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) ((unsigned long)(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 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
116 * Since cpu_power is a 'constant', we can use a reciprocal divide.
118 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
120 return reciprocal_divide(load, sg->reciprocal_cpu_power);
124 * Each time a sched group cpu_power is changed,
125 * we must compute its reciprocal value
127 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
129 sg->__cpu_power += val;
130 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
134 static inline int rt_policy(int policy)
136 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
141 static inline int task_has_rt_policy(struct task_struct *p)
143 return rt_policy(p->policy);
147 * This is the priority-queue data structure of the RT scheduling class:
149 struct rt_prio_array {
150 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
151 struct list_head queue[MAX_RT_PRIO];
154 #ifdef CONFIG_FAIR_GROUP_SCHED
158 /* task group related information */
160 /* schedulable entities of this group on each cpu */
161 struct sched_entity **se;
162 /* runqueue "owned" by this group on each cpu */
163 struct cfs_rq **cfs_rq;
164 unsigned long shares;
165 /* spinlock to serialize modification to shares */
169 /* Default task group's sched entity on each cpu */
170 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
171 /* Default task group's cfs_rq on each cpu */
172 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
174 static struct sched_entity *init_sched_entity_p[NR_CPUS];
175 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
177 /* Default task group.
178 * Every task in system belong to this group at bootup.
180 struct task_group init_task_group = {
181 .se = init_sched_entity_p,
182 .cfs_rq = init_cfs_rq_p,
185 #ifdef CONFIG_FAIR_USER_SCHED
186 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
188 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
191 static int init_task_group_load = INIT_TASK_GRP_LOAD;
193 /* return group to which a task belongs */
194 static inline struct task_group *task_group(struct task_struct *p)
196 struct task_group *tg;
198 #ifdef CONFIG_FAIR_USER_SCHED
201 tg = &init_task_group;
207 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
208 static inline void set_task_cfs_rq(struct task_struct *p)
210 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
211 p->se.parent = task_group(p)->se[task_cpu(p)];
216 static inline void set_task_cfs_rq(struct task_struct *p) { }
218 #endif /* CONFIG_FAIR_GROUP_SCHED */
220 /* CFS-related fields in a runqueue */
222 struct load_weight load;
223 unsigned long nr_running;
228 struct rb_root tasks_timeline;
229 struct rb_node *rb_leftmost;
230 struct rb_node *rb_load_balance_curr;
231 /* 'curr' points to currently running entity on this cfs_rq.
232 * It is set to NULL otherwise (i.e when none are currently running).
234 struct sched_entity *curr;
236 unsigned long nr_spread_over;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
241 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
242 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
243 * (like users, containers etc.)
245 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
246 * list is used during load balance.
248 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
249 struct task_group *tg; /* group that "owns" this runqueue */
254 /* Real-Time classes' related field in a runqueue: */
256 struct rt_prio_array active;
257 int rt_load_balance_idx;
258 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
262 * This is the main, per-CPU runqueue data structure.
264 * Locking rule: those places that want to lock multiple runqueues
265 * (such as the load balancing or the thread migration code), lock
266 * acquire operations must be ordered by ascending &runqueue.
269 spinlock_t lock; /* runqueue lock */
272 * nr_running and cpu_load should be in the same cacheline because
273 * remote CPUs use both these fields when doing load calculation.
275 unsigned long nr_running;
276 #define CPU_LOAD_IDX_MAX 5
277 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
278 unsigned char idle_at_tick;
280 unsigned char in_nohz_recently;
282 struct load_weight load; /* capture load from *all* tasks on this cpu */
283 unsigned long nr_load_updates;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
293 * This is part of a global counter where only the total sum
294 * over all CPUs matters. A task can increase this counter on
295 * one CPU and if it got migrated afterwards it may decrease
296 * it on another CPU. Always updated under the runqueue lock:
298 unsigned long nr_uninterruptible;
300 struct task_struct *curr, *idle;
301 unsigned long next_balance;
302 struct mm_struct *prev_mm;
304 u64 clock, prev_clock_raw;
307 unsigned int clock_warps, clock_overflows;
309 unsigned int clock_deep_idle_events;
315 struct sched_domain *sd;
317 /* For active balancing */
320 int cpu; /* cpu of this runqueue */
322 struct task_struct *migration_thread;
323 struct list_head migration_queue;
326 #ifdef CONFIG_SCHEDSTATS
328 struct sched_info rq_sched_info;
330 /* sys_sched_yield() stats */
331 unsigned long yld_exp_empty;
332 unsigned long yld_act_empty;
333 unsigned long yld_both_empty;
334 unsigned long yld_count;
336 /* schedule() stats */
337 unsigned long sched_switch;
338 unsigned long sched_count;
339 unsigned long sched_goidle;
341 /* try_to_wake_up() stats */
342 unsigned long ttwu_count;
343 unsigned long ttwu_local;
346 unsigned long bkl_count;
348 struct lock_class_key rq_lock_key;
351 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
352 static DEFINE_MUTEX(sched_hotcpu_mutex);
354 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
356 rq->curr->sched_class->check_preempt_curr(rq, p);
359 static inline int cpu_of(struct rq *rq)
369 * Update the per-runqueue clock, as finegrained as the platform can give
370 * us, but without assuming monotonicity, etc.:
372 static void __update_rq_clock(struct rq *rq)
374 u64 prev_raw = rq->prev_clock_raw;
375 u64 now = sched_clock();
376 s64 delta = now - prev_raw;
377 u64 clock = rq->clock;
379 #ifdef CONFIG_SCHED_DEBUG
380 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
383 * Protect against sched_clock() occasionally going backwards:
385 if (unlikely(delta < 0)) {
390 * Catch too large forward jumps too:
392 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
393 if (clock < rq->tick_timestamp + TICK_NSEC)
394 clock = rq->tick_timestamp + TICK_NSEC;
397 rq->clock_overflows++;
399 if (unlikely(delta > rq->clock_max_delta))
400 rq->clock_max_delta = delta;
405 rq->prev_clock_raw = now;
409 static void update_rq_clock(struct rq *rq)
411 if (likely(smp_processor_id() == cpu_of(rq)))
412 __update_rq_clock(rq);
416 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
417 * See detach_destroy_domains: synchronize_sched for details.
419 * The domain tree of any CPU may only be accessed from within
420 * preempt-disabled sections.
422 #define for_each_domain(cpu, __sd) \
423 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
425 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
426 #define this_rq() (&__get_cpu_var(runqueues))
427 #define task_rq(p) cpu_rq(task_cpu(p))
428 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
431 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
433 #ifdef CONFIG_SCHED_DEBUG
434 # define const_debug __read_mostly
436 # define const_debug static const
440 * Debugging: various feature bits
443 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
444 SCHED_FEAT_START_DEBIT = 2,
445 SCHED_FEAT_TREE_AVG = 4,
446 SCHED_FEAT_APPROX_AVG = 8,
447 SCHED_FEAT_WAKEUP_PREEMPT = 16,
448 SCHED_FEAT_PREEMPT_RESTRICT = 32,
451 const_debug unsigned int sysctl_sched_features =
452 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
453 SCHED_FEAT_START_DEBIT *1 |
454 SCHED_FEAT_TREE_AVG *0 |
455 SCHED_FEAT_APPROX_AVG *0 |
456 SCHED_FEAT_WAKEUP_PREEMPT *1 |
457 SCHED_FEAT_PREEMPT_RESTRICT *1;
459 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
462 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
463 * clock constructed from sched_clock():
465 unsigned long long cpu_clock(int cpu)
467 unsigned long long now;
471 local_irq_save(flags);
475 local_irq_restore(flags);
479 EXPORT_SYMBOL_GPL(cpu_clock);
481 #ifndef prepare_arch_switch
482 # define prepare_arch_switch(next) do { } while (0)
484 #ifndef finish_arch_switch
485 # define finish_arch_switch(prev) do { } while (0)
488 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
489 static inline int task_running(struct rq *rq, struct task_struct *p)
491 return rq->curr == p;
494 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
498 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
500 #ifdef CONFIG_DEBUG_SPINLOCK
501 /* this is a valid case when another task releases the spinlock */
502 rq->lock.owner = current;
505 * If we are tracking spinlock dependencies then we have to
506 * fix up the runqueue lock - which gets 'carried over' from
509 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
511 spin_unlock_irq(&rq->lock);
514 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
515 static inline int task_running(struct rq *rq, struct task_struct *p)
520 return rq->curr == p;
524 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
528 * We can optimise this out completely for !SMP, because the
529 * SMP rebalancing from interrupt is the only thing that cares
534 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
535 spin_unlock_irq(&rq->lock);
537 spin_unlock(&rq->lock);
541 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
545 * After ->oncpu is cleared, the task can be moved to a different CPU.
546 * We must ensure this doesn't happen until the switch is completely
552 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
556 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
559 * __task_rq_lock - lock the runqueue a given task resides on.
560 * Must be called interrupts disabled.
562 static inline struct rq *__task_rq_lock(struct task_struct *p)
566 struct rq *rq = task_rq(p);
567 spin_lock(&rq->lock);
568 if (likely(rq == task_rq(p)))
570 spin_unlock(&rq->lock);
575 * task_rq_lock - lock the runqueue a given task resides on and disable
576 * interrupts. Note the ordering: we can safely lookup the task_rq without
577 * explicitly disabling preemption.
