4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
57 #include <asm/unistd.h>
60 * Scheduler clock - returns current time in nanosec units.
61 * This is default implementation.
62 * Architectures and sub-architectures can override this.
64 unsigned long long __attribute__((weak)) sched_clock(void)
66 return (unsigned long long)jiffies * (1000000000 / HZ);
70 * Convert user-nice values [ -20 ... 0 ... 19 ]
71 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
74 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
75 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
76 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
79 * 'User priority' is the nice value converted to something we
80 * can work with better when scaling various scheduler parameters,
81 * it's a [ 0 ... 39 ] range.
83 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
84 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
85 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
88 * Some helpers for converting nanosecond timing to jiffy resolution
90 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
91 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 * These are the 'tuning knobs' of the scheduler:
96 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
97 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
98 * Timeslices get refilled after they expire.
100 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
101 #define DEF_TIMESLICE (100 * HZ / 1000)
102 #define ON_RUNQUEUE_WEIGHT 30
103 #define CHILD_PENALTY 95
104 #define PARENT_PENALTY 100
105 #define EXIT_WEIGHT 3
106 #define PRIO_BONUS_RATIO 25
107 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
108 #define INTERACTIVE_DELTA 2
109 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
110 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
111 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
114 * If a task is 'interactive' then we reinsert it in the active
115 * array after it has expired its current timeslice. (it will not
116 * continue to run immediately, it will still roundrobin with
117 * other interactive tasks.)
119 * This part scales the interactivity limit depending on niceness.
121 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
122 * Here are a few examples of different nice levels:
124 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
125 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
126 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
127 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
130 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
131 * priority range a task can explore, a value of '1' means the
132 * task is rated interactive.)
134 * Ie. nice +19 tasks can never get 'interactive' enough to be
135 * reinserted into the active array. And only heavily CPU-hog nice -20
136 * tasks will be expired. Default nice 0 tasks are somewhere between,
137 * it takes some effort for them to get interactive, but it's not
141 #define CURRENT_BONUS(p) \
142 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
145 #define GRANULARITY (10 * HZ / 1000 ? : 1)
148 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
149 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
156 #define SCALE(v1,v1_max,v2_max) \
157 (v1) * (v2_max) / (v1_max)
160 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
163 #define TASK_INTERACTIVE(p) \
164 ((p)->prio <= (p)->static_prio - DELTA(p))
166 #define INTERACTIVE_SLEEP(p) \
167 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
168 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
170 #define TASK_PREEMPTS_CURR(p, rq) \
171 ((p)->prio < (rq)->curr->prio)
173 #define SCALE_PRIO(x, prio) \
174 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
176 static unsigned int static_prio_timeslice(int static_prio)
178 if (static_prio < NICE_TO_PRIO(0))
179 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
181 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
185 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
186 * to time slice values: [800ms ... 100ms ... 5ms]
188 * The higher a thread's priority, the bigger timeslices
189 * it gets during one round of execution. But even the lowest
190 * priority thread gets MIN_TIMESLICE worth of execution time.
193 static inline unsigned int task_timeslice(struct task_struct *p)
195 return static_prio_timeslice(p->static_prio);
199 * These are the runqueue data structures:
203 unsigned int nr_active;
204 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
205 struct list_head queue[MAX_PRIO];
209 * This is the main, per-CPU runqueue data structure.
211 * Locking rule: those places that want to lock multiple runqueues
212 * (such as the load balancing or the thread migration code), lock
213 * acquire operations must be ordered by ascending &runqueue.
219 * nr_running and cpu_load should be in the same cacheline because
220 * remote CPUs use both these fields when doing load calculation.
222 unsigned long nr_running;
223 unsigned long raw_weighted_load;
225 unsigned long cpu_load[3];
226 unsigned char idle_at_tick;
228 unsigned char in_nohz_recently;
231 unsigned long long nr_switches;
234 * This is part of a global counter where only the total sum
235 * over all CPUs matters. A task can increase this counter on
236 * one CPU and if it got migrated afterwards it may decrease
237 * it on another CPU. Always updated under the runqueue lock:
239 unsigned long nr_uninterruptible;
241 unsigned long expired_timestamp;
242 /* Cached timestamp set by update_cpu_clock() */
243 unsigned long long most_recent_timestamp;
244 struct task_struct *curr, *idle;
245 unsigned long next_balance;
246 struct mm_struct *prev_mm;
247 struct prio_array *active, *expired, arrays[2];
248 int best_expired_prio;
252 struct sched_domain *sd;
254 /* For active balancing */
257 int cpu; /* cpu of this runqueue */
259 struct task_struct *migration_thread;
260 struct list_head migration_queue;
263 #ifdef CONFIG_SCHEDSTATS
265 struct sched_info rq_sched_info;
267 /* sys_sched_yield() stats */
268 unsigned long yld_exp_empty;
269 unsigned long yld_act_empty;
270 unsigned long yld_both_empty;
271 unsigned long yld_cnt;
273 /* schedule() stats */
274 unsigned long sched_switch;
275 unsigned long sched_cnt;
276 unsigned long sched_goidle;
278 /* try_to_wake_up() stats */
279 unsigned long ttwu_cnt;
280 unsigned long ttwu_local;
282 struct lock_class_key rq_lock_key;
285 static DEFINE_PER_CPU(struct rq, runqueues);
287 static inline int cpu_of(struct rq *rq)
297 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
298 * See detach_destroy_domains: synchronize_sched for details.
300 * The domain tree of any CPU may only be accessed from within
301 * preempt-disabled sections.
303 #define for_each_domain(cpu, __sd) \
304 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
306 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
307 #define this_rq() (&__get_cpu_var(runqueues))
308 #define task_rq(p) cpu_rq(task_cpu(p))
309 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
311 #ifndef prepare_arch_switch
312 # define prepare_arch_switch(next) do { } while (0)
314 #ifndef finish_arch_switch
315 # define finish_arch_switch(prev) do { } while (0)
318 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
319 static inline int task_running(struct rq *rq, struct task_struct *p)
321 return rq->curr == p;
324 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
328 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
330 #ifdef CONFIG_DEBUG_SPINLOCK
331 /* this is a valid case when another task releases the spinlock */
332 rq->lock.owner = current;
335 * If we are tracking spinlock dependencies then we have to
336 * fix up the runqueue lock - which gets 'carried over' from
339 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
341 spin_unlock_irq(&rq->lock);
344 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
345 static inline int task_running(struct rq *rq, struct task_struct *p)
350 return rq->curr == p;
354 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
358 * We can optimise this out completely for !SMP, because the
359 * SMP rebalancing from interrupt is the only thing that cares
364 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
365 spin_unlock_irq(&rq->lock);
367 spin_unlock(&rq->lock);
371 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
375 * After ->oncpu is cleared, the task can be moved to a different CPU.
376 * We must ensure this doesn't happen until the switch is completely
382 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
386 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
389 * __task_rq_lock - lock the runqueue a given task resides on.
390 * Must be called interrupts disabled.
392 static inline struct rq *__task_rq_lock(struct task_struct *p)
399 spin_lock(&rq->lock);
400 if (unlikely(rq != task_rq(p))) {
401 spin_unlock(&rq->lock);
402 goto repeat_lock_task;
408 * task_rq_lock - lock the runqueue a given task resides on and disable
409 * interrupts. Note the ordering: we can safely lookup the task_rq without
410 * explicitly disabling preemption.
412 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
418 local_irq_save(*flags);
420 spin_lock(&rq->lock);
421 if (unlikely(rq != task_rq(p))) {
422 spin_unlock_irqrestore(&rq->lock, *flags);
423 goto repeat_lock_task;
428 static inline void __task_rq_unlock(struct rq *rq)
431 spin_unlock(&rq->lock);
434 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
437 spin_unlock_irqrestore(&rq->lock, *flags);
440 #ifdef CONFIG_SCHEDSTATS
442 * bump this up when changing the output format or the meaning of an existing
443 * format, so that tools can adapt (or abort)
445 #define SCHEDSTAT_VERSION 14
447 static int show_schedstat(struct seq_file *seq, void *v)
451 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
452 seq_printf(seq, "timestamp %lu\n", jiffies);
453 for_each_online_cpu(cpu) {
454 struct rq *rq = cpu_rq(cpu);
456 struct sched_domain *sd;
460 /* runqueue-specific stats */
462 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
463 cpu, rq->yld_both_empty,
464 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
465 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
466 rq->ttwu_cnt, rq->ttwu_local,
467 rq->rq_sched_info.cpu_time,
468 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
470 seq_printf(seq, "\n");
473 /* domain-specific stats */
475 for_each_domain(cpu, sd) {
476 enum idle_type itype;
477 char mask_str[NR_CPUS];
479 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
480 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
481 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
483 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu "
486 sd->lb_balanced[itype],
487 sd->lb_failed[itype],
488 sd->lb_imbalance[itype],
489 sd->lb_gained[itype],
490 sd->lb_hot_gained[itype],
491 sd->lb_nobusyq[itype],
492 sd->lb_nobusyg[itype]);
494 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
496 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
497 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
498 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
499 sd->ttwu_wake_remote, sd->ttwu_move_affine,
500 sd->ttwu_move_balance);
508 static int schedstat_open(struct inode *inode, struct file *file)
510 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
511 char *buf = kmalloc(size, GFP_KERNEL);
517 res = single_open(file, show_schedstat, NULL);
519 m = file->private_data;
527 const struct file_operations proc_schedstat_operations = {
528 .open = schedstat_open,
531 .release = single_release,
535 * Expects runqueue lock to be held for atomicity of update
538 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
541 rq->rq_sched_info.run_delay += delta_jiffies;
542 rq->rq_sched_info.pcnt++;
547 * Expects runqueue lock to be held for atomicity of update
550 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
553 rq->rq_sched_info.cpu_time += delta_jiffies;
555 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
556 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
557 #else /* !CONFIG_SCHEDSTATS */
559 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
562 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
564 # define schedstat_inc(rq, field) do { } while (0)
565 # define schedstat_add(rq, field, amt) do { } while (0)
569 * this_rq_lock - lock this runqueue and disable interrupts.
571 static inline struct rq *this_rq_lock(void)
578 spin_lock(&rq->lock);
583 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
585 * Called when a process is dequeued from the active array and given
586 * the cpu. We should note that with the exception of interactive
587 * tasks, the expired queue will become the active queue after the active
588 * queue is empty, without explicitly dequeuing and requeuing tasks in the
589 * expired queue. (Interactive tasks may be requeued directly to the
590 * active queue, thus delaying tasks in the expired queue from running;
591 * see scheduler_tick()).
593 * This function is only called from sched_info_arrive(), rather than
594 * dequeue_task(). Even though a task may be queued and dequeued multiple
595 * times as it is shuffled about, we're really interested in knowing how
596 * long it was from the *first* time it was queued to the time that it
599 static inline void sched_info_dequeued(struct task_struct *t)
601 t->sched_info.last_queued = 0;
605 * Called when a task finally hits the cpu. We can now calculate how
606 * long it was waiting to run. We also note when it began so that we
607 * can keep stats on how long its timeslice is.
609 static void sched_info_arrive(struct task_struct *t)
611 unsigned long now = jiffies, delta_jiffies = 0;
613 if (t->sched_info.last_queued)
614 delta_jiffies = now - t->sched_info.last_queued;
615 sched_info_dequeued(t);
616 t->sched_info.run_delay += delta_jiffies;
617 t->sched_info.last_arrival = now;
618 t->sched_info.pcnt++;
620 rq_sched_info_arrive(task_rq(t), delta_jiffies);
624 * Called when a process is queued into either the active or expired
625 * array. The time is noted and later used to determine how long we
626 * had to wait for us to reach the cpu. Since the expired queue will
627 * become the active queue after active queue is empty, without dequeuing
628 * and requeuing any tasks, we are interested in queuing to either. It
629 * is unusual but not impossible for tasks to be dequeued and immediately
630 * requeued in the same or another array: this can happen in sched_yield(),
631 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
634 * This function is only called from enqueue_task(), but also only updates
635 * the timestamp if it is already not set. It's assumed that
636 * sched_info_dequeued() will clear that stamp when appropriate.
638 static inline void sched_info_queued(struct task_struct *t)
640 if (unlikely(sched_info_on()))
641 if (!t->sched_info.last_queued)
642 t->sched_info.last_queued = jiffies;
646 * Called when a process ceases being the active-running process, either
647 * voluntarily or involuntarily. Now we can calculate how long we ran.
649 static inline void sched_info_depart(struct task_struct *t)
651 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
653 t->sched_info.cpu_time += delta_jiffies;
654 rq_sched_info_depart(task_rq(t), delta_jiffies);
658 * Called when tasks are switched involuntarily due, typically, to expiring
659 * their time slice. (This may also be called when switching to or from
660 * the idle task.) We are only called when prev != next.
663 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
665 struct rq *rq = task_rq(prev);
668 * prev now departs the cpu. It's not interesting to record
669 * stats about how efficient we were at scheduling the idle
672 if (prev != rq->idle)
673 sched_info_depart(prev);
675 if (next != rq->idle)
676 sched_info_arrive(next);
679 sched_info_switch(struct task_struct *prev, struct task_struct *next)
681 if (unlikely(sched_info_on()))
682 __sched_info_switch(prev, next);
685 #define sched_info_queued(t) do { } while (0)
686 #define sched_info_switch(t, next) do { } while (0)
687 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
690 * Adding/removing a task to/from a priority array:
692 static void dequeue_task(struct task_struct *p, struct prio_array *array)
695 list_del(&p->run_list);
696 if (list_empty(array->queue + p->prio))
697 __clear_bit(p->prio, array->bitmap);
700 static void enqueue_task(struct task_struct *p, struct prio_array *array)
702 sched_info_queued(p);
703 list_add_tail(&p->run_list, array->queue + p->prio);
704 __set_bit(p->prio, array->bitmap);
710 * Put task to the end of the run list without the overhead of dequeue
711 * followed by enqueue.
713 static void requeue_task(struct task_struct *p, struct prio_array *array)
715 list_move_tail(&p->run_list, array->queue + p->prio);
719 enqueue_task_head(struct task_struct *p, struct prio_array *array)
721 list_add(&p->run_list, array->queue + p->prio);
722 __set_bit(p->prio, array->bitmap);
728 * __normal_prio - return the priority that is based on the static
729 * priority but is modified by bonuses/penalties.
731 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
732 * into the -5 ... 0 ... +5 bonus/penalty range.
734 * We use 25% of the full 0...39 priority range so that:
736 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
737 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
739 * Both properties are important to certain workloads.
742 static inline int __normal_prio(struct task_struct *p)
746 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
748 prio = p->static_prio - bonus;
749 if (prio < MAX_RT_PRIO)
751 if (prio > MAX_PRIO-1)
757 * To aid in avoiding the subversion of "niceness" due to uneven distribution
758 * of tasks with abnormal "nice" values across CPUs the contribution that
759 * each task makes to its run queue's load is weighted according to its
760 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
761 * scaled version of the new time slice allocation that they receive on time
766 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
767 * If static_prio_timeslice() is ever changed to break this assumption then
768 * this code will need modification
770 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
771 #define LOAD_WEIGHT(lp) \
772 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
773 #define PRIO_TO_LOAD_WEIGHT(prio) \
774 LOAD_WEIGHT(static_prio_timeslice(prio))
775 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
776 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
778 static void set_load_weight(struct task_struct *p)
780 if (has_rt_policy(p)) {
782 if (p == task_rq(p)->migration_thread)
784 * The migration thread does the actual balancing.
785 * Giving its load any weight will skew balancing
791 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
793 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
797 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
799 rq->raw_weighted_load += p->load_weight;
803 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
805 rq->raw_weighted_load -= p->load_weight;
808 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
811 inc_raw_weighted_load(rq, p);
814 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
817 dec_raw_weighted_load(rq, p);
821 * Calculate the expected normal priority: i.e. priority
822 * without taking RT-inheritance into account. Might be
823 * boosted by interactivity modifiers. Changes upon fork,
824 * setprio syscalls, and whenever the interactivity
825 * estimator recalculates.
827 static inline int normal_prio(struct task_struct *p)
831 if (has_rt_policy(p))
832 prio = MAX_RT_PRIO-1 - p->rt_priority;
834 prio = __normal_prio(p);
839 * Calculate the current priority, i.e. the priority
840 * taken into account by the scheduler. This value might
841 * be boosted by RT tasks, or might be boosted by
842 * interactivity modifiers. Will be RT if the task got
843 * RT-boosted. If not then it returns p->normal_prio.
845 static int effective_prio(struct task_struct *p)
847 p->normal_prio = normal_prio(p);
849 * If we are RT tasks or we were boosted to RT priority,
850 * keep the priority unchanged. Otherwise, update priority
851 * to the normal priority:
853 if (!rt_prio(p->prio))
854 return p->normal_prio;
859 * __activate_task - move a task to the runqueue.
861 static void __activate_task(struct task_struct *p, struct rq *rq)
863 struct prio_array *target = rq->active;
866 target = rq->expired;
867 enqueue_task(p, target);
868 inc_nr_running(p, rq);
872 * __activate_idle_task - move idle task to the _front_ of runqueue.
874 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
876 enqueue_task_head(p, rq->active);
877 inc_nr_running(p, rq);
881 * Recalculate p->normal_prio and p->prio after having slept,
882 * updating the sleep-average too:
884 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
886 /* Caller must always ensure 'now >= p->timestamp' */
887 unsigned long sleep_time = now - p->timestamp;
892 if (likely(sleep_time > 0)) {
894 * This ceiling is set to the lowest priority that would allow
895 * a task to be reinserted into the active array on timeslice
898 unsigned long ceiling = INTERACTIVE_SLEEP(p);
900 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
902 * Prevents user tasks from achieving best priority
903 * with one single large enough sleep.
905 p->sleep_avg = ceiling;
907 * Using INTERACTIVE_SLEEP() as a ceiling places a
908 * nice(0) task 1ms sleep away from promotion, and
909 * gives it 700ms to round-robin with no chance of
910 * being demoted. This is more than generous, so
911 * mark this sleep as non-interactive to prevent the
912 * on-runqueue bonus logic from intervening should
913 * this task not receive cpu immediately.
915 p->sleep_type = SLEEP_NONINTERACTIVE;
918 * Tasks waking from uninterruptible sleep are
919 * limited in their sleep_avg rise as they
920 * are likely to be waiting on I/O
922 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
923 if (p->sleep_avg >= ceiling)
925 else if (p->sleep_avg + sleep_time >=
927 p->sleep_avg = ceiling;
933 * This code gives a bonus to interactive tasks.
935 * The boost works by updating the 'average sleep time'
936 * value here, based on ->timestamp. The more time a
937 * task spends sleeping, the higher the average gets -
938 * and the higher the priority boost gets as well.
940 p->sleep_avg += sleep_time;
943 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
944 p->sleep_avg = NS_MAX_SLEEP_AVG;
947 return effective_prio(p);
951 * activate_task - move a task to the runqueue and do priority recalculation
953 * Update all the scheduling statistics stuff. (sleep average
954 * calculation, priority modifiers, etc.)
956 static void activate_task(struct task_struct *p, struct rq *rq, int local)
958 unsigned long long now;
966 /* Compensate for drifting sched_clock */
967 struct rq *this_rq = this_rq();
968 now = (now - this_rq->most_recent_timestamp)
969 + rq->most_recent_timestamp;
974 * Sleep time is in units of nanosecs, so shift by 20 to get a
975 * milliseconds-range estimation of the amount of time that the task
978 if (unlikely(prof_on == SLEEP_PROFILING)) {
979 if (p->state == TASK_UNINTERRUPTIBLE)
980 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
981 (now - p->timestamp) >> 20);
984 p->prio = recalc_task_prio(p, now);
987 * This checks to make sure it's not an uninterruptible task
988 * that is now waking up.
