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
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak)) sched_clock(void)
75 return (unsigned long long)jiffies * (1000000000 / HZ);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio)
145 if (static_prio == NICE_TO_PRIO(19))
148 if (static_prio < NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
151 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
154 static inline int rt_policy(int policy)
156 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
161 static inline int task_has_rt_policy(struct task_struct *p)
163 return rt_policy(p->policy);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array {
170 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
171 struct list_head queue[MAX_RT_PRIO];
175 struct load_weight load;
178 /* CFS-related fields in a runqueue */
180 struct load_weight load;
181 unsigned long nr_running;
187 unsigned long wait_runtime_overruns, wait_runtime_underruns;
189 struct rb_root tasks_timeline;
190 struct rb_node *rb_leftmost;
191 struct rb_node *rb_load_balance_curr;
192 #ifdef CONFIG_FAIR_GROUP_SCHED
193 /* 'curr' points to currently running entity on this cfs_rq.
194 * It is set to NULL otherwise (i.e when none are currently running).
196 struct sched_entity *curr;
197 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
199 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
200 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
201 * (like users, containers etc.)
203 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
204 * list is used during load balance.
206 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
210 /* Real-Time classes' related field in a runqueue: */
212 struct rt_prio_array active;
213 int rt_load_balance_idx;
214 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
218 * This is the main, per-CPU runqueue data structure.
220 * Locking rule: those places that want to lock multiple runqueues
221 * (such as the load balancing or the thread migration code), lock
222 * acquire operations must be ordered by ascending &runqueue.
225 spinlock_t lock; /* runqueue lock */
228 * nr_running and cpu_load should be in the same cacheline because
229 * remote CPUs use both these fields when doing load calculation.
231 unsigned long nr_running;
232 #define CPU_LOAD_IDX_MAX 5
233 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
234 unsigned char idle_at_tick;
236 unsigned char in_nohz_recently;
238 struct load_stat ls; /* capture load from *all* tasks on this cpu */
239 unsigned long nr_load_updates;
243 #ifdef CONFIG_FAIR_GROUP_SCHED
244 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
249 * This is part of a global counter where only the total sum
250 * over all CPUs matters. A task can increase this counter on
251 * one CPU and if it got migrated afterwards it may decrease
252 * it on another CPU. Always updated under the runqueue lock:
254 unsigned long nr_uninterruptible;
256 struct task_struct *curr, *idle;
257 unsigned long next_balance;
258 struct mm_struct *prev_mm;
260 u64 clock, prev_clock_raw;
263 unsigned int clock_warps, clock_overflows;
265 unsigned int clock_deep_idle_events;
271 struct sched_domain *sd;
273 /* For active balancing */
276 int cpu; /* cpu of this runqueue */
278 struct task_struct *migration_thread;
279 struct list_head migration_queue;
282 #ifdef CONFIG_SCHEDSTATS
284 struct sched_info rq_sched_info;
286 /* sys_sched_yield() stats */
287 unsigned long yld_exp_empty;
288 unsigned long yld_act_empty;
289 unsigned long yld_both_empty;
290 unsigned long yld_cnt;
292 /* schedule() stats */
293 unsigned long sched_switch;
294 unsigned long sched_cnt;
295 unsigned long sched_goidle;
297 /* try_to_wake_up() stats */
298 unsigned long ttwu_cnt;
299 unsigned long ttwu_local;
301 struct lock_class_key rq_lock_key;
304 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
305 static DEFINE_MUTEX(sched_hotcpu_mutex);
307 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
309 rq->curr->sched_class->check_preempt_curr(rq, p);
312 static inline int cpu_of(struct rq *rq)
322 * Update the per-runqueue clock, as finegrained as the platform can give
323 * us, but without assuming monotonicity, etc.:
325 static void __update_rq_clock(struct rq *rq)
327 u64 prev_raw = rq->prev_clock_raw;
328 u64 now = sched_clock();
329 s64 delta = now - prev_raw;
330 u64 clock = rq->clock;
332 #ifdef CONFIG_SCHED_DEBUG
333 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
336 * Protect against sched_clock() occasionally going backwards:
338 if (unlikely(delta < 0)) {
343 * Catch too large forward jumps too:
345 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
346 if (clock < rq->tick_timestamp + TICK_NSEC)
347 clock = rq->tick_timestamp + TICK_NSEC;
350 rq->clock_overflows++;
352 if (unlikely(delta > rq->clock_max_delta))
353 rq->clock_max_delta = delta;
358 rq->prev_clock_raw = now;
362 static void update_rq_clock(struct rq *rq)
364 if (likely(smp_processor_id() == cpu_of(rq)))
365 __update_rq_clock(rq);
369 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
370 * See detach_destroy_domains: synchronize_sched for details.
372 * The domain tree of any CPU may only be accessed from within
373 * preempt-disabled sections.
375 #define for_each_domain(cpu, __sd) \
376 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
378 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
379 #define this_rq() (&__get_cpu_var(runqueues))
380 #define task_rq(p) cpu_rq(task_cpu(p))
381 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
384 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
385 * clock constructed from sched_clock():
387 unsigned long long cpu_clock(int cpu)
389 unsigned long long now;
393 local_irq_save(flags);
397 local_irq_restore(flags);
402 #ifdef CONFIG_FAIR_GROUP_SCHED
403 /* Change a task's ->cfs_rq if it moves across CPUs */
404 static inline void set_task_cfs_rq(struct task_struct *p)
406 p->se.cfs_rq = &task_rq(p)->cfs;
409 static inline void set_task_cfs_rq(struct task_struct *p)
414 #ifndef prepare_arch_switch
415 # define prepare_arch_switch(next) do { } while (0)
417 #ifndef finish_arch_switch
418 # define finish_arch_switch(prev) do { } while (0)
421 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
422 static inline int task_running(struct rq *rq, struct task_struct *p)
424 return rq->curr == p;
427 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
431 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
433 #ifdef CONFIG_DEBUG_SPINLOCK
434 /* this is a valid case when another task releases the spinlock */
435 rq->lock.owner = current;
438 * If we are tracking spinlock dependencies then we have to
439 * fix up the runqueue lock - which gets 'carried over' from
442 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
444 spin_unlock_irq(&rq->lock);
447 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
448 static inline int task_running(struct rq *rq, struct task_struct *p)
453 return rq->curr == p;
457 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
461 * We can optimise this out completely for !SMP, because the
462 * SMP rebalancing from interrupt is the only thing that cares
467 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
468 spin_unlock_irq(&rq->lock);
470 spin_unlock(&rq->lock);
474 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
478 * After ->oncpu is cleared, the task can be moved to a different CPU.
479 * We must ensure this doesn't happen until the switch is completely
485 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
489 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
492 * __task_rq_lock - lock the runqueue a given task resides on.
493 * Must be called interrupts disabled.
495 static inline struct rq *__task_rq_lock(struct task_struct *p)
502 spin_lock(&rq->lock);
503 if (unlikely(rq != task_rq(p))) {
504 spin_unlock(&rq->lock);
505 goto repeat_lock_task;
511 * task_rq_lock - lock the runqueue a given task resides on and disable
512 * interrupts. Note the ordering: we can safely lookup the task_rq without
513 * explicitly disabling preemption.
515 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
521 local_irq_save(*flags);
523 spin_lock(&rq->lock);
524 if (unlikely(rq != task_rq(p))) {
525 spin_unlock_irqrestore(&rq->lock, *flags);
526 goto repeat_lock_task;
531 static inline void __task_rq_unlock(struct rq *rq)
534 spin_unlock(&rq->lock);
537 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
540 spin_unlock_irqrestore(&rq->lock, *flags);
544 * this_rq_lock - lock this runqueue and disable interrupts.
546 static inline struct rq *this_rq_lock(void)
553 spin_lock(&rq->lock);
559 * We are going deep-idle (irqs are disabled):
561 void sched_clock_idle_sleep_event(void)
563 struct rq *rq = cpu_rq(smp_processor_id());
565 spin_lock(&rq->lock);
566 __update_rq_clock(rq);
567 spin_unlock(&rq->lock);
568 rq->clock_deep_idle_events++;
570 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
573 * We just idled delta nanoseconds (called with irqs disabled):
575 void sched_clock_idle_wakeup_event(u64 delta_ns)
577 struct rq *rq = cpu_rq(smp_processor_id());
578 u64 now = sched_clock();
580 rq->idle_clock += delta_ns;
582 * Override the previous timestamp and ignore all
583 * sched_clock() deltas that occured while we idled,
584 * and use the PM-provided delta_ns to advance the
587 spin_lock(&rq->lock);
588 rq->prev_clock_raw = now;
589 rq->clock += delta_ns;
590 spin_unlock(&rq->lock);
592 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
595 * resched_task - mark a task 'to be rescheduled now'.
597 * On UP this means the setting of the need_resched flag, on SMP it
598 * might also involve a cross-CPU call to trigger the scheduler on
603 #ifndef tsk_is_polling
604 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
607 static void resched_task(struct task_struct *p)
611 assert_spin_locked(&task_rq(p)->lock);
613 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
616 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
619 if (cpu == smp_processor_id())
622 /* NEED_RESCHED must be visible before we test polling */
624 if (!tsk_is_polling(p))
625 smp_send_reschedule(cpu);
628 static void resched_cpu(int cpu)
630 struct rq *rq = cpu_rq(cpu);
633 if (!spin_trylock_irqsave(&rq->lock, flags))
635 resched_task(cpu_curr(cpu));
636 spin_unlock_irqrestore(&rq->lock, flags);
639 static inline void resched_task(struct task_struct *p)
641 assert_spin_locked(&task_rq(p)->lock);
642 set_tsk_need_resched(p);
646 static u64 div64_likely32(u64 divident, unsigned long divisor)
648 #if BITS_PER_LONG == 32
649 if (likely(divident <= 0xffffffffULL))
650 return (u32)divident / divisor;
651 do_div(divident, divisor);
655 return divident / divisor;
659 #if BITS_PER_LONG == 32
660 # define WMULT_CONST (~0UL)
662 # define WMULT_CONST (1UL << 32)
665 #define WMULT_SHIFT 32
668 * Shift right and round:
670 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
673 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
674 struct load_weight *lw)
678 if (unlikely(!lw->inv_weight))
679 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
681 tmp = (u64)delta_exec * weight;
683 * Check whether we'd overflow the 64-bit multiplication:
685 if (unlikely(tmp > WMULT_CONST))
686 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
689 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
691 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
694 static inline unsigned long
695 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
697 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
700 static void update_load_add(struct load_weight *lw, unsigned long inc)
706 static void update_load_sub(struct load_weight *lw, unsigned long dec)
713 * To aid in avoiding the subversion of "niceness" due to uneven distribution
714 * of tasks with abnormal "nice" values across CPUs the contribution that
715 * each task makes to its run queue's load is weighted according to its
716 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
717 * scaled version of the new time slice allocation that they receive on time
721 #define WEIGHT_IDLEPRIO 2
722 #define WMULT_IDLEPRIO (1 << 31)
725 * Nice levels are multiplicative, with a gentle 10% change for every
726 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
727 * nice 1, it will get ~10% less CPU time than another CPU-bound task
728 * that remained on nice 0.
730 * The "10% effect" is relative and cumulative: from _any_ nice level,
731 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
732 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
733 * If a task goes up by ~10% and another task goes down by ~10% then
734 * the relative distance between them is ~25%.)
736 static const int prio_to_weight[40] = {
737 /* -20 */ 88761, 71755, 56483, 46273, 36291,
738 /* -15 */ 29154, 23254, 18705, 14949, 11916,
739 /* -10 */ 9548, 7620, 6100, 4904, 3906,
740 /* -5 */ 3121, 2501, 1991, 1586, 1277,
741 /* 0 */ 1024, 820, 655, 526, 423,
742 /* 5 */ 335, 272, 215, 172, 137,
743 /* 10 */ 110, 87, 70, 56, 45,
744 /* 15 */ 36, 29, 23, 18, 15,
748 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
750 * In cases where the weight does not change often, we can use the
751 * precalculated inverse to speed up arithmetics by turning divisions
752 * into multiplications:
754 static const u32 prio_to_wmult[40] = {
755 /* -20 */ 48388, 59856, 76040, 92818, 118348,
756 /* -15 */ 147320, 184698, 229616, 287308, 360437,
757 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
758 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
759 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
760 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
761 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
762 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
765 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
768 * runqueue iterator, to support SMP load-balancing between different
769 * scheduling classes, without having to expose their internal data
770 * structures to the load-balancing proper:
774 struct task_struct *(*start)(void *);
775 struct task_struct *(*next)(void *);
778 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
779 unsigned long max_nr_move, unsigned long max_load_move,
780 struct sched_domain *sd, enum cpu_idle_type idle,
781 int *all_pinned, unsigned long *load_moved,
782 int *this_best_prio, struct rq_iterator *iterator);
784 #include "sched_stats.h"
785 #include "sched_rt.c"
786 #include "sched_fair.c"
787 #include "sched_idletask.c"
788 #ifdef CONFIG_SCHED_DEBUG
789 # include "sched_debug.c"
792 #define sched_class_highest (&rt_sched_class)
795 * Update delta_exec, delta_fair fields for rq.
797 * delta_fair clock advances at a rate inversely proportional to
798 * total load (rq->ls.load.weight) on the runqueue, while
799 * delta_exec advances at the same rate as wall-clock (provided
802 * delta_exec / delta_fair is a measure of the (smoothened) load on this
803 * runqueue over any given interval. This (smoothened) load is used
804 * during load balance.
806 * This function is called /before/ updating rq->ls.load
807 * and when switching tasks.
809 static inline void inc_load(struct rq *rq, const struct task_struct *p)
811 update_load_add(&rq->ls.load, p->se.load.weight);
814 static inline void dec_load(struct rq *rq, const struct task_struct *p)
816 update_load_sub(&rq->ls.load, p->se.load.weight);
819 static void inc_nr_running(struct task_struct *p, struct rq *rq)
825 static void dec_nr_running(struct task_struct *p, struct rq *rq)
831 static void set_load_weight(struct task_struct *p)
833 p->se.wait_runtime = 0;
835 if (task_has_rt_policy(p)) {
836 p->se.load.weight = prio_to_weight[0] * 2;
837 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
842 * SCHED_IDLE tasks get minimal weight:
844 if (p->policy == SCHED_IDLE) {
845 p->se.load.weight = WEIGHT_IDLEPRIO;
846 p->se.load.inv_weight = WMULT_IDLEPRIO;
850 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
851 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
854 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
856 sched_info_queued(p);
857 p->sched_class->enqueue_task(rq, p, wakeup);
861 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
863 p->sched_class->dequeue_task(rq, p, sleep);
868 * __normal_prio - return the priority that is based on the static prio
870 static inline int __normal_prio(struct task_struct *p)
872 return p->static_prio;
876 * Calculate the expected normal priority: i.e. priority
877 * without taking RT-inheritance into account. Might be
878 * boosted by interactivity modifiers. Changes upon fork,
879 * setprio syscalls, and whenever the interactivity
880 * estimator recalculates.
882 static inline int normal_prio(struct task_struct *p)
886 if (task_has_rt_policy(p))
887 prio = MAX_RT_PRIO-1 - p->rt_priority;
889 prio = __normal_prio(p);
894 * Calculate the current priority, i.e. the priority
895 * taken into account by the scheduler. This value might
896 * be boosted by RT tasks, or might be boosted by
897 * interactivity modifiers. Will be RT if the task got
898 * RT-boosted. If not then it returns p->normal_prio.
900 static int effective_prio(struct task_struct *p)
902 p->normal_prio = normal_prio(p);
904 * If we are RT tasks or we were boosted to RT priority,
905 * keep the priority unchanged. Otherwise, update priority
906 * to the normal priority:
908 if (!rt_prio(p->prio))
909 return p->normal_prio;
914 * activate_task - move a task to the runqueue.
916 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
918 if (p->state == TASK_UNINTERRUPTIBLE)
919 rq->nr_uninterruptible--;
921 enqueue_task(rq, p, wakeup);
922 inc_nr_running(p, rq);
926 * activate_idle_task - move idle task to the _front_ of runqueue.
928 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
932 if (p->state == TASK_UNINTERRUPTIBLE)
933 rq->nr_uninterruptible--;
935 enqueue_task(rq, p, 0);
936 inc_nr_running(p, rq);
940 * deactivate_task - remove a task from the runqueue.
942 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
944 if (p->state == TASK_UNINTERRUPTIBLE)
945 rq->nr_uninterruptible++;
947 dequeue_task(rq, p, sleep);
948 dec_nr_running(p, rq);
952 * task_curr - is this task currently executing on a CPU?
953 * @p: the task in question.
955 inline int task_curr(const struct task_struct *p)
957 return cpu_curr(task_cpu(p)) == p;
960 /* Used instead of source_load when we know the type == 0 */
961 unsigned long weighted_cpuload(const int cpu)
963 return cpu_rq(cpu)->ls.load.weight;
966 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
969 task_thread_info(p)->cpu = cpu;
976 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
978 int old_cpu = task_cpu(p);
979 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
980 u64 clock_offset, fair_clock_offset;
982 clock_offset = old_rq->clock - new_rq->clock;
983 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
985 if (p->se.wait_start_fair)
986 p->se.wait_start_fair -= fair_clock_offset;
987 if (p->se.sleep_start_fair)
988 p->se.sleep_start_fair -= fair_clock_offset;
990 #ifdef CONFIG_SCHEDSTATS
991 if (p->se.wait_start)
992 p->se.wait_start -= clock_offset;
993 if (p->se.sleep_start)
994 p->se.sleep_start -= clock_offset;
995 if (p->se.block_start)
996 p->se.block_start -= clock_offset;
999 __set_task_cpu(p, new_cpu);
1002 struct migration_req {
1003 struct list_head list;
1005 struct task_struct *task;
1008 struct completion done;
1012 * The task's runqueue lock must be held.
1013 * Returns true if you have to wait for migration thread.
1016 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1018 struct rq *rq = task_rq(p);
1021 * If the task is not on a runqueue (and not running), then
1022 * it is sufficient to simply update the task's cpu field.
1024 if (!p->se.on_rq && !task_running(rq, p)) {
1025 set_task_cpu(p, dest_cpu);
1029 init_completion(&req->done);
1031 req->dest_cpu = dest_cpu;
1032 list_add(&req->list, &rq->migration_queue);
1038 * wait_task_inactive - wait for a thread to unschedule.