579 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
585 local_irq_save(*flags);
587 spin_lock(&rq->lock);
588 if (likely(rq == task_rq(p)))
590 spin_unlock_irqrestore(&rq->lock, *flags);
594 static void __task_rq_unlock(struct rq *rq)
597 spin_unlock(&rq->lock);
600 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
603 spin_unlock_irqrestore(&rq->lock, *flags);
607 * this_rq_lock - lock this runqueue and disable interrupts.
609 static struct rq *this_rq_lock(void)
616 spin_lock(&rq->lock);
622 * We are going deep-idle (irqs are disabled):
624 void sched_clock_idle_sleep_event(void)
626 struct rq *rq = cpu_rq(smp_processor_id());
628 spin_lock(&rq->lock);
629 __update_rq_clock(rq);
630 spin_unlock(&rq->lock);
631 rq->clock_deep_idle_events++;
633 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
636 * We just idled delta nanoseconds (called with irqs disabled):
638 void sched_clock_idle_wakeup_event(u64 delta_ns)
640 struct rq *rq = cpu_rq(smp_processor_id());
641 u64 now = sched_clock();
643 rq->idle_clock += delta_ns;
645 * Override the previous timestamp and ignore all
646 * sched_clock() deltas that occured while we idled,
647 * and use the PM-provided delta_ns to advance the
650 spin_lock(&rq->lock);
651 rq->prev_clock_raw = now;
652 rq->clock += delta_ns;
653 spin_unlock(&rq->lock);
655 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
658 * resched_task - mark a task 'to be rescheduled now'.
660 * On UP this means the setting of the need_resched flag, on SMP it
661 * might also involve a cross-CPU call to trigger the scheduler on
666 #ifndef tsk_is_polling
667 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
670 static void resched_task(struct task_struct *p)
674 assert_spin_locked(&task_rq(p)->lock);
676 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
679 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
682 if (cpu == smp_processor_id())
685 /* NEED_RESCHED must be visible before we test polling */
687 if (!tsk_is_polling(p))
688 smp_send_reschedule(cpu);
691 static void resched_cpu(int cpu)
693 struct rq *rq = cpu_rq(cpu);
696 if (!spin_trylock_irqsave(&rq->lock, flags))
698 resched_task(cpu_curr(cpu));
699 spin_unlock_irqrestore(&rq->lock, flags);
702 static inline void resched_task(struct task_struct *p)
704 assert_spin_locked(&task_rq(p)->lock);
705 set_tsk_need_resched(p);
709 #if BITS_PER_LONG == 32
710 # define WMULT_CONST (~0UL)
712 # define WMULT_CONST (1UL << 32)
715 #define WMULT_SHIFT 32
718 * Shift right and round:
720 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
723 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
724 struct load_weight *lw)
728 if (unlikely(!lw->inv_weight))
729 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
731 tmp = (u64)delta_exec * weight;
733 * Check whether we'd overflow the 64-bit multiplication:
735 if (unlikely(tmp > WMULT_CONST))
736 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
739 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
741 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
744 static inline unsigned long
745 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
747 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
750 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
755 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
761 * To aid in avoiding the subversion of "niceness" due to uneven distribution
762 * of tasks with abnormal "nice" values across CPUs the contribution that
763 * each task makes to its run queue's load is weighted according to its
764 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
765 * scaled version of the new time slice allocation that they receive on time
769 #define WEIGHT_IDLEPRIO 2
770 #define WMULT_IDLEPRIO (1 << 31)
773 * Nice levels are multiplicative, with a gentle 10% change for every
774 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
775 * nice 1, it will get ~10% less CPU time than another CPU-bound task
776 * that remained on nice 0.
778 * The "10% effect" is relative and cumulative: from _any_ nice level,
779 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
780 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
781 * If a task goes up by ~10% and another task goes down by ~10% then
782 * the relative distance between them is ~25%.)
784 static const int prio_to_weight[40] = {
785 /* -20 */ 88761, 71755, 56483, 46273, 36291,
786 /* -15 */ 29154, 23254, 18705, 14949, 11916,
787 /* -10 */ 9548, 7620, 6100, 4904, 3906,
788 /* -5 */ 3121, 2501, 1991, 1586, 1277,
789 /* 0 */ 1024, 820, 655, 526, 423,
790 /* 5 */ 335, 272, 215, 172, 137,
791 /* 10 */ 110, 87, 70, 56, 45,
792 /* 15 */ 36, 29, 23, 18, 15,
796 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
798 * In cases where the weight does not change often, we can use the
799 * precalculated inverse to speed up arithmetics by turning divisions
800 * into multiplications:
802 static const u32 prio_to_wmult[40] = {
803 /* -20 */ 48388, 59856, 76040, 92818, 118348,
804 /* -15 */ 147320, 184698, 229616, 287308, 360437,
805 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
806 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
807 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
808 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
809 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
810 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
813 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
816 * runqueue iterator, to support SMP load-balancing between different
817 * scheduling classes, without having to expose their internal data
818 * structures to the load-balancing proper:
822 struct task_struct *(*start)(void *);
823 struct task_struct *(*next)(void *);
826 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
827 unsigned long max_nr_move, unsigned long max_load_move,
828 struct sched_domain *sd, enum cpu_idle_type idle,
829 int *all_pinned, unsigned long *load_moved,
830 int *this_best_prio, struct rq_iterator *iterator);
832 #include "sched_stats.h"
833 #include "sched_idletask.c"
834 #include "sched_fair.c"
835 #include "sched_rt.c"
836 #ifdef CONFIG_SCHED_DEBUG
837 # include "sched_debug.c"
840 #define sched_class_highest (&rt_sched_class)
843 * Update delta_exec, delta_fair fields for rq.
845 * delta_fair clock advances at a rate inversely proportional to
846 * total load (rq->load.weight) on the runqueue, while
847 * delta_exec advances at the same rate as wall-clock (provided
850 * delta_exec / delta_fair is a measure of the (smoothened) load on this
851 * runqueue over any given interval. This (smoothened) load is used
852 * during load balance.
854 * This function is called /before/ updating rq->load
855 * and when switching tasks.
857 static inline void inc_load(struct rq *rq, const struct task_struct *p)
859 update_load_add(&rq->load, p->se.load.weight);
862 static inline void dec_load(struct rq *rq, const struct task_struct *p)
864 update_load_sub(&rq->load, p->se.load.weight);
867 static void inc_nr_running(struct task_struct *p, struct rq *rq)
873 static void dec_nr_running(struct task_struct *p, struct rq *rq)
879 static void set_load_weight(struct task_struct *p)
881 if (task_has_rt_policy(p)) {
882 p->se.load.weight = prio_to_weight[0] * 2;
883 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
888 * SCHED_IDLE tasks get minimal weight:
890 if (p->policy == SCHED_IDLE) {
891 p->se.load.weight = WEIGHT_IDLEPRIO;
892 p->se.load.inv_weight = WMULT_IDLEPRIO;
896 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
897 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
900 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
902 sched_info_queued(p);
903 p->sched_class->enqueue_task(rq, p, wakeup);
907 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
909 p->sched_class->dequeue_task(rq, p, sleep);
914 * __normal_prio - return the priority that is based on the static prio
916 static inline int __normal_prio(struct task_struct *p)
918 return p->static_prio;
922 * Calculate the expected normal priority: i.e. priority
923 * without taking RT-inheritance into account. Might be
924 * boosted by interactivity modifiers. Changes upon fork,
925 * setprio syscalls, and whenever the interactivity
926 * estimator recalculates.
928 static inline int normal_prio(struct task_struct *p)
932 if (task_has_rt_policy(p))
933 prio = MAX_RT_PRIO-1 - p->rt_priority;
935 prio = __normal_prio(p);
940 * Calculate the current priority, i.e. the priority
941 * taken into account by the scheduler. This value might
942 * be boosted by RT tasks, or might be boosted by
943 * interactivity modifiers. Will be RT if the task got
944 * RT-boosted. If not then it returns p->normal_prio.
946 static int effective_prio(struct task_struct *p)
948 p->normal_prio = normal_prio(p);
950 * If we are RT tasks or we were boosted to RT priority,
951 * keep the priority unchanged. Otherwise, update priority
952 * to the normal priority:
954 if (!rt_prio(p->prio))
955 return p->normal_prio;
960 * activate_task - move a task to the runqueue.
962 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
964 if (p->state == TASK_UNINTERRUPTIBLE)
965 rq->nr_uninterruptible--;
967 enqueue_task(rq, p, wakeup);
968 inc_nr_running(p, rq);
972 * deactivate_task - remove a task from the runqueue.
974 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
976 if (p->state == TASK_UNINTERRUPTIBLE)
977 rq->nr_uninterruptible++;
979 dequeue_task(rq, p, sleep);
980 dec_nr_running(p, rq);
984 * task_curr - is this task currently executing on a CPU?
985 * @p: the task in question.
987 inline int task_curr(const struct task_struct *p)
989 return cpu_curr(task_cpu(p)) == p;
992 /* Used instead of source_load when we know the type == 0 */
993 unsigned long weighted_cpuload(const int cpu)
995 return cpu_rq(cpu)->load.weight;
998 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1001 task_thread_info(p)->cpu = cpu;
1008 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1010 int old_cpu = task_cpu(p);
1011 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1012 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1013 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1016 clock_offset = old_rq->clock - new_rq->clock;
1018 #ifdef CONFIG_SCHEDSTATS
1019 if (p->se.wait_start)
1020 p->se.wait_start -= clock_offset;
1021 if (p->se.sleep_start)
1022 p->se.sleep_start -= clock_offset;
1023 if (p->se.block_start)
1024 p->se.block_start -= clock_offset;
1026 p->se.vruntime -= old_cfsrq->min_vruntime -
1027 new_cfsrq->min_vruntime;
1029 __set_task_cpu(p, new_cpu);
1032 struct migration_req {
1033 struct list_head list;
1035 struct task_struct *task;
1038 struct completion done;
1042 * The task's runqueue lock must be held.
