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 /* 'curr' points to currently running entity on this cfs_rq.
193 * It is set to NULL otherwise (i.e when none are currently running).
195 struct sched_entity *curr;
196 #ifdef CONFIG_FAIR_GROUP_SCHED
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 inline void update_load_add(struct load_weight *lw, unsigned long inc)
703 lw->inv_weight = WMULT_CONST / lw->weight;
706 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
709 if (likely(lw->weight))
710 lw->inv_weight = WMULT_CONST / lw->weight;
714 * To aid in avoiding the subversion of "niceness" due to uneven distribution
715 * of tasks with abnormal "nice" values across CPUs the contribution that
716 * each task makes to its run queue's load is weighted according to its
717 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
718 * scaled version of the new time slice allocation that they receive on time
722 #define WEIGHT_IDLEPRIO 2
723 #define WMULT_IDLEPRIO (1 << 31)
726 * Nice levels are multiplicative, with a gentle 10% change for every
727 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
728 * nice 1, it will get ~10% less CPU time than another CPU-bound task
729 * that remained on nice 0.
731 * The "10% effect" is relative and cumulative: from _any_ nice level,
732 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
733 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
734 * If a task goes up by ~10% and another task goes down by ~10% then
735 * the relative distance between them is ~25%.)
737 static const int prio_to_weight[40] = {
738 /* -20 */ 88761, 71755, 56483, 46273, 36291,
739 /* -15 */ 29154, 23254, 18705, 14949, 11916,
740 /* -10 */ 9548, 7620, 6100, 4904, 3906,
741 /* -5 */ 3121, 2501, 1991, 1586, 1277,
742 /* 0 */ 1024, 820, 655, 526, 423,
743 /* 5 */ 335, 272, 215, 172, 137,
744 /* 10 */ 110, 87, 70, 56, 45,
745 /* 15 */ 36, 29, 23, 18, 15,
749 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
751 * In cases where the weight does not change often, we can use the
752 * precalculated inverse to speed up arithmetics by turning divisions
753 * into multiplications:
755 static const u32 prio_to_wmult[40] = {
756 /* -20 */ 48388, 59856, 76040, 92818, 118348,
757 /* -15 */ 147320, 184698, 229616, 287308, 360437,
758 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
759 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
760 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
761 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
762 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
763 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
766 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
769 * runqueue iterator, to support SMP load-balancing between different
770 * scheduling classes, without having to expose their internal data
771 * structures to the load-balancing proper:
775 struct task_struct *(*start)(void *);
776 struct task_struct *(*next)(void *);
779 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
780 unsigned long max_nr_move, unsigned long max_load_move,
781 struct sched_domain *sd, enum cpu_idle_type idle,
782 int *all_pinned, unsigned long *load_moved,
783 int *this_best_prio, struct rq_iterator *iterator);
785 #include "sched_stats.h"
786 #include "sched_rt.c"
787 #include "sched_fair.c"
788 #include "sched_idletask.c"
789 #ifdef CONFIG_SCHED_DEBUG
790 # include "sched_debug.c"
793 #define sched_class_highest (&rt_sched_class)
796 * Update delta_exec, delta_fair fields for rq.
798 * delta_fair clock advances at a rate inversely proportional to
799 * total load (rq->ls.load.weight) on the runqueue, while
800 * delta_exec advances at the same rate as wall-clock (provided
803 * delta_exec / delta_fair is a measure of the (smoothened) load on this
804 * runqueue over any given interval. This (smoothened) load is used
805 * during load balance.
807 * This function is called /before/ updating rq->ls.load
808 * and when switching tasks.
810 static inline void inc_load(struct rq *rq, const struct task_struct *p)
812 update_load_add(&rq->ls.load, p->se.load.weight);
815 static inline void dec_load(struct rq *rq, const struct task_struct *p)
817 update_load_sub(&rq->ls.load, p->se.load.weight);
820 static void inc_nr_running(struct task_struct *p, struct rq *rq)
826 static void dec_nr_running(struct task_struct *p, struct rq *rq)
832 static void set_load_weight(struct task_struct *p)
834 p->se.wait_runtime = 0;
836 if (task_has_rt_policy(p)) {
837 p->se.load.weight = prio_to_weight[0] * 2;
838 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
843 * SCHED_IDLE tasks get minimal weight:
845 if (p->policy == SCHED_IDLE) {
846 p->se.load.weight = WEIGHT_IDLEPRIO;
847 p->se.load.inv_weight = WMULT_IDLEPRIO;
851 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
852 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
855 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
857 sched_info_queued(p);
858 p->sched_class->enqueue_task(rq, p, wakeup);
862 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
864 p->sched_class->dequeue_task(rq, p, sleep);
869 * __normal_prio - return the priority that is based on the static prio
871 static inline int __normal_prio(struct task_struct *p)
873 return p->static_prio;
877 * Calculate the expected normal priority: i.e. priority
878 * without taking RT-inheritance into account. Might be
879 * boosted by interactivity modifiers. Changes upon fork,
880 * setprio syscalls, and whenever the interactivity
881 * estimator recalculates.
883 static inline int normal_prio(struct task_struct *p)
887 if (task_has_rt_policy(p))
888 prio = MAX_RT_PRIO-1 - p->rt_priority;
890 prio = __normal_prio(p);
895 * Calculate the current priority, i.e. the priority
896 * taken into account by the scheduler. This value might
897 * be boosted by RT tasks, or might be boosted by
898 * interactivity modifiers. Will be RT if the task got
899 * RT-boosted. If not then it returns p->normal_prio.
901 static int effective_prio(struct task_struct *p)
903 p->normal_prio = normal_prio(p);
905 * If we are RT tasks or we were boosted to RT priority,
906 * keep the priority unchanged. Otherwise, update priority
907 * to the normal priority:
909 if (!rt_prio(p->prio))
910 return p->normal_prio;
915 * activate_task - move a task to the runqueue.
917 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
919 if (p->state == TASK_UNINTERRUPTIBLE)
920 rq->nr_uninterruptible--;
922 enqueue_task(rq, p, wakeup);
923 inc_nr_running(p, rq);
927 * activate_idle_task - move idle task to the _front_ of runqueue.
929 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
933 if (p->state == TASK_UNINTERRUPTIBLE)
934 rq->nr_uninterruptible--;
936 enqueue_task(rq, p, 0);
937 inc_nr_running(p, rq);
941 * deactivate_task - remove a task from the runqueue.
943 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
945 if (p->state == TASK_UNINTERRUPTIBLE)
946 rq->nr_uninterruptible++;
948 dequeue_task(rq, p, sleep);
949 dec_nr_running(p, rq);
953 * task_curr - is this task currently executing on a CPU?
954 * @p: the task in question.
956 inline int task_curr(const struct task_struct *p)
958 return cpu_curr(task_cpu(p)) == p;
961 /* Used instead of source_load when we know the type == 0 */
962 unsigned long weighted_cpuload(const int cpu)
964 return cpu_rq(cpu)->ls.load.weight;
967 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
970 task_thread_info(p)->cpu = cpu;
977 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
979 int old_cpu = task_cpu(p);
980 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
981 u64 clock_offset, fair_clock_offset;
983 clock_offset = old_rq->clock - new_rq->clock;
984 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
986 if (p->se.wait_start_fair)
987 p->se.wait_start_fair -= fair_clock_offset;
988 if (p->se.sleep_start_fair)
989 p->se.sleep_start_fair -= fair_clock_offset;
991 #ifdef CONFIG_SCHEDSTATS
992 if (p->se.wait_start)
993 p->se.wait_start -= clock_offset;
994 if (p->se.sleep_start)
995 p->se.sleep_start -= clock_offset;
996 if (p->se.block_start)
997 p->se.block_start -= clock_offset;
1000 __set_task_cpu(p, new_cpu);
1003 struct migration_req {
1004 struct list_head list;
1006 struct task_struct *task;
1009 struct completion done;
1013 * The task's runqueue lock must be held.
1014 * Returns true if you have to wait for migration thread.
1017 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1019 struct rq *rq = task_rq(p);
1022 * If the task is not on a runqueue (and not running), then
1023 * it is sufficient to simply update the task's cpu field.
1025 if (!p->se.on_rq && !task_running(rq, p)) {
1026 set_task_cpu(p, dest_cpu);
1030 init_completion(&req->done);
1032 req->dest_cpu = dest_cpu;
1033 list_add(&req->list, &rq->migration_queue);
1039 * wait_task_inactive - wait for a thread to unschedule.
1041 * The caller must ensure that the task *will* unschedule sometime soon,
1042 * else this function might spin for a *long* time. This function can't
1043 * be called with interrupts off, or it may introduce deadlock with
1044 * smp_call_function() if an IPI is sent by the same process we are
1045 * waiting to become inactive.
1047 void wait_task_inactive(struct task_struct *p)
1049 unsigned long flags;
1055 * We do the initial early heuristics without holding
1056 * any task-queue locks at all. We'll only try to get
1057 * the runqueue lock when things look like they will
1063 * If the task is actively running on another CPU
1064 * still, just relax and busy-wait without holding
1067 * NOTE! Since we don't hold any locks, it's not
1068 * even sure that "rq" stays as the right runqueue!
1069 * But we don't care, since "task_running()" will
1070 * return false if the runqueue has changed and p
1071 * is actually now running somewhere else!
1073 while (task_running(rq, p))
1077 * Ok, time to look more closely! We need the rq
1078 * lock now, to be *sure*. If we're wrong, we'll
1079 * just go back and repeat.
1081 rq = task_rq_lock(p, &flags);
1082 running = task_running(rq, p);
1083 on_rq = p->se.on_rq;
1084 task_rq_unlock(rq, &flags);
1087 * Was it really running after all now that we
1088 * checked with the proper locks actually held?
1090 * Oops. Go back and try again..
1092 if (unlikely(running)) {
1098 * It's not enough that it's not actively running,
1099 * it must be off the runqueue _entirely_, and not
1102 * So if it wa still runnable (but just not actively
1103 * running right now), it's preempted, and we should
1104 * yield - it could be a while.
1106 if (unlikely(on_rq)) {
1112 * Ahh, all good. It wasn't running, and it wasn't
1113 * runnable, which means that it will never become
1114 * running in the future either. We're all done!
1119 * kick_process - kick a running thread to enter/exit the kernel
1120 * @p: the to-be-kicked thread
1122 * Cause a process which is running on another CPU to enter
1123 * kernel-mode, without any delay. (to get signals handled.)
1125 * NOTE: this function doesnt have to take the runqueue lock,
1126 * because all it wants to ensure is that the remote task enters
1127 * the kernel. If the IPI races and the task has been migrated
1128 * to another CPU then no harm is done and the purpose has been
1131 void kick_process(struct task_struct *p)
1137 if ((cpu != smp_processor_id()) && task_curr(p))
1138 smp_send_reschedule(cpu);
1143 * Return a low guess at the load of a migration-source cpu weighted
1144 * according to the scheduling class and "nice" value.
1146 * We want to under-estimate the load of migration sources, to
1147 * balance conservatively.
1149 static inline unsigned long source_load(int cpu, int type)
1151 struct rq *rq = cpu_rq(cpu);
1152 unsigned long total = weighted_cpuload(cpu);
1157 return min(rq->cpu_load[type-1], total);
1161 * Return a high guess at the load of a migration-target cpu weighted
1162 * according to the scheduling class and "nice" value.
1164 static inline unsigned long target_load(int cpu, int type)
1166 struct rq *rq = cpu_rq(cpu);
1167 unsigned long total = weighted_cpuload(cpu);
1172 return max(rq->cpu_load[type-1], total);
1176 * Return the average load per task on the cpu's run queue
1178 static inline unsigned long cpu_avg_load_per_task(int cpu)
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long total = weighted_cpuload(cpu);
1182 unsigned long n = rq->nr_running;
1184 return n ? total / n : SCHED_LOAD_SCALE;
1188 * find_idlest_group finds and returns the least busy CPU group within the
1191 static struct sched_group *
1192 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1194 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1195 unsigned long min_load = ULONG_MAX, this_load = 0;
1196 int load_idx = sd->forkexec_idx;
1197 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1200 unsigned long load, avg_load;
1204 /* Skip over this group if it has no CPUs allowed */
1205 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1208 local_group = cpu_isset(this_cpu, group->cpumask);
1210 /* Tally up the load of all CPUs in the group */
1213 for_each_cpu_mask(i, group->cpumask) {
1214 /* Bias balancing toward cpus of our domain */
1216 load = source_load(i, load_idx);
1218 load = target_load(i, load_idx);
1223 /* Adjust by relative CPU power of the group */
1224 avg_load = sg_div_cpu_power(group,
1225 avg_load * SCHED_LOAD_SCALE);
1228 this_load = avg_load;
1230 } else if (avg_load < min_load) {
1231 min_load = avg_load;
1235 group = group->next;
1236 } while (group != sd->groups);
1238 if (!idlest || 100*this_load < imbalance*min_load)
1244 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1247 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1250 unsigned long load, min_load = ULONG_MAX;
1254 /* Traverse only the allowed CPUs */
1255 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1257 for_each_cpu_mask(i, tmp) {
1258 load = weighted_cpuload(i);
1260 if (load < min_load || (load == min_load && i == this_cpu)) {
1270 * sched_balance_self: balance the current task (running on cpu) in domains
1271 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1274 * Balance, ie. select the least loaded group.
1276 * Returns the target CPU number, or the same CPU if no balancing is needed.
1278 * preempt must be disabled.
1280 static int sched_balance_self(int cpu, int flag)
1282 struct task_struct *t = current;
1283 struct sched_domain *tmp, *sd = NULL;
1285 for_each_domain(cpu, tmp) {
1287 * If power savings logic is enabled for a domain, stop there.
1289 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1291 if (tmp->flags & flag)
1297 struct sched_group *group;
1298 int new_cpu, weight;
1300 if (!(sd->flags & flag)) {
1306 group = find_idlest_group(sd, t, cpu);
1312 new_cpu = find_idlest_cpu(group, t, cpu);
1313 if (new_cpu == -1 || new_cpu == cpu) {
1314 /* Now try balancing at a lower domain level of cpu */
1319 /* Now try balancing at a lower domain level of new_cpu */
1322 weight = cpus_weight(span);
1323 for_each_domain(cpu, tmp) {
1324 if (weight <= cpus_weight(tmp->span))
1326 if (tmp->flags & flag)
1329 /* while loop will break here if sd == NULL */
1335 #endif /* CONFIG_SMP */
1338 * wake_idle() will wake a task on an idle cpu if task->cpu is
1339 * not idle and an idle cpu is available. The span of cpus to
1340 * search starts with cpus closest then further out as needed,
1341 * so we always favor a closer, idle cpu.
1343 * Returns the CPU we should wake onto.
1345 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1346 static int wake_idle(int cpu, struct task_struct *p)
1349 struct sched_domain *sd;
1353 * If it is idle, then it is the best cpu to run this task.
1355 * This cpu is also the best, if it has more than one task already.
1356 * Siblings must be also busy(in most cases) as they didn't already
1357 * pickup the extra load from this cpu and hence we need not check
1358 * sibling runqueue info. This will avoid the checks and cache miss
1359 * penalities associated with that.
1361 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1364 for_each_domain(cpu, sd) {
1365 if (sd->flags & SD_WAKE_IDLE) {
1366 cpus_and(tmp, sd->span, p->cpus_allowed);
1367 for_each_cpu_mask(i, tmp) {
1378 static inline int wake_idle(int cpu, struct task_struct *p)
1385 * try_to_wake_up - wake up a thread
1386 * @p: the to-be-woken-up thread
1387 * @state: the mask of task states that can be woken
1388 * @sync: do a synchronous wakeup?
1390 * Put it on the run-queue if it's not already there. The "current"
1391 * thread is always on the run-queue (except when the actual
1392 * re-schedule is in progress), and as such you're allowed to do
1393 * the simpler "current->state = TASK_RUNNING" to mark yourself
1394 * runnable without the overhead of this.
1396 * returns failure only if the task is already active.
1398 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1400 int cpu, this_cpu, success = 0;
1401 unsigned long flags;
1405 struct sched_domain *sd, *this_sd = NULL;
1406 unsigned long load, this_load;
1410 rq = task_rq_lock(p, &flags);
1411 old_state = p->state;
1412 if (!(old_state & state))
1419 this_cpu = smp_processor_id();
1422 if (unlikely(task_running(rq, p)))
1427 schedstat_inc(rq, ttwu_cnt);
1428 if (cpu == this_cpu) {
1429 schedstat_inc(rq, ttwu_local);
1433 for_each_domain(this_cpu, sd) {
1434 if (cpu_isset(cpu, sd->span)) {
1435 schedstat_inc(sd, ttwu_wake_remote);
1441 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1445 * Check for affine wakeup and passive balancing possibilities.
