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>
57 #include <asm/unistd.h>
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
64 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
73 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
78 * Some helpers for converting nanosecond timing to jiffy resolution
80 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
84 * These are the 'tuning knobs' of the scheduler:
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
90 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91 #define DEF_TIMESLICE (100 * HZ / 1000)
92 #define ON_RUNQUEUE_WEIGHT 30
93 #define CHILD_PENALTY 95
94 #define PARENT_PENALTY 100
96 #define PRIO_BONUS_RATIO 25
97 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98 #define INTERACTIVE_DELTA 2
99 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
101 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
109 * This part scales the interactivity limit depending on niceness.
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
131 #define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
135 #define GRANULARITY (10 * HZ / 1000 ? : 1)
138 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
142 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
146 #define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
153 #define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
156 #define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
160 #define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
163 #define SCALE_PRIO(x, prio) \
164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
166 static unsigned int static_prio_timeslice(int static_prio)
168 if (static_prio < NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
171 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
183 static inline unsigned int task_timeslice(struct task_struct *p)
185 return static_prio_timeslice(p->static_prio);
189 * These are the runqueue data structures:
193 unsigned int nr_active;
194 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
195 struct list_head queue[MAX_PRIO];
199 * This is the main, per-CPU runqueue data structure.
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
212 unsigned long nr_running;
213 unsigned long raw_weighted_load;
215 unsigned long cpu_load[3];
217 unsigned long long nr_switches;
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
225 unsigned long nr_uninterruptible;
227 unsigned long expired_timestamp;
228 /* Cached timestamp set by update_cpu_clock() */
229 unsigned long long most_recent_timestamp;
230 struct task_struct *curr, *idle;
231 unsigned long next_balance;
232 struct mm_struct *prev_mm;
233 struct prio_array *active, *expired, arrays[2];
234 int best_expired_prio;
238 struct sched_domain *sd;
240 /* For active balancing */
243 int cpu; /* cpu of this runqueue */
245 struct task_struct *migration_thread;
246 struct list_head migration_queue;
249 #ifdef CONFIG_SCHEDSTATS
251 struct sched_info rq_sched_info;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty;
255 unsigned long yld_act_empty;
256 unsigned long yld_both_empty;
257 unsigned long yld_cnt;
259 /* schedule() stats */
260 unsigned long sched_switch;
261 unsigned long sched_cnt;
262 unsigned long sched_goidle;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt;
266 unsigned long ttwu_local;
268 struct lock_class_key rq_lock_key;
271 static DEFINE_PER_CPU(struct rq, runqueues);
273 static inline int cpu_of(struct rq *rq)
283 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
284 * See detach_destroy_domains: synchronize_sched for details.
286 * The domain tree of any CPU may only be accessed from within
287 * preempt-disabled sections.
289 #define for_each_domain(cpu, __sd) \
290 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
292 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
293 #define this_rq() (&__get_cpu_var(runqueues))
294 #define task_rq(p) cpu_rq(task_cpu(p))
295 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
297 #ifndef prepare_arch_switch
298 # define prepare_arch_switch(next) do { } while (0)
300 #ifndef finish_arch_switch
301 # define finish_arch_switch(prev) do { } while (0)
304 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
305 static inline int task_running(struct rq *rq, struct task_struct *p)
307 return rq->curr == p;
310 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
314 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
316 #ifdef CONFIG_DEBUG_SPINLOCK
317 /* this is a valid case when another task releases the spinlock */
318 rq->lock.owner = current;
321 * If we are tracking spinlock dependencies then we have to
322 * fix up the runqueue lock - which gets 'carried over' from
325 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
327 spin_unlock_irq(&rq->lock);
330 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
331 static inline int task_running(struct rq *rq, struct task_struct *p)
336 return rq->curr == p;
340 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
344 * We can optimise this out completely for !SMP, because the
345 * SMP rebalancing from interrupt is the only thing that cares
350 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
351 spin_unlock_irq(&rq->lock);
353 spin_unlock(&rq->lock);
357 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
361 * After ->oncpu is cleared, the task can be moved to a different CPU.
362 * We must ensure this doesn't happen until the switch is completely
368 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
372 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
375 * __task_rq_lock - lock the runqueue a given task resides on.
376 * Must be called interrupts disabled.
378 static inline struct rq *__task_rq_lock(struct task_struct *p)
385 spin_lock(&rq->lock);
386 if (unlikely(rq != task_rq(p))) {
387 spin_unlock(&rq->lock);
388 goto repeat_lock_task;
394 * task_rq_lock - lock the runqueue a given task resides on and disable
395 * interrupts. Note the ordering: we can safely lookup the task_rq without
396 * explicitly disabling preemption.
398 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
404 local_irq_save(*flags);
406 spin_lock(&rq->lock);
407 if (unlikely(rq != task_rq(p))) {
408 spin_unlock_irqrestore(&rq->lock, *flags);
409 goto repeat_lock_task;
414 static inline void __task_rq_unlock(struct rq *rq)
417 spin_unlock(&rq->lock);
420 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
423 spin_unlock_irqrestore(&rq->lock, *flags);
426 #ifdef CONFIG_SCHEDSTATS
428 * bump this up when changing the output format or the meaning of an existing
429 * format, so that tools can adapt (or abort)
431 #define SCHEDSTAT_VERSION 14
433 static int show_schedstat(struct seq_file *seq, void *v)
437 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
438 seq_printf(seq, "timestamp %lu\n", jiffies);
439 for_each_online_cpu(cpu) {
440 struct rq *rq = cpu_rq(cpu);
442 struct sched_domain *sd;
446 /* runqueue-specific stats */
448 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
449 cpu, rq->yld_both_empty,
450 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
451 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
452 rq->ttwu_cnt, rq->ttwu_local,
453 rq->rq_sched_info.cpu_time,
454 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
456 seq_printf(seq, "\n");
459 /* domain-specific stats */
461 for_each_domain(cpu, sd) {
462 enum idle_type itype;
463 char mask_str[NR_CPUS];
465 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
466 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
467 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
469 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
471 sd->lb_balanced[itype],
472 sd->lb_failed[itype],
473 sd->lb_imbalance[itype],
474 sd->lb_gained[itype],
475 sd->lb_hot_gained[itype],
476 sd->lb_nobusyq[itype],
477 sd->lb_nobusyg[itype]);
479 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
480 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
481 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
482 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
483 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
491 static int schedstat_open(struct inode *inode, struct file *file)
493 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
494 char *buf = kmalloc(size, GFP_KERNEL);
500 res = single_open(file, show_schedstat, NULL);
502 m = file->private_data;
510 const struct file_operations proc_schedstat_operations = {
511 .open = schedstat_open,
514 .release = single_release,
518 * Expects runqueue lock to be held for atomicity of update
521 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
524 rq->rq_sched_info.run_delay += delta_jiffies;
525 rq->rq_sched_info.pcnt++;
530 * Expects runqueue lock to be held for atomicity of update
533 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
536 rq->rq_sched_info.cpu_time += delta_jiffies;
538 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
539 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
540 #else /* !CONFIG_SCHEDSTATS */
542 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
545 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
547 # define schedstat_inc(rq, field) do { } while (0)
548 # define schedstat_add(rq, field, amt) do { } while (0)
552 * this_rq_lock - lock this runqueue and disable interrupts.
554 static inline struct rq *this_rq_lock(void)
561 spin_lock(&rq->lock);
566 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
568 * Called when a process is dequeued from the active array and given
569 * the cpu. We should note that with the exception of interactive
570 * tasks, the expired queue will become the active queue after the active
571 * queue is empty, without explicitly dequeuing and requeuing tasks in the
572 * expired queue. (Interactive tasks may be requeued directly to the
573 * active queue, thus delaying tasks in the expired queue from running;
574 * see scheduler_tick()).
576 * This function is only called from sched_info_arrive(), rather than
577 * dequeue_task(). Even though a task may be queued and dequeued multiple
578 * times as it is shuffled about, we're really interested in knowing how
579 * long it was from the *first* time it was queued to the time that it
582 static inline void sched_info_dequeued(struct task_struct *t)
584 t->sched_info.last_queued = 0;
588 * Called when a task finally hits the cpu. We can now calculate how
589 * long it was waiting to run. We also note when it began so that we
590 * can keep stats on how long its timeslice is.
592 static void sched_info_arrive(struct task_struct *t)
594 unsigned long now = jiffies, delta_jiffies = 0;
596 if (t->sched_info.last_queued)
597 delta_jiffies = now - t->sched_info.last_queued;
598 sched_info_dequeued(t);
599 t->sched_info.run_delay += delta_jiffies;
600 t->sched_info.last_arrival = now;
601 t->sched_info.pcnt++;
603 rq_sched_info_arrive(task_rq(t), delta_jiffies);
607 * Called when a process is queued into either the active or expired
608 * array. The time is noted and later used to determine how long we
609 * had to wait for us to reach the cpu. Since the expired queue will
610 * become the active queue after active queue is empty, without dequeuing
611 * and requeuing any tasks, we are interested in queuing to either. It
612 * is unusual but not impossible for tasks to be dequeued and immediately
613 * requeued in the same or another array: this can happen in sched_yield(),
614 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
617 * This function is only called from enqueue_task(), but also only updates
618 * the timestamp if it is already not set. It's assumed that
619 * sched_info_dequeued() will clear that stamp when appropriate.
621 static inline void sched_info_queued(struct task_struct *t)
623 if (unlikely(sched_info_on()))
624 if (!t->sched_info.last_queued)
625 t->sched_info.last_queued = jiffies;
629 * Called when a process ceases being the active-running process, either
630 * voluntarily or involuntarily. Now we can calculate how long we ran.
632 static inline void sched_info_depart(struct task_struct *t)
634 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
636 t->sched_info.cpu_time += delta_jiffies;
637 rq_sched_info_depart(task_rq(t), delta_jiffies);
641 * Called when tasks are switched involuntarily due, typically, to expiring
642 * their time slice. (This may also be called when switching to or from
643 * the idle task.) We are only called when prev != next.
646 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
648 struct rq *rq = task_rq(prev);
651 * prev now departs the cpu. It's not interesting to record
652 * stats about how efficient we were at scheduling the idle
655 if (prev != rq->idle)
656 sched_info_depart(prev);
658 if (next != rq->idle)
659 sched_info_arrive(next);
662 sched_info_switch(struct task_struct *prev, struct task_struct *next)
664 if (unlikely(sched_info_on()))
665 __sched_info_switch(prev, next);
668 #define sched_info_queued(t) do { } while (0)
669 #define sched_info_switch(t, next) do { } while (0)
670 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
673 * Adding/removing a task to/from a priority array:
675 static void dequeue_task(struct task_struct *p, struct prio_array *array)
678 list_del(&p->run_list);
679 if (list_empty(array->queue + p->prio))
680 __clear_bit(p->prio, array->bitmap);
683 static void enqueue_task(struct task_struct *p, struct prio_array *array)
685 sched_info_queued(p);
686 list_add_tail(&p->run_list, array->queue + p->prio);
687 __set_bit(p->prio, array->bitmap);
693 * Put task to the end of the run list without the overhead of dequeue
694 * followed by enqueue.
696 static void requeue_task(struct task_struct *p, struct prio_array *array)
698 list_move_tail(&p->run_list, array->queue + p->prio);
702 enqueue_task_head(struct task_struct *p, struct prio_array *array)
704 list_add(&p->run_list, array->queue + p->prio);
705 __set_bit(p->prio, array->bitmap);
711 * __normal_prio - return the priority that is based on the static
712 * priority but is modified by bonuses/penalties.
714 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
715 * into the -5 ... 0 ... +5 bonus/penalty range.
717 * We use 25% of the full 0...39 priority range so that:
719 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
720 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
722 * Both properties are important to certain workloads.
725 static inline int __normal_prio(struct task_struct *p)
729 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
731 prio = p->static_prio - bonus;
732 if (prio < MAX_RT_PRIO)
734 if (prio > MAX_PRIO-1)
740 * To aid in avoiding the subversion of "niceness" due to uneven distribution
741 * of tasks with abnormal "nice" values across CPUs the contribution that
742 * each task makes to its run queue's load is weighted according to its
743 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
744 * scaled version of the new time slice allocation that they receive on time
749 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
750 * If static_prio_timeslice() is ever changed to break this assumption then
751 * this code will need modification
753 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
754 #define LOAD_WEIGHT(lp) \
755 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
756 #define PRIO_TO_LOAD_WEIGHT(prio) \
757 LOAD_WEIGHT(static_prio_timeslice(prio))
758 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
759 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
761 static void set_load_weight(struct task_struct *p)
763 if (has_rt_policy(p)) {
765 if (p == task_rq(p)->migration_thread)
767 * The migration thread does the actual balancing.
768 * Giving its load any weight will skew balancing
774 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
776 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
780 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
782 rq->raw_weighted_load += p->load_weight;
786 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
788 rq->raw_weighted_load -= p->load_weight;
791 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
794 inc_raw_weighted_load(rq, p);
797 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
800 dec_raw_weighted_load(rq, p);
804 * Calculate the expected normal priority: i.e. priority
805 * without taking RT-inheritance into account. Might be
806 * boosted by interactivity modifiers. Changes upon fork,
807 * setprio syscalls, and whenever the interactivity
808 * estimator recalculates.
810 static inline int normal_prio(struct task_struct *p)
814 if (has_rt_policy(p))
815 prio = MAX_RT_PRIO-1 - p->rt_priority;
817 prio = __normal_prio(p);
822 * Calculate the current priority, i.e. the priority
823 * taken into account by the scheduler. This value might
824 * be boosted by RT tasks, or might be boosted by
825 * interactivity modifiers. Will be RT if the task got
826 * RT-boosted. If not then it returns p->normal_prio.
828 static int effective_prio(struct task_struct *p)
830 p->normal_prio = normal_prio(p);
832 * If we are RT tasks or we were boosted to RT priority,
833 * keep the priority unchanged. Otherwise, update priority
834 * to the normal priority:
836 if (!rt_prio(p->prio))
837 return p->normal_prio;
842 * __activate_task - move a task to the runqueue.
844 static void __activate_task(struct task_struct *p, struct rq *rq)
846 struct prio_array *target = rq->active;
849 target = rq->expired;
850 enqueue_task(p, target);
851 inc_nr_running(p, rq);
855 * __activate_idle_task - move idle task to the _front_ of runqueue.
857 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
859 enqueue_task_head(p, rq->active);
860 inc_nr_running(p, rq);
864 * Recalculate p->normal_prio and p->prio after having slept,
865 * updating the sleep-average too:
867 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
869 /* Caller must always ensure 'now >= p->timestamp' */
870 unsigned long sleep_time = now - p->timestamp;
875 if (likely(sleep_time > 0)) {
877 * This ceiling is set to the lowest priority that would allow
878 * a task to be reinserted into the active array on timeslice
881 unsigned long ceiling = INTERACTIVE_SLEEP(p);
883 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
885 * Prevents user tasks from achieving best priority
886 * with one single large enough sleep.
888 p->sleep_avg = ceiling;
890 * Using INTERACTIVE_SLEEP() as a ceiling places a
891 * nice(0) task 1ms sleep away from promotion, and
892 * gives it 700ms to round-robin with no chance of
893 * being demoted. This is more than generous, so
894 * mark this sleep as non-interactive to prevent the
895 * on-runqueue bonus logic from intervening should
896 * this task not receive cpu immediately.
898 p->sleep_type = SLEEP_NONINTERACTIVE;
901 * Tasks waking from uninterruptible sleep are
902 * limited in their sleep_avg rise as they
903 * are likely to be waiting on I/O
905 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
906 if (p->sleep_avg >= ceiling)
908 else if (p->sleep_avg + sleep_time >=
910 p->sleep_avg = ceiling;
916 * This code gives a bonus to interactive tasks.
918 * The boost works by updating the 'average sleep time'
919 * value here, based on ->timestamp. The more time a
920 * task spends sleeping, the higher the average gets -
921 * and the higher the priority boost gets as well.
923 p->sleep_avg += sleep_time;
926 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
927 p->sleep_avg = NS_MAX_SLEEP_AVG;
930 return effective_prio(p);
934 * activate_task - move a task to the runqueue and do priority recalculation
936 * Update all the scheduling statistics stuff. (sleep average
937 * calculation, priority modifiers, etc.)
939 static void activate_task(struct task_struct *p, struct rq *rq, int local)
941 unsigned long long now;
946 /* Compensate for drifting sched_clock */
947 struct rq *this_rq = this_rq();
948 now = (now - this_rq->most_recent_timestamp)
949 + rq->most_recent_timestamp;
954 * Sleep time is in units of nanosecs, so shift by 20 to get a
955 * milliseconds-range estimation of the amount of time that the task
958 if (unlikely(prof_on == SLEEP_PROFILING)) {
959 if (p->state == TASK_UNINTERRUPTIBLE)
960 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
961 (now - p->timestamp) >> 20);
965 p->prio = recalc_task_prio(p, now);
968 * This checks to make sure it's not an uninterruptible task
969 * that is now waking up.
971 if (p->sleep_type == SLEEP_NORMAL) {
973 * Tasks which were woken up by interrupts (ie. hw events)
974 * are most likely of interactive nature. So we give them
975 * the credit of extending their sleep time to the period
976 * of time they spend on the runqueue, waiting for execution
977 * on a CPU, first time around:
980 p->sleep_type = SLEEP_INTERRUPTED;
983 * Normal first-time wakeups get a credit too for
984 * on-runqueue time, but it will be weighted down:
986 p->sleep_type = SLEEP_INTERACTIVE;
991 __activate_task(p, rq);
995 * deactivate_task - remove a task from the runqueue.
997 static void deactivate_task(struct task_struct *p, struct rq *rq)
999 dec_nr_running(p, rq);
1000 dequeue_task(p, p->array);
1005 * resched_task - mark a task 'to be rescheduled now'.
1007 * On UP this means the setting of the need_resched flag, on SMP it
1008 * might also involve a cross-CPU call to trigger the scheduler on
1013 #ifndef tsk_is_polling
1014 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1017 static void resched_task(struct task_struct *p)
1021 assert_spin_locked(&task_rq(p)->lock);
1023 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1026 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1029 if (cpu == smp_processor_id())
1032 /* NEED_RESCHED must be visible before we test polling */
1034 if (!tsk_is_polling(p))
1035 smp_send_reschedule(cpu);
1038 static inline void resched_task(struct task_struct *p)
1040 assert_spin_locked(&task_rq(p)->lock);
1041 set_tsk_need_resched(p);
1046 * task_curr - is this task currently executing on a CPU?
1047 * @p: the task in question.
1049 inline int task_curr(const struct task_struct *p)
1051 return cpu_curr(task_cpu(p)) == p;
1054 /* Used instead of source_load when we know the type == 0 */
1055 unsigned long weighted_cpuload(const int cpu)
1057 return cpu_rq(cpu)->raw_weighted_load;
1061 struct migration_req {
1062 struct list_head list;
1064 struct task_struct *task;
1067 struct completion done;
1071 * The task's runqueue lock must be held.
1072 * Returns true if you have to wait for migration thread.
1075 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1077 struct rq *rq = task_rq(p);
1080 * If the task is not on a runqueue (and not running), then
1081 * it is sufficient to simply update the task's cpu field.
1083 if (!p->array && !task_running(rq, p)) {
1084 set_task_cpu(p, dest_cpu);
1088 init_completion(&req->done);
1090 req->dest_cpu = dest_cpu;
1091 list_add(&req->list, &rq->migration_queue);
1097 * wait_task_inactive - wait for a thread to unschedule.