1043 * Returns true if you have to wait for migration thread.
1046 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1048 struct rq *rq = task_rq(p);
1051 * If the task is not on a runqueue (and not running), then
1052 * it is sufficient to simply update the task's cpu field.
1054 if (!p->se.on_rq && !task_running(rq, p)) {
1055 set_task_cpu(p, dest_cpu);
1059 init_completion(&req->done);
1061 req->dest_cpu = dest_cpu;
1062 list_add(&req->list, &rq->migration_queue);
1068 * wait_task_inactive - wait for a thread to unschedule.
1070 * The caller must ensure that the task *will* unschedule sometime soon,
1071 * else this function might spin for a *long* time. This function can't
1072 * be called with interrupts off, or it may introduce deadlock with
1073 * smp_call_function() if an IPI is sent by the same process we are
1074 * waiting to become inactive.
1076 void wait_task_inactive(struct task_struct *p)
1078 unsigned long flags;
1084 * We do the initial early heuristics without holding
1085 * any task-queue locks at all. We'll only try to get
1086 * the runqueue lock when things look like they will
1092 * If the task is actively running on another CPU
1093 * still, just relax and busy-wait without holding
1096 * NOTE! Since we don't hold any locks, it's not
1097 * even sure that "rq" stays as the right runqueue!
1098 * But we don't care, since "task_running()" will
1099 * return false if the runqueue has changed and p
1100 * is actually now running somewhere else!
1102 while (task_running(rq, p))
1106 * Ok, time to look more closely! We need the rq
1107 * lock now, to be *sure*. If we're wrong, we'll
1108 * just go back and repeat.
1110 rq = task_rq_lock(p, &flags);
1111 running = task_running(rq, p);
1112 on_rq = p->se.on_rq;
1113 task_rq_unlock(rq, &flags);
1116 * Was it really running after all now that we
1117 * checked with the proper locks actually held?
1119 * Oops. Go back and try again..
1121 if (unlikely(running)) {
1127 * It's not enough that it's not actively running,
1128 * it must be off the runqueue _entirely_, and not
1131 * So if it wa still runnable (but just not actively
1132 * running right now), it's preempted, and we should
1133 * yield - it could be a while.
1135 if (unlikely(on_rq)) {
1136 schedule_timeout_uninterruptible(1);
1141 * Ahh, all good. It wasn't running, and it wasn't
1142 * runnable, which means that it will never become
1143 * running in the future either. We're all done!
1150 * kick_process - kick a running thread to enter/exit the kernel
1151 * @p: the to-be-kicked thread
1153 * Cause a process which is running on another CPU to enter
1154 * kernel-mode, without any delay. (to get signals handled.)
1156 * NOTE: this function doesnt have to take the runqueue lock,
1157 * because all it wants to ensure is that the remote task enters
1158 * the kernel. If the IPI races and the task has been migrated
1159 * to another CPU then no harm is done and the purpose has been
1162 void kick_process(struct task_struct *p)
1168 if ((cpu != smp_processor_id()) && task_curr(p))
1169 smp_send_reschedule(cpu);
1174 * Return a low guess at the load of a migration-source cpu weighted
1175 * according to the scheduling class and "nice" value.
1177 * We want to under-estimate the load of migration sources, to
1178 * balance conservatively.
1180 static unsigned long source_load(int cpu, int type)
1182 struct rq *rq = cpu_rq(cpu);
1183 unsigned long total = weighted_cpuload(cpu);
1188 return min(rq->cpu_load[type-1], total);
1192 * Return a high guess at the load of a migration-target cpu weighted
1193 * according to the scheduling class and "nice" value.
1195 static unsigned long target_load(int cpu, int type)
1197 struct rq *rq = cpu_rq(cpu);
1198 unsigned long total = weighted_cpuload(cpu);
1203 return max(rq->cpu_load[type-1], total);
1207 * Return the average load per task on the cpu's run queue
1209 static inline unsigned long cpu_avg_load_per_task(int cpu)
1211 struct rq *rq = cpu_rq(cpu);
1212 unsigned long total = weighted_cpuload(cpu);
1213 unsigned long n = rq->nr_running;
1215 return n ? total / n : SCHED_LOAD_SCALE;
1219 * find_idlest_group finds and returns the least busy CPU group within the
1222 static struct sched_group *
1223 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1225 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1226 unsigned long min_load = ULONG_MAX, this_load = 0;
1227 int load_idx = sd->forkexec_idx;
1228 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1231 unsigned long load, avg_load;
1235 /* Skip over this group if it has no CPUs allowed */
1236 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1239 local_group = cpu_isset(this_cpu, group->cpumask);
1241 /* Tally up the load of all CPUs in the group */
1244 for_each_cpu_mask(i, group->cpumask) {
1245 /* Bias balancing toward cpus of our domain */
1247 load = source_load(i, load_idx);
1249 load = target_load(i, load_idx);
1254 /* Adjust by relative CPU power of the group */
1255 avg_load = sg_div_cpu_power(group,
1256 avg_load * SCHED_LOAD_SCALE);
1259 this_load = avg_load;
1261 } else if (avg_load < min_load) {
1262 min_load = avg_load;
1265 } while (group = group->next, group != sd->groups);
1267 if (!idlest || 100*this_load < imbalance*min_load)
1273 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1276 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1279 unsigned long load, min_load = ULONG_MAX;
1283 /* Traverse only the allowed CPUs */
1284 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1286 for_each_cpu_mask(i, tmp) {
1287 load = weighted_cpuload(i);
1289 if (load < min_load || (load == min_load && i == this_cpu)) {
1299 * sched_balance_self: balance the current task (running on cpu) in domains
1300 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1303 * Balance, ie. select the least loaded group.
1305 * Returns the target CPU number, or the same CPU if no balancing is needed.
1307 * preempt must be disabled.
1309 static int sched_balance_self(int cpu, int flag)
1311 struct task_struct *t = current;
1312 struct sched_domain *tmp, *sd = NULL;
1314 for_each_domain(cpu, tmp) {
1316 * If power savings logic is enabled for a domain, stop there.
1318 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1320 if (tmp->flags & flag)
1326 struct sched_group *group;
1327 int new_cpu, weight;
1329 if (!(sd->flags & flag)) {
1335 group = find_idlest_group(sd, t, cpu);
1341 new_cpu = find_idlest_cpu(group, t, cpu);
1342 if (new_cpu == -1 || new_cpu == cpu) {
1343 /* Now try balancing at a lower domain level of cpu */
1348 /* Now try balancing at a lower domain level of new_cpu */
1351 weight = cpus_weight(span);
1352 for_each_domain(cpu, tmp) {
1353 if (weight <= cpus_weight(tmp->span))
1355 if (tmp->flags & flag)
1358 /* while loop will break here if sd == NULL */
1364 #endif /* CONFIG_SMP */
1367 * wake_idle() will wake a task on an idle cpu if task->cpu is
1368 * not idle and an idle cpu is available. The span of cpus to
1369 * search starts with cpus closest then further out as needed,
1370 * so we always favor a closer, idle cpu.
1372 * Returns the CPU we should wake onto.
1374 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1375 static int wake_idle(int cpu, struct task_struct *p)
1378 struct sched_domain *sd;
1382 * If it is idle, then it is the best cpu to run this task.
1384 * This cpu is also the best, if it has more than one task already.
1385 * Siblings must be also busy(in most cases) as they didn't already
1386 * pickup the extra load from this cpu and hence we need not check
1387 * sibling runqueue info. This will avoid the checks and cache miss
1388 * penalities associated with that.
1390 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1393 for_each_domain(cpu, sd) {
1394 if (sd->flags & SD_WAKE_IDLE) {
1395 cpus_and(tmp, sd->span, p->cpus_allowed);
1396 for_each_cpu_mask(i, tmp) {
1407 static inline int wake_idle(int cpu, struct task_struct *p)
1414 * try_to_wake_up - wake up a thread
1415 * @p: the to-be-woken-up thread
1416 * @state: the mask of task states that can be woken
1417 * @sync: do a synchronous wakeup?
1419 * Put it on the run-queue if it's not already there. The "current"
1420 * thread is always on the run-queue (except when the actual
1421 * re-schedule is in progress), and as such you're allowed to do
1422 * the simpler "current->state = TASK_RUNNING" to mark yourself
1423 * runnable without the overhead of this.
1425 * returns failure only if the task is already active.
1427 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1429 int cpu, this_cpu, success = 0;
1430 unsigned long flags;
1434 struct sched_domain *sd, *this_sd = NULL;
1435 unsigned long load, this_load;
1439 rq = task_rq_lock(p, &flags);
1440 old_state = p->state;
1441 if (!(old_state & state))
1448 this_cpu = smp_processor_id();
1451 if (unlikely(task_running(rq, p)))
1456 schedstat_inc(rq, ttwu_count);
1457 if (cpu == this_cpu) {
1458 schedstat_inc(rq, ttwu_local);
1462 for_each_domain(this_cpu, sd) {
1463 if (cpu_isset(cpu, sd->span)) {
1464 schedstat_inc(sd, ttwu_wake_remote);
1470 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1474 * Check for affine wakeup and passive balancing possibilities.
