4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 #define NICE_0_LOAD SCHED_LOAD_SCALE
95 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
98 * These are the 'tuning knobs' of the scheduler:
100 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
101 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
102 * Timeslices get refilled after they expire.
104 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
105 #define DEF_TIMESLICE (100 * HZ / 1000)
106 #define ON_RUNQUEUE_WEIGHT 30
107 #define CHILD_PENALTY 95
108 #define PARENT_PENALTY 100
109 #define EXIT_WEIGHT 3
110 #define PRIO_BONUS_RATIO 25
111 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
112 #define INTERACTIVE_DELTA 2
113 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
114 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
115 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
118 * If a task is 'interactive' then we reinsert it in the active
119 * array after it has expired its current timeslice. (it will not
120 * continue to run immediately, it will still roundrobin with
121 * other interactive tasks.)
123 * This part scales the interactivity limit depending on niceness.
125 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
126 * Here are a few examples of different nice levels:
128 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
129 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
130 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
131 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
132 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
134 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
135 * priority range a task can explore, a value of '1' means the
136 * task is rated interactive.)
138 * Ie. nice +19 tasks can never get 'interactive' enough to be
139 * reinserted into the active array. And only heavily CPU-hog nice -20
140 * tasks will be expired. Default nice 0 tasks are somewhere between,
141 * it takes some effort for them to get interactive, but it's not
145 #define CURRENT_BONUS(p) \
146 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
149 #define GRANULARITY (10 * HZ / 1000 ? : 1)
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
156 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
157 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
160 #define SCALE(v1,v1_max,v2_max) \
161 (v1) * (v2_max) / (v1_max)
164 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
167 #define TASK_INTERACTIVE(p) \
168 ((p)->prio <= (p)->static_prio - DELTA(p))
170 #define INTERACTIVE_SLEEP(p) \
171 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
172 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
174 #define TASK_PREEMPTS_CURR(p, rq) \
175 ((p)->prio < (rq)->curr->prio)
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio)
182 if (static_prio < NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
185 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
190 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
191 * Since cpu_power is a 'constant', we can use a reciprocal divide.
193 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
195 return reciprocal_divide(load, sg->reciprocal_cpu_power);
199 * Each time a sched group cpu_power is changed,
200 * we must compute its reciprocal value
202 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
204 sg->__cpu_power += val;
205 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
210 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
211 * to time slice values: [800ms ... 100ms ... 5ms]
213 * The higher a thread's priority, the bigger timeslices
214 * it gets during one round of execution. But even the lowest
215 * priority thread gets MIN_TIMESLICE worth of execution time.
218 static inline unsigned int task_timeslice(struct task_struct *p)
220 return static_prio_timeslice(p->static_prio);
223 static inline int rt_policy(int policy)
225 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
230 static inline int task_has_rt_policy(struct task_struct *p)
232 return rt_policy(p->policy);
236 * This is the priority-queue data structure of the RT scheduling class:
238 struct rt_prio_array {
239 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
240 struct list_head queue[MAX_RT_PRIO];
244 struct load_weight load;
245 u64 load_update_start, load_update_last;
246 unsigned long delta_fair, delta_exec, delta_stat;
249 /* CFS-related fields in a runqueue */
251 struct load_weight load;
252 unsigned long nr_running;
258 unsigned long wait_runtime_overruns, wait_runtime_underruns;
260 struct rb_root tasks_timeline;
261 struct rb_node *rb_leftmost;
262 struct rb_node *rb_load_balance_curr;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* 'curr' points to currently running entity on this cfs_rq.
265 * It is set to NULL otherwise (i.e when none are currently running).
267 struct sched_entity *curr;
268 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
270 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
271 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
272 * (like users, containers etc.)
274 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
275 * list is used during load balance.
277 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
281 /* Real-Time classes' related field in a runqueue: */
283 struct rt_prio_array active;
284 int rt_load_balance_idx;
285 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
289 * The prio-array type of the old scheduler:
292 unsigned int nr_active;
293 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
294 struct list_head queue[MAX_PRIO];
298 * This is the main, per-CPU runqueue data structure.
300 * Locking rule: those places that want to lock multiple runqueues
301 * (such as the load balancing or the thread migration code), lock
302 * acquire operations must be ordered by ascending &runqueue.
305 spinlock_t lock; /* runqueue lock */
308 * nr_running and cpu_load should be in the same cacheline because
309 * remote CPUs use both these fields when doing load calculation.
311 unsigned long nr_running;
312 unsigned long raw_weighted_load;
313 #define CPU_LOAD_IDX_MAX 5
314 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
315 unsigned char idle_at_tick;
317 unsigned char in_nohz_recently;
319 struct load_stat ls; /* capture load from *all* tasks on this cpu */
320 unsigned long nr_load_updates;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
330 * This is part of a global counter where only the total sum
331 * over all CPUs matters. A task can increase this counter on
332 * one CPU and if it got migrated afterwards it may decrease
333 * it on another CPU. Always updated under the runqueue lock:
335 unsigned long nr_uninterruptible;
337 unsigned long expired_timestamp;
338 unsigned long long most_recent_timestamp;
340 struct task_struct *curr, *idle;
341 unsigned long next_balance;
342 struct mm_struct *prev_mm;
344 struct prio_array *active, *expired, arrays[2];
345 int best_expired_prio;
347 u64 clock, prev_clock_raw;
350 unsigned int clock_warps, clock_overflows;
351 unsigned int clock_unstable_events;
353 struct sched_class *load_balance_class;
358 struct sched_domain *sd;
360 /* For active balancing */
363 int cpu; /* cpu of this runqueue */
365 struct task_struct *migration_thread;
366 struct list_head migration_queue;
369 #ifdef CONFIG_SCHEDSTATS
371 struct sched_info rq_sched_info;
373 /* sys_sched_yield() stats */
374 unsigned long yld_exp_empty;
375 unsigned long yld_act_empty;
376 unsigned long yld_both_empty;
377 unsigned long yld_cnt;
379 /* schedule() stats */
380 unsigned long sched_switch;
381 unsigned long sched_cnt;
382 unsigned long sched_goidle;
384 /* try_to_wake_up() stats */
385 unsigned long ttwu_cnt;
386 unsigned long ttwu_local;
388 struct lock_class_key rq_lock_key;
391 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
392 static DEFINE_MUTEX(sched_hotcpu_mutex);
394 static inline int cpu_of(struct rq *rq)
404 * Per-runqueue clock, as finegrained as the platform can give us:
406 static unsigned long long __rq_clock(struct rq *rq)
408 u64 prev_raw = rq->prev_clock_raw;
409 u64 now = sched_clock();
410 s64 delta = now - prev_raw;
411 u64 clock = rq->clock;
414 * Protect against sched_clock() occasionally going backwards:
416 if (unlikely(delta < 0)) {
421 * Catch too large forward jumps too:
423 if (unlikely(delta > 2*TICK_NSEC)) {
425 rq->clock_overflows++;
427 if (unlikely(delta > rq->clock_max_delta))
428 rq->clock_max_delta = delta;
433 rq->prev_clock_raw = now;
439 static inline unsigned long long rq_clock(struct rq *rq)
441 int this_cpu = smp_processor_id();
443 if (this_cpu == cpu_of(rq))
444 return __rq_clock(rq);
450 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
451 * See detach_destroy_domains: synchronize_sched for details.
453 * The domain tree of any CPU may only be accessed from within
454 * preempt-disabled sections.
456 #define for_each_domain(cpu, __sd) \
457 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
459 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
460 #define this_rq() (&__get_cpu_var(runqueues))
461 #define task_rq(p) cpu_rq(task_cpu(p))
462 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
464 #ifdef CONFIG_FAIR_GROUP_SCHED
465 /* Change a task's ->cfs_rq if it moves across CPUs */
466 static inline void set_task_cfs_rq(struct task_struct *p)
468 p->se.cfs_rq = &task_rq(p)->cfs;
471 static inline void set_task_cfs_rq(struct task_struct *p)
476 #ifndef prepare_arch_switch
477 # define prepare_arch_switch(next) do { } while (0)
479 #ifndef finish_arch_switch
480 # define finish_arch_switch(prev) do { } while (0)
483 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
484 static inline int task_running(struct rq *rq, struct task_struct *p)
486 return rq->curr == p;
489 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
493 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
495 #ifdef CONFIG_DEBUG_SPINLOCK
496 /* this is a valid case when another task releases the spinlock */
497 rq->lock.owner = current;
500 * If we are tracking spinlock dependencies then we have to
501 * fix up the runqueue lock - which gets 'carried over' from
504 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
506 spin_unlock_irq(&rq->lock);
509 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
510 static inline int task_running(struct rq *rq, struct task_struct *p)
515 return rq->curr == p;
519 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
523 * We can optimise this out completely for !SMP, because the
524 * SMP rebalancing from interrupt is the only thing that cares
529 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
530 spin_unlock_irq(&rq->lock);
532 spin_unlock(&rq->lock);
536 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
540 * After ->oncpu is cleared, the task can be moved to a different CPU.
541 * We must ensure this doesn't happen until the switch is completely
547 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
551 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
554 * __task_rq_lock - lock the runqueue a given task resides on.
555 * Must be called interrupts disabled.
557 static inline struct rq *__task_rq_lock(struct task_struct *p)
564 spin_lock(&rq->lock);
565 if (unlikely(rq != task_rq(p))) {
566 spin_unlock(&rq->lock);
567 goto repeat_lock_task;
573 * task_rq_lock - lock the runqueue a given task resides on and disable
574 * interrupts. Note the ordering: we can safely lookup the task_rq without
575 * explicitly disabling preemption.
577 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
583 local_irq_save(*flags);
585 spin_lock(&rq->lock);
586 if (unlikely(rq != task_rq(p))) {
587 spin_unlock_irqrestore(&rq->lock, *flags);
588 goto repeat_lock_task;
593 static inline void __task_rq_unlock(struct rq *rq)
596 spin_unlock(&rq->lock);
599 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
602 spin_unlock_irqrestore(&rq->lock, *flags);
606 * this_rq_lock - lock this runqueue and disable interrupts.
608 static inline struct rq *this_rq_lock(void)
615 spin_lock(&rq->lock);
621 * resched_task - mark a task 'to be rescheduled now'.
623 * On UP this means the setting of the need_resched flag, on SMP it
624 * might also involve a cross-CPU call to trigger the scheduler on
629 #ifndef tsk_is_polling
630 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
633 static void resched_task(struct task_struct *p)
637 assert_spin_locked(&task_rq(p)->lock);
639 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
642 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
645 if (cpu == smp_processor_id())
648 /* NEED_RESCHED must be visible before we test polling */
650 if (!tsk_is_polling(p))
651 smp_send_reschedule(cpu);
654 static void resched_cpu(int cpu)
656 struct rq *rq = cpu_rq(cpu);
659 if (!spin_trylock_irqsave(&rq->lock, flags))
661 resched_task(cpu_curr(cpu));
662 spin_unlock_irqrestore(&rq->lock, flags);
665 static inline void resched_task(struct task_struct *p)
667 assert_spin_locked(&task_rq(p)->lock);
668 set_tsk_need_resched(p);
672 #include "sched_stats.h"
674 static u64 div64_likely32(u64 divident, unsigned long divisor)
676 #if BITS_PER_LONG == 32
677 if (likely(divident <= 0xffffffffULL))
678 return (u32)divident / divisor;
679 do_div(divident, divisor);
683 return divident / divisor;
687 #if BITS_PER_LONG == 32
688 # define WMULT_CONST (~0UL)
690 # define WMULT_CONST (1UL << 32)
693 #define WMULT_SHIFT 32
695 static inline unsigned long
696 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
697 struct load_weight *lw)
701 if (unlikely(!lw->inv_weight))
702 lw->inv_weight = WMULT_CONST / lw->weight;
704 tmp = (u64)delta_exec * weight;
706 * Check whether we'd overflow the 64-bit multiplication:
708 if (unlikely(tmp > WMULT_CONST)) {
709 tmp = ((tmp >> WMULT_SHIFT/2) * lw->inv_weight)
712 tmp = (tmp * lw->inv_weight) >> WMULT_SHIFT;
715 return (unsigned long)min(tmp, (u64)sysctl_sched_runtime_limit);
718 static inline unsigned long
719 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
721 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
724 static void update_load_add(struct load_weight *lw, unsigned long inc)
730 static void update_load_sub(struct load_weight *lw, unsigned long dec)
736 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
738 if (rq->curr != rq->idle && ls->load.weight) {
739 ls->delta_exec += ls->delta_stat;
740 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
746 * Update delta_exec, delta_fair fields for rq.
748 * delta_fair clock advances at a rate inversely proportional to
749 * total load (rq->ls.load.weight) on the runqueue, while
750 * delta_exec advances at the same rate as wall-clock (provided
753 * delta_exec / delta_fair is a measure of the (smoothened) load on this
754 * runqueue over any given interval. This (smoothened) load is used
755 * during load balance.
757 * This function is called /before/ updating rq->ls.load
758 * and when switching tasks.
760 static void update_curr_load(struct rq *rq, u64 now)
762 struct load_stat *ls = &rq->ls;
765 start = ls->load_update_start;
766 ls->load_update_start = now;
767 ls->delta_stat += now - start;
769 * Stagger updates to ls->delta_fair. Very frequent updates
772 if (ls->delta_stat >= sysctl_sched_stat_granularity)
773 __update_curr_load(rq, ls);
777 * To aid in avoiding the subversion of "niceness" due to uneven distribution
778 * of tasks with abnormal "nice" values across CPUs the contribution that
779 * each task makes to its run queue's load is weighted according to its
780 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
781 * scaled version of the new time slice allocation that they receive on time
786 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
787 * If static_prio_timeslice() is ever changed to break this assumption then
788 * this code will need modification
790 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
791 #define LOAD_WEIGHT(lp) \
792 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
793 #define PRIO_TO_LOAD_WEIGHT(prio) \
794 LOAD_WEIGHT(static_prio_timeslice(prio))
795 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
796 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
799 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
801 rq->raw_weighted_load += p->load_weight;
805 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
807 rq->raw_weighted_load -= p->load_weight;
810 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
813 inc_raw_weighted_load(rq, p);
816 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
819 dec_raw_weighted_load(rq, p);
822 static void set_load_weight(struct task_struct *p)
824 if (task_has_rt_policy(p)) {
826 if (p == task_rq(p)->migration_thread)
828 * The migration thread does the actual balancing.
829 * Giving its load any weight will skew balancing
835 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
837 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
841 * Adding/removing a task to/from a priority array:
843 static void dequeue_task(struct task_struct *p, struct prio_array *array)
846 list_del(&p->run_list);
847 if (list_empty(array->queue + p->prio))
848 __clear_bit(p->prio, array->bitmap);
851 static void enqueue_task(struct task_struct *p, struct prio_array *array)
853 sched_info_queued(p);
854 list_add_tail(&p->run_list, array->queue + p->prio);
855 __set_bit(p->prio, array->bitmap);
861 * Put task to the end of the run list without the overhead of dequeue
862 * followed by enqueue.
864 static void requeue_task(struct task_struct *p, struct prio_array *array)
866 list_move_tail(&p->run_list, array->queue + p->prio);
870 enqueue_task_head(struct task_struct *p, struct prio_array *array)
872 list_add(&p->run_list, array->queue + p->prio);
873 __set_bit(p->prio, array->bitmap);
879 * __normal_prio - return the priority that is based on the static
880 * priority but is modified by bonuses/penalties.
882 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
883 * into the -5 ... 0 ... +5 bonus/penalty range.
885 * We use 25% of the full 0...39 priority range so that:
887 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
888 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
890 * Both properties are important to certain workloads.
893 static inline int __normal_prio(struct task_struct *p)
899 prio = p->static_prio - bonus;
900 if (prio < MAX_RT_PRIO)
902 if (prio > MAX_PRIO-1)
908 * Calculate the expected normal priority: i.e. priority
909 * without taking RT-inheritance into account. Might be
910 * boosted by interactivity modifiers. Changes upon fork,
911 * setprio syscalls, and whenever the interactivity
912 * estimator recalculates.
914 static inline int normal_prio(struct task_struct *p)
918 if (task_has_rt_policy(p))
919 prio = MAX_RT_PRIO-1 - p->rt_priority;
921 prio = __normal_prio(p);
926 * Calculate the current priority, i.e. the priority
927 * taken into account by the scheduler. This value might
928 * be boosted by RT tasks, or might be boosted by
929 * interactivity modifiers. Will be RT if the task got
930 * RT-boosted. If not then it returns p->normal_prio.
932 static int effective_prio(struct task_struct *p)
934 p->normal_prio = normal_prio(p);
936 * If we are RT tasks or we were boosted to RT priority,
937 * keep the priority unchanged. Otherwise, update priority
938 * to the normal priority:
940 if (!rt_prio(p->prio))
941 return p->normal_prio;
946 * __activate_task - move a task to the runqueue.
948 static void __activate_task(struct task_struct *p, struct rq *rq)
950 struct prio_array *target = rq->active;
953 target = rq->expired;
954 enqueue_task(p, target);
955 inc_nr_running(p, rq);
959 * __activate_idle_task - move idle task to the _front_ of runqueue.
961 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
963 enqueue_task_head(p, rq->active);
964 inc_nr_running(p, rq);
968 * Recalculate p->normal_prio and p->prio after having slept,
969 * updating the sleep-average too:
971 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
973 return effective_prio(p);
977 * activate_task - move a task to the runqueue and do priority recalculation
979 * Update all the scheduling statistics stuff. (sleep average
980 * calculation, priority modifiers, etc.)
982 static void activate_task(struct task_struct *p, struct rq *rq, int local)
984 unsigned long long now;
992 /* Compensate for drifting sched_clock */
993 struct rq *this_rq = this_rq();
994 now = (now - this_rq->most_recent_timestamp)
995 + rq->most_recent_timestamp;
1000 * Sleep time is in units of nanosecs, so shift by 20 to get a
1001 * milliseconds-range estimation of the amount of time that the task
1004 if (unlikely(prof_on == SLEEP_PROFILING)) {
1005 if (p->state == TASK_UNINTERRUPTIBLE)
1006 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
1007 (now - p->timestamp) >> 20);
1010 p->prio = recalc_task_prio(p, now);
1013 __activate_task(p, rq);
1017 * deactivate_task - remove a task from the runqueue.
