4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
95 * These are the 'tuning knobs' of the scheduler:
97 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
98 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
99 * Timeslices get refilled after they expire.
101 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
102 #define DEF_TIMESLICE (100 * HZ / 1000)
103 #define ON_RUNQUEUE_WEIGHT 30
104 #define CHILD_PENALTY 95
105 #define PARENT_PENALTY 100
106 #define EXIT_WEIGHT 3
107 #define PRIO_BONUS_RATIO 25
108 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
109 #define INTERACTIVE_DELTA 2
110 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
111 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
112 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
115 * If a task is 'interactive' then we reinsert it in the active
116 * array after it has expired its current timeslice. (it will not
117 * continue to run immediately, it will still roundrobin with
118 * other interactive tasks.)
120 * This part scales the interactivity limit depending on niceness.
122 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
123 * Here are a few examples of different nice levels:
125 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
126 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
127 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
129 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
131 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
132 * priority range a task can explore, a value of '1' means the
133 * task is rated interactive.)
135 * Ie. nice +19 tasks can never get 'interactive' enough to be
136 * reinserted into the active array. And only heavily CPU-hog nice -20
137 * tasks will be expired. Default nice 0 tasks are somewhere between,
138 * it takes some effort for them to get interactive, but it's not
142 #define CURRENT_BONUS(p) \
143 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
146 #define GRANULARITY (10 * HZ / 1000 ? : 1)
149 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
150 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
153 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
154 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
157 #define SCALE(v1,v1_max,v2_max) \
158 (v1) * (v2_max) / (v1_max)
161 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
164 #define TASK_INTERACTIVE(p) \
165 ((p)->prio <= (p)->static_prio - DELTA(p))
167 #define INTERACTIVE_SLEEP(p) \
168 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
169 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
171 #define TASK_PREEMPTS_CURR(p, rq) \
172 ((p)->prio < (rq)->curr->prio)
174 #define SCALE_PRIO(x, prio) \
175 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
177 static unsigned int static_prio_timeslice(int static_prio)
179 if (static_prio < NICE_TO_PRIO(0))
180 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
182 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
187 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
188 * Since cpu_power is a 'constant', we can use a reciprocal divide.
190 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
192 return reciprocal_divide(load, sg->reciprocal_cpu_power);
196 * Each time a sched group cpu_power is changed,
197 * we must compute its reciprocal value
199 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
201 sg->__cpu_power += val;
202 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
207 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
208 * to time slice values: [800ms ... 100ms ... 5ms]
210 * The higher a thread's priority, the bigger timeslices
211 * it gets during one round of execution. But even the lowest
212 * priority thread gets MIN_TIMESLICE worth of execution time.
215 static inline unsigned int task_timeslice(struct task_struct *p)
217 return static_prio_timeslice(p->static_prio);
221 * These are the runqueue data structures:
225 unsigned int nr_active;
226 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
227 struct list_head queue[MAX_PRIO];
231 * This is the main, per-CPU runqueue data structure.
233 * Locking rule: those places that want to lock multiple runqueues
234 * (such as the load balancing or the thread migration code), lock
235 * acquire operations must be ordered by ascending &runqueue.
241 * nr_running and cpu_load should be in the same cacheline because
242 * remote CPUs use both these fields when doing load calculation.
244 unsigned long nr_running;
245 unsigned long raw_weighted_load;
247 unsigned long cpu_load[3];
248 unsigned char idle_at_tick;
250 unsigned char in_nohz_recently;
253 unsigned long long nr_switches;
256 * This is part of a global counter where only the total sum
257 * over all CPUs matters. A task can increase this counter on
258 * one CPU and if it got migrated afterwards it may decrease
259 * it on another CPU. Always updated under the runqueue lock:
261 unsigned long nr_uninterruptible;
263 unsigned long expired_timestamp;
264 /* Cached timestamp set by update_cpu_clock() */
265 unsigned long long most_recent_timestamp;
266 struct task_struct *curr, *idle;
267 unsigned long next_balance;
268 struct mm_struct *prev_mm;
269 struct prio_array *active, *expired, arrays[2];
270 int best_expired_prio;
274 struct sched_domain *sd;
276 /* For active balancing */
279 int cpu; /* cpu of this runqueue */
281 struct task_struct *migration_thread;
282 struct list_head migration_queue;
285 #ifdef CONFIG_SCHEDSTATS
287 struct sched_info rq_sched_info;
289 /* sys_sched_yield() stats */
290 unsigned long yld_exp_empty;
291 unsigned long yld_act_empty;
292 unsigned long yld_both_empty;
293 unsigned long yld_cnt;
295 /* schedule() stats */
296 unsigned long sched_switch;
297 unsigned long sched_cnt;
298 unsigned long sched_goidle;
300 /* try_to_wake_up() stats */
301 unsigned long ttwu_cnt;
302 unsigned long ttwu_local;
304 struct lock_class_key rq_lock_key;
307 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
308 static DEFINE_MUTEX(sched_hotcpu_mutex);
310 static inline int cpu_of(struct rq *rq)
320 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
321 * See detach_destroy_domains: synchronize_sched for details.
323 * The domain tree of any CPU may only be accessed from within
324 * preempt-disabled sections.
326 #define for_each_domain(cpu, __sd) \
327 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
329 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
330 #define this_rq() (&__get_cpu_var(runqueues))
331 #define task_rq(p) cpu_rq(task_cpu(p))
332 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
334 #ifndef prepare_arch_switch
335 # define prepare_arch_switch(next) do { } while (0)
337 #ifndef finish_arch_switch
338 # define finish_arch_switch(prev) do { } while (0)
341 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
342 static inline int task_running(struct rq *rq, struct task_struct *p)
344 return rq->curr == p;
347 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
351 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
353 #ifdef CONFIG_DEBUG_SPINLOCK
354 /* this is a valid case when another task releases the spinlock */
355 rq->lock.owner = current;
358 * If we are tracking spinlock dependencies then we have to
359 * fix up the runqueue lock - which gets 'carried over' from
362 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
364 spin_unlock_irq(&rq->lock);
367 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
368 static inline int task_running(struct rq *rq, struct task_struct *p)
373 return rq->curr == p;
377 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
381 * We can optimise this out completely for !SMP, because the
382 * SMP rebalancing from interrupt is the only thing that cares
387 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
388 spin_unlock_irq(&rq->lock);
390 spin_unlock(&rq->lock);
394 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
398 * After ->oncpu is cleared, the task can be moved to a different CPU.
399 * We must ensure this doesn't happen until the switch is completely
405 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
409 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
412 * __task_rq_lock - lock the runqueue a given task resides on.
413 * Must be called interrupts disabled.
415 static inline struct rq *__task_rq_lock(struct task_struct *p)
422 spin_lock(&rq->lock);
423 if (unlikely(rq != task_rq(p))) {
424 spin_unlock(&rq->lock);
425 goto repeat_lock_task;
431 * task_rq_lock - lock the runqueue a given task resides on and disable
432 * interrupts. Note the ordering: we can safely lookup the task_rq without
433 * explicitly disabling preemption.
435 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
441 local_irq_save(*flags);
443 spin_lock(&rq->lock);
444 if (unlikely(rq != task_rq(p))) {
445 spin_unlock_irqrestore(&rq->lock, *flags);
446 goto repeat_lock_task;
451 static inline void __task_rq_unlock(struct rq *rq)
454 spin_unlock(&rq->lock);
457 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
460 spin_unlock_irqrestore(&rq->lock, *flags);
463 #ifdef CONFIG_SCHEDSTATS
465 * bump this up when changing the output format or the meaning of an existing
466 * format, so that tools can adapt (or abort)
468 #define SCHEDSTAT_VERSION 14
470 static int show_schedstat(struct seq_file *seq, void *v)
474 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
475 seq_printf(seq, "timestamp %lu\n", jiffies);
476 for_each_online_cpu(cpu) {
477 struct rq *rq = cpu_rq(cpu);
479 struct sched_domain *sd;
483 /* runqueue-specific stats */
485 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
486 cpu, rq->yld_both_empty,
487 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
488 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
489 rq->ttwu_cnt, rq->ttwu_local,
490 rq->rq_sched_info.cpu_time,
491 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
493 seq_printf(seq, "\n");
496 /* domain-specific stats */
498 for_each_domain(cpu, sd) {
499 enum idle_type itype;
500 char mask_str[NR_CPUS];
502 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
503 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
504 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
506 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu "
509 sd->lb_balanced[itype],
510 sd->lb_failed[itype],
511 sd->lb_imbalance[itype],
512 sd->lb_gained[itype],
513 sd->lb_hot_gained[itype],
514 sd->lb_nobusyq[itype],
515 sd->lb_nobusyg[itype]);
517 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
519 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
520 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
521 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
522 sd->ttwu_wake_remote, sd->ttwu_move_affine,
523 sd->ttwu_move_balance);
531 static int schedstat_open(struct inode *inode, struct file *file)
533 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
534 char *buf = kmalloc(size, GFP_KERNEL);
540 res = single_open(file, show_schedstat, NULL);
542 m = file->private_data;
550 const struct file_operations proc_schedstat_operations = {
551 .open = schedstat_open,
554 .release = single_release,
558 * Expects runqueue lock to be held for atomicity of update
561 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
564 rq->rq_sched_info.run_delay += delta_jiffies;
565 rq->rq_sched_info.pcnt++;
570 * Expects runqueue lock to be held for atomicity of update
573 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
576 rq->rq_sched_info.cpu_time += delta_jiffies;
578 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
579 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
580 #else /* !CONFIG_SCHEDSTATS */
582 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
585 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
587 # define schedstat_inc(rq, field) do { } while (0)
588 # define schedstat_add(rq, field, amt) do { } while (0)
592 * this_rq_lock - lock this runqueue and disable interrupts.
594 static inline struct rq *this_rq_lock(void)
601 spin_lock(&rq->lock);
606 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
608 * Called when a process is dequeued from the active array and given
609 * the cpu. We should note that with the exception of interactive
610 * tasks, the expired queue will become the active queue after the active
611 * queue is empty, without explicitly dequeuing and requeuing tasks in the
612 * expired queue. (Interactive tasks may be requeued directly to the
613 * active queue, thus delaying tasks in the expired queue from running;
614 * see scheduler_tick()).
616 * This function is only called from sched_info_arrive(), rather than
617 * dequeue_task(). Even though a task may be queued and dequeued multiple
618 * times as it is shuffled about, we're really interested in knowing how
619 * long it was from the *first* time it was queued to the time that it
622 static inline void sched_info_dequeued(struct task_struct *t)
624 t->sched_info.last_queued = 0;
628 * Called when a task finally hits the cpu. We can now calculate how
629 * long it was waiting to run. We also note when it began so that we
630 * can keep stats on how long its timeslice is.
632 static void sched_info_arrive(struct task_struct *t)
634 unsigned long now = jiffies, delta_jiffies = 0;
636 if (t->sched_info.last_queued)
637 delta_jiffies = now - t->sched_info.last_queued;
638 sched_info_dequeued(t);
639 t->sched_info.run_delay += delta_jiffies;
640 t->sched_info.last_arrival = now;
641 t->sched_info.pcnt++;
643 rq_sched_info_arrive(task_rq(t), delta_jiffies);
647 * Called when a process is queued into either the active or expired
648 * array. The time is noted and later used to determine how long we
649 * had to wait for us to reach the cpu. Since the expired queue will
650 * become the active queue after active queue is empty, without dequeuing
651 * and requeuing any tasks, we are interested in queuing to either. It
652 * is unusual but not impossible for tasks to be dequeued and immediately
653 * requeued in the same or another array: this can happen in sched_yield(),
654 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
657 * This function is only called from enqueue_task(), but also only updates
658 * the timestamp if it is already not set. It's assumed that
659 * sched_info_dequeued() will clear that stamp when appropriate.
661 static inline void sched_info_queued(struct task_struct *t)
663 if (unlikely(sched_info_on()))
664 if (!t->sched_info.last_queued)
665 t->sched_info.last_queued = jiffies;
669 * Called when a process ceases being the active-running process, either
670 * voluntarily or involuntarily. Now we can calculate how long we ran.
672 static inline void sched_info_depart(struct task_struct *t)
674 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
676 t->sched_info.cpu_time += delta_jiffies;
677 rq_sched_info_depart(task_rq(t), delta_jiffies);
681 * Called when tasks are switched involuntarily due, typically, to expiring
682 * their time slice. (This may also be called when switching to or from
683 * the idle task.) We are only called when prev != next.
686 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
688 struct rq *rq = task_rq(prev);
691 * prev now departs the cpu. It's not interesting to record
692 * stats about how efficient we were at scheduling the idle
695 if (prev != rq->idle)
696 sched_info_depart(prev);
698 if (next != rq->idle)
699 sched_info_arrive(next);
702 sched_info_switch(struct task_struct *prev, struct task_struct *next)
704 if (unlikely(sched_info_on()))
705 __sched_info_switch(prev, next);
708 #define sched_info_queued(t) do { } while (0)
709 #define sched_info_switch(t, next) do { } while (0)
710 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
713 * Adding/removing a task to/from a priority array:
715 static void dequeue_task(struct task_struct *p, struct prio_array *array)
718 list_del(&p->run_list);
719 if (list_empty(array->queue + p->prio))
720 __clear_bit(p->prio, array->bitmap);
723 static void enqueue_task(struct task_struct *p, struct prio_array *array)
725 sched_info_queued(p);
726 list_add_tail(&p->run_list, array->queue + p->prio);
727 __set_bit(p->prio, array->bitmap);
733 * Put task to the end of the run list without the overhead of dequeue
734 * followed by enqueue.
736 static void requeue_task(struct task_struct *p, struct prio_array *array)
738 list_move_tail(&p->run_list, array->queue + p->prio);
742 enqueue_task_head(struct task_struct *p, struct prio_array *array)
744 list_add(&p->run_list, array->queue + p->prio);
745 __set_bit(p->prio, array->bitmap);
751 * __normal_prio - return the priority that is based on the static
752 * priority but is modified by bonuses/penalties.
754 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
755 * into the -5 ... 0 ... +5 bonus/penalty range.
757 * We use 25% of the full 0...39 priority range so that:
759 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
760 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
762 * Both properties are important to certain workloads.
765 static inline int __normal_prio(struct task_struct *p)
769 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
771 prio = p->static_prio - bonus;
772 if (prio < MAX_RT_PRIO)
774 if (prio > MAX_PRIO-1)
780 * To aid in avoiding the subversion of "niceness" due to uneven distribution
781 * of tasks with abnormal "nice" values across CPUs the contribution that
782 * each task makes to its run queue's load is weighted according to its
783 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
784 * scaled version of the new time slice allocation that they receive on time
789 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
790 * If static_prio_timeslice() is ever changed to break this assumption then
791 * this code will need modification
793 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
794 #define LOAD_WEIGHT(lp) \
795 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
796 #define PRIO_TO_LOAD_WEIGHT(prio) \
797 LOAD_WEIGHT(static_prio_timeslice(prio))
798 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
799 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
801 static void set_load_weight(struct task_struct *p)
803 if (has_rt_policy(p)) {
805 if (p == task_rq(p)->migration_thread)
807 * The migration thread does the actual balancing.
808 * Giving its load any weight will skew balancing
814 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
816 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
820 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
822 rq->raw_weighted_load += p->load_weight;
826 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
828 rq->raw_weighted_load -= p->load_weight;
831 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
834 inc_raw_weighted_load(rq, p);
837 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
840 dec_raw_weighted_load(rq, p);
844 * Calculate the expected normal priority: i.e. priority
845 * without taking RT-inheritance into account. Might be
846 * boosted by interactivity modifiers. Changes upon fork,
847 * setprio syscalls, and whenever the interactivity
848 * estimator recalculates.
850 static inline int normal_prio(struct task_struct *p)
854 if (has_rt_policy(p))
855 prio = MAX_RT_PRIO-1 - p->rt_priority;
857 prio = __normal_prio(p);
862 * Calculate the current priority, i.e. the priority
863 * taken into account by the scheduler. This value might
864 * be boosted by RT tasks, or might be boosted by
865 * interactivity modifiers. Will be RT if the task got
866 * RT-boosted. If not then it returns p->normal_prio.
868 static int effective_prio(struct task_struct *p)
870 p->normal_prio = normal_prio(p);
872 * If we are RT tasks or we were boosted to RT priority,
873 * keep the priority unchanged. Otherwise, update priority
874 * to the normal priority:
876 if (!rt_prio(p->prio))
877 return p->normal_prio;
882 * __activate_task - move a task to the runqueue.
884 static void __activate_task(struct task_struct *p, struct rq *rq)
886 struct prio_array *target = rq->active;
889 target = rq->expired;
890 enqueue_task(p, target);
891 inc_nr_running(p, rq);
895 * __activate_idle_task - move idle task to the _front_ of runqueue.
897 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
899 enqueue_task_head(p, rq->active);
900 inc_nr_running(p, rq);
904 * Recalculate p->normal_prio and p->prio after having slept,
905 * updating the sleep-average too:
907 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
909 /* Caller must always ensure 'now >= p->timestamp' */
910 unsigned long sleep_time = now - p->timestamp;
915 if (likely(sleep_time > 0)) {
917 * This ceiling is set to the lowest priority that would allow
918 * a task to be reinserted into the active array on timeslice
921 unsigned long ceiling = INTERACTIVE_SLEEP(p);
923 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
925 * Prevents user tasks from achieving best priority
926 * with one single large enough sleep.
928 p->sleep_avg = ceiling;
930 * Using INTERACTIVE_SLEEP() as a ceiling places a
931 * nice(0) task 1ms sleep away from promotion, and
932 * gives it 700ms to round-robin with no chance of
933 * being demoted. This is more than generous, so
934 * mark this sleep as non-interactive to prevent the
935 * on-runqueue bonus logic from intervening should
936 * this task not receive cpu immediately.
938 p->sleep_type = SLEEP_NONINTERACTIVE;
941 * Tasks waking from uninterruptible sleep are
942 * limited in their sleep_avg rise as they
943 * are likely to be waiting on I/O
945 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
946 if (p->sleep_avg >= ceiling)
948 else if (p->sleep_avg + sleep_time >=
950 p->sleep_avg = ceiling;
956 * This code gives a bonus to interactive tasks.
958 * The boost works by updating the 'average sleep time'
959 * value here, based on ->timestamp. The more time a
960 * task spends sleeping, the higher the average gets -
961 * and the higher the priority boost gets as well.
963 p->sleep_avg += sleep_time;
966 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
967 p->sleep_avg = NS_MAX_SLEEP_AVG;
970 return effective_prio(p);
974 * activate_task - move a task to the runqueue and do priority recalculation
976 * Update all the scheduling statistics stuff. (sleep average
977 * calculation, priority modifiers, etc.)
979 static void activate_task(struct task_struct *p, struct rq *rq, int local)
981 unsigned long long now;
989 /* Compensate for drifting sched_clock */
990 struct rq *this_rq = this_rq();
991 now = (now - this_rq->most_recent_timestamp)
992 + rq->most_recent_timestamp;
997 * Sleep time is in units of nanosecs, so shift by 20 to get a
998 * milliseconds-range estimation of the amount of time that the task
1001 if (unlikely(prof_on == SLEEP_PROFILING)) {
1002 if (p->state == TASK_UNINTERRUPTIBLE)
1003 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
1004 (now - p->timestamp) >> 20);
1007 p->prio = recalc_task_prio(p, now);
1010 * This checks to make sure it's not an uninterruptible task
1011 * that is now waking up.
1013 if (p->sleep_type == SLEEP_NORMAL) {
1015 * Tasks which were woken up by interrupts (ie. hw events)
1016 * are most likely of interactive nature. So we give them
1017 * the credit of extending their sleep time to the period
1018 * of time they spend on the runqueue, waiting for execution
1019 * on a CPU, first time around:
1022 p->sleep_type = SLEEP_INTERRUPTED;
1025 * Normal first-time wakeups get a credit too for
1026 * on-runqueue time, but it will be weighted down:
1028 p->sleep_type = SLEEP_INTERACTIVE;
1033 __activate_task(p, rq);
1037 * deactivate_task - remove a task from the runqueue.
1039 static void deactivate_task(struct task_struct *p, struct rq *rq)
1041 dec_nr_running(p, rq);
1042 dequeue_task(p, p->array);
1047 * resched_task - mark a task 'to be rescheduled now'.
