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);
309 static inline int cpu_of(struct rq *rq)
319 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
320 * See detach_destroy_domains: synchronize_sched for details.
322 * The domain tree of any CPU may only be accessed from within
323 * preempt-disabled sections.
325 #define for_each_domain(cpu, __sd) \
326 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
328 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
329 #define this_rq() (&__get_cpu_var(runqueues))
330 #define task_rq(p) cpu_rq(task_cpu(p))
331 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
333 #ifndef prepare_arch_switch
334 # define prepare_arch_switch(next) do { } while (0)
336 #ifndef finish_arch_switch
337 # define finish_arch_switch(prev) do { } while (0)
340 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
341 static inline int task_running(struct rq *rq, struct task_struct *p)
343 return rq->curr == p;
346 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
350 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
352 #ifdef CONFIG_DEBUG_SPINLOCK
353 /* this is a valid case when another task releases the spinlock */
354 rq->lock.owner = current;
357 * If we are tracking spinlock dependencies then we have to
358 * fix up the runqueue lock - which gets 'carried over' from
361 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
363 spin_unlock_irq(&rq->lock);
366 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
367 static inline int task_running(struct rq *rq, struct task_struct *p)
372 return rq->curr == p;
376 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
380 * We can optimise this out completely for !SMP, because the
381 * SMP rebalancing from interrupt is the only thing that cares
386 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
387 spin_unlock_irq(&rq->lock);
389 spin_unlock(&rq->lock);
393 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
397 * After ->oncpu is cleared, the task can be moved to a different CPU.
398 * We must ensure this doesn't happen until the switch is completely
404 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
408 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
411 * __task_rq_lock - lock the runqueue a given task resides on.
412 * Must be called interrupts disabled.
414 static inline struct rq *__task_rq_lock(struct task_struct *p)
421 spin_lock(&rq->lock);
422 if (unlikely(rq != task_rq(p))) {
423 spin_unlock(&rq->lock);
424 goto repeat_lock_task;
430 * task_rq_lock - lock the runqueue a given task resides on and disable
431 * interrupts. Note the ordering: we can safely lookup the task_rq without
432 * explicitly disabling preemption.
434 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
440 local_irq_save(*flags);
442 spin_lock(&rq->lock);
443 if (unlikely(rq != task_rq(p))) {
444 spin_unlock_irqrestore(&rq->lock, *flags);
445 goto repeat_lock_task;
450 static inline void __task_rq_unlock(struct rq *rq)
453 spin_unlock(&rq->lock);
456 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
459 spin_unlock_irqrestore(&rq->lock, *flags);
462 #ifdef CONFIG_SCHEDSTATS
464 * bump this up when changing the output format or the meaning of an existing
465 * format, so that tools can adapt (or abort)
467 #define SCHEDSTAT_VERSION 14
469 static int show_schedstat(struct seq_file *seq, void *v)
473 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
474 seq_printf(seq, "timestamp %lu\n", jiffies);
475 for_each_online_cpu(cpu) {
476 struct rq *rq = cpu_rq(cpu);
478 struct sched_domain *sd;
482 /* runqueue-specific stats */
484 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
485 cpu, rq->yld_both_empty,
486 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
487 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
488 rq->ttwu_cnt, rq->ttwu_local,
489 rq->rq_sched_info.cpu_time,
490 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
492 seq_printf(seq, "\n");
495 /* domain-specific stats */
497 for_each_domain(cpu, sd) {
498 enum idle_type itype;
499 char mask_str[NR_CPUS];
501 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
502 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
503 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
505 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu "
508 sd->lb_balanced[itype],
509 sd->lb_failed[itype],
510 sd->lb_imbalance[itype],
511 sd->lb_gained[itype],
512 sd->lb_hot_gained[itype],
513 sd->lb_nobusyq[itype],
514 sd->lb_nobusyg[itype]);
516 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
518 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
519 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
520 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
521 sd->ttwu_wake_remote, sd->ttwu_move_affine,
522 sd->ttwu_move_balance);
530 static int schedstat_open(struct inode *inode, struct file *file)
532 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
533 char *buf = kmalloc(size, GFP_KERNEL);
539 res = single_open(file, show_schedstat, NULL);
541 m = file->private_data;
549 const struct file_operations proc_schedstat_operations = {
550 .open = schedstat_open,
553 .release = single_release,
557 * Expects runqueue lock to be held for atomicity of update
560 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
563 rq->rq_sched_info.run_delay += delta_jiffies;
564 rq->rq_sched_info.pcnt++;
569 * Expects runqueue lock to be held for atomicity of update
572 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
575 rq->rq_sched_info.cpu_time += delta_jiffies;
577 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
578 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
579 #else /* !CONFIG_SCHEDSTATS */
581 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
584 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
586 # define schedstat_inc(rq, field) do { } while (0)
587 # define schedstat_add(rq, field, amt) do { } while (0)
591 * this_rq_lock - lock this runqueue and disable interrupts.
593 static inline struct rq *this_rq_lock(void)
600 spin_lock(&rq->lock);
605 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
607 * Called when a process is dequeued from the active array and given
608 * the cpu. We should note that with the exception of interactive
609 * tasks, the expired queue will become the active queue after the active
610 * queue is empty, without explicitly dequeuing and requeuing tasks in the
611 * expired queue. (Interactive tasks may be requeued directly to the
612 * active queue, thus delaying tasks in the expired queue from running;
613 * see scheduler_tick()).
615 * This function is only called from sched_info_arrive(), rather than
616 * dequeue_task(). Even though a task may be queued and dequeued multiple
617 * times as it is shuffled about, we're really interested in knowing how
618 * long it was from the *first* time it was queued to the time that it
621 static inline void sched_info_dequeued(struct task_struct *t)
623 t->sched_info.last_queued = 0;
627 * Called when a task finally hits the cpu. We can now calculate how
628 * long it was waiting to run. We also note when it began so that we
629 * can keep stats on how long its timeslice is.
631 static void sched_info_arrive(struct task_struct *t)
633 unsigned long now = jiffies, delta_jiffies = 0;
635 if (t->sched_info.last_queued)
636 delta_jiffies = now - t->sched_info.last_queued;
637 sched_info_dequeued(t);
638 t->sched_info.run_delay += delta_jiffies;
639 t->sched_info.last_arrival = now;
640 t->sched_info.pcnt++;
642 rq_sched_info_arrive(task_rq(t), delta_jiffies);
646 * Called when a process is queued into either the active or expired
647 * array. The time is noted and later used to determine how long we
648 * had to wait for us to reach the cpu. Since the expired queue will
649 * become the active queue after active queue is empty, without dequeuing
650 * and requeuing any tasks, we are interested in queuing to either. It
651 * is unusual but not impossible for tasks to be dequeued and immediately
652 * requeued in the same or another array: this can happen in sched_yield(),
653 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
656 * This function is only called from enqueue_task(), but also only updates
657 * the timestamp if it is already not set. It's assumed that
658 * sched_info_dequeued() will clear that stamp when appropriate.
660 static inline void sched_info_queued(struct task_struct *t)
662 if (unlikely(sched_info_on()))
663 if (!t->sched_info.last_queued)
664 t->sched_info.last_queued = jiffies;
668 * Called when a process ceases being the active-running process, either
669 * voluntarily or involuntarily. Now we can calculate how long we ran.
671 static inline void sched_info_depart(struct task_struct *t)
673 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
675 t->sched_info.cpu_time += delta_jiffies;
676 rq_sched_info_depart(task_rq(t), delta_jiffies);
680 * Called when tasks are switched involuntarily due, typically, to expiring
681 * their time slice. (This may also be called when switching to or from
682 * the idle task.) We are only called when prev != next.
685 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
687 struct rq *rq = task_rq(prev);
690 * prev now departs the cpu. It's not interesting to record
691 * stats about how efficient we were at scheduling the idle
694 if (prev != rq->idle)
695 sched_info_depart(prev);
697 if (next != rq->idle)
698 sched_info_arrive(next);
701 sched_info_switch(struct task_struct *prev, struct task_struct *next)
703 if (unlikely(sched_info_on()))
704 __sched_info_switch(prev, next);
707 #define sched_info_queued(t) do { } while (0)
708 #define sched_info_switch(t, next) do { } while (0)
709 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
712 * Adding/removing a task to/from a priority array:
714 static void dequeue_task(struct task_struct *p, struct prio_array *array)
717 list_del(&p->run_list);
718 if (list_empty(array->queue + p->prio))
719 __clear_bit(p->prio, array->bitmap);
722 static void enqueue_task(struct task_struct *p, struct prio_array *array)
724 sched_info_queued(p);
725 list_add_tail(&p->run_list, array->queue + p->prio);
726 __set_bit(p->prio, array->bitmap);
732 * Put task to the end of the run list without the overhead of dequeue
733 * followed by enqueue.
735 static void requeue_task(struct task_struct *p, struct prio_array *array)
737 list_move_tail(&p->run_list, array->queue + p->prio);
741 enqueue_task_head(struct task_struct *p, struct prio_array *array)
743 list_add(&p->run_list, array->queue + p->prio);
744 __set_bit(p->prio, array->bitmap);
750 * __normal_prio - return the priority that is based on the static
751 * priority but is modified by bonuses/penalties.
753 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
754 * into the -5 ... 0 ... +5 bonus/penalty range.
756 * We use 25% of the full 0...39 priority range so that:
758 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
759 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
761 * Both properties are important to certain workloads.
764 static inline int __normal_prio(struct task_struct *p)
768 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
770 prio = p->static_prio - bonus;
771 if (prio < MAX_RT_PRIO)
773 if (prio > MAX_PRIO-1)
779 * To aid in avoiding the subversion of "niceness" due to uneven distribution
780 * of tasks with abnormal "nice" values across CPUs the contribution that
781 * each task makes to its run queue's load is weighted according to its
782 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
783 * scaled version of the new time slice allocation that they receive on time
788 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
789 * If static_prio_timeslice() is ever changed to break this assumption then
790 * this code will need modification
792 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
793 #define LOAD_WEIGHT(lp) \
794 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
795 #define PRIO_TO_LOAD_WEIGHT(prio) \
796 LOAD_WEIGHT(static_prio_timeslice(prio))
797 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
798 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
800 static void set_load_weight(struct task_struct *p)
802 if (has_rt_policy(p)) {
804 if (p == task_rq(p)->migration_thread)
806 * The migration thread does the actual balancing.
807 * Giving its load any weight will skew balancing
813 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
815 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
819 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
821 rq->raw_weighted_load += p->load_weight;
825 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
827 rq->raw_weighted_load -= p->load_weight;
830 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
833 inc_raw_weighted_load(rq, p);
836 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
839 dec_raw_weighted_load(rq, p);
843 * Calculate the expected normal priority: i.e. priority
844 * without taking RT-inheritance into account. Might be
845 * boosted by interactivity modifiers. Changes upon fork,
846 * setprio syscalls, and whenever the interactivity
847 * estimator recalculates.
849 static inline int normal_prio(struct task_struct *p)
853 if (has_rt_policy(p))
854 prio = MAX_RT_PRIO-1 - p->rt_priority;
856 prio = __normal_prio(p);
861 * Calculate the current priority, i.e. the priority
862 * taken into account by the scheduler. This value might
863 * be boosted by RT tasks, or might be boosted by
864 * interactivity modifiers. Will be RT if the task got
865 * RT-boosted. If not then it returns p->normal_prio.
867 static int effective_prio(struct task_struct *p)
869 p->normal_prio = normal_prio(p);
871 * If we are RT tasks or we were boosted to RT priority,
872 * keep the priority unchanged. Otherwise, update priority
873 * to the normal priority:
875 if (!rt_prio(p->prio))
876 return p->normal_prio;
881 * __activate_task - move a task to the runqueue.
883 static void __activate_task(struct task_struct *p, struct rq *rq)
885 struct prio_array *target = rq->active;
888 target = rq->expired;
889 enqueue_task(p, target);
890 inc_nr_running(p, rq);
894 * __activate_idle_task - move idle task to the _front_ of runqueue.
896 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
898 enqueue_task_head(p, rq->active);
899 inc_nr_running(p, rq);
903 * Recalculate p->normal_prio and p->prio after having slept,
904 * updating the sleep-average too:
906 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
908 /* Caller must always ensure 'now >= p->timestamp' */
909 unsigned long sleep_time = now - p->timestamp;
914 if (likely(sleep_time > 0)) {
916 * This ceiling is set to the lowest priority that would allow
917 * a task to be reinserted into the active array on timeslice
920 unsigned long ceiling = INTERACTIVE_SLEEP(p);
922 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
924 * Prevents user tasks from achieving best priority
925 * with one single large enough sleep.
927 p->sleep_avg = ceiling;
929 * Using INTERACTIVE_SLEEP() as a ceiling places a
930 * nice(0) task 1ms sleep away from promotion, and
931 * gives it 700ms to round-robin with no chance of
932 * being demoted. This is more than generous, so
933 * mark this sleep as non-interactive to prevent the
934 * on-runqueue bonus logic from intervening should
935 * this task not receive cpu immediately.
937 p->sleep_type = SLEEP_NONINTERACTIVE;
940 * Tasks waking from uninterruptible sleep are
941 * limited in their sleep_avg rise as they
942 * are likely to be waiting on I/O
944 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
945 if (p->sleep_avg >= ceiling)
947 else if (p->sleep_avg + sleep_time >=
949 p->sleep_avg = ceiling;
955 * This code gives a bonus to interactive tasks.
957 * The boost works by updating the 'average sleep time'
958 * value here, based on ->timestamp. The more time a
959 * task spends sleeping, the higher the average gets -
960 * and the higher the priority boost gets as well.
962 p->sleep_avg += sleep_time;
965 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
966 p->sleep_avg = NS_MAX_SLEEP_AVG;
969 return effective_prio(p);
973 * activate_task - move a task to the runqueue and do priority recalculation
975 * Update all the scheduling statistics stuff. (sleep average
976 * calculation, priority modifiers, etc.)
978 static void activate_task(struct task_struct *p, struct rq *rq, int local)
980 unsigned long long now;
988 /* Compensate for drifting sched_clock */
989 struct rq *this_rq = this_rq();
990 now = (now - this_rq->most_recent_timestamp)
991 + rq->most_recent_timestamp;
996 * Sleep time is in units of nanosecs, so shift by 20 to get a
997 * milliseconds-range estimation of the amount of time that the task
1000 if (unlikely(prof_on == SLEEP_PROFILING)) {
1001 if (p->state == TASK_UNINTERRUPTIBLE)
1002 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
1003 (now - p->timestamp) >> 20);
1006 p->prio = recalc_task_prio(p, now);
1009 * This checks to make sure it's not an uninterruptible task
1010 * that is now waking up.
1012 if (p->sleep_type == SLEEP_NORMAL) {
1014 * Tasks which were woken up by interrupts (ie. hw events)
1015 * are most likely of interactive nature. So we give them
1016 * the credit of extending their sleep time to the period
1017 * of time they spend on the runqueue, waiting for execution
1018 * on a CPU, first time around:
1021 p->sleep_type = SLEEP_INTERRUPTED;
1024 * Normal first-time wakeups get a credit too for
1025 * on-runqueue time, but it will be weighted down:
1027 p->sleep_type = SLEEP_INTERACTIVE;
1032 __activate_task(p, rq);
1036 * deactivate_task - remove a task from the runqueue.
1038 static void deactivate_task(struct task_struct *p, struct rq *rq)
1040 dec_nr_running(p, rq);
1041 dequeue_task(p, p->array);
1046 * resched_task - mark a task 'to be rescheduled now'.
1048 * On UP this means the setting of the need_resched flag, on SMP it
1049 * might also involve a cross-CPU call to trigger the scheduler on
1054 #ifndef tsk_is_polling
1055 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1058 static void resched_task(struct task_struct *p)
1062 assert_spin_locked(&task_rq(p)->lock);
1064 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1067 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1070 if (cpu == smp_processor_id())
1073 /* NEED_RESCHED must be visible before we test polling */
1075 if (!tsk_is_polling(p))
1076 smp_send_reschedule(cpu);
1079 static void resched_cpu(int cpu)
1081 struct rq *rq = cpu_rq(cpu);
1082 unsigned long flags;
1084 if (!spin_trylock_irqsave(&rq->lock, flags))
1086 resched_task(cpu_curr(cpu));
1087 spin_unlock_irqrestore(&rq->lock, flags);
1090 static inline void resched_task(struct task_struct *p)
1092 assert_spin_locked(&task_rq(p)->lock);
1093 set_tsk_need_resched(p);
1098 * task_curr - is this task currently executing on a CPU?
1099 * @p: the task in question.
1101 inline int task_curr(const struct task_struct *p)
1103 return cpu_curr(task_cpu(p)) == p;
1106 /* Used instead of source_load when we know the type == 0 */
1107 unsigned long weighted_cpuload(const int cpu)
1109 return cpu_rq(cpu)->raw_weighted_load;
1113 struct migration_req {
1114 struct list_head list;
1116 struct task_struct *task;
1119 struct completion done;
1123 * The task's runqueue lock must be held.
1124 * Returns true if you have to wait for migration thread.
1127 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1129 struct rq *rq = task_rq(p);
1132 * If the task is not on a runqueue (and not running), then
1133 * it is sufficient to simply update the task's cpu field.
1135 if (!p->array && !task_running(rq, p)) {
1136 set_task_cpu(p, dest_cpu);
1140 init_completion(&req->done);
1142 req->dest_cpu = dest_cpu;
1143 list_add(&req->list, &rq->migration_queue);
1149 * wait_task_inactive - wait for a thread to unschedule.
1151 * The caller must ensure that the task *will* unschedule sometime soon,
1152 * else this function might spin for a *long* time. This function can't
1153 * be called with interrupts off, or it may introduce deadlock with
1154 * smp_call_function() if an IPI is sent by the same process we are
1155 * waiting to become inactive.
1157 void wait_task_inactive(struct task_struct *p)
1159 unsigned long flags;
1164 rq = task_rq_lock(p, &flags);
1165 /* Must be off runqueue entirely, not preempted. */
1166 if (unlikely(p->array || task_running(rq, p))) {
1167 /* If it's preempted, we yield. It could be a while. */
1168 preempted = !task_running(rq, p);
1169 task_rq_unlock(rq, &flags);
1175 task_rq_unlock(rq, &flags);
1179 * kick_process - kick a running thread to enter/exit the kernel
1180 * @p: the to-be-kicked thread
1182 * Cause a process which is running on another CPU to enter
1183 * kernel-mode, without any delay. (to get signals handled.)
1185 * NOTE: this function doesnt have to take the runqueue lock,
1186 * because all it wants to ensure is that the remote task enters
1187 * the kernel. If the IPI races and the task has been migrated
1188 * to another CPU then no harm is done and the purpose has been
1191 void kick_process(struct task_struct *p)
1197 if ((cpu != smp_processor_id()) && task_curr(p))
1198 smp_send_reschedule(cpu);
1203 * Return a low guess at the load of a migration-source cpu weighted
1204 * according to the scheduling class and "nice" value.
1206 * We want to under-estimate the load of migration sources, to
1207 * balance conservatively.
1209 static inline unsigned long source_load(int cpu, int type)
1211 struct rq *rq = cpu_rq(cpu);
1214 return rq->raw_weighted_load;
1216 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1220 * Return a high guess at the load of a migration-target cpu weighted
1221 * according to the scheduling class and "nice" value.