1448 int idx = this_sd->wake_idx;
1449 unsigned int imbalance;
1451 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1453 load = source_load(cpu, idx);
1454 this_load = target_load(this_cpu, idx);
1456 new_cpu = this_cpu; /* Wake to this CPU if we can */
1458 if (this_sd->flags & SD_WAKE_AFFINE) {
1459 unsigned long tl = this_load;
1460 unsigned long tl_per_task;
1462 tl_per_task = cpu_avg_load_per_task(this_cpu);
1465 * If sync wakeup then subtract the (maximum possible)
1466 * effect of the currently running task from the load
1467 * of the current CPU:
1470 tl -= current->se.load.weight;
1473 tl + target_load(cpu, idx) <= tl_per_task) ||
1474 100*(tl + p->se.load.weight) <= imbalance*load) {
1476 * This domain has SD_WAKE_AFFINE and
1477 * p is cache cold in this domain, and
1478 * there is no bad imbalance.
1480 schedstat_inc(this_sd, ttwu_move_affine);
1486 * Start passive balancing when half the imbalance_pct
1489 if (this_sd->flags & SD_WAKE_BALANCE) {
1490 if (imbalance*this_load <= 100*load) {
1491 schedstat_inc(this_sd, ttwu_move_balance);
1497 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1499 new_cpu = wake_idle(new_cpu, p);
1500 if (new_cpu != cpu) {
1501 set_task_cpu(p, new_cpu);
1502 task_rq_unlock(rq, &flags);
1503 /* might preempt at this point */
1504 rq = task_rq_lock(p, &flags);
1505 old_state = p->state;
1506 if (!(old_state & state))
1511 this_cpu = smp_processor_id();
1516 #endif /* CONFIG_SMP */
1517 update_rq_clock(rq);
1518 activate_task(rq, p, 1);
1520 * Sync wakeups (i.e. those types of wakeups where the waker
1521 * has indicated that it will leave the CPU in short order)
1522 * don't trigger a preemption, if the woken up task will run on
1523 * this cpu. (in this case the 'I will reschedule' promise of
1524 * the waker guarantees that the freshly woken up task is going
1525 * to be considered on this CPU.)
1527 if (!sync || cpu != this_cpu)
1528 check_preempt_curr(rq, p);
1532 p->state = TASK_RUNNING;
1534 task_rq_unlock(rq, &flags);
1539 int fastcall wake_up_process(struct task_struct *p)
1541 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1542 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1544 EXPORT_SYMBOL(wake_up_process);
1546 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1548 return try_to_wake_up(p, state, 0);
1552 * Perform scheduler related setup for a newly forked process p.
1553 * p is forked by current.
1555 * __sched_fork() is basic setup used by init_idle() too:
1557 static void __sched_fork(struct task_struct *p)
1559 p->se.wait_start_fair = 0;
1560 p->se.exec_start = 0;
1561 p->se.sum_exec_runtime = 0;
1562 p->se.prev_sum_exec_runtime = 0;
1563 p->se.wait_runtime = 0;
1564 p->se.sleep_start_fair = 0;
1566 #ifdef CONFIG_SCHEDSTATS
1567 p->se.wait_start = 0;
1568 p->se.sum_wait_runtime = 0;
1569 p->se.sum_sleep_runtime = 0;
1570 p->se.sleep_start = 0;
1571 p->se.block_start = 0;
1572 p->se.sleep_max = 0;
1573 p->se.block_max = 0;
1575 p->se.slice_max = 0;
1577 p->se.wait_runtime_overruns = 0;
1578 p->se.wait_runtime_underruns = 0;
1581 INIT_LIST_HEAD(&p->run_list);
1584 #ifdef CONFIG_PREEMPT_NOTIFIERS
1585 INIT_HLIST_HEAD(&p->preempt_notifiers);
1589 * We mark the process as running here, but have not actually
1590 * inserted it onto the runqueue yet. This guarantees that
1591 * nobody will actually run it, and a signal or other external
1592 * event cannot wake it up and insert it on the runqueue either.
1594 p->state = TASK_RUNNING;
1598 * fork()/clone()-time setup:
1600 void sched_fork(struct task_struct *p, int clone_flags)
1602 int cpu = get_cpu();
1607 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1609 __set_task_cpu(p, cpu);
1612 * Make sure we do not leak PI boosting priority to the child:
1614 p->prio = current->normal_prio;
1616 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1617 if (likely(sched_info_on()))
1618 memset(&p->sched_info, 0, sizeof(p->sched_info));
1620 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1623 #ifdef CONFIG_PREEMPT
1624 /* Want to start with kernel preemption disabled. */
1625 task_thread_info(p)->preempt_count = 1;
1631 * wake_up_new_task - wake up a newly created task for the first time.
1633 * This function will do some initial scheduler statistics housekeeping
1634 * that must be done for every newly created context, then puts the task
1635 * on the runqueue and wakes it.
1637 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1639 unsigned long flags;
1643 rq = task_rq_lock(p, &flags);
1644 BUG_ON(p->state != TASK_RUNNING);
1645 this_cpu = smp_processor_id(); /* parent's CPU */
1646 update_rq_clock(rq);
1648 p->prio = effective_prio(p);
1650 if (rt_prio(p->prio))
1651 p->sched_class = &rt_sched_class;
1653 p->sched_class = &fair_sched_class;
1655 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1656 !current->se.on_rq) {
1657 activate_task(rq, p, 0);
1660 * Let the scheduling class do new task startup
1661 * management (if any):
1663 p->sched_class->task_new(rq, p);
1664 inc_nr_running(p, rq);
1666 check_preempt_curr(rq, p);
1667 task_rq_unlock(rq, &flags);
1670 #ifdef CONFIG_PREEMPT_NOTIFIERS
1673 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1674 * @notifier: notifier struct to register
1676 void preempt_notifier_register(struct preempt_notifier *notifier)
1678 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1680 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1683 * preempt_notifier_unregister - no longer interested in preemption notifications
1684 * @notifier: notifier struct to unregister
1686 * This is safe to call from within a preemption notifier.
1688 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1690 hlist_del(¬ifier->link);
1692 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1694 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1696 struct preempt_notifier *notifier;
1697 struct hlist_node *node;
1699 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1700 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1704 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1705 struct task_struct *next)
1707 struct preempt_notifier *notifier;
1708 struct hlist_node *node;
1710 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1711 notifier->ops->sched_out(notifier, next);
1716 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1721 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1722 struct task_struct *next)
1729 * prepare_task_switch - prepare to switch tasks
1730 * @rq: the runqueue preparing to switch
1731 * @prev: the current task that is being switched out
1732 * @next: the task we are going to switch to.
1734 * This is called with the rq lock held and interrupts off. It must
1735 * be paired with a subsequent finish_task_switch after the context
1738 * prepare_task_switch sets up locking and calls architecture specific
1742 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1743 struct task_struct *next)
1745 fire_sched_out_preempt_notifiers(prev, next);
1746 prepare_lock_switch(rq, next);
1747 prepare_arch_switch(next);
1751 * finish_task_switch - clean up after a task-switch
1752 * @rq: runqueue associated with task-switch
1753 * @prev: the thread we just switched away from.
1755 * finish_task_switch must be called after the context switch, paired
1756 * with a prepare_task_switch call before the context switch.
1757 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1758 * and do any other architecture-specific cleanup actions.
1760 * Note that we may have delayed dropping an mm in context_switch(). If
1761 * so, we finish that here outside of the runqueue lock. (Doing it
1762 * with the lock held can cause deadlocks; see schedule() for
1765 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1766 __releases(rq->lock)
1768 struct mm_struct *mm = rq->prev_mm;
1774 * A task struct has one reference for the use as "current".
1775 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1776 * schedule one last time. The schedule call will never return, and
1777 * the scheduled task must drop that reference.
1778 * The test for TASK_DEAD must occur while the runqueue locks are
1779 * still held, otherwise prev could be scheduled on another cpu, die
1780 * there before we look at prev->state, and then the reference would
1782 * Manfred Spraul <manfred@colorfullife.com>
1784 prev_state = prev->state;
1785 finish_arch_switch(prev);
1786 finish_lock_switch(rq, prev);
1787 fire_sched_in_preempt_notifiers(current);
1790 if (unlikely(prev_state == TASK_DEAD)) {
1792 * Remove function-return probe instances associated with this
1793 * task and put them back on the free list.
1795 kprobe_flush_task(prev);
1796 put_task_struct(prev);
1801 * schedule_tail - first thing a freshly forked thread must call.
1802 * @prev: the thread we just switched away from.
1804 asmlinkage void schedule_tail(struct task_struct *prev)
1805 __releases(rq->lock)
1807 struct rq *rq = this_rq();
1809 finish_task_switch(rq, prev);
1810 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1811 /* In this case, finish_task_switch does not reenable preemption */
1814 if (current->set_child_tid)
1815 put_user(current->pid, current->set_child_tid);
1819 * context_switch - switch to the new MM and the new
1820 * thread's register state.
1823 context_switch(struct rq *rq, struct task_struct *prev,
1824 struct task_struct *next)
1826 struct mm_struct *mm, *oldmm;
1828 prepare_task_switch(rq, prev, next);
1830 oldmm = prev->active_mm;
1832 * For paravirt, this is coupled with an exit in switch_to to
1833 * combine the page table reload and the switch backend into
1836 arch_enter_lazy_cpu_mode();
1838 if (unlikely(!mm)) {
1839 next->active_mm = oldmm;
1840 atomic_inc(&oldmm->mm_count);
1841 enter_lazy_tlb(oldmm, next);
1843 switch_mm(oldmm, mm, next);
1845 if (unlikely(!prev->mm)) {
1846 prev->active_mm = NULL;
1847 rq->prev_mm = oldmm;
1850 * Since the runqueue lock will be released by the next
1851 * task (which is an invalid locking op but in the case
1852 * of the scheduler it's an obvious special-case), so we
1853 * do an early lockdep release here:
1855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1856 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1859 /* Here we just switch the register state and the stack. */
1860 switch_to(prev, next, prev);
1864 * this_rq must be evaluated again because prev may have moved
1865 * CPUs since it called schedule(), thus the 'rq' on its stack
1866 * frame will be invalid.
1868 finish_task_switch(this_rq(), prev);
1872 * nr_running, nr_uninterruptible and nr_context_switches:
1874 * externally visible scheduler statistics: current number of runnable
1875 * threads, current number of uninterruptible-sleeping threads, total
1876 * number of context switches performed since bootup.
1878 unsigned long nr_running(void)
1880 unsigned long i, sum = 0;
1882 for_each_online_cpu(i)
1883 sum += cpu_rq(i)->nr_running;
1888 unsigned long nr_uninterruptible(void)
1890 unsigned long i, sum = 0;
1892 for_each_possible_cpu(i)
1893 sum += cpu_rq(i)->nr_uninterruptible;
1896 * Since we read the counters lockless, it might be slightly
1897 * inaccurate. Do not allow it to go below zero though:
1899 if (unlikely((long)sum < 0))
1905 unsigned long long nr_context_switches(void)
1908 unsigned long long sum = 0;
1910 for_each_possible_cpu(i)
1911 sum += cpu_rq(i)->nr_switches;
1916 unsigned long nr_iowait(void)
1918 unsigned long i, sum = 0;
1920 for_each_possible_cpu(i)
1921 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1926 unsigned long nr_active(void)
1928 unsigned long i, running = 0, uninterruptible = 0;
1930 for_each_online_cpu(i) {
1931 running += cpu_rq(i)->nr_running;
1932 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1935 if (unlikely((long)uninterruptible < 0))
1936 uninterruptible = 0;
1938 return running + uninterruptible;
1942 * Update rq->cpu_load[] statistics. This function is usually called every
1943 * scheduler tick (TICK_NSEC).
1945 static void update_cpu_load(struct rq *this_rq)
1947 unsigned long this_load = this_rq->ls.load.weight;
1950 this_rq->nr_load_updates++;
1952 /* Update our load: */
1953 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1954 unsigned long old_load, new_load;
1956 /* scale is effectively 1 << i now, and >> i divides by scale */
1958 old_load = this_rq->cpu_load[i];
1959 new_load = this_load;
1961 * Round up the averaging division if load is increasing. This
1962 * prevents us from getting stuck on 9 if the load is 10, for
1965 if (new_load > old_load)
1966 new_load += scale-1;
1967 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
1974 * double_rq_lock - safely lock two runqueues
1976 * Note this does not disable interrupts like task_rq_lock,
1977 * you need to do so manually before calling.
1979 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1980 __acquires(rq1->lock)
1981 __acquires(rq2->lock)
1983 BUG_ON(!irqs_disabled());
1985 spin_lock(&rq1->lock);
1986 __acquire(rq2->lock); /* Fake it out ;) */
1989 spin_lock(&rq1->lock);
1990 spin_lock(&rq2->lock);
1992 spin_lock(&rq2->lock);
1993 spin_lock(&rq1->lock);
1996 update_rq_clock(rq1);
1997 update_rq_clock(rq2);
2001 * double_rq_unlock - safely unlock two runqueues
2003 * Note this does not restore interrupts like task_rq_unlock,
2004 * you need to do so manually after calling.
2006 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2007 __releases(rq1->lock)
2008 __releases(rq2->lock)
2010 spin_unlock(&rq1->lock);
2012 spin_unlock(&rq2->lock);
2014 __release(rq2->lock);
2018 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2020 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2021 __releases(this_rq->lock)
2022 __acquires(busiest->lock)
2023 __acquires(this_rq->lock)
2025 if (unlikely(!irqs_disabled())) {
2026 /* printk() doesn't work good under rq->lock */
2027 spin_unlock(&this_rq->lock);
2030 if (unlikely(!spin_trylock(&busiest->lock))) {
2031 if (busiest < this_rq) {
2032 spin_unlock(&this_rq->lock);
2033 spin_lock(&busiest->lock);
2034 spin_lock(&this_rq->lock);
2036 spin_lock(&busiest->lock);
2041 * If dest_cpu is allowed for this process, migrate the task to it.
2042 * This is accomplished by forcing the cpu_allowed mask to only
2043 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2044 * the cpu_allowed mask is restored.
2046 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2048 struct migration_req req;
2049 unsigned long flags;
2052 rq = task_rq_lock(p, &flags);
2053 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2054 || unlikely(cpu_is_offline(dest_cpu)))
2057 /* force the process onto the specified CPU */
2058 if (migrate_task(p, dest_cpu, &req)) {
2059 /* Need to wait for migration thread (might exit: take ref). */
2060 struct task_struct *mt = rq->migration_thread;
2062 get_task_struct(mt);
2063 task_rq_unlock(rq, &flags);
2064 wake_up_process(mt);
2065 put_task_struct(mt);
2066 wait_for_completion(&req.done);
2071 task_rq_unlock(rq, &flags);
2075 * sched_exec - execve() is a valuable balancing opportunity, because at
2076 * this point the task has the smallest effective memory and cache footprint.
2078 void sched_exec(void)
2080 int new_cpu, this_cpu = get_cpu();
2081 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2083 if (new_cpu != this_cpu)
2084 sched_migrate_task(current, new_cpu);
2088 * pull_task - move a task from a remote runqueue to the local runqueue.
2089 * Both runqueues must be locked.
2091 static void pull_task(struct rq *src_rq, struct task_struct *p,
2092 struct rq *this_rq, int this_cpu)
2094 deactivate_task(src_rq, p, 0);
2095 set_task_cpu(p, this_cpu);
2096 activate_task(this_rq, p, 0);
2098 * Note that idle threads have a prio of MAX_PRIO, for this test
2099 * to be always true for them.
2101 check_preempt_curr(this_rq, p);
2105 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2108 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2109 struct sched_domain *sd, enum cpu_idle_type idle,
2113 * We do not migrate tasks that are:
2114 * 1) running (obviously), or
2115 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2116 * 3) are cache-hot on their current CPU.
2118 if (!cpu_isset(this_cpu, p->cpus_allowed))
2122 if (task_running(rq, p))
2128 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2129 unsigned long max_nr_move, unsigned long max_load_move,
2130 struct sched_domain *sd, enum cpu_idle_type idle,
2131 int *all_pinned, unsigned long *load_moved,
2132 int *this_best_prio, struct rq_iterator *iterator)
2134 int pulled = 0, pinned = 0, skip_for_load;
2135 struct task_struct *p;
2136 long rem_load_move = max_load_move;
2138 if (max_nr_move == 0 || max_load_move == 0)
2144 * Start the load-balancing iterator:
2146 p = iterator->start(iterator->arg);
2151 * To help distribute high priority tasks accross CPUs we don't
2152 * skip a task if it will be the highest priority task (i.e. smallest
2153 * prio value) on its new queue regardless of its load weight
2155 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2156 SCHED_LOAD_SCALE_FUZZ;
2157 if ((skip_for_load && p->prio >= *this_best_prio) ||
2158 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2159 p = iterator->next(iterator->arg);
2163 pull_task(busiest, p, this_rq, this_cpu);
2165 rem_load_move -= p->se.load.weight;
2168 * We only want to steal up to the prescribed number of tasks
2169 * and the prescribed amount of weighted load.
2171 if (pulled < max_nr_move && rem_load_move > 0) {
2172 if (p->prio < *this_best_prio)
2173 *this_best_prio = p->prio;
2174 p = iterator->next(iterator->arg);
2179 * Right now, this is the only place pull_task() is called,
2180 * so we can safely collect pull_task() stats here rather than
2181 * inside pull_task().
2183 schedstat_add(sd, lb_gained[idle], pulled);
2186 *all_pinned = pinned;
2187 *load_moved = max_load_move - rem_load_move;
2192 * move_tasks tries to move up to max_load_move weighted load from busiest to
2193 * this_rq, as part of a balancing operation within domain "sd".