1049 * On UP this means the setting of the need_resched flag, on SMP it
1050 * might also involve a cross-CPU call to trigger the scheduler on
1055 #ifndef tsk_is_polling
1056 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1059 static void resched_task(struct task_struct *p)
1063 assert_spin_locked(&task_rq(p)->lock);
1065 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1068 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1071 if (cpu == smp_processor_id())
1074 /* NEED_RESCHED must be visible before we test polling */
1076 if (!tsk_is_polling(p))
1077 smp_send_reschedule(cpu);
1080 static void resched_cpu(int cpu)
1082 struct rq *rq = cpu_rq(cpu);
1083 unsigned long flags;
1085 if (!spin_trylock_irqsave(&rq->lock, flags))
1087 resched_task(cpu_curr(cpu));
1088 spin_unlock_irqrestore(&rq->lock, flags);
1091 static inline void resched_task(struct task_struct *p)
1093 assert_spin_locked(&task_rq(p)->lock);
1094 set_tsk_need_resched(p);
1099 * task_curr - is this task currently executing on a CPU?
1100 * @p: the task in question.
1102 inline int task_curr(const struct task_struct *p)
1104 return cpu_curr(task_cpu(p)) == p;
1107 /* Used instead of source_load when we know the type == 0 */
1108 unsigned long weighted_cpuload(const int cpu)
1110 return cpu_rq(cpu)->raw_weighted_load;
1114 struct migration_req {
1115 struct list_head list;
1117 struct task_struct *task;
1120 struct completion done;
1124 * The task's runqueue lock must be held.
1125 * Returns true if you have to wait for migration thread.
1128 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1130 struct rq *rq = task_rq(p);
1133 * If the task is not on a runqueue (and not running), then
1134 * it is sufficient to simply update the task's cpu field.
1136 if (!p->array && !task_running(rq, p)) {
1137 set_task_cpu(p, dest_cpu);
1141 init_completion(&req->done);
1143 req->dest_cpu = dest_cpu;
1144 list_add(&req->list, &rq->migration_queue);
1150 * wait_task_inactive - wait for a thread to unschedule.
1152 * The caller must ensure that the task *will* unschedule sometime soon,
1153 * else this function might spin for a *long* time. This function can't
1154 * be called with interrupts off, or it may introduce deadlock with
1155 * smp_call_function() if an IPI is sent by the same process we are
1156 * waiting to become inactive.
1158 void wait_task_inactive(struct task_struct *p)
1160 unsigned long flags;
1165 rq = task_rq_lock(p, &flags);
1166 /* Must be off runqueue entirely, not preempted. */
1167 if (unlikely(p->array || task_running(rq, p))) {
1168 /* If it's preempted, we yield. It could be a while. */
1169 preempted = !task_running(rq, p);
1170 task_rq_unlock(rq, &flags);
1176 task_rq_unlock(rq, &flags);
1180 * kick_process - kick a running thread to enter/exit the kernel
1181 * @p: the to-be-kicked thread
1183 * Cause a process which is running on another CPU to enter
1184 * kernel-mode, without any delay. (to get signals handled.)
1186 * NOTE: this function doesnt have to take the runqueue lock,
1187 * because all it wants to ensure is that the remote task enters
1188 * the kernel. If the IPI races and the task has been migrated
1189 * to another CPU then no harm is done and the purpose has been
1192 void kick_process(struct task_struct *p)
1198 if ((cpu != smp_processor_id()) && task_curr(p))
1199 smp_send_reschedule(cpu);
1204 * Return a low guess at the load of a migration-source cpu weighted
1205 * according to the scheduling class and "nice" value.
1207 * We want to under-estimate the load of migration sources, to
1208 * balance conservatively.
1210 static inline unsigned long source_load(int cpu, int type)
1212 struct rq *rq = cpu_rq(cpu);
1215 return rq->raw_weighted_load;
1217 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1221 * Return a high guess at the load of a migration-target cpu weighted
1222 * according to the scheduling class and "nice" value.
1224 static inline unsigned long target_load(int cpu, int type)
1226 struct rq *rq = cpu_rq(cpu);
1229 return rq->raw_weighted_load;
1231 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1235 * Return the average load per task on the cpu's run queue
1237 static inline unsigned long cpu_avg_load_per_task(int cpu)
1239 struct rq *rq = cpu_rq(cpu);
1240 unsigned long n = rq->nr_running;
1242 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1246 * find_idlest_group finds and returns the least busy CPU group within the
1249 static struct sched_group *
1250 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1252 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1253 unsigned long min_load = ULONG_MAX, this_load = 0;
1254 int load_idx = sd->forkexec_idx;
1255 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1258 unsigned long load, avg_load;
1262 /* Skip over this group if it has no CPUs allowed */
1263 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1266 local_group = cpu_isset(this_cpu, group->cpumask);
1268 /* Tally up the load of all CPUs in the group */
1271 for_each_cpu_mask(i, group->cpumask) {
1272 /* Bias balancing toward cpus of our domain */
1274 load = source_load(i, load_idx);
1276 load = target_load(i, load_idx);
1281 /* Adjust by relative CPU power of the group */
1282 avg_load = sg_div_cpu_power(group,
1283 avg_load * SCHED_LOAD_SCALE);
1286 this_load = avg_load;
1288 } else if (avg_load < min_load) {
1289 min_load = avg_load;
1293 group = group->next;
1294 } while (group != sd->groups);
1296 if (!idlest || 100*this_load < imbalance*min_load)
1302 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1305 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1308 unsigned long load, min_load = ULONG_MAX;
1312 /* Traverse only the allowed CPUs */
1313 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1315 for_each_cpu_mask(i, tmp) {
1316 load = weighted_cpuload(i);
1318 if (load < min_load || (load == min_load && i == this_cpu)) {
1328 find_idlest_cpu_nodomain(struct task_struct *p, int this_cpu)
1331 unsigned long load, min_load = ULONG_MAX;
1335 /* Traverse only the allowed CPUs */
1336 cpus_and(tmp, cpu_online_map, p->cpus_allowed);
1338 for_each_cpu_mask(i, tmp) {
1339 load = target_load(i, 1);
1341 if (load < min_load) {
1350 * sched_balance_self: balance the current task (running on cpu) in domains
1351 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1354 * Balance, ie. select the least loaded group.
1356 * Returns the target CPU number, or the same CPU if no balancing is needed.
1358 * preempt must be disabled.
1361 int affinity_load_balancing = 0;
1363 static int sched_balance_self(int cpu, int flag)
1365 struct task_struct *t = current;
1366 struct sched_domain *tmp, *sd = NULL;
1368 if (affinity_load_balancing && !cpus_full(t->cpus_allowed))
1369 return find_idlest_cpu_nodomain(t, cpu);
1371 for_each_domain(cpu, tmp) {
1373 * If power savings logic is enabled for a domain, stop there.
1375 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1377 if (tmp->flags & flag)
1383 struct sched_group *group;
1384 int new_cpu, weight;
1386 if (!(sd->flags & flag)) {
1392 group = find_idlest_group(sd, t, cpu);
1398 new_cpu = find_idlest_cpu(group, t, cpu);
1399 if (new_cpu == -1 || new_cpu == cpu) {
1400 /* Now try balancing at a lower domain level of cpu */
1405 /* Now try balancing at a lower domain level of new_cpu */
1408 weight = cpus_weight(span);
1409 for_each_domain(cpu, tmp) {
1410 if (weight <= cpus_weight(tmp->span))
1412 if (tmp->flags & flag)
1415 /* while loop will break here if sd == NULL */
1421 #endif /* CONFIG_SMP */
1424 * wake_idle() will wake a task on an idle cpu if task->cpu is
1425 * not idle and an idle cpu is available. The span of cpus to
1426 * search starts with cpus closest then further out as needed,
1427 * so we always favor a closer, idle cpu.
1429 * Returns the CPU we should wake onto.
1431 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1432 static int wake_idle(int cpu, struct task_struct *p)
1435 struct sched_domain *sd;
1439 * If it is idle, then it is the best cpu to run this task.
1441 * This cpu is also the best, if it has more than one task already.
1442 * Siblings must be also busy(in most cases) as they didn't already
1443 * pickup the extra load from this cpu and hence we need not check
1444 * sibling runqueue info. This will avoid the checks and cache miss
1445 * penalities associated with that.
1447 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1450 for_each_domain(cpu, sd) {
1451 if (sd->flags & SD_WAKE_IDLE) {
1452 cpus_and(tmp, sd->span, p->cpus_allowed);
1453 for_each_cpu_mask(i, tmp) {
1464 static inline int wake_idle(int cpu, struct task_struct *p)
1471 * try_to_wake_up - wake up a thread
1472 * @p: the to-be-woken-up thread
1473 * @state: the mask of task states that can be woken
1474 * @sync: do a synchronous wakeup?
1476 * Put it on the run-queue if it's not already there. The "current"
1477 * thread is always on the run-queue (except when the actual
1478 * re-schedule is in progress), and as such you're allowed to do
1479 * the simpler "current->state = TASK_RUNNING" to mark yourself
1480 * runnable without the overhead of this.
1482 * returns failure only if the task is already active.
1484 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1486 int cpu, this_cpu, success = 0;
1487 unsigned long flags;
1491 struct sched_domain *sd, *this_sd = NULL;
1492 unsigned long load, this_load;
1496 rq = task_rq_lock(p, &flags);
1497 old_state = p->state;
1498 if (!(old_state & state))
1505 this_cpu = smp_processor_id();
1508 if (unlikely(task_running(rq, p)))
1513 schedstat_inc(rq, ttwu_cnt);
1514 if (cpu == this_cpu) {
1515 schedstat_inc(rq, ttwu_local);
1519 for_each_domain(this_cpu, sd) {
1520 if (cpu_isset(cpu, sd->span)) {
1521 schedstat_inc(sd, ttwu_wake_remote);
1527 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1531 * Check for affine wakeup and passive balancing possibilities.
1534 int idx = this_sd->wake_idx;
1535 unsigned int imbalance;
1537 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1539 load = source_load(cpu, idx);
1540 this_load = target_load(this_cpu, idx);
1542 new_cpu = this_cpu; /* Wake to this CPU if we can */
1544 if (this_sd->flags & SD_WAKE_AFFINE) {
1545 unsigned long tl = this_load;
1546 unsigned long tl_per_task;
1548 tl_per_task = cpu_avg_load_per_task(this_cpu);
1551 * If sync wakeup then subtract the (maximum possible)
1552 * effect of the currently running task from the load
1553 * of the current CPU:
1556 tl -= current->load_weight;
1559 tl + target_load(cpu, idx) <= tl_per_task) ||
1560 100*(tl + p->load_weight) <= imbalance*load) {
1562 * This domain has SD_WAKE_AFFINE and
1563 * p is cache cold in this domain, and
1564 * there is no bad imbalance.
1566 schedstat_inc(this_sd, ttwu_move_affine);
1572 * Start passive balancing when half the imbalance_pct
1575 if (this_sd->flags & SD_WAKE_BALANCE) {
1576 if (imbalance*this_load <= 100*load) {
1577 schedstat_inc(this_sd, ttwu_move_balance);
1583 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1585 new_cpu = wake_idle(new_cpu, p);
1586 if (new_cpu != cpu) {
1587 set_task_cpu(p, new_cpu);
1588 task_rq_unlock(rq, &flags);
1589 /* might preempt at this point */
1590 rq = task_rq_lock(p, &flags);
1591 old_state = p->state;
1592 if (!(old_state & state))
1597 this_cpu = smp_processor_id();
1602 #endif /* CONFIG_SMP */
1603 if (old_state == TASK_UNINTERRUPTIBLE) {
1604 rq->nr_uninterruptible--;
1606 * Tasks on involuntary sleep don't earn
1607 * sleep_avg beyond just interactive state.
1609 p->sleep_type = SLEEP_NONINTERACTIVE;
1613 * Tasks that have marked their sleep as noninteractive get
1614 * woken up with their sleep average not weighted in an
1617 if (old_state & TASK_NONINTERACTIVE)
1618 p->sleep_type = SLEEP_NONINTERACTIVE;
1621 activate_task(p, rq, cpu == this_cpu);
1623 * Sync wakeups (i.e. those types of wakeups where the waker
1624 * has indicated that it will leave the CPU in short order)
1625 * don't trigger a preemption, if the woken up task will run on
1626 * this cpu. (in this case the 'I will reschedule' promise of
1627 * the waker guarantees that the freshly woken up task is going
1628 * to be considered on this CPU.)
1630 if (!sync || cpu != this_cpu) {
1631 if (TASK_PREEMPTS_CURR(p, rq))
1632 resched_task(rq->curr);
1637 p->state = TASK_RUNNING;
1639 task_rq_unlock(rq, &flags);
1644 int fastcall wake_up_process(struct task_struct *p)
1646 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1647 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1649 EXPORT_SYMBOL(wake_up_process);
1651 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1653 return try_to_wake_up(p, state, 0);
1656 static void task_running_tick(struct rq *rq, struct task_struct *p);
1658 * Perform scheduler related setup for a newly forked process p.
1659 * p is forked by current.
1661 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1663 int cpu = get_cpu();
1666 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1668 set_task_cpu(p, cpu);
1671 * We mark the process as running here, but have not actually
1672 * inserted it onto the runqueue yet. This guarantees that
1673 * nobody will actually run it, and a signal or other external
1674 * event cannot wake it up and insert it on the runqueue either.
1676 p->state = TASK_RUNNING;
1679 * Make sure we do not leak PI boosting priority to the child:
1681 p->prio = current->normal_prio;
1683 INIT_LIST_HEAD(&p->run_list);
1685 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1686 if (unlikely(sched_info_on()))
1687 memset(&p->sched_info, 0, sizeof(p->sched_info));
1689 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1692 #ifdef CONFIG_PREEMPT
1693 /* Want to start with kernel preemption disabled. */
1694 task_thread_info(p)->preempt_count = 1;
1697 * Share the timeslice between parent and child, thus the
1698 * total amount of pending timeslices in the system doesn't change,
1699 * resulting in more scheduling fairness.
1701 local_irq_disable();
1702 p->time_slice = (current->time_slice + 1) >> 1;
1704 * The remainder of the first timeslice might be recovered by
1705 * the parent if the child exits early enough.
1707 p->first_time_slice = 1;
1708 current->time_slice >>= 1;
1709 p->timestamp = sched_clock();
1710 if (unlikely(!current->time_slice)) {
1712 * This case is rare, it happens when the parent has only
1713 * a single jiffy left from its timeslice. Taking the
1714 * runqueue lock is not a problem.
1716 current->time_slice = 1;
1717 task_running_tick(cpu_rq(cpu), current);
1724 * wake_up_new_task - wake up a newly created task for the first time.
1726 * This function will do some initial scheduler statistics housekeeping
1727 * that must be done for every newly created context, then puts the task
1728 * on the runqueue and wakes it.
1730 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1732 struct rq *rq, *this_rq;
1733 unsigned long flags;
1736 rq = task_rq_lock(p, &flags);
1737 BUG_ON(p->state != TASK_RUNNING);
1738 this_cpu = smp_processor_id();
1742 * We decrease the sleep average of forking parents
1743 * and children as well, to keep max-interactive tasks
1744 * from forking tasks that are max-interactive. The parent
1745 * (current) is done further down, under its lock.
1747 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1748 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1750 p->prio = effective_prio(p);
1752 if (likely(cpu == this_cpu)) {
1753 if (!(clone_flags & CLONE_VM)) {
1755 * The VM isn't cloned, so we're in a good position to
1756 * do child-runs-first in anticipation of an exec. This
1757 * usually avoids a lot of COW overhead.
1759 if (unlikely(!current->array))
1760 __activate_task(p, rq);
1762 p->prio = current->prio;
1763 p->normal_prio = current->normal_prio;
1764 list_add_tail(&p->run_list, ¤t->run_list);
1765 p->array = current->array;
1766 p->array->nr_active++;
1767 inc_nr_running(p, rq);
1771 /* Run child last */
1772 __activate_task(p, rq);
1774 * We skip the following code due to cpu == this_cpu
1776 * task_rq_unlock(rq, &flags);
1777 * this_rq = task_rq_lock(current, &flags);
1781 this_rq = cpu_rq(this_cpu);
1784 * Not the local CPU - must adjust timestamp. This should
1785 * get optimised away in the !CONFIG_SMP case.
1787 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1788 + rq->most_recent_timestamp;
1789 __activate_task(p, rq);
1790 if (TASK_PREEMPTS_CURR(p, rq))
1791 resched_task(rq->curr);
1794 * Parent and child are on different CPUs, now get the
1795 * parent runqueue to update the parent's ->sleep_avg:
1797 task_rq_unlock(rq, &flags);
1798 this_rq = task_rq_lock(current, &flags);
1800 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1801 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1802 task_rq_unlock(this_rq, &flags);
1806 * Potentially available exiting-child timeslices are
1807 * retrieved here - this way the parent does not get
1808 * penalized for creating too many threads.
1810 * (this cannot be used to 'generate' timeslices
1811 * artificially, because any timeslice recovered here
1812 * was given away by the parent in the first place.)
1814 void fastcall sched_exit(struct task_struct *p)
1816 unsigned long flags;
1820 * If the child was a (relative-) CPU hog then decrease
1821 * the sleep_avg of the parent as well.
1823 rq = task_rq_lock(p->parent, &flags);
1824 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1825 p->parent->time_slice += p->time_slice;
1826 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1827 p->parent->time_slice = task_timeslice(p);
1829 if (p->sleep_avg < p->parent->sleep_avg)
1830 p->parent->sleep_avg = p->parent->sleep_avg /
1831 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1833 task_rq_unlock(rq, &flags);
1837 * prepare_task_switch - prepare to switch tasks
1838 * @rq: the runqueue preparing to switch
1839 * @next: the task we are going to switch to.
1841 * This is called with the rq lock held and interrupts off. It must
1842 * be paired with a subsequent finish_task_switch after the context
1845 * prepare_task_switch sets up locking and calls architecture specific
1848 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1850 prepare_lock_switch(rq, next);
1851 prepare_arch_switch(next);
1855 * finish_task_switch - clean up after a task-switch
1856 * @rq: runqueue associated with task-switch
1857 * @prev: the thread we just switched away from.
1859 * finish_task_switch must be called after the context switch, paired
1860 * with a prepare_task_switch call before the context switch.
1861 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1862 * and do any other architecture-specific cleanup actions.
1864 * Note that we may have delayed dropping an mm in context_switch(). If
1865 * so, we finish that here outside of the runqueue lock. (Doing it
1866 * with the lock held can cause deadlocks; see schedule() for
1869 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1870 __releases(rq->lock)
1872 struct mm_struct *mm = rq->prev_mm;
1878 * A task struct has one reference for the use as "current".
1879 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1880 * schedule one last time. The schedule call will never return, and
1881 * the scheduled task must drop that reference.
1882 * The test for TASK_DEAD must occur while the runqueue locks are
1883 * still held, otherwise prev could be scheduled on another cpu, die
1884 * there before we look at prev->state, and then the reference would
1886 * Manfred Spraul <manfred@colorfullife.com>
1888 prev_state = prev->state;
1889 finish_arch_switch(prev);
1890 finish_lock_switch(rq, prev);
1893 if (unlikely(prev_state == TASK_DEAD)) {
1895 * Remove function-return probe instances associated with this
1896 * task and put them back on the free list.
1898 kprobe_flush_task(prev);
1899 put_task_struct(prev);
1904 * schedule_tail - first thing a freshly forked thread must call.
1905 * @prev: the thread we just switched away from.
1907 asmlinkage void schedule_tail(struct task_struct *prev)
1908 __releases(rq->lock)
1910 struct rq *rq = this_rq();
1912 finish_task_switch(rq, prev);
1913 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1914 /* In this case, finish_task_switch does not reenable preemption */
1917 if (current->set_child_tid)
1918 put_user(current->pid, current->set_child_tid);
1922 * context_switch - switch to the new MM and the new
1923 * thread's register state.
1925 static inline struct task_struct *
1926 context_switch(struct rq *rq, struct task_struct *prev,
1927 struct task_struct *next)
1929 struct mm_struct *mm = next->mm;
1930 struct mm_struct *oldmm = prev->active_mm;
1933 * For paravirt, this is coupled with an exit in switch_to to
1934 * combine the page table reload and the switch backend into
1937 arch_enter_lazy_cpu_mode();
1940 next->active_mm = oldmm;
1941 atomic_inc(&oldmm->mm_count);
1942 enter_lazy_tlb(oldmm, next);
1944 switch_mm(oldmm, mm, next);
1947 prev->active_mm = NULL;
1948 WARN_ON(rq->prev_mm);
1949 rq->prev_mm = oldmm;
1952 * Since the runqueue lock will be released by the next
1953 * task (which is an invalid locking op but in the case
1954 * of the scheduler it's an obvious special-case), so we
1955 * do an early lockdep release here:
1957 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1958 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1961 /* Here we just switch the register state and the stack. */
1962 switch_to(prev, next, prev);
1968 * nr_running, nr_uninterruptible and nr_context_switches:
1970 * externally visible scheduler statistics: current number of runnable
1971 * threads, current number of uninterruptible-sleeping threads, total
1972 * number of context switches performed since bootup.