1223 static inline unsigned long target_load(int cpu, int type)
1225 struct rq *rq = cpu_rq(cpu);
1228 return rq->raw_weighted_load;
1230 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1234 * Return the average load per task on the cpu's run queue
1236 static inline unsigned long cpu_avg_load_per_task(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1239 unsigned long n = rq->nr_running;
1241 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1245 * find_idlest_group finds and returns the least busy CPU group within the
1248 static struct sched_group *
1249 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1251 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1252 unsigned long min_load = ULONG_MAX, this_load = 0;
1253 int load_idx = sd->forkexec_idx;
1254 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1257 unsigned long load, avg_load;
1261 /* Skip over this group if it has no CPUs allowed */
1262 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1265 local_group = cpu_isset(this_cpu, group->cpumask);
1267 /* Tally up the load of all CPUs in the group */
1270 for_each_cpu_mask(i, group->cpumask) {
1271 /* Bias balancing toward cpus of our domain */
1273 load = source_load(i, load_idx);
1275 load = target_load(i, load_idx);
1280 /* Adjust by relative CPU power of the group */
1281 avg_load = sg_div_cpu_power(group,
1282 avg_load * SCHED_LOAD_SCALE);
1285 this_load = avg_load;
1287 } else if (avg_load < min_load) {
1288 min_load = avg_load;
1292 group = group->next;
1293 } while (group != sd->groups);
1295 if (!idlest || 100*this_load < imbalance*min_load)
1301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1304 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1307 unsigned long load, min_load = ULONG_MAX;
1311 /* Traverse only the allowed CPUs */
1312 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1314 for_each_cpu_mask(i, tmp) {
1315 load = weighted_cpuload(i);
1317 if (load < min_load || (load == min_load && i == this_cpu)) {
1327 * sched_balance_self: balance the current task (running on cpu) in domains
1328 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1331 * Balance, ie. select the least loaded group.
1333 * Returns the target CPU number, or the same CPU if no balancing is needed.
1335 * preempt must be disabled.
1337 static int sched_balance_self(int cpu, int flag)
1339 struct task_struct *t = current;
1340 struct sched_domain *tmp, *sd = NULL;
1342 for_each_domain(cpu, tmp) {
1344 * If power savings logic is enabled for a domain, stop there.
1346 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1348 if (tmp->flags & flag)
1354 struct sched_group *group;
1355 int new_cpu, weight;
1357 if (!(sd->flags & flag)) {
1363 group = find_idlest_group(sd, t, cpu);
1369 new_cpu = find_idlest_cpu(group, t, cpu);
1370 if (new_cpu == -1 || new_cpu == cpu) {
1371 /* Now try balancing at a lower domain level of cpu */
1376 /* Now try balancing at a lower domain level of new_cpu */
1379 weight = cpus_weight(span);
1380 for_each_domain(cpu, tmp) {
1381 if (weight <= cpus_weight(tmp->span))
1383 if (tmp->flags & flag)
1386 /* while loop will break here if sd == NULL */
1392 #endif /* CONFIG_SMP */
1395 * wake_idle() will wake a task on an idle cpu if task->cpu is
1396 * not idle and an idle cpu is available. The span of cpus to
1397 * search starts with cpus closest then further out as needed,
1398 * so we always favor a closer, idle cpu.
1400 * Returns the CPU we should wake onto.
1402 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1403 static int wake_idle(int cpu, struct task_struct *p)
1406 struct sched_domain *sd;
1412 for_each_domain(cpu, sd) {
1413 if (sd->flags & SD_WAKE_IDLE) {
1414 cpus_and(tmp, sd->span, p->cpus_allowed);
1415 for_each_cpu_mask(i, tmp) {
1426 static inline int wake_idle(int cpu, struct task_struct *p)
1433 * try_to_wake_up - wake up a thread
1434 * @p: the to-be-woken-up thread
1435 * @state: the mask of task states that can be woken
1436 * @sync: do a synchronous wakeup?
1438 * Put it on the run-queue if it's not already there. The "current"
1439 * thread is always on the run-queue (except when the actual
1440 * re-schedule is in progress), and as such you're allowed to do
1441 * the simpler "current->state = TASK_RUNNING" to mark yourself
1442 * runnable without the overhead of this.
1444 * returns failure only if the task is already active.
1446 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1448 int cpu, this_cpu, success = 0;
1449 unsigned long flags;
1453 struct sched_domain *sd, *this_sd = NULL;
1454 unsigned long load, this_load;
1458 rq = task_rq_lock(p, &flags);
1459 old_state = p->state;
1460 if (!(old_state & state))
1467 this_cpu = smp_processor_id();
1470 if (unlikely(task_running(rq, p)))
1475 schedstat_inc(rq, ttwu_cnt);
1476 if (cpu == this_cpu) {
1477 schedstat_inc(rq, ttwu_local);
1481 for_each_domain(this_cpu, sd) {
1482 if (cpu_isset(cpu, sd->span)) {
1483 schedstat_inc(sd, ttwu_wake_remote);
1489 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1493 * Check for affine wakeup and passive balancing possibilities.
1496 int idx = this_sd->wake_idx;
1497 unsigned int imbalance;
1499 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1501 load = source_load(cpu, idx);
1502 this_load = target_load(this_cpu, idx);
1504 new_cpu = this_cpu; /* Wake to this CPU if we can */
1506 if (this_sd->flags & SD_WAKE_AFFINE) {
1507 unsigned long tl = this_load;
1508 unsigned long tl_per_task;
1510 tl_per_task = cpu_avg_load_per_task(this_cpu);
1513 * If sync wakeup then subtract the (maximum possible)
1514 * effect of the currently running task from the load
1515 * of the current CPU:
1518 tl -= current->load_weight;
1521 tl + target_load(cpu, idx) <= tl_per_task) ||
1522 100*(tl + p->load_weight) <= imbalance*load) {
1524 * This domain has SD_WAKE_AFFINE and
1525 * p is cache cold in this domain, and
1526 * there is no bad imbalance.
1528 schedstat_inc(this_sd, ttwu_move_affine);
1534 * Start passive balancing when half the imbalance_pct
1537 if (this_sd->flags & SD_WAKE_BALANCE) {
1538 if (imbalance*this_load <= 100*load) {
1539 schedstat_inc(this_sd, ttwu_move_balance);
1545 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1547 new_cpu = wake_idle(new_cpu, p);
1548 if (new_cpu != cpu) {
1549 set_task_cpu(p, new_cpu);
1550 task_rq_unlock(rq, &flags);
1551 /* might preempt at this point */
1552 rq = task_rq_lock(p, &flags);
1553 old_state = p->state;
1554 if (!(old_state & state))
1559 this_cpu = smp_processor_id();
1564 #endif /* CONFIG_SMP */
1565 if (old_state == TASK_UNINTERRUPTIBLE) {
1566 rq->nr_uninterruptible--;
1568 * Tasks on involuntary sleep don't earn
1569 * sleep_avg beyond just interactive state.
1571 p->sleep_type = SLEEP_NONINTERACTIVE;
1575 * Tasks that have marked their sleep as noninteractive get
1576 * woken up with their sleep average not weighted in an
1579 if (old_state & TASK_NONINTERACTIVE)
1580 p->sleep_type = SLEEP_NONINTERACTIVE;
1583 activate_task(p, rq, cpu == this_cpu);
1585 * Sync wakeups (i.e. those types of wakeups where the waker
1586 * has indicated that it will leave the CPU in short order)
1587 * don't trigger a preemption, if the woken up task will run on
1588 * this cpu. (in this case the 'I will reschedule' promise of
1589 * the waker guarantees that the freshly woken up task is going
1590 * to be considered on this CPU.)
1592 if (!sync || cpu != this_cpu) {
1593 if (TASK_PREEMPTS_CURR(p, rq))
1594 resched_task(rq->curr);
1599 p->state = TASK_RUNNING;
1601 task_rq_unlock(rq, &flags);
1606 int fastcall wake_up_process(struct task_struct *p)
1608 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1609 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1611 EXPORT_SYMBOL(wake_up_process);
1613 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1615 return try_to_wake_up(p, state, 0);
1618 static void task_running_tick(struct rq *rq, struct task_struct *p);
1620 * Perform scheduler related setup for a newly forked process p.
1621 * p is forked by current.
1623 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1625 int cpu = get_cpu();
1628 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1630 set_task_cpu(p, cpu);
1633 * We mark the process as running here, but have not actually
1634 * inserted it onto the runqueue yet. This guarantees that
1635 * nobody will actually run it, and a signal or other external
1636 * event cannot wake it up and insert it on the runqueue either.
1638 p->state = TASK_RUNNING;
1641 * Make sure we do not leak PI boosting priority to the child:
1643 p->prio = current->normal_prio;
1645 INIT_LIST_HEAD(&p->run_list);
1647 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1648 if (unlikely(sched_info_on()))
1649 memset(&p->sched_info, 0, sizeof(p->sched_info));
1651 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1654 #ifdef CONFIG_PREEMPT
1655 /* Want to start with kernel preemption disabled. */
1656 task_thread_info(p)->preempt_count = 1;
1659 * Share the timeslice between parent and child, thus the
1660 * total amount of pending timeslices in the system doesn't change,
1661 * resulting in more scheduling fairness.
1663 local_irq_disable();
1664 p->time_slice = (current->time_slice + 1) >> 1;
1666 * The remainder of the first timeslice might be recovered by
1667 * the parent if the child exits early enough.
1669 p->first_time_slice = 1;
1670 current->time_slice >>= 1;
1671 p->timestamp = sched_clock();
1672 if (unlikely(!current->time_slice)) {
1674 * This case is rare, it happens when the parent has only
1675 * a single jiffy left from its timeslice. Taking the
1676 * runqueue lock is not a problem.
1678 current->time_slice = 1;
1679 task_running_tick(cpu_rq(cpu), current);
1686 * wake_up_new_task - wake up a newly created task for the first time.
1688 * This function will do some initial scheduler statistics housekeeping
1689 * that must be done for every newly created context, then puts the task
1690 * on the runqueue and wakes it.
1692 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1694 struct rq *rq, *this_rq;
1695 unsigned long flags;
1698 rq = task_rq_lock(p, &flags);
1699 BUG_ON(p->state != TASK_RUNNING);
1700 this_cpu = smp_processor_id();
1704 * We decrease the sleep average of forking parents
1705 * and children as well, to keep max-interactive tasks
1706 * from forking tasks that are max-interactive. The parent
1707 * (current) is done further down, under its lock.
1709 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1710 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1712 p->prio = effective_prio(p);
1714 if (likely(cpu == this_cpu)) {
1715 if (!(clone_flags & CLONE_VM)) {
1717 * The VM isn't cloned, so we're in a good position to
1718 * do child-runs-first in anticipation of an exec. This
1719 * usually avoids a lot of COW overhead.
1721 if (unlikely(!current->array))
1722 __activate_task(p, rq);
1724 p->prio = current->prio;
1725 p->normal_prio = current->normal_prio;
1726 list_add_tail(&p->run_list, ¤t->run_list);
1727 p->array = current->array;
1728 p->array->nr_active++;
1729 inc_nr_running(p, rq);
1733 /* Run child last */
1734 __activate_task(p, rq);
1736 * We skip the following code due to cpu == this_cpu
1738 * task_rq_unlock(rq, &flags);
1739 * this_rq = task_rq_lock(current, &flags);
1743 this_rq = cpu_rq(this_cpu);
1746 * Not the local CPU - must adjust timestamp. This should
1747 * get optimised away in the !CONFIG_SMP case.
1749 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1750 + rq->most_recent_timestamp;
1751 __activate_task(p, rq);
1752 if (TASK_PREEMPTS_CURR(p, rq))
1753 resched_task(rq->curr);
1756 * Parent and child are on different CPUs, now get the
1757 * parent runqueue to update the parent's ->sleep_avg:
1759 task_rq_unlock(rq, &flags);
1760 this_rq = task_rq_lock(current, &flags);
1762 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1763 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1764 task_rq_unlock(this_rq, &flags);
1768 * Potentially available exiting-child timeslices are
1769 * retrieved here - this way the parent does not get
1770 * penalized for creating too many threads.
1772 * (this cannot be used to 'generate' timeslices
1773 * artificially, because any timeslice recovered here
1774 * was given away by the parent in the first place.)
1776 void fastcall sched_exit(struct task_struct *p)
1778 unsigned long flags;
1782 * If the child was a (relative-) CPU hog then decrease
1783 * the sleep_avg of the parent as well.
1785 rq = task_rq_lock(p->parent, &flags);
1786 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1787 p->parent->time_slice += p->time_slice;
1788 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1789 p->parent->time_slice = task_timeslice(p);
1791 if (p->sleep_avg < p->parent->sleep_avg)
1792 p->parent->sleep_avg = p->parent->sleep_avg /
1793 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1795 task_rq_unlock(rq, &flags);
1799 * prepare_task_switch - prepare to switch tasks
1800 * @rq: the runqueue preparing to switch
1801 * @next: the task we are going to switch to.
1803 * This is called with the rq lock held and interrupts off. It must
1804 * be paired with a subsequent finish_task_switch after the context
1807 * prepare_task_switch sets up locking and calls architecture specific
1810 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1812 prepare_lock_switch(rq, next);
1813 prepare_arch_switch(next);
1817 * finish_task_switch - clean up after a task-switch
1818 * @rq: runqueue associated with task-switch
1819 * @prev: the thread we just switched away from.
1821 * finish_task_switch must be called after the context switch, paired
1822 * with a prepare_task_switch call before the context switch.
1823 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1824 * and do any other architecture-specific cleanup actions.
1826 * Note that we may have delayed dropping an mm in context_switch(). If
1827 * so, we finish that here outside of the runqueue lock. (Doing it
1828 * with the lock held can cause deadlocks; see schedule() for
1831 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1832 __releases(rq->lock)
1834 struct mm_struct *mm = rq->prev_mm;
1840 * A task struct has one reference for the use as "current".
1841 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1842 * schedule one last time. The schedule call will never return, and
1843 * the scheduled task must drop that reference.
1844 * The test for TASK_DEAD must occur while the runqueue locks are
1845 * still held, otherwise prev could be scheduled on another cpu, die
1846 * there before we look at prev->state, and then the reference would
1848 * Manfred Spraul <manfred@colorfullife.com>
1850 prev_state = prev->state;
1851 finish_arch_switch(prev);
1852 finish_lock_switch(rq, prev);
1855 if (unlikely(prev_state == TASK_DEAD)) {
1857 * Remove function-return probe instances associated with this
1858 * task and put them back on the free list.
1860 kprobe_flush_task(prev);
1861 put_task_struct(prev);
1866 * schedule_tail - first thing a freshly forked thread must call.
1867 * @prev: the thread we just switched away from.
1869 asmlinkage void schedule_tail(struct task_struct *prev)
1870 __releases(rq->lock)
1872 struct rq *rq = this_rq();
1874 finish_task_switch(rq, prev);
1875 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1876 /* In this case, finish_task_switch does not reenable preemption */
1879 if (current->set_child_tid)
1880 put_user(current->pid, current->set_child_tid);
1884 * context_switch - switch to the new MM and the new
1885 * thread's register state.
1887 static inline struct task_struct *
1888 context_switch(struct rq *rq, struct task_struct *prev,
1889 struct task_struct *next)
1891 struct mm_struct *mm = next->mm;
1892 struct mm_struct *oldmm = prev->active_mm;
1895 * For paravirt, this is coupled with an exit in switch_to to
1896 * combine the page table reload and the switch backend into
1899 arch_enter_lazy_cpu_mode();
1902 next->active_mm = oldmm;
1903 atomic_inc(&oldmm->mm_count);
1904 enter_lazy_tlb(oldmm, next);
1906 switch_mm(oldmm, mm, next);
1909 prev->active_mm = NULL;
1910 WARN_ON(rq->prev_mm);
1911 rq->prev_mm = oldmm;
1914 * Since the runqueue lock will be released by the next
1915 * task (which is an invalid locking op but in the case
1916 * of the scheduler it's an obvious special-case), so we
1917 * do an early lockdep release here:
1919 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1920 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1923 /* Here we just switch the register state and the stack. */
1924 switch_to(prev, next, prev);
1930 * nr_running, nr_uninterruptible and nr_context_switches:
1932 * externally visible scheduler statistics: current number of runnable
1933 * threads, current number of uninterruptible-sleeping threads, total
1934 * number of context switches performed since bootup.
1936 unsigned long nr_running(void)
1938 unsigned long i, sum = 0;
1940 for_each_online_cpu(i)
1941 sum += cpu_rq(i)->nr_running;
1946 unsigned long nr_uninterruptible(void)
1948 unsigned long i, sum = 0;
1950 for_each_possible_cpu(i)
1951 sum += cpu_rq(i)->nr_uninterruptible;
1954 * Since we read the counters lockless, it might be slightly
1955 * inaccurate. Do not allow it to go below zero though:
1957 if (unlikely((long)sum < 0))
1963 unsigned long long nr_context_switches(void)
1966 unsigned long long sum = 0;
1968 for_each_possible_cpu(i)
1969 sum += cpu_rq(i)->nr_switches;
1974 unsigned long nr_iowait(void)
1976 unsigned long i, sum = 0;
1978 for_each_possible_cpu(i)
1979 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1984 unsigned long nr_active(void)
1986 unsigned long i, running = 0, uninterruptible = 0;
1988 for_each_online_cpu(i) {
1989 running += cpu_rq(i)->nr_running;
1990 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1993 if (unlikely((long)uninterruptible < 0))
1994 uninterruptible = 0;
1996 return running + uninterruptible;
2002 * Is this task likely cache-hot:
2005 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
2007 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
2011 * double_rq_lock - safely lock two runqueues
2013 * Note this does not disable interrupts like task_rq_lock,
2014 * you need to do so manually before calling.
2016 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2017 __acquires(rq1->lock)
2018 __acquires(rq2->lock)
2020 BUG_ON(!irqs_disabled());
2022 spin_lock(&rq1->lock);
2023 __acquire(rq2->lock); /* Fake it out ;) */
2026 spin_lock(&rq1->lock);
2027 spin_lock(&rq2->lock);
2029 spin_lock(&rq2->lock);
2030 spin_lock(&rq1->lock);
2036 * double_rq_unlock - safely unlock two runqueues
2038 * Note this does not restore interrupts like task_rq_unlock,
2039 * you need to do so manually after calling.
2041 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2042 __releases(rq1->lock)
2043 __releases(rq2->lock)
2045 spin_unlock(&rq1->lock);
2047 spin_unlock(&rq2->lock);
2049 __release(rq2->lock);
2053 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2055 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2056 __releases(this_rq->lock)
2057 __acquires(busiest->lock)
2058 __acquires(this_rq->lock)
2060 if (unlikely(!irqs_disabled())) {
2061 /* printk() doesn't work good under rq->lock */
2062 spin_unlock(&this_rq->lock);
2065 if (unlikely(!spin_trylock(&busiest->lock))) {
2066 if (busiest < this_rq) {
2067 spin_unlock(&this_rq->lock);
2068 spin_lock(&busiest->lock);
2069 spin_lock(&this_rq->lock);
2071 spin_lock(&busiest->lock);
2076 * If dest_cpu is allowed for this process, migrate the task to it.
2077 * This is accomplished by forcing the cpu_allowed mask to only
2078 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2079 * the cpu_allowed mask is restored.