1019 static void deactivate_task(struct task_struct *p, struct rq *rq)
1021 dec_nr_running(p, rq);
1022 dequeue_task(p, p->array);
1027 * task_curr - is this task currently executing on a CPU?
1028 * @p: the task in question.
1030 inline int task_curr(const struct task_struct *p)
1032 return cpu_curr(task_cpu(p)) == p;
1035 /* Used instead of source_load when we know the type == 0 */
1036 unsigned long weighted_cpuload(const int cpu)
1038 return cpu_rq(cpu)->raw_weighted_load;
1043 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1045 task_thread_info(p)->cpu = cpu;
1048 struct migration_req {
1049 struct list_head list;
1051 struct task_struct *task;
1054 struct completion done;
1058 * The task's runqueue lock must be held.
1059 * Returns true if you have to wait for migration thread.
1062 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1064 struct rq *rq = task_rq(p);
1067 * If the task is not on a runqueue (and not running), then
1068 * it is sufficient to simply update the task's cpu field.
1070 if (!p->array && !task_running(rq, p)) {
1071 set_task_cpu(p, dest_cpu);
1075 init_completion(&req->done);
1077 req->dest_cpu = dest_cpu;
1078 list_add(&req->list, &rq->migration_queue);
1084 * wait_task_inactive - wait for a thread to unschedule.
1086 * The caller must ensure that the task *will* unschedule sometime soon,
1087 * else this function might spin for a *long* time. This function can't
1088 * be called with interrupts off, or it may introduce deadlock with
1089 * smp_call_function() if an IPI is sent by the same process we are
1090 * waiting to become inactive.
1092 void wait_task_inactive(struct task_struct *p)
1094 unsigned long flags;
1096 struct prio_array *array;
1101 * We do the initial early heuristics without holding
1102 * any task-queue locks at all. We'll only try to get
1103 * the runqueue lock when things look like they will
1109 * If the task is actively running on another CPU
1110 * still, just relax and busy-wait without holding
1113 * NOTE! Since we don't hold any locks, it's not
1114 * even sure that "rq" stays as the right runqueue!
1115 * But we don't care, since "task_running()" will
1116 * return false if the runqueue has changed and p
1117 * is actually now running somewhere else!
1119 while (task_running(rq, p))
1123 * Ok, time to look more closely! We need the rq
1124 * lock now, to be *sure*. If we're wrong, we'll
1125 * just go back and repeat.
1127 rq = task_rq_lock(p, &flags);
1128 running = task_running(rq, p);
1130 task_rq_unlock(rq, &flags);
1133 * Was it really running after all now that we
1134 * checked with the proper locks actually held?
1136 * Oops. Go back and try again..
1138 if (unlikely(running)) {
1144 * It's not enough that it's not actively running,
1145 * it must be off the runqueue _entirely_, and not
1148 * So if it wa still runnable (but just not actively
1149 * running right now), it's preempted, and we should
1150 * yield - it could be a while.
1152 if (unlikely(array)) {
1158 * Ahh, all good. It wasn't running, and it wasn't
1159 * runnable, which means that it will never become
1160 * running in the future either. We're all done!
1165 * kick_process - kick a running thread to enter/exit the kernel
1166 * @p: the to-be-kicked thread
1168 * Cause a process which is running on another CPU to enter
1169 * kernel-mode, without any delay. (to get signals handled.)
1171 * NOTE: this function doesnt have to take the runqueue lock,
1172 * because all it wants to ensure is that the remote task enters
1173 * the kernel. If the IPI races and the task has been migrated
1174 * to another CPU then no harm is done and the purpose has been
1177 void kick_process(struct task_struct *p)
1183 if ((cpu != smp_processor_id()) && task_curr(p))
1184 smp_send_reschedule(cpu);
1189 * Return a low guess at the load of a migration-source cpu weighted
1190 * according to the scheduling class and "nice" value.
1192 * We want to under-estimate the load of migration sources, to
1193 * balance conservatively.
1195 static inline unsigned long source_load(int cpu, int type)
1197 struct rq *rq = cpu_rq(cpu);
1200 return rq->raw_weighted_load;
1202 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1206 * Return a high guess at the load of a migration-target cpu weighted
1207 * according to the scheduling class and "nice" value.
1209 static inline unsigned long target_load(int cpu, int type)
1211 struct rq *rq = cpu_rq(cpu);
1214 return rq->raw_weighted_load;
1216 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1220 * Return the average load per task on the cpu's run queue
1222 static inline unsigned long cpu_avg_load_per_task(int cpu)
1224 struct rq *rq = cpu_rq(cpu);
1225 unsigned long n = rq->nr_running;
1227 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1231 * find_idlest_group finds and returns the least busy CPU group within the
1234 static struct sched_group *
1235 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1237 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1238 unsigned long min_load = ULONG_MAX, this_load = 0;
1239 int load_idx = sd->forkexec_idx;
1240 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1243 unsigned long load, avg_load;
1247 /* Skip over this group if it has no CPUs allowed */
1248 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1251 local_group = cpu_isset(this_cpu, group->cpumask);
1253 /* Tally up the load of all CPUs in the group */
1256 for_each_cpu_mask(i, group->cpumask) {
1257 /* Bias balancing toward cpus of our domain */
1259 load = source_load(i, load_idx);
1261 load = target_load(i, load_idx);
1266 /* Adjust by relative CPU power of the group */
1267 avg_load = sg_div_cpu_power(group,
1268 avg_load * SCHED_LOAD_SCALE);
1271 this_load = avg_load;
1273 } else if (avg_load < min_load) {
1274 min_load = avg_load;
1278 group = group->next;
1279 } while (group != sd->groups);
1281 if (!idlest || 100*this_load < imbalance*min_load)
1287 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1290 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1293 unsigned long load, min_load = ULONG_MAX;
1297 /* Traverse only the allowed CPUs */
1298 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1300 for_each_cpu_mask(i, tmp) {
1301 load = weighted_cpuload(i);
1303 if (load < min_load || (load == min_load && i == this_cpu)) {
1313 * sched_balance_self: balance the current task (running on cpu) in domains
1314 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1317 * Balance, ie. select the least loaded group.
1319 * Returns the target CPU number, or the same CPU if no balancing is needed.
1321 * preempt must be disabled.
1323 static int sched_balance_self(int cpu, int flag)
1325 struct task_struct *t = current;
1326 struct sched_domain *tmp, *sd = NULL;
1328 for_each_domain(cpu, tmp) {
1330 * If power savings logic is enabled for a domain, stop there.
1332 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1334 if (tmp->flags & flag)
1340 struct sched_group *group;
1341 int new_cpu, weight;
1343 if (!(sd->flags & flag)) {
1349 group = find_idlest_group(sd, t, cpu);
1355 new_cpu = find_idlest_cpu(group, t, cpu);
1356 if (new_cpu == -1 || new_cpu == cpu) {
1357 /* Now try balancing at a lower domain level of cpu */
1362 /* Now try balancing at a lower domain level of new_cpu */
1365 weight = cpus_weight(span);
1366 for_each_domain(cpu, tmp) {
1367 if (weight <= cpus_weight(tmp->span))
1369 if (tmp->flags & flag)
1372 /* while loop will break here if sd == NULL */
1378 #endif /* CONFIG_SMP */
1381 * wake_idle() will wake a task on an idle cpu if task->cpu is
1382 * not idle and an idle cpu is available. The span of cpus to
1383 * search starts with cpus closest then further out as needed,
1384 * so we always favor a closer, idle cpu.
1386 * Returns the CPU we should wake onto.
1388 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1389 static int wake_idle(int cpu, struct task_struct *p)
1392 struct sched_domain *sd;
1396 * If it is idle, then it is the best cpu to run this task.
1398 * This cpu is also the best, if it has more than one task already.
1399 * Siblings must be also busy(in most cases) as they didn't already
1400 * pickup the extra load from this cpu and hence we need not check
1401 * sibling runqueue info. This will avoid the checks and cache miss
1402 * penalities associated with that.
1404 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1407 for_each_domain(cpu, sd) {
1408 if (sd->flags & SD_WAKE_IDLE) {
1409 cpus_and(tmp, sd->span, p->cpus_allowed);
1410 for_each_cpu_mask(i, tmp) {
1421 static inline int wake_idle(int cpu, struct task_struct *p)
1428 * try_to_wake_up - wake up a thread
1429 * @p: the to-be-woken-up thread
1430 * @state: the mask of task states that can be woken
1431 * @sync: do a synchronous wakeup?
1433 * Put it on the run-queue if it's not already there. The "current"
1434 * thread is always on the run-queue (except when the actual
1435 * re-schedule is in progress), and as such you're allowed to do
1436 * the simpler "current->state = TASK_RUNNING" to mark yourself
1437 * runnable without the overhead of this.
1439 * returns failure only if the task is already active.
1441 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1443 int cpu, this_cpu, success = 0;
1444 unsigned long flags;
1448 struct sched_domain *sd, *this_sd = NULL;
1449 unsigned long load, this_load;
1453 rq = task_rq_lock(p, &flags);
1454 old_state = p->state;
1455 if (!(old_state & state))
1462 this_cpu = smp_processor_id();
1465 if (unlikely(task_running(rq, p)))
1470 schedstat_inc(rq, ttwu_cnt);
1471 if (cpu == this_cpu) {
1472 schedstat_inc(rq, ttwu_local);
1476 for_each_domain(this_cpu, sd) {
1477 if (cpu_isset(cpu, sd->span)) {
1478 schedstat_inc(sd, ttwu_wake_remote);
1484 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1488 * Check for affine wakeup and passive balancing possibilities.
1491 int idx = this_sd->wake_idx;
1492 unsigned int imbalance;
1494 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1496 load = source_load(cpu, idx);
1497 this_load = target_load(this_cpu, idx);
1499 new_cpu = this_cpu; /* Wake to this CPU if we can */
1501 if (this_sd->flags & SD_WAKE_AFFINE) {
1502 unsigned long tl = this_load;
1503 unsigned long tl_per_task;
1505 tl_per_task = cpu_avg_load_per_task(this_cpu);
1508 * If sync wakeup then subtract the (maximum possible)
1509 * effect of the currently running task from the load
1510 * of the current CPU:
1513 tl -= current->load_weight;
1516 tl + target_load(cpu, idx) <= tl_per_task) ||
1517 100*(tl + p->load_weight) <= imbalance*load) {
1519 * This domain has SD_WAKE_AFFINE and
1520 * p is cache cold in this domain, and
1521 * there is no bad imbalance.
1523 schedstat_inc(this_sd, ttwu_move_affine);
1529 * Start passive balancing when half the imbalance_pct
1532 if (this_sd->flags & SD_WAKE_BALANCE) {
1533 if (imbalance*this_load <= 100*load) {
1534 schedstat_inc(this_sd, ttwu_move_balance);
1540 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1542 new_cpu = wake_idle(new_cpu, p);
1543 if (new_cpu != cpu) {
1544 set_task_cpu(p, new_cpu);
1545 task_rq_unlock(rq, &flags);
1546 /* might preempt at this point */
1547 rq = task_rq_lock(p, &flags);
1548 old_state = p->state;
1549 if (!(old_state & state))
1554 this_cpu = smp_processor_id();
1559 #endif /* CONFIG_SMP */
1560 if (old_state == TASK_UNINTERRUPTIBLE)
1561 rq->nr_uninterruptible--;
1563 activate_task(p, rq, cpu == this_cpu);
1565 * Sync wakeups (i.e. those types of wakeups where the waker
1566 * has indicated that it will leave the CPU in short order)
1567 * don't trigger a preemption, if the woken up task will run on
1568 * this cpu. (in this case the 'I will reschedule' promise of
1569 * the waker guarantees that the freshly woken up task is going
1570 * to be considered on this CPU.)
1572 if (!sync || cpu != this_cpu) {
1573 if (TASK_PREEMPTS_CURR(p, rq))
1574 resched_task(rq->curr);
1579 p->state = TASK_RUNNING;
1581 task_rq_unlock(rq, &flags);
1586 int fastcall wake_up_process(struct task_struct *p)
1588 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1589 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1591 EXPORT_SYMBOL(wake_up_process);
1593 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1595 return try_to_wake_up(p, state, 0);
1598 static void task_running_tick(struct rq *rq, struct task_struct *p);
1600 * Perform scheduler related setup for a newly forked process p.
1601 * p is forked by current.
1603 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1605 int cpu = get_cpu();
1608 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1610 set_task_cpu(p, cpu);
1613 * We mark the process as running here, but have not actually
1614 * inserted it onto the runqueue yet. This guarantees that
1615 * nobody will actually run it, and a signal or other external
1616 * event cannot wake it up and insert it on the runqueue either.
1618 p->state = TASK_RUNNING;
1621 * Make sure we do not leak PI boosting priority to the child:
1623 p->prio = current->normal_prio;
1625 INIT_LIST_HEAD(&p->run_list);
1627 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1628 if (unlikely(sched_info_on()))
1629 memset(&p->sched_info, 0, sizeof(p->sched_info));
1631 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1634 #ifdef CONFIG_PREEMPT
1635 /* Want to start with kernel preemption disabled. */
1636 task_thread_info(p)->preempt_count = 1;
1639 * Share the timeslice between parent and child, thus the
1640 * total amount of pending timeslices in the system doesn't change,
1641 * resulting in more scheduling fairness.
1643 local_irq_disable();
1644 p->time_slice = (current->time_slice + 1) >> 1;
1646 * The remainder of the first timeslice might be recovered by
1647 * the parent if the child exits early enough.
1649 p->first_time_slice = 1;
1650 current->time_slice >>= 1;
1651 p->timestamp = sched_clock();
1652 if (unlikely(!current->time_slice)) {
1654 * This case is rare, it happens when the parent has only
1655 * a single jiffy left from its timeslice. Taking the
1656 * runqueue lock is not a problem.
1658 current->time_slice = 1;
1659 task_running_tick(cpu_rq(cpu), current);
1666 * wake_up_new_task - wake up a newly created task for the first time.
1668 * This function will do some initial scheduler statistics housekeeping
1669 * that must be done for every newly created context, then puts the task
1670 * on the runqueue and wakes it.
1672 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1674 struct rq *rq, *this_rq;
1675 unsigned long flags;
1678 rq = task_rq_lock(p, &flags);
1679 BUG_ON(p->state != TASK_RUNNING);
1680 this_cpu = smp_processor_id();
1684 * We decrease the sleep average of forking parents
1685 * and children as well, to keep max-interactive tasks
1686 * from forking tasks that are max-interactive. The parent
1687 * (current) is done further down, under its lock.
1689 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1690 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1692 p->prio = effective_prio(p);
1694 if (likely(cpu == this_cpu)) {
1695 if (!(clone_flags & CLONE_VM)) {
1697 * The VM isn't cloned, so we're in a good position to
1698 * do child-runs-first in anticipation of an exec. This
1699 * usually avoids a lot of COW overhead.
1701 if (unlikely(!current->array))
1702 __activate_task(p, rq);
1704 p->prio = current->prio;
1705 p->normal_prio = current->normal_prio;
1706 list_add_tail(&p->run_list, ¤t->run_list);
1707 p->array = current->array;
1708 p->array->nr_active++;
1709 inc_nr_running(p, rq);
1713 /* Run child last */
1714 __activate_task(p, rq);
1716 * We skip the following code due to cpu == this_cpu
1718 * task_rq_unlock(rq, &flags);
1719 * this_rq = task_rq_lock(current, &flags);
1723 this_rq = cpu_rq(this_cpu);
1726 * Not the local CPU - must adjust timestamp. This should
1727 * get optimised away in the !CONFIG_SMP case.
1729 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1730 + rq->most_recent_timestamp;
1731 __activate_task(p, rq);
1732 if (TASK_PREEMPTS_CURR(p, rq))
1733 resched_task(rq->curr);
1736 * Parent and child are on different CPUs, now get the
1737 * parent runqueue to update the parent's ->sleep_avg:
1739 task_rq_unlock(rq, &flags);
1740 this_rq = task_rq_lock(current, &flags);
1742 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1743 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1744 task_rq_unlock(this_rq, &flags);
1748 * prepare_task_switch - prepare to switch tasks
1749 * @rq: the runqueue preparing to switch
1750 * @next: the task we are going to switch to.
1752 * This is called with the rq lock held and interrupts off. It must
1753 * be paired with a subsequent finish_task_switch after the context
1756 * prepare_task_switch sets up locking and calls architecture specific
1759 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1761 prepare_lock_switch(rq, next);
1762 prepare_arch_switch(next);
1766 * finish_task_switch - clean up after a task-switch
1767 * @rq: runqueue associated with task-switch
1768 * @prev: the thread we just switched away from.
1770 * finish_task_switch must be called after the context switch, paired
1771 * with a prepare_task_switch call before the context switch.
1772 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1773 * and do any other architecture-specific cleanup actions.
1775 * Note that we may have delayed dropping an mm in context_switch(). If
1776 * so, we finish that here outside of the runqueue lock. (Doing it
1777 * with the lock held can cause deadlocks; see schedule() for
1780 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1781 __releases(rq->lock)
1783 struct mm_struct *mm = rq->prev_mm;
1789 * A task struct has one reference for the use as "current".
1790 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1791 * schedule one last time. The schedule call will never return, and
1792 * the scheduled task must drop that reference.
1793 * The test for TASK_DEAD must occur while the runqueue locks are
1794 * still held, otherwise prev could be scheduled on another cpu, die
1795 * there before we look at prev->state, and then the reference would
1797 * Manfred Spraul <manfred@colorfullife.com>
1799 prev_state = prev->state;
1800 finish_arch_switch(prev);
1801 finish_lock_switch(rq, prev);
1804 if (unlikely(prev_state == TASK_DEAD)) {
1806 * Remove function-return probe instances associated with this
1807 * task and put them back on the free list.
1809 kprobe_flush_task(prev);
1810 put_task_struct(prev);
1815 * schedule_tail - first thing a freshly forked thread must call.
1816 * @prev: the thread we just switched away from.