1477 int idx = this_sd->wake_idx;
1478 unsigned int imbalance;
1480 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1482 load = source_load(cpu, idx);
1483 this_load = target_load(this_cpu, idx);
1485 new_cpu = this_cpu; /* Wake to this CPU if we can */
1487 if (this_sd->flags & SD_WAKE_AFFINE) {
1488 unsigned long tl = this_load;
1489 unsigned long tl_per_task;
1491 tl_per_task = cpu_avg_load_per_task(this_cpu);
1494 * If sync wakeup then subtract the (maximum possible)
1495 * effect of the currently running task from the load
1496 * of the current CPU:
1499 tl -= current->se.load.weight;
1502 tl + target_load(cpu, idx) <= tl_per_task) ||
1503 100*(tl + p->se.load.weight) <= imbalance*load) {
1505 * This domain has SD_WAKE_AFFINE and
1506 * p is cache cold in this domain, and
1507 * there is no bad imbalance.
1509 schedstat_inc(this_sd, ttwu_move_affine);
1515 * Start passive balancing when half the imbalance_pct
1518 if (this_sd->flags & SD_WAKE_BALANCE) {
1519 if (imbalance*this_load <= 100*load) {
1520 schedstat_inc(this_sd, ttwu_move_balance);
1526 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1528 new_cpu = wake_idle(new_cpu, p);
1529 if (new_cpu != cpu) {
1530 set_task_cpu(p, new_cpu);
1531 task_rq_unlock(rq, &flags);
1532 /* might preempt at this point */
1533 rq = task_rq_lock(p, &flags);
1534 old_state = p->state;
1535 if (!(old_state & state))
1540 this_cpu = smp_processor_id();
1545 #endif /* CONFIG_SMP */
1546 update_rq_clock(rq);
1547 activate_task(rq, p, 1);
1549 * Sync wakeups (i.e. those types of wakeups where the waker
1550 * has indicated that it will leave the CPU in short order)
1551 * don't trigger a preemption, if the woken up task will run on
1552 * this cpu. (in this case the 'I will reschedule' promise of
1553 * the waker guarantees that the freshly woken up task is going
1554 * to be considered on this CPU.)
1556 if (!sync || cpu != this_cpu)
1557 check_preempt_curr(rq, p);
1561 p->state = TASK_RUNNING;
1563 task_rq_unlock(rq, &flags);
1568 int fastcall wake_up_process(struct task_struct *p)
1570 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1571 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1573 EXPORT_SYMBOL(wake_up_process);
1575 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1577 return try_to_wake_up(p, state, 0);
1581 * Perform scheduler related setup for a newly forked process p.
1582 * p is forked by current.
1584 * __sched_fork() is basic setup used by init_idle() too:
1586 static void __sched_fork(struct task_struct *p)
1588 p->se.exec_start = 0;
1589 p->se.sum_exec_runtime = 0;
1590 p->se.prev_sum_exec_runtime = 0;
1592 #ifdef CONFIG_SCHEDSTATS
1593 p->se.wait_start = 0;
1594 p->se.sum_sleep_runtime = 0;
1595 p->se.sleep_start = 0;
1596 p->se.block_start = 0;
1597 p->se.sleep_max = 0;
1598 p->se.block_max = 0;
1600 p->se.slice_max = 0;
1604 INIT_LIST_HEAD(&p->run_list);
1607 #ifdef CONFIG_PREEMPT_NOTIFIERS
1608 INIT_HLIST_HEAD(&p->preempt_notifiers);
1612 * We mark the process as running here, but have not actually
1613 * inserted it onto the runqueue yet. This guarantees that
1614 * nobody will actually run it, and a signal or other external
1615 * event cannot wake it up and insert it on the runqueue either.
1617 p->state = TASK_RUNNING;
1621 * fork()/clone()-time setup:
1623 void sched_fork(struct task_struct *p, int clone_flags)
1625 int cpu = get_cpu();
1630 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1632 set_task_cpu(p, cpu);
1635 * Make sure we do not leak PI boosting priority to the child:
1637 p->prio = current->normal_prio;
1638 if (!rt_prio(p->prio))
1639 p->sched_class = &fair_sched_class;
1641 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1642 if (likely(sched_info_on()))
1643 memset(&p->sched_info, 0, sizeof(p->sched_info));
1645 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1648 #ifdef CONFIG_PREEMPT
1649 /* Want to start with kernel preemption disabled. */
1650 task_thread_info(p)->preempt_count = 1;
1656 * wake_up_new_task - wake up a newly created task for the first time.
1658 * This function will do some initial scheduler statistics housekeeping
1659 * that must be done for every newly created context, then puts the task
1660 * on the runqueue and wakes it.
1662 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1664 unsigned long flags;
1667 rq = task_rq_lock(p, &flags);
1668 BUG_ON(p->state != TASK_RUNNING);
1669 update_rq_clock(rq);
1671 p->prio = effective_prio(p);
1673 if (!p->sched_class->task_new || !current->se.on_rq || !rq->cfs.curr) {
1674 activate_task(rq, p, 0);
1677 * Let the scheduling class do new task startup
1678 * management (if any):
1680 p->sched_class->task_new(rq, p);
1681 inc_nr_running(p, rq);
1683 check_preempt_curr(rq, p);
1684 task_rq_unlock(rq, &flags);
1687 #ifdef CONFIG_PREEMPT_NOTIFIERS
1690 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1691 * @notifier: notifier struct to register
1693 void preempt_notifier_register(struct preempt_notifier *notifier)
1695 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1697 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1700 * preempt_notifier_unregister - no longer interested in preemption notifications
1701 * @notifier: notifier struct to unregister
1703 * This is safe to call from within a preemption notifier.
1705 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1707 hlist_del(¬ifier->link);
1709 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1711 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1713 struct preempt_notifier *notifier;
1714 struct hlist_node *node;
1716 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1717 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1721 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1722 struct task_struct *next)
1724 struct preempt_notifier *notifier;
1725 struct hlist_node *node;
1727 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1728 notifier->ops->sched_out(notifier, next);
1733 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1738 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1739 struct task_struct *next)
1746 * prepare_task_switch - prepare to switch tasks
1747 * @rq: the runqueue preparing to switch
1748 * @prev: the current task that is being switched out
1749 * @next: the task we are going to switch to.
1751 * This is called with the rq lock held and interrupts off. It must
1752 * be paired with a subsequent finish_task_switch after the context
1755 * prepare_task_switch sets up locking and calls architecture specific
1759 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1760 struct task_struct *next)
1762 fire_sched_out_preempt_notifiers(prev, next);
1763 prepare_lock_switch(rq, next);
1764 prepare_arch_switch(next);
1768 * finish_task_switch - clean up after a task-switch
1769 * @rq: runqueue associated with task-switch
1770 * @prev: the thread we just switched away from.
1772 * finish_task_switch must be called after the context switch, paired
1773 * with a prepare_task_switch call before the context switch.
1774 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1775 * and do any other architecture-specific cleanup actions.
1777 * Note that we may have delayed dropping an mm in context_switch(). If
1778 * so, we finish that here outside of the runqueue lock. (Doing it
1779 * with the lock held can cause deadlocks; see schedule() for
1782 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1783 __releases(rq->lock)
1785 struct mm_struct *mm = rq->prev_mm;
1791 * A task struct has one reference for the use as "current".
1792 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1793 * schedule one last time. The schedule call will never return, and
1794 * the scheduled task must drop that reference.
1795 * The test for TASK_DEAD must occur while the runqueue locks are
1796 * still held, otherwise prev could be scheduled on another cpu, die
1797 * there before we look at prev->state, and then the reference would
1799 * Manfred Spraul <manfred@colorfullife.com>
1801 prev_state = prev->state;
1802 finish_arch_switch(prev);
1803 finish_lock_switch(rq, prev);
1804 fire_sched_in_preempt_notifiers(current);
1807 if (unlikely(prev_state == TASK_DEAD)) {
1809 * Remove function-return probe instances associated with this
1810 * task and put them back on the free list.
1812 kprobe_flush_task(prev);
1813 put_task_struct(prev);
1818 * schedule_tail - first thing a freshly forked thread must call.
1819 * @prev: the thread we just switched away from.
1821 asmlinkage void schedule_tail(struct task_struct *prev)
1822 __releases(rq->lock)
1824 struct rq *rq = this_rq();
1826 finish_task_switch(rq, prev);
1827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1828 /* In this case, finish_task_switch does not reenable preemption */
1831 if (current->set_child_tid)
1832 put_user(current->pid, current->set_child_tid);
1836 * context_switch - switch to the new MM and the new
1837 * thread's register state.