1040 * The caller must ensure that the task *will* unschedule sometime soon,
1041 * else this function might spin for a *long* time. This function can't
1042 * be called with interrupts off, or it may introduce deadlock with
1043 * smp_call_function() if an IPI is sent by the same process we are
1044 * waiting to become inactive.
1046 void wait_task_inactive(struct task_struct *p)
1048 unsigned long flags;
1054 * We do the initial early heuristics without holding
1055 * any task-queue locks at all. We'll only try to get
1056 * the runqueue lock when things look like they will
1062 * If the task is actively running on another CPU
1063 * still, just relax and busy-wait without holding
1066 * NOTE! Since we don't hold any locks, it's not
1067 * even sure that "rq" stays as the right runqueue!
1068 * But we don't care, since "task_running()" will
1069 * return false if the runqueue has changed and p
1070 * is actually now running somewhere else!
1072 while (task_running(rq, p))
1076 * Ok, time to look more closely! We need the rq
1077 * lock now, to be *sure*. If we're wrong, we'll
1078 * just go back and repeat.
1080 rq = task_rq_lock(p, &flags);
1081 running = task_running(rq, p);
1082 on_rq = p->se.on_rq;
1083 task_rq_unlock(rq, &flags);
1086 * Was it really running after all now that we
1087 * checked with the proper locks actually held?
1089 * Oops. Go back and try again..
1091 if (unlikely(running)) {
1097 * It's not enough that it's not actively running,
1098 * it must be off the runqueue _entirely_, and not
1101 * So if it wa still runnable (but just not actively
1102 * running right now), it's preempted, and we should
1103 * yield - it could be a while.
1105 if (unlikely(on_rq)) {
1111 * Ahh, all good. It wasn't running, and it wasn't
1112 * runnable, which means that it will never become
1113 * running in the future either. We're all done!
1118 * kick_process - kick a running thread to enter/exit the kernel
1119 * @p: the to-be-kicked thread
1121 * Cause a process which is running on another CPU to enter
1122 * kernel-mode, without any delay. (to get signals handled.)
1124 * NOTE: this function doesnt have to take the runqueue lock,
1125 * because all it wants to ensure is that the remote task enters
1126 * the kernel. If the IPI races and the task has been migrated
1127 * to another CPU then no harm is done and the purpose has been
1130 void kick_process(struct task_struct *p)
1136 if ((cpu != smp_processor_id()) && task_curr(p))
1137 smp_send_reschedule(cpu);
1142 * Return a low guess at the load of a migration-source cpu weighted
1143 * according to the scheduling class and "nice" value.
1145 * We want to under-estimate the load of migration sources, to
1146 * balance conservatively.
1148 static inline unsigned long source_load(int cpu, int type)
1150 struct rq *rq = cpu_rq(cpu);
1151 unsigned long total = weighted_cpuload(cpu);
1156 return min(rq->cpu_load[type-1], total);
1160 * Return a high guess at the load of a migration-target cpu weighted
1161 * according to the scheduling class and "nice" value.
1163 static inline unsigned long target_load(int cpu, int type)
1165 struct rq *rq = cpu_rq(cpu);
1166 unsigned long total = weighted_cpuload(cpu);
1171 return max(rq->cpu_load[type-1], total);
1175 * Return the average load per task on the cpu's run queue
1177 static inline unsigned long cpu_avg_load_per_task(int cpu)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long total = weighted_cpuload(cpu);
1181 unsigned long n = rq->nr_running;
1183 return n ? total / n : SCHED_LOAD_SCALE;
1187 * find_idlest_group finds and returns the least busy CPU group within the
1190 static struct sched_group *
1191 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1193 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1194 unsigned long min_load = ULONG_MAX, this_load = 0;
1195 int load_idx = sd->forkexec_idx;
1196 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1199 unsigned long load, avg_load;
1203 /* Skip over this group if it has no CPUs allowed */
1204 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1207 local_group = cpu_isset(this_cpu, group->cpumask);
1209 /* Tally up the load of all CPUs in the group */
1212 for_each_cpu_mask(i, group->cpumask) {
1213 /* Bias balancing toward cpus of our domain */
1215 load = source_load(i, load_idx);
1217 load = target_load(i, load_idx);
1222 /* Adjust by relative CPU power of the group */
1223 avg_load = sg_div_cpu_power(group,
1224 avg_load * SCHED_LOAD_SCALE);
1227 this_load = avg_load;
1229 } else if (avg_load < min_load) {
1230 min_load = avg_load;
1234 group = group->next;
1235 } while (group != sd->groups);
1237 if (!idlest || 100*this_load < imbalance*min_load)
1243 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1246 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1249 unsigned long load, min_load = ULONG_MAX;
1253 /* Traverse only the allowed CPUs */
1254 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1256 for_each_cpu_mask(i, tmp) {
1257 load = weighted_cpuload(i);
1259 if (load < min_load || (load == min_load && i == this_cpu)) {
1269 * sched_balance_self: balance the current task (running on cpu) in domains
1270 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1273 * Balance, ie. select the least loaded group.
1275 * Returns the target CPU number, or the same CPU if no balancing is needed.
1277 * preempt must be disabled.
1279 static int sched_balance_self(int cpu, int flag)
1281 struct task_struct *t = current;
1282 struct sched_domain *tmp, *sd = NULL;
1284 for_each_domain(cpu, tmp) {
1286 * If power savings logic is enabled for a domain, stop there.
1288 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1290 if (tmp->flags & flag)
1296 struct sched_group *group;
1297 int new_cpu, weight;
1299 if (!(sd->flags & flag)) {
1305 group = find_idlest_group(sd, t, cpu);
1311 new_cpu = find_idlest_cpu(group, t, cpu);
1312 if (new_cpu == -1 || new_cpu == cpu) {
1313 /* Now try balancing at a lower domain level of cpu */
1318 /* Now try balancing at a lower domain level of new_cpu */
1321 weight = cpus_weight(span);
1322 for_each_domain(cpu, tmp) {
1323 if (weight <= cpus_weight(tmp->span))
1325 if (tmp->flags & flag)
1328 /* while loop will break here if sd == NULL */
1334 #endif /* CONFIG_SMP */
1337 * wake_idle() will wake a task on an idle cpu if task->cpu is
1338 * not idle and an idle cpu is available. The span of cpus to
1339 * search starts with cpus closest then further out as needed,
1340 * so we always favor a closer, idle cpu.
1342 * Returns the CPU we should wake onto.
1344 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1345 static int wake_idle(int cpu, struct task_struct *p)
1348 struct sched_domain *sd;
1352 * If it is idle, then it is the best cpu to run this task.
1354 * This cpu is also the best, if it has more than one task already.
1355 * Siblings must be also busy(in most cases) as they didn't already
1356 * pickup the extra load from this cpu and hence we need not check
1357 * sibling runqueue info. This will avoid the checks and cache miss
1358 * penalities associated with that.
1360 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1363 for_each_domain(cpu, sd) {
1364 if (sd->flags & SD_WAKE_IDLE) {
1365 cpus_and(tmp, sd->span, p->cpus_allowed);
1366 for_each_cpu_mask(i, tmp) {
1377 static inline int wake_idle(int cpu, struct task_struct *p)
1384 * try_to_wake_up - wake up a thread
1385 * @p: the to-be-woken-up thread
1386 * @state: the mask of task states that can be woken
1387 * @sync: do a synchronous wakeup?
1389 * Put it on the run-queue if it's not already there. The "current"
1390 * thread is always on the run-queue (except when the actual
1391 * re-schedule is in progress), and as such you're allowed to do
1392 * the simpler "current->state = TASK_RUNNING" to mark yourself
1393 * runnable without the overhead of this.
1395 * returns failure only if the task is already active.
1397 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1399 int cpu, this_cpu, success = 0;
1400 unsigned long flags;
1404 struct sched_domain *sd, *this_sd = NULL;
1405 unsigned long load, this_load;
1409 rq = task_rq_lock(p, &flags);
1410 old_state = p->state;
1411 if (!(old_state & state))
1418 this_cpu = smp_processor_id();
1421 if (unlikely(task_running(rq, p)))
1426 schedstat_inc(rq, ttwu_cnt);
1427 if (cpu == this_cpu) {
1428 schedstat_inc(rq, ttwu_local);
1432 for_each_domain(this_cpu, sd) {
1433 if (cpu_isset(cpu, sd->span)) {
1434 schedstat_inc(sd, ttwu_wake_remote);
1440 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1444 * Check for affine wakeup and passive balancing possibilities.
1447 int idx = this_sd->wake_idx;
1448 unsigned int imbalance;
1450 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1452 load = source_load(cpu, idx);
1453 this_load = target_load(this_cpu, idx);
1455 new_cpu = this_cpu; /* Wake to this CPU if we can */
1457 if (this_sd->flags & SD_WAKE_AFFINE) {
1458 unsigned long tl = this_load;
1459 unsigned long tl_per_task;
1461 tl_per_task = cpu_avg_load_per_task(this_cpu);
1464 * If sync wakeup then subtract the (maximum possible)
1465 * effect of the currently running task from the load
1466 * of the current CPU:
1469 tl -= current->se.load.weight;
1472 tl + target_load(cpu, idx) <= tl_per_task) ||
1473 100*(tl + p->se.load.weight) <= imbalance*load) {
1475 * This domain has SD_WAKE_AFFINE and
1476 * p is cache cold in this domain, and
1477 * there is no bad imbalance.
1479 schedstat_inc(this_sd, ttwu_move_affine);
1485 * Start passive balancing when half the imbalance_pct
1488 if (this_sd->flags & SD_WAKE_BALANCE) {
1489 if (imbalance*this_load <= 100*load) {
1490 schedstat_inc(this_sd, ttwu_move_balance);
1496 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1498 new_cpu = wake_idle(new_cpu, p);
1499 if (new_cpu != cpu) {
1500 set_task_cpu(p, new_cpu);
1501 task_rq_unlock(rq, &flags);
1502 /* might preempt at this point */
1503 rq = task_rq_lock(p, &flags);
1504 old_state = p->state;
1505 if (!(old_state & state))
1510 this_cpu = smp_processor_id();
1515 #endif /* CONFIG_SMP */
1516 update_rq_clock(rq);
1517 activate_task(rq, p, 1);
1519 * Sync wakeups (i.e. those types of wakeups where the waker
1520 * has indicated that it will leave the CPU in short order)
1521 * don't trigger a preemption, if the woken up task will run on
1522 * this cpu. (in this case the 'I will reschedule' promise of
1523 * the waker guarantees that the freshly woken up task is going
1524 * to be considered on this CPU.)
1526 if (!sync || cpu != this_cpu)
1527 check_preempt_curr(rq, p);
1531 p->state = TASK_RUNNING;
1533 task_rq_unlock(rq, &flags);
1538 int fastcall wake_up_process(struct task_struct *p)
1540 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1541 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1543 EXPORT_SYMBOL(wake_up_process);
1545 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1547 return try_to_wake_up(p, state, 0);
1551 * Perform scheduler related setup for a newly forked process p.
1552 * p is forked by current.
1554 * __sched_fork() is basic setup used by init_idle() too:
1556 static void __sched_fork(struct task_struct *p)
1558 p->se.wait_start_fair = 0;
1559 p->se.exec_start = 0;
1560 p->se.sum_exec_runtime = 0;
1561 p->se.prev_sum_exec_runtime = 0;
1562 p->se.wait_runtime = 0;
1563 p->se.sleep_start_fair = 0;
1565 #ifdef CONFIG_SCHEDSTATS
1566 p->se.wait_start = 0;
1567 p->se.sum_wait_runtime = 0;
1568 p->se.sum_sleep_runtime = 0;
1569 p->se.sleep_start = 0;
1570 p->se.block_start = 0;
1571 p->se.sleep_max = 0;
1572 p->se.block_max = 0;
1574 p->se.slice_max = 0;
1576 p->se.wait_runtime_overruns = 0;
1577 p->se.wait_runtime_underruns = 0;
1580 INIT_LIST_HEAD(&p->run_list);
1583 #ifdef CONFIG_PREEMPT_NOTIFIERS
1584 INIT_HLIST_HEAD(&p->preempt_notifiers);
1588 * We mark the process as running here, but have not actually
1589 * inserted it onto the runqueue yet. This guarantees that
1590 * nobody will actually run it, and a signal or other external
1591 * event cannot wake it up and insert it on the runqueue either.
1593 p->state = TASK_RUNNING;
1597 * fork()/clone()-time setup:
1599 void sched_fork(struct task_struct *p, int clone_flags)
1601 int cpu = get_cpu();
1606 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1608 __set_task_cpu(p, cpu);
1611 * Make sure we do not leak PI boosting priority to the child:
1613 p->prio = current->normal_prio;
1615 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1616 if (likely(sched_info_on()))
1617 memset(&p->sched_info, 0, sizeof(p->sched_info));
1619 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1622 #ifdef CONFIG_PREEMPT
1623 /* Want to start with kernel preemption disabled. */
1624 task_thread_info(p)->preempt_count = 1;
1630 * wake_up_new_task - wake up a newly created task for the first time.
1632 * This function will do some initial scheduler statistics housekeeping
1633 * that must be done for every newly created context, then puts the task
1634 * on the runqueue and wakes it.
1636 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1638 unsigned long flags;
1642 rq = task_rq_lock(p, &flags);
1643 BUG_ON(p->state != TASK_RUNNING);
1644 this_cpu = smp_processor_id(); /* parent's CPU */
1645 update_rq_clock(rq);
1647 p->prio = effective_prio(p);
1649 if (rt_prio(p->prio))
1650 p->sched_class = &rt_sched_class;
1652 p->sched_class = &fair_sched_class;
1654 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1655 !current->se.on_rq) {
1656 activate_task(rq, p, 0);
1659 * Let the scheduling class do new task startup
1660 * management (if any):
1662 p->sched_class->task_new(rq, p);
1663 inc_nr_running(p, rq);
1665 check_preempt_curr(rq, p);
1666 task_rq_unlock(rq, &flags);
1669 #ifdef CONFIG_PREEMPT_NOTIFIERS
1672 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1673 * @notifier: notifier struct to register
1675 void preempt_notifier_register(struct preempt_notifier *notifier)
1677 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1679 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1682 * preempt_notifier_unregister - no longer interested in preemption notifications
1683 * @notifier: notifier struct to unregister
1685 * This is safe to call from within a preemption notifier.
1687 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1689 hlist_del(¬ifier->link);
1691 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1693 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1695 struct preempt_notifier *notifier;
1696 struct hlist_node *node;
1698 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1699 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1703 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1704 struct task_struct *next)
1706 struct preempt_notifier *notifier;
1707 struct hlist_node *node;
1709 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1710 notifier->ops->sched_out(notifier, next);
1715 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1720 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1721 struct task_struct *next)
1728 * prepare_task_switch - prepare to switch tasks
1729 * @rq: the runqueue preparing to switch
1730 * @prev: the current task that is being switched out
1731 * @next: the task we are going to switch to.
1733 * This is called with the rq lock held and interrupts off. It must
1734 * be paired with a subsequent finish_task_switch after the context
1737 * prepare_task_switch sets up locking and calls architecture specific
1741 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1742 struct task_struct *next)
1744 fire_sched_out_preempt_notifiers(prev, next);
1745 prepare_lock_switch(rq, next);
1746 prepare_arch_switch(next);
1750 * finish_task_switch - clean up after a task-switch
1751 * @rq: runqueue associated with task-switch
1752 * @prev: the thread we just switched away from.
1754 * finish_task_switch must be called after the context switch, paired
1755 * with a prepare_task_switch call before the context switch.
1756 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1757 * and do any other architecture-specific cleanup actions.
1759 * Note that we may have delayed dropping an mm in context_switch(). If
1760 * so, we finish that here outside of the runqueue lock. (Doing it
1761 * with the lock held can cause deadlocks; see schedule() for
1764 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1765 __releases(rq->lock)
1767 struct mm_struct *mm = rq->prev_mm;
1773 * A task struct has one reference for the use as "current".
1774 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1775 * schedule one last time. The schedule call will never return, and
1776 * the scheduled task must drop that reference.
1777 * The test for TASK_DEAD must occur while the runqueue locks are
1778 * still held, otherwise prev could be scheduled on another cpu, die
1779 * there before we look at prev->state, and then the reference would
1781 * Manfred Spraul <manfred@colorfullife.com>
1783 prev_state = prev->state;
1784 finish_arch_switch(prev);
1785 finish_lock_switch(rq, prev);
1786 fire_sched_in_preempt_notifiers(current);
1789 if (unlikely(prev_state == TASK_DEAD)) {
1791 * Remove function-return probe instances associated with this
1792 * task and put them back on the free list.
1794 kprobe_flush_task(prev);
1795 put_task_struct(prev);
1800 * schedule_tail - first thing a freshly forked thread must call.
1801 * @prev: the thread we just switched away from.
1803 asmlinkage void schedule_tail(struct task_struct *prev)
1804 __releases(rq->lock)
1806 struct rq *rq = this_rq();
1808 finish_task_switch(rq, prev);
1809 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1810 /* In this case, finish_task_switch does not reenable preemption */
1813 if (current->set_child_tid)
1814 put_user(current->pid, current->set_child_tid);
1818 * context_switch - switch to the new MM and the new
1819 * thread's register state.
1822 context_switch(struct rq *rq, struct task_struct *prev,
1823 struct task_struct *next)
1825 struct mm_struct *mm, *oldmm;
1827 prepare_task_switch(rq, prev, next);
1829 oldmm = prev->active_mm;
1831 * For paravirt, this is coupled with an exit in switch_to to
1832 * combine the page table reload and the switch backend into
1835 arch_enter_lazy_cpu_mode();
1837 if (unlikely(!mm)) {
1838 next->active_mm = oldmm;
1839 atomic_inc(&oldmm->mm_count);
1840 enter_lazy_tlb(oldmm, next);
1842 switch_mm(oldmm, mm, next);
1844 if (unlikely(!prev->mm)) {
1845 prev->active_mm = NULL;
1846 rq->prev_mm = oldmm;
1849 * Since the runqueue lock will be released by the next
1850 * task (which is an invalid locking op but in the case
1851 * of the scheduler it's an obvious special-case), so we
1852 * do an early lockdep release here:
1854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1855 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1858 /* Here we just switch the register state and the stack. */
1859 switch_to(prev, next, prev);
1863 * this_rq must be evaluated again because prev may have moved
1864 * CPUs since it called schedule(), thus the 'rq' on its stack
1865 * frame will be invalid.
