_each_domain(cpu, tmp) {
1402 * If power savings logic is enabled for a domain, stop there.
1404 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1406 if (tmp->flags & flag)
1412 struct sched_group *group;
1413 int new_cpu, weight;
1415 if (!(sd->flags & flag)) {
1421 group = find_idlest_group(sd, t, cpu);
1427 new_cpu = find_idlest_cpu(group, t, cpu);
1428 if (new_cpu == -1 || new_cpu == cpu) {
1429 /* Now try balancing at a lower domain level of cpu */
1434 /* Now try balancing at a lower domain level of new_cpu */
1437 weight = cpus_weight(span);
1438 for_each_domain(cpu, tmp) {
1439 if (weight <= cpus_weight(tmp->span))
1441 if (tmp->flags & flag)
1444 /* while loop will break here if sd == NULL */
1450 #endif /* CONFIG_SMP */
1453 * wake_idle() will wake a task on an idle cpu if task->cpu is
1454 * not idle and an idle cpu is available. The span of cpus to
1455 * search starts with cpus closest then further out as needed,
1456 * so we always favor a closer, idle cpu.
1458 * Returns the CPU we should wake onto.
1460 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1461 static int wake_idle(int cpu, struct task_struct *p)
1464 struct sched_domain *sd;
1468 * If it is idle, then it is the best cpu to run this task.
1470 * This cpu is also the best, if it has more than one task already.
1471 * Siblings must be also busy(in most cases) as they didn't already
1472 * pickup the extra load from this cpu and hence we need not check
1473 * sibling runqueue info. This will avoid the checks and cache miss
1474 * penalities associated with that.
1476 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1479 for_each_domain(cpu, sd) {
1480 if (sd->flags & SD_WAKE_IDLE) {
1481 cpus_and(tmp, sd->span, p->cpus_allowed);
1482 for_each_cpu_mask(i, tmp) {
1493 static inline int wake_idle(int cpu, struct task_struct *p)
1500 * try_to_wake_up - wake up a thread
1501 * @p: the to-be-woken-up thread
1502 * @state: the mask of task states that can be woken
1503 * @sync: do a synchronous wakeup?
1505 * Put it on the run-queue if it's not already there. The "current"
1506 * thread is always on the run-queue (except when the actual
1507 * re-schedule is in progress), and as such you're allowed to do
1508 * the simpler "current->state = TASK_RUNNING" to mark yourself
1509 * runnable without the overhead of this.
1511 * returns failure only if the task is already active.
1513 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1515 int cpu, this_cpu, success = 0;
1516 unsigned long flags;
1520 struct sched_domain *sd, *this_sd = NULL;
1521 unsigned long load, this_load;
1525 rq = task_rq_lock(p, &flags);
1526 old_state = p->state;
1527 if (!(old_state & state))
1534 this_cpu = smp_processor_id();
1537 if (unlikely(task_running(rq, p)))
1542 schedstat_inc(rq, ttwu_cnt);
1543 if (cpu == this_cpu) {
1544 schedstat_inc(rq, ttwu_local);
1548 for_each_domain(this_cpu, sd) {
1549 if (cpu_isset(cpu, sd->span)) {
1550 schedstat_inc(sd, ttwu_wake_remote);
1556 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1560 * Check for affine wakeup and passive balancing possibilities.
1563 int idx = this_sd->wake_idx;
1564 unsigned int imbalance;
1566 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1568 load = source_load(cpu, idx);
1569 this_load = target_load(this_cpu, idx);
1571 new_cpu = this_cpu; /* Wake to this CPU if we can */
1573 if (this_sd->flags & SD_WAKE_AFFINE) {
1574 unsigned long tl = this_load;
1575 unsigned long tl_per_task;
1577 tl_per_task = cpu_avg_load_per_task(this_cpu);
1580 * If sync wakeup then subtract the (maximum possible)
1581 * effect of the currently running task from the load
1582 * of the current CPU:
1585 tl -= current->load_weight;
1588 tl + target_load(cpu, idx) <= tl_per_task) ||
1589 100*(tl + p->load_weight) <= imbalance*load) {
1591 * This domain has SD_WAKE_AFFINE and
1592 * p is cache cold in this domain, and
1593 * there is no bad imbalance.
1595 schedstat_inc(this_sd, ttwu_move_affine);
1601 * Start passive balancing when half the imbalance_pct
1604 if (this_sd->flags & SD_WAKE_BALANCE) {
1605 if (imbalance*this_load <= 100*load) {
1606 schedstat_inc(this_sd, ttwu_move_balance);
1612 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1614 new_cpu = wake_idle(new_cpu, p);
1615 if (new_cpu != cpu) {
1616 set_task_cpu(p, new_cpu);
1617 task_rq_unlock(rq, &flags);
1618 /* might preempt at this point */
1619 rq = task_rq_lock(p, &flags);
1620 old_state = p->state;
1621 if (!(old_state & state))
1626 this_cpu = smp_processor_id();
1631 #endif /* CONFIG_SMP */
1632 if (old_state == TASK_UNINTERRUPTIBLE) {
1633 rq->nr_uninterruptible--;
1635 * Tasks on involuntary sleep don't earn
1636 * sleep_avg beyond just interactive state.
1638 p->sleep_type = SLEEP_NONINTERACTIVE;
1642 * Tasks that have marked their sleep as noninteractive get
1643 * woken up with their sleep average not weighted in an
1646 if (old_state & TASK_NONINTERACTIVE)
1647 p->sleep_type = SLEEP_NONINTERACTIVE;
1650 activate_task(p, rq, cpu == this_cpu);
1652 * Sync wakeups (i.e. those types of wakeups where the waker
1653 * has indicated that it will leave the CPU in short order)
1654 * don't trigger a preemption, if the woken up task will run on
1655 * this cpu. (in this case the 'I will reschedule' promise of
1656 * the waker guarantees that the freshly woken up task is going
1657 * to be considered on this CPU.)
1659 if (!sync || cpu != this_cpu) {
1660 if (TASK_PREEMPTS_CURR(p, rq))
1661 resched_task(rq->curr);
1666 p->state = TASK_RUNNING;
1668 task_rq_unlock(rq, &flags);
1673 int fastcall wake_up_process(struct task_struct *p)
1675 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1676 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1678 EXPORT_SYMBOL(wake_up_process);
1680 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1682 return try_to_wake_up(p, state, 0);
1685 static void task_running_tick(struct rq *rq, struct task_struct *p);
1687 * Perform scheduler related setup for a newly forked process p.
1688 * p is forked by current.
1690 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1692 int cpu = get_cpu();
1695 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1697 set_task_cpu(p, cpu);
1700 * We mark the process as running here, but have not actually
1701 * inserted it onto the runqueue yet. This guarantees that
1702 * nobody will actually run it, and a signal or other external
1703 * event cannot wake it up and insert it on the runqueue either.
1705 p->state = TASK_RUNNING;
1708 * Make sure we do not leak PI boosting priority to the child:
1710 p->prio = current->normal_prio;
1712 INIT_LIST_HEAD(&p->run_list);
1714 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1715 if (unlikely(sched_info_on()))
1716 memset(&p->sched_info, 0, sizeof(p->sched_info));
1718 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1721 #ifdef CONFIG_PREEMPT
1722 /* Want to start with kernel preemption disabled. */
1723 task_thread_info(p)->preempt_count = 1;
1726 * Share the timeslice between parent and child, thus the
1727 * total amount of pending timeslices in the system doesn't change,
1728 * resulting in more scheduling fairness.
1730 local_irq_disable();
1731 p->time_slice = (current->time_slice + 1) >> 1;
1733 * The remainder of the first timeslice might be recovered by
1734 * the parent if the child exits early enough.
1736 p->first_time_slice = 1;
1737 current->time_slice >>= 1;
1738 p->timestamp = sched_clock();
1739 if (unlikely(!current->time_slice)) {
1741 * This case is rare, it happens when the parent has only
1742 * a single jiffy left from its timeslice. Taking the
1743 * runqueue lock is not a problem.
1745 current->time_slice = 1;
1746 task_running_tick(cpu_rq(cpu), current);
1753 * wake_up_new_task - wake up a newly created task for the first time.
1755 * This function will do some initial scheduler statistics housekeeping
1756 * that must be done for every newly created context, then puts the task
1757 * on the runqueue and wakes it.
1759 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1761 struct rq *rq, *this_rq;
1762 unsigned long flags;
1765 rq = task_rq_lock(p, &flags);
1766 BUG_ON(p->state != TASK_RUNNING);
1767 this_cpu = smp_processor_id();
1771 * We decrease the sleep average of forking parents
1772 * and children as well, to keep max-interactive tasks
1773 * from forking tasks that are max-interactive. The parent
1774 * (current) is done further down, under its lock.
1776 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1777 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1779 p->prio = effective_prio(p);
1781 if (likely(cpu == this_cpu)) {
1782 if (!(clone_flags & CLONE_VM)) {
1784 * The VM isn't cloned, so we're in a good position to
1785 * do child-runs-first in anticipation of an exec. This
1786 * usually avoids a lot of COW overhead.
1788 if (unlikely(!current->array))
1789 __activate_task(p, rq);
1791 p->prio = current->prio;
1792 p->normal_prio = current->normal_prio;
1793 list_add_tail(&p->run_list, ¤t->run_list);
1794 p->array = current->array;
1795 p->array->nr_active++;
1796 inc_nr_running(p, rq);
1800 /* Run child last */
1801 __activate_task(p, rq);
1803 * We skip the following code due to cpu == this_cpu
1805 * task_rq_unlock(rq, &flags);
1806 * this_rq = task_rq_lock(current, &flags);
1810 this_rq = cpu_rq(this_cpu);
1813 * Not the local CPU - must adjust timestamp. This should
1814 * get optimised away in the !CONFIG_SMP case.
1816 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1817 + rq->most_recent_timestamp;
1818 __activate_task(p, rq);
1819 if (TASK_PREEMPTS_CURR(p, rq))
1820 resched_task(rq->curr);
1823 * Parent and child are on different CPUs, now get the
1824 * parent runqueue to update the parent's ->sleep_avg:
1826 task_rq_unlock(rq, &flags);
1827 this_rq = task_rq_lock(current, &flags);
1829 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1830 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1831 task_rq_unlock(this_rq, &flags);
1835 * prepare_task_switch - prepare to switch tasks
1836 * @rq: the runqueue preparing to switch
1837 * @next: the task we are going to switch to.
1839 * This is called with the rq lock held and interrupts off. It must
1840 * be paired with a subsequent finish_task_switch after the context
1843 * prepare_task_switch sets up locking and calls architecture specific
1846 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1848 prepare_lock_switch(rq, next);
1849 prepare_arch_switch(next);
1853 * finish_task_switch - clean up after a task-switch
1854 * @rq: runqueue associated with task-switch
1855 * @prev: the thread we just switched away from.
1857 * finish_task_switch must be called after the context switch, paired
1858 * with a prepare_task_switch call before the context switch.
1859 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1860 * and do any other architecture-specific cleanup actions.
1862 * Note that we may have delayed dropping an mm in context_switch(). If
1863 * so, we finish that here outside of the runqueue lock. (Doing it
1864 * with the lock held can cause deadlocks; see schedule() for
1867 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1868 __releases(rq->lock)
1870 struct mm_struct *mm = rq->prev_mm;
1876 * A task struct has one reference for the use as "current".
1877 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1878 * schedule one last time. The schedule call will never return, and
1879 * the scheduled task must drop that reference.
1880 * The test for TASK_DEAD must occur while the runqueue locks are
1881 * still held, otherwise prev could be scheduled on another cpu, die
1882 * there before we look at prev->state, and then the reference would
1884 * Manfred Spraul <manfred@colorfullife.com>
1886 prev_state = prev->state;
1887 finish_arch_switch(prev);
1888 finish_lock_switch(rq, prev);
1891 if (unlikely(prev_state == TASK_DEAD)) {
1893 * Remove function-return probe instances associated with this
1894 * task and put them back on the free list.
1896 kprobe_flush_task(prev);
1897 put_task_struct(prev);
1902 * schedule_tail - first thing a freshly forked thread must call.
1903 * @prev: the thread we just switched away from.
1905 asmlinkage void schedule_tail(struct task_struct *prev)
1906 __releases(rq->lock)
1908 struct rq *rq = this_rq();
1910 finish_task_switch(rq, prev);
1911 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1912 /* In this case, finish_task_switch does not reenable preemption */
1915 if (current->set_child_tid)
1916 put_user(current->pid, current->set_child_tid);
1920 * context_switch - switch to the new MM and the new
1921 * thread's register state.
1923 static inline struct task_struct *
1924 context_switch(struct rq *rq, struct task_struct *prev,
1925 struct task_struct *next)
1927 struct mm_struct *mm = next->mm;
1928 struct mm_struct *oldmm = prev->active_mm;
1931 * For paravirt, this is coupled with an exit in switch_to to
1932 * combine the page table reload and the switch backend into
1935 arch_enter_lazy_cpu_mode();
1938 next->active_mm = oldmm;
1939 atomic_inc(&oldmm->mm_count);
1940 enter_lazy_tlb(oldmm, next);
1942 switch_mm(oldmm, mm, next);
1945 prev->active_mm = NULL;
1946 WARN_ON(rq->prev_mm);
1947 rq->prev_mm = oldmm;
1950 * Since the runqueue lock will be released by the next
1951 * task (which is an invalid locking op but in the case
1952 * of the scheduler it's an obvious special-case), so we
1953 * do an early lockdep release here:
1955 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1956 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1959 /* Here we just switch the register state and the stack. */
1960 switch_to(prev, next, prev);
1966 * nr_running, nr_uninterruptible and nr_context_switches:
1968 * externally visible scheduler statistics: current number of runnable
1969 * threads, current number of uninterruptible-sleeping threads, total
1970 * number of context switches performed since bootup.
1972 unsigned long nr_running(void)
1974 unsigned long i, sum = 0;
1976 for_each_online_cpu(i)
1977 sum += cpu_rq(i)->nr_running;
1982 unsigned long nr_uninterruptible(void)
1984 unsigned long i, sum = 0;
1986 for_each_possible_cpu(i)
1987 sum += cpu_rq(i)->nr_uninterruptible;
1990 * Since we read the counters lockless, it might be slightly
1991 * inaccurate. Do not allow it to go below zero though:
1993 if (unlikely((long)sum < 0))
1999 unsigned long long nr_context_switches(void)
2002 unsigned long long sum = 0;
2004 for_each_possible_cpu(i)
2005 sum += cpu_rq(i)->nr_switches;
2010 unsigned long nr_iowait(void)
2012 unsigned long i, sum = 0;
2014 for_each_possible_cpu(i)
2015 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2020 unsigned long nr_active(void)
2022 unsigned long i, running = 0, uninterruptible = 0;
2024 for_each_online_cpu(i) {
2025 running += cpu_rq(i)->nr_running;
2026 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2029 if (unlikely((long)uninterruptible < 0))
2030 uninterruptible = 0;
2032 return running + uninterruptible;
2038 * Is this task likely cache-hot:
2041 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
2043 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
2047 * double_rq_lock - safely lock two runqueues
2049 * Note this does not disable interrupts like task_rq_lock,
2050 * you need to do so manually before calling.