1840 context_switch(struct rq *rq, struct task_struct *prev,
1841 struct task_struct *next)
1843 struct mm_struct *mm, *oldmm;
1845 prepare_task_switch(rq, prev, next);
1847 oldmm = prev->active_mm;
1849 * For paravirt, this is coupled with an exit in switch_to to
1850 * combine the page table reload and the switch backend into
1853 arch_enter_lazy_cpu_mode();
1855 if (unlikely(!mm)) {
1856 next->active_mm = oldmm;
1857 atomic_inc(&oldmm->mm_count);
1858 enter_lazy_tlb(oldmm, next);
1860 switch_mm(oldmm, mm, next);
1862 if (unlikely(!prev->mm)) {
1863 prev->active_mm = NULL;
1864 rq->prev_mm = oldmm;
1867 * Since the runqueue lock will be released by the next
1868 * task (which is an invalid locking op but in the case
1869 * of the scheduler it's an obvious special-case), so we
1870 * do an early lockdep release here:
1872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1873 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1876 /* Here we just switch the register state and the stack. */
1877 switch_to(prev, next, prev);
1881 * this_rq must be evaluated again because prev may have moved
1882 * CPUs since it called schedule(), thus the 'rq' on its stack
1883 * frame will be invalid.
1885 finish_task_switch(this_rq(), prev);
1889 * nr_running, nr_uninterruptible and nr_context_switches:
1891 * externally visible scheduler statistics: current number of runnable
1892 * threads, current number of uninterruptible-sleeping threads, total
1893 * number of context switches performed since bootup.
1895 unsigned long nr_running(void)
1897 unsigned long i, sum = 0;
1899 for_each_online_cpu(i)
1900 sum += cpu_rq(i)->nr_running;
1905 unsigned long nr_uninterruptible(void)
1907 unsigned long i, sum = 0;
1909 for_each_possible_cpu(i)
1910 sum += cpu_rq(i)->nr_uninterruptible;
1913 * Since we read the counters lockless, it might be slightly
1914 * inaccurate. Do not allow it to go below zero though:
1916 if (unlikely((long)sum < 0))
1922 unsigned long long nr_context_switches(void)
1925 unsigned long long sum = 0;
1927 for_each_possible_cpu(i)
1928 sum += cpu_rq(i)->nr_switches;
1933 unsigned long nr_iowait(void)
1935 unsigned long i, sum = 0;
1937 for_each_possible_cpu(i)
1938 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1943 unsigned long nr_active(void)
1945 unsigned long i, running = 0, uninterruptible = 0;
1947 for_each_online_cpu(i) {
1948 running += cpu_rq(i)->nr_running;
1949 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1952 if (unlikely((long)uninterruptible < 0))
1953 uninterruptible = 0;
1955 return running + uninterruptible;
1959 * Update rq->cpu_load[] statistics. This function is usually called every
1960 * scheduler tick (TICK_NSEC).
1962 static void update_cpu_load(struct rq *this_rq)
1964 unsigned long this_load = this_rq->load.weight;
1967 this_rq->nr_load_updates++;
1969 /* Update our load: */
1970 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1971 unsigned long old_load, new_load;
1973 /* scale is effectively 1 << i now, and >> i divides by scale */
1975 old_load = this_rq->cpu_load[i];
1976 new_load = this_load;
1978 * Round up the averaging division if load is increasing. This
1979 * prevents us from getting stuck on 9 if the load is 10, for
1982 if (new_load > old_load)
1983 new_load += scale-1;
1984 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
1991 * double_rq_lock - safely lock two runqueues
1993 * Note this does not disable interrupts like task_rq_lock,
1994 * you need to do so manually before calling.
1996 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1997 __acquires(rq1->lock)
1998 __acquires(rq2->lock)
2000 BUG_ON(!irqs_disabled());
2002 spin_lock(&rq1->lock);
2003 __acquire(rq2->lock); /* Fake it out ;) */
2006 spin_lock(&rq1->lock);
2007 spin_lock(&rq2->lock);
2009 spin_lock(&rq2->lock);
2010 spin_lock(&rq1->lock);
2013 update_rq_clock(rq1);
2014 update_rq_clock(rq2);
2018 * double_rq_unlock - safely unlock two runqueues
2020 * Note this does not restore interrupts like task_rq_unlock,
2021 * you need to do so manually after calling.
2023 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2024 __releases(rq1->lock)
2025 __releases(rq2->lock)
2027 spin_unlock(&rq1->lock);
2029 spin_unlock(&rq2->lock);
2031 __release(rq2->lock);
2035 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2037 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2038 __releases(this_rq->lock)
2039 __acquires(busiest->lock)
2040 __acquires(this_rq->lock)
2042 if (unlikely(!irqs_disabled())) {
2043 /* printk() doesn't work good under rq->lock */
2044 spin_unlock(&this_rq->lock);
2047 if (unlikely(!spin_trylock(&busiest->lock))) {
2048 if (busiest < this_rq) {
2049 spin_unlock(&this_rq->lock);
2050 spin_lock(&busiest->lock);
2051 spin_lock(&this_rq->lock);
2053 spin_lock(&busiest->lock);
2058 * If dest_cpu is allowed for this process, migrate the task to it.
2059 * This is accomplished by forcing the cpu_allowed mask to only
2060 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2061 * the cpu_allowed mask is restored.
2063 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2065 struct migration_req req;
2066 unsigned long flags;
2069 rq = task_rq_lock(p, &flags);
2070 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2071 || unlikely(cpu_is_offline(dest_cpu)))
2074 /* force the process onto the specified CPU */
2075 if (migrate_task(p, dest_cpu, &req)) {
2076 /* Need to wait for migration thread (might exit: take ref). */
2077 struct task_struct *mt = rq->migration_thread;
2079 get_task_struct(mt);
2080 task_rq_unlock(rq, &flags);
2081 wake_up_process(mt);
2082 put_task_struct(mt);
2083 wait_for_completion(&req.done);
2088 task_rq_unlock(rq, &flags);
2092 * sched_exec - execve() is a valuable balancing opportunity, because at
2093 * this point the task has the smallest effective memory and cache footprint.
2095 void sched_exec(void)
2097 int new_cpu, this_cpu = get_cpu();
2098 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2100 if (new_cpu != this_cpu)
2101 sched_migrate_task(current, new_cpu);
2105 * pull_task - move a task from a remote runqueue to the local runqueue.
2106 * Both runqueues must be locked.
2108 static void pull_task(struct rq *src_rq, struct task_struct *p,
2109 struct rq *this_rq, int this_cpu)
2111 deactivate_task(src_rq, p, 0);
2112 set_task_cpu(p, this_cpu);
2113 activate_task(this_rq, p, 0);
2115 * Note that idle threads have a prio of MAX_PRIO, for this test
2116 * to be always true for them.
2118 check_preempt_curr(this_rq, p);
2122 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2125 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2126 struct sched_domain *sd, enum cpu_idle_type idle,
2130 * We do not migrate tasks that are:
2131 * 1) running (obviously), or
2132 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2133 * 3) are cache-hot on their current CPU.
2135 if (!cpu_isset(this_cpu, p->cpus_allowed))
2139 if (task_running(rq, p))
2145 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2146 unsigned long max_nr_move, unsigned long max_load_move,
2147 struct sched_domain *sd, enum cpu_idle_type idle,
2148 int *all_pinned, unsigned long *load_moved,
2149 int *this_best_prio, struct rq_iterator *iterator)
2151 int pulled = 0, pinned = 0, skip_for_load;
2152 struct task_struct *p;
2153 long rem_load_move = max_load_move;
2155 if (max_nr_move == 0 || max_load_move == 0)
2161 * Start the load-balancing iterator:
2163 p = iterator->start(iterator->arg);
2168 * To help distribute high priority tasks accross CPUs we don't
2169 * skip a task if it will be the highest priority task (i.e. smallest
2170 * prio value) on its new queue regardless of its load weight
2172 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2173 SCHED_LOAD_SCALE_FUZZ;
2174 if ((skip_for_load && p->prio >= *this_best_prio) ||
2175 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2176 p = iterator->next(iterator->arg);
2180 pull_task(busiest, p, this_rq, this_cpu);
2182 rem_load_move -= p->se.load.weight;
2185 * We only want to steal up to the prescribed number of tasks
2186 * and the prescribed amount of weighted load.
2188 if (pulled < max_nr_move && rem_load_move > 0) {
2189 if (p->prio < *this_best_prio)
2190 *this_best_prio = p->prio;
2191 p = iterator->next(iterator->arg);
2196 * Right now, this is the only place pull_task() is called,
2197 * so we can safely collect pull_task() stats here rather than
2198 * inside pull_task().
2200 schedstat_add(sd, lb_gained[idle], pulled);
2203 *all_pinned = pinned;
2204 *load_moved = max_load_move - rem_load_move;
2209 * move_tasks tries to move up to max_load_move weighted load from busiest to
2210 * this_rq, as part of a balancing operation within domain "sd".
2211 * Returns 1 if successful and 0 otherwise.
2213 * Called with both runqueues locked.
2215 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2216 unsigned long max_load_move,
2217 struct sched_domain *sd, enum cpu_idle_type idle,
2220 const struct sched_class *class = sched_class_highest;
2221 unsigned long total_load_moved = 0;
2222 int this_best_prio = this_rq->curr->prio;
2226 class->load_balance(this_rq, this_cpu, busiest,
2227 ULONG_MAX, max_load_move - total_load_moved,
2228 sd, idle, all_pinned, &this_best_prio);
2229 class = class->next;
2230 } while (class && max_load_move > total_load_moved);
2232 return total_load_moved > 0;
2236 * move_one_task tries to move exactly one task from busiest to this_rq, as
2237 * part of active balancing operations within "domain".
2238 * Returns 1 if successful and 0 otherwise.
2240 * Called with both runqueues locked.
