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 12
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,
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 long sum_nr_running, sum_weighted_load;
2283 local_group = cpu_isset(this_cpu, group->cpumask);
2285 /* Tally up the load of all CPUs in the group */
2286 sum_weighted_load = sum_nr_running = avg_load = 0;
2288 for_each_cpu_mask(i, group->cpumask) {
2291 if (!cpu_isset(i, *cpus))
2296 if (*sd_idle && !idle_cpu(i))
2299 /* Bias balancing toward cpus of our domain */
2301 load = target_load(i, load_idx);
2303 load = source_load(i, load_idx);
2306 sum_nr_running += rq->nr_running;
2307 sum_weighted_load += rq->raw_weighted_load;
2310 total_load += avg_load;
2311 total_pwr += group->cpu_power;
2313 /* Adjust by relative CPU power of the group */
2314 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2316 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2319 this_load = avg_load;
2321 this_nr_running = sum_nr_running;
2322 this_load_per_task = sum_weighted_load;
2323 } else if (avg_load > max_load &&
2324 sum_nr_running > group_capacity) {
2325 max_load = avg_load;
2327 busiest_nr_running = sum_nr_running;
2328 busiest_load_per_task = sum_weighted_load;
2331 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2333 * Busy processors will not participate in power savings
2336 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2340 * If the local group is idle or completely loaded
2341 * no need to do power savings balance at this domain
2343 if (local_group && (this_nr_running >= group_capacity ||
2345 power_savings_balance = 0;
2348 * If a group is already running at full capacity or idle,
2349 * don't include that group in power savings calculations
2351 if (!power_savings_balance || sum_nr_running >= group_capacity
2356 * Calculate the group which has the least non-idle load.
2357 * This is the group from where we need to pick up the load
2360 if ((sum_nr_running < min_nr_running) ||
2361 (sum_nr_running == min_nr_running &&
2362 first_cpu(group->cpumask) <
2363 first_cpu(group_min->cpumask))) {
2365 min_nr_running = sum_nr_running;
2366 min_load_per_task = sum_weighted_load /
2371 * Calculate the group which is almost near its
2372 * capacity but still has some space to pick up some load
2373 * from other group and save more power
2375 if (sum_nr_running <= group_capacity - 1) {
2376 if (sum_nr_running > leader_nr_running ||
2377 (sum_nr_running == leader_nr_running &&
2378 first_cpu(group->cpumask) >
2379 first_cpu(group_leader->cpumask))) {
2380 group_leader = group;
2381 leader_nr_running = sum_nr_running;
2386 group = group->next;
2387 } while (group != sd->groups);
2389 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2392 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2394 if (this_load >= avg_load ||
2395 100*max_load <= sd->imbalance_pct*this_load)
2398 busiest_load_per_task /= busiest_nr_running;
2400 * We're trying to get all the cpus to the average_load, so we don't
2401 * want to push ourselves above the average load, nor do we wish to
2402 * reduce the max loaded cpu below the average load, as either of these
2403 * actions would just result in more rebalancing later, and ping-pong
2404 * tasks around. Thus we look for the minimum possible imbalance.
2405 * Negative imbalances (*we* are more loaded than anyone else) will
2406 * be counted as no imbalance for these purposes -- we can't fix that
2407 * by pulling tasks to us. Be careful of negative numbers as they'll
2408 * appear as very large values with unsigned longs.
2410 if (max_load <= busiest_load_per_task)
2414 * In the presence of smp nice balancing, certain scenarios can have
2415 * max load less than avg load(as we skip the groups at or below
2416 * its cpu_power, while calculating max_load..)
2418 if (max_load < avg_load) {
2420 goto small_imbalance;
2423 /* Don't want to pull so many tasks that a group would go idle */
2424 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2426 /* How much load to actually move to equalise the imbalance */
2427 *imbalance = min(max_pull * busiest->cpu_power,
2428 (avg_load - this_load) * this->cpu_power)
2432 * if *imbalance is less than the average load per runnable task
2433 * there is no gaurantee that any tasks will be moved so we'll have
2434 * a think about bumping its value to force at least one task to be
2437 if (*imbalance < busiest_load_per_task) {
2438 unsigned long tmp, pwr_now, pwr_move;
2442 pwr_move = pwr_now = 0;
2444 if (this_nr_running) {
2445 this_load_per_task /= this_nr_running;
2446 if (busiest_load_per_task > this_load_per_task)
2449 this_load_per_task = SCHED_LOAD_SCALE;
2451 if (max_load - this_load >= busiest_load_per_task * imbn) {
2452 *imbalance = busiest_load_per_task;
2457 * OK, we don't have enough imbalance to justify moving tasks,
2458 * however we may be able to increase total CPU power used by
2462 pwr_now += busiest->cpu_power *
2463 min(busiest_load_per_task, max_load);
2464 pwr_now += this->cpu_power *
2465 min(this_load_per_task, this_load);
2466 pwr_now /= SCHED_LOAD_SCALE;
2468 /* Amount of load we'd subtract */
2469 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2471 pwr_move += busiest->cpu_power *
2472 min(busiest_load_per_task, max_load - tmp);
2474 /* Amount of load we'd add */
2475 if (max_load*busiest->cpu_power <
2476 busiest_load_per_task*SCHED_LOAD_SCALE)
2477 tmp = max_load*busiest->cpu_power/this->cpu_power;
2479 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2480 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2481 pwr_move /= SCHED_LOAD_SCALE;
2483 /* Move if we gain throughput */
2484 if (pwr_move <= pwr_now)
2487 *imbalance = busiest_load_per_task;
2493 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2494 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2497 if (this == group_leader && group_leader != group_min) {
2498 *imbalance = min_load_per_task;
2508 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2511 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2512 unsigned long imbalance, cpumask_t *cpus)
2514 struct rq *busiest = NULL, *rq;
2515 unsigned long max_load = 0;
2518 for_each_cpu_mask(i, group->cpumask) {
2520 if (!cpu_isset(i, *cpus))
2525 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2528 if (rq->raw_weighted_load > max_load) {
2529 max_load = rq->raw_weighted_load;
2538 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2539 * so long as it is large enough.
2541 #define MAX_PINNED_INTERVAL 512
2543 static inline unsigned long minus_1_or_zero(unsigned long n)
2545 return n > 0 ? n - 1 : 0;
2549 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2550 * tasks if there is an imbalance.
2552 static int load_balance(int this_cpu, struct rq *this_rq,
2553 struct sched_domain *sd, enum idle_type idle)
2555 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2556 struct sched_group *group;
2557 unsigned long imbalance;
2559 cpumask_t cpus = CPU_MASK_ALL;
2560 unsigned long flags;
2563 * When power savings policy is enabled for the parent domain, idle
2564 * sibling can pick up load irrespective of busy siblings. In this case,
2565 * let the state of idle sibling percolate up as IDLE, instead of
2566 * portraying it as NOT_IDLE.
2568 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2569 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2572 schedstat_inc(sd, lb_cnt[idle]);
2575 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2578 schedstat_inc(sd, lb_nobusyg[idle]);
2582 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2584 schedstat_inc(sd, lb_nobusyq[idle]);
2588 BUG_ON(busiest == this_rq);
2590 schedstat_add(sd, lb_imbalance[idle], imbalance);
2593 if (busiest->nr_running > 1) {
2595 * Attempt to move tasks. If find_busiest_group has found
2596 * an imbalance but busiest->nr_running <= 1, the group is
2597 * still unbalanced. nr_moved simply stays zero, so it is
2598 * correctly treated as an imbalance.
2600 local_irq_save(flags);
2601 double_rq_lock(this_rq, busiest);
2602 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2603 minus_1_or_zero(busiest->nr_running),
2604 imbalance, sd, idle, &all_pinned);
2605 double_rq_unlock(this_rq, busiest);
2606 local_irq_restore(flags);
2608 /* All tasks on this runqueue were pinned by CPU affinity */
2609 if (unlikely(all_pinned)) {
2610 cpu_clear(cpu_of(busiest), cpus);
2611 if (!cpus_empty(cpus))
2618 schedstat_inc(sd, lb_failed[idle]);
2619 sd->nr_balance_failed++;
2621 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2623 spin_lock_irqsave(&busiest->lock, flags);
2625 /* don't kick the migration_thread, if the curr
2626 * task on busiest cpu can't be moved to this_cpu
2628 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2629 spin_unlock_irqrestore(&busiest->lock, flags);
2631 goto out_one_pinned;
2634 if (!busiest->active_balance) {
2635 busiest->active_balance = 1;
2636 busiest->push_cpu = this_cpu;
2639 spin_unlock_irqrestore(&busiest->lock, flags);
2641 wake_up_process(busiest->migration_thread);
2644 * We've kicked active balancing, reset the failure
2647 sd->nr_balance_failed = sd->cache_nice_tries+1;
2650 sd->nr_balance_failed = 0;
2652 if (likely(!active_balance)) {
2653 /* We were unbalanced, so reset the balancing interval */
2654 sd->balance_interval = sd->min_interval;
2657 * If we've begun active balancing, start to back off. This
2658 * case may not be covered by the all_pinned logic if there
2659 * is only 1 task on the busy runqueue (because we don't call
2662 if (sd->balance_interval < sd->max_interval)
2663 sd->balance_interval *= 2;
2666 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2667 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2672 schedstat_inc(sd, lb_balanced[idle]);
2674 sd->nr_balance_failed = 0;
2677 /* tune up the balancing interval */
2678 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2679 (sd->balance_interval < sd->max_interval))
2680 sd->balance_interval *= 2;
2682 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2683 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2689 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2690 * tasks if there is an imbalance.
2692 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2693 * this_rq is locked.
2696 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2698 struct sched_group *group;
2699 struct rq *busiest = NULL;
2700 unsigned long imbalance;
2703 cpumask_t cpus = CPU_MASK_ALL;
2706 * When power savings policy is enabled for the parent domain, idle
2707 * sibling can pick up load irrespective of busy siblings. In this case,
2708 * let the state of idle sibling percolate up as IDLE, instead of
2709 * portraying it as NOT_IDLE.
2711 if (sd->flags & SD_SHARE_CPUPOWER &&
2712 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2715 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2717 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2720 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2724 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2727 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2731 BUG_ON(busiest == this_rq);
2733 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2736 if (busiest->nr_running > 1) {
2737 /* Attempt to move tasks */
2738 double_lock_balance(this_rq, busiest);
2739 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2740 minus_1_or_zero(busiest->nr_running),
2741 imbalance, sd, NEWLY_IDLE, NULL);
2742 spin_unlock(&busiest->lock);
2745 cpu_clear(cpu_of(busiest), cpus);
2746 if (!cpus_empty(cpus))
2752 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2753 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2754 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2757 sd->nr_balance_failed = 0;
2762 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2763 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2764 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2766 sd->nr_balance_failed = 0;
2772 * idle_balance is called by schedule() if this_cpu is about to become
2773 * idle. Attempts to pull tasks from other CPUs.
2775 static void idle_balance(int this_cpu, struct rq *this_rq)
2777 struct sched_domain *sd;
2778 int pulled_task = 0;
2779 unsigned long next_balance = jiffies + 60 * HZ;
2781 for_each_domain(this_cpu, sd) {
2782 if (sd->flags & SD_BALANCE_NEWIDLE) {
2783 /* If we've pulled tasks over stop searching: */
2784 pulled_task = load_balance_newidle(this_cpu,
2786 if (time_after(next_balance,
2787 sd->last_balance + sd->balance_interval))
2788 next_balance = sd->last_balance
2789 + sd->balance_interval;
2796 * We are going idle. next_balance may be set based on
2797 * a busy processor. So reset next_balance.
2799 this_rq->next_balance = next_balance;
2803 * active_load_balance is run by migration threads. It pushes running tasks
2804 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2805 * running on each physical CPU where possible, and avoids physical /
2806 * logical imbalances.
2808 * Called with busiest_rq locked.
2810 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2812 int target_cpu = busiest_rq->push_cpu;
2813 struct sched_domain *sd;
2814 struct rq *target_rq;
2816 /* Is there any task to move? */
2817 if (busiest_rq->nr_running <= 1)
2820 target_rq = cpu_rq(target_cpu);
2823 * This condition is "impossible", if it occurs
2824 * we need to fix it. Originally reported by
2825 * Bjorn Helgaas on a 128-cpu setup.
2827 BUG_ON(busiest_rq == target_rq);
2829 /* move a task from busiest_rq to target_rq */
2830 double_lock_balance(busiest_rq, target_rq);
2832 /* Search for an sd spanning us and the target CPU. */
2833 for_each_domain(target_cpu, sd) {
2834 if ((sd->flags & SD_LOAD_BALANCE) &&
2835 cpu_isset(busiest_cpu, sd->span))
2840 schedstat_inc(sd, alb_cnt);
2842 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2843 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2845 schedstat_inc(sd, alb_pushed);
2847 schedstat_inc(sd, alb_failed);
2849 spin_unlock(&target_rq->lock);
2852 static void update_load(struct rq *this_rq)
2854 unsigned long this_load;
2857 this_load = this_rq->raw_weighted_load;
2859 /* Update our load: */
2860 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2861 unsigned long old_load, new_load;
2863 old_load = this_rq->cpu_load[i];
2864 new_load = this_load;
2866 * Round up the averaging division if load is increasing. This
2867 * prevents us from getting stuck on 9 if the load is 10, for
2870 if (new_load > old_load)
2871 new_load += scale-1;
2872 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2877 * run_rebalance_domains is triggered when needed from the scheduler tick.
2879 * It checks each scheduling domain to see if it is due to be balanced,
2880 * and initiates a balancing operation if so.
