4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
57 #include <asm/unistd.h>
60 * Convert user-nice values [ -20 ... 0 ... 19 ]
61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
64 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
65 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
66 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
69 * 'User priority' is the nice value converted to something we
70 * can work with better when scaling various scheduler parameters,
71 * it's a [ 0 ... 39 ] range.
73 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
74 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
75 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
78 * Some helpers for converting nanosecond timing to jiffy resolution
80 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
81 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
84 * These are the 'tuning knobs' of the scheduler:
86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
88 * Timeslices get refilled after they expire.
90 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
91 #define DEF_TIMESLICE (100 * HZ / 1000)
92 #define ON_RUNQUEUE_WEIGHT 30
93 #define CHILD_PENALTY 95
94 #define PARENT_PENALTY 100
96 #define PRIO_BONUS_RATIO 25
97 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
98 #define INTERACTIVE_DELTA 2
99 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
100 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
101 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
104 * If a task is 'interactive' then we reinsert it in the active
105 * array after it has expired its current timeslice. (it will not
106 * continue to run immediately, it will still roundrobin with
107 * other interactive tasks.)
109 * This part scales the interactivity limit depending on niceness.
111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
112 * Here are a few examples of different nice levels:
114 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
115 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
116 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
118 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
121 * priority range a task can explore, a value of '1' means the
122 * task is rated interactive.)
124 * Ie. nice +19 tasks can never get 'interactive' enough to be
125 * reinserted into the active array. And only heavily CPU-hog nice -20
126 * tasks will be expired. Default nice 0 tasks are somewhere between,
127 * it takes some effort for them to get interactive, but it's not
131 #define CURRENT_BONUS(p) \
132 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
135 #define GRANULARITY (10 * HZ / 1000 ? : 1)
138 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
139 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
142 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
143 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
146 #define SCALE(v1,v1_max,v2_max) \
147 (v1) * (v2_max) / (v1_max)
150 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
153 #define TASK_INTERACTIVE(p) \
154 ((p)->prio <= (p)->static_prio - DELTA(p))
156 #define INTERACTIVE_SLEEP(p) \
157 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
158 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
160 #define TASK_PREEMPTS_CURR(p, rq) \
161 ((p)->prio < (rq)->curr->prio)
163 #define SCALE_PRIO(x, prio) \
164 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
166 static unsigned int static_prio_timeslice(int static_prio)
168 if (static_prio < NICE_TO_PRIO(0))
169 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
171 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
175 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
176 * to time slice values: [800ms ... 100ms ... 5ms]
178 * The higher a thread's priority, the bigger timeslices
179 * it gets during one round of execution. But even the lowest
180 * priority thread gets MIN_TIMESLICE worth of execution time.
183 static inline unsigned int task_timeslice(struct task_struct *p)
185 return static_prio_timeslice(p->static_prio);
189 * These are the runqueue data structures:
193 unsigned int nr_active;
194 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
195 struct list_head queue[MAX_PRIO];
199 * This is the main, per-CPU runqueue data structure.
201 * Locking rule: those places that want to lock multiple runqueues
202 * (such as the load balancing or the thread migration code), lock
203 * acquire operations must be ordered by ascending &runqueue.
209 * nr_running and cpu_load should be in the same cacheline because
210 * remote CPUs use both these fields when doing load calculation.
212 unsigned long nr_running;
213 unsigned long raw_weighted_load;
215 unsigned long cpu_load[3];
217 unsigned long long nr_switches;
220 * This is part of a global counter where only the total sum
221 * over all CPUs matters. A task can increase this counter on
222 * one CPU and if it got migrated afterwards it may decrease
223 * it on another CPU. Always updated under the runqueue lock:
225 unsigned long nr_uninterruptible;
227 unsigned long expired_timestamp;
228 unsigned long long timestamp_last_tick;
229 struct task_struct *curr, *idle;
230 unsigned long next_balance;
231 struct mm_struct *prev_mm;
232 struct prio_array *active, *expired, arrays[2];
233 int best_expired_prio;
237 struct sched_domain *sd;
239 /* For active balancing */
242 int cpu; /* cpu of this runqueue */
244 struct task_struct *migration_thread;
245 struct list_head migration_queue;
248 #ifdef CONFIG_SCHEDSTATS
250 struct sched_info rq_sched_info;
252 /* sys_sched_yield() stats */
253 unsigned long yld_exp_empty;
254 unsigned long yld_act_empty;
255 unsigned long yld_both_empty;
256 unsigned long yld_cnt;
258 /* schedule() stats */
259 unsigned long sched_switch;
260 unsigned long sched_cnt;
261 unsigned long sched_goidle;
263 /* try_to_wake_up() stats */
264 unsigned long ttwu_cnt;
265 unsigned long ttwu_local;
267 struct lock_class_key rq_lock_key;
270 static DEFINE_PER_CPU(struct rq, runqueues);
272 static inline int cpu_of(struct rq *rq)
282 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
283 * See detach_destroy_domains: synchronize_sched for details.
285 * The domain tree of any CPU may only be accessed from within
286 * preempt-disabled sections.
288 #define for_each_domain(cpu, __sd) \
289 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
291 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
292 #define this_rq() (&__get_cpu_var(runqueues))
293 #define task_rq(p) cpu_rq(task_cpu(p))
294 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
296 #ifndef prepare_arch_switch
297 # define prepare_arch_switch(next) do { } while (0)
299 #ifndef finish_arch_switch
300 # define finish_arch_switch(prev) do { } while (0)
303 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
304 static inline int task_running(struct rq *rq, struct task_struct *p)
306 return rq->curr == p;
309 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
313 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
315 #ifdef CONFIG_DEBUG_SPINLOCK
316 /* this is a valid case when another task releases the spinlock */
317 rq->lock.owner = current;
320 * If we are tracking spinlock dependencies then we have to
321 * fix up the runqueue lock - which gets 'carried over' from
324 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
326 spin_unlock_irq(&rq->lock);
329 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
330 static inline int task_running(struct rq *rq, struct task_struct *p)
335 return rq->curr == p;
339 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
343 * We can optimise this out completely for !SMP, because the
344 * SMP rebalancing from interrupt is the only thing that cares
349 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
350 spin_unlock_irq(&rq->lock);
352 spin_unlock(&rq->lock);
356 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
360 * After ->oncpu is cleared, the task can be moved to a different CPU.
361 * We must ensure this doesn't happen until the switch is completely
367 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
371 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
374 * __task_rq_lock - lock the runqueue a given task resides on.
375 * Must be called interrupts disabled.
377 static inline struct rq *__task_rq_lock(struct task_struct *p)
384 spin_lock(&rq->lock);
385 if (unlikely(rq != task_rq(p))) {
386 spin_unlock(&rq->lock);
387 goto repeat_lock_task;
393 * task_rq_lock - lock the runqueue a given task resides on and disable
394 * interrupts. Note the ordering: we can safely lookup the task_rq without
395 * explicitly disabling preemption.
397 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
403 local_irq_save(*flags);
405 spin_lock(&rq->lock);
406 if (unlikely(rq != task_rq(p))) {
407 spin_unlock_irqrestore(&rq->lock, *flags);
408 goto repeat_lock_task;
413 static inline void __task_rq_unlock(struct rq *rq)
416 spin_unlock(&rq->lock);
419 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
422 spin_unlock_irqrestore(&rq->lock, *flags);
425 #ifdef CONFIG_SCHEDSTATS
427 * bump this up when changing the output format or the meaning of an existing
428 * format, so that tools can adapt (or abort)
430 #define SCHEDSTAT_VERSION 12
432 static int show_schedstat(struct seq_file *seq, void *v)
436 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
437 seq_printf(seq, "timestamp %lu\n", jiffies);
438 for_each_online_cpu(cpu) {
439 struct rq *rq = cpu_rq(cpu);
441 struct sched_domain *sd;
445 /* runqueue-specific stats */
447 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
448 cpu, rq->yld_both_empty,
449 rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
450 rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
451 rq->ttwu_cnt, rq->ttwu_local,
452 rq->rq_sched_info.cpu_time,
453 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
455 seq_printf(seq, "\n");
458 /* domain-specific stats */
460 for_each_domain(cpu, sd) {
461 enum idle_type itype;
462 char mask_str[NR_CPUS];
464 cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
465 seq_printf(seq, "domain%d %s", dcnt++, mask_str);
466 for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
468 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
470 sd->lb_balanced[itype],
471 sd->lb_failed[itype],
472 sd->lb_imbalance[itype],
473 sd->lb_gained[itype],
474 sd->lb_hot_gained[itype],
475 sd->lb_nobusyq[itype],
476 sd->lb_nobusyg[itype]);
478 seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
479 sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
480 sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
481 sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
482 sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
490 static int schedstat_open(struct inode *inode, struct file *file)
492 unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
493 char *buf = kmalloc(size, GFP_KERNEL);
499 res = single_open(file, show_schedstat, NULL);
501 m = file->private_data;
509 const struct file_operations proc_schedstat_operations = {
510 .open = schedstat_open,
513 .release = single_release,
517 * Expects runqueue lock to be held for atomicity of update
520 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
523 rq->rq_sched_info.run_delay += delta_jiffies;
524 rq->rq_sched_info.pcnt++;
529 * Expects runqueue lock to be held for atomicity of update
532 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
535 rq->rq_sched_info.cpu_time += delta_jiffies;
537 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
538 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
539 #else /* !CONFIG_SCHEDSTATS */
541 rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
544 rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
546 # define schedstat_inc(rq, field) do { } while (0)
547 # define schedstat_add(rq, field, amt) do { } while (0)
551 * this_rq_lock - lock this runqueue and disable interrupts.
553 static inline struct rq *this_rq_lock(void)
560 spin_lock(&rq->lock);
565 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
567 * Called when a process is dequeued from the active array and given
568 * the cpu. We should note that with the exception of interactive
569 * tasks, the expired queue will become the active queue after the active
570 * queue is empty, without explicitly dequeuing and requeuing tasks in the
571 * expired queue. (Interactive tasks may be requeued directly to the
572 * active queue, thus delaying tasks in the expired queue from running;
573 * see scheduler_tick()).
575 * This function is only called from sched_info_arrive(), rather than
576 * dequeue_task(). Even though a task may be queued and dequeued multiple
577 * times as it is shuffled about, we're really interested in knowing how
578 * long it was from the *first* time it was queued to the time that it
581 static inline void sched_info_dequeued(struct task_struct *t)
583 t->sched_info.last_queued = 0;
587 * Called when a task finally hits the cpu. We can now calculate how
588 * long it was waiting to run. We also note when it began so that we
589 * can keep stats on how long its timeslice is.
591 static void sched_info_arrive(struct task_struct *t)
593 unsigned long now = jiffies, delta_jiffies = 0;
595 if (t->sched_info.last_queued)
596 delta_jiffies = now - t->sched_info.last_queued;
597 sched_info_dequeued(t);
598 t->sched_info.run_delay += delta_jiffies;
599 t->sched_info.last_arrival = now;
600 t->sched_info.pcnt++;
602 rq_sched_info_arrive(task_rq(t), delta_jiffies);
606 * Called when a process is queued into either the active or expired
607 * array. The time is noted and later used to determine how long we
608 * had to wait for us to reach the cpu. Since the expired queue will
609 * become the active queue after active queue is empty, without dequeuing
610 * and requeuing any tasks, we are interested in queuing to either. It
611 * is unusual but not impossible for tasks to be dequeued and immediately
612 * requeued in the same or another array: this can happen in sched_yield(),
613 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
616 * This function is only called from enqueue_task(), but also only updates
617 * the timestamp if it is already not set. It's assumed that
618 * sched_info_dequeued() will clear that stamp when appropriate.
620 static inline void sched_info_queued(struct task_struct *t)
622 if (unlikely(sched_info_on()))
623 if (!t->sched_info.last_queued)
624 t->sched_info.last_queued = jiffies;
628 * Called when a process ceases being the active-running process, either
629 * voluntarily or involuntarily. Now we can calculate how long we ran.
631 static inline void sched_info_depart(struct task_struct *t)
633 unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
635 t->sched_info.cpu_time += delta_jiffies;
636 rq_sched_info_depart(task_rq(t), delta_jiffies);
640 * Called when tasks are switched involuntarily due, typically, to expiring
641 * their time slice. (This may also be called when switching to or from
642 * the idle task.) We are only called when prev != next.
645 __sched_info_switch(struct task_struct *prev, struct task_struct *next)
647 struct rq *rq = task_rq(prev);
650 * prev now departs the cpu. It's not interesting to record
651 * stats about how efficient we were at scheduling the idle
654 if (prev != rq->idle)
655 sched_info_depart(prev);
657 if (next != rq->idle)
658 sched_info_arrive(next);
661 sched_info_switch(struct task_struct *prev, struct task_struct *next)
663 if (unlikely(sched_info_on()))
664 __sched_info_switch(prev, next);
667 #define sched_info_queued(t) do { } while (0)
668 #define sched_info_switch(t, next) do { } while (0)
669 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
672 * Adding/removing a task to/from a priority array:
674 static void dequeue_task(struct task_struct *p, struct prio_array *array)
677 list_del(&p->run_list);
678 if (list_empty(array->queue + p->prio))
679 __clear_bit(p->prio, array->bitmap);
682 static void enqueue_task(struct task_struct *p, struct prio_array *array)
684 sched_info_queued(p);
685 list_add_tail(&p->run_list, array->queue + p->prio);
686 __set_bit(p->prio, array->bitmap);
692 * Put task to the end of the run list without the overhead of dequeue
693 * followed by enqueue.
695 static void requeue_task(struct task_struct *p, struct prio_array *array)
697 list_move_tail(&p->run_list, array->queue + p->prio);
701 enqueue_task_head(struct task_struct *p, struct prio_array *array)
703 list_add(&p->run_list, array->queue + p->prio);
704 __set_bit(p->prio, array->bitmap);
710 * __normal_prio - return the priority that is based on the static
711 * priority but is modified by bonuses/penalties.
713 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
714 * into the -5 ... 0 ... +5 bonus/penalty range.
716 * We use 25% of the full 0...39 priority range so that:
718 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
719 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
721 * Both properties are important to certain workloads.
724 static inline int __normal_prio(struct task_struct *p)
728 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
730 prio = p->static_prio - bonus;
731 if (prio < MAX_RT_PRIO)
733 if (prio > MAX_PRIO-1)
739 * To aid in avoiding the subversion of "niceness" due to uneven distribution
740 * of tasks with abnormal "nice" values across CPUs the contribution that
741 * each task makes to its run queue's load is weighted according to its
742 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
743 * scaled version of the new time slice allocation that they receive on time
748 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
749 * If static_prio_timeslice() is ever changed to break this assumption then
750 * this code will need modification
752 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
753 #define LOAD_WEIGHT(lp) \
754 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
755 #define PRIO_TO_LOAD_WEIGHT(prio) \
756 LOAD_WEIGHT(static_prio_timeslice(prio))
757 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
758 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
760 static void set_load_weight(struct task_struct *p)
762 if (has_rt_policy(p)) {
764 if (p == task_rq(p)->migration_thread)
766 * The migration thread does the actual balancing.
767 * Giving its load any weight will skew balancing
773 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
775 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
779 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
781 rq->raw_weighted_load += p->load_weight;
785 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
787 rq->raw_weighted_load -= p->load_weight;
790 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
793 inc_raw_weighted_load(rq, p);
796 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
799 dec_raw_weighted_load(rq, p);
803 * Calculate the expected normal priority: i.e. priority
804 * without taking RT-inheritance into account. Might be
805 * boosted by interactivity modifiers. Changes upon fork,
806 * setprio syscalls, and whenever the interactivity
807 * estimator recalculates.
809 static inline int normal_prio(struct task_struct *p)
813 if (has_rt_policy(p))
814 prio = MAX_RT_PRIO-1 - p->rt_priority;
816 prio = __normal_prio(p);
821 * Calculate the current priority, i.e. the priority
822 * taken into account by the scheduler. This value might
823 * be boosted by RT tasks, or might be boosted by
824 * interactivity modifiers. Will be RT if the task got
825 * RT-boosted. If not then it returns p->normal_prio.
827 static int effective_prio(struct task_struct *p)
829 p->normal_prio = normal_prio(p);
831 * If we are RT tasks or we were boosted to RT priority,
832 * keep the priority unchanged. Otherwise, update priority
833 * to the normal priority:
835 if (!rt_prio(p->prio))
836 return p->normal_prio;
841 * __activate_task - move a task to the runqueue.
843 static void __activate_task(struct task_struct *p, struct rq *rq)
845 struct prio_array *target = rq->active;
848 target = rq->expired;
849 enqueue_task(p, target);
850 inc_nr_running(p, rq);
854 * __activate_idle_task - move idle task to the _front_ of runqueue.
856 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
858 enqueue_task_head(p, rq->active);
859 inc_nr_running(p, rq);
863 * Recalculate p->normal_prio and p->prio after having slept,
864 * updating the sleep-average too:
866 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
868 /* Caller must always ensure 'now >= p->timestamp' */
869 unsigned long sleep_time = now - p->timestamp;
874 if (likely(sleep_time > 0)) {
876 * This ceiling is set to the lowest priority that would allow
877 * a task to be reinserted into the active array on timeslice
880 unsigned long ceiling = INTERACTIVE_SLEEP(p);
882 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
884 * Prevents user tasks from achieving best priority
885 * with one single large enough sleep.
887 p->sleep_avg = ceiling;
889 * Using INTERACTIVE_SLEEP() as a ceiling places a
890 * nice(0) task 1ms sleep away from promotion, and
891 * gives it 700ms to round-robin with no chance of
892 * being demoted. This is more than generous, so
893 * mark this sleep as non-interactive to prevent the
894 * on-runqueue bonus logic from intervening should
895 * this task not receive cpu immediately.
897 p->sleep_type = SLEEP_NONINTERACTIVE;
900 * Tasks waking from uninterruptible sleep are
901 * limited in their sleep_avg rise as they
902 * are likely to be waiting on I/O
904 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
905 if (p->sleep_avg >= ceiling)
907 else if (p->sleep_avg + sleep_time >=
909 p->sleep_avg = ceiling;
915 * This code gives a bonus to interactive tasks.
917 * The boost works by updating the 'average sleep time'
918 * value here, based on ->timestamp. The more time a
919 * task spends sleeping, the higher the average gets -
920 * and the higher the priority boost gets as well.
922 p->sleep_avg += sleep_time;
925 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
926 p->sleep_avg = NS_MAX_SLEEP_AVG;
929 return effective_prio(p);
933 * activate_task - move a task to the runqueue and do priority recalculation
935 * Update all the scheduling statistics stuff. (sleep average
936 * calculation, priority modifiers, etc.)
938 static void activate_task(struct task_struct *p, struct rq *rq, int local)
940 unsigned long long now;
945 /* Compensate for drifting sched_clock */
946 struct rq *this_rq = this_rq();
947 now = (now - this_rq->timestamp_last_tick)
948 + rq->timestamp_last_tick;
953 * Sleep time is in units of nanosecs, so shift by 20 to get a
954 * milliseconds-range estimation of the amount of time that the task
957 if (unlikely(prof_on == SLEEP_PROFILING)) {
958 if (p->state == TASK_UNINTERRUPTIBLE)
959 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
960 (now - p->timestamp) >> 20);
964 p->prio = recalc_task_prio(p, now);
967 * This checks to make sure it's not an uninterruptible task
968 * that is now waking up.
970 if (p->sleep_type == SLEEP_NORMAL) {
972 * Tasks which were woken up by interrupts (ie. hw events)
973 * are most likely of interactive nature. So we give them
974 * the credit of extending their sleep time to the period
975 * of time they spend on the runqueue, waiting for execution
976 * on a CPU, first time around:
979 p->sleep_type = SLEEP_INTERRUPTED;
982 * Normal first-time wakeups get a credit too for
983 * on-runqueue time, but it will be weighted down:
985 p->sleep_type = SLEEP_INTERACTIVE;
990 __activate_task(p, rq);
994 * deactivate_task - remove a task from the runqueue.