2242 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2243 struct sched_domain *sd, enum cpu_idle_type idle)
2245 const struct sched_class *class;
2246 int this_best_prio = MAX_PRIO;
2248 for (class = sched_class_highest; class; class = class->next)
2249 if (class->load_balance(this_rq, this_cpu, busiest,
2250 1, ULONG_MAX, sd, idle, NULL,
2258 * find_busiest_group finds and returns the busiest CPU group within the
2259 * domain. It calculates and returns the amount of weighted load which
2260 * should be moved to restore balance via the imbalance parameter.
2262 static struct sched_group *
2263 find_busiest_group(struct sched_domain *sd, int this_cpu,
2264 unsigned long *imbalance, enum cpu_idle_type idle,
2265 int *sd_idle, cpumask_t *cpus, int *balance)
2267 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2268 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2269 unsigned long max_pull;
2270 unsigned long busiest_load_per_task, busiest_nr_running;
2271 unsigned long this_load_per_task, this_nr_running;
2273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2274 int power_savings_balance = 1;
2275 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2276 unsigned long min_nr_running = ULONG_MAX;
2277 struct sched_group *group_min = NULL, *group_leader = NULL;
2280 max_load = this_load = total_load = total_pwr = 0;
2281 busiest_load_per_task = busiest_nr_running = 0;
2282 this_load_per_task = this_nr_running = 0;
2283 if (idle == CPU_NOT_IDLE)
2284 load_idx = sd->busy_idx;
2285 else if (idle == CPU_NEWLY_IDLE)
2286 load_idx = sd->newidle_idx;
2288 load_idx = sd->idle_idx;
2291 unsigned long load, group_capacity;
2294 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2295 unsigned long sum_nr_running, sum_weighted_load;
2297 local_group = cpu_isset(this_cpu, group->cpumask);
2300 balance_cpu = first_cpu(group->cpumask);
2302 /* Tally up the load of all CPUs in the group */
2303 sum_weighted_load = sum_nr_running = avg_load = 0;
2305 for_each_cpu_mask(i, group->cpumask) {
2308 if (!cpu_isset(i, *cpus))
2313 if (*sd_idle && rq->nr_running)
2316 /* Bias balancing toward cpus of our domain */
2318 if (idle_cpu(i) && !first_idle_cpu) {
2323 load = target_load(i, load_idx);
2325 load = source_load(i, load_idx);
2328 sum_nr_running += rq->nr_running;
2329 sum_weighted_load += weighted_cpuload(i);
2333 * First idle cpu or the first cpu(busiest) in this sched group
2334 * is eligible for doing load balancing at this and above
2335 * domains. In the newly idle case, we will allow all the cpu's
2336 * to do the newly idle load balance.
2338 if (idle != CPU_NEWLY_IDLE && local_group &&
2339 balance_cpu != this_cpu && balance) {
2344 total_load += avg_load;
2345 total_pwr += group->__cpu_power;
2347 /* Adjust by relative CPU power of the group */
2348 avg_load = sg_div_cpu_power(group,
2349 avg_load * SCHED_LOAD_SCALE);
2351 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2354 this_load = avg_load;
2356 this_nr_running = sum_nr_running;
2357 this_load_per_task = sum_weighted_load;
2358 } else if (avg_load > max_load &&
2359 sum_nr_running > group_capacity) {
2360 max_load = avg_load;
2362 busiest_nr_running = sum_nr_running;
2363 busiest_load_per_task = sum_weighted_load;
2366 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2368 * Busy processors will not participate in power savings
2371 if (idle == CPU_NOT_IDLE ||
2372 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2376 * If the local group is idle or completely loaded
2377 * no need to do power savings balance at this domain
2379 if (local_group && (this_nr_running >= group_capacity ||
2381 power_savings_balance = 0;
2384 * If a group is already running at full capacity or idle,
2385 * don't include that group in power savings calculations
2387 if (!power_savings_balance || sum_nr_running >= group_capacity
2392 * Calculate the group which has the least non-idle load.
2393 * This is the group from where we need to pick up the load
2396 if ((sum_nr_running < min_nr_running) ||
2397 (sum_nr_running == min_nr_running &&
2398 first_cpu(group->cpumask) <
2399 first_cpu(group_min->cpumask))) {
2401 min_nr_running = sum_nr_running;
2402 min_load_per_task = sum_weighted_load /
2407 * Calculate the group which is almost near its
2408 * capacity but still has some space to pick up some load
2409 * from other group and save more power
2411 if (sum_nr_running <= group_capacity - 1) {
2412 if (sum_nr_running > leader_nr_running ||
2413 (sum_nr_running == leader_nr_running &&
2414 first_cpu(group->cpumask) >
2415 first_cpu(group_leader->cpumask))) {
2416 group_leader = group;
2417 leader_nr_running = sum_nr_running;
2422 group = group->next;
2423 } while (group != sd->groups);
2425 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2428 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2430 if (this_load >= avg_load ||
2431 100*max_load <= sd->imbalance_pct*this_load)
2434 busiest_load_per_task /= busiest_nr_running;
2436 * We're trying to get all the cpus to the average_load, so we don't
2437 * want to push ourselves above the average load, nor do we wish to
2438 * reduce the max loaded cpu below the average load, as either of these
2439 * actions would just result in more rebalancing later, and ping-pong
2440 * tasks around. Thus we look for the minimum possible imbalance.
2441 * Negative imbalances (*we* are more loaded than anyone else) will
2442 * be counted as no imbalance for these purposes -- we can't fix that
2443 * by pulling tasks to us. Be careful of negative numbers as they'll
2444 * appear as very large values with unsigned longs.
2446 if (max_load <= busiest_load_per_task)
2450 * In the presence of smp nice balancing, certain scenarios can have
2451 * max load less than avg load(as we skip the groups at or below
2452 * its cpu_power, while calculating max_load..)
2454 if (max_load < avg_load) {
2456 goto small_imbalance;
2459 /* Don't want to pull so many tasks that a group would go idle */
2460 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2462 /* How much load to actually move to equalise the imbalance */
2463 *imbalance = min(max_pull * busiest->__cpu_power,
2464 (avg_load - this_load) * this->__cpu_power)
2468 * if *imbalance is less than the average load per runnable task
2469 * there is no gaurantee that any tasks will be moved so we'll have
2470 * a think about bumping its value to force at least one task to be
2473 if (*imbalance < busiest_load_per_task) {
2474 unsigned long tmp, pwr_now, pwr_move;
2478 pwr_move = pwr_now = 0;
2480 if (this_nr_running) {
2481 this_load_per_task /= this_nr_running;
2482 if (busiest_load_per_task > this_load_per_task)
2485 this_load_per_task = SCHED_LOAD_SCALE;
2487 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2488 busiest_load_per_task * imbn) {
2489 *imbalance = busiest_load_per_task;
2494 * OK, we don't have enough imbalance to justify moving tasks,
2495 * however we may be able to increase total CPU power used by
2499 pwr_now += busiest->__cpu_power *
2500 min(busiest_load_per_task, max_load);
2501 pwr_now += this->__cpu_power *
2502 min(this_load_per_task, this_load);
2503 pwr_now /= SCHED_LOAD_SCALE;
2505 /* Amount of load we'd subtract */
2506 tmp = sg_div_cpu_power(busiest,
2507 busiest_load_per_task * SCHED_LOAD_SCALE);
2509 pwr_move += busiest->__cpu_power *
2510 min(busiest_load_per_task, max_load - tmp);
2512 /* Amount of load we'd add */
2513 if (max_load * busiest->__cpu_power <
2514 busiest_load_per_task * SCHED_LOAD_SCALE)
2515 tmp = sg_div_cpu_power(this,
2516 max_load * busiest->__cpu_power);
2518 tmp = sg_div_cpu_power(this,
2519 busiest_load_per_task * SCHED_LOAD_SCALE);
2520 pwr_move += this->__cpu_power *
2521 min(this_load_per_task, this_load + tmp);
2522 pwr_move /= SCHED_LOAD_SCALE;
2524 /* Move if we gain throughput */
2525 if (pwr_move > pwr_now)
2526 *imbalance = busiest_load_per_task;
2532 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2533 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2536 if (this == group_leader && group_leader != group_min) {
2537 *imbalance = min_load_per_task;
2547 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2550 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2551 unsigned long imbalance, cpumask_t *cpus)
2553 struct rq *busiest = NULL, *rq;
2554 unsigned long max_load = 0;
2557 for_each_cpu_mask(i, group->cpumask) {
2560 if (!cpu_isset(i, *cpus))
2564 wl = weighted_cpuload(i);
2566 if (rq->nr_running == 1 && wl > imbalance)
2569 if (wl > max_load) {
2579 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2580 * so long as it is large enough.
2582 #define MAX_PINNED_INTERVAL 512
2585 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2586 * tasks if there is an imbalance.
2588 static int load_balance(int this_cpu, struct rq *this_rq,
2589 struct sched_domain *sd, enum cpu_idle_type idle,
2592 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2593 struct sched_group *group;
2594 unsigned long imbalance;
2596 cpumask_t cpus = CPU_MASK_ALL;
2597 unsigned long flags;
2600 * When power savings policy is enabled for the parent domain, idle
2601 * sibling can pick up load irrespective of busy siblings. In this case,
2602 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2603 * portraying it as CPU_NOT_IDLE.