1867 finish_task_switch(this_rq(), prev);
1871 * nr_running, nr_uninterruptible and nr_context_switches:
1873 * externally visible scheduler statistics: current number of runnable
1874 * threads, current number of uninterruptible-sleeping threads, total
1875 * number of context switches performed since bootup.
1877 unsigned long nr_running(void)
1879 unsigned long i, sum = 0;
1881 for_each_online_cpu(i)
1882 sum += cpu_rq(i)->nr_running;
1887 unsigned long nr_uninterruptible(void)
1889 unsigned long i, sum = 0;
1891 for_each_possible_cpu(i)
1892 sum += cpu_rq(i)->nr_uninterruptible;
1895 * Since we read the counters lockless, it might be slightly
1896 * inaccurate. Do not allow it to go below zero though:
1898 if (unlikely((long)sum < 0))
1904 unsigned long long nr_context_switches(void)
1907 unsigned long long sum = 0;
1909 for_each_possible_cpu(i)
1910 sum += cpu_rq(i)->nr_switches;
1915 unsigned long nr_iowait(void)
1917 unsigned long i, sum = 0;
1919 for_each_possible_cpu(i)
1920 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1925 unsigned long nr_active(void)
1927 unsigned long i, running = 0, uninterruptible = 0;
1929 for_each_online_cpu(i) {
1930 running += cpu_rq(i)->nr_running;
1931 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1934 if (unlikely((long)uninterruptible < 0))
1935 uninterruptible = 0;
1937 return running + uninterruptible;
1941 * Update rq->cpu_load[] statistics. This function is usually called every
1942 * scheduler tick (TICK_NSEC).
1944 static void update_cpu_load(struct rq *this_rq)
1946 unsigned long this_load = this_rq->ls.load.weight;
1949 this_rq->nr_load_updates++;
1951 /* Update our load: */
1952 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1953 unsigned long old_load, new_load;
1955 /* scale is effectively 1 << i now, and >> i divides by scale */
1957 old_load = this_rq->cpu_load[i];
1958 new_load = this_load;
1960 * Round up the averaging division if load is increasing. This
1961 * prevents us from getting stuck on 9 if the load is 10, for
1964 if (new_load > old_load)
1965 new_load += scale-1;
1966 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
1973 * double_rq_lock - safely lock two runqueues
1975 * Note this does not disable interrupts like task_rq_lock,
1976 * you need to do so manually before calling.
1978 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1979 __acquires(rq1->lock)
1980 __acquires(rq2->lock)
1982 BUG_ON(!irqs_disabled());
1984 spin_lock(&rq1->lock);
1985 __acquire(rq2->lock); /* Fake it out ;) */
1988 spin_lock(&rq1->lock);
1989 spin_lock(&rq2->lock);
1991 spin_lock(&rq2->lock);
1992 spin_lock(&rq1->lock);
1995 update_rq_clock(rq1);
1996 update_rq_clock(rq2);
2000 * double_rq_unlock - safely unlock two runqueues
2002 * Note this does not restore interrupts like task_rq_unlock,
2003 * you need to do so manually after calling.
2005 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2006 __releases(rq1->lock)
2007 __releases(rq2->lock)
2009 spin_unlock(&rq1->lock);
2011 spin_unlock(&rq2->lock);
2013 __release(rq2->lock);
2017 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2019 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2020 __releases(this_rq->lock)
2021 __acquires(busiest->lock)
2022 __acquires(this_rq->lock)
2024 if (unlikely(!irqs_disabled())) {
2025 /* printk() doesn't work good under rq->lock */
2026 spin_unlock(&this_rq->lock);
2029 if (unlikely(!spin_trylock(&busiest->lock))) {
2030 if (busiest < this_rq) {
2031 spin_unlock(&this_rq->lock);
2032 spin_lock(&busiest->lock);
2033 spin_lock(&this_rq->lock);
2035 spin_lock(&busiest->lock);
2040 * If dest_cpu is allowed for this process, migrate the task to it.
2041 * This is accomplished by forcing the cpu_allowed mask to only
2042 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2043 * the cpu_allowed mask is restored.
2045 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2047 struct migration_req req;
2048 unsigned long flags;
2051 rq = task_rq_lock(p, &flags);
2052 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2053 || unlikely(cpu_is_offline(dest_cpu)))
2056 /* force the process onto the specified CPU */
2057 if (migrate_task(p, dest_cpu, &req)) {
2058 /* Need to wait for migration thread (might exit: take ref). */
2059 struct task_struct *mt = rq->migration_thread;
2061 get_task_struct(mt);
2062 task_rq_unlock(rq, &flags);
2063 wake_up_process(mt);
2064 put_task_struct(mt);
2065 wait_for_completion(&req.done);
2070 task_rq_unlock(rq, &flags);
2074 * sched_exec - execve() is a valuable balancing opportunity, because at
2075 * this point the task has the smallest effective memory and cache footprint.
2077 void sched_exec(void)
2079 int new_cpu, this_cpu = get_cpu();
2080 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2082 if (new_cpu != this_cpu)
2083 sched_migrate_task(current, new_cpu);
2087 * pull_task - move a task from a remote runqueue to the local runqueue.
2088 * Both runqueues must be locked.
2090 static void pull_task(struct rq *src_rq, struct task_struct *p,
2091 struct rq *this_rq, int this_cpu)
2093 deactivate_task(src_rq, p, 0);
2094 set_task_cpu(p, this_cpu);
2095 activate_task(this_rq, p, 0);
2097 * Note that idle threads have a prio of MAX_PRIO, for this test
2098 * to be always true for them.
2100 check_preempt_curr(this_rq, p);
2104 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2107 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2108 struct sched_domain *sd, enum cpu_idle_type idle,
2112 * We do not migrate tasks that are:
2113 * 1) running (obviously), or
2114 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2115 * 3) are cache-hot on their current CPU.
2117 if (!cpu_isset(this_cpu, p->cpus_allowed))
2121 if (task_running(rq, p))
2127 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2128 unsigned long max_nr_move, unsigned long max_load_move,
2129 struct sched_domain *sd, enum cpu_idle_type idle,
2130 int *all_pinned, unsigned long *load_moved,
2131 int *this_best_prio, struct rq_iterator *iterator)
2133 int pulled = 0, pinned = 0, skip_for_load;
2134 struct task_struct *p;
2135 long rem_load_move = max_load_move;
2137 if (max_nr_move == 0 || max_load_move == 0)
2143 * Start the load-balancing iterator:
2145 p = iterator->start(iterator->arg);
2150 * To help distribute high priority tasks accross CPUs we don't
2151 * skip a task if it will be the highest priority task (i.e. smallest
2152 * prio value) on its new queue regardless of its load weight
2154 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2155 SCHED_LOAD_SCALE_FUZZ;
2156 if ((skip_for_load && p->prio >= *this_best_prio) ||
2157 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2158 p = iterator->next(iterator->arg);
2162 pull_task(busiest, p, this_rq, this_cpu);
2164 rem_load_move -= p->se.load.weight;
2167 * We only want to steal up to the prescribed number of tasks
2168 * and the prescribed amount of weighted load.
2170 if (pulled < max_nr_move && rem_load_move > 0) {
2171 if (p->prio < *this_best_prio)
2172 *this_best_prio = p->prio;
2173 p = iterator->next(iterator->arg);
2178 * Right now, this is the only place pull_task() is called,
2179 * so we can safely collect pull_task() stats here rather than
2180 * inside pull_task().
2182 schedstat_add(sd, lb_gained[idle], pulled);
2185 *all_pinned = pinned;
2186 *load_moved = max_load_move - rem_load_move;
2191 * move_tasks tries to move up to max_load_move weighted load from busiest to
2192 * this_rq, as part of a balancing operation within domain "sd".
2193 * Returns 1 if successful and 0 otherwise.
2195 * Called with both runqueues locked.
2197 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2198 unsigned long max_load_move,
2199 struct sched_domain *sd, enum cpu_idle_type idle,
2202 struct sched_class *class = sched_class_highest;
2203 unsigned long total_load_moved = 0;
2204 int this_best_prio = this_rq->curr->prio;
2208 class->load_balance(this_rq, this_cpu, busiest,
2209 ULONG_MAX, max_load_move - total_load_moved,
2210 sd, idle, all_pinned, &this_best_prio);
2211 class = class->next;
2212 } while (class && max_load_move > total_load_moved);
2214 return total_load_moved > 0;
2218 * move_one_task tries to move exactly one task from busiest to this_rq, as
2219 * part of active balancing operations within "domain".
2220 * Returns 1 if successful and 0 otherwise.
2222 * Called with both runqueues locked.
2224 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2225 struct sched_domain *sd, enum cpu_idle_type idle)
2227 struct sched_class *class;
2228 int this_best_prio = MAX_PRIO;
2230 for (class = sched_class_highest; class; class = class->next)
2231 if (class->load_balance(this_rq, this_cpu, busiest,
2232 1, ULONG_MAX, sd, idle, NULL,
2240 * find_busiest_group finds and returns the busiest CPU group within the
2241 * domain. It calculates and returns the amount of weighted load which
2242 * should be moved to restore balance via the imbalance parameter.
2244 static struct sched_group *
2245 find_busiest_group(struct sched_domain *sd, int this_cpu,
2246 unsigned long *imbalance, enum cpu_idle_type idle,
2247 int *sd_idle, cpumask_t *cpus, int *balance)
2249 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2250 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2251 unsigned long max_pull;
2252 unsigned long busiest_load_per_task, busiest_nr_running;
2253 unsigned long this_load_per_task, this_nr_running;
2255 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2256 int power_savings_balance = 1;
2257 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2258 unsigned long min_nr_running = ULONG_MAX;
2259 struct sched_group *group_min = NULL, *group_leader = NULL;
2262 max_load = this_load = total_load = total_pwr = 0;
2263 busiest_load_per_task = busiest_nr_running = 0;
2264 this_load_per_task = this_nr_running = 0;
2265 if (idle == CPU_NOT_IDLE)
2266 load_idx = sd->busy_idx;
2267 else if (idle == CPU_NEWLY_IDLE)
2268 load_idx = sd->newidle_idx;
2270 load_idx = sd->idle_idx;
2273 unsigned long load, group_capacity;
2276 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2277 unsigned long sum_nr_running, sum_weighted_load;
2279 local_group = cpu_isset(this_cpu, group->cpumask);
2282 balance_cpu = first_cpu(group->cpumask);
2284 /* Tally up the load of all CPUs in the group */
2285 sum_weighted_load = sum_nr_running = avg_load = 0;
2287 for_each_cpu_mask(i, group->cpumask) {
2290 if (!cpu_isset(i, *cpus))
2295 if (*sd_idle && rq->nr_running)
2298 /* Bias balancing toward cpus of our domain */
2300 if (idle_cpu(i) && !first_idle_cpu) {
2305 load = target_load(i, load_idx);
2307 load = source_load(i, load_idx);
2310 sum_nr_running += rq->nr_running;
2311 sum_weighted_load += weighted_cpuload(i);
2315 * First idle cpu or the first cpu(busiest) in this sched group
2316 * is eligible for doing load balancing at this and above
2317 * domains. In the newly idle case, we will allow all the cpu's
2318 * to do the newly idle load balance.
2320 if (idle != CPU_NEWLY_IDLE && local_group &&
2321 balance_cpu != this_cpu && balance) {
2326 total_load += avg_load;
2327 total_pwr += group->__cpu_power;
2329 /* Adjust by relative CPU power of the group */
2330 avg_load = sg_div_cpu_power(group,
2331 avg_load * SCHED_LOAD_SCALE);
2333 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2336 this_load = avg_load;
2338 this_nr_running = sum_nr_running;
2339 this_load_per_task = sum_weighted_load;
2340 } else if (avg_load > max_load &&
2341 sum_nr_running > group_capacity) {
2342 max_load = avg_load;
2344 busiest_nr_running = sum_nr_running;
2345 busiest_load_per_task = sum_weighted_load;
2348 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2350 * Busy processors will not participate in power savings
2353 if (idle == CPU_NOT_IDLE ||
2354 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2358 * If the local group is idle or completely loaded
2359 * no need to do power savings balance at this domain
2361 if (local_group && (this_nr_running >= group_capacity ||
2363 power_savings_balance = 0;
2366 * If a group is already running at full capacity or idle,
2367 * don't include that group in power savings calculations
2369 if (!power_savings_balance || sum_nr_running >= group_capacity
2374 * Calculate the group which has the least non-idle load.
2375 * This is the group from where we need to pick up the load
2378 if ((sum_nr_running < min_nr_running) ||
2379 (sum_nr_running == min_nr_running &&
2380 first_cpu(group->cpumask) <
2381 first_cpu(group_min->cpumask))) {
2383 min_nr_running = sum_nr_running;
2384 min_load_per_task = sum_weighted_load /
2389 * Calculate the group which is almost near its
2390 * capacity but still has some space to pick up some load
2391 * from other group and save more power
2393 if (sum_nr_running <= group_capacity - 1) {
2394 if (sum_nr_running > leader_nr_running ||
2395 (sum_nr_running == leader_nr_running &&
2396 first_cpu(group->cpumask) >
2397 first_cpu(group_leader->cpumask))) {
2398 group_leader = group;
2399 leader_nr_running = sum_nr_running;
2404 group = group->next;
2405 } while (group != sd->groups);
2407 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2410 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2412 if (this_load >= avg_load ||
2413 100*max_load <= sd->imbalance_pct*this_load)
2416 busiest_load_per_task /= busiest_nr_running;
2418 * We're trying to get all the cpus to the average_load, so we don't
2419 * want to push ourselves above the average load, nor do we wish to
2420 * reduce the max loaded cpu below the average load, as either of these
2421 * actions would just result in more rebalancing later, and ping-pong
2422 * tasks around. Thus we look for the minimum possible imbalance.
2423 * Negative imbalances (*we* are more loaded than anyone else) will
2424 * be counted as no imbalance for these purposes -- we can't fix that
2425 * by pulling tasks to us. Be careful of negative numbers as they'll
2426 * appear as very large values with unsigned longs.
2428 if (max_load <= busiest_load_per_task)
2432 * In the presence of smp nice balancing, certain scenarios can have
2433 * max load less than avg load(as we skip the groups at or below
2434 * its cpu_power, while calculating max_load..)
2436 if (max_load < avg_load) {
2438 goto small_imbalance;
2441 /* Don't want to pull so many tasks that a group would go idle */
2442 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2444 /* How much load to actually move to equalise the imbalance */
2445 *imbalance = min(max_pull * busiest->__cpu_power,
2446 (avg_load - this_load) * this->__cpu_power)
2450 * if *imbalance is less than the average load per runnable task
2451 * there is no gaurantee that any tasks will be moved so we'll have
2452 * a think about bumping its value to force at least one task to be
2455 if (*imbalance < busiest_load_per_task) {
2456 unsigned long tmp, pwr_now, pwr_move;
2460 pwr_move = pwr_now = 0;
2462 if (this_nr_running) {
2463 this_load_per_task /= this_nr_running;
2464 if (busiest_load_per_task > this_load_per_task)
2467 this_load_per_task = SCHED_LOAD_SCALE;
2469 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2470 busiest_load_per_task * imbn) {
2471 *imbalance = busiest_load_per_task;
2476 * OK, we don't have enough imbalance to justify moving tasks,
2477 * however we may be able to increase total CPU power used by
2481 pwr_now += busiest->__cpu_power *
2482 min(busiest_load_per_task, max_load);
2483 pwr_now += this->__cpu_power *
2484 min(this_load_per_task, this_load);
2485 pwr_now /= SCHED_LOAD_SCALE;
2487 /* Amount of load we'd subtract */
2488 tmp = sg_div_cpu_power(busiest,
2489 busiest_load_per_task * SCHED_LOAD_SCALE);
2491 pwr_move += busiest->__cpu_power *
2492 min(busiest_load_per_task, max_load - tmp);
2494 /* Amount of load we'd add */
2495 if (max_load * busiest->__cpu_power <
2496 busiest_load_per_task * SCHED_LOAD_SCALE)
2497 tmp = sg_div_cpu_power(this,
2498 max_load * busiest->__cpu_power);
2500 tmp = sg_div_cpu_power(this,
2501 busiest_load_per_task * SCHED_LOAD_SCALE);
2502 pwr_move += this->__cpu_power *
2503 min(this_load_per_task, this_load + tmp);
2504 pwr_move /= SCHED_LOAD_SCALE;
2506 /* Move if we gain throughput */
2507 if (pwr_move > pwr_now)
2508 *imbalance = busiest_load_per_task;
2514 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2515 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2518 if (this == group_leader && group_leader != group_min) {
2519 *imbalance = min_load_per_task;
2529 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2532 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2533 unsigned long imbalance, cpumask_t *cpus)
2535 struct rq *busiest = NULL, *rq;
2536 unsigned long max_load = 0;
2539 for_each_cpu_mask(i, group->cpumask) {
2542 if (!cpu_isset(i, *cpus))
2546 wl = weighted_cpuload(i);
2548 if (rq->nr_running == 1 && wl > imbalance)
2551 if (wl > max_load) {
2561 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2562 * so long as it is large enough.
2564 #define MAX_PINNED_INTERVAL 512
2567 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2568 * tasks if there is an imbalance.
2570 static int load_balance(int this_cpu, struct rq *this_rq,
2571 struct sched_domain *sd, enum cpu_idle_type idle,
2574 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2575 struct sched_group *group;
2576 unsigned long imbalance;
2578 cpumask_t cpus = CPU_MASK_ALL;
2579 unsigned long flags;
2582 * When power savings policy is enabled for the parent domain, idle
2583 * sibling can pick up load irrespective of busy siblings. In this case,
2584 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2585 * portraying it as CPU_NOT_IDLE.
2587 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2588 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2591 schedstat_inc(sd, lb_cnt[idle]);
2594 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2601 schedstat_inc(sd, lb_nobusyg[idle]);
2605 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2607 schedstat_inc(sd, lb_nobusyq[idle]);
2611 BUG_ON(busiest == this_rq);
2613 schedstat_add(sd, lb_imbalance[idle], imbalance);
2616 if (busiest->nr_running > 1) {
2618 * Attempt to move tasks. If find_busiest_group has found
2619 * an imbalance but busiest->nr_running <= 1, the group is
2620 * still unbalanced. ld_moved simply stays zero, so it is
2621 * correctly treated as an imbalance.