1099 * The caller must ensure that the task *will* unschedule sometime soon,
1100 * else this function might spin for a *long* time. This function can't
1101 * be called with interrupts off, or it may introduce deadlock with
1102 * smp_call_function() if an IPI is sent by the same process we are
1103 * waiting to become inactive.
1105 void wait_task_inactive(struct task_struct *p)
1107 unsigned long flags;
1112 rq = task_rq_lock(p, &flags);
1113 /* Must be off runqueue entirely, not preempted. */
1114 if (unlikely(p->array || task_running(rq, p))) {
1115 /* If it's preempted, we yield. It could be a while. */
1116 preempted = !task_running(rq, p);
1117 task_rq_unlock(rq, &flags);
1123 task_rq_unlock(rq, &flags);
1127 * kick_process - kick a running thread to enter/exit the kernel
1128 * @p: the to-be-kicked thread
1130 * Cause a process which is running on another CPU to enter
1131 * kernel-mode, without any delay. (to get signals handled.)
1133 * NOTE: this function doesnt have to take the runqueue lock,
1134 * because all it wants to ensure is that the remote task enters
1135 * the kernel. If the IPI races and the task has been migrated
1136 * to another CPU then no harm is done and the purpose has been
1139 void kick_process(struct task_struct *p)
1145 if ((cpu != smp_processor_id()) && task_curr(p))
1146 smp_send_reschedule(cpu);
1151 * Return a low guess at the load of a migration-source cpu weighted
1152 * according to the scheduling class and "nice" value.
1154 * We want to under-estimate the load of migration sources, to
1155 * balance conservatively.
1157 static inline unsigned long source_load(int cpu, int type)
1159 struct rq *rq = cpu_rq(cpu);
1162 return rq->raw_weighted_load;
1164 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1168 * Return a high guess at the load of a migration-target cpu weighted
1169 * according to the scheduling class and "nice" value.
1171 static inline unsigned long target_load(int cpu, int type)
1173 struct rq *rq = cpu_rq(cpu);
1176 return rq->raw_weighted_load;
1178 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1182 * Return the average load per task on the cpu's run queue
1184 static inline unsigned long cpu_avg_load_per_task(int cpu)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long n = rq->nr_running;
1189 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1193 * find_idlest_group finds and returns the least busy CPU group within the
1196 static struct sched_group *
1197 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1199 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1200 unsigned long min_load = ULONG_MAX, this_load = 0;
1201 int load_idx = sd->forkexec_idx;
1202 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1205 unsigned long load, avg_load;
1209 /* Skip over this group if it has no CPUs allowed */
1210 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1213 local_group = cpu_isset(this_cpu, group->cpumask);
1215 /* Tally up the load of all CPUs in the group */
1218 for_each_cpu_mask(i, group->cpumask) {
1219 /* Bias balancing toward cpus of our domain */
1221 load = source_load(i, load_idx);
1223 load = target_load(i, load_idx);
1228 /* Adjust by relative CPU power of the group */
1229 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1232 this_load = avg_load;
1234 } else if (avg_load < min_load) {
1235 min_load = avg_load;
1239 group = group->next;
1240 } while (group != sd->groups);
1242 if (!idlest || 100*this_load < imbalance*min_load)
1248 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1251 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1254 unsigned long load, min_load = ULONG_MAX;
1258 /* Traverse only the allowed CPUs */
1259 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1261 for_each_cpu_mask(i, tmp) {
1262 load = weighted_cpuload(i);
1264 if (load < min_load || (load == min_load && i == this_cpu)) {
1274 * sched_balance_self: balance the current task (running on cpu) in domains
1275 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1278 * Balance, ie. select the least loaded group.
1280 * Returns the target CPU number, or the same CPU if no balancing is needed.
1282 * preempt must be disabled.
1284 static int sched_balance_self(int cpu, int flag)
1286 struct task_struct *t = current;
1287 struct sched_domain *tmp, *sd = NULL;
1289 for_each_domain(cpu, tmp) {
1291 * If power savings logic is enabled for a domain, stop there.
1293 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1295 if (tmp->flags & flag)
1301 struct sched_group *group;
1302 int new_cpu, weight;
1304 if (!(sd->flags & flag)) {
1310 group = find_idlest_group(sd, t, cpu);
1316 new_cpu = find_idlest_cpu(group, t, cpu);
1317 if (new_cpu == -1 || new_cpu == cpu) {
1318 /* Now try balancing at a lower domain level of cpu */
1323 /* Now try balancing at a lower domain level of new_cpu */
1326 weight = cpus_weight(span);
1327 for_each_domain(cpu, tmp) {
1328 if (weight <= cpus_weight(tmp->span))
1330 if (tmp->flags & flag)
1333 /* while loop will break here if sd == NULL */
1339 #endif /* CONFIG_SMP */
1342 * wake_idle() will wake a task on an idle cpu if task->cpu is
1343 * not idle and an idle cpu is available. The span of cpus to
1344 * search starts with cpus closest then further out as needed,
1345 * so we always favor a closer, idle cpu.
1347 * Returns the CPU we should wake onto.
1349 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1350 static int wake_idle(int cpu, struct task_struct *p)
1353 struct sched_domain *sd;
1359 for_each_domain(cpu, sd) {
1360 if (sd->flags & SD_WAKE_IDLE) {
1361 cpus_and(tmp, sd->span, p->cpus_allowed);
1362 for_each_cpu_mask(i, tmp) {
1373 static inline int wake_idle(int cpu, struct task_struct *p)
1380 * try_to_wake_up - wake up a thread
1381 * @p: the to-be-woken-up thread
1382 * @state: the mask of task states that can be woken
1383 * @sync: do a synchronous wakeup?
1385 * Put it on the run-queue if it's not already there. The "current"
1386 * thread is always on the run-queue (except when the actual
1387 * re-schedule is in progress), and as such you're allowed to do
1388 * the simpler "current->state = TASK_RUNNING" to mark yourself
1389 * runnable without the overhead of this.
1391 * returns failure only if the task is already active.
1393 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1395 int cpu, this_cpu, success = 0;
1396 unsigned long flags;
1400 struct sched_domain *sd, *this_sd = NULL;
1401 unsigned long load, this_load;
1405 rq = task_rq_lock(p, &flags);
1406 old_state = p->state;
1407 if (!(old_state & state))
1414 this_cpu = smp_processor_id();
1417 if (unlikely(task_running(rq, p)))
1422 schedstat_inc(rq, ttwu_cnt);
1423 if (cpu == this_cpu) {
1424 schedstat_inc(rq, ttwu_local);
1428 for_each_domain(this_cpu, sd) {
1429 if (cpu_isset(cpu, sd->span)) {
1430 schedstat_inc(sd, ttwu_wake_remote);
1436 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1440 * Check for affine wakeup and passive balancing possibilities.
1443 int idx = this_sd->wake_idx;
1444 unsigned int imbalance;
1446 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1448 load = source_load(cpu, idx);
1449 this_load = target_load(this_cpu, idx);
1451 new_cpu = this_cpu; /* Wake to this CPU if we can */
1453 if (this_sd->flags & SD_WAKE_AFFINE) {
1454 unsigned long tl = this_load;
1455 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1458 * If sync wakeup then subtract the (maximum possible)
1459 * effect of the currently running task from the load
1460 * of the current CPU:
1463 tl -= current->load_weight;
1466 tl + target_load(cpu, idx) <= tl_per_task) ||
1467 100*(tl + p->load_weight) <= imbalance*load) {
1469 * This domain has SD_WAKE_AFFINE and
1470 * p is cache cold in this domain, and
1471 * there is no bad imbalance.
1473 schedstat_inc(this_sd, ttwu_move_affine);
1479 * Start passive balancing when half the imbalance_pct
1482 if (this_sd->flags & SD_WAKE_BALANCE) {
1483 if (imbalance*this_load <= 100*load) {
1484 schedstat_inc(this_sd, ttwu_move_balance);
1490 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1492 new_cpu = wake_idle(new_cpu, p);
1493 if (new_cpu != cpu) {
1494 set_task_cpu(p, new_cpu);
1495 task_rq_unlock(rq, &flags);
1496 /* might preempt at this point */
1497 rq = task_rq_lock(p, &flags);
1498 old_state = p->state;
1499 if (!(old_state & state))
1504 this_cpu = smp_processor_id();
1509 #endif /* CONFIG_SMP */
1510 if (old_state == TASK_UNINTERRUPTIBLE) {
1511 rq->nr_uninterruptible--;
1513 * Tasks on involuntary sleep don't earn
1514 * sleep_avg beyond just interactive state.
1516 p->sleep_type = SLEEP_NONINTERACTIVE;
1520 * Tasks that have marked their sleep as noninteractive get
1521 * woken up with their sleep average not weighted in an
1524 if (old_state & TASK_NONINTERACTIVE)
1525 p->sleep_type = SLEEP_NONINTERACTIVE;
1528 activate_task(p, rq, cpu == this_cpu);
1530 * Sync wakeups (i.e. those types of wakeups where the waker
1531 * has indicated that it will leave the CPU in short order)
1532 * don't trigger a preemption, if the woken up task will run on
1533 * this cpu. (in this case the 'I will reschedule' promise of
1534 * the waker guarantees that the freshly woken up task is going
1535 * to be considered on this CPU.)
1537 if (!sync || cpu != this_cpu) {
1538 if (TASK_PREEMPTS_CURR(p, rq))
1539 resched_task(rq->curr);
1544 p->state = TASK_RUNNING;
1546 task_rq_unlock(rq, &flags);
1551 int fastcall wake_up_process(struct task_struct *p)
1553 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1554 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1556 EXPORT_SYMBOL(wake_up_process);
1558 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1560 return try_to_wake_up(p, state, 0);
1564 * Perform scheduler related setup for a newly forked process p.
1565 * p is forked by current.
1567 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1569 int cpu = get_cpu();
1572 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1574 set_task_cpu(p, cpu);
1577 * We mark the process as running here, but have not actually
1578 * inserted it onto the runqueue yet. This guarantees that
1579 * nobody will actually run it, and a signal or other external
1580 * event cannot wake it up and insert it on the runqueue either.
1582 p->state = TASK_RUNNING;
1585 * Make sure we do not leak PI boosting priority to the child:
1587 p->prio = current->normal_prio;
1589 INIT_LIST_HEAD(&p->run_list);
1591 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1592 if (unlikely(sched_info_on()))
1593 memset(&p->sched_info, 0, sizeof(p->sched_info));
1595 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1598 #ifdef CONFIG_PREEMPT
1599 /* Want to start with kernel preemption disabled. */
1600 task_thread_info(p)->preempt_count = 1;
1603 * Share the timeslice between parent and child, thus the
1604 * total amount of pending timeslices in the system doesn't change,
1605 * resulting in more scheduling fairness.
1607 local_irq_disable();
1608 p->time_slice = (current->time_slice + 1) >> 1;
1610 * The remainder of the first timeslice might be recovered by
1611 * the parent if the child exits early enough.
1613 p->first_time_slice = 1;
1614 current->time_slice >>= 1;
1615 p->timestamp = sched_clock();
1616 if (unlikely(!current->time_slice)) {
1618 * This case is rare, it happens when the parent has only
1619 * a single jiffy left from its timeslice. Taking the
1620 * runqueue lock is not a problem.
1622 current->time_slice = 1;
1630 * wake_up_new_task - wake up a newly created task for the first time.
1632 * This function will do some initial scheduler statistics housekeeping
1633 * that must be done for every newly created context, then puts the task
1634 * on the runqueue and wakes it.
1636 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1638 struct rq *rq, *this_rq;
1639 unsigned long flags;
1642 rq = task_rq_lock(p, &flags);
1643 BUG_ON(p->state != TASK_RUNNING);
1644 this_cpu = smp_processor_id();
1648 * We decrease the sleep average of forking parents
1649 * and children as well, to keep max-interactive tasks
1650 * from forking tasks that are max-interactive. The parent
1651 * (current) is done further down, under its lock.
1653 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1654 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1656 p->prio = effective_prio(p);
1658 if (likely(cpu == this_cpu)) {
1659 if (!(clone_flags & CLONE_VM)) {
1661 * The VM isn't cloned, so we're in a good position to
1662 * do child-runs-first in anticipation of an exec. This
1663 * usually avoids a lot of COW overhead.
1665 if (unlikely(!current->array))
1666 __activate_task(p, rq);
1668 p->prio = current->prio;
1669 p->normal_prio = current->normal_prio;
1670 list_add_tail(&p->run_list, ¤t->run_list);
1671 p->array = current->array;
1672 p->array->nr_active++;
1673 inc_nr_running(p, rq);
1677 /* Run child last */
1678 __activate_task(p, rq);
1680 * We skip the following code due to cpu == this_cpu
1682 * task_rq_unlock(rq, &flags);
1683 * this_rq = task_rq_lock(current, &flags);
1687 this_rq = cpu_rq(this_cpu);
1690 * Not the local CPU - must adjust timestamp. This should
1691 * get optimised away in the !CONFIG_SMP case.
1693 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1694 + rq->most_recent_timestamp;
1695 __activate_task(p, rq);
1696 if (TASK_PREEMPTS_CURR(p, rq))
1697 resched_task(rq->curr);
1700 * Parent and child are on different CPUs, now get the
1701 * parent runqueue to update the parent's ->sleep_avg:
1703 task_rq_unlock(rq, &flags);
1704 this_rq = task_rq_lock(current, &flags);
1706 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1707 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1708 task_rq_unlock(this_rq, &flags);
1712 * Potentially available exiting-child timeslices are
1713 * retrieved here - this way the parent does not get
1714 * penalized for creating too many threads.
1716 * (this cannot be used to 'generate' timeslices
1717 * artificially, because any timeslice recovered here
1718 * was given away by the parent in the first place.)
1720 void fastcall sched_exit(struct task_struct *p)
1722 unsigned long flags;
1726 * If the child was a (relative-) CPU hog then decrease
1727 * the sleep_avg of the parent as well.
1729 rq = task_rq_lock(p->parent, &flags);
1730 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1731 p->parent->time_slice += p->time_slice;
1732 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1733 p->parent->time_slice = task_timeslice(p);
1735 if (p->sleep_avg < p->parent->sleep_avg)
1736 p->parent->sleep_avg = p->parent->sleep_avg /
1737 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1739 task_rq_unlock(rq, &flags);
1743 * prepare_task_switch - prepare to switch tasks
1744 * @rq: the runqueue preparing to switch
1745 * @next: the task we are going to switch to.
1747 * This is called with the rq lock held and interrupts off. It must
1748 * be paired with a subsequent finish_task_switch after the context
1751 * prepare_task_switch sets up locking and calls architecture specific
1754 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1756 prepare_lock_switch(rq, next);
1757 prepare_arch_switch(next);
1761 * finish_task_switch - clean up after a task-switch
1762 * @rq: runqueue associated with task-switch
1763 * @prev: the thread we just switched away from.
1765 * finish_task_switch must be called after the context switch, paired
1766 * with a prepare_task_switch call before the context switch.
1767 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1768 * and do any other architecture-specific cleanup actions.
1770 * Note that we may have delayed dropping an mm in context_switch(). If
1771 * so, we finish that here outside of the runqueue lock. (Doing it
1772 * with the lock held can cause deadlocks; see schedule() for
1775 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1776 __releases(rq->lock)
1778 struct mm_struct *mm = rq->prev_mm;
1784 * A task struct has one reference for the use as "current".
1785 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1786 * schedule one last time. The schedule call will never return, and
1787 * the scheduled task must drop that reference.
1788 * The test for TASK_DEAD must occur while the runqueue locks are
1789 * still held, otherwise prev could be scheduled on another cpu, die
1790 * there before we look at prev->state, and then the reference would
1792 * Manfred Spraul <manfred@colorfullife.com>
1794 prev_state = prev->state;
1795 finish_arch_switch(prev);
1796 finish_lock_switch(rq, prev);
1799 if (unlikely(prev_state == TASK_DEAD)) {
1801 * Remove function-return probe instances associated with this
1802 * task and put them back on the free list.
1804 kprobe_flush_task(prev);
1805 put_task_struct(prev);
1810 * schedule_tail - first thing a freshly forked thread must call.
1811 * @prev: the thread we just switched away from.
1813 asmlinkage void schedule_tail(struct task_struct *prev)
1814 __releases(rq->lock)
1816 struct rq *rq = this_rq();
1818 finish_task_switch(rq, prev);
1819 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1820 /* In this case, finish_task_switch does not reenable preemption */
1823 if (current->set_child_tid)
1824 put_user(current->pid, current->set_child_tid);
1828 * context_switch - switch to the new MM and the new
1829 * thread's register state.
1831 static inline struct task_struct *
1832 context_switch(struct rq *rq, struct task_struct *prev,
1833 struct task_struct *next)
1835 struct mm_struct *mm = next->mm;
1836 struct mm_struct *oldmm = prev->active_mm;
1839 next->active_mm = oldmm;
1840 atomic_inc(&oldmm->mm_count);
1841 enter_lazy_tlb(oldmm, next);
1843 switch_mm(oldmm, mm, next);
1846 prev->active_mm = NULL;
1847 WARN_ON(rq->prev_mm);
1848 rq->prev_mm = oldmm;
1851 * Since the runqueue lock will be released by the next
1852 * task (which is an invalid locking op but in the case
1853 * of the scheduler it's an obvious special-case), so we
1854 * do an early lockdep release here:
1856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1857 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1860 /* Here we just switch the register state and the stack. */
1861 switch_to(prev, next, prev);
1867 * nr_running, nr_uninterruptible and nr_context_switches:
1869 * externally visible scheduler statistics: current number of runnable
1870 * threads, current number of uninterruptible-sleeping threads, total
1871 * number of context switches performed since bootup.
1873 unsigned long nr_running(void)
1875 unsigned long i, sum = 0;
1877 for_each_online_cpu(i)
1878 sum += cpu_rq(i)->nr_running;
1883 unsigned long nr_uninterruptible(void)
1885 unsigned long i, sum = 0;
1887 for_each_possible_cpu(i)
1888 sum += cpu_rq(i)->nr_uninterruptible;
1891 * Since we read the counters lockless, it might be slightly
1892 * inaccurate. Do not allow it to go below zero though:
1894 if (unlikely((long)sum < 0))
1900 unsigned long long nr_context_switches(void)
1903 unsigned long long sum = 0;
1905 for_each_possible_cpu(i)
1906 sum += cpu_rq(i)->nr_switches;
1911 unsigned long nr_iowait(void)
1913 unsigned long i, sum = 0;
1915 for_each_possible_cpu(i)
1916 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1921 unsigned long nr_active(void)
1923 unsigned long i, running = 0, uninterruptible = 0;
1925 for_each_online_cpu(i) {
1926 running += cpu_rq(i)->nr_running;
1927 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1930 if (unlikely((long)uninterruptible < 0))
1931 uninterruptible = 0;
1933 return running + uninterruptible;
1939 * Is this task likely cache-hot:
1942 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1944 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1948 * double_rq_lock - safely lock two runqueues
1950 * Note this does not disable interrupts like task_rq_lock,
1951 * you need to do so manually before calling.