2605 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2606 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2609 schedstat_inc(sd, lb_count[idle]);
2612 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2619 schedstat_inc(sd, lb_nobusyg[idle]);
2623 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2625 schedstat_inc(sd, lb_nobusyq[idle]);
2629 BUG_ON(busiest == this_rq);
2631 schedstat_add(sd, lb_imbalance[idle], imbalance);
2634 if (busiest->nr_running > 1) {
2636 * Attempt to move tasks. If find_busiest_group has found
2637 * an imbalance but busiest->nr_running <= 1, the group is
2638 * still unbalanced. ld_moved simply stays zero, so it is
2639 * correctly treated as an imbalance.
2641 local_irq_save(flags);
2642 double_rq_lock(this_rq, busiest);
2643 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2644 imbalance, sd, idle, &all_pinned);
2645 double_rq_unlock(this_rq, busiest);
2646 local_irq_restore(flags);
2649 * some other cpu did the load balance for us.
2651 if (ld_moved && this_cpu != smp_processor_id())
2652 resched_cpu(this_cpu);
2654 /* All tasks on this runqueue were pinned by CPU affinity */
2655 if (unlikely(all_pinned)) {
2656 cpu_clear(cpu_of(busiest), cpus);
2657 if (!cpus_empty(cpus))
2664 schedstat_inc(sd, lb_failed[idle]);
2665 sd->nr_balance_failed++;
2667 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2669 spin_lock_irqsave(&busiest->lock, flags);
2671 /* don't kick the migration_thread, if the curr
2672 * task on busiest cpu can't be moved to this_cpu
2674 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2675 spin_unlock_irqrestore(&busiest->lock, flags);
2677 goto out_one_pinned;
2680 if (!busiest->active_balance) {
2681 busiest->active_balance = 1;
2682 busiest->push_cpu = this_cpu;
2685 spin_unlock_irqrestore(&busiest->lock, flags);
2687 wake_up_process(busiest->migration_thread);
2690 * We've kicked active balancing, reset the failure
2693 sd->nr_balance_failed = sd->cache_nice_tries+1;
2696 sd->nr_balance_failed = 0;
2698 if (likely(!active_balance)) {
2699 /* We were unbalanced, so reset the balancing interval */
2700 sd->balance_interval = sd->min_interval;
2703 * If we've begun active balancing, start to back off. This
2704 * case may not be covered by the all_pinned logic if there
2705 * is only 1 task on the busy runqueue (because we don't call
2708 if (sd->balance_interval < sd->max_interval)
2709 sd->balance_interval *= 2;
2712 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2713 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2718 schedstat_inc(sd, lb_balanced[idle]);
2720 sd->nr_balance_failed = 0;
2723 /* tune up the balancing interval */
2724 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2725 (sd->balance_interval < sd->max_interval))
2726 sd->balance_interval *= 2;
2728 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2729 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2735 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2736 * tasks if there is an imbalance.
2738 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2739 * this_rq is locked.
2742 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2744 struct sched_group *group;
2745 struct rq *busiest = NULL;
2746 unsigned long imbalance;
2750 cpumask_t cpus = CPU_MASK_ALL;
2753 * When power savings policy is enabled for the parent domain, idle
2754 * sibling can pick up load irrespective of busy siblings. In this case,
2755 * let the state of idle sibling percolate up as IDLE, instead of
2756 * portraying it as CPU_NOT_IDLE.
2758 if (sd->flags & SD_SHARE_CPUPOWER &&
2759 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2762 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2764 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2765 &sd_idle, &cpus, NULL);
2767 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2771 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2774 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2778 BUG_ON(busiest == this_rq);
2780 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2783 if (busiest->nr_running > 1) {
2784 /* Attempt to move tasks */
2785 double_lock_balance(this_rq, busiest);
2786 /* this_rq->clock is already updated */
2787 update_rq_clock(busiest);
2788 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2789 imbalance, sd, CPU_NEWLY_IDLE,
2791 spin_unlock(&busiest->lock);
2793 if (unlikely(all_pinned)) {
2794 cpu_clear(cpu_of(busiest), cpus);
2795 if (!cpus_empty(cpus))
2801 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2802 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2803 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2806 sd->nr_balance_failed = 0;
2811 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2812 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2813 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2815 sd->nr_balance_failed = 0;
2821 * idle_balance is called by schedule() if this_cpu is about to become
2822 * idle. Attempts to pull tasks from other CPUs.
2824 static void idle_balance(int this_cpu, struct rq *this_rq)
2826 struct sched_domain *sd;
2827 int pulled_task = -1;
2828 unsigned long next_balance = jiffies + HZ;
2830 for_each_domain(this_cpu, sd) {
2831 unsigned long interval;
2833 if (!(sd->flags & SD_LOAD_BALANCE))
2836 if (sd->flags & SD_BALANCE_NEWIDLE)
2837 /* If we've pulled tasks over stop searching: */
2838 pulled_task = load_balance_newidle(this_cpu,
2841 interval = msecs_to_jiffies(sd->balance_interval);
2842 if (time_after(next_balance, sd->last_balance + interval))
2843 next_balance = sd->last_balance + interval;
2847 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2849 * We are going idle. next_balance may be set based on
2850 * a busy processor. So reset next_balance.
2852 this_rq->next_balance = next_balance;
2857 * active_load_balance is run by migration threads. It pushes running tasks
2858 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2859 * running on each physical CPU where possible, and avoids physical /
2860 * logical imbalances.
2862 * Called with busiest_rq locked.
2864 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2866 int target_cpu = busiest_rq->push_cpu;
2867 struct sched_domain *sd;
2868 struct rq *target_rq;
2870 /* Is there any task to move? */
2871 if (busiest_rq->nr_running <= 1)
2874 target_rq = cpu_rq(target_cpu);
2877 * This condition is "impossible", if it occurs
2878 * we need to fix it. Originally reported by
2879 * Bjorn Helgaas on a 128-cpu setup.
2881 BUG_ON(busiest_rq == target_rq);
2883 /* move a task from busiest_rq to target_rq */
2884 double_lock_balance(busiest_rq, target_rq);
2885 update_rq_clock(busiest_rq);
2886 update_rq_clock(target_rq);
2888 /* Search for an sd spanning us and the target CPU. */
2889 for_each_domain(target_cpu, sd) {
2890 if ((sd->flags & SD_LOAD_BALANCE) &&
2891 cpu_isset(busiest_cpu, sd->span))
2896 schedstat_inc(sd, alb_count);
2898 if (move_one_task(target_rq, target_cpu, busiest_rq,
2900 schedstat_inc(sd, alb_pushed);
2902 schedstat_inc(sd, alb_failed);
2904 spin_unlock(&target_rq->lock);
2909 atomic_t load_balancer;
2911 } nohz ____cacheline_aligned = {
2912 .load_balancer = ATOMIC_INIT(-1),
2913 .cpu_mask = CPU_MASK_NONE,
2917 * This routine will try to nominate the ilb (idle load balancing)
2918 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2919 * load balancing on behalf of all those cpus. If all the cpus in the system
2920 * go into this tickless mode, then there will be no ilb owner (as there is
2921 * no need for one) and all the cpus will sleep till the next wakeup event
2924 * For the ilb owner, tick is not stopped. And this tick will be used
2925 * for idle load balancing. ilb owner will still be part of
2928 * While stopping the tick, this cpu will become the ilb owner if there
2929 * is no other owner. And will be the owner till that cpu becomes busy
2930 * or if all cpus in the system stop their ticks at which point
2931 * there is no need for ilb owner.
2933 * When the ilb owner becomes busy, it nominates another owner, during the
2934 * next busy scheduler_tick()
2936 int select_nohz_load_balancer(int stop_tick)
2938 int cpu = smp_processor_id();
2941 cpu_set(cpu, nohz.cpu_mask);
2942 cpu_rq(cpu)->in_nohz_recently = 1;
2945 * If we are going offline and still the leader, give up!
2947 if (cpu_is_offline(cpu) &&
2948 atomic_read(&nohz.load_balancer) == cpu) {
2949 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2954 /* time for ilb owner also to sleep */
2955 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2956 if (atomic_read(&nohz.load_balancer) == cpu)
2957 atomic_set(&nohz.load_balancer, -1);
2961 if (atomic_read(&nohz.load_balancer) == -1) {
2962 /* make me the ilb owner */
2963 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2965 } else if (atomic_read(&nohz.load_balancer) == cpu)
2968 if (!cpu_isset(cpu, nohz.cpu_mask))
2971 cpu_clear(cpu, nohz.cpu_mask);
2973 if (atomic_read(&nohz.load_balancer) == cpu)
2974 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2981 static DEFINE_SPINLOCK(balancing);
2984 * It checks each scheduling domain to see if it is due to be balanced,
2985 * and initiates a balancing operation if so.
2987 * Balancing parameters are set up in arch_init_sched_domains.