2081 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2083 struct migration_req req;
2084 unsigned long flags;
2087 rq = task_rq_lock(p, &flags);
2088 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2089 || unlikely(cpu_is_offline(dest_cpu)))
2092 /* force the process onto the specified CPU */
2093 if (migrate_task(p, dest_cpu, &req)) {
2094 /* Need to wait for migration thread (might exit: take ref). */
2095 struct task_struct *mt = rq->migration_thread;
2097 get_task_struct(mt);
2098 task_rq_unlock(rq, &flags);
2099 wake_up_process(mt);
2100 put_task_struct(mt);
2101 wait_for_completion(&req.done);
2106 task_rq_unlock(rq, &flags);
2110 * sched_exec - execve() is a valuable balancing opportunity, because at
2111 * this point the task has the smallest effective memory and cache footprint.
2113 void sched_exec(void)
2115 int new_cpu, this_cpu = get_cpu();
2116 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2118 if (new_cpu != this_cpu)
2119 sched_migrate_task(current, new_cpu);
2123 * pull_task - move a task from a remote runqueue to the local runqueue.
2124 * Both runqueues must be locked.
2126 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2127 struct task_struct *p, struct rq *this_rq,
2128 struct prio_array *this_array, int this_cpu)
2130 dequeue_task(p, src_array);
2131 dec_nr_running(p, src_rq);
2132 set_task_cpu(p, this_cpu);
2133 inc_nr_running(p, this_rq);
2134 enqueue_task(p, this_array);
2135 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2136 + this_rq->most_recent_timestamp;
2138 * Note that idle threads have a prio of MAX_PRIO, for this test
2139 * to be always true for them.
2141 if (TASK_PREEMPTS_CURR(p, this_rq))
2142 resched_task(this_rq->curr);
2146 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2149 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2150 struct sched_domain *sd, enum idle_type idle,
2154 * We do not migrate tasks that are:
2155 * 1) running (obviously), or
2156 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2157 * 3) are cache-hot on their current CPU.
2159 if (!cpu_isset(this_cpu, p->cpus_allowed))
2163 if (task_running(rq, p))
2167 * Aggressive migration if:
2168 * 1) task is cache cold, or
2169 * 2) too many balance attempts have failed.
2172 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2173 #ifdef CONFIG_SCHEDSTATS
2174 if (task_hot(p, rq->most_recent_timestamp, sd))
2175 schedstat_inc(sd, lb_hot_gained[idle]);
2180 if (task_hot(p, rq->most_recent_timestamp, sd))
2185 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2188 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2189 * load from busiest to this_rq, as part of a balancing operation within
2190 * "domain". Returns the number of tasks moved.
2192 * Called with both runqueues locked.
2194 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2195 unsigned long max_nr_move, unsigned long max_load_move,
2196 struct sched_domain *sd, enum idle_type idle,
2199 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2200 best_prio_seen, skip_for_load;
2201 struct prio_array *array, *dst_array;
2202 struct list_head *head, *curr;
2203 struct task_struct *tmp;
2206 if (max_nr_move == 0 || max_load_move == 0)
2209 rem_load_move = max_load_move;
2211 this_best_prio = rq_best_prio(this_rq);
2212 best_prio = rq_best_prio(busiest);
2214 * Enable handling of the case where there is more than one task
2215 * with the best priority. If the current running task is one
2216 * of those with prio==best_prio we know it won't be moved
2217 * and therefore it's safe to override the skip (based on load) of
2218 * any task we find with that prio.
2220 best_prio_seen = best_prio == busiest->curr->prio;
2223 * We first consider expired tasks. Those will likely not be
2224 * executed in the near future, and they are most likely to
2225 * be cache-cold, thus switching CPUs has the least effect
2228 if (busiest->expired->nr_active) {
2229 array = busiest->expired;
2230 dst_array = this_rq->expired;
2232 array = busiest->active;
2233 dst_array = this_rq->active;
2237 /* Start searching at priority 0: */
2241 idx = sched_find_first_bit(array->bitmap);
2243 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2244 if (idx >= MAX_PRIO) {
2245 if (array == busiest->expired && busiest->active->nr_active) {
2246 array = busiest->active;
2247 dst_array = this_rq->active;
2253 head = array->queue + idx;
2256 tmp = list_entry(curr, struct task_struct, run_list);
2261 * To help distribute high priority tasks accross CPUs we don't
2262 * skip a task if it will be the highest priority task (i.e. smallest
2263 * prio value) on its new queue regardless of its load weight
2265 skip_for_load = tmp->load_weight > rem_load_move;
2266 if (skip_for_load && idx < this_best_prio)
2267 skip_for_load = !best_prio_seen && idx == best_prio;
2268 if (skip_for_load ||
2269 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2271 best_prio_seen |= idx == best_prio;
2278 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2280 rem_load_move -= tmp->load_weight;
2283 * We only want to steal up to the prescribed number of tasks
2284 * and the prescribed amount of weighted load.
2286 if (pulled < max_nr_move && rem_load_move > 0) {
2287 if (idx < this_best_prio)
2288 this_best_prio = idx;
2296 * Right now, this is the only place pull_task() is called,
2297 * so we can safely collect pull_task() stats here rather than
2298 * inside pull_task().
2300 schedstat_add(sd, lb_gained[idle], pulled);
2303 *all_pinned = pinned;
2308 * find_busiest_group finds and returns the busiest CPU group within the
2309 * domain. It calculates and returns the amount of weighted load which
2310 * should be moved to restore balance via the imbalance parameter.
2312 static struct sched_group *
2313 find_busiest_group(struct sched_domain *sd, int this_cpu,
2314 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2315 cpumask_t *cpus, int *balance)
2317 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2318 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2319 unsigned long max_pull;
2320 unsigned long busiest_load_per_task, busiest_nr_running;
2321 unsigned long this_load_per_task, this_nr_running;
2323 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2324 int power_savings_balance = 1;
2325 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2326 unsigned long min_nr_running = ULONG_MAX;
2327 struct sched_group *group_min = NULL, *group_leader = NULL;
2330 max_load = this_load = total_load = total_pwr = 0;
2331 busiest_load_per_task = busiest_nr_running = 0;
2332 this_load_per_task = this_nr_running = 0;
2333 if (idle == NOT_IDLE)
2334 load_idx = sd->busy_idx;
2335 else if (idle == NEWLY_IDLE)
2336 load_idx = sd->newidle_idx;
2338 load_idx = sd->idle_idx;
2341 unsigned long load, group_capacity;
2344 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2345 unsigned long sum_nr_running, sum_weighted_load;
2347 local_group = cpu_isset(this_cpu, group->cpumask);
2350 balance_cpu = first_cpu(group->cpumask);
2352 /* Tally up the load of all CPUs in the group */
2353 sum_weighted_load = sum_nr_running = avg_load = 0;
2355 for_each_cpu_mask(i, group->cpumask) {
2358 if (!cpu_isset(i, *cpus))
2363 if (*sd_idle && !idle_cpu(i))
2366 /* Bias balancing toward cpus of our domain */
2368 if (idle_cpu(i) && !first_idle_cpu) {
2373 load = target_load(i, load_idx);
2375 load = source_load(i, load_idx);
2378 sum_nr_running += rq->nr_running;
2379 sum_weighted_load += rq->raw_weighted_load;
2383 * First idle cpu or the first cpu(busiest) in this sched group
2384 * is eligible for doing load balancing at this and above
2387 if (local_group && balance_cpu != this_cpu && balance) {
2392 total_load += avg_load;
2393 total_pwr += group->__cpu_power;
2395 /* Adjust by relative CPU power of the group */
2396 avg_load = sg_div_cpu_power(group,
2397 avg_load * SCHED_LOAD_SCALE);
2399 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2402 this_load = avg_load;
2404 this_nr_running = sum_nr_running;
2405 this_load_per_task = sum_weighted_load;
2406 } else if (avg_load > max_load &&
2407 sum_nr_running > group_capacity) {
2408 max_load = avg_load;
2410 busiest_nr_running = sum_nr_running;
2411 busiest_load_per_task = sum_weighted_load;
2414 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2416 * Busy processors will not participate in power savings
2419 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2423 * If the local group is idle or completely loaded
2424 * no need to do power savings balance at this domain
2426 if (local_group && (this_nr_running >= group_capacity ||
2428 power_savings_balance = 0;
2431 * If a group is already running at full capacity or idle,
2432 * don't include that group in power savings calculations
2434 if (!power_savings_balance || sum_nr_running >= group_capacity
2439 * Calculate the group which has the least non-idle load.
2440 * This is the group from where we need to pick up the load
2443 if ((sum_nr_running < min_nr_running) ||
2444 (sum_nr_running == min_nr_running &&
2445 first_cpu(group->cpumask) <
2446 first_cpu(group_min->cpumask))) {
2448 min_nr_running = sum_nr_running;
2449 min_load_per_task = sum_weighted_load /
2454 * Calculate the group which is almost near its
2455 * capacity but still has some space to pick up some load
2456 * from other group and save more power
2458 if (sum_nr_running <= group_capacity - 1) {
2459 if (sum_nr_running > leader_nr_running ||
2460 (sum_nr_running == leader_nr_running &&
2461 first_cpu(group->cpumask) >
2462 first_cpu(group_leader->cpumask))) {
2463 group_leader = group;
2464 leader_nr_running = sum_nr_running;
2469 group = group->next;
2470 } while (group != sd->groups);
2472 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2475 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2477 if (this_load >= avg_load ||
2478 100*max_load <= sd->imbalance_pct*this_load)
2481 busiest_load_per_task /= busiest_nr_running;
2483 * We're trying to get all the cpus to the average_load, so we don't
2484 * want to push ourselves above the average load, nor do we wish to
2485 * reduce the max loaded cpu below the average load, as either of these
2486 * actions would just result in more rebalancing later, and ping-pong
2487 * tasks around. Thus we look for the minimum possible imbalance.
2488 * Negative imbalances (*we* are more loaded than anyone else) will
2489 * be counted as no imbalance for these purposes -- we can't fix that
2490 * by pulling tasks to us. Be careful of negative numbers as they'll
2491 * appear as very large values with unsigned longs.
2493 if (max_load <= busiest_load_per_task)
2497 * In the presence of smp nice balancing, certain scenarios can have
2498 * max load less than avg load(as we skip the groups at or below
2499 * its cpu_power, while calculating max_load..)
2501 if (max_load < avg_load) {
2503 goto small_imbalance;
2506 /* Don't want to pull so many tasks that a group would go idle */
2507 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2509 /* How much load to actually move to equalise the imbalance */
2510 *imbalance = min(max_pull * busiest->__cpu_power,
2511 (avg_load - this_load) * this->__cpu_power)
2515 * if *imbalance is less than the average load per runnable task
2516 * there is no gaurantee that any tasks will be moved so we'll have
2517 * a think about bumping its value to force at least one task to be
2520 if (*imbalance < busiest_load_per_task) {
2521 unsigned long tmp, pwr_now, pwr_move;
2525 pwr_move = pwr_now = 0;
2527 if (this_nr_running) {
2528 this_load_per_task /= this_nr_running;
2529 if (busiest_load_per_task > this_load_per_task)
2532 this_load_per_task = SCHED_LOAD_SCALE;
2534 if (max_load - this_load >= busiest_load_per_task * imbn) {
2535 *imbalance = busiest_load_per_task;
2540 * OK, we don't have enough imbalance to justify moving tasks,
2541 * however we may be able to increase total CPU power used by
2545 pwr_now += busiest->__cpu_power *
2546 min(busiest_load_per_task, max_load);
2547 pwr_now += this->__cpu_power *
2548 min(this_load_per_task, this_load);
2549 pwr_now /= SCHED_LOAD_SCALE;
2551 /* Amount of load we'd subtract */
2552 tmp = sg_div_cpu_power(busiest,
2553 busiest_load_per_task * SCHED_LOAD_SCALE);
2555 pwr_move += busiest->__cpu_power *
2556 min(busiest_load_per_task, max_load - tmp);
2558 /* Amount of load we'd add */
2559 if (max_load * busiest->__cpu_power <
2560 busiest_load_per_task * SCHED_LOAD_SCALE)
2561 tmp = sg_div_cpu_power(this,
2562 max_load * busiest->__cpu_power);
2564 tmp = sg_div_cpu_power(this,
2565 busiest_load_per_task * SCHED_LOAD_SCALE);
2566 pwr_move += this->__cpu_power *
2567 min(this_load_per_task, this_load + tmp);
2568 pwr_move /= SCHED_LOAD_SCALE;
2570 /* Move if we gain throughput */
2571 if (pwr_move <= pwr_now)
2574 *imbalance = busiest_load_per_task;
2580 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2581 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2584 if (this == group_leader && group_leader != group_min) {
2585 *imbalance = min_load_per_task;
2595 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2598 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2599 unsigned long imbalance, cpumask_t *cpus)
2601 struct rq *busiest = NULL, *rq;
2602 unsigned long max_load = 0;
2605 for_each_cpu_mask(i, group->cpumask) {
2607 if (!cpu_isset(i, *cpus))
2612 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2615 if (rq->raw_weighted_load > max_load) {
2616 max_load = rq->raw_weighted_load;
2625 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2626 * so long as it is large enough.
2628 #define MAX_PINNED_INTERVAL 512
2630 static inline unsigned long minus_1_or_zero(unsigned long n)
2632 return n > 0 ? n - 1 : 0;
2636 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2637 * tasks if there is an imbalance.
2639 static int load_balance(int this_cpu, struct rq *this_rq,
2640 struct sched_domain *sd, enum idle_type idle,
2643 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2644 struct sched_group *group;
2645 unsigned long imbalance;
2647 cpumask_t cpus = CPU_MASK_ALL;
2648 unsigned long flags;
2651 * When power savings policy is enabled for the parent domain, idle
2652 * sibling can pick up load irrespective of busy siblings. In this case,
2653 * let the state of idle sibling percolate up as IDLE, instead of
2654 * portraying it as NOT_IDLE.
2656 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2657 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2660 schedstat_inc(sd, lb_cnt[idle]);
2663 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2670 schedstat_inc(sd, lb_nobusyg[idle]);
2674 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2676 schedstat_inc(sd, lb_nobusyq[idle]);
2680 BUG_ON(busiest == this_rq);
2682 schedstat_add(sd, lb_imbalance[idle], imbalance);
2685 if (busiest->nr_running > 1) {
2687 * Attempt to move tasks. If find_busiest_group has found
2688 * an imbalance but busiest->nr_running <= 1, the group is
2689 * still unbalanced. nr_moved simply stays zero, so it is
2690 * correctly treated as an imbalance.
2692 local_irq_save(flags);
2693 double_rq_lock(this_rq, busiest);
2694 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2695 minus_1_or_zero(busiest->nr_running),
2696 imbalance, sd, idle, &all_pinned);
2697 double_rq_unlock(this_rq, busiest);
2698 local_irq_restore(flags);
2701 * some other cpu did the load balance for us.
2703 if (nr_moved && this_cpu != smp_processor_id())
2704 resched_cpu(this_cpu);
2706 /* All tasks on this runqueue were pinned by CPU affinity */
2707 if (unlikely(all_pinned)) {
2708 cpu_clear(cpu_of(busiest), cpus);
2709 if (!cpus_empty(cpus))
2716 schedstat_inc(sd, lb_failed[idle]);
2717 sd->nr_balance_failed++;
2719 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2721 spin_lock_irqsave(&busiest->lock, flags);
2723 /* don't kick the migration_thread, if the curr
2724 * task on busiest cpu can't be moved to this_cpu
2726 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2727 spin_unlock_irqrestore(&busiest->lock, flags);
2729 goto out_one_pinned;
2732 if (!busiest->active_balance) {
2733 busiest->active_balance = 1;
2734 busiest->push_cpu = this_cpu;
2737 spin_unlock_irqrestore(&busiest->lock, flags);
2739 wake_up_process(busiest->migration_thread);
2742 * We've kicked active balancing, reset the failure
2745 sd->nr_balance_failed = sd->cache_nice_tries+1;
2748 sd->nr_balance_failed = 0;
2750 if (likely(!active_balance)) {
2751 /* We were unbalanced, so reset the balancing interval */
2752 sd->balance_interval = sd->min_interval;
2755 * If we've begun active balancing, start to back off. This
2756 * case may not be covered by the all_pinned logic if there
2757 * is only 1 task on the busy runqueue (because we don't call
2760 if (sd->balance_interval < sd->max_interval)
2761 sd->balance_interval *= 2;
2764 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2765 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2770 schedstat_inc(sd, lb_balanced[idle]);
2772 sd->nr_balance_failed = 0;
2775 /* tune up the balancing interval */
2776 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2777 (sd->balance_interval < sd->max_interval))
2778 sd->balance_interval *= 2;
2780 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2781 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2787 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2788 * tasks if there is an imbalance.
2790 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2791 * this_rq is locked.
2794 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2796 struct sched_group *group;
2797 struct rq *busiest = NULL;
2798 unsigned long imbalance;
2801 cpumask_t cpus = CPU_MASK_ALL;
2804 * When power savings policy is enabled for the parent domain, idle
2805 * sibling can pick up load irrespective of busy siblings. In this case,
2806 * let the state of idle sibling percolate up as IDLE, instead of
2807 * portraying it as NOT_IDLE.
2809 if (sd->flags & SD_SHARE_CPUPOWER &&
2810 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2813 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2815 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2816 &sd_idle, &cpus, NULL);
2818 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2822 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2825 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2829 BUG_ON(busiest == this_rq);
2831 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2834 if (busiest->nr_running > 1) {
2835 /* Attempt to move tasks */
2836 double_lock_balance(this_rq, busiest);
2837 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2838 minus_1_or_zero(busiest->nr_running),
2839 imbalance, sd, NEWLY_IDLE, NULL);
2840 spin_unlock(&busiest->lock);
2843 cpu_clear(cpu_of(busiest), cpus);
2844 if (!cpus_empty(cpus))
2850 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2851 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2852 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2855 sd->nr_balance_failed = 0;
2860 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2861 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2862 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2864 sd->nr_balance_failed = 0;
2870 * idle_balance is called by schedule() if this_cpu is about to become
2871 * idle. Attempts to pull tasks from other CPUs.
2873 static void idle_balance(int this_cpu, struct rq *this_rq)
2875 struct sched_domain *sd;
2876 int pulled_task = 0;
2877 unsigned long next_balance = jiffies + 60 * HZ;
2879 for_each_domain(this_cpu, sd) {
2880 if (sd->flags & SD_BALANCE_NEWIDLE) {
2881 /* If we've pulled tasks over stop searching: */
2882 pulled_task = load_balance_newidle(this_cpu,
2884 if (time_after(next_balance,
2885 sd->last_balance + sd->balance_interval))
2886 next_balance = sd->last_balance
2887 + sd->balance_interval;
2894 * We are going idle. next_balance may be set based on
2895 * a busy processor. So reset next_balance.
2897 this_rq->next_balance = next_balance;
2901 * active_load_balance is run by migration threads. It pushes running tasks
2902 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2903 * running on each physical CPU where possible, and avoids physical /
2904 * logical imbalances.
2906 * Called with busiest_rq locked.
2908 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2910 int target_cpu = busiest_rq->push_cpu;
2911 struct sched_domain *sd;
2912 struct rq *target_rq;
2914 /* Is there any task to move? */
2915 if (busiest_rq->nr_running <= 1)
2918 target_rq = cpu_rq(target_cpu);
2921 * This condition is "impossible", if it occurs
2922 * we need to fix it. Originally reported by
2923 * Bjorn Helgaas on a 128-cpu setup.