2194 * Returns 1 if successful and 0 otherwise.
2196 * Called with both runqueues locked.
2198 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2199 unsigned long max_load_move,
2200 struct sched_domain *sd, enum cpu_idle_type idle,
2203 struct sched_class *class = sched_class_highest;
2204 unsigned long total_load_moved = 0;
2205 int this_best_prio = this_rq->curr->prio;
2209 class->load_balance(this_rq, this_cpu, busiest,
2210 ULONG_MAX, max_load_move - total_load_moved,
2211 sd, idle, all_pinned, &this_best_prio);
2212 class = class->next;
2213 } while (class && max_load_move > total_load_moved);
2215 return total_load_moved > 0;
2219 * move_one_task tries to move exactly one task from busiest to this_rq, as
2220 * part of active balancing operations within "domain".
2221 * Returns 1 if successful and 0 otherwise.
2223 * Called with both runqueues locked.
2225 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2226 struct sched_domain *sd, enum cpu_idle_type idle)
2228 struct sched_class *class;
2229 int this_best_prio = MAX_PRIO;
2231 for (class = sched_class_highest; class; class = class->next)
2232 if (class->load_balance(this_rq, this_cpu, busiest,
2233 1, ULONG_MAX, sd, idle, NULL,
2241 * find_busiest_group finds and returns the busiest CPU group within the
2242 * domain. It calculates and returns the amount of weighted load which
2243 * should be moved to restore balance via the imbalance parameter.
2245 static struct sched_group *
2246 find_busiest_group(struct sched_domain *sd, int this_cpu,
2247 unsigned long *imbalance, enum cpu_idle_type idle,
2248 int *sd_idle, cpumask_t *cpus, int *balance)
2250 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2251 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2252 unsigned long max_pull;
2253 unsigned long busiest_load_per_task, busiest_nr_running;
2254 unsigned long this_load_per_task, this_nr_running;
2256 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2257 int power_savings_balance = 1;
2258 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2259 unsigned long min_nr_running = ULONG_MAX;
2260 struct sched_group *group_min = NULL, *group_leader = NULL;
2263 max_load = this_load = total_load = total_pwr = 0;
2264 busiest_load_per_task = busiest_nr_running = 0;
2265 this_load_per_task = this_nr_running = 0;
2266 if (idle == CPU_NOT_IDLE)
2267 load_idx = sd->busy_idx;
2268 else if (idle == CPU_NEWLY_IDLE)
2269 load_idx = sd->newidle_idx;
2271 load_idx = sd->idle_idx;
2274 unsigned long load, group_capacity;
2277 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2278 unsigned long sum_nr_running, sum_weighted_load;
2280 local_group = cpu_isset(this_cpu, group->cpumask);
2283 balance_cpu = first_cpu(group->cpumask);
2285 /* Tally up the load of all CPUs in the group */
2286 sum_weighted_load = sum_nr_running = avg_load = 0;
2288 for_each_cpu_mask(i, group->cpumask) {
2291 if (!cpu_isset(i, *cpus))
2296 if (*sd_idle && rq->nr_running)
2299 /* Bias balancing toward cpus of our domain */
2301 if (idle_cpu(i) && !first_idle_cpu) {
2306 load = target_load(i, load_idx);
2308 load = source_load(i, load_idx);
2311 sum_nr_running += rq->nr_running;
2312 sum_weighted_load += weighted_cpuload(i);
2316 * First idle cpu or the first cpu(busiest) in this sched group
2317 * is eligible for doing load balancing at this and above
2318 * domains. In the newly idle case, we will allow all the cpu's
2319 * to do the newly idle load balance.
2321 if (idle != CPU_NEWLY_IDLE && local_group &&
2322 balance_cpu != this_cpu && balance) {
2327 total_load += avg_load;
2328 total_pwr += group->__cpu_power;
2330 /* Adjust by relative CPU power of the group */
2331 avg_load = sg_div_cpu_power(group,
2332 avg_load * SCHED_LOAD_SCALE);
2334 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2337 this_load = avg_load;
2339 this_nr_running = sum_nr_running;
2340 this_load_per_task = sum_weighted_load;
2341 } else if (avg_load > max_load &&
2342 sum_nr_running > group_capacity) {
2343 max_load = avg_load;
2345 busiest_nr_running = sum_nr_running;
2346 busiest_load_per_task = sum_weighted_load;
2349 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2351 * Busy processors will not participate in power savings
2354 if (idle == CPU_NOT_IDLE ||
2355 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2359 * If the local group is idle or completely loaded
2360 * no need to do power savings balance at this domain
2362 if (local_group && (this_nr_running >= group_capacity ||
2364 power_savings_balance = 0;
2367 * If a group is already running at full capacity or idle,
2368 * don't include that group in power savings calculations
2370 if (!power_savings_balance || sum_nr_running >= group_capacity
2375 * Calculate the group which has the least non-idle load.
2376 * This is the group from where we need to pick up the load
2379 if ((sum_nr_running < min_nr_running) ||
2380 (sum_nr_running == min_nr_running &&
2381 first_cpu(group->cpumask) <
2382 first_cpu(group_min->cpumask))) {
2384 min_nr_running = sum_nr_running;
2385 min_load_per_task = sum_weighted_load /
2390 * Calculate the group which is almost near its
2391 * capacity but still has some space to pick up some load
2392 * from other group and save more power
2394 if (sum_nr_running <= group_capacity - 1) {
2395 if (sum_nr_running > leader_nr_running ||
2396 (sum_nr_running == leader_nr_running &&
2397 first_cpu(group->cpumask) >
2398 first_cpu(group_leader->cpumask))) {
2399 group_leader = group;
2400 leader_nr_running = sum_nr_running;
2405 group = group->next;
2406 } while (group != sd->groups);
2408 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2411 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2413 if (this_load >= avg_load ||
2414 100*max_load <= sd->imbalance_pct*this_load)
2417 busiest_load_per_task /= busiest_nr_running;
2419 * We're trying to get all the cpus to the average_load, so we don't
2420 * want to push ourselves above the average load, nor do we wish to
2421 * reduce the max loaded cpu below the average load, as either of these
2422 * actions would just result in more rebalancing later, and ping-pong
2423 * tasks around. Thus we look for the minimum possible imbalance.
2424 * Negative imbalances (*we* are more loaded than anyone else) will
2425 * be counted as no imbalance for these purposes -- we can't fix that
2426 * by pulling tasks to us. Be careful of negative numbers as they'll
2427 * appear as very large values with unsigned longs.
2429 if (max_load <= busiest_load_per_task)
2433 * In the presence of smp nice balancing, certain scenarios can have
2434 * max load less than avg load(as we skip the groups at or below
2435 * its cpu_power, while calculating max_load..)
2437 if (max_load < avg_load) {
2439 goto small_imbalance;
2442 /* Don't want to pull so many tasks that a group would go idle */
2443 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2445 /* How much load to actually move to equalise the imbalance */
2446 *imbalance = min(max_pull * busiest->__cpu_power,
2447 (avg_load - this_load) * this->__cpu_power)
2451 * if *imbalance is less than the average load per runnable task
2452 * there is no gaurantee that any tasks will be moved so we'll have
2453 * a think about bumping its value to force at least one task to be
2456 if (*imbalance < busiest_load_per_task) {
2457 unsigned long tmp, pwr_now, pwr_move;
2461 pwr_move = pwr_now = 0;
2463 if (this_nr_running) {
2464 this_load_per_task /= this_nr_running;
2465 if (busiest_load_per_task > this_load_per_task)
2468 this_load_per_task = SCHED_LOAD_SCALE;
2470 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2471 busiest_load_per_task * imbn) {
2472 *imbalance = busiest_load_per_task;
2477 * OK, we don't have enough imbalance to justify moving tasks,
2478 * however we may be able to increase total CPU power used by
2482 pwr_now += busiest->__cpu_power *
2483 min(busiest_load_per_task, max_load);
2484 pwr_now += this->__cpu_power *
2485 min(this_load_per_task, this_load);
2486 pwr_now /= SCHED_LOAD_SCALE;
2488 /* Amount of load we'd subtract */
2489 tmp = sg_div_cpu_power(busiest,
2490 busiest_load_per_task * SCHED_LOAD_SCALE);
2492 pwr_move += busiest->__cpu_power *
2493 min(busiest_load_per_task, max_load - tmp);
2495 /* Amount of load we'd add */
2496 if (max_load * busiest->__cpu_power <
2497 busiest_load_per_task * SCHED_LOAD_SCALE)
2498 tmp = sg_div_cpu_power(this,
2499 max_load * busiest->__cpu_power);
2501 tmp = sg_div_cpu_power(this,
2502 busiest_load_per_task * SCHED_LOAD_SCALE);
2503 pwr_move += this->__cpu_power *
2504 min(this_load_per_task, this_load + tmp);
2505 pwr_move /= SCHED_LOAD_SCALE;
2507 /* Move if we gain throughput */
2508 if (pwr_move > pwr_now)
2509 *imbalance = busiest_load_per_task;
2515 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2516 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2519 if (this == group_leader && group_leader != group_min) {
2520 *imbalance = min_load_per_task;
2530 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2533 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2534 unsigned long imbalance, cpumask_t *cpus)
2536 struct rq *busiest = NULL, *rq;
2537 unsigned long max_load = 0;
2540 for_each_cpu_mask(i, group->cpumask) {
2543 if (!cpu_isset(i, *cpus))
2547 wl = weighted_cpuload(i);
2549 if (rq->nr_running == 1 && wl > imbalance)
2552 if (wl > max_load) {
2562 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2563 * so long as it is large enough.
2565 #define MAX_PINNED_INTERVAL 512
2568 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2569 * tasks if there is an imbalance.
2571 static int load_balance(int this_cpu, struct rq *this_rq,
2572 struct sched_domain *sd, enum cpu_idle_type idle,
2575 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2576 struct sched_group *group;
2577 unsigned long imbalance;
2579 cpumask_t cpus = CPU_MASK_ALL;
2580 unsigned long flags;
2583 * When power savings policy is enabled for the parent domain, idle
2584 * sibling can pick up load irrespective of busy siblings. In this case,
2585 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2586 * portraying it as CPU_NOT_IDLE.
2588 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2589 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2592 schedstat_inc(sd, lb_cnt[idle]);
2595 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2602 schedstat_inc(sd, lb_nobusyg[idle]);
2606 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2608 schedstat_inc(sd, lb_nobusyq[idle]);
2612 BUG_ON(busiest == this_rq);
2614 schedstat_add(sd, lb_imbalance[idle], imbalance);
2617 if (busiest->nr_running > 1) {
2619 * Attempt to move tasks. If find_busiest_group has found
2620 * an imbalance but busiest->nr_running <= 1, the group is
2621 * still unbalanced. ld_moved simply stays zero, so it is
2622 * correctly treated as an imbalance.
2624 local_irq_save(flags);
2625 double_rq_lock(this_rq, busiest);
2626 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2627 imbalance, sd, idle, &all_pinned);
2628 double_rq_unlock(this_rq, busiest);
2629 local_irq_restore(flags);
2632 * some other cpu did the load balance for us.
2634 if (ld_moved && this_cpu != smp_processor_id())
2635 resched_cpu(this_cpu);
2637 /* All tasks on this runqueue were pinned by CPU affinity */
2638 if (unlikely(all_pinned)) {
2639 cpu_clear(cpu_of(busiest), cpus);
2640 if (!cpus_empty(cpus))
2647 schedstat_inc(sd, lb_failed[idle]);
2648 sd->nr_balance_failed++;
2650 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2652 spin_lock_irqsave(&busiest->lock, flags);
2654 /* don't kick the migration_thread, if the curr
2655 * task on busiest cpu can't be moved to this_cpu
2657 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2658 spin_unlock_irqrestore(&busiest->lock, flags);
2660 goto out_one_pinned;
2663 if (!busiest->active_balance) {
2664 busiest->active_balance = 1;
2665 busiest->push_cpu = this_cpu;
2668 spin_unlock_irqrestore(&busiest->lock, flags);
2670 wake_up_process(busiest->migration_thread);
2673 * We've kicked active balancing, reset the failure
2676 sd->nr_balance_failed = sd->cache_nice_tries+1;
2679 sd->nr_balance_failed = 0;
2681 if (likely(!active_balance)) {
2682 /* We were unbalanced, so reset the balancing interval */
2683 sd->balance_interval = sd->min_interval;
2686 * If we've begun active balancing, start to back off. This
2687 * case may not be covered by the all_pinned logic if there
2688 * is only 1 task on the busy runqueue (because we don't call
2691 if (sd->balance_interval < sd->max_interval)
2692 sd->balance_interval *= 2;
2695 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2696 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2701 schedstat_inc(sd, lb_balanced[idle]);
2703 sd->nr_balance_failed = 0;
2706 /* tune up the balancing interval */
2707 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2708 (sd->balance_interval < sd->max_interval))
2709 sd->balance_interval *= 2;
2711 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2712 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2718 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2719 * tasks if there is an imbalance.
2721 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2722 * this_rq is locked.
2725 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2727 struct sched_group *group;
2728 struct rq *busiest = NULL;
2729 unsigned long imbalance;
2733 cpumask_t cpus = CPU_MASK_ALL;
2736 * When power savings policy is enabled for the parent domain, idle
2737 * sibling can pick up load irrespective of busy siblings. In this case,
2738 * let the state of idle sibling percolate up as IDLE, instead of
2739 * portraying it as CPU_NOT_IDLE.
2741 if (sd->flags & SD_SHARE_CPUPOWER &&
2742 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2745 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2747 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2748 &sd_idle, &cpus, NULL);
2750 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2754 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2757 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2761 BUG_ON(busiest == this_rq);
2763 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2766 if (busiest->nr_running > 1) {
2767 /* Attempt to move tasks */
2768 double_lock_balance(this_rq, busiest);
2769 /* this_rq->clock is already updated */
2770 update_rq_clock(busiest);
2771 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2772 imbalance, sd, CPU_NEWLY_IDLE,
2774 spin_unlock(&busiest->lock);
2776 if (unlikely(all_pinned)) {
2777 cpu_clear(cpu_of(busiest), cpus);
2778 if (!cpus_empty(cpus))
2784 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2785 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2786 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2789 sd->nr_balance_failed = 0;
2794 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2795 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2796 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2798 sd->nr_balance_failed = 0;
2804 * idle_balance is called by schedule() if this_cpu is about to become
2805 * idle. Attempts to pull tasks from other CPUs.
2807 static void idle_balance(int this_cpu, struct rq *this_rq)
2809 struct sched_domain *sd;
2810 int pulled_task = -1;
2811 unsigned long next_balance = jiffies + HZ;
2813 for_each_domain(this_cpu, sd) {
2814 unsigned long interval;
2816 if (!(sd->flags & SD_LOAD_BALANCE))
2819 if (sd->flags & SD_BALANCE_NEWIDLE)
2820 /* If we've pulled tasks over stop searching: */
2821 pulled_task = load_balance_newidle(this_cpu,
2824 interval = msecs_to_jiffies(sd->balance_interval);
2825 if (time_after(next_balance, sd->last_balance + interval))
2826 next_balance = sd->last_balance + interval;
2830 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2832 * We are going idle. next_balance may be set based on
2833 * a busy processor. So reset next_balance.
2835 this_rq->next_balance = next_balance;
2840 * active_load_balance is run by migration threads. It pushes running tasks
2841 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2842 * running on each physical CPU where possible, and avoids physical /
2843 * logical imbalances.
2845 * Called with busiest_rq locked.
2847 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2849 int target_cpu = busiest_rq->push_cpu;
2850 struct sched_domain *sd;
2851 struct rq *target_rq;
2853 /* Is there any task to move? */
2854 if (busiest_rq->nr_running <= 1)
2857 target_rq = cpu_rq(target_cpu);
2860 * This condition is "impossible", if it occurs
2861 * we need to fix it. Originally reported by
2862 * Bjorn Helgaas on a 128-cpu setup.
2864 BUG_ON(busiest_rq == target_rq);
2866 /* move a task from busiest_rq to target_rq */
2867 double_lock_balance(busiest_rq, target_rq);
2868 update_rq_clock(busiest_rq);
2869 update_rq_clock(target_rq);
2871 /* Search for an sd spanning us and the target CPU. */
2872 for_each_domain(target_cpu, sd) {
2873 if ((sd->flags & SD_LOAD_BALANCE) &&
2874 cpu_isset(busiest_cpu, sd->span))
2879 schedstat_inc(sd, alb_cnt);
2881 if (move_one_task(target_rq, target_cpu, busiest_rq,
2883 schedstat_inc(sd, alb_pushed);
2885 schedstat_inc(sd, alb_failed);
2887 spin_unlock(&target_rq->lock);
2892 atomic_t load_balancer;
2894 } nohz ____cacheline_aligned = {
2895 .load_balancer = ATOMIC_INIT(-1),
2896 .cpu_mask = CPU_MASK_NONE,
2900 * This routine will try to nominate the ilb (idle load balancing)
2901 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2902 * load balancing on behalf of all those cpus. If all the cpus in the system
2903 * go into this tickless mode, then there will be no ilb owner (as there is
2904 * no need for one) and all the cpus will sleep till the next wakeup event
2907 * For the ilb owner, tick is not stopped. And this tick will be used
2908 * for idle load balancing. ilb owner will still be part of
2911 * While stopping the tick, this cpu will become the ilb owner if there
2912 * is no other owner. And will be the owner till that cpu becomes busy
2913 * or if all cpus in the system stop their ticks at which point
2914 * there is no need for ilb owner.