1953 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1954 __acquires(rq1->lock)
1955 __acquires(rq2->lock)
1957 BUG_ON(!irqs_disabled());
1959 spin_lock(&rq1->lock);
1960 __acquire(rq2->lock); /* Fake it out ;) */
1963 spin_lock(&rq1->lock);
1964 spin_lock(&rq2->lock);
1966 spin_lock(&rq2->lock);
1967 spin_lock(&rq1->lock);
1973 * double_rq_unlock - safely unlock two runqueues
1975 * Note this does not restore interrupts like task_rq_unlock,
1976 * you need to do so manually after calling.
1978 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1979 __releases(rq1->lock)
1980 __releases(rq2->lock)
1982 spin_unlock(&rq1->lock);
1984 spin_unlock(&rq2->lock);
1986 __release(rq2->lock);
1990 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1992 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1993 __releases(this_rq->lock)
1994 __acquires(busiest->lock)
1995 __acquires(this_rq->lock)
1997 if (unlikely(!irqs_disabled())) {
1998 /* printk() doesn't work good under rq->lock */
1999 spin_unlock(&this_rq->lock);
2002 if (unlikely(!spin_trylock(&busiest->lock))) {
2003 if (busiest < this_rq) {
2004 spin_unlock(&this_rq->lock);
2005 spin_lock(&busiest->lock);
2006 spin_lock(&this_rq->lock);
2008 spin_lock(&busiest->lock);
2013 * If dest_cpu is allowed for this process, migrate the task to it.
2014 * This is accomplished by forcing the cpu_allowed mask to only
2015 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2016 * the cpu_allowed mask is restored.
2018 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2020 struct migration_req req;
2021 unsigned long flags;
2024 rq = task_rq_lock(p, &flags);
2025 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2026 || unlikely(cpu_is_offline(dest_cpu)))
2029 /* force the process onto the specified CPU */
2030 if (migrate_task(p, dest_cpu, &req)) {
2031 /* Need to wait for migration thread (might exit: take ref). */
2032 struct task_struct *mt = rq->migration_thread;
2034 get_task_struct(mt);
2035 task_rq_unlock(rq, &flags);
2036 wake_up_process(mt);
2037 put_task_struct(mt);
2038 wait_for_completion(&req.done);
2043 task_rq_unlock(rq, &flags);
2047 * sched_exec - execve() is a valuable balancing opportunity, because at
2048 * this point the task has the smallest effective memory and cache footprint.
2050 void sched_exec(void)
2052 int new_cpu, this_cpu = get_cpu();
2053 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2055 if (new_cpu != this_cpu)
2056 sched_migrate_task(current, new_cpu);
2060 * pull_task - move a task from a remote runqueue to the local runqueue.
2061 * Both runqueues must be locked.
2063 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2064 struct task_struct *p, struct rq *this_rq,
2065 struct prio_array *this_array, int this_cpu)
2067 dequeue_task(p, src_array);
2068 dec_nr_running(p, src_rq);
2069 set_task_cpu(p, this_cpu);
2070 inc_nr_running(p, this_rq);
2071 enqueue_task(p, this_array);
2072 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2073 + this_rq->most_recent_timestamp;
2075 * Note that idle threads have a prio of MAX_PRIO, for this test
2076 * to be always true for them.
2078 if (TASK_PREEMPTS_CURR(p, this_rq))
2079 resched_task(this_rq->curr);
2083 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2086 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2087 struct sched_domain *sd, enum idle_type idle,
2091 * We do not migrate tasks that are:
2092 * 1) running (obviously), or
2093 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2094 * 3) are cache-hot on their current CPU.
2096 if (!cpu_isset(this_cpu, p->cpus_allowed))
2100 if (task_running(rq, p))
2104 * Aggressive migration if:
2105 * 1) task is cache cold, or
2106 * 2) too many balance attempts have failed.
2109 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2110 #ifdef CONFIG_SCHEDSTATS
2111 if (task_hot(p, rq->most_recent_timestamp, sd))
2112 schedstat_inc(sd, lb_hot_gained[idle]);
2117 if (task_hot(p, rq->most_recent_timestamp, sd))
2122 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2125 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2126 * load from busiest to this_rq, as part of a balancing operation within
2127 * "domain". Returns the number of tasks moved.
2129 * Called with both runqueues locked.
2131 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2132 unsigned long max_nr_move, unsigned long max_load_move,
2133 struct sched_domain *sd, enum idle_type idle,
2136 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2137 best_prio_seen, skip_for_load;
2138 struct prio_array *array, *dst_array;
2139 struct list_head *head, *curr;
2140 struct task_struct *tmp;
2143 if (max_nr_move == 0 || max_load_move == 0)
2146 rem_load_move = max_load_move;
2148 this_best_prio = rq_best_prio(this_rq);
2149 best_prio = rq_best_prio(busiest);
2151 * Enable handling of the case where there is more than one task
2152 * with the best priority. If the current running task is one
2153 * of those with prio==best_prio we know it won't be moved
2154 * and therefore it's safe to override the skip (based on load) of
2155 * any task we find with that prio.
2157 best_prio_seen = best_prio == busiest->curr->prio;
2160 * We first consider expired tasks. Those will likely not be
2161 * executed in the near future, and they are most likely to
2162 * be cache-cold, thus switching CPUs has the least effect
2165 if (busiest->expired->nr_active) {
2166 array = busiest->expired;
2167 dst_array = this_rq->expired;
2169 array = busiest->active;
2170 dst_array = this_rq->active;
2174 /* Start searching at priority 0: */
2178 idx = sched_find_first_bit(array->bitmap);
2180 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2181 if (idx >= MAX_PRIO) {
2182 if (array == busiest->expired && busiest->active->nr_active) {
2183 array = busiest->active;
2184 dst_array = this_rq->active;
2190 head = array->queue + idx;
2193 tmp = list_entry(curr, struct task_struct, run_list);
2198 * To help distribute high priority tasks accross CPUs we don't
2199 * skip a task if it will be the highest priority task (i.e. smallest
2200 * prio value) on its new queue regardless of its load weight
2202 skip_for_load = tmp->load_weight > rem_load_move;
2203 if (skip_for_load && idx < this_best_prio)
2204 skip_for_load = !best_prio_seen && idx == best_prio;
2205 if (skip_for_load ||
2206 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2208 best_prio_seen |= idx == best_prio;
2215 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2217 rem_load_move -= tmp->load_weight;
2220 * We only want to steal up to the prescribed number of tasks
2221 * and the prescribed amount of weighted load.
2223 if (pulled < max_nr_move && rem_load_move > 0) {
2224 if (idx < this_best_prio)
2225 this_best_prio = idx;
2233 * Right now, this is the only place pull_task() is called,
2234 * so we can safely collect pull_task() stats here rather than
2235 * inside pull_task().
2237 schedstat_add(sd, lb_gained[idle], pulled);
2240 *all_pinned = pinned;
2245 * find_busiest_group finds and returns the busiest CPU group within the
2246 * domain. It calculates and returns the amount of weighted load which
2247 * should be moved to restore balance via the imbalance parameter.
2249 static struct sched_group *
2250 find_busiest_group(struct sched_domain *sd, int this_cpu,
2251 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2252 cpumask_t *cpus, int *balance)
2254 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2255 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2256 unsigned long max_pull;
2257 unsigned long busiest_load_per_task, busiest_nr_running;
2258 unsigned long this_load_per_task, this_nr_running;
2260 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2261 int power_savings_balance = 1;
2262 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2263 unsigned long min_nr_running = ULONG_MAX;
2264 struct sched_group *group_min = NULL, *group_leader = NULL;
2267 max_load = this_load = total_load = total_pwr = 0;
2268 busiest_load_per_task = busiest_nr_running = 0;
2269 this_load_per_task = this_nr_running = 0;
2270 if (idle == NOT_IDLE)
2271 load_idx = sd->busy_idx;
2272 else if (idle == NEWLY_IDLE)
2273 load_idx = sd->newidle_idx;
2275 load_idx = sd->idle_idx;
2278 unsigned long load, group_capacity;
2281 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2282 unsigned long sum_nr_running, sum_weighted_load;
2284 local_group = cpu_isset(this_cpu, group->cpumask);
2287 balance_cpu = first_cpu(group->cpumask);
2289 /* Tally up the load of all CPUs in the group */
2290 sum_weighted_load = sum_nr_running = avg_load = 0;
2292 for_each_cpu_mask(i, group->cpumask) {
2295 if (!cpu_isset(i, *cpus))
2300 if (*sd_idle && !idle_cpu(i))
2303 /* Bias balancing toward cpus of our domain */
2305 if (idle_cpu(i) && !first_idle_cpu) {
2310 load = target_load(i, load_idx);
2312 load = source_load(i, load_idx);
2315 sum_nr_running += rq->nr_running;
2316 sum_weighted_load += rq->raw_weighted_load;
2320 * First idle cpu or the first cpu(busiest) in this sched group
2321 * is eligible for doing load balancing at this and above
2324 if (local_group && balance_cpu != this_cpu && balance) {
2329 total_load += avg_load;
2330 total_pwr += group->cpu_power;
2332 /* Adjust by relative CPU power of the group */
2333 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2335 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2338 this_load = avg_load;
2340 this_nr_running = sum_nr_running;
2341 this_load_per_task = sum_weighted_load;
2342 } else if (avg_load > max_load &&
2343 sum_nr_running > group_capacity) {
2344 max_load = avg_load;
2346 busiest_nr_running = sum_nr_running;
2347 busiest_load_per_task = sum_weighted_load;
2350 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2352 * Busy processors will not participate in power savings
2355 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2359 * If the local group is idle or completely loaded
2360 * no need to do power savings balance at this domain
2362 if (local_group && (this_nr_running >= group_capacity ||
2364 power_savings_balance = 0;
2367 * If a group is already running at full capacity or idle,
2368 * don't include that group in power savings calculations
2370 if (!power_savings_balance || sum_nr_running >= group_capacity
2375 * Calculate the group which has the least non-idle load.
2376 * This is the group from where we need to pick up the load
2379 if ((sum_nr_running < min_nr_running) ||
2380 (sum_nr_running == min_nr_running &&
2381 first_cpu(group->cpumask) <
2382 first_cpu(group_min->cpumask))) {
2384 min_nr_running = sum_nr_running;
2385 min_load_per_task = sum_weighted_load /
2390 * Calculate the group which is almost near its
2391 * capacity but still has some space to pick up some load
2392 * from other group and save more power
2394 if (sum_nr_running <= group_capacity - 1) {
2395 if (sum_nr_running > leader_nr_running ||
2396 (sum_nr_running == leader_nr_running &&
2397 first_cpu(group->cpumask) >
2398 first_cpu(group_leader->cpumask))) {
2399 group_leader = group;
2400 leader_nr_running = sum_nr_running;
2405 group = group->next;
2406 } while (group != sd->groups);
2408 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2411 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2413 if (this_load >= avg_load ||
2414 100*max_load <= sd->imbalance_pct*this_load)
2417 busiest_load_per_task /= busiest_nr_running;
2419 * We're trying to get all the cpus to the average_load, so we don't
2420 * want to push ourselves above the average load, nor do we wish to
2421 * reduce the max loaded cpu below the average load, as either of these
2422 * actions would just result in more rebalancing later, and ping-pong
2423 * tasks around. Thus we look for the minimum possible imbalance.
2424 * Negative imbalances (*we* are more loaded than anyone else) will
2425 * be counted as no imbalance for these purposes -- we can't fix that
2426 * by pulling tasks to us. Be careful of negative numbers as they'll
2427 * appear as very large values with unsigned longs.
2429 if (max_load <= busiest_load_per_task)
2433 * In the presence of smp nice balancing, certain scenarios can have
2434 * max load less than avg load(as we skip the groups at or below
2435 * its cpu_power, while calculating max_load..)
2437 if (max_load < avg_load) {
2439 goto small_imbalance;
2442 /* Don't want to pull so many tasks that a group would go idle */
2443 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2445 /* How much load to actually move to equalise the imbalance */
2446 *imbalance = min(max_pull * busiest->cpu_power,
2447 (avg_load - this_load) * this->cpu_power)
2451 * if *imbalance is less than the average load per runnable task
2452 * there is no gaurantee that any tasks will be moved so we'll have
2453 * a think about bumping its value to force at least one task to be
2456 if (*imbalance < busiest_load_per_task) {
2457 unsigned long tmp, pwr_now, pwr_move;
2461 pwr_move = pwr_now = 0;
2463 if (this_nr_running) {
2464 this_load_per_task /= this_nr_running;
2465 if (busiest_load_per_task > this_load_per_task)
2468 this_load_per_task = SCHED_LOAD_SCALE;
2470 if (max_load - this_load >= busiest_load_per_task * imbn) {
2471 *imbalance = busiest_load_per_task;
2476 * OK, we don't have enough imbalance to justify moving tasks,
2477 * however we may be able to increase total CPU power used by
2481 pwr_now += busiest->cpu_power *
2482 min(busiest_load_per_task, max_load);
2483 pwr_now += this->cpu_power *
2484 min(this_load_per_task, this_load);
2485 pwr_now /= SCHED_LOAD_SCALE;
2487 /* Amount of load we'd subtract */
2488 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2490 pwr_move += busiest->cpu_power *
2491 min(busiest_load_per_task, max_load - tmp);
2493 /* Amount of load we'd add */
2494 if (max_load*busiest->cpu_power <
2495 busiest_load_per_task*SCHED_LOAD_SCALE)
2496 tmp = max_load*busiest->cpu_power/this->cpu_power;
2498 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2499 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2500 pwr_move /= SCHED_LOAD_SCALE;
2502 /* Move if we gain throughput */
2503 if (pwr_move <= pwr_now)
2506 *imbalance = busiest_load_per_task;
2512 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2513 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2516 if (this == group_leader && group_leader != group_min) {
2517 *imbalance = min_load_per_task;
2527 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2530 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2531 unsigned long imbalance, cpumask_t *cpus)
2533 struct rq *busiest = NULL, *rq;
2534 unsigned long max_load = 0;
2537 for_each_cpu_mask(i, group->cpumask) {
2539 if (!cpu_isset(i, *cpus))
2544 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2547 if (rq->raw_weighted_load > max_load) {
2548 max_load = rq->raw_weighted_load;
2557 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2558 * so long as it is large enough.
2560 #define MAX_PINNED_INTERVAL 512
2562 static inline unsigned long minus_1_or_zero(unsigned long n)
2564 return n > 0 ? n - 1 : 0;
2568 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2569 * tasks if there is an imbalance.
2571 static int load_balance(int this_cpu, struct rq *this_rq,
2572 struct sched_domain *sd, enum idle_type idle,
2575 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2576 struct sched_group *group;
2577 unsigned long imbalance;
2579 cpumask_t cpus = CPU_MASK_ALL;
2580 unsigned long flags;
2583 * When power savings policy is enabled for the parent domain, idle
2584 * sibling can pick up load irrespective of busy siblings. In this case,
2585 * let the state of idle sibling percolate up as IDLE, instead of
2586 * portraying it as NOT_IDLE.
2588 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2589 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2592 schedstat_inc(sd, lb_cnt[idle]);
2595 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2602 schedstat_inc(sd, lb_nobusyg[idle]);
2606 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2608 schedstat_inc(sd, lb_nobusyq[idle]);
2612 BUG_ON(busiest == this_rq);
2614 schedstat_add(sd, lb_imbalance[idle], imbalance);
2617 if (busiest->nr_running > 1) {
2619 * Attempt to move tasks. If find_busiest_group has found
2620 * an imbalance but busiest->nr_running <= 1, the group is
2621 * still unbalanced. nr_moved simply stays zero, so it is
2622 * correctly treated as an imbalance.
2624 local_irq_save(flags);
2625 double_rq_lock(this_rq, busiest);
2626 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2627 minus_1_or_zero(busiest->nr_running),
2628 imbalance, sd, idle, &all_pinned);
2629 double_rq_unlock(this_rq, busiest);
2630 local_irq_restore(flags);
2632 /* All tasks on this runqueue were pinned by CPU affinity */
2633 if (unlikely(all_pinned)) {
2634 cpu_clear(cpu_of(busiest), cpus);
2635 if (!cpus_empty(cpus))
2642 schedstat_inc(sd, lb_failed[idle]);
2643 sd->nr_balance_failed++;
2645 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2647 spin_lock_irqsave(&busiest->lock, flags);
2649 /* don't kick the migration_thread, if the curr
2650 * task on busiest cpu can't be moved to this_cpu
2652 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2653 spin_unlock_irqrestore(&busiest->lock, flags);
2655 goto out_one_pinned;
2658 if (!busiest->active_balance) {
2659 busiest->active_balance = 1;
2660 busiest->push_cpu = this_cpu;
2663 spin_unlock_irqrestore(&busiest->lock, flags);
2665 wake_up_process(busiest->migration_thread);
2668 * We've kicked active balancing, reset the failure
2671 sd->nr_balance_failed = sd->cache_nice_tries+1;
2674 sd->nr_balance_failed = 0;
2676 if (likely(!active_balance)) {
2677 /* We were unbalanced, so reset the balancing interval */
2678 sd->balance_interval = sd->min_interval;
2681 * If we've begun active balancing, start to back off. This
2682 * case may not be covered by the all_pinned logic if there
2683 * is only 1 task on the busy runqueue (because we don't call
2686 if (sd->balance_interval < sd->max_interval)
2687 sd->balance_interval *= 2;
2690 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2691 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2696 schedstat_inc(sd, lb_balanced[idle]);
2698 sd->nr_balance_failed = 0;
2701 /* tune up the balancing interval */
2702 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2703 (sd->balance_interval < sd->max_interval))
2704 sd->balance_interval *= 2;
2706 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2707 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2713 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2714 * tasks if there is an imbalance.
2716 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2717 * this_rq is locked.
2720 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2722 struct sched_group *group;
2723 struct rq *busiest = NULL;
2724 unsigned long imbalance;
2727 cpumask_t cpus = CPU_MASK_ALL;
2730 * When power savings policy is enabled for the parent domain, idle
2731 * sibling can pick up load irrespective of busy siblings. In this case,
2732 * let the state of idle sibling percolate up as IDLE, instead of
2733 * portraying it as NOT_IDLE.
2735 if (sd->flags & SD_SHARE_CPUPOWER &&
2736 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2739 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2741 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2742 &sd_idle, &cpus, NULL);
2744 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2748 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2751 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2755 BUG_ON(busiest == this_rq);
2757 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2760 if (busiest->nr_running > 1) {
2761 /* Attempt to move tasks */
2762 double_lock_balance(this_rq, busiest);
2763 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2764 minus_1_or_zero(busiest->nr_running),
2765 imbalance, sd, NEWLY_IDLE, NULL);
2766 spin_unlock(&busiest->lock);
2769 cpu_clear(cpu_of(busiest), cpus);
2770 if (!cpus_empty(cpus))
2776 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2777 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2778 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2781 sd->nr_balance_failed = 0;
2786 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2787 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2788 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2790 sd->nr_balance_failed = 0;
2796 * idle_balance is called by schedule() if this_cpu is about to become
2797 * idle. Attempts to pull tasks from other CPUs.
2799 static void idle_balance(int this_cpu, struct rq *this_rq)
2801 struct sched_domain *sd;
2802 int pulled_task = 0;
2803 unsigned long next_balance = jiffies + 60 * HZ;
2805 for_each_domain(this_cpu, sd) {
2806 if (sd->flags & SD_BALANCE_NEWIDLE) {
2807 /* If we've pulled tasks over stop searching: */
2808 pulled_task = load_balance_newidle(this_cpu,
2810 if (time_after(next_balance,
2811 sd->last_balance + sd->balance_interval))
2812 next_balance = sd->last_balance
2813 + sd->balance_interval;
2820 * We are going idle. next_balance may be set based on
2821 * a busy processor. So reset next_balance.