2052 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2053 __acquires(rq1->lock)
2054 __acquires(rq2->lock)
2056 BUG_ON(!irqs_disabled());
2058 spin_lock(&rq1->lock);
2059 __acquire(rq2->lock); /* Fake it out ;) */
2062 spin_lock(&rq1->lock);
2063 spin_lock(&rq2->lock);
2065 spin_lock(&rq2->lock);
2066 spin_lock(&rq1->lock);
2072 * double_rq_unlock - safely unlock two runqueues
2074 * Note this does not restore interrupts like task_rq_unlock,
2075 * you need to do so manually after calling.
2077 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2078 __releases(rq1->lock)
2079 __releases(rq2->lock)
2081 spin_unlock(&rq1->lock);
2083 spin_unlock(&rq2->lock);
2085 __release(rq2->lock);
2089 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2091 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2092 __releases(this_rq->lock)
2093 __acquires(busiest->lock)
2094 __acquires(this_rq->lock)
2096 if (unlikely(!irqs_disabled())) {
2097 /* printk() doesn't work good under rq->lock */
2098 spin_unlock(&this_rq->lock);
2101 if (unlikely(!spin_trylock(&busiest->lock))) {
2102 if (busiest < this_rq) {
2103 spin_unlock(&this_rq->lock);
2104 spin_lock(&busiest->lock);
2105 spin_lock(&this_rq->lock);
2107 spin_lock(&busiest->lock);
2112 * If dest_cpu is allowed for this process, migrate the task to it.
2113 * This is accomplished by forcing the cpu_allowed mask to only
2114 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2115 * the cpu_allowed mask is restored.
2117 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2119 struct migration_req req;
2120 unsigned long flags;
2123 rq = task_rq_lock(p, &flags);
2124 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2125 || unlikely(cpu_is_offline(dest_cpu)))
2128 /* force the process onto the specified CPU */
2129 if (migrate_task(p, dest_cpu, &req)) {
2130 /* Need to wait for migration thread (might exit: take ref). */
2131 struct task_struct *mt = rq->migration_thread;
2133 get_task_struct(mt);
2134 task_rq_unlock(rq, &flags);
2135 wake_up_process(mt);
2136 put_task_struct(mt);
2137 wait_for_completion(&req.done);
2142 task_rq_unlock(rq, &flags);
2146 * sched_exec - execve() is a valuable balancing opportunity, because at
2147 * this point the task has the smallest effective memory and cache footprint.
2149 void sched_exec(void)
2151 int new_cpu, this_cpu = get_cpu();
2152 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2154 if (new_cpu != this_cpu)
2155 sched_migrate_task(current, new_cpu);
2159 * pull_task - move a task from a remote runqueue to the local runqueue.
2160 * Both runqueues must be locked.
2162 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2163 struct task_struct *p, struct rq *this_rq,
2164 struct prio_array *this_array, int this_cpu)
2166 dequeue_task(p, src_array);
2167 dec_nr_running(p, src_rq);
2168 set_task_cpu(p, this_cpu);
2169 inc_nr_running(p, this_rq);
2170 enqueue_task(p, this_array);
2171 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2172 + this_rq->most_recent_timestamp;
2174 * Note that idle threads have a prio of MAX_PRIO, for this test
2175 * to be always true for them.
2177 if (TASK_PREEMPTS_CURR(p, this_rq))
2178 resched_task(this_rq->curr);
2182 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2185 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2186 struct sched_domain *sd, enum cpu_idle_type idle,
2190 * We do not migrate tasks that are:
2191 * 1) running (obviously), or
2192 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2193 * 3) are cache-hot on their current CPU.
2195 if (!cpu_isset(this_cpu, p->cpus_allowed))
2199 if (task_running(rq, p))
2203 * Aggressive migration if:
2204 * 1) task is cache cold, or
2205 * 2) too many balance attempts have failed.
2208 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2209 #ifdef CONFIG_SCHEDSTATS
2210 if (task_hot(p, rq->most_recent_timestamp, sd))
2211 schedstat_inc(sd, lb_hot_gained[idle]);
2216 if (task_hot(p, rq->most_recent_timestamp, sd))
2221 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2224 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2225 * load from busiest to this_rq, as part of a balancing operation within
2226 * "domain". Returns the number of tasks moved.
2228 * Called with both runqueues locked.
2230 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2231 unsigned long max_nr_move, unsigned long max_load_move,
2232 struct sched_domain *sd, enum cpu_idle_type idle,
2235 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2236 best_prio_seen, skip_for_load;
2237 struct prio_array *array, *dst_array;
2238 struct list_head *head, *curr;
2239 struct task_struct *tmp;
2242 if (max_nr_move == 0 || max_load_move == 0)
2245 rem_load_move = max_load_move;
2247 this_best_prio = rq_best_prio(this_rq);
2248 best_prio = rq_best_prio(busiest);
2250 * Enable handling of the case where there is more than one task
2251 * with the best priority. If the current running task is one
2252 * of those with prio==best_prio we know it won't be moved
2253 * and therefore it's safe to override the skip (based on load) of
2254 * any task we find with that prio.
2256 best_prio_seen = best_prio == busiest->curr->prio;
2259 * We first consider expired tasks. Those will likely not be
2260 * executed in the near future, and they are most likely to
2261 * be cache-cold, thus switching CPUs has the least effect
2264 if (busiest->expired->nr_active) {
2265 array = busiest->expired;
2266 dst_array = this_rq->expired;
2268 array = busiest->active;
2269 dst_array = this_rq->active;
2273 /* Start searching at priority 0: */
2277 idx = sched_find_first_bit(array->bitmap);
2279 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2280 if (idx >= MAX_PRIO) {
2281 if (array == busiest->expired && busiest->active->nr_active) {
2282 array = busiest->active;
2283 dst_array = this_rq->active;
2289 head = array->queue + idx;
2292 tmp = list_entry(curr, struct task_struct, run_list);
2297 * To help distribute high priority tasks accross CPUs we don't
2298 * skip a task if it will be the highest priority task (i.e. smallest
2299 * prio value) on its new queue regardless of its load weight
2301 skip_for_load = tmp->load_weight > rem_load_move;
2302 if (skip_for_load && idx < this_best_prio)
2303 skip_for_load = !best_prio_seen && idx == best_prio;
2304 if (skip_for_load ||
2305 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2307 best_prio_seen |= idx == best_prio;
2314 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2316 rem_load_move -= tmp->load_weight;
2319 * We only want to steal up to the prescribed number of tasks
2320 * and the prescribed amount of weighted load.
2322 if (pulled < max_nr_move && rem_load_move > 0) {
2323 if (idx < this_best_prio)
2324 this_best_prio = idx;
2332 * Right now, this is the only place pull_task() is called,
2333 * so we can safely collect pull_task() stats here rather than
2334 * inside pull_task().
2336 schedstat_add(sd, lb_gained[idle], pulled);
2339 *all_pinned = pinned;
2344 * find_busiest_group finds and returns the busiest CPU group within the
2345 * domain. It calculates and returns the amount of weighted load which
2346 * should be moved to restore balance via the imbalance parameter.
2348 static struct sched_group *
2349 find_busiest_group(struct sched_domain *sd, int this_cpu,
2350 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2351 cpumask_t *cpus, int *balance)
2353 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2354 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2355 unsigned long max_pull;
2356 unsigned long busiest_load_per_task, busiest_nr_running;
2357 unsigned long this_load_per_task, this_nr_running;
2359 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2360 int power_savings_balance = 1;
2361 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2362 unsigned long min_nr_running = ULONG_MAX;
2363 struct sched_group *group_min = NULL, *group_leader = NULL;
2366 max_load = this_load = total_load = total_pwr = 0;
2367 busiest_load_per_task = busiest_nr_running = 0;
2368 this_load_per_task = this_nr_running = 0;
2369 if (idle == CPU_NOT_IDLE)
2370 load_idx = sd->busy_idx;
2371 else if (idle == CPU_NEWLY_IDLE)
2372 load_idx = sd->newidle_idx;
2374 load_idx = sd->idle_idx;
2377 unsigned long load, group_capacity;
2380 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2381 unsigned long sum_nr_running, sum_weighted_load;
2383 local_group = cpu_isset(this_cpu, group->cpumask);
2386 balance_cpu = first_cpu(group->cpumask);
2388 /* Tally up the load of all CPUs in the group */
2389 sum_weighted_load = sum_nr_running = avg_load = 0;
2391 for_each_cpu_mask(i, group->cpumask) {
2394 if (!cpu_isset(i, *cpus))
2399 if (*sd_idle && !idle_cpu(i))
2402 /* Bias balancing toward cpus of our domain */
2404 if (idle_cpu(i) && !first_idle_cpu) {
2409 load = target_load(i, load_idx);
2411 load = source_load(i, load_idx);
2414 sum_nr_running += rq->nr_running;
2415 sum_weighted_load += rq->raw_weighted_load;
2419 * First idle cpu or the first cpu(busiest) in this sched group
2420 * is eligible for doing load balancing at this and above
2423 if (local_group && balance_cpu != this_cpu && balance) {
2428 total_load += avg_load;
2429 total_pwr += group->__cpu_power;
2431 /* Adjust by relative CPU power of the group */
2432 avg_load = sg_div_cpu_power(group,
2433 avg_load * SCHED_LOAD_SCALE);
2435 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2438 this_load = avg_load;
2440 this_nr_running = sum_nr_running;
2441 this_load_per_task = sum_weighted_load;
2442 } else if (avg_load > max_load &&
2443 sum_nr_running > group_capacity) {
2444 max_load = avg_load;
2446 busiest_nr_running = sum_nr_running;
2447 busiest_load_per_task = sum_weighted_load;
2450 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2452 * Busy processors will not participate in power savings
2455 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2459 * If the local group is idle or completely loaded
2460 * no need to do power savings balance at this domain
2462 if (local_group && (this_nr_running >= group_capacity ||
2464 power_savings_balance = 0;
2467 * If a group is already running at full capacity or idle,
2468 * don't include that group in power savings calculations
2470 if (!power_savings_balance || sum_nr_running >= group_capacity
2475 * Calculate the group which has the least non-idle load.
2476 * This is the group from where we need to pick up the load
2479 if ((sum_nr_running < min_nr_running) ||
2480 (sum_nr_running == min_nr_running &&
2481 first_cpu(group->cpumask) <
2482 first_cpu(group_min->cpumask))) {
2484 min_nr_running = sum_nr_running;
2485 min_load_per_task = sum_weighted_load /
2490 * Calculate the group which is almost near its
2491 * capacity but still has some space to pick up some load
2492 * from other group and save more power
2494 if (sum_nr_running <= group_capacity - 1) {
2495 if (sum_nr_running > leader_nr_running ||
2496 (sum_nr_running == leader_nr_running &&
2497 first_cpu(group->cpumask) >
2498 first_cpu(group_leader->cpumask))) {
2499 group_leader = group;
2500 leader_nr_running = sum_nr_running;
2505 group = group->next;
2506 } while (group != sd->groups);
2508 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2511 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2513 if (this_load >= avg_load ||
2514 100*max_load <= sd->imbalance_pct*this_load)
2517 busiest_load_per_task /= busiest_nr_running;
2519 * We're trying to get all the cpus to the average_load, so we don't
2520 * want to push ourselves above the average load, nor do we wish to
2521 * reduce the max loaded cpu below the average load, as either of these
2522 * actions would just result in more rebalancing later, and ping-pong
2523 * tasks around. Thus we look for the minimum possible imbalance.
2524 * Negative imbalances (*we* are more loaded than anyone else) will
2525 * be counted as no imbalance for these purposes -- we can't fix that
2526 * by pulling tasks to us. Be careful of negative numbers as they'll
2527 * appear as very large values with unsigned longs.
2529 if (max_load <= busiest_load_per_task)
2533 * In the presence of smp nice balancing, certain scenarios can have
2534 * max load less than avg load(as we skip the groups at or below
2535 * its cpu_power, while calculating max_load..)
2537 if (max_load < avg_load) {
2539 goto small_imbalance;
2542 /* Don't want to pull so many tasks that a group would go idle */
2543 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2545 /* How much load to actually move to equalise the imbalance */
2546 *imbalance = min(max_pull * busiest->__cpu_power,
2547 (avg_load - this_load) * this->__cpu_power)
2551 * if *imbalance is less than the average load per runnable task
2552 * there is no gaurantee that any tasks will be moved so we'll have
2553 * a think about bumping its value to force at least one task to be
2556 if (*imbalance < busiest_load_per_task) {
2557 unsigned long tmp, pwr_now, pwr_move;
2561 pwr_move = pwr_now = 0;
2563 if (this_nr_running) {
2564 this_load_per_task /= this_nr_running;
2565 if (busiest_load_per_task > this_load_per_task)
2568 this_load_per_task = SCHED_LOAD_SCALE;
2570 if (max_load - this_load >= busiest_load_per_task * imbn) {
2571 *imbalance = busiest_load_per_task;
2576 * OK, we don't have enough imbalance to justify moving tasks,
2577 * however we may be able to increase total CPU power used by
2581 pwr_now += busiest->__cpu_power *
2582 min(busiest_load_per_task, max_load);
2583 pwr_now += this->__cpu_power *
2584 min(this_load_per_task, this_load);
2585 pwr_now /= SCHED_LOAD_SCALE;
2587 /* Amount of load we'd subtract */
2588 tmp = sg_div_cpu_power(busiest,
2589 busiest_load_per_task * SCHED_LOAD_SCALE);
2591 pwr_move += busiest->__cpu_power *
2592 min(busiest_load_per_task, max_load - tmp);
2594 /* Amount of load we'd add */
2595 if (max_load * busiest->__cpu_power <
2596 busiest_load_per_task * SCHED_LOAD_SCALE)
2597 tmp = sg_div_cpu_power(this,
2598 max_load * busiest->__cpu_power);
2600 tmp = sg_div_cpu_power(this,
2601 busiest_load_per_task * SCHED_LOAD_SCALE);
2602 pwr_move += this->__cpu_power *
2603 min(this_load_per_task, this_load + tmp);
2604 pwr_move /= SCHED_LOAD_SCALE;
2606 /* Move if we gain throughput */
2607 if (pwr_move <= pwr_now)
2610 *imbalance = busiest_load_per_task;
2616 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2617 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2620 if (this == group_leader && group_leader != group_min) {
2621 *imbalance = min_load_per_task;
2631 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2634 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2635 unsigned long imbalance, cpumask_t *cpus)
2637 struct rq *busiest = NULL, *rq;
2638 unsigned long max_load = 0;
2641 for_each_cpu_mask(i, group->cpumask) {
2643 if (!cpu_isset(i, *cpus))
2648 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2651 if (rq->raw_weighted_load > max_load) {
2652 max_load = rq->raw_weighted_load;
2661 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2662 * so long as it is large enough.
2664 #define MAX_PINNED_INTERVAL 512
2666 static inline unsigned long minus_1_or_zero(unsigned long n)
2668 return n > 0 ? n - 1 : 0;
2672 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2673 * tasks if there is an imbalance.
2675 static int load_balance(int this_cpu, struct rq *this_rq,
2676 struct sched_domain *sd, enum cpu_idle_type idle,
2679 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2680 struct sched_group *group;
2681 unsigned long imbalance;
2683 cpumask_t cpus = CPU_MASK_ALL;
2684 unsigned long flags;
2687 * When power savings policy is enabled for the parent domain, idle
2688 * sibling can pick up load irrespective of busy siblings. In this case,
2689 * let the state of idle sibling percolate up as IDLE, instead of
2690 * portraying it as CPU_NOT_IDLE.