2916 * When the ilb owner becomes busy, it nominates another owner, during the
2917 * next busy scheduler_tick()
2919 int select_nohz_load_balancer(int stop_tick)
2921 int cpu = smp_processor_id();
2924 cpu_set(cpu, nohz.cpu_mask);
2925 cpu_rq(cpu)->in_nohz_recently = 1;
2928 * If we are going offline and still the leader, give up!
2930 if (cpu_is_offline(cpu) &&
2931 atomic_read(&nohz.load_balancer) == cpu) {
2932 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2937 /* time for ilb owner also to sleep */
2938 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2939 if (atomic_read(&nohz.load_balancer) == cpu)
2940 atomic_set(&nohz.load_balancer, -1);
2944 if (atomic_read(&nohz.load_balancer) == -1) {
2945 /* make me the ilb owner */
2946 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2948 } else if (atomic_read(&nohz.load_balancer) == cpu)
2951 if (!cpu_isset(cpu, nohz.cpu_mask))
2954 cpu_clear(cpu, nohz.cpu_mask);
2956 if (atomic_read(&nohz.load_balancer) == cpu)
2957 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2964 static DEFINE_SPINLOCK(balancing);
2967 * It checks each scheduling domain to see if it is due to be balanced,
2968 * and initiates a balancing operation if so.
2970 * Balancing parameters are set up in arch_init_sched_domains.
2972 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2975 struct rq *rq = cpu_rq(cpu);
2976 unsigned long interval;
2977 struct sched_domain *sd;
2978 /* Earliest time when we have to do rebalance again */
2979 unsigned long next_balance = jiffies + 60*HZ;
2980 int update_next_balance = 0;
2982 for_each_domain(cpu, sd) {
2983 if (!(sd->flags & SD_LOAD_BALANCE))
2986 interval = sd->balance_interval;
2987 if (idle != CPU_IDLE)
2988 interval *= sd->busy_factor;
2990 /* scale ms to jiffies */
2991 interval = msecs_to_jiffies(interval);
2992 if (unlikely(!interval))
2994 if (interval > HZ*NR_CPUS/10)
2995 interval = HZ*NR_CPUS/10;
2998 if (sd->flags & SD_SERIALIZE) {
2999 if (!spin_trylock(&balancing))
3003 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3004 if (load_balance(cpu, rq, sd, idle, &balance)) {
3006 * We've pulled tasks over so either we're no
3007 * longer idle, or one of our SMT siblings is
3010 idle = CPU_NOT_IDLE;
3012 sd->last_balance = jiffies;
3014 if (sd->flags & SD_SERIALIZE)
3015 spin_unlock(&balancing);
3017 if (time_after(next_balance, sd->last_balance + interval)) {
3018 next_balance = sd->last_balance + interval;
3019 update_next_balance = 1;
3023 * Stop the load balance at this level. There is another
3024 * CPU in our sched group which is doing load balancing more
3032 * next_balance will be updated only when there is a need.
3033 * When the cpu is attached to null domain for ex, it will not be
3036 if (likely(update_next_balance))
3037 rq->next_balance = next_balance;
3041 * run_rebalance_domains is triggered when needed from the scheduler tick.
3042 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3043 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3045 static void run_rebalance_domains(struct softirq_action *h)
3047 int this_cpu = smp_processor_id();
3048 struct rq *this_rq = cpu_rq(this_cpu);
3049 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3050 CPU_IDLE : CPU_NOT_IDLE;
3052 rebalance_domains(this_cpu, idle);
3056 * If this cpu is the owner for idle load balancing, then do the
3057 * balancing on behalf of the other idle cpus whose ticks are
3060 if (this_rq->idle_at_tick &&
3061 atomic_read(&nohz.load_balancer) == this_cpu) {
3062 cpumask_t cpus = nohz.cpu_mask;
3066 cpu_clear(this_cpu, cpus);
3067 for_each_cpu_mask(balance_cpu, cpus) {
3069 * If this cpu gets work to do, stop the load balancing
3070 * work being done for other cpus. Next load
3071 * balancing owner will pick it up.
3076 rebalance_domains(balance_cpu, CPU_IDLE);
3078 rq = cpu_rq(balance_cpu);
3079 if (time_after(this_rq->next_balance, rq->next_balance))
3080 this_rq->next_balance = rq->next_balance;
3087 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3089 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3090 * idle load balancing owner or decide to stop the periodic load balancing,
3091 * if the whole system is idle.
3093 static inline void trigger_load_balance(struct rq *rq, int cpu)
3097 * If we were in the nohz mode recently and busy at the current
3098 * scheduler tick, then check if we need to nominate new idle
3101 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3102 rq->in_nohz_recently = 0;
3104 if (atomic_read(&nohz.load_balancer) == cpu) {
3105 cpu_clear(cpu, nohz.cpu_mask);
3106 atomic_set(&nohz.load_balancer, -1);
3109 if (atomic_read(&nohz.load_balancer) == -1) {
3111 * simple selection for now: Nominate the
3112 * first cpu in the nohz list to be the next
3115 * TBD: Traverse the sched domains and nominate
3116 * the nearest cpu in the nohz.cpu_mask.
3118 int ilb = first_cpu(nohz.cpu_mask);
3126 * If this cpu is idle and doing idle load balancing for all the
3127 * cpus with ticks stopped, is it time for that to stop?
3129 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3130 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3136 * If this cpu is idle and the idle load balancing is done by
3137 * someone else, then no need raise the SCHED_SOFTIRQ
3139 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3140 cpu_isset(cpu, nohz.cpu_mask))
3143 if (time_after_eq(jiffies, rq->next_balance))
3144 raise_softirq(SCHED_SOFTIRQ);
3147 #else /* CONFIG_SMP */
3150 * on UP we do not need to balance between CPUs:
3152 static inline void idle_balance(int cpu, struct rq *rq)
3156 /* Avoid "used but not defined" warning on UP */
3157 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3158 unsigned long max_nr_move, unsigned long max_load_move,
3159 struct sched_domain *sd, enum cpu_idle_type idle,
3160 int *all_pinned, unsigned long *load_moved,
3161 int *this_best_prio, struct rq_iterator *iterator)
3170 DEFINE_PER_CPU(struct kernel_stat, kstat);
3172 EXPORT_PER_CPU_SYMBOL(kstat);
3175 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3176 * that have not yet been banked in case the task is currently running.
3178 unsigned long long task_sched_runtime(struct task_struct *p)
3180 unsigned long flags;
3184 rq = task_rq_lock(p, &flags);
3185 ns = p->se.sum_exec_runtime;
3186 if (rq->curr == p) {
3187 update_rq_clock(rq);
3188 delta_exec = rq->clock - p->se.exec_start;
3189 if ((s64)delta_exec > 0)
3192 task_rq_unlock(rq, &flags);
3198 * Account user cpu time to a process.
3199 * @p: the process that the cpu time gets accounted to
3200 * @hardirq_offset: the offset to subtract from hardirq_count()
3201 * @cputime: the cpu time spent in user space since the last update
3203 void account_user_time(struct task_struct *p, cputime_t cputime)
3205 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3208 p->utime = cputime_add(p->utime, cputime);
3210 /* Add user time to cpustat. */
3211 tmp = cputime_to_cputime64(cputime);
3212 if (TASK_NICE(p) > 0)
3213 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3215 cpustat->user = cputime64_add(cpustat->user, tmp);
3219 * Account system cpu time to a process.
3220 * @p: the process that the cpu time gets accounted to
3221 * @hardirq_offset: the offset to subtract from hardirq_count()
3222 * @cputime: the cpu time spent in kernel space since the last update
3224 void account_system_time(struct task_struct *p, int hardirq_offset,
3227 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3228 struct rq *rq = this_rq();
3231 p->stime = cputime_add(p->stime, cputime);
3233 /* Add system time to cpustat. */
3234 tmp = cputime_to_cputime64(cputime);
3235 if (hardirq_count() - hardirq_offset)
3236 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3237 else if (softirq_count())
3238 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3239 else if (p != rq->idle)
3240 cpustat->system = cputime64_add(cpustat->system, tmp);
3241 else if (atomic_read(&rq->nr_iowait) > 0)
3242 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3244 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3245 /* Account for system time used */
3246 acct_update_integrals(p);
3250 * Account for involuntary wait time.
3251 * @p: the process from which the cpu time has been stolen
3252 * @steal: the cpu time spent in involuntary wait
3254 void account_steal_time(struct task_struct *p, cputime_t steal)
3256 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3257 cputime64_t tmp = cputime_to_cputime64(steal);
3258 struct rq *rq = this_rq();
3260 if (p == rq->idle) {
3261 p->stime = cputime_add(p->stime, steal);
3262 if (atomic_read(&rq->nr_iowait) > 0)
3263 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3265 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3267 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3271 * This function gets called by the timer code, with HZ frequency.
3272 * We call it with interrupts disabled.
3274 * It also gets called by the fork code, when changing the parent's
3277 void scheduler_tick(void)
3279 int cpu = smp_processor_id();
3280 struct rq *rq = cpu_rq(cpu);
3281 struct task_struct *curr = rq->curr;
3282 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3284 spin_lock(&rq->lock);
3285 __update_rq_clock(rq);
3287 * Let rq->clock advance by at least TICK_NSEC:
3289 if (unlikely(rq->clock < next_tick))
3290 rq->clock = next_tick;
3291 rq->tick_timestamp = rq->clock;
3292 update_cpu_load(rq);
3293 if (curr != rq->idle) /* FIXME: needed? */
3294 curr->sched_class->task_tick(rq, curr);
3295 spin_unlock(&rq->lock);
3298 rq->idle_at_tick = idle_cpu(cpu);
3299 trigger_load_balance(rq, cpu);
3303 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3305 void fastcall add_preempt_count(int val)
3310 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3312 preempt_count() += val;
3314 * Spinlock count overflowing soon?
3316 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3319 EXPORT_SYMBOL(add_preempt_count);
3321 void fastcall sub_preempt_count(int val)
3326 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3329 * Is the spinlock portion underflowing?
3331 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3332 !(preempt_count() & PREEMPT_MASK)))
3335 preempt_count() -= val;
3337 EXPORT_SYMBOL(sub_preempt_count);
3342 * Print scheduling while atomic bug:
3344 static noinline void __schedule_bug(struct task_struct *prev)
3346 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3347 prev->comm, preempt_count(), prev->pid);
3348 debug_show_held_locks(prev);
3349 if (irqs_disabled())
3350 print_irqtrace_events(prev);
3355 * Various schedule()-time debugging checks and statistics:
3357 static inline void schedule_debug(struct task_struct *prev)
3360 * Test if we are atomic. Since do_exit() needs to call into
3361 * schedule() atomically, we ignore that path for now.
3362 * Otherwise, whine if we are scheduling when we should not be.
3364 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3365 __schedule_bug(prev);
3367 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3369 schedstat_inc(this_rq(), sched_cnt);
3373 * Pick up the highest-prio task:
3375 static inline struct task_struct *
3376 pick_next_task(struct rq *rq, struct task_struct *prev)
3378 struct sched_class *class;
3379 struct task_struct *p;
3382 * Optimization: we know that if all tasks are in
3383 * the fair class we can call that function directly:
3385 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3386 p = fair_sched_class.pick_next_task(rq);
3391 class = sched_class_highest;
3393 p = class->pick_next_task(rq);
3397 * Will never be NULL as the idle class always
3398 * returns a non-NULL p:
3400 class = class->next;
3405 * schedule() is the main scheduler function.
3407 asmlinkage void __sched schedule(void)
3409 struct task_struct *prev, *next;
3416 cpu = smp_processor_id();
3420 switch_count = &prev->nivcsw;
3422 release_kernel_lock(prev);
3423 need_resched_nonpreemptible:
3425 schedule_debug(prev);
3427 spin_lock_irq(&rq->lock);
3428 clear_tsk_need_resched(prev);
3429 __update_rq_clock(rq);
3431 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3432 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3433 unlikely(signal_pending(prev)))) {
3434 prev->state = TASK_RUNNING;
3436 deactivate_task(rq, prev, 1);
3438 switch_count = &prev->nvcsw;
3441 if (unlikely(!rq->nr_running))
3442 idle_balance(cpu, rq);
3444 prev->sched_class->put_prev_task(rq, prev);
3445 next = pick_next_task(rq, prev);
3447 sched_info_switch(prev, next);
3449 if (likely(prev != next)) {
3454 context_switch(rq, prev, next); /* unlocks the rq */
3456 spin_unlock_irq(&rq->lock);
3458 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3459 cpu = smp_processor_id();
3461 goto need_resched_nonpreemptible;
3463 preempt_enable_no_resched();
3464 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3467 EXPORT_SYMBOL(schedule);
3469 #ifdef CONFIG_PREEMPT
3471 * this is the entry point to schedule() from in-kernel preemption
3472 * off of preempt_enable. Kernel preemptions off return from interrupt
3473 * occur there and call schedule directly.
3475 asmlinkage void __sched preempt_schedule(void)
3477 struct thread_info *ti = current_thread_info();
3478 #ifdef CONFIG_PREEMPT_BKL
3479 struct task_struct *task = current;
3480 int saved_lock_depth;
3483 * If there is a non-zero preempt_count or interrupts are disabled,
3484 * we do not want to preempt the current task. Just return..
3486 if (likely(ti->preempt_count || irqs_disabled()))
3490 add_preempt_count(PREEMPT_ACTIVE);
3492 * We keep the big kernel semaphore locked, but we
3493 * clear ->lock_depth so that schedule() doesnt
3494 * auto-release the semaphore:
3496 #ifdef CONFIG_PREEMPT_BKL
3497 saved_lock_depth = task->lock_depth;
3498 task->lock_depth = -1;
3501 #ifdef CONFIG_PREEMPT_BKL
3502 task->lock_depth = saved_lock_depth;
3504 sub_preempt_count(PREEMPT_ACTIVE);
3506 /* we could miss a preemption opportunity between schedule and now */
3508 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3511 EXPORT_SYMBOL(preempt_schedule);
3514 * this is the entry point to schedule() from kernel preemption
3515 * off of irq context.
3516 * Note, that this is called and return with irqs disabled. This will
3517 * protect us against recursive calling from irq.
3519 asmlinkage void __sched preempt_schedule_irq(void)
3521 struct thread_info *ti = current_thread_info();
3522 #ifdef CONFIG_PREEMPT_BKL
3523 struct task_struct *task = current;
3524 int saved_lock_depth;
3526 /* Catch callers which need to be fixed */
3527 BUG_ON(ti->preempt_count || !irqs_disabled());
3530 add_preempt_count(PREEMPT_ACTIVE);
3532 * We keep the big kernel semaphore locked, but we
3533 * clear ->lock_depth so that schedule() doesnt
3534 * auto-release the semaphore:
3536 #ifdef CONFIG_PREEMPT_BKL
3537 saved_lock_depth = task->lock_depth;
3538 task->lock_depth = -1;
3542 local_irq_disable();
3543 #ifdef CONFIG_PREEMPT_BKL
3544 task->lock_depth = saved_lock_depth;
3546 sub_preempt_count(PREEMPT_ACTIVE);
3548 /* we could miss a preemption opportunity between schedule and now */
3550 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3554 #endif /* CONFIG_PREEMPT */
3556 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3559 return try_to_wake_up(curr->private, mode, sync);
3561 EXPORT_SYMBOL(default_wake_function);
3564 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3565 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3566 * number) then we wake all the non-exclusive tasks and one exclusive task.
3568 * There are circumstances in which we can try to wake a task which has already
3569 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3570 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3572 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3573 int nr_exclusive, int sync, void *key)
3575 wait_queue_t *curr, *next;
3577 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3578 unsigned flags = curr->flags;
3580 if (curr->func(curr, mode, sync, key) &&
3581 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3587 * __wake_up - wake up threads blocked on a waitqueue.
3589 * @mode: which threads
3590 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3591 * @key: is directly passed to the wakeup function
3593 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3594 int nr_exclusive, void *key)
3596 unsigned long flags;
3598 spin_lock_irqsave(&q->lock, flags);
3599 __wake_up_common(q, mode, nr_exclusive, 0, key);
3600 spin_unlock_irqrestore(&q->lock, flags);
3602 EXPORT_SYMBOL(__wake_up);
3605 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3607 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3609 __wake_up_common(q, mode, 1, 0, NULL);
3613 * __wake_up_sync - wake up threads blocked on a waitqueue.
3615 * @mode: which threads
3616 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3618 * The sync wakeup differs that the waker knows that it will schedule
3619 * away soon, so while the target thread will be woken up, it will not
3620 * be migrated to another CPU - ie. the two threads are 'synchronized'
3621 * with each other. This can prevent needless bouncing between CPUs.
3623 * On UP it can prevent extra preemption.