2823 this_rq->next_balance = next_balance;
2827 * active_load_balance is run by migration threads. It pushes running tasks
2828 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2829 * running on each physical CPU where possible, and avoids physical /
2830 * logical imbalances.
2832 * Called with busiest_rq locked.
2834 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2836 int target_cpu = busiest_rq->push_cpu;
2837 struct sched_domain *sd;
2838 struct rq *target_rq;
2840 /* Is there any task to move? */
2841 if (busiest_rq->nr_running <= 1)
2844 target_rq = cpu_rq(target_cpu);
2847 * This condition is "impossible", if it occurs
2848 * we need to fix it. Originally reported by
2849 * Bjorn Helgaas on a 128-cpu setup.
2851 BUG_ON(busiest_rq == target_rq);
2853 /* move a task from busiest_rq to target_rq */
2854 double_lock_balance(busiest_rq, target_rq);
2856 /* Search for an sd spanning us and the target CPU. */
2857 for_each_domain(target_cpu, sd) {
2858 if ((sd->flags & SD_LOAD_BALANCE) &&
2859 cpu_isset(busiest_cpu, sd->span))
2864 schedstat_inc(sd, alb_cnt);
2866 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2867 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2869 schedstat_inc(sd, alb_pushed);
2871 schedstat_inc(sd, alb_failed);
2873 spin_unlock(&target_rq->lock);
2876 static void update_load(struct rq *this_rq)
2878 unsigned long this_load;
2881 this_load = this_rq->raw_weighted_load;
2883 /* Update our load: */
2884 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2885 unsigned long old_load, new_load;
2887 old_load = this_rq->cpu_load[i];
2888 new_load = this_load;
2890 * Round up the averaging division if load is increasing. This
2891 * prevents us from getting stuck on 9 if the load is 10, for
2894 if (new_load > old_load)
2895 new_load += scale-1;
2896 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2901 * run_rebalance_domains is triggered when needed from the scheduler tick.
2903 * It checks each scheduling domain to see if it is due to be balanced,
2904 * and initiates a balancing operation if so.
2906 * Balancing parameters are set up in arch_init_sched_domains.
2908 static DEFINE_SPINLOCK(balancing);
2910 static void run_rebalance_domains(struct softirq_action *h)
2912 int this_cpu = smp_processor_id(), balance = 1;
2913 struct rq *this_rq = cpu_rq(this_cpu);
2914 unsigned long interval;
2915 struct sched_domain *sd;
2917 * We are idle if there are no processes running. This
2918 * is valid even if we are the idle process (SMT).
2920 enum idle_type idle = !this_rq->nr_running ?
2921 SCHED_IDLE : NOT_IDLE;
2922 /* Earliest time when we have to call run_rebalance_domains again */
2923 unsigned long next_balance = jiffies + 60*HZ;
2925 for_each_domain(this_cpu, sd) {
2926 if (!(sd->flags & SD_LOAD_BALANCE))
2929 interval = sd->balance_interval;
2930 if (idle != SCHED_IDLE)
2931 interval *= sd->busy_factor;
2933 /* scale ms to jiffies */
2934 interval = msecs_to_jiffies(interval);
2935 if (unlikely(!interval))
2938 if (sd->flags & SD_SERIALIZE) {
2939 if (!spin_trylock(&balancing))
2943 if (time_after_eq(jiffies, sd->last_balance + interval)) {
2944 if (load_balance(this_cpu, this_rq, sd, idle, &balance)) {
2946 * We've pulled tasks over so either we're no
2947 * longer idle, or one of our SMT siblings is
2952 sd->last_balance = jiffies;
2954 if (sd->flags & SD_SERIALIZE)
2955 spin_unlock(&balancing);
2957 if (time_after(next_balance, sd->last_balance + interval))
2958 next_balance = sd->last_balance + interval;
2961 * Stop the load balance at this level. There is another
2962 * CPU in our sched group which is doing load balancing more
2968 this_rq->next_balance = next_balance;
2972 * on UP we do not need to balance between CPUs:
2974 static inline void idle_balance(int cpu, struct rq *rq)
2979 static inline void wake_priority_sleeper(struct rq *rq)
2981 #ifdef CONFIG_SCHED_SMT
2982 if (!rq->nr_running)
2985 spin_lock(&rq->lock);
2987 * If an SMT sibling task has been put to sleep for priority
2988 * reasons reschedule the idle task to see if it can now run.
2991 resched_task(rq->idle);
2992 spin_unlock(&rq->lock);
2996 DEFINE_PER_CPU(struct kernel_stat, kstat);
2998 EXPORT_PER_CPU_SYMBOL(kstat);
3001 * This is called on clock ticks and on context switches.
3002 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3005 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3007 p->sched_time += now - p->last_ran;
3008 p->last_ran = rq->most_recent_timestamp = now;
3012 * Return current->sched_time plus any more ns on the sched_clock
3013 * that have not yet been banked.
3015 unsigned long long current_sched_time(const struct task_struct *p)
3017 unsigned long long ns;
3018 unsigned long flags;
3020 local_irq_save(flags);
3021 ns = p->sched_time + sched_clock() - p->last_ran;
3022 local_irq_restore(flags);
3028 * We place interactive tasks back into the active array, if possible.
3030 * To guarantee that this does not starve expired tasks we ignore the
3031 * interactivity of a task if the first expired task had to wait more
3032 * than a 'reasonable' amount of time. This deadline timeout is
3033 * load-dependent, as the frequency of array switched decreases with
3034 * increasing number of running tasks. We also ignore the interactivity
3035 * if a better static_prio task has expired:
3037 static inline int expired_starving(struct rq *rq)
3039 if (rq->curr->static_prio > rq->best_expired_prio)
3041 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3043 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3049 * Account user cpu time to a process.
3050 * @p: the process that the cpu time gets accounted to
3051 * @hardirq_offset: the offset to subtract from hardirq_count()
3052 * @cputime: the cpu time spent in user space since the last update
3054 void account_user_time(struct task_struct *p, cputime_t cputime)
3056 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3059 p->utime = cputime_add(p->utime, cputime);
3061 /* Add user time to cpustat. */
3062 tmp = cputime_to_cputime64(cputime);
3063 if (TASK_NICE(p) > 0)
3064 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3066 cpustat->user = cputime64_add(cpustat->user, tmp);
3070 * Account system cpu time to a process.
3071 * @p: the process that the cpu time gets accounted to
3072 * @hardirq_offset: the offset to subtract from hardirq_count()
3073 * @cputime: the cpu time spent in kernel space since the last update
3075 void account_system_time(struct task_struct *p, int hardirq_offset,
3078 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3079 struct rq *rq = this_rq();
3082 p->stime = cputime_add(p->stime, cputime);
3084 /* Add system time to cpustat. */
3085 tmp = cputime_to_cputime64(cputime);
3086 if (hardirq_count() - hardirq_offset)
3087 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3088 else if (softirq_count())
3089 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3090 else if (p != rq->idle)
3091 cpustat->system = cputime64_add(cpustat->system, tmp);
3092 else if (atomic_read(&rq->nr_iowait) > 0)
3093 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3095 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3096 /* Account for system time used */
3097 acct_update_integrals(p);
3101 * Account for involuntary wait time.
3102 * @p: the process from which the cpu time has been stolen
3103 * @steal: the cpu time spent in involuntary wait
3105 void account_steal_time(struct task_struct *p, cputime_t steal)
3107 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3108 cputime64_t tmp = cputime_to_cputime64(steal);
3109 struct rq *rq = this_rq();
3111 if (p == rq->idle) {
3112 p->stime = cputime_add(p->stime, steal);
3113 if (atomic_read(&rq->nr_iowait) > 0)
3114 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3116 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3118 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3121 static void task_running_tick(struct rq *rq, struct task_struct *p)
3123 if (p->array != rq->active) {
3124 /* Task has expired but was not scheduled yet */
3125 set_tsk_need_resched(p);
3128 spin_lock(&rq->lock);
3130 * The task was running during this tick - update the
3131 * time slice counter. Note: we do not update a thread's
3132 * priority until it either goes to sleep or uses up its
3133 * timeslice. This makes it possible for interactive tasks
3134 * to use up their timeslices at their highest priority levels.
3138 * RR tasks need a special form of timeslice management.
3139 * FIFO tasks have no timeslices.
3141 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3142 p->time_slice = task_timeslice(p);
3143 p->first_time_slice = 0;
3144 set_tsk_need_resched(p);
3146 /* put it at the end of the queue: */
3147 requeue_task(p, rq->active);
3151 if (!--p->time_slice) {
3152 dequeue_task(p, rq->active);
3153 set_tsk_need_resched(p);
3154 p->prio = effective_prio(p);
3155 p->time_slice = task_timeslice(p);
3156 p->first_time_slice = 0;
3158 if (!rq->expired_timestamp)
3159 rq->expired_timestamp = jiffies;
3160 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3161 enqueue_task(p, rq->expired);
3162 if (p->static_prio < rq->best_expired_prio)
3163 rq->best_expired_prio = p->static_prio;
3165 enqueue_task(p, rq->active);
3168 * Prevent a too long timeslice allowing a task to monopolize
3169 * the CPU. We do this by splitting up the timeslice into
3172 * Note: this does not mean the task's timeslices expire or
3173 * get lost in any way, they just might be preempted by
3174 * another task of equal priority. (one with higher
3175 * priority would have preempted this task already.) We
3176 * requeue this task to the end of the list on this priority
3177 * level, which is in essence a round-robin of tasks with
3180 * This only applies to tasks in the interactive
3181 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3183 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3184 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3185 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3186 (p->array == rq->active)) {
3188 requeue_task(p, rq->active);
3189 set_tsk_need_resched(p);
3193 spin_unlock(&rq->lock);
3197 * This function gets called by the timer code, with HZ frequency.
3198 * We call it with interrupts disabled.
3200 * It also gets called by the fork code, when changing the parent's
3203 void scheduler_tick(void)
3205 unsigned long long now = sched_clock();
3206 struct task_struct *p = current;
3207 int cpu = smp_processor_id();
3208 struct rq *rq = cpu_rq(cpu);
3210 update_cpu_clock(p, rq, now);
3213 /* Task on the idle queue */
3214 wake_priority_sleeper(rq);
3216 task_running_tick(rq, p);
3219 if (time_after_eq(jiffies, rq->next_balance))
3220 raise_softirq(SCHED_SOFTIRQ);
3224 #ifdef CONFIG_SCHED_SMT
3225 static inline void wakeup_busy_runqueue(struct rq *rq)
3227 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3228 if (rq->curr == rq->idle && rq->nr_running)
3229 resched_task(rq->idle);
3233 * Called with interrupt disabled and this_rq's runqueue locked.
3235 static void wake_sleeping_dependent(int this_cpu)
3237 struct sched_domain *tmp, *sd = NULL;
3240 for_each_domain(this_cpu, tmp) {
3241 if (tmp->flags & SD_SHARE_CPUPOWER) {
3250 for_each_cpu_mask(i, sd->span) {
3251 struct rq *smt_rq = cpu_rq(i);
3255 if (unlikely(!spin_trylock(&smt_rq->lock)))
3258 wakeup_busy_runqueue(smt_rq);
3259 spin_unlock(&smt_rq->lock);
3264 * number of 'lost' timeslices this task wont be able to fully
3265 * utilize, if another task runs on a sibling. This models the
3266 * slowdown effect of other tasks running on siblings:
3268 static inline unsigned long
3269 smt_slice(struct task_struct *p, struct sched_domain *sd)
3271 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3275 * To minimise lock contention and not have to drop this_rq's runlock we only
3276 * trylock the sibling runqueues and bypass those runqueues if we fail to
3277 * acquire their lock. As we only trylock the normal locking order does not
3278 * need to be obeyed.
3281 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3283 struct sched_domain *tmp, *sd = NULL;
3286 /* kernel/rt threads do not participate in dependent sleeping */
3287 if (!p->mm || rt_task(p))
3290 for_each_domain(this_cpu, tmp) {
3291 if (tmp->flags & SD_SHARE_CPUPOWER) {
3300 for_each_cpu_mask(i, sd->span) {
3301 struct task_struct *smt_curr;
3308 if (unlikely(!spin_trylock(&smt_rq->lock)))
3311 smt_curr = smt_rq->curr;
3317 * If a user task with lower static priority than the
3318 * running task on the SMT sibling is trying to schedule,
3319 * delay it till there is proportionately less timeslice
3320 * left of the sibling task to prevent a lower priority
3321 * task from using an unfair proportion of the
3322 * physical cpu's resources. -ck
3324 if (rt_task(smt_curr)) {
3326 * With real time tasks we run non-rt tasks only
3327 * per_cpu_gain% of the time.
3329 if ((jiffies % DEF_TIMESLICE) >
3330 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3333 if (smt_curr->static_prio < p->static_prio &&
3334 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3335 smt_slice(smt_curr, sd) > task_timeslice(p))
3339 spin_unlock(&smt_rq->lock);
3344 static inline void wake_sleeping_dependent(int this_cpu)
3348 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3354 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3356 void fastcall add_preempt_count(int val)
3361 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3363 preempt_count() += val;
3365 * Spinlock count overflowing soon?
3367 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3369 EXPORT_SYMBOL(add_preempt_count);
3371 void fastcall sub_preempt_count(int val)
3376 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3379 * Is the spinlock portion underflowing?
3381 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3382 !(preempt_count() & PREEMPT_MASK)))
3385 preempt_count() -= val;
3387 EXPORT_SYMBOL(sub_preempt_count);
3391 static inline int interactive_sleep(enum sleep_type sleep_type)
3393 return (sleep_type == SLEEP_INTERACTIVE ||
3394 sleep_type == SLEEP_INTERRUPTED);
3398 * schedule() is the main scheduler function.
3400 asmlinkage void __sched schedule(void)
3402 struct task_struct *prev, *next;
3403 struct prio_array *array;
3404 struct list_head *queue;
3405 unsigned long long now;
3406 unsigned long run_time;
3407 int cpu, idx, new_prio;
3412 * Test if we are atomic. Since do_exit() needs to call into
3413 * schedule() atomically, we ignore that path for now.
3414 * Otherwise, whine if we are scheduling when we should not be.
3416 if (unlikely(in_atomic() && !current->exit_state)) {
3417 printk(KERN_ERR "BUG: scheduling while atomic: "
3419 current->comm, preempt_count(), current->pid);
3420 debug_show_held_locks(current);
3423 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3428 release_kernel_lock(prev);
3429 need_resched_nonpreemptible:
3433 * The idle thread is not allowed to schedule!
3434 * Remove this check after it has been exercised a bit.
3436 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3437 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3441 schedstat_inc(rq, sched_cnt);
3442 now = sched_clock();
3443 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3444 run_time = now - prev->timestamp;
3445 if (unlikely((long long)(now - prev->timestamp) < 0))
3448 run_time = NS_MAX_SLEEP_AVG;
3451 * Tasks charged proportionately less run_time at high sleep_avg to
3452 * delay them losing their interactive status
3454 run_time /= (CURRENT_BONUS(prev) ? : 1);
3456 spin_lock_irq(&rq->lock);
3458 switch_count = &prev->nivcsw;
3459 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3460 switch_count = &prev->nvcsw;
3461 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3462 unlikely(signal_pending(prev))))
3463 prev->state = TASK_RUNNING;
3465 if (prev->state == TASK_UNINTERRUPTIBLE)
3466 rq->nr_uninterruptible++;
3467 deactivate_task(prev, rq);
3471 cpu = smp_processor_id();
3472 if (unlikely(!rq->nr_running)) {
3473 idle_balance(cpu, rq);
3474 if (!rq->nr_running) {
3476 rq->expired_timestamp = 0;
3477 wake_sleeping_dependent(cpu);
3483 if (unlikely(!array->nr_active)) {
3485 * Switch the active and expired arrays.
3487 schedstat_inc(rq, sched_switch);
3488 rq->active = rq->expired;
3489 rq->expired = array;
3491 rq->expired_timestamp = 0;
3492 rq->best_expired_prio = MAX_PRIO;
3495 idx = sched_find_first_bit(array->bitmap);
3496 queue = array->queue + idx;
3497 next = list_entry(queue->next, struct task_struct, run_list);
3499 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3500 unsigned long long delta = now - next->timestamp;
3501 if (unlikely((long long)(now - next->timestamp) < 0))
3504 if (next->sleep_type == SLEEP_INTERACTIVE)
3505 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3507 array = next->array;
3508 new_prio = recalc_task_prio(next, next->timestamp + delta);
3510 if (unlikely(next->prio != new_prio)) {
3511 dequeue_task(next, array);
3512 next->prio = new_prio;
3513 enqueue_task(next, array);
3516 next->sleep_type = SLEEP_NORMAL;
3517 if (dependent_sleeper(cpu, rq, next))
3520 if (next == rq->idle)
3521 schedstat_inc(rq, sched_goidle);
3523 prefetch_stack(next);
3524 clear_tsk_need_resched(prev);
3525 rcu_qsctr_inc(task_cpu(prev));
3527 update_cpu_clock(prev, rq, now);
3529 prev->sleep_avg -= run_time;
3530 if ((long)prev->sleep_avg <= 0)
3531 prev->sleep_avg = 0;
3532 prev->timestamp = prev->last_ran = now;
3534 sched_info_switch(prev, next);
3535 if (likely(prev != next)) {
3536 next->timestamp = now;
3541 prepare_task_switch(rq, next);
3542 prev = context_switch(rq, prev, next);
3545 * this_rq must be evaluated again because prev may have moved
3546 * CPUs since it called schedule(), thus the 'rq' on its stack
3547 * frame will be invalid.
3549 finish_task_switch(this_rq(), prev);
3551 spin_unlock_irq(&rq->lock);
3554 if (unlikely(reacquire_kernel_lock(prev) < 0))
3555 goto need_resched_nonpreemptible;
3556 preempt_enable_no_resched();
3557 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3560 EXPORT_SYMBOL(schedule);
3562 #ifdef CONFIG_PREEMPT
3564 * this is the entry point to schedule() from in-kernel preemption
3565 * off of preempt_enable. Kernel preemptions off return from interrupt
3566 * occur there and call schedule directly.
3568 asmlinkage void __sched preempt_schedule(void)
3570 struct thread_info *ti = current_thread_info();
3571 #ifdef CONFIG_PREEMPT_BKL
3572 struct task_struct *task = current;
3573 int saved_lock_depth;
3576 * If there is a non-zero preempt_count or interrupts are disabled,
3577 * we do not want to preempt the current task. Just return..
3579 if (likely(ti->preempt_count || irqs_disabled()))
3583 add_preempt_count(PREEMPT_ACTIVE);
3585 * We keep the big kernel semaphore locked, but we
3586 * clear ->lock_depth so that schedule() doesnt
3587 * auto-release the semaphore:
3589 #ifdef CONFIG_PREEMPT_BKL
3590 saved_lock_depth = task->lock_depth;
3591 task->lock_depth = -1;
3594 #ifdef CONFIG_PREEMPT_BKL
3595 task->lock_depth = saved_lock_depth;
3597 sub_preempt_count(PREEMPT_ACTIVE);
3599 /* we could miss a preemption opportunity between schedule and now */
3601 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3604 EXPORT_SYMBOL(preempt_schedule);
3607 * this is the entry point to schedule() from kernel preemption
3608 * off of irq context.