1818 asmlinkage void schedule_tail(struct task_struct *prev)
1819 __releases(rq->lock)
1821 struct rq *rq = this_rq();
1823 finish_task_switch(rq, prev);
1824 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1825 /* In this case, finish_task_switch does not reenable preemption */
1828 if (current->set_child_tid)
1829 put_user(current->pid, current->set_child_tid);
1833 * context_switch - switch to the new MM and the new
1834 * thread's register state.
1836 static inline struct task_struct *
1837 context_switch(struct rq *rq, struct task_struct *prev,
1838 struct task_struct *next)
1840 struct mm_struct *mm = next->mm;
1841 struct mm_struct *oldmm = prev->active_mm;
1844 * For paravirt, this is coupled with an exit in switch_to to
1845 * combine the page table reload and the switch backend into
1848 arch_enter_lazy_cpu_mode();
1851 next->active_mm = oldmm;
1852 atomic_inc(&oldmm->mm_count);
1853 enter_lazy_tlb(oldmm, next);
1855 switch_mm(oldmm, mm, next);
1858 prev->active_mm = NULL;
1859 WARN_ON(rq->prev_mm);
1860 rq->prev_mm = oldmm;
1863 * Since the runqueue lock will be released by the next
1864 * task (which is an invalid locking op but in the case
1865 * of the scheduler it's an obvious special-case), so we
1866 * do an early lockdep release here:
1868 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1869 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1872 /* Here we just switch the register state and the stack. */
1873 switch_to(prev, next, prev);
1879 * nr_running, nr_uninterruptible and nr_context_switches:
1881 * externally visible scheduler statistics: current number of runnable
1882 * threads, current number of uninterruptible-sleeping threads, total
1883 * number of context switches performed since bootup.
1885 unsigned long nr_running(void)
1887 unsigned long i, sum = 0;
1889 for_each_online_cpu(i)
1890 sum += cpu_rq(i)->nr_running;
1895 unsigned long nr_uninterruptible(void)
1897 unsigned long i, sum = 0;
1899 for_each_possible_cpu(i)
1900 sum += cpu_rq(i)->nr_uninterruptible;
1903 * Since we read the counters lockless, it might be slightly
1904 * inaccurate. Do not allow it to go below zero though:
1906 if (unlikely((long)sum < 0))
1912 unsigned long long nr_context_switches(void)
1915 unsigned long long sum = 0;
1917 for_each_possible_cpu(i)
1918 sum += cpu_rq(i)->nr_switches;
1923 unsigned long nr_iowait(void)
1925 unsigned long i, sum = 0;
1927 for_each_possible_cpu(i)
1928 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1933 unsigned long nr_active(void)
1935 unsigned long i, running = 0, uninterruptible = 0;
1937 for_each_online_cpu(i) {
1938 running += cpu_rq(i)->nr_running;
1939 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1942 if (unlikely((long)uninterruptible < 0))
1943 uninterruptible = 0;
1945 return running + uninterruptible;
1951 * Is this task likely cache-hot:
1954 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1956 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1960 * double_rq_lock - safely lock two runqueues
1962 * Note this does not disable interrupts like task_rq_lock,
1963 * you need to do so manually before calling.
1965 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1966 __acquires(rq1->lock)
1967 __acquires(rq2->lock)
1969 BUG_ON(!irqs_disabled());
1971 spin_lock(&rq1->lock);
1972 __acquire(rq2->lock); /* Fake it out ;) */
1975 spin_lock(&rq1->lock);
1976 spin_lock(&rq2->lock);
1978 spin_lock(&rq2->lock);
1979 spin_lock(&rq1->lock);
1985 * double_rq_unlock - safely unlock two runqueues
1987 * Note this does not restore interrupts like task_rq_unlock,
1988 * you need to do so manually after calling.
1990 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1991 __releases(rq1->lock)
1992 __releases(rq2->lock)
1994 spin_unlock(&rq1->lock);
1996 spin_unlock(&rq2->lock);
1998 __release(rq2->lock);
2002 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2004 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2005 __releases(this_rq->lock)
2006 __acquires(busiest->lock)
2007 __acquires(this_rq->lock)
2009 if (unlikely(!irqs_disabled())) {
2010 /* printk() doesn't work good under rq->lock */
2011 spin_unlock(&this_rq->lock);
2014 if (unlikely(!spin_trylock(&busiest->lock))) {
2015 if (busiest < this_rq) {
2016 spin_unlock(&this_rq->lock);
2017 spin_lock(&busiest->lock);
2018 spin_lock(&this_rq->lock);
2020 spin_lock(&busiest->lock);
2025 * If dest_cpu is allowed for this process, migrate the task to it.
2026 * This is accomplished by forcing the cpu_allowed mask to only
2027 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2028 * the cpu_allowed mask is restored.
2030 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2032 struct migration_req req;
2033 unsigned long flags;
2036 rq = task_rq_lock(p, &flags);
2037 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2038 || unlikely(cpu_is_offline(dest_cpu)))
2041 /* force the process onto the specified CPU */
2042 if (migrate_task(p, dest_cpu, &req)) {
2043 /* Need to wait for migration thread (might exit: take ref). */
2044 struct task_struct *mt = rq->migration_thread;
2046 get_task_struct(mt);
2047 task_rq_unlock(rq, &flags);
2048 wake_up_process(mt);
2049 put_task_struct(mt);
2050 wait_for_completion(&req.done);
2055 task_rq_unlock(rq, &flags);
2059 * sched_exec - execve() is a valuable balancing opportunity, because at
2060 * this point the task has the smallest effective memory and cache footprint.
2062 void sched_exec(void)
2064 int new_cpu, this_cpu = get_cpu();
2065 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2067 if (new_cpu != this_cpu)
2068 sched_migrate_task(current, new_cpu);
2072 * pull_task - move a task from a remote runqueue to the local runqueue.
2073 * Both runqueues must be locked.
2075 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2076 struct task_struct *p, struct rq *this_rq,
2077 struct prio_array *this_array, int this_cpu)
2079 dequeue_task(p, src_array);
2080 dec_nr_running(p, src_rq);
2081 set_task_cpu(p, this_cpu);
2082 inc_nr_running(p, this_rq);
2083 enqueue_task(p, this_array);
2084 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2085 + this_rq->most_recent_timestamp;
2087 * Note that idle threads have a prio of MAX_PRIO, for this test
2088 * to be always true for them.
2090 if (TASK_PREEMPTS_CURR(p, this_rq))
2091 resched_task(this_rq->curr);
2095 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2098 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2099 struct sched_domain *sd, enum cpu_idle_type idle,
2103 * We do not migrate tasks that are:
2104 * 1) running (obviously), or
2105 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2106 * 3) are cache-hot on their current CPU.
2108 if (!cpu_isset(this_cpu, p->cpus_allowed))
2112 if (task_running(rq, p))
2116 * Aggressive migration if:
2117 * 1) task is cache cold, or
2118 * 2) too many balance attempts have failed.
2121 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2122 #ifdef CONFIG_SCHEDSTATS
2123 if (task_hot(p, rq->most_recent_timestamp, sd))
2124 schedstat_inc(sd, lb_hot_gained[idle]);
2129 if (task_hot(p, rq->most_recent_timestamp, sd))
2134 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2137 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2138 * load from busiest to this_rq, as part of a balancing operation within
2139 * "domain". Returns the number of tasks moved.
2141 * Called with both runqueues locked.
2143 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2144 unsigned long max_nr_move, unsigned long max_load_move,
2145 struct sched_domain *sd, enum cpu_idle_type idle,
2148 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2149 best_prio_seen, skip_for_load;
2150 struct prio_array *array, *dst_array;
2151 struct list_head *head, *curr;
2152 struct task_struct *tmp;
2155 if (max_nr_move == 0 || max_load_move == 0)
2158 rem_load_move = max_load_move;
2160 this_best_prio = rq_best_prio(this_rq);
2161 best_prio = rq_best_prio(busiest);
2163 * Enable handling of the case where there is more than one task
2164 * with the best priority. If the current running task is one
2165 * of those with prio==best_prio we know it won't be moved
2166 * and therefore it's safe to override the skip (based on load) of
2167 * any task we find with that prio.
2169 best_prio_seen = best_prio == busiest->curr->prio;
2172 * We first consider expired tasks. Those will likely not be
2173 * executed in the near future, and they are most likely to
2174 * be cache-cold, thus switching CPUs has the least effect
2177 if (busiest->expired->nr_active) {
2178 array = busiest->expired;
2179 dst_array = this_rq->expired;
2181 array = busiest->active;
2182 dst_array = this_rq->active;
2186 /* Start searching at priority 0: */
2190 idx = sched_find_first_bit(array->bitmap);
2192 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2193 if (idx >= MAX_PRIO) {
2194 if (array == busiest->expired && busiest->active->nr_active) {
2195 array = busiest->active;
2196 dst_array = this_rq->active;
2202 head = array->queue + idx;
2205 tmp = list_entry(curr, struct task_struct, run_list);
2210 * To help distribute high priority tasks accross CPUs we don't
2211 * skip a task if it will be the highest priority task (i.e. smallest
2212 * prio value) on its new queue regardless of its load weight
2214 skip_for_load = tmp->load_weight > rem_load_move;
2215 if (skip_for_load && idx < this_best_prio)
2216 skip_for_load = !best_prio_seen && idx == best_prio;
2217 if (skip_for_load ||
2218 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2220 best_prio_seen |= idx == best_prio;
2227 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2229 rem_load_move -= tmp->load_weight;
2232 * We only want to steal up to the prescribed number of tasks
2233 * and the prescribed amount of weighted load.
2235 if (pulled < max_nr_move && rem_load_move > 0) {
2236 if (idx < this_best_prio)
2237 this_best_prio = idx;
2245 * Right now, this is the only place pull_task() is called,
2246 * so we can safely collect pull_task() stats here rather than
2247 * inside pull_task().
2249 schedstat_add(sd, lb_gained[idle], pulled);
2252 *all_pinned = pinned;
2257 * find_busiest_group finds and returns the busiest CPU group within the
2258 * domain. It calculates and returns the amount of weighted load which
2259 * should be moved to restore balance via the imbalance parameter.
2261 static struct sched_group *
2262 find_busiest_group(struct sched_domain *sd, int this_cpu,
2263 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2264 cpumask_t *cpus, int *balance)
2266 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2267 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2268 unsigned long max_pull;
2269 unsigned long busiest_load_per_task, busiest_nr_running;
2270 unsigned long this_load_per_task, this_nr_running;
2272 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2273 int power_savings_balance = 1;
2274 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2275 unsigned long min_nr_running = ULONG_MAX;
2276 struct sched_group *group_min = NULL, *group_leader = NULL;
2279 max_load = this_load = total_load = total_pwr = 0;
2280 busiest_load_per_task = busiest_nr_running = 0;
2281 this_load_per_task = this_nr_running = 0;
2282 if (idle == CPU_NOT_IDLE)
2283 load_idx = sd->busy_idx;
2284 else if (idle == CPU_NEWLY_IDLE)
2285 load_idx = sd->newidle_idx;
2287 load_idx = sd->idle_idx;
2290 unsigned long load, group_capacity;
2293 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2294 unsigned long sum_nr_running, sum_weighted_load;
2296 local_group = cpu_isset(this_cpu, group->cpumask);
2299 balance_cpu = first_cpu(group->cpumask);
2301 /* Tally up the load of all CPUs in the group */
2302 sum_weighted_load = sum_nr_running = avg_load = 0;
2304 for_each_cpu_mask(i, group->cpumask) {
2307 if (!cpu_isset(i, *cpus))
2312 if (*sd_idle && !idle_cpu(i))
2315 /* Bias balancing toward cpus of our domain */
2317 if (idle_cpu(i) && !first_idle_cpu) {
2322 load = target_load(i, load_idx);
2324 load = source_load(i, load_idx);
2327 sum_nr_running += rq->nr_running;
2328 sum_weighted_load += rq->raw_weighted_load;
2332 * First idle cpu or the first cpu(busiest) in this sched group
2333 * is eligible for doing load balancing at this and above
2336 if (local_group && balance_cpu != this_cpu && balance) {
2341 total_load += avg_load;
2342 total_pwr += group->__cpu_power;
2344 /* Adjust by relative CPU power of the group */
2345 avg_load = sg_div_cpu_power(group,
2346 avg_load * SCHED_LOAD_SCALE);
2348 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2351 this_load = avg_load;
2353 this_nr_running = sum_nr_running;
2354 this_load_per_task = sum_weighted_load;
2355 } else if (avg_load > max_load &&
2356 sum_nr_running > group_capacity) {
2357 max_load = avg_load;
2359 busiest_nr_running = sum_nr_running;
2360 busiest_load_per_task = sum_weighted_load;
2363 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2365 * Busy processors will not participate in power savings
2368 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2372 * If the local group is idle or completely loaded
2373 * no need to do power savings balance at this domain
2375 if (local_group && (this_nr_running >= group_capacity ||
2377 power_savings_balance = 0;
2380 * If a group is already running at full capacity or idle,
2381 * don't include that group in power savings calculations
2383 if (!power_savings_balance || sum_nr_running >= group_capacity
2388 * Calculate the group which has the least non-idle load.
2389 * This is the group from where we need to pick up the load
2392 if ((sum_nr_running < min_nr_running) ||
2393 (sum_nr_running == min_nr_running &&
2394 first_cpu(group->cpumask) <
2395 first_cpu(group_min->cpumask))) {
2397 min_nr_running = sum_nr_running;
2398 min_load_per_task = sum_weighted_load /
2403 * Calculate the group which is almost near its
2404 * capacity but still has some space to pick up some load
2405 * from other group and save more power
2407 if (sum_nr_running <= group_capacity - 1) {
2408 if (sum_nr_running > leader_nr_running ||
2409 (sum_nr_running == leader_nr_running &&
2410 first_cpu(group->cpumask) >
2411 first_cpu(group_leader->cpumask))) {
2412 group_leader = group;
2413 leader_nr_running = sum_nr_running;
2418 group = group->next;
2419 } while (group != sd->groups);
2421 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2424 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2426 if (this_load >= avg_load ||
2427 100*max_load <= sd->imbalance_pct*this_load)
2430 busiest_load_per_task /= busiest_nr_running;
2432 * We're trying to get all the cpus to the average_load, so we don't
2433 * want to push ourselves above the average load, nor do we wish to
2434 * reduce the max loaded cpu below the average load, as either of these
2435 * actions would just result in more rebalancing later, and ping-pong
2436 * tasks around. Thus we look for the minimum possible imbalance.
2437 * Negative imbalances (*we* are more loaded than anyone else) will
2438 * be counted as no imbalance for these purposes -- we can't fix that
2439 * by pulling tasks to us. Be careful of negative numbers as they'll
2440 * appear as very large values with unsigned longs.
2442 if (max_load <= busiest_load_per_task)
2446 * In the presence of smp nice balancing, certain scenarios can have
2447 * max load less than avg load(as we skip the groups at or below
2448 * its cpu_power, while calculating max_load..)
2450 if (max_load < avg_load) {
2452 goto small_imbalance;
2455 /* Don't want to pull so many tasks that a group would go idle */
2456 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2458 /* How much load to actually move to equalise the imbalance */
2459 *imbalance = min(max_pull * busiest->__cpu_power,
2460 (avg_load - this_load) * this->__cpu_power)
2464 * if *imbalance is less than the average load per runnable task
2465 * there is no gaurantee that any tasks will be moved so we'll have
2466 * a think about bumping its value to force at least one task to be
2469 if (*imbalance < busiest_load_per_task) {
2470 unsigned long tmp, pwr_now, pwr_move;
2474 pwr_move = pwr_now = 0;
2476 if (this_nr_running) {
2477 this_load_per_task /= this_nr_running;
2478 if (busiest_load_per_task > this_load_per_task)
2481 this_load_per_task = SCHED_LOAD_SCALE;
2483 if (max_load - this_load >= busiest_load_per_task * imbn) {
2484 *imbalance = busiest_load_per_task;
2489 * OK, we don't have enough imbalance to justify moving tasks,
2490 * however we may be able to increase total CPU power used by
2494 pwr_now += busiest->__cpu_power *
2495 min(busiest_load_per_task, max_load);
2496 pwr_now += this->__cpu_power *
2497 min(this_load_per_task, this_load);
2498 pwr_now /= SCHED_LOAD_SCALE;
2500 /* Amount of load we'd subtract */
2501 tmp = sg_div_cpu_power(busiest,
2502 busiest_load_per_task * SCHED_LOAD_SCALE);
2504 pwr_move += busiest->__cpu_power *
2505 min(busiest_load_per_task, max_load - tmp);
2507 /* Amount of load we'd add */
2508 if (max_load * busiest->__cpu_power <
2509 busiest_load_per_task * SCHED_LOAD_SCALE)
2510 tmp = sg_div_cpu_power(this,
2511 max_load * busiest->__cpu_power);
2513 tmp = sg_div_cpu_power(this,
2514 busiest_load_per_task * SCHED_LOAD_SCALE);
2515 pwr_move += this->__cpu_power *
2516 min(this_load_per_task, this_load + tmp);
2517 pwr_move /= SCHED_LOAD_SCALE;
2519 /* Move if we gain throughput */
2520 if (pwr_move <= pwr_now)
2523 *imbalance = busiest_load_per_task;
2529 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2530 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2533 if (this == group_leader && group_leader != group_min) {
2534 *imbalance = min_load_per_task;
2544 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2547 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2548 unsigned long imbalance, cpumask_t *cpus)
2550 struct rq *busiest = NULL, *rq;
2551 unsigned long max_load = 0;
2554 for_each_cpu_mask(i, group->cpumask) {
2556 if (!cpu_isset(i, *cpus))
2561 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2564 if (rq->raw_weighted_load > max_load) {
2565 max_load = rq->raw_weighted_load;
2574 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2575 * so long as it is large enough.
2577 #define MAX_PINNED_INTERVAL 512
2579 static inline unsigned long minus_1_or_zero(unsigned long n)
2581 return n > 0 ? n - 1 : 0;
2585 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2586 * tasks if there is an imbalance.
2588 static int load_balance(int this_cpu, struct rq *this_rq,
2589 struct sched_domain *sd, enum cpu_idle_type idle,
2592 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2593 struct sched_group *group;
2594 unsigned long imbalance;
2596 cpumask_t cpus = CPU_MASK_ALL;
2597 unsigned long flags;
2600 * When power savings policy is enabled for the parent domain, idle
2601 * sibling can pick up load irrespective of busy siblings. In this case,
2602 * let the state of idle sibling percolate up as IDLE, instead of
2603 * portraying it as CPU_NOT_IDLE.