2692 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2693 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2696 schedstat_inc(sd, lb_cnt[idle]);
2699 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2706 schedstat_inc(sd, lb_nobusyg[idle]);
2710 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2712 schedstat_inc(sd, lb_nobusyq[idle]);
2716 BUG_ON(busiest == this_rq);
2718 schedstat_add(sd, lb_imbalance[idle], imbalance);
2721 if (busiest->nr_running > 1) {
2723 * Attempt to move tasks. If find_busiest_group has found
2724 * an imbalance but busiest->nr_running <= 1, the group is
2725 * still unbalanced. nr_moved simply stays zero, so it is
2726 * correctly treated as an imbalance.
2728 local_irq_save(flags);
2729 double_rq_lock(this_rq, busiest);
2730 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2731 minus_1_or_zero(busiest->nr_running),
2732 imbalance, sd, idle, &all_pinned);
2733 double_rq_unlock(this_rq, busiest);
2734 local_irq_restore(flags);
2737 * some other cpu did the load balance for us.
2739 if (nr_moved && this_cpu != smp_processor_id())
2740 resched_cpu(this_cpu);
2742 /* All tasks on this runqueue were pinned by CPU affinity */
2743 if (unlikely(all_pinned)) {
2744 cpu_clear(cpu_of(busiest), cpus);
2745 if (!cpus_empty(cpus))
2752 schedstat_inc(sd, lb_failed[idle]);
2753 sd->nr_balance_failed++;
2755 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2757 spin_lock_irqsave(&busiest->lock, flags);
2759 /* don't kick the migration_thread, if the curr
2760 * task on busiest cpu can't be moved to this_cpu
2762 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2763 spin_unlock_irqrestore(&busiest->lock, flags);
2765 goto out_one_pinned;
2768 if (!busiest->active_balance) {
2769 busiest->active_balance = 1;
2770 busiest->push_cpu = this_cpu;
2773 spin_unlock_irqrestore(&busiest->lock, flags);
2775 wake_up_process(busiest->migration_thread);
2778 * We've kicked active balancing, reset the failure
2781 sd->nr_balance_failed = sd->cache_nice_tries+1;
2784 sd->nr_balance_failed = 0;
2786 if (likely(!active_balance)) {
2787 /* We were unbalanced, so reset the balancing interval */
2788 sd->balance_interval = sd->min_interval;
2791 * If we've begun active balancing, start to back off. This
2792 * case may not be covered by the all_pinned logic if there
2793 * is only 1 task on the busy runqueue (because we don't call
2796 if (sd->balance_interval < sd->max_interval)
2797 sd->balance_interval *= 2;
2800 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2801 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2806 schedstat_inc(sd, lb_balanced[idle]);
2808 sd->nr_balance_failed = 0;
2811 /* tune up the balancing interval */
2812 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2813 (sd->balance_interval < sd->max_interval))
2814 sd->balance_interval *= 2;
2816 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2817 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2823 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2824 * tasks if there is an imbalance.
2826 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2827 * this_rq is locked.
2830 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2832 struct sched_group *group;
2833 struct rq *busiest = NULL;
2834 unsigned long imbalance;
2837 cpumask_t cpus = CPU_MASK_ALL;
2840 * When power savings policy is enabled for the parent domain, idle
2841 * sibling can pick up load irrespective of busy siblings. In this case,
2842 * let the state of idle sibling percolate up as IDLE, instead of
2843 * portraying it as CPU_NOT_IDLE.
2845 if (sd->flags & SD_SHARE_CPUPOWER &&
2846 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2849 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2851 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2852 &sd_idle, &cpus, NULL);
2854 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2858 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2861 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2865 BUG_ON(busiest == this_rq);
2867 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2870 if (busiest->nr_running > 1) {
2871 /* Attempt to move tasks */
2872 double_lock_balance(this_rq, busiest);
2873 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2874 minus_1_or_zero(busiest->nr_running),
2875 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2876 spin_unlock(&busiest->lock);
2879 cpu_clear(cpu_of(busiest), cpus);
2880 if (!cpus_empty(cpus))
2886 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2887 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2888 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2891 sd->nr_balance_failed = 0;
2896 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2897 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2898 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2900 sd->nr_balance_failed = 0;
2906 * idle_balance is called by schedule() if this_cpu is about to become
2907 * idle. Attempts to pull tasks from other CPUs.
2909 static void idle_balance(int this_cpu, struct rq *this_rq)
2911 struct sched_domain *sd;
2912 int pulled_task = 0;
2913 unsigned long next_balance = jiffies + 60 * HZ;
2915 for_each_domain(this_cpu, sd) {
2916 unsigned long interval;
2918 if (!(sd->flags & SD_LOAD_BALANCE))
2921 if (sd->flags & SD_BALANCE_NEWIDLE)
2922 /* If we've pulled tasks over stop searching: */
2923 pulled_task = load_balance_newidle(this_cpu,
2926 interval = msecs_to_jiffies(sd->balance_interval);
2927 if (time_after(next_balance, sd->last_balance + interval))
2928 next_balance = sd->last_balance + interval;
2934 * We are going idle. next_balance may be set based on
2935 * a busy processor. So reset next_balance.
2937 this_rq->next_balance = next_balance;
2941 * active_load_balance is run by migration threads. It pushes running tasks
2942 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2943 * running on each physical CPU where possible, and avoids physical /
2944 * logical imbalances.
2946 * Called with busiest_rq locked.
2948 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2950 int target_cpu = busiest_rq->push_cpu;
2951 struct sched_domain *sd;
2952 struct rq *target_rq;
2954 /* Is there any task to move? */
2955 if (busiest_rq->nr_running <= 1)
2958 target_rq = cpu_rq(target_cpu);
2961 * This condition is "impossible", if it occurs
2962 * we need to fix it. Originally reported by
2963 * Bjorn Helgaas on a 128-cpu setup.
2965 BUG_ON(busiest_rq == target_rq);
2967 /* move a task from busiest_rq to target_rq */
2968 double_lock_balance(busiest_rq, target_rq);
2970 /* Search for an sd spanning us and the target CPU. */
2971 for_each_domain(target_cpu, sd) {
2972 if ((sd->flags & SD_LOAD_BALANCE) &&
2973 cpu_isset(busiest_cpu, sd->span))
2978 schedstat_inc(sd, alb_cnt);
2980 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2981 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2983 schedstat_inc(sd, alb_pushed);
2985 schedstat_inc(sd, alb_failed);
2987 spin_unlock(&target_rq->lock);
2990 static void update_load(struct rq *this_rq)
2992 unsigned long this_load;
2993 unsigned int i, scale;
2995 this_load = this_rq->raw_weighted_load;
2997 /* Update our load: */
2998 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2999 unsigned long old_load, new_load;
3001 /* scale is effectively 1 << i now, and >> i divides by scale */
3003 old_load = this_rq->cpu_load[i];
3004 new_load = this_load;
3006 * Round up the averaging division if load is increasing. This
3007 * prevents us from getting stuck on 9 if the load is 10, for
3010 if (new_load > old_load)
3011 new_load += scale-1;
3012 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3018 atomic_t load_balancer;
3020 } nohz ____cacheline_aligned = {
3021 .load_balancer = ATOMIC_INIT(-1),
3022 .cpu_mask = CPU_MASK_NONE,
3026 * This routine will try to nominate the ilb (idle load balancing)
3027 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3028 * load balancing on behalf of all those cpus. If all the cpus in the system
3029 * go into this tickless mode, then there will be no ilb owner (as there is
3030 * no need for one) and all the cpus will sleep till the next wakeup event
3033 * For the ilb owner, tick is not stopped. And this tick will be used
3034 * for idle load balancing. ilb owner will still be part of
3037 * While stopping the tick, this cpu will become the ilb owner if there
3038 * is no other owner. And will be the owner till that cpu becomes busy
3039 * or if all cpus in the system stop their ticks at which point
3040 * there is no need for ilb owner.
3042 * When the ilb owner becomes busy, it nominates another owner, during the
3043 * next busy scheduler_tick()
3045 int select_nohz_load_balancer(int stop_tick)
3047 int cpu = smp_processor_id();
3050 cpu_set(cpu, nohz.cpu_mask);
3051 cpu_rq(cpu)->in_nohz_recently = 1;
3054 * If we are going offline and still the leader, give up!
3056 if (cpu_is_offline(cpu) &&
3057 atomic_read(&nohz.load_balancer) == cpu) {
3058 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3063 /* time for ilb owner also to sleep */
3064 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3065 if (atomic_read(&nohz.load_balancer) == cpu)
3066 atomic_set(&nohz.load_balancer, -1);
3070 if (atomic_read(&nohz.load_balancer) == -1) {
3071 /* make me the ilb owner */
3072 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3074 } else if (atomic_read(&nohz.load_balancer) == cpu)
3077 if (!cpu_isset(cpu, nohz.cpu_mask))
3080 cpu_clear(cpu, nohz.cpu_mask);
3082 if (atomic_read(&nohz.load_balancer) == cpu)
3083 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3090 static DEFINE_SPINLOCK(balancing);
3093 * It checks each scheduling domain to see if it is due to be balanced,
3094 * and initiates a balancing operation if so.
3096 * Balancing parameters are set up in arch_init_sched_domains.
3098 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3101 struct rq *rq = cpu_rq(cpu);
3102 unsigned long interval;
3103 struct sched_domain *sd;
3104 /* Earliest time when we have to do rebalance again */
3105 unsigned long next_balance = jiffies + 60*HZ;
3107 for_each_domain(cpu, sd) {
3108 if (!(sd->flags & SD_LOAD_BALANCE))
3111 interval = sd->balance_interval;
3112 if (idle != CPU_IDLE)
3113 interval *= sd->busy_factor;
3115 /* scale ms to jiffies */
3116 interval = msecs_to_jiffies(interval);
3117 if (unlikely(!interval))
3120 if (sd->flags & SD_SERIALIZE) {
3121 if (!spin_trylock(&balancing))
3125 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3126 if (load_balance(cpu, rq, sd, idle, &balance)) {
3128 * We've pulled tasks over so either we're no
3129 * longer idle, or one of our SMT siblings is
3132 idle = CPU_NOT_IDLE;
3134 sd->last_balance = jiffies;
3136 if (sd->flags & SD_SERIALIZE)
3137 spin_unlock(&balancing);
3139 if (time_after(next_balance, sd->last_balance + interval))
3140 next_balance = sd->last_balance + interval;
3143 * Stop the load balance at this level. There is another
3144 * CPU in our sched group which is doing load balancing more
3150 rq->next_balance = next_balance;
3154 * run_rebalance_domains is triggered when needed from the scheduler tick.
3155 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3156 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3158 static void run_rebalance_domains(struct softirq_action *h)
3160 int local_cpu = smp_processor_id();
3161 struct rq *local_rq = cpu_rq(local_cpu);
3162 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3164 rebalance_domains(local_cpu, idle);
3168 * If this cpu is the owner for idle load balancing, then do the
3169 * balancing on behalf of the other idle cpus whose ticks are
3172 if (local_rq->idle_at_tick &&
3173 atomic_read(&nohz.load_balancer) == local_cpu) {
3174 cpumask_t cpus = nohz.cpu_mask;
3178 cpu_clear(local_cpu, cpus);
3179 for_each_cpu_mask(balance_cpu, cpus) {
3181 * If this cpu gets work to do, stop the load balancing
3182 * work being done for other cpus. Next load
3183 * balancing owner will pick it up.
3188 rebalance_domains(balance_cpu, CPU_IDLE);
3190 rq = cpu_rq(balance_cpu);
3191 if (time_after(local_rq->next_balance, rq->next_balance))
3192 local_rq->next_balance = rq->next_balance;
3199 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3201 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3202 * idle load balancing owner or decide to stop the periodic load balancing,
3203 * if the whole system is idle.
3205 static inline void trigger_load_balance(int cpu)
3207 struct rq *rq = cpu_rq(cpu);
3210 * If we were in the nohz mode recently and busy at the current
3211 * scheduler tick, then check if we need to nominate new idle
3214 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3215 rq->in_nohz_recently = 0;
3217 if (atomic_read(&nohz.load_balancer) == cpu) {
3218 cpu_clear(cpu, nohz.cpu_mask);
3219 atomic_set(&nohz.load_balancer, -1);
3222 if (atomic_read(&nohz.load_balancer) == -1) {
3224 * simple selection for now: Nominate the
3225 * first cpu in the nohz list to be the next
3228 * TBD: Traverse the sched domains and nominate
3229 * the nearest cpu in the nohz.cpu_mask.
3231 int ilb = first_cpu(nohz.cpu_mask);
3239 * If this cpu is idle and doing idle load balancing for all the
3240 * cpus with ticks stopped, is it time for that to stop?
3242 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3243 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3249 * If this cpu is idle and the idle load balancing is done by
3250 * someone else, then no need raise the SCHED_SOFTIRQ
3252 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3253 cpu_isset(cpu, nohz.cpu_mask))
3256 if (time_after_eq(jiffies, rq->next_balance))
3257 raise_softirq(SCHED_SOFTIRQ);
3261 * on UP we do not need to balance between CPUs:
3263 static inline void idle_balance(int cpu, struct rq *rq)
3268 DEFINE_PER_CPU(struct kernel_stat, kstat);
3270 EXPORT_PER_CPU_SYMBOL(kstat);
3273 * This is called on clock ticks and on context switches.
3274 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3277 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3279 p->sched_time += now - p->last_ran;
3280 p->last_ran = rq->most_recent_timestamp = now;
3284 * Return current->sched_time plus any more ns on the sched_clock
3285 * that have not yet been banked.
3287 unsigned long long current_sched_time(const struct task_struct *p)
3289 unsigned long long ns;
3290 unsigned long flags;
3292 local_irq_save(flags);
3293 ns = p->sched_time + sched_clock() - p->last_ran;
3294 local_irq_restore(flags);
3300 * We place interactive tasks back into the active array, if possible.
3302 * To guarantee that this does not starve expired tasks we ignore the
3303 * interactivity of a task if the first expired task had to wait more
3304 * than a 'reasonable' amount of time. This deadline timeout is
3305 * load-dependent, as the frequency of array switched decreases with
3306 * increasing number of running tasks. We also ignore the interactivity
3307 * if a better static_prio task has expired:
3309 static inline int expired_starving(struct rq *rq)
3311 if (rq->curr->static_prio > rq->best_expired_prio)
3313 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3315 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3321 * Account user cpu time to a process.
3322 * @p: the process that the cpu time gets accounted to
3323 * @hardirq_offset: the offset to subtract from hardirq_count()
3324 * @cputime: the cpu time spent in user space since the last update
3326 void account_user_time(struct task_struct *p, cputime_t cputime)
3328 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3331 p->utime = cputime_add(p->utime, cputime);
3333 /* Add user time to cpustat. */
3334 tmp = cputime_to_cputime64(cputime);
3335 if (TASK_NICE(p) > 0)
3336 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3338 cpustat->user = cputime64_add(cpustat->user, tmp);
3342 * Account system cpu time to a process.
3343 * @p: the process that the cpu time gets accounted to
3344 * @hardirq_offset: the offset to subtract from hardirq_count()
3345 * @cputime: the cpu time spent in kernel space since the last update
3347 void account_system_time(struct task_struct *p, int hardirq_offset,
3350 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3351 struct rq *rq = this_rq();
3354 p->stime = cputime_add(p->stime, cputime);
3356 /* Add system time to cpustat. */
3357 tmp = cputime_to_cputime64(cputime);
3358 if (hardirq_count() - hardirq_offset)
3359 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3360 else if (softirq_count())
3361 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3362 else if (p != rq->idle)
3363 cpustat->system = cputime64_add(cpustat->system, tmp);
3364 else if (atomic_read(&rq->nr_iowait) > 0)
3365 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3367 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3368 /* Account for system time used */
3369 acct_update_integrals(p);
3373 * Account for involuntary wait time.