2623 local_irq_save(flags);
2624 double_rq_lock(this_rq, busiest);
2625 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2626 imbalance, sd, idle, &all_pinned);
2627 double_rq_unlock(this_rq, busiest);
2628 local_irq_restore(flags);
2631 * some other cpu did the load balance for us.
2633 if (ld_moved && this_cpu != smp_processor_id())
2634 resched_cpu(this_cpu);
2636 /* All tasks on this runqueue were pinned by CPU affinity */
2637 if (unlikely(all_pinned)) {
2638 cpu_clear(cpu_of(busiest), cpus);
2639 if (!cpus_empty(cpus))
2646 schedstat_inc(sd, lb_failed[idle]);
2647 sd->nr_balance_failed++;
2649 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2651 spin_lock_irqsave(&busiest->lock, flags);
2653 /* don't kick the migration_thread, if the curr
2654 * task on busiest cpu can't be moved to this_cpu
2656 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2657 spin_unlock_irqrestore(&busiest->lock, flags);
2659 goto out_one_pinned;
2662 if (!busiest->active_balance) {
2663 busiest->active_balance = 1;
2664 busiest->push_cpu = this_cpu;
2667 spin_unlock_irqrestore(&busiest->lock, flags);
2669 wake_up_process(busiest->migration_thread);
2672 * We've kicked active balancing, reset the failure
2675 sd->nr_balance_failed = sd->cache_nice_tries+1;
2678 sd->nr_balance_failed = 0;
2680 if (likely(!active_balance)) {
2681 /* We were unbalanced, so reset the balancing interval */
2682 sd->balance_interval = sd->min_interval;
2685 * If we've begun active balancing, start to back off. This
2686 * case may not be covered by the all_pinned logic if there
2687 * is only 1 task on the busy runqueue (because we don't call
2690 if (sd->balance_interval < sd->max_interval)
2691 sd->balance_interval *= 2;
2694 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2695 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2700 schedstat_inc(sd, lb_balanced[idle]);
2702 sd->nr_balance_failed = 0;
2705 /* tune up the balancing interval */
2706 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2707 (sd->balance_interval < sd->max_interval))
2708 sd->balance_interval *= 2;
2710 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2711 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2717 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2718 * tasks if there is an imbalance.
2720 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2721 * this_rq is locked.
2724 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2726 struct sched_group *group;
2727 struct rq *busiest = NULL;
2728 unsigned long imbalance;
2732 cpumask_t cpus = CPU_MASK_ALL;
2735 * When power savings policy is enabled for the parent domain, idle
2736 * sibling can pick up load irrespective of busy siblings. In this case,
2737 * let the state of idle sibling percolate up as IDLE, instead of
2738 * portraying it as CPU_NOT_IDLE.
2740 if (sd->flags & SD_SHARE_CPUPOWER &&
2741 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2744 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2746 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2747 &sd_idle, &cpus, NULL);
2749 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2753 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2756 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2760 BUG_ON(busiest == this_rq);
2762 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2765 if (busiest->nr_running > 1) {
2766 /* Attempt to move tasks */
2767 double_lock_balance(this_rq, busiest);
2768 /* this_rq->clock is already updated */
2769 update_rq_clock(busiest);
2770 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2771 imbalance, sd, CPU_NEWLY_IDLE,
2773 spin_unlock(&busiest->lock);
2775 if (unlikely(all_pinned)) {
2776 cpu_clear(cpu_of(busiest), cpus);
2777 if (!cpus_empty(cpus))
2783 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2784 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2785 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2788 sd->nr_balance_failed = 0;
2793 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2794 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2795 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2797 sd->nr_balance_failed = 0;
2803 * idle_balance is called by schedule() if this_cpu is about to become
2804 * idle. Attempts to pull tasks from other CPUs.
2806 static void idle_balance(int this_cpu, struct rq *this_rq)
2808 struct sched_domain *sd;
2809 int pulled_task = -1;
2810 unsigned long next_balance = jiffies + HZ;
2812 for_each_domain(this_cpu, sd) {
2813 unsigned long interval;
2815 if (!(sd->flags & SD_LOAD_BALANCE))
2818 if (sd->flags & SD_BALANCE_NEWIDLE)
2819 /* If we've pulled tasks over stop searching: */
2820 pulled_task = load_balance_newidle(this_cpu,
2823 interval = msecs_to_jiffies(sd->balance_interval);
2824 if (time_after(next_balance, sd->last_balance + interval))
2825 next_balance = sd->last_balance + interval;
2829 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2831 * We are going idle. next_balance may be set based on
2832 * a busy processor. So reset next_balance.
2834 this_rq->next_balance = next_balance;
2839 * active_load_balance is run by migration threads. It pushes running tasks
2840 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2841 * running on each physical CPU where possible, and avoids physical /
2842 * logical imbalances.
2844 * Called with busiest_rq locked.
2846 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2848 int target_cpu = busiest_rq->push_cpu;
2849 struct sched_domain *sd;
2850 struct rq *target_rq;
2852 /* Is there any task to move? */
2853 if (busiest_rq->nr_running <= 1)
2856 target_rq = cpu_rq(target_cpu);
2859 * This condition is "impossible", if it occurs
2860 * we need to fix it. Originally reported by
2861 * Bjorn Helgaas on a 128-cpu setup.
2863 BUG_ON(busiest_rq == target_rq);
2865 /* move a task from busiest_rq to target_rq */
2866 double_lock_balance(busiest_rq, target_rq);
2867 update_rq_clock(busiest_rq);
2868 update_rq_clock(target_rq);
2870 /* Search for an sd spanning us and the target CPU. */
2871 for_each_domain(target_cpu, sd) {
2872 if ((sd->flags & SD_LOAD_BALANCE) &&
2873 cpu_isset(busiest_cpu, sd->span))
2878 schedstat_inc(sd, alb_cnt);
2880 if (move_one_task(target_rq, target_cpu, busiest_rq,
2882 schedstat_inc(sd, alb_pushed);
2884 schedstat_inc(sd, alb_failed);
2886 spin_unlock(&target_rq->lock);
2891 atomic_t load_balancer;
2893 } nohz ____cacheline_aligned = {
2894 .load_balancer = ATOMIC_INIT(-1),
2895 .cpu_mask = CPU_MASK_NONE,
2899 * This routine will try to nominate the ilb (idle load balancing)
2900 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2901 * load balancing on behalf of all those cpus. If all the cpus in the system
2902 * go into this tickless mode, then there will be no ilb owner (as there is
2903 * no need for one) and all the cpus will sleep till the next wakeup event
2906 * For the ilb owner, tick is not stopped. And this tick will be used
2907 * for idle load balancing. ilb owner will still be part of
2910 * While stopping the tick, this cpu will become the ilb owner if there
2911 * is no other owner. And will be the owner till that cpu becomes busy
2912 * or if all cpus in the system stop their ticks at which point
2913 * there is no need for ilb owner.
2915 * When the ilb owner becomes busy, it nominates another owner, during the
2916 * next busy scheduler_tick()
2918 int select_nohz_load_balancer(int stop_tick)
2920 int cpu = smp_processor_id();
2923 cpu_set(cpu, nohz.cpu_mask);
2924 cpu_rq(cpu)->in_nohz_recently = 1;
2927 * If we are going offline and still the leader, give up!
2929 if (cpu_is_offline(cpu) &&
2930 atomic_read(&nohz.load_balancer) == cpu) {
2931 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2936 /* time for ilb owner also to sleep */
2937 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2938 if (atomic_read(&nohz.load_balancer) == cpu)
2939 atomic_set(&nohz.load_balancer, -1);
2943 if (atomic_read(&nohz.load_balancer) == -1) {
2944 /* make me the ilb owner */
2945 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2947 } else if (atomic_read(&nohz.load_balancer) == cpu)
2950 if (!cpu_isset(cpu, nohz.cpu_mask))
2953 cpu_clear(cpu, nohz.cpu_mask);
2955 if (atomic_read(&nohz.load_balancer) == cpu)
2956 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2963 static DEFINE_SPINLOCK(balancing);
2966 * It checks each scheduling domain to see if it is due to be balanced,
2967 * and initiates a balancing operation if so.
2969 * Balancing parameters are set up in arch_init_sched_domains.
2971 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2974 struct rq *rq = cpu_rq(cpu);
2975 unsigned long interval;
2976 struct sched_domain *sd;
2977 /* Earliest time when we have to do rebalance again */
2978 unsigned long next_balance = jiffies + 60*HZ;
2979 int update_next_balance = 0;
2981 for_each_domain(cpu, sd) {
2982 if (!(sd->flags & SD_LOAD_BALANCE))
2985 interval = sd->balance_interval;
2986 if (idle != CPU_IDLE)
2987 interval *= sd->busy_factor;
2989 /* scale ms to jiffies */
2990 interval = msecs_to_jiffies(interval);
2991 if (unlikely(!interval))
2993 if (interval > HZ*NR_CPUS/10)
2994 interval = HZ*NR_CPUS/10;
2997 if (sd->flags & SD_SERIALIZE) {
2998 if (!spin_trylock(&balancing))
3002 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3003 if (load_balance(cpu, rq, sd, idle, &balance)) {
3005 * We've pulled tasks over so either we're no
3006 * longer idle, or one of our SMT siblings is
3009 idle = CPU_NOT_IDLE;
3011 sd->last_balance = jiffies;
3013 if (sd->flags & SD_SERIALIZE)
3014 spin_unlock(&balancing);
3016 if (time_after(next_balance, sd->last_balance + interval)) {
3017 next_balance = sd->last_balance + interval;
3018 update_next_balance = 1;
3022 * Stop the load balance at this level. There is another
3023 * CPU in our sched group which is doing load balancing more
3031 * next_balance will be updated only when there is a need.
3032 * When the cpu is attached to null domain for ex, it will not be
3035 if (likely(update_next_balance))
3036 rq->next_balance = next_balance;
3040 * run_rebalance_domains is triggered when needed from the scheduler tick.
3041 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3042 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3044 static void run_rebalance_domains(struct softirq_action *h)
3046 int this_cpu = smp_processor_id();
3047 struct rq *this_rq = cpu_rq(this_cpu);
3048 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3049 CPU_IDLE : CPU_NOT_IDLE;
3051 rebalance_domains(this_cpu, idle);
3055 * If this cpu is the owner for idle load balancing, then do the
3056 * balancing on behalf of the other idle cpus whose ticks are
3059 if (this_rq->idle_at_tick &&
3060 atomic_read(&nohz.load_balancer) == this_cpu) {
3061 cpumask_t cpus = nohz.cpu_mask;
3065 cpu_clear(this_cpu, cpus);
3066 for_each_cpu_mask(balance_cpu, cpus) {
3068 * If this cpu gets work to do, stop the load balancing
3069 * work being done for other cpus. Next load
3070 * balancing owner will pick it up.
3075 rebalance_domains(balance_cpu, CPU_IDLE);
3077 rq = cpu_rq(balance_cpu);
3078 if (time_after(this_rq->next_balance, rq->next_balance))
3079 this_rq->next_balance = rq->next_balance;
3086 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3088 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3089 * idle load balancing owner or decide to stop the periodic load balancing,
3090 * if the whole system is idle.
3092 static inline void trigger_load_balance(struct rq *rq, int cpu)
3096 * If we were in the nohz mode recently and busy at the current
3097 * scheduler tick, then check if we need to nominate new idle
3100 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3101 rq->in_nohz_recently = 0;
3103 if (atomic_read(&nohz.load_balancer) == cpu) {
3104 cpu_clear(cpu, nohz.cpu_mask);
3105 atomic_set(&nohz.load_balancer, -1);
3108 if (atomic_read(&nohz.load_balancer) == -1) {
3110 * simple selection for now: Nominate the
3111 * first cpu in the nohz list to be the next
3114 * TBD: Traverse the sched domains and nominate
3115 * the nearest cpu in the nohz.cpu_mask.
3117 int ilb = first_cpu(nohz.cpu_mask);
3125 * If this cpu is idle and doing idle load balancing for all the
3126 * cpus with ticks stopped, is it time for that to stop?
3128 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3129 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3135 * If this cpu is idle and the idle load balancing is done by
3136 * someone else, then no need raise the SCHED_SOFTIRQ
3138 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3139 cpu_isset(cpu, nohz.cpu_mask))
3142 if (time_after_eq(jiffies, rq->next_balance))
3143 raise_softirq(SCHED_SOFTIRQ);
3146 #else /* CONFIG_SMP */
3149 * on UP we do not need to balance between CPUs:
3151 static inline void idle_balance(int cpu, struct rq *rq)
3155 /* Avoid "used but not defined" warning on UP */
3156 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3157 unsigned long max_nr_move, unsigned long max_load_move,
3158 struct sched_domain *sd, enum cpu_idle_type idle,
3159 int *all_pinned, unsigned long *load_moved,
3160 int *this_best_prio, struct rq_iterator *iterator)
3169 DEFINE_PER_CPU(struct kernel_stat, kstat);
3171 EXPORT_PER_CPU_SYMBOL(kstat);
3174 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3175 * that have not yet been banked in case the task is currently running.
3177 unsigned long long task_sched_runtime(struct task_struct *p)
3179 unsigned long flags;
3183 rq = task_rq_lock(p, &flags);
3184 ns = p->se.sum_exec_runtime;
3185 if (rq->curr == p) {
3186 update_rq_clock(rq);
3187 delta_exec = rq->clock - p->se.exec_start;
3188 if ((s64)delta_exec > 0)
3191 task_rq_unlock(rq, &flags);
3197 * Account user cpu time to a process.
3198 * @p: the process that the cpu time gets accounted to
3199 * @hardirq_offset: the offset to subtract from hardirq_count()
3200 * @cputime: the cpu time spent in user space since the last update
3202 void account_user_time(struct task_struct *p, cputime_t cputime)
3204 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3207 p->utime = cputime_add(p->utime, cputime);
3209 /* Add user time to cpustat. */
3210 tmp = cputime_to_cputime64(cputime);
3211 if (TASK_NICE(p) > 0)
3212 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3214 cpustat->user = cputime64_add(cpustat->user, tmp);
3218 * Account system cpu time to a process.
3219 * @p: the process that the cpu time gets accounted to
3220 * @hardirq_offset: the offset to subtract from hardirq_count()
3221 * @cputime: the cpu time spent in kernel space since the last update
3223 void account_system_time(struct task_struct *p, int hardirq_offset,
3226 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3227 struct rq *rq = this_rq();
3230 p->stime = cputime_add(p->stime, cputime);
3232 /* Add system time to cpustat. */
3233 tmp = cputime_to_cputime64(cputime);
3234 if (hardirq_count() - hardirq_offset)
3235 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3236 else if (softirq_count())
3237 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3238 else if (p != rq->idle)
3239 cpustat->system = cputime64_add(cpustat->system, tmp);
3240 else if (atomic_read(&rq->nr_iowait) > 0)
3241 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3243 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3244 /* Account for system time used */
3245 acct_update_integrals(p);
3249 * Account for involuntary wait time.
3250 * @p: the process from which the cpu time has been stolen
3251 * @steal: the cpu time spent in involuntary wait
3253 void account_steal_time(struct task_struct *p, cputime_t steal)
3255 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3256 cputime64_t tmp = cputime_to_cputime64(steal);
3257 struct rq *rq = this_rq();
3259 if (p == rq->idle) {
3260 p->stime = cputime_add(p->stime, steal);
3261 if (atomic_read(&rq->nr_iowait) > 0)
3262 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3264 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3266 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3270 * This function gets called by the timer code, with HZ frequency.
3271 * We call it with interrupts disabled.
3273 * It also gets called by the fork code, when changing the parent's
3276 void scheduler_tick(void)
3278 int cpu = smp_processor_id();
3279 struct rq *rq = cpu_rq(cpu);
3280 struct task_struct *curr = rq->curr;
3281 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3283 spin_lock(&rq->lock);
3284 __update_rq_clock(rq);
3286 * Let rq->clock advance by at least TICK_NSEC:
3288 if (unlikely(rq->clock < next_tick))
3289 rq->clock = next_tick;
3290 rq->tick_timestamp = rq->clock;
3291 update_cpu_load(rq);
3292 if (curr != rq->idle) /* FIXME: needed? */
3293 curr->sched_class->task_tick(rq, curr);
3294 spin_unlock(&rq->lock);
3297 rq->idle_at_tick = idle_cpu(cpu);
3298 trigger_load_balance(rq, cpu);
3302 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3304 void fastcall add_preempt_count(int val)
3309 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3311 preempt_count() += val;
3313 * Spinlock count overflowing soon?
3315 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3318 EXPORT_SYMBOL(add_preempt_count);
3320 void fastcall sub_preempt_count(int val)
3325 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3328 * Is the spinlock portion underflowing?
3330 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3331 !(preempt_count() & PREEMPT_MASK)))
3334 preempt_count() -= val;
3336 EXPORT_SYMBOL(sub_preempt_count);
3341 * Print scheduling while atomic bug:
3343 static noinline void __schedule_bug(struct task_struct *prev)
3345 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3346 prev->comm, preempt_count(), prev->pid);
3347 debug_show_held_locks(prev);
3348 if (irqs_disabled())
3349 print_irqtrace_events(prev);
3354 * Various schedule()-time debugging checks and statistics:
3356 static inline void schedule_debug(struct task_struct *prev)
3359 * Test if we are atomic. Since do_exit() needs to call into
3360 * schedule() atomically, we ignore that path for now.
3361 * Otherwise, whine if we are scheduling when we should not be.
3363 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3364 __schedule_bug(prev);
3366 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3368 schedstat_inc(this_rq(), sched_cnt);
3372 * Pick up the highest-prio task:
3374 static inline struct task_struct *
3375 pick_next_task(struct rq *rq, struct task_struct *prev)
3377 struct sched_class *class;
3378 struct task_struct *p;
3381 * Optimization: we know that if all tasks are in
3382 * the fair class we can call that function directly:
3384 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3385 p = fair_sched_class.pick_next_task(rq);
3390 class = sched_class_highest;
3392 p = class->pick_next_task(rq);
3396 * Will never be NULL as the idle class always
3397 * returns a non-NULL p:
3399 class = class->next;
3404 * schedule() is the main scheduler function.