2605 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2606 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2609 schedstat_inc(sd, lb_cnt[idle]);
2612 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2619 schedstat_inc(sd, lb_nobusyg[idle]);
2623 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2625 schedstat_inc(sd, lb_nobusyq[idle]);
2629 BUG_ON(busiest == this_rq);
2631 schedstat_add(sd, lb_imbalance[idle], imbalance);
2634 if (busiest->nr_running > 1) {
2636 * Attempt to move tasks. If find_busiest_group has found
2637 * an imbalance but busiest->nr_running <= 1, the group is
2638 * still unbalanced. nr_moved simply stays zero, so it is
2639 * correctly treated as an imbalance.
2641 local_irq_save(flags);
2642 double_rq_lock(this_rq, busiest);
2643 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2644 minus_1_or_zero(busiest->nr_running),
2645 imbalance, sd, idle, &all_pinned);
2646 double_rq_unlock(this_rq, busiest);
2647 local_irq_restore(flags);
2650 * some other cpu did the load balance for us.
2652 if (nr_moved && this_cpu != smp_processor_id())
2653 resched_cpu(this_cpu);
2655 /* All tasks on this runqueue were pinned by CPU affinity */
2656 if (unlikely(all_pinned)) {
2657 cpu_clear(cpu_of(busiest), cpus);
2658 if (!cpus_empty(cpus))
2665 schedstat_inc(sd, lb_failed[idle]);
2666 sd->nr_balance_failed++;
2668 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2670 spin_lock_irqsave(&busiest->lock, flags);
2672 /* don't kick the migration_thread, if the curr
2673 * task on busiest cpu can't be moved to this_cpu
2675 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2676 spin_unlock_irqrestore(&busiest->lock, flags);
2678 goto out_one_pinned;
2681 if (!busiest->active_balance) {
2682 busiest->active_balance = 1;
2683 busiest->push_cpu = this_cpu;
2686 spin_unlock_irqrestore(&busiest->lock, flags);
2688 wake_up_process(busiest->migration_thread);
2691 * We've kicked active balancing, reset the failure
2694 sd->nr_balance_failed = sd->cache_nice_tries+1;
2697 sd->nr_balance_failed = 0;
2699 if (likely(!active_balance)) {
2700 /* We were unbalanced, so reset the balancing interval */
2701 sd->balance_interval = sd->min_interval;
2704 * If we've begun active balancing, start to back off. This
2705 * case may not be covered by the all_pinned logic if there
2706 * is only 1 task on the busy runqueue (because we don't call
2709 if (sd->balance_interval < sd->max_interval)
2710 sd->balance_interval *= 2;
2713 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2714 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2719 schedstat_inc(sd, lb_balanced[idle]);
2721 sd->nr_balance_failed = 0;
2724 /* tune up the balancing interval */
2725 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2726 (sd->balance_interval < sd->max_interval))
2727 sd->balance_interval *= 2;
2729 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2730 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2736 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2737 * tasks if there is an imbalance.
2739 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2740 * this_rq is locked.
2743 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2745 struct sched_group *group;
2746 struct rq *busiest = NULL;
2747 unsigned long imbalance;
2750 cpumask_t cpus = CPU_MASK_ALL;
2753 * When power savings policy is enabled for the parent domain, idle
2754 * sibling can pick up load irrespective of busy siblings. In this case,
2755 * let the state of idle sibling percolate up as IDLE, instead of
2756 * portraying it as CPU_NOT_IDLE.
2758 if (sd->flags & SD_SHARE_CPUPOWER &&
2759 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2762 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2764 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2765 &sd_idle, &cpus, NULL);
2767 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2771 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2774 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2778 BUG_ON(busiest == this_rq);
2780 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2783 if (busiest->nr_running > 1) {
2784 /* Attempt to move tasks */
2785 double_lock_balance(this_rq, busiest);
2786 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2787 minus_1_or_zero(busiest->nr_running),
2788 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2789 spin_unlock(&busiest->lock);
2792 cpu_clear(cpu_of(busiest), cpus);
2793 if (!cpus_empty(cpus))
2799 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2800 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2801 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2804 sd->nr_balance_failed = 0;
2809 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2810 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2811 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2813 sd->nr_balance_failed = 0;
2819 * idle_balance is called by schedule() if this_cpu is about to become
2820 * idle. Attempts to pull tasks from other CPUs.
2822 static void idle_balance(int this_cpu, struct rq *this_rq)
2824 struct sched_domain *sd;
2825 int pulled_task = 0;
2826 unsigned long next_balance = jiffies + 60 * HZ;
2828 for_each_domain(this_cpu, sd) {
2829 unsigned long interval;
2831 if (!(sd->flags & SD_LOAD_BALANCE))
2834 if (sd->flags & SD_BALANCE_NEWIDLE)
2835 /* If we've pulled tasks over stop searching: */
2836 pulled_task = load_balance_newidle(this_cpu,
2839 interval = msecs_to_jiffies(sd->balance_interval);
2840 if (time_after(next_balance, sd->last_balance + interval))
2841 next_balance = sd->last_balance + interval;
2847 * We are going idle. next_balance may be set based on
2848 * a busy processor. So reset next_balance.
2850 this_rq->next_balance = next_balance;
2854 * active_load_balance is run by migration threads. It pushes running tasks
2855 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2856 * running on each physical CPU where possible, and avoids physical /
2857 * logical imbalances.
2859 * Called with busiest_rq locked.
2861 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2863 int target_cpu = busiest_rq->push_cpu;
2864 struct sched_domain *sd;
2865 struct rq *target_rq;
2867 /* Is there any task to move? */
2868 if (busiest_rq->nr_running <= 1)
2871 target_rq = cpu_rq(target_cpu);
2874 * This condition is "impossible", if it occurs
2875 * we need to fix it. Originally reported by
2876 * Bjorn Helgaas on a 128-cpu setup.
2878 BUG_ON(busiest_rq == target_rq);
2880 /* move a task from busiest_rq to target_rq */
2881 double_lock_balance(busiest_rq, target_rq);
2883 /* Search for an sd spanning us and the target CPU. */
2884 for_each_domain(target_cpu, sd) {
2885 if ((sd->flags & SD_LOAD_BALANCE) &&
2886 cpu_isset(busiest_cpu, sd->span))
2891 schedstat_inc(sd, alb_cnt);
2893 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2894 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2896 schedstat_inc(sd, alb_pushed);
2898 schedstat_inc(sd, alb_failed);
2900 spin_unlock(&target_rq->lock);
2903 static void update_load(struct rq *this_rq)
2905 unsigned long this_load;
2906 unsigned int i, scale;
2908 this_load = this_rq->raw_weighted_load;
2910 /* Update our load: */
2911 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
2912 unsigned long old_load, new_load;
2914 /* scale is effectively 1 << i now, and >> i divides by scale */
2916 old_load = this_rq->cpu_load[i];
2917 new_load = this_load;
2919 * Round up the averaging division if load is increasing. This
2920 * prevents us from getting stuck on 9 if the load is 10, for
2923 if (new_load > old_load)
2924 new_load += scale-1;
2925 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2931 atomic_t load_balancer;
2933 } nohz ____cacheline_aligned = {
2934 .load_balancer = ATOMIC_INIT(-1),
2935 .cpu_mask = CPU_MASK_NONE,
2939 * This routine will try to nominate the ilb (idle load balancing)
2940 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2941 * load balancing on behalf of all those cpus. If all the cpus in the system
2942 * go into this tickless mode, then there will be no ilb owner (as there is
2943 * no need for one) and all the cpus will sleep till the next wakeup event
2946 * For the ilb owner, tick is not stopped. And this tick will be used
2947 * for idle load balancing. ilb owner will still be part of
2950 * While stopping the tick, this cpu will become the ilb owner if there
2951 * is no other owner. And will be the owner till that cpu becomes busy
2952 * or if all cpus in the system stop their ticks at which point
2953 * there is no need for ilb owner.
2955 * When the ilb owner becomes busy, it nominates another owner, during the
2956 * next busy scheduler_tick()
2958 int select_nohz_load_balancer(int stop_tick)
2960 int cpu = smp_processor_id();
2963 cpu_set(cpu, nohz.cpu_mask);
2964 cpu_rq(cpu)->in_nohz_recently = 1;
2967 * If we are going offline and still the leader, give up!
2969 if (cpu_is_offline(cpu) &&
2970 atomic_read(&nohz.load_balancer) == cpu) {
2971 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2976 /* time for ilb owner also to sleep */
2977 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2978 if (atomic_read(&nohz.load_balancer) == cpu)
2979 atomic_set(&nohz.load_balancer, -1);
2983 if (atomic_read(&nohz.load_balancer) == -1) {
2984 /* make me the ilb owner */
2985 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2987 } else if (atomic_read(&nohz.load_balancer) == cpu)
2990 if (!cpu_isset(cpu, nohz.cpu_mask))
2993 cpu_clear(cpu, nohz.cpu_mask);
2995 if (atomic_read(&nohz.load_balancer) == cpu)
2996 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3003 static DEFINE_SPINLOCK(balancing);
3006 * It checks each scheduling domain to see if it is due to be balanced,
3007 * and initiates a balancing operation if so.
3009 * Balancing parameters are set up in arch_init_sched_domains.
3011 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3014 struct rq *rq = cpu_rq(cpu);
3015 unsigned long interval;
3016 struct sched_domain *sd;
3017 /* Earliest time when we have to do rebalance again */
3018 unsigned long next_balance = jiffies + 60*HZ;
3020 for_each_domain(cpu, sd) {
3021 if (!(sd->flags & SD_LOAD_BALANCE))
3024 interval = sd->balance_interval;
3025 if (idle != CPU_IDLE)
3026 interval *= sd->busy_factor;
3028 /* scale ms to jiffies */
3029 interval = msecs_to_jiffies(interval);
3030 if (unlikely(!interval))
3033 if (sd->flags & SD_SERIALIZE) {
3034 if (!spin_trylock(&balancing))
3038 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3039 if (load_balance(cpu, rq, sd, idle, &balance)) {
3041 * We've pulled tasks over so either we're no
3042 * longer idle, or one of our SMT siblings is
3045 idle = CPU_NOT_IDLE;
3047 sd->last_balance = jiffies;
3049 if (sd->flags & SD_SERIALIZE)
3050 spin_unlock(&balancing);
3052 if (time_after(next_balance, sd->last_balance + interval))
3053 next_balance = sd->last_balance + interval;
3056 * Stop the load balance at this level. There is another
3057 * CPU in our sched group which is doing load balancing more
3063 rq->next_balance = next_balance;
3067 * run_rebalance_domains is triggered when needed from the scheduler tick.
3068 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3069 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3071 static void run_rebalance_domains(struct softirq_action *h)
3073 int local_cpu = smp_processor_id();
3074 struct rq *local_rq = cpu_rq(local_cpu);
3075 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3077 rebalance_domains(local_cpu, idle);
3081 * If this cpu is the owner for idle load balancing, then do the
3082 * balancing on behalf of the other idle cpus whose ticks are
3085 if (local_rq->idle_at_tick &&
3086 atomic_read(&nohz.load_balancer) == local_cpu) {
3087 cpumask_t cpus = nohz.cpu_mask;
3091 cpu_clear(local_cpu, cpus);
3092 for_each_cpu_mask(balance_cpu, cpus) {
3094 * If this cpu gets work to do, stop the load balancing
3095 * work being done for other cpus. Next load
3096 * balancing owner will pick it up.
3101 rebalance_domains(balance_cpu, CPU_IDLE);
3103 rq = cpu_rq(balance_cpu);
3104 if (time_after(local_rq->next_balance, rq->next_balance))
3105 local_rq->next_balance = rq->next_balance;
3112 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3114 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3115 * idle load balancing owner or decide to stop the periodic load balancing,
3116 * if the whole system is idle.
3118 static inline void trigger_load_balance(int cpu)
3120 struct rq *rq = cpu_rq(cpu);
3123 * If we were in the nohz mode recently and busy at the current
3124 * scheduler tick, then check if we need to nominate new idle
3127 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3128 rq->in_nohz_recently = 0;
3130 if (atomic_read(&nohz.load_balancer) == cpu) {
3131 cpu_clear(cpu, nohz.cpu_mask);
3132 atomic_set(&nohz.load_balancer, -1);
3135 if (atomic_read(&nohz.load_balancer) == -1) {
3137 * simple selection for now: Nominate the
3138 * first cpu in the nohz list to be the next
3141 * TBD: Traverse the sched domains and nominate
3142 * the nearest cpu in the nohz.cpu_mask.
3144 int ilb = first_cpu(nohz.cpu_mask);
3152 * If this cpu is idle and doing idle load balancing for all the
3153 * cpus with ticks stopped, is it time for that to stop?
3155 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3156 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3162 * If this cpu is idle and the idle load balancing is done by
3163 * someone else, then no need raise the SCHED_SOFTIRQ
3165 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3166 cpu_isset(cpu, nohz.cpu_mask))
3169 if (time_after_eq(jiffies, rq->next_balance))
3170 raise_softirq(SCHED_SOFTIRQ);
3174 * on UP we do not need to balance between CPUs:
3176 static inline void idle_balance(int cpu, struct rq *rq)
3181 DEFINE_PER_CPU(struct kernel_stat, kstat);
3183 EXPORT_PER_CPU_SYMBOL(kstat);
3186 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3187 * that have not yet been banked in case the task is currently running.
3189 unsigned long long task_sched_runtime(struct task_struct *p)
3191 unsigned long flags;
3195 rq = task_rq_lock(p, &flags);
3196 ns = p->se.sum_exec_runtime;
3197 if (rq->curr == p) {
3198 delta_exec = rq_clock(rq) - p->se.exec_start;
3199 if ((s64)delta_exec > 0)
3202 task_rq_unlock(rq, &flags);
3208 * Account user cpu time to a process.
3209 * @p: the process that the cpu time gets accounted to
3210 * @hardirq_offset: the offset to subtract from hardirq_count()
3211 * @cputime: the cpu time spent in user space since the last update
3213 void account_user_time(struct task_struct *p, cputime_t cputime)
3215 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3218 p->utime = cputime_add(p->utime, cputime);
3220 /* Add user time to cpustat. */
3221 tmp = cputime_to_cputime64(cputime);
3222 if (TASK_NICE(p) > 0)
3223 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3225 cpustat->user = cputime64_add(cpustat->user, tmp);
3229 * Account system cpu time to a process.
3230 * @p: the process that the cpu time gets accounted to
3231 * @hardirq_offset: the offset to subtract from hardirq_count()
3232 * @cputime: the cpu time spent in kernel space since the last update
3234 void account_system_time(struct task_struct *p, int hardirq_offset,
3237 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3238 struct rq *rq = this_rq();
3241 p->stime = cputime_add(p->stime, cputime);
3243 /* Add system time to cpustat. */
3244 tmp = cputime_to_cputime64(cputime);
3245 if (hardirq_count() - hardirq_offset)
3246 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3247 else if (softirq_count())
3248 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3249 else if (p != rq->idle)
3250 cpustat->system = cputime64_add(cpustat->system, tmp);
3251 else if (atomic_read(&rq->nr_iowait) > 0)
3252 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3254 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3255 /* Account for system time used */
3256 acct_update_integrals(p);
3260 * Account for involuntary wait time.
3261 * @p: the process from which the cpu time has been stolen
3262 * @steal: the cpu time spent in involuntary wait
3264 void account_steal_time(struct task_struct *p, cputime_t steal)
3266 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3267 cputime64_t tmp = cputime_to_cputime64(steal);
3268 struct rq *rq = this_rq();
3270 if (p == rq->idle) {
3271 p->stime = cputime_add(p->stime, steal);
3272 if (atomic_read(&rq->nr_iowait) > 0)
3273 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3275 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3277 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3280 static void task_running_tick(struct rq *rq, struct task_struct *p)
3282 if (p->array != rq->active) {
3283 /* Task has expired but was not scheduled yet */
3284 set_tsk_need_resched(p);
3287 spin_lock(&rq->lock);
3289 * The task was running during this tick - update the
3290 * time slice counter. Note: we do not update a thread's
3291 * priority until it either goes to sleep or uses up its
3292 * timeslice. This makes it possible for interactive tasks
3293 * to use up their timeslices at their highest priority levels.
3297 * RR tasks need a special form of timeslice management.
3298 * FIFO tasks have no timeslices.
3300 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3301 p->time_slice = task_timeslice(p);
3302 p->first_time_slice = 0;
3303 set_tsk_need_resched(p);
3305 /* put it at the end of the queue: */
3306 requeue_task(p, rq->active);
3310 if (!--p->time_slice) {
3311 dequeue_task(p, rq->active);
3312 set_tsk_need_resched(p);
3313 p->prio = effective_prio(p);
3314 p->time_slice = task_timeslice(p);
3315 p->first_time_slice = 0;
3317 if (!rq->expired_timestamp)
3318 rq->expired_timestamp = jiffies;
3319 if (!TASK_INTERACTIVE(p)) {
3320 enqueue_task(p, rq->expired);
3321 if (p->static_prio < rq->best_expired_prio)
3322 rq->best_expired_prio = p->static_prio;
3324 enqueue_task(p, rq->active);
3327 * Prevent a too long timeslice allowing a task to monopolize
3328 * the CPU. We do this by splitting up the timeslice into
3331 * Note: this does not mean the task's timeslices expire or
3332 * get lost in any way, they just might be preempted by
3333 * another task of equal priority. (one with higher
3334 * priority would have preempted this task already.) We
3335 * requeue this task to the end of the list on this priority
3336 * level, which is in essence a round-robin of tasks with
3339 * This only applies to tasks in the interactive
3340 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3342 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3343 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3344 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3345 (p->array == rq->active)) {
3347 requeue_task(p, rq->active);
3348 set_tsk_need_resched(p);
3352 spin_unlock(&rq->lock);
3356 * This function gets called by the timer code, with HZ frequency.
3357 * We call it with interrupts disabled.