990 if (p->sleep_type == SLEEP_NORMAL) {
992 * Tasks which were woken up by interrupts (ie. hw events)
993 * are most likely of interactive nature. So we give them
994 * the credit of extending their sleep time to the period
995 * of time they spend on the runqueue, waiting for execution
996 * on a CPU, first time around:
999 p->sleep_type = SLEEP_INTERRUPTED;
1002 * Normal first-time wakeups get a credit too for
1003 * on-runqueue time, but it will be weighted down:
1005 p->sleep_type = SLEEP_INTERACTIVE;
1010 __activate_task(p, rq);
1014 * deactivate_task - remove a task from the runqueue.
1016 static void deactivate_task(struct task_struct *p, struct rq *rq)
1018 dec_nr_running(p, rq);
1019 dequeue_task(p, p->array);
1024 * resched_task - mark a task 'to be rescheduled now'.
1026 * On UP this means the setting of the need_resched flag, on SMP it
1027 * might also involve a cross-CPU call to trigger the scheduler on
1032 #ifndef tsk_is_polling
1033 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1036 static void resched_task(struct task_struct *p)
1040 assert_spin_locked(&task_rq(p)->lock);
1042 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1045 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1048 if (cpu == smp_processor_id())
1051 /* NEED_RESCHED must be visible before we test polling */
1053 if (!tsk_is_polling(p))
1054 smp_send_reschedule(cpu);
1057 static void resched_cpu(int cpu)
1059 struct rq *rq = cpu_rq(cpu);
1060 unsigned long flags;
1062 if (!spin_trylock_irqsave(&rq->lock, flags))
1064 resched_task(cpu_curr(cpu));
1065 spin_unlock_irqrestore(&rq->lock, flags);
1068 static inline void resched_task(struct task_struct *p)
1070 assert_spin_locked(&task_rq(p)->lock);
1071 set_tsk_need_resched(p);
1076 * task_curr - is this task currently executing on a CPU?
1077 * @p: the task in question.
1079 inline int task_curr(const struct task_struct *p)
1081 return cpu_curr(task_cpu(p)) == p;
1084 /* Used instead of source_load when we know the type == 0 */
1085 unsigned long weighted_cpuload(const int cpu)
1087 return cpu_rq(cpu)->raw_weighted_load;
1091 struct migration_req {
1092 struct list_head list;
1094 struct task_struct *task;
1097 struct completion done;
1101 * The task's runqueue lock must be held.
1102 * Returns true if you have to wait for migration thread.
1105 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1107 struct rq *rq = task_rq(p);
1110 * If the task is not on a runqueue (and not running), then
1111 * it is sufficient to simply update the task's cpu field.
1113 if (!p->array && !task_running(rq, p)) {
1114 set_task_cpu(p, dest_cpu);
1118 init_completion(&req->done);
1120 req->dest_cpu = dest_cpu;
1121 list_add(&req->list, &rq->migration_queue);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * The caller must ensure that the task *will* unschedule sometime soon,
1130 * else this function might spin for a *long* time. This function can't
1131 * be called with interrupts off, or it may introduce deadlock with
1132 * smp_call_function() if an IPI is sent by the same process we are
1133 * waiting to become inactive.
1135 void wait_task_inactive(struct task_struct *p)
1137 unsigned long flags;
1142 rq = task_rq_lock(p, &flags);
1143 /* Must be off runqueue entirely, not preempted. */
1144 if (unlikely(p->array || task_running(rq, p))) {
1145 /* If it's preempted, we yield. It could be a while. */
1146 preempted = !task_running(rq, p);
1147 task_rq_unlock(rq, &flags);
1153 task_rq_unlock(rq, &flags);
1157 * kick_process - kick a running thread to enter/exit the kernel
1158 * @p: the to-be-kicked thread
1160 * Cause a process which is running on another CPU to enter
1161 * kernel-mode, without any delay. (to get signals handled.)
1163 * NOTE: this function doesnt have to take the runqueue lock,
1164 * because all it wants to ensure is that the remote task enters
1165 * the kernel. If the IPI races and the task has been migrated
1166 * to another CPU then no harm is done and the purpose has been
1169 void kick_process(struct task_struct *p)
1175 if ((cpu != smp_processor_id()) && task_curr(p))
1176 smp_send_reschedule(cpu);
1181 * Return a low guess at the load of a migration-source cpu weighted
1182 * according to the scheduling class and "nice" value.
1184 * We want to under-estimate the load of migration sources, to
1185 * balance conservatively.
1187 static inline unsigned long source_load(int cpu, int type)
1189 struct rq *rq = cpu_rq(cpu);
1192 return rq->raw_weighted_load;
1194 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1198 * Return a high guess at the load of a migration-target cpu weighted
1199 * according to the scheduling class and "nice" value.
1201 static inline unsigned long target_load(int cpu, int type)
1203 struct rq *rq = cpu_rq(cpu);
1206 return rq->raw_weighted_load;
1208 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1212 * Return the average load per task on the cpu's run queue
1214 static inline unsigned long cpu_avg_load_per_task(int cpu)
1216 struct rq *rq = cpu_rq(cpu);
1217 unsigned long n = rq->nr_running;
1219 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1223 * find_idlest_group finds and returns the least busy CPU group within the
1226 static struct sched_group *
1227 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1229 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1230 unsigned long min_load = ULONG_MAX, this_load = 0;
1231 int load_idx = sd->forkexec_idx;
1232 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1235 unsigned long load, avg_load;
1239 /* Skip over this group if it has no CPUs allowed */
1240 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1243 local_group = cpu_isset(this_cpu, group->cpumask);
1245 /* Tally up the load of all CPUs in the group */
1248 for_each_cpu_mask(i, group->cpumask) {
1249 /* Bias balancing toward cpus of our domain */
1251 load = source_load(i, load_idx);
1253 load = target_load(i, load_idx);
1258 /* Adjust by relative CPU power of the group */
1259 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1262 this_load = avg_load;
1264 } else if (avg_load < min_load) {
1265 min_load = avg_load;
1269 group = group->next;
1270 } while (group != sd->groups);
1272 if (!idlest || 100*this_load < imbalance*min_load)
1278 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1281 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1284 unsigned long load, min_load = ULONG_MAX;
1288 /* Traverse only the allowed CPUs */
1289 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1291 for_each_cpu_mask(i, tmp) {
1292 load = weighted_cpuload(i);
1294 if (load < min_load || (load == min_load && i == this_cpu)) {
1304 * sched_balance_self: balance the current task (running on cpu) in domains
1305 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1308 * Balance, ie. select the least loaded group.
1310 * Returns the target CPU number, or the same CPU if no balancing is needed.
1312 * preempt must be disabled.
1314 static int sched_balance_self(int cpu, int flag)
1316 struct task_struct *t = current;
1317 struct sched_domain *tmp, *sd = NULL;
1319 for_each_domain(cpu, tmp) {
1321 * If power savings logic is enabled for a domain, stop there.
1323 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1325 if (tmp->flags & flag)
1331 struct sched_group *group;
1332 int new_cpu, weight;
1334 if (!(sd->flags & flag)) {
1340 group = find_idlest_group(sd, t, cpu);
1346 new_cpu = find_idlest_cpu(group, t, cpu);
1347 if (new_cpu == -1 || new_cpu == cpu) {
1348 /* Now try balancing at a lower domain level of cpu */
1353 /* Now try balancing at a lower domain level of new_cpu */
1356 weight = cpus_weight(span);
1357 for_each_domain(cpu, tmp) {
1358 if (weight <= cpus_weight(tmp->span))
1360 if (tmp->flags & flag)
1363 /* while loop will break here if sd == NULL */
1369 #endif /* CONFIG_SMP */
1372 * wake_idle() will wake a task on an idle cpu if task->cpu is
1373 * not idle and an idle cpu is available. The span of cpus to
1374 * search starts with cpus closest then further out as needed,
1375 * so we always favor a closer, idle cpu.
1377 * Returns the CPU we should wake onto.
1379 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1380 static int wake_idle(int cpu, struct task_struct *p)
1383 struct sched_domain *sd;
1389 for_each_domain(cpu, sd) {
1390 if (sd->flags & SD_WAKE_IDLE) {
1391 cpus_and(tmp, sd->span, p->cpus_allowed);
1392 for_each_cpu_mask(i, tmp) {
1403 static inline int wake_idle(int cpu, struct task_struct *p)
1410 * try_to_wake_up - wake up a thread
1411 * @p: the to-be-woken-up thread
1412 * @state: the mask of task states that can be woken
1413 * @sync: do a synchronous wakeup?
1415 * Put it on the run-queue if it's not already there. The "current"
1416 * thread is always on the run-queue (except when the actual
1417 * re-schedule is in progress), and as such you're allowed to do
1418 * the simpler "current->state = TASK_RUNNING" to mark yourself
1419 * runnable without the overhead of this.
1421 * returns failure only if the task is already active.
1423 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1425 int cpu, this_cpu, success = 0;
1426 unsigned long flags;
1430 struct sched_domain *sd, *this_sd = NULL;
1431 unsigned long load, this_load;
1435 rq = task_rq_lock(p, &flags);
1436 old_state = p->state;
1437 if (!(old_state & state))
1444 this_cpu = smp_processor_id();
1447 if (unlikely(task_running(rq, p)))
1452 schedstat_inc(rq, ttwu_cnt);
1453 if (cpu == this_cpu) {
1454 schedstat_inc(rq, ttwu_local);
1458 for_each_domain(this_cpu, sd) {
1459 if (cpu_isset(cpu, sd->span)) {
1460 schedstat_inc(sd, ttwu_wake_remote);
1466 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1470 * Check for affine wakeup and passive balancing possibilities.
1473 int idx = this_sd->wake_idx;
1474 unsigned int imbalance;
1476 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1478 load = source_load(cpu, idx);
1479 this_load = target_load(this_cpu, idx);
1481 new_cpu = this_cpu; /* Wake to this CPU if we can */
1483 if (this_sd->flags & SD_WAKE_AFFINE) {
1484 unsigned long tl = this_load;
1485 unsigned long tl_per_task;
1487 tl_per_task = cpu_avg_load_per_task(this_cpu);
1490 * If sync wakeup then subtract the (maximum possible)
1491 * effect of the currently running task from the load
1492 * of the current CPU:
1495 tl -= current->load_weight;
1498 tl + target_load(cpu, idx) <= tl_per_task) ||
1499 100*(tl + p->load_weight) <= imbalance*load) {
1501 * This domain has SD_WAKE_AFFINE and
1502 * p is cache cold in this domain, and
1503 * there is no bad imbalance.
1505 schedstat_inc(this_sd, ttwu_move_affine);
1511 * Start passive balancing when half the imbalance_pct
1514 if (this_sd->flags & SD_WAKE_BALANCE) {
1515 if (imbalance*this_load <= 100*load) {
1516 schedstat_inc(this_sd, ttwu_move_balance);
1522 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1524 new_cpu = wake_idle(new_cpu, p);
1525 if (new_cpu != cpu) {
1526 set_task_cpu(p, new_cpu);
1527 task_rq_unlock(rq, &flags);
1528 /* might preempt at this point */
1529 rq = task_rq_lock(p, &flags);
1530 old_state = p->state;
1531 if (!(old_state & state))
1536 this_cpu = smp_processor_id();
1541 #endif /* CONFIG_SMP */
1542 if (old_state == TASK_UNINTERRUPTIBLE) {
1543 rq->nr_uninterruptible--;
1545 * Tasks on involuntary sleep don't earn
1546 * sleep_avg beyond just interactive state.
1548 p->sleep_type = SLEEP_NONINTERACTIVE;
1552 * Tasks that have marked their sleep as noninteractive get
1553 * woken up with their sleep average not weighted in an
1556 if (old_state & TASK_NONINTERACTIVE)
1557 p->sleep_type = SLEEP_NONINTERACTIVE;
1560 activate_task(p, rq, cpu == this_cpu);
1562 * Sync wakeups (i.e. those types of wakeups where the waker
1563 * has indicated that it will leave the CPU in short order)
1564 * don't trigger a preemption, if the woken up task will run on
1565 * this cpu. (in this case the 'I will reschedule' promise of
1566 * the waker guarantees that the freshly woken up task is going
1567 * to be considered on this CPU.)
1569 if (!sync || cpu != this_cpu) {
1570 if (TASK_PREEMPTS_CURR(p, rq))
1571 resched_task(rq->curr);
1576 p->state = TASK_RUNNING;
1578 task_rq_unlock(rq, &flags);
1583 int fastcall wake_up_process(struct task_struct *p)
1585 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1586 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1588 EXPORT_SYMBOL(wake_up_process);
1590 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1592 return try_to_wake_up(p, state, 0);
1595 static void task_running_tick(struct rq *rq, struct task_struct *p);
1597 * Perform scheduler related setup for a newly forked process p.
1598 * p is forked by current.
1600 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1602 int cpu = get_cpu();
1605 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1607 set_task_cpu(p, cpu);
1610 * We mark the process as running here, but have not actually
1611 * inserted it onto the runqueue yet. This guarantees that
1612 * nobody will actually run it, and a signal or other external
1613 * event cannot wake it up and insert it on the runqueue either.
1615 p->state = TASK_RUNNING;
1618 * Make sure we do not leak PI boosting priority to the child:
1620 p->prio = current->normal_prio;
1622 INIT_LIST_HEAD(&p->run_list);
1624 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1625 if (unlikely(sched_info_on()))
1626 memset(&p->sched_info, 0, sizeof(p->sched_info));
1628 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1631 #ifdef CONFIG_PREEMPT
1632 /* Want to start with kernel preemption disabled. */
1633 task_thread_info(p)->preempt_count = 1;
1636 * Share the timeslice between parent and child, thus the
1637 * total amount of pending timeslices in the system doesn't change,
1638 * resulting in more scheduling fairness.
1640 local_irq_disable();
1641 p->time_slice = (current->time_slice + 1) >> 1;
1643 * The remainder of the first timeslice might be recovered by
1644 * the parent if the child exits early enough.
1646 p->first_time_slice = 1;
1647 current->time_slice >>= 1;
1648 p->timestamp = sched_clock();
1649 if (unlikely(!current->time_slice)) {
1651 * This case is rare, it happens when the parent has only
1652 * a single jiffy left from its timeslice. Taking the
1653 * runqueue lock is not a problem.
1655 current->time_slice = 1;
1656 task_running_tick(cpu_rq(cpu), current);
1663 * wake_up_new_task - wake up a newly created task for the first time.
1665 * This function will do some initial scheduler statistics housekeeping
1666 * that must be done for every newly created context, then puts the task
1667 * on the runqueue and wakes it.
1669 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1671 struct rq *rq, *this_rq;
1672 unsigned long flags;
1675 rq = task_rq_lock(p, &flags);
1676 BUG_ON(p->state != TASK_RUNNING);
1677 this_cpu = smp_processor_id();
1681 * We decrease the sleep average of forking parents
1682 * and children as well, to keep max-interactive tasks
1683 * from forking tasks that are max-interactive. The parent
1684 * (current) is done further down, under its lock.
1686 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1687 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1689 p->prio = effective_prio(p);
1691 if (likely(cpu == this_cpu)) {
1692 if (!(clone_flags & CLONE_VM)) {
1694 * The VM isn't cloned, so we're in a good position to
1695 * do child-runs-first in anticipation of an exec. This
1696 * usually avoids a lot of COW overhead.
1698 if (unlikely(!current->array))
1699 __activate_task(p, rq);
1701 p->prio = current->prio;
1702 p->normal_prio = current->normal_prio;
1703 list_add_tail(&p->run_list, ¤t->run_list);
1704 p->array = current->array;
1705 p->array->nr_active++;
1706 inc_nr_running(p, rq);
1710 /* Run child last */
1711 __activate_task(p, rq);
1713 * We skip the following code due to cpu == this_cpu
1715 * task_rq_unlock(rq, &flags);
1716 * this_rq = task_rq_lock(current, &flags);
1720 this_rq = cpu_rq(this_cpu);
1723 * Not the local CPU - must adjust timestamp. This should
1724 * get optimised away in the !CONFIG_SMP case.
1726 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1727 + rq->most_recent_timestamp;
1728 __activate_task(p, rq);
1729 if (TASK_PREEMPTS_CURR(p, rq))
1730 resched_task(rq->curr);
1733 * Parent and child are on different CPUs, now get the
1734 * parent runqueue to update the parent's ->sleep_avg:
1736 task_rq_unlock(rq, &flags);
1737 this_rq = task_rq_lock(current, &flags);
1739 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1740 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1741 task_rq_unlock(this_rq, &flags);
1745 * Potentially available exiting-child timeslices are
1746 * retrieved here - this way the parent does not get
1747 * penalized for creating too many threads.
1749 * (this cannot be used to 'generate' timeslices
1750 * artificially, because any timeslice recovered here
1751 * was given away by the parent in the first place.)
1753 void fastcall sched_exit(struct task_struct *p)
1755 unsigned long flags;
1759 * If the child was a (relative-) CPU hog then decrease
1760 * the sleep_avg of the parent as well.
1762 rq = task_rq_lock(p->parent, &flags);
1763 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1764 p->parent->time_slice += p->time_slice;
1765 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1766 p->parent->time_slice = task_timeslice(p);
1768 if (p->sleep_avg < p->parent->sleep_avg)
1769 p->parent->sleep_avg = p->parent->sleep_avg /
1770 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1772 task_rq_unlock(rq, &flags);
1776 * prepare_task_switch - prepare to switch tasks
1777 * @rq: the runqueue preparing to switch
1778 * @next: the task we are going to switch to.
1780 * This is called with the rq lock held and interrupts off. It must
1781 * be paired with a subsequent finish_task_switch after the context
1784 * prepare_task_switch sets up locking and calls architecture specific
1787 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1789 prepare_lock_switch(rq, next);
1790 prepare_arch_switch(next);
1794 * finish_task_switch - clean up after a task-switch
1795 * @rq: runqueue associated with task-switch
1796 * @prev: the thread we just switched away from.
1798 * finish_task_switch must be called after the context switch, paired
1799 * with a prepare_task_switch call before the context switch.
1800 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1801 * and do any other architecture-specific cleanup actions.
1803 * Note that we may have delayed dropping an mm in context_switch(). If
1804 * so, we finish that here outside of the runqueue lock. (Doing it
1805 * with the lock held can cause deadlocks; see schedule() for
1808 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1809 __releases(rq->lock)
1811 struct mm_struct *mm = rq->prev_mm;
1817 * A task struct has one reference for the use as "current".
1818 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1819 * schedule one last time. The schedule call will never return, and
1820 * the scheduled task must drop that reference.
1821 * The test for TASK_DEAD must occur while the runqueue locks are
1822 * still held, otherwise prev could be scheduled on another cpu, die
1823 * there before we look at prev->state, and then the reference would
1825 * Manfred Spraul <manfred@colorfullife.com>
1827 prev_state = prev->state;
1828 finish_arch_switch(prev);
1829 finish_lock_switch(rq, prev);
1832 if (unlikely(prev_state == TASK_DEAD)) {
1834 * Remove function-return probe instances associated with this
1835 * task and put them back on the free list.
1837 kprobe_flush_task(prev);
1838 put_task_struct(prev);
1843 * schedule_tail - first thing a freshly forked thread must call.
1844 * @prev: the thread we just switched away from.
1846 asmlinkage void schedule_tail(struct task_struct *prev)
1847 __releases(rq->lock)
1849 struct rq *rq = this_rq();
1851 finish_task_switch(rq, prev);
1852 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1853 /* In this case, finish_task_switch does not reenable preemption */
1856 if (current->set_child_tid)
1857 put_user(current->pid, current->set_child_tid);
1861 * context_switch - switch to the new MM and the new
1862 * thread's register state.