1974 unsigned long nr_running(void)
1976 unsigned long i, sum = 0;
1978 for_each_online_cpu(i)
1979 sum += cpu_rq(i)->nr_running;
1984 unsigned long nr_uninterruptible(void)
1986 unsigned long i, sum = 0;
1988 for_each_possible_cpu(i)
1989 sum += cpu_rq(i)->nr_uninterruptible;
1992 * Since we read the counters lockless, it might be slightly
1993 * inaccurate. Do not allow it to go below zero though:
1995 if (unlikely((long)sum < 0))
2001 unsigned long long nr_context_switches(void)
2004 unsigned long long sum = 0;
2006 for_each_possible_cpu(i)
2007 sum += cpu_rq(i)->nr_switches;
2012 unsigned long nr_iowait(void)
2014 unsigned long i, sum = 0;
2016 for_each_possible_cpu(i)
2017 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2022 unsigned long nr_active(void)
2024 unsigned long i, running = 0, uninterruptible = 0;
2026 for_each_online_cpu(i) {
2027 running += cpu_rq(i)->nr_running;
2028 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2031 if (unlikely((long)uninterruptible < 0))
2032 uninterruptible = 0;
2034 return running + uninterruptible;
2040 * Is this task likely cache-hot:
2043 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
2045 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
2049 * double_rq_lock - safely lock two runqueues
2051 * Note this does not disable interrupts like task_rq_lock,
2052 * you need to do so manually before calling.
2054 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2055 __acquires(rq1->lock)
2056 __acquires(rq2->lock)
2058 BUG_ON(!irqs_disabled());
2060 spin_lock(&rq1->lock);
2061 __acquire(rq2->lock); /* Fake it out ;) */
2064 spin_lock(&rq1->lock);
2065 spin_lock(&rq2->lock);
2067 spin_lock(&rq2->lock);
2068 spin_lock(&rq1->lock);
2074 * double_rq_unlock - safely unlock two runqueues
2076 * Note this does not restore interrupts like task_rq_unlock,
2077 * you need to do so manually after calling.
2079 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2080 __releases(rq1->lock)
2081 __releases(rq2->lock)
2083 spin_unlock(&rq1->lock);
2085 spin_unlock(&rq2->lock);
2087 __release(rq2->lock);
2091 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2093 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2094 __releases(this_rq->lock)
2095 __acquires(busiest->lock)
2096 __acquires(this_rq->lock)
2098 if (unlikely(!irqs_disabled())) {
2099 /* printk() doesn't work good under rq->lock */
2100 spin_unlock(&this_rq->lock);
2103 if (unlikely(!spin_trylock(&busiest->lock))) {
2104 if (busiest < this_rq) {
2105 spin_unlock(&this_rq->lock);
2106 spin_lock(&busiest->lock);
2107 spin_lock(&this_rq->lock);
2109 spin_lock(&busiest->lock);
2114 * If dest_cpu is allowed for this process, migrate the task to it.
2115 * This is accomplished by forcing the cpu_allowed mask to only
2116 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2117 * the cpu_allowed mask is restored.
2119 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2121 struct migration_req req;
2122 unsigned long flags;
2125 rq = task_rq_lock(p, &flags);
2126 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2127 || unlikely(cpu_is_offline(dest_cpu)))
2130 /* force the process onto the specified CPU */
2131 if (migrate_task(p, dest_cpu, &req)) {
2132 /* Need to wait for migration thread (might exit: take ref). */
2133 struct task_struct *mt = rq->migration_thread;
2135 get_task_struct(mt);
2136 task_rq_unlock(rq, &flags);
2137 wake_up_process(mt);
2138 put_task_struct(mt);
2139 wait_for_completion(&req.done);
2144 task_rq_unlock(rq, &flags);
2148 * sched_exec - execve() is a valuable balancing opportunity, because at
2149 * this point the task has the smallest effective memory and cache footprint.
2151 void sched_exec(void)
2153 int new_cpu, this_cpu = get_cpu();
2154 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2156 if (new_cpu != this_cpu)
2157 sched_migrate_task(current, new_cpu);
2161 * pull_task - move a task from a remote runqueue to the local runqueue.
2162 * Both runqueues must be locked.
2164 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2165 struct task_struct *p, struct rq *this_rq,
2166 struct prio_array *this_array, int this_cpu)
2168 dequeue_task(p, src_array);
2169 dec_nr_running(p, src_rq);
2170 set_task_cpu(p, this_cpu);
2171 inc_nr_running(p, this_rq);
2172 enqueue_task(p, this_array);
2173 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2174 + this_rq->most_recent_timestamp;
2176 * Note that idle threads have a prio of MAX_PRIO, for this test
2177 * to be always true for them.
2179 if (TASK_PREEMPTS_CURR(p, this_rq))
2180 resched_task(this_rq->curr);
2184 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2187 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2188 struct sched_domain *sd, enum idle_type idle,
2192 * We do not migrate tasks that are:
2193 * 1) running (obviously), or
2194 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2195 * 3) are cache-hot on their current CPU.
2197 if (!cpu_isset(this_cpu, p->cpus_allowed))
2201 if (task_running(rq, p))
2205 * Aggressive migration if:
2206 * 1) task is cache cold, or
2207 * 2) too many balance attempts have failed.
2210 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2211 #ifdef CONFIG_SCHEDSTATS
2212 if (task_hot(p, rq->most_recent_timestamp, sd))
2213 schedstat_inc(sd, lb_hot_gained[idle]);
2218 if (task_hot(p, rq->most_recent_timestamp, sd))
2223 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2226 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2227 * load from busiest to this_rq, as part of a balancing operation within
2228 * "domain". Returns the number of tasks moved.
2230 * Called with both runqueues locked.
2232 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2233 unsigned long max_nr_move, unsigned long max_load_move,
2234 struct sched_domain *sd, enum idle_type idle,
2237 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2238 best_prio_seen, skip_for_load;
2239 struct prio_array *array, *dst_array;
2240 struct list_head *head, *curr;
2241 struct task_struct *tmp;
2244 if (max_nr_move == 0 || max_load_move == 0)
2247 rem_load_move = max_load_move;
2249 this_best_prio = rq_best_prio(this_rq);
2250 best_prio = rq_best_prio(busiest);
2252 * Enable handling of the case where there is more than one task
2253 * with the best priority. If the current running task is one
2254 * of those with prio==best_prio we know it won't be moved
2255 * and therefore it's safe to override the skip (based on load) of
2256 * any task we find with that prio.
2258 best_prio_seen = best_prio == busiest->curr->prio;
2261 * We first consider expired tasks. Those will likely not be
2262 * executed in the near future, and they are most likely to
2263 * be cache-cold, thus switching CPUs has the least effect
2266 if (busiest->expired->nr_active) {
2267 array = busiest->expired;
2268 dst_array = this_rq->expired;
2270 array = busiest->active;
2271 dst_array = this_rq->active;
2275 /* Start searching at priority 0: */
2279 idx = sched_find_first_bit(array->bitmap);
2281 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2282 if (idx >= MAX_PRIO) {
2283 if (array == busiest->expired && busiest->active->nr_active) {
2284 array = busiest->active;
2285 dst_array = this_rq->active;
2291 head = array->queue + idx;
2294 tmp = list_entry(curr, struct task_struct, run_list);
2299 * To help distribute high priority tasks accross CPUs we don't
2300 * skip a task if it will be the highest priority task (i.e. smallest
2301 * prio value) on its new queue regardless of its load weight
2303 skip_for_load = tmp->load_weight > rem_load_move;
2304 if (skip_for_load && idx < this_best_prio)
2305 skip_for_load = !best_prio_seen && idx == best_prio;
2306 if (skip_for_load ||
2307 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2309 best_prio_seen |= idx == best_prio;
2316 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2318 rem_load_move -= tmp->load_weight;
2321 * We only want to steal up to the prescribed number of tasks
2322 * and the prescribed amount of weighted load.
2324 if (pulled < max_nr_move && rem_load_move > 0) {
2325 if (idx < this_best_prio)
2326 this_best_prio = idx;
2334 * Right now, this is the only place pull_task() is called,
2335 * so we can safely collect pull_task() stats here rather than
2336 * inside pull_task().
2338 schedstat_add(sd, lb_gained[idle], pulled);
2341 *all_pinned = pinned;
2346 * find_busiest_group finds and returns the busiest CPU group within the
2347 * domain. It calculates and returns the amount of weighted load which
2348 * should be moved to restore balance via the imbalance parameter.
2350 static struct sched_group *
2351 find_busiest_group(struct sched_domain *sd, int this_cpu,
2352 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2353 cpumask_t *cpus, int *balance)
2355 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2356 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2357 unsigned long max_pull;
2358 unsigned long busiest_load_per_task, busiest_nr_running;
2359 unsigned long this_load_per_task, this_nr_running;
2361 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2362 int power_savings_balance = 1;
2363 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2364 unsigned long min_nr_running = ULONG_MAX;
2365 struct sched_group *group_min = NULL, *group_leader = NULL;
2368 max_load = this_load = total_load = total_pwr = 0;
2369 busiest_load_per_task = busiest_nr_running = 0;
2370 this_load_per_task = this_nr_running = 0;
2371 if (idle == NOT_IDLE)
2372 load_idx = sd->busy_idx;
2373 else if (idle == NEWLY_IDLE)
2374 load_idx = sd->newidle_idx;
2376 load_idx = sd->idle_idx;
2379 unsigned long load, group_capacity;
2382 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2383 unsigned long sum_nr_running, sum_weighted_load;
2385 local_group = cpu_isset(this_cpu, group->cpumask);
2388 balance_cpu = first_cpu(group->cpumask);
2390 /* Tally up the load of all CPUs in the group */
2391 sum_weighted_load = sum_nr_running = avg_load = 0;
2393 for_each_cpu_mask(i, group->cpumask) {
2396 if (!cpu_isset(i, *cpus))
2401 if (*sd_idle && !idle_cpu(i))
2404 /* Bias balancing toward cpus of our domain */
2406 if (idle_cpu(i) && !first_idle_cpu) {
2411 load = target_load(i, load_idx);
2413 load = source_load(i, load_idx);
2416 sum_nr_running += rq->nr_running;
2417 sum_weighted_load += rq->raw_weighted_load;
2421 * First idle cpu or the first cpu(busiest) in this sched group
2422 * is eligible for doing load balancing at this and above
2425 if (local_group && balance_cpu != this_cpu && balance) {
2430 total_load += avg_load;
2431 total_pwr += group->__cpu_power;
2433 /* Adjust by relative CPU power of the group */
2434 avg_load = sg_div_cpu_power(group,
2435 avg_load * SCHED_LOAD_SCALE);
2437 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2440 this_load = avg_load;
2442 this_nr_running = sum_nr_running;
2443 this_load_per_task = sum_weighted_load;
2444 } else if (avg_load > max_load &&
2445 sum_nr_running > group_capacity) {
2446 max_load = avg_load;
2448 busiest_nr_running = sum_nr_running;
2449 busiest_load_per_task = sum_weighted_load;
2452 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2454 * Busy processors will not participate in power savings
2457 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2461 * If the local group is idle or completely loaded
2462 * no need to do power savings balance at this domain
2464 if (local_group && (this_nr_running >= group_capacity ||
2466 power_savings_balance = 0;
2469 * If a group is already running at full capacity or idle,
2470 * don't include that group in power savings calculations
2472 if (!power_savings_balance || sum_nr_running >= group_capacity
2477 * Calculate the group which has the least non-idle load.
2478 * This is the group from where we need to pick up the load
2481 if ((sum_nr_running < min_nr_running) ||
2482 (sum_nr_running == min_nr_running &&
2483 first_cpu(group->cpumask) <
2484 first_cpu(group_min->cpumask))) {
2486 min_nr_running = sum_nr_running;
2487 min_load_per_task = sum_weighted_load /
2492 * Calculate the group which is almost near its
2493 * capacity but still has some space to pick up some load
2494 * from other group and save more power
2496 if (sum_nr_running <= group_capacity - 1) {
2497 if (sum_nr_running > leader_nr_running ||
2498 (sum_nr_running == leader_nr_running &&
2499 first_cpu(group->cpumask) >
2500 first_cpu(group_leader->cpumask))) {
2501 group_leader = group;
2502 leader_nr_running = sum_nr_running;
2507 group = group->next;
2508 } while (group != sd->groups);
2510 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2513 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2515 if (this_load >= avg_load ||
2516 100*max_load <= sd->imbalance_pct*this_load)
2519 busiest_load_per_task /= busiest_nr_running;
2521 * We're trying to get all the cpus to the average_load, so we don't
2522 * want to push ourselves above the average load, nor do we wish to
2523 * reduce the max loaded cpu below the average load, as either of these
2524 * actions would just result in more rebalancing later, and ping-pong
2525 * tasks around. Thus we look for the minimum possible imbalance.
2526 * Negative imbalances (*we* are more loaded than anyone else) will
2527 * be counted as no imbalance for these purposes -- we can't fix that
2528 * by pulling tasks to us. Be careful of negative numbers as they'll
2529 * appear as very large values with unsigned longs.
2531 if (max_load <= busiest_load_per_task)
2535 * In the presence of smp nice balancing, certain scenarios can have
2536 * max load less than avg load(as we skip the groups at or below
2537 * its cpu_power, while calculating max_load..)
2539 if (max_load < avg_load) {
2541 goto small_imbalance;
2544 /* Don't want to pull so many tasks that a group would go idle */
2545 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2547 /* How much load to actually move to equalise the imbalance */
2548 *imbalance = min(max_pull * busiest->__cpu_power,
2549 (avg_load - this_load) * this->__cpu_power)
2553 * if *imbalance is less than the average load per runnable task
2554 * there is no gaurantee that any tasks will be moved so we'll have
2555 * a think about bumping its value to force at least one task to be
2558 if (*imbalance < busiest_load_per_task) {
2559 unsigned long tmp, pwr_now, pwr_move;
2563 pwr_move = pwr_now = 0;
2565 if (this_nr_running) {
2566 this_load_per_task /= this_nr_running;
2567 if (busiest_load_per_task > this_load_per_task)
2570 this_load_per_task = SCHED_LOAD_SCALE;
2572 if (max_load - this_load >= busiest_load_per_task * imbn) {
2573 *imbalance = busiest_load_per_task;
2578 * OK, we don't have enough imbalance to justify moving tasks,
2579 * however we may be able to increase total CPU power used by
2583 pwr_now += busiest->__cpu_power *
2584 min(busiest_load_per_task, max_load);
2585 pwr_now += this->__cpu_power *
2586 min(this_load_per_task, this_load);
2587 pwr_now /= SCHED_LOAD_SCALE;
2589 /* Amount of load we'd subtract */
2590 tmp = sg_div_cpu_power(busiest,
2591 busiest_load_per_task * SCHED_LOAD_SCALE);
2593 pwr_move += busiest->__cpu_power *
2594 min(busiest_load_per_task, max_load - tmp);
2596 /* Amount of load we'd add */
2597 if (max_load * busiest->__cpu_power <
2598 busiest_load_per_task * SCHED_LOAD_SCALE)
2599 tmp = sg_div_cpu_power(this,
2600 max_load * busiest->__cpu_power);
2602 tmp = sg_div_cpu_power(this,
2603 busiest_load_per_task * SCHED_LOAD_SCALE);
2604 pwr_move += this->__cpu_power *
2605 min(this_load_per_task, this_load + tmp);
2606 pwr_move /= SCHED_LOAD_SCALE;
2608 /* Move if we gain throughput */
2609 if (pwr_move <= pwr_now)
2612 *imbalance = busiest_load_per_task;
2618 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2619 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2622 if (this == group_leader && group_leader != group_min) {
2623 *imbalance = min_load_per_task;
2633 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2636 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2637 unsigned long imbalance, cpumask_t *cpus)
2639 struct rq *busiest = NULL, *rq;
2640 unsigned long max_load = 0;
2643 for_each_cpu_mask(i, group->cpumask) {
2645 if (!cpu_isset(i, *cpus))
2650 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2653 if (rq->raw_weighted_load > max_load) {
2654 max_load = rq->raw_weighted_load;
2663 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2664 * so long as it is large enough.
2666 #define MAX_PINNED_INTERVAL 512
2668 static inline unsigned long minus_1_or_zero(unsigned long n)
2670 return n > 0 ? n - 1 : 0;
2674 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2675 * tasks if there is an imbalance.
2677 static int load_balance(int this_cpu, struct rq *this_rq,
2678 struct sched_domain *sd, enum idle_type idle,
2681 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2682 struct sched_group *group;
2683 unsigned long imbalance;
2685 cpumask_t cpus = CPU_MASK_ALL;
2686 unsigned long flags;
2689 * When power savings policy is enabled for the parent domain, idle
2690 * sibling can pick up load irrespective of busy siblings. In this case,
2691 * let the state of idle sibling percolate up as IDLE, instead of
2692 * portraying it as NOT_IDLE.
2694 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2695 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2698 schedstat_inc(sd, lb_cnt[idle]);
2701 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2708 schedstat_inc(sd, lb_nobusyg[idle]);
2712 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2714 schedstat_inc(sd, lb_nobusyq[idle]);
2718 BUG_ON(busiest == this_rq);
2720 schedstat_add(sd, lb_imbalance[idle], imbalance);
2723 if (busiest->nr_running > 1) {
2725 * Attempt to move tasks. If find_busiest_group has found
2726 * an imbalance but busiest->nr_running <= 1, the group is
2727 * still unbalanced. nr_moved simply stays zero, so it is
2728 * correctly treated as an imbalance.
2730 local_irq_save(flags);
2731 double_rq_lock(this_rq, busiest);
2732 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2733 minus_1_or_zero(busiest->nr_running),
2734 imbalance, sd, idle, &all_pinned);
2735 double_rq_unlock(this_rq, busiest);
2736 local_irq_restore(flags);
2739 * some other cpu did the load balance for us.
2741 if (nr_moved && this_cpu != smp_processor_id())
2742 resched_cpu(this_cpu);
2744 /* All tasks on this runqueue were pinned by CPU affinity */
2745 if (unlikely(all_pinned)) {
2746 cpu_clear(cpu_of(busiest), cpus);
2747 if (!cpus_empty(cpus))
2754 schedstat_inc(sd, lb_failed[idle]);
2755 sd->nr_balance_failed++;
2757 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2759 spin_lock_irqsave(&busiest->lock, flags);
2761 /* don't kick the migration_thread, if the curr
2762 * task on busiest cpu can't be moved to this_cpu
2764 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2765 spin_unlock_irqrestore(&busiest->lock, flags);
2767 goto out_one_pinned;
2770 if (!busiest->active_balance) {
2771 busiest->active_balance = 1;
2772 busiest->push_cpu = this_cpu;
2775 spin_unlock_irqrestore(&busiest->lock, flags);
2777 wake_up_process(busiest->migration_thread);
2780 * We've kicked active balancing, reset the failure
2783 sd->nr_balance_failed = sd->cache_nice_tries+1;
2786 sd->nr_balance_failed = 0;
2788 if (likely(!active_balance)) {
2789 /* We were unbalanced, so reset the balancing interval */
2790 sd->balance_interval = sd->min_interval;
2793 * If we've begun active balancing, start to back off. This
2794 * case may not be covered by the all_pinned logic if there
2795 * is only 1 task on the busy runqueue (because we don't call
2798 if (sd->balance_interval < sd->max_interval)
2799 sd->balance_interval *= 2;
2802 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2803 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2808 schedstat_inc(sd, lb_balanced[idle]);
2810 sd->nr_balance_failed = 0;
2813 /* tune up the balancing interval */
2814 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2815 (sd->balance_interval < sd->max_interval))
2816 sd->balance_interval *= 2;
2818 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2819 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2825 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2826 * tasks if there is an imbalance.
2828 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2829 * this_rq is locked.
2832 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2834 struct sched_group *group;
2835 struct rq *busiest = NULL;
2836 unsigned long imbalance;
2839 cpumask_t cpus = CPU_MASK_ALL;
2842 * When power savings policy is enabled for the parent domain, idle
2843 * sibling can pick up load irrespective of busy siblings. In this case,
2844 * let the state of idle sibling percolate up as IDLE, instead of
2845 * portraying it as NOT_IDLE.