2882 * Balancing parameters are set up in arch_init_sched_domains.
2884 static DEFINE_SPINLOCK(balancing);
2886 static void run_rebalance_domains(struct softirq_action *h)
2888 int this_cpu = smp_processor_id();
2889 struct rq *this_rq = cpu_rq(this_cpu);
2890 unsigned long interval;
2891 struct sched_domain *sd;
2893 * We are idle if there are no processes running. This
2894 * is valid even if we are the idle process (SMT).
2896 enum idle_type idle = !this_rq->nr_running ?
2897 SCHED_IDLE : NOT_IDLE;
2898 /* Earliest time when we have to call run_rebalance_domains again */
2899 unsigned long next_balance = jiffies + 60*HZ;
2901 for_each_domain(this_cpu, sd) {
2902 if (!(sd->flags & SD_LOAD_BALANCE))
2905 interval = sd->balance_interval;
2906 if (idle != SCHED_IDLE)
2907 interval *= sd->busy_factor;
2909 /* scale ms to jiffies */
2910 interval = msecs_to_jiffies(interval);
2911 if (unlikely(!interval))
2914 if (sd->flags & SD_SERIALIZE) {
2915 if (!spin_trylock(&balancing))
2919 if (time_after_eq(jiffies, sd->last_balance + interval)) {
2920 if (load_balance(this_cpu, this_rq, sd, idle)) {
2922 * We've pulled tasks over so either we're no
2923 * longer idle, or one of our SMT siblings is
2928 sd->last_balance = jiffies;
2930 if (sd->flags & SD_SERIALIZE)
2931 spin_unlock(&balancing);
2933 if (time_after(next_balance, sd->last_balance + interval))
2934 next_balance = sd->last_balance + interval;
2936 this_rq->next_balance = next_balance;
2940 * on UP we do not need to balance between CPUs:
2942 static inline void idle_balance(int cpu, struct rq *rq)
2947 static inline void wake_priority_sleeper(struct rq *rq)
2949 #ifdef CONFIG_SCHED_SMT
2950 if (!rq->nr_running)
2953 spin_lock(&rq->lock);
2955 * If an SMT sibling task has been put to sleep for priority
2956 * reasons reschedule the idle task to see if it can now run.
2959 resched_task(rq->idle);
2960 spin_unlock(&rq->lock);
2964 DEFINE_PER_CPU(struct kernel_stat, kstat);
2966 EXPORT_PER_CPU_SYMBOL(kstat);
2969 * This is called on clock ticks and on context switches.
2970 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2973 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
2975 p->sched_time += now - p->last_ran;
2976 p->last_ran = rq->most_recent_timestamp = now;
2980 * Return current->sched_time plus any more ns on the sched_clock
2981 * that have not yet been banked.
2983 unsigned long long current_sched_time(const struct task_struct *p)
2985 unsigned long long ns;
2986 unsigned long flags;
2988 local_irq_save(flags);
2989 ns = p->sched_time + sched_clock() - p->last_ran;
2990 local_irq_restore(flags);
2996 * We place interactive tasks back into the active array, if possible.
2998 * To guarantee that this does not starve expired tasks we ignore the
2999 * interactivity of a task if the first expired task had to wait more
3000 * than a 'reasonable' amount of time. This deadline timeout is
3001 * load-dependent, as the frequency of array switched decreases with
3002 * increasing number of running tasks. We also ignore the interactivity
3003 * if a better static_prio task has expired:
3005 static inline int expired_starving(struct rq *rq)
3007 if (rq->curr->static_prio > rq->best_expired_prio)
3009 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3011 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3017 * Account user cpu time to a process.
3018 * @p: the process that the cpu time gets accounted to
3019 * @hardirq_offset: the offset to subtract from hardirq_count()
3020 * @cputime: the cpu time spent in user space since the last update
3022 void account_user_time(struct task_struct *p, cputime_t cputime)
3024 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3027 p->utime = cputime_add(p->utime, cputime);
3029 /* Add user time to cpustat. */
3030 tmp = cputime_to_cputime64(cputime);
3031 if (TASK_NICE(p) > 0)
3032 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3034 cpustat->user = cputime64_add(cpustat->user, tmp);
3038 * Account system cpu time to a process.
3039 * @p: the process that the cpu time gets accounted to
3040 * @hardirq_offset: the offset to subtract from hardirq_count()
3041 * @cputime: the cpu time spent in kernel space since the last update
3043 void account_system_time(struct task_struct *p, int hardirq_offset,
3046 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3047 struct rq *rq = this_rq();
3050 p->stime = cputime_add(p->stime, cputime);
3052 /* Add system time to cpustat. */
3053 tmp = cputime_to_cputime64(cputime);
3054 if (hardirq_count() - hardirq_offset)
3055 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3056 else if (softirq_count())
3057 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3058 else if (p != rq->idle)
3059 cpustat->system = cputime64_add(cpustat->system, tmp);
3060 else if (atomic_read(&rq->nr_iowait) > 0)
3061 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3063 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3064 /* Account for system time used */
3065 acct_update_integrals(p);
3069 * Account for involuntary wait time.
3070 * @p: the process from which the cpu time has been stolen
3071 * @steal: the cpu time spent in involuntary wait
3073 void account_steal_time(struct task_struct *p, cputime_t steal)
3075 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3076 cputime64_t tmp = cputime_to_cputime64(steal);
3077 struct rq *rq = this_rq();
3079 if (p == rq->idle) {
3080 p->stime = cputime_add(p->stime, steal);
3081 if (atomic_read(&rq->nr_iowait) > 0)
3082 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3084 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3086 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3089 static void task_running_tick(struct rq *rq, struct task_struct *p)
3091 if (p->array != rq->active) {
3092 /* Task has expired but was not scheduled yet */
3093 set_tsk_need_resched(p);
3096 spin_lock(&rq->lock);
3098 * The task was running during this tick - update the
3099 * time slice counter. Note: we do not update a thread's
3100 * priority until it either goes to sleep or uses up its
3101 * timeslice. This makes it possible for interactive tasks
3102 * to use up their timeslices at their highest priority levels.
3106 * RR tasks need a special form of timeslice management.
3107 * FIFO tasks have no timeslices.
3109 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3110 p->time_slice = task_timeslice(p);
3111 p->first_time_slice = 0;
3112 set_tsk_need_resched(p);
3114 /* put it at the end of the queue: */
3115 requeue_task(p, rq->active);
3119 if (!--p->time_slice) {
3120 dequeue_task(p, rq->active);
3121 set_tsk_need_resched(p);
3122 p->prio = effective_prio(p);
3123 p->time_slice = task_timeslice(p);
3124 p->first_time_slice = 0;
3126 if (!rq->expired_timestamp)
3127 rq->expired_timestamp = jiffies;
3128 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3129 enqueue_task(p, rq->expired);
3130 if (p->static_prio < rq->best_expired_prio)
3131 rq->best_expired_prio = p->static_prio;
3133 enqueue_task(p, rq->active);
3136 * Prevent a too long timeslice allowing a task to monopolize
3137 * the CPU. We do this by splitting up the timeslice into
3140 * Note: this does not mean the task's timeslices expire or
3141 * get lost in any way, they just might be preempted by
3142 * another task of equal priority. (one with higher
3143 * priority would have preempted this task already.) We
3144 * requeue this task to the end of the list on this priority
3145 * level, which is in essence a round-robin of tasks with
3148 * This only applies to tasks in the interactive
3149 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3151 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3152 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3153 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3154 (p->array == rq->active)) {
3156 requeue_task(p, rq->active);
3157 set_tsk_need_resched(p);
3161 spin_unlock(&rq->lock);
3165 * This function gets called by the timer code, with HZ frequency.
3166 * We call it with interrupts disabled.
3168 * It also gets called by the fork code, when changing the parent's
3171 void scheduler_tick(void)
3173 unsigned long long now = sched_clock();
3174 struct task_struct *p = current;
3175 int cpu = smp_processor_id();
3176 struct rq *rq = cpu_rq(cpu);
3178 update_cpu_clock(p, rq, now);
3181 /* Task on the idle queue */
3182 wake_priority_sleeper(rq);
3184 task_running_tick(rq, p);
3187 if (time_after_eq(jiffies, rq->next_balance))
3188 raise_softirq(SCHED_SOFTIRQ);
3192 #ifdef CONFIG_SCHED_SMT
3193 static inline void wakeup_busy_runqueue(struct rq *rq)
3195 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3196 if (rq->curr == rq->idle && rq->nr_running)
3197 resched_task(rq->idle);
3201 * Called with interrupt disabled and this_rq's runqueue locked.
3203 static void wake_sleeping_dependent(int this_cpu)
3205 struct sched_domain *tmp, *sd = NULL;
3208 for_each_domain(this_cpu, tmp) {
3209 if (tmp->flags & SD_SHARE_CPUPOWER) {
3218 for_each_cpu_mask(i, sd->span) {
3219 struct rq *smt_rq = cpu_rq(i);
3223 if (unlikely(!spin_trylock(&smt_rq->lock)))
3226 wakeup_busy_runqueue(smt_rq);
3227 spin_unlock(&smt_rq->lock);
3232 * number of 'lost' timeslices this task wont be able to fully
3233 * utilize, if another task runs on a sibling. This models the
3234 * slowdown effect of other tasks running on siblings:
3236 static inline unsigned long
3237 smt_slice(struct task_struct *p, struct sched_domain *sd)
3239 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3243 * To minimise lock contention and not have to drop this_rq's runlock we only
3244 * trylock the sibling runqueues and bypass those runqueues if we fail to
3245 * acquire their lock. As we only trylock the normal locking order does not
3246 * need to be obeyed.
3249 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3251 struct sched_domain *tmp, *sd = NULL;
3254 /* kernel/rt threads do not participate in dependent sleeping */
3255 if (!p->mm || rt_task(p))
3258 for_each_domain(this_cpu, tmp) {
3259 if (tmp->flags & SD_SHARE_CPUPOWER) {
3268 for_each_cpu_mask(i, sd->span) {
3269 struct task_struct *smt_curr;
3276 if (unlikely(!spin_trylock(&smt_rq->lock)))
3279 smt_curr = smt_rq->curr;
3285 * If a user task with lower static priority than the
3286 * running task on the SMT sibling is trying to schedule,
3287 * delay it till there is proportionately less timeslice
3288 * left of the sibling task to prevent a lower priority
3289 * task from using an unfair proportion of the
3290 * physical cpu's resources. -ck
3292 if (rt_task(smt_curr)) {
3294 * With real time tasks we run non-rt tasks only
3295 * per_cpu_gain% of the time.
3297 if ((jiffies % DEF_TIMESLICE) >
3298 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3301 if (smt_curr->static_prio < p->static_prio &&
3302 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3303 smt_slice(smt_curr, sd) > task_timeslice(p))
3307 spin_unlock(&smt_rq->lock);
3312 static inline void wake_sleeping_dependent(int this_cpu)
3316 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3322 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3324 void fastcall add_preempt_count(int val)
3329 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3331 preempt_count() += val;
3333 * Spinlock count overflowing soon?
3335 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3337 EXPORT_SYMBOL(add_preempt_count);
3339 void fastcall sub_preempt_count(int val)
3344 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3347 * Is the spinlock portion underflowing?
3349 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3350 !(preempt_count() & PREEMPT_MASK)))
3353 preempt_count() -= val;
3355 EXPORT_SYMBOL(sub_preempt_count);
3359 static inline int interactive_sleep(enum sleep_type sleep_type)
3361 return (sleep_type == SLEEP_INTERACTIVE ||
3362 sleep_type == SLEEP_INTERRUPTED);
3366 * schedule() is the main scheduler function.
3368 asmlinkage void __sched schedule(void)
3370 struct task_struct *prev, *next;
3371 struct prio_array *array;
3372 struct list_head *queue;
3373 unsigned long long now;
3374 unsigned long run_time;
3375 int cpu, idx, new_prio;
3380 * Test if we are atomic. Since do_exit() needs to call into
3381 * schedule() atomically, we ignore that path for now.
3382 * Otherwise, whine if we are scheduling when we should not be.
3384 if (unlikely(in_atomic() && !current->exit_state)) {
3385 printk(KERN_ERR "BUG: scheduling while atomic: "
3387 current->comm, preempt_count(), current->pid);
3388 debug_show_held_locks(current);
3391 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3396 release_kernel_lock(prev);
3397 need_resched_nonpreemptible:
3401 * The idle thread is not allowed to schedule!
3402 * Remove this check after it has been exercised a bit.
3404 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3405 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3409 schedstat_inc(rq, sched_cnt);
3410 now = sched_clock();
3411 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3412 run_time = now - prev->timestamp;
3413 if (unlikely((long long)(now - prev->timestamp) < 0))
3416 run_time = NS_MAX_SLEEP_AVG;
3419 * Tasks charged proportionately less run_time at high sleep_avg to
3420 * delay them losing their interactive status
3422 run_time /= (CURRENT_BONUS(prev) ? : 1);
3424 spin_lock_irq(&rq->lock);
3426 switch_count = &prev->nivcsw;
3427 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3428 switch_count = &prev->nvcsw;
3429 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3430 unlikely(signal_pending(prev))))
3431 prev->state = TASK_RUNNING;
3433 if (prev->state == TASK_UNINTERRUPTIBLE)
3434 rq->nr_uninterruptible++;
3435 deactivate_task(prev, rq);
3439 cpu = smp_processor_id();
3440 if (unlikely(!rq->nr_running)) {
3441 idle_balance(cpu, rq);
3442 if (!rq->nr_running) {
3444 rq->expired_timestamp = 0;
3445 wake_sleeping_dependent(cpu);
3451 if (unlikely(!array->nr_active)) {
3453 * Switch the active and expired arrays.