3406 asmlinkage void __sched schedule(void)
3408 struct task_struct *prev, *next;
3415 cpu = smp_processor_id();
3419 switch_count = &prev->nivcsw;
3421 release_kernel_lock(prev);
3422 need_resched_nonpreemptible:
3424 schedule_debug(prev);
3426 spin_lock_irq(&rq->lock);
3427 clear_tsk_need_resched(prev);
3428 __update_rq_clock(rq);
3430 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3431 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3432 unlikely(signal_pending(prev)))) {
3433 prev->state = TASK_RUNNING;
3435 deactivate_task(rq, prev, 1);
3437 switch_count = &prev->nvcsw;
3440 if (unlikely(!rq->nr_running))
3441 idle_balance(cpu, rq);
3443 prev->sched_class->put_prev_task(rq, prev);
3444 next = pick_next_task(rq, prev);
3446 sched_info_switch(prev, next);
3448 if (likely(prev != next)) {
3453 context_switch(rq, prev, next); /* unlocks the rq */
3455 spin_unlock_irq(&rq->lock);
3457 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3458 cpu = smp_processor_id();
3460 goto need_resched_nonpreemptible;
3462 preempt_enable_no_resched();
3463 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3466 EXPORT_SYMBOL(schedule);
3468 #ifdef CONFIG_PREEMPT
3470 * this is the entry point to schedule() from in-kernel preemption
3471 * off of preempt_enable. Kernel preemptions off return from interrupt
3472 * occur there and call schedule directly.
3474 asmlinkage void __sched preempt_schedule(void)
3476 struct thread_info *ti = current_thread_info();
3477 #ifdef CONFIG_PREEMPT_BKL
3478 struct task_struct *task = current;
3479 int saved_lock_depth;
3482 * If there is a non-zero preempt_count or interrupts are disabled,
3483 * we do not want to preempt the current task. Just return..
3485 if (likely(ti->preempt_count || irqs_disabled()))
3489 add_preempt_count(PREEMPT_ACTIVE);
3491 * We keep the big kernel semaphore locked, but we
3492 * clear ->lock_depth so that schedule() doesnt
3493 * auto-release the semaphore:
3495 #ifdef CONFIG_PREEMPT_BKL
3496 saved_lock_depth = task->lock_depth;
3497 task->lock_depth = -1;
3500 #ifdef CONFIG_PREEMPT_BKL
3501 task->lock_depth = saved_lock_depth;
3503 sub_preempt_count(PREEMPT_ACTIVE);
3505 /* we could miss a preemption opportunity between schedule and now */
3507 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3510 EXPORT_SYMBOL(preempt_schedule);
3513 * this is the entry point to schedule() from kernel preemption
3514 * off of irq context.
3515 * Note, that this is called and return with irqs disabled. This will
3516 * protect us against recursive calling from irq.
3518 asmlinkage void __sched preempt_schedule_irq(void)
3520 struct thread_info *ti = current_thread_info();
3521 #ifdef CONFIG_PREEMPT_BKL
3522 struct task_struct *task = current;
3523 int saved_lock_depth;
3525 /* Catch callers which need to be fixed */
3526 BUG_ON(ti->preempt_count || !irqs_disabled());
3529 add_preempt_count(PREEMPT_ACTIVE);
3531 * We keep the big kernel semaphore locked, but we
3532 * clear ->lock_depth so that schedule() doesnt
3533 * auto-release the semaphore:
3535 #ifdef CONFIG_PREEMPT_BKL
3536 saved_lock_depth = task->lock_depth;
3537 task->lock_depth = -1;
3541 local_irq_disable();
3542 #ifdef CONFIG_PREEMPT_BKL
3543 task->lock_depth = saved_lock_depth;
3545 sub_preempt_count(PREEMPT_ACTIVE);
3547 /* we could miss a preemption opportunity between schedule and now */
3549 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3553 #endif /* CONFIG_PREEMPT */
3555 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3558 return try_to_wake_up(curr->private, mode, sync);
3560 EXPORT_SYMBOL(default_wake_function);
3563 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3564 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3565 * number) then we wake all the non-exclusive tasks and one exclusive task.
3567 * There are circumstances in which we can try to wake a task which has already
3568 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3569 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3571 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3572 int nr_exclusive, int sync, void *key)
3574 wait_queue_t *curr, *next;
3576 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3577 unsigned flags = curr->flags;
3579 if (curr->func(curr, mode, sync, key) &&
3580 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3586 * __wake_up - wake up threads blocked on a waitqueue.
3588 * @mode: which threads
3589 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3590 * @key: is directly passed to the wakeup function
3592 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3593 int nr_exclusive, void *key)
3595 unsigned long flags;
3597 spin_lock_irqsave(&q->lock, flags);
3598 __wake_up_common(q, mode, nr_exclusive, 0, key);
3599 spin_unlock_irqrestore(&q->lock, flags);
3601 EXPORT_SYMBOL(__wake_up);
3604 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3606 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3608 __wake_up_common(q, mode, 1, 0, NULL);
3612 * __wake_up_sync - wake up threads blocked on a waitqueue.
3614 * @mode: which threads
3615 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3617 * The sync wakeup differs that the waker knows that it will schedule
3618 * away soon, so while the target thread will be woken up, it will not
3619 * be migrated to another CPU - ie. the two threads are 'synchronized'
3620 * with each other. This can prevent needless bouncing between CPUs.
3622 * On UP it can prevent extra preemption.
3625 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3627 unsigned long flags;
3633 if (unlikely(!nr_exclusive))
3636 spin_lock_irqsave(&q->lock, flags);
3637 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3638 spin_unlock_irqrestore(&q->lock, flags);
3640 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3642 void fastcall complete(struct completion *x)
3644 unsigned long flags;
3646 spin_lock_irqsave(&x->wait.lock, flags);
3648 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3650 spin_unlock_irqrestore(&x->wait.lock, flags);
3652 EXPORT_SYMBOL(complete);
3654 void fastcall complete_all(struct completion *x)
3656 unsigned long flags;
3658 spin_lock_irqsave(&x->wait.lock, flags);
3659 x->done += UINT_MAX/2;
3660 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3662 spin_unlock_irqrestore(&x->wait.lock, flags);
3664 EXPORT_SYMBOL(complete_all);
3666 void fastcall __sched wait_for_completion(struct completion *x)
3670 spin_lock_irq(&x->wait.lock);
3672 DECLARE_WAITQUEUE(wait, current);
3674 wait.flags |= WQ_FLAG_EXCLUSIVE;
3675 __add_wait_queue_tail(&x->wait, &wait);
3677 __set_current_state(TASK_UNINTERRUPTIBLE);
3678 spin_unlock_irq(&x->wait.lock);
3680 spin_lock_irq(&x->wait.lock);
3682 __remove_wait_queue(&x->wait, &wait);
3685 spin_unlock_irq(&x->wait.lock);
3687 EXPORT_SYMBOL(wait_for_completion);
3689 unsigned long fastcall __sched
3690 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3694 spin_lock_irq(&x->wait.lock);
3696 DECLARE_WAITQUEUE(wait, current);
3698 wait.flags |= WQ_FLAG_EXCLUSIVE;
3699 __add_wait_queue_tail(&x->wait, &wait);
3701 __set_current_state(TASK_UNINTERRUPTIBLE);
3702 spin_unlock_irq(&x->wait.lock);
3703 timeout = schedule_timeout(timeout);
3704 spin_lock_irq(&x->wait.lock);
3706 __remove_wait_queue(&x->wait, &wait);
3710 __remove_wait_queue(&x->wait, &wait);
3714 spin_unlock_irq(&x->wait.lock);
3717 EXPORT_SYMBOL(wait_for_completion_timeout);
3719 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3725 spin_lock_irq(&x->wait.lock);
3727 DECLARE_WAITQUEUE(wait, current);
3729 wait.flags |= WQ_FLAG_EXCLUSIVE;
3730 __add_wait_queue_tail(&x->wait, &wait);
3732 if (signal_pending(current)) {
3734 __remove_wait_queue(&x->wait, &wait);
3737 __set_current_state(TASK_INTERRUPTIBLE);
3738 spin_unlock_irq(&x->wait.lock);
3740 spin_lock_irq(&x->wait.lock);
3742 __remove_wait_queue(&x->wait, &wait);
3746 spin_unlock_irq(&x->wait.lock);
3750 EXPORT_SYMBOL(wait_for_completion_interruptible);
3752 unsigned long fastcall __sched
3753 wait_for_completion_interruptible_timeout(struct completion *x,
3754 unsigned long timeout)
3758 spin_lock_irq(&x->wait.lock);
3760 DECLARE_WAITQUEUE(wait, current);
3762 wait.flags |= WQ_FLAG_EXCLUSIVE;
3763 __add_wait_queue_tail(&x->wait, &wait);
3765 if (signal_pending(current)) {
3766 timeout = -ERESTARTSYS;
3767 __remove_wait_queue(&x->wait, &wait);
3770 __set_current_state(TASK_INTERRUPTIBLE);
3771 spin_unlock_irq(&x->wait.lock);
3772 timeout = schedule_timeout(timeout);
3773 spin_lock_irq(&x->wait.lock);
3775 __remove_wait_queue(&x->wait, &wait);
3779 __remove_wait_queue(&x->wait, &wait);
3783 spin_unlock_irq(&x->wait.lock);
3786 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3789 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3791 spin_lock_irqsave(&q->lock, *flags);
3792 __add_wait_queue(q, wait);
3793 spin_unlock(&q->lock);
3797 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3799 spin_lock_irq(&q->lock);
3800 __remove_wait_queue(q, wait);
3801 spin_unlock_irqrestore(&q->lock, *flags);
3804 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3806 unsigned long flags;
3809 init_waitqueue_entry(&wait, current);
3811 current->state = TASK_INTERRUPTIBLE;
3813 sleep_on_head(q, &wait, &flags);
3815 sleep_on_tail(q, &wait, &flags);
3817 EXPORT_SYMBOL(interruptible_sleep_on);
3820 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3822 unsigned long flags;
3825 init_waitqueue_entry(&wait, current);
3827 current->state = TASK_INTERRUPTIBLE;
3829 sleep_on_head(q, &wait, &flags);
3830 timeout = schedule_timeout(timeout);
3831 sleep_on_tail(q, &wait, &flags);
3835 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3837 void __sched sleep_on(wait_queue_head_t *q)
3839 unsigned long flags;
3842 init_waitqueue_entry(&wait, current);
3844 current->state = TASK_UNINTERRUPTIBLE;
3846 sleep_on_head(q, &wait, &flags);
3848 sleep_on_tail(q, &wait, &flags);
3850 EXPORT_SYMBOL(sleep_on);
3852 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3854 unsigned long flags;
3857 init_waitqueue_entry(&wait, current);
3859 current->state = TASK_UNINTERRUPTIBLE;
3861 sleep_on_head(q, &wait, &flags);
3862 timeout = schedule_timeout(timeout);
3863 sleep_on_tail(q, &wait, &flags);
3867 EXPORT_SYMBOL(sleep_on_timeout);
3869 #ifdef CONFIG_RT_MUTEXES
3872 * rt_mutex_setprio - set the current priority of a task
3874 * @prio: prio value (kernel-internal form)
3876 * This function changes the 'effective' priority of a task. It does
3877 * not touch ->normal_prio like __setscheduler().
3879 * Used by the rt_mutex code to implement priority inheritance logic.
3881 void rt_mutex_setprio(struct task_struct *p, int prio)
3883 unsigned long flags;
3887 BUG_ON(prio < 0 || prio > MAX_PRIO);
3889 rq = task_rq_lock(p, &flags);
3890 update_rq_clock(rq);
3893 on_rq = p->se.on_rq;
3895 dequeue_task(rq, p, 0);
3898 p->sched_class = &rt_sched_class;
3900 p->sched_class = &fair_sched_class;
3905 enqueue_task(rq, p, 0);
3907 * Reschedule if we are currently running on this runqueue and
3908 * our priority decreased, or if we are not currently running on
3909 * this runqueue and our priority is higher than the current's
3911 if (task_running(rq, p)) {
3912 if (p->prio > oldprio)
3913 resched_task(rq->curr);
3915 check_preempt_curr(rq, p);
3918 task_rq_unlock(rq, &flags);
3923 void set_user_nice(struct task_struct *p, long nice)
3925 int old_prio, delta, on_rq;
3926 unsigned long flags;
3929 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3932 * We have to be careful, if called from sys_setpriority(),
3933 * the task might be in the middle of scheduling on another CPU.
3935 rq = task_rq_lock(p, &flags);
3936 update_rq_clock(rq);
3938 * The RT priorities are set via sched_setscheduler(), but we still
3939 * allow the 'normal' nice value to be set - but as expected
3940 * it wont have any effect on scheduling until the task is
3941 * SCHED_FIFO/SCHED_RR:
3943 if (task_has_rt_policy(p)) {
3944 p->static_prio = NICE_TO_PRIO(nice);
3947 on_rq = p->se.on_rq;
3949 dequeue_task(rq, p, 0);
3953 p->static_prio = NICE_TO_PRIO(nice);
3956 p->prio = effective_prio(p);
3957 delta = p->prio - old_prio;
3960 enqueue_task(rq, p, 0);
3963 * If the task increased its priority or is running and
3964 * lowered its priority, then reschedule its CPU:
3966 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3967 resched_task(rq->curr);
3970 task_rq_unlock(rq, &flags);
3972 EXPORT_SYMBOL(set_user_nice);
3975 * can_nice - check if a task can reduce its nice value
3979 int can_nice(const struct task_struct *p, const int nice)
3981 /* convert nice value [19,-20] to rlimit style value [1,40] */
3982 int nice_rlim = 20 - nice;
3984 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3985 capable(CAP_SYS_NICE));
3988 #ifdef __ARCH_WANT_SYS_NICE
3991 * sys_nice - change the priority of the current process.
3992 * @increment: priority increment
3994 * sys_setpriority is a more generic, but much slower function that
3995 * does similar things.
3997 asmlinkage long sys_nice(int increment)
4002 * Setpriority might change our priority at the same moment.
4003 * We don't have to worry. Conceptually one call occurs first
4004 * and we have a single winner.
4006 if (increment < -40)
4011 nice = PRIO_TO_NICE(current->static_prio) + increment;
4017 if (increment < 0 && !can_nice(current, nice))
4020 retval = security_task_setnice(current, nice);
4024 set_user_nice(current, nice);
4031 * task_prio - return the priority value of a given task.
4032 * @p: the task in question.
4034 * This is the priority value as seen by users in /proc.
4035 * RT tasks are offset by -200. Normal tasks are centered
4036 * around 0, value goes from -16 to +15.
4038 int task_prio(const struct task_struct *p)
4040 return p->prio - MAX_RT_PRIO;
4044 * task_nice - return the nice value of a given task.
4045 * @p: the task in question.
4047 int task_nice(const struct task_struct *p)
4049 return TASK_NICE(p);
4051 EXPORT_SYMBOL_GPL(task_nice);
4054 * idle_cpu - is a given cpu idle currently?
4055 * @cpu: the processor in question.
4057 int idle_cpu(int cpu)
4059 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4063 * idle_task - return the idle task for a given cpu.
4064 * @cpu: the processor in question.
4066 struct task_struct *idle_task(int cpu)
4068 return cpu_rq(cpu)->idle;
4072 * find_process_by_pid - find a process with a matching PID value.
4073 * @pid: the pid in question.
4075 static inline struct task_struct *find_process_by_pid(pid_t pid)
4077 return pid ? find_task_by_pid(pid) : current;
4080 /* Actually do priority change: must hold rq lock. */
4082 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4084 BUG_ON(p->se.on_rq);
4087 switch (p->policy) {
4091 p->sched_class = &fair_sched_class;
4095 p->sched_class = &rt_sched_class;
4099 p->rt_priority = prio;
4100 p->normal_prio = normal_prio(p);
4101 /* we are holding p->pi_lock already */
4102 p->prio = rt_mutex_getprio(p);
4107 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4108 * @p: the task in question.
4109 * @policy: new policy.
4110 * @param: structure containing the new RT priority.
4112 * NOTE that the task may be already dead.
4114 int sched_setscheduler(struct task_struct *p, int policy,
4115 struct sched_param *param)
4117 int retval, oldprio, oldpolicy = -1, on_rq;
4118 unsigned long flags;
4121 /* may grab non-irq protected spin_locks */
4122 BUG_ON(in_interrupt());
4124 /* double check policy once rq lock held */
4126 policy = oldpolicy = p->policy;
4127 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4128 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4129 policy != SCHED_IDLE)
4132 * Valid priorities for SCHED_FIFO and SCHED_RR are
4133 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4134 * SCHED_BATCH and SCHED_IDLE is 0.
4136 if (param->sched_priority < 0 ||
4137 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4138 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4140 if (rt_policy(policy) != (param->sched_priority != 0))
4144 * Allow unprivileged RT tasks to decrease priority:
4146 if (!capable(CAP_SYS_NICE)) {
4147 if (rt_policy(policy)) {
4148 unsigned long rlim_rtprio;
4150 if (!lock_task_sighand(p, &flags))
4152 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4153 unlock_task_sighand(p, &flags);
4155 /* can't set/change the rt policy */
4156 if (policy != p->policy && !rlim_rtprio)
4159 /* can't increase priority */
4160 if (param->sched_priority > p->rt_priority &&
4161 param->sched_priority > rlim_rtprio)
4165 * Like positive nice levels, dont allow tasks to
4166 * move out of SCHED_IDLE either:
4168 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4171 /* can't change other user's priorities */
4172 if ((current->euid != p->euid) &&
4173 (current->euid != p->uid))
4177 retval = security_task_setscheduler(p, policy, param);
4181 * make sure no PI-waiters arrive (or leave) while we are
4182 * changing the priority of the task:
4184 spin_lock_irqsave(&p->pi_lock, flags);
4186 * To be able to change p->policy safely, the apropriate
4187 * runqueue lock must be held.
4189 rq = __task_rq_lock(p);
4190 /* recheck policy now with rq lock held */
4191 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4192 policy = oldpolicy = -1;
4193 __task_rq_unlock(rq);
4194 spin_unlock_irqrestore(&p->pi_lock, flags);
4197 update_rq_clock(rq);
4198 on_rq = p->se.on_rq;
4200 deactivate_task(rq, p, 0);
4202 __setscheduler(rq, p, policy, param->sched_priority);
4204 activate_task(rq, p, 0);
4206 * Reschedule if we are currently running on this runqueue and
4207 * our priority decreased, or if we are not currently running on
4208 * this runqueue and our priority is higher than the current's
4210 if (task_running(rq, p)) {
4211 if (p->prio > oldprio)
4212 resched_task(rq->curr);
4214 check_preempt_curr(rq, p);
4217 __task_rq_unlock(rq);
4218 spin_unlock_irqrestore(&p->pi_lock, flags);
4220 rt_mutex_adjust_pi(p);
4224 EXPORT_SYMBOL_GPL(sched_setscheduler);
4227 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4229 struct sched_param lparam;
4230 struct task_struct *p;
4233 if (!param || pid < 0)
4235 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4240 p = find_process_by_pid(pid);
4242 retval = sched_setscheduler(p, policy, &lparam);
4249 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4250 * @pid: the pid in question.