3626 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3628 unsigned long flags;
3634 if (unlikely(!nr_exclusive))
3637 spin_lock_irqsave(&q->lock, flags);
3638 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3639 spin_unlock_irqrestore(&q->lock, flags);
3641 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3643 void fastcall complete(struct completion *x)
3645 unsigned long flags;
3647 spin_lock_irqsave(&x->wait.lock, flags);
3649 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3651 spin_unlock_irqrestore(&x->wait.lock, flags);
3653 EXPORT_SYMBOL(complete);
3655 void fastcall complete_all(struct completion *x)
3657 unsigned long flags;
3659 spin_lock_irqsave(&x->wait.lock, flags);
3660 x->done += UINT_MAX/2;
3661 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3663 spin_unlock_irqrestore(&x->wait.lock, flags);
3665 EXPORT_SYMBOL(complete_all);
3667 void fastcall __sched wait_for_completion(struct completion *x)
3671 spin_lock_irq(&x->wait.lock);
3673 DECLARE_WAITQUEUE(wait, current);
3675 wait.flags |= WQ_FLAG_EXCLUSIVE;
3676 __add_wait_queue_tail(&x->wait, &wait);
3678 __set_current_state(TASK_UNINTERRUPTIBLE);
3679 spin_unlock_irq(&x->wait.lock);
3681 spin_lock_irq(&x->wait.lock);
3683 __remove_wait_queue(&x->wait, &wait);
3686 spin_unlock_irq(&x->wait.lock);
3688 EXPORT_SYMBOL(wait_for_completion);
3690 unsigned long fastcall __sched
3691 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3695 spin_lock_irq(&x->wait.lock);
3697 DECLARE_WAITQUEUE(wait, current);
3699 wait.flags |= WQ_FLAG_EXCLUSIVE;
3700 __add_wait_queue_tail(&x->wait, &wait);
3702 __set_current_state(TASK_UNINTERRUPTIBLE);
3703 spin_unlock_irq(&x->wait.lock);
3704 timeout = schedule_timeout(timeout);
3705 spin_lock_irq(&x->wait.lock);
3707 __remove_wait_queue(&x->wait, &wait);
3711 __remove_wait_queue(&x->wait, &wait);
3715 spin_unlock_irq(&x->wait.lock);
3718 EXPORT_SYMBOL(wait_for_completion_timeout);
3720 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3726 spin_lock_irq(&x->wait.lock);
3728 DECLARE_WAITQUEUE(wait, current);
3730 wait.flags |= WQ_FLAG_EXCLUSIVE;
3731 __add_wait_queue_tail(&x->wait, &wait);
3733 if (signal_pending(current)) {
3735 __remove_wait_queue(&x->wait, &wait);
3738 __set_current_state(TASK_INTERRUPTIBLE);
3739 spin_unlock_irq(&x->wait.lock);
3741 spin_lock_irq(&x->wait.lock);
3743 __remove_wait_queue(&x->wait, &wait);
3747 spin_unlock_irq(&x->wait.lock);
3751 EXPORT_SYMBOL(wait_for_completion_interruptible);
3753 unsigned long fastcall __sched
3754 wait_for_completion_interruptible_timeout(struct completion *x,
3755 unsigned long timeout)
3759 spin_lock_irq(&x->wait.lock);
3761 DECLARE_WAITQUEUE(wait, current);
3763 wait.flags |= WQ_FLAG_EXCLUSIVE;
3764 __add_wait_queue_tail(&x->wait, &wait);
3766 if (signal_pending(current)) {
3767 timeout = -ERESTARTSYS;
3768 __remove_wait_queue(&x->wait, &wait);
3771 __set_current_state(TASK_INTERRUPTIBLE);
3772 spin_unlock_irq(&x->wait.lock);
3773 timeout = schedule_timeout(timeout);
3774 spin_lock_irq(&x->wait.lock);
3776 __remove_wait_queue(&x->wait, &wait);
3780 __remove_wait_queue(&x->wait, &wait);
3784 spin_unlock_irq(&x->wait.lock);
3787 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3790 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3792 spin_lock_irqsave(&q->lock, *flags);
3793 __add_wait_queue(q, wait);
3794 spin_unlock(&q->lock);
3798 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3800 spin_lock_irq(&q->lock);
3801 __remove_wait_queue(q, wait);
3802 spin_unlock_irqrestore(&q->lock, *flags);
3805 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3807 unsigned long flags;
3810 init_waitqueue_entry(&wait, current);
3812 current->state = TASK_INTERRUPTIBLE;
3814 sleep_on_head(q, &wait, &flags);
3816 sleep_on_tail(q, &wait, &flags);
3818 EXPORT_SYMBOL(interruptible_sleep_on);
3821 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3823 unsigned long flags;
3826 init_waitqueue_entry(&wait, current);
3828 current->state = TASK_INTERRUPTIBLE;
3830 sleep_on_head(q, &wait, &flags);
3831 timeout = schedule_timeout(timeout);
3832 sleep_on_tail(q, &wait, &flags);
3836 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3838 void __sched sleep_on(wait_queue_head_t *q)
3840 unsigned long flags;
3843 init_waitqueue_entry(&wait, current);
3845 current->state = TASK_UNINTERRUPTIBLE;
3847 sleep_on_head(q, &wait, &flags);
3849 sleep_on_tail(q, &wait, &flags);
3851 EXPORT_SYMBOL(sleep_on);
3853 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3855 unsigned long flags;
3858 init_waitqueue_entry(&wait, current);
3860 current->state = TASK_UNINTERRUPTIBLE;
3862 sleep_on_head(q, &wait, &flags);
3863 timeout = schedule_timeout(timeout);
3864 sleep_on_tail(q, &wait, &flags);
3868 EXPORT_SYMBOL(sleep_on_timeout);
3870 #ifdef CONFIG_RT_MUTEXES
3873 * rt_mutex_setprio - set the current priority of a task
3875 * @prio: prio value (kernel-internal form)
3877 * This function changes the 'effective' priority of a task. It does
3878 * not touch ->normal_prio like __setscheduler().
3880 * Used by the rt_mutex code to implement priority inheritance logic.
3882 void rt_mutex_setprio(struct task_struct *p, int prio)
3884 unsigned long flags;
3888 BUG_ON(prio < 0 || prio > MAX_PRIO);
3890 rq = task_rq_lock(p, &flags);
3891 update_rq_clock(rq);
3894 on_rq = p->se.on_rq;
3896 dequeue_task(rq, p, 0);
3899 p->sched_class = &rt_sched_class;
3901 p->sched_class = &fair_sched_class;
3906 enqueue_task(rq, p, 0);
3908 * Reschedule if we are currently running on this runqueue and
3909 * our priority decreased, or if we are not currently running on
3910 * this runqueue and our priority is higher than the current's
3912 if (task_running(rq, p)) {
3913 if (p->prio > oldprio)
3914 resched_task(rq->curr);
3916 check_preempt_curr(rq, p);
3919 task_rq_unlock(rq, &flags);
3924 void set_user_nice(struct task_struct *p, long nice)
3926 int old_prio, delta, on_rq;
3927 unsigned long flags;
3930 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3933 * We have to be careful, if called from sys_setpriority(),
3934 * the task might be in the middle of scheduling on another CPU.
3936 rq = task_rq_lock(p, &flags);
3937 update_rq_clock(rq);
3939 * The RT priorities are set via sched_setscheduler(), but we still
3940 * allow the 'normal' nice value to be set - but as expected
3941 * it wont have any effect on scheduling until the task is
3942 * SCHED_FIFO/SCHED_RR:
3944 if (task_has_rt_policy(p)) {
3945 p->static_prio = NICE_TO_PRIO(nice);
3948 on_rq = p->se.on_rq;
3950 dequeue_task(rq, p, 0);
3954 p->static_prio = NICE_TO_PRIO(nice);
3957 p->prio = effective_prio(p);
3958 delta = p->prio - old_prio;
3961 enqueue_task(rq, p, 0);
3964 * If the task increased its priority or is running and
3965 * lowered its priority, then reschedule its CPU:
3967 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3968 resched_task(rq->curr);
3971 task_rq_unlock(rq, &flags);
3973 EXPORT_SYMBOL(set_user_nice);
3976 * can_nice - check if a task can reduce its nice value
3980 int can_nice(const struct task_struct *p, const int nice)
3982 /* convert nice value [19,-20] to rlimit style value [1,40] */
3983 int nice_rlim = 20 - nice;
3985 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3986 capable(CAP_SYS_NICE));
3989 #ifdef __ARCH_WANT_SYS_NICE
3992 * sys_nice - change the priority of the current process.
3993 * @increment: priority increment
3995 * sys_setpriority is a more generic, but much slower function that
3996 * does similar things.
3998 asmlinkage long sys_nice(int increment)
4003 * Setpriority might change our priority at the same moment.
4004 * We don't have to worry. Conceptually one call occurs first
4005 * and we have a single winner.
4007 if (increment < -40)
4012 nice = PRIO_TO_NICE(current->static_prio) + increment;
4018 if (increment < 0 && !can_nice(current, nice))
4021 retval = security_task_setnice(current, nice);
4025 set_user_nice(current, nice);
4032 * task_prio - return the priority value of a given task.
4033 * @p: the task in question.
4035 * This is the priority value as seen by users in /proc.
4036 * RT tasks are offset by -200. Normal tasks are centered
4037 * around 0, value goes from -16 to +15.
4039 int task_prio(const struct task_struct *p)
4041 return p->prio - MAX_RT_PRIO;
4045 * task_nice - return the nice value of a given task.
4046 * @p: the task in question.
4048 int task_nice(const struct task_struct *p)
4050 return TASK_NICE(p);
4052 EXPORT_SYMBOL_GPL(task_nice);
4055 * idle_cpu - is a given cpu idle currently?
4056 * @cpu: the processor in question.
4058 int idle_cpu(int cpu)
4060 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4064 * idle_task - return the idle task for a given cpu.
4065 * @cpu: the processor in question.
4067 struct task_struct *idle_task(int cpu)
4069 return cpu_rq(cpu)->idle;
4073 * find_process_by_pid - find a process with a matching PID value.
4074 * @pid: the pid in question.
4076 static inline struct task_struct *find_process_by_pid(pid_t pid)
4078 return pid ? find_task_by_pid(pid) : current;
4081 /* Actually do priority change: must hold rq lock. */
4083 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4085 BUG_ON(p->se.on_rq);
4088 switch (p->policy) {
4092 p->sched_class = &fair_sched_class;
4096 p->sched_class = &rt_sched_class;
4100 p->rt_priority = prio;
4101 p->normal_prio = normal_prio(p);
4102 /* we are holding p->pi_lock already */
4103 p->prio = rt_mutex_getprio(p);
4108 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4109 * @p: the task in question.
4110 * @policy: new policy.
4111 * @param: structure containing the new RT priority.
4113 * NOTE that the task may be already dead.
4115 int sched_setscheduler(struct task_struct *p, int policy,
4116 struct sched_param *param)
4118 int retval, oldprio, oldpolicy = -1, on_rq;
4119 unsigned long flags;
4122 /* may grab non-irq protected spin_locks */
4123 BUG_ON(in_interrupt());
4125 /* double check policy once rq lock held */
4127 policy = oldpolicy = p->policy;
4128 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4129 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4130 policy != SCHED_IDLE)
4133 * Valid priorities for SCHED_FIFO and SCHED_RR are
4134 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4135 * SCHED_BATCH and SCHED_IDLE is 0.
4137 if (param->sched_priority < 0 ||
4138 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4139 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4141 if (rt_policy(policy) != (param->sched_priority != 0))
4145 * Allow unprivileged RT tasks to decrease priority:
4147 if (!capable(CAP_SYS_NICE)) {
4148 if (rt_policy(policy)) {
4149 unsigned long rlim_rtprio;
4151 if (!lock_task_sighand(p, &flags))
4153 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4154 unlock_task_sighand(p, &flags);
4156 /* can't set/change the rt policy */
4157 if (policy != p->policy && !rlim_rtprio)
4160 /* can't increase priority */
4161 if (param->sched_priority > p->rt_priority &&
4162 param->sched_priority > rlim_rtprio)
4166 * Like positive nice levels, dont allow tasks to
4167 * move out of SCHED_IDLE either:
4169 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4172 /* can't change other user's priorities */
4173 if ((current->euid != p->euid) &&
4174 (current->euid != p->uid))
4178 retval = security_task_setscheduler(p, policy, param);
4182 * make sure no PI-waiters arrive (or leave) while we are
4183 * changing the priority of the task:
4185 spin_lock_irqsave(&p->pi_lock, flags);
4187 * To be able to change p->policy safely, the apropriate
4188 * runqueue lock must be held.
4190 rq = __task_rq_lock(p);
4191 /* recheck policy now with rq lock held */
4192 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4193 policy = oldpolicy = -1;
4194 __task_rq_unlock(rq);
4195 spin_unlock_irqrestore(&p->pi_lock, flags);
4198 update_rq_clock(rq);
4199 on_rq = p->se.on_rq;
4201 deactivate_task(rq, p, 0);
4203 __setscheduler(rq, p, policy, param->sched_priority);
4205 activate_task(rq, p, 0);
4207 * Reschedule if we are currently running on this runqueue and
4208 * our priority decreased, or if we are not currently running on
4209 * this runqueue and our priority is higher than the current's
4211 if (task_running(rq, p)) {
4212 if (p->prio > oldprio)
4213 resched_task(rq->curr);
4215 check_preempt_curr(rq, p);
4218 __task_rq_unlock(rq);
4219 spin_unlock_irqrestore(&p->pi_lock, flags);
4221 rt_mutex_adjust_pi(p);
4225 EXPORT_SYMBOL_GPL(sched_setscheduler);
4228 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4230 struct sched_param lparam;
4231 struct task_struct *p;
4234 if (!param || pid < 0)
4236 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4241 p = find_process_by_pid(pid);
4243 retval = sched_setscheduler(p, policy, &lparam);
4250 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4251 * @pid: the pid in question.
4252 * @policy: new policy.
4253 * @param: structure containing the new RT priority.
4255 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4256 struct sched_param __user *param)
4258 /* negative values for policy are not valid */
4262 return do_sched_setscheduler(pid, policy, param);
4266 * sys_sched_setparam - set/change the RT priority of a thread
4267 * @pid: the pid in question.
4268 * @param: structure containing the new RT priority.
4270 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4272 return do_sched_setscheduler(pid, -1, param);
4276 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4277 * @pid: the pid in question.
4279 asmlinkage long sys_sched_getscheduler(pid_t pid)
4281 struct task_struct *p;
4282 int retval = -EINVAL;
4288 read_lock(&tasklist_lock);
4289 p = find_process_by_pid(pid);
4291 retval = security_task_getscheduler(p);
4295 read_unlock(&tasklist_lock);
4302 * sys_sched_getscheduler - get the RT priority of a thread
4303 * @pid: the pid in question.
4304 * @param: structure containing the RT priority.
4306 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4308 struct sched_param lp;
4309 struct task_struct *p;
4310 int retval = -EINVAL;
4312 if (!param || pid < 0)
4315 read_lock(&tasklist_lock);
4316 p = find_process_by_pid(pid);
4321 retval = security_task_getscheduler(p);
4325 lp.sched_priority = p->rt_priority;
4326 read_unlock(&tasklist_lock);
4329 * This one might sleep, we cannot do it with a spinlock held ...
4331 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4337 read_unlock(&tasklist_lock);
4341 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4343 cpumask_t cpus_allowed;
4344 struct task_struct *p;
4347 mutex_lock(&sched_hotcpu_mutex);
4348 read_lock(&tasklist_lock);
4350 p = find_process_by_pid(pid);
4352 read_unlock(&tasklist_lock);
4353 mutex_unlock(&sched_hotcpu_mutex);
4358 * It is not safe to call set_cpus_allowed with the
4359 * tasklist_lock held. We will bump the task_struct's
4360 * usage count and then drop tasklist_lock.
4363 read_unlock(&tasklist_lock);
4366 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4367 !capable(CAP_SYS_NICE))
4370 retval = security_task_setscheduler(p, 0, NULL);
4374 cpus_allowed = cpuset_cpus_allowed(p);
4375 cpus_and(new_mask, new_mask, cpus_allowed);
4376 retval = set_cpus_allowed(p, new_mask);
4380 mutex_unlock(&sched_hotcpu_mutex);
4384 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4385 cpumask_t *new_mask)
4387 if (len < sizeof(cpumask_t)) {
4388 memset(new_mask, 0, sizeof(cpumask_t));
4389 } else if (len > sizeof(cpumask_t)) {
4390 len = sizeof(cpumask_t);
4392 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4396 * sys_sched_setaffinity - set the cpu affinity of a process
4397 * @pid: pid of the process
4398 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4399 * @user_mask_ptr: user-space pointer to the new cpu mask
4401 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4402 unsigned long __user *user_mask_ptr)
4407 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4411 return sched_setaffinity(pid, new_mask);
4415 * Represents all cpu's present in the system
4416 * In systems capable of hotplug, this map could dynamically grow
4417 * as new cpu's are detected in the system via any platform specific
4418 * method, such as ACPI for e.g.