3609 * Note, that this is called and return with irqs disabled. This will
3610 * protect us against recursive calling from irq.
3612 asmlinkage void __sched preempt_schedule_irq(void)
3614 struct thread_info *ti = current_thread_info();
3615 #ifdef CONFIG_PREEMPT_BKL
3616 struct task_struct *task = current;
3617 int saved_lock_depth;
3619 /* Catch callers which need to be fixed */
3620 BUG_ON(ti->preempt_count || !irqs_disabled());
3623 add_preempt_count(PREEMPT_ACTIVE);
3625 * We keep the big kernel semaphore locked, but we
3626 * clear ->lock_depth so that schedule() doesnt
3627 * auto-release the semaphore:
3629 #ifdef CONFIG_PREEMPT_BKL
3630 saved_lock_depth = task->lock_depth;
3631 task->lock_depth = -1;
3635 local_irq_disable();
3636 #ifdef CONFIG_PREEMPT_BKL
3637 task->lock_depth = saved_lock_depth;
3639 sub_preempt_count(PREEMPT_ACTIVE);
3641 /* we could miss a preemption opportunity between schedule and now */
3643 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3647 #endif /* CONFIG_PREEMPT */
3649 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3652 return try_to_wake_up(curr->private, mode, sync);
3654 EXPORT_SYMBOL(default_wake_function);
3657 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3658 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3659 * number) then we wake all the non-exclusive tasks and one exclusive task.
3661 * There are circumstances in which we can try to wake a task which has already
3662 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3663 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3665 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3666 int nr_exclusive, int sync, void *key)
3668 struct list_head *tmp, *next;
3670 list_for_each_safe(tmp, next, &q->task_list) {
3671 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3672 unsigned flags = curr->flags;
3674 if (curr->func(curr, mode, sync, key) &&
3675 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3681 * __wake_up - wake up threads blocked on a waitqueue.
3683 * @mode: which threads
3684 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3685 * @key: is directly passed to the wakeup function
3687 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3688 int nr_exclusive, void *key)
3690 unsigned long flags;
3692 spin_lock_irqsave(&q->lock, flags);
3693 __wake_up_common(q, mode, nr_exclusive, 0, key);
3694 spin_unlock_irqrestore(&q->lock, flags);
3696 EXPORT_SYMBOL(__wake_up);
3699 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3701 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3703 __wake_up_common(q, mode, 1, 0, NULL);
3707 * __wake_up_sync - wake up threads blocked on a waitqueue.
3709 * @mode: which threads
3710 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3712 * The sync wakeup differs that the waker knows that it will schedule
3713 * away soon, so while the target thread will be woken up, it will not
3714 * be migrated to another CPU - ie. the two threads are 'synchronized'
3715 * with each other. This can prevent needless bouncing between CPUs.
3717 * On UP it can prevent extra preemption.
3720 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3722 unsigned long flags;
3728 if (unlikely(!nr_exclusive))
3731 spin_lock_irqsave(&q->lock, flags);
3732 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3733 spin_unlock_irqrestore(&q->lock, flags);
3735 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3737 void fastcall complete(struct completion *x)
3739 unsigned long flags;
3741 spin_lock_irqsave(&x->wait.lock, flags);
3743 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3745 spin_unlock_irqrestore(&x->wait.lock, flags);
3747 EXPORT_SYMBOL(complete);
3749 void fastcall complete_all(struct completion *x)
3751 unsigned long flags;
3753 spin_lock_irqsave(&x->wait.lock, flags);
3754 x->done += UINT_MAX/2;
3755 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3757 spin_unlock_irqrestore(&x->wait.lock, flags);
3759 EXPORT_SYMBOL(complete_all);
3761 void fastcall __sched wait_for_completion(struct completion *x)
3765 spin_lock_irq(&x->wait.lock);
3767 DECLARE_WAITQUEUE(wait, current);
3769 wait.flags |= WQ_FLAG_EXCLUSIVE;
3770 __add_wait_queue_tail(&x->wait, &wait);
3772 __set_current_state(TASK_UNINTERRUPTIBLE);
3773 spin_unlock_irq(&x->wait.lock);
3775 spin_lock_irq(&x->wait.lock);
3777 __remove_wait_queue(&x->wait, &wait);
3780 spin_unlock_irq(&x->wait.lock);
3782 EXPORT_SYMBOL(wait_for_completion);
3784 unsigned long fastcall __sched
3785 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3789 spin_lock_irq(&x->wait.lock);
3791 DECLARE_WAITQUEUE(wait, current);
3793 wait.flags |= WQ_FLAG_EXCLUSIVE;
3794 __add_wait_queue_tail(&x->wait, &wait);
3796 __set_current_state(TASK_UNINTERRUPTIBLE);
3797 spin_unlock_irq(&x->wait.lock);
3798 timeout = schedule_timeout(timeout);
3799 spin_lock_irq(&x->wait.lock);
3801 __remove_wait_queue(&x->wait, &wait);
3805 __remove_wait_queue(&x->wait, &wait);
3809 spin_unlock_irq(&x->wait.lock);
3812 EXPORT_SYMBOL(wait_for_completion_timeout);
3814 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3820 spin_lock_irq(&x->wait.lock);
3822 DECLARE_WAITQUEUE(wait, current);
3824 wait.flags |= WQ_FLAG_EXCLUSIVE;
3825 __add_wait_queue_tail(&x->wait, &wait);
3827 if (signal_pending(current)) {
3829 __remove_wait_queue(&x->wait, &wait);
3832 __set_current_state(TASK_INTERRUPTIBLE);
3833 spin_unlock_irq(&x->wait.lock);
3835 spin_lock_irq(&x->wait.lock);
3837 __remove_wait_queue(&x->wait, &wait);
3841 spin_unlock_irq(&x->wait.lock);
3845 EXPORT_SYMBOL(wait_for_completion_interruptible);
3847 unsigned long fastcall __sched
3848 wait_for_completion_interruptible_timeout(struct completion *x,
3849 unsigned long timeout)
3853 spin_lock_irq(&x->wait.lock);
3855 DECLARE_WAITQUEUE(wait, current);
3857 wait.flags |= WQ_FLAG_EXCLUSIVE;
3858 __add_wait_queue_tail(&x->wait, &wait);
3860 if (signal_pending(current)) {
3861 timeout = -ERESTARTSYS;
3862 __remove_wait_queue(&x->wait, &wait);
3865 __set_current_state(TASK_INTERRUPTIBLE);
3866 spin_unlock_irq(&x->wait.lock);
3867 timeout = schedule_timeout(timeout);
3868 spin_lock_irq(&x->wait.lock);
3870 __remove_wait_queue(&x->wait, &wait);
3874 __remove_wait_queue(&x->wait, &wait);
3878 spin_unlock_irq(&x->wait.lock);
3881 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3884 #define SLEEP_ON_VAR \
3885 unsigned long flags; \
3886 wait_queue_t wait; \
3887 init_waitqueue_entry(&wait, current);
3889 #define SLEEP_ON_HEAD \
3890 spin_lock_irqsave(&q->lock,flags); \
3891 __add_wait_queue(q, &wait); \
3892 spin_unlock(&q->lock);
3894 #define SLEEP_ON_TAIL \
3895 spin_lock_irq(&q->lock); \
3896 __remove_wait_queue(q, &wait); \
3897 spin_unlock_irqrestore(&q->lock, flags);
3899 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3903 current->state = TASK_INTERRUPTIBLE;
3909 EXPORT_SYMBOL(interruptible_sleep_on);
3911 long fastcall __sched
3912 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3916 current->state = TASK_INTERRUPTIBLE;
3919 timeout = schedule_timeout(timeout);
3924 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3926 void fastcall __sched sleep_on(wait_queue_head_t *q)
3930 current->state = TASK_UNINTERRUPTIBLE;
3936 EXPORT_SYMBOL(sleep_on);
3938 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3942 current->state = TASK_UNINTERRUPTIBLE;
3945 timeout = schedule_timeout(timeout);
3951 EXPORT_SYMBOL(sleep_on_timeout);
3953 #ifdef CONFIG_RT_MUTEXES
3956 * rt_mutex_setprio - set the current priority of a task
3958 * @prio: prio value (kernel-internal form)
3960 * This function changes the 'effective' priority of a task. It does
3961 * not touch ->normal_prio like __setscheduler().
3963 * Used by the rt_mutex code to implement priority inheritance logic.
3965 void rt_mutex_setprio(struct task_struct *p, int prio)
3967 struct prio_array *array;
3968 unsigned long flags;
3972 BUG_ON(prio < 0 || prio > MAX_PRIO);
3974 rq = task_rq_lock(p, &flags);
3979 dequeue_task(p, array);
3984 * If changing to an RT priority then queue it
3985 * in the active array!
3989 enqueue_task(p, array);
3991 * Reschedule if we are currently running on this runqueue and
3992 * our priority decreased, or if we are not currently running on
3993 * this runqueue and our priority is higher than the current's
3995 if (task_running(rq, p)) {
3996 if (p->prio > oldprio)
3997 resched_task(rq->curr);
3998 } else if (TASK_PREEMPTS_CURR(p, rq))
3999 resched_task(rq->curr);
4001 task_rq_unlock(rq, &flags);
4006 void set_user_nice(struct task_struct *p, long nice)
4008 struct prio_array *array;
4009 int old_prio, delta;
4010 unsigned long flags;
4013 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4016 * We have to be careful, if called from sys_setpriority(),
4017 * the task might be in the middle of scheduling on another CPU.
4019 rq = task_rq_lock(p, &flags);
4021 * The RT priorities are set via sched_setscheduler(), but we still
4022 * allow the 'normal' nice value to be set - but as expected
4023 * it wont have any effect on scheduling until the task is
4024 * not SCHED_NORMAL/SCHED_BATCH:
4026 if (has_rt_policy(p)) {
4027 p->static_prio = NICE_TO_PRIO(nice);
4032 dequeue_task(p, array);
4033 dec_raw_weighted_load(rq, p);
4036 p->static_prio = NICE_TO_PRIO(nice);
4039 p->prio = effective_prio(p);
4040 delta = p->prio - old_prio;
4043 enqueue_task(p, array);
4044 inc_raw_weighted_load(rq, p);
4046 * If the task increased its priority or is running and
4047 * lowered its priority, then reschedule its CPU:
4049 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4050 resched_task(rq->curr);
4053 task_rq_unlock(rq, &flags);
4055 EXPORT_SYMBOL(set_user_nice);
4058 * can_nice - check if a task can reduce its nice value
4062 int can_nice(const struct task_struct *p, const int nice)
4064 /* convert nice value [19,-20] to rlimit style value [1,40] */
4065 int nice_rlim = 20 - nice;
4067 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4068 capable(CAP_SYS_NICE));
4071 #ifdef __ARCH_WANT_SYS_NICE
4074 * sys_nice - change the priority of the current process.
4075 * @increment: priority increment
4077 * sys_setpriority is a more generic, but much slower function that
4078 * does similar things.
4080 asmlinkage long sys_nice(int increment)
4085 * Setpriority might change our priority at the same moment.
4086 * We don't have to worry. Conceptually one call occurs first
4087 * and we have a single winner.
4089 if (increment < -40)
4094 nice = PRIO_TO_NICE(current->static_prio) + increment;
4100 if (increment < 0 && !can_nice(current, nice))
4103 retval = security_task_setnice(current, nice);
4107 set_user_nice(current, nice);
4114 * task_prio - return the priority value of a given task.
4115 * @p: the task in question.
4117 * This is the priority value as seen by users in /proc.
4118 * RT tasks are offset by -200. Normal tasks are centered
4119 * around 0, value goes from -16 to +15.
4121 int task_prio(const struct task_struct *p)
4123 return p->prio - MAX_RT_PRIO;
4127 * task_nice - return the nice value of a given task.
4128 * @p: the task in question.
4130 int task_nice(const struct task_struct *p)
4132 return TASK_NICE(p);
4134 EXPORT_SYMBOL_GPL(task_nice);
4137 * idle_cpu - is a given cpu idle currently?
4138 * @cpu: the processor in question.
4140 int idle_cpu(int cpu)
4142 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4146 * idle_task - return the idle task for a given cpu.
4147 * @cpu: the processor in question.
4149 struct task_struct *idle_task(int cpu)
4151 return cpu_rq(cpu)->idle;
4155 * find_process_by_pid - find a process with a matching PID value.
4156 * @pid: the pid in question.
4158 static inline struct task_struct *find_process_by_pid(pid_t pid)
4160 return pid ? find_task_by_pid(pid) : current;
4163 /* Actually do priority change: must hold rq lock. */
4164 static void __setscheduler(struct task_struct *p, int policy, int prio)
4169 p->rt_priority = prio;
4170 p->normal_prio = normal_prio(p);
4171 /* we are holding p->pi_lock already */
4172 p->prio = rt_mutex_getprio(p);
4174 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4176 if (policy == SCHED_BATCH)
4182 * sched_setscheduler - change the scheduling policy and/or RT priority of
4184 * @p: the task in question.
4185 * @policy: new policy.
4186 * @param: structure containing the new RT priority.
4188 * NOTE: the task may be already dead
4190 int sched_setscheduler(struct task_struct *p, int policy,
4191 struct sched_param *param)
4193 int retval, oldprio, oldpolicy = -1;
4194 struct prio_array *array;
4195 unsigned long flags;
4198 /* may grab non-irq protected spin_locks */
4199 BUG_ON(in_interrupt());
4201 /* double check policy once rq lock held */
4203 policy = oldpolicy = p->policy;
4204 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4205 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4208 * Valid priorities for SCHED_FIFO and SCHED_RR are
4209 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4212 if (param->sched_priority < 0 ||
4213 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4214 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4216 if (is_rt_policy(policy) != (param->sched_priority != 0))
4220 * Allow unprivileged RT tasks to decrease priority:
4222 if (!capable(CAP_SYS_NICE)) {
4223 if (is_rt_policy(policy)) {
4224 unsigned long rlim_rtprio;
4225 unsigned long flags;
4227 if (!lock_task_sighand(p, &flags))
4229 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4230 unlock_task_sighand(p, &flags);
4232 /* can't set/change the rt policy */
4233 if (policy != p->policy && !rlim_rtprio)
4236 /* can't increase priority */
4237 if (param->sched_priority > p->rt_priority &&
4238 param->sched_priority > rlim_rtprio)
4242 /* can't change other user's priorities */
4243 if ((current->euid != p->euid) &&
4244 (current->euid != p->uid))
4248 retval = security_task_setscheduler(p, policy, param);
4252 * make sure no PI-waiters arrive (or leave) while we are
4253 * changing the priority of the task:
4255 spin_lock_irqsave(&p->pi_lock, flags);
4257 * To be able to change p->policy safely, the apropriate
4258 * runqueue lock must be held.
4260 rq = __task_rq_lock(p);
4261 /* recheck policy now with rq lock held */
4262 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4263 policy = oldpolicy = -1;
4264 __task_rq_unlock(rq);
4265 spin_unlock_irqrestore(&p->pi_lock, flags);
4270 deactivate_task(p, rq);
4272 __setscheduler(p, policy, param->sched_priority);
4274 __activate_task(p, rq);
4276 * Reschedule if we are currently running on this runqueue and
4277 * our priority decreased, or if we are not currently running on
4278 * this runqueue and our priority is higher than the current's
4280 if (task_running(rq, p)) {
4281 if (p->prio > oldprio)
4282 resched_task(rq->curr);
4283 } else if (TASK_PREEMPTS_CURR(p, rq))
4284 resched_task(rq->curr);
4286 __task_rq_unlock(rq);
4287 spin_unlock_irqrestore(&p->pi_lock, flags);
4289 rt_mutex_adjust_pi(p);
4293 EXPORT_SYMBOL_GPL(sched_setscheduler);
4296 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4298 struct sched_param lparam;
4299 struct task_struct *p;
4302 if (!param || pid < 0)
4304 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4309 p = find_process_by_pid(pid);
4311 retval = sched_setscheduler(p, policy, &lparam);
4318 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4319 * @pid: the pid in question.
4320 * @policy: new policy.
4321 * @param: structure containing the new RT priority.
4323 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4324 struct sched_param __user *param)
4326 /* negative values for policy are not valid */
4330 return do_sched_setscheduler(pid, policy, param);
4334 * sys_sched_setparam - set/change the RT priority of a thread
4335 * @pid: the pid in question.
4336 * @param: structure containing the new RT priority.
4338 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4340 return do_sched_setscheduler(pid, -1, param);
4344 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4345 * @pid: the pid in question.
4347 asmlinkage long sys_sched_getscheduler(pid_t pid)
4349 struct task_struct *p;
4350 int retval = -EINVAL;
4356 read_lock(&tasklist_lock);
4357 p = find_process_by_pid(pid);
4359 retval = security_task_getscheduler(p);
4363 read_unlock(&tasklist_lock);
4370 * sys_sched_getscheduler - get the RT priority of a thread
4371 * @pid: the pid in question.
4372 * @param: structure containing the RT priority.
4374 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4376 struct sched_param lp;
4377 struct task_struct *p;
4378 int retval = -EINVAL;
4380 if (!param || pid < 0)
4383 read_lock(&tasklist_lock);
4384 p = find_process_by_pid(pid);
4389 retval = security_task_getscheduler(p);
4393 lp.sched_priority = p->rt_priority;
4394 read_unlock(&tasklist_lock);
4397 * This one might sleep, we cannot do it with a spinlock held ...
4399 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4405 read_unlock(&tasklist_lock);
4409 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4411 cpumask_t cpus_allowed;
4412 struct task_struct *p;
4416 read_lock(&tasklist_lock);
4418 p = find_process_by_pid(pid);
4420 read_unlock(&tasklist_lock);
4421 unlock_cpu_hotplug();
4426 * It is not safe to call set_cpus_allowed with the
4427 * tasklist_lock held. We will bump the task_struct's
4428 * usage count and then drop tasklist_lock.
4431 read_unlock(&tasklist_lock);
4434 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4435 !capable(CAP_SYS_NICE))
4438 retval = security_task_setscheduler(p, 0, NULL);
4442 cpus_allowed = cpuset_cpus_allowed(p);
4443 cpus_and(new_mask, new_mask, cpus_allowed);
4444 retval = set_cpus_allowed(p, new_mask);
4448 unlock_cpu_hotplug();
4452 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4453 cpumask_t *new_mask)
4455 if (len < sizeof(cpumask_t)) {
4456 memset(new_mask, 0, sizeof(cpumask_t));
4457 } else if (len > sizeof(cpumask_t)) {
4458 len = sizeof(cpumask_t);
4460 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4464 * sys_sched_setaffinity - set the cpu affinity of a process
4465 * @pid: pid of the process
4466 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4467 * @user_mask_ptr: user-space pointer to the new cpu mask
4469 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4470 unsigned long __user *user_mask_ptr)
4475 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4479 return sched_setaffinity(pid, new_mask);
4483 * Represents all cpu's present in the system
4484 * In systems capable of hotplug, this map could dynamically grow
4485 * as new cpu's are detected in the system via any platform specific
4486 * method, such as ACPI for e.g.