4251 * @policy: new policy.
4252 * @param: structure containing the new RT priority.
4254 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4255 struct sched_param __user *param)
4257 /* negative values for policy are not valid */
4261 return do_sched_setscheduler(pid, policy, param);
4265 * sys_sched_setparam - set/change the RT priority of a thread
4266 * @pid: the pid in question.
4267 * @param: structure containing the new RT priority.
4269 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4271 return do_sched_setscheduler(pid, -1, param);
4275 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4276 * @pid: the pid in question.
4278 asmlinkage long sys_sched_getscheduler(pid_t pid)
4280 struct task_struct *p;
4281 int retval = -EINVAL;
4287 read_lock(&tasklist_lock);
4288 p = find_process_by_pid(pid);
4290 retval = security_task_getscheduler(p);
4294 read_unlock(&tasklist_lock);
4301 * sys_sched_getscheduler - get the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the RT priority.
4305 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4307 struct sched_param lp;
4308 struct task_struct *p;
4309 int retval = -EINVAL;
4311 if (!param || pid < 0)
4314 read_lock(&tasklist_lock);
4315 p = find_process_by_pid(pid);
4320 retval = security_task_getscheduler(p);
4324 lp.sched_priority = p->rt_priority;
4325 read_unlock(&tasklist_lock);
4328 * This one might sleep, we cannot do it with a spinlock held ...
4330 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4336 read_unlock(&tasklist_lock);
4340 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4342 cpumask_t cpus_allowed;
4343 struct task_struct *p;
4346 mutex_lock(&sched_hotcpu_mutex);
4347 read_lock(&tasklist_lock);
4349 p = find_process_by_pid(pid);
4351 read_unlock(&tasklist_lock);
4352 mutex_unlock(&sched_hotcpu_mutex);
4357 * It is not safe to call set_cpus_allowed with the
4358 * tasklist_lock held. We will bump the task_struct's
4359 * usage count and then drop tasklist_lock.
4362 read_unlock(&tasklist_lock);
4365 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4366 !capable(CAP_SYS_NICE))
4369 retval = security_task_setscheduler(p, 0, NULL);
4373 cpus_allowed = cpuset_cpus_allowed(p);
4374 cpus_and(new_mask, new_mask, cpus_allowed);
4375 retval = set_cpus_allowed(p, new_mask);
4379 mutex_unlock(&sched_hotcpu_mutex);
4383 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4384 cpumask_t *new_mask)
4386 if (len < sizeof(cpumask_t)) {
4387 memset(new_mask, 0, sizeof(cpumask_t));
4388 } else if (len > sizeof(cpumask_t)) {
4389 len = sizeof(cpumask_t);
4391 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4395 * sys_sched_setaffinity - set the cpu affinity of a process
4396 * @pid: pid of the process
4397 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4398 * @user_mask_ptr: user-space pointer to the new cpu mask
4400 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4401 unsigned long __user *user_mask_ptr)
4406 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4410 return sched_setaffinity(pid, new_mask);
4414 * Represents all cpu's present in the system
4415 * In systems capable of hotplug, this map could dynamically grow
4416 * as new cpu's are detected in the system via any platform specific
4417 * method, such as ACPI for e.g.
4420 cpumask_t cpu_present_map __read_mostly;
4421 EXPORT_SYMBOL(cpu_present_map);
4424 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4425 EXPORT_SYMBOL(cpu_online_map);
4427 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4428 EXPORT_SYMBOL(cpu_possible_map);
4431 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4433 struct task_struct *p;
4436 mutex_lock(&sched_hotcpu_mutex);
4437 read_lock(&tasklist_lock);
4440 p = find_process_by_pid(pid);
4444 retval = security_task_getscheduler(p);
4448 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4451 read_unlock(&tasklist_lock);
4452 mutex_unlock(&sched_hotcpu_mutex);
4458 * sys_sched_getaffinity - get the cpu affinity of a process
4459 * @pid: pid of the process
4460 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4461 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4463 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4464 unsigned long __user *user_mask_ptr)
4469 if (len < sizeof(cpumask_t))
4472 ret = sched_getaffinity(pid, &mask);
4476 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4479 return sizeof(cpumask_t);
4483 * sys_sched_yield - yield the current processor to other threads.
4485 * This function yields the current CPU to other tasks. If there are no
4486 * other threads running on this CPU then this function will return.
4488 asmlinkage long sys_sched_yield(void)
4490 struct rq *rq = this_rq_lock();
4492 schedstat_inc(rq, yld_cnt);
4493 current->sched_class->yield_task(rq, current);
4496 * Since we are going to call schedule() anyway, there's
4497 * no need to preempt or enable interrupts:
4499 __release(rq->lock);
4500 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4501 _raw_spin_unlock(&rq->lock);
4502 preempt_enable_no_resched();
4509 static void __cond_resched(void)
4511 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4512 __might_sleep(__FILE__, __LINE__);
4515 * The BKS might be reacquired before we have dropped
4516 * PREEMPT_ACTIVE, which could trigger a second
4517 * cond_resched() call.
4520 add_preempt_count(PREEMPT_ACTIVE);
4522 sub_preempt_count(PREEMPT_ACTIVE);
4523 } while (need_resched());
4526 int __sched cond_resched(void)
4528 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4529 system_state == SYSTEM_RUNNING) {
4535 EXPORT_SYMBOL(cond_resched);
4538 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4539 * call schedule, and on return reacquire the lock.
4541 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4542 * operations here to prevent schedule() from being called twice (once via
4543 * spin_unlock(), once by hand).
4545 int cond_resched_lock(spinlock_t *lock)
4549 if (need_lockbreak(lock)) {
4555 if (need_resched() && system_state == SYSTEM_RUNNING) {
4556 spin_release(&lock->dep_map, 1, _THIS_IP_);
4557 _raw_spin_unlock(lock);
4558 preempt_enable_no_resched();
4565 EXPORT_SYMBOL(cond_resched_lock);
4567 int __sched cond_resched_softirq(void)
4569 BUG_ON(!in_softirq());
4571 if (need_resched() && system_state == SYSTEM_RUNNING) {
4579 EXPORT_SYMBOL(cond_resched_softirq);
4582 * yield - yield the current processor to other threads.
4584 * This is a shortcut for kernel-space yielding - it marks the
4585 * thread runnable and calls sys_sched_yield().
4587 void __sched yield(void)
4589 set_current_state(TASK_RUNNING);
4592 EXPORT_SYMBOL(yield);
4595 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4596 * that process accounting knows that this is a task in IO wait state.
4598 * But don't do that if it is a deliberate, throttling IO wait (this task
4599 * has set its backing_dev_info: the queue against which it should throttle)
4601 void __sched io_schedule(void)
4603 struct rq *rq = &__raw_get_cpu_var(runqueues);
4605 delayacct_blkio_start();
4606 atomic_inc(&rq->nr_iowait);
4608 atomic_dec(&rq->nr_iowait);
4609 delayacct_blkio_end();
4611 EXPORT_SYMBOL(io_schedule);
4613 long __sched io_schedule_timeout(long timeout)
4615 struct rq *rq = &__raw_get_cpu_var(runqueues);
4618 delayacct_blkio_start();
4619 atomic_inc(&rq->nr_iowait);
4620 ret = schedule_timeout(timeout);
4621 atomic_dec(&rq->nr_iowait);
4622 delayacct_blkio_end();
4627 * sys_sched_get_priority_max - return maximum RT priority.
4628 * @policy: scheduling class.
4630 * this syscall returns the maximum rt_priority that can be used
4631 * by a given scheduling class.
4633 asmlinkage long sys_sched_get_priority_max(int policy)
4640 ret = MAX_USER_RT_PRIO-1;
4652 * sys_sched_get_priority_min - return minimum RT priority.
4653 * @policy: scheduling class.
4655 * this syscall returns the minimum rt_priority that can be used
4656 * by a given scheduling class.
4658 asmlinkage long sys_sched_get_priority_min(int policy)
4676 * sys_sched_rr_get_interval - return the default timeslice of a process.
4677 * @pid: pid of the process.
4678 * @interval: userspace pointer to the timeslice value.
4680 * this syscall writes the default timeslice value of a given process
4681 * into the user-space timespec buffer. A value of '0' means infinity.
4684 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4686 struct task_struct *p;
4687 int retval = -EINVAL;
4694 read_lock(&tasklist_lock);
4695 p = find_process_by_pid(pid);
4699 retval = security_task_getscheduler(p);
4703 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4704 0 : static_prio_timeslice(p->static_prio), &t);
4705 read_unlock(&tasklist_lock);
4706 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4710 read_unlock(&tasklist_lock);
4714 static const char stat_nam[] = "RSDTtZX";
4716 static void show_task(struct task_struct *p)
4718 unsigned long free = 0;
4721 state = p->state ? __ffs(p->state) + 1 : 0;
4722 printk("%-13.13s %c", p->comm,
4723 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4724 #if BITS_PER_LONG == 32
4725 if (state == TASK_RUNNING)
4726 printk(" running ");
4728 printk(" %08lx ", thread_saved_pc(p));
4730 if (state == TASK_RUNNING)
4731 printk(" running task ");
4733 printk(" %016lx ", thread_saved_pc(p));
4735 #ifdef CONFIG_DEBUG_STACK_USAGE
4737 unsigned long *n = end_of_stack(p);
4740 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4743 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4745 if (state != TASK_RUNNING)
4746 show_stack(p, NULL);
4749 void show_state_filter(unsigned long state_filter)
4751 struct task_struct *g, *p;
4753 #if BITS_PER_LONG == 32
4755 " task PC stack pid father\n");
4758 " task PC stack pid father\n");
4760 read_lock(&tasklist_lock);
4761 do_each_thread(g, p) {
4763 * reset the NMI-timeout, listing all files on a slow
4764 * console might take alot of time:
4766 touch_nmi_watchdog();
4767 if (!state_filter || (p->state & state_filter))
4769 } while_each_thread(g, p);
4771 touch_all_softlockup_watchdogs();
4773 #ifdef CONFIG_SCHED_DEBUG
4774 sysrq_sched_debug_show();
4776 read_unlock(&tasklist_lock);
4778 * Only show locks if all tasks are dumped:
4780 if (state_filter == -1)
4781 debug_show_all_locks();
4784 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4786 idle->sched_class = &idle_sched_class;
4790 * init_idle - set up an idle thread for a given CPU
4791 * @idle: task in question
4792 * @cpu: cpu the idle task belongs to
4794 * NOTE: this function does not set the idle thread's NEED_RESCHED
4795 * flag, to make booting more robust.
4797 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4799 struct rq *rq = cpu_rq(cpu);
4800 unsigned long flags;
4803 idle->se.exec_start = sched_clock();
4805 idle->prio = idle->normal_prio = MAX_PRIO;
4806 idle->cpus_allowed = cpumask_of_cpu(cpu);
4807 __set_task_cpu(idle, cpu);
4809 spin_lock_irqsave(&rq->lock, flags);
4810 rq->curr = rq->idle = idle;
4811 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4814 spin_unlock_irqrestore(&rq->lock, flags);
4816 /* Set the preempt count _outside_ the spinlocks! */
4817 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4818 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4820 task_thread_info(idle)->preempt_count = 0;
4823 * The idle tasks have their own, simple scheduling class:
4825 idle->sched_class = &idle_sched_class;
4829 * In a system that switches off the HZ timer nohz_cpu_mask
4830 * indicates which cpus entered this state. This is used
4831 * in the rcu update to wait only for active cpus. For system
4832 * which do not switch off the HZ timer nohz_cpu_mask should
4833 * always be CPU_MASK_NONE.
4835 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4839 * This is how migration works:
4841 * 1) we queue a struct migration_req structure in the source CPU's
4842 * runqueue and wake up that CPU's migration thread.
4843 * 2) we down() the locked semaphore => thread blocks.
4844 * 3) migration thread wakes up (implicitly it forces the migrated
4845 * thread off the CPU)
4846 * 4) it gets the migration request and checks whether the migrated
4847 * task is still in the wrong runqueue.
4848 * 5) if it's in the wrong runqueue then the migration thread removes
4849 * it and puts it into the right queue.
4850 * 6) migration thread up()s the semaphore.
4851 * 7) we wake up and the migration is done.
4855 * Change a given task's CPU affinity. Migrate the thread to a
4856 * proper CPU and schedule it away if the CPU it's executing on
4857 * is removed from the allowed bitmask.
4859 * NOTE: the caller must have a valid reference to the task, the
4860 * task must not exit() & deallocate itself prematurely. The
4861 * call is not atomic; no spinlocks may be held.
4863 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4865 struct migration_req req;
4866 unsigned long flags;
4870 rq = task_rq_lock(p, &flags);
4871 if (!cpus_intersects(new_mask, cpu_online_map)) {
4876 p->cpus_allowed = new_mask;
4877 /* Can the task run on the task's current CPU? If so, we're done */
4878 if (cpu_isset(task_cpu(p), new_mask))
4881 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4882 /* Need help from migration thread: drop lock and wait. */
4883 task_rq_unlock(rq, &flags);
4884 wake_up_process(rq->migration_thread);
4885 wait_for_completion(&req.done);
4886 tlb_migrate_finish(p->mm);
4890 task_rq_unlock(rq, &flags);
4894 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4897 * Move (not current) task off this cpu, onto dest cpu. We're doing
4898 * this because either it can't run here any more (set_cpus_allowed()
4899 * away from this CPU, or CPU going down), or because we're
4900 * attempting to rebalance this task on exec (sched_exec).
4902 * So we race with normal scheduler movements, but that's OK, as long
4903 * as the task is no longer on this CPU.
4905 * Returns non-zero if task was successfully migrated.
4907 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4909 struct rq *rq_dest, *rq_src;
4912 if (unlikely(cpu_is_offline(dest_cpu)))
4915 rq_src = cpu_rq(src_cpu);
4916 rq_dest = cpu_rq(dest_cpu);
4918 double_rq_lock(rq_src, rq_dest);
4919 /* Already moved. */
4920 if (task_cpu(p) != src_cpu)
4922 /* Affinity changed (again). */
4923 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4926 on_rq = p->se.on_rq;
4928 deactivate_task(rq_src, p, 0);
4930 set_task_cpu(p, dest_cpu);
4932 activate_task(rq_dest, p, 0);
4933 check_preempt_curr(rq_dest, p);
4937 double_rq_unlock(rq_src, rq_dest);
4942 * migration_thread - this is a highprio system thread that performs
4943 * thread migration by bumping thread off CPU then 'pushing' onto
4946 static int migration_thread(void *data)
4948 int cpu = (long)data;
4952 BUG_ON(rq->migration_thread != current);
4954 set_current_state(TASK_INTERRUPTIBLE);
4955 while (!kthread_should_stop()) {
4956 struct migration_req *req;
4957 struct list_head *head;
4959 spin_lock_irq(&rq->lock);
4961 if (cpu_is_offline(cpu)) {
4962 spin_unlock_irq(&rq->lock);
4966 if (rq->active_balance) {
4967 active_load_balance(rq, cpu);
4968 rq->active_balance = 0;
4971 head = &rq->migration_queue;
4973 if (list_empty(head)) {
4974 spin_unlock_irq(&rq->lock);
4976 set_current_state(TASK_INTERRUPTIBLE);
4979 req = list_entry(head->next, struct migration_req, list);
4980 list_del_init(head->next);
4982 spin_unlock(&rq->lock);
4983 __migrate_task(req->task, cpu, req->dest_cpu);
4986 complete(&req->done);
4988 __set_current_state(TASK_RUNNING);
4992 /* Wait for kthread_stop */
4993 set_current_state(TASK_INTERRUPTIBLE);
4994 while (!kthread_should_stop()) {
4996 set_current_state(TASK_INTERRUPTIBLE);
4998 __set_current_state(TASK_RUNNING);
5002 #ifdef CONFIG_HOTPLUG_CPU
5004 * Figure out where task on dead CPU should go, use force if neccessary.
5005 * NOTE: interrupts should be disabled by the caller
5007 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5009 unsigned long flags;
5016 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5017 cpus_and(mask, mask, p->cpus_allowed);
5018 dest_cpu = any_online_cpu(mask);
5020 /* On any allowed CPU? */
5021 if (dest_cpu == NR_CPUS)
5022 dest_cpu = any_online_cpu(p->cpus_allowed);
5024 /* No more Mr. Nice Guy. */
5025 if (dest_cpu == NR_CPUS) {
5026 rq = task_rq_lock(p, &flags);
5027 cpus_setall(p->cpus_allowed);
5028 dest_cpu = any_online_cpu(p->cpus_allowed);
5029 task_rq_unlock(rq, &flags);
5032 * Don't tell them about moving exiting tasks or
5033 * kernel threads (both mm NULL), since they never
5036 if (p->mm && printk_ratelimit())
5037 printk(KERN_INFO "process %d (%s) no "
5038 "longer affine to cpu%d\n",
5039 p->pid, p->comm, dead_cpu);
5041 if (!__migrate_task(p, dead_cpu, dest_cpu))
5046 * While a dead CPU has no uninterruptible tasks queued at this point,
5047 * it might still have a nonzero ->nr_uninterruptible counter, because
5048 * for performance reasons the counter is not stricly tracking tasks to
5049 * their home CPUs. So we just add the counter to another CPU's counter,
5050 * to keep the global sum constant after CPU-down:
5052 static void migrate_nr_uninterruptible(struct rq *rq_src)
5054 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5055 unsigned long flags;
5057 local_irq_save(flags);
5058 double_rq_lock(rq_src, rq_dest);
5059 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5060 rq_src->nr_uninterruptible = 0;
5061 double_rq_unlock(rq_src, rq_dest);
5062 local_irq_restore(flags);
5065 /* Run through task list and migrate tasks from the dead cpu. */
5066 static void migrate_live_tasks(int src_cpu)
5068 struct task_struct *p, *t;
5070 write_lock_irq(&tasklist_lock);
5072 do_each_thread(t, p) {
5076 if (task_cpu(p) == src_cpu)
5077 move_task_off_dead_cpu(src_cpu, p);
5078 } while_each_thread(t, p);
5080 write_unlock_irq(&tasklist_lock);
5084 * Schedules idle task to be the next runnable task on current CPU.