1864 static inline struct task_struct *
1865 context_switch(struct rq *rq, struct task_struct *prev,
1866 struct task_struct *next)
1868 struct mm_struct *mm = next->mm;
1869 struct mm_struct *oldmm = prev->active_mm;
1872 * For paravirt, this is coupled with an exit in switch_to to
1873 * combine the page table reload and the switch backend into
1876 arch_enter_lazy_cpu_mode();
1879 next->active_mm = oldmm;
1880 atomic_inc(&oldmm->mm_count);
1881 enter_lazy_tlb(oldmm, next);
1883 switch_mm(oldmm, mm, next);
1886 prev->active_mm = NULL;
1887 WARN_ON(rq->prev_mm);
1888 rq->prev_mm = oldmm;
1891 * Since the runqueue lock will be released by the next
1892 * task (which is an invalid locking op but in the case
1893 * of the scheduler it's an obvious special-case), so we
1894 * do an early lockdep release here:
1896 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1897 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1900 /* Here we just switch the register state and the stack. */
1901 switch_to(prev, next, prev);
1907 * nr_running, nr_uninterruptible and nr_context_switches:
1909 * externally visible scheduler statistics: current number of runnable
1910 * threads, current number of uninterruptible-sleeping threads, total
1911 * number of context switches performed since bootup.
1913 unsigned long nr_running(void)
1915 unsigned long i, sum = 0;
1917 for_each_online_cpu(i)
1918 sum += cpu_rq(i)->nr_running;
1923 unsigned long nr_uninterruptible(void)
1925 unsigned long i, sum = 0;
1927 for_each_possible_cpu(i)
1928 sum += cpu_rq(i)->nr_uninterruptible;
1931 * Since we read the counters lockless, it might be slightly
1932 * inaccurate. Do not allow it to go below zero though:
1934 if (unlikely((long)sum < 0))
1940 unsigned long long nr_context_switches(void)
1943 unsigned long long sum = 0;
1945 for_each_possible_cpu(i)
1946 sum += cpu_rq(i)->nr_switches;
1951 unsigned long nr_iowait(void)
1953 unsigned long i, sum = 0;
1955 for_each_possible_cpu(i)
1956 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1961 unsigned long nr_active(void)
1963 unsigned long i, running = 0, uninterruptible = 0;
1965 for_each_online_cpu(i) {
1966 running += cpu_rq(i)->nr_running;
1967 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1970 if (unlikely((long)uninterruptible < 0))
1971 uninterruptible = 0;
1973 return running + uninterruptible;
1979 * Is this task likely cache-hot:
1982 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1984 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1988 * double_rq_lock - safely lock two runqueues
1990 * Note this does not disable interrupts like task_rq_lock,
1991 * you need to do so manually before calling.
1993 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1994 __acquires(rq1->lock)
1995 __acquires(rq2->lock)
1997 BUG_ON(!irqs_disabled());
1999 spin_lock(&rq1->lock);
2000 __acquire(rq2->lock); /* Fake it out ;) */
2003 spin_lock(&rq1->lock);
2004 spin_lock(&rq2->lock);
2006 spin_lock(&rq2->lock);
2007 spin_lock(&rq1->lock);
2013 * double_rq_unlock - safely unlock two runqueues
2015 * Note this does not restore interrupts like task_rq_unlock,
2016 * you need to do so manually after calling.
2018 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2019 __releases(rq1->lock)
2020 __releases(rq2->lock)
2022 spin_unlock(&rq1->lock);
2024 spin_unlock(&rq2->lock);
2026 __release(rq2->lock);
2030 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2032 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2033 __releases(this_rq->lock)
2034 __acquires(busiest->lock)
2035 __acquires(this_rq->lock)
2037 if (unlikely(!irqs_disabled())) {
2038 /* printk() doesn't work good under rq->lock */
2039 spin_unlock(&this_rq->lock);
2042 if (unlikely(!spin_trylock(&busiest->lock))) {
2043 if (busiest < this_rq) {
2044 spin_unlock(&this_rq->lock);
2045 spin_lock(&busiest->lock);
2046 spin_lock(&this_rq->lock);
2048 spin_lock(&busiest->lock);
2053 * If dest_cpu is allowed for this process, migrate the task to it.
2054 * This is accomplished by forcing the cpu_allowed mask to only
2055 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2056 * the cpu_allowed mask is restored.
2058 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2060 struct migration_req req;
2061 unsigned long flags;
2064 rq = task_rq_lock(p, &flags);
2065 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2066 || unlikely(cpu_is_offline(dest_cpu)))
2069 /* force the process onto the specified CPU */
2070 if (migrate_task(p, dest_cpu, &req)) {
2071 /* Need to wait for migration thread (might exit: take ref). */
2072 struct task_struct *mt = rq->migration_thread;
2074 get_task_struct(mt);
2075 task_rq_unlock(rq, &flags);
2076 wake_up_process(mt);
2077 put_task_struct(mt);
2078 wait_for_completion(&req.done);
2083 task_rq_unlock(rq, &flags);
2087 * sched_exec - execve() is a valuable balancing opportunity, because at
2088 * this point the task has the smallest effective memory and cache footprint.
2090 void sched_exec(void)
2092 int new_cpu, this_cpu = get_cpu();
2093 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2095 if (new_cpu != this_cpu)
2096 sched_migrate_task(current, new_cpu);
2100 * pull_task - move a task from a remote runqueue to the local runqueue.
2101 * Both runqueues must be locked.
2103 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2104 struct task_struct *p, struct rq *this_rq,
2105 struct prio_array *this_array, int this_cpu)
2107 dequeue_task(p, src_array);
2108 dec_nr_running(p, src_rq);
2109 set_task_cpu(p, this_cpu);
2110 inc_nr_running(p, this_rq);
2111 enqueue_task(p, this_array);
2112 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2113 + this_rq->most_recent_timestamp;
2115 * Note that idle threads have a prio of MAX_PRIO, for this test
2116 * to be always true for them.
2118 if (TASK_PREEMPTS_CURR(p, this_rq))
2119 resched_task(this_rq->curr);
2123 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2126 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2127 struct sched_domain *sd, enum idle_type idle,
2131 * We do not migrate tasks that are:
2132 * 1) running (obviously), or
2133 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2134 * 3) are cache-hot on their current CPU.
2136 if (!cpu_isset(this_cpu, p->cpus_allowed))
2140 if (task_running(rq, p))
2144 * Aggressive migration if:
2145 * 1) task is cache cold, or
2146 * 2) too many balance attempts have failed.
2149 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2150 #ifdef CONFIG_SCHEDSTATS
2151 if (task_hot(p, rq->most_recent_timestamp, sd))
2152 schedstat_inc(sd, lb_hot_gained[idle]);
2157 if (task_hot(p, rq->most_recent_timestamp, sd))
2162 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2165 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2166 * load from busiest to this_rq, as part of a balancing operation within
2167 * "domain". Returns the number of tasks moved.
2169 * Called with both runqueues locked.
2171 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2172 unsigned long max_nr_move, unsigned long max_load_move,
2173 struct sched_domain *sd, enum idle_type idle,
2176 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2177 best_prio_seen, skip_for_load;
2178 struct prio_array *array, *dst_array;
2179 struct list_head *head, *curr;
2180 struct task_struct *tmp;
2183 if (max_nr_move == 0 || max_load_move == 0)
2186 rem_load_move = max_load_move;
2188 this_best_prio = rq_best_prio(this_rq);
2189 best_prio = rq_best_prio(busiest);
2191 * Enable handling of the case where there is more than one task
2192 * with the best priority. If the current running task is one
2193 * of those with prio==best_prio we know it won't be moved
2194 * and therefore it's safe to override the skip (based on load) of
2195 * any task we find with that prio.
2197 best_prio_seen = best_prio == busiest->curr->prio;
2200 * We first consider expired tasks. Those will likely not be
2201 * executed in the near future, and they are most likely to
2202 * be cache-cold, thus switching CPUs has the least effect
2205 if (busiest->expired->nr_active) {
2206 array = busiest->expired;
2207 dst_array = this_rq->expired;
2209 array = busiest->active;
2210 dst_array = this_rq->active;
2214 /* Start searching at priority 0: */
2218 idx = sched_find_first_bit(array->bitmap);
2220 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2221 if (idx >= MAX_PRIO) {
2222 if (array == busiest->expired && busiest->active->nr_active) {
2223 array = busiest->active;
2224 dst_array = this_rq->active;
2230 head = array->queue + idx;
2233 tmp = list_entry(curr, struct task_struct, run_list);
2238 * To help distribute high priority tasks accross CPUs we don't
2239 * skip a task if it will be the highest priority task (i.e. smallest
2240 * prio value) on its new queue regardless of its load weight
2242 skip_for_load = tmp->load_weight > rem_load_move;
2243 if (skip_for_load && idx < this_best_prio)
2244 skip_for_load = !best_prio_seen && idx == best_prio;
2245 if (skip_for_load ||
2246 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2248 best_prio_seen |= idx == best_prio;
2255 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2257 rem_load_move -= tmp->load_weight;
2260 * We only want to steal up to the prescribed number of tasks
2261 * and the prescribed amount of weighted load.
2263 if (pulled < max_nr_move && rem_load_move > 0) {
2264 if (idx < this_best_prio)
2265 this_best_prio = idx;
2273 * Right now, this is the only place pull_task() is called,
2274 * so we can safely collect pull_task() stats here rather than
2275 * inside pull_task().
2277 schedstat_add(sd, lb_gained[idle], pulled);
2280 *all_pinned = pinned;
2285 * find_busiest_group finds and returns the busiest CPU group within the
2286 * domain. It calculates and returns the amount of weighted load which
2287 * should be moved to restore balance via the imbalance parameter.
2289 static struct sched_group *
2290 find_busiest_group(struct sched_domain *sd, int this_cpu,
2291 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2292 cpumask_t *cpus, int *balance)
2294 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2295 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2296 unsigned long max_pull;
2297 unsigned long busiest_load_per_task, busiest_nr_running;
2298 unsigned long this_load_per_task, this_nr_running;
2300 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2301 int power_savings_balance = 1;
2302 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2303 unsigned long min_nr_running = ULONG_MAX;
2304 struct sched_group *group_min = NULL, *group_leader = NULL;
2307 max_load = this_load = total_load = total_pwr = 0;
2308 busiest_load_per_task = busiest_nr_running = 0;
2309 this_load_per_task = this_nr_running = 0;
2310 if (idle == NOT_IDLE)
2311 load_idx = sd->busy_idx;
2312 else if (idle == NEWLY_IDLE)
2313 load_idx = sd->newidle_idx;
2315 load_idx = sd->idle_idx;
2318 unsigned long load, group_capacity;
2321 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2322 unsigned long sum_nr_running, sum_weighted_load;
2324 local_group = cpu_isset(this_cpu, group->cpumask);
2327 balance_cpu = first_cpu(group->cpumask);
2329 /* Tally up the load of all CPUs in the group */
2330 sum_weighted_load = sum_nr_running = avg_load = 0;
2332 for_each_cpu_mask(i, group->cpumask) {
2335 if (!cpu_isset(i, *cpus))
2340 if (*sd_idle && !idle_cpu(i))
2343 /* Bias balancing toward cpus of our domain */
2345 if (idle_cpu(i) && !first_idle_cpu) {
2350 load = target_load(i, load_idx);
2352 load = source_load(i, load_idx);
2355 sum_nr_running += rq->nr_running;
2356 sum_weighted_load += rq->raw_weighted_load;
2360 * First idle cpu or the first cpu(busiest) in this sched group
2361 * is eligible for doing load balancing at this and above
2364 if (local_group && balance_cpu != this_cpu && balance) {
2369 total_load += avg_load;
2370 total_pwr += group->cpu_power;
2372 /* Adjust by relative CPU power of the group */
2373 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2375 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2378 this_load = avg_load;
2380 this_nr_running = sum_nr_running;
2381 this_load_per_task = sum_weighted_load;
2382 } else if (avg_load > max_load &&
2383 sum_nr_running > group_capacity) {
2384 max_load = avg_load;
2386 busiest_nr_running = sum_nr_running;
2387 busiest_load_per_task = sum_weighted_load;
2390 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2392 * Busy processors will not participate in power savings
2395 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2399 * If the local group is idle or completely loaded
2400 * no need to do power savings balance at this domain
2402 if (local_group && (this_nr_running >= group_capacity ||
2404 power_savings_balance = 0;
2407 * If a group is already running at full capacity or idle,
2408 * don't include that group in power savings calculations
2410 if (!power_savings_balance || sum_nr_running >= group_capacity
2415 * Calculate the group which has the least non-idle load.
2416 * This is the group from where we need to pick up the load
2419 if ((sum_nr_running < min_nr_running) ||
2420 (sum_nr_running == min_nr_running &&
2421 first_cpu(group->cpumask) <
2422 first_cpu(group_min->cpumask))) {
2424 min_nr_running = sum_nr_running;
2425 min_load_per_task = sum_weighted_load /
2430 * Calculate the group which is almost near its
2431 * capacity but still has some space to pick up some load
2432 * from other group and save more power
2434 if (sum_nr_running <= group_capacity - 1) {
2435 if (sum_nr_running > leader_nr_running ||
2436 (sum_nr_running == leader_nr_running &&
2437 first_cpu(group->cpumask) >
2438 first_cpu(group_leader->cpumask))) {
2439 group_leader = group;
2440 leader_nr_running = sum_nr_running;
2445 group = group->next;
2446 } while (group != sd->groups);
2448 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2451 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2453 if (this_load >= avg_load ||
2454 100*max_load <= sd->imbalance_pct*this_load)
2457 busiest_load_per_task /= busiest_nr_running;
2459 * We're trying to get all the cpus to the average_load, so we don't
2460 * want to push ourselves above the average load, nor do we wish to
2461 * reduce the max loaded cpu below the average load, as either of these
2462 * actions would just result in more rebalancing later, and ping-pong
2463 * tasks around. Thus we look for the minimum possible imbalance.
2464 * Negative imbalances (*we* are more loaded than anyone else) will
2465 * be counted as no imbalance for these purposes -- we can't fix that
2466 * by pulling tasks to us. Be careful of negative numbers as they'll
2467 * appear as very large values with unsigned longs.
2469 if (max_load <= busiest_load_per_task)
2473 * In the presence of smp nice balancing, certain scenarios can have
2474 * max load less than avg load(as we skip the groups at or below
2475 * its cpu_power, while calculating max_load..)
2477 if (max_load < avg_load) {
2479 goto small_imbalance;
2482 /* Don't want to pull so many tasks that a group would go idle */
2483 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2485 /* How much load to actually move to equalise the imbalance */
2486 *imbalance = min(max_pull * busiest->cpu_power,
2487 (avg_load - this_load) * this->cpu_power)
2491 * if *imbalance is less than the average load per runnable task
2492 * there is no gaurantee that any tasks will be moved so we'll have
2493 * a think about bumping its value to force at least one task to be
2496 if (*imbalance < busiest_load_per_task) {
2497 unsigned long tmp, pwr_now, pwr_move;
2501 pwr_move = pwr_now = 0;
2503 if (this_nr_running) {
2504 this_load_per_task /= this_nr_running;
2505 if (busiest_load_per_task > this_load_per_task)
2508 this_load_per_task = SCHED_LOAD_SCALE;
2510 if (max_load - this_load >= busiest_load_per_task * imbn) {
2511 *imbalance = busiest_load_per_task;
2516 * OK, we don't have enough imbalance to justify moving tasks,
2517 * however we may be able to increase total CPU power used by
2521 pwr_now += busiest->cpu_power *
2522 min(busiest_load_per_task, max_load);
2523 pwr_now += this->cpu_power *
2524 min(this_load_per_task, this_load);
2525 pwr_now /= SCHED_LOAD_SCALE;
2527 /* Amount of load we'd subtract */
2528 tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
2531 pwr_move += busiest->cpu_power *
2532 min(busiest_load_per_task, max_load - tmp);
2534 /* Amount of load we'd add */
2535 if (max_load * busiest->cpu_power <
2536 busiest_load_per_task * SCHED_LOAD_SCALE)
2537 tmp = max_load * busiest->cpu_power / this->cpu_power;
2539 tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
2541 pwr_move += this->cpu_power *
2542 min(this_load_per_task, this_load + tmp);
2543 pwr_move /= SCHED_LOAD_SCALE;
2545 /* Move if we gain throughput */
2546 if (pwr_move <= pwr_now)
2549 *imbalance = busiest_load_per_task;
2555 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2556 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2559 if (this == group_leader && group_leader != group_min) {
2560 *imbalance = min_load_per_task;
2570 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2573 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2574 unsigned long imbalance, cpumask_t *cpus)
2576 struct rq *busiest = NULL, *rq;
2577 unsigned long max_load = 0;
2580 for_each_cpu_mask(i, group->cpumask) {
2582 if (!cpu_isset(i, *cpus))
2587 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2590 if (rq->raw_weighted_load > max_load) {
2591 max_load = rq->raw_weighted_load;
2600 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2601 * so long as it is large enough.
2603 #define MAX_PINNED_INTERVAL 512
2605 static inline unsigned long minus_1_or_zero(unsigned long n)
2607 return n > 0 ? n - 1 : 0;
2611 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2612 * tasks if there is an imbalance.
2614 static int load_balance(int this_cpu, struct rq *this_rq,
2615 struct sched_domain *sd, enum idle_type idle,
2618 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2619 struct sched_group *group;
2620 unsigned long imbalance;
2622 cpumask_t cpus = CPU_MASK_ALL;
2623 unsigned long flags;
2626 * When power savings policy is enabled for the parent domain, idle
2627 * sibling can pick up load irrespective of busy siblings. In this case,
2628 * let the state of idle sibling percolate up as IDLE, instead of
2629 * portraying it as NOT_IDLE.
2631 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2632 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2635 schedstat_inc(sd, lb_cnt[idle]);
2638 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2645 schedstat_inc(sd, lb_nobusyg[idle]);
2649 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2651 schedstat_inc(sd, lb_nobusyq[idle]);
2655 BUG_ON(busiest == this_rq);
2657 schedstat_add(sd, lb_imbalance[idle], imbalance);
2660 if (busiest->nr_running > 1) {
2662 * Attempt to move tasks. If find_busiest_group has found
2663 * an imbalance but busiest->nr_running <= 1, the group is
2664 * still unbalanced. nr_moved simply stays zero, so it is
2665 * correctly treated as an imbalance.
2667 local_irq_save(flags);
2668 double_rq_lock(this_rq, busiest);
2669 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2670 minus_1_or_zero(busiest->nr_running),
2671 imbalance, sd, idle, &all_pinned);
2672 double_rq_unlock(this_rq, busiest);
2673 local_irq_restore(flags);
2676 * some other cpu did the load balance for us.
2678 if (nr_moved && this_cpu != smp_processor_id())
2679 resched_cpu(this_cpu);
2681 /* All tasks on this runqueue were pinned by CPU affinity */
2682 if (unlikely(all_pinned)) {
2683 cpu_clear(cpu_of(busiest), cpus);
2684 if (!cpus_empty(cpus))
2691 schedstat_inc(sd, lb_failed[idle]);
2692 sd->nr_balance_failed++;
2694 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2696 spin_lock_irqsave(&busiest->lock, flags);
2698 /* don't kick the migration_thread, if the curr
2699 * task on busiest cpu can't be moved to this_cpu
2701 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2702 spin_unlock_irqrestore(&busiest->lock, flags);
2704 goto out_one_pinned;
2707 if (!busiest->active_balance) {
2708 busiest->active_balance = 1;
2709 busiest->push_cpu = this_cpu;
2712 spin_unlock_irqrestore(&busiest->lock, flags);
2714 wake_up_process(busiest->migration_thread);
2717 * We've kicked active balancing, reset the failure
2720 sd->nr_balance_failed = sd->cache_nice_tries+1;
2723 sd->nr_balance_failed = 0;
2725 if (likely(!active_balance)) {
2726 /* We were unbalanced, so reset the balancing interval */
2727 sd->balance_interval = sd->min_interval;
2730 * If we've begun active balancing, start to back off. This
2731 * case may not be covered by the all_pinned logic if there
2732 * is only 1 task on the busy runqueue (because we don't call
2735 if (sd->balance_interval < sd->max_interval)
2736 sd->balance_interval *= 2;
2739 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2740 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2745 schedstat_inc(sd, lb_balanced[idle]);
2747 sd->nr_balance_failed = 0;
2750 /* tune up the balancing interval */
2751 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2752 (sd->balance_interval < sd->max_interval))
2753 sd->balance_interval *= 2;
2755 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2756 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2762 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2763 * tasks if there is an imbalance.