4489 cpumask_t cpu_present_map __read_mostly;
4490 EXPORT_SYMBOL(cpu_present_map);
4493 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4494 EXPORT_SYMBOL(cpu_online_map);
4496 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4497 EXPORT_SYMBOL(cpu_possible_map);
4500 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4502 struct task_struct *p;
4506 read_lock(&tasklist_lock);
4509 p = find_process_by_pid(pid);
4513 retval = security_task_getscheduler(p);
4517 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4520 read_unlock(&tasklist_lock);
4521 unlock_cpu_hotplug();
4529 * sys_sched_getaffinity - get the cpu affinity of a process
4530 * @pid: pid of the process
4531 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4532 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4534 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4535 unsigned long __user *user_mask_ptr)
4540 if (len < sizeof(cpumask_t))
4543 ret = sched_getaffinity(pid, &mask);
4547 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4550 return sizeof(cpumask_t);
4554 * sys_sched_yield - yield the current processor to other threads.
4556 * this function yields the current CPU by moving the calling thread
4557 * to the expired array. If there are no other threads running on this
4558 * CPU then this function will return.
4560 asmlinkage long sys_sched_yield(void)
4562 struct rq *rq = this_rq_lock();
4563 struct prio_array *array = current->array, *target = rq->expired;
4565 schedstat_inc(rq, yld_cnt);
4567 * We implement yielding by moving the task into the expired
4570 * (special rule: RT tasks will just roundrobin in the active
4573 if (rt_task(current))
4574 target = rq->active;
4576 if (array->nr_active == 1) {
4577 schedstat_inc(rq, yld_act_empty);
4578 if (!rq->expired->nr_active)
4579 schedstat_inc(rq, yld_both_empty);
4580 } else if (!rq->expired->nr_active)
4581 schedstat_inc(rq, yld_exp_empty);
4583 if (array != target) {
4584 dequeue_task(current, array);
4585 enqueue_task(current, target);
4588 * requeue_task is cheaper so perform that if possible.
4590 requeue_task(current, array);
4593 * Since we are going to call schedule() anyway, there's
4594 * no need to preempt or enable interrupts:
4596 __release(rq->lock);
4597 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4598 _raw_spin_unlock(&rq->lock);
4599 preempt_enable_no_resched();
4606 static inline int __resched_legal(int expected_preempt_count)
4608 if (unlikely(preempt_count() != expected_preempt_count))
4610 if (unlikely(system_state != SYSTEM_RUNNING))
4615 static void __cond_resched(void)
4617 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4618 __might_sleep(__FILE__, __LINE__);
4621 * The BKS might be reacquired before we have dropped
4622 * PREEMPT_ACTIVE, which could trigger a second
4623 * cond_resched() call.
4626 add_preempt_count(PREEMPT_ACTIVE);
4628 sub_preempt_count(PREEMPT_ACTIVE);
4629 } while (need_resched());
4632 int __sched cond_resched(void)
4634 if (need_resched() && __resched_legal(0)) {
4640 EXPORT_SYMBOL(cond_resched);
4643 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4644 * call schedule, and on return reacquire the lock.
4646 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4647 * operations here to prevent schedule() from being called twice (once via
4648 * spin_unlock(), once by hand).
4650 int cond_resched_lock(spinlock_t *lock)
4654 if (need_lockbreak(lock)) {
4660 if (need_resched() && __resched_legal(1)) {
4661 spin_release(&lock->dep_map, 1, _THIS_IP_);
4662 _raw_spin_unlock(lock);
4663 preempt_enable_no_resched();
4670 EXPORT_SYMBOL(cond_resched_lock);
4672 int __sched cond_resched_softirq(void)
4674 BUG_ON(!in_softirq());
4676 if (need_resched() && __resched_legal(0)) {
4677 raw_local_irq_disable();
4679 raw_local_irq_enable();
4686 EXPORT_SYMBOL(cond_resched_softirq);
4689 * yield - yield the current processor to other threads.
4691 * this is a shortcut for kernel-space yielding - it marks the
4692 * thread runnable and calls sys_sched_yield().
4694 void __sched yield(void)
4696 set_current_state(TASK_RUNNING);
4699 EXPORT_SYMBOL(yield);
4702 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4703 * that process accounting knows that this is a task in IO wait state.
4705 * But don't do that if it is a deliberate, throttling IO wait (this task
4706 * has set its backing_dev_info: the queue against which it should throttle)
4708 void __sched io_schedule(void)
4710 struct rq *rq = &__raw_get_cpu_var(runqueues);
4712 delayacct_blkio_start();
4713 atomic_inc(&rq->nr_iowait);
4715 atomic_dec(&rq->nr_iowait);
4716 delayacct_blkio_end();
4718 EXPORT_SYMBOL(io_schedule);
4720 long __sched io_schedule_timeout(long timeout)
4722 struct rq *rq = &__raw_get_cpu_var(runqueues);
4725 delayacct_blkio_start();
4726 atomic_inc(&rq->nr_iowait);
4727 ret = schedule_timeout(timeout);
4728 atomic_dec(&rq->nr_iowait);
4729 delayacct_blkio_end();
4734 * sys_sched_get_priority_max - return maximum RT priority.
4735 * @policy: scheduling class.
4737 * this syscall returns the maximum rt_priority that can be used
4738 * by a given scheduling class.
4740 asmlinkage long sys_sched_get_priority_max(int policy)
4747 ret = MAX_USER_RT_PRIO-1;
4758 * sys_sched_get_priority_min - return minimum RT priority.
4759 * @policy: scheduling class.
4761 * this syscall returns the minimum rt_priority that can be used
4762 * by a given scheduling class.
4764 asmlinkage long sys_sched_get_priority_min(int policy)
4781 * sys_sched_rr_get_interval - return the default timeslice of a process.
4782 * @pid: pid of the process.
4783 * @interval: userspace pointer to the timeslice value.
4785 * this syscall writes the default timeslice value of a given process
4786 * into the user-space timespec buffer. A value of '0' means infinity.
4789 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4791 struct task_struct *p;
4792 int retval = -EINVAL;
4799 read_lock(&tasklist_lock);
4800 p = find_process_by_pid(pid);
4804 retval = security_task_getscheduler(p);
4808 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4809 0 : task_timeslice(p), &t);
4810 read_unlock(&tasklist_lock);
4811 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4815 read_unlock(&tasklist_lock);
4819 static inline struct task_struct *eldest_child(struct task_struct *p)
4821 if (list_empty(&p->children))
4823 return list_entry(p->children.next,struct task_struct,sibling);
4826 static inline struct task_struct *older_sibling(struct task_struct *p)
4828 if (p->sibling.prev==&p->parent->children)
4830 return list_entry(p->sibling.prev,struct task_struct,sibling);
4833 static inline struct task_struct *younger_sibling(struct task_struct *p)
4835 if (p->sibling.next==&p->parent->children)
4837 return list_entry(p->sibling.next,struct task_struct,sibling);
4840 static const char stat_nam[] = "RSDTtZX";
4842 static void show_task(struct task_struct *p)
4844 struct task_struct *relative;
4845 unsigned long free = 0;
4848 state = p->state ? __ffs(p->state) + 1 : 0;
4849 printk("%-13.13s %c", p->comm,
4850 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4851 #if (BITS_PER_LONG == 32)
4852 if (state == TASK_RUNNING)
4853 printk(" running ");
4855 printk(" %08lX ", thread_saved_pc(p));
4857 if (state == TASK_RUNNING)
4858 printk(" running task ");
4860 printk(" %016lx ", thread_saved_pc(p));
4862 #ifdef CONFIG_DEBUG_STACK_USAGE
4864 unsigned long *n = end_of_stack(p);
4867 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4870 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4871 if ((relative = eldest_child(p)))
4872 printk("%5d ", relative->pid);
4875 if ((relative = younger_sibling(p)))
4876 printk("%7d", relative->pid);
4879 if ((relative = older_sibling(p)))
4880 printk(" %5d", relative->pid);
4884 printk(" (L-TLB)\n");
4886 printk(" (NOTLB)\n");
4888 if (state != TASK_RUNNING)
4889 show_stack(p, NULL);
4892 void show_state_filter(unsigned long state_filter)
4894 struct task_struct *g, *p;
4896 #if (BITS_PER_LONG == 32)
4899 printk(" task PC stack pid father child younger older\n");
4903 printk(" task PC stack pid father child younger older\n");
4905 read_lock(&tasklist_lock);
4906 do_each_thread(g, p) {
4908 * reset the NMI-timeout, listing all files on a slow
4909 * console might take alot of time:
4911 touch_nmi_watchdog();
4912 if (p->state & state_filter)
4914 } while_each_thread(g, p);
4916 read_unlock(&tasklist_lock);
4918 * Only show locks if all tasks are dumped:
4920 if (state_filter == -1)
4921 debug_show_all_locks();
4925 * init_idle - set up an idle thread for a given CPU
4926 * @idle: task in question
4927 * @cpu: cpu the idle task belongs to
4929 * NOTE: this function does not set the idle thread's NEED_RESCHED
4930 * flag, to make booting more robust.
4932 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4934 struct rq *rq = cpu_rq(cpu);
4935 unsigned long flags;
4937 idle->timestamp = sched_clock();
4938 idle->sleep_avg = 0;
4940 idle->prio = idle->normal_prio = MAX_PRIO;
4941 idle->state = TASK_RUNNING;
4942 idle->cpus_allowed = cpumask_of_cpu(cpu);
4943 set_task_cpu(idle, cpu);
4945 spin_lock_irqsave(&rq->lock, flags);
4946 rq->curr = rq->idle = idle;
4947 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4950 spin_unlock_irqrestore(&rq->lock, flags);
4952 /* Set the preempt count _outside_ the spinlocks! */
4953 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4954 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4956 task_thread_info(idle)->preempt_count = 0;
4961 * In a system that switches off the HZ timer nohz_cpu_mask
4962 * indicates which cpus entered this state. This is used
4963 * in the rcu update to wait only for active cpus. For system
4964 * which do not switch off the HZ timer nohz_cpu_mask should
4965 * always be CPU_MASK_NONE.
4967 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4971 * This is how migration works:
4973 * 1) we queue a struct migration_req structure in the source CPU's
4974 * runqueue and wake up that CPU's migration thread.
4975 * 2) we down() the locked semaphore => thread blocks.
4976 * 3) migration thread wakes up (implicitly it forces the migrated
4977 * thread off the CPU)
4978 * 4) it gets the migration request and checks whether the migrated
4979 * task is still in the wrong runqueue.
4980 * 5) if it's in the wrong runqueue then the migration thread removes
4981 * it and puts it into the right queue.
4982 * 6) migration thread up()s the semaphore.
4983 * 7) we wake up and the migration is done.
4987 * Change a given task's CPU affinity. Migrate the thread to a
4988 * proper CPU and schedule it away if the CPU it's executing on
4989 * is removed from the allowed bitmask.
4991 * NOTE: the caller must have a valid reference to the task, the
4992 * task must not exit() & deallocate itself prematurely. The
4993 * call is not atomic; no spinlocks may be held.
4995 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4997 struct migration_req req;
4998 unsigned long flags;
5002 rq = task_rq_lock(p, &flags);
5003 if (!cpus_intersects(new_mask, cpu_online_map)) {
5008 p->cpus_allowed = new_mask;
5009 /* Can the task run on the task's current CPU? If so, we're done */
5010 if (cpu_isset(task_cpu(p), new_mask))
5013 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5014 /* Need help from migration thread: drop lock and wait. */
5015 task_rq_unlock(rq, &flags);
5016 wake_up_process(rq->migration_thread);
5017 wait_for_completion(&req.done);
5018 tlb_migrate_finish(p->mm);
5022 task_rq_unlock(rq, &flags);
5026 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5029 * Move (not current) task off this cpu, onto dest cpu. We're doing
5030 * this because either it can't run here any more (set_cpus_allowed()
5031 * away from this CPU, or CPU going down), or because we're
5032 * attempting to rebalance this task on exec (sched_exec).
5034 * So we race with normal scheduler movements, but that's OK, as long
5035 * as the task is no longer on this CPU.
5037 * Returns non-zero if task was successfully migrated.
5039 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5041 struct rq *rq_dest, *rq_src;
5044 if (unlikely(cpu_is_offline(dest_cpu)))
5047 rq_src = cpu_rq(src_cpu);
5048 rq_dest = cpu_rq(dest_cpu);
5050 double_rq_lock(rq_src, rq_dest);
5051 /* Already moved. */
5052 if (task_cpu(p) != src_cpu)
5054 /* Affinity changed (again). */
5055 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5058 set_task_cpu(p, dest_cpu);
5061 * Sync timestamp with rq_dest's before activating.
5062 * The same thing could be achieved by doing this step
5063 * afterwards, and pretending it was a local activate.
5064 * This way is cleaner and logically correct.
5066 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5067 + rq_dest->most_recent_timestamp;
5068 deactivate_task(p, rq_src);
5069 __activate_task(p, rq_dest);
5070 if (TASK_PREEMPTS_CURR(p, rq_dest))
5071 resched_task(rq_dest->curr);
5075 double_rq_unlock(rq_src, rq_dest);
5080 * migration_thread - this is a highprio system thread that performs
5081 * thread migration by bumping thread off CPU then 'pushing' onto
5084 static int migration_thread(void *data)
5086 int cpu = (long)data;
5090 BUG_ON(rq->migration_thread != current);
5092 set_current_state(TASK_INTERRUPTIBLE);
5093 while (!kthread_should_stop()) {
5094 struct migration_req *req;
5095 struct list_head *head;
5099 spin_lock_irq(&rq->lock);
5101 if (cpu_is_offline(cpu)) {
5102 spin_unlock_irq(&rq->lock);
5106 if (rq->active_balance) {
5107 active_load_balance(rq, cpu);
5108 rq->active_balance = 0;
5111 head = &rq->migration_queue;
5113 if (list_empty(head)) {
5114 spin_unlock_irq(&rq->lock);
5116 set_current_state(TASK_INTERRUPTIBLE);
5119 req = list_entry(head->next, struct migration_req, list);
5120 list_del_init(head->next);
5122 spin_unlock(&rq->lock);
5123 __migrate_task(req->task, cpu, req->dest_cpu);
5126 complete(&req->done);
5128 __set_current_state(TASK_RUNNING);
5132 /* Wait for kthread_stop */
5133 set_current_state(TASK_INTERRUPTIBLE);
5134 while (!kthread_should_stop()) {
5136 set_current_state(TASK_INTERRUPTIBLE);
5138 __set_current_state(TASK_RUNNING);
5142 #ifdef CONFIG_HOTPLUG_CPU
5144 * Figure out where task on dead CPU should go, use force if neccessary.
5145 * NOTE: interrupts should be disabled by the caller
5147 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5149 unsigned long flags;
5156 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5157 cpus_and(mask, mask, p->cpus_allowed);
5158 dest_cpu = any_online_cpu(mask);
5160 /* On any allowed CPU? */
5161 if (dest_cpu == NR_CPUS)
5162 dest_cpu = any_online_cpu(p->cpus_allowed);
5164 /* No more Mr. Nice Guy. */
5165 if (dest_cpu == NR_CPUS) {
5166 rq = task_rq_lock(p, &flags);
5167 cpus_setall(p->cpus_allowed);
5168 dest_cpu = any_online_cpu(p->cpus_allowed);
5169 task_rq_unlock(rq, &flags);
5172 * Don't tell them about moving exiting tasks or
5173 * kernel threads (both mm NULL), since they never
5176 if (p->mm && printk_ratelimit())
5177 printk(KERN_INFO "process %d (%s) no "
5178 "longer affine to cpu%d\n",
5179 p->pid, p->comm, dead_cpu);
5181 if (!__migrate_task(p, dead_cpu, dest_cpu))
5186 * While a dead CPU has no uninterruptible tasks queued at this point,
5187 * it might still have a nonzero ->nr_uninterruptible counter, because
5188 * for performance reasons the counter is not stricly tracking tasks to
5189 * their home CPUs. So we just add the counter to another CPU's counter,
5190 * to keep the global sum constant after CPU-down:
5192 static void migrate_nr_uninterruptible(struct rq *rq_src)
5194 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5195 unsigned long flags;
5197 local_irq_save(flags);
5198 double_rq_lock(rq_src, rq_dest);
5199 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5200 rq_src->nr_uninterruptible = 0;
5201 double_rq_unlock(rq_src, rq_dest);
5202 local_irq_restore(flags);
5205 /* Run through task list and migrate tasks from the dead cpu. */
5206 static void migrate_live_tasks(int src_cpu)
5208 struct task_struct *p, *t;
5210 write_lock_irq(&tasklist_lock);
5212 do_each_thread(t, p) {
5216 if (task_cpu(p) == src_cpu)
5217 move_task_off_dead_cpu(src_cpu, p);
5218 } while_each_thread(t, p);
5220 write_unlock_irq(&tasklist_lock);
5223 /* Schedules idle task to be the next runnable task on current CPU.
5224 * It does so by boosting its priority to highest possible and adding it to
5225 * the _front_ of the runqueue. Used by CPU offline code.
5227 void sched_idle_next(void)
5229 int this_cpu = smp_processor_id();
5230 struct rq *rq = cpu_rq(this_cpu);
5231 struct task_struct *p = rq->idle;
5232 unsigned long flags;
5234 /* cpu has to be offline */
5235 BUG_ON(cpu_online(this_cpu));
5238 * Strictly not necessary since rest of the CPUs are stopped by now
5239 * and interrupts disabled on the current cpu.
5241 spin_lock_irqsave(&rq->lock, flags);
5243 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5245 /* Add idle task to the _front_ of its priority queue: */
5246 __activate_idle_task(p, rq);
5248 spin_unlock_irqrestore(&rq->lock, flags);
5252 * Ensures that the idle task is using init_mm right before its cpu goes
5255 void idle_task_exit(void)
5257 struct mm_struct *mm = current->active_mm;
5259 BUG_ON(cpu_online(smp_processor_id()));
5262 switch_mm(mm, &init_mm, current);
5266 /* called under rq->lock with disabled interrupts */
5267 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5269 struct rq *rq = cpu_rq(dead_cpu);
5271 /* Must be exiting, otherwise would be on tasklist. */
5272 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5274 /* Cannot have done final schedule yet: would have vanished. */
5275 BUG_ON(p->state == TASK_DEAD);
5280 * Drop lock around migration; if someone else moves it,
5281 * that's OK. No task can be added to this CPU, so iteration is
5283 * NOTE: interrupts should be left disabled --dev@
5285 spin_unlock(&rq->lock);
5286 move_task_off_dead_cpu(dead_cpu, p);
5287 spin_lock(&rq->lock);
5292 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5293 static void migrate_dead_tasks(unsigned int dead_cpu)
5295 struct rq *rq = cpu_rq(dead_cpu);
5296 unsigned int arr, i;
5298 for (arr = 0; arr < 2; arr++) {
5299 for (i = 0; i < MAX_PRIO; i++) {
5300 struct list_head *list = &rq->arrays[arr].queue[i];
5302 while (!list_empty(list))
5303 migrate_dead(dead_cpu, list_entry(list->next,
5304 struct task_struct, run_list));
5308 #endif /* CONFIG_HOTPLUG_CPU */
5311 * migration_call - callback that gets triggered when a CPU is added.
5312 * Here we can start up the necessary migration thread for the new CPU.