996 static void deactivate_task(struct task_struct *p, struct rq *rq)
998 dec_nr_running(p, rq);
999 dequeue_task(p, p->array);
1004 * resched_task - mark a task 'to be rescheduled now'.
1006 * On UP this means the setting of the need_resched flag, on SMP it
1007 * might also involve a cross-CPU call to trigger the scheduler on
1012 #ifndef tsk_is_polling
1013 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1016 static void resched_task(struct task_struct *p)
1020 assert_spin_locked(&task_rq(p)->lock);
1022 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1025 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1028 if (cpu == smp_processor_id())
1031 /* NEED_RESCHED must be visible before we test polling */
1033 if (!tsk_is_polling(p))
1034 smp_send_reschedule(cpu);
1037 static inline void resched_task(struct task_struct *p)
1039 assert_spin_locked(&task_rq(p)->lock);
1040 set_tsk_need_resched(p);
1045 * task_curr - is this task currently executing on a CPU?
1046 * @p: the task in question.
1048 inline int task_curr(const struct task_struct *p)
1050 return cpu_curr(task_cpu(p)) == p;
1053 /* Used instead of source_load when we know the type == 0 */
1054 unsigned long weighted_cpuload(const int cpu)
1056 return cpu_rq(cpu)->raw_weighted_load;
1060 struct migration_req {
1061 struct list_head list;
1063 struct task_struct *task;
1066 struct completion done;
1070 * The task's runqueue lock must be held.
1071 * Returns true if you have to wait for migration thread.
1074 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1076 struct rq *rq = task_rq(p);
1079 * If the task is not on a runqueue (and not running), then
1080 * it is sufficient to simply update the task's cpu field.
1082 if (!p->array && !task_running(rq, p)) {
1083 set_task_cpu(p, dest_cpu);
1087 init_completion(&req->done);
1089 req->dest_cpu = dest_cpu;
1090 list_add(&req->list, &rq->migration_queue);
1096 * wait_task_inactive - wait for a thread to unschedule.
1098 * The caller must ensure that the task *will* unschedule sometime soon,
1099 * else this function might spin for a *long* time. This function can't
1100 * be called with interrupts off, or it may introduce deadlock with
1101 * smp_call_function() if an IPI is sent by the same process we are
1102 * waiting to become inactive.
1104 void wait_task_inactive(struct task_struct *p)
1106 unsigned long flags;
1111 rq = task_rq_lock(p, &flags);
1112 /* Must be off runqueue entirely, not preempted. */
1113 if (unlikely(p->array || task_running(rq, p))) {
1114 /* If it's preempted, we yield. It could be a while. */
1115 preempted = !task_running(rq, p);
1116 task_rq_unlock(rq, &flags);
1122 task_rq_unlock(rq, &flags);
1126 * kick_process - kick a running thread to enter/exit the kernel
1127 * @p: the to-be-kicked thread
1129 * Cause a process which is running on another CPU to enter
1130 * kernel-mode, without any delay. (to get signals handled.)
1132 * NOTE: this function doesnt have to take the runqueue lock,
1133 * because all it wants to ensure is that the remote task enters
1134 * the kernel. If the IPI races and the task has been migrated
1135 * to another CPU then no harm is done and the purpose has been
1138 void kick_process(struct task_struct *p)
1144 if ((cpu != smp_processor_id()) && task_curr(p))
1145 smp_send_reschedule(cpu);
1150 * Return a low guess at the load of a migration-source cpu weighted
1151 * according to the scheduling class and "nice" value.
1153 * We want to under-estimate the load of migration sources, to
1154 * balance conservatively.
1156 static inline unsigned long source_load(int cpu, int type)
1158 struct rq *rq = cpu_rq(cpu);
1161 return rq->raw_weighted_load;
1163 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1167 * Return a high guess at the load of a migration-target cpu weighted
1168 * according to the scheduling class and "nice" value.
1170 static inline unsigned long target_load(int cpu, int type)
1172 struct rq *rq = cpu_rq(cpu);
1175 return rq->raw_weighted_load;
1177 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1181 * Return the average load per task on the cpu's run queue
1183 static inline unsigned long cpu_avg_load_per_task(int cpu)
1185 struct rq *rq = cpu_rq(cpu);
1186 unsigned long n = rq->nr_running;
1188 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1192 * find_idlest_group finds and returns the least busy CPU group within the
1195 static struct sched_group *
1196 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1198 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1199 unsigned long min_load = ULONG_MAX, this_load = 0;
1200 int load_idx = sd->forkexec_idx;
1201 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1204 unsigned long load, avg_load;
1208 /* Skip over this group if it has no CPUs allowed */
1209 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1212 local_group = cpu_isset(this_cpu, group->cpumask);
1214 /* Tally up the load of all CPUs in the group */
1217 for_each_cpu_mask(i, group->cpumask) {
1218 /* Bias balancing toward cpus of our domain */
1220 load = source_load(i, load_idx);
1222 load = target_load(i, load_idx);
1227 /* Adjust by relative CPU power of the group */
1228 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1231 this_load = avg_load;
1233 } else if (avg_load < min_load) {
1234 min_load = avg_load;
1238 group = group->next;
1239 } while (group != sd->groups);
1241 if (!idlest || 100*this_load < imbalance*min_load)
1247 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1250 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1253 unsigned long load, min_load = ULONG_MAX;
1257 /* Traverse only the allowed CPUs */
1258 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1260 for_each_cpu_mask(i, tmp) {
1261 load = weighted_cpuload(i);
1263 if (load < min_load || (load == min_load && i == this_cpu)) {
1273 * sched_balance_self: balance the current task (running on cpu) in domains
1274 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1277 * Balance, ie. select the least loaded group.
1279 * Returns the target CPU number, or the same CPU if no balancing is needed.
1281 * preempt must be disabled.
1283 static int sched_balance_self(int cpu, int flag)
1285 struct task_struct *t = current;
1286 struct sched_domain *tmp, *sd = NULL;
1288 for_each_domain(cpu, tmp) {
1290 * If power savings logic is enabled for a domain, stop there.
1292 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1294 if (tmp->flags & flag)
1300 struct sched_group *group;
1301 int new_cpu, weight;
1303 if (!(sd->flags & flag)) {
1309 group = find_idlest_group(sd, t, cpu);
1315 new_cpu = find_idlest_cpu(group, t, cpu);
1316 if (new_cpu == -1 || new_cpu == cpu) {
1317 /* Now try balancing at a lower domain level of cpu */
1322 /* Now try balancing at a lower domain level of new_cpu */
1325 weight = cpus_weight(span);
1326 for_each_domain(cpu, tmp) {
1327 if (weight <= cpus_weight(tmp->span))
1329 if (tmp->flags & flag)
1332 /* while loop will break here if sd == NULL */
1338 #endif /* CONFIG_SMP */
1341 * wake_idle() will wake a task on an idle cpu if task->cpu is
1342 * not idle and an idle cpu is available. The span of cpus to
1343 * search starts with cpus closest then further out as needed,
1344 * so we always favor a closer, idle cpu.
1346 * Returns the CPU we should wake onto.
1348 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1349 static int wake_idle(int cpu, struct task_struct *p)
1352 struct sched_domain *sd;
1358 for_each_domain(cpu, sd) {
1359 if (sd->flags & SD_WAKE_IDLE) {
1360 cpus_and(tmp, sd->span, p->cpus_allowed);
1361 for_each_cpu_mask(i, tmp) {
1372 static inline int wake_idle(int cpu, struct task_struct *p)
1379 * try_to_wake_up - wake up a thread
1380 * @p: the to-be-woken-up thread
1381 * @state: the mask of task states that can be woken
1382 * @sync: do a synchronous wakeup?
1384 * Put it on the run-queue if it's not already there. The "current"
1385 * thread is always on the run-queue (except when the actual
1386 * re-schedule is in progress), and as such you're allowed to do
1387 * the simpler "current->state = TASK_RUNNING" to mark yourself
1388 * runnable without the overhead of this.
1390 * returns failure only if the task is already active.
1392 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1394 int cpu, this_cpu, success = 0;
1395 unsigned long flags;
1399 struct sched_domain *sd, *this_sd = NULL;
1400 unsigned long load, this_load;
1404 rq = task_rq_lock(p, &flags);
1405 old_state = p->state;
1406 if (!(old_state & state))
1413 this_cpu = smp_processor_id();
1416 if (unlikely(task_running(rq, p)))
1421 schedstat_inc(rq, ttwu_cnt);
1422 if (cpu == this_cpu) {
1423 schedstat_inc(rq, ttwu_local);
1427 for_each_domain(this_cpu, sd) {
1428 if (cpu_isset(cpu, sd->span)) {
1429 schedstat_inc(sd, ttwu_wake_remote);
1435 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1439 * Check for affine wakeup and passive balancing possibilities.
1442 int idx = this_sd->wake_idx;
1443 unsigned int imbalance;
1445 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1447 load = source_load(cpu, idx);
1448 this_load = target_load(this_cpu, idx);
1450 new_cpu = this_cpu; /* Wake to this CPU if we can */
1452 if (this_sd->flags & SD_WAKE_AFFINE) {
1453 unsigned long tl = this_load;
1454 unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1457 * If sync wakeup then subtract the (maximum possible)
1458 * effect of the currently running task from the load
1459 * of the current CPU:
1462 tl -= current->load_weight;
1465 tl + target_load(cpu, idx) <= tl_per_task) ||
1466 100*(tl + p->load_weight) <= imbalance*load) {
1468 * This domain has SD_WAKE_AFFINE and
1469 * p is cache cold in this domain, and
1470 * there is no bad imbalance.
1472 schedstat_inc(this_sd, ttwu_move_affine);
1478 * Start passive balancing when half the imbalance_pct
1481 if (this_sd->flags & SD_WAKE_BALANCE) {
1482 if (imbalance*this_load <= 100*load) {
1483 schedstat_inc(this_sd, ttwu_move_balance);
1489 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1491 new_cpu = wake_idle(new_cpu, p);
1492 if (new_cpu != cpu) {
1493 set_task_cpu(p, new_cpu);
1494 task_rq_unlock(rq, &flags);
1495 /* might preempt at this point */
1496 rq = task_rq_lock(p, &flags);
1497 old_state = p->state;
1498 if (!(old_state & state))
1503 this_cpu = smp_processor_id();
1508 #endif /* CONFIG_SMP */
1509 if (old_state == TASK_UNINTERRUPTIBLE) {
1510 rq->nr_uninterruptible--;
1512 * Tasks on involuntary sleep don't earn
1513 * sleep_avg beyond just interactive state.
1515 p->sleep_type = SLEEP_NONINTERACTIVE;
1519 * Tasks that have marked their sleep as noninteractive get
1520 * woken up with their sleep average not weighted in an
1523 if (old_state & TASK_NONINTERACTIVE)
1524 p->sleep_type = SLEEP_NONINTERACTIVE;
1527 activate_task(p, rq, cpu == this_cpu);
1529 * Sync wakeups (i.e. those types of wakeups where the waker
1530 * has indicated that it will leave the CPU in short order)
1531 * don't trigger a preemption, if the woken up task will run on
1532 * this cpu. (in this case the 'I will reschedule' promise of
1533 * the waker guarantees that the freshly woken up task is going
1534 * to be considered on this CPU.)
1536 if (!sync || cpu != this_cpu) {
1537 if (TASK_PREEMPTS_CURR(p, rq))
1538 resched_task(rq->curr);
1543 p->state = TASK_RUNNING;
1545 task_rq_unlock(rq, &flags);
1550 int fastcall wake_up_process(struct task_struct *p)
1552 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1553 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1555 EXPORT_SYMBOL(wake_up_process);
1557 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1559 return try_to_wake_up(p, state, 0);
1563 * Perform scheduler related setup for a newly forked process p.
1564 * p is forked by current.
1566 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1568 int cpu = get_cpu();
1571 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1573 set_task_cpu(p, cpu);
1576 * We mark the process as running here, but have not actually
1577 * inserted it onto the runqueue yet. This guarantees that
1578 * nobody will actually run it, and a signal or other external
1579 * event cannot wake it up and insert it on the runqueue either.
1581 p->state = TASK_RUNNING;
1584 * Make sure we do not leak PI boosting priority to the child:
1586 p->prio = current->normal_prio;
1588 INIT_LIST_HEAD(&p->run_list);
1590 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1591 if (unlikely(sched_info_on()))
1592 memset(&p->sched_info, 0, sizeof(p->sched_info));
1594 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1597 #ifdef CONFIG_PREEMPT
1598 /* Want to start with kernel preemption disabled. */
1599 task_thread_info(p)->preempt_count = 1;
1602 * Share the timeslice between parent and child, thus the
1603 * total amount of pending timeslices in the system doesn't change,
1604 * resulting in more scheduling fairness.
1606 local_irq_disable();
1607 p->time_slice = (current->time_slice + 1) >> 1;
1609 * The remainder of the first timeslice might be recovered by
1610 * the parent if the child exits early enough.
1612 p->first_time_slice = 1;
1613 current->time_slice >>= 1;
1614 p->timestamp = sched_clock();
1615 if (unlikely(!current->time_slice)) {
1617 * This case is rare, it happens when the parent has only
1618 * a single jiffy left from its timeslice. Taking the
1619 * runqueue lock is not a problem.
1621 current->time_slice = 1;
1629 * wake_up_new_task - wake up a newly created task for the first time.
1631 * This function will do some initial scheduler statistics housekeeping
1632 * that must be done for every newly created context, then puts the task
1633 * on the runqueue and wakes it.
1635 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1637 struct rq *rq, *this_rq;
1638 unsigned long flags;
1641 rq = task_rq_lock(p, &flags);
1642 BUG_ON(p->state != TASK_RUNNING);
1643 this_cpu = smp_processor_id();
1647 * We decrease the sleep average of forking parents
1648 * and children as well, to keep max-interactive tasks
1649 * from forking tasks that are max-interactive. The parent
1650 * (current) is done further down, under its lock.
1652 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1653 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1655 p->prio = effective_prio(p);
1657 if (likely(cpu == this_cpu)) {
1658 if (!(clone_flags & CLONE_VM)) {
1660 * The VM isn't cloned, so we're in a good position to
1661 * do child-runs-first in anticipation of an exec. This
1662 * usually avoids a lot of COW overhead.
1664 if (unlikely(!current->array))
1665 __activate_task(p, rq);
1667 p->prio = current->prio;
1668 p->normal_prio = current->normal_prio;
1669 list_add_tail(&p->run_list, ¤t->run_list);
1670 p->array = current->array;
1671 p->array->nr_active++;
1672 inc_nr_running(p, rq);
1676 /* Run child last */
1677 __activate_task(p, rq);
1679 * We skip the following code due to cpu == this_cpu
1681 * task_rq_unlock(rq, &flags);
1682 * this_rq = task_rq_lock(current, &flags);
1686 this_rq = cpu_rq(this_cpu);
1689 * Not the local CPU - must adjust timestamp. This should
1690 * get optimised away in the !CONFIG_SMP case.
1692 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1693 + rq->timestamp_last_tick;
1694 __activate_task(p, rq);
1695 if (TASK_PREEMPTS_CURR(p, rq))
1696 resched_task(rq->curr);
1699 * Parent and child are on different CPUs, now get the
1700 * parent runqueue to update the parent's ->sleep_avg:
1702 task_rq_unlock(rq, &flags);
1703 this_rq = task_rq_lock(current, &flags);
1705 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1706 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1707 task_rq_unlock(this_rq, &flags);
1711 * Potentially available exiting-child timeslices are
1712 * retrieved here - this way the parent does not get
1713 * penalized for creating too many threads.
1715 * (this cannot be used to 'generate' timeslices
1716 * artificially, because any timeslice recovered here
1717 * was given away by the parent in the first place.)
1719 void fastcall sched_exit(struct task_struct *p)
1721 unsigned long flags;
1725 * If the child was a (relative-) CPU hog then decrease
1726 * the sleep_avg of the parent as well.
1728 rq = task_rq_lock(p->parent, &flags);
1729 if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1730 p->parent->time_slice += p->time_slice;
1731 if (unlikely(p->parent->time_slice > task_timeslice(p)))
1732 p->parent->time_slice = task_timeslice(p);
1734 if (p->sleep_avg < p->parent->sleep_avg)
1735 p->parent->sleep_avg = p->parent->sleep_avg /
1736 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1738 task_rq_unlock(rq, &flags);
1742 * prepare_task_switch - prepare to switch tasks
1743 * @rq: the runqueue preparing to switch
1744 * @next: the task we are going to switch to.
1746 * This is called with the rq lock held and interrupts off. It must
1747 * be paired with a subsequent finish_task_switch after the context
1750 * prepare_task_switch sets up locking and calls architecture specific
1753 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1755 prepare_lock_switch(rq, next);
1756 prepare_arch_switch(next);
1760 * finish_task_switch - clean up after a task-switch
1761 * @rq: runqueue associated with task-switch
1762 * @prev: the thread we just switched away from.
1764 * finish_task_switch must be called after the context switch, paired
1765 * with a prepare_task_switch call before the context switch.
1766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1767 * and do any other architecture-specific cleanup actions.
1769 * Note that we may have delayed dropping an mm in context_switch(). If
1770 * so, we finish that here outside of the runqueue lock. (Doing it
1771 * with the lock held can cause deadlocks; see schedule() for
1774 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1775 __releases(rq->lock)
1777 struct mm_struct *mm = rq->prev_mm;
1783 * A task struct has one reference for the use as "current".
1784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1785 * schedule one last time. The schedule call will never return, and
1786 * the scheduled task must drop that reference.
1787 * The test for TASK_DEAD must occur while the runqueue locks are
1788 * still held, otherwise prev could be scheduled on another cpu, die
1789 * there before we look at prev->state, and then the reference would
1791 * Manfred Spraul <manfred@colorfullife.com>
1793 prev_state = prev->state;
1794 finish_arch_switch(prev);
1795 finish_lock_switch(rq, prev);
1798 if (unlikely(prev_state == TASK_DEAD)) {
1800 * Remove function-return probe instances associated with this
1801 * task and put them back on the free list.
1803 kprobe_flush_task(prev);
1804 put_task_struct(prev);
1809 * schedule_tail - first thing a freshly forked thread must call.
1810 * @prev: the thread we just switched away from.
1812 asmlinkage void schedule_tail(struct task_struct *prev)
1813 __releases(rq->lock)
1815 struct rq *rq = this_rq();
1817 finish_task_switch(rq, prev);
1818 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1819 /* In this case, finish_task_switch does not reenable preemption */
1822 if (current->set_child_tid)
1823 put_user(current->pid, current->set_child_tid);
1827 * context_switch - switch to the new MM and the new
1828 * thread's register state.
1830 static inline struct task_struct *
1831 context_switch(struct rq *rq, struct task_struct *prev,
1832 struct task_struct *next)
1834 struct mm_struct *mm = next->mm;
1835 struct mm_struct *oldmm = prev->active_mm;
1838 next->active_mm = oldmm;
1839 atomic_inc(&oldmm->mm_count);
1840 enter_lazy_tlb(oldmm, next);
1842 switch_mm(oldmm, mm, next);
1845 prev->active_mm = NULL;
1846 WARN_ON(rq->prev_mm);
1847 rq->prev_mm = oldmm;
1850 * Since the runqueue lock will be released by the next
1851 * task (which is an invalid locking op but in the case
1852 * of the scheduler it's an obvious special-case), so we
1853 * do an early lockdep release here:
1855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1856 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1859 /* Here we just switch the register state and the stack. */
1860 switch_to(prev, next, prev);
1866 * nr_running, nr_uninterruptible and nr_context_switches:
1868 * externally visible scheduler statistics: current number of runnable
1869 * threads, current number of uninterruptible-sleeping threads, total
1870 * number of context switches performed since bootup.