3455 schedstat_inc(rq, sched_switch);
3456 rq->active = rq->expired;
3457 rq->expired = array;
3459 rq->expired_timestamp = 0;
3460 rq->best_expired_prio = MAX_PRIO;
3463 idx = sched_find_first_bit(array->bitmap);
3464 queue = array->queue + idx;
3465 next = list_entry(queue->next, struct task_struct, run_list);
3467 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3468 unsigned long long delta = now - next->timestamp;
3469 if (unlikely((long long)(now - next->timestamp) < 0))
3472 if (next->sleep_type == SLEEP_INTERACTIVE)
3473 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3475 array = next->array;
3476 new_prio = recalc_task_prio(next, next->timestamp + delta);
3478 if (unlikely(next->prio != new_prio)) {
3479 dequeue_task(next, array);
3480 next->prio = new_prio;
3481 enqueue_task(next, array);
3484 next->sleep_type = SLEEP_NORMAL;
3485 if (dependent_sleeper(cpu, rq, next))
3488 if (next == rq->idle)
3489 schedstat_inc(rq, sched_goidle);
3491 prefetch_stack(next);
3492 clear_tsk_need_resched(prev);
3493 rcu_qsctr_inc(task_cpu(prev));
3495 update_cpu_clock(prev, rq, now);
3497 prev->sleep_avg -= run_time;
3498 if ((long)prev->sleep_avg <= 0)
3499 prev->sleep_avg = 0;
3500 prev->timestamp = prev->last_ran = now;
3502 sched_info_switch(prev, next);
3503 if (likely(prev != next)) {
3504 next->timestamp = now;
3509 prepare_task_switch(rq, next);
3510 prev = context_switch(rq, prev, next);
3513 * this_rq must be evaluated again because prev may have moved
3514 * CPUs since it called schedule(), thus the 'rq' on its stack
3515 * frame will be invalid.
3517 finish_task_switch(this_rq(), prev);
3519 spin_unlock_irq(&rq->lock);
3522 if (unlikely(reacquire_kernel_lock(prev) < 0))
3523 goto need_resched_nonpreemptible;
3524 preempt_enable_no_resched();
3525 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3528 EXPORT_SYMBOL(schedule);
3530 #ifdef CONFIG_PREEMPT
3532 * this is the entry point to schedule() from in-kernel preemption
3533 * off of preempt_enable. Kernel preemptions off return from interrupt
3534 * occur there and call schedule directly.
3536 asmlinkage void __sched preempt_schedule(void)
3538 struct thread_info *ti = current_thread_info();
3539 #ifdef CONFIG_PREEMPT_BKL
3540 struct task_struct *task = current;
3541 int saved_lock_depth;
3544 * If there is a non-zero preempt_count or interrupts are disabled,
3545 * we do not want to preempt the current task. Just return..
3547 if (likely(ti->preempt_count || irqs_disabled()))
3551 add_preempt_count(PREEMPT_ACTIVE);
3553 * We keep the big kernel semaphore locked, but we
3554 * clear ->lock_depth so that schedule() doesnt
3555 * auto-release the semaphore:
3557 #ifdef CONFIG_PREEMPT_BKL
3558 saved_lock_depth = task->lock_depth;
3559 task->lock_depth = -1;
3562 #ifdef CONFIG_PREEMPT_BKL
3563 task->lock_depth = saved_lock_depth;
3565 sub_preempt_count(PREEMPT_ACTIVE);
3567 /* we could miss a preemption opportunity between schedule and now */
3569 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3572 EXPORT_SYMBOL(preempt_schedule);
3575 * this is the entry point to schedule() from kernel preemption
3576 * off of irq context.
3577 * Note, that this is called and return with irqs disabled. This will
3578 * protect us against recursive calling from irq.
3580 asmlinkage void __sched preempt_schedule_irq(void)
3582 struct thread_info *ti = current_thread_info();
3583 #ifdef CONFIG_PREEMPT_BKL
3584 struct task_struct *task = current;
3585 int saved_lock_depth;
3587 /* Catch callers which need to be fixed */
3588 BUG_ON(ti->preempt_count || !irqs_disabled());
3591 add_preempt_count(PREEMPT_ACTIVE);
3593 * We keep the big kernel semaphore locked, but we
3594 * clear ->lock_depth so that schedule() doesnt
3595 * auto-release the semaphore:
3597 #ifdef CONFIG_PREEMPT_BKL
3598 saved_lock_depth = task->lock_depth;
3599 task->lock_depth = -1;
3603 local_irq_disable();
3604 #ifdef CONFIG_PREEMPT_BKL
3605 task->lock_depth = saved_lock_depth;
3607 sub_preempt_count(PREEMPT_ACTIVE);
3609 /* we could miss a preemption opportunity between schedule and now */
3611 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3615 #endif /* CONFIG_PREEMPT */
3617 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3620 return try_to_wake_up(curr->private, mode, sync);
3622 EXPORT_SYMBOL(default_wake_function);
3625 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3626 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3627 * number) then we wake all the non-exclusive tasks and one exclusive task.
3629 * There are circumstances in which we can try to wake a task which has already
3630 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3631 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3633 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3634 int nr_exclusive, int sync, void *key)
3636 struct list_head *tmp, *next;
3638 list_for_each_safe(tmp, next, &q->task_list) {
3639 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3640 unsigned flags = curr->flags;
3642 if (curr->func(curr, mode, sync, key) &&
3643 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3649 * __wake_up - wake up threads blocked on a waitqueue.
3651 * @mode: which threads
3652 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3653 * @key: is directly passed to the wakeup function
3655 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3656 int nr_exclusive, void *key)
3658 unsigned long flags;
3660 spin_lock_irqsave(&q->lock, flags);
3661 __wake_up_common(q, mode, nr_exclusive, 0, key);
3662 spin_unlock_irqrestore(&q->lock, flags);
3664 EXPORT_SYMBOL(__wake_up);
3667 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3669 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3671 __wake_up_common(q, mode, 1, 0, NULL);
3675 * __wake_up_sync - wake up threads blocked on a waitqueue.
3677 * @mode: which threads
3678 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3680 * The sync wakeup differs that the waker knows that it will schedule
3681 * away soon, so while the target thread will be woken up, it will not
3682 * be migrated to another CPU - ie. the two threads are 'synchronized'
3683 * with each other. This can prevent needless bouncing between CPUs.
3685 * On UP it can prevent extra preemption.
3688 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3690 unsigned long flags;
3696 if (unlikely(!nr_exclusive))
3699 spin_lock_irqsave(&q->lock, flags);
3700 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3701 spin_unlock_irqrestore(&q->lock, flags);
3703 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3705 void fastcall complete(struct completion *x)
3707 unsigned long flags;
3709 spin_lock_irqsave(&x->wait.lock, flags);
3711 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3713 spin_unlock_irqrestore(&x->wait.lock, flags);
3715 EXPORT_SYMBOL(complete);
3717 void fastcall complete_all(struct completion *x)
3719 unsigned long flags;
3721 spin_lock_irqsave(&x->wait.lock, flags);
3722 x->done += UINT_MAX/2;
3723 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3725 spin_unlock_irqrestore(&x->wait.lock, flags);
3727 EXPORT_SYMBOL(complete_all);
3729 void fastcall __sched wait_for_completion(struct completion *x)
3733 spin_lock_irq(&x->wait.lock);
3735 DECLARE_WAITQUEUE(wait, current);
3737 wait.flags |= WQ_FLAG_EXCLUSIVE;
3738 __add_wait_queue_tail(&x->wait, &wait);
3740 __set_current_state(TASK_UNINTERRUPTIBLE);
3741 spin_unlock_irq(&x->wait.lock);
3743 spin_lock_irq(&x->wait.lock);
3745 __remove_wait_queue(&x->wait, &wait);
3748 spin_unlock_irq(&x->wait.lock);
3750 EXPORT_SYMBOL(wait_for_completion);
3752 unsigned long fastcall __sched
3753 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3757 spin_lock_irq(&x->wait.lock);
3759 DECLARE_WAITQUEUE(wait, current);
3761 wait.flags |= WQ_FLAG_EXCLUSIVE;
3762 __add_wait_queue_tail(&x->wait, &wait);
3764 __set_current_state(TASK_UNINTERRUPTIBLE);
3765 spin_unlock_irq(&x->wait.lock);
3766 timeout = schedule_timeout(timeout);
3767 spin_lock_irq(&x->wait.lock);
3769 __remove_wait_queue(&x->wait, &wait);
3773 __remove_wait_queue(&x->wait, &wait);
3777 spin_unlock_irq(&x->wait.lock);
3780 EXPORT_SYMBOL(wait_for_completion_timeout);
3782 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3788 spin_lock_irq(&x->wait.lock);
3790 DECLARE_WAITQUEUE(wait, current);
3792 wait.flags |= WQ_FLAG_EXCLUSIVE;
3793 __add_wait_queue_tail(&x->wait, &wait);
3795 if (signal_pending(current)) {
3797 __remove_wait_queue(&x->wait, &wait);
3800 __set_current_state(TASK_INTERRUPTIBLE);
3801 spin_unlock_irq(&x->wait.lock);
3803 spin_lock_irq(&x->wait.lock);
3805 __remove_wait_queue(&x->wait, &wait);
3809 spin_unlock_irq(&x->wait.lock);
3813 EXPORT_SYMBOL(wait_for_completion_interruptible);
3815 unsigned long fastcall __sched
3816 wait_for_completion_interruptible_timeout(struct completion *x,
3817 unsigned long timeout)
3821 spin_lock_irq(&x->wait.lock);
3823 DECLARE_WAITQUEUE(wait, current);
3825 wait.flags |= WQ_FLAG_EXCLUSIVE;
3826 __add_wait_queue_tail(&x->wait, &wait);
3828 if (signal_pending(current)) {
3829 timeout = -ERESTARTSYS;
3830 __remove_wait_queue(&x->wait, &wait);
3833 __set_current_state(TASK_INTERRUPTIBLE);
3834 spin_unlock_irq(&x->wait.lock);
3835 timeout = schedule_timeout(timeout);
3836 spin_lock_irq(&x->wait.lock);
3838 __remove_wait_queue(&x->wait, &wait);
3842 __remove_wait_queue(&x->wait, &wait);
3846 spin_unlock_irq(&x->wait.lock);
3849 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3852 #define SLEEP_ON_VAR \
3853 unsigned long flags; \
3854 wait_queue_t wait; \
3855 init_waitqueue_entry(&wait, current);
3857 #define SLEEP_ON_HEAD \
3858 spin_lock_irqsave(&q->lock,flags); \
3859 __add_wait_queue(q, &wait); \
3860 spin_unlock(&q->lock);
3862 #define SLEEP_ON_TAIL \
3863 spin_lock_irq(&q->lock); \
3864 __remove_wait_queue(q, &wait); \
3865 spin_unlock_irqrestore(&q->lock, flags);
3867 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3871 current->state = TASK_INTERRUPTIBLE;
3877 EXPORT_SYMBOL(interruptible_sleep_on);
3879 long fastcall __sched
3880 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3884 current->state = TASK_INTERRUPTIBLE;
3887 timeout = schedule_timeout(timeout);
3892 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3894 void fastcall __sched sleep_on(wait_queue_head_t *q)
3898 current->state = TASK_UNINTERRUPTIBLE;
3904 EXPORT_SYMBOL(sleep_on);
3906 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3910 current->state = TASK_UNINTERRUPTIBLE;
3913 timeout = schedule_timeout(timeout);
3919 EXPORT_SYMBOL(sleep_on_timeout);
3921 #ifdef CONFIG_RT_MUTEXES
3924 * rt_mutex_setprio - set the current priority of a task
3926 * @prio: prio value (kernel-internal form)
3928 * This function changes the 'effective' priority of a task. It does
3929 * not touch ->normal_prio like __setscheduler().
3931 * Used by the rt_mutex code to implement priority inheritance logic.
3933 void rt_mutex_setprio(struct task_struct *p, int prio)
3935 struct prio_array *array;
3936 unsigned long flags;
3940 BUG_ON(prio < 0 || prio > MAX_PRIO);
3942 rq = task_rq_lock(p, &flags);
3947 dequeue_task(p, array);
3952 * If changing to an RT priority then queue it
3953 * in the active array!
3957 enqueue_task(p, array);
3959 * Reschedule if we are currently running on this runqueue and
3960 * our priority decreased, or if we are not currently running on
3961 * this runqueue and our priority is higher than the current's
3963 if (task_running(rq, p)) {
3964 if (p->prio > oldprio)
3965 resched_task(rq->curr);
3966 } else if (TASK_PREEMPTS_CURR(p, rq))
3967 resched_task(rq->curr);
3969 task_rq_unlock(rq, &flags);
3974 void set_user_nice(struct task_struct *p, long nice)
3976 struct prio_array *array;
3977 int old_prio, delta;
3978 unsigned long flags;
3981 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3984 * We have to be careful, if called from sys_setpriority(),
3985 * the task might be in the middle of scheduling on another CPU.