2925 BUG_ON(busiest_rq == target_rq);
2927 /* move a task from busiest_rq to target_rq */
2928 double_lock_balance(busiest_rq, target_rq);
2930 /* Search for an sd spanning us and the target CPU. */
2931 for_each_domain(target_cpu, sd) {
2932 if ((sd->flags & SD_LOAD_BALANCE) &&
2933 cpu_isset(busiest_cpu, sd->span))
2938 schedstat_inc(sd, alb_cnt);
2940 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2941 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2943 schedstat_inc(sd, alb_pushed);
2945 schedstat_inc(sd, alb_failed);
2947 spin_unlock(&target_rq->lock);
2950 static void update_load(struct rq *this_rq)
2952 unsigned long this_load;
2953 unsigned int i, scale;
2955 this_load = this_rq->raw_weighted_load;
2957 /* Update our load: */
2958 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2959 unsigned long old_load, new_load;
2961 /* scale is effectively 1 << i now, and >> i divides by scale */
2963 old_load = this_rq->cpu_load[i];
2964 new_load = this_load;
2966 * Round up the averaging division if load is increasing. This
2967 * prevents us from getting stuck on 9 if the load is 10, for
2970 if (new_load > old_load)
2971 new_load += scale-1;
2972 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2978 atomic_t load_balancer;
2980 } nohz ____cacheline_aligned = {
2981 .load_balancer = ATOMIC_INIT(-1),
2982 .cpu_mask = CPU_MASK_NONE,
2986 * This routine will try to nominate the ilb (idle load balancing)
2987 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2988 * load balancing on behalf of all those cpus. If all the cpus in the system
2989 * go into this tickless mode, then there will be no ilb owner (as there is
2990 * no need for one) and all the cpus will sleep till the next wakeup event
2993 * For the ilb owner, tick is not stopped. And this tick will be used
2994 * for idle load balancing. ilb owner will still be part of
2997 * While stopping the tick, this cpu will become the ilb owner if there
2998 * is no other owner. And will be the owner till that cpu becomes busy
2999 * or if all cpus in the system stop their ticks at which point
3000 * there is no need for ilb owner.
3002 * When the ilb owner becomes busy, it nominates another owner, during the
3003 * next busy scheduler_tick()
3005 int select_nohz_load_balancer(int stop_tick)
3007 int cpu = smp_processor_id();
3010 cpu_set(cpu, nohz.cpu_mask);
3011 cpu_rq(cpu)->in_nohz_recently = 1;
3014 * If we are going offline and still the leader, give up!
3016 if (cpu_is_offline(cpu) &&
3017 atomic_read(&nohz.load_balancer) == cpu) {
3018 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3023 /* time for ilb owner also to sleep */
3024 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3025 if (atomic_read(&nohz.load_balancer) == cpu)
3026 atomic_set(&nohz.load_balancer, -1);
3030 if (atomic_read(&nohz.load_balancer) == -1) {
3031 /* make me the ilb owner */
3032 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3034 } else if (atomic_read(&nohz.load_balancer) == cpu)
3037 if (!cpu_isset(cpu, nohz.cpu_mask))
3040 cpu_clear(cpu, nohz.cpu_mask);
3042 if (atomic_read(&nohz.load_balancer) == cpu)
3043 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3050 static DEFINE_SPINLOCK(balancing);
3053 * It checks each scheduling domain to see if it is due to be balanced,
3054 * and initiates a balancing operation if so.
3056 * Balancing parameters are set up in arch_init_sched_domains.
3058 static inline void rebalance_domains(int cpu, enum idle_type idle)
3061 struct rq *rq = cpu_rq(cpu);
3062 unsigned long interval;
3063 struct sched_domain *sd;
3064 /* Earliest time when we have to do rebalance again */
3065 unsigned long next_balance = jiffies + 60*HZ;
3067 for_each_domain(cpu, sd) {
3068 if (!(sd->flags & SD_LOAD_BALANCE))
3071 interval = sd->balance_interval;
3072 if (idle != SCHED_IDLE)
3073 interval *= sd->busy_factor;
3075 /* scale ms to jiffies */
3076 interval = msecs_to_jiffies(interval);
3077 if (unlikely(!interval))
3080 if (sd->flags & SD_SERIALIZE) {
3081 if (!spin_trylock(&balancing))
3085 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3086 if (load_balance(cpu, rq, sd, idle, &balance)) {
3088 * We've pulled tasks over so either we're no
3089 * longer idle, or one of our SMT siblings is
3094 sd->last_balance = jiffies;
3096 if (sd->flags & SD_SERIALIZE)
3097 spin_unlock(&balancing);
3099 if (time_after(next_balance, sd->last_balance + interval))
3100 next_balance = sd->last_balance + interval;
3103 * Stop the load balance at this level. There is another
3104 * CPU in our sched group which is doing load balancing more
3110 rq->next_balance = next_balance;
3114 * run_rebalance_domains is triggered when needed from the scheduler tick.
3115 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3116 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3118 static void run_rebalance_domains(struct softirq_action *h)
3120 int local_cpu = smp_processor_id();
3121 struct rq *local_rq = cpu_rq(local_cpu);
3122 enum idle_type idle = local_rq->idle_at_tick ? SCHED_IDLE : NOT_IDLE;
3124 rebalance_domains(local_cpu, idle);
3128 * If this cpu is the owner for idle load balancing, then do the
3129 * balancing on behalf of the other idle cpus whose ticks are
3132 if (local_rq->idle_at_tick &&
3133 atomic_read(&nohz.load_balancer) == local_cpu) {
3134 cpumask_t cpus = nohz.cpu_mask;
3138 cpu_clear(local_cpu, cpus);
3139 for_each_cpu_mask(balance_cpu, cpus) {
3141 * If this cpu gets work to do, stop the load balancing
3142 * work being done for other cpus. Next load
3143 * balancing owner will pick it up.
3148 rebalance_domains(balance_cpu, SCHED_IDLE);
3150 rq = cpu_rq(balance_cpu);
3151 if (time_after(local_rq->next_balance, rq->next_balance))
3152 local_rq->next_balance = rq->next_balance;
3159 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3161 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3162 * idle load balancing owner or decide to stop the periodic load balancing,
3163 * if the whole system is idle.
3165 static inline void trigger_load_balance(int cpu)
3167 struct rq *rq = cpu_rq(cpu);
3170 * If we were in the nohz mode recently and busy at the current
3171 * scheduler tick, then check if we need to nominate new idle
3174 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3175 rq->in_nohz_recently = 0;
3177 if (atomic_read(&nohz.load_balancer) == cpu) {
3178 cpu_clear(cpu, nohz.cpu_mask);
3179 atomic_set(&nohz.load_balancer, -1);
3182 if (atomic_read(&nohz.load_balancer) == -1) {
3184 * simple selection for now: Nominate the
3185 * first cpu in the nohz list to be the next
3188 * TBD: Traverse the sched domains and nominate
3189 * the nearest cpu in the nohz.cpu_mask.
3191 int ilb = first_cpu(nohz.cpu_mask);
3199 * If this cpu is idle and doing idle load balancing for all the
3200 * cpus with ticks stopped, is it time for that to stop?
3202 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3203 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3209 * If this cpu is idle and the idle load balancing is done by
3210 * someone else, then no need raise the SCHED_SOFTIRQ
3212 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3213 cpu_isset(cpu, nohz.cpu_mask))
3216 if (time_after_eq(jiffies, rq->next_balance))
3217 raise_softirq(SCHED_SOFTIRQ);
3221 * on UP we do not need to balance between CPUs:
3223 static inline void idle_balance(int cpu, struct rq *rq)
3228 DEFINE_PER_CPU(struct kernel_stat, kstat);
3230 EXPORT_PER_CPU_SYMBOL(kstat);
3233 * This is called on clock ticks and on context switches.
3234 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3237 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3239 p->sched_time += now - p->last_ran;
3240 p->last_ran = rq->most_recent_timestamp = now;
3244 * Return current->sched_time plus any more ns on the sched_clock
3245 * that have not yet been banked.
3247 unsigned long long current_sched_time(const struct task_struct *p)
3249 unsigned long long ns;
3250 unsigned long flags;
3252 local_irq_save(flags);
3253 ns = p->sched_time + sched_clock() - p->last_ran;
3254 local_irq_restore(flags);
3260 * We place interactive tasks back into the active array, if possible.
3262 * To guarantee that this does not starve expired tasks we ignore the
3263 * interactivity of a task if the first expired task had to wait more
3264 * than a 'reasonable' amount of time. This deadline timeout is
3265 * load-dependent, as the frequency of array switched decreases with
3266 * increasing number of running tasks. We also ignore the interactivity
3267 * if a better static_prio task has expired:
3269 static inline int expired_starving(struct rq *rq)
3271 if (rq->curr->static_prio > rq->best_expired_prio)
3273 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3275 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3281 * Account user cpu time to a process.
3282 * @p: the process that the cpu time gets accounted to
3283 * @hardirq_offset: the offset to subtract from hardirq_count()
3284 * @cputime: the cpu time spent in user space since the last update
3286 void account_user_time(struct task_struct *p, cputime_t cputime)
3288 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3291 p->utime = cputime_add(p->utime, cputime);
3293 /* Add user time to cpustat. */
3294 tmp = cputime_to_cputime64(cputime);
3295 if (TASK_NICE(p) > 0)
3296 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3298 cpustat->user = cputime64_add(cpustat->user, tmp);
3302 * Account system cpu time to a process.
3303 * @p: the process that the cpu time gets accounted to
3304 * @hardirq_offset: the offset to subtract from hardirq_count()
3305 * @cputime: the cpu time spent in kernel space since the last update
3307 void account_system_time(struct task_struct *p, int hardirq_offset,
3310 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3311 struct rq *rq = this_rq();
3314 p->stime = cputime_add(p->stime, cputime);
3316 /* Add system time to cpustat. */
3317 tmp = cputime_to_cputime64(cputime);
3318 if (hardirq_count() - hardirq_offset)
3319 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3320 else if (softirq_count())
3321 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3322 else if (p != rq->idle)
3323 cpustat->system = cputime64_add(cpustat->system, tmp);
3324 else if (atomic_read(&rq->nr_iowait) > 0)
3325 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3327 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3328 /* Account for system time used */
3329 acct_update_integrals(p);
3333 * Account for involuntary wait time.
3334 * @p: the process from which the cpu time has been stolen
3335 * @steal: the cpu time spent in involuntary wait
3337 void account_steal_time(struct task_struct *p, cputime_t steal)
3339 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3340 cputime64_t tmp = cputime_to_cputime64(steal);
3341 struct rq *rq = this_rq();
3343 if (p == rq->idle) {
3344 p->stime = cputime_add(p->stime, steal);
3345 if (atomic_read(&rq->nr_iowait) > 0)
3346 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3348 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3350 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3353 static void task_running_tick(struct rq *rq, struct task_struct *p)
3355 if (p->array != rq->active) {
3356 /* Task has expired but was not scheduled yet */
3357 set_tsk_need_resched(p);
3360 spin_lock(&rq->lock);
3362 * The task was running during this tick - update the
3363 * time slice counter. Note: we do not update a thread's
3364 * priority until it either goes to sleep or uses up its
3365 * timeslice. This makes it possible for interactive tasks
3366 * to use up their timeslices at their highest priority levels.
3370 * RR tasks need a special form of timeslice management.
3371 * FIFO tasks have no timeslices.
3373 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3374 p->time_slice = task_timeslice(p);
3375 p->first_time_slice = 0;
3376 set_tsk_need_resched(p);
3378 /* put it at the end of the queue: */
3379 requeue_task(p, rq->active);
3383 if (!--p->time_slice) {
3384 dequeue_task(p, rq->active);
3385 set_tsk_need_resched(p);
3386 p->prio = effective_prio(p);
3387 p->time_slice = task_timeslice(p);
3388 p->first_time_slice = 0;
3390 if (!rq->expired_timestamp)
3391 rq->expired_timestamp = jiffies;
3392 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3393 enqueue_task(p, rq->expired);
3394 if (p->static_prio < rq->best_expired_prio)
3395 rq->best_expired_prio = p->static_prio;
3397 enqueue_task(p, rq->active);
3400 * Prevent a too long timeslice allowing a task to monopolize
3401 * the CPU. We do this by splitting up the timeslice into
3404 * Note: this does not mean the task's timeslices expire or
3405 * get lost in any way, they just might be preempted by
3406 * another task of equal priority. (one with higher
3407 * priority would have preempted this task already.) We
3408 * requeue this task to the end of the list on this priority
3409 * level, which is in essence a round-robin of tasks with
3412 * This only applies to tasks in the interactive
3413 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3415 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3416 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3417 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3418 (p->array == rq->active)) {
3420 requeue_task(p, rq->active);
3421 set_tsk_need_resched(p);
3425 spin_unlock(&rq->lock);
3429 * This function gets called by the timer code, with HZ frequency.
3430 * We call it with interrupts disabled.
3432 * It also gets called by the fork code, when changing the parent's
3435 void scheduler_tick(void)
3437 unsigned long long now = sched_clock();
3438 struct task_struct *p = current;
3439 int cpu = smp_processor_id();
3440 int idle_at_tick = idle_cpu(cpu);
3441 struct rq *rq = cpu_rq(cpu);
3443 update_cpu_clock(p, rq, now);
3446 task_running_tick(rq, p);
3449 rq->idle_at_tick = idle_at_tick;
3450 trigger_load_balance(cpu);
3454 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3456 void fastcall add_preempt_count(int val)
3461 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3463 preempt_count() += val;
3465 * Spinlock count overflowing soon?
3467 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3470 EXPORT_SYMBOL(add_preempt_count);
3472 void fastcall sub_preempt_count(int val)
3477 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3480 * Is the spinlock portion underflowing?
3482 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3483 !(preempt_count() & PREEMPT_MASK)))
3486 preempt_count() -= val;
3488 EXPORT_SYMBOL(sub_preempt_count);
3492 static inline int interactive_sleep(enum sleep_type sleep_type)
3494 return (sleep_type == SLEEP_INTERACTIVE ||
3495 sleep_type == SLEEP_INTERRUPTED);
3499 * schedule() is the main scheduler function.
3501 asmlinkage void __sched schedule(void)
3503 struct task_struct *prev, *next;
3504 struct prio_array *array;
3505 struct list_head *queue;
3506 unsigned long long now;
3507 unsigned long run_time;
3508 int cpu, idx, new_prio;
3513 * Test if we are atomic. Since do_exit() needs to call into
3514 * schedule() atomically, we ignore that path for now.
3515 * Otherwise, whine if we are scheduling when we should not be.
3517 if (unlikely(in_atomic() && !current->exit_state)) {
3518 printk(KERN_ERR "BUG: scheduling while atomic: "
3520 current->comm, preempt_count(), current->pid);
3521 debug_show_held_locks(current);
3522 if (irqs_disabled())
3523 print_irqtrace_events(current);
3526 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3531 release_kernel_lock(prev);
3532 need_resched_nonpreemptible:
3536 * The idle thread is not allowed to schedule!
3537 * Remove this check after it has been exercised a bit.
3539 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3540 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3544 schedstat_inc(rq, sched_cnt);
3545 now = sched_clock();
3546 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3547 run_time = now - prev->timestamp;
3548 if (unlikely((long long)(now - prev->timestamp) < 0))
3551 run_time = NS_MAX_SLEEP_AVG;
3554 * Tasks charged proportionately less run_time at high sleep_avg to
3555 * delay them losing their interactive status
3557 run_time /= (CURRENT_BONUS(prev) ? : 1);
3559 spin_lock_irq(&rq->lock);
3561 switch_count = &prev->nivcsw;
3562 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3563 switch_count = &prev->nvcsw;
3564 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3565 unlikely(signal_pending(prev))))
3566 prev->state = TASK_RUNNING;
3568 if (prev->state == TASK_UNINTERRUPTIBLE)
3569 rq->nr_uninterruptible++;
3570 deactivate_task(prev, rq);
3574 cpu = smp_processor_id();
3575 if (unlikely(!rq->nr_running)) {
3576 idle_balance(cpu, rq);
3577 if (!rq->nr_running) {
3579 rq->expired_timestamp = 0;
3585 if (unlikely(!array->nr_active)) {
3587 * Switch the active and expired arrays.
3589 schedstat_inc(rq, sched_switch);
3590 rq->active = rq->expired;
3591 rq->expired = array;
3593 rq->expired_timestamp = 0;
3594 rq->best_expired_prio = MAX_PRIO;
3597 idx = sched_find_first_bit(array->bitmap);
3598 queue = array->queue + idx;
3599 next = list_entry(queue->next, struct task_struct, run_list);
3601 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3602 unsigned long long delta = now - next->timestamp;
3603 if (unlikely((long long)(now - next->timestamp) < 0))
3606 if (next->sleep_type == SLEEP_INTERACTIVE)
3607 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3609 array = next->array;
3610 new_prio = recalc_task_prio(next, next->timestamp + delta);
3612 if (unlikely(next->prio != new_prio)) {
3613 dequeue_task(next, array);
3614 next->prio = new_prio;
3615 enqueue_task(next, array);
3618 next->sleep_type = SLEEP_NORMAL;
3620 if (next == rq->idle)
3621 schedstat_inc(rq, sched_goidle);
3623 prefetch_stack(next);
3624 clear_tsk_need_resched(prev);
3625 rcu_qsctr_inc(task_cpu(prev));
3627 update_cpu_clock(prev, rq, now);
3629 prev->sleep_avg -= run_time;
3630 if ((long)prev->sleep_avg <= 0)
3631 prev->sleep_avg = 0;
3632 prev->timestamp = prev->last_ran = now;
3634 sched_info_switch(prev, next);
3635 if (likely(prev != next)) {
3636 next->timestamp = next->last_ran = now;
3641 prepare_task_switch(rq, next);
3642 prev = context_switch(rq, prev, next);
3645 * this_rq must be evaluated again because prev may have moved
3646 * CPUs since it called schedule(), thus the 'rq' on its stack
3647 * frame will be invalid.
3649 finish_task_switch(this_rq(), prev);
3651 spin_unlock_irq(&rq->lock);
3654 if (unlikely(reacquire_kernel_lock(prev) < 0))
3655 goto need_resched_nonpreemptible;
3656 preempt_enable_no_resched();
3657 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3660 EXPORT_SYMBOL(schedule);
3662 #ifdef CONFIG_PREEMPT
3664 * this is the entry point to schedule() from in-kernel preemption
3665 * off of preempt_enable. Kernel preemptions off return from interrupt
3666 * occur there and call schedule directly.
3668 asmlinkage void __sched preempt_schedule(void)
3670 struct thread_info *ti = current_thread_info();
3671 #ifdef CONFIG_PREEMPT_BKL
3672 struct task_struct *task = current;
3673 int saved_lock_depth;
3676 * If there is a non-zero preempt_count or interrupts are disabled,
3677 * we do not want to preempt the current task. Just return..