3374 * @p: the process from which the cpu time has been stolen
3375 * @steal: the cpu time spent in involuntary wait
3377 void account_steal_time(struct task_struct *p, cputime_t steal)
3379 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3380 cputime64_t tmp = cputime_to_cputime64(steal);
3381 struct rq *rq = this_rq();
3383 if (p == rq->idle) {
3384 p->stime = cputime_add(p->stime, steal);
3385 if (atomic_read(&rq->nr_iowait) > 0)
3386 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3388 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3390 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3393 static void task_running_tick(struct rq *rq, struct task_struct *p)
3395 if (p->array != rq->active) {
3396 /* Task has expired but was not scheduled yet */
3397 set_tsk_need_resched(p);
3400 spin_lock(&rq->lock);
3402 * The task was running during this tick - update the
3403 * time slice counter. Note: we do not update a thread's
3404 * priority until it either goes to sleep or uses up its
3405 * timeslice. This makes it possible for interactive tasks
3406 * to use up their timeslices at their highest priority levels.
3410 * RR tasks need a special form of timeslice management.
3411 * FIFO tasks have no timeslices.
3413 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3414 p->time_slice = task_timeslice(p);
3415 p->first_time_slice = 0;
3416 set_tsk_need_resched(p);
3418 /* put it at the end of the queue: */
3419 requeue_task(p, rq->active);
3423 if (!--p->time_slice) {
3424 dequeue_task(p, rq->active);
3425 set_tsk_need_resched(p);
3426 p->prio = effective_prio(p);
3427 p->time_slice = task_timeslice(p);
3428 p->first_time_slice = 0;
3430 if (!rq->expired_timestamp)
3431 rq->expired_timestamp = jiffies;
3432 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3433 enqueue_task(p, rq->expired);
3434 if (p->static_prio < rq->best_expired_prio)
3435 rq->best_expired_prio = p->static_prio;
3437 enqueue_task(p, rq->active);
3440 * Prevent a too long timeslice allowing a task to monopolize
3441 * the CPU. We do this by splitting up the timeslice into
3444 * Note: this does not mean the task's timeslices expire or
3445 * get lost in any way, they just might be preempted by
3446 * another task of equal priority. (one with higher
3447 * priority would have preempted this task already.) We
3448 * requeue this task to the end of the list on this priority
3449 * level, which is in essence a round-robin of tasks with
3452 * This only applies to tasks in the interactive
3453 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3455 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3456 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3457 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3458 (p->array == rq->active)) {
3460 requeue_task(p, rq->active);
3461 set_tsk_need_resched(p);
3465 spin_unlock(&rq->lock);
3469 * This function gets called by the timer code, with HZ frequency.
3470 * We call it with interrupts disabled.
3472 * It also gets called by the fork code, when changing the parent's
3475 void scheduler_tick(void)
3477 unsigned long long now = sched_clock();
3478 struct task_struct *p = current;
3479 int cpu = smp_processor_id();
3480 int idle_at_tick = idle_cpu(cpu);
3481 struct rq *rq = cpu_rq(cpu);
3483 update_cpu_clock(p, rq, now);
3486 task_running_tick(rq, p);
3489 rq->idle_at_tick = idle_at_tick;
3490 trigger_load_balance(cpu);
3494 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3496 void fastcall add_preempt_count(int val)
3501 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3503 preempt_count() += val;
3505 * Spinlock count overflowing soon?
3507 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3510 EXPORT_SYMBOL(add_preempt_count);
3512 void fastcall sub_preempt_count(int val)
3517 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3520 * Is the spinlock portion underflowing?
3522 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3523 !(preempt_count() & PREEMPT_MASK)))
3526 preempt_count() -= val;
3528 EXPORT_SYMBOL(sub_preempt_count);
3532 static inline int interactive_sleep(enum sleep_type sleep_type)
3534 return (sleep_type == SLEEP_INTERACTIVE ||
3535 sleep_type == SLEEP_INTERRUPTED);
3539 * schedule() is the main scheduler function.
3541 asmlinkage void __sched schedule(void)
3543 struct task_struct *prev, *next;
3544 struct prio_array *array;
3545 struct list_head *queue;
3546 unsigned long long now;
3547 unsigned long run_time;
3548 int cpu, idx, new_prio;
3553 * Test if we are atomic. Since do_exit() needs to call into
3554 * schedule() atomically, we ignore that path for now.
3555 * Otherwise, whine if we are scheduling when we should not be.
3557 if (unlikely(in_atomic() && !current->exit_state)) {
3558 printk(KERN_ERR "BUG: scheduling while atomic: "
3560 current->comm, preempt_count(), current->pid);
3561 debug_show_held_locks(current);
3562 if (irqs_disabled())
3563 print_irqtrace_events(current);
3566 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3571 release_kernel_lock(prev);
3572 need_resched_nonpreemptible:
3576 * The idle thread is not allowed to schedule!
3577 * Remove this check after it has been exercised a bit.
3579 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3580 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3584 schedstat_inc(rq, sched_cnt);
3585 now = sched_clock();
3586 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3587 run_time = now - prev->timestamp;
3588 if (unlikely((long long)(now - prev->timestamp) < 0))
3591 run_time = NS_MAX_SLEEP_AVG;
3594 * Tasks charged proportionately less run_time at high sleep_avg to
3595 * delay them losing their interactive status
3597 run_time /= (CURRENT_BONUS(prev) ? : 1);
3599 spin_lock_irq(&rq->lock);
3601 switch_count = &prev->nivcsw;
3602 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3603 switch_count = &prev->nvcsw;
3604 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3605 unlikely(signal_pending(prev))))
3606 prev->state = TASK_RUNNING;
3608 if (prev->state == TASK_UNINTERRUPTIBLE)
3609 rq->nr_uninterruptible++;
3610 deactivate_task(prev, rq);
3614 cpu = smp_processor_id();
3615 if (unlikely(!rq->nr_running)) {
3616 idle_balance(cpu, rq);
3617 if (!rq->nr_running) {
3619 rq->expired_timestamp = 0;
3625 if (unlikely(!array->nr_active)) {
3627 * Switch the active and expired arrays.
3629 schedstat_inc(rq, sched_switch);
3630 rq->active = rq->expired;
3631 rq->expired = array;
3633 rq->expired_timestamp = 0;
3634 rq->best_expired_prio = MAX_PRIO;
3637 idx = sched_find_first_bit(array->bitmap);
3638 queue = array->queue + idx;
3639 next = list_entry(queue->next, struct task_struct, run_list);
3641 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3642 unsigned long long delta = now - next->timestamp;
3643 if (unlikely((long long)(now - next->timestamp) < 0))
3646 if (next->sleep_type == SLEEP_INTERACTIVE)
3647 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3649 array = next->array;
3650 new_prio = recalc_task_prio(next, next->timestamp + delta);
3652 if (unlikely(next->prio != new_prio)) {
3653 dequeue_task(next, array);
3654 next->prio = new_prio;
3655 enqueue_task(next, array);
3658 next->sleep_type = SLEEP_NORMAL;
3660 if (next == rq->idle)
3661 schedstat_inc(rq, sched_goidle);
3663 prefetch_stack(next);
3664 clear_tsk_need_resched(prev);
3665 rcu_qsctr_inc(task_cpu(prev));
3667 update_cpu_clock(prev, rq, now);
3669 prev->sleep_avg -= run_time;
3670 if ((long)prev->sleep_avg <= 0)
3671 prev->sleep_avg = 0;
3672 prev->timestamp = prev->last_ran = now;
3674 sched_info_switch(prev, next);
3675 if (likely(prev != next)) {
3676 next->timestamp = next->last_ran = now;
3681 prepare_task_switch(rq, next);
3682 prev = context_switch(rq, prev, next);
3685 * this_rq must be evaluated again because prev may have moved
3686 * CPUs since it called schedule(), thus the 'rq' on its stack
3687 * frame will be invalid.
3689 finish_task_switch(this_rq(), prev);
3691 spin_unlock_irq(&rq->lock);
3694 if (unlikely(reacquire_kernel_lock(prev) < 0))
3695 goto need_resched_nonpreemptible;
3696 preempt_enable_no_resched();
3697 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3700 EXPORT_SYMBOL(schedule);
3702 #ifdef CONFIG_PREEMPT
3704 * this is the entry point to schedule() from in-kernel preemption
3705 * off of preempt_enable. Kernel preemptions off return from interrupt
3706 * occur there and call schedule directly.
3708 asmlinkage void __sched preempt_schedule(void)
3710 struct thread_info *ti = current_thread_info();
3711 #ifdef CONFIG_PREEMPT_BKL
3712 struct task_struct *task = current;
3713 int saved_lock_depth;
3716 * If there is a non-zero preempt_count or interrupts are disabled,
3717 * we do not want to preempt the current task. Just return..
3719 if (likely(ti->preempt_count || irqs_disabled()))
3723 add_preempt_count(PREEMPT_ACTIVE);
3725 * We keep the big kernel semaphore locked, but we
3726 * clear ->lock_depth so that schedule() doesnt
3727 * auto-release the semaphore:
3729 #ifdef CONFIG_PREEMPT_BKL
3730 saved_lock_depth = task->lock_depth;
3731 task->lock_depth = -1;
3734 #ifdef CONFIG_PREEMPT_BKL
3735 task->lock_depth = saved_lock_depth;
3737 sub_preempt_count(PREEMPT_ACTIVE);
3739 /* we could miss a preemption opportunity between schedule and now */
3741 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3744 EXPORT_SYMBOL(preempt_schedule);
3747 * this is the entry point to schedule() from kernel preemption
3748 * off of irq context.
3749 * Note, that this is called and return with irqs disabled. This will
3750 * protect us against recursive calling from irq.
3752 asmlinkage void __sched preempt_schedule_irq(void)
3754 struct thread_info *ti = current_thread_info();
3755 #ifdef CONFIG_PREEMPT_BKL
3756 struct task_struct *task = current;
3757 int saved_lock_depth;
3759 /* Catch callers which need to be fixed */
3760 BUG_ON(ti->preempt_count || !irqs_disabled());
3763 add_preempt_count(PREEMPT_ACTIVE);
3765 * We keep the big kernel semaphore locked, but we
3766 * clear ->lock_depth so that schedule() doesnt
3767 * auto-release the semaphore:
3769 #ifdef CONFIG_PREEMPT_BKL
3770 saved_lock_depth = task->lock_depth;
3771 task->lock_depth = -1;
3775 local_irq_disable();
3776 #ifdef CONFIG_PREEMPT_BKL
3777 task->lock_depth = saved_lock_depth;
3779 sub_preempt_count(PREEMPT_ACTIVE);
3781 /* we could miss a preemption opportunity between schedule and now */
3783 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3787 #endif /* CONFIG_PREEMPT */
3789 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3792 return try_to_wake_up(curr->private, mode, sync);
3794 EXPORT_SYMBOL(default_wake_function);
3797 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3798 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3799 * number) then we wake all the non-exclusive tasks and one exclusive task.
3801 * There are circumstances in which we can try to wake a task which has already
3802 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3803 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3805 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3806 int nr_exclusive, int sync, void *key)
3808 struct list_head *tmp, *next;
3810 list_for_each_safe(tmp, next, &q->task_list) {
3811 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3812 unsigned flags = curr->flags;
3814 if (curr->func(curr, mode, sync, key) &&
3815 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3821 * __wake_up - wake up threads blocked on a waitqueue.
3823 * @mode: which threads
3824 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3825 * @key: is directly passed to the wakeup function
3827 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3828 int nr_exclusive, void *key)
3830 unsigned long flags;
3832 spin_lock_irqsave(&q->lock, flags);
3833 __wake_up_common(q, mode, nr_exclusive, 0, key);
3834 spin_unlock_irqrestore(&q->lock, flags);
3836 EXPORT_SYMBOL(__wake_up);
3839 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3841 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3843 __wake_up_common(q, mode, 1, 0, NULL);
3847 * __wake_up_sync - wake up threads blocked on a waitqueue.
3849 * @mode: which threads
3850 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3852 * The sync wakeup differs that the waker knows that it will schedule
3853 * away soon, so while the target thread will be woken up, it will not
3854 * be migrated to another CPU - ie. the two threads are 'synchronized'
3855 * with each other. This can prevent needless bouncing between CPUs.
3857 * On UP it can prevent extra preemption.
3860 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3862 unsigned long flags;
3868 if (unlikely(!nr_exclusive))
3871 spin_lock_irqsave(&q->lock, flags);
3872 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3873 spin_unlock_irqrestore(&q->lock, flags);
3875 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3877 void fastcall complete(struct completion *x)
3879 unsigned long flags;
3881 spin_lock_irqsave(&x->wait.lock, flags);
3883 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3885 spin_unlock_irqrestore(&x->wait.lock, flags);
3887 EXPORT_SYMBOL(complete);
3889 void fastcall complete_all(struct completion *x)
3891 unsigned long flags;
3893 spin_lock_irqsave(&x->wait.lock, flags);
3894 x->done += UINT_MAX/2;
3895 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3897 spin_unlock_irqrestore(&x->wait.lock, flags);
3899 EXPORT_SYMBOL(complete_all);
3901 void fastcall __sched wait_for_completion(struct completion *x)
3905 spin_lock_irq(&x->wait.lock);
3907 DECLARE_WAITQUEUE(wait, current);
3909 wait.flags |= WQ_FLAG_EXCLUSIVE;
3910 __add_wait_queue_tail(&x->wait, &wait);
3912 __set_current_state(TASK_UNINTERRUPTIBLE);
3913 spin_unlock_irq(&x->wait.lock);
3915 spin_lock_irq(&x->wait.lock);
3917 __remove_wait_queue(&x->wait, &wait);
3920 spin_unlock_irq(&x->wait.lock);
3922 EXPORT_SYMBOL(wait_for_completion);
3924 unsigned long fastcall __sched
3925 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3929 spin_lock_irq(&x->wait.lock);
3931 DECLARE_WAITQUEUE(wait, current);
3933 wait.flags |= WQ_FLAG_EXCLUSIVE;
3934 __add_wait_queue_tail(&x->wait, &wait);
3936 __set_current_state(TASK_UNINTERRUPTIBLE);
3937 spin_unlock_irq(&x->wait.lock);
3938 timeout = schedule_timeout(timeout);
3939 spin_lock_irq(&x->wait.lock);
3941 __remove_wait_queue(&x->wait, &wait);
3945 __remove_wait_queue(&x->wait, &wait);
3949 spin_unlock_irq(&x->wait.lock);
3952 EXPORT_SYMBOL(wait_for_completion_timeout);
3954 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3960 spin_lock_irq(&x->wait.lock);
3962 DECLARE_WAITQUEUE(wait, current);
3964 wait.flags |= WQ_FLAG_EXCLUSIVE;
3965 __add_wait_queue_tail(&x->wait, &wait);
3967 if (signal_pending(current)) {
3969 __remove_wait_queue(&x->wait, &wait);
3972 __set_current_state(TASK_INTERRUPTIBLE);
3973 spin_unlock_irq(&x->wait.lock);
3975 spin_lock_irq(&x->wait.lock);
3977 __remove_wait_queue(&x->wait, &wait);
3981 spin_unlock_irq(&x->wait.lock);
3985 EXPORT_SYMBOL(wait_for_completion_interruptible);
3987 unsigned long fastcall __sched
3988 wait_for_completion_interruptible_timeout(struct completion *x,
3989 unsigned long timeout)
3993 spin_lock_irq(&x->wait.lock);
3995 DECLARE_WAITQUEUE(wait, current);
3997 wait.flags |= WQ_FLAG_EXCLUSIVE;
3998 __add_wait_queue_tail(&x->wait, &wait);
4000 if (signal_pending(current)) {
4001 timeout = -ERESTARTSYS;
4002 __remove_wait_queue(&x->wait, &wait);
4005 __set_current_state(TASK_INTERRUPTIBLE);
4006 spin_unlock_irq(&x->wait.lock);
4007 timeout = schedule_timeout(timeout);
4008 spin_lock_irq(&x->wait.lock);
4010 __remove_wait_queue(&x->wait, &wait);
4014 __remove_wait_queue(&x->wait, &wait);
4018 spin_unlock_irq(&x->wait.lock);
4021 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4024 #define SLEEP_ON_VAR \
4025 unsigned long flags; \
4026 wait_queue_t wait; \
4027 init_waitqueue_entry(&wait, current);
4029 #define SLEEP_ON_HEAD \
4030 spin_lock_irqsave(&q->lock,flags); \
4031 __add_wait_queue(q, &wait); \
4032 spin_unlock(&q->lock);
4034 #define SLEEP_ON_TAIL \
4035 spin_lock_irq(&q->lock); \
4036 __remove_wait_queue(q, &wait); \
4037 spin_unlock_irqrestore(&q->lock, flags);
4039 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
4043 current->state = TASK_INTERRUPTIBLE;
4049 EXPORT_SYMBOL(interruptible_sleep_on);
4051 long fastcall __sched
4052 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4056 current->state = TASK_INTERRUPTIBLE;
4059 timeout = schedule_timeout(timeout);
4064 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4066 void fastcall __sched sleep_on(wait_queue_head_t *q)
4070 current->state = TASK_UNINTERRUPTIBLE;
4076 EXPORT_SYMBOL(sleep_on);
4078 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4082 current->state = TASK_UNINTERRUPTIBLE;
4085 timeout = schedule_timeout(timeout);
4091 EXPORT_SYMBOL(sleep_on_timeout);
4093 #ifdef CONFIG_RT_MUTEXES
4096 * rt_mutex_setprio - set the current priority of a task
4098 * @prio: prio value (kernel-internal form)
4100 * This function changes the 'effective' priority of a task. It does
4101 * not touch ->normal_prio like __setscheduler().