5314 static int __cpuinit
5315 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5317 struct task_struct *p;
5318 int cpu = (long)hcpu;
5319 unsigned long flags;
5323 case CPU_UP_PREPARE:
5324 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5327 p->flags |= PF_NOFREEZE;
5328 kthread_bind(p, cpu);
5329 /* Must be high prio: stop_machine expects to yield to it. */
5330 rq = task_rq_lock(p, &flags);
5331 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5332 task_rq_unlock(rq, &flags);
5333 cpu_rq(cpu)->migration_thread = p;
5337 /* Strictly unneccessary, as first user will wake it. */
5338 wake_up_process(cpu_rq(cpu)->migration_thread);
5341 #ifdef CONFIG_HOTPLUG_CPU
5342 case CPU_UP_CANCELED:
5343 if (!cpu_rq(cpu)->migration_thread)
5345 /* Unbind it from offline cpu so it can run. Fall thru. */
5346 kthread_bind(cpu_rq(cpu)->migration_thread,
5347 any_online_cpu(cpu_online_map));
5348 kthread_stop(cpu_rq(cpu)->migration_thread);
5349 cpu_rq(cpu)->migration_thread = NULL;
5353 migrate_live_tasks(cpu);
5355 kthread_stop(rq->migration_thread);
5356 rq->migration_thread = NULL;
5357 /* Idle task back to normal (off runqueue, low prio) */
5358 rq = task_rq_lock(rq->idle, &flags);
5359 deactivate_task(rq->idle, rq);
5360 rq->idle->static_prio = MAX_PRIO;
5361 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5362 migrate_dead_tasks(cpu);
5363 task_rq_unlock(rq, &flags);
5364 migrate_nr_uninterruptible(rq);
5365 BUG_ON(rq->nr_running != 0);
5367 /* No need to migrate the tasks: it was best-effort if
5368 * they didn't do lock_cpu_hotplug(). Just wake up
5369 * the requestors. */
5370 spin_lock_irq(&rq->lock);
5371 while (!list_empty(&rq->migration_queue)) {
5372 struct migration_req *req;
5374 req = list_entry(rq->migration_queue.next,
5375 struct migration_req, list);
5376 list_del_init(&req->list);
5377 complete(&req->done);
5379 spin_unlock_irq(&rq->lock);
5386 /* Register at highest priority so that task migration (migrate_all_tasks)
5387 * happens before everything else.
5389 static struct notifier_block __cpuinitdata migration_notifier = {
5390 .notifier_call = migration_call,
5394 int __init migration_init(void)
5396 void *cpu = (void *)(long)smp_processor_id();
5399 /* Start one for the boot CPU: */
5400 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5401 BUG_ON(err == NOTIFY_BAD);
5402 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5403 register_cpu_notifier(&migration_notifier);
5410 #undef SCHED_DOMAIN_DEBUG
5411 #ifdef SCHED_DOMAIN_DEBUG
5412 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5417 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5421 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5426 struct sched_group *group = sd->groups;
5427 cpumask_t groupmask;
5429 cpumask_scnprintf(str, NR_CPUS, sd->span);
5430 cpus_clear(groupmask);
5433 for (i = 0; i < level + 1; i++)
5435 printk("domain %d: ", level);
5437 if (!(sd->flags & SD_LOAD_BALANCE)) {
5438 printk("does not load-balance\n");
5440 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5444 printk("span %s\n", str);
5446 if (!cpu_isset(cpu, sd->span))
5447 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5448 if (!cpu_isset(cpu, group->cpumask))
5449 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5452 for (i = 0; i < level + 2; i++)
5458 printk(KERN_ERR "ERROR: group is NULL\n");
5462 if (!group->cpu_power) {
5464 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5467 if (!cpus_weight(group->cpumask)) {
5469 printk(KERN_ERR "ERROR: empty group\n");
5472 if (cpus_intersects(groupmask, group->cpumask)) {
5474 printk(KERN_ERR "ERROR: repeated CPUs\n");
5477 cpus_or(groupmask, groupmask, group->cpumask);
5479 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5482 group = group->next;
5483 } while (group != sd->groups);
5486 if (!cpus_equal(sd->span, groupmask))
5487 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5493 if (!cpus_subset(groupmask, sd->span))
5494 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5500 # define sched_domain_debug(sd, cpu) do { } while (0)
5503 static int sd_degenerate(struct sched_domain *sd)
5505 if (cpus_weight(sd->span) == 1)
5508 /* Following flags need at least 2 groups */
5509 if (sd->flags & (SD_LOAD_BALANCE |
5510 SD_BALANCE_NEWIDLE |
5514 SD_SHARE_PKG_RESOURCES)) {
5515 if (sd->groups != sd->groups->next)
5519 /* Following flags don't use groups */
5520 if (sd->flags & (SD_WAKE_IDLE |
5529 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5531 unsigned long cflags = sd->flags, pflags = parent->flags;
5533 if (sd_degenerate(parent))
5536 if (!cpus_equal(sd->span, parent->span))
5539 /* Does parent contain flags not in child? */
5540 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5541 if (cflags & SD_WAKE_AFFINE)
5542 pflags &= ~SD_WAKE_BALANCE;
5543 /* Flags needing groups don't count if only 1 group in parent */
5544 if (parent->groups == parent->groups->next) {
5545 pflags &= ~(SD_LOAD_BALANCE |
5546 SD_BALANCE_NEWIDLE |
5550 SD_SHARE_PKG_RESOURCES);
5552 if (~cflags & pflags)
5559 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5560 * hold the hotplug lock.
5562 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5564 struct rq *rq = cpu_rq(cpu);
5565 struct sched_domain *tmp;
5567 /* Remove the sched domains which do not contribute to scheduling. */
5568 for (tmp = sd; tmp; tmp = tmp->parent) {
5569 struct sched_domain *parent = tmp->parent;
5572 if (sd_parent_degenerate(tmp, parent)) {
5573 tmp->parent = parent->parent;
5575 parent->parent->child = tmp;
5579 if (sd && sd_degenerate(sd)) {
5585 sched_domain_debug(sd, cpu);
5587 rcu_assign_pointer(rq->sd, sd);
5590 /* cpus with isolated domains */
5591 static cpumask_t __cpuinitdata cpu_isolated_map = CPU_MASK_NONE;
5593 /* Setup the mask of cpus configured for isolated domains */
5594 static int __init isolated_cpu_setup(char *str)
5596 int ints[NR_CPUS], i;
5598 str = get_options(str, ARRAY_SIZE(ints), ints);
5599 cpus_clear(cpu_isolated_map);
5600 for (i = 1; i <= ints[0]; i++)
5601 if (ints[i] < NR_CPUS)
5602 cpu_set(ints[i], cpu_isolated_map);
5606 __setup ("isolcpus=", isolated_cpu_setup);
5609 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5610 * to a function which identifies what group(along with sched group) a CPU
5611 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5612 * (due to the fact that we keep track of groups covered with a cpumask_t).
5614 * init_sched_build_groups will build a circular linked list of the groups
5615 * covered by the given span, and will set each group's ->cpumask correctly,
5616 * and ->cpu_power to 0.
5619 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5620 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5621 struct sched_group **sg))
5623 struct sched_group *first = NULL, *last = NULL;
5624 cpumask_t covered = CPU_MASK_NONE;
5627 for_each_cpu_mask(i, span) {
5628 struct sched_group *sg;
5629 int group = group_fn(i, cpu_map, &sg);
5632 if (cpu_isset(i, covered))
5635 sg->cpumask = CPU_MASK_NONE;
5638 for_each_cpu_mask(j, span) {
5639 if (group_fn(j, cpu_map, NULL) != group)
5642 cpu_set(j, covered);
5643 cpu_set(j, sg->cpumask);
5654 #define SD_NODES_PER_DOMAIN 16
5657 * Self-tuning task migration cost measurement between source and target CPUs.
5659 * This is done by measuring the cost of manipulating buffers of varying
5660 * sizes. For a given buffer-size here are the steps that are taken:
5662 * 1) the source CPU reads+dirties a shared buffer
5663 * 2) the target CPU reads+dirties the same shared buffer
5665 * We measure how long they take, in the following 4 scenarios:
5667 * - source: CPU1, target: CPU2 | cost1
5668 * - source: CPU2, target: CPU1 | cost2
5669 * - source: CPU1, target: CPU1 | cost3
5670 * - source: CPU2, target: CPU2 | cost4
5672 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5673 * the cost of migration.
5675 * We then start off from a small buffer-size and iterate up to larger
5676 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5677 * doing a maximum search for the cost. (The maximum cost for a migration
5678 * normally occurs when the working set size is around the effective cache
5681 #define SEARCH_SCOPE 2
5682 #define MIN_CACHE_SIZE (64*1024U)
5683 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5684 #define ITERATIONS 1
5685 #define SIZE_THRESH 130
5686 #define COST_THRESH 130
5689 * The migration cost is a function of 'domain distance'. Domain
5690 * distance is the number of steps a CPU has to iterate down its
5691 * domain tree to share a domain with the other CPU. The farther
5692 * two CPUs are from each other, the larger the distance gets.
5694 * Note that we use the distance only to cache measurement results,
5695 * the distance value is not used numerically otherwise. When two
5696 * CPUs have the same distance it is assumed that the migration
5697 * cost is the same. (this is a simplification but quite practical)
5699 #define MAX_DOMAIN_DISTANCE 32
5701 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5702 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5704 * Architectures may override the migration cost and thus avoid
5705 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5706 * virtualized hardware:
5708 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5709 CONFIG_DEFAULT_MIGRATION_COST
5716 * Allow override of migration cost - in units of microseconds.
5717 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5718 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5720 static int __init migration_cost_setup(char *str)
5722 int ints[MAX_DOMAIN_DISTANCE+1], i;
5724 str = get_options(str, ARRAY_SIZE(ints), ints);
5726 printk("#ints: %d\n", ints[0]);
5727 for (i = 1; i <= ints[0]; i++) {
5728 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5729 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5734 __setup ("migration_cost=", migration_cost_setup);
5737 * Global multiplier (divisor) for migration-cutoff values,
5738 * in percentiles. E.g. use a value of 150 to get 1.5 times
5739 * longer cache-hot cutoff times.
5741 * (We scale it from 100 to 128 to long long handling easier.)
5744 #define MIGRATION_FACTOR_SCALE 128
5746 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5748 static int __init setup_migration_factor(char *str)
5750 get_option(&str, &migration_factor);
5751 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5755 __setup("migration_factor=", setup_migration_factor);
5758 * Estimated distance of two CPUs, measured via the number of domains
5759 * we have to pass for the two CPUs to be in the same span:
5761 static unsigned long domain_distance(int cpu1, int cpu2)
5763 unsigned long distance = 0;
5764 struct sched_domain *sd;
5766 for_each_domain(cpu1, sd) {
5767 WARN_ON(!cpu_isset(cpu1, sd->span));
5768 if (cpu_isset(cpu2, sd->span))
5772 if (distance >= MAX_DOMAIN_DISTANCE) {
5774 distance = MAX_DOMAIN_DISTANCE-1;
5780 static unsigned int migration_debug;
5782 static int __init setup_migration_debug(char *str)
5784 get_option(&str, &migration_debug);
5788 __setup("migration_debug=", setup_migration_debug);
5791 * Maximum cache-size that the scheduler should try to measure.
5792 * Architectures with larger caches should tune this up during
5793 * bootup. Gets used in the domain-setup code (i.e. during SMP
5796 unsigned int max_cache_size;
5798 static int __init setup_max_cache_size(char *str)
5800 get_option(&str, &max_cache_size);
5804 __setup("max_cache_size=", setup_max_cache_size);
5807 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5808 * is the operation that is timed, so we try to generate unpredictable
5809 * cachemisses that still end up filling the L2 cache:
5811 static void touch_cache(void *__cache, unsigned long __size)
5813 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5815 unsigned long *cache = __cache;
5818 for (i = 0; i < size/6; i += 8) {
5821 case 1: cache[size-1-i]++;
5822 case 2: cache[chunk1-i]++;
5823 case 3: cache[chunk1+i]++;
5824 case 4: cache[chunk2-i]++;
5825 case 5: cache[chunk2+i]++;
5831 * Measure the cache-cost of one task migration. Returns in units of nsec.
5833 static unsigned long long
5834 measure_one(void *cache, unsigned long size, int source, int target)
5836 cpumask_t mask, saved_mask;
5837 unsigned long long t0, t1, t2, t3, cost;
5839 saved_mask = current->cpus_allowed;
5842 * Flush source caches to RAM and invalidate them:
5847 * Migrate to the source CPU:
5849 mask = cpumask_of_cpu(source);
5850 set_cpus_allowed(current, mask);
5851 WARN_ON(smp_processor_id() != source);
5854 * Dirty the working set:
5857 touch_cache(cache, size);
5861 * Migrate to the target CPU, dirty the L2 cache and access
5862 * the shared buffer. (which represents the working set
5863 * of a migrated task.)
5865 mask = cpumask_of_cpu(target);
5866 set_cpus_allowed(current, mask);
5867 WARN_ON(smp_processor_id() != target);
5870 touch_cache(cache, size);
5873 cost = t1-t0 + t3-t2;
5875 if (migration_debug >= 2)
5876 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5877 source, target, t1-t0, t1-t0, t3-t2, cost);
5879 * Flush target caches to RAM and invalidate them:
5883 set_cpus_allowed(current, saved_mask);
5889 * Measure a series of task migrations and return the average
5890 * result. Since this code runs early during bootup the system
5891 * is 'undisturbed' and the average latency makes sense.
5893 * The algorithm in essence auto-detects the relevant cache-size,
5894 * so it will properly detect different cachesizes for different
5895 * cache-hierarchies, depending on how the CPUs are connected.
5897 * Architectures can prime the upper limit of the search range via
5898 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5900 static unsigned long long
5901 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5903 unsigned long long cost1, cost2;
5907 * Measure the migration cost of 'size' bytes, over an
5908 * average of 10 runs:
5910 * (We perturb the cache size by a small (0..4k)
5911 * value to compensate size/alignment related artifacts.
5912 * We also subtract the cost of the operation done on
5918 * dry run, to make sure we start off cache-cold on cpu1,
5919 * and to get any vmalloc pagefaults in advance:
5921 measure_one(cache, size, cpu1, cpu2);
5922 for (i = 0; i < ITERATIONS; i++)
5923 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5925 measure_one(cache, size, cpu2, cpu1);
5926 for (i = 0; i < ITERATIONS; i++)
5927 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5930 * (We measure the non-migrating [cached] cost on both
5931 * cpu1 and cpu2, to handle CPUs with different speeds)
5935 measure_one(cache, size, cpu1, cpu1);
5936 for (i = 0; i < ITERATIONS; i++)
5937 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5939 measure_one(cache, size, cpu2, cpu2);
5940 for (i = 0; i < ITERATIONS; i++)
5941 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5944 * Get the per-iteration migration cost:
5946 do_div(cost1, 2*ITERATIONS);
5947 do_div(cost2, 2*ITERATIONS);
5949 return cost1 - cost2;
5952 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5954 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5955 unsigned int max_size, size, size_found = 0;
5956 long long cost = 0, prev_cost;
5960 * Search from max_cache_size*5 down to 64K - the real relevant
5961 * cachesize has to lie somewhere inbetween.
5963 if (max_cache_size) {
5964 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5965 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5968 * Since we have no estimation about the relevant
5971 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5972 size = MIN_CACHE_SIZE;
5975 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5976 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5981 * Allocate the working set:
5983 cache = vmalloc(max_size);
5985 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5986 return 1000000; /* return 1 msec on very small boxen */
5989 while (size <= max_size) {
5991 cost = measure_cost(cpu1, cpu2, cache, size);
5997 if (max_cost < cost) {
6003 * Calculate average fluctuation, we use this to prevent
6004 * noise from triggering an early break out of the loop:
6006 fluct = abs(cost - prev_cost);
6007 avg_fluct = (avg_fluct + fluct)/2;
6009 if (migration_debug)
6010 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
6012 (long)cost / 1000000,
6013 ((long)cost / 100000) % 10,
6014 (long)max_cost / 1000000,
6015 ((long)max_cost / 100000) % 10,
6016 domain_distance(cpu1, cpu2),
6020 * If we iterated at least 20% past the previous maximum,
6021 * and the cost has dropped by more than 20% already,
6022 * (taking fluctuations into account) then we assume to
6023 * have found the maximum and break out of the loop early:
6025 if (size_found && (size*100 > size_found*SIZE_THRESH))
6026 if (cost+avg_fluct <= 0 ||
6027 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6029 if (migration_debug)
6030 printk("-> found max.\n");
6034 * Increase the cachesize in 10% steps:
6036 size = size * 10 / 9;
6039 if (migration_debug)
6040 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6041 cpu1, cpu2, size_found, max_cost);
6046 * A task is considered 'cache cold' if at least 2 times
6047 * the worst-case cost of migration has passed.
6049 * (this limit is only listened to if the load-balancing
6050 * situation is 'nice' - if there is a large imbalance we
6051 * ignore it for the sake of CPU utilization and
6052 * processing fairness.)
6054 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6057 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6059 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6060 unsigned long j0, j1, distance, max_distance = 0;
6061 struct sched_domain *sd;
6066 * First pass - calculate the cacheflush times:
6068 for_each_cpu_mask(cpu1, *cpu_map) {
6069 for_each_cpu_mask(cpu2, *cpu_map) {
6072 distance = domain_distance(cpu1, cpu2);
6073 max_distance = max(max_distance, distance);
6075 * No result cached yet?
6077 if (migration_cost[distance] == -1LL)
6078 migration_cost[distance] =
6079 measure_migration_cost(cpu1, cpu2);
6083 * Second pass - update the sched domain hierarchy with
6084 * the new cache-hot-time estimations:
6086 for_each_cpu_mask(cpu, *cpu_map) {
6088 for_each_domain(cpu, sd) {
6089 sd->cache_hot_time = migration_cost[distance];
6096 if (migration_debug)
6097 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6105 if (system_state == SYSTEM_BOOTING) {
6106 if (num_online_cpus() > 1) {
6107 printk("migration_cost=");
6108 for (distance = 0; distance <= max_distance; distance++) {
6111 printk("%ld", (long)migration_cost[distance] / 1000);
6117 if (migration_debug)
6118 printk("migration: %ld seconds\n", (j1-j0)/HZ);
6121 * Move back to the original CPU. NUMA-Q gets confused
6122 * if we migrate to another quad during bootup.
6124 if (raw_smp_processor_id() != orig_cpu) {
6125 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6126 saved_mask = current->cpus_allowed;
6128 set_cpus_allowed(current, mask);
6129 set_cpus_allowed(current, saved_mask);
6136 * find_next_best_node - find the next node to include in a sched_domain
6137 * @node: node whose sched_domain we're building
6138 * @used_nodes: nodes already in the sched_domain
6140 * Find the next node to include in a given scheduling domain. Simply
6141 * finds the closest node not already in the @used_nodes map.
6143 * Should use nodemask_t.