1872 unsigned long nr_running(void)
1874 unsigned long i, sum = 0;
1876 for_each_online_cpu(i)
1877 sum += cpu_rq(i)->nr_running;
1882 unsigned long nr_uninterruptible(void)
1884 unsigned long i, sum = 0;
1886 for_each_possible_cpu(i)
1887 sum += cpu_rq(i)->nr_uninterruptible;
1890 * Since we read the counters lockless, it might be slightly
1891 * inaccurate. Do not allow it to go below zero though:
1893 if (unlikely((long)sum < 0))
1899 unsigned long long nr_context_switches(void)
1902 unsigned long long sum = 0;
1904 for_each_possible_cpu(i)
1905 sum += cpu_rq(i)->nr_switches;
1910 unsigned long nr_iowait(void)
1912 unsigned long i, sum = 0;
1914 for_each_possible_cpu(i)
1915 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1920 unsigned long nr_active(void)
1922 unsigned long i, running = 0, uninterruptible = 0;
1924 for_each_online_cpu(i) {
1925 running += cpu_rq(i)->nr_running;
1926 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1929 if (unlikely((long)uninterruptible < 0))
1930 uninterruptible = 0;
1932 return running + uninterruptible;
1938 * Is this task likely cache-hot:
1941 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1943 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1947 * double_rq_lock - safely lock two runqueues
1949 * Note this does not disable interrupts like task_rq_lock,
1950 * you need to do so manually before calling.
1952 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1953 __acquires(rq1->lock)
1954 __acquires(rq2->lock)
1956 BUG_ON(!irqs_disabled());
1958 spin_lock(&rq1->lock);
1959 __acquire(rq2->lock); /* Fake it out ;) */
1962 spin_lock(&rq1->lock);
1963 spin_lock(&rq2->lock);
1965 spin_lock(&rq2->lock);
1966 spin_lock(&rq1->lock);
1972 * double_rq_unlock - safely unlock two runqueues
1974 * Note this does not restore interrupts like task_rq_unlock,
1975 * you need to do so manually after calling.
1977 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1978 __releases(rq1->lock)
1979 __releases(rq2->lock)
1981 spin_unlock(&rq1->lock);
1983 spin_unlock(&rq2->lock);
1985 __release(rq2->lock);
1989 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1991 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1992 __releases(this_rq->lock)
1993 __acquires(busiest->lock)
1994 __acquires(this_rq->lock)
1996 if (unlikely(!irqs_disabled())) {
1997 /* printk() doesn't work good under rq->lock */
1998 spin_unlock(&this_rq->lock);
2001 if (unlikely(!spin_trylock(&busiest->lock))) {
2002 if (busiest < this_rq) {
2003 spin_unlock(&this_rq->lock);
2004 spin_lock(&busiest->lock);
2005 spin_lock(&this_rq->lock);
2007 spin_lock(&busiest->lock);
2012 * If dest_cpu is allowed for this process, migrate the task to it.
2013 * This is accomplished by forcing the cpu_allowed mask to only
2014 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2015 * the cpu_allowed mask is restored.
2017 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2019 struct migration_req req;
2020 unsigned long flags;
2023 rq = task_rq_lock(p, &flags);
2024 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2025 || unlikely(cpu_is_offline(dest_cpu)))
2028 /* force the process onto the specified CPU */
2029 if (migrate_task(p, dest_cpu, &req)) {
2030 /* Need to wait for migration thread (might exit: take ref). */
2031 struct task_struct *mt = rq->migration_thread;
2033 get_task_struct(mt);
2034 task_rq_unlock(rq, &flags);
2035 wake_up_process(mt);
2036 put_task_struct(mt);
2037 wait_for_completion(&req.done);
2042 task_rq_unlock(rq, &flags);
2046 * sched_exec - execve() is a valuable balancing opportunity, because at
2047 * this point the task has the smallest effective memory and cache footprint.
2049 void sched_exec(void)
2051 int new_cpu, this_cpu = get_cpu();
2052 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2054 if (new_cpu != this_cpu)
2055 sched_migrate_task(current, new_cpu);
2059 * pull_task - move a task from a remote runqueue to the local runqueue.
2060 * Both runqueues must be locked.
2062 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2063 struct task_struct *p, struct rq *this_rq,
2064 struct prio_array *this_array, int this_cpu)
2066 dequeue_task(p, src_array);
2067 dec_nr_running(p, src_rq);
2068 set_task_cpu(p, this_cpu);
2069 inc_nr_running(p, this_rq);
2070 enqueue_task(p, this_array);
2071 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
2072 + this_rq->timestamp_last_tick;
2074 * Note that idle threads have a prio of MAX_PRIO, for this test
2075 * to be always true for them.
2077 if (TASK_PREEMPTS_CURR(p, this_rq))
2078 resched_task(this_rq->curr);
2082 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2085 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2086 struct sched_domain *sd, enum idle_type idle,
2090 * We do not migrate tasks that are:
2091 * 1) running (obviously), or
2092 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2093 * 3) are cache-hot on their current CPU.
2095 if (!cpu_isset(this_cpu, p->cpus_allowed))
2099 if (task_running(rq, p))
2103 * Aggressive migration if:
2104 * 1) task is cache cold, or
2105 * 2) too many balance attempts have failed.
2108 if (sd->nr_balance_failed > sd->cache_nice_tries)
2111 if (task_hot(p, rq->timestamp_last_tick, sd))
2116 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2119 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2120 * load from busiest to this_rq, as part of a balancing operation within
2121 * "domain". Returns the number of tasks moved.
2123 * Called with both runqueues locked.
2125 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2126 unsigned long max_nr_move, unsigned long max_load_move,
2127 struct sched_domain *sd, enum idle_type idle,
2130 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2131 best_prio_seen, skip_for_load;
2132 struct prio_array *array, *dst_array;
2133 struct list_head *head, *curr;
2134 struct task_struct *tmp;
2137 if (max_nr_move == 0 || max_load_move == 0)
2140 rem_load_move = max_load_move;
2142 this_best_prio = rq_best_prio(this_rq);
2143 best_prio = rq_best_prio(busiest);
2145 * Enable handling of the case where there is more than one task
2146 * with the best priority. If the current running task is one
2147 * of those with prio==best_prio we know it won't be moved
2148 * and therefore it's safe to override the skip (based on load) of
2149 * any task we find with that prio.
2151 best_prio_seen = best_prio == busiest->curr->prio;
2154 * We first consider expired tasks. Those will likely not be
2155 * executed in the near future, and they are most likely to
2156 * be cache-cold, thus switching CPUs has the least effect
2159 if (busiest->expired->nr_active) {
2160 array = busiest->expired;
2161 dst_array = this_rq->expired;
2163 array = busiest->active;
2164 dst_array = this_rq->active;
2168 /* Start searching at priority 0: */
2172 idx = sched_find_first_bit(array->bitmap);
2174 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2175 if (idx >= MAX_PRIO) {
2176 if (array == busiest->expired && busiest->active->nr_active) {
2177 array = busiest->active;
2178 dst_array = this_rq->active;
2184 head = array->queue + idx;
2187 tmp = list_entry(curr, struct task_struct, run_list);
2192 * To help distribute high priority tasks accross CPUs we don't
2193 * skip a task if it will be the highest priority task (i.e. smallest
2194 * prio value) on its new queue regardless of its load weight
2196 skip_for_load = tmp->load_weight > rem_load_move;
2197 if (skip_for_load && idx < this_best_prio)
2198 skip_for_load = !best_prio_seen && idx == best_prio;
2199 if (skip_for_load ||
2200 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2202 best_prio_seen |= idx == best_prio;
2209 #ifdef CONFIG_SCHEDSTATS
2210 if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2211 schedstat_inc(sd, lb_hot_gained[idle]);
2214 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2216 rem_load_move -= tmp->load_weight;
2219 * We only want to steal up to the prescribed number of tasks
2220 * and the prescribed amount of weighted load.
2222 if (pulled < max_nr_move && rem_load_move > 0) {
2223 if (idx < this_best_prio)
2224 this_best_prio = idx;
2232 * Right now, this is the only place pull_task() is called,
2233 * so we can safely collect pull_task() stats here rather than
2234 * inside pull_task().
2236 schedstat_add(sd, lb_gained[idle], pulled);
2239 *all_pinned = pinned;
2244 * find_busiest_group finds and returns the busiest CPU group within the
2245 * domain. It calculates and returns the amount of weighted load which
2246 * should be moved to restore balance via the imbalance parameter.
2248 static struct sched_group *
2249 find_busiest_group(struct sched_domain *sd, int this_cpu,
2250 unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2253 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2254 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2255 unsigned long max_pull;
2256 unsigned long busiest_load_per_task, busiest_nr_running;
2257 unsigned long this_load_per_task, this_nr_running;
2259 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2260 int power_savings_balance = 1;
2261 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2262 unsigned long min_nr_running = ULONG_MAX;
2263 struct sched_group *group_min = NULL, *group_leader = NULL;
2266 max_load = this_load = total_load = total_pwr = 0;
2267 busiest_load_per_task = busiest_nr_running = 0;
2268 this_load_per_task = this_nr_running = 0;
2269 if (idle == NOT_IDLE)
2270 load_idx = sd->busy_idx;
2271 else if (idle == NEWLY_IDLE)
2272 load_idx = sd->newidle_idx;
2274 load_idx = sd->idle_idx;
2277 unsigned long load, group_capacity;
2280 unsigned long sum_nr_running, sum_weighted_load;
2282 local_group = cpu_isset(this_cpu, group->cpumask);
2284 /* Tally up the load of all CPUs in the group */
2285 sum_weighted_load = sum_nr_running = avg_load = 0;
2287 for_each_cpu_mask(i, group->cpumask) {
2290 if (!cpu_isset(i, *cpus))
2295 if (*sd_idle && !idle_cpu(i))
2298 /* Bias balancing toward cpus of our domain */
2300 load = target_load(i, load_idx);
2302 load = source_load(i, load_idx);
2305 sum_nr_running += rq->nr_running;
2306 sum_weighted_load += rq->raw_weighted_load;
2309 total_load += avg_load;
2310 total_pwr += group->cpu_power;
2312 /* Adjust by relative CPU power of the group */
2313 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2315 group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2318 this_load = avg_load;
2320 this_nr_running = sum_nr_running;
2321 this_load_per_task = sum_weighted_load;
2322 } else if (avg_load > max_load &&
2323 sum_nr_running > group_capacity) {
2324 max_load = avg_load;
2326 busiest_nr_running = sum_nr_running;
2327 busiest_load_per_task = sum_weighted_load;
2330 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2332 * Busy processors will not participate in power savings
2335 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2339 * If the local group is idle or completely loaded
2340 * no need to do power savings balance at this domain
2342 if (local_group && (this_nr_running >= group_capacity ||
2344 power_savings_balance = 0;
2347 * If a group is already running at full capacity or idle,
2348 * don't include that group in power savings calculations
2350 if (!power_savings_balance || sum_nr_running >= group_capacity
2355 * Calculate the group which has the least non-idle load.
2356 * This is the group from where we need to pick up the load
2359 if ((sum_nr_running < min_nr_running) ||
2360 (sum_nr_running == min_nr_running &&
2361 first_cpu(group->cpumask) <
2362 first_cpu(group_min->cpumask))) {
2364 min_nr_running = sum_nr_running;
2365 min_load_per_task = sum_weighted_load /
2370 * Calculate the group which is almost near its
2371 * capacity but still has some space to pick up some load
2372 * from other group and save more power
2374 if (sum_nr_running <= group_capacity - 1) {
2375 if (sum_nr_running > leader_nr_running ||
2376 (sum_nr_running == leader_nr_running &&
2377 first_cpu(group->cpumask) >
2378 first_cpu(group_leader->cpumask))) {
2379 group_leader = group;
2380 leader_nr_running = sum_nr_running;
2385 group = group->next;
2386 } while (group != sd->groups);
2388 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2391 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2393 if (this_load >= avg_load ||
2394 100*max_load <= sd->imbalance_pct*this_load)
2397 busiest_load_per_task /= busiest_nr_running;
2399 * We're trying to get all the cpus to the average_load, so we don't
2400 * want to push ourselves above the average load, nor do we wish to
2401 * reduce the max loaded cpu below the average load, as either of these
2402 * actions would just result in more rebalancing later, and ping-pong
2403 * tasks around. Thus we look for the minimum possible imbalance.
2404 * Negative imbalances (*we* are more loaded than anyone else) will
2405 * be counted as no imbalance for these purposes -- we can't fix that
2406 * by pulling tasks to us. Be careful of negative numbers as they'll
2407 * appear as very large values with unsigned longs.
2409 if (max_load <= busiest_load_per_task)
2413 * In the presence of smp nice balancing, certain scenarios can have
2414 * max load less than avg load(as we skip the groups at or below
2415 * its cpu_power, while calculating max_load..)
2417 if (max_load < avg_load) {
2419 goto small_imbalance;
2422 /* Don't want to pull so many tasks that a group would go idle */
2423 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2425 /* How much load to actually move to equalise the imbalance */
2426 *imbalance = min(max_pull * busiest->cpu_power,
2427 (avg_load - this_load) * this->cpu_power)
2431 * if *imbalance is less than the average load per runnable task
2432 * there is no gaurantee that any tasks will be moved so we'll have
2433 * a think about bumping its value to force at least one task to be
2436 if (*imbalance < busiest_load_per_task) {
2437 unsigned long tmp, pwr_now, pwr_move;
2441 pwr_move = pwr_now = 0;
2443 if (this_nr_running) {
2444 this_load_per_task /= this_nr_running;
2445 if (busiest_load_per_task > this_load_per_task)
2448 this_load_per_task = SCHED_LOAD_SCALE;
2450 if (max_load - this_load >= busiest_load_per_task * imbn) {
2451 *imbalance = busiest_load_per_task;
2456 * OK, we don't have enough imbalance to justify moving tasks,
2457 * however we may be able to increase total CPU power used by
2461 pwr_now += busiest->cpu_power *
2462 min(busiest_load_per_task, max_load);
2463 pwr_now += this->cpu_power *
2464 min(this_load_per_task, this_load);
2465 pwr_now /= SCHED_LOAD_SCALE;
2467 /* Amount of load we'd subtract */
2468 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2470 pwr_move += busiest->cpu_power *
2471 min(busiest_load_per_task, max_load - tmp);
2473 /* Amount of load we'd add */
2474 if (max_load*busiest->cpu_power <
2475 busiest_load_per_task*SCHED_LOAD_SCALE)
2476 tmp = max_load*busiest->cpu_power/this->cpu_power;
2478 tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2479 pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2480 pwr_move /= SCHED_LOAD_SCALE;
2482 /* Move if we gain throughput */
2483 if (pwr_move <= pwr_now)
2486 *imbalance = busiest_load_per_task;
2492 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2493 if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2496 if (this == group_leader && group_leader != group_min) {
2497 *imbalance = min_load_per_task;
2507 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2510 find_busiest_queue(struct sched_group *group, enum idle_type idle,
2511 unsigned long imbalance, cpumask_t *cpus)
2513 struct rq *busiest = NULL, *rq;
2514 unsigned long max_load = 0;
2517 for_each_cpu_mask(i, group->cpumask) {
2519 if (!cpu_isset(i, *cpus))
2524 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2527 if (rq->raw_weighted_load > max_load) {
2528 max_load = rq->raw_weighted_load;
2537 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2538 * so long as it is large enough.
2540 #define MAX_PINNED_INTERVAL 512
2542 static inline unsigned long minus_1_or_zero(unsigned long n)
2544 return n > 0 ? n - 1 : 0;
2548 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2549 * tasks if there is an imbalance.
2551 static int load_balance(int this_cpu, struct rq *this_rq,
2552 struct sched_domain *sd, enum idle_type idle)
2554 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2555 struct sched_group *group;
2556 unsigned long imbalance;
2558 cpumask_t cpus = CPU_MASK_ALL;
2559 unsigned long flags;
2562 * When power savings policy is enabled for the parent domain, idle
2563 * sibling can pick up load irrespective of busy siblings. In this case,
2564 * let the state of idle sibling percolate up as IDLE, instead of
2565 * portraying it as NOT_IDLE.
2567 if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2568 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2571 schedstat_inc(sd, lb_cnt[idle]);
2574 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2577 schedstat_inc(sd, lb_nobusyg[idle]);
2581 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2583 schedstat_inc(sd, lb_nobusyq[idle]);
2587 BUG_ON(busiest == this_rq);
2589 schedstat_add(sd, lb_imbalance[idle], imbalance);
2592 if (busiest->nr_running > 1) {
2594 * Attempt to move tasks. If find_busiest_group has found
2595 * an imbalance but busiest->nr_running <= 1, the group is
2596 * still unbalanced. nr_moved simply stays zero, so it is
2597 * correctly treated as an imbalance.
2599 local_irq_save(flags);
2600 double_rq_lock(this_rq, busiest);
2601 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2602 minus_1_or_zero(busiest->nr_running),
2603 imbalance, sd, idle, &all_pinned);
2604 double_rq_unlock(this_rq, busiest);
2605 local_irq_restore(flags);
2607 /* All tasks on this runqueue were pinned by CPU affinity */
2608 if (unlikely(all_pinned)) {
2609 cpu_clear(cpu_of(busiest), cpus);
2610 if (!cpus_empty(cpus))
2617 schedstat_inc(sd, lb_failed[idle]);
2618 sd->nr_balance_failed++;
2620 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2622 spin_lock_irqsave(&busiest->lock, flags);
2624 /* don't kick the migration_thread, if the curr
2625 * task on busiest cpu can't be moved to this_cpu
2627 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2628 spin_unlock_irqrestore(&busiest->lock, flags);
2630 goto out_one_pinned;
2633 if (!busiest->active_balance) {
2634 busiest->active_balance = 1;
2635 busiest->push_cpu = this_cpu;
2638 spin_unlock_irqrestore(&busiest->lock, flags);
2640 wake_up_process(busiest->migration_thread);
2643 * We've kicked active balancing, reset the failure
2646 sd->nr_balance_failed = sd->cache_nice_tries+1;
2649 sd->nr_balance_failed = 0;
2651 if (likely(!active_balance)) {
2652 /* We were unbalanced, so reset the balancing interval */
2653 sd->balance_interval = sd->min_interval;
2656 * If we've begun active balancing, start to back off. This
2657 * case may not be covered by the all_pinned logic if there
2658 * is only 1 task on the busy runqueue (because we don't call
2661 if (sd->balance_interval < sd->max_interval)
2662 sd->balance_interval *= 2;
2665 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2666 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2671 schedstat_inc(sd, lb_balanced[idle]);
2673 sd->nr_balance_failed = 0;
2676 /* tune up the balancing interval */
2677 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2678 (sd->balance_interval < sd->max_interval))
2679 sd->balance_interval *= 2;
2681 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2682 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2688 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2689 * tasks if there is an imbalance.
2691 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2692 * this_rq is locked.
2695 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2697 struct sched_group *group;
2698 struct rq *busiest = NULL;
2699 unsigned long imbalance;
2702 cpumask_t cpus = CPU_MASK_ALL;
2705 * When power savings policy is enabled for the parent domain, idle
2706 * sibling can pick up load irrespective of busy siblings. In this case,
2707 * let the state of idle sibling percolate up as IDLE, instead of
2708 * portraying it as NOT_IDLE.