2765 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2766 * this_rq is locked.
2769 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2771 struct sched_group *group;
2772 struct rq *busiest = NULL;
2773 unsigned long imbalance;
2776 cpumask_t cpus = CPU_MASK_ALL;
2779 * When power savings policy is enabled for the parent domain, idle
2780 * sibling can pick up load irrespective of busy siblings. In this case,
2781 * let the state of idle sibling percolate up as IDLE, instead of
2782 * portraying it as NOT_IDLE.
2784 if (sd->flags & SD_SHARE_CPUPOWER &&
2785 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2788 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2790 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2791 &sd_idle, &cpus, NULL);
2793 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2797 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2800 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2804 BUG_ON(busiest == this_rq);
2806 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2809 if (busiest->nr_running > 1) {
2810 /* Attempt to move tasks */
2811 double_lock_balance(this_rq, busiest);
2812 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2813 minus_1_or_zero(busiest->nr_running),
2814 imbalance, sd, NEWLY_IDLE, NULL);
2815 spin_unlock(&busiest->lock);
2818 cpu_clear(cpu_of(busiest), cpus);
2819 if (!cpus_empty(cpus))
2825 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2826 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2827 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2830 sd->nr_balance_failed = 0;
2835 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2836 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2837 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2839 sd->nr_balance_failed = 0;
2845 * idle_balance is called by schedule() if this_cpu is about to become
2846 * idle. Attempts to pull tasks from other CPUs.
2848 static void idle_balance(int this_cpu, struct rq *this_rq)
2850 struct sched_domain *sd;
2851 int pulled_task = 0;
2852 unsigned long next_balance = jiffies + 60 * HZ;
2854 for_each_domain(this_cpu, sd) {
2855 if (sd->flags & SD_BALANCE_NEWIDLE) {
2856 /* If we've pulled tasks over stop searching: */
2857 pulled_task = load_balance_newidle(this_cpu,
2859 if (time_after(next_balance,
2860 sd->last_balance + sd->balance_interval))
2861 next_balance = sd->last_balance
2862 + sd->balance_interval;
2869 * We are going idle. next_balance may be set based on
2870 * a busy processor. So reset next_balance.
2872 this_rq->next_balance = next_balance;
2876 * active_load_balance is run by migration threads. It pushes running tasks
2877 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2878 * running on each physical CPU where possible, and avoids physical /
2879 * logical imbalances.
2881 * Called with busiest_rq locked.
2883 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2885 int target_cpu = busiest_rq->push_cpu;
2886 struct sched_domain *sd;
2887 struct rq *target_rq;
2889 /* Is there any task to move? */
2890 if (busiest_rq->nr_running <= 1)
2893 target_rq = cpu_rq(target_cpu);
2896 * This condition is "impossible", if it occurs
2897 * we need to fix it. Originally reported by
2898 * Bjorn Helgaas on a 128-cpu setup.
2900 BUG_ON(busiest_rq == target_rq);
2902 /* move a task from busiest_rq to target_rq */
2903 double_lock_balance(busiest_rq, target_rq);
2905 /* Search for an sd spanning us and the target CPU. */
2906 for_each_domain(target_cpu, sd) {
2907 if ((sd->flags & SD_LOAD_BALANCE) &&
2908 cpu_isset(busiest_cpu, sd->span))
2913 schedstat_inc(sd, alb_cnt);
2915 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2916 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2918 schedstat_inc(sd, alb_pushed);
2920 schedstat_inc(sd, alb_failed);
2922 spin_unlock(&target_rq->lock);
2925 static void update_load(struct rq *this_rq)
2927 unsigned long this_load;
2928 unsigned int i, scale;
2930 this_load = this_rq->raw_weighted_load;
2932 /* Update our load: */
2933 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2934 unsigned long old_load, new_load;
2936 /* scale is effectively 1 << i now, and >> i divides by scale */
2938 old_load = this_rq->cpu_load[i];
2939 new_load = this_load;
2941 * Round up the averaging division if load is increasing. This
2942 * prevents us from getting stuck on 9 if the load is 10, for
2945 if (new_load > old_load)
2946 new_load += scale-1;
2947 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2953 atomic_t load_balancer;
2955 } nohz ____cacheline_aligned = {
2956 .load_balancer = ATOMIC_INIT(-1),
2957 .cpu_mask = CPU_MASK_NONE,
2961 * This routine will try to nominate the ilb (idle load balancing)
2962 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2963 * load balancing on behalf of all those cpus. If all the cpus in the system
2964 * go into this tickless mode, then there will be no ilb owner (as there is
2965 * no need for one) and all the cpus will sleep till the next wakeup event
2968 * For the ilb owner, tick is not stopped. And this tick will be used
2969 * for idle load balancing. ilb owner will still be part of
2972 * While stopping the tick, this cpu will become the ilb owner if there
2973 * is no other owner. And will be the owner till that cpu becomes busy
2974 * or if all cpus in the system stop their ticks at which point
2975 * there is no need for ilb owner.
2977 * When the ilb owner becomes busy, it nominates another owner, during the
2978 * next busy scheduler_tick()
2980 int select_nohz_load_balancer(int stop_tick)
2982 int cpu = smp_processor_id();
2985 cpu_set(cpu, nohz.cpu_mask);
2986 cpu_rq(cpu)->in_nohz_recently = 1;
2989 * If we are going offline and still the leader, give up!
2991 if (cpu_is_offline(cpu) &&
2992 atomic_read(&nohz.load_balancer) == cpu) {
2993 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2998 /* time for ilb owner also to sleep */
2999 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3000 if (atomic_read(&nohz.load_balancer) == cpu)
3001 atomic_set(&nohz.load_balancer, -1);
3005 if (atomic_read(&nohz.load_balancer) == -1) {
3006 /* make me the ilb owner */
3007 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3009 } else if (atomic_read(&nohz.load_balancer) == cpu)
3012 if (!cpu_isset(cpu, nohz.cpu_mask))
3015 cpu_clear(cpu, nohz.cpu_mask);
3017 if (atomic_read(&nohz.load_balancer) == cpu)
3018 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3025 static DEFINE_SPINLOCK(balancing);
3028 * It checks each scheduling domain to see if it is due to be balanced,
3029 * and initiates a balancing operation if so.
3031 * Balancing parameters are set up in arch_init_sched_domains.
3033 static inline void rebalance_domains(int cpu, enum idle_type idle)
3036 struct rq *rq = cpu_rq(cpu);
3037 unsigned long interval;
3038 struct sched_domain *sd;
3039 /* Earliest time when we have to do rebalance again */
3040 unsigned long next_balance = jiffies + 60*HZ;
3042 for_each_domain(cpu, sd) {
3043 if (!(sd->flags & SD_LOAD_BALANCE))
3046 interval = sd->balance_interval;
3047 if (idle != SCHED_IDLE)
3048 interval *= sd->busy_factor;
3050 /* scale ms to jiffies */
3051 interval = msecs_to_jiffies(interval);
3052 if (unlikely(!interval))
3055 if (sd->flags & SD_SERIALIZE) {
3056 if (!spin_trylock(&balancing))
3060 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3061 if (load_balance(cpu, rq, sd, idle, &balance)) {
3063 * We've pulled tasks over so either we're no
3064 * longer idle, or one of our SMT siblings is
3069 sd->last_balance = jiffies;
3071 if (sd->flags & SD_SERIALIZE)
3072 spin_unlock(&balancing);
3074 if (time_after(next_balance, sd->last_balance + interval))
3075 next_balance = sd->last_balance + interval;
3078 * Stop the load balance at this level. There is another
3079 * CPU in our sched group which is doing load balancing more
3085 rq->next_balance = next_balance;
3089 * run_rebalance_domains is triggered when needed from the scheduler tick.
3090 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3091 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3093 static void run_rebalance_domains(struct softirq_action *h)
3095 int local_cpu = smp_processor_id();
3096 struct rq *local_rq = cpu_rq(local_cpu);
3097 enum idle_type idle = local_rq->idle_at_tick ? SCHED_IDLE : NOT_IDLE;
3099 rebalance_domains(local_cpu, idle);
3103 * If this cpu is the owner for idle load balancing, then do the
3104 * balancing on behalf of the other idle cpus whose ticks are
3107 if (local_rq->idle_at_tick &&
3108 atomic_read(&nohz.load_balancer) == local_cpu) {
3109 cpumask_t cpus = nohz.cpu_mask;
3113 cpu_clear(local_cpu, cpus);
3114 for_each_cpu_mask(balance_cpu, cpus) {
3116 * If this cpu gets work to do, stop the load balancing
3117 * work being done for other cpus. Next load
3118 * balancing owner will pick it up.
3123 rebalance_domains(balance_cpu, SCHED_IDLE);
3125 rq = cpu_rq(balance_cpu);
3126 if (time_after(local_rq->next_balance, rq->next_balance))
3127 local_rq->next_balance = rq->next_balance;
3134 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3136 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3137 * idle load balancing owner or decide to stop the periodic load balancing,
3138 * if the whole system is idle.
3140 static inline void trigger_load_balance(int cpu)
3142 struct rq *rq = cpu_rq(cpu);
3145 * If we were in the nohz mode recently and busy at the current
3146 * scheduler tick, then check if we need to nominate new idle
3149 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3150 rq->in_nohz_recently = 0;
3152 if (atomic_read(&nohz.load_balancer) == cpu) {
3153 cpu_clear(cpu, nohz.cpu_mask);
3154 atomic_set(&nohz.load_balancer, -1);
3157 if (atomic_read(&nohz.load_balancer) == -1) {
3159 * simple selection for now: Nominate the
3160 * first cpu in the nohz list to be the next
3163 * TBD: Traverse the sched domains and nominate
3164 * the nearest cpu in the nohz.cpu_mask.
3166 int ilb = first_cpu(nohz.cpu_mask);
3174 * If this cpu is idle and doing idle load balancing for all the
3175 * cpus with ticks stopped, is it time for that to stop?
3177 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3178 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3184 * If this cpu is idle and the idle load balancing is done by
3185 * someone else, then no need raise the SCHED_SOFTIRQ
3187 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3188 cpu_isset(cpu, nohz.cpu_mask))
3191 if (time_after_eq(jiffies, rq->next_balance))
3192 raise_softirq(SCHED_SOFTIRQ);
3196 * on UP we do not need to balance between CPUs:
3198 static inline void idle_balance(int cpu, struct rq *rq)
3203 DEFINE_PER_CPU(struct kernel_stat, kstat);
3205 EXPORT_PER_CPU_SYMBOL(kstat);
3208 * This is called on clock ticks and on context switches.
3209 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3212 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
3214 p->sched_time += now - p->last_ran;
3215 p->last_ran = rq->most_recent_timestamp = now;
3219 * Return current->sched_time plus any more ns on the sched_clock
3220 * that have not yet been banked.
3222 unsigned long long current_sched_time(const struct task_struct *p)
3224 unsigned long long ns;
3225 unsigned long flags;
3227 local_irq_save(flags);
3228 ns = p->sched_time + sched_clock() - p->last_ran;
3229 local_irq_restore(flags);
3235 * We place interactive tasks back into the active array, if possible.
3237 * To guarantee that this does not starve expired tasks we ignore the
3238 * interactivity of a task if the first expired task had to wait more
3239 * than a 'reasonable' amount of time. This deadline timeout is
3240 * load-dependent, as the frequency of array switched decreases with
3241 * increasing number of running tasks. We also ignore the interactivity
3242 * if a better static_prio task has expired:
3244 static inline int expired_starving(struct rq *rq)
3246 if (rq->curr->static_prio > rq->best_expired_prio)
3248 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3250 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3256 * Account user cpu time to a process.
3257 * @p: the process that the cpu time gets accounted to
3258 * @hardirq_offset: the offset to subtract from hardirq_count()
3259 * @cputime: the cpu time spent in user space since the last update
3261 void account_user_time(struct task_struct *p, cputime_t cputime)
3263 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3266 p->utime = cputime_add(p->utime, cputime);
3268 /* Add user time to cpustat. */
3269 tmp = cputime_to_cputime64(cputime);
3270 if (TASK_NICE(p) > 0)
3271 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3273 cpustat->user = cputime64_add(cpustat->user, tmp);
3277 * Account system cpu time to a process.
3278 * @p: the process that the cpu time gets accounted to
3279 * @hardirq_offset: the offset to subtract from hardirq_count()
3280 * @cputime: the cpu time spent in kernel space since the last update
3282 void account_system_time(struct task_struct *p, int hardirq_offset,
3285 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3286 struct rq *rq = this_rq();
3289 p->stime = cputime_add(p->stime, cputime);
3291 /* Add system time to cpustat. */
3292 tmp = cputime_to_cputime64(cputime);
3293 if (hardirq_count() - hardirq_offset)
3294 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3295 else if (softirq_count())
3296 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3297 else if (p != rq->idle)
3298 cpustat->system = cputime64_add(cpustat->system, tmp);
3299 else if (atomic_read(&rq->nr_iowait) > 0)
3300 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3302 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3303 /* Account for system time used */
3304 acct_update_integrals(p);
3308 * Account for involuntary wait time.
3309 * @p: the process from which the cpu time has been stolen
3310 * @steal: the cpu time spent in involuntary wait
3312 void account_steal_time(struct task_struct *p, cputime_t steal)
3314 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3315 cputime64_t tmp = cputime_to_cputime64(steal);
3316 struct rq *rq = this_rq();
3318 if (p == rq->idle) {
3319 p->stime = cputime_add(p->stime, steal);
3320 if (atomic_read(&rq->nr_iowait) > 0)
3321 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3323 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3325 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3328 static void task_running_tick(struct rq *rq, struct task_struct *p)
3330 if (p->array != rq->active) {
3331 /* Task has expired but was not scheduled yet */
3332 set_tsk_need_resched(p);
3335 spin_lock(&rq->lock);
3337 * The task was running during this tick - update the
3338 * time slice counter. Note: we do not update a thread's
3339 * priority until it either goes to sleep or uses up its
3340 * timeslice. This makes it possible for interactive tasks
3341 * to use up their timeslices at their highest priority levels.
3345 * RR tasks need a special form of timeslice management.
3346 * FIFO tasks have no timeslices.
3348 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3349 p->time_slice = task_timeslice(p);
3350 p->first_time_slice = 0;
3351 set_tsk_need_resched(p);
3353 /* put it at the end of the queue: */
3354 requeue_task(p, rq->active);
3358 if (!--p->time_slice) {
3359 dequeue_task(p, rq->active);
3360 set_tsk_need_resched(p);
3361 p->prio = effective_prio(p);
3362 p->time_slice = task_timeslice(p);
3363 p->first_time_slice = 0;
3365 if (!rq->expired_timestamp)
3366 rq->expired_timestamp = jiffies;
3367 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3368 enqueue_task(p, rq->expired);
3369 if (p->static_prio < rq->best_expired_prio)
3370 rq->best_expired_prio = p->static_prio;
3372 enqueue_task(p, rq->active);
3375 * Prevent a too long timeslice allowing a task to monopolize
3376 * the CPU. We do this by splitting up the timeslice into
3379 * Note: this does not mean the task's timeslices expire or
3380 * get lost in any way, they just might be preempted by
3381 * another task of equal priority. (one with higher
3382 * priority would have preempted this task already.) We
3383 * requeue this task to the end of the list on this priority
3384 * level, which is in essence a round-robin of tasks with
3387 * This only applies to tasks in the interactive
3388 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3390 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3391 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3392 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3393 (p->array == rq->active)) {
3395 requeue_task(p, rq->active);
3396 set_tsk_need_resched(p);
3400 spin_unlock(&rq->lock);
3404 * This function gets called by the timer code, with HZ frequency.
3405 * We call it with interrupts disabled.
3407 * It also gets called by the fork code, when changing the parent's
3410 void scheduler_tick(void)
3412 unsigned long long now = sched_clock();
3413 struct task_struct *p = current;
3414 int cpu = smp_processor_id();
3415 int idle_at_tick = idle_cpu(cpu);
3416 struct rq *rq = cpu_rq(cpu);
3418 update_cpu_clock(p, rq, now);
3421 task_running_tick(rq, p);
3424 rq->idle_at_tick = idle_at_tick;
3425 trigger_load_balance(cpu);
3429 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3431 void fastcall add_preempt_count(int val)
3436 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3438 preempt_count() += val;
3440 * Spinlock count overflowing soon?
3442 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3445 EXPORT_SYMBOL(add_preempt_count);
3447 void fastcall sub_preempt_count(int val)
3452 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3455 * Is the spinlock portion underflowing?
3457 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3458 !(preempt_count() & PREEMPT_MASK)))
3461 preempt_count() -= val;
3463 EXPORT_SYMBOL(sub_preempt_count);
3467 static inline int interactive_sleep(enum sleep_type sleep_type)
3469 return (sleep_type == SLEEP_INTERACTIVE ||
3470 sleep_type == SLEEP_INTERRUPTED);
3474 * schedule() is the main scheduler function.
3476 asmlinkage void __sched schedule(void)
3478 struct task_struct *prev, *next;
3479 struct prio_array *array;
3480 struct list_head *queue;
3481 unsigned long long now;
3482 unsigned long run_time;
3483 int cpu, idx, new_prio;
3488 * Test if we are atomic. Since do_exit() needs to call into
3489 * schedule() atomically, we ignore that path for now.
3490 * Otherwise, whine if we are scheduling when we should not be.
3492 if (unlikely(in_atomic() && !current->exit_state)) {
3493 printk(KERN_ERR "BUG: scheduling while atomic: "
3495 current->comm, preempt_count(), current->pid);
3496 debug_show_held_locks(current);
3497 if (irqs_disabled())
3498 print_irqtrace_events(current);
3501 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3506 release_kernel_lock(prev);
3507 need_resched_nonpreemptible:
3511 * The idle thread is not allowed to schedule!
3512 * Remove this check after it has been exercised a bit.
3514 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3515 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3519 schedstat_inc(rq, sched_cnt);
3520 now = sched_clock();
3521 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3522 run_time = now - prev->timestamp;
3523 if (unlikely((long long)(now - prev->timestamp) < 0))
3526 run_time = NS_MAX_SLEEP_AVG;
3529 * Tasks charged proportionately less run_time at high sleep_avg to
3530 * delay them losing their interactive status
3532 run_time /= (CURRENT_BONUS(prev) ? : 1);
3534 spin_lock_irq(&rq->lock);
3536 switch_count = &prev->nivcsw;
3537 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3538 switch_count = &prev->nvcsw;
3539 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3540 unlikely(signal_pending(prev))))
3541 prev->state = TASK_RUNNING;
3543 if (prev->state == TASK_UNINTERRUPTIBLE)
3544 rq->nr_uninterruptible++;
3545 deactivate_task(prev, rq);
3549 cpu = smp_processor_id();
3550 if (unlikely(!rq->nr_running)) {
3551 idle_balance(cpu, rq);
3552 if (!rq->nr_running) {
3554 rq->expired_timestamp = 0;
3560 if (unlikely(!array->nr_active)) {
3562 * Switch the active and expired arrays.
3564 schedstat_inc(rq, sched_switch);
3565 rq->active = rq->expired;
3566 rq->expired = array;
3568 rq->expired_timestamp = 0;
3569 rq->best_expired_prio = MAX_PRIO;
3572 idx = sched_find_first_bit(array->bitmap);
3573 queue = array->queue + idx;
3574 next = list_entry(queue->next, struct task_struct, run_list);
3576 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3577 unsigned long long delta = now - next->timestamp;
3578 if (unlikely((long long)(now - next->timestamp) < 0))
3581 if (next->sleep_type == SLEEP_INTERACTIVE)
3582 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3584 array = next->array;
3585 new_prio = recalc_task_prio(next, next->timestamp + delta);
3587 if (unlikely(next->prio != new_prio)) {
3588 dequeue_task(next, array);
3589 next->prio = new_prio;
3590 enqueue_task(next, array);
3593 next->sleep_type = SLEEP_NORMAL;
3595 if (next == rq->idle)
3596 schedstat_inc(rq, sched_goidle);
3598 prefetch_stack(next);
3599 clear_tsk_need_resched(prev);
3600 rcu_qsctr_inc(task_cpu(prev));
3602 update_cpu_clock(prev, rq, now);
3604 prev->sleep_avg -= run_time;
3605 if ((long)prev->sleep_avg <= 0)
3606 prev->sleep_avg = 0;
3607 prev->timestamp = prev->last_ran = now;
3609 sched_info_switch(prev, next);
3610 if (likely(prev != next)) {
3611 next->timestamp = next->last_ran = now;
3616 prepare_task_switch(rq, next);
3617 prev = context_switch(rq, prev, next);
3620 * this_rq must be evaluated again because prev may have moved
3621 * CPUs since it called schedule(), thus the 'rq' on its stack
3622 * frame will be invalid.