2847 if (sd->flags & SD_SHARE_CPUPOWER &&
2848 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2851 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2853 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2854 &sd_idle, &cpus, NULL);
2856 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2860 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2863 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2867 BUG_ON(busiest == this_rq);
2869 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2872 if (busiest->nr_running > 1) {
2873 /* Attempt to move tasks */
2874 double_lock_balance(this_rq, busiest);
2875 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2876 minus_1_or_zero(busiest->nr_running),
2877 imbalance, sd, NEWLY_IDLE, NULL);
2878 spin_unlock(&busiest->lock);
2881 cpu_clear(cpu_of(busiest), cpus);
2882 if (!cpus_empty(cpus))
2888 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2889 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2890 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2893 sd->nr_balance_failed = 0;
2898 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2899 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2900 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2902 sd->nr_balance_failed = 0;
2908 * idle_balance is called by schedule() if this_cpu is about to become
2909 * idle. Attempts to pull tasks from other CPUs.
2911 static void idle_balance(int this_cpu, struct rq *this_rq)
2913 struct sched_domain *sd;
2914 int pulled_task = 0;
2915 unsigned long next_balance = jiffies + 60 * HZ;
2917 for_each_domain(this_cpu, sd) {
2918 if (sd->flags & SD_BALANCE_NEWIDLE) {
2919 /* If we've pulled tasks over stop searching: */
2920 pulled_task = load_balance_newidle(this_cpu,
2922 if (time_after(next_balance,
2923 sd->last_balance + sd->balance_interval))
2924 next_balance = sd->last_balance
2925 + sd->balance_interval;
2932 * We are going idle. next_balance may be set based on
2933 * a busy processor. So reset next_balance.
2935 this_rq->next_balance = next_balance;
2939 * active_load_balance is run by migration threads. It pushes running tasks
2940 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2941 * running on each physical CPU where possible, and avoids physical /
2942 * logical imbalances.
2944 * Called with busiest_rq locked.
2946 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2948 int target_cpu = busiest_rq->push_cpu;
2949 struct sched_domain *sd;
2950 struct rq *target_rq;
2952 /* Is there any task to move? */
2953 if (busiest_rq->nr_running <= 1)
2956 target_rq = cpu_rq(target_cpu);
2959 * This condition is "impossible", if it occurs
2960 * we need to fix it. Originally reported by
2961 * Bjorn Helgaas on a 128-cpu setup.
2963 BUG_ON(busiest_rq == target_rq);
2965 /* move a task from busiest_rq to target_rq */
2966 double_lock_balance(busiest_rq, target_rq);
2968 /* Search for an sd spanning us and the target CPU. */
2969 for_each_domain(target_cpu, sd) {
2970 if ((sd->flags & SD_LOAD_BALANCE) &&
2971 cpu_isset(busiest_cpu, sd->span))
2976 schedstat_inc(sd, alb_cnt);
2978 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2979 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2981 schedstat_inc(sd, alb_pushed);
2983 schedstat_inc(sd, alb_failed);
2985 spin_unlock(&target_rq->lock);
2988 static void update_load(struct rq *this_rq)
2990 unsigned long this_load;
2991 unsigned int i, scale;
2993 this_load = this_rq->raw_weighted_load;
2995 /* Update our load: */
2996 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2997 unsigned long old_load, new_load;
2999 /* scale is effectively 1 << i now, and >> i divides by scale */
3001 old_load = this_rq->cpu_load[i];
3002 new_load = this_load;
3004 * Round up the averaging division if load is increasing. This
3005 * prevents us from getting stuck on 9 if the load is 10, for
3008 if (new_load > old_load)
3009 new_load += scale-1;
3010 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3016 atomic_t load_balancer;
3018 } nohz ____cacheline_aligned = {
3019 .load_balancer = ATOMIC_INIT(-1),
3020 .cpu_mask = CPU_MASK_NONE,
3024 * This routine will try to nominate the ilb (idle load balancing)
3025 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3026 * load balancing on behalf of all those cpus. If all the cpus in the system
3027 * go into this tickless mode, then there will be no ilb owner (as there is
3028 * no need for one) and all the cpus will sleep till the next wakeup event
3031 * For the ilb owner, tick is not stopped. And this tick will be used
3032 * for idle load balancing. ilb owner will still be part of
3035 * While stopping the tick, this cpu will become the ilb owner if there
3036 * is no other owner. And will be the owner till that cpu becomes busy
3037 * or if all cpus in the system stop their ticks at which point
3038 * there is no need for ilb owner.
3040 * When the ilb owner becomes busy, it nominates another owner, during the
3041 * next busy scheduler_tick()
3043 int select_nohz_load_balancer(int stop_tick)
3045 int cpu = smp_processor_id();
3048 cpu_set(cpu, nohz.cpu_mask);
3049 cpu_rq(cpu)->in_nohz_recently = 1;
3052 * If we are going offline and still the leader, give up!
3054 if (cpu_is_offline(cpu) &&
3055 atomic_read(&nohz.load_balancer) == cpu) {
3056 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3061 /* time for ilb owner also to sleep */
3062 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3063 if (atomic_read(&nohz.load_balancer) == cpu)
3064 atomic_set(&nohz.load_balancer, -1);
3068 if (atomic_read(&nohz.load_balancer) == -1) {
3069 /* make me the ilb owner */
3070 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3072 } else if (atomic_read(&nohz.load_balancer) == cpu)
3075 if (!cpu_isset(cpu, nohz.cpu_mask))
3078 cpu_clear(cpu, nohz.cpu_mask);
3080 if (atomic_read(&nohz.load_balancer) == cpu)
3081 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3088 static DEFINE_SPINLOCK(balancing);
3091 * It checks each scheduling domain to see if it is due to be balanced,
3092 * and initiates a balancing operation if so.
3094 * Balancing parameters are set up in arch_init_sched_domains.
3096 static inline void rebalance_domains(int cpu, enum idle_type idle)
3099 struct rq *rq = cpu_rq(cpu);
3100 unsigned long interval;
3101 struct sched_domain *sd;
3102 /* Earliest time when we have to do rebalance again */
3103 unsigned long next_balance = jiffies + 60*HZ;
3105 for_each_domain(cpu, sd) {
3106 if (!(sd->flags & SD_LOAD_BALANCE))
3109 interval = sd->balance_interval;
3110 if (idle != SCHED_IDLE)
3111 interval *= sd->busy_factor;
3113 /* scale ms to jiffies */
3114 interval = msecs_to_jiffies(interval);
3115 if (unlikely(!interval))
3118 if (sd->flags & SD_SERIALIZE) {
3119 if (!spin_trylock(&balancing))
3123 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3124 if (load_balance(cpu, rq, sd, idle, &balance)) {
3126 * We've pulled tasks over so either we're no
3127 * longer idle, or one of our SMT siblings is
3132 sd->last_balance = jiffies;
3134 if (sd->flags & SD_SERIALIZE)
3135 spin_unlock(&balancing);
3137 if (time_after(next_balance, sd->last_balance + interval))
3138 next_balance = sd->last_balance + interval;
3141 * Stop the load balance at this level. There is another
3142 * CPU in our sched group which is doing load balancing more
3148 rq->next_balance = next_balance;
3152 * run_rebalance_domains is triggered when needed from the scheduler tick.
3153 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3154 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3156 static void run_rebalance_domains(struct softirq_action *h)
3158 int local_cpu = smp_processor_id();
3159 struct rq *local_rq = cpu_rq(local_cpu);
3160 enum idle_type idle = local_rq->idle_at_tick ? SCHED_IDLE : NOT_IDLE;
3162 rebalance_domains(local_cpu, idle);
3166 * If this cpu is the owner for idle load balancing, then do the
3167 * balancing on behalf of the other idle cpus whose ticks are
3170 if (local_rq->idle_at_tick &&
3171 atomic_read(&nohz.load_balancer) == local_cpu) {
3172 cpumask_t cpus = nohz.cpu_mask;
3176 cpu_clear(local_cpu, cpus);
3177 for_each_cpu_mask(balance_cpu, cpus) {
3179 * If this cpu gets work to do, stop the load balancing
3180 * work being done for other cpus. Next load
3181 * balancing owner will pick it up.
3186 rebalance_domains(balance_cpu, SCHED_IDLE);
3188 rq = cpu_rq(balance_cpu);
3189 if (time_after(local_rq->next_balance, rq->next_balance))
3190 local_rq->next_balance = rq->next_balance;
3197 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3199 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3200 * idle load balancing owner or decide to stop the periodic load balancing,
3201 * if the whole system is idle.
3203 static inline void trigger_load_balance(int cpu)
3205 struct rq *rq = cpu_rq(cpu);
3208 * If we were in the nohz mode recently and busy at the current
3209 * scheduler tick, then check if we need to nominate new idle
3212 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3213 rq->in_nohz_recently = 0;
3215 if (atomic_read(&nohz.load_balancer) == cpu) {
3216 cpu_clear(cpu, nohz.cpu_mask);
3217 atomic_set(&nohz.load_balancer, -1);
3220 if (atomic_read(&nohz.load_balancer) == -1) {
3222 * simple selection for now: Nominate the
3223 * first cpu in the nohz list to be the next
3226 * TBD: Traverse the sched domains and nominate
3227 * the nearest cpu in the nohz.cpu_mask.
3229 int ilb = first_cpu(nohz.cpu_mask);
3237 * If this cpu is idle and doing idle load balancing for all the
3238 * cpus with ticks stopped, is it time for that to stop?
3240 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3241 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3247 * If this cpu is idle and the idle load balancing is done by
3248 * someone else, then no need raise the SCHED_SOFTIRQ
3250 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3251 cpu_isset(cpu, nohz.cpu_mask))
3254 if (time_after_eq(jiffies, rq->next_balance))
3255 raise_softirq(SCHED_SOFTIRQ);
3259 * on UP we do not need to balance between CPUs:
3261 static inline void idle_balance(int cpu, struct rq *rq)
3266 DEFINE_PER_CPU(struct kernel_stat, kstat);
3268 EXPORT_PER_CPU_SYMBOL(kstat);
3271 * This is called on clock ticks and on context switches.
3272 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3275 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3277 p->sched_time += now - p->last_ran;
3278 p->last_ran = rq->most_recent_timestamp = now;
3282 * Return current->sched_time plus any more ns on the sched_clock
3283 * that have not yet been banked.
3285 unsigned long long current_sched_time(const struct task_struct *p)
3287 unsigned long long ns;
3288 unsigned long flags;
3290 local_irq_save(flags);
3291 ns = p->sched_time + sched_clock() - p->last_ran;
3292 local_irq_restore(flags);
3298 * We place interactive tasks back into the active array, if possible.
3300 * To guarantee that this does not starve expired tasks we ignore the
3301 * interactivity of a task if the first expired task had to wait more
3302 * than a 'reasonable' amount of time. This deadline timeout is
3303 * load-dependent, as the frequency of array switched decreases with
3304 * increasing number of running tasks. We also ignore the interactivity
3305 * if a better static_prio task has expired:
3307 static inline int expired_starving(struct rq *rq)
3309 if (rq->curr->static_prio > rq->best_expired_prio)
3311 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3313 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3319 * Account user cpu time to a process.
3320 * @p: the process that the cpu time gets accounted to
3321 * @hardirq_offset: the offset to subtract from hardirq_count()
3322 * @cputime: the cpu time spent in user space since the last update
3324 void account_user_time(struct task_struct *p, cputime_t cputime)
3326 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3329 p->utime = cputime_add(p->utime, cputime);
3331 /* Add user time to cpustat. */
3332 tmp = cputime_to_cputime64(cputime);
3333 if (TASK_NICE(p) > 0)
3334 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3336 cpustat->user = cputime64_add(cpustat->user, tmp);
3340 * Account system cpu time to a process.
3341 * @p: the process that the cpu time gets accounted to
3342 * @hardirq_offset: the offset to subtract from hardirq_count()
3343 * @cputime: the cpu time spent in kernel space since the last update
3345 void account_system_time(struct task_struct *p, int hardirq_offset,
3348 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3349 struct rq *rq = this_rq();
3352 p->stime = cputime_add(p->stime, cputime);
3354 /* Add system time to cpustat. */
3355 tmp = cputime_to_cputime64(cputime);
3356 if (hardirq_count() - hardirq_offset)
3357 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3358 else if (softirq_count())
3359 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3360 else if (p != rq->idle)
3361 cpustat->system = cputime64_add(cpustat->system, tmp);
3362 else if (atomic_read(&rq->nr_iowait) > 0)
3363 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3365 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3366 /* Account for system time used */
3367 acct_update_integrals(p);
3371 * Account for involuntary wait time.
3372 * @p: the process from which the cpu time has been stolen
3373 * @steal: the cpu time spent in involuntary wait
3375 void account_steal_time(struct task_struct *p, cputime_t steal)
3377 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3378 cputime64_t tmp = cputime_to_cputime64(steal);
3379 struct rq *rq = this_rq();
3381 if (p == rq->idle) {
3382 p->stime = cputime_add(p->stime, steal);
3383 if (atomic_read(&rq->nr_iowait) > 0)
3384 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3386 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3388 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3391 static void task_running_tick(struct rq *rq, struct task_struct *p)
3393 if (p->array != rq->active) {
3394 /* Task has expired but was not scheduled yet */
3395 set_tsk_need_resched(p);
3398 spin_lock(&rq->lock);
3400 * The task was running during this tick - update the
3401 * time slice counter. Note: we do not update a thread's
3402 * priority until it either goes to sleep or uses up its
3403 * timeslice. This makes it possible for interactive tasks
3404 * to use up their timeslices at their highest priority levels.
3408 * RR tasks need a special form of timeslice management.
3409 * FIFO tasks have no timeslices.
3411 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3412 p->time_slice = task_timeslice(p);
3413 p->first_time_slice = 0;
3414 set_tsk_need_resched(p);
3416 /* put it at the end of the queue: */
3417 requeue_task(p, rq->active);
3421 if (!--p->time_slice) {
3422 dequeue_task(p, rq->active);
3423 set_tsk_need_resched(p);
3424 p->prio = effective_prio(p);
3425 p->time_slice = task_timeslice(p);
3426 p->first_time_slice = 0;
3428 if (!rq->expired_timestamp)
3429 rq->expired_timestamp = jiffies;
3430 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3431 enqueue_task(p, rq->expired);
3432 if (p->static_prio < rq->best_expired_prio)
3433 rq->best_expired_prio = p->static_prio;
3435 enqueue_task(p, rq->active);
3438 * Prevent a too long timeslice allowing a task to monopolize
3439 * the CPU. We do this by splitting up the timeslice into
3442 * Note: this does not mean the task's timeslices expire or
3443 * get lost in any way, they just might be preempted by
3444 * another task of equal priority. (one with higher
3445 * priority would have preempted this task already.) We
3446 * requeue this task to the end of the list on this priority
3447 * level, which is in essence a round-robin of tasks with
3450 * This only applies to tasks in the interactive
3451 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3453 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3454 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3455 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3456 (p->array == rq->active)) {
3458 requeue_task(p, rq->active);
3459 set_tsk_need_resched(p);
3463 spin_unlock(&rq->lock);
3467 * This function gets called by the timer code, with HZ frequency.
3468 * We call it with interrupts disabled.
3470 * It also gets called by the fork code, when changing the parent's
3473 void scheduler_tick(void)
3475 unsigned long long now = sched_clock();
3476 struct task_struct *p = current;
3477 int cpu = smp_processor_id();
3478 int idle_at_tick = idle_cpu(cpu);
3479 struct rq *rq = cpu_rq(cpu);
3481 update_cpu_clock(p, rq, now);
3484 task_running_tick(rq, p);
3487 rq->idle_at_tick = idle_at_tick;
3488 trigger_load_balance(cpu);
3492 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3494 void fastcall add_preempt_count(int val)
3499 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3501 preempt_count() += val;
3503 * Spinlock count overflowing soon?
3505 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3508 EXPORT_SYMBOL(add_preempt_count);
3510 void fastcall sub_preempt_count(int val)
3515 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3518 * Is the spinlock portion underflowing?
3520 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3521 !(preempt_count() & PREEMPT_MASK)))
3524 preempt_count() -= val;
3526 EXPORT_SYMBOL(sub_preempt_count);
3530 static inline int interactive_sleep(enum sleep_type sleep_type)
3532 return (sleep_type == SLEEP_INTERACTIVE ||
3533 sleep_type == SLEEP_INTERRUPTED);
3537 * schedule() is the main scheduler function.
3539 asmlinkage void __sched schedule(void)
3541 struct task_struct *prev, *next;
3542 struct prio_array *array;
3543 struct list_head *queue;
3544 unsigned long long now;
3545 unsigned long run_time;
3546 int cpu, idx, new_prio;
3551 * Test if we are atomic. Since do_exit() needs to call into
3552 * schedule() atomically, we ignore that path for now.
3553 * Otherwise, whine if we are scheduling when we should not be.
3555 if (unlikely(in_atomic() && !current->exit_state)) {
3556 printk(KERN_ERR "BUG: scheduling while atomic: "
3558 current->comm, preempt_count(), current->pid);
3559 debug_show_held_locks(current);
3560 if (irqs_disabled())
3561 print_irqtrace_events(current);
3564 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3569 release_kernel_lock(prev);
3570 need_resched_nonpreemptible:
3574 * The idle thread is not allowed to schedule!
3575 * Remove this check after it has been exercised a bit.
3577 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3578 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3582 schedstat_inc(rq, sched_cnt);
3583 now = sched_clock();
3584 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3585 run_time = now - prev->timestamp;
3586 if (unlikely((long long)(now - prev->timestamp) < 0))
3589 run_time = NS_MAX_SLEEP_AVG;
3592 * Tasks charged proportionately less run_time at high sleep_avg to
3593 * delay them losing their interactive status
3595 run_time /= (CURRENT_BONUS(prev) ? : 1);
3597 spin_lock_irq(&rq->lock);
3599 switch_count = &prev->nivcsw;
3600 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3601 switch_count = &prev->nvcsw;
3602 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3603 unlikely(signal_pending(prev))))
3604 prev->state = TASK_RUNNING;
3606 if (prev->state == TASK_UNINTERRUPTIBLE)
3607 rq->nr_uninterruptible++;
3608 deactivate_task(prev, rq);
3612 cpu = smp_processor_id();
3613 if (unlikely(!rq->nr_running)) {
3614 idle_balance(cpu, rq);
3615 if (!rq->nr_running) {
3617 rq->expired_timestamp = 0;
3623 if (unlikely(!array->nr_active)) {
3625 * Switch the active and expired arrays.
3627 schedstat_inc(rq, sched_switch);
3628 rq->active = rq->expired;
3629 rq->expired = array;
3631 rq->expired_timestamp = 0;
3632 rq->best_expired_prio = MAX_PRIO;
3635 idx = sched_find_first_bit(array->bitmap);
3636 queue = array->queue + idx;
3637 next = list_entry(queue->next, struct task_struct, run_list);
3639 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3640 unsigned long long delta = now - next->timestamp;
3641 if (unlikely((long long)(now - next->timestamp) < 0))
3644 if (next->sleep_type == SLEEP_INTERACTIVE)
3645 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3647 array = next->array;
3648 new_prio = recalc_task_prio(next, next->timestamp + delta);
3650 if (unlikely(next->prio != new_prio)) {
3651 dequeue_task(next, array);
3652 next->prio = new_prio;
3653 enqueue_task(next, array);
3656 next->sleep_type = SLEEP_NORMAL;
3658 if (next == rq->idle)
3659 schedstat_inc(rq, sched_goidle);
3661 prefetch_stack(next);
3662 clear_tsk_need_resched(prev);
3663 rcu_qsctr_inc(task_cpu(prev));
3665 update_cpu_clock(prev, rq, now);
3667 prev->sleep_avg -= run_time;
3668 if ((long)prev->sleep_avg <= 0)
3669 prev->sleep_avg = 0;
3670 prev->timestamp = prev->last_ran = now;
3672 sched_info_switch(prev, next);
3673 if (likely(prev != next)) {
3674 next->timestamp = next->last_ran = now;
3679 prepare_task_switch(rq, next);
3680 prev = context_switch(rq, prev, next);
3683 * this_rq must be evaluated again because prev may have moved
3684 * CPUs since it called schedule(), thus the 'rq' on its stack
3685 * frame will be invalid.
3687 finish_task_switch(this_rq(), prev);
3689 spin_unlock_irq(&rq->lock);
3692 if (unlikely(reacquire_kernel_lock(prev) < 0))
3693 goto need_resched_nonpreemptible;
3694 preempt_enable_no_resched();
3695 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3698 EXPORT_SYMBOL(schedule);
3700 #ifdef CONFIG_PREEMPT
3702 * this is the entry point to schedule() from in-kernel preemption
3703 * off of preempt_enable. Kernel preemptions off return from interrupt
3704 * occur there and call schedule directly.