2710 if (sd->flags & SD_SHARE_CPUPOWER &&
2711 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2714 schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2716 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2719 schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2723 busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2726 schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2730 BUG_ON(busiest == this_rq);
2732 schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2735 if (busiest->nr_running > 1) {
2736 /* Attempt to move tasks */
2737 double_lock_balance(this_rq, busiest);
2738 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2739 minus_1_or_zero(busiest->nr_running),
2740 imbalance, sd, NEWLY_IDLE, NULL);
2741 spin_unlock(&busiest->lock);
2744 cpu_clear(cpu_of(busiest), cpus);
2745 if (!cpus_empty(cpus))
2751 schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2752 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2753 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2756 sd->nr_balance_failed = 0;
2761 schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2762 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2763 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2765 sd->nr_balance_failed = 0;
2771 * idle_balance is called by schedule() if this_cpu is about to become
2772 * idle. Attempts to pull tasks from other CPUs.
2774 static void idle_balance(int this_cpu, struct rq *this_rq)
2776 struct sched_domain *sd;
2777 int pulled_task = 0;
2778 unsigned long next_balance = jiffies + 60 * HZ;
2780 for_each_domain(this_cpu, sd) {
2781 if (sd->flags & SD_BALANCE_NEWIDLE) {
2782 /* If we've pulled tasks over stop searching: */
2783 pulled_task = load_balance_newidle(this_cpu,
2785 if (time_after(next_balance,
2786 sd->last_balance + sd->balance_interval))
2787 next_balance = sd->last_balance
2788 + sd->balance_interval;
2795 * We are going idle. next_balance may be set based on
2796 * a busy processor. So reset next_balance.
2798 this_rq->next_balance = next_balance;
2802 * active_load_balance is run by migration threads. It pushes running tasks
2803 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2804 * running on each physical CPU where possible, and avoids physical /
2805 * logical imbalances.
2807 * Called with busiest_rq locked.
2809 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2811 int target_cpu = busiest_rq->push_cpu;
2812 struct sched_domain *sd;
2813 struct rq *target_rq;
2815 /* Is there any task to move? */
2816 if (busiest_rq->nr_running <= 1)
2819 target_rq = cpu_rq(target_cpu);
2822 * This condition is "impossible", if it occurs
2823 * we need to fix it. Originally reported by
2824 * Bjorn Helgaas on a 128-cpu setup.
2826 BUG_ON(busiest_rq == target_rq);
2828 /* move a task from busiest_rq to target_rq */
2829 double_lock_balance(busiest_rq, target_rq);
2831 /* Search for an sd spanning us and the target CPU. */
2832 for_each_domain(target_cpu, sd) {
2833 if ((sd->flags & SD_LOAD_BALANCE) &&
2834 cpu_isset(busiest_cpu, sd->span))
2839 schedstat_inc(sd, alb_cnt);
2841 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2842 RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2844 schedstat_inc(sd, alb_pushed);
2846 schedstat_inc(sd, alb_failed);
2848 spin_unlock(&target_rq->lock);
2851 static void update_load(struct rq *this_rq)
2853 unsigned long this_load;
2856 this_load = this_rq->raw_weighted_load;
2858 /* Update our load: */
2859 for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2860 unsigned long old_load, new_load;
2862 old_load = this_rq->cpu_load[i];
2863 new_load = this_load;
2865 * Round up the averaging division if load is increasing. This
2866 * prevents us from getting stuck on 9 if the load is 10, for
2869 if (new_load > old_load)
2870 new_load += scale-1;
2871 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2876 * run_rebalance_domains is triggered when needed from the scheduler tick.
2878 * It checks each scheduling domain to see if it is due to be balanced,
2879 * and initiates a balancing operation if so.
2881 * Balancing parameters are set up in arch_init_sched_domains.
2884 static void run_rebalance_domains(struct softirq_action *h)
2886 int this_cpu = smp_processor_id();
2887 struct rq *this_rq = cpu_rq(this_cpu);
2888 unsigned long interval;
2889 struct sched_domain *sd;
2891 * We are idle if there are no processes running. This
2892 * is valid even if we are the idle process (SMT).
2894 enum idle_type idle = !this_rq->nr_running ?
2895 SCHED_IDLE : NOT_IDLE;
2896 /* Earliest time when we have to call run_rebalance_domains again */
2897 unsigned long next_balance = jiffies + 60*HZ;
2899 for_each_domain(this_cpu, sd) {
2900 if (!(sd->flags & SD_LOAD_BALANCE))
2903 interval = sd->balance_interval;
2904 if (idle != SCHED_IDLE)
2905 interval *= sd->busy_factor;
2907 /* scale ms to jiffies */
2908 interval = msecs_to_jiffies(interval);
2909 if (unlikely(!interval))
2912 if (time_after_eq(jiffies, sd->last_balance + interval)) {
2913 if (load_balance(this_cpu, this_rq, sd, idle)) {
2915 * We've pulled tasks over so either we're no
2916 * longer idle, or one of our SMT siblings is
2921 sd->last_balance = jiffies;
2923 if (time_after(next_balance, sd->last_balance + interval))
2924 next_balance = sd->last_balance + interval;
2926 this_rq->next_balance = next_balance;
2930 * on UP we do not need to balance between CPUs:
2932 static inline void idle_balance(int cpu, struct rq *rq)
2937 static inline void wake_priority_sleeper(struct rq *rq)
2939 #ifdef CONFIG_SCHED_SMT
2940 if (!rq->nr_running)
2943 spin_lock(&rq->lock);
2945 * If an SMT sibling task has been put to sleep for priority
2946 * reasons reschedule the idle task to see if it can now run.
2949 resched_task(rq->idle);
2950 spin_unlock(&rq->lock);
2954 DEFINE_PER_CPU(struct kernel_stat, kstat);
2956 EXPORT_PER_CPU_SYMBOL(kstat);
2959 * This is called on clock ticks and on context switches.
2960 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2963 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
2965 p->sched_time += now - max(p->timestamp, rq->timestamp_last_tick);
2969 * Return current->sched_time plus any more ns on the sched_clock
2970 * that have not yet been banked.
2972 unsigned long long current_sched_time(const struct task_struct *p)
2974 unsigned long long ns;
2975 unsigned long flags;
2977 local_irq_save(flags);
2978 ns = max(p->timestamp, task_rq(p)->timestamp_last_tick);
2979 ns = p->sched_time + sched_clock() - ns;
2980 local_irq_restore(flags);
2986 * We place interactive tasks back into the active array, if possible.
2988 * To guarantee that this does not starve expired tasks we ignore the
2989 * interactivity of a task if the first expired task had to wait more
2990 * than a 'reasonable' amount of time. This deadline timeout is
2991 * load-dependent, as the frequency of array switched decreases with
2992 * increasing number of running tasks. We also ignore the interactivity
2993 * if a better static_prio task has expired:
2995 static inline int expired_starving(struct rq *rq)
2997 if (rq->curr->static_prio > rq->best_expired_prio)
2999 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3001 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3007 * Account user cpu time to a process.
3008 * @p: the process that the cpu time gets accounted to
3009 * @hardirq_offset: the offset to subtract from hardirq_count()
3010 * @cputime: the cpu time spent in user space since the last update
3012 void account_user_time(struct task_struct *p, cputime_t cputime)
3014 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3017 p->utime = cputime_add(p->utime, cputime);
3019 /* Add user time to cpustat. */
3020 tmp = cputime_to_cputime64(cputime);
3021 if (TASK_NICE(p) > 0)
3022 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3024 cpustat->user = cputime64_add(cpustat->user, tmp);
3028 * Account system cpu time to a process.
3029 * @p: the process that the cpu time gets accounted to
3030 * @hardirq_offset: the offset to subtract from hardirq_count()
3031 * @cputime: the cpu time spent in kernel space since the last update
3033 void account_system_time(struct task_struct *p, int hardirq_offset,
3036 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3037 struct rq *rq = this_rq();
3040 p->stime = cputime_add(p->stime, cputime);
3042 /* Add system time to cpustat. */
3043 tmp = cputime_to_cputime64(cputime);
3044 if (hardirq_count() - hardirq_offset)
3045 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3046 else if (softirq_count())
3047 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3048 else if (p != rq->idle)
3049 cpustat->system = cputime64_add(cpustat->system, tmp);
3050 else if (atomic_read(&rq->nr_iowait) > 0)
3051 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3053 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3054 /* Account for system time used */
3055 acct_update_integrals(p);
3059 * Account for involuntary wait time.
3060 * @p: the process from which the cpu time has been stolen
3061 * @steal: the cpu time spent in involuntary wait
3063 void account_steal_time(struct task_struct *p, cputime_t steal)
3065 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3066 cputime64_t tmp = cputime_to_cputime64(steal);
3067 struct rq *rq = this_rq();
3069 if (p == rq->idle) {
3070 p->stime = cputime_add(p->stime, steal);
3071 if (atomic_read(&rq->nr_iowait) > 0)
3072 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3074 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3076 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3079 static void task_running_tick(struct rq *rq, struct task_struct *p)
3081 if (p->array != rq->active) {
3082 /* Task has expired but was not scheduled yet */
3083 set_tsk_need_resched(p);
3086 spin_lock(&rq->lock);
3088 * The task was running during this tick - update the
3089 * time slice counter. Note: we do not update a thread's
3090 * priority until it either goes to sleep or uses up its
3091 * timeslice. This makes it possible for interactive tasks
3092 * to use up their timeslices at their highest priority levels.
3096 * RR tasks need a special form of timeslice management.
3097 * FIFO tasks have no timeslices.
3099 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3100 p->time_slice = task_timeslice(p);
3101 p->first_time_slice = 0;
3102 set_tsk_need_resched(p);
3104 /* put it at the end of the queue: */
3105 requeue_task(p, rq->active);
3109 if (!--p->time_slice) {
3110 dequeue_task(p, rq->active);
3111 set_tsk_need_resched(p);
3112 p->prio = effective_prio(p);
3113 p->time_slice = task_timeslice(p);
3114 p->first_time_slice = 0;
3116 if (!rq->expired_timestamp)
3117 rq->expired_timestamp = jiffies;
3118 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3119 enqueue_task(p, rq->expired);
3120 if (p->static_prio < rq->best_expired_prio)
3121 rq->best_expired_prio = p->static_prio;
3123 enqueue_task(p, rq->active);
3126 * Prevent a too long timeslice allowing a task to monopolize
3127 * the CPU. We do this by splitting up the timeslice into
3130 * Note: this does not mean the task's timeslices expire or
3131 * get lost in any way, they just might be preempted by
3132 * another task of equal priority. (one with higher
3133 * priority would have preempted this task already.) We
3134 * requeue this task to the end of the list on this priority
3135 * level, which is in essence a round-robin of tasks with
3138 * This only applies to tasks in the interactive
3139 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3141 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3142 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3143 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3144 (p->array == rq->active)) {
3146 requeue_task(p, rq->active);
3147 set_tsk_need_resched(p);
3151 spin_unlock(&rq->lock);
3155 * This function gets called by the timer code, with HZ frequency.
3156 * We call it with interrupts disabled.
3158 * It also gets called by the fork code, when changing the parent's
3161 void scheduler_tick(void)
3163 unsigned long long now = sched_clock();
3164 struct task_struct *p = current;
3165 int cpu = smp_processor_id();
3166 struct rq *rq = cpu_rq(cpu);
3168 update_cpu_clock(p, rq, now);
3170 rq->timestamp_last_tick = now;
3173 /* Task on the idle queue */
3174 wake_priority_sleeper(rq);
3176 task_running_tick(rq, p);
3179 if (time_after_eq(jiffies, rq->next_balance))
3180 raise_softirq(SCHED_SOFTIRQ);
3184 #ifdef CONFIG_SCHED_SMT
3185 static inline void wakeup_busy_runqueue(struct rq *rq)
3187 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3188 if (rq->curr == rq->idle && rq->nr_running)
3189 resched_task(rq->idle);
3193 * Called with interrupt disabled and this_rq's runqueue locked.
3195 static void wake_sleeping_dependent(int this_cpu)
3197 struct sched_domain *tmp, *sd = NULL;
3200 for_each_domain(this_cpu, tmp) {
3201 if (tmp->flags & SD_SHARE_CPUPOWER) {
3210 for_each_cpu_mask(i, sd->span) {
3211 struct rq *smt_rq = cpu_rq(i);
3215 if (unlikely(!spin_trylock(&smt_rq->lock)))
3218 wakeup_busy_runqueue(smt_rq);
3219 spin_unlock(&smt_rq->lock);
3224 * number of 'lost' timeslices this task wont be able to fully
3225 * utilize, if another task runs on a sibling. This models the
3226 * slowdown effect of other tasks running on siblings:
3228 static inline unsigned long
3229 smt_slice(struct task_struct *p, struct sched_domain *sd)
3231 return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3235 * To minimise lock contention and not have to drop this_rq's runlock we only
3236 * trylock the sibling runqueues and bypass those runqueues if we fail to
3237 * acquire their lock. As we only trylock the normal locking order does not
3238 * need to be obeyed.
3241 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3243 struct sched_domain *tmp, *sd = NULL;
3246 /* kernel/rt threads do not participate in dependent sleeping */
3247 if (!p->mm || rt_task(p))
3250 for_each_domain(this_cpu, tmp) {
3251 if (tmp->flags & SD_SHARE_CPUPOWER) {
3260 for_each_cpu_mask(i, sd->span) {
3261 struct task_struct *smt_curr;
3268 if (unlikely(!spin_trylock(&smt_rq->lock)))
3271 smt_curr = smt_rq->curr;
3277 * If a user task with lower static priority than the
3278 * running task on the SMT sibling is trying to schedule,
3279 * delay it till there is proportionately less timeslice
3280 * left of the sibling task to prevent a lower priority
3281 * task from using an unfair proportion of the
3282 * physical cpu's resources. -ck
3284 if (rt_task(smt_curr)) {
3286 * With real time tasks we run non-rt tasks only
3287 * per_cpu_gain% of the time.
3289 if ((jiffies % DEF_TIMESLICE) >
3290 (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3293 if (smt_curr->static_prio < p->static_prio &&
3294 !TASK_PREEMPTS_CURR(p, smt_rq) &&
3295 smt_slice(smt_curr, sd) > task_timeslice(p))
3299 spin_unlock(&smt_rq->lock);
3304 static inline void wake_sleeping_dependent(int this_cpu)
3308 dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3314 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3316 void fastcall add_preempt_count(int val)
3321 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3323 preempt_count() += val;
3325 * Spinlock count overflowing soon?
3327 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3329 EXPORT_SYMBOL(add_preempt_count);
3331 void fastcall sub_preempt_count(int val)
3336 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3339 * Is the spinlock portion underflowing?
3341 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3342 !(preempt_count() & PREEMPT_MASK)))
3345 preempt_count() -= val;
3347 EXPORT_SYMBOL(sub_preempt_count);
3351 static inline int interactive_sleep(enum sleep_type sleep_type)
3353 return (sleep_type == SLEEP_INTERACTIVE ||
3354 sleep_type == SLEEP_INTERRUPTED);
3358 * schedule() is the main scheduler function.
3360 asmlinkage void __sched schedule(void)
3362 struct task_struct *prev, *next;
3363 struct prio_array *array;
3364 struct list_head *queue;
3365 unsigned long long now;
3366 unsigned long run_time;
3367 int cpu, idx, new_prio;
3372 * Test if we are atomic. Since do_exit() needs to call into
3373 * schedule() atomically, we ignore that path for now.
3374 * Otherwise, whine if we are scheduling when we should not be.
3376 if (unlikely(in_atomic() && !current->exit_state)) {
3377 printk(KERN_ERR "BUG: scheduling while atomic: "
3379 current->comm, preempt_count(), current->pid);
3380 debug_show_held_locks(current);
3383 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3388 release_kernel_lock(prev);
3389 need_resched_nonpreemptible:
3393 * The idle thread is not allowed to schedule!
3394 * Remove this check after it has been exercised a bit.
3396 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3397 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3401 schedstat_inc(rq, sched_cnt);
3402 now = sched_clock();
3403 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3404 run_time = now - prev->timestamp;
3405 if (unlikely((long long)(now - prev->timestamp) < 0))
3408 run_time = NS_MAX_SLEEP_AVG;
3411 * Tasks charged proportionately less run_time at high sleep_avg to
3412 * delay them losing their interactive status
3414 run_time /= (CURRENT_BONUS(prev) ? : 1);
3416 spin_lock_irq(&rq->lock);
3418 switch_count = &prev->nivcsw;
3419 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3420 switch_count = &prev->nvcsw;
3421 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3422 unlikely(signal_pending(prev))))
3423 prev->state = TASK_RUNNING;
3425 if (prev->state == TASK_UNINTERRUPTIBLE)
3426 rq->nr_uninterruptible++;
3427 deactivate_task(prev, rq);
3431 cpu = smp_processor_id();
3432 if (unlikely(!rq->nr_running)) {
3433 idle_balance(cpu, rq);
3434 if (!rq->nr_running) {
3436 rq->expired_timestamp = 0;
3437 wake_sleeping_dependent(cpu);
3443 if (unlikely(!array->nr_active)) {
3445 * Switch the active and expired arrays.
3447 schedstat_inc(rq, sched_switch);
3448 rq->active = rq->expired;
3449 rq->expired = array;
3451 rq->expired_timestamp = 0;
3452 rq->best_expired_prio = MAX_PRIO;
3455 idx = sched_find_first_bit(array->bitmap);
3456 queue = array->queue + idx;
3457 next = list_entry(queue->next, struct task_struct, run_list);
3459 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3460 unsigned long long delta = now - next->timestamp;
3461 if (unlikely((long long)(now - next->timestamp) < 0))
3464 if (next->sleep_type == SLEEP_INTERACTIVE)
3465 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3467 array = next->array;
3468 new_prio = recalc_task_prio(next, next->timestamp + delta);
3470 if (unlikely(next->prio != new_prio)) {
3471 dequeue_task(next, array);
3472 next->prio = new_prio;
3473 enqueue_task(next, array);
3476 next->sleep_type = SLEEP_NORMAL;
3477 if (dependent_sleeper(cpu, rq, next))
3480 if (next == rq->idle)
3481 schedstat_inc(rq, sched_goidle);
3483 prefetch_stack(next);
3484 clear_tsk_need_resched(prev);
3485 rcu_qsctr_inc(task_cpu(prev));
3487 update_cpu_clock(prev, rq, now);
3489 prev->sleep_avg -= run_time;
3490 if ((long)prev->sleep_avg <= 0)
3491 prev->sleep_avg = 0;
3492 prev->timestamp = prev->last_ran = now;
3494 sched_info_switch(prev, next);
3495 if (likely(prev != next)) {
3496 next->timestamp = now;
3501 prepare_task_switch(rq, next);
3502 prev = context_switch(rq, prev, next);
3505 * this_rq must be evaluated again because prev may have moved
3506 * CPUs since it called schedule(), thus the 'rq' on its stack
3507 * frame will be invalid.
3509 finish_task_switch(this_rq(), prev);
3511 spin_unlock_irq(&rq->lock);
3514 if (unlikely(reacquire_kernel_lock(prev) < 0))
3515 goto need_resched_nonpreemptible;
3516 preempt_enable_no_resched();
3517 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3520 EXPORT_SYMBOL(schedule);
3522 #ifdef CONFIG_PREEMPT
3524 * this is the entry point to schedule() from in-kernel preemption
3525 * off of preempt_enable. Kernel preemptions off return from interrupt
3526 * occur there and call schedule directly.
3528 asmlinkage void __sched preempt_schedule(void)
3530 struct thread_info *ti = current_thread_info();
3531 #ifdef CONFIG_PREEMPT_BKL
3532 struct task_struct *task = current;
3533 int saved_lock_depth;
3536 * If there is a non-zero preempt_count or interrupts are disabled,
3537 * we do not want to preempt the current task. Just return..