4421 cpumask_t cpu_present_map __read_mostly;
4422 EXPORT_SYMBOL(cpu_present_map);
4425 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4426 EXPORT_SYMBOL(cpu_online_map);
4428 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4429 EXPORT_SYMBOL(cpu_possible_map);
4432 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4434 struct task_struct *p;
4437 mutex_lock(&sched_hotcpu_mutex);
4438 read_lock(&tasklist_lock);
4441 p = find_process_by_pid(pid);
4445 retval = security_task_getscheduler(p);
4449 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4452 read_unlock(&tasklist_lock);
4453 mutex_unlock(&sched_hotcpu_mutex);
4459 * sys_sched_getaffinity - get the cpu affinity of a process
4460 * @pid: pid of the process
4461 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4462 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4464 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4465 unsigned long __user *user_mask_ptr)
4470 if (len < sizeof(cpumask_t))
4473 ret = sched_getaffinity(pid, &mask);
4477 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4480 return sizeof(cpumask_t);
4484 * sys_sched_yield - yield the current processor to other threads.
4486 * This function yields the current CPU to other tasks. If there are no
4487 * other threads running on this CPU then this function will return.
4489 asmlinkage long sys_sched_yield(void)
4491 struct rq *rq = this_rq_lock();
4493 schedstat_inc(rq, yld_cnt);
4494 current->sched_class->yield_task(rq, current);
4497 * Since we are going to call schedule() anyway, there's
4498 * no need to preempt or enable interrupts:
4500 __release(rq->lock);
4501 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4502 _raw_spin_unlock(&rq->lock);
4503 preempt_enable_no_resched();
4510 static void __cond_resched(void)
4512 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4513 __might_sleep(__FILE__, __LINE__);
4516 * The BKS might be reacquired before we have dropped
4517 * PREEMPT_ACTIVE, which could trigger a second
4518 * cond_resched() call.
4521 add_preempt_count(PREEMPT_ACTIVE);
4523 sub_preempt_count(PREEMPT_ACTIVE);
4524 } while (need_resched());
4527 int __sched cond_resched(void)
4529 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4530 system_state == SYSTEM_RUNNING) {
4536 EXPORT_SYMBOL(cond_resched);
4539 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4540 * call schedule, and on return reacquire the lock.
4542 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4543 * operations here to prevent schedule() from being called twice (once via
4544 * spin_unlock(), once by hand).
4546 int cond_resched_lock(spinlock_t *lock)
4550 if (need_lockbreak(lock)) {
4556 if (need_resched() && system_state == SYSTEM_RUNNING) {
4557 spin_release(&lock->dep_map, 1, _THIS_IP_);
4558 _raw_spin_unlock(lock);
4559 preempt_enable_no_resched();
4566 EXPORT_SYMBOL(cond_resched_lock);
4568 int __sched cond_resched_softirq(void)
4570 BUG_ON(!in_softirq());
4572 if (need_resched() && system_state == SYSTEM_RUNNING) {
4580 EXPORT_SYMBOL(cond_resched_softirq);
4583 * yield - yield the current processor to other threads.
4585 * This is a shortcut for kernel-space yielding - it marks the
4586 * thread runnable and calls sys_sched_yield().
4588 void __sched yield(void)
4590 set_current_state(TASK_RUNNING);
4593 EXPORT_SYMBOL(yield);
4596 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4597 * that process accounting knows that this is a task in IO wait state.
4599 * But don't do that if it is a deliberate, throttling IO wait (this task
4600 * has set its backing_dev_info: the queue against which it should throttle)
4602 void __sched io_schedule(void)
4604 struct rq *rq = &__raw_get_cpu_var(runqueues);
4606 delayacct_blkio_start();
4607 atomic_inc(&rq->nr_iowait);
4609 atomic_dec(&rq->nr_iowait);
4610 delayacct_blkio_end();
4612 EXPORT_SYMBOL(io_schedule);
4614 long __sched io_schedule_timeout(long timeout)
4616 struct rq *rq = &__raw_get_cpu_var(runqueues);
4619 delayacct_blkio_start();
4620 atomic_inc(&rq->nr_iowait);
4621 ret = schedule_timeout(timeout);
4622 atomic_dec(&rq->nr_iowait);
4623 delayacct_blkio_end();
4628 * sys_sched_get_priority_max - return maximum RT priority.
4629 * @policy: scheduling class.
4631 * this syscall returns the maximum rt_priority that can be used
4632 * by a given scheduling class.
4634 asmlinkage long sys_sched_get_priority_max(int policy)
4641 ret = MAX_USER_RT_PRIO-1;
4653 * sys_sched_get_priority_min - return minimum RT priority.
4654 * @policy: scheduling class.
4656 * this syscall returns the minimum rt_priority that can be used
4657 * by a given scheduling class.
4659 asmlinkage long sys_sched_get_priority_min(int policy)
4677 * sys_sched_rr_get_interval - return the default timeslice of a process.
4678 * @pid: pid of the process.
4679 * @interval: userspace pointer to the timeslice value.
4681 * this syscall writes the default timeslice value of a given process
4682 * into the user-space timespec buffer. A value of '0' means infinity.
4685 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4687 struct task_struct *p;
4688 int retval = -EINVAL;
4695 read_lock(&tasklist_lock);
4696 p = find_process_by_pid(pid);
4700 retval = security_task_getscheduler(p);
4704 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4705 0 : static_prio_timeslice(p->static_prio), &t);
4706 read_unlock(&tasklist_lock);
4707 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4711 read_unlock(&tasklist_lock);
4715 static const char stat_nam[] = "RSDTtZX";
4717 static void show_task(struct task_struct *p)
4719 unsigned long free = 0;
4722 state = p->state ? __ffs(p->state) + 1 : 0;
4723 printk("%-13.13s %c", p->comm,
4724 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4725 #if BITS_PER_LONG == 32
4726 if (state == TASK_RUNNING)
4727 printk(" running ");
4729 printk(" %08lx ", thread_saved_pc(p));
4731 if (state == TASK_RUNNING)
4732 printk(" running task ");
4734 printk(" %016lx ", thread_saved_pc(p));
4736 #ifdef CONFIG_DEBUG_STACK_USAGE
4738 unsigned long *n = end_of_stack(p);
4741 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4744 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4746 if (state != TASK_RUNNING)
4747 show_stack(p, NULL);
4750 void show_state_filter(unsigned long state_filter)
4752 struct task_struct *g, *p;
4754 #if BITS_PER_LONG == 32
4756 " task PC stack pid father\n");
4759 " task PC stack pid father\n");
4761 read_lock(&tasklist_lock);
4762 do_each_thread(g, p) {
4764 * reset the NMI-timeout, listing all files on a slow
4765 * console might take alot of time:
4767 touch_nmi_watchdog();
4768 if (!state_filter || (p->state & state_filter))
4770 } while_each_thread(g, p);
4772 touch_all_softlockup_watchdogs();
4774 #ifdef CONFIG_SCHED_DEBUG
4775 sysrq_sched_debug_show();
4777 read_unlock(&tasklist_lock);
4779 * Only show locks if all tasks are dumped:
4781 if (state_filter == -1)
4782 debug_show_all_locks();
4785 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4787 idle->sched_class = &idle_sched_class;
4791 * init_idle - set up an idle thread for a given CPU
4792 * @idle: task in question
4793 * @cpu: cpu the idle task belongs to
4795 * NOTE: this function does not set the idle thread's NEED_RESCHED
4796 * flag, to make booting more robust.
4798 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4800 struct rq *rq = cpu_rq(cpu);
4801 unsigned long flags;
4804 idle->se.exec_start = sched_clock();
4806 idle->prio = idle->normal_prio = MAX_PRIO;
4807 idle->cpus_allowed = cpumask_of_cpu(cpu);
4808 __set_task_cpu(idle, cpu);
4810 spin_lock_irqsave(&rq->lock, flags);
4811 rq->curr = rq->idle = idle;
4812 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4815 spin_unlock_irqrestore(&rq->lock, flags);
4817 /* Set the preempt count _outside_ the spinlocks! */
4818 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4819 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4821 task_thread_info(idle)->preempt_count = 0;
4824 * The idle tasks have their own, simple scheduling class:
4826 idle->sched_class = &idle_sched_class;
4830 * In a system that switches off the HZ timer nohz_cpu_mask
4831 * indicates which cpus entered this state. This is used
4832 * in the rcu update to wait only for active cpus. For system
4833 * which do not switch off the HZ timer nohz_cpu_mask should
4834 * always be CPU_MASK_NONE.
4836 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4840 * This is how migration works:
4842 * 1) we queue a struct migration_req structure in the source CPU's
4843 * runqueue and wake up that CPU's migration thread.
4844 * 2) we down() the locked semaphore => thread blocks.
4845 * 3) migration thread wakes up (implicitly it forces the migrated
4846 * thread off the CPU)
4847 * 4) it gets the migration request and checks whether the migrated
4848 * task is still in the wrong runqueue.
4849 * 5) if it's in the wrong runqueue then the migration thread removes
4850 * it and puts it into the right queue.
4851 * 6) migration thread up()s the semaphore.
4852 * 7) we wake up and the migration is done.
4856 * Change a given task's CPU affinity. Migrate the thread to a
4857 * proper CPU and schedule it away if the CPU it's executing on
4858 * is removed from the allowed bitmask.
4860 * NOTE: the caller must have a valid reference to the task, the
4861 * task must not exit() & deallocate itself prematurely. The
4862 * call is not atomic; no spinlocks may be held.
4864 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4866 struct migration_req req;
4867 unsigned long flags;
4871 rq = task_rq_lock(p, &flags);
4872 if (!cpus_intersects(new_mask, cpu_online_map)) {
4877 p->cpus_allowed = new_mask;
4878 /* Can the task run on the task's current CPU? If so, we're done */
4879 if (cpu_isset(task_cpu(p), new_mask))
4882 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4883 /* Need help from migration thread: drop lock and wait. */
4884 task_rq_unlock(rq, &flags);
4885 wake_up_process(rq->migration_thread);
4886 wait_for_completion(&req.done);
4887 tlb_migrate_finish(p->mm);
4891 task_rq_unlock(rq, &flags);
4895 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4898 * Move (not current) task off this cpu, onto dest cpu. We're doing
4899 * this because either it can't run here any more (set_cpus_allowed()
4900 * away from this CPU, or CPU going down), or because we're
4901 * attempting to rebalance this task on exec (sched_exec).
4903 * So we race with normal scheduler movements, but that's OK, as long
4904 * as the task is no longer on this CPU.
4906 * Returns non-zero if task was successfully migrated.
4908 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4910 struct rq *rq_dest, *rq_src;
4913 if (unlikely(cpu_is_offline(dest_cpu)))
4916 rq_src = cpu_rq(src_cpu);
4917 rq_dest = cpu_rq(dest_cpu);
4919 double_rq_lock(rq_src, rq_dest);
4920 /* Already moved. */
4921 if (task_cpu(p) != src_cpu)
4923 /* Affinity changed (again). */
4924 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4927 on_rq = p->se.on_rq;
4929 deactivate_task(rq_src, p, 0);
4931 set_task_cpu(p, dest_cpu);
4933 activate_task(rq_dest, p, 0);
4934 check_preempt_curr(rq_dest, p);
4938 double_rq_unlock(rq_src, rq_dest);
4943 * migration_thread - this is a highprio system thread that performs
4944 * thread migration by bumping thread off CPU then 'pushing' onto
4947 static int migration_thread(void *data)
4949 int cpu = (long)data;
4953 BUG_ON(rq->migration_thread != current);
4955 set_current_state(TASK_INTERRUPTIBLE);
4956 while (!kthread_should_stop()) {
4957 struct migration_req *req;
4958 struct list_head *head;
4960 spin_lock_irq(&rq->lock);
4962 if (cpu_is_offline(cpu)) {
4963 spin_unlock_irq(&rq->lock);
4967 if (rq->active_balance) {
4968 active_load_balance(rq, cpu);
4969 rq->active_balance = 0;
4972 head = &rq->migration_queue;
4974 if (list_empty(head)) {
4975 spin_unlock_irq(&rq->lock);
4977 set_current_state(TASK_INTERRUPTIBLE);
4980 req = list_entry(head->next, struct migration_req, list);
4981 list_del_init(head->next);
4983 spin_unlock(&rq->lock);
4984 __migrate_task(req->task, cpu, req->dest_cpu);
4987 complete(&req->done);
4989 __set_current_state(TASK_RUNNING);
4993 /* Wait for kthread_stop */
4994 set_current_state(TASK_INTERRUPTIBLE);
4995 while (!kthread_should_stop()) {
4997 set_current_state(TASK_INTERRUPTIBLE);
4999 __set_current_state(TASK_RUNNING);
5003 #ifdef CONFIG_HOTPLUG_CPU
5005 * Figure out where task on dead CPU should go, use force if neccessary.
5006 * NOTE: interrupts should be disabled by the caller
5008 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5010 unsigned long flags;
5017 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5018 cpus_and(mask, mask, p->cpus_allowed);
5019 dest_cpu = any_online_cpu(mask);
5021 /* On any allowed CPU? */
5022 if (dest_cpu == NR_CPUS)
5023 dest_cpu = any_online_cpu(p->cpus_allowed);
5025 /* No more Mr. Nice Guy. */
5026 if (dest_cpu == NR_CPUS) {
5027 rq = task_rq_lock(p, &flags);
5028 cpus_setall(p->cpus_allowed);
5029 dest_cpu = any_online_cpu(p->cpus_allowed);
5030 task_rq_unlock(rq, &flags);
5033 * Don't tell them about moving exiting tasks or
5034 * kernel threads (both mm NULL), since they never
5037 if (p->mm && printk_ratelimit())
5038 printk(KERN_INFO "process %d (%s) no "
5039 "longer affine to cpu%d\n",
5040 p->pid, p->comm, dead_cpu);
5042 if (!__migrate_task(p, dead_cpu, dest_cpu))
5047 * While a dead CPU has no uninterruptible tasks queued at this point,
5048 * it might still have a nonzero ->nr_uninterruptible counter, because
5049 * for performance reasons the counter is not stricly tracking tasks to
5050 * their home CPUs. So we just add the counter to another CPU's counter,
5051 * to keep the global sum constant after CPU-down:
5053 static void migrate_nr_uninterruptible(struct rq *rq_src)
5055 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5056 unsigned long flags;
5058 local_irq_save(flags);
5059 double_rq_lock(rq_src, rq_dest);
5060 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5061 rq_src->nr_uninterruptible = 0;
5062 double_rq_unlock(rq_src, rq_dest);
5063 local_irq_restore(flags);
5066 /* Run through task list and migrate tasks from the dead cpu. */
5067 static void migrate_live_tasks(int src_cpu)
5069 struct task_struct *p, *t;
5071 write_lock_irq(&tasklist_lock);
5073 do_each_thread(t, p) {
5077 if (task_cpu(p) == src_cpu)
5078 move_task_off_dead_cpu(src_cpu, p);
5079 } while_each_thread(t, p);
5081 write_unlock_irq(&tasklist_lock);
5085 * Schedules idle task to be the next runnable task on current CPU.
5086 * It does so by boosting its priority to highest possible and adding it to
5087 * the _front_ of the runqueue. Used by CPU offline code.
5089 void sched_idle_next(void)
5091 int this_cpu = smp_processor_id();
5092 struct rq *rq = cpu_rq(this_cpu);
5093 struct task_struct *p = rq->idle;
5094 unsigned long flags;
5096 /* cpu has to be offline */
5097 BUG_ON(cpu_online(this_cpu));
5100 * Strictly not necessary since rest of the CPUs are stopped by now
5101 * and interrupts disabled on the current cpu.