3624 finish_task_switch(this_rq(), prev);
3626 spin_unlock_irq(&rq->lock);
3629 if (unlikely(reacquire_kernel_lock(prev) < 0))
3630 goto need_resched_nonpreemptible;
3631 preempt_enable_no_resched();
3632 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3635 EXPORT_SYMBOL(schedule);
3637 #ifdef CONFIG_PREEMPT
3639 * this is the entry point to schedule() from in-kernel preemption
3640 * off of preempt_enable. Kernel preemptions off return from interrupt
3641 * occur there and call schedule directly.
3643 asmlinkage void __sched preempt_schedule(void)
3645 struct thread_info *ti = current_thread_info();
3646 #ifdef CONFIG_PREEMPT_BKL
3647 struct task_struct *task = current;
3648 int saved_lock_depth;
3651 * If there is a non-zero preempt_count or interrupts are disabled,
3652 * we do not want to preempt the current task. Just return..
3654 if (likely(ti->preempt_count || irqs_disabled()))
3658 add_preempt_count(PREEMPT_ACTIVE);
3660 * We keep the big kernel semaphore locked, but we
3661 * clear ->lock_depth so that schedule() doesnt
3662 * auto-release the semaphore:
3664 #ifdef CONFIG_PREEMPT_BKL
3665 saved_lock_depth = task->lock_depth;
3666 task->lock_depth = -1;
3669 #ifdef CONFIG_PREEMPT_BKL
3670 task->lock_depth = saved_lock_depth;
3672 sub_preempt_count(PREEMPT_ACTIVE);
3674 /* we could miss a preemption opportunity between schedule and now */
3676 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3679 EXPORT_SYMBOL(preempt_schedule);
3682 * this is the entry point to schedule() from kernel preemption
3683 * off of irq context.
3684 * Note, that this is called and return with irqs disabled. This will
3685 * protect us against recursive calling from irq.
3687 asmlinkage void __sched preempt_schedule_irq(void)
3689 struct thread_info *ti = current_thread_info();
3690 #ifdef CONFIG_PREEMPT_BKL
3691 struct task_struct *task = current;
3692 int saved_lock_depth;
3694 /* Catch callers which need to be fixed */
3695 BUG_ON(ti->preempt_count || !irqs_disabled());
3698 add_preempt_count(PREEMPT_ACTIVE);
3700 * We keep the big kernel semaphore locked, but we
3701 * clear ->lock_depth so that schedule() doesnt
3702 * auto-release the semaphore:
3704 #ifdef CONFIG_PREEMPT_BKL
3705 saved_lock_depth = task->lock_depth;
3706 task->lock_depth = -1;
3710 local_irq_disable();
3711 #ifdef CONFIG_PREEMPT_BKL
3712 task->lock_depth = saved_lock_depth;
3714 sub_preempt_count(PREEMPT_ACTIVE);
3716 /* we could miss a preemption opportunity between schedule and now */
3718 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3722 #endif /* CONFIG_PREEMPT */
3724 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3727 return try_to_wake_up(curr->private, mode, sync);
3729 EXPORT_SYMBOL(default_wake_function);
3732 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3733 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3734 * number) then we wake all the non-exclusive tasks and one exclusive task.
3736 * There are circumstances in which we can try to wake a task which has already
3737 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3738 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3740 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3741 int nr_exclusive, int sync, void *key)
3743 struct list_head *tmp, *next;
3745 list_for_each_safe(tmp, next, &q->task_list) {
3746 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3747 unsigned flags = curr->flags;
3749 if (curr->func(curr, mode, sync, key) &&
3750 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3756 * __wake_up - wake up threads blocked on a waitqueue.
3758 * @mode: which threads
3759 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3760 * @key: is directly passed to the wakeup function
3762 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3763 int nr_exclusive, void *key)
3765 unsigned long flags;
3767 spin_lock_irqsave(&q->lock, flags);
3768 __wake_up_common(q, mode, nr_exclusive, 0, key);
3769 spin_unlock_irqrestore(&q->lock, flags);
3771 EXPORT_SYMBOL(__wake_up);
3774 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3776 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3778 __wake_up_common(q, mode, 1, 0, NULL);
3782 * __wake_up_sync - wake up threads blocked on a waitqueue.
3784 * @mode: which threads
3785 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3787 * The sync wakeup differs that the waker knows that it will schedule
3788 * away soon, so while the target thread will be woken up, it will not
3789 * be migrated to another CPU - ie. the two threads are 'synchronized'
3790 * with each other. This can prevent needless bouncing between CPUs.
3792 * On UP it can prevent extra preemption.
3795 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3797 unsigned long flags;
3803 if (unlikely(!nr_exclusive))
3806 spin_lock_irqsave(&q->lock, flags);
3807 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3808 spin_unlock_irqrestore(&q->lock, flags);
3810 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3812 void fastcall complete(struct completion *x)
3814 unsigned long flags;
3816 spin_lock_irqsave(&x->wait.lock, flags);
3818 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3820 spin_unlock_irqrestore(&x->wait.lock, flags);
3822 EXPORT_SYMBOL(complete);
3824 void fastcall complete_all(struct completion *x)
3826 unsigned long flags;
3828 spin_lock_irqsave(&x->wait.lock, flags);
3829 x->done += UINT_MAX/2;
3830 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3832 spin_unlock_irqrestore(&x->wait.lock, flags);
3834 EXPORT_SYMBOL(complete_all);
3836 void fastcall __sched wait_for_completion(struct completion *x)
3840 spin_lock_irq(&x->wait.lock);
3842 DECLARE_WAITQUEUE(wait, current);
3844 wait.flags |= WQ_FLAG_EXCLUSIVE;
3845 __add_wait_queue_tail(&x->wait, &wait);
3847 __set_current_state(TASK_UNINTERRUPTIBLE);
3848 spin_unlock_irq(&x->wait.lock);
3850 spin_lock_irq(&x->wait.lock);
3852 __remove_wait_queue(&x->wait, &wait);
3855 spin_unlock_irq(&x->wait.lock);
3857 EXPORT_SYMBOL(wait_for_completion);
3859 unsigned long fastcall __sched
3860 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3864 spin_lock_irq(&x->wait.lock);
3866 DECLARE_WAITQUEUE(wait, current);
3868 wait.flags |= WQ_FLAG_EXCLUSIVE;
3869 __add_wait_queue_tail(&x->wait, &wait);
3871 __set_current_state(TASK_UNINTERRUPTIBLE);
3872 spin_unlock_irq(&x->wait.lock);
3873 timeout = schedule_timeout(timeout);
3874 spin_lock_irq(&x->wait.lock);
3876 __remove_wait_queue(&x->wait, &wait);
3880 __remove_wait_queue(&x->wait, &wait);
3884 spin_unlock_irq(&x->wait.lock);
3887 EXPORT_SYMBOL(wait_for_completion_timeout);
3889 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3895 spin_lock_irq(&x->wait.lock);
3897 DECLARE_WAITQUEUE(wait, current);
3899 wait.flags |= WQ_FLAG_EXCLUSIVE;
3900 __add_wait_queue_tail(&x->wait, &wait);
3902 if (signal_pending(current)) {
3904 __remove_wait_queue(&x->wait, &wait);
3907 __set_current_state(TASK_INTERRUPTIBLE);
3908 spin_unlock_irq(&x->wait.lock);
3910 spin_lock_irq(&x->wait.lock);
3912 __remove_wait_queue(&x->wait, &wait);
3916 spin_unlock_irq(&x->wait.lock);
3920 EXPORT_SYMBOL(wait_for_completion_interruptible);
3922 unsigned long fastcall __sched
3923 wait_for_completion_interruptible_timeout(struct completion *x,
3924 unsigned long timeout)
3928 spin_lock_irq(&x->wait.lock);
3930 DECLARE_WAITQUEUE(wait, current);
3932 wait.flags |= WQ_FLAG_EXCLUSIVE;
3933 __add_wait_queue_tail(&x->wait, &wait);
3935 if (signal_pending(current)) {
3936 timeout = -ERESTARTSYS;
3937 __remove_wait_queue(&x->wait, &wait);
3940 __set_current_state(TASK_INTERRUPTIBLE);
3941 spin_unlock_irq(&x->wait.lock);
3942 timeout = schedule_timeout(timeout);
3943 spin_lock_irq(&x->wait.lock);
3945 __remove_wait_queue(&x->wait, &wait);
3949 __remove_wait_queue(&x->wait, &wait);
3953 spin_unlock_irq(&x->wait.lock);
3956 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3959 #define SLEEP_ON_VAR \
3960 unsigned long flags; \
3961 wait_queue_t wait; \
3962 init_waitqueue_entry(&wait, current);
3964 #define SLEEP_ON_HEAD \
3965 spin_lock_irqsave(&q->lock,flags); \
3966 __add_wait_queue(q, &wait); \
3967 spin_unlock(&q->lock);
3969 #define SLEEP_ON_TAIL \
3970 spin_lock_irq(&q->lock); \
3971 __remove_wait_queue(q, &wait); \
3972 spin_unlock_irqrestore(&q->lock, flags);
3974 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3978 current->state = TASK_INTERRUPTIBLE;
3984 EXPORT_SYMBOL(interruptible_sleep_on);
3986 long fastcall __sched
3987 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3991 current->state = TASK_INTERRUPTIBLE;
3994 timeout = schedule_timeout(timeout);
3999 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4001 void fastcall __sched sleep_on(wait_queue_head_t *q)
4005 current->state = TASK_UNINTERRUPTIBLE;
4011 EXPORT_SYMBOL(sleep_on);
4013 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4017 current->state = TASK_UNINTERRUPTIBLE;
4020 timeout = schedule_timeout(timeout);
4026 EXPORT_SYMBOL(sleep_on_timeout);
4028 #ifdef CONFIG_RT_MUTEXES
4031 * rt_mutex_setprio - set the current priority of a task
4033 * @prio: prio value (kernel-internal form)
4035 * This function changes the 'effective' priority of a task. It does
4036 * not touch ->normal_prio like __setscheduler().
4038 * Used by the rt_mutex code to implement priority inheritance logic.
4040 void rt_mutex_setprio(struct task_struct *p, int prio)
4042 struct prio_array *array;
4043 unsigned long flags;
4047 BUG_ON(prio < 0 || prio > MAX_PRIO);
4049 rq = task_rq_lock(p, &flags);
4054 dequeue_task(p, array);
4059 * If changing to an RT priority then queue it
4060 * in the active array!
4064 enqueue_task(p, array);
4066 * Reschedule if we are currently running on this runqueue and
4067 * our priority decreased, or if we are not currently running on
4068 * this runqueue and our priority is higher than the current's
4070 if (task_running(rq, p)) {
4071 if (p->prio > oldprio)
4072 resched_task(rq->curr);
4073 } else if (TASK_PREEMPTS_CURR(p, rq))
4074 resched_task(rq->curr);
4076 task_rq_unlock(rq, &flags);
4081 void set_user_nice(struct task_struct *p, long nice)
4083 struct prio_array *array;
4084 int old_prio, delta;
4085 unsigned long flags;
4088 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4091 * We have to be careful, if called from sys_setpriority(),
4092 * the task might be in the middle of scheduling on another CPU.
4094 rq = task_rq_lock(p, &flags);
4096 * The RT priorities are set via sched_setscheduler(), but we still
4097 * allow the 'normal' nice value to be set - but as expected
4098 * it wont have any effect on scheduling until the task is
4099 * not SCHED_NORMAL/SCHED_BATCH:
4101 if (has_rt_policy(p)) {
4102 p->static_prio = NICE_TO_PRIO(nice);
4107 dequeue_task(p, array);
4108 dec_raw_weighted_load(rq, p);
4111 p->static_prio = NICE_TO_PRIO(nice);
4114 p->prio = effective_prio(p);
4115 delta = p->prio - old_prio;
4118 enqueue_task(p, array);
4119 inc_raw_weighted_load(rq, p);
4121 * If the task increased its priority or is running and
4122 * lowered its priority, then reschedule its CPU:
4124 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4125 resched_task(rq->curr);
4128 task_rq_unlock(rq, &flags);
4130 EXPORT_SYMBOL(set_user_nice);
4133 * can_nice - check if a task can reduce its nice value
4137 int can_nice(const struct task_struct *p, const int nice)
4139 /* convert nice value [19,-20] to rlimit style value [1,40] */
4140 int nice_rlim = 20 - nice;
4142 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4143 capable(CAP_SYS_NICE));
4146 #ifdef __ARCH_WANT_SYS_NICE
4149 * sys_nice - change the priority of the current process.
4150 * @increment: priority increment
4152 * sys_setpriority is a more generic, but much slower function that
4153 * does similar things.
4155 asmlinkage long sys_nice(int increment)
4160 * Setpriority might change our priority at the same moment.
4161 * We don't have to worry. Conceptually one call occurs first
4162 * and we have a single winner.
4164 if (increment < -40)
4169 nice = PRIO_TO_NICE(current->static_prio) + increment;
4175 if (increment < 0 && !can_nice(current, nice))
4178 retval = security_task_setnice(current, nice);
4182 set_user_nice(current, nice);
4189 * task_prio - return the priority value of a given task.
4190 * @p: the task in question.
4192 * This is the priority value as seen by users in /proc.
4193 * RT tasks are offset by -200. Normal tasks are centered
4194 * around 0, value goes from -16 to +15.
4196 int task_prio(const struct task_struct *p)
4198 return p->prio - MAX_RT_PRIO;
4202 * task_nice - return the nice value of a given task.
4203 * @p: the task in question.
4205 int task_nice(const struct task_struct *p)
4207 return TASK_NICE(p);
4209 EXPORT_SYMBOL_GPL(task_nice);
4212 * idle_cpu - is a given cpu idle currently?
4213 * @cpu: the processor in question.
4215 int idle_cpu(int cpu)
4217 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4221 * idle_task - return the idle task for a given cpu.
4222 * @cpu: the processor in question.
4224 struct task_struct *idle_task(int cpu)
4226 return cpu_rq(cpu)->idle;
4230 * find_process_by_pid - find a process with a matching PID value.
4231 * @pid: the pid in question.
4233 static inline struct task_struct *find_process_by_pid(pid_t pid)
4235 return pid ? find_task_by_pid(pid) : current;
4238 /* Actually do priority change: must hold rq lock. */
4239 static void __setscheduler(struct task_struct *p, int policy, int prio)
4244 p->rt_priority = prio;
4245 p->normal_prio = normal_prio(p);
4246 /* we are holding p->pi_lock already */
4247 p->prio = rt_mutex_getprio(p);
4249 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4251 if (policy == SCHED_BATCH)
4257 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4258 * @p: the task in question.
4259 * @policy: new policy.
4260 * @param: structure containing the new RT priority.
4262 * NOTE that the task may be already dead.
4264 int sched_setscheduler(struct task_struct *p, int policy,
4265 struct sched_param *param)
4267 int retval, oldprio, oldpolicy = -1;
4268 struct prio_array *array;
4269 unsigned long flags;
4272 /* may grab non-irq protected spin_locks */
4273 BUG_ON(in_interrupt());
4275 /* double check policy once rq lock held */
4277 policy = oldpolicy = p->policy;
4278 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4279 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4282 * Valid priorities for SCHED_FIFO and SCHED_RR are
4283 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4286 if (param->sched_priority < 0 ||
4287 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4288 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4290 if (is_rt_policy(policy) != (param->sched_priority != 0))
4294 * Allow unprivileged RT tasks to decrease priority:
4296 if (!capable(CAP_SYS_NICE)) {
4297 if (is_rt_policy(policy)) {
4298 unsigned long rlim_rtprio;
4299 unsigned long flags;
4301 if (!lock_task_sighand(p, &flags))
4303 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4304 unlock_task_sighand(p, &flags);
4306 /* can't set/change the rt policy */
4307 if (policy != p->policy && !rlim_rtprio)
4310 /* can't increase priority */
4311 if (param->sched_priority > p->rt_priority &&
4312 param->sched_priority > rlim_rtprio)
4316 /* can't change other user's priorities */
4317 if ((current->euid != p->euid) &&
4318 (current->euid != p->uid))
4322 retval = security_task_setscheduler(p, policy, param);
4326 * make sure no PI-waiters arrive (or leave) while we are
4327 * changing the priority of the task:
4329 spin_lock_irqsave(&p->pi_lock, flags);
4331 * To be able to change p->policy safely, the apropriate
4332 * runqueue lock must be held.
4334 rq = __task_rq_lock(p);
4335 /* recheck policy now with rq lock held */
4336 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4337 policy = oldpolicy = -1;
4338 __task_rq_unlock(rq);
4339 spin_unlock_irqrestore(&p->pi_lock, flags);
4344 deactivate_task(p, rq);
4346 __setscheduler(p, policy, param->sched_priority);
4348 __activate_task(p, rq);
4350 * Reschedule if we are currently running on this runqueue and
4351 * our priority decreased, or if we are not currently running on
4352 * this runqueue and our priority is higher than the current's
4354 if (task_running(rq, p)) {
4355 if (p->prio > oldprio)
4356 resched_task(rq->curr);
4357 } else if (TASK_PREEMPTS_CURR(p, rq))
4358 resched_task(rq->curr);
4360 __task_rq_unlock(rq);
4361 spin_unlock_irqrestore(&p->pi_lock, flags);
4363 rt_mutex_adjust_pi(p);
4367 EXPORT_SYMBOL_GPL(sched_setscheduler);
4370 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4372 struct sched_param lparam;
4373 struct task_struct *p;
4376 if (!param || pid < 0)
4378 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4383 p = find_process_by_pid(pid);
4385 retval = sched_setscheduler(p, policy, &lparam);
4392 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4393 * @pid: the pid in question.
4394 * @policy: new policy.
4395 * @param: structure containing the new RT priority.
4397 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4398 struct sched_param __user *param)
4400 /* negative values for policy are not valid */
4404 return do_sched_setscheduler(pid, policy, param);
4408 * sys_sched_setparam - set/change the RT priority of a thread
4409 * @pid: the pid in question.
4410 * @param: structure containing the new RT priority.
4412 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4414 return do_sched_setscheduler(pid, -1, param);
4418 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4419 * @pid: the pid in question.
4421 asmlinkage long sys_sched_getscheduler(pid_t pid)
4423 struct task_struct *p;
4424 int retval = -EINVAL;
4430 read_lock(&tasklist_lock);
4431 p = find_process_by_pid(pid);
4433 retval = security_task_getscheduler(p);
4437 read_unlock(&tasklist_lock);
4444 * sys_sched_getscheduler - get the RT priority of a thread
4445 * @pid: the pid in question.
4446 * @param: structure containing the RT priority.
4448 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4450 struct sched_param lp;
4451 struct task_struct *p;
4452 int retval = -EINVAL;
4454 if (!param || pid < 0)
4457 read_lock(&tasklist_lock);
4458 p = find_process_by_pid(pid);
4463 retval = security_task_getscheduler(p);
4467 lp.sched_priority = p->rt_priority;
4468 read_unlock(&tasklist_lock);
4471 * This one might sleep, we cannot do it with a spinlock held ...
4473 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4479 read_unlock(&tasklist_lock);
4483 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4485 cpumask_t cpus_allowed;
4486 struct task_struct *p;
4490 read_lock(&tasklist_lock);
4492 p = find_process_by_pid(pid);
4494 read_unlock(&tasklist_lock);
4495 unlock_cpu_hotplug();
4500 * It is not safe to call set_cpus_allowed with the
4501 * tasklist_lock held. We will bump the task_struct's
4502 * usage count and then drop tasklist_lock.