3987 rq = task_rq_lock(p, &flags);
3989 * The RT priorities are set via sched_setscheduler(), but we still
3990 * allow the 'normal' nice value to be set - but as expected
3991 * it wont have any effect on scheduling until the task is
3992 * not SCHED_NORMAL/SCHED_BATCH:
3994 if (has_rt_policy(p)) {
3995 p->static_prio = NICE_TO_PRIO(nice);
4000 dequeue_task(p, array);
4001 dec_raw_weighted_load(rq, p);
4004 p->static_prio = NICE_TO_PRIO(nice);
4007 p->prio = effective_prio(p);
4008 delta = p->prio - old_prio;
4011 enqueue_task(p, array);
4012 inc_raw_weighted_load(rq, p);
4014 * If the task increased its priority or is running and
4015 * lowered its priority, then reschedule its CPU:
4017 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4018 resched_task(rq->curr);
4021 task_rq_unlock(rq, &flags);
4023 EXPORT_SYMBOL(set_user_nice);
4026 * can_nice - check if a task can reduce its nice value
4030 int can_nice(const struct task_struct *p, const int nice)
4032 /* convert nice value [19,-20] to rlimit style value [1,40] */
4033 int nice_rlim = 20 - nice;
4035 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4036 capable(CAP_SYS_NICE));
4039 #ifdef __ARCH_WANT_SYS_NICE
4042 * sys_nice - change the priority of the current process.
4043 * @increment: priority increment
4045 * sys_setpriority is a more generic, but much slower function that
4046 * does similar things.
4048 asmlinkage long sys_nice(int increment)
4053 * Setpriority might change our priority at the same moment.
4054 * We don't have to worry. Conceptually one call occurs first
4055 * and we have a single winner.
4057 if (increment < -40)
4062 nice = PRIO_TO_NICE(current->static_prio) + increment;
4068 if (increment < 0 && !can_nice(current, nice))
4071 retval = security_task_setnice(current, nice);
4075 set_user_nice(current, nice);
4082 * task_prio - return the priority value of a given task.
4083 * @p: the task in question.
4085 * This is the priority value as seen by users in /proc.
4086 * RT tasks are offset by -200. Normal tasks are centered
4087 * around 0, value goes from -16 to +15.
4089 int task_prio(const struct task_struct *p)
4091 return p->prio - MAX_RT_PRIO;
4095 * task_nice - return the nice value of a given task.
4096 * @p: the task in question.
4098 int task_nice(const struct task_struct *p)
4100 return TASK_NICE(p);
4102 EXPORT_SYMBOL_GPL(task_nice);
4105 * idle_cpu - is a given cpu idle currently?
4106 * @cpu: the processor in question.
4108 int idle_cpu(int cpu)
4110 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4114 * idle_task - return the idle task for a given cpu.
4115 * @cpu: the processor in question.
4117 struct task_struct *idle_task(int cpu)
4119 return cpu_rq(cpu)->idle;
4123 * find_process_by_pid - find a process with a matching PID value.
4124 * @pid: the pid in question.
4126 static inline struct task_struct *find_process_by_pid(pid_t pid)
4128 return pid ? find_task_by_pid(pid) : current;
4131 /* Actually do priority change: must hold rq lock. */
4132 static void __setscheduler(struct task_struct *p, int policy, int prio)
4137 p->rt_priority = prio;
4138 p->normal_prio = normal_prio(p);
4139 /* we are holding p->pi_lock already */
4140 p->prio = rt_mutex_getprio(p);
4142 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4144 if (policy == SCHED_BATCH)
4150 * sched_setscheduler - change the scheduling policy and/or RT priority of
4152 * @p: the task in question.
4153 * @policy: new policy.
4154 * @param: structure containing the new RT priority.
4156 * NOTE: the task may be already dead
4158 int sched_setscheduler(struct task_struct *p, int policy,
4159 struct sched_param *param)
4161 int retval, oldprio, oldpolicy = -1;
4162 struct prio_array *array;
4163 unsigned long flags;
4166 /* may grab non-irq protected spin_locks */
4167 BUG_ON(in_interrupt());
4169 /* double check policy once rq lock held */
4171 policy = oldpolicy = p->policy;
4172 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4173 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4176 * Valid priorities for SCHED_FIFO and SCHED_RR are
4177 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4180 if (param->sched_priority < 0 ||
4181 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4182 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4184 if (is_rt_policy(policy) != (param->sched_priority != 0))
4188 * Allow unprivileged RT tasks to decrease priority:
4190 if (!capable(CAP_SYS_NICE)) {
4191 if (is_rt_policy(policy)) {
4192 unsigned long rlim_rtprio;
4193 unsigned long flags;
4195 if (!lock_task_sighand(p, &flags))
4197 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4198 unlock_task_sighand(p, &flags);
4200 /* can't set/change the rt policy */
4201 if (policy != p->policy && !rlim_rtprio)
4204 /* can't increase priority */
4205 if (param->sched_priority > p->rt_priority &&
4206 param->sched_priority > rlim_rtprio)
4210 /* can't change other user's priorities */
4211 if ((current->euid != p->euid) &&
4212 (current->euid != p->uid))
4216 retval = security_task_setscheduler(p, policy, param);
4220 * make sure no PI-waiters arrive (or leave) while we are
4221 * changing the priority of the task:
4223 spin_lock_irqsave(&p->pi_lock, flags);
4225 * To be able to change p->policy safely, the apropriate
4226 * runqueue lock must be held.
4228 rq = __task_rq_lock(p);
4229 /* recheck policy now with rq lock held */
4230 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4231 policy = oldpolicy = -1;
4232 __task_rq_unlock(rq);
4233 spin_unlock_irqrestore(&p->pi_lock, flags);
4238 deactivate_task(p, rq);
4240 __setscheduler(p, policy, param->sched_priority);
4242 __activate_task(p, rq);
4244 * Reschedule if we are currently running on this runqueue and
4245 * our priority decreased, or if we are not currently running on
4246 * this runqueue and our priority is higher than the current's
4248 if (task_running(rq, p)) {
4249 if (p->prio > oldprio)
4250 resched_task(rq->curr);
4251 } else if (TASK_PREEMPTS_CURR(p, rq))
4252 resched_task(rq->curr);
4254 __task_rq_unlock(rq);
4255 spin_unlock_irqrestore(&p->pi_lock, flags);
4257 rt_mutex_adjust_pi(p);
4261 EXPORT_SYMBOL_GPL(sched_setscheduler);
4264 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4266 struct sched_param lparam;
4267 struct task_struct *p;
4270 if (!param || pid < 0)
4272 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4277 p = find_process_by_pid(pid);
4279 retval = sched_setscheduler(p, policy, &lparam);
4286 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4287 * @pid: the pid in question.
4288 * @policy: new policy.
4289 * @param: structure containing the new RT priority.
4291 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4292 struct sched_param __user *param)
4294 /* negative values for policy are not valid */
4298 return do_sched_setscheduler(pid, policy, param);
4302 * sys_sched_setparam - set/change the RT priority of a thread
4303 * @pid: the pid in question.
4304 * @param: structure containing the new RT priority.
4306 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4308 return do_sched_setscheduler(pid, -1, param);
4312 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4313 * @pid: the pid in question.
4315 asmlinkage long sys_sched_getscheduler(pid_t pid)
4317 struct task_struct *p;
4318 int retval = -EINVAL;
4324 read_lock(&tasklist_lock);
4325 p = find_process_by_pid(pid);
4327 retval = security_task_getscheduler(p);
4331 read_unlock(&tasklist_lock);
4338 * sys_sched_getscheduler - get the RT priority of a thread
4339 * @pid: the pid in question.
4340 * @param: structure containing the RT priority.
4342 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4344 struct sched_param lp;
4345 struct task_struct *p;
4346 int retval = -EINVAL;
4348 if (!param || pid < 0)
4351 read_lock(&tasklist_lock);
4352 p = find_process_by_pid(pid);
4357 retval = security_task_getscheduler(p);
4361 lp.sched_priority = p->rt_priority;
4362 read_unlock(&tasklist_lock);
4365 * This one might sleep, we cannot do it with a spinlock held ...
4367 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4373 read_unlock(&tasklist_lock);
4377 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4379 cpumask_t cpus_allowed;
4380 struct task_struct *p;
4384 read_lock(&tasklist_lock);
4386 p = find_process_by_pid(pid);
4388 read_unlock(&tasklist_lock);
4389 unlock_cpu_hotplug();
4394 * It is not safe to call set_cpus_allowed with the
4395 * tasklist_lock held. We will bump the task_struct's
4396 * usage count and then drop tasklist_lock.
4399 read_unlock(&tasklist_lock);
4402 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4403 !capable(CAP_SYS_NICE))
4406 retval = security_task_setscheduler(p, 0, NULL);
4410 cpus_allowed = cpuset_cpus_allowed(p);
4411 cpus_and(new_mask, new_mask, cpus_allowed);
4412 retval = set_cpus_allowed(p, new_mask);
4416 unlock_cpu_hotplug();
4420 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4421 cpumask_t *new_mask)
4423 if (len < sizeof(cpumask_t)) {
4424 memset(new_mask, 0, sizeof(cpumask_t));
4425 } else if (len > sizeof(cpumask_t)) {
4426 len = sizeof(cpumask_t);
4428 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4432 * sys_sched_setaffinity - set the cpu affinity of a process
4433 * @pid: pid of the process
4434 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4435 * @user_mask_ptr: user-space pointer to the new cpu mask
4437 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4438 unsigned long __user *user_mask_ptr)
4443 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4447 return sched_setaffinity(pid, new_mask);
4451 * Represents all cpu's present in the system
4452 * In systems capable of hotplug, this map could dynamically grow
4453 * as new cpu's are detected in the system via any platform specific
4454 * method, such as ACPI for e.g.
4457 cpumask_t cpu_present_map __read_mostly;
4458 EXPORT_SYMBOL(cpu_present_map);
4461 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4462 EXPORT_SYMBOL(cpu_online_map);
4464 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4465 EXPORT_SYMBOL(cpu_possible_map);
4468 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4470 struct task_struct *p;
4474 read_lock(&tasklist_lock);
4477 p = find_process_by_pid(pid);
4481 retval = security_task_getscheduler(p);
4485 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4488 read_unlock(&tasklist_lock);
4489 unlock_cpu_hotplug();
4497 * sys_sched_getaffinity - get the cpu affinity of a process
4498 * @pid: pid of the process
4499 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4500 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4502 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4503 unsigned long __user *user_mask_ptr)
4508 if (len < sizeof(cpumask_t))
4511 ret = sched_getaffinity(pid, &mask);
4515 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4518 return sizeof(cpumask_t);
4522 * sys_sched_yield - yield the current processor to other threads.
4524 * this function yields the current CPU by moving the calling thread
4525 * to the expired array. If there are no other threads running on this
4526 * CPU then this function will return.
4528 asmlinkage long sys_sched_yield(void)
4530 struct rq *rq = this_rq_lock();
4531 struct prio_array *array = current->array, *target = rq->expired;
4533 schedstat_inc(rq, yld_cnt);
4535 * We implement yielding by moving the task into the expired
4538 * (special rule: RT tasks will just roundrobin in the active
4541 if (rt_task(current))
4542 target = rq->active;
4544 if (array->nr_active == 1) {
4545 schedstat_inc(rq, yld_act_empty);
4546 if (!rq->expired->nr_active)
4547 schedstat_inc(rq, yld_both_empty);
4548 } else if (!rq->expired->nr_active)
4549 schedstat_inc(rq, yld_exp_empty);
4551 if (array != target) {
4552 dequeue_task(current, array);
4553 enqueue_task(current, target);
4556 * requeue_task is cheaper so perform that if possible.
4558 requeue_task(current, array);
4561 * Since we are going to call schedule() anyway, there's
4562 * no need to preempt or enable interrupts:
4564 __release(rq->lock);
4565 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4566 _raw_spin_unlock(&rq->lock);
4567 preempt_enable_no_resched();
4574 static inline int __resched_legal(int expected_preempt_count)
4576 if (unlikely(preempt_count() != expected_preempt_count))
4578 if (unlikely(system_state != SYSTEM_RUNNING))
4583 static void __cond_resched(void)
4585 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4586 __might_sleep(__FILE__, __LINE__);
4589 * The BKS might be reacquired before we have dropped
4590 * PREEMPT_ACTIVE, which could trigger a second
4591 * cond_resched() call.
4594 add_preempt_count(PREEMPT_ACTIVE);
4596 sub_preempt_count(PREEMPT_ACTIVE);
4597 } while (need_resched());
4600 int __sched cond_resched(void)
4602 if (need_resched() && __resched_legal(0)) {
4608 EXPORT_SYMBOL(cond_resched);
4611 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4612 * call schedule, and on return reacquire the lock.
4614 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4615 * operations here to prevent schedule() from being called twice (once via
4616 * spin_unlock(), once by hand).
4618 int cond_resched_lock(spinlock_t *lock)
4622 if (need_lockbreak(lock)) {
4628 if (need_resched() && __resched_legal(1)) {
4629 spin_release(&lock->dep_map, 1, _THIS_IP_);
4630 _raw_spin_unlock(lock);
4631 preempt_enable_no_resched();
4638 EXPORT_SYMBOL(cond_resched_lock);
4640 int __sched cond_resched_softirq(void)
4642 BUG_ON(!in_softirq());
4644 if (need_resched() && __resched_legal(0)) {
4645 raw_local_irq_disable();
4647 raw_local_irq_enable();
4654 EXPORT_SYMBOL(cond_resched_softirq);
4657 * yield - yield the current processor to other threads.
4659 * this is a shortcut for kernel-space yielding - it marks the
4660 * thread runnable and calls sys_sched_yield().
4662 void __sched yield(void)
4664 set_current_state(TASK_RUNNING);
4667 EXPORT_SYMBOL(yield);
4670 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4671 * that process accounting knows that this is a task in IO wait state.