3359 * It also gets called by the fork code, when changing the parent's
3362 void scheduler_tick(void)
3364 struct task_struct *p = current;
3365 int cpu = smp_processor_id();
3366 int idle_at_tick = idle_cpu(cpu);
3367 struct rq *rq = cpu_rq(cpu);
3370 task_running_tick(rq, p);
3373 rq->idle_at_tick = idle_at_tick;
3374 trigger_load_balance(cpu);
3378 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3380 void fastcall add_preempt_count(int val)
3385 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3387 preempt_count() += val;
3389 * Spinlock count overflowing soon?
3391 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3394 EXPORT_SYMBOL(add_preempt_count);
3396 void fastcall sub_preempt_count(int val)
3401 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3404 * Is the spinlock portion underflowing?
3406 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3407 !(preempt_count() & PREEMPT_MASK)))
3410 preempt_count() -= val;
3412 EXPORT_SYMBOL(sub_preempt_count);
3417 * schedule() is the main scheduler function.
3419 asmlinkage void __sched schedule(void)
3421 struct task_struct *prev, *next;
3422 struct prio_array *array;
3423 struct list_head *queue;
3424 unsigned long long now;
3425 unsigned long run_time;
3431 * Test if we are atomic. Since do_exit() needs to call into
3432 * schedule() atomically, we ignore that path for now.
3433 * Otherwise, whine if we are scheduling when we should not be.
3435 if (unlikely(in_atomic() && !current->exit_state)) {
3436 printk(KERN_ERR "BUG: scheduling while atomic: "
3438 current->comm, preempt_count(), current->pid);
3439 debug_show_held_locks(current);
3440 if (irqs_disabled())
3441 print_irqtrace_events(current);
3444 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3449 release_kernel_lock(prev);
3450 need_resched_nonpreemptible:
3454 * The idle thread is not allowed to schedule!
3455 * Remove this check after it has been exercised a bit.
3457 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3458 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3462 schedstat_inc(rq, sched_cnt);
3463 now = sched_clock();
3464 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3465 run_time = now - prev->timestamp;
3466 if (unlikely((long long)(now - prev->timestamp) < 0))
3469 run_time = NS_MAX_SLEEP_AVG;
3472 * Tasks charged proportionately less run_time at high sleep_avg to
3473 * delay them losing their interactive status
3475 run_time /= (CURRENT_BONUS(prev) ? : 1);
3477 spin_lock_irq(&rq->lock);
3479 switch_count = &prev->nivcsw;
3480 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3481 switch_count = &prev->nvcsw;
3482 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3483 unlikely(signal_pending(prev))))
3484 prev->state = TASK_RUNNING;
3486 if (prev->state == TASK_UNINTERRUPTIBLE)
3487 rq->nr_uninterruptible++;
3488 deactivate_task(prev, rq);
3492 cpu = smp_processor_id();
3493 if (unlikely(!rq->nr_running)) {
3494 idle_balance(cpu, rq);
3495 if (!rq->nr_running) {
3497 rq->expired_timestamp = 0;
3503 if (unlikely(!array->nr_active)) {
3505 * Switch the active and expired arrays.
3507 schedstat_inc(rq, sched_switch);
3508 rq->active = rq->expired;
3509 rq->expired = array;
3511 rq->expired_timestamp = 0;
3512 rq->best_expired_prio = MAX_PRIO;
3515 idx = sched_find_first_bit(array->bitmap);
3516 queue = array->queue + idx;
3517 next = list_entry(queue->next, struct task_struct, run_list);
3520 if (next == rq->idle)
3521 schedstat_inc(rq, sched_goidle);
3523 prefetch_stack(next);
3524 clear_tsk_need_resched(prev);
3525 rcu_qsctr_inc(task_cpu(prev));
3527 prev->timestamp = prev->last_ran = now;
3529 sched_info_switch(prev, next);
3530 if (likely(prev != next)) {
3531 next->timestamp = next->last_ran = now;
3536 prepare_task_switch(rq, next);
3537 prev = context_switch(rq, prev, next);
3540 * this_rq must be evaluated again because prev may have moved
3541 * CPUs since it called schedule(), thus the 'rq' on its stack
3542 * frame will be invalid.
3544 finish_task_switch(this_rq(), prev);
3546 spin_unlock_irq(&rq->lock);
3549 if (unlikely(reacquire_kernel_lock(prev) < 0))
3550 goto need_resched_nonpreemptible;
3551 preempt_enable_no_resched();
3552 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3555 EXPORT_SYMBOL(schedule);
3557 #ifdef CONFIG_PREEMPT
3559 * this is the entry point to schedule() from in-kernel preemption
3560 * off of preempt_enable. Kernel preemptions off return from interrupt
3561 * occur there and call schedule directly.
3563 asmlinkage void __sched preempt_schedule(void)
3565 struct thread_info *ti = current_thread_info();
3566 #ifdef CONFIG_PREEMPT_BKL
3567 struct task_struct *task = current;
3568 int saved_lock_depth;
3571 * If there is a non-zero preempt_count or interrupts are disabled,
3572 * we do not want to preempt the current task. Just return..
3574 if (likely(ti->preempt_count || irqs_disabled()))
3578 add_preempt_count(PREEMPT_ACTIVE);
3580 * We keep the big kernel semaphore locked, but we
3581 * clear ->lock_depth so that schedule() doesnt
3582 * auto-release the semaphore:
3584 #ifdef CONFIG_PREEMPT_BKL
3585 saved_lock_depth = task->lock_depth;
3586 task->lock_depth = -1;
3589 #ifdef CONFIG_PREEMPT_BKL
3590 task->lock_depth = saved_lock_depth;
3592 sub_preempt_count(PREEMPT_ACTIVE);
3594 /* we could miss a preemption opportunity between schedule and now */
3596 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3599 EXPORT_SYMBOL(preempt_schedule);
3602 * this is the entry point to schedule() from kernel preemption
3603 * off of irq context.
3604 * Note, that this is called and return with irqs disabled. This will
3605 * protect us against recursive calling from irq.
3607 asmlinkage void __sched preempt_schedule_irq(void)
3609 struct thread_info *ti = current_thread_info();
3610 #ifdef CONFIG_PREEMPT_BKL
3611 struct task_struct *task = current;
3612 int saved_lock_depth;
3614 /* Catch callers which need to be fixed */
3615 BUG_ON(ti->preempt_count || !irqs_disabled());
3618 add_preempt_count(PREEMPT_ACTIVE);
3620 * We keep the big kernel semaphore locked, but we
3621 * clear ->lock_depth so that schedule() doesnt
3622 * auto-release the semaphore:
3624 #ifdef CONFIG_PREEMPT_BKL
3625 saved_lock_depth = task->lock_depth;
3626 task->lock_depth = -1;
3630 local_irq_disable();
3631 #ifdef CONFIG_PREEMPT_BKL
3632 task->lock_depth = saved_lock_depth;
3634 sub_preempt_count(PREEMPT_ACTIVE);
3636 /* we could miss a preemption opportunity between schedule and now */
3638 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3642 #endif /* CONFIG_PREEMPT */
3644 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3647 return try_to_wake_up(curr->private, mode, sync);
3649 EXPORT_SYMBOL(default_wake_function);
3652 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3653 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3654 * number) then we wake all the non-exclusive tasks and one exclusive task.
3656 * There are circumstances in which we can try to wake a task which has already
3657 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3658 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3660 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3661 int nr_exclusive, int sync, void *key)
3663 struct list_head *tmp, *next;
3665 list_for_each_safe(tmp, next, &q->task_list) {
3666 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3667 unsigned flags = curr->flags;
3669 if (curr->func(curr, mode, sync, key) &&
3670 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3676 * __wake_up - wake up threads blocked on a waitqueue.
3678 * @mode: which threads
3679 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3680 * @key: is directly passed to the wakeup function
3682 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3683 int nr_exclusive, void *key)
3685 unsigned long flags;
3687 spin_lock_irqsave(&q->lock, flags);
3688 __wake_up_common(q, mode, nr_exclusive, 0, key);
3689 spin_unlock_irqrestore(&q->lock, flags);
3691 EXPORT_SYMBOL(__wake_up);
3694 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3696 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3698 __wake_up_common(q, mode, 1, 0, NULL);
3702 * __wake_up_sync - wake up threads blocked on a waitqueue.
3704 * @mode: which threads
3705 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3707 * The sync wakeup differs that the waker knows that it will schedule
3708 * away soon, so while the target thread will be woken up, it will not
3709 * be migrated to another CPU - ie. the two threads are 'synchronized'
3710 * with each other. This can prevent needless bouncing between CPUs.
3712 * On UP it can prevent extra preemption.
3715 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3717 unsigned long flags;
3723 if (unlikely(!nr_exclusive))
3726 spin_lock_irqsave(&q->lock, flags);
3727 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3728 spin_unlock_irqrestore(&q->lock, flags);
3730 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3732 void fastcall complete(struct completion *x)
3734 unsigned long flags;
3736 spin_lock_irqsave(&x->wait.lock, flags);
3738 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3740 spin_unlock_irqrestore(&x->wait.lock, flags);
3742 EXPORT_SYMBOL(complete);
3744 void fastcall complete_all(struct completion *x)
3746 unsigned long flags;
3748 spin_lock_irqsave(&x->wait.lock, flags);
3749 x->done += UINT_MAX/2;
3750 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3752 spin_unlock_irqrestore(&x->wait.lock, flags);
3754 EXPORT_SYMBOL(complete_all);
3756 void fastcall __sched wait_for_completion(struct completion *x)
3760 spin_lock_irq(&x->wait.lock);
3762 DECLARE_WAITQUEUE(wait, current);
3764 wait.flags |= WQ_FLAG_EXCLUSIVE;
3765 __add_wait_queue_tail(&x->wait, &wait);
3767 __set_current_state(TASK_UNINTERRUPTIBLE);
3768 spin_unlock_irq(&x->wait.lock);
3770 spin_lock_irq(&x->wait.lock);
3772 __remove_wait_queue(&x->wait, &wait);
3775 spin_unlock_irq(&x->wait.lock);
3777 EXPORT_SYMBOL(wait_for_completion);
3779 unsigned long fastcall __sched
3780 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3784 spin_lock_irq(&x->wait.lock);
3786 DECLARE_WAITQUEUE(wait, current);
3788 wait.flags |= WQ_FLAG_EXCLUSIVE;
3789 __add_wait_queue_tail(&x->wait, &wait);
3791 __set_current_state(TASK_UNINTERRUPTIBLE);
3792 spin_unlock_irq(&x->wait.lock);
3793 timeout = schedule_timeout(timeout);
3794 spin_lock_irq(&x->wait.lock);
3796 __remove_wait_queue(&x->wait, &wait);
3800 __remove_wait_queue(&x->wait, &wait);
3804 spin_unlock_irq(&x->wait.lock);
3807 EXPORT_SYMBOL(wait_for_completion_timeout);
3809 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3815 spin_lock_irq(&x->wait.lock);
3817 DECLARE_WAITQUEUE(wait, current);
3819 wait.flags |= WQ_FLAG_EXCLUSIVE;
3820 __add_wait_queue_tail(&x->wait, &wait);
3822 if (signal_pending(current)) {
3824 __remove_wait_queue(&x->wait, &wait);
3827 __set_current_state(TASK_INTERRUPTIBLE);
3828 spin_unlock_irq(&x->wait.lock);
3830 spin_lock_irq(&x->wait.lock);
3832 __remove_wait_queue(&x->wait, &wait);
3836 spin_unlock_irq(&x->wait.lock);
3840 EXPORT_SYMBOL(wait_for_completion_interruptible);
3842 unsigned long fastcall __sched
3843 wait_for_completion_interruptible_timeout(struct completion *x,
3844 unsigned long timeout)
3848 spin_lock_irq(&x->wait.lock);
3850 DECLARE_WAITQUEUE(wait, current);
3852 wait.flags |= WQ_FLAG_EXCLUSIVE;
3853 __add_wait_queue_tail(&x->wait, &wait);
3855 if (signal_pending(current)) {
3856 timeout = -ERESTARTSYS;
3857 __remove_wait_queue(&x->wait, &wait);
3860 __set_current_state(TASK_INTERRUPTIBLE);
3861 spin_unlock_irq(&x->wait.lock);
3862 timeout = schedule_timeout(timeout);
3863 spin_lock_irq(&x->wait.lock);
3865 __remove_wait_queue(&x->wait, &wait);
3869 __remove_wait_queue(&x->wait, &wait);
3873 spin_unlock_irq(&x->wait.lock);
3876 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3879 #define SLEEP_ON_VAR \
3880 unsigned long flags; \
3881 wait_queue_t wait; \
3882 init_waitqueue_entry(&wait, current);
3884 #define SLEEP_ON_HEAD \
3885 spin_lock_irqsave(&q->lock,flags); \
3886 __add_wait_queue(q, &wait); \
3887 spin_unlock(&q->lock);
3889 #define SLEEP_ON_TAIL \
3890 spin_lock_irq(&q->lock); \
3891 __remove_wait_queue(q, &wait); \
3892 spin_unlock_irqrestore(&q->lock, flags);
3894 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3898 current->state = TASK_INTERRUPTIBLE;
3904 EXPORT_SYMBOL(interruptible_sleep_on);
3906 long fastcall __sched
3907 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3911 current->state = TASK_INTERRUPTIBLE;
3914 timeout = schedule_timeout(timeout);
3919 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3921 void fastcall __sched sleep_on(wait_queue_head_t *q)
3925 current->state = TASK_UNINTERRUPTIBLE;
3931 EXPORT_SYMBOL(sleep_on);
3933 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3937 current->state = TASK_UNINTERRUPTIBLE;
3940 timeout = schedule_timeout(timeout);
3946 EXPORT_SYMBOL(sleep_on_timeout);
3948 #ifdef CONFIG_RT_MUTEXES
3951 * rt_mutex_setprio - set the current priority of a task
3953 * @prio: prio value (kernel-internal form)
3955 * This function changes the 'effective' priority of a task. It does
3956 * not touch ->normal_prio like __setscheduler().
3958 * Used by the rt_mutex code to implement priority inheritance logic.
3960 void rt_mutex_setprio(struct task_struct *p, int prio)
3962 struct prio_array *array;
3963 unsigned long flags;
3967 BUG_ON(prio < 0 || prio > MAX_PRIO);
3969 rq = task_rq_lock(p, &flags);
3974 dequeue_task(p, array);
3979 * If changing to an RT priority then queue it
3980 * in the active array!
3984 enqueue_task(p, array);
3986 * Reschedule if we are currently running on this runqueue and
3987 * our priority decreased, or if we are not currently running on
3988 * this runqueue and our priority is higher than the current's
3990 if (task_running(rq, p)) {
3991 if (p->prio > oldprio)
3992 resched_task(rq->curr);
3993 } else if (TASK_PREEMPTS_CURR(p, rq))
3994 resched_task(rq->curr);
3996 task_rq_unlock(rq, &flags);
4001 void set_user_nice(struct task_struct *p, long nice)
4003 struct prio_array *array;
4004 int old_prio, delta;
4005 unsigned long flags;
4008 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4011 * We have to be careful, if called from sys_setpriority(),
4012 * the task might be in the middle of scheduling on another CPU.
4014 rq = task_rq_lock(p, &flags);
4016 * The RT priorities are set via sched_setscheduler(), but we still
4017 * allow the 'normal' nice value to be set - but as expected
4018 * it wont have any effect on scheduling until the task is
4019 * not SCHED_NORMAL/SCHED_BATCH:
4021 if (task_has_rt_policy(p)) {
4022 p->static_prio = NICE_TO_PRIO(nice);
4027 dequeue_task(p, array);
4028 dec_raw_weighted_load(rq, p);
4031 p->static_prio = NICE_TO_PRIO(nice);
4034 p->prio = effective_prio(p);
4035 delta = p->prio - old_prio;
4038 enqueue_task(p, array);
4039 inc_raw_weighted_load(rq, p);
4041 * If the task increased its priority or is running and
4042 * lowered its priority, then reschedule its CPU:
4044 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4045 resched_task(rq->curr);
4048 task_rq_unlock(rq, &flags);
4050 EXPORT_SYMBOL(set_user_nice);
4053 * can_nice - check if a task can reduce its nice value
4057 int can_nice(const struct task_struct *p, const int nice)
4059 /* convert nice value [19,-20] to rlimit style value [1,40] */
4060 int nice_rlim = 20 - nice;
4062 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4063 capable(CAP_SYS_NICE));
4066 #ifdef __ARCH_WANT_SYS_NICE
4069 * sys_nice - change the priority of the current process.
4070 * @increment: priority increment
4072 * sys_setpriority is a more generic, but much slower function that
4073 * does similar things.
4075 asmlinkage long sys_nice(int increment)
4080 * Setpriority might change our priority at the same moment.
4081 * We don't have to worry. Conceptually one call occurs first
4082 * and we have a single winner.
4084 if (increment < -40)
4089 nice = PRIO_TO_NICE(current->static_prio) + increment;
4095 if (increment < 0 && !can_nice(current, nice))
4098 retval = security_task_setnice(current, nice);
4102 set_user_nice(current, nice);
4109 * task_prio - return the priority value of a given task.
4110 * @p: the task in question.
4112 * This is the priority value as seen by users in /proc.
4113 * RT tasks are offset by -200. Normal tasks are centered
4114 * around 0, value goes from -16 to +15.
4116 int task_prio(const struct task_struct *p)
4118 return p->prio - MAX_RT_PRIO;
4122 * task_nice - return the nice value of a given task.
4123 * @p: the task in question.
4125 int task_nice(const struct task_struct *p)
4127 return TASK_NICE(p);
4129 EXPORT_SYMBOL_GPL(task_nice);
4132 * idle_cpu - is a given cpu idle currently?
4133 * @cpu: the processor in question.
4135 int idle_cpu(int cpu)
4137 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4141 * idle_task - return the idle task for a given cpu.
4142 * @cpu: the processor in question.
4144 struct task_struct *idle_task(int cpu)
4146 return cpu_rq(cpu)->idle;
4150 * find_process_by_pid - find a process with a matching PID value.
4151 * @pid: the pid in question.
4153 static inline struct task_struct *find_process_by_pid(pid_t pid)
4155 return pid ? find_task_by_pid(pid) : current;
4158 /* Actually do priority change: must hold rq lock. */
4159 static void __setscheduler(struct task_struct *p, int policy, int prio)
4164 p->rt_priority = prio;
4165 p->normal_prio = normal_prio(p);
4166 /* we are holding p->pi_lock already */
4167 p->prio = rt_mutex_getprio(p);
4172 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4173 * @p: the task in question.