4505 read_unlock(&tasklist_lock);
4508 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4509 !capable(CAP_SYS_NICE))
4512 retval = security_task_setscheduler(p, 0, NULL);
4516 cpus_allowed = cpuset_cpus_allowed(p);
4517 cpus_and(new_mask, new_mask, cpus_allowed);
4518 retval = set_cpus_allowed(p, new_mask);
4522 unlock_cpu_hotplug();
4526 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4527 cpumask_t *new_mask)
4529 if (len < sizeof(cpumask_t)) {
4530 memset(new_mask, 0, sizeof(cpumask_t));
4531 } else if (len > sizeof(cpumask_t)) {
4532 len = sizeof(cpumask_t);
4534 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4538 * sys_sched_setaffinity - set the cpu affinity of a process
4539 * @pid: pid of the process
4540 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4541 * @user_mask_ptr: user-space pointer to the new cpu mask
4543 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4544 unsigned long __user *user_mask_ptr)
4549 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4553 return sched_setaffinity(pid, new_mask);
4557 * Represents all cpu's present in the system
4558 * In systems capable of hotplug, this map could dynamically grow
4559 * as new cpu's are detected in the system via any platform specific
4560 * method, such as ACPI for e.g.
4563 cpumask_t cpu_present_map __read_mostly;
4564 EXPORT_SYMBOL(cpu_present_map);
4567 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4568 EXPORT_SYMBOL(cpu_online_map);
4570 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4571 EXPORT_SYMBOL(cpu_possible_map);
4574 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4576 struct task_struct *p;
4580 read_lock(&tasklist_lock);
4583 p = find_process_by_pid(pid);
4587 retval = security_task_getscheduler(p);
4591 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4594 read_unlock(&tasklist_lock);
4595 unlock_cpu_hotplug();
4603 * sys_sched_getaffinity - get the cpu affinity of a process
4604 * @pid: pid of the process
4605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4606 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4608 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4609 unsigned long __user *user_mask_ptr)
4614 if (len < sizeof(cpumask_t))
4617 ret = sched_getaffinity(pid, &mask);
4621 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4624 return sizeof(cpumask_t);
4628 * sys_sched_yield - yield the current processor to other threads.
4630 * This function yields the current CPU by moving the calling thread
4631 * to the expired array. If there are no other threads running on this
4632 * CPU then this function will return.
4634 asmlinkage long sys_sched_yield(void)
4636 struct rq *rq = this_rq_lock();
4637 struct prio_array *array = current->array, *target = rq->expired;
4639 schedstat_inc(rq, yld_cnt);
4641 * We implement yielding by moving the task into the expired
4644 * (special rule: RT tasks will just roundrobin in the active
4647 if (rt_task(current))
4648 target = rq->active;
4650 if (array->nr_active == 1) {
4651 schedstat_inc(rq, yld_act_empty);
4652 if (!rq->expired->nr_active)
4653 schedstat_inc(rq, yld_both_empty);
4654 } else if (!rq->expired->nr_active)
4655 schedstat_inc(rq, yld_exp_empty);
4657 if (array != target) {
4658 dequeue_task(current, array);
4659 enqueue_task(current, target);
4662 * requeue_task is cheaper so perform that if possible.
4664 requeue_task(current, array);
4667 * Since we are going to call schedule() anyway, there's
4668 * no need to preempt or enable interrupts:
4670 __release(rq->lock);
4671 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4672 _raw_spin_unlock(&rq->lock);
4673 preempt_enable_no_resched();
4680 static void __cond_resched(void)
4682 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4683 __might_sleep(__FILE__, __LINE__);
4686 * The BKS might be reacquired before we have dropped
4687 * PREEMPT_ACTIVE, which could trigger a second
4688 * cond_resched() call.
4691 add_preempt_count(PREEMPT_ACTIVE);
4693 sub_preempt_count(PREEMPT_ACTIVE);
4694 } while (need_resched());
4697 int __sched cond_resched(void)
4699 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4700 system_state == SYSTEM_RUNNING) {
4706 EXPORT_SYMBOL(cond_resched);
4709 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4710 * call schedule, and on return reacquire the lock.
4712 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4713 * operations here to prevent schedule() from being called twice (once via
4714 * spin_unlock(), once by hand).
4716 int cond_resched_lock(spinlock_t *lock)
4720 if (need_lockbreak(lock)) {
4726 if (need_resched() && system_state == SYSTEM_RUNNING) {
4727 spin_release(&lock->dep_map, 1, _THIS_IP_);
4728 _raw_spin_unlock(lock);
4729 preempt_enable_no_resched();
4736 EXPORT_SYMBOL(cond_resched_lock);
4738 int __sched cond_resched_softirq(void)
4740 BUG_ON(!in_softirq());
4742 if (need_resched() && system_state == SYSTEM_RUNNING) {
4743 raw_local_irq_disable();
4745 raw_local_irq_enable();
4752 EXPORT_SYMBOL(cond_resched_softirq);
4755 * yield - yield the current processor to other threads.
4757 * This is a shortcut for kernel-space yielding - it marks the
4758 * thread runnable and calls sys_sched_yield().
4760 void __sched yield(void)
4762 set_current_state(TASK_RUNNING);
4765 EXPORT_SYMBOL(yield);
4768 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4769 * that process accounting knows that this is a task in IO wait state.
4771 * But don't do that if it is a deliberate, throttling IO wait (this task
4772 * has set its backing_dev_info: the queue against which it should throttle)
4774 void __sched io_schedule(void)
4776 struct rq *rq = &__raw_get_cpu_var(runqueues);
4778 delayacct_blkio_start();
4779 atomic_inc(&rq->nr_iowait);
4781 atomic_dec(&rq->nr_iowait);
4782 delayacct_blkio_end();
4784 EXPORT_SYMBOL(io_schedule);
4786 long __sched io_schedule_timeout(long timeout)
4788 struct rq *rq = &__raw_get_cpu_var(runqueues);
4791 delayacct_blkio_start();
4792 atomic_inc(&rq->nr_iowait);
4793 ret = schedule_timeout(timeout);
4794 atomic_dec(&rq->nr_iowait);
4795 delayacct_blkio_end();
4800 * sys_sched_get_priority_max - return maximum RT priority.
4801 * @policy: scheduling class.
4803 * this syscall returns the maximum rt_priority that can be used
4804 * by a given scheduling class.
4806 asmlinkage long sys_sched_get_priority_max(int policy)
4813 ret = MAX_USER_RT_PRIO-1;
4824 * sys_sched_get_priority_min - return minimum RT priority.
4825 * @policy: scheduling class.
4827 * this syscall returns the minimum rt_priority that can be used
4828 * by a given scheduling class.
4830 asmlinkage long sys_sched_get_priority_min(int policy)
4847 * sys_sched_rr_get_interval - return the default timeslice of a process.
4848 * @pid: pid of the process.
4849 * @interval: userspace pointer to the timeslice value.
4851 * this syscall writes the default timeslice value of a given process
4852 * into the user-space timespec buffer. A value of '0' means infinity.
4855 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4857 struct task_struct *p;
4858 int retval = -EINVAL;
4865 read_lock(&tasklist_lock);
4866 p = find_process_by_pid(pid);
4870 retval = security_task_getscheduler(p);
4874 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4875 0 : task_timeslice(p), &t);
4876 read_unlock(&tasklist_lock);
4877 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4881 read_unlock(&tasklist_lock);
4885 static const char stat_nam[] = "RSDTtZX";
4887 static void show_task(struct task_struct *p)
4889 unsigned long free = 0;
4892 state = p->state ? __ffs(p->state) + 1 : 0;
4893 printk("%-13.13s %c", p->comm,
4894 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4895 #if (BITS_PER_LONG == 32)
4896 if (state == TASK_RUNNING)
4897 printk(" running ");
4899 printk(" %08lX ", thread_saved_pc(p));
4901 if (state == TASK_RUNNING)
4902 printk(" running task ");
4904 printk(" %016lx ", thread_saved_pc(p));
4906 #ifdef CONFIG_DEBUG_STACK_USAGE
4908 unsigned long *n = end_of_stack(p);
4911 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4914 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4916 printk(" (L-TLB)\n");
4918 printk(" (NOTLB)\n");
4920 if (state != TASK_RUNNING)
4921 show_stack(p, NULL);
4924 void show_state_filter(unsigned long state_filter)
4926 struct task_struct *g, *p;
4928 #if (BITS_PER_LONG == 32)
4931 printk(" task PC stack pid father child younger older\n");
4935 printk(" task PC stack pid father child younger older\n");
4937 read_lock(&tasklist_lock);
4938 do_each_thread(g, p) {
4940 * reset the NMI-timeout, listing all files on a slow
4941 * console might take alot of time:
4943 touch_nmi_watchdog();
4944 if (!state_filter || (p->state & state_filter))
4946 } while_each_thread(g, p);
4948 touch_all_softlockup_watchdogs();
4950 read_unlock(&tasklist_lock);
4952 * Only show locks if all tasks are dumped:
4954 if (state_filter == -1)
4955 debug_show_all_locks();
4959 * init_idle - set up an idle thread for a given CPU
4960 * @idle: task in question
4961 * @cpu: cpu the idle task belongs to
4963 * NOTE: this function does not set the idle thread's NEED_RESCHED
4964 * flag, to make booting more robust.
4966 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4968 struct rq *rq = cpu_rq(cpu);
4969 unsigned long flags;
4971 idle->timestamp = sched_clock();
4972 idle->sleep_avg = 0;
4974 idle->prio = idle->normal_prio = MAX_PRIO;
4975 idle->state = TASK_RUNNING;
4976 idle->cpus_allowed = cpumask_of_cpu(cpu);
4977 set_task_cpu(idle, cpu);
4979 spin_lock_irqsave(&rq->lock, flags);
4980 rq->curr = rq->idle = idle;
4981 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4984 spin_unlock_irqrestore(&rq->lock, flags);
4986 /* Set the preempt count _outside_ the spinlocks! */
4987 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4988 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4990 task_thread_info(idle)->preempt_count = 0;
4995 * In a system that switches off the HZ timer nohz_cpu_mask
4996 * indicates which cpus entered this state. This is used
4997 * in the rcu update to wait only for active cpus. For system
4998 * which do not switch off the HZ timer nohz_cpu_mask should
4999 * always be CPU_MASK_NONE.
5001 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5005 * This is how migration works:
5007 * 1) we queue a struct migration_req structure in the source CPU's
5008 * runqueue and wake up that CPU's migration thread.
5009 * 2) we down() the locked semaphore => thread blocks.
5010 * 3) migration thread wakes up (implicitly it forces the migrated
5011 * thread off the CPU)
5012 * 4) it gets the migration request and checks whether the migrated
5013 * task is still in the wrong runqueue.
5014 * 5) if it's in the wrong runqueue then the migration thread removes
5015 * it and puts it into the right queue.
5016 * 6) migration thread up()s the semaphore.
5017 * 7) we wake up and the migration is done.
5021 * Change a given task's CPU affinity. Migrate the thread to a
5022 * proper CPU and schedule it away if the CPU it's executing on
5023 * is removed from the allowed bitmask.
5025 * NOTE: the caller must have a valid reference to the task, the
5026 * task must not exit() & deallocate itself prematurely. The
5027 * call is not atomic; no spinlocks may be held.
5029 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5031 struct migration_req req;
5032 unsigned long flags;
5036 rq = task_rq_lock(p, &flags);
5037 if (!cpus_intersects(new_mask, cpu_online_map)) {
5042 p->cpus_allowed = new_mask;
5043 /* Can the task run on the task's current CPU? If so, we're done */
5044 if (cpu_isset(task_cpu(p), new_mask))
5047 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5048 /* Need help from migration thread: drop lock and wait. */
5049 task_rq_unlock(rq, &flags);
5050 wake_up_process(rq->migration_thread);
5051 wait_for_completion(&req.done);
5052 tlb_migrate_finish(p->mm);
5056 task_rq_unlock(rq, &flags);
5060 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5063 * Move (not current) task off this cpu, onto dest cpu. We're doing
5064 * this because either it can't run here any more (set_cpus_allowed()
5065 * away from this CPU, or CPU going down), or because we're
5066 * attempting to rebalance this task on exec (sched_exec).
5068 * So we race with normal scheduler movements, but that's OK, as long
5069 * as the task is no longer on this CPU.
5071 * Returns non-zero if task was successfully migrated.
5073 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5075 struct rq *rq_dest, *rq_src;
5078 if (unlikely(cpu_is_offline(dest_cpu)))
5081 rq_src = cpu_rq(src_cpu);
5082 rq_dest = cpu_rq(dest_cpu);
5084 double_rq_lock(rq_src, rq_dest);
5085 /* Already moved. */
5086 if (task_cpu(p) != src_cpu)
5088 /* Affinity changed (again). */
5089 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5092 set_task_cpu(p, dest_cpu);
5095 * Sync timestamp with rq_dest's before activating.
5096 * The same thing could be achieved by doing this step
5097 * afterwards, and pretending it was a local activate.
5098 * This way is cleaner and logically correct.
5100 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5101 + rq_dest->most_recent_timestamp;
5102 deactivate_task(p, rq_src);
5103 __activate_task(p, rq_dest);
5104 if (TASK_PREEMPTS_CURR(p, rq_dest))
5105 resched_task(rq_dest->curr);
5109 double_rq_unlock(rq_src, rq_dest);
5114 * migration_thread - this is a highprio system thread that performs
5115 * thread migration by bumping thread off CPU then 'pushing' onto
5118 static int migration_thread(void *data)
5120 int cpu = (long)data;
5124 BUG_ON(rq->migration_thread != current);
5126 set_current_state(TASK_INTERRUPTIBLE);
5127 while (!kthread_should_stop()) {
5128 struct migration_req *req;
5129 struct list_head *head;
5133 spin_lock_irq(&rq->lock);
5135 if (cpu_is_offline(cpu)) {
5136 spin_unlock_irq(&rq->lock);
5140 if (rq->active_balance) {
5141 active_load_balance(rq, cpu);
5142 rq->active_balance = 0;
5145 head = &rq->migration_queue;
5147 if (list_empty(head)) {
5148 spin_unlock_irq(&rq->lock);
5150 set_current_state(TASK_INTERRUPTIBLE);
5153 req = list_entry(head->next, struct migration_req, list);
5154 list_del_init(head->next);
5156 spin_unlock(&rq->lock);
5157 __migrate_task(req->task, cpu, req->dest_cpu);
5160 complete(&req->done);
5162 __set_current_state(TASK_RUNNING);
5166 /* Wait for kthread_stop */
5167 set_current_state(TASK_INTERRUPTIBLE);
5168 while (!kthread_should_stop()) {
5170 set_current_state(TASK_INTERRUPTIBLE);
5172 __set_current_state(TASK_RUNNING);
5176 #ifdef CONFIG_HOTPLUG_CPU
5178 * Figure out where task on dead CPU should go, use force if neccessary.
5179 * NOTE: interrupts should be disabled by the caller
5181 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5183 unsigned long flags;
5190 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5191 cpus_and(mask, mask, p->cpus_allowed);
5192 dest_cpu = any_online_cpu(mask);
5194 /* On any allowed CPU? */
5195 if (dest_cpu == NR_CPUS)
5196 dest_cpu = any_online_cpu(p->cpus_allowed);
5198 /* No more Mr. Nice Guy. */
5199 if (dest_cpu == NR_CPUS) {
5200 rq = task_rq_lock(p, &flags);
5201 cpus_setall(p->cpus_allowed);
5202 dest_cpu = any_online_cpu(p->cpus_allowed);
5203 task_rq_unlock(rq, &flags);
5206 * Don't tell them about moving exiting tasks or
5207 * kernel threads (both mm NULL), since they never
5210 if (p->mm && printk_ratelimit())
5211 printk(KERN_INFO "process %d (%s) no "
5212 "longer affine to cpu%d\n",
5213 p->pid, p->comm, dead_cpu);
5215 if (!__migrate_task(p, dead_cpu, dest_cpu))
5220 * While a dead CPU has no uninterruptible tasks queued at this point,
5221 * it might still have a nonzero ->nr_uninterruptible counter, because
5222 * for performance reasons the counter is not stricly tracking tasks to
5223 * their home CPUs. So we just add the counter to another CPU's counter,
5224 * to keep the global sum constant after CPU-down:
5226 static void migrate_nr_uninterruptible(struct rq *rq_src)
5228 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5229 unsigned long flags;
5231 local_irq_save(flags);
5232 double_rq_lock(rq_src, rq_dest);
5233 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5234 rq_src->nr_uninterruptible = 0;
5235 double_rq_unlock(rq_src, rq_dest);
5236 local_irq_restore(flags);
5239 /* Run through task list and migrate tasks from the dead cpu. */
5240 static void migrate_live_tasks(int src_cpu)
5242 struct task_struct *p, *t;
5244 write_lock_irq(&tasklist_lock);
5246 do_each_thread(t, p) {
5250 if (task_cpu(p) == src_cpu)
5251 move_task_off_dead_cpu(src_cpu, p);
5252 } while_each_thread(t, p);
5254 write_unlock_irq(&tasklist_lock);
5257 /* Schedules idle task to be the next runnable task on current CPU.
5258 * It does so by boosting its priority to highest possible and adding it to
5259 * the _front_ of the runqueue. Used by CPU offline code.
5261 void sched_idle_next(void)
5263 int this_cpu = smp_processor_id();
5264 struct rq *rq = cpu_rq(this_cpu);
5265 struct task_struct *p = rq->idle;
5266 unsigned long flags;
5268 /* cpu has to be offline */
5269 BUG_ON(cpu_online(this_cpu));
5272 * Strictly not necessary since rest of the CPUs are stopped by now
5273 * and interrupts disabled on the current cpu.
5275 spin_lock_irqsave(&rq->lock, flags);
5277 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5279 /* Add idle task to the _front_ of its priority queue: */
5280 __activate_idle_task(p, rq);
5282 spin_unlock_irqrestore(&rq->lock, flags);
5286 * Ensures that the idle task is using init_mm right before its cpu goes
5289 void idle_task_exit(void)
5291 struct mm_struct *mm = current->active_mm;
5293 BUG_ON(cpu_online(smp_processor_id()));
5296 switch_mm(mm, &init_mm, current);
5300 /* called under rq->lock with disabled interrupts */
5301 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5303 struct rq *rq = cpu_rq(dead_cpu);
5305 /* Must be exiting, otherwise would be on tasklist. */
5306 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5308 /* Cannot have done final schedule yet: would have vanished. */
5309 BUG_ON(p->state == TASK_DEAD);
5314 * Drop lock around migration; if someone else moves it,
5315 * that's OK. No task can be added to this CPU, so iteration is
5317 * NOTE: interrupts should be left disabled --dev@
5319 spin_unlock(&rq->lock);
5320 move_task_off_dead_cpu(dead_cpu, p);
5321 spin_lock(&rq->lock);
5326 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5327 static void migrate_dead_tasks(unsigned int dead_cpu)
5329 struct rq *rq = cpu_rq(dead_cpu);
5330 unsigned int arr, i;
5332 for (arr = 0; arr < 2; arr++) {
5333 for (i = 0; i < MAX_PRIO; i++) {
5334 struct list_head *list = &rq->arrays[arr].queue[i];
5336 while (!list_empty(list))
5337 migrate_dead(dead_cpu, list_entry(list->next,
5338 struct task_struct, run_list));
5342 #endif /* CONFIG_HOTPLUG_CPU */
5345 * migration_call - callback that gets triggered when a CPU is added.
5346 * Here we can start up the necessary migration thread for the new CPU.