4103 * Used by the rt_mutex code to implement priority inheritance logic.
4105 void rt_mutex_setprio(struct task_struct *p, int prio)
4107 struct prio_array *array;
4108 unsigned long flags;
4112 BUG_ON(prio < 0 || prio > MAX_PRIO);
4114 rq = task_rq_lock(p, &flags);
4119 dequeue_task(p, array);
4124 * If changing to an RT priority then queue it
4125 * in the active array!
4129 enqueue_task(p, array);
4131 * Reschedule if we are currently running on this runqueue and
4132 * our priority decreased, or if we are not currently running on
4133 * this runqueue and our priority is higher than the current's
4135 if (task_running(rq, p)) {
4136 if (p->prio > oldprio)
4137 resched_task(rq->curr);
4138 } else if (TASK_PREEMPTS_CURR(p, rq))
4139 resched_task(rq->curr);
4141 task_rq_unlock(rq, &flags);
4146 void set_user_nice(struct task_struct *p, long nice)
4148 struct prio_array *array;
4149 int old_prio, delta;
4150 unsigned long flags;
4153 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4156 * We have to be careful, if called from sys_setpriority(),
4157 * the task might be in the middle of scheduling on another CPU.
4159 rq = task_rq_lock(p, &flags);
4161 * The RT priorities are set via sched_setscheduler(), but we still
4162 * allow the 'normal' nice value to be set - but as expected
4163 * it wont have any effect on scheduling until the task is
4164 * not SCHED_NORMAL/SCHED_BATCH:
4166 if (has_rt_policy(p)) {
4167 p->static_prio = NICE_TO_PRIO(nice);
4172 dequeue_task(p, array);
4173 dec_raw_weighted_load(rq, p);
4176 p->static_prio = NICE_TO_PRIO(nice);
4179 p->prio = effective_prio(p);
4180 delta = p->prio - old_prio;
4183 enqueue_task(p, array);
4184 inc_raw_weighted_load(rq, p);
4186 * If the task increased its priority or is running and
4187 * lowered its priority, then reschedule its CPU:
4189 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4190 resched_task(rq->curr);
4193 task_rq_unlock(rq, &flags);
4195 EXPORT_SYMBOL(set_user_nice);
4198 * can_nice - check if a task can reduce its nice value
4202 int can_nice(const struct task_struct *p, const int nice)
4204 /* convert nice value [19,-20] to rlimit style value [1,40] */
4205 int nice_rlim = 20 - nice;
4207 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4208 capable(CAP_SYS_NICE));
4211 #ifdef __ARCH_WANT_SYS_NICE
4214 * sys_nice - change the priority of the current process.
4215 * @increment: priority increment
4217 * sys_setpriority is a more generic, but much slower function that
4218 * does similar things.
4220 asmlinkage long sys_nice(int increment)
4225 * Setpriority might change our priority at the same moment.
4226 * We don't have to worry. Conceptually one call occurs first
4227 * and we have a single winner.
4229 if (increment < -40)
4234 nice = PRIO_TO_NICE(current->static_prio) + increment;
4240 if (increment < 0 && !can_nice(current, nice))
4243 retval = security_task_setnice(current, nice);
4247 set_user_nice(current, nice);
4254 * task_prio - return the priority value of a given task.
4255 * @p: the task in question.
4257 * This is the priority value as seen by users in /proc.
4258 * RT tasks are offset by -200. Normal tasks are centered
4259 * around 0, value goes from -16 to +15.
4261 int task_prio(const struct task_struct *p)
4263 return p->prio - MAX_RT_PRIO;
4267 * task_nice - return the nice value of a given task.
4268 * @p: the task in question.
4270 int task_nice(const struct task_struct *p)
4272 return TASK_NICE(p);
4274 EXPORT_SYMBOL_GPL(task_nice);
4277 * idle_cpu - is a given cpu idle currently?
4278 * @cpu: the processor in question.
4280 int idle_cpu(int cpu)
4282 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4286 * idle_task - return the idle task for a given cpu.
4287 * @cpu: the processor in question.
4289 struct task_struct *idle_task(int cpu)
4291 return cpu_rq(cpu)->idle;
4295 * find_process_by_pid - find a process with a matching PID value.
4296 * @pid: the pid in question.
4298 static inline struct task_struct *find_process_by_pid(pid_t pid)
4300 return pid ? find_task_by_pid(pid) : current;
4303 /* Actually do priority change: must hold rq lock. */
4304 static void __setscheduler(struct task_struct *p, int policy, int prio)
4309 p->rt_priority = prio;
4310 p->normal_prio = normal_prio(p);
4311 /* we are holding p->pi_lock already */
4312 p->prio = rt_mutex_getprio(p);
4314 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4316 if (policy == SCHED_BATCH)
4322 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4323 * @p: the task in question.
4324 * @policy: new policy.
4325 * @param: structure containing the new RT priority.
4327 * NOTE that the task may be already dead.
4329 int sched_setscheduler(struct task_struct *p, int policy,
4330 struct sched_param *param)
4332 int retval, oldprio, oldpolicy = -1;
4333 struct prio_array *array;
4334 unsigned long flags;
4337 /* may grab non-irq protected spin_locks */
4338 BUG_ON(in_interrupt());
4340 /* double check policy once rq lock held */
4342 policy = oldpolicy = p->policy;
4343 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4344 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4347 * Valid priorities for SCHED_FIFO and SCHED_RR are
4348 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4351 if (param->sched_priority < 0 ||
4352 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4353 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4355 if (is_rt_policy(policy) != (param->sched_priority != 0))
4359 * Allow unprivileged RT tasks to decrease priority:
4361 if (!capable(CAP_SYS_NICE)) {
4362 if (is_rt_policy(policy)) {
4363 unsigned long rlim_rtprio;
4364 unsigned long flags;
4366 if (!lock_task_sighand(p, &flags))
4368 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4369 unlock_task_sighand(p, &flags);
4371 /* can't set/change the rt policy */
4372 if (policy != p->policy && !rlim_rtprio)
4375 /* can't increase priority */
4376 if (param->sched_priority > p->rt_priority &&
4377 param->sched_priority > rlim_rtprio)
4381 /* can't change other user's priorities */
4382 if ((current->euid != p->euid) &&
4383 (current->euid != p->uid))
4387 retval = security_task_setscheduler(p, policy, param);
4391 * make sure no PI-waiters arrive (or leave) while we are
4392 * changing the priority of the task:
4394 spin_lock_irqsave(&p->pi_lock, flags);
4396 * To be able to change p->policy safely, the apropriate
4397 * runqueue lock must be held.
4399 rq = __task_rq_lock(p);
4400 /* recheck policy now with rq lock held */
4401 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4402 policy = oldpolicy = -1;
4403 __task_rq_unlock(rq);
4404 spin_unlock_irqrestore(&p->pi_lock, flags);
4409 deactivate_task(p, rq);
4411 __setscheduler(p, policy, param->sched_priority);
4413 __activate_task(p, rq);
4415 * Reschedule if we are currently running on this runqueue and
4416 * our priority decreased, or if we are not currently running on
4417 * this runqueue and our priority is higher than the current's
4419 if (task_running(rq, p)) {
4420 if (p->prio > oldprio)
4421 resched_task(rq->curr);
4422 } else if (TASK_PREEMPTS_CURR(p, rq))
4423 resched_task(rq->curr);
4425 __task_rq_unlock(rq);
4426 spin_unlock_irqrestore(&p->pi_lock, flags);
4428 rt_mutex_adjust_pi(p);
4432 EXPORT_SYMBOL_GPL(sched_setscheduler);
4435 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4437 struct sched_param lparam;
4438 struct task_struct *p;
4441 if (!param || pid < 0)
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4448 p = find_process_by_pid(pid);
4450 retval = sched_setscheduler(p, policy, &lparam);
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4462 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4463 struct sched_param __user *param)
4465 /* negative values for policy are not valid */
4469 return do_sched_setscheduler(pid, policy, param);
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4477 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4479 return do_sched_setscheduler(pid, -1, param);
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4486 asmlinkage long sys_sched_getscheduler(pid_t pid)
4488 struct task_struct *p;
4489 int retval = -EINVAL;
4495 read_lock(&tasklist_lock);
4496 p = find_process_by_pid(pid);
4498 retval = security_task_getscheduler(p);
4502 read_unlock(&tasklist_lock);
4509 * sys_sched_getscheduler - get the RT priority of a thread
4510 * @pid: the pid in question.
4511 * @param: structure containing the RT priority.
4513 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4515 struct sched_param lp;
4516 struct task_struct *p;
4517 int retval = -EINVAL;
4519 if (!param || pid < 0)
4522 read_lock(&tasklist_lock);
4523 p = find_process_by_pid(pid);
4528 retval = security_task_getscheduler(p);
4532 lp.sched_priority = p->rt_priority;
4533 read_unlock(&tasklist_lock);
4536 * This one might sleep, we cannot do it with a spinlock held ...
4538 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4544 read_unlock(&tasklist_lock);
4548 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4550 cpumask_t cpus_allowed;
4551 struct task_struct *p;
4554 mutex_lock(&sched_hotcpu_mutex);
4555 read_lock(&tasklist_lock);
4557 p = find_process_by_pid(pid);
4559 read_unlock(&tasklist_lock);
4560 mutex_unlock(&sched_hotcpu_mutex);
4565 * It is not safe to call set_cpus_allowed with the
4566 * tasklist_lock held. We will bump the task_struct's
4567 * usage count and then drop tasklist_lock.
4570 read_unlock(&tasklist_lock);
4573 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4574 !capable(CAP_SYS_NICE))
4577 retval = security_task_setscheduler(p, 0, NULL);
4581 cpus_allowed = cpuset_cpus_allowed(p);
4582 cpus_and(new_mask, new_mask, cpus_allowed);
4583 retval = set_cpus_allowed(p, new_mask);
4587 mutex_unlock(&sched_hotcpu_mutex);
4591 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4592 cpumask_t *new_mask)
4594 if (len < sizeof(cpumask_t)) {
4595 memset(new_mask, 0, sizeof(cpumask_t));
4596 } else if (len > sizeof(cpumask_t)) {
4597 len = sizeof(cpumask_t);
4599 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4603 * sys_sched_setaffinity - set the cpu affinity of a process
4604 * @pid: pid of the process
4605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4606 * @user_mask_ptr: user-space pointer to the new cpu mask
4608 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4609 unsigned long __user *user_mask_ptr)
4614 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4618 return sched_setaffinity(pid, new_mask);
4622 * Represents all cpu's present in the system
4623 * In systems capable of hotplug, this map could dynamically grow
4624 * as new cpu's are detected in the system via any platform specific
4625 * method, such as ACPI for e.g.
4628 cpumask_t cpu_present_map __read_mostly;
4629 EXPORT_SYMBOL(cpu_present_map);
4632 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4633 EXPORT_SYMBOL(cpu_online_map);
4635 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4636 EXPORT_SYMBOL(cpu_possible_map);
4639 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4641 struct task_struct *p;
4644 mutex_lock(&sched_hotcpu_mutex);
4645 read_lock(&tasklist_lock);
4648 p = find_process_by_pid(pid);
4652 retval = security_task_getscheduler(p);
4656 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4659 read_unlock(&tasklist_lock);
4660 mutex_unlock(&sched_hotcpu_mutex);
4668 * sys_sched_getaffinity - get the cpu affinity of a process
4669 * @pid: pid of the process
4670 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4671 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4673 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4674 unsigned long __user *user_mask_ptr)
4679 if (len < sizeof(cpumask_t))
4682 ret = sched_getaffinity(pid, &mask);
4686 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4689 return sizeof(cpumask_t);
4693 * sys_sched_yield - yield the current processor to other threads.
4695 * This function yields the current CPU by moving the calling thread
4696 * to the expired array. If there are no other threads running on this
4697 * CPU then this function will return.
4699 asmlinkage long sys_sched_yield(void)
4701 struct rq *rq = this_rq_lock();
4702 struct prio_array *array = current->array, *target = rq->expired;
4704 schedstat_inc(rq, yld_cnt);
4706 * We implement yielding by moving the task into the expired
4709 * (special rule: RT tasks will just roundrobin in the active
4712 if (rt_task(current))
4713 target = rq->active;
4715 if (array->nr_active == 1) {
4716 schedstat_inc(rq, yld_act_empty);
4717 if (!rq->expired->nr_active)
4718 schedstat_inc(rq, yld_both_empty);
4719 } else if (!rq->expired->nr_active)
4720 schedstat_inc(rq, yld_exp_empty);
4722 if (array != target) {
4723 dequeue_task(current, array);
4724 enqueue_task(current, target);
4727 * requeue_task is cheaper so perform that if possible.