4673 * But don't do that if it is a deliberate, throttling IO wait (this task
4674 * has set its backing_dev_info: the queue against which it should throttle)
4676 void __sched io_schedule(void)
4678 struct rq *rq = &__raw_get_cpu_var(runqueues);
4680 delayacct_blkio_start();
4681 atomic_inc(&rq->nr_iowait);
4683 atomic_dec(&rq->nr_iowait);
4684 delayacct_blkio_end();
4686 EXPORT_SYMBOL(io_schedule);
4688 long __sched io_schedule_timeout(long timeout)
4690 struct rq *rq = &__raw_get_cpu_var(runqueues);
4693 delayacct_blkio_start();
4694 atomic_inc(&rq->nr_iowait);
4695 ret = schedule_timeout(timeout);
4696 atomic_dec(&rq->nr_iowait);
4697 delayacct_blkio_end();
4702 * sys_sched_get_priority_max - return maximum RT priority.
4703 * @policy: scheduling class.
4705 * this syscall returns the maximum rt_priority that can be used
4706 * by a given scheduling class.
4708 asmlinkage long sys_sched_get_priority_max(int policy)
4715 ret = MAX_USER_RT_PRIO-1;
4726 * sys_sched_get_priority_min - return minimum RT priority.
4727 * @policy: scheduling class.
4729 * this syscall returns the minimum rt_priority that can be used
4730 * by a given scheduling class.
4732 asmlinkage long sys_sched_get_priority_min(int policy)
4749 * sys_sched_rr_get_interval - return the default timeslice of a process.
4750 * @pid: pid of the process.
4751 * @interval: userspace pointer to the timeslice value.
4753 * this syscall writes the default timeslice value of a given process
4754 * into the user-space timespec buffer. A value of '0' means infinity.
4757 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4759 struct task_struct *p;
4760 int retval = -EINVAL;
4767 read_lock(&tasklist_lock);
4768 p = find_process_by_pid(pid);
4772 retval = security_task_getscheduler(p);
4776 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4777 0 : task_timeslice(p), &t);
4778 read_unlock(&tasklist_lock);
4779 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4783 read_unlock(&tasklist_lock);
4787 static inline struct task_struct *eldest_child(struct task_struct *p)
4789 if (list_empty(&p->children))
4791 return list_entry(p->children.next,struct task_struct,sibling);
4794 static inline struct task_struct *older_sibling(struct task_struct *p)
4796 if (p->sibling.prev==&p->parent->children)
4798 return list_entry(p->sibling.prev,struct task_struct,sibling);
4801 static inline struct task_struct *younger_sibling(struct task_struct *p)
4803 if (p->sibling.next==&p->parent->children)
4805 return list_entry(p->sibling.next,struct task_struct,sibling);
4808 static const char stat_nam[] = "RSDTtZX";
4810 static void show_task(struct task_struct *p)
4812 struct task_struct *relative;
4813 unsigned long free = 0;
4816 state = p->state ? __ffs(p->state) + 1 : 0;
4817 printk("%-13.13s %c", p->comm,
4818 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4819 #if (BITS_PER_LONG == 32)
4820 if (state == TASK_RUNNING)
4821 printk(" running ");
4823 printk(" %08lX ", thread_saved_pc(p));
4825 if (state == TASK_RUNNING)
4826 printk(" running task ");
4828 printk(" %016lx ", thread_saved_pc(p));
4830 #ifdef CONFIG_DEBUG_STACK_USAGE
4832 unsigned long *n = end_of_stack(p);
4835 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4838 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4839 if ((relative = eldest_child(p)))
4840 printk("%5d ", relative->pid);
4843 if ((relative = younger_sibling(p)))
4844 printk("%7d", relative->pid);
4847 if ((relative = older_sibling(p)))
4848 printk(" %5d", relative->pid);
4852 printk(" (L-TLB)\n");
4854 printk(" (NOTLB)\n");
4856 if (state != TASK_RUNNING)
4857 show_stack(p, NULL);
4860 void show_state_filter(unsigned long state_filter)
4862 struct task_struct *g, *p;
4864 #if (BITS_PER_LONG == 32)
4867 printk(" task PC stack pid father child younger older\n");
4871 printk(" task PC stack pid father child younger older\n");
4873 read_lock(&tasklist_lock);
4874 do_each_thread(g, p) {
4876 * reset the NMI-timeout, listing all files on a slow
4877 * console might take alot of time:
4879 touch_nmi_watchdog();
4880 if (p->state & state_filter)
4882 } while_each_thread(g, p);
4884 read_unlock(&tasklist_lock);
4886 * Only show locks if all tasks are dumped:
4888 if (state_filter == -1)
4889 debug_show_all_locks();
4893 * init_idle - set up an idle thread for a given CPU
4894 * @idle: task in question
4895 * @cpu: cpu the idle task belongs to
4897 * NOTE: this function does not set the idle thread's NEED_RESCHED
4898 * flag, to make booting more robust.
4900 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4902 struct rq *rq = cpu_rq(cpu);
4903 unsigned long flags;
4905 idle->timestamp = sched_clock();
4906 idle->sleep_avg = 0;
4908 idle->prio = idle->normal_prio = MAX_PRIO;
4909 idle->state = TASK_RUNNING;
4910 idle->cpus_allowed = cpumask_of_cpu(cpu);
4911 set_task_cpu(idle, cpu);
4913 spin_lock_irqsave(&rq->lock, flags);
4914 rq->curr = rq->idle = idle;
4915 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4918 spin_unlock_irqrestore(&rq->lock, flags);
4920 /* Set the preempt count _outside_ the spinlocks! */
4921 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4922 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4924 task_thread_info(idle)->preempt_count = 0;
4929 * In a system that switches off the HZ timer nohz_cpu_mask
4930 * indicates which cpus entered this state. This is used
4931 * in the rcu update to wait only for active cpus. For system
4932 * which do not switch off the HZ timer nohz_cpu_mask should
4933 * always be CPU_MASK_NONE.
4935 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4939 * This is how migration works:
4941 * 1) we queue a struct migration_req structure in the source CPU's
4942 * runqueue and wake up that CPU's migration thread.
4943 * 2) we down() the locked semaphore => thread blocks.
4944 * 3) migration thread wakes up (implicitly it forces the migrated
4945 * thread off the CPU)
4946 * 4) it gets the migration request and checks whether the migrated
4947 * task is still in the wrong runqueue.
4948 * 5) if it's in the wrong runqueue then the migration thread removes
4949 * it and puts it into the right queue.
4950 * 6) migration thread up()s the semaphore.
4951 * 7) we wake up and the migration is done.
4955 * Change a given task's CPU affinity. Migrate the thread to a
4956 * proper CPU and schedule it away if the CPU it's executing on
4957 * is removed from the allowed bitmask.
4959 * NOTE: the caller must have a valid reference to the task, the
4960 * task must not exit() & deallocate itself prematurely. The
4961 * call is not atomic; no spinlocks may be held.
4963 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4965 struct migration_req req;
4966 unsigned long flags;
4970 rq = task_rq_lock(p, &flags);
4971 if (!cpus_intersects(new_mask, cpu_online_map)) {
4976 p->cpus_allowed = new_mask;
4977 /* Can the task run on the task's current CPU? If so, we're done */
4978 if (cpu_isset(task_cpu(p), new_mask))
4981 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4982 /* Need help from migration thread: drop lock and wait. */
4983 task_rq_unlock(rq, &flags);
4984 wake_up_process(rq->migration_thread);
4985 wait_for_completion(&req.done);
4986 tlb_migrate_finish(p->mm);
4990 task_rq_unlock(rq, &flags);
4994 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4997 * Move (not current) task off this cpu, onto dest cpu. We're doing
4998 * this because either it can't run here any more (set_cpus_allowed()
4999 * away from this CPU, or CPU going down), or because we're
5000 * attempting to rebalance this task on exec (sched_exec).
5002 * So we race with normal scheduler movements, but that's OK, as long
5003 * as the task is no longer on this CPU.
5005 * Returns non-zero if task was successfully migrated.
5007 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5009 struct rq *rq_dest, *rq_src;
5012 if (unlikely(cpu_is_offline(dest_cpu)))
5015 rq_src = cpu_rq(src_cpu);
5016 rq_dest = cpu_rq(dest_cpu);
5018 double_rq_lock(rq_src, rq_dest);
5019 /* Already moved. */
5020 if (task_cpu(p) != src_cpu)
5022 /* Affinity changed (again). */
5023 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5026 set_task_cpu(p, dest_cpu);
5029 * Sync timestamp with rq_dest's before activating.
5030 * The same thing could be achieved by doing this step
5031 * afterwards, and pretending it was a local activate.
5032 * This way is cleaner and logically correct.
5034 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5035 + rq_dest->most_recent_timestamp;
5036 deactivate_task(p, rq_src);
5037 __activate_task(p, rq_dest);
5038 if (TASK_PREEMPTS_CURR(p, rq_dest))
5039 resched_task(rq_dest->curr);
5043 double_rq_unlock(rq_src, rq_dest);
5048 * migration_thread - this is a highprio system thread that performs
5049 * thread migration by bumping thread off CPU then 'pushing' onto
5052 static int migration_thread(void *data)
5054 int cpu = (long)data;
5058 BUG_ON(rq->migration_thread != current);
5060 set_current_state(TASK_INTERRUPTIBLE);
5061 while (!kthread_should_stop()) {
5062 struct migration_req *req;
5063 struct list_head *head;
5067 spin_lock_irq(&rq->lock);
5069 if (cpu_is_offline(cpu)) {
5070 spin_unlock_irq(&rq->lock);
5074 if (rq->active_balance) {
5075 active_load_balance(rq, cpu);
5076 rq->active_balance = 0;
5079 head = &rq->migration_queue;
5081 if (list_empty(head)) {
5082 spin_unlock_irq(&rq->lock);
5084 set_current_state(TASK_INTERRUPTIBLE);
5087 req = list_entry(head->next, struct migration_req, list);
5088 list_del_init(head->next);
5090 spin_unlock(&rq->lock);
5091 __migrate_task(req->task, cpu, req->dest_cpu);
5094 complete(&req->done);
5096 __set_current_state(TASK_RUNNING);
5100 /* Wait for kthread_stop */
5101 set_current_state(TASK_INTERRUPTIBLE);
5102 while (!kthread_should_stop()) {
5104 set_current_state(TASK_INTERRUPTIBLE);
5106 __set_current_state(TASK_RUNNING);
5110 #ifdef CONFIG_HOTPLUG_CPU
5112 * Figure out where task on dead CPU should go, use force if neccessary.
5113 * NOTE: interrupts should be disabled by the caller
5115 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5117 unsigned long flags;
5124 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5125 cpus_and(mask, mask, p->cpus_allowed);
5126 dest_cpu = any_online_cpu(mask);
5128 /* On any allowed CPU? */
5129 if (dest_cpu == NR_CPUS)
5130 dest_cpu = any_online_cpu(p->cpus_allowed);
5132 /* No more Mr. Nice Guy. */
5133 if (dest_cpu == NR_CPUS) {
5134 rq = task_rq_lock(p, &flags);
5135 cpus_setall(p->cpus_allowed);
5136 dest_cpu = any_online_cpu(p->cpus_allowed);
5137 task_rq_unlock(rq, &flags);
5140 * Don't tell them about moving exiting tasks or
5141 * kernel threads (both mm NULL), since they never
5144 if (p->mm && printk_ratelimit())
5145 printk(KERN_INFO "process %d (%s) no "
5146 "longer affine to cpu%d\n",
5147 p->pid, p->comm, dead_cpu);
5149 if (!__migrate_task(p, dead_cpu, dest_cpu))
5154 * While a dead CPU has no uninterruptible tasks queued at this point,
5155 * it might still have a nonzero ->nr_uninterruptible counter, because
5156 * for performance reasons the counter is not stricly tracking tasks to
5157 * their home CPUs. So we just add the counter to another CPU's counter,
5158 * to keep the global sum constant after CPU-down:
5160 static void migrate_nr_uninterruptible(struct rq *rq_src)
5162 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5163 unsigned long flags;
5165 local_irq_save(flags);
5166 double_rq_lock(rq_src, rq_dest);
5167 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5168 rq_src->nr_uninterruptible = 0;
5169 double_rq_unlock(rq_src, rq_dest);
5170 local_irq_restore(flags);
5173 /* Run through task list and migrate tasks from the dead cpu. */
5174 static void migrate_live_tasks(int src_cpu)
5176 struct task_struct *p, *t;
5178 write_lock_irq(&tasklist_lock);
5180 do_each_thread(t, p) {
5184 if (task_cpu(p) == src_cpu)
5185 move_task_off_dead_cpu(src_cpu, p);
5186 } while_each_thread(t, p);
5188 write_unlock_irq(&tasklist_lock);
5191 /* Schedules idle task to be the next runnable task on current CPU.
5192 * It does so by boosting its priority to highest possible and adding it to
5193 * the _front_ of the runqueue. Used by CPU offline code.
5195 void sched_idle_next(void)
5197 int this_cpu = smp_processor_id();
5198 struct rq *rq = cpu_rq(this_cpu);
5199 struct task_struct *p = rq->idle;
5200 unsigned long flags;
5202 /* cpu has to be offline */
5203 BUG_ON(cpu_online(this_cpu));
5206 * Strictly not necessary since rest of the CPUs are stopped by now
5207 * and interrupts disabled on the current cpu.