4174 * @policy: new policy.
4175 * @param: structure containing the new RT priority.
4177 * NOTE that the task may be already dead.
4179 int sched_setscheduler(struct task_struct *p, int policy,
4180 struct sched_param *param)
4182 int retval, oldprio, oldpolicy = -1;
4183 struct prio_array *array;
4184 unsigned long flags;
4187 /* may grab non-irq protected spin_locks */
4188 BUG_ON(in_interrupt());
4190 /* double check policy once rq lock held */
4192 policy = oldpolicy = p->policy;
4193 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4194 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4197 * Valid priorities for SCHED_FIFO and SCHED_RR are
4198 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4201 if (param->sched_priority < 0 ||
4202 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4203 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4205 if (rt_policy(policy) != (param->sched_priority != 0))
4209 * Allow unprivileged RT tasks to decrease priority:
4211 if (!capable(CAP_SYS_NICE)) {
4212 if (rt_policy(policy)) {
4213 unsigned long rlim_rtprio;
4214 unsigned long flags;
4216 if (!lock_task_sighand(p, &flags))
4218 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4219 unlock_task_sighand(p, &flags);
4221 /* can't set/change the rt policy */
4222 if (policy != p->policy && !rlim_rtprio)
4225 /* can't increase priority */
4226 if (param->sched_priority > p->rt_priority &&
4227 param->sched_priority > rlim_rtprio)
4231 /* can't change other user's priorities */
4232 if ((current->euid != p->euid) &&
4233 (current->euid != p->uid))
4237 retval = security_task_setscheduler(p, policy, param);
4241 * make sure no PI-waiters arrive (or leave) while we are
4242 * changing the priority of the task:
4244 spin_lock_irqsave(&p->pi_lock, flags);
4246 * To be able to change p->policy safely, the apropriate
4247 * runqueue lock must be held.
4249 rq = __task_rq_lock(p);
4250 /* recheck policy now with rq lock held */
4251 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4252 policy = oldpolicy = -1;
4253 __task_rq_unlock(rq);
4254 spin_unlock_irqrestore(&p->pi_lock, flags);
4259 deactivate_task(p, rq);
4261 __setscheduler(p, policy, param->sched_priority);
4263 __activate_task(p, rq);
4265 * Reschedule if we are currently running on this runqueue and
4266 * our priority decreased, or if we are not currently running on
4267 * this runqueue and our priority is higher than the current's
4269 if (task_running(rq, p)) {
4270 if (p->prio > oldprio)
4271 resched_task(rq->curr);
4272 } else if (TASK_PREEMPTS_CURR(p, rq))
4273 resched_task(rq->curr);
4275 __task_rq_unlock(rq);
4276 spin_unlock_irqrestore(&p->pi_lock, flags);
4278 rt_mutex_adjust_pi(p);
4282 EXPORT_SYMBOL_GPL(sched_setscheduler);
4285 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4287 struct sched_param lparam;
4288 struct task_struct *p;
4291 if (!param || pid < 0)
4293 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4298 p = find_process_by_pid(pid);
4300 retval = sched_setscheduler(p, policy, &lparam);
4307 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4308 * @pid: the pid in question.
4309 * @policy: new policy.
4310 * @param: structure containing the new RT priority.
4312 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4313 struct sched_param __user *param)
4315 /* negative values for policy are not valid */
4319 return do_sched_setscheduler(pid, policy, param);
4323 * sys_sched_setparam - set/change the RT priority of a thread
4324 * @pid: the pid in question.
4325 * @param: structure containing the new RT priority.
4327 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4329 return do_sched_setscheduler(pid, -1, param);
4333 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4334 * @pid: the pid in question.
4336 asmlinkage long sys_sched_getscheduler(pid_t pid)
4338 struct task_struct *p;
4339 int retval = -EINVAL;
4345 read_lock(&tasklist_lock);
4346 p = find_process_by_pid(pid);
4348 retval = security_task_getscheduler(p);
4352 read_unlock(&tasklist_lock);
4359 * sys_sched_getscheduler - get the RT priority of a thread
4360 * @pid: the pid in question.
4361 * @param: structure containing the RT priority.
4363 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4365 struct sched_param lp;
4366 struct task_struct *p;
4367 int retval = -EINVAL;
4369 if (!param || pid < 0)
4372 read_lock(&tasklist_lock);
4373 p = find_process_by_pid(pid);
4378 retval = security_task_getscheduler(p);
4382 lp.sched_priority = p->rt_priority;
4383 read_unlock(&tasklist_lock);
4386 * This one might sleep, we cannot do it with a spinlock held ...
4388 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4394 read_unlock(&tasklist_lock);
4398 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4400 cpumask_t cpus_allowed;
4401 struct task_struct *p;
4404 mutex_lock(&sched_hotcpu_mutex);
4405 read_lock(&tasklist_lock);
4407 p = find_process_by_pid(pid);
4409 read_unlock(&tasklist_lock);
4410 mutex_unlock(&sched_hotcpu_mutex);
4415 * It is not safe to call set_cpus_allowed with the
4416 * tasklist_lock held. We will bump the task_struct's
4417 * usage count and then drop tasklist_lock.
4420 read_unlock(&tasklist_lock);
4423 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4424 !capable(CAP_SYS_NICE))
4427 retval = security_task_setscheduler(p, 0, NULL);
4431 cpus_allowed = cpuset_cpus_allowed(p);
4432 cpus_and(new_mask, new_mask, cpus_allowed);
4433 retval = set_cpus_allowed(p, new_mask);
4437 mutex_unlock(&sched_hotcpu_mutex);
4441 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4442 cpumask_t *new_mask)
4444 if (len < sizeof(cpumask_t)) {
4445 memset(new_mask, 0, sizeof(cpumask_t));
4446 } else if (len > sizeof(cpumask_t)) {
4447 len = sizeof(cpumask_t);
4449 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4453 * sys_sched_setaffinity - set the cpu affinity of a process
4454 * @pid: pid of the process
4455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4456 * @user_mask_ptr: user-space pointer to the new cpu mask
4458 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4459 unsigned long __user *user_mask_ptr)
4464 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4468 return sched_setaffinity(pid, new_mask);
4472 * Represents all cpu's present in the system
4473 * In systems capable of hotplug, this map could dynamically grow
4474 * as new cpu's are detected in the system via any platform specific
4475 * method, such as ACPI for e.g.
4478 cpumask_t cpu_present_map __read_mostly;
4479 EXPORT_SYMBOL(cpu_present_map);
4482 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4483 EXPORT_SYMBOL(cpu_online_map);
4485 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4486 EXPORT_SYMBOL(cpu_possible_map);
4489 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4491 struct task_struct *p;
4494 mutex_lock(&sched_hotcpu_mutex);
4495 read_lock(&tasklist_lock);
4498 p = find_process_by_pid(pid);
4502 retval = security_task_getscheduler(p);
4506 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4509 read_unlock(&tasklist_lock);
4510 mutex_unlock(&sched_hotcpu_mutex);
4518 * sys_sched_getaffinity - get the cpu affinity of a process
4519 * @pid: pid of the process
4520 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4521 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4523 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4524 unsigned long __user *user_mask_ptr)
4529 if (len < sizeof(cpumask_t))
4532 ret = sched_getaffinity(pid, &mask);
4536 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4539 return sizeof(cpumask_t);
4543 * sys_sched_yield - yield the current processor to other threads.
4545 * This function yields the current CPU by moving the calling thread
4546 * to the expired array. If there are no other threads running on this
4547 * CPU then this function will return.
4549 asmlinkage long sys_sched_yield(void)
4551 struct rq *rq = this_rq_lock();
4552 struct prio_array *array = current->array, *target = rq->expired;
4554 schedstat_inc(rq, yld_cnt);
4556 * We implement yielding by moving the task into the expired
4559 * (special rule: RT tasks will just roundrobin in the active
4562 if (rt_task(current))
4563 target = rq->active;
4565 if (array->nr_active == 1) {
4566 schedstat_inc(rq, yld_act_empty);
4567 if (!rq->expired->nr_active)
4568 schedstat_inc(rq, yld_both_empty);
4569 } else if (!rq->expired->nr_active)
4570 schedstat_inc(rq, yld_exp_empty);
4572 if (array != target) {
4573 dequeue_task(current, array);
4574 enqueue_task(current, target);
4577 * requeue_task is cheaper so perform that if possible.
4579 requeue_task(current, array);
4582 * Since we are going to call schedule() anyway, there's
4583 * no need to preempt or enable interrupts:
4585 __release(rq->lock);
4586 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4587 _raw_spin_unlock(&rq->lock);
4588 preempt_enable_no_resched();
4595 static void __cond_resched(void)
4597 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4598 __might_sleep(__FILE__, __LINE__);
4601 * The BKS might be reacquired before we have dropped
4602 * PREEMPT_ACTIVE, which could trigger a second
4603 * cond_resched() call.
4606 add_preempt_count(PREEMPT_ACTIVE);
4608 sub_preempt_count(PREEMPT_ACTIVE);
4609 } while (need_resched());
4612 int __sched cond_resched(void)
4614 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4615 system_state == SYSTEM_RUNNING) {
4621 EXPORT_SYMBOL(cond_resched);
4624 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4625 * call schedule, and on return reacquire the lock.
4627 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4628 * operations here to prevent schedule() from being called twice (once via
4629 * spin_unlock(), once by hand).
4631 int cond_resched_lock(spinlock_t *lock)
4635 if (need_lockbreak(lock)) {
4641 if (need_resched() && system_state == SYSTEM_RUNNING) {
4642 spin_release(&lock->dep_map, 1, _THIS_IP_);
4643 _raw_spin_unlock(lock);
4644 preempt_enable_no_resched();
4651 EXPORT_SYMBOL(cond_resched_lock);
4653 int __sched cond_resched_softirq(void)
4655 BUG_ON(!in_softirq());
4657 if (need_resched() && system_state == SYSTEM_RUNNING) {
4665 EXPORT_SYMBOL(cond_resched_softirq);
4668 * yield - yield the current processor to other threads.
4670 * This is a shortcut for kernel-space yielding - it marks the
4671 * thread runnable and calls sys_sched_yield().
4673 void __sched yield(void)
4675 set_current_state(TASK_RUNNING);
4678 EXPORT_SYMBOL(yield);
4681 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4682 * that process accounting knows that this is a task in IO wait state.
4684 * But don't do that if it is a deliberate, throttling IO wait (this task
4685 * has set its backing_dev_info: the queue against which it should throttle)
4687 void __sched io_schedule(void)
4689 struct rq *rq = &__raw_get_cpu_var(runqueues);
4691 delayacct_blkio_start();
4692 atomic_inc(&rq->nr_iowait);
4694 atomic_dec(&rq->nr_iowait);
4695 delayacct_blkio_end();
4697 EXPORT_SYMBOL(io_schedule);
4699 long __sched io_schedule_timeout(long timeout)
4701 struct rq *rq = &__raw_get_cpu_var(runqueues);
4704 delayacct_blkio_start();
4705 atomic_inc(&rq->nr_iowait);
4706 ret = schedule_timeout(timeout);
4707 atomic_dec(&rq->nr_iowait);
4708 delayacct_blkio_end();
4713 * sys_sched_get_priority_max - return maximum RT priority.
4714 * @policy: scheduling class.
4716 * this syscall returns the maximum rt_priority that can be used
4717 * by a given scheduling class.
4719 asmlinkage long sys_sched_get_priority_max(int policy)
4726 ret = MAX_USER_RT_PRIO-1;
4737 * sys_sched_get_priority_min - return minimum RT priority.
4738 * @policy: scheduling class.
4740 * this syscall returns the minimum rt_priority that can be used
4741 * by a given scheduling class.
4743 asmlinkage long sys_sched_get_priority_min(int policy)
4760 * sys_sched_rr_get_interval - return the default timeslice of a process.
4761 * @pid: pid of the process.
4762 * @interval: userspace pointer to the timeslice value.
4764 * this syscall writes the default timeslice value of a given process
4765 * into the user-space timespec buffer. A value of '0' means infinity.
4768 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4770 struct task_struct *p;
4771 int retval = -EINVAL;
4778 read_lock(&tasklist_lock);
4779 p = find_process_by_pid(pid);
4783 retval = security_task_getscheduler(p);
4787 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4788 0 : task_timeslice(p), &t);
4789 read_unlock(&tasklist_lock);
4790 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4794 read_unlock(&tasklist_lock);
4798 static const char stat_nam[] = "RSDTtZX";
4800 static void show_task(struct task_struct *p)
4802 unsigned long free = 0;
4805 state = p->state ? __ffs(p->state) + 1 : 0;
4806 printk("%-13.13s %c", p->comm,
4807 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4808 #if (BITS_PER_LONG == 32)
4809 if (state == TASK_RUNNING)
4810 printk(" running ");
4812 printk(" %08lX ", thread_saved_pc(p));
4814 if (state == TASK_RUNNING)
4815 printk(" running task ");
4817 printk(" %016lx ", thread_saved_pc(p));
4819 #ifdef CONFIG_DEBUG_STACK_USAGE
4821 unsigned long *n = end_of_stack(p);
4824 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4827 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4829 printk(" (L-TLB)\n");
4831 printk(" (NOTLB)\n");
4833 if (state != TASK_RUNNING)
4834 show_stack(p, NULL);
4837 void show_state_filter(unsigned long state_filter)
4839 struct task_struct *g, *p;
4841 #if (BITS_PER_LONG == 32)
4844 printk(" task PC stack pid father child younger older\n");
4848 printk(" task PC stack pid father child younger older\n");
4850 read_lock(&tasklist_lock);
4851 do_each_thread(g, p) {
4853 * reset the NMI-timeout, listing all files on a slow
4854 * console might take alot of time:
4856 touch_nmi_watchdog();
4857 if (!state_filter || (p->state & state_filter))
4859 } while_each_thread(g, p);
4861 touch_all_softlockup_watchdogs();
4863 read_unlock(&tasklist_lock);
4865 * Only show locks if all tasks are dumped:
4867 if (state_filter == -1)
4868 debug_show_all_locks();
4871 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4877 * init_idle - set up an idle thread for a given CPU
4878 * @idle: task in question
4879 * @cpu: cpu the idle task belongs to
4881 * NOTE: this function does not set the idle thread's NEED_RESCHED
4882 * flag, to make booting more robust.
4884 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4886 struct rq *rq = cpu_rq(cpu);
4887 unsigned long flags;
4889 idle->timestamp = sched_clock();
4891 idle->prio = idle->normal_prio = MAX_PRIO;
4892 idle->state = TASK_RUNNING;
4893 idle->cpus_allowed = cpumask_of_cpu(cpu);
4894 set_task_cpu(idle, cpu);
4896 spin_lock_irqsave(&rq->lock, flags);
4897 rq->curr = rq->idle = idle;
4898 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4901 spin_unlock_irqrestore(&rq->lock, flags);
4903 /* Set the preempt count _outside_ the spinlocks! */
4904 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4905 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4907 task_thread_info(idle)->preempt_count = 0;
4912 * In a system that switches off the HZ timer nohz_cpu_mask
4913 * indicates which cpus entered this state. This is used
4914 * in the rcu update to wait only for active cpus. For system
4915 * which do not switch off the HZ timer nohz_cpu_mask should
4916 * always be CPU_MASK_NONE.
4918 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4922 * This is how migration works:
4924 * 1) we queue a struct migration_req structure in the source CPU's
4925 * runqueue and wake up that CPU's migration thread.
4926 * 2) we down() the locked semaphore => thread blocks.
4927 * 3) migration thread wakes up (implicitly it forces the migrated
4928 * thread off the CPU)
4929 * 4) it gets the migration request and checks whether the migrated
4930 * task is still in the wrong runqueue.
4931 * 5) if it's in the wrong runqueue then the migration thread removes
4932 * it and puts it into the right queue.
4933 * 6) migration thread up()s the semaphore.
4934 * 7) we wake up and the migration is done.
4938 * Change a given task's CPU affinity. Migrate the thread to a
4939 * proper CPU and schedule it away if the CPU it's executing on
4940 * is removed from the allowed bitmask.
4942 * NOTE: the caller must have a valid reference to the task, the
4943 * task must not exit() & deallocate itself prematurely. The
4944 * call is not atomic; no spinlocks may be held.
4946 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4948 struct migration_req req;
4949 unsigned long flags;
4953 rq = task_rq_lock(p, &flags);
4954 if (!cpus_intersects(new_mask, cpu_online_map)) {
4959 p->cpus_allowed = new_mask;
4960 /* Can the task run on the task's current CPU? If so, we're done */
4961 if (cpu_isset(task_cpu(p), new_mask))
4964 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4965 /* Need help from migration thread: drop lock and wait. */
4966 task_rq_unlock(rq, &flags);
4967 wake_up_process(rq->migration_thread);
4968 wait_for_completion(&req.done);
4969 tlb_migrate_finish(p->mm);
4973 task_rq_unlock(rq, &flags);
4977 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4980 * Move (not current) task off this cpu, onto dest cpu. We're doing
4981 * this because either it can't run here any more (set_cpus_allowed()
4982 * away from this CPU, or CPU going down), or because we're
4983 * attempting to rebalance this task on exec (sched_exec).
4985 * So we race with normal scheduler movements, but that's OK, as long
4986 * as the task is no longer on this CPU.
4988 * Returns non-zero if task was successfully migrated.
4990 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4992 struct rq *rq_dest, *rq_src;
4995 if (unlikely(cpu_is_offline(dest_cpu)))
4998 rq_src = cpu_rq(src_cpu);
4999 rq_dest = cpu_rq(dest_cpu);
5001 double_rq_lock(rq_src, rq_dest);
5002 /* Already moved. */
5003 if (task_cpu(p) != src_cpu)
5005 /* Affinity changed (again). */
5006 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5009 set_task_cpu(p, dest_cpu);
5012 * Sync timestamp with rq_dest's before activating.
5013 * The same thing could be achieved by doing this step
5014 * afterwards, and pretending it was a local activate.
5015 * This way is cleaner and logically correct.