3539 if (likely(ti->preempt_count || irqs_disabled()))
3543 add_preempt_count(PREEMPT_ACTIVE);
3545 * We keep the big kernel semaphore locked, but we
3546 * clear ->lock_depth so that schedule() doesnt
3547 * auto-release the semaphore:
3549 #ifdef CONFIG_PREEMPT_BKL
3550 saved_lock_depth = task->lock_depth;
3551 task->lock_depth = -1;
3554 #ifdef CONFIG_PREEMPT_BKL
3555 task->lock_depth = saved_lock_depth;
3557 sub_preempt_count(PREEMPT_ACTIVE);
3559 /* we could miss a preemption opportunity between schedule and now */
3561 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3564 EXPORT_SYMBOL(preempt_schedule);
3567 * this is the entry point to schedule() from kernel preemption
3568 * off of irq context.
3569 * Note, that this is called and return with irqs disabled. This will
3570 * protect us against recursive calling from irq.
3572 asmlinkage void __sched preempt_schedule_irq(void)
3574 struct thread_info *ti = current_thread_info();
3575 #ifdef CONFIG_PREEMPT_BKL
3576 struct task_struct *task = current;
3577 int saved_lock_depth;
3579 /* Catch callers which need to be fixed */
3580 BUG_ON(ti->preempt_count || !irqs_disabled());
3583 add_preempt_count(PREEMPT_ACTIVE);
3585 * We keep the big kernel semaphore locked, but we
3586 * clear ->lock_depth so that schedule() doesnt
3587 * auto-release the semaphore:
3589 #ifdef CONFIG_PREEMPT_BKL
3590 saved_lock_depth = task->lock_depth;
3591 task->lock_depth = -1;
3595 local_irq_disable();
3596 #ifdef CONFIG_PREEMPT_BKL
3597 task->lock_depth = saved_lock_depth;
3599 sub_preempt_count(PREEMPT_ACTIVE);
3601 /* we could miss a preemption opportunity between schedule and now */
3603 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3607 #endif /* CONFIG_PREEMPT */
3609 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3612 return try_to_wake_up(curr->private, mode, sync);
3614 EXPORT_SYMBOL(default_wake_function);
3617 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3618 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3619 * number) then we wake all the non-exclusive tasks and one exclusive task.
3621 * There are circumstances in which we can try to wake a task which has already
3622 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3623 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3625 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3626 int nr_exclusive, int sync, void *key)
3628 struct list_head *tmp, *next;
3630 list_for_each_safe(tmp, next, &q->task_list) {
3631 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3632 unsigned flags = curr->flags;
3634 if (curr->func(curr, mode, sync, key) &&
3635 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3641 * __wake_up - wake up threads blocked on a waitqueue.
3643 * @mode: which threads
3644 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3645 * @key: is directly passed to the wakeup function
3647 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3648 int nr_exclusive, void *key)
3650 unsigned long flags;
3652 spin_lock_irqsave(&q->lock, flags);
3653 __wake_up_common(q, mode, nr_exclusive, 0, key);
3654 spin_unlock_irqrestore(&q->lock, flags);
3656 EXPORT_SYMBOL(__wake_up);
3659 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3661 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3663 __wake_up_common(q, mode, 1, 0, NULL);
3667 * __wake_up_sync - wake up threads blocked on a waitqueue.
3669 * @mode: which threads
3670 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3672 * The sync wakeup differs that the waker knows that it will schedule
3673 * away soon, so while the target thread will be woken up, it will not
3674 * be migrated to another CPU - ie. the two threads are 'synchronized'
3675 * with each other. This can prevent needless bouncing between CPUs.
3677 * On UP it can prevent extra preemption.
3680 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3682 unsigned long flags;
3688 if (unlikely(!nr_exclusive))
3691 spin_lock_irqsave(&q->lock, flags);
3692 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3693 spin_unlock_irqrestore(&q->lock, flags);
3695 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3697 void fastcall complete(struct completion *x)
3699 unsigned long flags;
3701 spin_lock_irqsave(&x->wait.lock, flags);
3703 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3705 spin_unlock_irqrestore(&x->wait.lock, flags);
3707 EXPORT_SYMBOL(complete);
3709 void fastcall complete_all(struct completion *x)
3711 unsigned long flags;
3713 spin_lock_irqsave(&x->wait.lock, flags);
3714 x->done += UINT_MAX/2;
3715 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3717 spin_unlock_irqrestore(&x->wait.lock, flags);
3719 EXPORT_SYMBOL(complete_all);
3721 void fastcall __sched wait_for_completion(struct completion *x)
3725 spin_lock_irq(&x->wait.lock);
3727 DECLARE_WAITQUEUE(wait, current);
3729 wait.flags |= WQ_FLAG_EXCLUSIVE;
3730 __add_wait_queue_tail(&x->wait, &wait);
3732 __set_current_state(TASK_UNINTERRUPTIBLE);
3733 spin_unlock_irq(&x->wait.lock);
3735 spin_lock_irq(&x->wait.lock);
3737 __remove_wait_queue(&x->wait, &wait);
3740 spin_unlock_irq(&x->wait.lock);
3742 EXPORT_SYMBOL(wait_for_completion);
3744 unsigned long fastcall __sched
3745 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3749 spin_lock_irq(&x->wait.lock);
3751 DECLARE_WAITQUEUE(wait, current);
3753 wait.flags |= WQ_FLAG_EXCLUSIVE;
3754 __add_wait_queue_tail(&x->wait, &wait);
3756 __set_current_state(TASK_UNINTERRUPTIBLE);
3757 spin_unlock_irq(&x->wait.lock);
3758 timeout = schedule_timeout(timeout);
3759 spin_lock_irq(&x->wait.lock);
3761 __remove_wait_queue(&x->wait, &wait);
3765 __remove_wait_queue(&x->wait, &wait);
3769 spin_unlock_irq(&x->wait.lock);
3772 EXPORT_SYMBOL(wait_for_completion_timeout);
3774 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3780 spin_lock_irq(&x->wait.lock);
3782 DECLARE_WAITQUEUE(wait, current);
3784 wait.flags |= WQ_FLAG_EXCLUSIVE;
3785 __add_wait_queue_tail(&x->wait, &wait);
3787 if (signal_pending(current)) {
3789 __remove_wait_queue(&x->wait, &wait);
3792 __set_current_state(TASK_INTERRUPTIBLE);
3793 spin_unlock_irq(&x->wait.lock);
3795 spin_lock_irq(&x->wait.lock);
3797 __remove_wait_queue(&x->wait, &wait);
3801 spin_unlock_irq(&x->wait.lock);
3805 EXPORT_SYMBOL(wait_for_completion_interruptible);
3807 unsigned long fastcall __sched
3808 wait_for_completion_interruptible_timeout(struct completion *x,
3809 unsigned long timeout)
3813 spin_lock_irq(&x->wait.lock);
3815 DECLARE_WAITQUEUE(wait, current);
3817 wait.flags |= WQ_FLAG_EXCLUSIVE;
3818 __add_wait_queue_tail(&x->wait, &wait);
3820 if (signal_pending(current)) {
3821 timeout = -ERESTARTSYS;
3822 __remove_wait_queue(&x->wait, &wait);
3825 __set_current_state(TASK_INTERRUPTIBLE);
3826 spin_unlock_irq(&x->wait.lock);
3827 timeout = schedule_timeout(timeout);
3828 spin_lock_irq(&x->wait.lock);
3830 __remove_wait_queue(&x->wait, &wait);
3834 __remove_wait_queue(&x->wait, &wait);
3838 spin_unlock_irq(&x->wait.lock);
3841 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3844 #define SLEEP_ON_VAR \
3845 unsigned long flags; \
3846 wait_queue_t wait; \
3847 init_waitqueue_entry(&wait, current);
3849 #define SLEEP_ON_HEAD \
3850 spin_lock_irqsave(&q->lock,flags); \
3851 __add_wait_queue(q, &wait); \
3852 spin_unlock(&q->lock);
3854 #define SLEEP_ON_TAIL \
3855 spin_lock_irq(&q->lock); \
3856 __remove_wait_queue(q, &wait); \
3857 spin_unlock_irqrestore(&q->lock, flags);
3859 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3863 current->state = TASK_INTERRUPTIBLE;
3869 EXPORT_SYMBOL(interruptible_sleep_on);
3871 long fastcall __sched
3872 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3876 current->state = TASK_INTERRUPTIBLE;
3879 timeout = schedule_timeout(timeout);
3884 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3886 void fastcall __sched sleep_on(wait_queue_head_t *q)
3890 current->state = TASK_UNINTERRUPTIBLE;
3896 EXPORT_SYMBOL(sleep_on);
3898 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3902 current->state = TASK_UNINTERRUPTIBLE;
3905 timeout = schedule_timeout(timeout);
3911 EXPORT_SYMBOL(sleep_on_timeout);
3913 #ifdef CONFIG_RT_MUTEXES
3916 * rt_mutex_setprio - set the current priority of a task
3918 * @prio: prio value (kernel-internal form)
3920 * This function changes the 'effective' priority of a task. It does
3921 * not touch ->normal_prio like __setscheduler().
3923 * Used by the rt_mutex code to implement priority inheritance logic.
3925 void rt_mutex_setprio(struct task_struct *p, int prio)
3927 struct prio_array *array;
3928 unsigned long flags;
3932 BUG_ON(prio < 0 || prio > MAX_PRIO);
3934 rq = task_rq_lock(p, &flags);
3939 dequeue_task(p, array);
3944 * If changing to an RT priority then queue it
3945 * in the active array!
3949 enqueue_task(p, array);
3951 * Reschedule if we are currently running on this runqueue and
3952 * our priority decreased, or if we are not currently running on
3953 * this runqueue and our priority is higher than the current's
3955 if (task_running(rq, p)) {
3956 if (p->prio > oldprio)
3957 resched_task(rq->curr);
3958 } else if (TASK_PREEMPTS_CURR(p, rq))
3959 resched_task(rq->curr);
3961 task_rq_unlock(rq, &flags);
3966 void set_user_nice(struct task_struct *p, long nice)
3968 struct prio_array *array;
3969 int old_prio, delta;
3970 unsigned long flags;
3973 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3976 * We have to be careful, if called from sys_setpriority(),
3977 * the task might be in the middle of scheduling on another CPU.
3979 rq = task_rq_lock(p, &flags);
3981 * The RT priorities are set via sched_setscheduler(), but we still
3982 * allow the 'normal' nice value to be set - but as expected
3983 * it wont have any effect on scheduling until the task is
3984 * not SCHED_NORMAL/SCHED_BATCH:
3986 if (has_rt_policy(p)) {
3987 p->static_prio = NICE_TO_PRIO(nice);
3992 dequeue_task(p, array);
3993 dec_raw_weighted_load(rq, p);
3996 p->static_prio = NICE_TO_PRIO(nice);
3999 p->prio = effective_prio(p);
4000 delta = p->prio - old_prio;
4003 enqueue_task(p, array);
4004 inc_raw_weighted_load(rq, p);
4006 * If the task increased its priority or is running and
4007 * lowered its priority, then reschedule its CPU:
4009 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4010 resched_task(rq->curr);
4013 task_rq_unlock(rq, &flags);
4015 EXPORT_SYMBOL(set_user_nice);
4018 * can_nice - check if a task can reduce its nice value
4022 int can_nice(const struct task_struct *p, const int nice)
4024 /* convert nice value [19,-20] to rlimit style value [1,40] */
4025 int nice_rlim = 20 - nice;
4027 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4028 capable(CAP_SYS_NICE));
4031 #ifdef __ARCH_WANT_SYS_NICE
4034 * sys_nice - change the priority of the current process.
4035 * @increment: priority increment
4037 * sys_setpriority is a more generic, but much slower function that
4038 * does similar things.
4040 asmlinkage long sys_nice(int increment)
4045 * Setpriority might change our priority at the same moment.
4046 * We don't have to worry. Conceptually one call occurs first
4047 * and we have a single winner.
4049 if (increment < -40)
4054 nice = PRIO_TO_NICE(current->static_prio) + increment;
4060 if (increment < 0 && !can_nice(current, nice))
4063 retval = security_task_setnice(current, nice);
4067 set_user_nice(current, nice);
4074 * task_prio - return the priority value of a given task.
4075 * @p: the task in question.
4077 * This is the priority value as seen by users in /proc.
4078 * RT tasks are offset by -200. Normal tasks are centered
4079 * around 0, value goes from -16 to +15.
4081 int task_prio(const struct task_struct *p)
4083 return p->prio - MAX_RT_PRIO;
4087 * task_nice - return the nice value of a given task.
4088 * @p: the task in question.
4090 int task_nice(const struct task_struct *p)
4092 return TASK_NICE(p);
4094 EXPORT_SYMBOL_GPL(task_nice);
4097 * idle_cpu - is a given cpu idle currently?
4098 * @cpu: the processor in question.
4100 int idle_cpu(int cpu)
4102 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4106 * idle_task - return the idle task for a given cpu.
4107 * @cpu: the processor in question.
4109 struct task_struct *idle_task(int cpu)
4111 return cpu_rq(cpu)->idle;
4115 * find_process_by_pid - find a process with a matching PID value.
4116 * @pid: the pid in question.
4118 static inline struct task_struct *find_process_by_pid(pid_t pid)
4120 return pid ? find_task_by_pid(pid) : current;
4123 /* Actually do priority change: must hold rq lock. */
4124 static void __setscheduler(struct task_struct *p, int policy, int prio)
4129 p->rt_priority = prio;
4130 p->normal_prio = normal_prio(p);
4131 /* we are holding p->pi_lock already */
4132 p->prio = rt_mutex_getprio(p);
4134 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4136 if (policy == SCHED_BATCH)
4142 * sched_setscheduler - change the scheduling policy and/or RT priority of
4144 * @p: the task in question.
4145 * @policy: new policy.
4146 * @param: structure containing the new RT priority.
4148 * NOTE: the task may be already dead
4150 int sched_setscheduler(struct task_struct *p, int policy,
4151 struct sched_param *param)
4153 int retval, oldprio, oldpolicy = -1;
4154 struct prio_array *array;
4155 unsigned long flags;
4158 /* may grab non-irq protected spin_locks */
4159 BUG_ON(in_interrupt());
4161 /* double check policy once rq lock held */
4163 policy = oldpolicy = p->policy;
4164 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4165 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4168 * Valid priorities for SCHED_FIFO and SCHED_RR are
4169 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4172 if (param->sched_priority < 0 ||
4173 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4174 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4176 if (is_rt_policy(policy) != (param->sched_priority != 0))
4180 * Allow unprivileged RT tasks to decrease priority:
4182 if (!capable(CAP_SYS_NICE)) {
4183 if (is_rt_policy(policy)) {
4184 unsigned long rlim_rtprio;
4185 unsigned long flags;
4187 if (!lock_task_sighand(p, &flags))
4189 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4190 unlock_task_sighand(p, &flags);
4192 /* can't set/change the rt policy */
4193 if (policy != p->policy && !rlim_rtprio)
4196 /* can't increase priority */
4197 if (param->sched_priority > p->rt_priority &&
4198 param->sched_priority > rlim_rtprio)
4202 /* can't change other user's priorities */
4203 if ((current->euid != p->euid) &&
4204 (current->euid != p->uid))
4208 retval = security_task_setscheduler(p, policy, param);
4212 * make sure no PI-waiters arrive (or leave) while we are
4213 * changing the priority of the task:
4215 spin_lock_irqsave(&p->pi_lock, flags);
4217 * To be able to change p->policy safely, the apropriate
4218 * runqueue lock must be held.
4220 rq = __task_rq_lock(p);
4221 /* recheck policy now with rq lock held */
4222 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4223 policy = oldpolicy = -1;
4224 __task_rq_unlock(rq);
4225 spin_unlock_irqrestore(&p->pi_lock, flags);
4230 deactivate_task(p, rq);
4232 __setscheduler(p, policy, param->sched_priority);
4234 __activate_task(p, rq);
4236 * Reschedule if we are currently running on this runqueue and
4237 * our priority decreased, or if we are not currently running on
4238 * this runqueue and our priority is higher than the current's
4240 if (task_running(rq, p)) {
4241 if (p->prio > oldprio)
4242 resched_task(rq->curr);
4243 } else if (TASK_PREEMPTS_CURR(p, rq))
4244 resched_task(rq->curr);
4246 __task_rq_unlock(rq);
4247 spin_unlock_irqrestore(&p->pi_lock, flags);
4249 rt_mutex_adjust_pi(p);
4253 EXPORT_SYMBOL_GPL(sched_setscheduler);
4256 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4258 struct sched_param lparam;
4259 struct task_struct *p;
4262 if (!param || pid < 0)
4264 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4269 p = find_process_by_pid(pid);
4271 retval = sched_setscheduler(p, policy, &lparam);
4278 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4279 * @pid: the pid in question.
4280 * @policy: new policy.
4281 * @param: structure containing the new RT priority.
4283 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4284 struct sched_param __user *param)
4286 /* negative values for policy are not valid */
4290 return do_sched_setscheduler(pid, policy, param);
4294 * sys_sched_setparam - set/change the RT priority of a thread
4295 * @pid: the pid in question.
4296 * @param: structure containing the new RT priority.
4298 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4300 return do_sched_setscheduler(pid, -1, param);
4304 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4305 * @pid: the pid in question.
4307 asmlinkage long sys_sched_getscheduler(pid_t pid)
4309 struct task_struct *p;
4310 int retval = -EINVAL;
4316 read_lock(&tasklist_lock);
4317 p = find_process_by_pid(pid);
4319 retval = security_task_getscheduler(p);
4323 read_unlock(&tasklist_lock);
4330 * sys_sched_getscheduler - get the RT priority of a thread
4331 * @pid: the pid in question.
4332 * @param: structure containing the RT priority.
4334 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4336 struct sched_param lp;
4337 struct task_struct *p;
4338 int retval = -EINVAL;
4340 if (!param || pid < 0)
4343 read_lock(&tasklist_lock);
4344 p = find_process_by_pid(pid);
4349 retval = security_task_getscheduler(p);
4353 lp.sched_priority = p->rt_priority;
4354 read_unlock(&tasklist_lock);
4357 * This one might sleep, we cannot do it with a spinlock held ...
4359 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4365 read_unlock(&tasklist_lock);
4369 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4371 cpumask_t cpus_allowed;
4372 struct task_struct *p;
4376 read_lock(&tasklist_lock);
4378 p = find_process_by_pid(pid);
4380 read_unlock(&tasklist_lock);
4381 unlock_cpu_hotplug();
4386 * It is not safe to call set_cpus_allowed with the
4387 * tasklist_lock held. We will bump the task_struct's
4388 * usage count and then drop tasklist_lock.
4391 read_unlock(&tasklist_lock);
4394 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4395 !capable(CAP_SYS_NICE))
4398 retval = security_task_setscheduler(p, 0, NULL);
4402 cpus_allowed = cpuset_cpus_allowed(p);
4403 cpus_and(new_mask, new_mask, cpus_allowed);
4404 retval = set_cpus_allowed(p, new_mask);
4408 unlock_cpu_hotplug();
4412 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4413 cpumask_t *new_mask)
4415 if (len < sizeof(cpumask_t)) {
4416 memset(new_mask, 0, sizeof(cpumask_t));
4417 } else if (len > sizeof(cpumask_t)) {
4418 len = sizeof(cpumask_t);
4420 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4424 * sys_sched_setaffinity - set the cpu affinity of a process
4425 * @pid: pid of the process
4426 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4427 * @user_mask_ptr: user-space pointer to the new cpu mask
4429 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4430 unsigned long __user *user_mask_ptr)
4435 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4439 return sched_setaffinity(pid, new_mask);
4443 * Represents all cpu's present in the system
4444 * In systems capable of hotplug, this map could dynamically grow
4445 * as new cpu's are detected in the system via any platform specific
4446 * method, such as ACPI for e.g.