5209 spin_lock_irqsave(&rq->lock, flags);
5211 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5213 /* Add idle task to the _front_ of its priority queue: */
5214 __activate_idle_task(p, rq);
5216 spin_unlock_irqrestore(&rq->lock, flags);
5220 * Ensures that the idle task is using init_mm right before its cpu goes
5223 void idle_task_exit(void)
5225 struct mm_struct *mm = current->active_mm;
5227 BUG_ON(cpu_online(smp_processor_id()));
5230 switch_mm(mm, &init_mm, current);
5234 /* called under rq->lock with disabled interrupts */
5235 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5237 struct rq *rq = cpu_rq(dead_cpu);
5239 /* Must be exiting, otherwise would be on tasklist. */
5240 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5242 /* Cannot have done final schedule yet: would have vanished. */
5243 BUG_ON(p->state == TASK_DEAD);
5248 * Drop lock around migration; if someone else moves it,
5249 * that's OK. No task can be added to this CPU, so iteration is
5251 * NOTE: interrupts should be left disabled --dev@
5253 spin_unlock(&rq->lock);
5254 move_task_off_dead_cpu(dead_cpu, p);
5255 spin_lock(&rq->lock);
5260 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5261 static void migrate_dead_tasks(unsigned int dead_cpu)
5263 struct rq *rq = cpu_rq(dead_cpu);
5264 unsigned int arr, i;
5266 for (arr = 0; arr < 2; arr++) {
5267 for (i = 0; i < MAX_PRIO; i++) {
5268 struct list_head *list = &rq->arrays[arr].queue[i];
5270 while (!list_empty(list))
5271 migrate_dead(dead_cpu, list_entry(list->next,
5272 struct task_struct, run_list));
5276 #endif /* CONFIG_HOTPLUG_CPU */
5279 * migration_call - callback that gets triggered when a CPU is added.
5280 * Here we can start up the necessary migration thread for the new CPU.
5282 static int __cpuinit
5283 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5285 struct task_struct *p;
5286 int cpu = (long)hcpu;
5287 unsigned long flags;
5291 case CPU_UP_PREPARE:
5292 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5295 p->flags |= PF_NOFREEZE;
5296 kthread_bind(p, cpu);
5297 /* Must be high prio: stop_machine expects to yield to it. */
5298 rq = task_rq_lock(p, &flags);
5299 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5300 task_rq_unlock(rq, &flags);
5301 cpu_rq(cpu)->migration_thread = p;
5305 /* Strictly unneccessary, as first user will wake it. */
5306 wake_up_process(cpu_rq(cpu)->migration_thread);
5309 #ifdef CONFIG_HOTPLUG_CPU
5310 case CPU_UP_CANCELED:
5311 if (!cpu_rq(cpu)->migration_thread)
5313 /* Unbind it from offline cpu so it can run. Fall thru. */
5314 kthread_bind(cpu_rq(cpu)->migration_thread,
5315 any_online_cpu(cpu_online_map));
5316 kthread_stop(cpu_rq(cpu)->migration_thread);
5317 cpu_rq(cpu)->migration_thread = NULL;
5321 migrate_live_tasks(cpu);
5323 kthread_stop(rq->migration_thread);
5324 rq->migration_thread = NULL;
5325 /* Idle task back to normal (off runqueue, low prio) */
5326 rq = task_rq_lock(rq->idle, &flags);
5327 deactivate_task(rq->idle, rq);
5328 rq->idle->static_prio = MAX_PRIO;
5329 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5330 migrate_dead_tasks(cpu);
5331 task_rq_unlock(rq, &flags);
5332 migrate_nr_uninterruptible(rq);
5333 BUG_ON(rq->nr_running != 0);
5335 /* No need to migrate the tasks: it was best-effort if
5336 * they didn't do lock_cpu_hotplug(). Just wake up
5337 * the requestors. */
5338 spin_lock_irq(&rq->lock);
5339 while (!list_empty(&rq->migration_queue)) {
5340 struct migration_req *req;
5342 req = list_entry(rq->migration_queue.next,
5343 struct migration_req, list);
5344 list_del_init(&req->list);
5345 complete(&req->done);
5347 spin_unlock_irq(&rq->lock);
5354 /* Register at highest priority so that task migration (migrate_all_tasks)
5355 * happens before everything else.
5357 static struct notifier_block __cpuinitdata migration_notifier = {
5358 .notifier_call = migration_call,
5362 int __init migration_init(void)
5364 void *cpu = (void *)(long)smp_processor_id();
5367 /* Start one for the boot CPU: */
5368 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5369 BUG_ON(err == NOTIFY_BAD);
5370 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5371 register_cpu_notifier(&migration_notifier);
5378 #undef SCHED_DOMAIN_DEBUG
5379 #ifdef SCHED_DOMAIN_DEBUG
5380 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5385 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5389 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5394 struct sched_group *group = sd->groups;
5395 cpumask_t groupmask;
5397 cpumask_scnprintf(str, NR_CPUS, sd->span);
5398 cpus_clear(groupmask);
5401 for (i = 0; i < level + 1; i++)
5403 printk("domain %d: ", level);
5405 if (!(sd->flags & SD_LOAD_BALANCE)) {
5406 printk("does not load-balance\n");
5408 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5412 printk("span %s\n", str);
5414 if (!cpu_isset(cpu, sd->span))
5415 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5416 if (!cpu_isset(cpu, group->cpumask))
5417 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5420 for (i = 0; i < level + 2; i++)
5426 printk(KERN_ERR "ERROR: group is NULL\n");
5430 if (!group->cpu_power) {
5432 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5435 if (!cpus_weight(group->cpumask)) {
5437 printk(KERN_ERR "ERROR: empty group\n");
5440 if (cpus_intersects(groupmask, group->cpumask)) {
5442 printk(KERN_ERR "ERROR: repeated CPUs\n");
5445 cpus_or(groupmask, groupmask, group->cpumask);
5447 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5450 group = group->next;
5451 } while (group != sd->groups);
5454 if (!cpus_equal(sd->span, groupmask))
5455 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5461 if (!cpus_subset(groupmask, sd->span))
5462 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5468 # define sched_domain_debug(sd, cpu) do { } while (0)
5471 static int sd_degenerate(struct sched_domain *sd)
5473 if (cpus_weight(sd->span) == 1)
5476 /* Following flags need at least 2 groups */
5477 if (sd->flags & (SD_LOAD_BALANCE |
5478 SD_BALANCE_NEWIDLE |
5482 SD_SHARE_PKG_RESOURCES)) {
5483 if (sd->groups != sd->groups->next)
5487 /* Following flags don't use groups */
5488 if (sd->flags & (SD_WAKE_IDLE |
5497 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5499 unsigned long cflags = sd->flags, pflags = parent->flags;
5501 if (sd_degenerate(parent))
5504 if (!cpus_equal(sd->span, parent->span))
5507 /* Does parent contain flags not in child? */
5508 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5509 if (cflags & SD_WAKE_AFFINE)
5510 pflags &= ~SD_WAKE_BALANCE;
5511 /* Flags needing groups don't count if only 1 group in parent */
5512 if (parent->groups == parent->groups->next) {
5513 pflags &= ~(SD_LOAD_BALANCE |
5514 SD_BALANCE_NEWIDLE |
5518 SD_SHARE_PKG_RESOURCES);
5520 if (~cflags & pflags)
5527 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5528 * hold the hotplug lock.
5530 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5532 struct rq *rq = cpu_rq(cpu);
5533 struct sched_domain *tmp;
5535 /* Remove the sched domains which do not contribute to scheduling. */
5536 for (tmp = sd; tmp; tmp = tmp->parent) {
5537 struct sched_domain *parent = tmp->parent;
5540 if (sd_parent_degenerate(tmp, parent)) {
5541 tmp->parent = parent->parent;
5543 parent->parent->child = tmp;
5547 if (sd && sd_degenerate(sd)) {
5553 sched_domain_debug(sd, cpu);
5555 rcu_assign_pointer(rq->sd, sd);
5558 /* cpus with isolated domains */
5559 static cpumask_t __cpuinitdata cpu_isolated_map = CPU_MASK_NONE;
5561 /* Setup the mask of cpus configured for isolated domains */
5562 static int __init isolated_cpu_setup(char *str)
5564 int ints[NR_CPUS], i;
5566 str = get_options(str, ARRAY_SIZE(ints), ints);
5567 cpus_clear(cpu_isolated_map);
5568 for (i = 1; i <= ints[0]; i++)
5569 if (ints[i] < NR_CPUS)
5570 cpu_set(ints[i], cpu_isolated_map);
5574 __setup ("isolcpus=", isolated_cpu_setup);
5577 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5578 * to a function which identifies what group(along with sched group) a CPU
5579 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5580 * (due to the fact that we keep track of groups covered with a cpumask_t).
5582 * init_sched_build_groups will build a circular linked list of the groups
5583 * covered by the given span, and will set each group's ->cpumask correctly,
5584 * and ->cpu_power to 0.
5587 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5588 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5589 struct sched_group **sg))
5591 struct sched_group *first = NULL, *last = NULL;
5592 cpumask_t covered = CPU_MASK_NONE;
5595 for_each_cpu_mask(i, span) {
5596 struct sched_group *sg;
5597 int group = group_fn(i, cpu_map, &sg);
5600 if (cpu_isset(i, covered))
5603 sg->cpumask = CPU_MASK_NONE;
5606 for_each_cpu_mask(j, span) {
5607 if (group_fn(j, cpu_map, NULL) != group)
5610 cpu_set(j, covered);
5611 cpu_set(j, sg->cpumask);
5622 #define SD_NODES_PER_DOMAIN 16
5625 * Self-tuning task migration cost measurement between source and target CPUs.
5627 * This is done by measuring the cost of manipulating buffers of varying
5628 * sizes. For a given buffer-size here are the steps that are taken:
5630 * 1) the source CPU reads+dirties a shared buffer
5631 * 2) the target CPU reads+dirties the same shared buffer
5633 * We measure how long they take, in the following 4 scenarios:
5635 * - source: CPU1, target: CPU2 | cost1
5636 * - source: CPU2, target: CPU1 | cost2
5637 * - source: CPU1, target: CPU1 | cost3
5638 * - source: CPU2, target: CPU2 | cost4
5640 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5641 * the cost of migration.
5643 * We then start off from a small buffer-size and iterate up to larger
5644 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5645 * doing a maximum search for the cost. (The maximum cost for a migration
5646 * normally occurs when the working set size is around the effective cache
5649 #define SEARCH_SCOPE 2
5650 #define MIN_CACHE_SIZE (64*1024U)
5651 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5652 #define ITERATIONS 1
5653 #define SIZE_THRESH 130
5654 #define COST_THRESH 130
5657 * The migration cost is a function of 'domain distance'. Domain
5658 * distance is the number of steps a CPU has to iterate down its
5659 * domain tree to share a domain with the other CPU. The farther
5660 * two CPUs are from each other, the larger the distance gets.
5662 * Note that we use the distance only to cache measurement results,
5663 * the distance value is not used numerically otherwise. When two
5664 * CPUs have the same distance it is assumed that the migration
5665 * cost is the same. (this is a simplification but quite practical)
5667 #define MAX_DOMAIN_DISTANCE 32
5669 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5670 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5672 * Architectures may override the migration cost and thus avoid
5673 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5674 * virtualized hardware:
5676 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5677 CONFIG_DEFAULT_MIGRATION_COST
5684 * Allow override of migration cost - in units of microseconds.
5685 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5686 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5688 static int __init migration_cost_setup(char *str)
5690 int ints[MAX_DOMAIN_DISTANCE+1], i;
5692 str = get_options(str, ARRAY_SIZE(ints), ints);
5694 printk("#ints: %d\n", ints[0]);
5695 for (i = 1; i <= ints[0]; i++) {
5696 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5697 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5702 __setup ("migration_cost=", migration_cost_setup);
5705 * Global multiplier (divisor) for migration-cutoff values,
5706 * in percentiles. E.g. use a value of 150 to get 1.5 times
5707 * longer cache-hot cutoff times.
5709 * (We scale it from 100 to 128 to long long handling easier.)
5712 #define MIGRATION_FACTOR_SCALE 128
5714 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5716 static int __init setup_migration_factor(char *str)
5718 get_option(&str, &migration_factor);
5719 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5723 __setup("migration_factor=", setup_migration_factor);
5726 * Estimated distance of two CPUs, measured via the number of domains
5727 * we have to pass for the two CPUs to be in the same span:
5729 static unsigned long domain_distance(int cpu1, int cpu2)
5731 unsigned long distance = 0;
5732 struct sched_domain *sd;
5734 for_each_domain(cpu1, sd) {
5735 WARN_ON(!cpu_isset(cpu1, sd->span));
5736 if (cpu_isset(cpu2, sd->span))
5740 if (distance >= MAX_DOMAIN_DISTANCE) {
5742 distance = MAX_DOMAIN_DISTANCE-1;
5748 static unsigned int migration_debug;
5750 static int __init setup_migration_debug(char *str)
5752 get_option(&str, &migration_debug);
5756 __setup("migration_debug=", setup_migration_debug);
5759 * Maximum cache-size that the scheduler should try to measure.
5760 * Architectures with larger caches should tune this up during
5761 * bootup. Gets used in the domain-setup code (i.e. during SMP
5764 unsigned int max_cache_size;
5766 static int __init setup_max_cache_size(char *str)
5768 get_option(&str, &max_cache_size);
5772 __setup("max_cache_size=", setup_max_cache_size);
5775 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5776 * is the operation that is timed, so we try to generate unpredictable
5777 * cachemisses that still end up filling the L2 cache:
5779 static void touch_cache(void *__cache, unsigned long __size)
5781 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5783 unsigned long *cache = __cache;
5786 for (i = 0; i < size/6; i += 8) {
5789 case 1: cache[size-1-i]++;
5790 case 2: cache[chunk1-i]++;
5791 case 3: cache[chunk1+i]++;
5792 case 4: cache[chunk2-i]++;
5793 case 5: cache[chunk2+i]++;
5799 * Measure the cache-cost of one task migration. Returns in units of nsec.