4729 requeue_task(current, array);
4732 * Since we are going to call schedule() anyway, there's
4733 * no need to preempt or enable interrupts:
4735 __release(rq->lock);
4736 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4737 _raw_spin_unlock(&rq->lock);
4738 preempt_enable_no_resched();
4745 static void __cond_resched(void)
4747 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4748 __might_sleep(__FILE__, __LINE__);
4751 * The BKS might be reacquired before we have dropped
4752 * PREEMPT_ACTIVE, which could trigger a second
4753 * cond_resched() call.
4756 add_preempt_count(PREEMPT_ACTIVE);
4758 sub_preempt_count(PREEMPT_ACTIVE);
4759 } while (need_resched());
4762 int __sched cond_resched(void)
4764 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4765 system_state == SYSTEM_RUNNING) {
4771 EXPORT_SYMBOL(cond_resched);
4774 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4775 * call schedule, and on return reacquire the lock.
4777 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4778 * operations here to prevent schedule() from being called twice (once via
4779 * spin_unlock(), once by hand).
4781 int cond_resched_lock(spinlock_t *lock)
4785 if (need_lockbreak(lock)) {
4791 if (need_resched() && system_state == SYSTEM_RUNNING) {
4792 spin_release(&lock->dep_map, 1, _THIS_IP_);
4793 _raw_spin_unlock(lock);
4794 preempt_enable_no_resched();
4801 EXPORT_SYMBOL(cond_resched_lock);
4803 int __sched cond_resched_softirq(void)
4805 BUG_ON(!in_softirq());
4807 if (need_resched() && system_state == SYSTEM_RUNNING) {
4815 EXPORT_SYMBOL(cond_resched_softirq);
4818 * yield - yield the current processor to other threads.
4820 * This is a shortcut for kernel-space yielding - it marks the
4821 * thread runnable and calls sys_sched_yield().
4823 void __sched yield(void)
4825 set_current_state(TASK_RUNNING);
4828 EXPORT_SYMBOL(yield);
4831 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4832 * that process accounting knows that this is a task in IO wait state.
4834 * But don't do that if it is a deliberate, throttling IO wait (this task
4835 * has set its backing_dev_info: the queue against which it should throttle)
4837 void __sched io_schedule(void)
4839 struct rq *rq = &__raw_get_cpu_var(runqueues);
4841 delayacct_blkio_start();
4842 atomic_inc(&rq->nr_iowait);
4844 atomic_dec(&rq->nr_iowait);
4845 delayacct_blkio_end();
4847 EXPORT_SYMBOL(io_schedule);
4849 long __sched io_schedule_timeout(long timeout)
4851 struct rq *rq = &__raw_get_cpu_var(runqueues);
4854 delayacct_blkio_start();
4855 atomic_inc(&rq->nr_iowait);
4856 ret = schedule_timeout(timeout);
4857 atomic_dec(&rq->nr_iowait);
4858 delayacct_blkio_end();
4863 * sys_sched_get_priority_max - return maximum RT priority.
4864 * @policy: scheduling class.
4866 * this syscall returns the maximum rt_priority that can be used
4867 * by a given scheduling class.
4869 asmlinkage long sys_sched_get_priority_max(int policy)
4876 ret = MAX_USER_RT_PRIO-1;
4887 * sys_sched_get_priority_min - return minimum RT priority.
4888 * @policy: scheduling class.
4890 * this syscall returns the minimum rt_priority that can be used
4891 * by a given scheduling class.
4893 asmlinkage long sys_sched_get_priority_min(int policy)
4910 * sys_sched_rr_get_interval - return the default timeslice of a process.
4911 * @pid: pid of the process.
4912 * @interval: userspace pointer to the timeslice value.
4914 * this syscall writes the default timeslice value of a given process
4915 * into the user-space timespec buffer. A value of '0' means infinity.
4918 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4920 struct task_struct *p;
4921 int retval = -EINVAL;
4928 read_lock(&tasklist_lock);
4929 p = find_process_by_pid(pid);
4933 retval = security_task_getscheduler(p);
4937 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4938 0 : task_timeslice(p), &t);
4939 read_unlock(&tasklist_lock);
4940 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4944 read_unlock(&tasklist_lock);
4948 static const char stat_nam[] = "RSDTtZX";
4950 static void show_task(struct task_struct *p)
4952 unsigned long free = 0;
4955 state = p->state ? __ffs(p->state) + 1 : 0;
4956 printk("%-13.13s %c", p->comm,
4957 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4958 #if (BITS_PER_LONG == 32)
4959 if (state == TASK_RUNNING)
4960 printk(" running ");
4962 printk(" %08lX ", thread_saved_pc(p));
4964 if (state == TASK_RUNNING)
4965 printk(" running task ");
4967 printk(" %016lx ", thread_saved_pc(p));
4969 #ifdef CONFIG_DEBUG_STACK_USAGE
4971 unsigned long *n = end_of_stack(p);
4974 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4977 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4979 printk(" (L-TLB)\n");
4981 printk(" (NOTLB)\n");
4983 if (state != TASK_RUNNING)
4984 show_stack(p, NULL);
4987 void show_state_filter(unsigned long state_filter)
4989 struct task_struct *g, *p;
4991 #if (BITS_PER_LONG == 32)
4994 printk(" task PC stack pid father child younger older\n");
4998 printk(" task PC stack pid father child younger older\n");
5000 read_lock(&tasklist_lock);
5001 do_each_thread(g, p) {
5003 * reset the NMI-timeout, listing all files on a slow
5004 * console might take alot of time:
5006 touch_nmi_watchdog();
5007 if (!state_filter || (p->state & state_filter))
5009 } while_each_thread(g, p);
5011 touch_all_softlockup_watchdogs();
5013 read_unlock(&tasklist_lock);
5015 * Only show locks if all tasks are dumped:
5017 if (state_filter == -1)
5018 debug_show_all_locks();
5021 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5027 * init_idle - set up an idle thread for a given CPU
5028 * @idle: task in question
5029 * @cpu: cpu the idle task belongs to
5031 * NOTE: this function does not set the idle thread's NEED_RESCHED
5032 * flag, to make booting more robust.
5034 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5036 struct rq *rq = cpu_rq(cpu);
5037 unsigned long flags;
5039 idle->timestamp = sched_clock();
5040 idle->sleep_avg = 0;
5042 idle->prio = idle->normal_prio = MAX_PRIO;
5043 idle->state = TASK_RUNNING;
5044 idle->cpus_allowed = cpumask_of_cpu(cpu);
5045 set_task_cpu(idle, cpu);
5047 spin_lock_irqsave(&rq->lock, flags);
5048 rq->curr = rq->idle = idle;
5049 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5052 spin_unlock_irqrestore(&rq->lock, flags);
5054 /* Set the preempt count _outside_ the spinlocks! */
5055 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5056 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5058 task_thread_info(idle)->preempt_count = 0;
5063 * In a system that switches off the HZ timer nohz_cpu_mask
5064 * indicates which cpus entered this state. This is used
5065 * in the rcu update to wait only for active cpus. For system
5066 * which do not switch off the HZ timer nohz_cpu_mask should
5067 * always be CPU_MASK_NONE.
5069 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5073 * This is how migration works:
5075 * 1) we queue a struct migration_req structure in the source CPU's
5076 * runqueue and wake up that CPU's migration thread.
5077 * 2) we down() the locked semaphore => thread blocks.
5078 * 3) migration thread wakes up (implicitly it forces the migrated
5079 * thread off the CPU)
5080 * 4) it gets the migration request and checks whether the migrated
5081 * task is still in the wrong runqueue.
5082 * 5) if it's in the wrong runqueue then the migration thread removes
5083 * it and puts it into the right queue.
5084 * 6) migration thread up()s the semaphore.
5085 * 7) we wake up and the migration is done.
5089 * Change a given task's CPU affinity. Migrate the thread to a
5090 * proper CPU and schedule it away if the CPU it's executing on
5091 * is removed from the allowed bitmask.
5093 * NOTE: the caller must have a valid reference to the task, the
5094 * task must not exit() & deallocate itself prematurely. The
5095 * call is not atomic; no spinlocks may be held.
5097 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5099 struct migration_req req;
5100 unsigned long flags;
5104 rq = task_rq_lock(p, &flags);
5105 if (!cpus_intersects(new_mask, cpu_online_map)) {
5110 p->cpus_allowed = new_mask;
5111 /* Can the task run on the task's current CPU? If so, we're done */
5112 if (cpu_isset(task_cpu(p), new_mask))
5115 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5116 /* Need help from migration thread: drop lock and wait. */
5117 task_rq_unlock(rq, &flags);
5118 wake_up_process(rq->migration_thread);
5119 wait_for_completion(&req.done);
5120 tlb_migrate_finish(p->mm);
5124 task_rq_unlock(rq, &flags);
5128 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5131 * Move (not current) task off this cpu, onto dest cpu. We're doing
5132 * this because either it can't run here any more (set_cpus_allowed()
5133 * away from this CPU, or CPU going down), or because we're
5134 * attempting to rebalance this task on exec (sched_exec).
5136 * So we race with normal scheduler movements, but that's OK, as long
5137 * as the task is no longer on this CPU.
5139 * Returns non-zero if task was successfully migrated.
5141 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5143 struct rq *rq_dest, *rq_src;
5146 if (unlikely(cpu_is_offline(dest_cpu)))
5149 rq_src = cpu_rq(src_cpu);
5150 rq_dest = cpu_rq(dest_cpu);
5152 double_rq_lock(rq_src, rq_dest);
5153 /* Already moved. */
5154 if (task_cpu(p) != src_cpu)
5156 /* Affinity changed (again). */
5157 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5160 set_task_cpu(p, dest_cpu);
5163 * Sync timestamp with rq_dest's before activating.
5164 * The same thing could be achieved by doing this step
5165 * afterwards, and pretending it was a local activate.
5166 * This way is cleaner and logically correct.
5168 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5169 + rq_dest->most_recent_timestamp;
5170 deactivate_task(p, rq_src);
5171 __activate_task(p, rq_dest);
5172 if (TASK_PREEMPTS_CURR(p, rq_dest))
5173 resched_task(rq_dest->curr);
5177 double_rq_unlock(rq_src, rq_dest);
5182 * migration_thread - this is a highprio system thread that performs
5183 * thread migration by bumping thread off CPU then 'pushing' onto
5186 static int migration_thread(void *data)
5188 int cpu = (long)data;
5192 BUG_ON(rq->migration_thread != current);
5194 set_current_state(TASK_INTERRUPTIBLE);
5195 while (!kthread_should_stop()) {
5196 struct migration_req *req;
5197 struct list_head *head;
5201 spin_lock_irq(&rq->lock);
5203 if (cpu_is_offline(cpu)) {
5204 spin_unlock_irq(&rq->lock);
5208 if (rq->active_balance) {
5209 active_load_balance(rq, cpu);
5210 rq->active_balance = 0;
5213 head = &rq->migration_queue;
5215 if (list_empty(head)) {
5216 spin_unlock_irq(&rq->lock);
5218 set_current_state(TASK_INTERRUPTIBLE);
5221 req = list_entry(head->next, struct migration_req, list);
5222 list_del_init(head->next);
5224 spin_unlock(&rq->lock);
5225 __migrate_task(req->task, cpu, req->dest_cpu);
5228 complete(&req->done);
5230 __set_current_state(TASK_RUNNING);
5234 /* Wait for kthread_stop */
5235 set_current_state(TASK_INTERRUPTIBLE);
5236 while (!kthread_should_stop()) {
5238 set_current_state(TASK_INTERRUPTIBLE);
5240 __set_current_state(TASK_RUNNING);
5244 #ifdef CONFIG_HOTPLUG_CPU
5246 * Figure out where task on dead CPU should go, use force if neccessary.
5247 * NOTE: interrupts should be disabled by the caller
5249 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5251 unsigned long flags;
5258 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5259 cpus_and(mask, mask, p->cpus_allowed);
5260 dest_cpu = any_online_cpu(mask);
5262 /* On any allowed CPU? */
5263 if (dest_cpu == NR_CPUS)
5264 dest_cpu = any_online_cpu(p->cpus_allowed);
5266 /* No more Mr. Nice Guy. */
5267 if (dest_cpu == NR_CPUS) {
5268 rq = task_rq_lock(p, &flags);
5269 cpus_setall(p->cpus_allowed);
5270 dest_cpu = any_online_cpu(p->cpus_allowed);
5271 task_rq_unlock(rq, &flags);
5274 * Don't tell them about moving exiting tasks or
5275 * kernel threads (both mm NULL), since they never
5278 if (p->mm && printk_ratelimit())
5279 printk(KERN_INFO "process %d (%s) no "
5280 "longer affine to cpu%d\n",
5281 p->pid, p->comm, dead_cpu);
5283 if (!__migrate_task(p, dead_cpu, dest_cpu))
5288 * While a dead CPU has no uninterruptible tasks queued at this point,
5289 * it might still have a nonzero ->nr_uninterruptible counter, because
5290 * for performance reasons the counter is not stricly tracking tasks to
5291 * their home CPUs. So we just add the counter to another CPU's counter,
5292 * to keep the global sum constant after CPU-down:
5294 static void migrate_nr_uninterruptible(struct rq *rq_src)
5296 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5297 unsigned long flags;
5299 local_irq_save(flags);
5300 double_rq_lock(rq_src, rq_dest);
5301 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5302 rq_src->nr_uninterruptible = 0;
5303 double_rq_unlock(rq_src, rq_dest);
5304 local_irq_restore(flags);
5307 /* Run through task list and migrate tasks from the dead cpu. */
5308 static void migrate_live_tasks(int src_cpu)
5310 struct task_struct *p, *t;
5312 write_lock_irq(&tasklist_lock);
5314 do_each_thread(t, p) {
5318 if (task_cpu(p) == src_cpu)
5319 move_task_off_dead_cpu(src_cpu, p);
5320 } while_each_thread(t, p);
5322 write_unlock_irq(&tasklist_lock);
5325 /* Schedules idle task to be the next runnable task on current CPU.
5326 * It does so by boosting its priority to highest possible and adding it to
5327 * the _front_ of the runqueue. Used by CPU offline code.
5329 void sched_idle_next(void)
5331 int this_cpu = smp_processor_id();
5332 struct rq *rq = cpu_rq(this_cpu);
5333 struct task_struct *p = rq->idle;
5334 unsigned long flags;
5336 /* cpu has to be offline */
5337 BUG_ON(cpu_online(this_cpu));
5340 * Strictly not necessary since rest of the CPUs are stopped by now
5341 * and interrupts disabled on the current cpu.
5343 spin_lock_irqsave(&rq->lock, flags);
5345 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5347 /* Add idle task to the _front_ of its priority queue: */
5348 __activate_idle_task(p, rq);
5350 spin_unlock_irqrestore(&rq->lock, flags);
5354 * Ensures that the idle task is using init_mm right before its cpu goes
5357 void idle_task_exit(void)
5359 struct mm_struct *mm = current->active_mm;
5361 BUG_ON(cpu_online(smp_processor_id()));
5364 switch_mm(mm, &init_mm, current);
5368 /* called under rq->lock with disabled interrupts */
5369 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5371 struct rq *rq = cpu_rq(dead_cpu);
5373 /* Must be exiting, otherwise would be on tasklist. */
5374 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5376 /* Cannot have done final schedule yet: would have vanished. */
5377 BUG_ON(p->state == TASK_DEAD);
5382 * Drop lock around migration; if someone else moves it,
5383 * that's OK. No task can be added to this CPU, so iteration is
5385 * NOTE: interrupts should be left disabled --dev@
5387 spin_unlock(&rq->lock);
5388 move_task_off_dead_cpu(dead_cpu, p);
5389 spin_lock(&rq->lock);
5394 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5395 static void migrate_dead_tasks(unsigned int dead_cpu)
5397 struct rq *rq = cpu_rq(dead_cpu);
5398 unsigned int arr, i;
5400 for (arr = 0; arr < 2; arr++) {
5401 for (i = 0; i < MAX_PRIO; i++) {
5402 struct list_head *list = &rq->arrays[arr].queue[i];
5404 while (!list_empty(list))
5405 migrate_dead(dead_cpu, list_entry(list->next,
5406 struct task_struct, run_list));
5410 #endif /* CONFIG_HOTPLUG_CPU */
5413 * migration_call - callback that gets triggered when a CPU is added.