3679 if (likely(ti->preempt_count || irqs_disabled()))
3683 add_preempt_count(PREEMPT_ACTIVE);
3685 * We keep the big kernel semaphore locked, but we
3686 * clear ->lock_depth so that schedule() doesnt
3687 * auto-release the semaphore:
3689 #ifdef CONFIG_PREEMPT_BKL
3690 saved_lock_depth = task->lock_depth;
3691 task->lock_depth = -1;
3694 #ifdef CONFIG_PREEMPT_BKL
3695 task->lock_depth = saved_lock_depth;
3697 sub_preempt_count(PREEMPT_ACTIVE);
3699 /* we could miss a preemption opportunity between schedule and now */
3701 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3704 EXPORT_SYMBOL(preempt_schedule);
3707 * this is the entry point to schedule() from kernel preemption
3708 * off of irq context.
3709 * Note, that this is called and return with irqs disabled. This will
3710 * protect us against recursive calling from irq.
3712 asmlinkage void __sched preempt_schedule_irq(void)
3714 struct thread_info *ti = current_thread_info();
3715 #ifdef CONFIG_PREEMPT_BKL
3716 struct task_struct *task = current;
3717 int saved_lock_depth;
3719 /* Catch callers which need to be fixed */
3720 BUG_ON(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;
3735 local_irq_disable();
3736 #ifdef CONFIG_PREEMPT_BKL
3737 task->lock_depth = saved_lock_depth;
3739 sub_preempt_count(PREEMPT_ACTIVE);
3741 /* we could miss a preemption opportunity between schedule and now */
3743 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3747 #endif /* CONFIG_PREEMPT */
3749 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3752 return try_to_wake_up(curr->private, mode, sync);
3754 EXPORT_SYMBOL(default_wake_function);
3757 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3758 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3759 * number) then we wake all the non-exclusive tasks and one exclusive task.
3761 * There are circumstances in which we can try to wake a task which has already
3762 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3763 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3765 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3766 int nr_exclusive, int sync, void *key)
3768 struct list_head *tmp, *next;
3770 list_for_each_safe(tmp, next, &q->task_list) {
3771 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3772 unsigned flags = curr->flags;
3774 if (curr->func(curr, mode, sync, key) &&
3775 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3781 * __wake_up - wake up threads blocked on a waitqueue.
3783 * @mode: which threads
3784 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3785 * @key: is directly passed to the wakeup function
3787 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3788 int nr_exclusive, void *key)
3790 unsigned long flags;
3792 spin_lock_irqsave(&q->lock, flags);
3793 __wake_up_common(q, mode, nr_exclusive, 0, key);
3794 spin_unlock_irqrestore(&q->lock, flags);
3796 EXPORT_SYMBOL(__wake_up);
3799 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3801 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3803 __wake_up_common(q, mode, 1, 0, NULL);
3807 * __wake_up_sync - wake up threads blocked on a waitqueue.
3809 * @mode: which threads
3810 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3812 * The sync wakeup differs that the waker knows that it will schedule
3813 * away soon, so while the target thread will be woken up, it will not
3814 * be migrated to another CPU - ie. the two threads are 'synchronized'
3815 * with each other. This can prevent needless bouncing between CPUs.
3817 * On UP it can prevent extra preemption.
3820 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3822 unsigned long flags;
3828 if (unlikely(!nr_exclusive))
3831 spin_lock_irqsave(&q->lock, flags);
3832 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3833 spin_unlock_irqrestore(&q->lock, flags);
3835 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3837 void fastcall complete(struct completion *x)
3839 unsigned long flags;
3841 spin_lock_irqsave(&x->wait.lock, flags);
3843 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3845 spin_unlock_irqrestore(&x->wait.lock, flags);
3847 EXPORT_SYMBOL(complete);
3849 void fastcall complete_all(struct completion *x)
3851 unsigned long flags;
3853 spin_lock_irqsave(&x->wait.lock, flags);
3854 x->done += UINT_MAX/2;
3855 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3857 spin_unlock_irqrestore(&x->wait.lock, flags);
3859 EXPORT_SYMBOL(complete_all);
3861 void fastcall __sched wait_for_completion(struct completion *x)
3865 spin_lock_irq(&x->wait.lock);
3867 DECLARE_WAITQUEUE(wait, current);
3869 wait.flags |= WQ_FLAG_EXCLUSIVE;
3870 __add_wait_queue_tail(&x->wait, &wait);
3872 __set_current_state(TASK_UNINTERRUPTIBLE);
3873 spin_unlock_irq(&x->wait.lock);
3875 spin_lock_irq(&x->wait.lock);
3877 __remove_wait_queue(&x->wait, &wait);
3880 spin_unlock_irq(&x->wait.lock);
3882 EXPORT_SYMBOL(wait_for_completion);
3884 unsigned long fastcall __sched
3885 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3889 spin_lock_irq(&x->wait.lock);
3891 DECLARE_WAITQUEUE(wait, current);
3893 wait.flags |= WQ_FLAG_EXCLUSIVE;
3894 __add_wait_queue_tail(&x->wait, &wait);
3896 __set_current_state(TASK_UNINTERRUPTIBLE);
3897 spin_unlock_irq(&x->wait.lock);
3898 timeout = schedule_timeout(timeout);
3899 spin_lock_irq(&x->wait.lock);
3901 __remove_wait_queue(&x->wait, &wait);
3905 __remove_wait_queue(&x->wait, &wait);
3909 spin_unlock_irq(&x->wait.lock);
3912 EXPORT_SYMBOL(wait_for_completion_timeout);
3914 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3920 spin_lock_irq(&x->wait.lock);
3922 DECLARE_WAITQUEUE(wait, current);
3924 wait.flags |= WQ_FLAG_EXCLUSIVE;
3925 __add_wait_queue_tail(&x->wait, &wait);
3927 if (signal_pending(current)) {
3929 __remove_wait_queue(&x->wait, &wait);
3932 __set_current_state(TASK_INTERRUPTIBLE);
3933 spin_unlock_irq(&x->wait.lock);
3935 spin_lock_irq(&x->wait.lock);
3937 __remove_wait_queue(&x->wait, &wait);
3941 spin_unlock_irq(&x->wait.lock);
3945 EXPORT_SYMBOL(wait_for_completion_interruptible);
3947 unsigned long fastcall __sched
3948 wait_for_completion_interruptible_timeout(struct completion *x,
3949 unsigned long timeout)
3953 spin_lock_irq(&x->wait.lock);
3955 DECLARE_WAITQUEUE(wait, current);
3957 wait.flags |= WQ_FLAG_EXCLUSIVE;
3958 __add_wait_queue_tail(&x->wait, &wait);
3960 if (signal_pending(current)) {
3961 timeout = -ERESTARTSYS;
3962 __remove_wait_queue(&x->wait, &wait);
3965 __set_current_state(TASK_INTERRUPTIBLE);
3966 spin_unlock_irq(&x->wait.lock);
3967 timeout = schedule_timeout(timeout);
3968 spin_lock_irq(&x->wait.lock);
3970 __remove_wait_queue(&x->wait, &wait);
3974 __remove_wait_queue(&x->wait, &wait);
3978 spin_unlock_irq(&x->wait.lock);
3981 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3984 #define SLEEP_ON_VAR \
3985 unsigned long flags; \
3986 wait_queue_t wait; \
3987 init_waitqueue_entry(&wait, current);
3989 #define SLEEP_ON_HEAD \
3990 spin_lock_irqsave(&q->lock,flags); \
3991 __add_wait_queue(q, &wait); \
3992 spin_unlock(&q->lock);
3994 #define SLEEP_ON_TAIL \
3995 spin_lock_irq(&q->lock); \
3996 __remove_wait_queue(q, &wait); \
3997 spin_unlock_irqrestore(&q->lock, flags);
3999 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
4003 current->state = TASK_INTERRUPTIBLE;
4009 EXPORT_SYMBOL(interruptible_sleep_on);
4011 long fastcall __sched
4012 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4016 current->state = TASK_INTERRUPTIBLE;
4019 timeout = schedule_timeout(timeout);
4024 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4026 void fastcall __sched sleep_on(wait_queue_head_t *q)
4030 current->state = TASK_UNINTERRUPTIBLE;
4036 EXPORT_SYMBOL(sleep_on);
4038 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4042 current->state = TASK_UNINTERRUPTIBLE;
4045 timeout = schedule_timeout(timeout);
4051 EXPORT_SYMBOL(sleep_on_timeout);
4053 #ifdef CONFIG_RT_MUTEXES
4056 * rt_mutex_setprio - set the current priority of a task
4058 * @prio: prio value (kernel-internal form)
4060 * This function changes the 'effective' priority of a task. It does
4061 * not touch ->normal_prio like __setscheduler().
4063 * Used by the rt_mutex code to implement priority inheritance logic.
4065 void rt_mutex_setprio(struct task_struct *p, int prio)
4067 struct prio_array *array;
4068 unsigned long flags;
4072 BUG_ON(prio < 0 || prio > MAX_PRIO);
4074 rq = task_rq_lock(p, &flags);
4079 dequeue_task(p, array);
4084 * If changing to an RT priority then queue it
4085 * in the active array!
4089 enqueue_task(p, array);
4091 * Reschedule if we are currently running on this runqueue and
4092 * our priority decreased, or if we are not currently running on
4093 * this runqueue and our priority is higher than the current's
4095 if (task_running(rq, p)) {
4096 if (p->prio > oldprio)
4097 resched_task(rq->curr);
4098 } else if (TASK_PREEMPTS_CURR(p, rq))
4099 resched_task(rq->curr);
4101 task_rq_unlock(rq, &flags);
4106 void set_user_nice(struct task_struct *p, long nice)
4108 struct prio_array *array;
4109 int old_prio, delta;
4110 unsigned long flags;
4113 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4116 * We have to be careful, if called from sys_setpriority(),
4117 * the task might be in the middle of scheduling on another CPU.
4119 rq = task_rq_lock(p, &flags);
4121 * The RT priorities are set via sched_setscheduler(), but we still
4122 * allow the 'normal' nice value to be set - but as expected
4123 * it wont have any effect on scheduling until the task is
4124 * not SCHED_NORMAL/SCHED_BATCH:
4126 if (has_rt_policy(p)) {
4127 p->static_prio = NICE_TO_PRIO(nice);
4132 dequeue_task(p, array);
4133 dec_raw_weighted_load(rq, p);
4136 p->static_prio = NICE_TO_PRIO(nice);
4139 p->prio = effective_prio(p);
4140 delta = p->prio - old_prio;
4143 enqueue_task(p, array);
4144 inc_raw_weighted_load(rq, p);
4146 * If the task increased its priority or is running and
4147 * lowered its priority, then reschedule its CPU:
4149 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4150 resched_task(rq->curr);
4153 task_rq_unlock(rq, &flags);
4155 EXPORT_SYMBOL(set_user_nice);
4158 * can_nice - check if a task can reduce its nice value
4162 int can_nice(const struct task_struct *p, const int nice)
4164 /* convert nice value [19,-20] to rlimit style value [1,40] */
4165 int nice_rlim = 20 - nice;
4167 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4168 capable(CAP_SYS_NICE));
4171 #ifdef __ARCH_WANT_SYS_NICE
4174 * sys_nice - change the priority of the current process.
4175 * @increment: priority increment
4177 * sys_setpriority is a more generic, but much slower function that
4178 * does similar things.
4180 asmlinkage long sys_nice(int increment)
4185 * Setpriority might change our priority at the same moment.
4186 * We don't have to worry. Conceptually one call occurs first
4187 * and we have a single winner.
4189 if (increment < -40)
4194 nice = PRIO_TO_NICE(current->static_prio) + increment;
4200 if (increment < 0 && !can_nice(current, nice))
4203 retval = security_task_setnice(current, nice);
4207 set_user_nice(current, nice);
4214 * task_prio - return the priority value of a given task.
4215 * @p: the task in question.
4217 * This is the priority value as seen by users in /proc.
4218 * RT tasks are offset by -200. Normal tasks are centered
4219 * around 0, value goes from -16 to +15.
4221 int task_prio(const struct task_struct *p)
4223 return p->prio - MAX_RT_PRIO;
4227 * task_nice - return the nice value of a given task.
4228 * @p: the task in question.
4230 int task_nice(const struct task_struct *p)
4232 return TASK_NICE(p);
4234 EXPORT_SYMBOL_GPL(task_nice);
4237 * idle_cpu - is a given cpu idle currently?
4238 * @cpu: the processor in question.
4240 int idle_cpu(int cpu)
4242 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4246 * idle_task - return the idle task for a given cpu.
4247 * @cpu: the processor in question.
4249 struct task_struct *idle_task(int cpu)
4251 return cpu_rq(cpu)->idle;
4255 * find_process_by_pid - find a process with a matching PID value.
4256 * @pid: the pid in question.
4258 static inline struct task_struct *find_process_by_pid(pid_t pid)
4260 return pid ? find_task_by_pid(pid) : current;
4263 /* Actually do priority change: must hold rq lock. */
4264 static void __setscheduler(struct task_struct *p, int policy, int prio)
4269 p->rt_priority = prio;
4270 p->normal_prio = normal_prio(p);
4271 /* we are holding p->pi_lock already */
4272 p->prio = rt_mutex_getprio(p);
4274 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4276 if (policy == SCHED_BATCH)
4282 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4283 * @p: the task in question.
4284 * @policy: new policy.
4285 * @param: structure containing the new RT priority.
4287 * NOTE that the task may be already dead.
4289 int sched_setscheduler(struct task_struct *p, int policy,
4290 struct sched_param *param)
4292 int retval, oldprio, oldpolicy = -1;
4293 struct prio_array *array;
4294 unsigned long flags;
4297 /* may grab non-irq protected spin_locks */
4298 BUG_ON(in_interrupt());
4300 /* double check policy once rq lock held */
4302 policy = oldpolicy = p->policy;
4303 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4304 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4307 * Valid priorities for SCHED_FIFO and SCHED_RR are
4308 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4311 if (param->sched_priority < 0 ||
4312 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4313 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4315 if (is_rt_policy(policy) != (param->sched_priority != 0))
4319 * Allow unprivileged RT tasks to decrease priority:
4321 if (!capable(CAP_SYS_NICE)) {
4322 if (is_rt_policy(policy)) {
4323 unsigned long rlim_rtprio;
4324 unsigned long flags;
4326 if (!lock_task_sighand(p, &flags))
4328 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4329 unlock_task_sighand(p, &flags);
4331 /* can't set/change the rt policy */
4332 if (policy != p->policy && !rlim_rtprio)
4335 /* can't increase priority */
4336 if (param->sched_priority > p->rt_priority &&
4337 param->sched_priority > rlim_rtprio)
4341 /* can't change other user's priorities */
4342 if ((current->euid != p->euid) &&
4343 (current->euid != p->uid))
4347 retval = security_task_setscheduler(p, policy, param);
4351 * make sure no PI-waiters arrive (or leave) while we are
4352 * changing the priority of the task:
4354 spin_lock_irqsave(&p->pi_lock, flags);
4356 * To be able to change p->policy safely, the apropriate
4357 * runqueue lock must be held.
4359 rq = __task_rq_lock(p);
4360 /* recheck policy now with rq lock held */
4361 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4362 policy = oldpolicy = -1;
4363 __task_rq_unlock(rq);
4364 spin_unlock_irqrestore(&p->pi_lock, flags);
4369 deactivate_task(p, rq);
4371 __setscheduler(p, policy, param->sched_priority);
4373 __activate_task(p, rq);
4375 * Reschedule if we are currently running on this runqueue and
4376 * our priority decreased, or if we are not currently running on
4377 * this runqueue and our priority is higher than the current's
4379 if (task_running(rq, p)) {
4380 if (p->prio > oldprio)
4381 resched_task(rq->curr);
4382 } else if (TASK_PREEMPTS_CURR(p, rq))
4383 resched_task(rq->curr);
4385 __task_rq_unlock(rq);
4386 spin_unlock_irqrestore(&p->pi_lock, flags);
4388 rt_mutex_adjust_pi(p);
4392 EXPORT_SYMBOL_GPL(sched_setscheduler);
4395 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4397 struct sched_param lparam;
4398 struct task_struct *p;
4401 if (!param || pid < 0)
4403 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4408 p = find_process_by_pid(pid);
4410 retval = sched_setscheduler(p, policy, &lparam);
4417 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4418 * @pid: the pid in question.
4419 * @policy: new policy.
4420 * @param: structure containing the new RT priority.
4422 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4423 struct sched_param __user *param)
4425 /* negative values for policy are not valid */
4429 return do_sched_setscheduler(pid, policy, param);
4433 * sys_sched_setparam - set/change the RT priority of a thread
4434 * @pid: the pid in question.
4435 * @param: structure containing the new RT priority.
4437 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4439 return do_sched_setscheduler(pid, -1, param);
4443 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4444 * @pid: the pid in question.
4446 asmlinkage long sys_sched_getscheduler(pid_t pid)
4448 struct task_struct *p;
4449 int retval = -EINVAL;
4455 read_lock(&tasklist_lock);
4456 p = find_process_by_pid(pid);
4458 retval = security_task_getscheduler(p);
4462 read_unlock(&tasklist_lock);
4469 * sys_sched_getscheduler - get the RT priority of a thread
4470 * @pid: the pid in question.
4471 * @param: structure containing the RT priority.
4473 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4475 struct sched_param lp;
4476 struct task_struct *p;
4477 int retval = -EINVAL;
4479 if (!param || pid < 0)
4482 read_lock(&tasklist_lock);
4483 p = find_process_by_pid(pid);
4488 retval = security_task_getscheduler(p);
4492 lp.sched_priority = p->rt_priority;
4493 read_unlock(&tasklist_lock);
4496 * This one might sleep, we cannot do it with a spinlock held ...
4498 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4504 read_unlock(&tasklist_lock);
4508 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4510 cpumask_t cpus_allowed;
4511 struct task_struct *p;
4515 read_lock(&tasklist_lock);
4517 p = find_process_by_pid(pid);
4519 read_unlock(&tasklist_lock);
4520 unlock_cpu_hotplug();
4525 * It is not safe to call set_cpus_allowed with the
4526 * tasklist_lock held. We will bump the task_struct's
4527 * usage count and then drop tasklist_lock.
4530 read_unlock(&tasklist_lock);
4533 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4534 !capable(CAP_SYS_NICE))
4537 retval = security_task_setscheduler(p, 0, NULL);
4541 cpus_allowed = cpuset_cpus_allowed(p);
4542 cpus_and(new_mask, new_mask, cpus_allowed);
4543 retval = set_cpus_allowed(p, new_mask);
4547 unlock_cpu_hotplug();
4551 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4552 cpumask_t *new_mask)
4554 if (len < sizeof(cpumask_t)) {
4555 memset(new_mask, 0, sizeof(cpumask_t));
4556 } else if (len > sizeof(cpumask_t)) {
4557 len = sizeof(cpumask_t);
4559 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4563 * sys_sched_setaffinity - set the cpu affinity of a process
4564 * @pid: pid of the process
4565 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4566 * @user_mask_ptr: user-space pointer to the new cpu mask
4568 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4569 unsigned long __user *user_mask_ptr)
4574 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4578 return sched_setaffinity(pid, new_mask);
4582 * Represents all cpu's present in the system
4583 * In systems capable of hotplug, this map could dynamically grow
4584 * as new cpu's are detected in the system via any platform specific
4585 * method, such as ACPI for e.g.