5085 * It does so by boosting its priority to highest possible and adding it to
5086 * the _front_ of the runqueue. Used by CPU offline code.
5088 void sched_idle_next(void)
5090 int this_cpu = smp_processor_id();
5091 struct rq *rq = cpu_rq(this_cpu);
5092 struct task_struct *p = rq->idle;
5093 unsigned long flags;
5095 /* cpu has to be offline */
5096 BUG_ON(cpu_online(this_cpu));
5099 * Strictly not necessary since rest of the CPUs are stopped by now
5100 * and interrupts disabled on the current cpu.
5102 spin_lock_irqsave(&rq->lock, flags);
5104 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5106 /* Add idle task to the _front_ of its priority queue: */
5107 activate_idle_task(p, rq);
5109 spin_unlock_irqrestore(&rq->lock, flags);
5113 * Ensures that the idle task is using init_mm right before its cpu goes
5116 void idle_task_exit(void)
5118 struct mm_struct *mm = current->active_mm;
5120 BUG_ON(cpu_online(smp_processor_id()));
5123 switch_mm(mm, &init_mm, current);
5127 /* called under rq->lock with disabled interrupts */
5128 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5130 struct rq *rq = cpu_rq(dead_cpu);
5132 /* Must be exiting, otherwise would be on tasklist. */
5133 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5135 /* Cannot have done final schedule yet: would have vanished. */
5136 BUG_ON(p->state == TASK_DEAD);
5141 * Drop lock around migration; if someone else moves it,
5142 * that's OK. No task can be added to this CPU, so iteration is
5144 * NOTE: interrupts should be left disabled --dev@
5146 spin_unlock(&rq->lock);
5147 move_task_off_dead_cpu(dead_cpu, p);
5148 spin_lock(&rq->lock);
5153 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5154 static void migrate_dead_tasks(unsigned int dead_cpu)
5156 struct rq *rq = cpu_rq(dead_cpu);
5157 struct task_struct *next;
5160 if (!rq->nr_running)
5162 update_rq_clock(rq);
5163 next = pick_next_task(rq, rq->curr);
5166 migrate_dead(dead_cpu, next);
5170 #endif /* CONFIG_HOTPLUG_CPU */
5172 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5174 static struct ctl_table sd_ctl_dir[] = {
5176 .procname = "sched_domain",
5182 static struct ctl_table sd_ctl_root[] = {
5184 .ctl_name = CTL_KERN,
5185 .procname = "kernel",
5187 .child = sd_ctl_dir,
5192 static struct ctl_table *sd_alloc_ctl_entry(int n)
5194 struct ctl_table *entry =
5195 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5198 memset(entry, 0, n * sizeof(struct ctl_table));
5204 set_table_entry(struct ctl_table *entry,
5205 const char *procname, void *data, int maxlen,
5206 mode_t mode, proc_handler *proc_handler)
5208 entry->procname = procname;
5210 entry->maxlen = maxlen;
5212 entry->proc_handler = proc_handler;
5215 static struct ctl_table *
5216 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5218 struct ctl_table *table = sd_alloc_ctl_entry(14);
5220 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5221 sizeof(long), 0644, proc_doulongvec_minmax);
5222 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5223 sizeof(long), 0644, proc_doulongvec_minmax);
5224 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5225 sizeof(int), 0644, proc_dointvec_minmax);
5226 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5227 sizeof(int), 0644, proc_dointvec_minmax);
5228 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5229 sizeof(int), 0644, proc_dointvec_minmax);
5230 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5231 sizeof(int), 0644, proc_dointvec_minmax);
5232 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5233 sizeof(int), 0644, proc_dointvec_minmax);
5234 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5235 sizeof(int), 0644, proc_dointvec_minmax);
5236 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5237 sizeof(int), 0644, proc_dointvec_minmax);
5238 set_table_entry(&table[10], "cache_nice_tries",
5239 &sd->cache_nice_tries,
5240 sizeof(int), 0644, proc_dointvec_minmax);
5241 set_table_entry(&table[12], "flags", &sd->flags,
5242 sizeof(int), 0644, proc_dointvec_minmax);
5247 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5249 struct ctl_table *entry, *table;
5250 struct sched_domain *sd;
5251 int domain_num = 0, i;
5254 for_each_domain(cpu, sd)
5256 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5259 for_each_domain(cpu, sd) {
5260 snprintf(buf, 32, "domain%d", i);
5261 entry->procname = kstrdup(buf, GFP_KERNEL);
5263 entry->child = sd_alloc_ctl_domain_table(sd);
5270 static struct ctl_table_header *sd_sysctl_header;
5271 static void init_sched_domain_sysctl(void)
5273 int i, cpu_num = num_online_cpus();
5274 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5277 sd_ctl_dir[0].child = entry;
5279 for (i = 0; i < cpu_num; i++, entry++) {
5280 snprintf(buf, 32, "cpu%d", i);
5281 entry->procname = kstrdup(buf, GFP_KERNEL);
5283 entry->child = sd_alloc_ctl_cpu_table(i);
5285 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5288 static void init_sched_domain_sysctl(void)
5294 * migration_call - callback that gets triggered when a CPU is added.
5295 * Here we can start up the necessary migration thread for the new CPU.
5297 static int __cpuinit
5298 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5300 struct task_struct *p;
5301 int cpu = (long)hcpu;
5302 unsigned long flags;
5306 case CPU_LOCK_ACQUIRE:
5307 mutex_lock(&sched_hotcpu_mutex);
5310 case CPU_UP_PREPARE:
5311 case CPU_UP_PREPARE_FROZEN:
5312 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5315 kthread_bind(p, cpu);
5316 /* Must be high prio: stop_machine expects to yield to it. */
5317 rq = task_rq_lock(p, &flags);
5318 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5319 task_rq_unlock(rq, &flags);
5320 cpu_rq(cpu)->migration_thread = p;
5324 case CPU_ONLINE_FROZEN:
5325 /* Strictly unneccessary, as first user will wake it. */
5326 wake_up_process(cpu_rq(cpu)->migration_thread);
5329 #ifdef CONFIG_HOTPLUG_CPU
5330 case CPU_UP_CANCELED:
5331 case CPU_UP_CANCELED_FROZEN:
5332 if (!cpu_rq(cpu)->migration_thread)
5334 /* Unbind it from offline cpu so it can run. Fall thru. */
5335 kthread_bind(cpu_rq(cpu)->migration_thread,
5336 any_online_cpu(cpu_online_map));
5337 kthread_stop(cpu_rq(cpu)->migration_thread);
5338 cpu_rq(cpu)->migration_thread = NULL;
5342 case CPU_DEAD_FROZEN:
5343 migrate_live_tasks(cpu);
5345 kthread_stop(rq->migration_thread);
5346 rq->migration_thread = NULL;
5347 /* Idle task back to normal (off runqueue, low prio) */
5348 rq = task_rq_lock(rq->idle, &flags);
5349 update_rq_clock(rq);
5350 deactivate_task(rq, rq->idle, 0);
5351 rq->idle->static_prio = MAX_PRIO;
5352 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5353 rq->idle->sched_class = &idle_sched_class;
5354 migrate_dead_tasks(cpu);
5355 task_rq_unlock(rq, &flags);
5356 migrate_nr_uninterruptible(rq);
5357 BUG_ON(rq->nr_running != 0);
5359 /* No need to migrate the tasks: it was best-effort if
5360 * they didn't take sched_hotcpu_mutex. Just wake up
5361 * the requestors. */
5362 spin_lock_irq(&rq->lock);
5363 while (!list_empty(&rq->migration_queue)) {
5364 struct migration_req *req;
5366 req = list_entry(rq->migration_queue.next,
5367 struct migration_req, list);
5368 list_del_init(&req->list);
5369 complete(&req->done);
5371 spin_unlock_irq(&rq->lock);
5374 case CPU_LOCK_RELEASE:
5375 mutex_unlock(&sched_hotcpu_mutex);
5381 /* Register at highest priority so that task migration (migrate_all_tasks)
5382 * happens before everything else.
5384 static struct notifier_block __cpuinitdata migration_notifier = {
5385 .notifier_call = migration_call,
5389 int __init migration_init(void)
5391 void *cpu = (void *)(long)smp_processor_id();
5394 /* Start one for the boot CPU: */
5395 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5396 BUG_ON(err == NOTIFY_BAD);
5397 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5398 register_cpu_notifier(&migration_notifier);
5406 /* Number of possible processor ids */
5407 int nr_cpu_ids __read_mostly = NR_CPUS;
5408 EXPORT_SYMBOL(nr_cpu_ids);
5410 #undef SCHED_DOMAIN_DEBUG
5411 #ifdef SCHED_DOMAIN_DEBUG
5412 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5417 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5421 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5426 struct sched_group *group = sd->groups;
5427 cpumask_t groupmask;
5429 cpumask_scnprintf(str, NR_CPUS, sd->span);
5430 cpus_clear(groupmask);
5433 for (i = 0; i < level + 1; i++)
5435 printk("domain %d: ", level);
5437 if (!(sd->flags & SD_LOAD_BALANCE)) {
5438 printk("does not load-balance\n");
5440 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5445 printk("span %s\n", str);
5447 if (!cpu_isset(cpu, sd->span))
5448 printk(KERN_ERR "ERROR: domain->span does not contain "
5450 if (!cpu_isset(cpu, group->cpumask))
5451 printk(KERN_ERR "ERROR: domain->groups does not contain"
5455 for (i = 0; i < level + 2; i++)
5461 printk(KERN_ERR "ERROR: group is NULL\n");
5465 if (!group->__cpu_power) {
5467 printk(KERN_ERR "ERROR: domain->cpu_power not "
5471 if (!cpus_weight(group->cpumask)) {
5473 printk(KERN_ERR "ERROR: empty group\n");
5476 if (cpus_intersects(groupmask, group->cpumask)) {
5478 printk(KERN_ERR "ERROR: repeated CPUs\n");
5481 cpus_or(groupmask, groupmask, group->cpumask);
5483 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5486 group = group->next;
5487 } while (group != sd->groups);
5490 if (!cpus_equal(sd->span, groupmask))
5491 printk(KERN_ERR "ERROR: groups don't span "
5499 if (!cpus_subset(groupmask, sd->span))
5500 printk(KERN_ERR "ERROR: parent span is not a superset "
5501 "of domain->span\n");
5506 # define sched_domain_debug(sd, cpu) do { } while (0)
5509 static int sd_degenerate(struct sched_domain *sd)
5511 if (cpus_weight(sd->span) == 1)
5514 /* Following flags need at least 2 groups */
5515 if (sd->flags & (SD_LOAD_BALANCE |
5516 SD_BALANCE_NEWIDLE |
5520 SD_SHARE_PKG_RESOURCES)) {
5521 if (sd->groups != sd->groups->next)
5525 /* Following flags don't use groups */
5526 if (sd->flags & (SD_WAKE_IDLE |
5535 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5537 unsigned long cflags = sd->flags, pflags = parent->flags;
5539 if (sd_degenerate(parent))
5542 if (!cpus_equal(sd->span, parent->span))
5545 /* Does parent contain flags not in child? */
5546 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5547 if (cflags & SD_WAKE_AFFINE)
5548 pflags &= ~SD_WAKE_BALANCE;
5549 /* Flags needing groups don't count if only 1 group in parent */
5550 if (parent->groups == parent->groups->next) {
5551 pflags &= ~(SD_LOAD_BALANCE |
5552 SD_BALANCE_NEWIDLE |
5556 SD_SHARE_PKG_RESOURCES);
5558 if (~cflags & pflags)
5565 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5566 * hold the hotplug lock.
5568 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5570 struct rq *rq = cpu_rq(cpu);
5571 struct sched_domain *tmp;
5573 /* Remove the sched domains which do not contribute to scheduling. */
5574 for (tmp = sd; tmp; tmp = tmp->parent) {
5575 struct sched_domain *parent = tmp->parent;
5578 if (sd_parent_degenerate(tmp, parent)) {
5579 tmp->parent = parent->parent;
5581 parent->parent->child = tmp;
5585 if (sd && sd_degenerate(sd)) {
5591 sched_domain_debug(sd, cpu);
5593 rcu_assign_pointer(rq->sd, sd);
5596 /* cpus with isolated domains */
5597 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5599 /* Setup the mask of cpus configured for isolated domains */
5600 static int __init isolated_cpu_setup(char *str)
5602 int ints[NR_CPUS], i;
5604 str = get_options(str, ARRAY_SIZE(ints), ints);
5605 cpus_clear(cpu_isolated_map);
5606 for (i = 1; i <= ints[0]; i++)
5607 if (ints[i] < NR_CPUS)
5608 cpu_set(ints[i], cpu_isolated_map);
5612 __setup ("isolcpus=", isolated_cpu_setup);
5615 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5616 * to a function which identifies what group(along with sched group) a CPU
5617 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5618 * (due to the fact that we keep track of groups covered with a cpumask_t).
5620 * init_sched_build_groups will build a circular linked list of the groups
5621 * covered by the given span, and will set each group's ->cpumask correctly,
5622 * and ->cpu_power to 0.
5625 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5626 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5627 struct sched_group **sg))
5629 struct sched_group *first = NULL, *last = NULL;
5630 cpumask_t covered = CPU_MASK_NONE;
5633 for_each_cpu_mask(i, span) {
5634 struct sched_group *sg;
5635 int group = group_fn(i, cpu_map, &sg);
5638 if (cpu_isset(i, covered))
5641 sg->cpumask = CPU_MASK_NONE;
5642 sg->__cpu_power = 0;
5644 for_each_cpu_mask(j, span) {
5645 if (group_fn(j, cpu_map, NULL) != group)
5648 cpu_set(j, covered);
5649 cpu_set(j, sg->cpumask);
5660 #define SD_NODES_PER_DOMAIN 16
5665 * find_next_best_node - find the next node to include in a sched_domain
5666 * @node: node whose sched_domain we're building
5667 * @used_nodes: nodes already in the sched_domain
5669 * Find the next node to include in a given scheduling domain. Simply
5670 * finds the closest node not already in the @used_nodes map.
5672 * Should use nodemask_t.
5674 static int find_next_best_node(int node, unsigned long *used_nodes)
5676 int i, n, val, min_val, best_node = 0;
5680 for (i = 0; i < MAX_NUMNODES; i++) {
5681 /* Start at @node */
5682 n = (node + i) % MAX_NUMNODES;
5684 if (!nr_cpus_node(n))
5687 /* Skip already used nodes */
5688 if (test_bit(n, used_nodes))
5691 /* Simple min distance search */
5692 val = node_distance(node, n);
5694 if (val < min_val) {
5700 set_bit(best_node, used_nodes);
5705 * sched_domain_node_span - get a cpumask for a node's sched_domain
5706 * @node: node whose cpumask we're constructing
5707 * @size: number of nodes to include in this span
5709 * Given a node, construct a good cpumask for its sched_domain to span. It
5710 * should be one that prevents unnecessary balancing, but also spreads tasks
5713 static cpumask_t sched_domain_node_span(int node)
5715 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5716 cpumask_t span, nodemask;
5720 bitmap_zero(used_nodes, MAX_NUMNODES);
5722 nodemask = node_to_cpumask(node);
5723 cpus_or(span, span, nodemask);
5724 set_bit(node, used_nodes);
5726 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5727 int next_node = find_next_best_node(node, used_nodes);
5729 nodemask = node_to_cpumask(next_node);
5730 cpus_or(span, span, nodemask);
5737 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5740 * SMT sched-domains:
5742 #ifdef CONFIG_SCHED_SMT
5743 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5744 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5746 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5747 struct sched_group **sg)
5750 *sg = &per_cpu(sched_group_cpus, cpu);
5756 * multi-core sched-domains:
5758 #ifdef CONFIG_SCHED_MC
5759 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5760 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5763 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5764 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5765 struct sched_group **sg)
5768 cpumask_t mask = cpu_sibling_map[cpu];
5769 cpus_and(mask, mask, *cpu_map);
5770 group = first_cpu(mask);
5772 *sg = &per_cpu(sched_group_core, group);
5775 #elif defined(CONFIG_SCHED_MC)
5776 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5777 struct sched_group **sg)
5780 *sg = &per_cpu(sched_group_core, cpu);
5785 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5786 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5788 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5789 struct sched_group **sg)
5792 #ifdef CONFIG_SCHED_MC
5793 cpumask_t mask = cpu_coregroup_map(cpu);
5794 cpus_and(mask, mask, *cpu_map);
5795 group = first_cpu(mask);
5796 #elif defined(CONFIG_SCHED_SMT)
5797 cpumask_t mask = cpu_sibling_map[cpu];
5798 cpus_and(mask, mask, *cpu_map);
5799 group = first_cpu(mask);
5804 *sg = &per_cpu(sched_group_phys, group);
5810 * The init_sched_build_groups can't handle what we want to do with node
5811 * groups, so roll our own. Now each node has its own list of groups which
5812 * gets dynamically allocated.
5814 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5815 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5817 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5818 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5820 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5821 struct sched_group **sg)
5823 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5826 cpus_and(nodemask, nodemask, *cpu_map);
5827 group = first_cpu(nodemask);
5830 *sg = &per_cpu(sched_group_allnodes, group);
5834 static void init_numa_sched_groups_power(struct sched_group *group_head)
5836 struct sched_group *sg = group_head;
5842 for_each_cpu_mask(j, sg->cpumask) {
5843 struct sched_domain *sd;
5845 sd = &per_cpu(phys_domains, j);
5846 if (j != first_cpu(sd->groups->cpumask)) {
5848 * Only add "power" once for each
5854 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5857 if (sg != group_head)
5863 /* Free memory allocated for various sched_group structures */
5864 static void free_sched_groups(const cpumask_t *cpu_map)
5868 for_each_cpu_mask(cpu, *cpu_map) {
5869 struct sched_group **sched_group_nodes
5870 = sched_group_nodes_bycpu[cpu];
5872 if (!sched_group_nodes)
5875 for (i = 0; i < MAX_NUMNODES; i++) {
5876 cpumask_t nodemask = node_to_cpumask(i);
5877 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5879 cpus_and(nodemask, nodemask, *cpu_map);
5880 if (cpus_empty(nodemask))
5890 if (oldsg != sched_group_nodes[i])
5893 kfree(sched_group_nodes);
5894 sched_group_nodes_bycpu[cpu] = NULL;
5898 static void free_sched_groups(const cpumask_t *cpu_map)
5904 * Initialize sched groups cpu_power.