5103 spin_lock_irqsave(&rq->lock, flags);
5105 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5107 /* Add idle task to the _front_ of its priority queue: */
5108 activate_idle_task(p, rq);
5110 spin_unlock_irqrestore(&rq->lock, flags);
5114 * Ensures that the idle task is using init_mm right before its cpu goes
5117 void idle_task_exit(void)
5119 struct mm_struct *mm = current->active_mm;
5121 BUG_ON(cpu_online(smp_processor_id()));
5124 switch_mm(mm, &init_mm, current);
5128 /* called under rq->lock with disabled interrupts */
5129 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5131 struct rq *rq = cpu_rq(dead_cpu);
5133 /* Must be exiting, otherwise would be on tasklist. */
5134 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5136 /* Cannot have done final schedule yet: would have vanished. */
5137 BUG_ON(p->state == TASK_DEAD);
5142 * Drop lock around migration; if someone else moves it,
5143 * that's OK. No task can be added to this CPU, so iteration is
5145 * NOTE: interrupts should be left disabled --dev@
5147 spin_unlock(&rq->lock);
5148 move_task_off_dead_cpu(dead_cpu, p);
5149 spin_lock(&rq->lock);
5154 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5155 static void migrate_dead_tasks(unsigned int dead_cpu)
5157 struct rq *rq = cpu_rq(dead_cpu);
5158 struct task_struct *next;
5161 if (!rq->nr_running)
5163 update_rq_clock(rq);
5164 next = pick_next_task(rq, rq->curr);
5167 migrate_dead(dead_cpu, next);
5171 #endif /* CONFIG_HOTPLUG_CPU */
5173 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5175 static struct ctl_table sd_ctl_dir[] = {
5177 .procname = "sched_domain",
5183 static struct ctl_table sd_ctl_root[] = {
5185 .ctl_name = CTL_KERN,
5186 .procname = "kernel",
5188 .child = sd_ctl_dir,
5193 static struct ctl_table *sd_alloc_ctl_entry(int n)
5195 struct ctl_table *entry =
5196 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5199 memset(entry, 0, n * sizeof(struct ctl_table));
5205 set_table_entry(struct ctl_table *entry,
5206 const char *procname, void *data, int maxlen,
5207 mode_t mode, proc_handler *proc_handler)
5209 entry->procname = procname;
5211 entry->maxlen = maxlen;
5213 entry->proc_handler = proc_handler;
5216 static struct ctl_table *
5217 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5219 struct ctl_table *table = sd_alloc_ctl_entry(14);
5221 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5222 sizeof(long), 0644, proc_doulongvec_minmax);
5223 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5224 sizeof(long), 0644, proc_doulongvec_minmax);
5225 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5226 sizeof(int), 0644, proc_dointvec_minmax);
5227 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5228 sizeof(int), 0644, proc_dointvec_minmax);
5229 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5230 sizeof(int), 0644, proc_dointvec_minmax);
5231 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5232 sizeof(int), 0644, proc_dointvec_minmax);
5233 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5234 sizeof(int), 0644, proc_dointvec_minmax);
5235 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5236 sizeof(int), 0644, proc_dointvec_minmax);
5237 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5238 sizeof(int), 0644, proc_dointvec_minmax);
5239 set_table_entry(&table[10], "cache_nice_tries",
5240 &sd->cache_nice_tries,
5241 sizeof(int), 0644, proc_dointvec_minmax);
5242 set_table_entry(&table[12], "flags", &sd->flags,
5243 sizeof(int), 0644, proc_dointvec_minmax);
5248 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5250 struct ctl_table *entry, *table;
5251 struct sched_domain *sd;
5252 int domain_num = 0, i;
5255 for_each_domain(cpu, sd)
5257 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5260 for_each_domain(cpu, sd) {
5261 snprintf(buf, 32, "domain%d", i);
5262 entry->procname = kstrdup(buf, GFP_KERNEL);
5264 entry->child = sd_alloc_ctl_domain_table(sd);
5271 static struct ctl_table_header *sd_sysctl_header;
5272 static void init_sched_domain_sysctl(void)
5274 int i, cpu_num = num_online_cpus();
5275 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5278 sd_ctl_dir[0].child = entry;
5280 for (i = 0; i < cpu_num; i++, entry++) {
5281 snprintf(buf, 32, "cpu%d", i);
5282 entry->procname = kstrdup(buf, GFP_KERNEL);
5284 entry->child = sd_alloc_ctl_cpu_table(i);
5286 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5289 static void init_sched_domain_sysctl(void)
5295 * migration_call - callback that gets triggered when a CPU is added.
5296 * Here we can start up the necessary migration thread for the new CPU.
5298 static int __cpuinit
5299 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5301 struct task_struct *p;
5302 int cpu = (long)hcpu;
5303 unsigned long flags;
5307 case CPU_LOCK_ACQUIRE:
5308 mutex_lock(&sched_hotcpu_mutex);
5311 case CPU_UP_PREPARE:
5312 case CPU_UP_PREPARE_FROZEN:
5313 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5316 kthread_bind(p, cpu);
5317 /* Must be high prio: stop_machine expects to yield to it. */
5318 rq = task_rq_lock(p, &flags);
5319 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5320 task_rq_unlock(rq, &flags);
5321 cpu_rq(cpu)->migration_thread = p;
5325 case CPU_ONLINE_FROZEN:
5326 /* Strictly unneccessary, as first user will wake it. */
5327 wake_up_process(cpu_rq(cpu)->migration_thread);
5330 #ifdef CONFIG_HOTPLUG_CPU
5331 case CPU_UP_CANCELED:
5332 case CPU_UP_CANCELED_FROZEN:
5333 if (!cpu_rq(cpu)->migration_thread)
5335 /* Unbind it from offline cpu so it can run. Fall thru. */
5336 kthread_bind(cpu_rq(cpu)->migration_thread,
5337 any_online_cpu(cpu_online_map));
5338 kthread_stop(cpu_rq(cpu)->migration_thread);
5339 cpu_rq(cpu)->migration_thread = NULL;
5343 case CPU_DEAD_FROZEN:
5344 migrate_live_tasks(cpu);
5346 kthread_stop(rq->migration_thread);
5347 rq->migration_thread = NULL;
5348 /* Idle task back to normal (off runqueue, low prio) */
5349 rq = task_rq_lock(rq->idle, &flags);
5350 update_rq_clock(rq);
5351 deactivate_task(rq, rq->idle, 0);
5352 rq->idle->static_prio = MAX_PRIO;
5353 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5354 rq->idle->sched_class = &idle_sched_class;
5355 migrate_dead_tasks(cpu);
5356 task_rq_unlock(rq, &flags);
5357 migrate_nr_uninterruptible(rq);
5358 BUG_ON(rq->nr_running != 0);
5360 /* No need to migrate the tasks: it was best-effort if
5361 * they didn't take sched_hotcpu_mutex. Just wake up
5362 * the requestors. */
5363 spin_lock_irq(&rq->lock);
5364 while (!list_empty(&rq->migration_queue)) {
5365 struct migration_req *req;
5367 req = list_entry(rq->migration_queue.next,
5368 struct migration_req, list);
5369 list_del_init(&req->list);
5370 complete(&req->done);
5372 spin_unlock_irq(&rq->lock);
5375 case CPU_LOCK_RELEASE:
5376 mutex_unlock(&sched_hotcpu_mutex);
5382 /* Register at highest priority so that task migration (migrate_all_tasks)
5383 * happens before everything else.
5385 static struct notifier_block __cpuinitdata migration_notifier = {
5386 .notifier_call = migration_call,
5390 int __init migration_init(void)
5392 void *cpu = (void *)(long)smp_processor_id();
5395 /* Start one for the boot CPU: */
5396 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5397 BUG_ON(err == NOTIFY_BAD);
5398 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5399 register_cpu_notifier(&migration_notifier);
5407 /* Number of possible processor ids */
5408 int nr_cpu_ids __read_mostly = NR_CPUS;
5409 EXPORT_SYMBOL(nr_cpu_ids);
5411 #undef SCHED_DOMAIN_DEBUG
5412 #ifdef SCHED_DOMAIN_DEBUG
5413 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5418 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5422 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5427 struct sched_group *group = sd->groups;
5428 cpumask_t groupmask;
5430 cpumask_scnprintf(str, NR_CPUS, sd->span);
5431 cpus_clear(groupmask);
5434 for (i = 0; i < level + 1; i++)
5436 printk("domain %d: ", level);
5438 if (!(sd->flags & SD_LOAD_BALANCE)) {
5439 printk("does not load-balance\n");
5441 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5446 printk("span %s\n", str);
5448 if (!cpu_isset(cpu, sd->span))
5449 printk(KERN_ERR "ERROR: domain->span does not contain "
5451 if (!cpu_isset(cpu, group->cpumask))
5452 printk(KERN_ERR "ERROR: domain->groups does not contain"
5456 for (i = 0; i < level + 2; i++)
5462 printk(KERN_ERR "ERROR: group is NULL\n");
5466 if (!group->__cpu_power) {
5468 printk(KERN_ERR "ERROR: domain->cpu_power not "
5472 if (!cpus_weight(group->cpumask)) {
5474 printk(KERN_ERR "ERROR: empty group\n");
5477 if (cpus_intersects(groupmask, group->cpumask)) {
5479 printk(KERN_ERR "ERROR: repeated CPUs\n");
5482 cpus_or(groupmask, groupmask, group->cpumask);
5484 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5487 group = group->next;
5488 } while (group != sd->groups);
5491 if (!cpus_equal(sd->span, groupmask))
5492 printk(KERN_ERR "ERROR: groups don't span "
5500 if (!cpus_subset(groupmask, sd->span))
5501 printk(KERN_ERR "ERROR: parent span is not a superset "
5502 "of domain->span\n");
5507 # define sched_domain_debug(sd, cpu) do { } while (0)
5510 static int sd_degenerate(struct sched_domain *sd)
5512 if (cpus_weight(sd->span) == 1)
5515 /* Following flags need at least 2 groups */
5516 if (sd->flags & (SD_LOAD_BALANCE |
5517 SD_BALANCE_NEWIDLE |
5521 SD_SHARE_PKG_RESOURCES)) {
5522 if (sd->groups != sd->groups->next)
5526 /* Following flags don't use groups */
5527 if (sd->flags & (SD_WAKE_IDLE |
5536 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5538 unsigned long cflags = sd->flags, pflags = parent->flags;
5540 if (sd_degenerate(parent))
5543 if (!cpus_equal(sd->span, parent->span))
5546 /* Does parent contain flags not in child? */
5547 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5548 if (cflags & SD_WAKE_AFFINE)
5549 pflags &= ~SD_WAKE_BALANCE;
5550 /* Flags needing groups don't count if only 1 group in parent */
5551 if (parent->groups == parent->groups->next) {
5552 pflags &= ~(SD_LOAD_BALANCE |
5553 SD_BALANCE_NEWIDLE |
5557 SD_SHARE_PKG_RESOURCES);
5559 if (~cflags & pflags)
5566 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5567 * hold the hotplug lock.
5569 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5571 struct rq *rq = cpu_rq(cpu);
5572 struct sched_domain *tmp;
5574 /* Remove the sched domains which do not contribute to scheduling. */
5575 for (tmp = sd; tmp; tmp = tmp->parent) {
5576 struct sched_domain *parent = tmp->parent;
5579 if (sd_parent_degenerate(tmp, parent)) {
5580 tmp->parent = parent->parent;
5582 parent->parent->child = tmp;
5586 if (sd && sd_degenerate(sd)) {
5592 sched_domain_debug(sd, cpu);
5594 rcu_assign_pointer(rq->sd, sd);
5597 /* cpus with isolated domains */
5598 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5600 /* Setup the mask of cpus configured for isolated domains */
5601 static int __init isolated_cpu_setup(char *str)
5603 int ints[NR_CPUS], i;
5605 str = get_options(str, ARRAY_SIZE(ints), ints);
5606 cpus_clear(cpu_isolated_map);
5607 for (i = 1; i <= ints[0]; i++)
5608 if (ints[i] < NR_CPUS)
5609 cpu_set(ints[i], cpu_isolated_map);
5613 __setup ("isolcpus=", isolated_cpu_setup);
5616 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5617 * to a function which identifies what group(along with sched group) a CPU
5618 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5619 * (due to the fact that we keep track of groups covered with a cpumask_t).
5621 * init_sched_build_groups will build a circular linked list of the groups
5622 * covered by the given span, and will set each group's ->cpumask correctly,
5623 * and ->cpu_power to 0.
5626 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5627 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5628 struct sched_group **sg))
5630 struct sched_group *first = NULL, *last = NULL;
5631 cpumask_t covered = CPU_MASK_NONE;
5634 for_each_cpu_mask(i, span) {
5635 struct sched_group *sg;
5636 int group = group_fn(i, cpu_map, &sg);
5639 if (cpu_isset(i, covered))
5642 sg->cpumask = CPU_MASK_NONE;
5643 sg->__cpu_power = 0;
5645 for_each_cpu_mask(j, span) {
5646 if (group_fn(j, cpu_map, NULL) != group)
5649 cpu_set(j, covered);
5650 cpu_set(j, sg->cpumask);
5661 #define SD_NODES_PER_DOMAIN 16
5666 * find_next_best_node - find the next node to include in a sched_domain
5667 * @node: node whose sched_domain we're building
5668 * @used_nodes: nodes already in the sched_domain
5670 * Find the next node to include in a given scheduling domain. Simply
5671 * finds the closest node not already in the @used_nodes map.
5673 * Should use nodemask_t.
5675 static int find_next_best_node(int node, unsigned long *used_nodes)
5677 int i, n, val, min_val, best_node = 0;
5681 for (i = 0; i < MAX_NUMNODES; i++) {
5682 /* Start at @node */
5683 n = (node + i) % MAX_NUMNODES;
5685 if (!nr_cpus_node(n))
5688 /* Skip already used nodes */
5689 if (test_bit(n, used_nodes))
5692 /* Simple min distance search */
5693 val = node_distance(node, n);
5695 if (val < min_val) {
5701 set_bit(best_node, used_nodes);
5706 * sched_domain_node_span - get a cpumask for a node's sched_domain
5707 * @node: node whose cpumask we're constructing
5708 * @size: number of nodes to include in this span
5710 * Given a node, construct a good cpumask for its sched_domain to span. It
5711 * should be one that prevents unnecessary balancing, but also spreads tasks
5714 static cpumask_t sched_domain_node_span(int node)
5716 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5717 cpumask_t span, nodemask;
5721 bitmap_zero(used_nodes, MAX_NUMNODES);
5723 nodemask = node_to_cpumask(node);
5724 cpus_or(span, span, nodemask);
5725 set_bit(node, used_nodes);
5727 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5728 int next_node = find_next_best_node(node, used_nodes);
5730 nodemask = node_to_cpumask(next_node);
5731 cpus_or(span, span, nodemask);
5738 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5741 * SMT sched-domains:
5743 #ifdef CONFIG_SCHED_SMT
5744 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5745 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5747 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5748 struct sched_group **sg)
5751 *sg = &per_cpu(sched_group_cpus, cpu);
5757 * multi-core sched-domains:
5759 #ifdef CONFIG_SCHED_MC
5760 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5761 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5764 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5765 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5766 struct sched_group **sg)
5769 cpumask_t mask = cpu_sibling_map[cpu];
5770 cpus_and(mask, mask, *cpu_map);
5771 group = first_cpu(mask);
5773 *sg = &per_cpu(sched_group_core, group);
5776 #elif defined(CONFIG_SCHED_MC)
5777 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5778 struct sched_group **sg)
5781 *sg = &per_cpu(sched_group_core, cpu);
5786 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5787 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5789 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5790 struct sched_group **sg)
5793 #ifdef CONFIG_SCHED_MC
5794 cpumask_t mask = cpu_coregroup_map(cpu);
5795 cpus_and(mask, mask, *cpu_map);
5796 group = first_cpu(mask);
5797 #elif defined(CONFIG_SCHED_SMT)
5798 cpumask_t mask = cpu_sibling_map[cpu];
5799 cpus_and(mask, mask, *cpu_map);
5800 group = first_cpu(mask);
5805 *sg = &per_cpu(sched_group_phys, group);
5811 * The init_sched_build_groups can't handle what we want to do with node
5812 * groups, so roll our own. Now each node has its own list of groups which
5813 * gets dynamically allocated.
5815 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5816 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5818 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5819 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5821 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5822 struct sched_group **sg)
5824 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5827 cpus_and(nodemask, nodemask, *cpu_map);
5828 group = first_cpu(nodemask);
5831 *sg = &per_cpu(sched_group_allnodes, group);
5835 static void init_numa_sched_groups_power(struct sched_group *group_head)
5837 struct sched_group *sg = group_head;
5843 for_each_cpu_mask(j, sg->cpumask) {
5844 struct sched_domain *sd;
5846 sd = &per_cpu(phys_domains, j);
5847 if (j != first_cpu(sd->groups->cpumask)) {
5849 * Only add "power" once for each
5855 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5858 if (sg != group_head)
5864 /* Free memory allocated for various sched_group structures */
5865 static void free_sched_groups(const cpumask_t *cpu_map)
5869 for_each_cpu_mask(cpu, *cpu_map) {
5870 struct sched_group **sched_group_nodes
5871 = sched_group_nodes_bycpu[cpu];
5873 if (!sched_group_nodes)
5876 for (i = 0; i < MAX_NUMNODES; i++) {
5877 cpumask_t nodemask = node_to_cpumask(i);
5878 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5880 cpus_and(nodemask, nodemask, *cpu_map);
5881 if (cpus_empty(nodemask))
5891 if (oldsg != sched_group_nodes[i])
5894 kfree(sched_group_nodes);
5895 sched_group_nodes_bycpu[cpu] = NULL;
5899 static void free_sched_groups(const cpumask_t *cpu_map)
5905 * Initialize sched groups cpu_power.
5907 * cpu_power indicates the capacity of sched group, which is used while
5908 * distributing the load between different sched groups in a sched domain.
5909 * Typically cpu_power for all the groups in a sched domain will be same unless
5910 * there are asymmetries in the topology. If there are asymmetries, group
5911 * having more cpu_power will pickup more load compared to the group having
5914 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5915 * the maximum number of tasks a group can handle in the presence of other idle
5916 * or lightly loaded groups in the same sched domain.
5918 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5920 struct sched_domain *child;
5921 struct sched_group *group;
5923 WARN_ON(!sd || !sd->groups);
5925 if (cpu != first_cpu(sd->groups->cpumask))
5930 sd->groups->__cpu_power = 0;
5933 * For perf policy, if the groups in child domain share resources
5934 * (for example cores sharing some portions of the cache hierarchy
5935 * or SMT), then set this domain groups cpu_power such that each group
5936 * can handle only one task, when there are other idle groups in the
5937 * same sched domain.