4588 cpumask_t cpu_present_map __read_mostly;
4589 EXPORT_SYMBOL(cpu_present_map);
4592 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4593 EXPORT_SYMBOL(cpu_online_map);
4595 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4596 EXPORT_SYMBOL(cpu_possible_map);
4599 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4601 struct task_struct *p;
4605 read_lock(&tasklist_lock);
4608 p = find_process_by_pid(pid);
4612 retval = security_task_getscheduler(p);
4616 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4619 read_unlock(&tasklist_lock);
4620 unlock_cpu_hotplug();
4628 * sys_sched_getaffinity - get the cpu affinity of a process
4629 * @pid: pid of the process
4630 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4631 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4633 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4634 unsigned long __user *user_mask_ptr)
4639 if (len < sizeof(cpumask_t))
4642 ret = sched_getaffinity(pid, &mask);
4646 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4649 return sizeof(cpumask_t);
4653 * sys_sched_yield - yield the current processor to other threads.
4655 * This function yields the current CPU by moving the calling thread
4656 * to the expired array. If there are no other threads running on this
4657 * CPU then this function will return.
4659 asmlinkage long sys_sched_yield(void)
4661 struct rq *rq = this_rq_lock();
4662 struct prio_array *array = current->array, *target = rq->expired;
4664 schedstat_inc(rq, yld_cnt);
4666 * We implement yielding by moving the task into the expired
4669 * (special rule: RT tasks will just roundrobin in the active
4672 if (rt_task(current))
4673 target = rq->active;
4675 if (array->nr_active == 1) {
4676 schedstat_inc(rq, yld_act_empty);
4677 if (!rq->expired->nr_active)
4678 schedstat_inc(rq, yld_both_empty);
4679 } else if (!rq->expired->nr_active)
4680 schedstat_inc(rq, yld_exp_empty);
4682 if (array != target) {
4683 dequeue_task(current, array);
4684 enqueue_task(current, target);
4687 * requeue_task is cheaper so perform that if possible.
4689 requeue_task(current, array);
4692 * Since we are going to call schedule() anyway, there's
4693 * no need to preempt or enable interrupts:
4695 __release(rq->lock);
4696 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4697 _raw_spin_unlock(&rq->lock);
4698 preempt_enable_no_resched();
4705 static void __cond_resched(void)
4707 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4708 __might_sleep(__FILE__, __LINE__);
4711 * The BKS might be reacquired before we have dropped
4712 * PREEMPT_ACTIVE, which could trigger a second
4713 * cond_resched() call.
4716 add_preempt_count(PREEMPT_ACTIVE);
4718 sub_preempt_count(PREEMPT_ACTIVE);
4719 } while (need_resched());
4722 int __sched cond_resched(void)
4724 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4725 system_state == SYSTEM_RUNNING) {
4731 EXPORT_SYMBOL(cond_resched);
4734 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4735 * call schedule, and on return reacquire the lock.
4737 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4738 * operations here to prevent schedule() from being called twice (once via
4739 * spin_unlock(), once by hand).
4741 int cond_resched_lock(spinlock_t *lock)
4745 if (need_lockbreak(lock)) {
4751 if (need_resched() && system_state == SYSTEM_RUNNING) {
4752 spin_release(&lock->dep_map, 1, _THIS_IP_);
4753 _raw_spin_unlock(lock);
4754 preempt_enable_no_resched();
4761 EXPORT_SYMBOL(cond_resched_lock);
4763 int __sched cond_resched_softirq(void)
4765 BUG_ON(!in_softirq());
4767 if (need_resched() && system_state == SYSTEM_RUNNING) {
4768 raw_local_irq_disable();
4770 raw_local_irq_enable();
4777 EXPORT_SYMBOL(cond_resched_softirq);
4780 * yield - yield the current processor to other threads.
4782 * This is a shortcut for kernel-space yielding - it marks the
4783 * thread runnable and calls sys_sched_yield().
4785 void __sched yield(void)
4787 set_current_state(TASK_RUNNING);
4790 EXPORT_SYMBOL(yield);
4793 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4794 * that process accounting knows that this is a task in IO wait state.
4796 * But don't do that if it is a deliberate, throttling IO wait (this task
4797 * has set its backing_dev_info: the queue against which it should throttle)
4799 void __sched io_schedule(void)
4801 struct rq *rq = &__raw_get_cpu_var(runqueues);
4803 delayacct_blkio_start();
4804 atomic_inc(&rq->nr_iowait);
4806 atomic_dec(&rq->nr_iowait);
4807 delayacct_blkio_end();
4809 EXPORT_SYMBOL(io_schedule);
4811 long __sched io_schedule_timeout(long timeout)
4813 struct rq *rq = &__raw_get_cpu_var(runqueues);
4816 delayacct_blkio_start();
4817 atomic_inc(&rq->nr_iowait);
4818 ret = schedule_timeout(timeout);
4819 atomic_dec(&rq->nr_iowait);
4820 delayacct_blkio_end();
4825 * sys_sched_get_priority_max - return maximum RT priority.
4826 * @policy: scheduling class.
4828 * this syscall returns the maximum rt_priority that can be used
4829 * by a given scheduling class.
4831 asmlinkage long sys_sched_get_priority_max(int policy)
4838 ret = MAX_USER_RT_PRIO-1;
4849 * sys_sched_get_priority_min - return minimum RT priority.
4850 * @policy: scheduling class.
4852 * this syscall returns the minimum rt_priority that can be used
4853 * by a given scheduling class.
4855 asmlinkage long sys_sched_get_priority_min(int policy)
4872 * sys_sched_rr_get_interval - return the default timeslice of a process.
4873 * @pid: pid of the process.
4874 * @interval: userspace pointer to the timeslice value.
4876 * this syscall writes the default timeslice value of a given process
4877 * into the user-space timespec buffer. A value of '0' means infinity.
4880 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4882 struct task_struct *p;
4883 int retval = -EINVAL;
4890 read_lock(&tasklist_lock);
4891 p = find_process_by_pid(pid);
4895 retval = security_task_getscheduler(p);
4899 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4900 0 : task_timeslice(p), &t);
4901 read_unlock(&tasklist_lock);
4902 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4906 read_unlock(&tasklist_lock);
4910 static const char stat_nam[] = "RSDTtZX";
4912 static void show_task(struct task_struct *p)
4914 unsigned long free = 0;
4917 state = p->state ? __ffs(p->state) + 1 : 0;
4918 printk("%-13.13s %c", p->comm,
4919 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4920 #if (BITS_PER_LONG == 32)
4921 if (state == TASK_RUNNING)
4922 printk(" running ");
4924 printk(" %08lX ", thread_saved_pc(p));
4926 if (state == TASK_RUNNING)
4927 printk(" running task ");
4929 printk(" %016lx ", thread_saved_pc(p));
4931 #ifdef CONFIG_DEBUG_STACK_USAGE
4933 unsigned long *n = end_of_stack(p);
4936 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4939 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4941 printk(" (L-TLB)\n");
4943 printk(" (NOTLB)\n");
4945 if (state != TASK_RUNNING)
4946 show_stack(p, NULL);
4949 void show_state_filter(unsigned long state_filter)
4951 struct task_struct *g, *p;
4953 #if (BITS_PER_LONG == 32)
4956 printk(" task PC stack pid father child younger older\n");
4960 printk(" task PC stack pid father child younger older\n");
4962 read_lock(&tasklist_lock);
4963 do_each_thread(g, p) {
4965 * reset the NMI-timeout, listing all files on a slow
4966 * console might take alot of time:
4968 touch_nmi_watchdog();
4969 if (!state_filter || (p->state & state_filter))
4971 } while_each_thread(g, p);
4973 touch_all_softlockup_watchdogs();
4975 read_unlock(&tasklist_lock);
4977 * Only show locks if all tasks are dumped:
4979 if (state_filter == -1)
4980 debug_show_all_locks();
4984 * init_idle - set up an idle thread for a given CPU
4985 * @idle: task in question
4986 * @cpu: cpu the idle task belongs to
4988 * NOTE: this function does not set the idle thread's NEED_RESCHED
4989 * flag, to make booting more robust.
4991 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4993 struct rq *rq = cpu_rq(cpu);
4994 unsigned long flags;
4996 idle->timestamp = sched_clock();
4997 idle->sleep_avg = 0;
4999 idle->prio = idle->normal_prio = MAX_PRIO;
5000 idle->state = TASK_RUNNING;
5001 idle->cpus_allowed = cpumask_of_cpu(cpu);
5002 set_task_cpu(idle, cpu);
5004 spin_lock_irqsave(&rq->lock, flags);
5005 rq->curr = rq->idle = idle;
5006 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5009 spin_unlock_irqrestore(&rq->lock, flags);
5011 /* Set the preempt count _outside_ the spinlocks! */
5012 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5013 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5015 task_thread_info(idle)->preempt_count = 0;
5020 * In a system that switches off the HZ timer nohz_cpu_mask
5021 * indicates which cpus entered this state. This is used
5022 * in the rcu update to wait only for active cpus. For system
5023 * which do not switch off the HZ timer nohz_cpu_mask should
5024 * always be CPU_MASK_NONE.
5026 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5030 * This is how migration works:
5032 * 1) we queue a struct migration_req structure in the source CPU's
5033 * runqueue and wake up that CPU's migration thread.
5034 * 2) we down() the locked semaphore => thread blocks.
5035 * 3) migration thread wakes up (implicitly it forces the migrated
5036 * thread off the CPU)
5037 * 4) it gets the migration request and checks whether the migrated
5038 * task is still in the wrong runqueue.
5039 * 5) if it's in the wrong runqueue then the migration thread removes
5040 * it and puts it into the right queue.
5041 * 6) migration thread up()s the semaphore.
5042 * 7) we wake up and the migration is done.
5046 * Change a given task's CPU affinity. Migrate the thread to a
5047 * proper CPU and schedule it away if the CPU it's executing on
5048 * is removed from the allowed bitmask.
5050 * NOTE: the caller must have a valid reference to the task, the
5051 * task must not exit() & deallocate itself prematurely. The
5052 * call is not atomic; no spinlocks may be held.
5054 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5056 struct migration_req req;
5057 unsigned long flags;
5061 rq = task_rq_lock(p, &flags);
5062 if (!cpus_intersects(new_mask, cpu_online_map)) {
5067 p->cpus_allowed = new_mask;
5068 /* Can the task run on the task's current CPU? If so, we're done */
5069 if (cpu_isset(task_cpu(p), new_mask))
5072 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5073 /* Need help from migration thread: drop lock and wait. */
5074 task_rq_unlock(rq, &flags);
5075 wake_up_process(rq->migration_thread);
5076 wait_for_completion(&req.done);
5077 tlb_migrate_finish(p->mm);
5081 task_rq_unlock(rq, &flags);
5085 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5088 * Move (not current) task off this cpu, onto dest cpu. We're doing
5089 * this because either it can't run here any more (set_cpus_allowed()
5090 * away from this CPU, or CPU going down), or because we're
5091 * attempting to rebalance this task on exec (sched_exec).
5093 * So we race with normal scheduler movements, but that's OK, as long
5094 * as the task is no longer on this CPU.
5096 * Returns non-zero if task was successfully migrated.
5098 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5100 struct rq *rq_dest, *rq_src;
5103 if (unlikely(cpu_is_offline(dest_cpu)))
5106 rq_src = cpu_rq(src_cpu);
5107 rq_dest = cpu_rq(dest_cpu);
5109 double_rq_lock(rq_src, rq_dest);
5110 /* Already moved. */
5111 if (task_cpu(p) != src_cpu)
5113 /* Affinity changed (again). */
5114 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5117 set_task_cpu(p, dest_cpu);
5120 * Sync timestamp with rq_dest's before activating.
5121 * The same thing could be achieved by doing this step
5122 * afterwards, and pretending it was a local activate.
5123 * This way is cleaner and logically correct.
5125 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5126 + rq_dest->most_recent_timestamp;
5127 deactivate_task(p, rq_src);
5128 __activate_task(p, rq_dest);
5129 if (TASK_PREEMPTS_CURR(p, rq_dest))
5130 resched_task(rq_dest->curr);
5134 double_rq_unlock(rq_src, rq_dest);
5139 * migration_thread - this is a highprio system thread that performs
5140 * thread migration by bumping thread off CPU then 'pushing' onto
5143 static int migration_thread(void *data)
5145 int cpu = (long)data;
5149 BUG_ON(rq->migration_thread != current);
5151 set_current_state(TASK_INTERRUPTIBLE);
5152 while (!kthread_should_stop()) {
5153 struct migration_req *req;
5154 struct list_head *head;
5158 spin_lock_irq(&rq->lock);
5160 if (cpu_is_offline(cpu)) {
5161 spin_unlock_irq(&rq->lock);
5165 if (rq->active_balance) {
5166 active_load_balance(rq, cpu);
5167 rq->active_balance = 0;
5170 head = &rq->migration_queue;
5172 if (list_empty(head)) {
5173 spin_unlock_irq(&rq->lock);
5175 set_current_state(TASK_INTERRUPTIBLE);
5178 req = list_entry(head->next, struct migration_req, list);
5179 list_del_init(head->next);
5181 spin_unlock(&rq->lock);
5182 __migrate_task(req->task, cpu, req->dest_cpu);
5185 complete(&req->done);
5187 __set_current_state(TASK_RUNNING);
5191 /* Wait for kthread_stop */
5192 set_current_state(TASK_INTERRUPTIBLE);
5193 while (!kthread_should_stop()) {
5195 set_current_state(TASK_INTERRUPTIBLE);
5197 __set_current_state(TASK_RUNNING);
5201 #ifdef CONFIG_HOTPLUG_CPU
5203 * Figure out where task on dead CPU should go, use force if neccessary.
5204 * NOTE: interrupts should be disabled by the caller
5206 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5208 unsigned long flags;
5215 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5216 cpus_and(mask, mask, p->cpus_allowed);
5217 dest_cpu = any_online_cpu(mask);
5219 /* On any allowed CPU? */
5220 if (dest_cpu == NR_CPUS)
5221 dest_cpu = any_online_cpu(p->cpus_allowed);
5223 /* No more Mr. Nice Guy. */
5224 if (dest_cpu == NR_CPUS) {
5225 rq = task_rq_lock(p, &flags);
5226 cpus_setall(p->cpus_allowed);
5227 dest_cpu = any_online_cpu(p->cpus_allowed);
5228 task_rq_unlock(rq, &flags);
5231 * Don't tell them about moving exiting tasks or
5232 * kernel threads (both mm NULL), since they never
5235 if (p->mm && printk_ratelimit())
5236 printk(KERN_INFO "process %d (%s) no "
5237 "longer affine to cpu%d\n",
5238 p->pid, p->comm, dead_cpu);
5240 if (!__migrate_task(p, dead_cpu, dest_cpu))
5245 * While a dead CPU has no uninterruptible tasks queued at this point,
5246 * it might still have a nonzero ->nr_uninterruptible counter, because
5247 * for performance reasons the counter is not stricly tracking tasks to
5248 * their home CPUs. So we just add the counter to another CPU's counter,
5249 * to keep the global sum constant after CPU-down:
5251 static void migrate_nr_uninterruptible(struct rq *rq_src)
5253 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5254 unsigned long flags;
5256 local_irq_save(flags);
5257 double_rq_lock(rq_src, rq_dest);
5258 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5259 rq_src->nr_uninterruptible = 0;
5260 double_rq_unlock(rq_src, rq_dest);
5261 local_irq_restore(flags);
5264 /* Run through task list and migrate tasks from the dead cpu. */
5265 static void migrate_live_tasks(int src_cpu)
5267 struct task_struct *p, *t;
5269 write_lock_irq(&tasklist_lock);
5271 do_each_thread(t, p) {
5275 if (task_cpu(p) == src_cpu)
5276 move_task_off_dead_cpu(src_cpu, p);
5277 } while_each_thread(t, p);
5279 write_unlock_irq(&tasklist_lock);
5282 /* Schedules idle task to be the next runnable task on current CPU.
5283 * It does so by boosting its priority to highest possible and adding it to
5284 * the _front_ of the runqueue. Used by CPU offline code.
5286 void sched_idle_next(void)
5288 int this_cpu = smp_processor_id();
5289 struct rq *rq = cpu_rq(this_cpu);
5290 struct task_struct *p = rq->idle;
5291 unsigned long flags;
5293 /* cpu has to be offline */
5294 BUG_ON(cpu_online(this_cpu));
5297 * Strictly not necessary since rest of the CPUs are stopped by now
5298 * and interrupts disabled on the current cpu.
5300 spin_lock_irqsave(&rq->lock, flags);
5302 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5304 /* Add idle task to the _front_ of its priority queue: */
5305 __activate_idle_task(p, rq);
5307 spin_unlock_irqrestore(&rq->lock, flags);
5311 * Ensures that the idle task is using init_mm right before its cpu goes
5314 void idle_task_exit(void)
5316 struct mm_struct *mm = current->active_mm;
5318 BUG_ON(cpu_online(smp_processor_id()));
5321 switch_mm(mm, &init_mm, current);
5325 /* called under rq->lock with disabled interrupts */
5326 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5328 struct rq *rq = cpu_rq(dead_cpu);
5330 /* Must be exiting, otherwise would be on tasklist. */
5331 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5333 /* Cannot have done final schedule yet: would have vanished. */
5334 BUG_ON(p->state == TASK_DEAD);
5339 * Drop lock around migration; if someone else moves it,
5340 * that's OK. No task can be added to this CPU, so iteration is
5342 * NOTE: interrupts should be left disabled --dev@
5344 spin_unlock(&rq->lock);
5345 move_task_off_dead_cpu(dead_cpu, p);
5346 spin_lock(&rq->lock);
5351 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5352 static void migrate_dead_tasks(unsigned int dead_cpu)
5354 struct rq *rq = cpu_rq(dead_cpu);
5355 unsigned int arr, i;
5357 for (arr = 0; arr < 2; arr++) {
5358 for (i = 0; i < MAX_PRIO; i++) {
5359 struct list_head *list = &rq->arrays[arr].queue[i];
5361 while (!list_empty(list))
5362 migrate_dead(dead_cpu, list_entry(list->next,
5363 struct task_struct, run_list));
5367 #endif /* CONFIG_HOTPLUG_CPU */
5370 * migration_call - callback that gets triggered when a CPU is added.
5371 * Here we can start up the necessary migration thread for the new CPU.
5373 static int __cpuinit
5374 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5376 struct task_struct *p;
5377 int cpu = (long)hcpu;
5378 unsigned long flags;
5382 case CPU_UP_PREPARE:
5383 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5386 p->flags |= PF_NOFREEZE;
5387 kthread_bind(p, cpu);
5388 /* Must be high prio: stop_machine expects to yield to it. */
5389 rq = task_rq_lock(p, &flags);
5390 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5391 task_rq_unlock(rq, &flags);
5392 cpu_rq(cpu)->migration_thread = p;
5396 /* Strictly unneccessary, as first user will wake it. */
5397 wake_up_process(cpu_rq(cpu)->migration_thread);
5400 #ifdef CONFIG_HOTPLUG_CPU
5401 case CPU_UP_CANCELED:
5402 if (!cpu_rq(cpu)->migration_thread)
5404 /* Unbind it from offline cpu so it can run. Fall thru. */
5405 kthread_bind(cpu_rq(cpu)->migration_thread,
5406 any_online_cpu(cpu_online_map));
5407 kthread_stop(cpu_rq(cpu)->migration_thread);
5408 cpu_rq(cpu)->migration_thread = NULL;
5412 migrate_live_tasks(cpu);
5414 kthread_stop(rq->migration_thread);
5415 rq->migration_thread = NULL;
5416 /* Idle task back to normal (off runqueue, low prio) */
5417 rq = task_rq_lock(rq->idle, &flags);
5418 deactivate_task(rq->idle, rq);
5419 rq->idle->static_prio = MAX_PRIO;
5420 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5421 migrate_dead_tasks(cpu);
5422 task_rq_unlock(rq, &flags);
5423 migrate_nr_uninterruptible(rq);
5424 BUG_ON(rq->nr_running != 0);
5426 /* No need to migrate the tasks: it was best-effort if
5427 * they didn't do lock_cpu_hotplug(). Just wake up
5428 * the requestors. */
5429 spin_lock_irq(&rq->lock);
5430 while (!list_empty(&rq->migration_queue)) {
5431 struct migration_req *req;
5433 req = list_entry(rq->migration_queue.next,
5434 struct migration_req, list);
5435 list_del_init(&req->list);
5436 complete(&req->done);
5438 spin_unlock_irq(&rq->lock);
5445 /* Register at highest priority so that task migration (migrate_all_tasks)
5446 * happens before everything else.