5017 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5018 + rq_dest->most_recent_timestamp;
5019 deactivate_task(p, rq_src);
5020 __activate_task(p, rq_dest);
5021 if (TASK_PREEMPTS_CURR(p, rq_dest))
5022 resched_task(rq_dest->curr);
5026 double_rq_unlock(rq_src, rq_dest);
5031 * migration_thread - this is a highprio system thread that performs
5032 * thread migration by bumping thread off CPU then 'pushing' onto
5035 static int migration_thread(void *data)
5037 int cpu = (long)data;
5041 BUG_ON(rq->migration_thread != current);
5043 set_current_state(TASK_INTERRUPTIBLE);
5044 while (!kthread_should_stop()) {
5045 struct migration_req *req;
5046 struct list_head *head;
5050 spin_lock_irq(&rq->lock);
5052 if (cpu_is_offline(cpu)) {
5053 spin_unlock_irq(&rq->lock);
5057 if (rq->active_balance) {
5058 active_load_balance(rq, cpu);
5059 rq->active_balance = 0;
5062 head = &rq->migration_queue;
5064 if (list_empty(head)) {
5065 spin_unlock_irq(&rq->lock);
5067 set_current_state(TASK_INTERRUPTIBLE);
5070 req = list_entry(head->next, struct migration_req, list);
5071 list_del_init(head->next);
5073 spin_unlock(&rq->lock);
5074 __migrate_task(req->task, cpu, req->dest_cpu);
5077 complete(&req->done);
5079 __set_current_state(TASK_RUNNING);
5083 /* Wait for kthread_stop */
5084 set_current_state(TASK_INTERRUPTIBLE);
5085 while (!kthread_should_stop()) {
5087 set_current_state(TASK_INTERRUPTIBLE);
5089 __set_current_state(TASK_RUNNING);
5093 #ifdef CONFIG_HOTPLUG_CPU
5095 * Figure out where task on dead CPU should go, use force if neccessary.
5096 * NOTE: interrupts should be disabled by the caller
5098 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5100 unsigned long flags;
5107 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5108 cpus_and(mask, mask, p->cpus_allowed);
5109 dest_cpu = any_online_cpu(mask);
5111 /* On any allowed CPU? */
5112 if (dest_cpu == NR_CPUS)
5113 dest_cpu = any_online_cpu(p->cpus_allowed);
5115 /* No more Mr. Nice Guy. */
5116 if (dest_cpu == NR_CPUS) {
5117 rq = task_rq_lock(p, &flags);
5118 cpus_setall(p->cpus_allowed);
5119 dest_cpu = any_online_cpu(p->cpus_allowed);
5120 task_rq_unlock(rq, &flags);
5123 * Don't tell them about moving exiting tasks or
5124 * kernel threads (both mm NULL), since they never
5127 if (p->mm && printk_ratelimit())
5128 printk(KERN_INFO "process %d (%s) no "
5129 "longer affine to cpu%d\n",
5130 p->pid, p->comm, dead_cpu);
5132 if (!__migrate_task(p, dead_cpu, dest_cpu))
5137 * While a dead CPU has no uninterruptible tasks queued at this point,
5138 * it might still have a nonzero ->nr_uninterruptible counter, because
5139 * for performance reasons the counter is not stricly tracking tasks to
5140 * their home CPUs. So we just add the counter to another CPU's counter,
5141 * to keep the global sum constant after CPU-down:
5143 static void migrate_nr_uninterruptible(struct rq *rq_src)
5145 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5146 unsigned long flags;
5148 local_irq_save(flags);
5149 double_rq_lock(rq_src, rq_dest);
5150 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5151 rq_src->nr_uninterruptible = 0;
5152 double_rq_unlock(rq_src, rq_dest);
5153 local_irq_restore(flags);
5156 /* Run through task list and migrate tasks from the dead cpu. */
5157 static void migrate_live_tasks(int src_cpu)
5159 struct task_struct *p, *t;
5161 write_lock_irq(&tasklist_lock);
5163 do_each_thread(t, p) {
5167 if (task_cpu(p) == src_cpu)
5168 move_task_off_dead_cpu(src_cpu, p);
5169 } while_each_thread(t, p);
5171 write_unlock_irq(&tasklist_lock);
5174 /* Schedules idle task to be the next runnable task on current CPU.
5175 * It does so by boosting its priority to highest possible and adding it to
5176 * the _front_ of the runqueue. Used by CPU offline code.
5178 void sched_idle_next(void)
5180 int this_cpu = smp_processor_id();
5181 struct rq *rq = cpu_rq(this_cpu);
5182 struct task_struct *p = rq->idle;
5183 unsigned long flags;
5185 /* cpu has to be offline */
5186 BUG_ON(cpu_online(this_cpu));
5189 * Strictly not necessary since rest of the CPUs are stopped by now
5190 * and interrupts disabled on the current cpu.
5192 spin_lock_irqsave(&rq->lock, flags);
5194 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5196 /* Add idle task to the _front_ of its priority queue: */
5197 __activate_idle_task(p, rq);
5199 spin_unlock_irqrestore(&rq->lock, flags);
5203 * Ensures that the idle task is using init_mm right before its cpu goes
5206 void idle_task_exit(void)
5208 struct mm_struct *mm = current->active_mm;
5210 BUG_ON(cpu_online(smp_processor_id()));
5213 switch_mm(mm, &init_mm, current);
5217 /* called under rq->lock with disabled interrupts */
5218 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5220 struct rq *rq = cpu_rq(dead_cpu);
5222 /* Must be exiting, otherwise would be on tasklist. */
5223 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5225 /* Cannot have done final schedule yet: would have vanished. */
5226 BUG_ON(p->state == TASK_DEAD);
5231 * Drop lock around migration; if someone else moves it,
5232 * that's OK. No task can be added to this CPU, so iteration is
5234 * NOTE: interrupts should be left disabled --dev@
5236 spin_unlock(&rq->lock);
5237 move_task_off_dead_cpu(dead_cpu, p);
5238 spin_lock(&rq->lock);
5243 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5244 static void migrate_dead_tasks(unsigned int dead_cpu)
5246 struct rq *rq = cpu_rq(dead_cpu);
5247 unsigned int arr, i;
5249 for (arr = 0; arr < 2; arr++) {
5250 for (i = 0; i < MAX_PRIO; i++) {
5251 struct list_head *list = &rq->arrays[arr].queue[i];
5253 while (!list_empty(list))
5254 migrate_dead(dead_cpu, list_entry(list->next,
5255 struct task_struct, run_list));
5259 #endif /* CONFIG_HOTPLUG_CPU */
5262 * migration_call - callback that gets triggered when a CPU is added.
5263 * Here we can start up the necessary migration thread for the new CPU.
5265 static int __cpuinit
5266 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5268 struct task_struct *p;
5269 int cpu = (long)hcpu;
5270 unsigned long flags;
5274 case CPU_LOCK_ACQUIRE:
5275 mutex_lock(&sched_hotcpu_mutex);
5278 case CPU_UP_PREPARE:
5279 case CPU_UP_PREPARE_FROZEN:
5280 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5283 p->flags |= PF_NOFREEZE;
5284 kthread_bind(p, cpu);
5285 /* Must be high prio: stop_machine expects to yield to it. */
5286 rq = task_rq_lock(p, &flags);
5287 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5288 task_rq_unlock(rq, &flags);
5289 cpu_rq(cpu)->migration_thread = p;
5293 case CPU_ONLINE_FROZEN:
5294 /* Strictly unneccessary, as first user will wake it. */
5295 wake_up_process(cpu_rq(cpu)->migration_thread);
5298 #ifdef CONFIG_HOTPLUG_CPU
5299 case CPU_UP_CANCELED:
5300 case CPU_UP_CANCELED_FROZEN:
5301 if (!cpu_rq(cpu)->migration_thread)
5303 /* Unbind it from offline cpu so it can run. Fall thru. */
5304 kthread_bind(cpu_rq(cpu)->migration_thread,
5305 any_online_cpu(cpu_online_map));
5306 kthread_stop(cpu_rq(cpu)->migration_thread);
5307 cpu_rq(cpu)->migration_thread = NULL;
5311 case CPU_DEAD_FROZEN:
5312 migrate_live_tasks(cpu);
5314 kthread_stop(rq->migration_thread);
5315 rq->migration_thread = NULL;
5316 /* Idle task back to normal (off runqueue, low prio) */
5317 rq = task_rq_lock(rq->idle, &flags);
5318 deactivate_task(rq->idle, rq);
5319 rq->idle->static_prio = MAX_PRIO;
5320 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5321 migrate_dead_tasks(cpu);
5322 task_rq_unlock(rq, &flags);
5323 migrate_nr_uninterruptible(rq);
5324 BUG_ON(rq->nr_running != 0);
5326 /* No need to migrate the tasks: it was best-effort if
5327 * they didn't take sched_hotcpu_mutex. Just wake up
5328 * the requestors. */
5329 spin_lock_irq(&rq->lock);
5330 while (!list_empty(&rq->migration_queue)) {
5331 struct migration_req *req;
5333 req = list_entry(rq->migration_queue.next,
5334 struct migration_req, list);
5335 list_del_init(&req->list);
5336 complete(&req->done);
5338 spin_unlock_irq(&rq->lock);
5341 case CPU_LOCK_RELEASE:
5342 mutex_unlock(&sched_hotcpu_mutex);
5348 /* Register at highest priority so that task migration (migrate_all_tasks)
5349 * happens before everything else.
5351 static struct notifier_block __cpuinitdata migration_notifier = {
5352 .notifier_call = migration_call,
5356 int __init migration_init(void)
5358 void *cpu = (void *)(long)smp_processor_id();
5361 /* Start one for the boot CPU: */
5362 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5363 BUG_ON(err == NOTIFY_BAD);
5364 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5365 register_cpu_notifier(&migration_notifier);
5373 /* Number of possible processor ids */
5374 int nr_cpu_ids __read_mostly = NR_CPUS;
5375 EXPORT_SYMBOL(nr_cpu_ids);
5377 #undef SCHED_DOMAIN_DEBUG
5378 #ifdef SCHED_DOMAIN_DEBUG
5379 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5384 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5388 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5393 struct sched_group *group = sd->groups;
5394 cpumask_t groupmask;
5396 cpumask_scnprintf(str, NR_CPUS, sd->span);
5397 cpus_clear(groupmask);
5400 for (i = 0; i < level + 1; i++)
5402 printk("domain %d: ", level);
5404 if (!(sd->flags & SD_LOAD_BALANCE)) {
5405 printk("does not load-balance\n");
5407 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5412 printk("span %s\n", str);
5414 if (!cpu_isset(cpu, sd->span))
5415 printk(KERN_ERR "ERROR: domain->span does not contain "
5417 if (!cpu_isset(cpu, group->cpumask))
5418 printk(KERN_ERR "ERROR: domain->groups does not contain"
5422 for (i = 0; i < level + 2; i++)
5428 printk(KERN_ERR "ERROR: group is NULL\n");
5432 if (!group->__cpu_power) {
5434 printk(KERN_ERR "ERROR: domain->cpu_power not "
5438 if (!cpus_weight(group->cpumask)) {
5440 printk(KERN_ERR "ERROR: empty group\n");
5443 if (cpus_intersects(groupmask, group->cpumask)) {
5445 printk(KERN_ERR "ERROR: repeated CPUs\n");
5448 cpus_or(groupmask, groupmask, group->cpumask);
5450 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5453 group = group->next;
5454 } while (group != sd->groups);
5457 if (!cpus_equal(sd->span, groupmask))
5458 printk(KERN_ERR "ERROR: groups don't span "
5466 if (!cpus_subset(groupmask, sd->span))
5467 printk(KERN_ERR "ERROR: parent span is not a superset "
5468 "of domain->span\n");
5473 # define sched_domain_debug(sd, cpu) do { } while (0)
5476 static int sd_degenerate(struct sched_domain *sd)
5478 if (cpus_weight(sd->span) == 1)
5481 /* Following flags need at least 2 groups */
5482 if (sd->flags & (SD_LOAD_BALANCE |
5483 SD_BALANCE_NEWIDLE |
5487 SD_SHARE_PKG_RESOURCES)) {
5488 if (sd->groups != sd->groups->next)
5492 /* Following flags don't use groups */
5493 if (sd->flags & (SD_WAKE_IDLE |
5502 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5504 unsigned long cflags = sd->flags, pflags = parent->flags;
5506 if (sd_degenerate(parent))
5509 if (!cpus_equal(sd->span, parent->span))
5512 /* Does parent contain flags not in child? */
5513 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5514 if (cflags & SD_WAKE_AFFINE)
5515 pflags &= ~SD_WAKE_BALANCE;
5516 /* Flags needing groups don't count if only 1 group in parent */
5517 if (parent->groups == parent->groups->next) {
5518 pflags &= ~(SD_LOAD_BALANCE |
5519 SD_BALANCE_NEWIDLE |
5523 SD_SHARE_PKG_RESOURCES);
5525 if (~cflags & pflags)
5532 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5533 * hold the hotplug lock.
5535 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5537 struct rq *rq = cpu_rq(cpu);
5538 struct sched_domain *tmp;
5540 /* Remove the sched domains which do not contribute to scheduling. */
5541 for (tmp = sd; tmp; tmp = tmp->parent) {
5542 struct sched_domain *parent = tmp->parent;
5545 if (sd_parent_degenerate(tmp, parent)) {
5546 tmp->parent = parent->parent;
5548 parent->parent->child = tmp;
5552 if (sd && sd_degenerate(sd)) {
5558 sched_domain_debug(sd, cpu);
5560 rcu_assign_pointer(rq->sd, sd);
5563 /* cpus with isolated domains */
5564 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5566 /* Setup the mask of cpus configured for isolated domains */
5567 static int __init isolated_cpu_setup(char *str)
5569 int ints[NR_CPUS], i;
5571 str = get_options(str, ARRAY_SIZE(ints), ints);
5572 cpus_clear(cpu_isolated_map);
5573 for (i = 1; i <= ints[0]; i++)
5574 if (ints[i] < NR_CPUS)
5575 cpu_set(ints[i], cpu_isolated_map);
5579 __setup ("isolcpus=", isolated_cpu_setup);
5582 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5583 * to a function which identifies what group(along with sched group) a CPU
5584 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5585 * (due to the fact that we keep track of groups covered with a cpumask_t).
5587 * init_sched_build_groups will build a circular linked list of the groups
5588 * covered by the given span, and will set each group's ->cpumask correctly,
5589 * and ->cpu_power to 0.
5592 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5593 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5594 struct sched_group **sg))
5596 struct sched_group *first = NULL, *last = NULL;
5597 cpumask_t covered = CPU_MASK_NONE;
5600 for_each_cpu_mask(i, span) {
5601 struct sched_group *sg;
5602 int group = group_fn(i, cpu_map, &sg);
5605 if (cpu_isset(i, covered))
5608 sg->cpumask = CPU_MASK_NONE;
5609 sg->__cpu_power = 0;
5611 for_each_cpu_mask(j, span) {
5612 if (group_fn(j, cpu_map, NULL) != group)
5615 cpu_set(j, covered);
5616 cpu_set(j, sg->cpumask);
5627 #define SD_NODES_PER_DOMAIN 16
5632 * find_next_best_node - find the next node to include in a sched_domain
5633 * @node: node whose sched_domain we're building
5634 * @used_nodes: nodes already in the sched_domain
5636 * Find the next node to include in a given scheduling domain. Simply
5637 * finds the closest node not already in the @used_nodes map.
5639 * Should use nodemask_t.
5641 static int find_next_best_node(int node, unsigned long *used_nodes)
5643 int i, n, val, min_val, best_node = 0;
5647 for (i = 0; i < MAX_NUMNODES; i++) {
5648 /* Start at @node */
5649 n = (node + i) % MAX_NUMNODES;
5651 if (!nr_cpus_node(n))
5654 /* Skip already used nodes */
5655 if (test_bit(n, used_nodes))
5658 /* Simple min distance search */
5659 val = node_distance(node, n);
5661 if (val < min_val) {
5667 set_bit(best_node, used_nodes);
5672 * sched_domain_node_span - get a cpumask for a node's sched_domain
5673 * @node: node whose cpumask we're constructing
5674 * @size: number of nodes to include in this span
5676 * Given a node, construct a good cpumask for its sched_domain to span. It
5677 * should be one that prevents unnecessary balancing, but also spreads tasks
5680 static cpumask_t sched_domain_node_span(int node)
5682 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5683 cpumask_t span, nodemask;
5687 bitmap_zero(used_nodes, MAX_NUMNODES);
5689 nodemask = node_to_cpumask(node);
5690 cpus_or(span, span, nodemask);
5691 set_bit(node, used_nodes);
5693 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5694 int next_node = find_next_best_node(node, used_nodes);
5696 nodemask = node_to_cpumask(next_node);
5697 cpus_or(span, span, nodemask);
5704 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5707 * SMT sched-domains:
5709 #ifdef CONFIG_SCHED_SMT
5710 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5711 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5713 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5714 struct sched_group **sg)
5717 *sg = &per_cpu(sched_group_cpus, cpu);
5723 * multi-core sched-domains:
5725 #ifdef CONFIG_SCHED_MC
5726 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5727 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5730 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5731 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5732 struct sched_group **sg)
5735 cpumask_t mask = cpu_sibling_map[cpu];
5736 cpus_and(mask, mask, *cpu_map);
5737 group = first_cpu(mask);
5739 *sg = &per_cpu(sched_group_core, group);
5742 #elif defined(CONFIG_SCHED_MC)
5743 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5744 struct sched_group **sg)
5747 *sg = &per_cpu(sched_group_core, cpu);
5752 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5753 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5755 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5756 struct sched_group **sg)
5759 #ifdef CONFIG_SCHED_MC
5760 cpumask_t mask = cpu_coregroup_map(cpu);
5761 cpus_and(mask, mask, *cpu_map);
5762 group = first_cpu(mask);
5763 #elif defined(CONFIG_SCHED_SMT)
5764 cpumask_t mask = cpu_sibling_map[cpu];
5765 cpus_and(mask, mask, *cpu_map);
5766 group = first_cpu(mask);
5771 *sg = &per_cpu(sched_group_phys, group);
5777 * The init_sched_build_groups can't handle what we want to do with node
5778 * groups, so roll our own. Now each node has its own list of groups which
5779 * gets dynamically allocated.