3706 asmlinkage void __sched preempt_schedule(void)
3708 struct thread_info *ti = current_thread_info();
3709 #ifdef CONFIG_PREEMPT_BKL
3710 struct task_struct *task = current;
3711 int saved_lock_depth;
3714 * If there is a non-zero preempt_count or interrupts are disabled,
3715 * we do not want to preempt the current task. Just return..
3717 if (likely(ti->preempt_count || irqs_disabled()))
3721 add_preempt_count(PREEMPT_ACTIVE);
3723 * We keep the big kernel semaphore locked, but we
3724 * clear ->lock_depth so that schedule() doesnt
3725 * auto-release the semaphore:
3727 #ifdef CONFIG_PREEMPT_BKL
3728 saved_lock_depth = task->lock_depth;
3729 task->lock_depth = -1;
3732 #ifdef CONFIG_PREEMPT_BKL
3733 task->lock_depth = saved_lock_depth;
3735 sub_preempt_count(PREEMPT_ACTIVE);
3737 /* we could miss a preemption opportunity between schedule and now */
3739 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3742 EXPORT_SYMBOL(preempt_schedule);
3745 * this is the entry point to schedule() from kernel preemption
3746 * off of irq context.
3747 * Note, that this is called and return with irqs disabled. This will
3748 * protect us against recursive calling from irq.
3750 asmlinkage void __sched preempt_schedule_irq(void)
3752 struct thread_info *ti = current_thread_info();
3753 #ifdef CONFIG_PREEMPT_BKL
3754 struct task_struct *task = current;
3755 int saved_lock_depth;
3757 /* Catch callers which need to be fixed */
3758 BUG_ON(ti->preempt_count || !irqs_disabled());
3761 add_preempt_count(PREEMPT_ACTIVE);
3763 * We keep the big kernel semaphore locked, but we
3764 * clear ->lock_depth so that schedule() doesnt
3765 * auto-release the semaphore:
3767 #ifdef CONFIG_PREEMPT_BKL
3768 saved_lock_depth = task->lock_depth;
3769 task->lock_depth = -1;
3773 local_irq_disable();
3774 #ifdef CONFIG_PREEMPT_BKL
3775 task->lock_depth = saved_lock_depth;
3777 sub_preempt_count(PREEMPT_ACTIVE);
3779 /* we could miss a preemption opportunity between schedule and now */
3781 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3785 #endif /* CONFIG_PREEMPT */
3787 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3790 return try_to_wake_up(curr->private, mode, sync);
3792 EXPORT_SYMBOL(default_wake_function);
3795 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3796 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3797 * number) then we wake all the non-exclusive tasks and one exclusive task.
3799 * There are circumstances in which we can try to wake a task which has already
3800 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3801 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3803 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3804 int nr_exclusive, int sync, void *key)
3806 struct list_head *tmp, *next;
3808 list_for_each_safe(tmp, next, &q->task_list) {
3809 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3810 unsigned flags = curr->flags;
3812 if (curr->func(curr, mode, sync, key) &&
3813 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3819 * __wake_up - wake up threads blocked on a waitqueue.
3821 * @mode: which threads
3822 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3823 * @key: is directly passed to the wakeup function
3825 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3826 int nr_exclusive, void *key)
3828 unsigned long flags;
3830 spin_lock_irqsave(&q->lock, flags);
3831 __wake_up_common(q, mode, nr_exclusive, 0, key);
3832 spin_unlock_irqrestore(&q->lock, flags);
3834 EXPORT_SYMBOL(__wake_up);
3837 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3839 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3841 __wake_up_common(q, mode, 1, 0, NULL);
3845 * __wake_up_sync - wake up threads blocked on a waitqueue.
3847 * @mode: which threads
3848 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3850 * The sync wakeup differs that the waker knows that it will schedule
3851 * away soon, so while the target thread will be woken up, it will not
3852 * be migrated to another CPU - ie. the two threads are 'synchronized'
3853 * with each other. This can prevent needless bouncing between CPUs.
3855 * On UP it can prevent extra preemption.
3858 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3860 unsigned long flags;
3866 if (unlikely(!nr_exclusive))
3869 spin_lock_irqsave(&q->lock, flags);
3870 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3871 spin_unlock_irqrestore(&q->lock, flags);
3873 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3875 void fastcall complete(struct completion *x)
3877 unsigned long flags;
3879 spin_lock_irqsave(&x->wait.lock, flags);
3881 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3883 spin_unlock_irqrestore(&x->wait.lock, flags);
3885 EXPORT_SYMBOL(complete);
3887 void fastcall complete_all(struct completion *x)
3889 unsigned long flags;
3891 spin_lock_irqsave(&x->wait.lock, flags);
3892 x->done += UINT_MAX/2;
3893 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3895 spin_unlock_irqrestore(&x->wait.lock, flags);
3897 EXPORT_SYMBOL(complete_all);
3899 void fastcall __sched wait_for_completion(struct completion *x)
3903 spin_lock_irq(&x->wait.lock);
3905 DECLARE_WAITQUEUE(wait, current);
3907 wait.flags |= WQ_FLAG_EXCLUSIVE;
3908 __add_wait_queue_tail(&x->wait, &wait);
3910 __set_current_state(TASK_UNINTERRUPTIBLE);
3911 spin_unlock_irq(&x->wait.lock);
3913 spin_lock_irq(&x->wait.lock);
3915 __remove_wait_queue(&x->wait, &wait);
3918 spin_unlock_irq(&x->wait.lock);
3920 EXPORT_SYMBOL(wait_for_completion);
3922 unsigned long fastcall __sched
3923 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3927 spin_lock_irq(&x->wait.lock);
3929 DECLARE_WAITQUEUE(wait, current);
3931 wait.flags |= WQ_FLAG_EXCLUSIVE;
3932 __add_wait_queue_tail(&x->wait, &wait);
3934 __set_current_state(TASK_UNINTERRUPTIBLE);
3935 spin_unlock_irq(&x->wait.lock);
3936 timeout = schedule_timeout(timeout);
3937 spin_lock_irq(&x->wait.lock);
3939 __remove_wait_queue(&x->wait, &wait);
3943 __remove_wait_queue(&x->wait, &wait);
3947 spin_unlock_irq(&x->wait.lock);
3950 EXPORT_SYMBOL(wait_for_completion_timeout);
3952 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3958 spin_lock_irq(&x->wait.lock);
3960 DECLARE_WAITQUEUE(wait, current);
3962 wait.flags |= WQ_FLAG_EXCLUSIVE;
3963 __add_wait_queue_tail(&x->wait, &wait);
3965 if (signal_pending(current)) {
3967 __remove_wait_queue(&x->wait, &wait);
3970 __set_current_state(TASK_INTERRUPTIBLE);
3971 spin_unlock_irq(&x->wait.lock);
3973 spin_lock_irq(&x->wait.lock);
3975 __remove_wait_queue(&x->wait, &wait);
3979 spin_unlock_irq(&x->wait.lock);
3983 EXPORT_SYMBOL(wait_for_completion_interruptible);
3985 unsigned long fastcall __sched
3986 wait_for_completion_interruptible_timeout(struct completion *x,
3987 unsigned long timeout)
3991 spin_lock_irq(&x->wait.lock);
3993 DECLARE_WAITQUEUE(wait, current);
3995 wait.flags |= WQ_FLAG_EXCLUSIVE;
3996 __add_wait_queue_tail(&x->wait, &wait);
3998 if (signal_pending(current)) {
3999 timeout = -ERESTARTSYS;
4000 __remove_wait_queue(&x->wait, &wait);
4003 __set_current_state(TASK_INTERRUPTIBLE);
4004 spin_unlock_irq(&x->wait.lock);
4005 timeout = schedule_timeout(timeout);
4006 spin_lock_irq(&x->wait.lock);
4008 __remove_wait_queue(&x->wait, &wait);
4012 __remove_wait_queue(&x->wait, &wait);
4016 spin_unlock_irq(&x->wait.lock);
4019 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4022 #define SLEEP_ON_VAR \
4023 unsigned long flags; \
4024 wait_queue_t wait; \
4025 init_waitqueue_entry(&wait, current);
4027 #define SLEEP_ON_HEAD \
4028 spin_lock_irqsave(&q->lock,flags); \
4029 __add_wait_queue(q, &wait); \
4030 spin_unlock(&q->lock);
4032 #define SLEEP_ON_TAIL \
4033 spin_lock_irq(&q->lock); \
4034 __remove_wait_queue(q, &wait); \
4035 spin_unlock_irqrestore(&q->lock, flags);
4037 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
4041 current->state = TASK_INTERRUPTIBLE;
4047 EXPORT_SYMBOL(interruptible_sleep_on);
4049 long fastcall __sched
4050 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4054 current->state = TASK_INTERRUPTIBLE;
4057 timeout = schedule_timeout(timeout);
4062 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4064 void fastcall __sched sleep_on(wait_queue_head_t *q)
4068 current->state = TASK_UNINTERRUPTIBLE;
4074 EXPORT_SYMBOL(sleep_on);
4076 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4080 current->state = TASK_UNINTERRUPTIBLE;
4083 timeout = schedule_timeout(timeout);
4089 EXPORT_SYMBOL(sleep_on_timeout);
4091 #ifdef CONFIG_RT_MUTEXES
4094 * rt_mutex_setprio - set the current priority of a task
4096 * @prio: prio value (kernel-internal form)
4098 * This function changes the 'effective' priority of a task. It does
4099 * not touch ->normal_prio like __setscheduler().
4101 * Used by the rt_mutex code to implement priority inheritance logic.
4103 void rt_mutex_setprio(struct task_struct *p, int prio)
4105 struct prio_array *array;
4106 unsigned long flags;
4110 BUG_ON(prio < 0 || prio > MAX_PRIO);
4112 rq = task_rq_lock(p, &flags);
4117 dequeue_task(p, array);
4122 * If changing to an RT priority then queue it
4123 * in the active array!
4127 enqueue_task(p, array);
4129 * Reschedule if we are currently running on this runqueue and
4130 * our priority decreased, or if we are not currently running on
4131 * this runqueue and our priority is higher than the current's
4133 if (task_running(rq, p)) {
4134 if (p->prio > oldprio)
4135 resched_task(rq->curr);
4136 } else if (TASK_PREEMPTS_CURR(p, rq))
4137 resched_task(rq->curr);
4139 task_rq_unlock(rq, &flags);
4144 void set_user_nice(struct task_struct *p, long nice)
4146 struct prio_array *array;
4147 int old_prio, delta;
4148 unsigned long flags;
4151 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4154 * We have to be careful, if called from sys_setpriority(),
4155 * the task might be in the middle of scheduling on another CPU.
4157 rq = task_rq_lock(p, &flags);
4159 * The RT priorities are set via sched_setscheduler(), but we still
4160 * allow the 'normal' nice value to be set - but as expected
4161 * it wont have any effect on scheduling until the task is
4162 * not SCHED_NORMAL/SCHED_BATCH:
4164 if (has_rt_policy(p)) {
4165 p->static_prio = NICE_TO_PRIO(nice);
4170 dequeue_task(p, array);
4171 dec_raw_weighted_load(rq, p);
4174 p->static_prio = NICE_TO_PRIO(nice);
4177 p->prio = effective_prio(p);
4178 delta = p->prio - old_prio;
4181 enqueue_task(p, array);
4182 inc_raw_weighted_load(rq, p);
4184 * If the task increased its priority or is running and
4185 * lowered its priority, then reschedule its CPU:
4187 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4188 resched_task(rq->curr);
4191 task_rq_unlock(rq, &flags);
4193 EXPORT_SYMBOL(set_user_nice);
4196 * can_nice - check if a task can reduce its nice value
4200 int can_nice(const struct task_struct *p, const int nice)
4202 /* convert nice value [19,-20] to rlimit style value [1,40] */
4203 int nice_rlim = 20 - nice;
4205 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4206 capable(CAP_SYS_NICE));
4209 #ifdef __ARCH_WANT_SYS_NICE
4212 * sys_nice - change the priority of the current process.
4213 * @increment: priority increment
4215 * sys_setpriority is a more generic, but much slower function that
4216 * does similar things.
4218 asmlinkage long sys_nice(int increment)
4223 * Setpriority might change our priority at the same moment.
4224 * We don't have to worry. Conceptually one call occurs first
4225 * and we have a single winner.
4227 if (increment < -40)
4232 nice = PRIO_TO_NICE(current->static_prio) + increment;
4238 if (increment < 0 && !can_nice(current, nice))
4241 retval = security_task_setnice(current, nice);
4245 set_user_nice(current, nice);
4252 * task_prio - return the priority value of a given task.
4253 * @p: the task in question.
4255 * This is the priority value as seen by users in /proc.
4256 * RT tasks are offset by -200. Normal tasks are centered
4257 * around 0, value goes from -16 to +15.
4259 int task_prio(const struct task_struct *p)
4261 return p->prio - MAX_RT_PRIO;
4265 * task_nice - return the nice value of a given task.
4266 * @p: the task in question.
4268 int task_nice(const struct task_struct *p)
4270 return TASK_NICE(p);
4272 EXPORT_SYMBOL_GPL(task_nice);
4275 * idle_cpu - is a given cpu idle currently?
4276 * @cpu: the processor in question.
4278 int idle_cpu(int cpu)
4280 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4284 * idle_task - return the idle task for a given cpu.
4285 * @cpu: the processor in question.
4287 struct task_struct *idle_task(int cpu)
4289 return cpu_rq(cpu)->idle;
4293 * find_process_by_pid - find a process with a matching PID value.
4294 * @pid: the pid in question.
4296 static inline struct task_struct *find_process_by_pid(pid_t pid)
4298 return pid ? find_task_by_pid(pid) : current;
4301 /* Actually do priority change: must hold rq lock. */
4302 static void __setscheduler(struct task_struct *p, int policy, int prio)
4307 p->rt_priority = prio;
4308 p->normal_prio = normal_prio(p);
4309 /* we are holding p->pi_lock already */
4310 p->prio = rt_mutex_getprio(p);
4312 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4314 if (policy == SCHED_BATCH)
4320 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4321 * @p: the task in question.
4322 * @policy: new policy.
4323 * @param: structure containing the new RT priority.
4325 * NOTE that the task may be already dead.
4327 int sched_setscheduler(struct task_struct *p, int policy,
4328 struct sched_param *param)
4330 int retval, oldprio, oldpolicy = -1;
4331 struct prio_array *array;
4332 unsigned long flags;
4335 /* may grab non-irq protected spin_locks */
4336 BUG_ON(in_interrupt());
4338 /* double check policy once rq lock held */
4340 policy = oldpolicy = p->policy;
4341 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4342 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4345 * Valid priorities for SCHED_FIFO and SCHED_RR are
4346 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4349 if (param->sched_priority < 0 ||
4350 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4351 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4353 if (is_rt_policy(policy) != (param->sched_priority != 0))
4357 * Allow unprivileged RT tasks to decrease priority:
4359 if (!capable(CAP_SYS_NICE)) {
4360 if (is_rt_policy(policy)) {
4361 unsigned long rlim_rtprio;
4362 unsigned long flags;
4364 if (!lock_task_sighand(p, &flags))
4366 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4367 unlock_task_sighand(p, &flags);
4369 /* can't set/change the rt policy */
4370 if (policy != p->policy && !rlim_rtprio)
4373 /* can't increase priority */
4374 if (param->sched_priority > p->rt_priority &&
4375 param->sched_priority > rlim_rtprio)
4379 /* can't change other user's priorities */
4380 if ((current->euid != p->euid) &&
4381 (current->euid != p->uid))
4385 retval = security_task_setscheduler(p, policy, param);
4389 * make sure no PI-waiters arrive (or leave) while we are
4390 * changing the priority of the task:
4392 spin_lock_irqsave(&p->pi_lock, flags);
4394 * To be able to change p->policy safely, the apropriate
4395 * runqueue lock must be held.
4397 rq = __task_rq_lock(p);
4398 /* recheck policy now with rq lock held */
4399 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4400 policy = oldpolicy = -1;
4401 __task_rq_unlock(rq);
4402 spin_unlock_irqrestore(&p->pi_lock, flags);
4407 deactivate_task(p, rq);
4409 __setscheduler(p, policy, param->sched_priority);
4411 __activate_task(p, rq);
4413 * Reschedule if we are currently running on this runqueue and
4414 * our priority decreased, or if we are not currently running on
4415 * this runqueue and our priority is higher than the current's
4417 if (task_running(rq, p)) {
4418 if (p->prio > oldprio)
4419 resched_task(rq->curr);
4420 } else if (TASK_PREEMPTS_CURR(p, rq))
4421 resched_task(rq->curr);
4423 __task_rq_unlock(rq);
4424 spin_unlock_irqrestore(&p->pi_lock, flags);
4426 rt_mutex_adjust_pi(p);
4430 EXPORT_SYMBOL_GPL(sched_setscheduler);
4433 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4435 struct sched_param lparam;
4436 struct task_struct *p;
4439 if (!param || pid < 0)
4441 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4446 p = find_process_by_pid(pid);
4448 retval = sched_setscheduler(p, policy, &lparam);
4455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4456 * @pid: the pid in question.
4457 * @policy: new policy.
4458 * @param: structure containing the new RT priority.
4460 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4461 struct sched_param __user *param)
4463 /* negative values for policy are not valid */
4467 return do_sched_setscheduler(pid, policy, param);
4471 * sys_sched_setparam - set/change the RT priority of a thread
4472 * @pid: the pid in question.
4473 * @param: structure containing the new RT priority.
4475 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4477 return do_sched_setscheduler(pid, -1, param);
4481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4482 * @pid: the pid in question.
4484 asmlinkage long sys_sched_getscheduler(pid_t pid)
4486 struct task_struct *p;
4487 int retval = -EINVAL;
4493 read_lock(&tasklist_lock);
4494 p = find_process_by_pid(pid);
4496 retval = security_task_getscheduler(p);
4500 read_unlock(&tasklist_lock);
4507 * sys_sched_getscheduler - get the RT priority of a thread
4508 * @pid: the pid in question.
4509 * @param: structure containing the RT priority.
4511 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4513 struct sched_param lp;
4514 struct task_struct *p;
4515 int retval = -EINVAL;
4517 if (!param || pid < 0)
4520 read_lock(&tasklist_lock);
4521 p = find_process_by_pid(pid);
4526 retval = security_task_getscheduler(p);
4530 lp.sched_priority = p->rt_priority;
4531 read_unlock(&tasklist_lock);
4534 * This one might sleep, we cannot do it with a spinlock held ...
4536 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4542 read_unlock(&tasklist_lock);
4546 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4548 cpumask_t cpus_allowed;
4549 struct task_struct *p;
4552 mutex_lock(&sched_hotcpu_mutex);
4553 read_lock(&tasklist_lock);
4555 p = find_process_by_pid(pid);
4557 read_unlock(&tasklist_lock);
4558 mutex_unlock(&sched_hotcpu_mutex);
4563 * It is not safe to call set_cpus_allowed with the
4564 * tasklist_lock held. We will bump the task_struct's
4565 * usage count and then drop tasklist_lock.
4568 read_unlock(&tasklist_lock);
4571 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4572 !capable(CAP_SYS_NICE))
4575 retval = security_task_setscheduler(p, 0, NULL);
4579 cpus_allowed = cpuset_cpus_allowed(p);
4580 cpus_and(new_mask, new_mask, cpus_allowed);
4581 retval = set_cpus_allowed(p, new_mask);
4585 mutex_unlock(&sched_hotcpu_mutex);
4589 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4590 cpumask_t *new_mask)
4592 if (len < sizeof(cpumask_t)) {
4593 memset(new_mask, 0, sizeof(cpumask_t));
4594 } else if (len > sizeof(cpumask_t)) {
4595 len = sizeof(cpumask_t);
4597 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4601 * sys_sched_setaffinity - set the cpu affinity of a process
4602 * @pid: pid of the process
4603 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4604 * @user_mask_ptr: user-space pointer to the new cpu mask
4606 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4607 unsigned long __user *user_mask_ptr)
4612 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4616 return sched_setaffinity(pid, new_mask);
4620 * Represents all cpu's present in the system
4621 * In systems capable of hotplug, this map could dynamically grow
4622 * as new cpu's are detected in the system via any platform specific
4623 * method, such as ACPI for e.g.