5348 static int __cpuinit
5349 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5351 struct task_struct *p;
5352 int cpu = (long)hcpu;
5353 unsigned long flags;
5357 case CPU_UP_PREPARE:
5358 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5361 p->flags |= PF_NOFREEZE;
5362 kthread_bind(p, cpu);
5363 /* Must be high prio: stop_machine expects to yield to it. */
5364 rq = task_rq_lock(p, &flags);
5365 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5366 task_rq_unlock(rq, &flags);
5367 cpu_rq(cpu)->migration_thread = p;
5371 /* Strictly unneccessary, as first user will wake it. */
5372 wake_up_process(cpu_rq(cpu)->migration_thread);
5375 #ifdef CONFIG_HOTPLUG_CPU
5376 case CPU_UP_CANCELED:
5377 if (!cpu_rq(cpu)->migration_thread)
5379 /* Unbind it from offline cpu so it can run. Fall thru. */
5380 kthread_bind(cpu_rq(cpu)->migration_thread,
5381 any_online_cpu(cpu_online_map));
5382 kthread_stop(cpu_rq(cpu)->migration_thread);
5383 cpu_rq(cpu)->migration_thread = NULL;
5387 migrate_live_tasks(cpu);
5389 kthread_stop(rq->migration_thread);
5390 rq->migration_thread = NULL;
5391 /* Idle task back to normal (off runqueue, low prio) */
5392 rq = task_rq_lock(rq->idle, &flags);
5393 deactivate_task(rq->idle, rq);
5394 rq->idle->static_prio = MAX_PRIO;
5395 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5396 migrate_dead_tasks(cpu);
5397 task_rq_unlock(rq, &flags);
5398 migrate_nr_uninterruptible(rq);
5399 BUG_ON(rq->nr_running != 0);
5401 /* No need to migrate the tasks: it was best-effort if
5402 * they didn't do lock_cpu_hotplug(). Just wake up
5403 * the requestors. */
5404 spin_lock_irq(&rq->lock);
5405 while (!list_empty(&rq->migration_queue)) {
5406 struct migration_req *req;
5408 req = list_entry(rq->migration_queue.next,
5409 struct migration_req, list);
5410 list_del_init(&req->list);
5411 complete(&req->done);
5413 spin_unlock_irq(&rq->lock);
5420 /* Register at highest priority so that task migration (migrate_all_tasks)
5421 * happens before everything else.
5423 static struct notifier_block __cpuinitdata migration_notifier = {
5424 .notifier_call = migration_call,
5428 int __init migration_init(void)
5430 void *cpu = (void *)(long)smp_processor_id();
5433 /* Start one for the boot CPU: */
5434 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5435 BUG_ON(err == NOTIFY_BAD);
5436 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5437 register_cpu_notifier(&migration_notifier);
5445 /* Number of possible processor ids */
5446 int nr_cpu_ids __read_mostly = NR_CPUS;
5447 EXPORT_SYMBOL(nr_cpu_ids);
5449 #undef SCHED_DOMAIN_DEBUG
5450 #ifdef SCHED_DOMAIN_DEBUG
5451 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5456 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5460 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5465 struct sched_group *group = sd->groups;
5466 cpumask_t groupmask;
5468 cpumask_scnprintf(str, NR_CPUS, sd->span);
5469 cpus_clear(groupmask);
5472 for (i = 0; i < level + 1; i++)
5474 printk("domain %d: ", level);
5476 if (!(sd->flags & SD_LOAD_BALANCE)) {
5477 printk("does not load-balance\n");
5479 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5484 printk("span %s\n", str);
5486 if (!cpu_isset(cpu, sd->span))
5487 printk(KERN_ERR "ERROR: domain->span does not contain "
5489 if (!cpu_isset(cpu, group->cpumask))
5490 printk(KERN_ERR "ERROR: domain->groups does not contain"
5494 for (i = 0; i < level + 2; i++)
5500 printk(KERN_ERR "ERROR: group is NULL\n");
5504 if (!group->cpu_power) {
5506 printk(KERN_ERR "ERROR: domain->cpu_power not "
5510 if (!cpus_weight(group->cpumask)) {
5512 printk(KERN_ERR "ERROR: empty group\n");
5515 if (cpus_intersects(groupmask, group->cpumask)) {
5517 printk(KERN_ERR "ERROR: repeated CPUs\n");
5520 cpus_or(groupmask, groupmask, group->cpumask);
5522 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5525 group = group->next;
5526 } while (group != sd->groups);
5529 if (!cpus_equal(sd->span, groupmask))
5530 printk(KERN_ERR "ERROR: groups don't span "
5538 if (!cpus_subset(groupmask, sd->span))
5539 printk(KERN_ERR "ERROR: parent span is not a superset "
5540 "of domain->span\n");
5545 # define sched_domain_debug(sd, cpu) do { } while (0)
5548 static int sd_degenerate(struct sched_domain *sd)
5550 if (cpus_weight(sd->span) == 1)
5553 /* Following flags need at least 2 groups */
5554 if (sd->flags & (SD_LOAD_BALANCE |
5555 SD_BALANCE_NEWIDLE |
5559 SD_SHARE_PKG_RESOURCES)) {
5560 if (sd->groups != sd->groups->next)
5564 /* Following flags don't use groups */
5565 if (sd->flags & (SD_WAKE_IDLE |
5574 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5576 unsigned long cflags = sd->flags, pflags = parent->flags;
5578 if (sd_degenerate(parent))
5581 if (!cpus_equal(sd->span, parent->span))
5584 /* Does parent contain flags not in child? */
5585 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5586 if (cflags & SD_WAKE_AFFINE)
5587 pflags &= ~SD_WAKE_BALANCE;
5588 /* Flags needing groups don't count if only 1 group in parent */
5589 if (parent->groups == parent->groups->next) {
5590 pflags &= ~(SD_LOAD_BALANCE |
5591 SD_BALANCE_NEWIDLE |
5595 SD_SHARE_PKG_RESOURCES);
5597 if (~cflags & pflags)
5604 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5605 * hold the hotplug lock.
5607 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5609 struct rq *rq = cpu_rq(cpu);
5610 struct sched_domain *tmp;
5612 /* Remove the sched domains which do not contribute to scheduling. */
5613 for (tmp = sd; tmp; tmp = tmp->parent) {
5614 struct sched_domain *parent = tmp->parent;
5617 if (sd_parent_degenerate(tmp, parent)) {
5618 tmp->parent = parent->parent;
5620 parent->parent->child = tmp;
5624 if (sd && sd_degenerate(sd)) {
5630 sched_domain_debug(sd, cpu);
5632 rcu_assign_pointer(rq->sd, sd);
5635 /* cpus with isolated domains */
5636 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5638 /* Setup the mask of cpus configured for isolated domains */
5639 static int __init isolated_cpu_setup(char *str)
5641 int ints[NR_CPUS], i;
5643 str = get_options(str, ARRAY_SIZE(ints), ints);
5644 cpus_clear(cpu_isolated_map);
5645 for (i = 1; i <= ints[0]; i++)
5646 if (ints[i] < NR_CPUS)
5647 cpu_set(ints[i], cpu_isolated_map);
5651 __setup ("isolcpus=", isolated_cpu_setup);
5654 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5655 * to a function which identifies what group(along with sched group) a CPU
5656 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5657 * (due to the fact that we keep track of groups covered with a cpumask_t).
5659 * init_sched_build_groups will build a circular linked list of the groups
5660 * covered by the given span, and will set each group's ->cpumask correctly,
5661 * and ->cpu_power to 0.
5664 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5665 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5666 struct sched_group **sg))
5668 struct sched_group *first = NULL, *last = NULL;
5669 cpumask_t covered = CPU_MASK_NONE;
5672 for_each_cpu_mask(i, span) {
5673 struct sched_group *sg;
5674 int group = group_fn(i, cpu_map, &sg);
5677 if (cpu_isset(i, covered))
5680 sg->cpumask = CPU_MASK_NONE;
5683 for_each_cpu_mask(j, span) {
5684 if (group_fn(j, cpu_map, NULL) != group)
5687 cpu_set(j, covered);
5688 cpu_set(j, sg->cpumask);
5699 #define SD_NODES_PER_DOMAIN 16
5702 * Self-tuning task migration cost measurement between source and target CPUs.
5704 * This is done by measuring the cost of manipulating buffers of varying
5705 * sizes. For a given buffer-size here are the steps that are taken:
5707 * 1) the source CPU reads+dirties a shared buffer
5708 * 2) the target CPU reads+dirties the same shared buffer
5710 * We measure how long they take, in the following 4 scenarios:
5712 * - source: CPU1, target: CPU2 | cost1
5713 * - source: CPU2, target: CPU1 | cost2
5714 * - source: CPU1, target: CPU1 | cost3
5715 * - source: CPU2, target: CPU2 | cost4
5717 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5718 * the cost of migration.
5720 * We then start off from a small buffer-size and iterate up to larger
5721 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5722 * doing a maximum search for the cost. (The maximum cost for a migration
5723 * normally occurs when the working set size is around the effective cache
5726 #define SEARCH_SCOPE 2
5727 #define MIN_CACHE_SIZE (64*1024U)
5728 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5729 #define ITERATIONS 1
5730 #define SIZE_THRESH 130
5731 #define COST_THRESH 130
5734 * The migration cost is a function of 'domain distance'. Domain
5735 * distance is the number of steps a CPU has to iterate down its
5736 * domain tree to share a domain with the other CPU. The farther
5737 * two CPUs are from each other, the larger the distance gets.
5739 * Note that we use the distance only to cache measurement results,
5740 * the distance value is not used numerically otherwise. When two
5741 * CPUs have the same distance it is assumed that the migration
5742 * cost is the same. (this is a simplification but quite practical)
5744 #define MAX_DOMAIN_DISTANCE 32
5746 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5747 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5749 * Architectures may override the migration cost and thus avoid
5750 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5751 * virtualized hardware:
5753 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5754 CONFIG_DEFAULT_MIGRATION_COST
5761 * Allow override of migration cost - in units of microseconds.
5762 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5763 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5765 static int __init migration_cost_setup(char *str)
5767 int ints[MAX_DOMAIN_DISTANCE+1], i;
5769 str = get_options(str, ARRAY_SIZE(ints), ints);
5771 printk("#ints: %d\n", ints[0]);
5772 for (i = 1; i <= ints[0]; i++) {
5773 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5774 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5779 __setup ("migration_cost=", migration_cost_setup);
5782 * Global multiplier (divisor) for migration-cutoff values,
5783 * in percentiles. E.g. use a value of 150 to get 1.5 times
5784 * longer cache-hot cutoff times.
5786 * (We scale it from 100 to 128 to long long handling easier.)
5789 #define MIGRATION_FACTOR_SCALE 128
5791 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5793 static int __init setup_migration_factor(char *str)
5795 get_option(&str, &migration_factor);
5796 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5800 __setup("migration_factor=", setup_migration_factor);
5803 * Estimated distance of two CPUs, measured via the number of domains
5804 * we have to pass for the two CPUs to be in the same span:
5806 static unsigned long domain_distance(int cpu1, int cpu2)
5808 unsigned long distance = 0;
5809 struct sched_domain *sd;
5811 for_each_domain(cpu1, sd) {
5812 WARN_ON(!cpu_isset(cpu1, sd->span));
5813 if (cpu_isset(cpu2, sd->span))
5817 if (distance >= MAX_DOMAIN_DISTANCE) {
5819 distance = MAX_DOMAIN_DISTANCE-1;
5825 static unsigned int migration_debug;
5827 static int __init setup_migration_debug(char *str)
5829 get_option(&str, &migration_debug);
5833 __setup("migration_debug=", setup_migration_debug);
5836 * Maximum cache-size that the scheduler should try to measure.
5837 * Architectures with larger caches should tune this up during
5838 * bootup. Gets used in the domain-setup code (i.e. during SMP
5841 unsigned int max_cache_size;
5843 static int __init setup_max_cache_size(char *str)
5845 get_option(&str, &max_cache_size);
5849 __setup("max_cache_size=", setup_max_cache_size);
5852 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5853 * is the operation that is timed, so we try to generate unpredictable
5854 * cachemisses that still end up filling the L2 cache:
5856 static void touch_cache(void *__cache, unsigned long __size)
5858 unsigned long size = __size / sizeof(long);
5859 unsigned long chunk1 = size / 3;
5860 unsigned long chunk2 = 2 * size / 3;
5861 unsigned long *cache = __cache;
5864 for (i = 0; i < size/6; i += 8) {
5867 case 1: cache[size-1-i]++;
5868 case 2: cache[chunk1-i]++;
5869 case 3: cache[chunk1+i]++;
5870 case 4: cache[chunk2-i]++;
5871 case 5: cache[chunk2+i]++;
5877 * Measure the cache-cost of one task migration. Returns in units of nsec.
5879 static unsigned long long
5880 measure_one(void *cache, unsigned long size, int source, int target)
5882 cpumask_t mask, saved_mask;
5883 unsigned long long t0, t1, t2, t3, cost;
5885 saved_mask = current->cpus_allowed;
5888 * Flush source caches to RAM and invalidate them:
5893 * Migrate to the source CPU:
5895 mask = cpumask_of_cpu(source);
5896 set_cpus_allowed(current, mask);
5897 WARN_ON(smp_processor_id() != source);
5900 * Dirty the working set:
5903 touch_cache(cache, size);
5907 * Migrate to the target CPU, dirty the L2 cache and access
5908 * the shared buffer. (which represents the working set
5909 * of a migrated task.)
5911 mask = cpumask_of_cpu(target);
5912 set_cpus_allowed(current, mask);
5913 WARN_ON(smp_processor_id() != target);
5916 touch_cache(cache, size);
5919 cost = t1-t0 + t3-t2;
5921 if (migration_debug >= 2)
5922 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5923 source, target, t1-t0, t1-t0, t3-t2, cost);
5925 * Flush target caches to RAM and invalidate them:
5929 set_cpus_allowed(current, saved_mask);
5935 * Measure a series of task migrations and return the average
5936 * result. Since this code runs early during bootup the system
5937 * is 'undisturbed' and the average latency makes sense.
5939 * The algorithm in essence auto-detects the relevant cache-size,
5940 * so it will properly detect different cachesizes for different
5941 * cache-hierarchies, depending on how the CPUs are connected.
5943 * Architectures can prime the upper limit of the search range via
5944 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5946 static unsigned long long
5947 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5949 unsigned long long cost1, cost2;
5953 * Measure the migration cost of 'size' bytes, over an
5954 * average of 10 runs:
5956 * (We perturb the cache size by a small (0..4k)
5957 * value to compensate size/alignment related artifacts.
5958 * We also subtract the cost of the operation done on
5964 * dry run, to make sure we start off cache-cold on cpu1,
5965 * and to get any vmalloc pagefaults in advance:
5967 measure_one(cache, size, cpu1, cpu2);
5968 for (i = 0; i < ITERATIONS; i++)
5969 cost1 += measure_one(cache, size - i * 1024, cpu1, cpu2);
5971 measure_one(cache, size, cpu2, cpu1);
5972 for (i = 0; i < ITERATIONS; i++)
5973 cost1 += measure_one(cache, size - i * 1024, cpu2, cpu1);
5976 * (We measure the non-migrating [cached] cost on both
5977 * cpu1 and cpu2, to handle CPUs with different speeds)
5981 measure_one(cache, size, cpu1, cpu1);
5982 for (i = 0; i < ITERATIONS; i++)
5983 cost2 += measure_one(cache, size - i * 1024, cpu1, cpu1);
5985 measure_one(cache, size, cpu2, cpu2);
5986 for (i = 0; i < ITERATIONS; i++)
5987 cost2 += measure_one(cache, size - i * 1024, cpu2, cpu2);
5990 * Get the per-iteration migration cost:
5992 do_div(cost1, 2 * ITERATIONS);
5993 do_div(cost2, 2 * ITERATIONS);
5995 return cost1 - cost2;
5998 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
6000 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
6001 unsigned int max_size, size, size_found = 0;
6002 long long cost = 0, prev_cost;
6006 * Search from max_cache_size*5 down to 64K - the real relevant
6007 * cachesize has to lie somewhere inbetween.
6009 if (max_cache_size) {
6010 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
6011 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
6014 * Since we have no estimation about the relevant
6017 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
6018 size = MIN_CACHE_SIZE;
6021 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
6022 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
6027 * Allocate the working set:
6029 cache = vmalloc(max_size);
6031 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size);
6032 return 1000000; /* return 1 msec on very small boxen */
6035 while (size <= max_size) {
6037 cost = measure_cost(cpu1, cpu2, cache, size);
6043 if (max_cost < cost) {
6049 * Calculate average fluctuation, we use this to prevent
6050 * noise from triggering an early break out of the loop:
6052 fluct = abs(cost - prev_cost);
6053 avg_fluct = (avg_fluct + fluct)/2;
6055 if (migration_debug)
6056 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6059 (long)cost / 1000000,
6060 ((long)cost / 100000) % 10,
6061 (long)max_cost / 1000000,
6062 ((long)max_cost / 100000) % 10,
6063 domain_distance(cpu1, cpu2),
6067 * If we iterated at least 20% past the previous maximum,
6068 * and the cost has dropped by more than 20% already,
6069 * (taking fluctuations into account) then we assume to
6070 * have found the maximum and break out of the loop early:
6072 if (size_found && (size*100 > size_found*SIZE_THRESH))
6073 if (cost+avg_fluct <= 0 ||
6074 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
6076 if (migration_debug)
6077 printk("-> found max.\n");
6081 * Increase the cachesize in 10% steps:
6083 size = size * 10 / 9;
6086 if (migration_debug)
6087 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6088 cpu1, cpu2, size_found, max_cost);
6093 * A task is considered 'cache cold' if at least 2 times
6094 * the worst-case cost of migration has passed.
6096 * (this limit is only listened to if the load-balancing
6097 * situation is 'nice' - if there is a large imbalance we
6098 * ignore it for the sake of CPU utilization and
6099 * processing fairness.)
6101 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6104 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6106 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6107 unsigned long j0, j1, distance, max_distance = 0;
6108 struct sched_domain *sd;
6113 * First pass - calculate the cacheflush times:
6115 for_each_cpu_mask(cpu1, *cpu_map) {
6116 for_each_cpu_mask(cpu2, *cpu_map) {
6119 distance = domain_distance(cpu1, cpu2);
6120 max_distance = max(max_distance, distance);
6122 * No result cached yet?
6124 if (migration_cost[distance] == -1LL)
6125 migration_cost[distance] =
6126 measure_migration_cost(cpu1, cpu2);
6130 * Second pass - update the sched domain hierarchy with
6131 * the new cache-hot-time estimations:
6133 for_each_cpu_mask(cpu, *cpu_map) {
6135 for_each_domain(cpu, sd) {
6136 sd->cache_hot_time = migration_cost[distance];
6143 if (migration_debug)
6144 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6152 if (system_state == SYSTEM_BOOTING && num_online_cpus() > 1) {
6153 printk("migration_cost=");
6154 for (distance = 0; distance <= max_distance; distance++) {
6157 printk("%ld", (long)migration_cost[distance] / 1000);
6162 if (migration_debug)
6163 printk("migration: %ld seconds\n", (j1-j0) / HZ);
6166 * Move back to the original CPU. NUMA-Q gets confused
6167 * if we migrate to another quad during bootup.
6169 if (raw_smp_processor_id() != orig_cpu) {
6170 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6171 saved_mask = current->cpus_allowed;
6173 set_cpus_allowed(current, mask);
6174 set_cpus_allowed(current, saved_mask);
6181 * find_next_best_node - find the next node to include in a sched_domain
6182 * @node: node whose sched_domain we're building
6183 * @used_nodes: nodes already in the sched_domain
6185 * Find the next node to include in a given scheduling domain. Simply
6186 * finds the closest node not already in the @used_nodes map.
6188 * Should use nodemask_t.