5781 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5782 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5784 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5785 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5787 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5788 struct sched_group **sg)
5790 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5793 cpus_and(nodemask, nodemask, *cpu_map);
5794 group = first_cpu(nodemask);
5797 *sg = &per_cpu(sched_group_allnodes, group);
5801 static void init_numa_sched_groups_power(struct sched_group *group_head)
5803 struct sched_group *sg = group_head;
5809 for_each_cpu_mask(j, sg->cpumask) {
5810 struct sched_domain *sd;
5812 sd = &per_cpu(phys_domains, j);
5813 if (j != first_cpu(sd->groups->cpumask)) {
5815 * Only add "power" once for each
5821 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5824 if (sg != group_head)
5830 /* Free memory allocated for various sched_group structures */
5831 static void free_sched_groups(const cpumask_t *cpu_map)
5835 for_each_cpu_mask(cpu, *cpu_map) {
5836 struct sched_group **sched_group_nodes
5837 = sched_group_nodes_bycpu[cpu];
5839 if (!sched_group_nodes)
5842 for (i = 0; i < MAX_NUMNODES; i++) {
5843 cpumask_t nodemask = node_to_cpumask(i);
5844 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5846 cpus_and(nodemask, nodemask, *cpu_map);
5847 if (cpus_empty(nodemask))
5857 if (oldsg != sched_group_nodes[i])
5860 kfree(sched_group_nodes);
5861 sched_group_nodes_bycpu[cpu] = NULL;
5865 static void free_sched_groups(const cpumask_t *cpu_map)
5871 * Initialize sched groups cpu_power.
5873 * cpu_power indicates the capacity of sched group, which is used while
5874 * distributing the load between different sched groups in a sched domain.
5875 * Typically cpu_power for all the groups in a sched domain will be same unless
5876 * there are asymmetries in the topology. If there are asymmetries, group
5877 * having more cpu_power will pickup more load compared to the group having
5880 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5881 * the maximum number of tasks a group can handle in the presence of other idle
5882 * or lightly loaded groups in the same sched domain.
5884 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5886 struct sched_domain *child;
5887 struct sched_group *group;
5889 WARN_ON(!sd || !sd->groups);
5891 if (cpu != first_cpu(sd->groups->cpumask))
5896 sd->groups->__cpu_power = 0;
5899 * For perf policy, if the groups in child domain share resources
5900 * (for example cores sharing some portions of the cache hierarchy
5901 * or SMT), then set this domain groups cpu_power such that each group
5902 * can handle only one task, when there are other idle groups in the
5903 * same sched domain.
5905 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5907 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5908 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5913 * add cpu_power of each child group to this groups cpu_power
5915 group = child->groups;
5917 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5918 group = group->next;
5919 } while (group != child->groups);
5923 * Build sched domains for a given set of cpus and attach the sched domains
5924 * to the individual cpus
5926 static int build_sched_domains(const cpumask_t *cpu_map)
5929 struct sched_domain *sd;
5931 struct sched_group **sched_group_nodes = NULL;
5932 int sd_allnodes = 0;
5935 * Allocate the per-node list of sched groups
5937 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
5939 if (!sched_group_nodes) {
5940 printk(KERN_WARNING "Can not alloc sched group node list\n");
5943 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5947 * Set up domains for cpus specified by the cpu_map.
5949 for_each_cpu_mask(i, *cpu_map) {
5950 struct sched_domain *sd = NULL, *p;
5951 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5953 cpus_and(nodemask, nodemask, *cpu_map);
5956 if (cpus_weight(*cpu_map)
5957 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5958 sd = &per_cpu(allnodes_domains, i);
5959 *sd = SD_ALLNODES_INIT;
5960 sd->span = *cpu_map;
5961 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
5967 sd = &per_cpu(node_domains, i);
5969 sd->span = sched_domain_node_span(cpu_to_node(i));
5973 cpus_and(sd->span, sd->span, *cpu_map);
5977 sd = &per_cpu(phys_domains, i);
5979 sd->span = nodemask;
5983 cpu_to_phys_group(i, cpu_map, &sd->groups);
5985 #ifdef CONFIG_SCHED_MC
5987 sd = &per_cpu(core_domains, i);
5989 sd->span = cpu_coregroup_map(i);
5990 cpus_and(sd->span, sd->span, *cpu_map);
5993 cpu_to_core_group(i, cpu_map, &sd->groups);
5996 #ifdef CONFIG_SCHED_SMT
5998 sd = &per_cpu(cpu_domains, i);
5999 *sd = SD_SIBLING_INIT;
6000 sd->span = cpu_sibling_map[i];
6001 cpus_and(sd->span, sd->span, *cpu_map);
6004 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6008 #ifdef CONFIG_SCHED_SMT
6009 /* Set up CPU (sibling) groups */
6010 for_each_cpu_mask(i, *cpu_map) {
6011 cpumask_t this_sibling_map = cpu_sibling_map[i];
6012 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6013 if (i != first_cpu(this_sibling_map))
6016 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6020 #ifdef CONFIG_SCHED_MC
6021 /* Set up multi-core groups */
6022 for_each_cpu_mask(i, *cpu_map) {
6023 cpumask_t this_core_map = cpu_coregroup_map(i);
6024 cpus_and(this_core_map, this_core_map, *cpu_map);
6025 if (i != first_cpu(this_core_map))
6027 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6032 /* Set up physical groups */
6033 for (i = 0; i < MAX_NUMNODES; i++) {
6034 cpumask_t nodemask = node_to_cpumask(i);
6036 cpus_and(nodemask, nodemask, *cpu_map);
6037 if (cpus_empty(nodemask))
6040 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6044 /* Set up node groups */
6046 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6048 for (i = 0; i < MAX_NUMNODES; i++) {
6049 /* Set up node groups */
6050 struct sched_group *sg, *prev;
6051 cpumask_t nodemask = node_to_cpumask(i);
6052 cpumask_t domainspan;
6053 cpumask_t covered = CPU_MASK_NONE;
6056 cpus_and(nodemask, nodemask, *cpu_map);
6057 if (cpus_empty(nodemask)) {
6058 sched_group_nodes[i] = NULL;
6062 domainspan = sched_domain_node_span(i);
6063 cpus_and(domainspan, domainspan, *cpu_map);
6065 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6067 printk(KERN_WARNING "Can not alloc domain group for "
6071 sched_group_nodes[i] = sg;
6072 for_each_cpu_mask(j, nodemask) {
6073 struct sched_domain *sd;
6074 sd = &per_cpu(node_domains, j);
6077 sg->__cpu_power = 0;
6078 sg->cpumask = nodemask;
6080 cpus_or(covered, covered, nodemask);
6083 for (j = 0; j < MAX_NUMNODES; j++) {
6084 cpumask_t tmp, notcovered;
6085 int n = (i + j) % MAX_NUMNODES;
6087 cpus_complement(notcovered, covered);
6088 cpus_and(tmp, notcovered, *cpu_map);
6089 cpus_and(tmp, tmp, domainspan);
6090 if (cpus_empty(tmp))
6093 nodemask = node_to_cpumask(n);
6094 cpus_and(tmp, tmp, nodemask);
6095 if (cpus_empty(tmp))
6098 sg = kmalloc_node(sizeof(struct sched_group),
6102 "Can not alloc domain group for node %d\n", j);
6105 sg->__cpu_power = 0;
6107 sg->next = prev->next;
6108 cpus_or(covered, covered, tmp);
6115 /* Calculate CPU power for physical packages and nodes */
6116 #ifdef CONFIG_SCHED_SMT
6117 for_each_cpu_mask(i, *cpu_map) {
6118 sd = &per_cpu(cpu_domains, i);
6119 init_sched_groups_power(i, sd);
6122 #ifdef CONFIG_SCHED_MC
6123 for_each_cpu_mask(i, *cpu_map) {
6124 sd = &per_cpu(core_domains, i);
6125 init_sched_groups_power(i, sd);
6129 for_each_cpu_mask(i, *cpu_map) {
6130 sd = &per_cpu(phys_domains, i);
6131 init_sched_groups_power(i, sd);
6135 for (i = 0; i < MAX_NUMNODES; i++)
6136 init_numa_sched_groups_power(sched_group_nodes[i]);
6139 struct sched_group *sg;
6141 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6142 init_numa_sched_groups_power(sg);
6146 /* Attach the domains */
6147 for_each_cpu_mask(i, *cpu_map) {
6148 struct sched_domain *sd;
6149 #ifdef CONFIG_SCHED_SMT
6150 sd = &per_cpu(cpu_domains, i);
6151 #elif defined(CONFIG_SCHED_MC)
6152 sd = &per_cpu(core_domains, i);
6154 sd = &per_cpu(phys_domains, i);
6156 cpu_attach_domain(sd, i);
6163 free_sched_groups(cpu_map);
6168 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6170 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6172 cpumask_t cpu_default_map;
6176 * Setup mask for cpus without special case scheduling requirements.
6177 * For now this just excludes isolated cpus, but could be used to
6178 * exclude other special cases in the future.
6180 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6182 err = build_sched_domains(&cpu_default_map);
6187 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6189 free_sched_groups(cpu_map);
6193 * Detach sched domains from a group of cpus specified in cpu_map
6194 * These cpus will now be attached to the NULL domain
6196 static void detach_destroy_domains(const cpumask_t *cpu_map)
6200 for_each_cpu_mask(i, *cpu_map)
6201 cpu_attach_domain(NULL, i);
6202 synchronize_sched();
6203 arch_destroy_sched_domains(cpu_map);
6207 * Partition sched domains as specified by the cpumasks below.
6208 * This attaches all cpus from the cpumasks to the NULL domain,
6209 * waits for a RCU quiescent period, recalculates sched
6210 * domain information and then attaches them back to the
6211 * correct sched domains
6212 * Call with hotplug lock held
6214 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6216 cpumask_t change_map;
6219 cpus_and(*partition1, *partition1, cpu_online_map);
6220 cpus_and(*partition2, *partition2, cpu_online_map);
6221 cpus_or(change_map, *partition1, *partition2);
6223 /* Detach sched domains from all of the affected cpus */
6224 detach_destroy_domains(&change_map);
6225 if (!cpus_empty(*partition1))
6226 err = build_sched_domains(partition1);
6227 if (!err && !cpus_empty(*partition2))
6228 err = build_sched_domains(partition2);
6233 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6234 int arch_reinit_sched_domains(void)
6238 mutex_lock(&sched_hotcpu_mutex);
6239 detach_destroy_domains(&cpu_online_map);
6240 err = arch_init_sched_domains(&cpu_online_map);
6241 mutex_unlock(&sched_hotcpu_mutex);
6246 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6250 if (buf[0] != '0' && buf[0] != '1')
6254 sched_smt_power_savings = (buf[0] == '1');
6256 sched_mc_power_savings = (buf[0] == '1');
6258 ret = arch_reinit_sched_domains();
6260 return ret ? ret : count;
6263 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6267 #ifdef CONFIG_SCHED_SMT
6269 err = sysfs_create_file(&cls->kset.kobj,
6270 &attr_sched_smt_power_savings.attr);
6272 #ifdef CONFIG_SCHED_MC
6273 if (!err && mc_capable())
6274 err = sysfs_create_file(&cls->kset.kobj,
6275 &attr_sched_mc_power_savings.attr);
6281 #ifdef CONFIG_SCHED_MC
6282 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6284 return sprintf(page, "%u\n", sched_mc_power_savings);
6286 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6287 const char *buf, size_t count)
6289 return sched_power_savings_store(buf, count, 0);
6291 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6292 sched_mc_power_savings_store);
6295 #ifdef CONFIG_SCHED_SMT
6296 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6298 return sprintf(page, "%u\n", sched_smt_power_savings);
6300 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6301 const char *buf, size_t count)
6303 return sched_power_savings_store(buf, count, 1);
6305 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6306 sched_smt_power_savings_store);
6310 * Force a reinitialization of the sched domains hierarchy. The domains
6311 * and groups cannot be updated in place without racing with the balancing
6312 * code, so we temporarily attach all running cpus to the NULL domain
6313 * which will prevent rebalancing while the sched domains are recalculated.
6315 static int update_sched_domains(struct notifier_block *nfb,
6316 unsigned long action, void *hcpu)
6319 case CPU_UP_PREPARE:
6320 case CPU_UP_PREPARE_FROZEN:
6321 case CPU_DOWN_PREPARE:
6322 case CPU_DOWN_PREPARE_FROZEN:
6323 detach_destroy_domains(&cpu_online_map);
6326 case CPU_UP_CANCELED:
6327 case CPU_UP_CANCELED_FROZEN:
6328 case CPU_DOWN_FAILED:
6329 case CPU_DOWN_FAILED_FROZEN:
6331 case CPU_ONLINE_FROZEN:
6333 case CPU_DEAD_FROZEN:
6335 * Fall through and re-initialise the domains.
6342 /* The hotplug lock is already held by cpu_up/cpu_down */
6343 arch_init_sched_domains(&cpu_online_map);
6348 void __init sched_init_smp(void)
6350 cpumask_t non_isolated_cpus;
6352 mutex_lock(&sched_hotcpu_mutex);
6353 arch_init_sched_domains(&cpu_online_map);
6354 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6355 if (cpus_empty(non_isolated_cpus))
6356 cpu_set(smp_processor_id(), non_isolated_cpus);
6357 mutex_unlock(&sched_hotcpu_mutex);
6358 /* XXX: Theoretical race here - CPU may be hotplugged now */
6359 hotcpu_notifier(update_sched_domains, 0);
6361 /* Move init over to a non-isolated CPU */
6362 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6366 void __init sched_init_smp(void)
6369 #endif /* CONFIG_SMP */
6371 int in_sched_functions(unsigned long addr)
6373 /* Linker adds these: start and end of __sched functions */
6374 extern char __sched_text_start[], __sched_text_end[];
6376 return in_lock_functions(addr) ||
6377 (addr >= (unsigned long)__sched_text_start
6378 && addr < (unsigned long)__sched_text_end);
6381 void __init sched_init(void)
6384 int highest_cpu = 0;
6386 for_each_possible_cpu(i) {
6387 struct prio_array *array;
6391 spin_lock_init(&rq->lock);
6392 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6394 rq->active = rq->arrays;
6395 rq->expired = rq->arrays + 1;
6396 rq->best_expired_prio = MAX_PRIO;
6400 for (j = 1; j < 3; j++)
6401 rq->cpu_load[j] = 0;
6402 rq->active_balance = 0;
6405 rq->migration_thread = NULL;
6406 INIT_LIST_HEAD(&rq->migration_queue);
6408 atomic_set(&rq->nr_iowait, 0);
6410 for (j = 0; j < 2; j++) {
6411 array = rq->arrays + j;
6412 for (k = 0; k < MAX_PRIO; k++) {
6413 INIT_LIST_HEAD(array->queue + k);
6414 __clear_bit(k, array->bitmap);
6416 // delimiter for bitsearch
6417 __set_bit(MAX_PRIO, array->bitmap);
6422 set_load_weight(&init_task);
6425 nr_cpu_ids = highest_cpu + 1;
6426 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6429 #ifdef CONFIG_RT_MUTEXES
6430 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6434 * The boot idle thread does lazy MMU switching as well:
6436 atomic_inc(&init_mm.mm_count);
6437 enter_lazy_tlb(&init_mm, current);
6440 * Make us the idle thread. Technically, schedule() should not be
6441 * called from this thread, however somewhere below it might be,
6442 * but because we are the idle thread, we just pick up running again
6443 * when this runqueue becomes "idle".
6445 init_idle(current, smp_processor_id());
6448 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6449 void __might_sleep(char *file, int line)
6452 static unsigned long prev_jiffy; /* ratelimiting */
6454 if ((in_atomic() || irqs_disabled()) &&
6455 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6456 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6458 prev_jiffy = jiffies;
6459 printk(KERN_ERR "BUG: sleeping function called from invalid"
6460 " context at %s:%d\n", file, line);
6461 printk("in_atomic():%d, irqs_disabled():%d\n",
6462 in_atomic(), irqs_disabled());
6463 debug_show_held_locks(current);
6464 if (irqs_disabled())
6465 print_irqtrace_events(current);
6470 EXPORT_SYMBOL(__might_sleep);
6473 #ifdef CONFIG_MAGIC_SYSRQ
6474 void normalize_rt_tasks(void)
6476 struct prio_array *array;
6477 struct task_struct *g, *p;
6478 unsigned long flags;
6481 read_lock_irq(&tasklist_lock);
6483 do_each_thread(g, p) {
6487 spin_lock_irqsave(&p->pi_lock, flags);
6488 rq = __task_rq_lock(p);
6492 deactivate_task(p, task_rq(p));
6493 __setscheduler(p, SCHED_NORMAL, 0);
6495 __activate_task(p, task_rq(p));
6496 resched_task(rq->curr);
6499 __task_rq_unlock(rq);
6500 spin_unlock_irqrestore(&p->pi_lock, flags);
6501 } while_each_thread(g, p);
6503 read_unlock_irq(&tasklist_lock);
6506 #endif /* CONFIG_MAGIC_SYSRQ */
6510 * These functions are only useful for the IA64 MCA handling.
6512 * They can only be called when the whole system has been
6513 * stopped - every CPU needs to be quiescent, and no scheduling
6514 * activity can take place. Using them for anything else would
6515 * be a serious bug, and as a result, they aren't even visible
6516 * under any other configuration.
6520 * curr_task - return the current task for a given cpu.
6521 * @cpu: the processor in question.
6523 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6525 struct task_struct *curr_task(int cpu)
6527 return cpu_curr(cpu);
6531 * set_curr_task - set the current task for a given cpu.
6532 * @cpu: the processor in question.
6533 * @p: the task pointer to set.
6535 * Description: This function must only be used when non-maskable interrupts
6536 * are serviced on a separate stack. It allows the architecture to switch the
6537 * notion of the current task on a cpu in a non-blocking manner. This function
6538 * must be called with all CPU's synchronized, and interrupts disabled, the
6539 * and caller must save the original value of the current task (see
6540 * curr_task() above) and restore that value before reenabling interrupts and
6541 * re-starting the system.
6543 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6545 void set_curr_task(int cpu, struct task_struct *p)