5801 static unsigned long long
5802 measure_one(void *cache, unsigned long size, int source, int target)
5804 cpumask_t mask, saved_mask;
5805 unsigned long long t0, t1, t2, t3, cost;
5807 saved_mask = current->cpus_allowed;
5810 * Flush source caches to RAM and invalidate them:
5815 * Migrate to the source CPU:
5817 mask = cpumask_of_cpu(source);
5818 set_cpus_allowed(current, mask);
5819 WARN_ON(smp_processor_id() != source);
5822 * Dirty the working set:
5825 touch_cache(cache, size);
5829 * Migrate to the target CPU, dirty the L2 cache and access
5830 * the shared buffer. (which represents the working set
5831 * of a migrated task.)
5833 mask = cpumask_of_cpu(target);
5834 set_cpus_allowed(current, mask);
5835 WARN_ON(smp_processor_id() != target);
5838 touch_cache(cache, size);
5841 cost = t1-t0 + t3-t2;
5843 if (migration_debug >= 2)
5844 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5845 source, target, t1-t0, t1-t0, t3-t2, cost);
5847 * Flush target caches to RAM and invalidate them:
5851 set_cpus_allowed(current, saved_mask);
5857 * Measure a series of task migrations and return the average
5858 * result. Since this code runs early during bootup the system
5859 * is 'undisturbed' and the average latency makes sense.
5861 * The algorithm in essence auto-detects the relevant cache-size,
5862 * so it will properly detect different cachesizes for different
5863 * cache-hierarchies, depending on how the CPUs are connected.
5865 * Architectures can prime the upper limit of the search range via
5866 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5868 static unsigned long long
5869 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5871 unsigned long long cost1, cost2;
5875 * Measure the migration cost of 'size' bytes, over an
5876 * average of 10 runs:
5878 * (We perturb the cache size by a small (0..4k)
5879 * value to compensate size/alignment related artifacts.
5880 * We also subtract the cost of the operation done on
5886 * dry run, to make sure we start off cache-cold on cpu1,
5887 * and to get any vmalloc pagefaults in advance:
5889 measure_one(cache, size, cpu1, cpu2);
5890 for (i = 0; i < ITERATIONS; i++)
5891 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5893 measure_one(cache, size, cpu2, cpu1);
5894 for (i = 0; i < ITERATIONS; i++)
5895 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5898 * (We measure the non-migrating [cached] cost on both
5899 * cpu1 and cpu2, to handle CPUs with different speeds)
5903 measure_one(cache, size, cpu1, cpu1);
5904 for (i = 0; i < ITERATIONS; i++)
5905 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5907 measure_one(cache, size, cpu2, cpu2);
5908 for (i = 0; i < ITERATIONS; i++)
5909 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5912 * Get the per-iteration migration cost:
5914 do_div(cost1, 2*ITERATIONS);
5915 do_div(cost2, 2*ITERATIONS);
5917 return cost1 - cost2;
5920 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5922 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5923 unsigned int max_size, size, size_found = 0;
5924 long long cost = 0, prev_cost;
5928 * Search from max_cache_size*5 down to 64K - the real relevant
5929 * cachesize has to lie somewhere inbetween.
5931 if (max_cache_size) {
5932 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5933 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5936 * Since we have no estimation about the relevant
5939 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5940 size = MIN_CACHE_SIZE;
5943 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5944 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5949 * Allocate the working set:
5951 cache = vmalloc(max_size);
5953 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5954 return 1000000; /* return 1 msec on very small boxen */
5957 while (size <= max_size) {
5959 cost = measure_cost(cpu1, cpu2, cache, size);
5965 if (max_cost < cost) {
5971 * Calculate average fluctuation, we use this to prevent
5972 * noise from triggering an early break out of the loop:
5974 fluct = abs(cost - prev_cost);
5975 avg_fluct = (avg_fluct + fluct)/2;
5977 if (migration_debug)
5978 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5980 (long)cost / 1000000,
5981 ((long)cost / 100000) % 10,
5982 (long)max_cost / 1000000,
5983 ((long)max_cost / 100000) % 10,
5984 domain_distance(cpu1, cpu2),
5988 * If we iterated at least 20% past the previous maximum,
5989 * and the cost has dropped by more than 20% already,
5990 * (taking fluctuations into account) then we assume to
5991 * have found the maximum and break out of the loop early:
5993 if (size_found && (size*100 > size_found*SIZE_THRESH))
5994 if (cost+avg_fluct <= 0 ||
5995 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5997 if (migration_debug)
5998 printk("-> found max.\n");
6002 * Increase the cachesize in 10% steps:
6004 size = size * 10 / 9;
6007 if (migration_debug)
6008 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6009 cpu1, cpu2, size_found, max_cost);
6014 * A task is considered 'cache cold' if at least 2 times
6015 * the worst-case cost of migration has passed.
6017 * (this limit is only listened to if the load-balancing
6018 * situation is 'nice' - if there is a large imbalance we
6019 * ignore it for the sake of CPU utilization and
6020 * processing fairness.)
6022 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6025 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6027 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6028 unsigned long j0, j1, distance, max_distance = 0;
6029 struct sched_domain *sd;
6034 * First pass - calculate the cacheflush times:
6036 for_each_cpu_mask(cpu1, *cpu_map) {
6037 for_each_cpu_mask(cpu2, *cpu_map) {
6040 distance = domain_distance(cpu1, cpu2);
6041 max_distance = max(max_distance, distance);
6043 * No result cached yet?
6045 if (migration_cost[distance] == -1LL)
6046 migration_cost[distance] =
6047 measure_migration_cost(cpu1, cpu2);
6051 * Second pass - update the sched domain hierarchy with
6052 * the new cache-hot-time estimations:
6054 for_each_cpu_mask(cpu, *cpu_map) {
6056 for_each_domain(cpu, sd) {
6057 sd->cache_hot_time = migration_cost[distance];
6064 if (migration_debug)
6065 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6073 if (system_state == SYSTEM_BOOTING) {
6074 if (num_online_cpus() > 1) {
6075 printk("migration_cost=");
6076 for (distance = 0; distance <= max_distance; distance++) {
6079 printk("%ld", (long)migration_cost[distance] / 1000);
6085 if (migration_debug)
6086 printk("migration: %ld seconds\n", (j1-j0)/HZ);
6089 * Move back to the original CPU. NUMA-Q gets confused
6090 * if we migrate to another quad during bootup.
6092 if (raw_smp_processor_id() != orig_cpu) {
6093 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6094 saved_mask = current->cpus_allowed;
6096 set_cpus_allowed(current, mask);
6097 set_cpus_allowed(current, saved_mask);
6104 * find_next_best_node - find the next node to include in a sched_domain
6105 * @node: node whose sched_domain we're building
6106 * @used_nodes: nodes already in the sched_domain
6108 * Find the next node to include in a given scheduling domain. Simply
6109 * finds the closest node not already in the @used_nodes map.
6111 * Should use nodemask_t.
6113 static int find_next_best_node(int node, unsigned long *used_nodes)
6115 int i, n, val, min_val, best_node = 0;
6119 for (i = 0; i < MAX_NUMNODES; i++) {
6120 /* Start at @node */
6121 n = (node + i) % MAX_NUMNODES;
6123 if (!nr_cpus_node(n))
6126 /* Skip already used nodes */
6127 if (test_bit(n, used_nodes))
6130 /* Simple min distance search */
6131 val = node_distance(node, n);
6133 if (val < min_val) {
6139 set_bit(best_node, used_nodes);
6144 * sched_domain_node_span - get a cpumask for a node's sched_domain
6145 * @node: node whose cpumask we're constructing
6146 * @size: number of nodes to include in this span
6148 * Given a node, construct a good cpumask for its sched_domain to span. It
6149 * should be one that prevents unnecessary balancing, but also spreads tasks
6152 static cpumask_t sched_domain_node_span(int node)
6154 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6155 cpumask_t span, nodemask;
6159 bitmap_zero(used_nodes, MAX_NUMNODES);
6161 nodemask = node_to_cpumask(node);
6162 cpus_or(span, span, nodemask);
6163 set_bit(node, used_nodes);
6165 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6166 int next_node = find_next_best_node(node, used_nodes);
6168 nodemask = node_to_cpumask(next_node);
6169 cpus_or(span, span, nodemask);
6176 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6179 * SMT sched-domains:
6181 #ifdef CONFIG_SCHED_SMT
6182 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6183 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6185 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6186 struct sched_group **sg)
6189 *sg = &per_cpu(sched_group_cpus, cpu);
6195 * multi-core sched-domains:
6197 #ifdef CONFIG_SCHED_MC
6198 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6199 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6202 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6203 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6204 struct sched_group **sg)
6207 cpumask_t mask = cpu_sibling_map[cpu];
6208 cpus_and(mask, mask, *cpu_map);
6209 group = first_cpu(mask);
6211 *sg = &per_cpu(sched_group_core, group);
6214 #elif defined(CONFIG_SCHED_MC)
6215 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6216 struct sched_group **sg)
6219 *sg = &per_cpu(sched_group_core, cpu);
6224 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6225 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6227 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6228 struct sched_group **sg)
6231 #ifdef CONFIG_SCHED_MC
6232 cpumask_t mask = cpu_coregroup_map(cpu);
6233 cpus_and(mask, mask, *cpu_map);
6234 group = first_cpu(mask);
6235 #elif defined(CONFIG_SCHED_SMT)
6236 cpumask_t mask = cpu_sibling_map[cpu];
6237 cpus_and(mask, mask, *cpu_map);
6238 group = first_cpu(mask);
6243 *sg = &per_cpu(sched_group_phys, group);
6249 * The init_sched_build_groups can't handle what we want to do with node
6250 * groups, so roll our own. Now each node has its own list of groups which
6251 * gets dynamically allocated.
6253 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6254 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6256 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6257 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6259 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6260 struct sched_group **sg)
6262 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6265 cpus_and(nodemask, nodemask, *cpu_map);
6266 group = first_cpu(nodemask);
6269 *sg = &per_cpu(sched_group_allnodes, group);
6273 static void init_numa_sched_groups_power(struct sched_group *group_head)
6275 struct sched_group *sg = group_head;
6281 for_each_cpu_mask(j, sg->cpumask) {
6282 struct sched_domain *sd;
6284 sd = &per_cpu(phys_domains, j);
6285 if (j != first_cpu(sd->groups->cpumask)) {
6287 * Only add "power" once for each
6293 sg->cpu_power += sd->groups->cpu_power;
6296 if (sg != group_head)
6302 /* Free memory allocated for various sched_group structures */
6303 static void free_sched_groups(const cpumask_t *cpu_map)
6307 for_each_cpu_mask(cpu, *cpu_map) {
6308 struct sched_group **sched_group_nodes
6309 = sched_group_nodes_bycpu[cpu];
6311 if (!sched_group_nodes)
6314 for (i = 0; i < MAX_NUMNODES; i++) {
6315 cpumask_t nodemask = node_to_cpumask(i);
6316 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6318 cpus_and(nodemask, nodemask, *cpu_map);
6319 if (cpus_empty(nodemask))
6329 if (oldsg != sched_group_nodes[i])
6332 kfree(sched_group_nodes);
6333 sched_group_nodes_bycpu[cpu] = NULL;
6337 static void free_sched_groups(const cpumask_t *cpu_map)
6343 * Initialize sched groups cpu_power.
6345 * cpu_power indicates the capacity of sched group, which is used while
6346 * distributing the load between different sched groups in a sched domain.
6347 * Typically cpu_power for all the groups in a sched domain will be same unless
6348 * there are asymmetries in the topology. If there are asymmetries, group
6349 * having more cpu_power will pickup more load compared to the group having
6352 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6353 * the maximum number of tasks a group can handle in the presence of other idle
6354 * or lightly loaded groups in the same sched domain.
6356 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6358 struct sched_domain *child;
6359 struct sched_group *group;
6361 WARN_ON(!sd || !sd->groups);
6363 if (cpu != first_cpu(sd->groups->cpumask))
6369 * For perf policy, if the groups in child domain share resources
6370 * (for example cores sharing some portions of the cache hierarchy
6371 * or SMT), then set this domain groups cpu_power such that each group
6372 * can handle only one task, when there are other idle groups in the
6373 * same sched domain.
6375 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6377 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6378 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6382 sd->groups->cpu_power = 0;
6385 * add cpu_power of each child group to this groups cpu_power
6387 group = child->groups;
6389 sd->groups->cpu_power += group->cpu_power;
6390 group = group->next;
6391 } while (group != child->groups);
6395 * Build sched domains for a given set of cpus and attach the sched domains
6396 * to the individual cpus
6398 static int build_sched_domains(const cpumask_t *cpu_map)
6401 struct sched_domain *sd;
6403 struct sched_group **sched_group_nodes = NULL;
6404 int sd_allnodes = 0;
6407 * Allocate the per-node list of sched groups
6409 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6411 if (!sched_group_nodes) {
6412 printk(KERN_WARNING "Can not alloc sched group node list\n");
6415 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6419 * Set up domains for cpus specified by the cpu_map.