5448 static struct notifier_block __cpuinitdata migration_notifier = {
5449 .notifier_call = migration_call,
5453 int __init migration_init(void)
5455 void *cpu = (void *)(long)smp_processor_id();
5458 /* Start one for the boot CPU: */
5459 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5460 BUG_ON(err == NOTIFY_BAD);
5461 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5462 register_cpu_notifier(&migration_notifier);
5470 /* Number of possible processor ids */
5471 int nr_cpu_ids __read_mostly = NR_CPUS;
5472 EXPORT_SYMBOL(nr_cpu_ids);
5474 #undef SCHED_DOMAIN_DEBUG
5475 #ifdef SCHED_DOMAIN_DEBUG
5476 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5481 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5485 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5490 struct sched_group *group = sd->groups;
5491 cpumask_t groupmask;
5493 cpumask_scnprintf(str, NR_CPUS, sd->span);
5494 cpus_clear(groupmask);
5497 for (i = 0; i < level + 1; i++)
5499 printk("domain %d: ", level);
5501 if (!(sd->flags & SD_LOAD_BALANCE)) {
5502 printk("does not load-balance\n");
5504 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5509 printk("span %s\n", str);
5511 if (!cpu_isset(cpu, sd->span))
5512 printk(KERN_ERR "ERROR: domain->span does not contain "
5514 if (!cpu_isset(cpu, group->cpumask))
5515 printk(KERN_ERR "ERROR: domain->groups does not contain"
5519 for (i = 0; i < level + 2; i++)
5525 printk(KERN_ERR "ERROR: group is NULL\n");
5529 if (!group->__cpu_power) {
5531 printk(KERN_ERR "ERROR: domain->cpu_power not "
5535 if (!cpus_weight(group->cpumask)) {
5537 printk(KERN_ERR "ERROR: empty group\n");
5540 if (cpus_intersects(groupmask, group->cpumask)) {
5542 printk(KERN_ERR "ERROR: repeated CPUs\n");
5545 cpus_or(groupmask, groupmask, group->cpumask);
5547 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5550 group = group->next;
5551 } while (group != sd->groups);
5554 if (!cpus_equal(sd->span, groupmask))
5555 printk(KERN_ERR "ERROR: groups don't span "
5563 if (!cpus_subset(groupmask, sd->span))
5564 printk(KERN_ERR "ERROR: parent span is not a superset "
5565 "of domain->span\n");
5570 # define sched_domain_debug(sd, cpu) do { } while (0)
5573 static int sd_degenerate(struct sched_domain *sd)
5575 if (cpus_weight(sd->span) == 1)
5578 /* Following flags need at least 2 groups */
5579 if (sd->flags & (SD_LOAD_BALANCE |
5580 SD_BALANCE_NEWIDLE |
5584 SD_SHARE_PKG_RESOURCES)) {
5585 if (sd->groups != sd->groups->next)
5589 /* Following flags don't use groups */
5590 if (sd->flags & (SD_WAKE_IDLE |
5599 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5601 unsigned long cflags = sd->flags, pflags = parent->flags;
5603 if (sd_degenerate(parent))
5606 if (!cpus_equal(sd->span, parent->span))
5609 /* Does parent contain flags not in child? */
5610 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5611 if (cflags & SD_WAKE_AFFINE)
5612 pflags &= ~SD_WAKE_BALANCE;
5613 /* Flags needing groups don't count if only 1 group in parent */
5614 if (parent->groups == parent->groups->next) {
5615 pflags &= ~(SD_LOAD_BALANCE |
5616 SD_BALANCE_NEWIDLE |
5620 SD_SHARE_PKG_RESOURCES);
5622 if (~cflags & pflags)
5629 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5630 * hold the hotplug lock.
5632 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5634 struct rq *rq = cpu_rq(cpu);
5635 struct sched_domain *tmp;
5637 /* Remove the sched domains which do not contribute to scheduling. */
5638 for (tmp = sd; tmp; tmp = tmp->parent) {
5639 struct sched_domain *parent = tmp->parent;
5642 if (sd_parent_degenerate(tmp, parent)) {
5643 tmp->parent = parent->parent;
5645 parent->parent->child = tmp;
5649 if (sd && sd_degenerate(sd)) {
5655 sched_domain_debug(sd, cpu);
5657 rcu_assign_pointer(rq->sd, sd);
5660 /* cpus with isolated domains */
5661 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5663 /* Setup the mask of cpus configured for isolated domains */
5664 static int __init isolated_cpu_setup(char *str)
5666 int ints[NR_CPUS], i;
5668 str = get_options(str, ARRAY_SIZE(ints), ints);
5669 cpus_clear(cpu_isolated_map);
5670 for (i = 1; i <= ints[0]; i++)
5671 if (ints[i] < NR_CPUS)
5672 cpu_set(ints[i], cpu_isolated_map);
5676 __setup ("isolcpus=", isolated_cpu_setup);
5679 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5680 * to a function which identifies what group(along with sched group) a CPU
5681 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5682 * (due to the fact that we keep track of groups covered with a cpumask_t).
5684 * init_sched_build_groups will build a circular linked list of the groups
5685 * covered by the given span, and will set each group's ->cpumask correctly,
5686 * and ->cpu_power to 0.
5689 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5690 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5691 struct sched_group **sg))
5693 struct sched_group *first = NULL, *last = NULL;
5694 cpumask_t covered = CPU_MASK_NONE;
5697 for_each_cpu_mask(i, span) {
5698 struct sched_group *sg;
5699 int group = group_fn(i, cpu_map, &sg);
5702 if (cpu_isset(i, covered))
5705 sg->cpumask = CPU_MASK_NONE;
5706 sg->__cpu_power = 0;
5708 for_each_cpu_mask(j, span) {
5709 if (group_fn(j, cpu_map, NULL) != group)
5712 cpu_set(j, covered);
5713 cpu_set(j, sg->cpumask);
5724 #define SD_NODES_PER_DOMAIN 16
5727 * Self-tuning task migration cost measurement between source and target CPUs.
5729 * This is done by measuring the cost of manipulating buffers of varying
5730 * sizes. For a given buffer-size here are the steps that are taken:
5732 * 1) the source CPU reads+dirties a shared buffer
5733 * 2) the target CPU reads+dirties the same shared buffer
5735 * We measure how long they take, in the following 4 scenarios:
5737 * - source: CPU1, target: CPU2 | cost1
5738 * - source: CPU2, target: CPU1 | cost2
5739 * - source: CPU1, target: CPU1 | cost3
5740 * - source: CPU2, target: CPU2 | cost4
5742 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5743 * the cost of migration.
5745 * We then start off from a small buffer-size and iterate up to larger
5746 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5747 * doing a maximum search for the cost. (The maximum cost for a migration
5748 * normally occurs when the working set size is around the effective cache
5751 #define SEARCH_SCOPE 2
5752 #define MIN_CACHE_SIZE (64*1024U)
5753 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5754 #define ITERATIONS 1
5755 #define SIZE_THRESH 130
5756 #define COST_THRESH 130
5759 * The migration cost is a function of 'domain distance'. Domain
5760 * distance is the number of steps a CPU has to iterate down its
5761 * domain tree to share a domain with the other CPU. The farther
5762 * two CPUs are from each other, the larger the distance gets.
5764 * Note that we use the distance only to cache measurement results,
5765 * the distance value is not used numerically otherwise. When two
5766 * CPUs have the same distance it is assumed that the migration
5767 * cost is the same. (this is a simplification but quite practical)
5769 #define MAX_DOMAIN_DISTANCE 32
5771 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5772 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5774 * Architectures may override the migration cost and thus avoid
5775 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5776 * virtualized hardware:
5778 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5779 CONFIG_DEFAULT_MIGRATION_COST
5786 * Allow override of migration cost - in units of microseconds.
5787 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5788 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5790 static int __init migration_cost_setup(char *str)
5792 int ints[MAX_DOMAIN_DISTANCE+1], i;
5794 str = get_options(str, ARRAY_SIZE(ints), ints);
5796 printk("#ints: %d\n", ints[0]);
5797 for (i = 1; i <= ints[0]; i++) {
5798 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5799 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5804 __setup ("migration_cost=", migration_cost_setup);
5807 * Global multiplier (divisor) for migration-cutoff values,
5808 * in percentiles. E.g. use a value of 150 to get 1.5 times
5809 * longer cache-hot cutoff times.
5811 * (We scale it from 100 to 128 to long long handling easier.)
5814 #define MIGRATION_FACTOR_SCALE 128
5816 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5818 static int __init setup_migration_factor(char *str)
5820 get_option(&str, &migration_factor);
5821 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5825 __setup("migration_factor=", setup_migration_factor);
5828 * Estimated distance of two CPUs, measured via the number of domains
5829 * we have to pass for the two CPUs to be in the same span:
5831 static unsigned long domain_distance(int cpu1, int cpu2)
5833 unsigned long distance = 0;
5834 struct sched_domain *sd;
5836 for_each_domain(cpu1, sd) {
5837 WARN_ON(!cpu_isset(cpu1, sd->span));
5838 if (cpu_isset(cpu2, sd->span))
5842 if (distance >= MAX_DOMAIN_DISTANCE) {
5844 distance = MAX_DOMAIN_DISTANCE-1;
5850 static unsigned int migration_debug;
5852 static int __init setup_migration_debug(char *str)
5854 get_option(&str, &migration_debug);
5858 __setup("migration_debug=", setup_migration_debug);
5861 * Maximum cache-size that the scheduler should try to measure.
5862 * Architectures with larger caches should tune this up during
5863 * bootup. Gets used in the domain-setup code (i.e. during SMP
5866 unsigned int max_cache_size;
5868 static int __init setup_max_cache_size(char *str)
5870 get_option(&str, &max_cache_size);
5874 __setup("max_cache_size=", setup_max_cache_size);
5877 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5878 * is the operation that is timed, so we try to generate unpredictable
5879 * cachemisses that still end up filling the L2 cache:
5881 static void touch_cache(void *__cache, unsigned long __size)
5883 unsigned long size = __size / sizeof(long);
5884 unsigned long chunk1 = size / 3;
5885 unsigned long chunk2 = 2 * size / 3;
5886 unsigned long *cache = __cache;
5889 for (i = 0; i < size/6; i += 8) {
5892 case 1: cache[size-1-i]++;
5893 case 2: cache[chunk1-i]++;
5894 case 3: cache[chunk1+i]++;
5895 case 4: cache[chunk2-i]++;
5896 case 5: cache[chunk2+i]++;
5902 * Measure the cache-cost of one task migration. Returns in units of nsec.
5904 static unsigned long long
5905 measure_one(void *cache, unsigned long size, int source, int target)
5907 cpumask_t mask, saved_mask;
5908 unsigned long long t0, t1, t2, t3, cost;
5910 saved_mask = current->cpus_allowed;
5913 * Flush source caches to RAM and invalidate them:
5918 * Migrate to the source CPU:
5920 mask = cpumask_of_cpu(source);
5921 set_cpus_allowed(current, mask);
5922 WARN_ON(smp_processor_id() != source);
5925 * Dirty the working set:
5928 touch_cache(cache, size);
5932 * Migrate to the target CPU, dirty the L2 cache and access
5933 * the shared buffer. (which represents the working set
5934 * of a migrated task.)
5936 mask = cpumask_of_cpu(target);
5937 set_cpus_allowed(current, mask);
5938 WARN_ON(smp_processor_id() != target);
5941 touch_cache(cache, size);
5944 cost = t1-t0 + t3-t2;
5946 if (migration_debug >= 2)
5947 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5948 source, target, t1-t0, t1-t0, t3-t2, cost);
5950 * Flush target caches to RAM and invalidate them:
5954 set_cpus_allowed(current, saved_mask);
5960 * Measure a series of task migrations and return the average
5961 * result. Since this code runs early during bootup the system
5962 * is 'undisturbed' and the average latency makes sense.
5964 * The algorithm in essence auto-detects the relevant cache-size,
5965 * so it will properly detect different cachesizes for different
5966 * cache-hierarchies, depending on how the CPUs are connected.
5968 * Architectures can prime the upper limit of the search range via
5969 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5971 static unsigned long long
5972 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5974 unsigned long long cost1, cost2;
5978 * Measure the migration cost of 'size' bytes, over an
5979 * average of 10 runs:
5981 * (We perturb the cache size by a small (0..4k)
5982 * value to compensate size/alignment related artifacts.
5983 * We also subtract the cost of the operation done on
5989 * dry run, to make sure we start off cache-cold on cpu1,
5990 * and to get any vmalloc pagefaults in advance:
5992 measure_one(cache, size, cpu1, cpu2);
5993 for (i = 0; i < ITERATIONS; i++)
5994 cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2);
5996 measure_one(cache, size, cpu2, cpu1);
5997 for (i = 0; i < ITERATIONS; i++)
5998 cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1);
6001 * (We measure the non-migrating [cached] cost on both
6002 * cpu1 and cpu2, to handle CPUs with different speeds)
6006 measure_one(cache, size, cpu1, cpu1);
6007 for (i = 0; i < ITERATIONS; i++)
6008 cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1);
6010 measure_one(cache, size, cpu2, cpu2);
6011 for (i = 0; i < ITERATIONS; i++)
6012 cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2);
6015 * Get the per-iteration migration cost:
6017 do_div(cost1, 2 * ITERATIONS);
6018 do_div(cost2, 2 * ITERATIONS);
6020 return cost1 - cost2;
6023 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
6025 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
6026 unsigned int max_size, size, size_found = 0;
6027 long long cost = 0, prev_cost;
6031 * Search from max_cache_size*5 down to 64K - the real relevant
6032 * cachesize has to lie somewhere inbetween.
6034 if (max_cache_size) {
6035 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
6036 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
6039 * Since we have no estimation about the relevant
6042 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
6043 size = MIN_CACHE_SIZE;
6046 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
6047 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
6052 * Allocate the working set:
6054 cache = vmalloc(max_size);
6056 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size);
6057 return 1000000; /* return 1 msec on very small boxen */
6060 while (size <= max_size) {
6062 cost = measure_cost(cpu1, cpu2, cache, size);
6068 if (max_cost < cost) {
6074 * Calculate average fluctuation, we use this to prevent
6075 * noise from triggering an early break out of the loop:
6077 fluct = abs(cost - prev_cost);
6078 avg_fluct = (avg_fluct + fluct)/2;
6080 if (migration_debug)
6081 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6084 (long)cost / 1000000,
6085 ((long)cost / 100000) % 10,
6086 (long)max_cost / 1000000,
6087 ((long)max_cost / 100000) % 10,
6088 domain_distance(cpu1, cpu2),
6092 * If we iterated at least 20% past the previous maximum,
6093 * and the cost has dropped by more than 20% already,
6094 * (taking fluctuations into account) then we assume to
6095 * have found the maximum and break out of the loop early:
6097 if (size_found && (size*100 > size_found*SIZE_THRESH))
6098 if (cost+avg_fluct <= 0 ||
6099 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6101 if (migration_debug)
6102 printk("-> found max.\n");
6106 * Increase the cachesize in 10% steps:
6108 size = size * 10 / 9;
6111 if (migration_debug)
6112 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6113 cpu1, cpu2, size_found, max_cost);
6118 * A task is considered 'cache cold' if at least 2 times
6119 * the worst-case cost of migration has passed.
6121 * (this limit is only listened to if the load-balancing
6122 * situation is 'nice' - if there is a large imbalance we
6123 * ignore it for the sake of CPU utilization and
6124 * processing fairness.)
6126 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6129 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6131 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6132 unsigned long j0, j1, distance, max_distance = 0;
6133 struct sched_domain *sd;
6138 * First pass - calculate the cacheflush times:
6140 for_each_cpu_mask(cpu1, *cpu_map) {
6141 for_each_cpu_mask(cpu2, *cpu_map) {
6144 distance = domain_distance(cpu1, cpu2);
6145 max_distance = max(max_distance, distance);
6147 * No result cached yet?
6149 if (migration_cost[distance] == -1LL)
6150 migration_cost[distance] =
6151 measure_migration_cost(cpu1, cpu2);
6155 * Second pass - update the sched domain hierarchy with
6156 * the new cache-hot-time estimations:
6158 for_each_cpu_mask(cpu, *cpu_map) {
6160 for_each_domain(cpu, sd) {
6161 sd->cache_hot_time = migration_cost[distance];
6168 if (migration_debug)
6169 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6177 if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) {
6178 printk("migration_cost=");
6179 for (distance = 0; distance <= max_distance; distance++) {
6182 printk("%ld", (long)migration_cost[distance] / 1000);
6187 if (migration_debug)
6188 printk("migration: %ld seconds\n", (j1-j0) / HZ);
6191 * Move back to the original CPU. NUMA-Q gets confused
6192 * if we migrate to another quad during bootup.
6194 if (raw_smp_processor_id() != orig_cpu) {
6195 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6196 saved_mask = current->cpus_allowed;
6198 set_cpus_allowed(current, mask);
6199 set_cpus_allowed(current, saved_mask);
6206 * find_next_best_node - find the next node to include in a sched_domain
6207 * @node: node whose sched_domain we're building
6208 * @used_nodes: nodes already in the sched_domain
6210 * Find the next node to include in a given scheduling domain. Simply
6211 * finds the closest node not already in the @used_nodes map.
6213 * Should use nodemask_t.