4626 cpumask_t cpu_present_map __read_mostly;
4627 EXPORT_SYMBOL(cpu_present_map);
4630 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4631 EXPORT_SYMBOL(cpu_online_map);
4633 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4634 EXPORT_SYMBOL(cpu_possible_map);
4637 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4639 struct task_struct *p;
4642 mutex_lock(&sched_hotcpu_mutex);
4643 read_lock(&tasklist_lock);
4646 p = find_process_by_pid(pid);
4650 retval = security_task_getscheduler(p);
4654 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4657 read_unlock(&tasklist_lock);
4658 mutex_unlock(&sched_hotcpu_mutex);
4666 * sys_sched_getaffinity - get the cpu affinity of a process
4667 * @pid: pid of the process
4668 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4669 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4671 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4672 unsigned long __user *user_mask_ptr)
4677 if (len < sizeof(cpumask_t))
4680 ret = sched_getaffinity(pid, &mask);
4684 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4687 return sizeof(cpumask_t);
4691 * sys_sched_yield - yield the current processor to other threads.
4693 * This function yields the current CPU by moving the calling thread
4694 * to the expired array. If there are no other threads running on this
4695 * CPU then this function will return.
4697 asmlinkage long sys_sched_yield(void)
4699 struct rq *rq = this_rq_lock();
4700 struct prio_array *array = current->array, *target = rq->expired;
4702 schedstat_inc(rq, yld_cnt);
4704 * We implement yielding by moving the task into the expired
4707 * (special rule: RT tasks will just roundrobin in the active
4710 if (rt_task(current))
4711 target = rq->active;
4713 if (array->nr_active == 1) {
4714 schedstat_inc(rq, yld_act_empty);
4715 if (!rq->expired->nr_active)
4716 schedstat_inc(rq, yld_both_empty);
4717 } else if (!rq->expired->nr_active)
4718 schedstat_inc(rq, yld_exp_empty);
4720 if (array != target) {
4721 dequeue_task(current, array);
4722 enqueue_task(current, target);
4725 * requeue_task is cheaper so perform that if possible.
4727 requeue_task(current, array);
4730 * Since we are going to call schedule() anyway, there's
4731 * no need to preempt or enable interrupts:
4733 __release(rq->lock);
4734 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4735 _raw_spin_unlock(&rq->lock);
4736 preempt_enable_no_resched();
4743 static void __cond_resched(void)
4745 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4746 __might_sleep(__FILE__, __LINE__);
4749 * The BKS might be reacquired before we have dropped
4750 * PREEMPT_ACTIVE, which could trigger a second
4751 * cond_resched() call.
4754 add_preempt_count(PREEMPT_ACTIVE);
4756 sub_preempt_count(PREEMPT_ACTIVE);
4757 } while (need_resched());
4760 int __sched cond_resched(void)
4762 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4763 system_state == SYSTEM_RUNNING) {
4769 EXPORT_SYMBOL(cond_resched);
4772 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4773 * call schedule, and on return reacquire the lock.
4775 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4776 * operations here to prevent schedule() from being called twice (once via
4777 * spin_unlock(), once by hand).
4779 int cond_resched_lock(spinlock_t *lock)
4783 if (need_lockbreak(lock)) {
4789 if (need_resched() && system_state == SYSTEM_RUNNING) {
4790 spin_release(&lock->dep_map, 1, _THIS_IP_);
4791 _raw_spin_unlock(lock);
4792 preempt_enable_no_resched();
4799 EXPORT_SYMBOL(cond_resched_lock);
4801 int __sched cond_resched_softirq(void)
4803 BUG_ON(!in_softirq());
4805 if (need_resched() && system_state == SYSTEM_RUNNING) {
4813 EXPORT_SYMBOL(cond_resched_softirq);
4816 * yield - yield the current processor to other threads.
4818 * This is a shortcut for kernel-space yielding - it marks the
4819 * thread runnable and calls sys_sched_yield().
4821 void __sched yield(void)
4823 set_current_state(TASK_RUNNING);
4826 EXPORT_SYMBOL(yield);
4829 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4830 * that process accounting knows that this is a task in IO wait state.
4832 * But don't do that if it is a deliberate, throttling IO wait (this task
4833 * has set its backing_dev_info: the queue against which it should throttle)
4835 void __sched io_schedule(void)
4837 struct rq *rq = &__raw_get_cpu_var(runqueues);
4839 delayacct_blkio_start();
4840 atomic_inc(&rq->nr_iowait);
4842 atomic_dec(&rq->nr_iowait);
4843 delayacct_blkio_end();
4845 EXPORT_SYMBOL(io_schedule);
4847 long __sched io_schedule_timeout(long timeout)
4849 struct rq *rq = &__raw_get_cpu_var(runqueues);
4852 delayacct_blkio_start();
4853 atomic_inc(&rq->nr_iowait);
4854 ret = schedule_timeout(timeout);
4855 atomic_dec(&rq->nr_iowait);
4856 delayacct_blkio_end();
4861 * sys_sched_get_priority_max - return maximum RT priority.
4862 * @policy: scheduling class.
4864 * this syscall returns the maximum rt_priority that can be used
4865 * by a given scheduling class.
4867 asmlinkage long sys_sched_get_priority_max(int policy)
4874 ret = MAX_USER_RT_PRIO-1;
4885 * sys_sched_get_priority_min - return minimum RT priority.
4886 * @policy: scheduling class.
4888 * this syscall returns the minimum rt_priority that can be used
4889 * by a given scheduling class.
4891 asmlinkage long sys_sched_get_priority_min(int policy)
4908 * sys_sched_rr_get_interval - return the default timeslice of a process.
4909 * @pid: pid of the process.
4910 * @interval: userspace pointer to the timeslice value.
4912 * this syscall writes the default timeslice value of a given process
4913 * into the user-space timespec buffer. A value of '0' means infinity.
4916 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4918 struct task_struct *p;
4919 int retval = -EINVAL;
4926 read_lock(&tasklist_lock);
4927 p = find_process_by_pid(pid);
4931 retval = security_task_getscheduler(p);
4935 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4936 0 : task_timeslice(p), &t);
4937 read_unlock(&tasklist_lock);
4938 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4942 read_unlock(&tasklist_lock);
4946 static const char stat_nam[] = "RSDTtZX";
4948 static void show_task(struct task_struct *p)
4950 unsigned long free = 0;
4953 state = p->state ? __ffs(p->state) + 1 : 0;
4954 printk("%-13.13s %c", p->comm,
4955 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4956 #if (BITS_PER_LONG == 32)
4957 if (state == TASK_RUNNING)
4958 printk(" running ");
4960 printk(" %08lX ", thread_saved_pc(p));
4962 if (state == TASK_RUNNING)
4963 printk(" running task ");
4965 printk(" %016lx ", thread_saved_pc(p));
4967 #ifdef CONFIG_DEBUG_STACK_USAGE
4969 unsigned long *n = end_of_stack(p);
4972 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4975 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4977 printk(" (L-TLB)\n");
4979 printk(" (NOTLB)\n");
4981 if (state != TASK_RUNNING)
4982 show_stack(p, NULL);
4985 void show_state_filter(unsigned long state_filter)
4987 struct task_struct *g, *p;
4989 #if (BITS_PER_LONG == 32)
4992 printk(" task PC stack pid father child younger older\n");
4996 printk(" task PC stack pid father child younger older\n");
4998 read_lock(&tasklist_lock);
4999 do_each_thread(g, p) {
5001 * reset the NMI-timeout, listing all files on a slow
5002 * console might take alot of time:
5004 touch_nmi_watchdog();
5005 if (!state_filter || (p->state & state_filter))
5007 } while_each_thread(g, p);
5009 touch_all_softlockup_watchdogs();
5011 read_unlock(&tasklist_lock);
5013 * Only show locks if all tasks are dumped:
5015 if (state_filter == -1)
5016 debug_show_all_locks();
5020 * init_idle - set up an idle thread for a given CPU
5021 * @idle: task in question
5022 * @cpu: cpu the idle task belongs to
5024 * NOTE: this function does not set the idle thread's NEED_RESCHED
5025 * flag, to make booting more robust.
5027 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5029 struct rq *rq = cpu_rq(cpu);
5030 unsigned long flags;
5032 idle->timestamp = sched_clock();
5033 idle->sleep_avg = 0;
5035 idle->prio = idle->normal_prio = MAX_PRIO;
5036 idle->state = TASK_RUNNING;
5037 idle->cpus_allowed = cpumask_of_cpu(cpu);
5038 set_task_cpu(idle, cpu);
5040 spin_lock_irqsave(&rq->lock, flags);
5041 rq->curr = rq->idle = idle;
5042 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5045 spin_unlock_irqrestore(&rq->lock, flags);
5047 /* Set the preempt count _outside_ the spinlocks! */
5048 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5049 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5051 task_thread_info(idle)->preempt_count = 0;
5056 * In a system that switches off the HZ timer nohz_cpu_mask
5057 * indicates which cpus entered this state. This is used
5058 * in the rcu update to wait only for active cpus. For system
5059 * which do not switch off the HZ timer nohz_cpu_mask should
5060 * always be CPU_MASK_NONE.
5062 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5066 * This is how migration works:
5068 * 1) we queue a struct migration_req structure in the source CPU's
5069 * runqueue and wake up that CPU's migration thread.
5070 * 2) we down() the locked semaphore => thread blocks.
5071 * 3) migration thread wakes up (implicitly it forces the migrated
5072 * thread off the CPU)
5073 * 4) it gets the migration request and checks whether the migrated
5074 * task is still in the wrong runqueue.
5075 * 5) if it's in the wrong runqueue then the migration thread removes
5076 * it and puts it into the right queue.
5077 * 6) migration thread up()s the semaphore.
5078 * 7) we wake up and the migration is done.
5082 * Change a given task's CPU affinity. Migrate the thread to a
5083 * proper CPU and schedule it away if the CPU it's executing on
5084 * is removed from the allowed bitmask.
5086 * NOTE: the caller must have a valid reference to the task, the
5087 * task must not exit() & deallocate itself prematurely. The
5088 * call is not atomic; no spinlocks may be held.
5090 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5092 struct migration_req req;
5093 unsigned long flags;
5097 rq = task_rq_lock(p, &flags);
5098 if (!cpus_intersects(new_mask, cpu_online_map)) {
5103 p->cpus_allowed = new_mask;
5104 /* Can the task run on the task's current CPU? If so, we're done */
5105 if (cpu_isset(task_cpu(p), new_mask))
5108 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5109 /* Need help from migration thread: drop lock and wait. */
5110 task_rq_unlock(rq, &flags);
5111 wake_up_process(rq->migration_thread);
5112 wait_for_completion(&req.done);
5113 tlb_migrate_finish(p->mm);
5117 task_rq_unlock(rq, &flags);
5121 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5124 * Move (not current) task off this cpu, onto dest cpu. We're doing
5125 * this because either it can't run here any more (set_cpus_allowed()
5126 * away from this CPU, or CPU going down), or because we're
5127 * attempting to rebalance this task on exec (sched_exec).
5129 * So we race with normal scheduler movements, but that's OK, as long
5130 * as the task is no longer on this CPU.
5132 * Returns non-zero if task was successfully migrated.
5134 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5136 struct rq *rq_dest, *rq_src;
5139 if (unlikely(cpu_is_offline(dest_cpu)))
5142 rq_src = cpu_rq(src_cpu);
5143 rq_dest = cpu_rq(dest_cpu);
5145 double_rq_lock(rq_src, rq_dest);
5146 /* Already moved. */
5147 if (task_cpu(p) != src_cpu)
5149 /* Affinity changed (again). */
5150 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5153 set_task_cpu(p, dest_cpu);
5156 * Sync timestamp with rq_dest's before activating.
5157 * The same thing could be achieved by doing this step
5158 * afterwards, and pretending it was a local activate.
5159 * This way is cleaner and logically correct.
5161 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5162 + rq_dest->most_recent_timestamp;
5163 deactivate_task(p, rq_src);
5164 __activate_task(p, rq_dest);
5165 if (TASK_PREEMPTS_CURR(p, rq_dest))
5166 resched_task(rq_dest->curr);
5170 double_rq_unlock(rq_src, rq_dest);
5175 * migration_thread - this is a highprio system thread that performs
5176 * thread migration by bumping thread off CPU then 'pushing' onto
5179 static int migration_thread(void *data)
5181 int cpu = (long)data;
5185 BUG_ON(rq->migration_thread != current);
5187 set_current_state(TASK_INTERRUPTIBLE);
5188 while (!kthread_should_stop()) {
5189 struct migration_req *req;
5190 struct list_head *head;
5194 spin_lock_irq(&rq->lock);
5196 if (cpu_is_offline(cpu)) {
5197 spin_unlock_irq(&rq->lock);
5201 if (rq->active_balance) {
5202 active_load_balance(rq, cpu);
5203 rq->active_balance = 0;
5206 head = &rq->migration_queue;
5208 if (list_empty(head)) {
5209 spin_unlock_irq(&rq->lock);
5211 set_current_state(TASK_INTERRUPTIBLE);
5214 req = list_entry(head->next, struct migration_req, list);
5215 list_del_init(head->next);
5217 spin_unlock(&rq->lock);
5218 __migrate_task(req->task, cpu, req->dest_cpu);
5221 complete(&req->done);
5223 __set_current_state(TASK_RUNNING);
5227 /* Wait for kthread_stop */
5228 set_current_state(TASK_INTERRUPTIBLE);
5229 while (!kthread_should_stop()) {
5231 set_current_state(TASK_INTERRUPTIBLE);
5233 __set_current_state(TASK_RUNNING);
5237 #ifdef CONFIG_HOTPLUG_CPU
5239 * Figure out where task on dead CPU should go, use force if neccessary.
5240 * NOTE: interrupts should be disabled by the caller
5242 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5244 unsigned long flags;
5251 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5252 cpus_and(mask, mask, p->cpus_allowed);
5253 dest_cpu = any_online_cpu(mask);
5255 /* On any allowed CPU? */
5256 if (dest_cpu == NR_CPUS)
5257 dest_cpu = any_online_cpu(p->cpus_allowed);
5259 /* No more Mr. Nice Guy. */
5260 if (dest_cpu == NR_CPUS) {
5261 rq = task_rq_lock(p, &flags);
5262 cpus_setall(p->cpus_allowed);
5263 dest_cpu = any_online_cpu(p->cpus_allowed);
5264 task_rq_unlock(rq, &flags);
5267 * Don't tell them about moving exiting tasks or
5268 * kernel threads (both mm NULL), since they never
5271 if (p->mm && printk_ratelimit())
5272 printk(KERN_INFO "process %d (%s) no "
5273 "longer affine to cpu%d\n",
5274 p->pid, p->comm, dead_cpu);
5276 if (!__migrate_task(p, dead_cpu, dest_cpu))
5281 * While a dead CPU has no uninterruptible tasks queued at this point,
5282 * it might still have a nonzero ->nr_uninterruptible counter, because
5283 * for performance reasons the counter is not stricly tracking tasks to
5284 * their home CPUs. So we just add the counter to another CPU's counter,
5285 * to keep the global sum constant after CPU-down:
5287 static void migrate_nr_uninterruptible(struct rq *rq_src)
5289 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5290 unsigned long flags;
5292 local_irq_save(flags);
5293 double_rq_lock(rq_src, rq_dest);
5294 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5295 rq_src->nr_uninterruptible = 0;
5296 double_rq_unlock(rq_src, rq_dest);
5297 local_irq_restore(flags);
5300 /* Run through task list and migrate tasks from the dead cpu. */
5301 static void migrate_live_tasks(int src_cpu)
5303 struct task_struct *p, *t;
5305 write_lock_irq(&tasklist_lock);
5307 do_each_thread(t, p) {
5311 if (task_cpu(p) == src_cpu)
5312 move_task_off_dead_cpu(src_cpu, p);
5313 } while_each_thread(t, p);
5315 write_unlock_irq(&tasklist_lock);
5318 /* Schedules idle task to be the next runnable task on current CPU.
5319 * It does so by boosting its priority to highest possible and adding it to
5320 * the _front_ of the runqueue. Used by CPU offline code.
5322 void sched_idle_next(void)
5324 int this_cpu = smp_processor_id();
5325 struct rq *rq = cpu_rq(this_cpu);
5326 struct task_struct *p = rq->idle;
5327 unsigned long flags;
5329 /* cpu has to be offline */
5330 BUG_ON(cpu_online(this_cpu));
5333 * Strictly not necessary since rest of the CPUs are stopped by now
5334 * and interrupts disabled on the current cpu.
5336 spin_lock_irqsave(&rq->lock, flags);
5338 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5340 /* Add idle task to the _front_ of its priority queue: */
5341 __activate_idle_task(p, rq);
5343 spin_unlock_irqrestore(&rq->lock, flags);
5347 * Ensures that the idle task is using init_mm right before its cpu goes
5350 void idle_task_exit(void)
5352 struct mm_struct *mm = current->active_mm;
5354 BUG_ON(cpu_online(smp_processor_id()));
5357 switch_mm(mm, &init_mm, current);
5361 /* called under rq->lock with disabled interrupts */
5362 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5364 struct rq *rq = cpu_rq(dead_cpu);
5366 /* Must be exiting, otherwise would be on tasklist. */
5367 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5369 /* Cannot have done final schedule yet: would have vanished. */
5370 BUG_ON(p->state == TASK_DEAD);
5375 * Drop lock around migration; if someone else moves it,
5376 * that's OK. No task can be added to this CPU, so iteration is
5378 * NOTE: interrupts should be left disabled --dev@
5380 spin_unlock(&rq->lock);
5381 move_task_off_dead_cpu(dead_cpu, p);
5382 spin_lock(&rq->lock);
5387 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5388 static void migrate_dead_tasks(unsigned int dead_cpu)
5390 struct rq *rq = cpu_rq(dead_cpu);
5391 unsigned int arr, i;
5393 for (arr = 0; arr < 2; arr++) {
5394 for (i = 0; i < MAX_PRIO; i++) {
5395 struct list_head *list = &rq->arrays[arr].queue[i];
5397 while (!list_empty(list))
5398 migrate_dead(dead_cpu, list_entry(list->next,
5399 struct task_struct, run_list));
5403 #endif /* CONFIG_HOTPLUG_CPU */
5406 * migration_call - callback that gets triggered when a CPU is added.
5407 * Here we can start up the necessary migration thread for the new CPU.
5409 static int __cpuinit
5410 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5412 struct task_struct *p;
5413 int cpu = (long)hcpu;
5414 unsigned long flags;
5418 case CPU_LOCK_ACQUIRE:
5419 mutex_lock(&sched_hotcpu_mutex);
5422 case CPU_UP_PREPARE:
5423 case CPU_UP_PREPARE_FROZEN:
5424 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5427 p->flags |= PF_NOFREEZE;
5428 kthread_bind(p, cpu);
5429 /* Must be high prio: stop_machine expects to yield to it. */
5430 rq = task_rq_lock(p, &flags);
5431 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5432 task_rq_unlock(rq, &flags);
5433 cpu_rq(cpu)->migration_thread = p;
5437 case CPU_ONLINE_FROZEN:
5438 /* Strictly unneccessary, as first user will wake it. */
5439 wake_up_process(cpu_rq(cpu)->migration_thread);
5442 #ifdef CONFIG_HOTPLUG_CPU
5443 case CPU_UP_CANCELED:
5444 case CPU_UP_CANCELED_FROZEN:
5445 if (!cpu_rq(cpu)->migration_thread)
5447 /* Unbind it from offline cpu so it can run. Fall thru. */
5448 kthread_bind(cpu_rq(cpu)->migration_thread,
5449 any_online_cpu(cpu_online_map));
5450 kthread_stop(cpu_rq(cpu)->migration_thread);
5451 cpu_rq(cpu)->migration_thread = NULL;
5455 case CPU_DEAD_FROZEN:
5456 migrate_live_tasks(cpu);
5458 kthread_stop(rq->migration_thread);
5459 rq->migration_thread = NULL;
5460 /* Idle task back to normal (off runqueue, low prio) */
5461 rq = task_rq_lock(rq->idle, &flags);
5462 deactivate_task(rq->idle, rq);
5463 rq->idle->static_prio = MAX_PRIO;
5464 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5465 migrate_dead_tasks(cpu);
5466 task_rq_unlock(rq, &flags);
5467 migrate_nr_uninterruptible(rq);
5468 BUG_ON(rq->nr_running != 0);
5470 /* No need to migrate the tasks: it was best-effort if
5471 * they didn't take sched_hotcpu_mutex. Just wake up
5472 * the requestors. */
5473 spin_lock_irq(&rq->lock);
5474 while (!list_empty(&rq->migration_queue)) {
5475 struct migration_req *req;
5477 req = list_entry(rq->migration_queue.next,
5478 struct migration_req, list);
5479 list_del_init(&req->list);
5480 complete(&req->done);
5482 spin_unlock_irq(&rq->lock);
5485 case CPU_LOCK_RELEASE:
5486 mutex_unlock(&sched_hotcpu_mutex);
5492 /* Register at highest priority so that task migration (migrate_all_tasks)
5493 * happens before everything else.