6421 for_each_cpu_mask(i, *cpu_map) {
6422 struct sched_domain *sd = NULL, *p;
6423 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6425 cpus_and(nodemask, nodemask, *cpu_map);
6428 if (cpus_weight(*cpu_map)
6429 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6430 sd = &per_cpu(allnodes_domains, i);
6431 *sd = SD_ALLNODES_INIT;
6432 sd->span = *cpu_map;
6433 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6439 sd = &per_cpu(node_domains, i);
6441 sd->span = sched_domain_node_span(cpu_to_node(i));
6445 cpus_and(sd->span, sd->span, *cpu_map);
6449 sd = &per_cpu(phys_domains, i);
6451 sd->span = nodemask;
6455 cpu_to_phys_group(i, cpu_map, &sd->groups);
6457 #ifdef CONFIG_SCHED_MC
6459 sd = &per_cpu(core_domains, i);
6461 sd->span = cpu_coregroup_map(i);
6462 cpus_and(sd->span, sd->span, *cpu_map);
6465 cpu_to_core_group(i, cpu_map, &sd->groups);
6468 #ifdef CONFIG_SCHED_SMT
6470 sd = &per_cpu(cpu_domains, i);
6471 *sd = SD_SIBLING_INIT;
6472 sd->span = cpu_sibling_map[i];
6473 cpus_and(sd->span, sd->span, *cpu_map);
6476 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6480 #ifdef CONFIG_SCHED_SMT
6481 /* Set up CPU (sibling) groups */
6482 for_each_cpu_mask(i, *cpu_map) {
6483 cpumask_t this_sibling_map = cpu_sibling_map[i];
6484 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6485 if (i != first_cpu(this_sibling_map))
6488 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6492 #ifdef CONFIG_SCHED_MC
6493 /* Set up multi-core groups */
6494 for_each_cpu_mask(i, *cpu_map) {
6495 cpumask_t this_core_map = cpu_coregroup_map(i);
6496 cpus_and(this_core_map, this_core_map, *cpu_map);
6497 if (i != first_cpu(this_core_map))
6499 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6504 /* Set up physical groups */
6505 for (i = 0; i < MAX_NUMNODES; i++) {
6506 cpumask_t nodemask = node_to_cpumask(i);
6508 cpus_and(nodemask, nodemask, *cpu_map);
6509 if (cpus_empty(nodemask))
6512 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6516 /* Set up node groups */
6518 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6520 for (i = 0; i < MAX_NUMNODES; i++) {
6521 /* Set up node groups */
6522 struct sched_group *sg, *prev;
6523 cpumask_t nodemask = node_to_cpumask(i);
6524 cpumask_t domainspan;
6525 cpumask_t covered = CPU_MASK_NONE;
6528 cpus_and(nodemask, nodemask, *cpu_map);
6529 if (cpus_empty(nodemask)) {
6530 sched_group_nodes[i] = NULL;
6534 domainspan = sched_domain_node_span(i);
6535 cpus_and(domainspan, domainspan, *cpu_map);
6537 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6539 printk(KERN_WARNING "Can not alloc domain group for "
6543 sched_group_nodes[i] = sg;
6544 for_each_cpu_mask(j, nodemask) {
6545 struct sched_domain *sd;
6546 sd = &per_cpu(node_domains, j);
6550 sg->cpumask = nodemask;
6552 cpus_or(covered, covered, nodemask);
6555 for (j = 0; j < MAX_NUMNODES; j++) {
6556 cpumask_t tmp, notcovered;
6557 int n = (i + j) % MAX_NUMNODES;
6559 cpus_complement(notcovered, covered);
6560 cpus_and(tmp, notcovered, *cpu_map);
6561 cpus_and(tmp, tmp, domainspan);
6562 if (cpus_empty(tmp))
6565 nodemask = node_to_cpumask(n);
6566 cpus_and(tmp, tmp, nodemask);
6567 if (cpus_empty(tmp))
6570 sg = kmalloc_node(sizeof(struct sched_group),
6574 "Can not alloc domain group for node %d\n", j);
6579 sg->next = prev->next;
6580 cpus_or(covered, covered, tmp);
6587 /* Calculate CPU power for physical packages and nodes */
6588 #ifdef CONFIG_SCHED_SMT
6589 for_each_cpu_mask(i, *cpu_map) {
6590 sd = &per_cpu(cpu_domains, i);
6591 init_sched_groups_power(i, sd);
6594 #ifdef CONFIG_SCHED_MC
6595 for_each_cpu_mask(i, *cpu_map) {
6596 sd = &per_cpu(core_domains, i);
6597 init_sched_groups_power(i, sd);
6601 for_each_cpu_mask(i, *cpu_map) {
6602 sd = &per_cpu(phys_domains, i);
6603 init_sched_groups_power(i, sd);
6607 for (i = 0; i < MAX_NUMNODES; i++)
6608 init_numa_sched_groups_power(sched_group_nodes[i]);
6611 struct sched_group *sg;
6613 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6614 init_numa_sched_groups_power(sg);
6618 /* Attach the domains */
6619 for_each_cpu_mask(i, *cpu_map) {
6620 struct sched_domain *sd;
6621 #ifdef CONFIG_SCHED_SMT
6622 sd = &per_cpu(cpu_domains, i);
6623 #elif defined(CONFIG_SCHED_MC)
6624 sd = &per_cpu(core_domains, i);
6626 sd = &per_cpu(phys_domains, i);
6628 cpu_attach_domain(sd, i);
6631 * Tune cache-hot values:
6633 calibrate_migration_costs(cpu_map);
6639 free_sched_groups(cpu_map);
6644 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6646 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6648 cpumask_t cpu_default_map;
6652 * Setup mask for cpus without special case scheduling requirements.
6653 * For now this just excludes isolated cpus, but could be used to
6654 * exclude other special cases in the future.
6656 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6658 err = build_sched_domains(&cpu_default_map);
6663 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6665 free_sched_groups(cpu_map);
6669 * Detach sched domains from a group of cpus specified in cpu_map
6670 * These cpus will now be attached to the NULL domain
6672 static void detach_destroy_domains(const cpumask_t *cpu_map)
6676 for_each_cpu_mask(i, *cpu_map)
6677 cpu_attach_domain(NULL, i);
6678 synchronize_sched();
6679 arch_destroy_sched_domains(cpu_map);
6683 * Partition sched domains as specified by the cpumasks below.
6684 * This attaches all cpus from the cpumasks to the NULL domain,
6685 * waits for a RCU quiescent period, recalculates sched
6686 * domain information and then attaches them back to the
6687 * correct sched domains
6688 * Call with hotplug lock held
6690 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6692 cpumask_t change_map;
6695 cpus_and(*partition1, *partition1, cpu_online_map);
6696 cpus_and(*partition2, *partition2, cpu_online_map);
6697 cpus_or(change_map, *partition1, *partition2);
6699 /* Detach sched domains from all of the affected cpus */
6700 detach_destroy_domains(&change_map);
6701 if (!cpus_empty(*partition1))
6702 err = build_sched_domains(partition1);
6703 if (!err && !cpus_empty(*partition2))
6704 err = build_sched_domains(partition2);
6709 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6710 int arch_reinit_sched_domains(void)
6715 detach_destroy_domains(&cpu_online_map);
6716 err = arch_init_sched_domains(&cpu_online_map);
6717 unlock_cpu_hotplug();
6722 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6726 if (buf[0] != '0' && buf[0] != '1')
6730 sched_smt_power_savings = (buf[0] == '1');
6732 sched_mc_power_savings = (buf[0] == '1');
6734 ret = arch_reinit_sched_domains();
6736 return ret ? ret : count;
6739 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6743 #ifdef CONFIG_SCHED_SMT
6745 err = sysfs_create_file(&cls->kset.kobj,
6746 &attr_sched_smt_power_savings.attr);
6748 #ifdef CONFIG_SCHED_MC
6749 if (!err && mc_capable())
6750 err = sysfs_create_file(&cls->kset.kobj,
6751 &attr_sched_mc_power_savings.attr);
6757 #ifdef CONFIG_SCHED_MC
6758 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6760 return sprintf(page, "%u\n", sched_mc_power_savings);
6762 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6763 const char *buf, size_t count)
6765 return sched_power_savings_store(buf, count, 0);
6767 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6768 sched_mc_power_savings_store);
6771 #ifdef CONFIG_SCHED_SMT
6772 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6774 return sprintf(page, "%u\n", sched_smt_power_savings);
6776 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6777 const char *buf, size_t count)
6779 return sched_power_savings_store(buf, count, 1);
6781 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6782 sched_smt_power_savings_store);
6786 * Force a reinitialization of the sched domains hierarchy. The domains
6787 * and groups cannot be updated in place without racing with the balancing
6788 * code, so we temporarily attach all running cpus to the NULL domain
6789 * which will prevent rebalancing while the sched domains are recalculated.
6791 static int update_sched_domains(struct notifier_block *nfb,
6792 unsigned long action, void *hcpu)
6795 case CPU_UP_PREPARE:
6796 case CPU_DOWN_PREPARE:
6797 detach_destroy_domains(&cpu_online_map);
6800 case CPU_UP_CANCELED:
6801 case CPU_DOWN_FAILED:
6805 * Fall through and re-initialise the domains.
6812 /* The hotplug lock is already held by cpu_up/cpu_down */
6813 arch_init_sched_domains(&cpu_online_map);
6818 void __init sched_init_smp(void)
6820 cpumask_t non_isolated_cpus;
6823 arch_init_sched_domains(&cpu_online_map);
6824 cpus_andnot(non_isolated_cpus, cpu_online_map, cpu_isolated_map);
6825 if (cpus_empty(non_isolated_cpus))
6826 cpu_set(smp_processor_id(), non_isolated_cpus);
6827 unlock_cpu_hotplug();
6828 /* XXX: Theoretical race here - CPU may be hotplugged now */
6829 hotcpu_notifier(update_sched_domains, 0);
6831 /* Move init over to a non-isolated CPU */
6832 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6836 void __init sched_init_smp(void)
6839 #endif /* CONFIG_SMP */
6841 int in_sched_functions(unsigned long addr)
6843 /* Linker adds these: start and end of __sched functions */
6844 extern char __sched_text_start[], __sched_text_end[];
6846 return in_lock_functions(addr) ||
6847 (addr >= (unsigned long)__sched_text_start
6848 && addr < (unsigned long)__sched_text_end);
6851 void __init sched_init(void)
6855 for_each_possible_cpu(i) {
6856 struct prio_array *array;
6860 spin_lock_init(&rq->lock);
6861 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6863 rq->active = rq->arrays;
6864 rq->expired = rq->arrays + 1;
6865 rq->best_expired_prio = MAX_PRIO;
6869 for (j = 1; j < 3; j++)
6870 rq->cpu_load[j] = 0;
6871 rq->active_balance = 0;
6874 rq->migration_thread = NULL;
6875 INIT_LIST_HEAD(&rq->migration_queue);
6877 atomic_set(&rq->nr_iowait, 0);
6879 for (j = 0; j < 2; j++) {
6880 array = rq->arrays + j;
6881 for (k = 0; k < MAX_PRIO; k++) {
6882 INIT_LIST_HEAD(array->queue + k);
6883 __clear_bit(k, array->bitmap);
6885 // delimiter for bitsearch
6886 __set_bit(MAX_PRIO, array->bitmap);
6890 set_load_weight(&init_task);
6893 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6896 #ifdef CONFIG_RT_MUTEXES
6897 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6901 * The boot idle thread does lazy MMU switching as well:
6903 atomic_inc(&init_mm.mm_count);
6904 enter_lazy_tlb(&init_mm, current);
6907 * Make us the idle thread. Technically, schedule() should not be
6908 * called from this thread, however somewhere below it might be,
6909 * but because we are the idle thread, we just pick up running again
6910 * when this runqueue becomes "idle".
6912 init_idle(current, smp_processor_id());
6915 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6916 void __might_sleep(char *file, int line)
6919 static unsigned long prev_jiffy; /* ratelimiting */
6921 if ((in_atomic() || irqs_disabled()) &&
6922 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6923 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6925 prev_jiffy = jiffies;
6926 printk(KERN_ERR "BUG: sleeping function called from invalid"
6927 " context at %s:%d\n", file, line);
6928 printk("in_atomic():%d, irqs_disabled():%d\n",
6929 in_atomic(), irqs_disabled());
6930 debug_show_held_locks(current);
6935 EXPORT_SYMBOL(__might_sleep);
6938 #ifdef CONFIG_MAGIC_SYSRQ
6939 void normalize_rt_tasks(void)
6941 struct prio_array *array;
6942 struct task_struct *p;
6943 unsigned long flags;
6946 read_lock_irq(&tasklist_lock);
6947 for_each_process(p) {
6951 spin_lock_irqsave(&p->pi_lock, flags);
6952 rq = __task_rq_lock(p);
6956 deactivate_task(p, task_rq(p));
6957 __setscheduler(p, SCHED_NORMAL, 0);
6959 __activate_task(p, task_rq(p));
6960 resched_task(rq->curr);
6963 __task_rq_unlock(rq);
6964 spin_unlock_irqrestore(&p->pi_lock, flags);
6966 read_unlock_irq(&tasklist_lock);
6969 #endif /* CONFIG_MAGIC_SYSRQ */
6973 * These functions are only useful for the IA64 MCA handling.
6975 * They can only be called when the whole system has been
6976 * stopped - every CPU needs to be quiescent, and no scheduling
6977 * activity can take place. Using them for anything else would
6978 * be a serious bug, and as a result, they aren't even visible
6979 * under any other configuration.
6983 * curr_task - return the current task for a given cpu.
6984 * @cpu: the processor in question.
6986 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6988 struct task_struct *curr_task(int cpu)
6990 return cpu_curr(cpu);
6994 * set_curr_task - set the current task for a given cpu.
6995 * @cpu: the processor in question.
6996 * @p: the task pointer to set.
6998 * Description: This function must only be used when non-maskable interrupts
6999 * are serviced on a separate stack. It allows the architecture to switch the
7000 * notion of the current task on a cpu in a non-blocking manner. This function
7001 * must be called with all CPU's synchronized, and interrupts disabled, the
7002 * and caller must save the original value of the current task (see
7003 * curr_task() above) and restore that value before reenabling interrupts and
7004 * re-starting the system.
7006 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7008 void set_curr_task(int cpu, struct task_struct *p)