4449 cpumask_t cpu_present_map __read_mostly;
4450 EXPORT_SYMBOL(cpu_present_map);
4453 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4454 EXPORT_SYMBOL(cpu_online_map);
4456 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4457 EXPORT_SYMBOL(cpu_possible_map);
4460 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4462 struct task_struct *p;
4466 read_lock(&tasklist_lock);
4469 p = find_process_by_pid(pid);
4473 retval = security_task_getscheduler(p);
4477 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4480 read_unlock(&tasklist_lock);
4481 unlock_cpu_hotplug();
4489 * sys_sched_getaffinity - get the cpu affinity of a process
4490 * @pid: pid of the process
4491 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4492 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4494 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4495 unsigned long __user *user_mask_ptr)
4500 if (len < sizeof(cpumask_t))
4503 ret = sched_getaffinity(pid, &mask);
4507 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4510 return sizeof(cpumask_t);
4514 * sys_sched_yield - yield the current processor to other threads.
4516 * this function yields the current CPU by moving the calling thread
4517 * to the expired array. If there are no other threads running on this
4518 * CPU then this function will return.
4520 asmlinkage long sys_sched_yield(void)
4522 struct rq *rq = this_rq_lock();
4523 struct prio_array *array = current->array, *target = rq->expired;
4525 schedstat_inc(rq, yld_cnt);
4527 * We implement yielding by moving the task into the expired
4530 * (special rule: RT tasks will just roundrobin in the active
4533 if (rt_task(current))
4534 target = rq->active;
4536 if (array->nr_active == 1) {
4537 schedstat_inc(rq, yld_act_empty);
4538 if (!rq->expired->nr_active)
4539 schedstat_inc(rq, yld_both_empty);
4540 } else if (!rq->expired->nr_active)
4541 schedstat_inc(rq, yld_exp_empty);
4543 if (array != target) {
4544 dequeue_task(current, array);
4545 enqueue_task(current, target);
4548 * requeue_task is cheaper so perform that if possible.
4550 requeue_task(current, array);
4553 * Since we are going to call schedule() anyway, there's
4554 * no need to preempt or enable interrupts:
4556 __release(rq->lock);
4557 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4558 _raw_spin_unlock(&rq->lock);
4559 preempt_enable_no_resched();
4566 static inline int __resched_legal(int expected_preempt_count)
4568 if (unlikely(preempt_count() != expected_preempt_count))
4570 if (unlikely(system_state != SYSTEM_RUNNING))
4575 static void __cond_resched(void)
4577 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4578 __might_sleep(__FILE__, __LINE__);
4581 * The BKS might be reacquired before we have dropped
4582 * PREEMPT_ACTIVE, which could trigger a second
4583 * cond_resched() call.
4586 add_preempt_count(PREEMPT_ACTIVE);
4588 sub_preempt_count(PREEMPT_ACTIVE);
4589 } while (need_resched());
4592 int __sched cond_resched(void)
4594 if (need_resched() && __resched_legal(0)) {
4600 EXPORT_SYMBOL(cond_resched);
4603 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4604 * call schedule, and on return reacquire the lock.
4606 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4607 * operations here to prevent schedule() from being called twice (once via
4608 * spin_unlock(), once by hand).
4610 int cond_resched_lock(spinlock_t *lock)
4614 if (need_lockbreak(lock)) {
4620 if (need_resched() && __resched_legal(1)) {
4621 spin_release(&lock->dep_map, 1, _THIS_IP_);
4622 _raw_spin_unlock(lock);
4623 preempt_enable_no_resched();
4630 EXPORT_SYMBOL(cond_resched_lock);
4632 int __sched cond_resched_softirq(void)
4634 BUG_ON(!in_softirq());
4636 if (need_resched() && __resched_legal(0)) {
4637 raw_local_irq_disable();
4639 raw_local_irq_enable();
4646 EXPORT_SYMBOL(cond_resched_softirq);
4649 * yield - yield the current processor to other threads.
4651 * this is a shortcut for kernel-space yielding - it marks the
4652 * thread runnable and calls sys_sched_yield().
4654 void __sched yield(void)
4656 set_current_state(TASK_RUNNING);
4659 EXPORT_SYMBOL(yield);
4662 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4663 * that process accounting knows that this is a task in IO wait state.
4665 * But don't do that if it is a deliberate, throttling IO wait (this task
4666 * has set its backing_dev_info: the queue against which it should throttle)
4668 void __sched io_schedule(void)
4670 struct rq *rq = &__raw_get_cpu_var(runqueues);
4672 delayacct_blkio_start();
4673 atomic_inc(&rq->nr_iowait);
4675 atomic_dec(&rq->nr_iowait);
4676 delayacct_blkio_end();
4678 EXPORT_SYMBOL(io_schedule);
4680 long __sched io_schedule_timeout(long timeout)
4682 struct rq *rq = &__raw_get_cpu_var(runqueues);
4685 delayacct_blkio_start();
4686 atomic_inc(&rq->nr_iowait);
4687 ret = schedule_timeout(timeout);
4688 atomic_dec(&rq->nr_iowait);
4689 delayacct_blkio_end();
4694 * sys_sched_get_priority_max - return maximum RT priority.
4695 * @policy: scheduling class.
4697 * this syscall returns the maximum rt_priority that can be used
4698 * by a given scheduling class.
4700 asmlinkage long sys_sched_get_priority_max(int policy)
4707 ret = MAX_USER_RT_PRIO-1;
4718 * sys_sched_get_priority_min - return minimum RT priority.
4719 * @policy: scheduling class.
4721 * this syscall returns the minimum rt_priority that can be used
4722 * by a given scheduling class.
4724 asmlinkage long sys_sched_get_priority_min(int policy)
4741 * sys_sched_rr_get_interval - return the default timeslice of a process.
4742 * @pid: pid of the process.
4743 * @interval: userspace pointer to the timeslice value.
4745 * this syscall writes the default timeslice value of a given process
4746 * into the user-space timespec buffer. A value of '0' means infinity.
4749 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4751 struct task_struct *p;
4752 int retval = -EINVAL;
4759 read_lock(&tasklist_lock);
4760 p = find_process_by_pid(pid);
4764 retval = security_task_getscheduler(p);
4768 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4769 0 : task_timeslice(p), &t);
4770 read_unlock(&tasklist_lock);
4771 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4775 read_unlock(&tasklist_lock);
4779 static inline struct task_struct *eldest_child(struct task_struct *p)
4781 if (list_empty(&p->children))
4783 return list_entry(p->children.next,struct task_struct,sibling);
4786 static inline struct task_struct *older_sibling(struct task_struct *p)
4788 if (p->sibling.prev==&p->parent->children)
4790 return list_entry(p->sibling.prev,struct task_struct,sibling);
4793 static inline struct task_struct *younger_sibling(struct task_struct *p)
4795 if (p->sibling.next==&p->parent->children)
4797 return list_entry(p->sibling.next,struct task_struct,sibling);
4800 static const char stat_nam[] = "RSDTtZX";
4802 static void show_task(struct task_struct *p)
4804 struct task_struct *relative;
4805 unsigned long free = 0;
4808 state = p->state ? __ffs(p->state) + 1 : 0;
4809 printk("%-13.13s %c", p->comm,
4810 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4811 #if (BITS_PER_LONG == 32)
4812 if (state == TASK_RUNNING)
4813 printk(" running ");
4815 printk(" %08lX ", thread_saved_pc(p));
4817 if (state == TASK_RUNNING)
4818 printk(" running task ");
4820 printk(" %016lx ", thread_saved_pc(p));
4822 #ifdef CONFIG_DEBUG_STACK_USAGE
4824 unsigned long *n = end_of_stack(p);
4827 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4830 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4831 if ((relative = eldest_child(p)))
4832 printk("%5d ", relative->pid);
4835 if ((relative = younger_sibling(p)))
4836 printk("%7d", relative->pid);
4839 if ((relative = older_sibling(p)))
4840 printk(" %5d", relative->pid);
4844 printk(" (L-TLB)\n");
4846 printk(" (NOTLB)\n");
4848 if (state != TASK_RUNNING)
4849 show_stack(p, NULL);
4852 void show_state_filter(unsigned long state_filter)
4854 struct task_struct *g, *p;
4856 #if (BITS_PER_LONG == 32)
4859 printk(" task PC stack pid father child younger older\n");
4863 printk(" task PC stack pid father child younger older\n");
4865 read_lock(&tasklist_lock);
4866 do_each_thread(g, p) {
4868 * reset the NMI-timeout, listing all files on a slow
4869 * console might take alot of time:
4871 touch_nmi_watchdog();
4872 if (p->state & state_filter)
4874 } while_each_thread(g, p);
4876 read_unlock(&tasklist_lock);
4878 * Only show locks if all tasks are dumped:
4880 if (state_filter == -1)
4881 debug_show_all_locks();
4885 * init_idle - set up an idle thread for a given CPU
4886 * @idle: task in question
4887 * @cpu: cpu the idle task belongs to
4889 * NOTE: this function does not set the idle thread's NEED_RESCHED
4890 * flag, to make booting more robust.
4892 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4894 struct rq *rq = cpu_rq(cpu);
4895 unsigned long flags;
4897 idle->timestamp = sched_clock();
4898 idle->sleep_avg = 0;
4900 idle->prio = idle->normal_prio = MAX_PRIO;
4901 idle->state = TASK_RUNNING;
4902 idle->cpus_allowed = cpumask_of_cpu(cpu);
4903 set_task_cpu(idle, cpu);
4905 spin_lock_irqsave(&rq->lock, flags);
4906 rq->curr = rq->idle = idle;
4907 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4910 spin_unlock_irqrestore(&rq->lock, flags);
4912 /* Set the preempt count _outside_ the spinlocks! */
4913 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4914 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4916 task_thread_info(idle)->preempt_count = 0;
4921 * In a system that switches off the HZ timer nohz_cpu_mask
4922 * indicates which cpus entered this state. This is used
4923 * in the rcu update to wait only for active cpus. For system
4924 * which do not switch off the HZ timer nohz_cpu_mask should
4925 * always be CPU_MASK_NONE.
4927 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4931 * This is how migration works:
4933 * 1) we queue a struct migration_req structure in the source CPU's
4934 * runqueue and wake up that CPU's migration thread.
4935 * 2) we down() the locked semaphore => thread blocks.
4936 * 3) migration thread wakes up (implicitly it forces the migrated
4937 * thread off the CPU)
4938 * 4) it gets the migration request and checks whether the migrated
4939 * task is still in the wrong runqueue.
4940 * 5) if it's in the wrong runqueue then the migration thread removes
4941 * it and puts it into the right queue.
4942 * 6) migration thread up()s the semaphore.
4943 * 7) we wake up and the migration is done.
4947 * Change a given task's CPU affinity. Migrate the thread to a
4948 * proper CPU and schedule it away if the CPU it's executing on
4949 * is removed from the allowed bitmask.
4951 * NOTE: the caller must have a valid reference to the task, the
4952 * task must not exit() & deallocate itself prematurely. The
4953 * call is not atomic; no spinlocks may be held.
4955 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4957 struct migration_req req;
4958 unsigned long flags;
4962 rq = task_rq_lock(p, &flags);
4963 if (!cpus_intersects(new_mask, cpu_online_map)) {
4968 p->cpus_allowed = new_mask;
4969 /* Can the task run on the task's current CPU? If so, we're done */
4970 if (cpu_isset(task_cpu(p), new_mask))
4973 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4974 /* Need help from migration thread: drop lock and wait. */
4975 task_rq_unlock(rq, &flags);
4976 wake_up_process(rq->migration_thread);
4977 wait_for_completion(&req.done);
4978 tlb_migrate_finish(p->mm);
4982 task_rq_unlock(rq, &flags);
4986 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4989 * Move (not current) task off this cpu, onto dest cpu. We're doing
4990 * this because either it can't run here any more (set_cpus_allowed()
4991 * away from this CPU, or CPU going down), or because we're
4992 * attempting to rebalance this task on exec (sched_exec).
4994 * So we race with normal scheduler movements, but that's OK, as long
4995 * as the task is no longer on this CPU.
4997 * Returns non-zero if task was successfully migrated.
4999 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5001 struct rq *rq_dest, *rq_src;
5004 if (unlikely(cpu_is_offline(dest_cpu)))
5007 rq_src = cpu_rq(src_cpu);
5008 rq_dest = cpu_rq(dest_cpu);
5010 double_rq_lock(rq_src, rq_dest);
5011 /* Already moved. */
5012 if (task_cpu(p) != src_cpu)
5014 /* Affinity changed (again). */
5015 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5018 set_task_cpu(p, dest_cpu);
5021 * Sync timestamp with rq_dest's before activating.
5022 * The same thing could be achieved by doing this step
5023 * afterwards, and pretending it was a local activate.
5024 * This way is cleaner and logically correct.
5026 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
5027 + rq_dest->timestamp_last_tick;
5028 deactivate_task(p, rq_src);
5029 __activate_task(p, rq_dest);
5030 if (TASK_PREEMPTS_CURR(p, rq_dest))
5031 resched_task(rq_dest->curr);
5035 double_rq_unlock(rq_src, rq_dest);
5040 * migration_thread - this is a highprio system thread that performs
5041 * thread migration by bumping thread off CPU then 'pushing' onto
5044 static int migration_thread(void *data)
5046 int cpu = (long)data;
5050 BUG_ON(rq->migration_thread != current);
5052 set_current_state(TASK_INTERRUPTIBLE);
5053 while (!kthread_should_stop()) {
5054 struct migration_req *req;
5055 struct list_head *head;
5059 spin_lock_irq(&rq->lock);
5061 if (cpu_is_offline(cpu)) {
5062 spin_unlock_irq(&rq->lock);
5066 if (rq->active_balance) {
5067 active_load_balance(rq, cpu);
5068 rq->active_balance = 0;
5071 head = &rq->migration_queue;
5073 if (list_empty(head)) {
5074 spin_unlock_irq(&rq->lock);
5076 set_current_state(TASK_INTERRUPTIBLE);
5079 req = list_entry(head->next, struct migration_req, list);
5080 list_del_init(head->next);
5082 spin_unlock(&rq->lock);
5083 __migrate_task(req->task, cpu, req->dest_cpu);
5086 complete(&req->done);
5088 __set_current_state(TASK_RUNNING);
5092 /* Wait for kthread_stop */
5093 set_current_state(TASK_INTERRUPTIBLE);
5094 while (!kthread_should_stop()) {
5096 set_current_state(TASK_INTERRUPTIBLE);
5098 __set_current_state(TASK_RUNNING);
5102 #ifdef CONFIG_HOTPLUG_CPU
5104 * Figure out where task on dead CPU should go, use force if neccessary.
5105 * NOTE: interrupts should be disabled by the caller
5107 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5109 unsigned long flags;
5116 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5117 cpus_and(mask, mask, p->cpus_allowed);
5118 dest_cpu = any_online_cpu(mask);
5120 /* On any allowed CPU? */
5121 if (dest_cpu == NR_CPUS)
5122 dest_cpu = any_online_cpu(p->cpus_allowed);
5124 /* No more Mr. Nice Guy. */
5125 if (dest_cpu == NR_CPUS) {
5126 rq = task_rq_lock(p, &flags);
5127 cpus_setall(p->cpus_allowed);
5128 dest_cpu = any_online_cpu(p->cpus_allowed);
5129 task_rq_unlock(rq, &flags);
5132 * Don't tell them about moving exiting tasks or
5133 * kernel threads (both mm NULL), since they never
5136 if (p->mm && printk_ratelimit())
5137 printk(KERN_INFO "process %d (%s) no "
5138 "longer affine to cpu%d\n",
5139 p->pid, p->comm, dead_cpu);
5141 if (!__migrate_task(p, dead_cpu, dest_cpu))
5146 * While a dead CPU has no uninterruptible tasks queued at this point,
5147 * it might still have a nonzero ->nr_uninterruptible counter, because
5148 * for performance reasons the counter is not stricly tracking tasks to
5149 * their home CPUs. So we just add the counter to another CPU's counter,
5150 * to keep the global sum constant after CPU-down:
5152 static void migrate_nr_uninterruptible(struct rq *rq_src)
5154 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5155 unsigned long flags;
5157 local_irq_save(flags);
5158 double_rq_lock(rq_src, rq_dest);
5159 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5160 rq_src->nr_uninterruptible = 0;
5161 double_rq_unlock(rq_src, rq_dest);
5162 local_irq_restore(flags);
5165 /* Run through task list and migrate tasks from the dead cpu. */
5166 static void migrate_live_tasks(int src_cpu)
5168 struct task_struct *p, *t;
5170 write_lock_irq(&tasklist_lock);
5172 do_each_thread(t, p) {
5176 if (task_cpu(p) == src_cpu)
5177 move_task_off_dead_cpu(src_cpu, p);
5178 } while_each_thread(t, p);
5180 write_unlock_irq(&tasklist_lock);
5183 /* Schedules idle task to be the next runnable task on current CPU.
5184 * It does so by boosting its priority to highest possible and adding it to
5185 * the _front_ of the runqueue. Used by CPU offline code.
5187 void sched_idle_next(void)
5189 int this_cpu = smp_processor_id();
5190 struct rq *rq = cpu_rq(this_cpu);
5191 struct task_struct *p = rq->idle;
5192 unsigned long flags;
5194 /* cpu has to be offline */
5195 BUG_ON(cpu_online(this_cpu));
5198 * Strictly not necessary since rest of the CPUs are stopped by now
5199 * and interrupts disabled on the current cpu.
5201 spin_lock_irqsave(&rq->lock, flags);
5203 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5205 /* Add idle task to the _front_ of its priority queue: */
5206 __activate_idle_task(p, rq);
5208 spin_unlock_irqrestore(&rq->lock, flags);
5212 * Ensures that the idle task is using init_mm right before its cpu goes
5215 void idle_task_exit(void)
5217 struct mm_struct *mm = current->active_mm;
5219 BUG_ON(cpu_online(smp_processor_id()));
5222 switch_mm(mm, &init_mm, current);
5226 /* called under rq->lock with disabled interrupts */
5227 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5229 struct rq *rq = cpu_rq(dead_cpu);
5231 /* Must be exiting, otherwise would be on tasklist. */
5232 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5234 /* Cannot have done final schedule yet: would have vanished. */
5235 BUG_ON(p->state == TASK_DEAD);
5240 * Drop lock around migration; if someone else moves it,
5241 * that's OK. No task can be added to this CPU, so iteration is
5243 * NOTE: interrupts should be left disabled --dev@
5245 spin_unlock(&rq->lock);
5246 move_task_off_dead_cpu(dead_cpu, p);
5247 spin_lock(&rq->lock);
5252 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5253 static void migrate_dead_tasks(unsigned int dead_cpu)
5255 struct rq *rq = cpu_rq(dead_cpu);
5256 unsigned int arr, i;
5258 for (arr = 0; arr < 2; arr++) {
5259 for (i = 0; i < MAX_PRIO; i++) {
5260 struct list_head *list = &rq->arrays[arr].queue[i];
5262 while (!list_empty(list))
5263 migrate_dead(dead_cpu, list_entry(list->next,
5264 struct task_struct, run_list));
5268 #endif /* CONFIG_HOTPLUG_CPU */
5271 * migration_call - callback that gets triggered when a CPU is added.
5272 * Here we can start up the necessary migration thread for the new CPU.