5414 * Here we can start up the necessary migration thread for the new CPU.
5416 static int __cpuinit
5417 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5419 struct task_struct *p;
5420 int cpu = (long)hcpu;
5421 unsigned long flags;
5425 case CPU_LOCK_ACQUIRE:
5426 mutex_lock(&sched_hotcpu_mutex);
5429 case CPU_UP_PREPARE:
5430 case CPU_UP_PREPARE_FROZEN:
5431 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5434 p->flags |= PF_NOFREEZE;
5435 kthread_bind(p, cpu);
5436 /* Must be high prio: stop_machine expects to yield to it. */
5437 rq = task_rq_lock(p, &flags);
5438 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5439 task_rq_unlock(rq, &flags);
5440 cpu_rq(cpu)->migration_thread = p;
5444 case CPU_ONLINE_FROZEN:
5445 /* Strictly unneccessary, as first user will wake it. */
5446 wake_up_process(cpu_rq(cpu)->migration_thread);
5449 #ifdef CONFIG_HOTPLUG_CPU
5450 case CPU_UP_CANCELED:
5451 case CPU_UP_CANCELED_FROZEN:
5452 if (!cpu_rq(cpu)->migration_thread)
5454 /* Unbind it from offline cpu so it can run. Fall thru. */
5455 kthread_bind(cpu_rq(cpu)->migration_thread,
5456 any_online_cpu(cpu_online_map));
5457 kthread_stop(cpu_rq(cpu)->migration_thread);
5458 cpu_rq(cpu)->migration_thread = NULL;
5462 case CPU_DEAD_FROZEN:
5463 migrate_live_tasks(cpu);
5465 kthread_stop(rq->migration_thread);
5466 rq->migration_thread = NULL;
5467 /* Idle task back to normal (off runqueue, low prio) */
5468 rq = task_rq_lock(rq->idle, &flags);
5469 deactivate_task(rq->idle, rq);
5470 rq->idle->static_prio = MAX_PRIO;
5471 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5472 migrate_dead_tasks(cpu);
5473 task_rq_unlock(rq, &flags);
5474 migrate_nr_uninterruptible(rq);
5475 BUG_ON(rq->nr_running != 0);
5477 /* No need to migrate the tasks: it was best-effort if
5478 * they didn't take sched_hotcpu_mutex. Just wake up
5479 * the requestors. */
5480 spin_lock_irq(&rq->lock);
5481 while (!list_empty(&rq->migration_queue)) {
5482 struct migration_req *req;
5484 req = list_entry(rq->migration_queue.next,
5485 struct migration_req, list);
5486 list_del_init(&req->list);
5487 complete(&req->done);
5489 spin_unlock_irq(&rq->lock);
5492 case CPU_LOCK_RELEASE:
5493 mutex_unlock(&sched_hotcpu_mutex);
5499 /* Register at highest priority so that task migration (migrate_all_tasks)
5500 * happens before everything else.
5502 static struct notifier_block __cpuinitdata migration_notifier = {
5503 .notifier_call = migration_call,
5507 int __init migration_init(void)
5509 void *cpu = (void *)(long)smp_processor_id();
5512 /* Start one for the boot CPU: */
5513 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5514 BUG_ON(err == NOTIFY_BAD);
5515 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5516 register_cpu_notifier(&migration_notifier);
5524 /* Number of possible processor ids */
5525 int nr_cpu_ids __read_mostly = NR_CPUS;
5526 EXPORT_SYMBOL(nr_cpu_ids);
5528 #undef SCHED_DOMAIN_DEBUG
5529 #ifdef SCHED_DOMAIN_DEBUG
5530 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5535 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5539 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5544 struct sched_group *group = sd->groups;
5545 cpumask_t groupmask;
5547 cpumask_scnprintf(str, NR_CPUS, sd->span);
5548 cpus_clear(groupmask);
5551 for (i = 0; i < level + 1; i++)
5553 printk("domain %d: ", level);
5555 if (!(sd->flags & SD_LOAD_BALANCE)) {
5556 printk("does not load-balance\n");
5558 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5563 printk("span %s\n", str);
5565 if (!cpu_isset(cpu, sd->span))
5566 printk(KERN_ERR "ERROR: domain->span does not contain "
5568 if (!cpu_isset(cpu, group->cpumask))
5569 printk(KERN_ERR "ERROR: domain->groups does not contain"
5573 for (i = 0; i < level + 2; i++)
5579 printk(KERN_ERR "ERROR: group is NULL\n");
5583 if (!group->__cpu_power) {
5585 printk(KERN_ERR "ERROR: domain->cpu_power not "
5589 if (!cpus_weight(group->cpumask)) {
5591 printk(KERN_ERR "ERROR: empty group\n");
5594 if (cpus_intersects(groupmask, group->cpumask)) {
5596 printk(KERN_ERR "ERROR: repeated CPUs\n");
5599 cpus_or(groupmask, groupmask, group->cpumask);
5601 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5604 group = group->next;
5605 } while (group != sd->groups);
5608 if (!cpus_equal(sd->span, groupmask))
5609 printk(KERN_ERR "ERROR: groups don't span "
5617 if (!cpus_subset(groupmask, sd->span))
5618 printk(KERN_ERR "ERROR: parent span is not a superset "
5619 "of domain->span\n");
5624 # define sched_domain_debug(sd, cpu) do { } while (0)
5627 static int sd_degenerate(struct sched_domain *sd)
5629 if (cpus_weight(sd->span) == 1)
5632 /* Following flags need at least 2 groups */
5633 if (sd->flags & (SD_LOAD_BALANCE |
5634 SD_BALANCE_NEWIDLE |
5638 SD_SHARE_PKG_RESOURCES)) {
5639 if (sd->groups != sd->groups->next)
5643 /* Following flags don't use groups */
5644 if (sd->flags & (SD_WAKE_IDLE |
5653 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5655 unsigned long cflags = sd->flags, pflags = parent->flags;
5657 if (sd_degenerate(parent))
5660 if (!cpus_equal(sd->span, parent->span))
5663 /* Does parent contain flags not in child? */
5664 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5665 if (cflags & SD_WAKE_AFFINE)
5666 pflags &= ~SD_WAKE_BALANCE;
5667 /* Flags needing groups don't count if only 1 group in parent */
5668 if (parent->groups == parent->groups->next) {
5669 pflags &= ~(SD_LOAD_BALANCE |
5670 SD_BALANCE_NEWIDLE |
5674 SD_SHARE_PKG_RESOURCES);
5676 if (~cflags & pflags)
5683 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5684 * hold the hotplug lock.
5686 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5688 struct rq *rq = cpu_rq(cpu);
5689 struct sched_domain *tmp;
5691 /* Remove the sched domains which do not contribute to scheduling. */
5692 for (tmp = sd; tmp; tmp = tmp->parent) {
5693 struct sched_domain *parent = tmp->parent;
5696 if (sd_parent_degenerate(tmp, parent)) {
5697 tmp->parent = parent->parent;
5699 parent->parent->child = tmp;
5703 if (sd && sd_degenerate(sd)) {
5709 sched_domain_debug(sd, cpu);
5711 rcu_assign_pointer(rq->sd, sd);
5714 /* cpus with isolated domains */
5715 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5717 /* Setup the mask of cpus configured for isolated domains */
5718 static int __init isolated_cpu_setup(char *str)
5720 int ints[NR_CPUS], i;
5722 str = get_options(str, ARRAY_SIZE(ints), ints);
5723 cpus_clear(cpu_isolated_map);
5724 for (i = 1; i <= ints[0]; i++)
5725 if (ints[i] < NR_CPUS)
5726 cpu_set(ints[i], cpu_isolated_map);
5730 __setup ("isolcpus=", isolated_cpu_setup);
5733 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5734 * to a function which identifies what group(along with sched group) a CPU
5735 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5736 * (due to the fact that we keep track of groups covered with a cpumask_t).
5738 * init_sched_build_groups will build a circular linked list of the groups
5739 * covered by the given span, and will set each group's ->cpumask correctly,
5740 * and ->cpu_power to 0.
5743 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5744 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5745 struct sched_group **sg))
5747 struct sched_group *first = NULL, *last = NULL;
5748 cpumask_t covered = CPU_MASK_NONE;
5751 for_each_cpu_mask(i, span) {
5752 struct sched_group *sg;
5753 int group = group_fn(i, cpu_map, &sg);
5756 if (cpu_isset(i, covered))
5759 sg->cpumask = CPU_MASK_NONE;
5760 sg->__cpu_power = 0;
5762 for_each_cpu_mask(j, span) {
5763 if (group_fn(j, cpu_map, NULL) != group)
5766 cpu_set(j, covered);
5767 cpu_set(j, sg->cpumask);
5778 #define SD_NODES_PER_DOMAIN 16
5783 * find_next_best_node - find the next node to include in a sched_domain
5784 * @node: node whose sched_domain we're building
5785 * @used_nodes: nodes already in the sched_domain
5787 * Find the next node to include in a given scheduling domain. Simply
5788 * finds the closest node not already in the @used_nodes map.
5790 * Should use nodemask_t.
5792 static int find_next_best_node(int node, unsigned long *used_nodes)
5794 int i, n, val, min_val, best_node = 0;
5798 for (i = 0; i < MAX_NUMNODES; i++) {
5799 /* Start at @node */
5800 n = (node + i) % MAX_NUMNODES;
5802 if (!nr_cpus_node(n))
5805 /* Skip already used nodes */
5806 if (test_bit(n, used_nodes))
5809 /* Simple min distance search */
5810 val = node_distance(node, n);
5812 if (val < min_val) {
5818 set_bit(best_node, used_nodes);
5823 * sched_domain_node_span - get a cpumask for a node's sched_domain
5824 * @node: node whose cpumask we're constructing
5825 * @size: number of nodes to include in this span
5827 * Given a node, construct a good cpumask for its sched_domain to span. It
5828 * should be one that prevents unnecessary balancing, but also spreads tasks
5831 static cpumask_t sched_domain_node_span(int node)
5833 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5834 cpumask_t span, nodemask;
5838 bitmap_zero(used_nodes, MAX_NUMNODES);
5840 nodemask = node_to_cpumask(node);
5841 cpus_or(span, span, nodemask);
5842 set_bit(node, used_nodes);
5844 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5845 int next_node = find_next_best_node(node, used_nodes);
5847 nodemask = node_to_cpumask(next_node);
5848 cpus_or(span, span, nodemask);
5855 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5858 * SMT sched-domains:
5860 #ifdef CONFIG_SCHED_SMT
5861 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5862 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5864 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5865 struct sched_group **sg)
5868 *sg = &per_cpu(sched_group_cpus, cpu);
5874 * multi-core sched-domains:
5876 #ifdef CONFIG_SCHED_MC
5877 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5878 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5881 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5882 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5883 struct sched_group **sg)
5886 cpumask_t mask = cpu_sibling_map[cpu];
5887 cpus_and(mask, mask, *cpu_map);
5888 group = first_cpu(mask);
5890 *sg = &per_cpu(sched_group_core, group);
5893 #elif defined(CONFIG_SCHED_MC)
5894 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5895 struct sched_group **sg)
5898 *sg = &per_cpu(sched_group_core, cpu);
5903 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5904 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5906 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5907 struct sched_group **sg)
5910 #ifdef CONFIG_SCHED_MC
5911 cpumask_t mask = cpu_coregroup_map(cpu);
5912 cpus_and(mask, mask, *cpu_map);
5913 group = first_cpu(mask);
5914 #elif defined(CONFIG_SCHED_SMT)
5915 cpumask_t mask = cpu_sibling_map[cpu];
5916 cpus_and(mask, mask, *cpu_map);
5917 group = first_cpu(mask);
5922 *sg = &per_cpu(sched_group_phys, group);
5928 * The init_sched_build_groups can't handle what we want to do with node
5929 * groups, so roll our own. Now each node has its own list of groups which
5930 * gets dynamically allocated.
5932 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5933 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5935 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5936 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5938 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5939 struct sched_group **sg)
5941 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5944 cpus_and(nodemask, nodemask, *cpu_map);
5945 group = first_cpu(nodemask);
5948 *sg = &per_cpu(sched_group_allnodes, group);
5952 static void init_numa_sched_groups_power(struct sched_group *group_head)
5954 struct sched_group *sg = group_head;
5960 for_each_cpu_mask(j, sg->cpumask) {
5961 struct sched_domain *sd;
5963 sd = &per_cpu(phys_domains, j);
5964 if (j != first_cpu(sd->groups->cpumask)) {
5966 * Only add "power" once for each
5972 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5975 if (sg != group_head)
5981 /* Free memory allocated for various sched_group structures */
5982 static void free_sched_groups(const cpumask_t *cpu_map)
5986 for_each_cpu_mask(cpu, *cpu_map) {
5987 struct sched_group **sched_group_nodes
5988 = sched_group_nodes_bycpu[cpu];
5990 if (!sched_group_nodes)
5993 for (i = 0; i < MAX_NUMNODES; i++) {
5994 cpumask_t nodemask = node_to_cpumask(i);
5995 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5997 cpus_and(nodemask, nodemask, *cpu_map);
5998 if (cpus_empty(nodemask))
6008 if (oldsg != sched_group_nodes[i])
6011 kfree(sched_group_nodes);
6012 sched_group_nodes_bycpu[cpu] = NULL;
6016 static void free_sched_groups(const cpumask_t *cpu_map)
6022 * Initialize sched groups cpu_power.
6024 * cpu_power indicates the capacity of sched group, which is used while
6025 * distributing the load between different sched groups in a sched domain.
6026 * Typically cpu_power for all the groups in a sched domain will be same unless
6027 * there are asymmetries in the topology. If there are asymmetries, group
6028 * having more cpu_power will pickup more load compared to the group having
6031 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6032 * the maximum number of tasks a group can handle in the presence of other idle
6033 * or lightly loaded groups in the same sched domain.
6035 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6037 struct sched_domain *child;
6038 struct sched_group *group;
6040 WARN_ON(!sd || !sd->groups);
6042 if (cpu != first_cpu(sd->groups->cpumask))
6047 sd->groups->__cpu_power = 0;
6050 * For perf policy, if the groups in child domain share resources
6051 * (for example cores sharing some portions of the cache hierarchy
6052 * or SMT), then set this domain groups cpu_power such that each group
6053 * can handle only one task, when there are other idle groups in the
6054 * same sched domain.