5906 * cpu_power indicates the capacity of sched group, which is used while
5907 * distributing the load between different sched groups in a sched domain.
5908 * Typically cpu_power for all the groups in a sched domain will be same unless
5909 * there are asymmetries in the topology. If there are asymmetries, group
5910 * having more cpu_power will pickup more load compared to the group having
5913 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5914 * the maximum number of tasks a group can handle in the presence of other idle
5915 * or lightly loaded groups in the same sched domain.
5917 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5919 struct sched_domain *child;
5920 struct sched_group *group;
5922 WARN_ON(!sd || !sd->groups);
5924 if (cpu != first_cpu(sd->groups->cpumask))
5929 sd->groups->__cpu_power = 0;
5932 * For perf policy, if the groups in child domain share resources
5933 * (for example cores sharing some portions of the cache hierarchy
5934 * or SMT), then set this domain groups cpu_power such that each group
5935 * can handle only one task, when there are other idle groups in the
5936 * same sched domain.
5938 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5940 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5941 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5946 * add cpu_power of each child group to this groups cpu_power
5948 group = child->groups;
5950 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5951 group = group->next;
5952 } while (group != child->groups);
5956 * Build sched domains for a given set of cpus and attach the sched domains
5957 * to the individual cpus
5959 static int build_sched_domains(const cpumask_t *cpu_map)
5963 struct sched_group **sched_group_nodes = NULL;
5964 int sd_allnodes = 0;
5967 * Allocate the per-node list of sched groups
5969 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
5971 if (!sched_group_nodes) {
5972 printk(KERN_WARNING "Can not alloc sched group node list\n");
5975 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5979 * Set up domains for cpus specified by the cpu_map.
5981 for_each_cpu_mask(i, *cpu_map) {
5982 struct sched_domain *sd = NULL, *p;
5983 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5985 cpus_and(nodemask, nodemask, *cpu_map);
5988 if (cpus_weight(*cpu_map) >
5989 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5990 sd = &per_cpu(allnodes_domains, i);
5991 *sd = SD_ALLNODES_INIT;
5992 sd->span = *cpu_map;
5993 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
5999 sd = &per_cpu(node_domains, i);
6001 sd->span = sched_domain_node_span(cpu_to_node(i));
6005 cpus_and(sd->span, sd->span, *cpu_map);
6009 sd = &per_cpu(phys_domains, i);
6011 sd->span = nodemask;
6015 cpu_to_phys_group(i, cpu_map, &sd->groups);
6017 #ifdef CONFIG_SCHED_MC
6019 sd = &per_cpu(core_domains, i);
6021 sd->span = cpu_coregroup_map(i);
6022 cpus_and(sd->span, sd->span, *cpu_map);
6025 cpu_to_core_group(i, cpu_map, &sd->groups);
6028 #ifdef CONFIG_SCHED_SMT
6030 sd = &per_cpu(cpu_domains, i);
6031 *sd = SD_SIBLING_INIT;
6032 sd->span = cpu_sibling_map[i];
6033 cpus_and(sd->span, sd->span, *cpu_map);
6036 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6040 #ifdef CONFIG_SCHED_SMT
6041 /* Set up CPU (sibling) groups */
6042 for_each_cpu_mask(i, *cpu_map) {
6043 cpumask_t this_sibling_map = cpu_sibling_map[i];
6044 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6045 if (i != first_cpu(this_sibling_map))
6048 init_sched_build_groups(this_sibling_map, cpu_map,
6053 #ifdef CONFIG_SCHED_MC
6054 /* Set up multi-core groups */
6055 for_each_cpu_mask(i, *cpu_map) {
6056 cpumask_t this_core_map = cpu_coregroup_map(i);
6057 cpus_and(this_core_map, this_core_map, *cpu_map);
6058 if (i != first_cpu(this_core_map))
6060 init_sched_build_groups(this_core_map, cpu_map,
6061 &cpu_to_core_group);
6065 /* Set up physical groups */
6066 for (i = 0; i < MAX_NUMNODES; i++) {
6067 cpumask_t nodemask = node_to_cpumask(i);
6069 cpus_and(nodemask, nodemask, *cpu_map);
6070 if (cpus_empty(nodemask))
6073 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6077 /* Set up node groups */
6079 init_sched_build_groups(*cpu_map, cpu_map,
6080 &cpu_to_allnodes_group);
6082 for (i = 0; i < MAX_NUMNODES; i++) {
6083 /* Set up node groups */
6084 struct sched_group *sg, *prev;
6085 cpumask_t nodemask = node_to_cpumask(i);
6086 cpumask_t domainspan;
6087 cpumask_t covered = CPU_MASK_NONE;
6090 cpus_and(nodemask, nodemask, *cpu_map);
6091 if (cpus_empty(nodemask)) {
6092 sched_group_nodes[i] = NULL;
6096 domainspan = sched_domain_node_span(i);
6097 cpus_and(domainspan, domainspan, *cpu_map);
6099 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6101 printk(KERN_WARNING "Can not alloc domain group for "
6105 sched_group_nodes[i] = sg;
6106 for_each_cpu_mask(j, nodemask) {
6107 struct sched_domain *sd;
6109 sd = &per_cpu(node_domains, j);
6112 sg->__cpu_power = 0;
6113 sg->cpumask = nodemask;
6115 cpus_or(covered, covered, nodemask);
6118 for (j = 0; j < MAX_NUMNODES; j++) {
6119 cpumask_t tmp, notcovered;
6120 int n = (i + j) % MAX_NUMNODES;
6122 cpus_complement(notcovered, covered);
6123 cpus_and(tmp, notcovered, *cpu_map);
6124 cpus_and(tmp, tmp, domainspan);
6125 if (cpus_empty(tmp))
6128 nodemask = node_to_cpumask(n);
6129 cpus_and(tmp, tmp, nodemask);
6130 if (cpus_empty(tmp))
6133 sg = kmalloc_node(sizeof(struct sched_group),
6137 "Can not alloc domain group for node %d\n", j);
6140 sg->__cpu_power = 0;
6142 sg->next = prev->next;
6143 cpus_or(covered, covered, tmp);
6150 /* Calculate CPU power for physical packages and nodes */
6151 #ifdef CONFIG_SCHED_SMT
6152 for_each_cpu_mask(i, *cpu_map) {
6153 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6155 init_sched_groups_power(i, sd);
6158 #ifdef CONFIG_SCHED_MC
6159 for_each_cpu_mask(i, *cpu_map) {
6160 struct sched_domain *sd = &per_cpu(core_domains, i);
6162 init_sched_groups_power(i, sd);
6166 for_each_cpu_mask(i, *cpu_map) {
6167 struct sched_domain *sd = &per_cpu(phys_domains, i);
6169 init_sched_groups_power(i, sd);
6173 for (i = 0; i < MAX_NUMNODES; i++)
6174 init_numa_sched_groups_power(sched_group_nodes[i]);
6177 struct sched_group *sg;
6179 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6180 init_numa_sched_groups_power(sg);
6184 /* Attach the domains */
6185 for_each_cpu_mask(i, *cpu_map) {
6186 struct sched_domain *sd;
6187 #ifdef CONFIG_SCHED_SMT
6188 sd = &per_cpu(cpu_domains, i);
6189 #elif defined(CONFIG_SCHED_MC)
6190 sd = &per_cpu(core_domains, i);
6192 sd = &per_cpu(phys_domains, i);
6194 cpu_attach_domain(sd, i);
6201 free_sched_groups(cpu_map);
6206 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6208 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6210 cpumask_t cpu_default_map;
6214 * Setup mask for cpus without special case scheduling requirements.
6215 * For now this just excludes isolated cpus, but could be used to
6216 * exclude other special cases in the future.
6218 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6220 err = build_sched_domains(&cpu_default_map);
6225 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6227 free_sched_groups(cpu_map);
6231 * Detach sched domains from a group of cpus specified in cpu_map
6232 * These cpus will now be attached to the NULL domain
6234 static void detach_destroy_domains(const cpumask_t *cpu_map)
6238 for_each_cpu_mask(i, *cpu_map)
6239 cpu_attach_domain(NULL, i);
6240 synchronize_sched();
6241 arch_destroy_sched_domains(cpu_map);
6245 * Partition sched domains as specified by the cpumasks below.
6246 * This attaches all cpus from the cpumasks to the NULL domain,
6247 * waits for a RCU quiescent period, recalculates sched
6248 * domain information and then attaches them back to the
6249 * correct sched domains
6250 * Call with hotplug lock held
6252 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6254 cpumask_t change_map;
6257 cpus_and(*partition1, *partition1, cpu_online_map);
6258 cpus_and(*partition2, *partition2, cpu_online_map);
6259 cpus_or(change_map, *partition1, *partition2);
6261 /* Detach sched domains from all of the affected cpus */
6262 detach_destroy_domains(&change_map);
6263 if (!cpus_empty(*partition1))
6264 err = build_sched_domains(partition1);
6265 if (!err && !cpus_empty(*partition2))
6266 err = build_sched_domains(partition2);
6271 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6272 static int arch_reinit_sched_domains(void)
6276 mutex_lock(&sched_hotcpu_mutex);
6277 detach_destroy_domains(&cpu_online_map);
6278 err = arch_init_sched_domains(&cpu_online_map);
6279 mutex_unlock(&sched_hotcpu_mutex);
6284 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6288 if (buf[0] != '0' && buf[0] != '1')
6292 sched_smt_power_savings = (buf[0] == '1');
6294 sched_mc_power_savings = (buf[0] == '1');
6296 ret = arch_reinit_sched_domains();
6298 return ret ? ret : count;
6301 #ifdef CONFIG_SCHED_MC
6302 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6304 return sprintf(page, "%u\n", sched_mc_power_savings);
6306 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6307 const char *buf, size_t count)
6309 return sched_power_savings_store(buf, count, 0);
6311 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6312 sched_mc_power_savings_store);
6315 #ifdef CONFIG_SCHED_SMT
6316 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6318 return sprintf(page, "%u\n", sched_smt_power_savings);
6320 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6321 const char *buf, size_t count)
6323 return sched_power_savings_store(buf, count, 1);
6325 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6326 sched_smt_power_savings_store);
6329 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6333 #ifdef CONFIG_SCHED_SMT
6335 err = sysfs_create_file(&cls->kset.kobj,
6336 &attr_sched_smt_power_savings.attr);
6338 #ifdef CONFIG_SCHED_MC
6339 if (!err && mc_capable())
6340 err = sysfs_create_file(&cls->kset.kobj,
6341 &attr_sched_mc_power_savings.attr);
6348 * Force a reinitialization of the sched domains hierarchy. The domains
6349 * and groups cannot be updated in place without racing with the balancing
6350 * code, so we temporarily attach all running cpus to the NULL domain
6351 * which will prevent rebalancing while the sched domains are recalculated.
6353 static int update_sched_domains(struct notifier_block *nfb,
6354 unsigned long action, void *hcpu)
6357 case CPU_UP_PREPARE:
6358 case CPU_UP_PREPARE_FROZEN:
6359 case CPU_DOWN_PREPARE:
6360 case CPU_DOWN_PREPARE_FROZEN:
6361 detach_destroy_domains(&cpu_online_map);
6364 case CPU_UP_CANCELED:
6365 case CPU_UP_CANCELED_FROZEN:
6366 case CPU_DOWN_FAILED:
6367 case CPU_DOWN_FAILED_FROZEN:
6369 case CPU_ONLINE_FROZEN:
6371 case CPU_DEAD_FROZEN:
6373 * Fall through and re-initialise the domains.
6380 /* The hotplug lock is already held by cpu_up/cpu_down */
6381 arch_init_sched_domains(&cpu_online_map);
6386 void __init sched_init_smp(void)
6388 cpumask_t non_isolated_cpus;
6390 mutex_lock(&sched_hotcpu_mutex);
6391 arch_init_sched_domains(&cpu_online_map);
6392 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6393 if (cpus_empty(non_isolated_cpus))
6394 cpu_set(smp_processor_id(), non_isolated_cpus);
6395 mutex_unlock(&sched_hotcpu_mutex);
6396 /* XXX: Theoretical race here - CPU may be hotplugged now */
6397 hotcpu_notifier(update_sched_domains, 0);
6399 init_sched_domain_sysctl();
6401 /* Move init over to a non-isolated CPU */
6402 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6406 void __init sched_init_smp(void)
6409 #endif /* CONFIG_SMP */
6411 int in_sched_functions(unsigned long addr)
6413 /* Linker adds these: start and end of __sched functions */
6414 extern char __sched_text_start[], __sched_text_end[];
6416 return in_lock_functions(addr) ||
6417 (addr >= (unsigned long)__sched_text_start
6418 && addr < (unsigned long)__sched_text_end);
6421 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6423 cfs_rq->tasks_timeline = RB_ROOT;
6424 cfs_rq->fair_clock = 1;
6425 #ifdef CONFIG_FAIR_GROUP_SCHED
6430 void __init sched_init(void)
6432 int highest_cpu = 0;
6436 * Link up the scheduling class hierarchy:
6438 rt_sched_class.next = &fair_sched_class;
6439 fair_sched_class.next = &idle_sched_class;
6440 idle_sched_class.next = NULL;
6442 for_each_possible_cpu(i) {
6443 struct rt_prio_array *array;
6447 spin_lock_init(&rq->lock);
6448 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6451 init_cfs_rq(&rq->cfs, rq);
6452 #ifdef CONFIG_FAIR_GROUP_SCHED
6453 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6454 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6457 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6458 rq->cpu_load[j] = 0;
6461 rq->active_balance = 0;
6462 rq->next_balance = jiffies;
6465 rq->migration_thread = NULL;
6466 INIT_LIST_HEAD(&rq->migration_queue);
6468 atomic_set(&rq->nr_iowait, 0);
6470 array = &rq->rt.active;
6471 for (j = 0; j < MAX_RT_PRIO; j++) {
6472 INIT_LIST_HEAD(array->queue + j);
6473 __clear_bit(j, array->bitmap);
6476 /* delimiter for bitsearch: */
6477 __set_bit(MAX_RT_PRIO, array->bitmap);
6480 set_load_weight(&init_task);
6482 #ifdef CONFIG_PREEMPT_NOTIFIERS
6483 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6487 nr_cpu_ids = highest_cpu + 1;
6488 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6491 #ifdef CONFIG_RT_MUTEXES
6492 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6496 * The boot idle thread does lazy MMU switching as well:
6498 atomic_inc(&init_mm.mm_count);
6499 enter_lazy_tlb(&init_mm, current);
6502 * Make us the idle thread. Technically, schedule() should not be
6503 * called from this thread, however somewhere below it might be,
6504 * but because we are the idle thread, we just pick up running again
6505 * when this runqueue becomes "idle".
6507 init_idle(current, smp_processor_id());
6509 * During early bootup we pretend to be a normal task:
6511 current->sched_class = &fair_sched_class;
6514 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6515 void __might_sleep(char *file, int line)
6518 static unsigned long prev_jiffy; /* ratelimiting */
6520 if ((in_atomic() || irqs_disabled()) &&
6521 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6522 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6524 prev_jiffy = jiffies;
6525 printk(KERN_ERR "BUG: sleeping function called from invalid"
6526 " context at %s:%d\n", file, line);
6527 printk("in_atomic():%d, irqs_disabled():%d\n",
6528 in_atomic(), irqs_disabled());
6529 debug_show_held_locks(current);
6530 if (irqs_disabled())
6531 print_irqtrace_events(current);
6536 EXPORT_SYMBOL(__might_sleep);
6539 #ifdef CONFIG_MAGIC_SYSRQ
6540 void normalize_rt_tasks(void)
6542 struct task_struct *g, *p;
6543 unsigned long flags;
6547 read_lock_irq(&tasklist_lock);
6548 do_each_thread(g, p) {
6550 p->se.wait_runtime = 0;
6551 p->se.exec_start = 0;
6552 p->se.wait_start_fair = 0;
6553 p->se.sleep_start_fair = 0;
6554 #ifdef CONFIG_SCHEDSTATS
6555 p->se.wait_start = 0;
6556 p->se.sleep_start = 0;
6557 p->se.block_start = 0;
6559 task_rq(p)->cfs.fair_clock = 0;
6560 task_rq(p)->clock = 0;
6564 * Renice negative nice level userspace
6567 if (TASK_NICE(p) < 0 && p->mm)
6568 set_user_nice(p, 0);
6572 spin_lock_irqsave(&p->pi_lock, flags);
6573 rq = __task_rq_lock(p);
6576 * Do not touch the migration thread:
6578 if (p == rq->migration_thread)
6582 update_rq_clock(rq);
6583 on_rq = p->se.on_rq;
6585 deactivate_task(rq, p, 0);
6586 __setscheduler(rq, p, SCHED_NORMAL, 0);
6588 activate_task(rq, p, 0);
6589 resched_task(rq->curr);
6594 __task_rq_unlock(rq);
6595 spin_unlock_irqrestore(&p->pi_lock, flags);
6596 } while_each_thread(g, p);
6598 read_unlock_irq(&tasklist_lock);
6601 #endif /* CONFIG_MAGIC_SYSRQ */
6605 * These functions are only useful for the IA64 MCA handling.
6607 * They can only be called when the whole system has been
6608 * stopped - every CPU needs to be quiescent, and no scheduling
6609 * activity can take place. Using them for anything else would
6610 * be a serious bug, and as a result, they aren't even visible
6611 * under any other configuration.
6615 * curr_task - return the current task for a given cpu.
6616 * @cpu: the processor in question.
6618 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6620 struct task_struct *curr_task(int cpu)
6622 return cpu_curr(cpu);
6626 * set_curr_task - set the current task for a given cpu.
6627 * @cpu: the processor in question.
6628 * @p: the task pointer to set.
6630 * Description: This function must only be used when non-maskable interrupts
6631 * are serviced on a separate stack. It allows the architecture to switch the
6632 * notion of the current task on a cpu in a non-blocking manner. This function
6633 * must be called with all CPU's synchronized, and interrupts disabled, the
6634 * and caller must save the original value of the current task (see
6635 * curr_task() above) and restore that value before reenabling interrupts and
6636 * re-starting the system.
6638 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6640 void set_curr_task(int cpu, struct task_struct *p)