6145 static int find_next_best_node(int node, unsigned long *used_nodes)
6147 int i, n, val, min_val, best_node = 0;
6151 for (i = 0; i < MAX_NUMNODES; i++) {
6152 /* Start at @node */
6153 n = (node + i) % MAX_NUMNODES;
6155 if (!nr_cpus_node(n))
6158 /* Skip already used nodes */
6159 if (test_bit(n, used_nodes))
6162 /* Simple min distance search */
6163 val = node_distance(node, n);
6165 if (val < min_val) {
6171 set_bit(best_node, used_nodes);
6176 * sched_domain_node_span - get a cpumask for a node's sched_domain
6177 * @node: node whose cpumask we're constructing
6178 * @size: number of nodes to include in this span
6180 * Given a node, construct a good cpumask for its sched_domain to span. It
6181 * should be one that prevents unnecessary balancing, but also spreads tasks
6184 static cpumask_t sched_domain_node_span(int node)
6186 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6187 cpumask_t span, nodemask;
6191 bitmap_zero(used_nodes, MAX_NUMNODES);
6193 nodemask = node_to_cpumask(node);
6194 cpus_or(span, span, nodemask);
6195 set_bit(node, used_nodes);
6197 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6198 int next_node = find_next_best_node(node, used_nodes);
6200 nodemask = node_to_cpumask(next_node);
6201 cpus_or(span, span, nodemask);
6208 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6211 * SMT sched-domains:
6213 #ifdef CONFIG_SCHED_SMT
6214 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6215 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6217 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6218 struct sched_group **sg)
6221 *sg = &per_cpu(sched_group_cpus, cpu);
6227 * multi-core sched-domains:
6229 #ifdef CONFIG_SCHED_MC
6230 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6231 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6234 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6235 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6236 struct sched_group **sg)
6239 cpumask_t mask = cpu_sibling_map[cpu];
6240 cpus_and(mask, mask, *cpu_map);
6241 group = first_cpu(mask);
6243 *sg = &per_cpu(sched_group_core, group);
6246 #elif defined(CONFIG_SCHED_MC)
6247 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6248 struct sched_group **sg)
6251 *sg = &per_cpu(sched_group_core, cpu);
6256 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6257 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6259 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6260 struct sched_group **sg)
6263 #ifdef CONFIG_SCHED_MC
6264 cpumask_t mask = cpu_coregroup_map(cpu);
6265 cpus_and(mask, mask, *cpu_map);
6266 group = first_cpu(mask);
6267 #elif defined(CONFIG_SCHED_SMT)
6268 cpumask_t mask = cpu_sibling_map[cpu];
6269 cpus_and(mask, mask, *cpu_map);
6270 group = first_cpu(mask);
6275 *sg = &per_cpu(sched_group_phys, group);
6281 * The init_sched_build_groups can't handle what we want to do with node
6282 * groups, so roll our own. Now each node has its own list of groups which
6283 * gets dynamically allocated.
6285 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6286 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6288 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6289 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6291 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6292 struct sched_group **sg)
6294 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6297 cpus_and(nodemask, nodemask, *cpu_map);
6298 group = first_cpu(nodemask);
6301 *sg = &per_cpu(sched_group_allnodes, group);
6305 static void init_numa_sched_groups_power(struct sched_group *group_head)
6307 struct sched_group *sg = group_head;
6313 for_each_cpu_mask(j, sg->cpumask) {
6314 struct sched_domain *sd;
6316 sd = &per_cpu(phys_domains, j);
6317 if (j != first_cpu(sd->groups->cpumask)) {
6319 * Only add "power" once for each
6325 sg->cpu_power += sd->groups->cpu_power;
6328 if (sg != group_head)
6334 /* Free memory allocated for various sched_group structures */
6335 static void free_sched_groups(const cpumask_t *cpu_map)
6339 for_each_cpu_mask(cpu, *cpu_map) {
6340 struct sched_group **sched_group_nodes
6341 = sched_group_nodes_bycpu[cpu];
6343 if (!sched_group_nodes)
6346 for (i = 0; i < MAX_NUMNODES; i++) {
6347 cpumask_t nodemask = node_to_cpumask(i);
6348 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6350 cpus_and(nodemask, nodemask, *cpu_map);
6351 if (cpus_empty(nodemask))
6361 if (oldsg != sched_group_nodes[i])
6364 kfree(sched_group_nodes);
6365 sched_group_nodes_bycpu[cpu] = NULL;
6369 static void free_sched_groups(const cpumask_t *cpu_map)
6375 * Initialize sched groups cpu_power.
6377 * cpu_power indicates the capacity of sched group, which is used while
6378 * distributing the load between different sched groups in a sched domain.
6379 * Typically cpu_power for all the groups in a sched domain will be same unless
6380 * there are asymmetries in the topology. If there are asymmetries, group
6381 * having more cpu_power will pickup more load compared to the group having
6384 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6385 * the maximum number of tasks a group can handle in the presence of other idle
6386 * or lightly loaded groups in the same sched domain.
6388 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6390 struct sched_domain *child;
6391 struct sched_group *group;
6393 WARN_ON(!sd || !sd->groups);
6395 if (cpu != first_cpu(sd->groups->cpumask))
6401 * For perf policy, if the groups in child domain share resources
6402 * (for example cores sharing some portions of the cache hierarchy
6403 * or SMT), then set this domain groups cpu_power such that each group
6404 * can handle only one task, when there are other idle groups in the
6405 * same sched domain.
6407 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6409 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6410 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6414 sd->groups->cpu_power = 0;
6417 * add cpu_power of each child group to this groups cpu_power
6419 group = child->groups;
6421 sd->groups->cpu_power += group->cpu_power;
6422 group = group->next;
6423 } while (group != child->groups);
6427 * Build sched domains for a given set of cpus and attach the sched domains
6428 * to the individual cpus
6430 static int build_sched_domains(const cpumask_t *cpu_map)
6433 struct sched_domain *sd;
6435 struct sched_group **sched_group_nodes = NULL;
6436 int sd_allnodes = 0;
6439 * Allocate the per-node list of sched groups
6441 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6443 if (!sched_group_nodes) {
6444 printk(KERN_WARNING "Can not alloc sched group node list\n");
6447 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6451 * Set up domains for cpus specified by the cpu_map.
6453 for_each_cpu_mask(i, *cpu_map) {
6454 struct sched_domain *sd = NULL, *p;
6455 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6457 cpus_and(nodemask, nodemask, *cpu_map);
6460 if (cpus_weight(*cpu_map)
6461 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6462 sd = &per_cpu(allnodes_domains, i);
6463 *sd = SD_ALLNODES_INIT;
6464 sd->span = *cpu_map;
6465 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6471 sd = &per_cpu(node_domains, i);
6473 sd->span = sched_domain_node_span(cpu_to_node(i));
6477 cpus_and(sd->span, sd->span, *cpu_map);
6481 sd = &per_cpu(phys_domains, i);
6483 sd->span = nodemask;
6487 cpu_to_phys_group(i, cpu_map, &sd->groups);
6489 #ifdef CONFIG_SCHED_MC
6491 sd = &per_cpu(core_domains, i);
6493 sd->span = cpu_coregroup_map(i);
6494 cpus_and(sd->span, sd->span, *cpu_map);
6497 cpu_to_core_group(i, cpu_map, &sd->groups);
6500 #ifdef CONFIG_SCHED_SMT
6502 sd = &per_cpu(cpu_domains, i);
6503 *sd = SD_SIBLING_INIT;
6504 sd->span = cpu_sibling_map[i];
6505 cpus_and(sd->span, sd->span, *cpu_map);
6508 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6512 #ifdef CONFIG_SCHED_SMT
6513 /* Set up CPU (sibling) groups */
6514 for_each_cpu_mask(i, *cpu_map) {
6515 cpumask_t this_sibling_map = cpu_sibling_map[i];
6516 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6517 if (i != first_cpu(this_sibling_map))
6520 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6524 #ifdef CONFIG_SCHED_MC
6525 /* Set up multi-core groups */
6526 for_each_cpu_mask(i, *cpu_map) {
6527 cpumask_t this_core_map = cpu_coregroup_map(i);
6528 cpus_and(this_core_map, this_core_map, *cpu_map);
6529 if (i != first_cpu(this_core_map))
6531 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6536 /* Set up physical groups */
6537 for (i = 0; i < MAX_NUMNODES; i++) {
6538 cpumask_t nodemask = node_to_cpumask(i);
6540 cpus_and(nodemask, nodemask, *cpu_map);
6541 if (cpus_empty(nodemask))
6544 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6548 /* Set up node groups */
6550 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6552 for (i = 0; i < MAX_NUMNODES; i++) {
6553 /* Set up node groups */
6554 struct sched_group *sg, *prev;
6555 cpumask_t nodemask = node_to_cpumask(i);
6556 cpumask_t domainspan;
6557 cpumask_t covered = CPU_MASK_NONE;
6560 cpus_and(nodemask, nodemask, *cpu_map);
6561 if (cpus_empty(nodemask)) {
6562 sched_group_nodes[i] = NULL;
6566 domainspan = sched_domain_node_span(i);
6567 cpus_and(domainspan, domainspan, *cpu_map);
6569 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6571 printk(KERN_WARNING "Can not alloc domain group for "
6575 sched_group_nodes[i] = sg;
6576 for_each_cpu_mask(j, nodemask) {
6577 struct sched_domain *sd;
6578 sd = &per_cpu(node_domains, j);
6582 sg->cpumask = nodemask;
6584 cpus_or(covered, covered, nodemask);
6587 for (j = 0; j < MAX_NUMNODES; j++) {
6588 cpumask_t tmp, notcovered;
6589 int n = (i + j) % MAX_NUMNODES;
6591 cpus_complement(notcovered, covered);
6592 cpus_and(tmp, notcovered, *cpu_map);
6593 cpus_and(tmp, tmp, domainspan);
6594 if (cpus_empty(tmp))
6597 nodemask = node_to_cpumask(n);
6598 cpus_and(tmp, tmp, nodemask);
6599 if (cpus_empty(tmp))
6602 sg = kmalloc_node(sizeof(struct sched_group),
6606 "Can not alloc domain group for node %d\n", j);
6611 sg->next = prev->next;
6612 cpus_or(covered, covered, tmp);
6619 /* Calculate CPU power for physical packages and nodes */
6620 #ifdef CONFIG_SCHED_SMT
6621 for_each_cpu_mask(i, *cpu_map) {
6622 sd = &per_cpu(cpu_domains, i);
6623 init_sched_groups_power(i, sd);
6626 #ifdef CONFIG_SCHED_MC
6627 for_each_cpu_mask(i, *cpu_map) {
6628 sd = &per_cpu(core_domains, i);
6629 init_sched_groups_power(i, sd);
6633 for_each_cpu_mask(i, *cpu_map) {
6634 sd = &per_cpu(phys_domains, i);
6635 init_sched_groups_power(i, sd);
6639 for (i = 0; i < MAX_NUMNODES; i++)
6640 init_numa_sched_groups_power(sched_group_nodes[i]);
6643 struct sched_group *sg;
6645 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6646 init_numa_sched_groups_power(sg);
6650 /* Attach the domains */
6651 for_each_cpu_mask(i, *cpu_map) {
6652 struct sched_domain *sd;
6653 #ifdef CONFIG_SCHED_SMT
6654 sd = &per_cpu(cpu_domains, i);
6655 #elif defined(CONFIG_SCHED_MC)
6656 sd = &per_cpu(core_domains, i);
6658 sd = &per_cpu(phys_domains, i);
6660 cpu_attach_domain(sd, i);
6663 * Tune cache-hot values:
6665 calibrate_migration_costs(cpu_map);
6671 free_sched_groups(cpu_map);
6676 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6678 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6680 cpumask_t cpu_default_map;
6684 * Setup mask for cpus without special case scheduling requirements.
6685 * For now this just excludes isolated cpus, but could be used to
6686 * exclude other special cases in the future.
6688 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6690 err = build_sched_domains(&cpu_default_map);
6695 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6697 free_sched_groups(cpu_map);
6701 * Detach sched domains from a group of cpus specified in cpu_map
6702 * These cpus will now be attached to the NULL domain
6704 static void detach_destroy_domains(const cpumask_t *cpu_map)
6708 for_each_cpu_mask(i, *cpu_map)
6709 cpu_attach_domain(NULL, i);
6710 synchronize_sched();
6711 arch_destroy_sched_domains(cpu_map);
6715 * Partition sched domains as specified by the cpumasks below.
6716 * This attaches all cpus from the cpumasks to the NULL domain,
6717 * waits for a RCU quiescent period, recalculates sched
6718 * domain information and then attaches them back to the
6719 * correct sched domains
6720 * Call with hotplug lock held
6722 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6724 cpumask_t change_map;
6727 cpus_and(*partition1, *partition1, cpu_online_map);
6728 cpus_and(*partition2, *partition2, cpu_online_map);
6729 cpus_or(change_map, *partition1, *partition2);
6731 /* Detach sched domains from all of the affected cpus */
6732 detach_destroy_domains(&change_map);
6733 if (!cpus_empty(*partition1))
6734 err = build_sched_domains(partition1);
6735 if (!err && !cpus_empty(*partition2))
6736 err = build_sched_domains(partition2);
6741 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6742 int arch_reinit_sched_domains(void)
6747 detach_destroy_domains(&cpu_online_map);
6748 err = arch_init_sched_domains(&cpu_online_map);
6749 unlock_cpu_hotplug();
6754 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6758 if (buf[0] != '0' && buf[0] != '1')
6762 sched_smt_power_savings = (buf[0] == '1');
6764 sched_mc_power_savings = (buf[0] == '1');
6766 ret = arch_reinit_sched_domains();
6768 return ret ? ret : count;
6771 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6775 #ifdef CONFIG_SCHED_SMT
6777 err = sysfs_create_file(&cls->kset.kobj,
6778 &attr_sched_smt_power_savings.attr);
6780 #ifdef CONFIG_SCHED_MC
6781 if (!err && mc_capable())
6782 err = sysfs_create_file(&cls->kset.kobj,
6783 &attr_sched_mc_power_savings.attr);
6789 #ifdef CONFIG_SCHED_MC
6790 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6792 return sprintf(page, "%u\n", sched_mc_power_savings);
6794 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6795 const char *buf, size_t count)
6797 return sched_power_savings_store(buf, count, 0);
6799 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6800 sched_mc_power_savings_store);
6803 #ifdef CONFIG_SCHED_SMT
6804 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6806 return sprintf(page, "%u\n", sched_smt_power_savings);
6808 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6809 const char *buf, size_t count)
6811 return sched_power_savings_store(buf, count, 1);
6813 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6814 sched_smt_power_savings_store);
6818 * Force a reinitialization of the sched domains hierarchy. The domains
6819 * and groups cannot be updated in place without racing with the balancing
6820 * code, so we temporarily attach all running cpus to the NULL domain
6821 * which will prevent rebalancing while the sched domains are recalculated.
6823 static int update_sched_domains(struct notifier_block *nfb,
6824 unsigned long action, void *hcpu)
6827 case CPU_UP_PREPARE:
6828 case CPU_DOWN_PREPARE:
6829 detach_destroy_domains(&cpu_online_map);
6832 case CPU_UP_CANCELED:
6833 case CPU_DOWN_FAILED:
6837 * Fall through and re-initialise the domains.
6844 /* The hotplug lock is already held by cpu_up/cpu_down */
6845 arch_init_sched_domains(&cpu_online_map);
6850 void __init sched_init_smp(void)
6852 cpumask_t non_isolated_cpus;
6855 arch_init_sched_domains(&cpu_online_map);
6856 cpus_andnot(non_isolated_cpus, cpu_online_map, cpu_isolated_map);
6857 if (cpus_empty(non_isolated_cpus))
6858 cpu_set(smp_processor_id(), non_isolated_cpus);
6859 unlock_cpu_hotplug();
6860 /* XXX: Theoretical race here - CPU may be hotplugged now */
6861 hotcpu_notifier(update_sched_domains, 0);
6863 /* Move init over to a non-isolated CPU */
6864 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6868 void __init sched_init_smp(void)
6871 #endif /* CONFIG_SMP */
6873 int in_sched_functions(unsigned long addr)
6875 /* Linker adds these: start and end of __sched functions */
6876 extern char __sched_text_start[], __sched_text_end[];
6878 return in_lock_functions(addr) ||
6879 (addr >= (unsigned long)__sched_text_start
6880 && addr < (unsigned long)__sched_text_end);
6883 void __init sched_init(void)
6887 for_each_possible_cpu(i) {
6888 struct prio_array *array;
6892 spin_lock_init(&rq->lock);
6893 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6895 rq->active = rq->arrays;
6896 rq->expired = rq->arrays + 1;
6897 rq->best_expired_prio = MAX_PRIO;
6901 for (j = 1; j < 3; j++)
6902 rq->cpu_load[j] = 0;
6903 rq->active_balance = 0;
6906 rq->migration_thread = NULL;
6907 INIT_LIST_HEAD(&rq->migration_queue);
6909 atomic_set(&rq->nr_iowait, 0);
6911 for (j = 0; j < 2; j++) {
6912 array = rq->arrays + j;
6913 for (k = 0; k < MAX_PRIO; k++) {
6914 INIT_LIST_HEAD(array->queue + k);
6915 __clear_bit(k, array->bitmap);
6917 // delimiter for bitsearch
6918 __set_bit(MAX_PRIO, array->bitmap);
6922 set_load_weight(&init_task);
6925 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6928 #ifdef CONFIG_RT_MUTEXES
6929 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6933 * The boot idle thread does lazy MMU switching as well:
6935 atomic_inc(&init_mm.mm_count);
6936 enter_lazy_tlb(&init_mm, current);
6939 * Make us the idle thread. Technically, schedule() should not be
6940 * called from this thread, however somewhere below it might be,
6941 * but because we are the idle thread, we just pick up running again
6942 * when this runqueue becomes "idle".
6944 init_idle(current, smp_processor_id());
6947 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6948 void __might_sleep(char *file, int line)
6951 static unsigned long prev_jiffy; /* ratelimiting */
6953 if ((in_atomic() || irqs_disabled()) &&
6954 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6955 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6957 prev_jiffy = jiffies;
6958 printk(KERN_ERR "BUG: sleeping function called from invalid"
6959 " context at %s:%d\n", file, line);
6960 printk("in_atomic():%d, irqs_disabled():%d\n",
6961 in_atomic(), irqs_disabled());
6962 debug_show_held_locks(current);
6967 EXPORT_SYMBOL(__might_sleep);
6970 #ifdef CONFIG_MAGIC_SYSRQ
6971 void normalize_rt_tasks(void)
6973 struct prio_array *array;
6974 struct task_struct *p;
6975 unsigned long flags;
6978 read_lock_irq(&tasklist_lock);
6979 for_each_process(p) {
6983 spin_lock_irqsave(&p->pi_lock, flags);
6984 rq = __task_rq_lock(p);
6988 deactivate_task(p, task_rq(p));
6989 __setscheduler(p, SCHED_NORMAL, 0);
6991 __activate_task(p, task_rq(p));
6992 resched_task(rq->curr);
6995 __task_rq_unlock(rq);
6996 spin_unlock_irqrestore(&p->pi_lock, flags);
6998 read_unlock_irq(&tasklist_lock);
7001 #endif /* CONFIG_MAGIC_SYSRQ */
7005 * These functions are only useful for the IA64 MCA handling.
7007 * They can only be called when the whole system has been
7008 * stopped - every CPU needs to be quiescent, and no scheduling
7009 * activity can take place. Using them for anything else would
7010 * be a serious bug, and as a result, they aren't even visible
7011 * under any other configuration.
7015 * curr_task - return the current task for a given cpu.
7016 * @cpu: the processor in question.
7018 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7020 struct task_struct *curr_task(int cpu)
7022 return cpu_curr(cpu);
7026 * set_curr_task - set the current task for a given cpu.
7027 * @cpu: the processor in question.
7028 * @p: the task pointer to set.
7030 * Description: This function must only be used when non-maskable interrupts
7031 * are serviced on a separate stack. It allows the architecture to switch the
7032 * notion of the current task on a cpu in a non-blocking manner. This function
7033 * must be called with all CPU's synchronized, and interrupts disabled, the
7034 * and caller must save the original value of the current task (see
7035 * curr_task() above) and restore that value before reenabling interrupts and
7036 * re-starting the system.
7038 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7040 void set_curr_task(int cpu, struct task_struct *p)