6056 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6058 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6059 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6064 * add cpu_power of each child group to this groups cpu_power
6066 group = child->groups;
6068 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6069 group = group->next;
6070 } while (group != child->groups);
6074 * Build sched domains for a given set of cpus and attach the sched domains
6075 * to the individual cpus
6077 static int build_sched_domains(const cpumask_t *cpu_map)
6080 struct sched_domain *sd;
6082 struct sched_group **sched_group_nodes = NULL;
6083 int sd_allnodes = 0;
6086 * Allocate the per-node list of sched groups
6088 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6090 if (!sched_group_nodes) {
6091 printk(KERN_WARNING "Can not alloc sched group node list\n");
6094 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6098 * Set up domains for cpus specified by the cpu_map.
6100 for_each_cpu_mask(i, *cpu_map) {
6101 struct sched_domain *sd = NULL, *p;
6102 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6104 cpus_and(nodemask, nodemask, *cpu_map);
6107 if (cpus_weight(*cpu_map)
6108 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6109 sd = &per_cpu(allnodes_domains, i);
6110 *sd = SD_ALLNODES_INIT;
6111 sd->span = *cpu_map;
6112 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6118 sd = &per_cpu(node_domains, i);
6120 sd->span = sched_domain_node_span(cpu_to_node(i));
6124 cpus_and(sd->span, sd->span, *cpu_map);
6128 sd = &per_cpu(phys_domains, i);
6130 sd->span = nodemask;
6134 cpu_to_phys_group(i, cpu_map, &sd->groups);
6136 #ifdef CONFIG_SCHED_MC
6138 sd = &per_cpu(core_domains, i);
6140 sd->span = cpu_coregroup_map(i);
6141 cpus_and(sd->span, sd->span, *cpu_map);
6144 cpu_to_core_group(i, cpu_map, &sd->groups);
6147 #ifdef CONFIG_SCHED_SMT
6149 sd = &per_cpu(cpu_domains, i);
6150 *sd = SD_SIBLING_INIT;
6151 sd->span = cpu_sibling_map[i];
6152 cpus_and(sd->span, sd->span, *cpu_map);
6155 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6159 #ifdef CONFIG_SCHED_SMT
6160 /* Set up CPU (sibling) groups */
6161 for_each_cpu_mask(i, *cpu_map) {
6162 cpumask_t this_sibling_map = cpu_sibling_map[i];
6163 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6164 if (i != first_cpu(this_sibling_map))
6167 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6171 #ifdef CONFIG_SCHED_MC
6172 /* Set up multi-core groups */
6173 for_each_cpu_mask(i, *cpu_map) {
6174 cpumask_t this_core_map = cpu_coregroup_map(i);
6175 cpus_and(this_core_map, this_core_map, *cpu_map);
6176 if (i != first_cpu(this_core_map))
6178 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6183 /* Set up physical groups */
6184 for (i = 0; i < MAX_NUMNODES; i++) {
6185 cpumask_t nodemask = node_to_cpumask(i);
6187 cpus_and(nodemask, nodemask, *cpu_map);
6188 if (cpus_empty(nodemask))
6191 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6195 /* Set up node groups */
6197 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6199 for (i = 0; i < MAX_NUMNODES; i++) {
6200 /* Set up node groups */
6201 struct sched_group *sg, *prev;
6202 cpumask_t nodemask = node_to_cpumask(i);
6203 cpumask_t domainspan;
6204 cpumask_t covered = CPU_MASK_NONE;
6207 cpus_and(nodemask, nodemask, *cpu_map);
6208 if (cpus_empty(nodemask)) {
6209 sched_group_nodes[i] = NULL;
6213 domainspan = sched_domain_node_span(i);
6214 cpus_and(domainspan, domainspan, *cpu_map);
6216 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6218 printk(KERN_WARNING "Can not alloc domain group for "
6222 sched_group_nodes[i] = sg;
6223 for_each_cpu_mask(j, nodemask) {
6224 struct sched_domain *sd;
6225 sd = &per_cpu(node_domains, j);
6228 sg->__cpu_power = 0;
6229 sg->cpumask = nodemask;
6231 cpus_or(covered, covered, nodemask);
6234 for (j = 0; j < MAX_NUMNODES; j++) {
6235 cpumask_t tmp, notcovered;
6236 int n = (i + j) % MAX_NUMNODES;
6238 cpus_complement(notcovered, covered);
6239 cpus_and(tmp, notcovered, *cpu_map);
6240 cpus_and(tmp, tmp, domainspan);
6241 if (cpus_empty(tmp))
6244 nodemask = node_to_cpumask(n);
6245 cpus_and(tmp, tmp, nodemask);
6246 if (cpus_empty(tmp))
6249 sg = kmalloc_node(sizeof(struct sched_group),
6253 "Can not alloc domain group for node %d\n", j);
6256 sg->__cpu_power = 0;
6258 sg->next = prev->next;
6259 cpus_or(covered, covered, tmp);
6266 /* Calculate CPU power for physical packages and nodes */
6267 #ifdef CONFIG_SCHED_SMT
6268 for_each_cpu_mask(i, *cpu_map) {
6269 sd = &per_cpu(cpu_domains, i);
6270 init_sched_groups_power(i, sd);
6273 #ifdef CONFIG_SCHED_MC
6274 for_each_cpu_mask(i, *cpu_map) {
6275 sd = &per_cpu(core_domains, i);
6276 init_sched_groups_power(i, sd);
6280 for_each_cpu_mask(i, *cpu_map) {
6281 sd = &per_cpu(phys_domains, i);
6282 init_sched_groups_power(i, sd);
6286 for (i = 0; i < MAX_NUMNODES; i++)
6287 init_numa_sched_groups_power(sched_group_nodes[i]);
6290 struct sched_group *sg;
6292 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6293 init_numa_sched_groups_power(sg);
6297 /* Attach the domains */
6298 for_each_cpu_mask(i, *cpu_map) {
6299 struct sched_domain *sd;
6300 #ifdef CONFIG_SCHED_SMT
6301 sd = &per_cpu(cpu_domains, i);
6302 #elif defined(CONFIG_SCHED_MC)
6303 sd = &per_cpu(core_domains, i);
6305 sd = &per_cpu(phys_domains, i);
6307 cpu_attach_domain(sd, i);
6314 free_sched_groups(cpu_map);
6319 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6321 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6323 cpumask_t cpu_default_map;
6327 * Setup mask for cpus without special case scheduling requirements.
6328 * For now this just excludes isolated cpus, but could be used to
6329 * exclude other special cases in the future.
6331 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6333 err = build_sched_domains(&cpu_default_map);
6338 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6340 free_sched_groups(cpu_map);
6344 * Detach sched domains from a group of cpus specified in cpu_map
6345 * These cpus will now be attached to the NULL domain
6347 static void detach_destroy_domains(const cpumask_t *cpu_map)
6351 for_each_cpu_mask(i, *cpu_map)
6352 cpu_attach_domain(NULL, i);
6353 synchronize_sched();
6354 arch_destroy_sched_domains(cpu_map);
6358 * Partition sched domains as specified by the cpumasks below.
6359 * This attaches all cpus from the cpumasks to the NULL domain,
6360 * waits for a RCU quiescent period, recalculates sched
6361 * domain information and then attaches them back to the
6362 * correct sched domains
6363 * Call with hotplug lock held
6365 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6367 cpumask_t change_map;
6370 cpus_and(*partition1, *partition1, cpu_online_map);
6371 cpus_and(*partition2, *partition2, cpu_online_map);
6372 cpus_or(change_map, *partition1, *partition2);
6374 /* Detach sched domains from all of the affected cpus */
6375 detach_destroy_domains(&change_map);
6376 if (!cpus_empty(*partition1))
6377 err = build_sched_domains(partition1);
6378 if (!err && !cpus_empty(*partition2))
6379 err = build_sched_domains(partition2);
6384 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6385 int arch_reinit_sched_domains(void)
6389 mutex_lock(&sched_hotcpu_mutex);
6390 detach_destroy_domains(&cpu_online_map);
6391 err = arch_init_sched_domains(&cpu_online_map);
6392 mutex_unlock(&sched_hotcpu_mutex);
6397 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6401 if (buf[0] != '0' && buf[0] != '1')
6405 sched_smt_power_savings = (buf[0] == '1');
6407 sched_mc_power_savings = (buf[0] == '1');
6409 ret = arch_reinit_sched_domains();
6411 return ret ? ret : count;
6414 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6418 #ifdef CONFIG_SCHED_SMT
6420 err = sysfs_create_file(&cls->kset.kobj,
6421 &attr_sched_smt_power_savings.attr);
6423 #ifdef CONFIG_SCHED_MC
6424 if (!err && mc_capable())
6425 err = sysfs_create_file(&cls->kset.kobj,
6426 &attr_sched_mc_power_savings.attr);
6432 #ifdef CONFIG_SCHED_MC
6433 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6435 return sprintf(page, "%u\n", sched_mc_power_savings);
6437 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6438 const char *buf, size_t count)
6440 return sched_power_savings_store(buf, count, 0);
6442 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6443 sched_mc_power_savings_store);
6446 #ifdef CONFIG_SCHED_SMT
6447 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6449 return sprintf(page, "%u\n", sched_smt_power_savings);
6451 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6452 const char *buf, size_t count)
6454 return sched_power_savings_store(buf, count, 1);
6456 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6457 sched_smt_power_savings_store);
6461 * Force a reinitialization of the sched domains hierarchy. The domains
6462 * and groups cannot be updated in place without racing with the balancing
6463 * code, so we temporarily attach all running cpus to the NULL domain
6464 * which will prevent rebalancing while the sched domains are recalculated.
6466 static int update_sched_domains(struct notifier_block *nfb,
6467 unsigned long action, void *hcpu)
6470 case CPU_UP_PREPARE:
6471 case CPU_UP_PREPARE_FROZEN:
6472 case CPU_DOWN_PREPARE:
6473 case CPU_DOWN_PREPARE_FROZEN:
6474 detach_destroy_domains(&cpu_online_map);
6477 case CPU_UP_CANCELED:
6478 case CPU_UP_CANCELED_FROZEN:
6479 case CPU_DOWN_FAILED:
6480 case CPU_DOWN_FAILED_FROZEN:
6482 case CPU_ONLINE_FROZEN:
6484 case CPU_DEAD_FROZEN:
6486 * Fall through and re-initialise the domains.
6493 /* The hotplug lock is already held by cpu_up/cpu_down */
6494 arch_init_sched_domains(&cpu_online_map);
6499 void __init sched_init_smp(void)
6501 cpumask_t non_isolated_cpus;
6503 mutex_lock(&sched_hotcpu_mutex);
6504 arch_init_sched_domains(&cpu_online_map);
6505 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6506 if (cpus_empty(non_isolated_cpus))
6507 cpu_set(smp_processor_id(), non_isolated_cpus);
6508 mutex_unlock(&sched_hotcpu_mutex);
6509 /* XXX: Theoretical race here - CPU may be hotplugged now */
6510 hotcpu_notifier(update_sched_domains, 0);
6512 /* Move init over to a non-isolated CPU */
6513 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6517 void __init sched_init_smp(void)
6520 #endif /* CONFIG_SMP */
6522 int in_sched_functions(unsigned long addr)
6524 /* Linker adds these: start and end of __sched functions */
6525 extern char __sched_text_start[], __sched_text_end[];
6527 return in_lock_functions(addr) ||
6528 (addr >= (unsigned long)__sched_text_start
6529 && addr < (unsigned long)__sched_text_end);
6532 void __init sched_init(void)
6535 int highest_cpu = 0;
6537 for_each_possible_cpu(i) {
6538 struct prio_array *array;
6542 spin_lock_init(&rq->lock);
6543 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6545 rq->active = rq->arrays;
6546 rq->expired = rq->arrays + 1;
6547 rq->best_expired_prio = MAX_PRIO;
6551 for (j = 1; j < 3; j++)
6552 rq->cpu_load[j] = 0;
6553 rq->active_balance = 0;
6556 rq->migration_thread = NULL;
6557 INIT_LIST_HEAD(&rq->migration_queue);
6559 atomic_set(&rq->nr_iowait, 0);
6561 for (j = 0; j < 2; j++) {
6562 array = rq->arrays + j;
6563 for (k = 0; k < MAX_PRIO; k++) {
6564 INIT_LIST_HEAD(array->queue + k);
6565 __clear_bit(k, array->bitmap);
6567 // delimiter for bitsearch
6568 __set_bit(MAX_PRIO, array->bitmap);
6573 set_load_weight(&init_task);
6576 nr_cpu_ids = highest_cpu + 1;
6577 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6580 #ifdef CONFIG_RT_MUTEXES
6581 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6585 * The boot idle thread does lazy MMU switching as well:
6587 atomic_inc(&init_mm.mm_count);
6588 enter_lazy_tlb(&init_mm, current);
6591 * Make us the idle thread. Technically, schedule() should not be
6592 * called from this thread, however somewhere below it might be,
6593 * but because we are the idle thread, we just pick up running again
6594 * when this runqueue becomes "idle".
6596 init_idle(current, smp_processor_id());
6599 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6600 void __might_sleep(char *file, int line)
6603 static unsigned long prev_jiffy; /* ratelimiting */
6605 if ((in_atomic() || irqs_disabled()) &&
6606 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6607 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6609 prev_jiffy = jiffies;
6610 printk(KERN_ERR "BUG: sleeping function called from invalid"
6611 " context at %s:%d\n", file, line);
6612 printk("in_atomic():%d, irqs_disabled():%d\n",
6613 in_atomic(), irqs_disabled());
6614 debug_show_held_locks(current);
6615 if (irqs_disabled())
6616 print_irqtrace_events(current);
6621 EXPORT_SYMBOL(__might_sleep);
6624 #ifdef CONFIG_MAGIC_SYSRQ
6625 void normalize_rt_tasks(void)
6627 struct prio_array *array;
6628 struct task_struct *g, *p;
6629 unsigned long flags;
6632 read_lock_irq(&tasklist_lock);
6634 do_each_thread(g, p) {
6638 spin_lock_irqsave(&p->pi_lock, flags);
6639 rq = __task_rq_lock(p);
6643 deactivate_task(p, task_rq(p));
6644 __setscheduler(p, SCHED_NORMAL, 0);
6646 __activate_task(p, task_rq(p));
6647 resched_task(rq->curr);
6650 __task_rq_unlock(rq);
6651 spin_unlock_irqrestore(&p->pi_lock, flags);
6652 } while_each_thread(g, p);
6654 read_unlock_irq(&tasklist_lock);
6657 #endif /* CONFIG_MAGIC_SYSRQ */
6661 * These functions are only useful for the IA64 MCA handling.
6663 * They can only be called when the whole system has been
6664 * stopped - every CPU needs to be quiescent, and no scheduling
6665 * activity can take place. Using them for anything else would
6666 * be a serious bug, and as a result, they aren't even visible
6667 * under any other configuration.
6671 * curr_task - return the current task for a given cpu.
6672 * @cpu: the processor in question.
6674 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6676 struct task_struct *curr_task(int cpu)
6678 return cpu_curr(cpu);
6682 * set_curr_task - set the current task for a given cpu.
6683 * @cpu: the processor in question.
6684 * @p: the task pointer to set.
6686 * Description: This function must only be used when non-maskable interrupts
6687 * are serviced on a separate stack. It allows the architecture to switch the
6688 * notion of the current task on a cpu in a non-blocking manner. This function
6689 * must be called with all CPU's synchronized, and interrupts disabled, the
6690 * and caller must save the original value of the current task (see
6691 * curr_task() above) and restore that value before reenabling interrupts and
6692 * re-starting the system.
6694 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6696 void set_curr_task(int cpu, struct task_struct *p)