5274 static int __cpuinit
5275 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5277 struct task_struct *p;
5278 int cpu = (long)hcpu;
5279 unsigned long flags;
5283 case CPU_UP_PREPARE:
5284 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5287 p->flags |= PF_NOFREEZE;
5288 kthread_bind(p, cpu);
5289 /* Must be high prio: stop_machine expects to yield to it. */
5290 rq = task_rq_lock(p, &flags);
5291 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5292 task_rq_unlock(rq, &flags);
5293 cpu_rq(cpu)->migration_thread = p;
5297 /* Strictly unneccessary, as first user will wake it. */
5298 wake_up_process(cpu_rq(cpu)->migration_thread);
5301 #ifdef CONFIG_HOTPLUG_CPU
5302 case CPU_UP_CANCELED:
5303 if (!cpu_rq(cpu)->migration_thread)
5305 /* Unbind it from offline cpu so it can run. Fall thru. */
5306 kthread_bind(cpu_rq(cpu)->migration_thread,
5307 any_online_cpu(cpu_online_map));
5308 kthread_stop(cpu_rq(cpu)->migration_thread);
5309 cpu_rq(cpu)->migration_thread = NULL;
5313 migrate_live_tasks(cpu);
5315 kthread_stop(rq->migration_thread);
5316 rq->migration_thread = NULL;
5317 /* Idle task back to normal (off runqueue, low prio) */
5318 rq = task_rq_lock(rq->idle, &flags);
5319 deactivate_task(rq->idle, rq);
5320 rq->idle->static_prio = MAX_PRIO;
5321 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5322 migrate_dead_tasks(cpu);
5323 task_rq_unlock(rq, &flags);
5324 migrate_nr_uninterruptible(rq);
5325 BUG_ON(rq->nr_running != 0);
5327 /* No need to migrate the tasks: it was best-effort if
5328 * they didn't do lock_cpu_hotplug(). Just wake up
5329 * the requestors. */
5330 spin_lock_irq(&rq->lock);
5331 while (!list_empty(&rq->migration_queue)) {
5332 struct migration_req *req;
5334 req = list_entry(rq->migration_queue.next,
5335 struct migration_req, list);
5336 list_del_init(&req->list);
5337 complete(&req->done);
5339 spin_unlock_irq(&rq->lock);
5346 /* Register at highest priority so that task migration (migrate_all_tasks)
5347 * happens before everything else.
5349 static struct notifier_block __cpuinitdata migration_notifier = {
5350 .notifier_call = migration_call,
5354 int __init migration_init(void)
5356 void *cpu = (void *)(long)smp_processor_id();
5359 /* Start one for the boot CPU: */
5360 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5361 BUG_ON(err == NOTIFY_BAD);
5362 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5363 register_cpu_notifier(&migration_notifier);
5370 #undef SCHED_DOMAIN_DEBUG
5371 #ifdef SCHED_DOMAIN_DEBUG
5372 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5377 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5381 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5386 struct sched_group *group = sd->groups;
5387 cpumask_t groupmask;
5389 cpumask_scnprintf(str, NR_CPUS, sd->span);
5390 cpus_clear(groupmask);
5393 for (i = 0; i < level + 1; i++)
5395 printk("domain %d: ", level);
5397 if (!(sd->flags & SD_LOAD_BALANCE)) {
5398 printk("does not load-balance\n");
5400 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
5404 printk("span %s\n", str);
5406 if (!cpu_isset(cpu, sd->span))
5407 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
5408 if (!cpu_isset(cpu, group->cpumask))
5409 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
5412 for (i = 0; i < level + 2; i++)
5418 printk(KERN_ERR "ERROR: group is NULL\n");
5422 if (!group->cpu_power) {
5424 printk(KERN_ERR "ERROR: domain->cpu_power not set\n");
5427 if (!cpus_weight(group->cpumask)) {
5429 printk(KERN_ERR "ERROR: empty group\n");
5432 if (cpus_intersects(groupmask, group->cpumask)) {
5434 printk(KERN_ERR "ERROR: repeated CPUs\n");
5437 cpus_or(groupmask, groupmask, group->cpumask);
5439 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5442 group = group->next;
5443 } while (group != sd->groups);
5446 if (!cpus_equal(sd->span, groupmask))
5447 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5453 if (!cpus_subset(groupmask, sd->span))
5454 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
5460 # define sched_domain_debug(sd, cpu) do { } while (0)
5463 static int sd_degenerate(struct sched_domain *sd)
5465 if (cpus_weight(sd->span) == 1)
5468 /* Following flags need at least 2 groups */
5469 if (sd->flags & (SD_LOAD_BALANCE |
5470 SD_BALANCE_NEWIDLE |
5474 SD_SHARE_PKG_RESOURCES)) {
5475 if (sd->groups != sd->groups->next)
5479 /* Following flags don't use groups */
5480 if (sd->flags & (SD_WAKE_IDLE |
5489 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5491 unsigned long cflags = sd->flags, pflags = parent->flags;
5493 if (sd_degenerate(parent))
5496 if (!cpus_equal(sd->span, parent->span))
5499 /* Does parent contain flags not in child? */
5500 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5501 if (cflags & SD_WAKE_AFFINE)
5502 pflags &= ~SD_WAKE_BALANCE;
5503 /* Flags needing groups don't count if only 1 group in parent */
5504 if (parent->groups == parent->groups->next) {
5505 pflags &= ~(SD_LOAD_BALANCE |
5506 SD_BALANCE_NEWIDLE |
5510 SD_SHARE_PKG_RESOURCES);
5512 if (~cflags & pflags)
5519 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5520 * hold the hotplug lock.
5522 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5524 struct rq *rq = cpu_rq(cpu);
5525 struct sched_domain *tmp;
5527 /* Remove the sched domains which do not contribute to scheduling. */
5528 for (tmp = sd; tmp; tmp = tmp->parent) {
5529 struct sched_domain *parent = tmp->parent;
5532 if (sd_parent_degenerate(tmp, parent)) {
5533 tmp->parent = parent->parent;
5535 parent->parent->child = tmp;
5539 if (sd && sd_degenerate(sd)) {
5545 sched_domain_debug(sd, cpu);
5547 rcu_assign_pointer(rq->sd, sd);
5550 /* cpus with isolated domains */
5551 static cpumask_t __cpuinitdata cpu_isolated_map = CPU_MASK_NONE;
5553 /* Setup the mask of cpus configured for isolated domains */
5554 static int __init isolated_cpu_setup(char *str)
5556 int ints[NR_CPUS], i;
5558 str = get_options(str, ARRAY_SIZE(ints), ints);
5559 cpus_clear(cpu_isolated_map);
5560 for (i = 1; i <= ints[0]; i++)
5561 if (ints[i] < NR_CPUS)
5562 cpu_set(ints[i], cpu_isolated_map);
5566 __setup ("isolcpus=", isolated_cpu_setup);
5569 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5570 * to a function which identifies what group(along with sched group) a CPU
5571 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5572 * (due to the fact that we keep track of groups covered with a cpumask_t).
5574 * init_sched_build_groups will build a circular linked list of the groups
5575 * covered by the given span, and will set each group's ->cpumask correctly,
5576 * and ->cpu_power to 0.
5579 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5580 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5581 struct sched_group **sg))
5583 struct sched_group *first = NULL, *last = NULL;
5584 cpumask_t covered = CPU_MASK_NONE;
5587 for_each_cpu_mask(i, span) {
5588 struct sched_group *sg;
5589 int group = group_fn(i, cpu_map, &sg);
5592 if (cpu_isset(i, covered))
5595 sg->cpumask = CPU_MASK_NONE;
5598 for_each_cpu_mask(j, span) {
5599 if (group_fn(j, cpu_map, NULL) != group)
5602 cpu_set(j, covered);
5603 cpu_set(j, sg->cpumask);
5614 #define SD_NODES_PER_DOMAIN 16
5617 * Self-tuning task migration cost measurement between source and target CPUs.
5619 * This is done by measuring the cost of manipulating buffers of varying
5620 * sizes. For a given buffer-size here are the steps that are taken:
5622 * 1) the source CPU reads+dirties a shared buffer
5623 * 2) the target CPU reads+dirties the same shared buffer
5625 * We measure how long they take, in the following 4 scenarios:
5627 * - source: CPU1, target: CPU2 | cost1
5628 * - source: CPU2, target: CPU1 | cost2
5629 * - source: CPU1, target: CPU1 | cost3
5630 * - source: CPU2, target: CPU2 | cost4
5632 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5633 * the cost of migration.
5635 * We then start off from a small buffer-size and iterate up to larger
5636 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5637 * doing a maximum search for the cost. (The maximum cost for a migration
5638 * normally occurs when the working set size is around the effective cache
5641 #define SEARCH_SCOPE 2
5642 #define MIN_CACHE_SIZE (64*1024U)
5643 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5644 #define ITERATIONS 1
5645 #define SIZE_THRESH 130
5646 #define COST_THRESH 130
5649 * The migration cost is a function of 'domain distance'. Domain
5650 * distance is the number of steps a CPU has to iterate down its
5651 * domain tree to share a domain with the other CPU. The farther
5652 * two CPUs are from each other, the larger the distance gets.
5654 * Note that we use the distance only to cache measurement results,
5655 * the distance value is not used numerically otherwise. When two
5656 * CPUs have the same distance it is assumed that the migration
5657 * cost is the same. (this is a simplification but quite practical)
5659 #define MAX_DOMAIN_DISTANCE 32
5661 static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5662 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] =
5664 * Architectures may override the migration cost and thus avoid
5665 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5666 * virtualized hardware:
5668 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5669 CONFIG_DEFAULT_MIGRATION_COST
5676 * Allow override of migration cost - in units of microseconds.
5677 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5678 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5680 static int __init migration_cost_setup(char *str)
5682 int ints[MAX_DOMAIN_DISTANCE+1], i;
5684 str = get_options(str, ARRAY_SIZE(ints), ints);
5686 printk("#ints: %d\n", ints[0]);
5687 for (i = 1; i <= ints[0]; i++) {
5688 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5689 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5694 __setup ("migration_cost=", migration_cost_setup);
5697 * Global multiplier (divisor) for migration-cutoff values,
5698 * in percentiles. E.g. use a value of 150 to get 1.5 times
5699 * longer cache-hot cutoff times.
5701 * (We scale it from 100 to 128 to long long handling easier.)
5704 #define MIGRATION_FACTOR_SCALE 128
5706 static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5708 static int __init setup_migration_factor(char *str)
5710 get_option(&str, &migration_factor);
5711 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5715 __setup("migration_factor=", setup_migration_factor);
5718 * Estimated distance of two CPUs, measured via the number of domains
5719 * we have to pass for the two CPUs to be in the same span:
5721 static unsigned long domain_distance(int cpu1, int cpu2)
5723 unsigned long distance = 0;
5724 struct sched_domain *sd;
5726 for_each_domain(cpu1, sd) {
5727 WARN_ON(!cpu_isset(cpu1, sd->span));
5728 if (cpu_isset(cpu2, sd->span))
5732 if (distance >= MAX_DOMAIN_DISTANCE) {
5734 distance = MAX_DOMAIN_DISTANCE-1;
5740 static unsigned int migration_debug;
5742 static int __init setup_migration_debug(char *str)
5744 get_option(&str, &migration_debug);
5748 __setup("migration_debug=", setup_migration_debug);
5751 * Maximum cache-size that the scheduler should try to measure.
5752 * Architectures with larger caches should tune this up during
5753 * bootup. Gets used in the domain-setup code (i.e. during SMP
5756 unsigned int max_cache_size;
5758 static int __init setup_max_cache_size(char *str)
5760 get_option(&str, &max_cache_size);
5764 __setup("max_cache_size=", setup_max_cache_size);
5767 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5768 * is the operation that is timed, so we try to generate unpredictable
5769 * cachemisses that still end up filling the L2 cache:
5771 static void touch_cache(void *__cache, unsigned long __size)
5773 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5775 unsigned long *cache = __cache;
5778 for (i = 0; i < size/6; i += 8) {
5781 case 1: cache[size-1-i]++;
5782 case 2: cache[chunk1-i]++;
5783 case 3: cache[chunk1+i]++;
5784 case 4: cache[chunk2-i]++;
5785 case 5: cache[chunk2+i]++;
5791 * Measure the cache-cost of one task migration. Returns in units of nsec.
5793 static unsigned long long
5794 measure_one(void *cache, unsigned long size, int source, int target)
5796 cpumask_t mask, saved_mask;
5797 unsigned long long t0, t1, t2, t3, cost;
5799 saved_mask = current->cpus_allowed;
5802 * Flush source caches to RAM and invalidate them:
5807 * Migrate to the source CPU:
5809 mask = cpumask_of_cpu(source);
5810 set_cpus_allowed(current, mask);
5811 WARN_ON(smp_processor_id() != source);
5814 * Dirty the working set:
5817 touch_cache(cache, size);
5821 * Migrate to the target CPU, dirty the L2 cache and access
5822 * the shared buffer. (which represents the working set
5823 * of a migrated task.)
5825 mask = cpumask_of_cpu(target);
5826 set_cpus_allowed(current, mask);
5827 WARN_ON(smp_processor_id() != target);
5830 touch_cache(cache, size);
5833 cost = t1-t0 + t3-t2;
5835 if (migration_debug >= 2)
5836 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5837 source, target, t1-t0, t1-t0, t3-t2, cost);
5839 * Flush target caches to RAM and invalidate them:
5843 set_cpus_allowed(current, saved_mask);
5849 * Measure a series of task migrations and return the average
5850 * result. Since this code runs early during bootup the system
5851 * is 'undisturbed' and the average latency makes sense.
5853 * The algorithm in essence auto-detects the relevant cache-size,
5854 * so it will properly detect different cachesizes for different
5855 * cache-hierarchies, depending on how the CPUs are connected.
5857 * Architectures can prime the upper limit of the search range via
5858 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5860 static unsigned long long
5861 measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5863 unsigned long long cost1, cost2;
5867 * Measure the migration cost of 'size' bytes, over an
5868 * average of 10 runs:
5870 * (We perturb the cache size by a small (0..4k)
5871 * value to compensate size/alignment related artifacts.
5872 * We also subtract the cost of the operation done on
5878 * dry run, to make sure we start off cache-cold on cpu1,
5879 * and to get any vmalloc pagefaults in advance:
5881 measure_one(cache, size, cpu1, cpu2);
5882 for (i = 0; i < ITERATIONS; i++)
5883 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5885 measure_one(cache, size, cpu2, cpu1);
5886 for (i = 0; i < ITERATIONS; i++)
5887 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5890 * (We measure the non-migrating [cached] cost on both
5891 * cpu1 and cpu2, to handle CPUs with different speeds)
5895 measure_one(cache, size, cpu1, cpu1);
5896 for (i = 0; i < ITERATIONS; i++)
5897 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5899 measure_one(cache, size, cpu2, cpu2);
5900 for (i = 0; i < ITERATIONS; i++)
5901 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5904 * Get the per-iteration migration cost:
5906 do_div(cost1, 2*ITERATIONS);
5907 do_div(cost2, 2*ITERATIONS);
5909 return cost1 - cost2;
5912 static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5914 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5915 unsigned int max_size, size, size_found = 0;
5916 long long cost = 0, prev_cost;
5920 * Search from max_cache_size*5 down to 64K - the real relevant
5921 * cachesize has to lie somewhere inbetween.
5923 if (max_cache_size) {
5924 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5925 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5928 * Since we have no estimation about the relevant
5931 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5932 size = MIN_CACHE_SIZE;
5935 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5936 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5941 * Allocate the working set:
5943 cache = vmalloc(max_size);
5945 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5946 return 1000000; /* return 1 msec on very small boxen */
5949 while (size <= max_size) {
5951 cost = measure_cost(cpu1, cpu2, cache, size);
5957 if (max_cost < cost) {
5963 * Calculate average fluctuation, we use this to prevent
5964 * noise from triggering an early break out of the loop:
5966 fluct = abs(cost - prev_cost);
5967 avg_fluct = (avg_fluct + fluct)/2;
5969 if (migration_debug)
5970 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5972 (long)cost / 1000000,
5973 ((long)cost / 100000) % 10,
5974 (long)max_cost / 1000000,
5975 ((long)max_cost / 100000) % 10,
5976 domain_distance(cpu1, cpu2),
5980 * If we iterated at least 20% past the previous maximum,
5981 * and the cost has dropped by more than 20% already,
5982 * (taking fluctuations into account) then we assume to
5983 * have found the maximum and break out of the loop early:
5985 if (size_found && (size*100 > size_found*SIZE_THRESH))
5986 if (cost+avg_fluct <= 0 ||
5987 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5989 if (migration_debug)
5990 printk("-> found max.\n");
5994 * Increase the cachesize in 10% steps:
5996 size = size * 10 / 9;
5999 if (migration_debug)
6000 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6001 cpu1, cpu2, size_found, max_cost);
6006 * A task is considered 'cache cold' if at least 2 times
6007 * the worst-case cost of migration has passed.
6009 * (this limit is only listened to if the load-balancing
6010 * situation is 'nice' - if there is a large imbalance we
6011 * ignore it for the sake of CPU utilization and
6012 * processing fairness.)
6014 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
6017 static void calibrate_migration_costs(const cpumask_t *cpu_map)
6019 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
6020 unsigned long j0, j1, distance, max_distance = 0;
6021 struct sched_domain *sd;
6026 * First pass - calculate the cacheflush times:
6028 for_each_cpu_mask(cpu1, *cpu_map) {
6029 for_each_cpu_mask(cpu2, *cpu_map) {
6032 distance = domain_distance(cpu1, cpu2);
6033 max_distance = max(max_distance, distance);
6035 * No result cached yet?
6037 if (migration_cost[distance] == -1LL)
6038 migration_cost[distance] =
6039 measure_migration_cost(cpu1, cpu2);
6043 * Second pass - update the sched domain hierarchy with
6044 * the new cache-hot-time estimations:
6046 for_each_cpu_mask(cpu, *cpu_map) {
6048 for_each_domain(cpu, sd) {
6049 sd->cache_hot_time = migration_cost[distance];
6056 if (migration_debug)
6057 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6065 if (system_state == SYSTEM_BOOTING) {
6066 if (num_online_cpus() > 1) {
6067 printk("migration_cost=");
6068 for (distance = 0; distance <= max_distance; distance++) {
6071 printk("%ld", (long)migration_cost[distance] / 1000);
6077 if (migration_debug)
6078 printk("migration: %ld seconds\n", (j1-j0)/HZ);
6081 * Move back to the original CPU. NUMA-Q gets confused
6082 * if we migrate to another quad during bootup.
6084 if (raw_smp_processor_id() != orig_cpu) {
6085 cpumask_t mask = cpumask_of_cpu(orig_cpu),
6086 saved_mask = current->cpus_allowed;
6088 set_cpus_allowed(current, mask);
6089 set_cpus_allowed(current, saved_mask);
6096 * find_next_best_node - find the next node to include in a sched_domain
6097 * @node: node whose sched_domain we're building
6098 * @used_nodes: nodes already in the sched_domain
6100 * Find the next node to include in a given scheduling domain. Simply
6101 * finds the closest node not already in the @used_nodes map.
6103 * Should use nodemask_t.