6215 static int find_next_best_node(int node, unsigned long *used_nodes)
6217 int i, n, val, min_val, best_node = 0;
6221 for (i = 0; i < MAX_NUMNODES; i++) {
6222 /* Start at @node */
6223 n = (node + i) % MAX_NUMNODES;
6225 if (!nr_cpus_node(n))
6228 /* Skip already used nodes */
6229 if (test_bit(n, used_nodes))
6232 /* Simple min distance search */
6233 val = node_distance(node, n);
6235 if (val < min_val) {
6241 set_bit(best_node, used_nodes);
6246 * sched_domain_node_span - get a cpumask for a node's sched_domain
6247 * @node: node whose cpumask we're constructing
6248 * @size: number of nodes to include in this span
6250 * Given a node, construct a good cpumask for its sched_domain to span. It
6251 * should be one that prevents unnecessary balancing, but also spreads tasks
6254 static cpumask_t sched_domain_node_span(int node)
6256 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6257 cpumask_t span, nodemask;
6261 bitmap_zero(used_nodes, MAX_NUMNODES);
6263 nodemask = node_to_cpumask(node);
6264 cpus_or(span, span, nodemask);
6265 set_bit(node, used_nodes);
6267 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6268 int next_node = find_next_best_node(node, used_nodes);
6270 nodemask = node_to_cpumask(next_node);
6271 cpus_or(span, span, nodemask);
6278 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6281 * SMT sched-domains:
6283 #ifdef CONFIG_SCHED_SMT
6284 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6285 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6287 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6288 struct sched_group **sg)
6291 *sg = &per_cpu(sched_group_cpus, cpu);
6297 * multi-core sched-domains:
6299 #ifdef CONFIG_SCHED_MC
6300 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6301 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6304 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6305 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6306 struct sched_group **sg)
6309 cpumask_t mask = cpu_sibling_map[cpu];
6310 cpus_and(mask, mask, *cpu_map);
6311 group = first_cpu(mask);
6313 *sg = &per_cpu(sched_group_core, group);
6316 #elif defined(CONFIG_SCHED_MC)
6317 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6318 struct sched_group **sg)
6321 *sg = &per_cpu(sched_group_core, cpu);
6326 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6327 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6329 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6330 struct sched_group **sg)
6333 #ifdef CONFIG_SCHED_MC
6334 cpumask_t mask = cpu_coregroup_map(cpu);
6335 cpus_and(mask, mask, *cpu_map);
6336 group = first_cpu(mask);
6337 #elif defined(CONFIG_SCHED_SMT)
6338 cpumask_t mask = cpu_sibling_map[cpu];
6339 cpus_and(mask, mask, *cpu_map);
6340 group = first_cpu(mask);
6345 *sg = &per_cpu(sched_group_phys, group);
6351 * The init_sched_build_groups can't handle what we want to do with node
6352 * groups, so roll our own. Now each node has its own list of groups which
6353 * gets dynamically allocated.
6355 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6356 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6358 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6359 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6361 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6362 struct sched_group **sg)
6364 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6367 cpus_and(nodemask, nodemask, *cpu_map);
6368 group = first_cpu(nodemask);
6371 *sg = &per_cpu(sched_group_allnodes, group);
6375 static void init_numa_sched_groups_power(struct sched_group *group_head)
6377 struct sched_group *sg = group_head;
6383 for_each_cpu_mask(j, sg->cpumask) {
6384 struct sched_domain *sd;
6386 sd = &per_cpu(phys_domains, j);
6387 if (j != first_cpu(sd->groups->cpumask)) {
6389 * Only add "power" once for each
6395 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6398 if (sg != group_head)
6404 /* Free memory allocated for various sched_group structures */
6405 static void free_sched_groups(const cpumask_t *cpu_map)
6409 for_each_cpu_mask(cpu, *cpu_map) {
6410 struct sched_group **sched_group_nodes
6411 = sched_group_nodes_bycpu[cpu];
6413 if (!sched_group_nodes)
6416 for (i = 0; i < MAX_NUMNODES; i++) {
6417 cpumask_t nodemask = node_to_cpumask(i);
6418 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6420 cpus_and(nodemask, nodemask, *cpu_map);
6421 if (cpus_empty(nodemask))
6431 if (oldsg != sched_group_nodes[i])
6434 kfree(sched_group_nodes);
6435 sched_group_nodes_bycpu[cpu] = NULL;
6439 static void free_sched_groups(const cpumask_t *cpu_map)
6445 * Initialize sched groups cpu_power.
6447 * cpu_power indicates the capacity of sched group, which is used while
6448 * distributing the load between different sched groups in a sched domain.
6449 * Typically cpu_power for all the groups in a sched domain will be same unless
6450 * there are asymmetries in the topology. If there are asymmetries, group
6451 * having more cpu_power will pickup more load compared to the group having
6454 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6455 * the maximum number of tasks a group can handle in the presence of other idle
6456 * or lightly loaded groups in the same sched domain.
6458 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6460 struct sched_domain *child;
6461 struct sched_group *group;
6463 WARN_ON(!sd || !sd->groups);
6465 if (cpu != first_cpu(sd->groups->cpumask))
6470 sd->groups->__cpu_power = 0;
6473 * For perf policy, if the groups in child domain share resources
6474 * (for example cores sharing some portions of the cache hierarchy
6475 * or SMT), then set this domain groups cpu_power such that each group
6476 * can handle only one task, when there are other idle groups in the
6477 * same sched domain.
6479 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6481 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6482 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6487 * add cpu_power of each child group to this groups cpu_power
6489 group = child->groups;
6491 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6492 group = group->next;
6493 } while (group != child->groups);
6497 * Build sched domains for a given set of cpus and attach the sched domains
6498 * to the individual cpus
6500 static int build_sched_domains(const cpumask_t *cpu_map)
6503 struct sched_domain *sd;
6505 struct sched_group **sched_group_nodes = NULL;
6506 int sd_allnodes = 0;
6509 * Allocate the per-node list of sched groups
6511 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6513 if (!sched_group_nodes) {
6514 printk(KERN_WARNING "Can not alloc sched group node list\n");
6517 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6521 * Set up domains for cpus specified by the cpu_map.
6523 for_each_cpu_mask(i, *cpu_map) {
6524 struct sched_domain *sd = NULL, *p;
6525 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6527 cpus_and(nodemask, nodemask, *cpu_map);
6530 if (cpus_weight(*cpu_map)
6531 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6532 sd = &per_cpu(allnodes_domains, i);
6533 *sd = SD_ALLNODES_INIT;
6534 sd->span = *cpu_map;
6535 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6541 sd = &per_cpu(node_domains, i);
6543 sd->span = sched_domain_node_span(cpu_to_node(i));
6547 cpus_and(sd->span, sd->span, *cpu_map);
6551 sd = &per_cpu(phys_domains, i);
6553 sd->span = nodemask;
6557 cpu_to_phys_group(i, cpu_map, &sd->groups);
6559 #ifdef CONFIG_SCHED_MC
6561 sd = &per_cpu(core_domains, i);
6563 sd->span = cpu_coregroup_map(i);
6564 cpus_and(sd->span, sd->span, *cpu_map);
6567 cpu_to_core_group(i, cpu_map, &sd->groups);
6570 #ifdef CONFIG_SCHED_SMT
6572 sd = &per_cpu(cpu_domains, i);
6573 *sd = SD_SIBLING_INIT;
6574 sd->span = cpu_sibling_map[i];
6575 cpus_and(sd->span, sd->span, *cpu_map);
6578 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6582 #ifdef CONFIG_SCHED_SMT
6583 /* Set up CPU (sibling) groups */
6584 for_each_cpu_mask(i, *cpu_map) {
6585 cpumask_t this_sibling_map = cpu_sibling_map[i];
6586 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6587 if (i != first_cpu(this_sibling_map))
6590 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6594 #ifdef CONFIG_SCHED_MC
6595 /* Set up multi-core groups */
6596 for_each_cpu_mask(i, *cpu_map) {
6597 cpumask_t this_core_map = cpu_coregroup_map(i);
6598 cpus_and(this_core_map, this_core_map, *cpu_map);
6599 if (i != first_cpu(this_core_map))
6601 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6606 /* Set up physical groups */
6607 for (i = 0; i < MAX_NUMNODES; i++) {
6608 cpumask_t nodemask = node_to_cpumask(i);
6610 cpus_and(nodemask, nodemask, *cpu_map);
6611 if (cpus_empty(nodemask))
6614 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6618 /* Set up node groups */
6620 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6622 for (i = 0; i < MAX_NUMNODES; i++) {
6623 /* Set up node groups */
6624 struct sched_group *sg, *prev;
6625 cpumask_t nodemask = node_to_cpumask(i);
6626 cpumask_t domainspan;
6627 cpumask_t covered = CPU_MASK_NONE;
6630 cpus_and(nodemask, nodemask, *cpu_map);
6631 if (cpus_empty(nodemask)) {
6632 sched_group_nodes[i] = NULL;
6636 domainspan = sched_domain_node_span(i);
6637 cpus_and(domainspan, domainspan, *cpu_map);
6639 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6641 printk(KERN_WARNING "Can not alloc domain group for "
6645 sched_group_nodes[i] = sg;
6646 for_each_cpu_mask(j, nodemask) {
6647 struct sched_domain *sd;
6648 sd = &per_cpu(node_domains, j);
6651 sg->__cpu_power = 0;
6652 sg->cpumask = nodemask;
6654 cpus_or(covered, covered, nodemask);
6657 for (j = 0; j < MAX_NUMNODES; j++) {
6658 cpumask_t tmp, notcovered;
6659 int n = (i + j) % MAX_NUMNODES;
6661 cpus_complement(notcovered, covered);
6662 cpus_and(tmp, notcovered, *cpu_map);
6663 cpus_and(tmp, tmp, domainspan);
6664 if (cpus_empty(tmp))
6667 nodemask = node_to_cpumask(n);
6668 cpus_and(tmp, tmp, nodemask);
6669 if (cpus_empty(tmp))
6672 sg = kmalloc_node(sizeof(struct sched_group),
6676 "Can not alloc domain group for node %d\n", j);
6679 sg->__cpu_power = 0;
6681 sg->next = prev->next;
6682 cpus_or(covered, covered, tmp);
6689 /* Calculate CPU power for physical packages and nodes */
6690 #ifdef CONFIG_SCHED_SMT
6691 for_each_cpu_mask(i, *cpu_map) {
6692 sd = &per_cpu(cpu_domains, i);
6693 init_sched_groups_power(i, sd);
6696 #ifdef CONFIG_SCHED_MC
6697 for_each_cpu_mask(i, *cpu_map) {
6698 sd = &per_cpu(core_domains, i);
6699 init_sched_groups_power(i, sd);
6703 for_each_cpu_mask(i, *cpu_map) {
6704 sd = &per_cpu(phys_domains, i);
6705 init_sched_groups_power(i, sd);
6709 for (i = 0; i < MAX_NUMNODES; i++)
6710 init_numa_sched_groups_power(sched_group_nodes[i]);
6713 struct sched_group *sg;
6715 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6716 init_numa_sched_groups_power(sg);
6720 /* Attach the domains */
6721 for_each_cpu_mask(i, *cpu_map) {
6722 struct sched_domain *sd;
6723 #ifdef CONFIG_SCHED_SMT
6724 sd = &per_cpu(cpu_domains, i);
6725 #elif defined(CONFIG_SCHED_MC)
6726 sd = &per_cpu(core_domains, i);
6728 sd = &per_cpu(phys_domains, i);
6730 cpu_attach_domain(sd, i);
6733 * Tune cache-hot values:
6735 calibrate_migration_costs(cpu_map);
6741 free_sched_groups(cpu_map);
6746 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6748 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6750 cpumask_t cpu_default_map;
6754 * Setup mask for cpus without special case scheduling requirements.
6755 * For now this just excludes isolated cpus, but could be used to
6756 * exclude other special cases in the future.
6758 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6760 err = build_sched_domains(&cpu_default_map);
6765 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6767 free_sched_groups(cpu_map);
6771 * Detach sched domains from a group of cpus specified in cpu_map
6772 * These cpus will now be attached to the NULL domain
6774 static void detach_destroy_domains(const cpumask_t *cpu_map)
6778 for_each_cpu_mask(i, *cpu_map)
6779 cpu_attach_domain(NULL, i);
6780 synchronize_sched();
6781 arch_destroy_sched_domains(cpu_map);
6785 * Partition sched domains as specified by the cpumasks below.
6786 * This attaches all cpus from the cpumasks to the NULL domain,
6787 * waits for a RCU quiescent period, recalculates sched
6788 * domain information and then attaches them back to the
6789 * correct sched domains
6790 * Call with hotplug lock held
6792 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6794 cpumask_t change_map;
6797 cpus_and(*partition1, *partition1, cpu_online_map);
6798 cpus_and(*partition2, *partition2, cpu_online_map);
6799 cpus_or(change_map, *partition1, *partition2);
6801 /* Detach sched domains from all of the affected cpus */
6802 detach_destroy_domains(&change_map);
6803 if (!cpus_empty(*partition1))
6804 err = build_sched_domains(partition1);
6805 if (!err && !cpus_empty(*partition2))
6806 err = build_sched_domains(partition2);
6811 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6812 int arch_reinit_sched_domains(void)
6817 detach_destroy_domains(&cpu_online_map);
6818 err = arch_init_sched_domains(&cpu_online_map);
6819 unlock_cpu_hotplug();
6824 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6828 if (buf[0] != '0' && buf[0] != '1')
6832 sched_smt_power_savings = (buf[0] == '1');
6834 sched_mc_power_savings = (buf[0] == '1');
6836 ret = arch_reinit_sched_domains();
6838 return ret ? ret : count;
6841 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6845 #ifdef CONFIG_SCHED_SMT
6847 err = sysfs_create_file(&cls->kset.kobj,
6848 &attr_sched_smt_power_savings.attr);
6850 #ifdef CONFIG_SCHED_MC
6851 if (!err && mc_capable())
6852 err = sysfs_create_file(&cls->kset.kobj,
6853 &attr_sched_mc_power_savings.attr);
6859 #ifdef CONFIG_SCHED_MC
6860 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6862 return sprintf(page, "%u\n", sched_mc_power_savings);
6864 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6865 const char *buf, size_t count)
6867 return sched_power_savings_store(buf, count, 0);
6869 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6870 sched_mc_power_savings_store);
6873 #ifdef CONFIG_SCHED_SMT
6874 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6876 return sprintf(page, "%u\n", sched_smt_power_savings);
6878 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6879 const char *buf, size_t count)
6881 return sched_power_savings_store(buf, count, 1);
6883 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6884 sched_smt_power_savings_store);
6888 * Force a reinitialization of the sched domains hierarchy. The domains
6889 * and groups cannot be updated in place without racing with the balancing
6890 * code, so we temporarily attach all running cpus to the NULL domain
6891 * which will prevent rebalancing while the sched domains are recalculated.
6893 static int update_sched_domains(struct notifier_block *nfb,
6894 unsigned long action, void *hcpu)
6897 case CPU_UP_PREPARE:
6898 case CPU_DOWN_PREPARE:
6899 detach_destroy_domains(&cpu_online_map);
6902 case CPU_UP_CANCELED:
6903 case CPU_DOWN_FAILED:
6907 * Fall through and re-initialise the domains.
6914 /* The hotplug lock is already held by cpu_up/cpu_down */
6915 arch_init_sched_domains(&cpu_online_map);
6920 void __init sched_init_smp(void)
6922 cpumask_t non_isolated_cpus;
6925 arch_init_sched_domains(&cpu_online_map);
6926 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6927 if (cpus_empty(non_isolated_cpus))
6928 cpu_set(smp_processor_id(), non_isolated_cpus);
6929 unlock_cpu_hotplug();
6930 /* XXX: Theoretical race here - CPU may be hotplugged now */
6931 hotcpu_notifier(update_sched_domains, 0);
6933 /* Move init over to a non-isolated CPU */
6934 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6938 void __init sched_init_smp(void)
6941 #endif /* CONFIG_SMP */
6943 int in_sched_functions(unsigned long addr)
6945 /* Linker adds these: start and end of __sched functions */
6946 extern char __sched_text_start[], __sched_text_end[];
6948 return in_lock_functions(addr) ||
6949 (addr >= (unsigned long)__sched_text_start
6950 && addr < (unsigned long)__sched_text_end);
6953 void __init sched_init(void)
6956 int highest_cpu = 0;
6958 for_each_possible_cpu(i) {
6959 struct prio_array *array;
6963 spin_lock_init(&rq->lock);
6964 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6966 rq->active = rq->arrays;
6967 rq->expired = rq->arrays + 1;
6968 rq->best_expired_prio = MAX_PRIO;
6972 for (j = 1; j < 3; j++)
6973 rq->cpu_load[j] = 0;
6974 rq->active_balance = 0;
6977 rq->migration_thread = NULL;
6978 INIT_LIST_HEAD(&rq->migration_queue);
6980 atomic_set(&rq->nr_iowait, 0);
6982 for (j = 0; j < 2; j++) {
6983 array = rq->arrays + j;
6984 for (k = 0; k < MAX_PRIO; k++) {
6985 INIT_LIST_HEAD(array->queue + k);
6986 __clear_bit(k, array->bitmap);
6988 // delimiter for bitsearch
6989 __set_bit(MAX_PRIO, array->bitmap);
6994 set_load_weight(&init_task);
6997 nr_cpu_ids = highest_cpu + 1;
6998 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7001 #ifdef CONFIG_RT_MUTEXES
7002 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7006 * The boot idle thread does lazy MMU switching as well:
7008 atomic_inc(&init_mm.mm_count);
7009 enter_lazy_tlb(&init_mm, current);
7012 * Make us the idle thread. Technically, schedule() should not be
7013 * called from this thread, however somewhere below it might be,
7014 * but because we are the idle thread, we just pick up running again
7015 * when this runqueue becomes "idle".
7017 init_idle(current, smp_processor_id());
7020 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7021 void __might_sleep(char *file, int line)
7024 static unsigned long prev_jiffy; /* ratelimiting */
7026 if ((in_atomic() || irqs_disabled()) &&
7027 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7028 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7030 prev_jiffy = jiffies;
7031 printk(KERN_ERR "BUG: sleeping function called from invalid"
7032 " context at %s:%d\n", file, line);
7033 printk("in_atomic():%d, irqs_disabled():%d\n",
7034 in_atomic(), irqs_disabled());
7035 debug_show_held_locks(current);
7036 if (irqs_disabled())
7037 print_irqtrace_events(current);
7042 EXPORT_SYMBOL(__might_sleep);
7045 #ifdef CONFIG_MAGIC_SYSRQ
7046 void normalize_rt_tasks(void)
7048 struct prio_array *array;
7049 struct task_struct *p;
7050 unsigned long flags;
7053 read_lock_irq(&tasklist_lock);
7054 for_each_process(p) {
7058 spin_lock_irqsave(&p->pi_lock, flags);
7059 rq = __task_rq_lock(p);
7063 deactivate_task(p, task_rq(p));
7064 __setscheduler(p, SCHED_NORMAL, 0);
7066 __activate_task(p, task_rq(p));
7067 resched_task(rq->curr);
7070 __task_rq_unlock(rq);
7071 spin_unlock_irqrestore(&p->pi_lock, flags);
7073 read_unlock_irq(&tasklist_lock);
7076 #endif /* CONFIG_MAGIC_SYSRQ */
7080 * These functions are only useful for the IA64 MCA handling.
7082 * They can only be called when the whole system has been
7083 * stopped - every CPU needs to be quiescent, and no scheduling
7084 * activity can take place. Using them for anything else would
7085 * be a serious bug, and as a result, they aren't even visible
7086 * under any other configuration.
7090 * curr_task - return the current task for a given cpu.
7091 * @cpu: the processor in question.
7093 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7095 struct task_struct *curr_task(int cpu)
7097 return cpu_curr(cpu);
7101 * set_curr_task - set the current task for a given cpu.
7102 * @cpu: the processor in question.
7103 * @p: the task pointer to set.
7105 * Description: This function must only be used when non-maskable interrupts
7106 * are serviced on a separate stack. It allows the architecture to switch the
7107 * notion of the current task on a cpu in a non-blocking manner. This function
7108 * must be called with all CPU's synchronized, and interrupts disabled, the
7109 * and caller must save the original value of the current task (see
7110 * curr_task() above) and restore that value before reenabling interrupts and
7111 * re-starting the system.
7113 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7115 void set_curr_task(int cpu, struct task_struct *p)