6190 static int find_next_best_node(int node, unsigned long *used_nodes)
6192 int i, n, val, min_val, best_node = 0;
6196 for (i = 0; i < MAX_NUMNODES; i++) {
6197 /* Start at @node */
6198 n = (node + i) % MAX_NUMNODES;
6200 if (!nr_cpus_node(n))
6203 /* Skip already used nodes */
6204 if (test_bit(n, used_nodes))
6207 /* Simple min distance search */
6208 val = node_distance(node, n);
6210 if (val < min_val) {
6216 set_bit(best_node, used_nodes);
6221 * sched_domain_node_span - get a cpumask for a node's sched_domain
6222 * @node: node whose cpumask we're constructing
6223 * @size: number of nodes to include in this span
6225 * Given a node, construct a good cpumask for its sched_domain to span. It
6226 * should be one that prevents unnecessary balancing, but also spreads tasks
6229 static cpumask_t sched_domain_node_span(int node)
6231 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6232 cpumask_t span, nodemask;
6236 bitmap_zero(used_nodes, MAX_NUMNODES);
6238 nodemask = node_to_cpumask(node);
6239 cpus_or(span, span, nodemask);
6240 set_bit(node, used_nodes);
6242 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6243 int next_node = find_next_best_node(node, used_nodes);
6245 nodemask = node_to_cpumask(next_node);
6246 cpus_or(span, span, nodemask);
6253 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6256 * SMT sched-domains:
6258 #ifdef CONFIG_SCHED_SMT
6259 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6260 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6262 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6263 struct sched_group **sg)
6266 *sg = &per_cpu(sched_group_cpus, cpu);
6272 * multi-core sched-domains:
6274 #ifdef CONFIG_SCHED_MC
6275 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6276 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6279 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6280 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6281 struct sched_group **sg)
6284 cpumask_t mask = cpu_sibling_map[cpu];
6285 cpus_and(mask, mask, *cpu_map);
6286 group = first_cpu(mask);
6288 *sg = &per_cpu(sched_group_core, group);
6291 #elif defined(CONFIG_SCHED_MC)
6292 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6293 struct sched_group **sg)
6296 *sg = &per_cpu(sched_group_core, cpu);
6301 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6302 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6304 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6305 struct sched_group **sg)
6308 #ifdef CONFIG_SCHED_MC
6309 cpumask_t mask = cpu_coregroup_map(cpu);
6310 cpus_and(mask, mask, *cpu_map);
6311 group = first_cpu(mask);
6312 #elif defined(CONFIG_SCHED_SMT)
6313 cpumask_t mask = cpu_sibling_map[cpu];
6314 cpus_and(mask, mask, *cpu_map);
6315 group = first_cpu(mask);
6320 *sg = &per_cpu(sched_group_phys, group);
6326 * The init_sched_build_groups can't handle what we want to do with node
6327 * groups, so roll our own. Now each node has its own list of groups which
6328 * gets dynamically allocated.
6330 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6331 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6333 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6334 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6336 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6337 struct sched_group **sg)
6339 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6342 cpus_and(nodemask, nodemask, *cpu_map);
6343 group = first_cpu(nodemask);
6346 *sg = &per_cpu(sched_group_allnodes, group);
6350 static void init_numa_sched_groups_power(struct sched_group *group_head)
6352 struct sched_group *sg = group_head;
6358 for_each_cpu_mask(j, sg->cpumask) {
6359 struct sched_domain *sd;
6361 sd = &per_cpu(phys_domains, j);
6362 if (j != first_cpu(sd->groups->cpumask)) {
6364 * Only add "power" once for each
6370 sg->cpu_power += sd->groups->cpu_power;
6373 if (sg != group_head)
6379 /* Free memory allocated for various sched_group structures */
6380 static void free_sched_groups(const cpumask_t *cpu_map)
6384 for_each_cpu_mask(cpu, *cpu_map) {
6385 struct sched_group **sched_group_nodes
6386 = sched_group_nodes_bycpu[cpu];
6388 if (!sched_group_nodes)
6391 for (i = 0; i < MAX_NUMNODES; i++) {
6392 cpumask_t nodemask = node_to_cpumask(i);
6393 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6395 cpus_and(nodemask, nodemask, *cpu_map);
6396 if (cpus_empty(nodemask))
6406 if (oldsg != sched_group_nodes[i])
6409 kfree(sched_group_nodes);
6410 sched_group_nodes_bycpu[cpu] = NULL;
6414 static void free_sched_groups(const cpumask_t *cpu_map)
6420 * Initialize sched groups cpu_power.
6422 * cpu_power indicates the capacity of sched group, which is used while
6423 * distributing the load between different sched groups in a sched domain.
6424 * Typically cpu_power for all the groups in a sched domain will be same unless
6425 * there are asymmetries in the topology. If there are asymmetries, group
6426 * having more cpu_power will pickup more load compared to the group having
6429 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6430 * the maximum number of tasks a group can handle in the presence of other idle
6431 * or lightly loaded groups in the same sched domain.
6433 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6435 struct sched_domain *child;
6436 struct sched_group *group;
6438 WARN_ON(!sd || !sd->groups);
6440 if (cpu != first_cpu(sd->groups->cpumask))
6446 * For perf policy, if the groups in child domain share resources
6447 * (for example cores sharing some portions of the cache hierarchy
6448 * or SMT), then set this domain groups cpu_power such that each group
6449 * can handle only one task, when there are other idle groups in the
6450 * same sched domain.
6452 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6454 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6455 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6459 sd->groups->cpu_power = 0;
6462 * add cpu_power of each child group to this groups cpu_power
6464 group = child->groups;
6466 sd->groups->cpu_power += group->cpu_power;
6467 group = group->next;
6468 } while (group != child->groups);
6472 * Build sched domains for a given set of cpus and attach the sched domains
6473 * to the individual cpus
6475 static int build_sched_domains(const cpumask_t *cpu_map)
6478 struct sched_domain *sd;
6480 struct sched_group **sched_group_nodes = NULL;
6481 int sd_allnodes = 0;
6484 * Allocate the per-node list of sched groups
6486 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6488 if (!sched_group_nodes) {
6489 printk(KERN_WARNING "Can not alloc sched group node list\n");
6492 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6496 * Set up domains for cpus specified by the cpu_map.
6498 for_each_cpu_mask(i, *cpu_map) {
6499 struct sched_domain *sd = NULL, *p;
6500 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6502 cpus_and(nodemask, nodemask, *cpu_map);
6505 if (cpus_weight(*cpu_map)
6506 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6507 sd = &per_cpu(allnodes_domains, i);
6508 *sd = SD_ALLNODES_INIT;
6509 sd->span = *cpu_map;
6510 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6516 sd = &per_cpu(node_domains, i);
6518 sd->span = sched_domain_node_span(cpu_to_node(i));
6522 cpus_and(sd->span, sd->span, *cpu_map);
6526 sd = &per_cpu(phys_domains, i);
6528 sd->span = nodemask;
6532 cpu_to_phys_group(i, cpu_map, &sd->groups);
6534 #ifdef CONFIG_SCHED_MC
6536 sd = &per_cpu(core_domains, i);
6538 sd->span = cpu_coregroup_map(i);
6539 cpus_and(sd->span, sd->span, *cpu_map);
6542 cpu_to_core_group(i, cpu_map, &sd->groups);
6545 #ifdef CONFIG_SCHED_SMT
6547 sd = &per_cpu(cpu_domains, i);
6548 *sd = SD_SIBLING_INIT;
6549 sd->span = cpu_sibling_map[i];
6550 cpus_and(sd->span, sd->span, *cpu_map);
6553 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6557 #ifdef CONFIG_SCHED_SMT
6558 /* Set up CPU (sibling) groups */
6559 for_each_cpu_mask(i, *cpu_map) {
6560 cpumask_t this_sibling_map = cpu_sibling_map[i];
6561 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6562 if (i != first_cpu(this_sibling_map))
6565 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6569 #ifdef CONFIG_SCHED_MC
6570 /* Set up multi-core groups */
6571 for_each_cpu_mask(i, *cpu_map) {
6572 cpumask_t this_core_map = cpu_coregroup_map(i);
6573 cpus_and(this_core_map, this_core_map, *cpu_map);
6574 if (i != first_cpu(this_core_map))
6576 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6581 /* Set up physical groups */
6582 for (i = 0; i < MAX_NUMNODES; i++) {
6583 cpumask_t nodemask = node_to_cpumask(i);
6585 cpus_and(nodemask, nodemask, *cpu_map);
6586 if (cpus_empty(nodemask))
6589 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6593 /* Set up node groups */
6595 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6597 for (i = 0; i < MAX_NUMNODES; i++) {
6598 /* Set up node groups */
6599 struct sched_group *sg, *prev;
6600 cpumask_t nodemask = node_to_cpumask(i);
6601 cpumask_t domainspan;
6602 cpumask_t covered = CPU_MASK_NONE;
6605 cpus_and(nodemask, nodemask, *cpu_map);
6606 if (cpus_empty(nodemask)) {
6607 sched_group_nodes[i] = NULL;
6611 domainspan = sched_domain_node_span(i);
6612 cpus_and(domainspan, domainspan, *cpu_map);
6614 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6616 printk(KERN_WARNING "Can not alloc domain group for "
6620 sched_group_nodes[i] = sg;
6621 for_each_cpu_mask(j, nodemask) {
6622 struct sched_domain *sd;
6623 sd = &per_cpu(node_domains, j);
6627 sg->cpumask = nodemask;
6629 cpus_or(covered, covered, nodemask);
6632 for (j = 0; j < MAX_NUMNODES; j++) {
6633 cpumask_t tmp, notcovered;
6634 int n = (i + j) % MAX_NUMNODES;
6636 cpus_complement(notcovered, covered);
6637 cpus_and(tmp, notcovered, *cpu_map);
6638 cpus_and(tmp, tmp, domainspan);
6639 if (cpus_empty(tmp))
6642 nodemask = node_to_cpumask(n);
6643 cpus_and(tmp, tmp, nodemask);
6644 if (cpus_empty(tmp))
6647 sg = kmalloc_node(sizeof(struct sched_group),
6651 "Can not alloc domain group for node %d\n", j);
6656 sg->next = prev->next;
6657 cpus_or(covered, covered, tmp);
6664 /* Calculate CPU power for physical packages and nodes */
6665 #ifdef CONFIG_SCHED_SMT
6666 for_each_cpu_mask(i, *cpu_map) {
6667 sd = &per_cpu(cpu_domains, i);
6668 init_sched_groups_power(i, sd);
6671 #ifdef CONFIG_SCHED_MC
6672 for_each_cpu_mask(i, *cpu_map) {
6673 sd = &per_cpu(core_domains, i);
6674 init_sched_groups_power(i, sd);
6678 for_each_cpu_mask(i, *cpu_map) {
6679 sd = &per_cpu(phys_domains, i);
6680 init_sched_groups_power(i, sd);
6684 for (i = 0; i < MAX_NUMNODES; i++)
6685 init_numa_sched_groups_power(sched_group_nodes[i]);
6688 struct sched_group *sg;
6690 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6691 init_numa_sched_groups_power(sg);
6695 /* Attach the domains */
6696 for_each_cpu_mask(i, *cpu_map) {
6697 struct sched_domain *sd;
6698 #ifdef CONFIG_SCHED_SMT
6699 sd = &per_cpu(cpu_domains, i);
6700 #elif defined(CONFIG_SCHED_MC)
6701 sd = &per_cpu(core_domains, i);
6703 sd = &per_cpu(phys_domains, i);
6705 cpu_attach_domain(sd, i);
6708 * Tune cache-hot values:
6710 calibrate_migration_costs(cpu_map);
6716 free_sched_groups(cpu_map);
6721 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6723 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6725 cpumask_t cpu_default_map;
6729 * Setup mask for cpus without special case scheduling requirements.
6730 * For now this just excludes isolated cpus, but could be used to
6731 * exclude other special cases in the future.
6733 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6735 err = build_sched_domains(&cpu_default_map);
6740 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6742 free_sched_groups(cpu_map);
6746 * Detach sched domains from a group of cpus specified in cpu_map
6747 * These cpus will now be attached to the NULL domain
6749 static void detach_destroy_domains(const cpumask_t *cpu_map)
6753 for_each_cpu_mask(i, *cpu_map)
6754 cpu_attach_domain(NULL, i);
6755 synchronize_sched();
6756 arch_destroy_sched_domains(cpu_map);
6760 * Partition sched domains as specified by the cpumasks below.
6761 * This attaches all cpus from the cpumasks to the NULL domain,
6762 * waits for a RCU quiescent period, recalculates sched
6763 * domain information and then attaches them back to the
6764 * correct sched domains
6765 * Call with hotplug lock held
6767 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6769 cpumask_t change_map;
6772 cpus_and(*partition1, *partition1, cpu_online_map);
6773 cpus_and(*partition2, *partition2, cpu_online_map);
6774 cpus_or(change_map, *partition1, *partition2);
6776 /* Detach sched domains from all of the affected cpus */
6777 detach_destroy_domains(&change_map);
6778 if (!cpus_empty(*partition1))
6779 err = build_sched_domains(partition1);
6780 if (!err && !cpus_empty(*partition2))
6781 err = build_sched_domains(partition2);
6786 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6787 int arch_reinit_sched_domains(void)
6792 detach_destroy_domains(&cpu_online_map);
6793 err = arch_init_sched_domains(&cpu_online_map);
6794 unlock_cpu_hotplug();
6799 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6803 if (buf[0] != '0' && buf[0] != '1')
6807 sched_smt_power_savings = (buf[0] == '1');
6809 sched_mc_power_savings = (buf[0] == '1');
6811 ret = arch_reinit_sched_domains();
6813 return ret ? ret : count;
6816 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6820 #ifdef CONFIG_SCHED_SMT
6822 err = sysfs_create_file(&cls->kset.kobj,
6823 &attr_sched_smt_power_savings.attr);
6825 #ifdef CONFIG_SCHED_MC
6826 if (!err && mc_capable())
6827 err = sysfs_create_file(&cls->kset.kobj,
6828 &attr_sched_mc_power_savings.attr);
6834 #ifdef CONFIG_SCHED_MC
6835 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6837 return sprintf(page, "%u\n", sched_mc_power_savings);
6839 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6840 const char *buf, size_t count)
6842 return sched_power_savings_store(buf, count, 0);
6844 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6845 sched_mc_power_savings_store);
6848 #ifdef CONFIG_SCHED_SMT
6849 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6851 return sprintf(page, "%u\n", sched_smt_power_savings);
6853 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6854 const char *buf, size_t count)
6856 return sched_power_savings_store(buf, count, 1);
6858 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6859 sched_smt_power_savings_store);
6863 * Force a reinitialization of the sched domains hierarchy. The domains
6864 * and groups cannot be updated in place without racing with the balancing
6865 * code, so we temporarily attach all running cpus to the NULL domain
6866 * which will prevent rebalancing while the sched domains are recalculated.
6868 static int update_sched_domains(struct notifier_block *nfb,
6869 unsigned long action, void *hcpu)
6872 case CPU_UP_PREPARE:
6873 case CPU_DOWN_PREPARE:
6874 detach_destroy_domains(&cpu_online_map);
6877 case CPU_UP_CANCELED:
6878 case CPU_DOWN_FAILED:
6882 * Fall through and re-initialise the domains.
6889 /* The hotplug lock is already held by cpu_up/cpu_down */
6890 arch_init_sched_domains(&cpu_online_map);
6895 void __init sched_init_smp(void)
6897 cpumask_t non_isolated_cpus;
6900 arch_init_sched_domains(&cpu_online_map);
6901 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6902 if (cpus_empty(non_isolated_cpus))
6903 cpu_set(smp_processor_id(), non_isolated_cpus);
6904 unlock_cpu_hotplug();
6905 /* XXX: Theoretical race here - CPU may be hotplugged now */
6906 hotcpu_notifier(update_sched_domains, 0);
6908 /* Move init over to a non-isolated CPU */
6909 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6913 void __init sched_init_smp(void)
6916 #endif /* CONFIG_SMP */
6918 int in_sched_functions(unsigned long addr)
6920 /* Linker adds these: start and end of __sched functions */
6921 extern char __sched_text_start[], __sched_text_end[];
6923 return in_lock_functions(addr) ||
6924 (addr >= (unsigned long)__sched_text_start
6925 && addr < (unsigned long)__sched_text_end);
6928 void __init sched_init(void)
6931 int highest_cpu = 0;
6933 for_each_possible_cpu(i) {
6934 struct prio_array *array;
6938 spin_lock_init(&rq->lock);
6939 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6941 rq->active = rq->arrays;
6942 rq->expired = rq->arrays + 1;
6943 rq->best_expired_prio = MAX_PRIO;
6947 for (j = 1; j < 3; j++)
6948 rq->cpu_load[j] = 0;
6949 rq->active_balance = 0;
6952 rq->migration_thread = NULL;
6953 INIT_LIST_HEAD(&rq->migration_queue);
6955 atomic_set(&rq->nr_iowait, 0);
6957 for (j = 0; j < 2; j++) {
6958 array = rq->arrays + j;
6959 for (k = 0; k < MAX_PRIO; k++) {
6960 INIT_LIST_HEAD(array->queue + k);
6961 __clear_bit(k, array->bitmap);
6963 // delimiter for bitsearch
6964 __set_bit(MAX_PRIO, array->bitmap);
6969 set_load_weight(&init_task);
6972 nr_cpu_ids = highest_cpu + 1;
6973 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6976 #ifdef CONFIG_RT_MUTEXES
6977 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6981 * The boot idle thread does lazy MMU switching as well:
6983 atomic_inc(&init_mm.mm_count);
6984 enter_lazy_tlb(&init_mm, current);
6987 * Make us the idle thread. Technically, schedule() should not be
6988 * called from this thread, however somewhere below it might be,
6989 * but because we are the idle thread, we just pick up running again
6990 * when this runqueue becomes "idle".
6992 init_idle(current, smp_processor_id());
6995 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6996 void __might_sleep(char *file, int line)
6999 static unsigned long prev_jiffy; /* ratelimiting */
7001 if ((in_atomic() || irqs_disabled()) &&
7002 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7003 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7005 prev_jiffy = jiffies;
7006 printk(KERN_ERR "BUG: sleeping function called from invalid"
7007 " context at %s:%d\n", file, line);
7008 printk("in_atomic():%d, irqs_disabled():%d\n",
7009 in_atomic(), irqs_disabled());
7010 debug_show_held_locks(current);
7011 if (irqs_disabled())
7012 print_irqtrace_events(current);
7017 EXPORT_SYMBOL(__might_sleep);
7020 #ifdef CONFIG_MAGIC_SYSRQ
7021 void normalize_rt_tasks(void)
7023 struct prio_array *array;
7024 struct task_struct *p;
7025 unsigned long flags;
7028 read_lock_irq(&tasklist_lock);
7029 for_each_process(p) {
7033 spin_lock_irqsave(&p->pi_lock, flags);
7034 rq = __task_rq_lock(p);
7038 deactivate_task(p, task_rq(p));
7039 __setscheduler(p, SCHED_NORMAL, 0);
7041 __activate_task(p, task_rq(p));
7042 resched_task(rq->curr);
7045 __task_rq_unlock(rq);
7046 spin_unlock_irqrestore(&p->pi_lock, flags);
7048 read_unlock_irq(&tasklist_lock);
7051 #endif /* CONFIG_MAGIC_SYSRQ */
7055 * These functions are only useful for the IA64 MCA handling.
7057 * They can only be called when the whole system has been
7058 * stopped - every CPU needs to be quiescent, and no scheduling
7059 * activity can take place. Using them for anything else would
7060 * be a serious bug, and as a result, they aren't even visible
7061 * under any other configuration.
7065 * curr_task - return the current task for a given cpu.
7066 * @cpu: the processor in question.
7068 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7070 struct task_struct *curr_task(int cpu)
7072 return cpu_curr(cpu);
7076 * set_curr_task - set the current task for a given cpu.
7077 * @cpu: the processor in question.
7078 * @p: the task pointer to set.
7080 * Description: This function must only be used when non-maskable interrupts
7081 * are serviced on a separate stack. It allows the architecture to switch the
7082 * notion of the current task on a cpu in a non-blocking manner. This function
7083 * must be called with all CPU's synchronized, and interrupts disabled, the
7084 * and caller must save the original value of the current task (see
7085 * curr_task() above) and restore that value before reenabling interrupts and
7086 * re-starting the system.
7088 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7090 void set_curr_task(int cpu, struct task_struct *p)