6105 static int find_next_best_node(int node, unsigned long *used_nodes)
6107 int i, n, val, min_val, best_node = 0;
6111 for (i = 0; i < MAX_NUMNODES; i++) {
6112 /* Start at @node */
6113 n = (node + i) % MAX_NUMNODES;
6115 if (!nr_cpus_node(n))
6118 /* Skip already used nodes */
6119 if (test_bit(n, used_nodes))
6122 /* Simple min distance search */
6123 val = node_distance(node, n);
6125 if (val < min_val) {
6131 set_bit(best_node, used_nodes);
6136 * sched_domain_node_span - get a cpumask for a node's sched_domain
6137 * @node: node whose cpumask we're constructing
6138 * @size: number of nodes to include in this span
6140 * Given a node, construct a good cpumask for its sched_domain to span. It
6141 * should be one that prevents unnecessary balancing, but also spreads tasks
6144 static cpumask_t sched_domain_node_span(int node)
6146 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6147 cpumask_t span, nodemask;
6151 bitmap_zero(used_nodes, MAX_NUMNODES);
6153 nodemask = node_to_cpumask(node);
6154 cpus_or(span, span, nodemask);
6155 set_bit(node, used_nodes);
6157 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6158 int next_node = find_next_best_node(node, used_nodes);
6160 nodemask = node_to_cpumask(next_node);
6161 cpus_or(span, span, nodemask);
6168 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6171 * SMT sched-domains:
6173 #ifdef CONFIG_SCHED_SMT
6174 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6175 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6177 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6178 struct sched_group **sg)
6181 *sg = &per_cpu(sched_group_cpus, cpu);
6187 * multi-core sched-domains:
6189 #ifdef CONFIG_SCHED_MC
6190 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6191 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6194 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6195 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6196 struct sched_group **sg)
6199 cpumask_t mask = cpu_sibling_map[cpu];
6200 cpus_and(mask, mask, *cpu_map);
6201 group = first_cpu(mask);
6203 *sg = &per_cpu(sched_group_core, group);
6206 #elif defined(CONFIG_SCHED_MC)
6207 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6208 struct sched_group **sg)
6211 *sg = &per_cpu(sched_group_core, cpu);
6216 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6217 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6219 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6220 struct sched_group **sg)
6223 #ifdef CONFIG_SCHED_MC
6224 cpumask_t mask = cpu_coregroup_map(cpu);
6225 cpus_and(mask, mask, *cpu_map);
6226 group = first_cpu(mask);
6227 #elif defined(CONFIG_SCHED_SMT)
6228 cpumask_t mask = cpu_sibling_map[cpu];
6229 cpus_and(mask, mask, *cpu_map);
6230 group = first_cpu(mask);
6235 *sg = &per_cpu(sched_group_phys, group);
6241 * The init_sched_build_groups can't handle what we want to do with node
6242 * groups, so roll our own. Now each node has its own list of groups which
6243 * gets dynamically allocated.
6245 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6246 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6248 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6249 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6251 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6252 struct sched_group **sg)
6254 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6257 cpus_and(nodemask, nodemask, *cpu_map);
6258 group = first_cpu(nodemask);
6261 *sg = &per_cpu(sched_group_allnodes, group);
6265 static void init_numa_sched_groups_power(struct sched_group *group_head)
6267 struct sched_group *sg = group_head;
6273 for_each_cpu_mask(j, sg->cpumask) {
6274 struct sched_domain *sd;
6276 sd = &per_cpu(phys_domains, j);
6277 if (j != first_cpu(sd->groups->cpumask)) {
6279 * Only add "power" once for each
6285 sg->cpu_power += sd->groups->cpu_power;
6288 if (sg != group_head)
6294 /* Free memory allocated for various sched_group structures */
6295 static void free_sched_groups(const cpumask_t *cpu_map)
6299 for_each_cpu_mask(cpu, *cpu_map) {
6300 struct sched_group **sched_group_nodes
6301 = sched_group_nodes_bycpu[cpu];
6303 if (!sched_group_nodes)
6306 for (i = 0; i < MAX_NUMNODES; i++) {
6307 cpumask_t nodemask = node_to_cpumask(i);
6308 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6310 cpus_and(nodemask, nodemask, *cpu_map);
6311 if (cpus_empty(nodemask))
6321 if (oldsg != sched_group_nodes[i])
6324 kfree(sched_group_nodes);
6325 sched_group_nodes_bycpu[cpu] = NULL;
6329 static void free_sched_groups(const cpumask_t *cpu_map)
6335 * Initialize sched groups cpu_power.
6337 * cpu_power indicates the capacity of sched group, which is used while
6338 * distributing the load between different sched groups in a sched domain.
6339 * Typically cpu_power for all the groups in a sched domain will be same unless
6340 * there are asymmetries in the topology. If there are asymmetries, group
6341 * having more cpu_power will pickup more load compared to the group having
6344 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6345 * the maximum number of tasks a group can handle in the presence of other idle
6346 * or lightly loaded groups in the same sched domain.
6348 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6350 struct sched_domain *child;
6351 struct sched_group *group;
6353 WARN_ON(!sd || !sd->groups);
6355 if (cpu != first_cpu(sd->groups->cpumask))
6361 * For perf policy, if the groups in child domain share resources
6362 * (for example cores sharing some portions of the cache hierarchy
6363 * or SMT), then set this domain groups cpu_power such that each group
6364 * can handle only one task, when there are other idle groups in the
6365 * same sched domain.
6367 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6369 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6370 sd->groups->cpu_power = SCHED_LOAD_SCALE;
6374 sd->groups->cpu_power = 0;
6377 * add cpu_power of each child group to this groups cpu_power
6379 group = child->groups;
6381 sd->groups->cpu_power += group->cpu_power;
6382 group = group->next;
6383 } while (group != child->groups);
6387 * Build sched domains for a given set of cpus and attach the sched domains
6388 * to the individual cpus
6390 static int build_sched_domains(const cpumask_t *cpu_map)
6393 struct sched_domain *sd;
6395 struct sched_group **sched_group_nodes = NULL;
6396 int sd_allnodes = 0;
6399 * Allocate the per-node list of sched groups
6401 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6403 if (!sched_group_nodes) {
6404 printk(KERN_WARNING "Can not alloc sched group node list\n");
6407 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6411 * Set up domains for cpus specified by the cpu_map.
6413 for_each_cpu_mask(i, *cpu_map) {
6414 struct sched_domain *sd = NULL, *p;
6415 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6417 cpus_and(nodemask, nodemask, *cpu_map);
6420 if (cpus_weight(*cpu_map)
6421 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6422 sd = &per_cpu(allnodes_domains, i);
6423 *sd = SD_ALLNODES_INIT;
6424 sd->span = *cpu_map;
6425 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6431 sd = &per_cpu(node_domains, i);
6433 sd->span = sched_domain_node_span(cpu_to_node(i));
6437 cpus_and(sd->span, sd->span, *cpu_map);
6441 sd = &per_cpu(phys_domains, i);
6443 sd->span = nodemask;
6447 cpu_to_phys_group(i, cpu_map, &sd->groups);
6449 #ifdef CONFIG_SCHED_MC
6451 sd = &per_cpu(core_domains, i);
6453 sd->span = cpu_coregroup_map(i);
6454 cpus_and(sd->span, sd->span, *cpu_map);
6457 cpu_to_core_group(i, cpu_map, &sd->groups);
6460 #ifdef CONFIG_SCHED_SMT
6462 sd = &per_cpu(cpu_domains, i);
6463 *sd = SD_SIBLING_INIT;
6464 sd->span = cpu_sibling_map[i];
6465 cpus_and(sd->span, sd->span, *cpu_map);
6468 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6472 #ifdef CONFIG_SCHED_SMT
6473 /* Set up CPU (sibling) groups */
6474 for_each_cpu_mask(i, *cpu_map) {
6475 cpumask_t this_sibling_map = cpu_sibling_map[i];
6476 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6477 if (i != first_cpu(this_sibling_map))
6480 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6484 #ifdef CONFIG_SCHED_MC
6485 /* Set up multi-core groups */
6486 for_each_cpu_mask(i, *cpu_map) {
6487 cpumask_t this_core_map = cpu_coregroup_map(i);
6488 cpus_and(this_core_map, this_core_map, *cpu_map);
6489 if (i != first_cpu(this_core_map))
6491 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6496 /* Set up physical groups */
6497 for (i = 0; i < MAX_NUMNODES; i++) {
6498 cpumask_t nodemask = node_to_cpumask(i);
6500 cpus_and(nodemask, nodemask, *cpu_map);
6501 if (cpus_empty(nodemask))
6504 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6508 /* Set up node groups */
6510 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6512 for (i = 0; i < MAX_NUMNODES; i++) {
6513 /* Set up node groups */
6514 struct sched_group *sg, *prev;
6515 cpumask_t nodemask = node_to_cpumask(i);
6516 cpumask_t domainspan;
6517 cpumask_t covered = CPU_MASK_NONE;
6520 cpus_and(nodemask, nodemask, *cpu_map);
6521 if (cpus_empty(nodemask)) {
6522 sched_group_nodes[i] = NULL;
6526 domainspan = sched_domain_node_span(i);
6527 cpus_and(domainspan, domainspan, *cpu_map);
6529 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6531 printk(KERN_WARNING "Can not alloc domain group for "
6535 sched_group_nodes[i] = sg;
6536 for_each_cpu_mask(j, nodemask) {
6537 struct sched_domain *sd;
6538 sd = &per_cpu(node_domains, j);
6542 sg->cpumask = nodemask;
6544 cpus_or(covered, covered, nodemask);
6547 for (j = 0; j < MAX_NUMNODES; j++) {
6548 cpumask_t tmp, notcovered;
6549 int n = (i + j) % MAX_NUMNODES;
6551 cpus_complement(notcovered, covered);
6552 cpus_and(tmp, notcovered, *cpu_map);
6553 cpus_and(tmp, tmp, domainspan);
6554 if (cpus_empty(tmp))
6557 nodemask = node_to_cpumask(n);
6558 cpus_and(tmp, tmp, nodemask);
6559 if (cpus_empty(tmp))
6562 sg = kmalloc_node(sizeof(struct sched_group),
6566 "Can not alloc domain group for node %d\n", j);
6571 sg->next = prev->next;
6572 cpus_or(covered, covered, tmp);
6579 /* Calculate CPU power for physical packages and nodes */
6580 #ifdef CONFIG_SCHED_SMT
6581 for_each_cpu_mask(i, *cpu_map) {
6582 sd = &per_cpu(cpu_domains, i);
6583 init_sched_groups_power(i, sd);
6586 #ifdef CONFIG_SCHED_MC
6587 for_each_cpu_mask(i, *cpu_map) {
6588 sd = &per_cpu(core_domains, i);
6589 init_sched_groups_power(i, sd);
6593 for_each_cpu_mask(i, *cpu_map) {
6594 sd = &per_cpu(phys_domains, i);
6595 init_sched_groups_power(i, sd);
6599 for (i = 0; i < MAX_NUMNODES; i++)
6600 init_numa_sched_groups_power(sched_group_nodes[i]);
6603 struct sched_group *sg;
6605 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6606 init_numa_sched_groups_power(sg);
6610 /* Attach the domains */
6611 for_each_cpu_mask(i, *cpu_map) {
6612 struct sched_domain *sd;
6613 #ifdef CONFIG_SCHED_SMT
6614 sd = &per_cpu(cpu_domains, i);
6615 #elif defined(CONFIG_SCHED_MC)
6616 sd = &per_cpu(core_domains, i);
6618 sd = &per_cpu(phys_domains, i);
6620 cpu_attach_domain(sd, i);
6623 * Tune cache-hot values:
6625 calibrate_migration_costs(cpu_map);
6631 free_sched_groups(cpu_map);
6636 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6638 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6640 cpumask_t cpu_default_map;
6644 * Setup mask for cpus without special case scheduling requirements.
6645 * For now this just excludes isolated cpus, but could be used to
6646 * exclude other special cases in the future.
6648 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6650 err = build_sched_domains(&cpu_default_map);
6655 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6657 free_sched_groups(cpu_map);
6661 * Detach sched domains from a group of cpus specified in cpu_map
6662 * These cpus will now be attached to the NULL domain
6664 static void detach_destroy_domains(const cpumask_t *cpu_map)
6668 for_each_cpu_mask(i, *cpu_map)
6669 cpu_attach_domain(NULL, i);
6670 synchronize_sched();
6671 arch_destroy_sched_domains(cpu_map);
6675 * Partition sched domains as specified by the cpumasks below.
6676 * This attaches all cpus from the cpumasks to the NULL domain,
6677 * waits for a RCU quiescent period, recalculates sched
6678 * domain information and then attaches them back to the
6679 * correct sched domains
6680 * Call with hotplug lock held
6682 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6684 cpumask_t change_map;
6687 cpus_and(*partition1, *partition1, cpu_online_map);
6688 cpus_and(*partition2, *partition2, cpu_online_map);
6689 cpus_or(change_map, *partition1, *partition2);
6691 /* Detach sched domains from all of the affected cpus */
6692 detach_destroy_domains(&change_map);
6693 if (!cpus_empty(*partition1))
6694 err = build_sched_domains(partition1);
6695 if (!err && !cpus_empty(*partition2))
6696 err = build_sched_domains(partition2);
6701 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6702 int arch_reinit_sched_domains(void)
6707 detach_destroy_domains(&cpu_online_map);
6708 err = arch_init_sched_domains(&cpu_online_map);
6709 unlock_cpu_hotplug();
6714 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6718 if (buf[0] != '0' && buf[0] != '1')
6722 sched_smt_power_savings = (buf[0] == '1');
6724 sched_mc_power_savings = (buf[0] == '1');
6726 ret = arch_reinit_sched_domains();
6728 return ret ? ret : count;
6731 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6735 #ifdef CONFIG_SCHED_SMT
6737 err = sysfs_create_file(&cls->kset.kobj,
6738 &attr_sched_smt_power_savings.attr);
6740 #ifdef CONFIG_SCHED_MC
6741 if (!err && mc_capable())
6742 err = sysfs_create_file(&cls->kset.kobj,
6743 &attr_sched_mc_power_savings.attr);
6749 #ifdef CONFIG_SCHED_MC
6750 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6752 return sprintf(page, "%u\n", sched_mc_power_savings);
6754 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6755 const char *buf, size_t count)
6757 return sched_power_savings_store(buf, count, 0);
6759 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6760 sched_mc_power_savings_store);
6763 #ifdef CONFIG_SCHED_SMT
6764 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6766 return sprintf(page, "%u\n", sched_smt_power_savings);
6768 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6769 const char *buf, size_t count)
6771 return sched_power_savings_store(buf, count, 1);
6773 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6774 sched_smt_power_savings_store);
6778 * Force a reinitialization of the sched domains hierarchy. The domains
6779 * and groups cannot be updated in place without racing with the balancing
6780 * code, so we temporarily attach all running cpus to the NULL domain
6781 * which will prevent rebalancing while the sched domains are recalculated.
6783 static int update_sched_domains(struct notifier_block *nfb,
6784 unsigned long action, void *hcpu)
6787 case CPU_UP_PREPARE:
6788 case CPU_DOWN_PREPARE:
6789 detach_destroy_domains(&cpu_online_map);
6792 case CPU_UP_CANCELED:
6793 case CPU_DOWN_FAILED:
6797 * Fall through and re-initialise the domains.
6804 /* The hotplug lock is already held by cpu_up/cpu_down */
6805 arch_init_sched_domains(&cpu_online_map);
6810 void __init sched_init_smp(void)
6812 cpumask_t non_isolated_cpus;
6815 arch_init_sched_domains(&cpu_online_map);
6816 cpus_andnot(non_isolated_cpus, cpu_online_map, cpu_isolated_map);
6817 if (cpus_empty(non_isolated_cpus))
6818 cpu_set(smp_processor_id(), non_isolated_cpus);
6819 unlock_cpu_hotplug();
6820 /* XXX: Theoretical race here - CPU may be hotplugged now */
6821 hotcpu_notifier(update_sched_domains, 0);
6823 /* Move init over to a non-isolated CPU */
6824 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6828 void __init sched_init_smp(void)
6831 #endif /* CONFIG_SMP */
6833 int in_sched_functions(unsigned long addr)
6835 /* Linker adds these: start and end of __sched functions */
6836 extern char __sched_text_start[], __sched_text_end[];
6838 return in_lock_functions(addr) ||
6839 (addr >= (unsigned long)__sched_text_start
6840 && addr < (unsigned long)__sched_text_end);
6843 void __init sched_init(void)
6847 for_each_possible_cpu(i) {
6848 struct prio_array *array;
6852 spin_lock_init(&rq->lock);
6853 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6855 rq->active = rq->arrays;
6856 rq->expired = rq->arrays + 1;
6857 rq->best_expired_prio = MAX_PRIO;
6861 for (j = 1; j < 3; j++)
6862 rq->cpu_load[j] = 0;
6863 rq->active_balance = 0;
6866 rq->migration_thread = NULL;
6867 INIT_LIST_HEAD(&rq->migration_queue);
6869 atomic_set(&rq->nr_iowait, 0);
6871 for (j = 0; j < 2; j++) {
6872 array = rq->arrays + j;
6873 for (k = 0; k < MAX_PRIO; k++) {
6874 INIT_LIST_HEAD(array->queue + k);
6875 __clear_bit(k, array->bitmap);
6877 // delimiter for bitsearch
6878 __set_bit(MAX_PRIO, array->bitmap);
6882 set_load_weight(&init_task);
6885 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6888 #ifdef CONFIG_RT_MUTEXES
6889 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6893 * The boot idle thread does lazy MMU switching as well:
6895 atomic_inc(&init_mm.mm_count);
6896 enter_lazy_tlb(&init_mm, current);
6899 * Make us the idle thread. Technically, schedule() should not be
6900 * called from this thread, however somewhere below it might be,
6901 * but because we are the idle thread, we just pick up running again
6902 * when this runqueue becomes "idle".
6904 init_idle(current, smp_processor_id());
6907 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6908 void __might_sleep(char *file, int line)
6911 static unsigned long prev_jiffy; /* ratelimiting */
6913 if ((in_atomic() || irqs_disabled()) &&
6914 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6915 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6917 prev_jiffy = jiffies;
6918 printk(KERN_ERR "BUG: sleeping function called from invalid"
6919 " context at %s:%d\n", file, line);
6920 printk("in_atomic():%d, irqs_disabled():%d\n",
6921 in_atomic(), irqs_disabled());
6922 debug_show_held_locks(current);
6927 EXPORT_SYMBOL(__might_sleep);
6930 #ifdef CONFIG_MAGIC_SYSRQ
6931 void normalize_rt_tasks(void)
6933 struct prio_array *array;
6934 struct task_struct *p;
6935 unsigned long flags;
6938 read_lock_irq(&tasklist_lock);
6939 for_each_process(p) {
6943 spin_lock_irqsave(&p->pi_lock, flags);
6944 rq = __task_rq_lock(p);
6948 deactivate_task(p, task_rq(p));
6949 __setscheduler(p, SCHED_NORMAL, 0);
6951 __activate_task(p, task_rq(p));
6952 resched_task(rq->curr);
6955 __task_rq_unlock(rq);
6956 spin_unlock_irqrestore(&p->pi_lock, flags);
6958 read_unlock_irq(&tasklist_lock);
6961 #endif /* CONFIG_MAGIC_SYSRQ */
6965 * These functions are only useful for the IA64 MCA handling.
6967 * They can only be called when the whole system has been
6968 * stopped - every CPU needs to be quiescent, and no scheduling
6969 * activity can take place. Using them for anything else would
6970 * be a serious bug, and as a result, they aren't even visible
6971 * under any other configuration.
6975 * curr_task - return the current task for a given cpu.
6976 * @cpu: the processor in question.
6978 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6980 struct task_struct *curr_task(int cpu)
6982 return cpu_curr(cpu);
6986 * set_curr_task - set the current task for a given cpu.
6987 * @cpu: the processor in question.
6988 * @p: the task pointer to set.
6990 * Description: This function must only be used when non-maskable interrupts
6991 * are serviced on a separate stack. It allows the architecture to switch the
6992 * notion of the current task on a cpu in a non-blocking manner. This function
6993 * must be called with all CPU's synchronized, and interrupts disabled, the
6994 * and caller must save the original value of the current task (see
6995 * curr_task() above) and restore that value before reenabling interrupts and
6996 * re-starting the system.
6998 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7000 void set_curr_task(int cpu, struct task_struct *p)