5939 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5941 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5942 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5947 * add cpu_power of each child group to this groups cpu_power
5949 group = child->groups;
5951 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5952 group = group->next;
5953 } while (group != child->groups);
5957 * Build sched domains for a given set of cpus and attach the sched domains
5958 * to the individual cpus
5960 static int build_sched_domains(const cpumask_t *cpu_map)
5964 struct sched_group **sched_group_nodes = NULL;
5965 int sd_allnodes = 0;
5968 * Allocate the per-node list of sched groups
5970 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
5972 if (!sched_group_nodes) {
5973 printk(KERN_WARNING "Can not alloc sched group node list\n");
5976 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5980 * Set up domains for cpus specified by the cpu_map.
5982 for_each_cpu_mask(i, *cpu_map) {
5983 struct sched_domain *sd = NULL, *p;
5984 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5986 cpus_and(nodemask, nodemask, *cpu_map);
5989 if (cpus_weight(*cpu_map) >
5990 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5991 sd = &per_cpu(allnodes_domains, i);
5992 *sd = SD_ALLNODES_INIT;
5993 sd->span = *cpu_map;
5994 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6000 sd = &per_cpu(node_domains, i);
6002 sd->span = sched_domain_node_span(cpu_to_node(i));
6006 cpus_and(sd->span, sd->span, *cpu_map);
6010 sd = &per_cpu(phys_domains, i);
6012 sd->span = nodemask;
6016 cpu_to_phys_group(i, cpu_map, &sd->groups);
6018 #ifdef CONFIG_SCHED_MC
6020 sd = &per_cpu(core_domains, i);
6022 sd->span = cpu_coregroup_map(i);
6023 cpus_and(sd->span, sd->span, *cpu_map);
6026 cpu_to_core_group(i, cpu_map, &sd->groups);
6029 #ifdef CONFIG_SCHED_SMT
6031 sd = &per_cpu(cpu_domains, i);
6032 *sd = SD_SIBLING_INIT;
6033 sd->span = cpu_sibling_map[i];
6034 cpus_and(sd->span, sd->span, *cpu_map);
6037 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6041 #ifdef CONFIG_SCHED_SMT
6042 /* Set up CPU (sibling) groups */
6043 for_each_cpu_mask(i, *cpu_map) {
6044 cpumask_t this_sibling_map = cpu_sibling_map[i];
6045 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6046 if (i != first_cpu(this_sibling_map))
6049 init_sched_build_groups(this_sibling_map, cpu_map,
6054 #ifdef CONFIG_SCHED_MC
6055 /* Set up multi-core groups */
6056 for_each_cpu_mask(i, *cpu_map) {
6057 cpumask_t this_core_map = cpu_coregroup_map(i);
6058 cpus_and(this_core_map, this_core_map, *cpu_map);
6059 if (i != first_cpu(this_core_map))
6061 init_sched_build_groups(this_core_map, cpu_map,
6062 &cpu_to_core_group);
6066 /* Set up physical groups */
6067 for (i = 0; i < MAX_NUMNODES; i++) {
6068 cpumask_t nodemask = node_to_cpumask(i);
6070 cpus_and(nodemask, nodemask, *cpu_map);
6071 if (cpus_empty(nodemask))
6074 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6078 /* Set up node groups */
6080 init_sched_build_groups(*cpu_map, cpu_map,
6081 &cpu_to_allnodes_group);
6083 for (i = 0; i < MAX_NUMNODES; i++) {
6084 /* Set up node groups */
6085 struct sched_group *sg, *prev;
6086 cpumask_t nodemask = node_to_cpumask(i);
6087 cpumask_t domainspan;
6088 cpumask_t covered = CPU_MASK_NONE;
6091 cpus_and(nodemask, nodemask, *cpu_map);
6092 if (cpus_empty(nodemask)) {
6093 sched_group_nodes[i] = NULL;
6097 domainspan = sched_domain_node_span(i);
6098 cpus_and(domainspan, domainspan, *cpu_map);
6100 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6102 printk(KERN_WARNING "Can not alloc domain group for "
6106 sched_group_nodes[i] = sg;
6107 for_each_cpu_mask(j, nodemask) {
6108 struct sched_domain *sd;
6110 sd = &per_cpu(node_domains, j);
6113 sg->__cpu_power = 0;
6114 sg->cpumask = nodemask;
6116 cpus_or(covered, covered, nodemask);
6119 for (j = 0; j < MAX_NUMNODES; j++) {
6120 cpumask_t tmp, notcovered;
6121 int n = (i + j) % MAX_NUMNODES;
6123 cpus_complement(notcovered, covered);
6124 cpus_and(tmp, notcovered, *cpu_map);
6125 cpus_and(tmp, tmp, domainspan);
6126 if (cpus_empty(tmp))
6129 nodemask = node_to_cpumask(n);
6130 cpus_and(tmp, tmp, nodemask);
6131 if (cpus_empty(tmp))
6134 sg = kmalloc_node(sizeof(struct sched_group),
6138 "Can not alloc domain group for node %d\n", j);
6141 sg->__cpu_power = 0;
6143 sg->next = prev->next;
6144 cpus_or(covered, covered, tmp);
6151 /* Calculate CPU power for physical packages and nodes */
6152 #ifdef CONFIG_SCHED_SMT
6153 for_each_cpu_mask(i, *cpu_map) {
6154 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6156 init_sched_groups_power(i, sd);
6159 #ifdef CONFIG_SCHED_MC
6160 for_each_cpu_mask(i, *cpu_map) {
6161 struct sched_domain *sd = &per_cpu(core_domains, i);
6163 init_sched_groups_power(i, sd);
6167 for_each_cpu_mask(i, *cpu_map) {
6168 struct sched_domain *sd = &per_cpu(phys_domains, i);
6170 init_sched_groups_power(i, sd);
6174 for (i = 0; i < MAX_NUMNODES; i++)
6175 init_numa_sched_groups_power(sched_group_nodes[i]);
6178 struct sched_group *sg;
6180 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6181 init_numa_sched_groups_power(sg);
6185 /* Attach the domains */
6186 for_each_cpu_mask(i, *cpu_map) {
6187 struct sched_domain *sd;
6188 #ifdef CONFIG_SCHED_SMT
6189 sd = &per_cpu(cpu_domains, i);
6190 #elif defined(CONFIG_SCHED_MC)
6191 sd = &per_cpu(core_domains, i);
6193 sd = &per_cpu(phys_domains, i);
6195 cpu_attach_domain(sd, i);
6202 free_sched_groups(cpu_map);
6207 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6209 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6211 cpumask_t cpu_default_map;
6215 * Setup mask for cpus without special case scheduling requirements.
6216 * For now this just excludes isolated cpus, but could be used to
6217 * exclude other special cases in the future.
6219 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6221 err = build_sched_domains(&cpu_default_map);
6226 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6228 free_sched_groups(cpu_map);
6232 * Detach sched domains from a group of cpus specified in cpu_map
6233 * These cpus will now be attached to the NULL domain
6235 static void detach_destroy_domains(const cpumask_t *cpu_map)
6239 for_each_cpu_mask(i, *cpu_map)
6240 cpu_attach_domain(NULL, i);
6241 synchronize_sched();
6242 arch_destroy_sched_domains(cpu_map);
6246 * Partition sched domains as specified by the cpumasks below.
6247 * This attaches all cpus from the cpumasks to the NULL domain,
6248 * waits for a RCU quiescent period, recalculates sched
6249 * domain information and then attaches them back to the
6250 * correct sched domains
6251 * Call with hotplug lock held
6253 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6255 cpumask_t change_map;
6258 cpus_and(*partition1, *partition1, cpu_online_map);
6259 cpus_and(*partition2, *partition2, cpu_online_map);
6260 cpus_or(change_map, *partition1, *partition2);
6262 /* Detach sched domains from all of the affected cpus */
6263 detach_destroy_domains(&change_map);
6264 if (!cpus_empty(*partition1))
6265 err = build_sched_domains(partition1);
6266 if (!err && !cpus_empty(*partition2))
6267 err = build_sched_domains(partition2);
6272 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6273 static int arch_reinit_sched_domains(void)
6277 mutex_lock(&sched_hotcpu_mutex);
6278 detach_destroy_domains(&cpu_online_map);
6279 err = arch_init_sched_domains(&cpu_online_map);
6280 mutex_unlock(&sched_hotcpu_mutex);
6285 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6289 if (buf[0] != '0' && buf[0] != '1')
6293 sched_smt_power_savings = (buf[0] == '1');
6295 sched_mc_power_savings = (buf[0] == '1');
6297 ret = arch_reinit_sched_domains();
6299 return ret ? ret : count;
6302 #ifdef CONFIG_SCHED_MC
6303 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6305 return sprintf(page, "%u\n", sched_mc_power_savings);
6307 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6308 const char *buf, size_t count)
6310 return sched_power_savings_store(buf, count, 0);
6312 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6313 sched_mc_power_savings_store);
6316 #ifdef CONFIG_SCHED_SMT
6317 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6319 return sprintf(page, "%u\n", sched_smt_power_savings);
6321 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6322 const char *buf, size_t count)
6324 return sched_power_savings_store(buf, count, 1);
6326 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6327 sched_smt_power_savings_store);
6330 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6334 #ifdef CONFIG_SCHED_SMT
6336 err = sysfs_create_file(&cls->kset.kobj,
6337 &attr_sched_smt_power_savings.attr);
6339 #ifdef CONFIG_SCHED_MC
6340 if (!err && mc_capable())
6341 err = sysfs_create_file(&cls->kset.kobj,
6342 &attr_sched_mc_power_savings.attr);
6349 * Force a reinitialization of the sched domains hierarchy. The domains
6350 * and groups cannot be updated in place without racing with the balancing
6351 * code, so we temporarily attach all running cpus to the NULL domain
6352 * which will prevent rebalancing while the sched domains are recalculated.
6354 static int update_sched_domains(struct notifier_block *nfb,
6355 unsigned long action, void *hcpu)
6358 case CPU_UP_PREPARE:
6359 case CPU_UP_PREPARE_FROZEN:
6360 case CPU_DOWN_PREPARE:
6361 case CPU_DOWN_PREPARE_FROZEN:
6362 detach_destroy_domains(&cpu_online_map);
6365 case CPU_UP_CANCELED:
6366 case CPU_UP_CANCELED_FROZEN:
6367 case CPU_DOWN_FAILED:
6368 case CPU_DOWN_FAILED_FROZEN:
6370 case CPU_ONLINE_FROZEN:
6372 case CPU_DEAD_FROZEN:
6374 * Fall through and re-initialise the domains.
6381 /* The hotplug lock is already held by cpu_up/cpu_down */
6382 arch_init_sched_domains(&cpu_online_map);
6387 void __init sched_init_smp(void)
6389 cpumask_t non_isolated_cpus;
6391 mutex_lock(&sched_hotcpu_mutex);
6392 arch_init_sched_domains(&cpu_online_map);
6393 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6394 if (cpus_empty(non_isolated_cpus))
6395 cpu_set(smp_processor_id(), non_isolated_cpus);
6396 mutex_unlock(&sched_hotcpu_mutex);
6397 /* XXX: Theoretical race here - CPU may be hotplugged now */
6398 hotcpu_notifier(update_sched_domains, 0);
6400 init_sched_domain_sysctl();
6402 /* Move init over to a non-isolated CPU */
6403 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6407 void __init sched_init_smp(void)
6410 #endif /* CONFIG_SMP */
6412 int in_sched_functions(unsigned long addr)
6414 /* Linker adds these: start and end of __sched functions */
6415 extern char __sched_text_start[], __sched_text_end[];
6417 return in_lock_functions(addr) ||
6418 (addr >= (unsigned long)__sched_text_start
6419 && addr < (unsigned long)__sched_text_end);
6422 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6424 cfs_rq->tasks_timeline = RB_ROOT;
6425 cfs_rq->fair_clock = 1;
6426 #ifdef CONFIG_FAIR_GROUP_SCHED
6431 void __init sched_init(void)
6433 int highest_cpu = 0;
6437 * Link up the scheduling class hierarchy:
6439 rt_sched_class.next = &fair_sched_class;
6440 fair_sched_class.next = &idle_sched_class;
6441 idle_sched_class.next = NULL;
6443 for_each_possible_cpu(i) {
6444 struct rt_prio_array *array;
6448 spin_lock_init(&rq->lock);
6449 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6452 init_cfs_rq(&rq->cfs, rq);
6453 #ifdef CONFIG_FAIR_GROUP_SCHED
6454 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6455 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6458 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6459 rq->cpu_load[j] = 0;
6462 rq->active_balance = 0;
6463 rq->next_balance = jiffies;
6466 rq->migration_thread = NULL;
6467 INIT_LIST_HEAD(&rq->migration_queue);
6469 atomic_set(&rq->nr_iowait, 0);
6471 array = &rq->rt.active;
6472 for (j = 0; j < MAX_RT_PRIO; j++) {
6473 INIT_LIST_HEAD(array->queue + j);
6474 __clear_bit(j, array->bitmap);
6477 /* delimiter for bitsearch: */
6478 __set_bit(MAX_RT_PRIO, array->bitmap);
6481 set_load_weight(&init_task);
6483 #ifdef CONFIG_PREEMPT_NOTIFIERS
6484 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6488 nr_cpu_ids = highest_cpu + 1;
6489 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6492 #ifdef CONFIG_RT_MUTEXES
6493 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6497 * The boot idle thread does lazy MMU switching as well:
6499 atomic_inc(&init_mm.mm_count);
6500 enter_lazy_tlb(&init_mm, current);
6503 * Make us the idle thread. Technically, schedule() should not be
6504 * called from this thread, however somewhere below it might be,
6505 * but because we are the idle thread, we just pick up running again
6506 * when this runqueue becomes "idle".
6508 init_idle(current, smp_processor_id());
6510 * During early bootup we pretend to be a normal task:
6512 current->sched_class = &fair_sched_class;
6515 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6516 void __might_sleep(char *file, int line)
6519 static unsigned long prev_jiffy; /* ratelimiting */
6521 if ((in_atomic() || irqs_disabled()) &&
6522 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6523 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6525 prev_jiffy = jiffies;
6526 printk(KERN_ERR "BUG: sleeping function called from invalid"
6527 " context at %s:%d\n", file, line);
6528 printk("in_atomic():%d, irqs_disabled():%d\n",
6529 in_atomic(), irqs_disabled());
6530 debug_show_held_locks(current);
6531 if (irqs_disabled())
6532 print_irqtrace_events(current);
6537 EXPORT_SYMBOL(__might_sleep);
6540 #ifdef CONFIG_MAGIC_SYSRQ
6541 void normalize_rt_tasks(void)
6543 struct task_struct *g, *p;
6544 unsigned long flags;
6548 read_lock_irq(&tasklist_lock);
6549 do_each_thread(g, p) {
6551 p->se.wait_runtime = 0;
6552 p->se.exec_start = 0;
6553 p->se.wait_start_fair = 0;
6554 p->se.sleep_start_fair = 0;
6555 #ifdef CONFIG_SCHEDSTATS
6556 p->se.wait_start = 0;
6557 p->se.sleep_start = 0;
6558 p->se.block_start = 0;
6560 task_rq(p)->cfs.fair_clock = 0;
6561 task_rq(p)->clock = 0;
6565 * Renice negative nice level userspace
6568 if (TASK_NICE(p) < 0 && p->mm)
6569 set_user_nice(p, 0);
6573 spin_lock_irqsave(&p->pi_lock, flags);
6574 rq = __task_rq_lock(p);
6577 * Do not touch the migration thread:
6579 if (p == rq->migration_thread)
6583 update_rq_clock(rq);
6584 on_rq = p->se.on_rq;
6586 deactivate_task(rq, p, 0);
6587 __setscheduler(rq, p, SCHED_NORMAL, 0);
6589 activate_task(rq, p, 0);
6590 resched_task(rq->curr);
6595 __task_rq_unlock(rq);
6596 spin_unlock_irqrestore(&p->pi_lock, flags);
6597 } while_each_thread(g, p);
6599 read_unlock_irq(&tasklist_lock);
6602 #endif /* CONFIG_MAGIC_SYSRQ */
6606 * These functions are only useful for the IA64 MCA handling.
6608 * They can only be called when the whole system has been
6609 * stopped - every CPU needs to be quiescent, and no scheduling
6610 * activity can take place. Using them for anything else would
6611 * be a serious bug, and as a result, they aren't even visible
6612 * under any other configuration.
6616 * curr_task - return the current task for a given cpu.
6617 * @cpu: the processor in question.
6619 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6621 struct task_struct *curr_task(int cpu)
6623 return cpu_curr(cpu);
6627 * set_curr_task - set the current task for a given cpu.
6628 * @cpu: the processor in question.
6629 * @p: the task pointer to set.
6631 * Description: This function must only be used when non-maskable interrupts
6632 * are serviced on a separate stack. It allows the architecture to switch the
6633 * notion of the current task on a cpu in a non-blocking manner. This function
6634 * must be called with all CPU's synchronized, and interrupts disabled, the
6635 * and caller must save the original value of the current task (see
6636 * curr_task() above) and restore that value before reenabling interrupts and
6637 * re-starting the system.
6639 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6641 void set_curr_task(int cpu, struct task_struct *p)