5495 static struct notifier_block __cpuinitdata migration_notifier = {
5496 .notifier_call = migration_call,
5500 int __init migration_init(void)
5502 void *cpu = (void *)(long)smp_processor_id();
5505 /* Start one for the boot CPU: */
5506 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5507 BUG_ON(err == NOTIFY_BAD);
5508 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5509 register_cpu_notifier(&migration_notifier);
5517 /* Number of possible processor ids */
5518 int nr_cpu_ids __read_mostly = NR_CPUS;
5519 EXPORT_SYMBOL(nr_cpu_ids);
5521 #undef SCHED_DOMAIN_DEBUG
5522 #ifdef SCHED_DOMAIN_DEBUG
5523 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5528 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5532 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5537 struct sched_group *group = sd->groups;
5538 cpumask_t groupmask;
5540 cpumask_scnprintf(str, NR_CPUS, sd->span);
5541 cpus_clear(groupmask);
5544 for (i = 0; i < level + 1; i++)
5546 printk("domain %d: ", level);
5548 if (!(sd->flags & SD_LOAD_BALANCE)) {
5549 printk("does not load-balance\n");
5551 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5556 printk("span %s\n", str);
5558 if (!cpu_isset(cpu, sd->span))
5559 printk(KERN_ERR "ERROR: domain->span does not contain "
5561 if (!cpu_isset(cpu, group->cpumask))
5562 printk(KERN_ERR "ERROR: domain->groups does not contain"
5566 for (i = 0; i < level + 2; i++)
5572 printk(KERN_ERR "ERROR: group is NULL\n");
5576 if (!group->__cpu_power) {
5578 printk(KERN_ERR "ERROR: domain->cpu_power not "
5582 if (!cpus_weight(group->cpumask)) {
5584 printk(KERN_ERR "ERROR: empty group\n");
5587 if (cpus_intersects(groupmask, group->cpumask)) {
5589 printk(KERN_ERR "ERROR: repeated CPUs\n");
5592 cpus_or(groupmask, groupmask, group->cpumask);
5594 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5597 group = group->next;
5598 } while (group != sd->groups);
5601 if (!cpus_equal(sd->span, groupmask))
5602 printk(KERN_ERR "ERROR: groups don't span "
5610 if (!cpus_subset(groupmask, sd->span))
5611 printk(KERN_ERR "ERROR: parent span is not a superset "
5612 "of domain->span\n");
5617 # define sched_domain_debug(sd, cpu) do { } while (0)
5620 static int sd_degenerate(struct sched_domain *sd)
5622 if (cpus_weight(sd->span) == 1)
5625 /* Following flags need at least 2 groups */
5626 if (sd->flags & (SD_LOAD_BALANCE |
5627 SD_BALANCE_NEWIDLE |
5631 SD_SHARE_PKG_RESOURCES)) {
5632 if (sd->groups != sd->groups->next)
5636 /* Following flags don't use groups */
5637 if (sd->flags & (SD_WAKE_IDLE |
5646 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5648 unsigned long cflags = sd->flags, pflags = parent->flags;
5650 if (sd_degenerate(parent))
5653 if (!cpus_equal(sd->span, parent->span))
5656 /* Does parent contain flags not in child? */
5657 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5658 if (cflags & SD_WAKE_AFFINE)
5659 pflags &= ~SD_WAKE_BALANCE;
5660 /* Flags needing groups don't count if only 1 group in parent */
5661 if (parent->groups == parent->groups->next) {
5662 pflags &= ~(SD_LOAD_BALANCE |
5663 SD_BALANCE_NEWIDLE |
5667 SD_SHARE_PKG_RESOURCES);
5669 if (~cflags & pflags)
5676 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5677 * hold the hotplug lock.
5679 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5681 struct rq *rq = cpu_rq(cpu);
5682 struct sched_domain *tmp;
5684 /* Remove the sched domains which do not contribute to scheduling. */
5685 for (tmp = sd; tmp; tmp = tmp->parent) {
5686 struct sched_domain *parent = tmp->parent;
5689 if (sd_parent_degenerate(tmp, parent)) {
5690 tmp->parent = parent->parent;
5692 parent->parent->child = tmp;
5696 if (sd && sd_degenerate(sd)) {
5702 sched_domain_debug(sd, cpu);
5704 rcu_assign_pointer(rq->sd, sd);
5707 /* cpus with isolated domains */
5708 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5710 /* Setup the mask of cpus configured for isolated domains */
5711 static int __init isolated_cpu_setup(char *str)
5713 int ints[NR_CPUS], i;
5715 str = get_options(str, ARRAY_SIZE(ints), ints);
5716 cpus_clear(cpu_isolated_map);
5717 for (i = 1; i <= ints[0]; i++)
5718 if (ints[i] < NR_CPUS)
5719 cpu_set(ints[i], cpu_isolated_map);
5723 __setup ("isolcpus=", isolated_cpu_setup);
5726 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5727 * to a function which identifies what group(along with sched group) a CPU
5728 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5729 * (due to the fact that we keep track of groups covered with a cpumask_t).
5731 * init_sched_build_groups will build a circular linked list of the groups
5732 * covered by the given span, and will set each group's ->cpumask correctly,
5733 * and ->cpu_power to 0.
5736 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5737 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5738 struct sched_group **sg))
5740 struct sched_group *first = NULL, *last = NULL;
5741 cpumask_t covered = CPU_MASK_NONE;
5744 for_each_cpu_mask(i, span) {
5745 struct sched_group *sg;
5746 int group = group_fn(i, cpu_map, &sg);
5749 if (cpu_isset(i, covered))
5752 sg->cpumask = CPU_MASK_NONE;
5753 sg->__cpu_power = 0;
5755 for_each_cpu_mask(j, span) {
5756 if (group_fn(j, cpu_map, NULL) != group)
5759 cpu_set(j, covered);
5760 cpu_set(j, sg->cpumask);
5771 #define SD_NODES_PER_DOMAIN 16
5774 * Self-tuning task migration cost measurement between source and target CPUs.
5776 * This is done by measuring the cost of manipulating buffers of varying
5777 * sizes. For a given buffer-size here are the steps that are taken:
5779 * 1) the source CPU reads+dirties a shared buffer
5780 * 2) the target CPU reads+dirties the same shared buffer
5782 * We measure how long they take, in the following 4 scenarios:
5784 * - source: CPU1, target: CPU2 | cost1
5785 * - source: CPU2, target: CPU1 | cost2
5786 * - source: CPU1, target: CPU1 | cost3
5787 * - source: CPU2, target: CPU2 | cost4
5789 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5790 * the cost of migration.
5792 * We then start off from a small buffer-size and iterate up to larger
5793 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5794 * doing a maximum search for the cost. (The maximum cost for a migration
5795 * normally occurs when the working set size is around the effective cache
5798 #define SEARCH_SCOPE 2
5799 #define MIN_CACHE_SIZE (64*1024U)
5800 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5801 #define ITERATIONS 1
5802 #define SIZE_THRESH 130
5803 #define COST_THRESH 130
5806 * The migration cost is a function of 'domain distance'. Domain
5807 * distance is the number of steps a CPU has to iterate down its
5808 * domain tree to share a domain with the other CPU. The farther
5809 * two CPUs are from each other, the larger the distance gets.
5811 * Note that we use the distance only to cache measurement results,
5812 * the distance value is not used numerically otherwise. When two
5813 * CPUs have the same distance it is assumed that the migration
5814 * cost is the same. (this is a simplification but quite practical)
5816 #define MAX_DOMAIN_DISTANCE 32
5818 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5819 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5821 * Architectures may override the migration cost and thus avoid
5822 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5823 * virtualized hardware:
5825 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5826 CONFIG_DEFAULT_MIGRATION_COST
5833 * Allow override of migration cost - in units of microseconds.
5834 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5835 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5837 static int __init migration_cost_setup(char *str)
5839 int ints[MAX_DOMAIN_DISTANCE+1], i;
5841 str = get_options(str, ARRAY_SIZE(ints), ints);
5843 printk("#ints: %d\n", ints[0]);
5844 for (i = 1; i <= ints[0]; i++) {
5845 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5846 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5851 __setup ("migration_cost=", migration_cost_setup);
5854 * Global multiplier (divisor) for migration-cutoff values,
5855 * in percentiles. E.g. use a value of 150 to get 1.5 times
5856 * longer cache-hot cutoff times.
5858 * (We scale it from 100 to 128 to long long handling easier.)
5861 #define MIGRATION_FACTOR_SCALE 128
5863 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5865 static int __init setup_migration_factor(char *str)
5867 get_option(&str, &migration_factor);
5868 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5872 __setup("migration_factor=", setup_migration_factor);
5875 * Estimated distance of two CPUs, measured via the number of domains
5876 * we have to pass for the two CPUs to be in the same span:
5878 static unsigned long domain_distance(int cpu1, int cpu2)
5880 unsigned long distance = 0;
5881 struct sched_domain *sd;
5883 for_each_domain(cpu1, sd) {
5884 WARN_ON(!cpu_isset(cpu1, sd->span));
5885 if (cpu_isset(cpu2, sd->span))
5889 if (distance >= MAX_DOMAIN_DISTANCE) {
5891 distance = MAX_DOMAIN_DISTANCE-1;
5897 static unsigned int migration_debug;
5899 static int __init setup_migration_debug(char *str)
5901 get_option(&str, &migration_debug);
5905 __setup("migration_debug=", setup_migration_debug);
5908 * Maximum cache-size that the scheduler should try to measure.
5909 * Architectures with larger caches should tune this up during
5910 * bootup. Gets used in the domain-setup code (i.e. during SMP
5913 unsigned int max_cache_size;
5915 static int __init setup_max_cache_size(char *str)
5917 get_option(&str, &max_cache_size);
5921 __setup("max_cache_size=", setup_max_cache_size);
5924 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5925 * is the operation that is timed, so we try to generate unpredictable
5926 * cachemisses that still end up filling the L2 cache:
5928 static void touch_cache(void *__cache, unsigned long __size)
5930 unsigned long size = __size / sizeof(long);
5931 unsigned long chunk1 = size / 3;
5932 unsigned long chunk2 = 2 * size / 3;
5933 unsigned long *cache = __cache;
5936 for (i = 0; i < size/6; i += 8) {
5939 case 1: cache[size-1-i]++;
5940 case 2: cache[chunk1-i]++;
5941 case 3: cache[chunk1+i]++;
5942 case 4: cache[chunk2-i]++;
5943 case 5: cache[chunk2+i]++;
5949 * Measure the cache-cost of one task migration. Returns in units of nsec.
5951 static unsigned long long
5952 measure_one(void *cache, unsigned long size, int source, int target)
5954 cpumask_t mask, saved_mask;
5955 unsigned long long t0, t1, t2, t3, cost;
5957 saved_mask = current->cpus_allowed;
5960 * Flush source caches to RAM and invalidate them:
5965 * Migrate to the source CPU:
5967 mask = cpumask_of_cpu(source);
5968 set_cpus_allowed(current, mask);
5969 WARN_ON(smp_processor_id() != source);
5972 * Dirty the working set:
5975 touch_cache(cache, size);
5979 * Migrate to the target CPU, dirty the L2 cache and access
5980 * the shared buffer. (which represents the working set
5981 * of a migrated task.)
5983 mask = cpumask_of_cpu(target);
5984 set_cpus_allowed(current, mask);
5985 WARN_ON(smp_processor_id() != target);
5988 touch_cache(cache, size);
5991 cost = t1-t0 + t3-t2;
5993 if (migration_debug >= 2)
5994 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5995 source, target, t1-t0, t1-t0, t3-t2, cost);
5997 * Flush target caches to RAM and invalidate them:
6001 set_cpus_allowed(current, saved_mask);
6007 * Measure a series of task migrations and return the average
6008 * result. Since this code runs early during bootup the system
6009 * is 'undisturbed' and the average latency makes sense.
6011 * The algorithm in essence auto-detects the relevant cache-size,
6012 * so it will properly detect different cachesizes for different
6013 * cache-hierarchies, depending on how the CPUs are connected.
6015 * Architectures can prime the upper limit of the search range via
6016 * max_cache_size, otherwise the search range defaults to 20MB...64K.
6018 static unsigned long long
6019 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
6021 unsigned long long cost1, cost2;
6025 * Measure the migration cost of 'size' bytes, over an
6026 * average of 10 runs:
6028 * (We perturb the cache size by a small (0..4k)
6029 * value to compensate size/alignment related artifacts.
6030 * We also subtract the cost of the operation done on
6036 * dry run, to make sure we start off cache-cold on cpu1,
6037 * and to get any vmalloc pagefaults in advance:
6039 measure_one(cache, size, cpu1, cpu2);
6040 for (i = 0; i < ITERATIONS; i++)
6041 cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2);
6043 measure_one(cache, size, cpu2, cpu1);
6044 for (i = 0; i < ITERATIONS; i++)
6045 cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1);
6048 * (We measure the non-migrating [cached] cost on both
6049 * cpu1 and cpu2, to handle CPUs with different speeds)
6053 measure_one(cache, size, cpu1, cpu1);
6054 for (i = 0; i < ITERATIONS; i++)
6055 cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1);
6057 measure_one(cache, size, cpu2, cpu2);
6058 for (i = 0; i < ITERATIONS; i++)
6059 cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2);
6062 * Get the per-iteration migration cost:
6064 do_div(cost1, 2 * ITERATIONS);
6065 do_div(cost2, 2 * ITERATIONS);
6067 return cost1 - cost2;
6070 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
6072 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
6073 unsigned int max_size, size, size_found = 0;
6074 long long cost = 0, prev_cost;
6078 * Search from max_cache_size*5 down to 64K - the real relevant
6079 * cachesize has to lie somewhere inbetween.
6081 if (max_cache_size) {
6082 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
6083 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
6086 * Since we have no estimation about the relevant
6089 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
6090 size = MIN_CACHE_SIZE;
6093 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
6094 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
6099 * Allocate the working set:
6101 cache = vmalloc(max_size);
6103 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size);
6104 return 1000000; /* return 1 msec on very small boxen */
6107 while (size <= max_size) {
6109 cost = measure_cost(cpu1, cpu2, cache, size);
6115 if (max_cost < cost) {
6121 * Calculate average fluctuation, we use this to prevent
6122 * noise from triggering an early break out of the loop:
6124 fluct = abs(cost - prev_cost);
6125 avg_fluct = (avg_fluct + fluct)/2;
6127 if (migration_debug)
6128 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6131 (long)cost / 1000000,
6132 ((long)cost / 100000) % 10,
6133 (long)max_cost / 1000000,
6134 ((long)max_cost / 100000) % 10,
6135 domain_distance(cpu1, cpu2),
6139 * If we iterated at least 20% past the previous maximum,
6140 * and the cost has dropped by more than 20% already,
6141 * (taking fluctuations into account) then we assume to
6142 * have found the maximum and break out of the loop early:
6144 if (size_found && (size*100 > size_found*SIZE_THRESH))
6145 if (cost+avg_fluct <= 0 ||
6146 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6148 if (migration_debug)
6149 printk("-> found max.\n");
6153 * Increase the cachesize in 10% steps:
6155 size = size * 10 / 9;
6158 if (migration_debug)
6159 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6160 cpu1, cpu2, size_found, max_cost);
6165 * A task is considered 'cache cold' if at least 2 times
6166 * the worst-case cost of migration has passed.
6168 * (this limit is only listened to if the load-balancing
6169 * situation is 'nice' - if there is a large imbalance we
6170 * ignore it for the sake of CPU utilization and
6171 * processing fairness.)
6173 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6176 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6178 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6179 unsigned long j0, j1, distance, max_distance = 0;
6180 struct sched_domain *sd;
6185 * First pass - calculate the cacheflush times:
6187 for_each_cpu_mask(cpu1, *cpu_map) {
6188 for_each_cpu_mask(cpu2, *cpu_map) {
6191 distance = domain_distance(cpu1, cpu2);
6192 max_distance = max(max_distance, distance);
6194 * No result cached yet?
6196 if (migration_cost[distance] == -1LL)
6197 migration_cost[distance] =
6198 measure_migration_cost(cpu1, cpu2);
6202 * Second pass - update the sched domain hierarchy with
6203 * the new cache-hot-time estimations:
6205 for_each_cpu_mask(cpu, *cpu_map) {
6207 for_each_domain(cpu, sd) {
6208 sd->cache_hot_time = migration_cost[distance];
6215 if (migration_debug)
6216 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6224 if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) {
6225 printk("migration_cost=");
6226 for (distance = 0; distance <= max_distance; distance++) {
6229 printk("%ld", (long)migration_cost[distance] / 1000);
6234 if (migration_debug)
6235 printk("migration: %ld seconds\n", (j1-j0) / HZ);
6238 * Move back to the original CPU. NUMA-Q gets confused
6239 * if we migrate to another quad during bootup.
6241 if (raw_smp_processor_id() != orig_cpu) {
6242 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6243 saved_mask = current->cpus_allowed;
6245 set_cpus_allowed(current, mask);
6246 set_cpus_allowed(current, saved_mask);
6253 * find_next_best_node - find the next node to include in a sched_domain
6254 * @node: node whose sched_domain we're building
6255 * @used_nodes: nodes already in the sched_domain
6257 * Find the next node to include in a given scheduling domain. Simply
6258 * finds the closest node not already in the @used_nodes map.
6260 * Should use nodemask_t.
6262 static int find_next_best_node(int node, unsigned long *used_nodes)
6264 int i, n, val, min_val, best_node = 0;
6268 for (i = 0; i < MAX_NUMNODES; i++) {
6269 /* Start at @node */
6270 n = (node + i) % MAX_NUMNODES;
6272 if (!nr_cpus_node(n))
6275 /* Skip already used nodes */
6276 if (test_bit(n, used_nodes))
6279 /* Simple min distance search */
6280 val = node_distance(node, n);
6282 if (val < min_val) {
6288 set_bit(best_node, used_nodes);
6293 * sched_domain_node_span - get a cpumask for a node's sched_domain
6294 * @node: node whose cpumask we're constructing
6295 * @size: number of nodes to include in this span
6297 * Given a node, construct a good cpumask for its sched_domain to span. It
6298 * should be one that prevents unnecessary balancing, but also spreads tasks
6301 static cpumask_t sched_domain_node_span(int node)
6303 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6304 cpumask_t span, nodemask;
6308 bitmap_zero(used_nodes, MAX_NUMNODES);
6310 nodemask = node_to_cpumask(node);
6311 cpus_or(span, span, nodemask);
6312 set_bit(node, used_nodes);
6314 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6315 int next_node = find_next_best_node(node, used_nodes);
6317 nodemask = node_to_cpumask(next_node);
6318 cpus_or(span, span, nodemask);
6325 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6328 * SMT sched-domains:
6330 #ifdef CONFIG_SCHED_SMT
6331 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6332 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6334 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6335 struct sched_group **sg)
6338 *sg = &per_cpu(sched_group_cpus, cpu);
6344 * multi-core sched-domains:
6346 #ifdef CONFIG_SCHED_MC
6347 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6348 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6351 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6352 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6353 struct sched_group **sg)
6356 cpumask_t mask = cpu_sibling_map[cpu];
6357 cpus_and(mask, mask, *cpu_map);
6358 group = first_cpu(mask);
6360 *sg = &per_cpu(sched_group_core, group);
6363 #elif defined(CONFIG_SCHED_MC)
6364 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6365 struct sched_group **sg)
6368 *sg = &per_cpu(sched_group_core, cpu);
6373 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6374 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6376 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6377 struct sched_group **sg)
6380 #ifdef CONFIG_SCHED_MC
6381 cpumask_t mask = cpu_coregroup_map(cpu);
6382 cpus_and(mask, mask, *cpu_map);
6383 group = first_cpu(mask);
6384 #elif defined(CONFIG_SCHED_SMT)
6385 cpumask_t mask = cpu_sibling_map[cpu];
6386 cpus_and(mask, mask, *cpu_map);
6387 group = first_cpu(mask);
6392 *sg = &per_cpu(sched_group_phys, group);
6398 * The init_sched_build_groups can't handle what we want to do with node
6399 * groups, so roll our own. Now each node has its own list of groups which
6400 * gets dynamically allocated.
6402 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6403 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6405 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6406 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6408 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6409 struct sched_group **sg)
6411 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6414 cpus_and(nodemask, nodemask, *cpu_map);
6415 group = first_cpu(nodemask);
6418 *sg = &per_cpu(sched_group_allnodes, group);
6422 static void init_numa_sched_groups_power(struct sched_group *group_head)
6424 struct sched_group *sg = group_head;
6430 for_each_cpu_mask(j, sg->cpumask) {
6431 struct sched_domain *sd;
6433 sd = &per_cpu(phys_domains, j);
6434 if (j != first_cpu(sd->groups->cpumask)) {
6436 * Only add "power" once for each
6442 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6445 if (sg != group_head)
6451 /* Free memory allocated for various sched_group structures */
6452 static void free_sched_groups(const cpumask_t *cpu_map)
6456 for_each_cpu_mask(cpu, *cpu_map) {
6457 struct sched_group **sched_group_nodes
6458 = sched_group_nodes_bycpu[cpu];
6460 if (!sched_group_nodes)
6463 for (i = 0; i < MAX_NUMNODES; i++) {
6464 cpumask_t nodemask = node_to_cpumask(i);
6465 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6467 cpus_and(nodemask, nodemask, *cpu_map);
6468 if (cpus_empty(nodemask))
6478 if (oldsg != sched_group_nodes[i])
6481 kfree(sched_group_nodes);
6482 sched_group_nodes_bycpu[cpu] = NULL;
6486 static void free_sched_groups(const cpumask_t *cpu_map)
6492 * Initialize sched groups cpu_power.
6494 * cpu_power indicates the capacity of sched group, which is used while
6495 * distributing the load between different sched groups in a sched domain.
6496 * Typically cpu_power for all the groups in a sched domain will be same unless
6497 * there are asymmetries in the topology. If there are asymmetries, group
6498 * having more cpu_power will pickup more load compared to the group having
6501 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6502 * the maximum number of tasks a group can handle in the presence of other idle
6503 * or lightly loaded groups in the same sched domain.
6505 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6507 struct sched_domain *child;
6508 struct sched_group *group;
6510 WARN_ON(!sd || !sd->groups);
6512 if (cpu != first_cpu(sd->groups->cpumask))
6517 sd->groups->__cpu_power = 0;
6520 * For perf policy, if the groups in child domain share resources
6521 * (for example cores sharing some portions of the cache hierarchy
6522 * or SMT), then set this domain groups cpu_power such that each group
6523 * can handle only one task, when there are other idle groups in the
6524 * same sched domain.
6526 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6528 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6529 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6534 * add cpu_power of each child group to this groups cpu_power
6536 group = child->groups;
6538 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6539 group = group->next;
6540 } while (group != child->groups);
6544 * Build sched domains for a given set of cpus and attach the sched domains
6545 * to the individual cpus
6547 static int build_sched_domains(const cpumask_t *cpu_map)
6550 struct sched_domain *sd;
6552 struct sched_group **sched_group_nodes = NULL;
6553 int sd_allnodes = 0;
6556 * Allocate the per-node list of sched groups
6558 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6560 if (!sched_group_nodes) {
6561 printk(KERN_WARNING "Can not alloc sched group node list\n");
6564 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6568 * Set up domains for cpus specified by the cpu_map.
6570 for_each_cpu_mask(i, *cpu_map) {
6571 struct sched_domain *sd = NULL, *p;
6572 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6574 cpus_and(nodemask, nodemask, *cpu_map);
6577 if (cpus_weight(*cpu_map)
6578 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6579 sd = &per_cpu(allnodes_domains, i);
6580 *sd = SD_ALLNODES_INIT;
6581 sd->span = *cpu_map;
6582 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6588 sd = &per_cpu(node_domains, i);
6590 sd->span = sched_domain_node_span(cpu_to_node(i));
6594 cpus_and(sd->span, sd->span, *cpu_map);
6598 sd = &per_cpu(phys_domains, i);
6600 sd->span = nodemask;
6604 cpu_to_phys_group(i, cpu_map, &sd->groups);
6606 #ifdef CONFIG_SCHED_MC
6608 sd = &per_cpu(core_domains, i);
6610 sd->span = cpu_coregroup_map(i);
6611 cpus_and(sd->span, sd->span, *cpu_map);
6614 cpu_to_core_group(i, cpu_map, &sd->groups);
6617 #ifdef CONFIG_SCHED_SMT
6619 sd = &per_cpu(cpu_domains, i);
6620 *sd = SD_SIBLING_INIT;
6621 sd->span = cpu_sibling_map[i];
6622 cpus_and(sd->span, sd->span, *cpu_map);
6625 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6629 #ifdef CONFIG_SCHED_SMT
6630 /* Set up CPU (sibling) groups */
6631 for_each_cpu_mask(i, *cpu_map) {
6632 cpumask_t this_sibling_map = cpu_sibling_map[i];
6633 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6634 if (i != first_cpu(this_sibling_map))
6637 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6641 #ifdef CONFIG_SCHED_MC
6642 /* Set up multi-core groups */
6643 for_each_cpu_mask(i, *cpu_map) {
6644 cpumask_t this_core_map = cpu_coregroup_map(i);
6645 cpus_and(this_core_map, this_core_map, *cpu_map);
6646 if (i != first_cpu(this_core_map))
6648 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6653 /* Set up physical groups */
6654 for (i = 0; i < MAX_NUMNODES; i++) {
6655 cpumask_t nodemask = node_to_cpumask(i);
6657 cpus_and(nodemask, nodemask, *cpu_map);
6658 if (cpus_empty(nodemask))
6661 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6665 /* Set up node groups */
6667 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6669 for (i = 0; i < MAX_NUMNODES; i++) {
6670 /* Set up node groups */
6671 struct sched_group *sg, *prev;
6672 cpumask_t nodemask = node_to_cpumask(i);
6673 cpumask_t domainspan;
6674 cpumask_t covered = CPU_MASK_NONE;
6677 cpus_and(nodemask, nodemask, *cpu_map);
6678 if (cpus_empty(nodemask)) {
6679 sched_group_nodes[i] = NULL;
6683 domainspan = sched_domain_node_span(i);
6684 cpus_and(domainspan, domainspan, *cpu_map);
6686 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6688 printk(KERN_WARNING "Can not alloc domain group for "
6692 sched_group_nodes[i] = sg;
6693 for_each_cpu_mask(j, nodemask) {
6694 struct sched_domain *sd;
6695 sd = &per_cpu(node_domains, j);
6698 sg->__cpu_power = 0;
6699 sg->cpumask = nodemask;
6701 cpus_or(covered, covered, nodemask);
6704 for (j = 0; j < MAX_NUMNODES; j++) {
6705 cpumask_t tmp, notcovered;
6706 int n = (i + j) % MAX_NUMNODES;
6708 cpus_complement(notcovered, covered);
6709 cpus_and(tmp, notcovered, *cpu_map);
6710 cpus_and(tmp, tmp, domainspan);
6711 if (cpus_empty(tmp))
6714 nodemask = node_to_cpumask(n);
6715 cpus_and(tmp, tmp, nodemask);
6716 if (cpus_empty(tmp))
6719 sg = kmalloc_node(sizeof(struct sched_group),
6723 "Can not alloc domain group for node %d\n", j);
6726 sg->__cpu_power = 0;
6728 sg->next = prev->next;
6729 cpus_or(covered, covered, tmp);
6736 /* Calculate CPU power for physical packages and nodes */
6737 #ifdef CONFIG_SCHED_SMT
6738 for_each_cpu_mask(i, *cpu_map) {
6739 sd = &per_cpu(cpu_domains, i);
6740 init_sched_groups_power(i, sd);
6743 #ifdef CONFIG_SCHED_MC
6744 for_each_cpu_mask(i, *cpu_map) {
6745 sd = &per_cpu(core_domains, i);
6746 init_sched_groups_power(i, sd);
6750 for_each_cpu_mask(i, *cpu_map) {
6751 sd = &per_cpu(phys_domains, i);
6752 init_sched_groups_power(i, sd);
6756 for (i = 0; i < MAX_NUMNODES; i++)
6757 init_numa_sched_groups_power(sched_group_nodes[i]);
6760 struct sched_group *sg;
6762 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6763 init_numa_sched_groups_power(sg);
6767 /* Attach the domains */
6768 for_each_cpu_mask(i, *cpu_map) {
6769 struct sched_domain *sd;
6770 #ifdef CONFIG_SCHED_SMT
6771 sd = &per_cpu(cpu_domains, i);
6772 #elif defined(CONFIG_SCHED_MC)
6773 sd = &per_cpu(core_domains, i);
6775 sd = &per_cpu(phys_domains, i);
6777 cpu_attach_domain(sd, i);
6780 * Tune cache-hot values:
6782 calibrate_migration_costs(cpu_map);
6788 free_sched_groups(cpu_map);
6793 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6795 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6797 cpumask_t cpu_default_map;
6801 * Setup mask for cpus without special case scheduling requirements.
6802 * For now this just excludes isolated cpus, but could be used to
6803 * exclude other special cases in the future.
6805 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6807 err = build_sched_domains(&cpu_default_map);
6812 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6814 free_sched_groups(cpu_map);
6818 * Detach sched domains from a group of cpus specified in cpu_map
6819 * These cpus will now be attached to the NULL domain
6821 static void detach_destroy_domains(const cpumask_t *cpu_map)
6825 for_each_cpu_mask(i, *cpu_map)
6826 cpu_attach_domain(NULL, i);
6827 synchronize_sched();
6828 arch_destroy_sched_domains(cpu_map);
6832 * Partition sched domains as specified by the cpumasks below.
6833 * This attaches all cpus from the cpumasks to the NULL domain,
6834 * waits for a RCU quiescent period, recalculates sched
6835 * domain information and then attaches them back to the
6836 * correct sched domains
6837 * Call with hotplug lock held
6839 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6841 cpumask_t change_map;
6844 cpus_and(*partition1, *partition1, cpu_online_map);
6845 cpus_and(*partition2, *partition2, cpu_online_map);
6846 cpus_or(change_map, *partition1, *partition2);
6848 /* Detach sched domains from all of the affected cpus */
6849 detach_destroy_domains(&change_map);
6850 if (!cpus_empty(*partition1))
6851 err = build_sched_domains(partition1);
6852 if (!err && !cpus_empty(*partition2))
6853 err = build_sched_domains(partition2);
6858 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6859 int arch_reinit_sched_domains(void)
6863 mutex_lock(&sched_hotcpu_mutex);
6864 detach_destroy_domains(&cpu_online_map);
6865 err = arch_init_sched_domains(&cpu_online_map);
6866 mutex_unlock(&sched_hotcpu_mutex);
6871 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6875 if (buf[0] != '0' && buf[0] != '1')
6879 sched_smt_power_savings = (buf[0] == '1');
6881 sched_mc_power_savings = (buf[0] == '1');
6883 ret = arch_reinit_sched_domains();
6885 return ret ? ret : count;
6888 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6892 #ifdef CONFIG_SCHED_SMT
6894 err = sysfs_create_file(&cls->kset.kobj,
6895 &attr_sched_smt_power_savings.attr);
6897 #ifdef CONFIG_SCHED_MC
6898 if (!err && mc_capable())
6899 err = sysfs_create_file(&cls->kset.kobj,
6900 &attr_sched_mc_power_savings.attr);
6906 #ifdef CONFIG_SCHED_MC
6907 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6909 return sprintf(page, "%u\n", sched_mc_power_savings);
6911 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6912 const char *buf, size_t count)
6914 return sched_power_savings_store(buf, count, 0);
6916 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6917 sched_mc_power_savings_store);
6920 #ifdef CONFIG_SCHED_SMT
6921 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6923 return sprintf(page, "%u\n", sched_smt_power_savings);
6925 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6926 const char *buf, size_t count)
6928 return sched_power_savings_store(buf, count, 1);
6930 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6931 sched_smt_power_savings_store);
6935 * Force a reinitialization of the sched domains hierarchy. The domains
6936 * and groups cannot be updated in place without racing with the balancing
6937 * code, so we temporarily attach all running cpus to the NULL domain
6938 * which will prevent rebalancing while the sched domains are recalculated.
6940 static int update_sched_domains(struct notifier_block *nfb,
6941 unsigned long action, void *hcpu)
6944 case CPU_UP_PREPARE:
6945 case CPU_UP_PREPARE_FROZEN:
6946 case CPU_DOWN_PREPARE:
6947 case CPU_DOWN_PREPARE_FROZEN:
6948 detach_destroy_domains(&cpu_online_map);
6951 case CPU_UP_CANCELED:
6952 case CPU_UP_CANCELED_FROZEN:
6953 case CPU_DOWN_FAILED:
6954 case CPU_DOWN_FAILED_FROZEN:
6956 case CPU_ONLINE_FROZEN:
6958 case CPU_DEAD_FROZEN:
6960 * Fall through and re-initialise the domains.
6967 /* The hotplug lock is already held by cpu_up/cpu_down */
6968 arch_init_sched_domains(&cpu_online_map);
6973 void __init sched_init_smp(void)
6975 cpumask_t non_isolated_cpus;
6977 mutex_lock(&sched_hotcpu_mutex);
6978 arch_init_sched_domains(&cpu_online_map);
6979 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6980 if (cpus_empty(non_isolated_cpus))
6981 cpu_set(smp_processor_id(), non_isolated_cpus);
6982 mutex_unlock(&sched_hotcpu_mutex);
6983 /* XXX: Theoretical race here - CPU may be hotplugged now */
6984 hotcpu_notifier(update_sched_domains, 0);
6986 /* Move init over to a non-isolated CPU */
6987 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6991 void __init sched_init_smp(void)
6994 #endif /* CONFIG_SMP */
6996 int in_sched_functions(unsigned long addr)
6998 /* Linker adds these: start and end of __sched functions */
6999 extern char __sched_text_start[], __sched_text_end[];
7001 return in_lock_functions(addr) ||
7002 (addr >= (unsigned long)__sched_text_start
7003 && addr < (unsigned long)__sched_text_end);
7006 void __init sched_init(void)
7009 int highest_cpu = 0;
7011 for_each_possible_cpu(i) {
7012 struct prio_array *array;
7016 spin_lock_init(&rq->lock);
7017 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7019 rq->active = rq->arrays;
7020 rq->expired = rq->arrays + 1;
7021 rq->best_expired_prio = MAX_PRIO;
7025 for (j = 1; j < 3; j++)
7026 rq->cpu_load[j] = 0;
7027 rq->active_balance = 0;
7030 rq->migration_thread = NULL;
7031 INIT_LIST_HEAD(&rq->migration_queue);
7033 atomic_set(&rq->nr_iowait, 0);
7035 for (j = 0; j < 2; j++) {
7036 array = rq->arrays + j;
7037 for (k = 0; k < MAX_PRIO; k++) {
7038 INIT_LIST_HEAD(array->queue + k);
7039 __clear_bit(k, array->bitmap);
7041 // delimiter for bitsearch
7042 __set_bit(MAX_PRIO, array->bitmap);
7047 set_load_weight(&init_task);
7050 nr_cpu_ids = highest_cpu + 1;
7051 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7054 #ifdef CONFIG_RT_MUTEXES
7055 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7059 * The boot idle thread does lazy MMU switching as well:
7061 atomic_inc(&init_mm.mm_count);
7062 enter_lazy_tlb(&init_mm, current);
7065 * Make us the idle thread. Technically, schedule() should not be
7066 * called from this thread, however somewhere below it might be,
7067 * but because we are the idle thread, we just pick up running again
7068 * when this runqueue becomes "idle".
7070 init_idle(current, smp_processor_id());
7073 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7074 void __might_sleep(char *file, int line)
7077 static unsigned long prev_jiffy; /* ratelimiting */
7079 if ((in_atomic() || irqs_disabled()) &&
7080 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7081 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7083 prev_jiffy = jiffies;
7084 printk(KERN_ERR "BUG: sleeping function called from invalid"
7085 " context at %s:%d\n", file, line);
7086 printk("in_atomic():%d, irqs_disabled():%d\n",
7087 in_atomic(), irqs_disabled());
7088 debug_show_held_locks(current);
7089 if (irqs_disabled())
7090 print_irqtrace_events(current);
7095 EXPORT_SYMBOL(__might_sleep);
7098 #ifdef CONFIG_MAGIC_SYSRQ
7099 void normalize_rt_tasks(void)
7101 struct prio_array *array;
7102 struct task_struct *p;
7103 unsigned long flags;
7106 read_lock_irq(&tasklist_lock);
7107 for_each_process(p) {
7111 spin_lock_irqsave(&p->pi_lock, flags);
7112 rq = __task_rq_lock(p);
7116 deactivate_task(p, task_rq(p));
7117 __setscheduler(p, SCHED_NORMAL, 0);
7119 __activate_task(p, task_rq(p));
7120 resched_task(rq->curr);
7123 __task_rq_unlock(rq);
7124 spin_unlock_irqrestore(&p->pi_lock, flags);
7126 read_unlock_irq(&tasklist_lock);
7129 #endif /* CONFIG_MAGIC_SYSRQ */
7131 #if defined(CONFIG_IA64) || defined(CONFIG_KDB)
7133 * These functions are only useful for the IA64 MCA handling.
7135 * They can only be called when the whole system has been
7136 * stopped - every CPU needs to be quiescent, and no scheduling
7137 * activity can take place. Using them for anything else would
7138 * be a serious bug, and as a result, they aren't even visible
7139 * under any other configuration.
7143 * curr_task - return the current task for a given cpu.
7144 * @cpu: the processor in question.
7146 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7148 struct task_struct *curr_task(int cpu)
7150 return cpu_curr(cpu);
7154 * set_curr_task - set the current task for a given cpu.
7155 * @cpu: the processor in question.
7156 * @p: the task pointer to set.
7158 * Description: This function must only be used when non-maskable interrupts
7159 * are serviced on a separate stack. It allows the architecture to switch the
7160 * notion of the current task on a cpu in a non-blocking manner. This function
7161 * must be called with all CPU's synchronized, and interrupts disabled, the
7162 * and caller must save the original value of the current task (see
7163 * curr_task() above) and restore that value before reenabling interrupts and
7164 * re-starting the system.
7166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7168 void set_curr_task(int cpu, struct task_struct *p)
7177 #include <linux/kdb.h>
7180 kdb_prio(char *name, struct prio_array *array, kdb_printf_t xxx_printf)
7184 xxx_printf(" %s nr_active:%d bitmap: 0x%lx 0x%lx 0x%lx\n",
7185 name, array->nr_active,
7186 array->bitmap[0], array->bitmap[1], array->bitmap[2]);
7188 pri = sched_find_first_bit(array->bitmap);
7189 if (pri != MAX_PRIO) {
7190 xxx_printf(" bitmap priorities:");
7191 while (pri != MAX_PRIO) {
7192 xxx_printf(" %d", pri);
7194 pri = find_next_bit(array->bitmap, MAX_PRIO, pri);
7199 for (pri = 0; pri < MAX_PRIO; pri++) {
7200 int printed_hdr = 0;
7201 struct list_head *head, *curr;
7203 head = array->queue + pri;
7205 while(curr != head) {
7206 struct task_struct *task;
7208 xxx_printf(" queue at priority=%d\n", pri);
7211 task = list_entry(curr, struct task_struct, run_list);
7212 xxx_printf(" 0x%p %d %s time_slice:%d\n",
7213 task, task->pid, task->comm,
7220 /* This code must be in sched.c because struct rq is only defined in this
7221 * source. To allow most of kdb to be modular, this code cannot call any kdb
7222 * functions directly, any external functions that it needs must be passed in
7227 kdb_runqueue(unsigned long cpu, kdb_printf_t xxx_printf)
7233 xxx_printf("CPU%ld lock:%s curr:0x%p(%d)(%s)",
7234 cpu, (spin_is_locked(&rq->lock))?"LOCKED":"free",
7235 rq->curr, rq->curr->pid, rq->curr->comm);
7236 if (rq->curr == rq->idle)
7237 xxx_printf(" is idle");
7240 xxx_printf(" cpu_load:%lu %lu %lu",
7241 rq->cpu_load[0], rq->cpu_load[1], rq->cpu_load[2]);
7243 xxx_printf(" nr_running:%lu nr_switches:%llu\n",
7244 rq->nr_running, rq->nr_switches);
7245 kdb_prio("active", rq->active, xxx_printf);
7246 kdb_prio("expired", rq->expired, xxx_printf);
7248 EXPORT_SYMBOL(kdb_runqueue);
7250 #endif /* CONFIG_KDB */