2989 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
2992 struct rq *rq = cpu_rq(cpu);
2993 unsigned long interval;
2994 struct sched_domain *sd;
2995 /* Earliest time when we have to do rebalance again */
2996 unsigned long next_balance = jiffies + 60*HZ;
2997 int update_next_balance = 0;
2999 for_each_domain(cpu, sd) {
3000 if (!(sd->flags & SD_LOAD_BALANCE))
3003 interval = sd->balance_interval;
3004 if (idle != CPU_IDLE)
3005 interval *= sd->busy_factor;
3007 /* scale ms to jiffies */
3008 interval = msecs_to_jiffies(interval);
3009 if (unlikely(!interval))
3011 if (interval > HZ*NR_CPUS/10)
3012 interval = HZ*NR_CPUS/10;
3015 if (sd->flags & SD_SERIALIZE) {
3016 if (!spin_trylock(&balancing))
3020 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3021 if (load_balance(cpu, rq, sd, idle, &balance)) {
3023 * We've pulled tasks over so either we're no
3024 * longer idle, or one of our SMT siblings is
3027 idle = CPU_NOT_IDLE;
3029 sd->last_balance = jiffies;
3031 if (sd->flags & SD_SERIALIZE)
3032 spin_unlock(&balancing);
3034 if (time_after(next_balance, sd->last_balance + interval)) {
3035 next_balance = sd->last_balance + interval;
3036 update_next_balance = 1;
3040 * Stop the load balance at this level. There is another
3041 * CPU in our sched group which is doing load balancing more
3049 * next_balance will be updated only when there is a need.
3050 * When the cpu is attached to null domain for ex, it will not be
3053 if (likely(update_next_balance))
3054 rq->next_balance = next_balance;
3058 * run_rebalance_domains is triggered when needed from the scheduler tick.
3059 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3060 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3062 static void run_rebalance_domains(struct softirq_action *h)
3064 int this_cpu = smp_processor_id();
3065 struct rq *this_rq = cpu_rq(this_cpu);
3066 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3067 CPU_IDLE : CPU_NOT_IDLE;
3069 rebalance_domains(this_cpu, idle);
3073 * If this cpu is the owner for idle load balancing, then do the
3074 * balancing on behalf of the other idle cpus whose ticks are
3077 if (this_rq->idle_at_tick &&
3078 atomic_read(&nohz.load_balancer) == this_cpu) {
3079 cpumask_t cpus = nohz.cpu_mask;
3083 cpu_clear(this_cpu, cpus);
3084 for_each_cpu_mask(balance_cpu, cpus) {
3086 * If this cpu gets work to do, stop the load balancing
3087 * work being done for other cpus. Next load
3088 * balancing owner will pick it up.
3093 rebalance_domains(balance_cpu, CPU_IDLE);
3095 rq = cpu_rq(balance_cpu);
3096 if (time_after(this_rq->next_balance, rq->next_balance))
3097 this_rq->next_balance = rq->next_balance;
3104 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3106 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3107 * idle load balancing owner or decide to stop the periodic load balancing,
3108 * if the whole system is idle.
3110 static inline void trigger_load_balance(struct rq *rq, int cpu)
3114 * If we were in the nohz mode recently and busy at the current
3115 * scheduler tick, then check if we need to nominate new idle
3118 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3119 rq->in_nohz_recently = 0;
3121 if (atomic_read(&nohz.load_balancer) == cpu) {
3122 cpu_clear(cpu, nohz.cpu_mask);
3123 atomic_set(&nohz.load_balancer, -1);
3126 if (atomic_read(&nohz.load_balancer) == -1) {
3128 * simple selection for now: Nominate the
3129 * first cpu in the nohz list to be the next
3132 * TBD: Traverse the sched domains and nominate
3133 * the nearest cpu in the nohz.cpu_mask.
3135 int ilb = first_cpu(nohz.cpu_mask);
3143 * If this cpu is idle and doing idle load balancing for all the
3144 * cpus with ticks stopped, is it time for that to stop?
3146 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3147 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3153 * If this cpu is idle and the idle load balancing is done by
3154 * someone else, then no need raise the SCHED_SOFTIRQ
3156 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3157 cpu_isset(cpu, nohz.cpu_mask))
3160 if (time_after_eq(jiffies, rq->next_balance))
3161 raise_softirq(SCHED_SOFTIRQ);
3164 #else /* CONFIG_SMP */
3167 * on UP we do not need to balance between CPUs:
3169 static inline void idle_balance(int cpu, struct rq *rq)
3173 /* Avoid "used but not defined" warning on UP */
3174 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3175 unsigned long max_nr_move, unsigned long max_load_move,
3176 struct sched_domain *sd, enum cpu_idle_type idle,
3177 int *all_pinned, unsigned long *load_moved,
3178 int *this_best_prio, struct rq_iterator *iterator)
3187 DEFINE_PER_CPU(struct kernel_stat, kstat);
3189 EXPORT_PER_CPU_SYMBOL(kstat);
3192 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3193 * that have not yet been banked in case the task is currently running.
3195 unsigned long long task_sched_runtime(struct task_struct *p)
3197 unsigned long flags;
3201 rq = task_rq_lock(p, &flags);
3202 ns = p->se.sum_exec_runtime;
3203 if (rq->curr == p) {
3204 update_rq_clock(rq);
3205 delta_exec = rq->clock - p->se.exec_start;
3206 if ((s64)delta_exec > 0)
3209 task_rq_unlock(rq, &flags);
3215 * Account user cpu time to a process.
3216 * @p: the process that the cpu time gets accounted to
3217 * @hardirq_offset: the offset to subtract from hardirq_count()
3218 * @cputime: the cpu time spent in user space since the last update
3220 void account_user_time(struct task_struct *p, cputime_t cputime)
3222 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3225 p->utime = cputime_add(p->utime, cputime);
3227 /* Add user time to cpustat. */
3228 tmp = cputime_to_cputime64(cputime);
3229 if (TASK_NICE(p) > 0)
3230 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3232 cpustat->user = cputime64_add(cpustat->user, tmp);
3236 * Account system cpu time to a process.
3237 * @p: the process that the cpu time gets accounted to
3238 * @hardirq_offset: the offset to subtract from hardirq_count()
3239 * @cputime: the cpu time spent in kernel space since the last update
3241 void account_system_time(struct task_struct *p, int hardirq_offset,
3244 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3245 struct rq *rq = this_rq();
3248 p->stime = cputime_add(p->stime, cputime);
3250 /* Add system time to cpustat. */
3251 tmp = cputime_to_cputime64(cputime);
3252 if (hardirq_count() - hardirq_offset)
3253 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3254 else if (softirq_count())
3255 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3256 else if (p != rq->idle)
3257 cpustat->system = cputime64_add(cpustat->system, tmp);
3258 else if (atomic_read(&rq->nr_iowait) > 0)
3259 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3261 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3262 /* Account for system time used */
3263 acct_update_integrals(p);
3267 * Account for involuntary wait time.
3268 * @p: the process from which the cpu time has been stolen
3269 * @steal: the cpu time spent in involuntary wait
3271 void account_steal_time(struct task_struct *p, cputime_t steal)
3273 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3274 cputime64_t tmp = cputime_to_cputime64(steal);
3275 struct rq *rq = this_rq();
3277 if (p == rq->idle) {
3278 p->stime = cputime_add(p->stime, steal);
3279 if (atomic_read(&rq->nr_iowait) > 0)
3280 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3282 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3284 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3288 * This function gets called by the timer code, with HZ frequency.
3289 * We call it with interrupts disabled.
3291 * It also gets called by the fork code, when changing the parent's
3294 void scheduler_tick(void)
3296 int cpu = smp_processor_id();
3297 struct rq *rq = cpu_rq(cpu);
3298 struct task_struct *curr = rq->curr;
3299 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3301 spin_lock(&rq->lock);
3302 __update_rq_clock(rq);
3304 * Let rq->clock advance by at least TICK_NSEC:
3306 if (unlikely(rq->clock < next_tick))
3307 rq->clock = next_tick;
3308 rq->tick_timestamp = rq->clock;
3309 update_cpu_load(rq);
3310 if (curr != rq->idle) /* FIXME: needed? */
3311 curr->sched_class->task_tick(rq, curr);
3312 spin_unlock(&rq->lock);
3315 rq->idle_at_tick = idle_cpu(cpu);
3316 trigger_load_balance(rq, cpu);
3320 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3322 void fastcall add_preempt_count(int val)
3327 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3329 preempt_count() += val;
3331 * Spinlock count overflowing soon?
3333 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3336 EXPORT_SYMBOL(add_preempt_count);
3338 void fastcall sub_preempt_count(int val)
3343 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3346 * Is the spinlock portion underflowing?
3348 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3349 !(preempt_count() & PREEMPT_MASK)))
3352 preempt_count() -= val;
3354 EXPORT_SYMBOL(sub_preempt_count);
3359 * Print scheduling while atomic bug:
3361 static noinline void __schedule_bug(struct task_struct *prev)
3363 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3364 prev->comm, preempt_count(), prev->pid);
3365 debug_show_held_locks(prev);
3366 if (irqs_disabled())
3367 print_irqtrace_events(prev);
3372 * Various schedule()-time debugging checks and statistics:
3374 static inline void schedule_debug(struct task_struct *prev)
3377 * Test if we are atomic. Since do_exit() needs to call into
3378 * schedule() atomically, we ignore that path for now.
3379 * Otherwise, whine if we are scheduling when we should not be.
3381 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3382 __schedule_bug(prev);
3384 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3386 schedstat_inc(this_rq(), sched_count);
3387 #ifdef CONFIG_SCHEDSTATS
3388 if (unlikely(prev->lock_depth >= 0)) {
3389 schedstat_inc(this_rq(), bkl_count);
3390 schedstat_inc(prev, sched_info.bkl_count);
3396 * Pick up the highest-prio task:
3398 static inline struct task_struct *
3399 pick_next_task(struct rq *rq, struct task_struct *prev)
3401 const struct sched_class *class;
3402 struct task_struct *p;
3405 * Optimization: we know that if all tasks are in
3406 * the fair class we can call that function directly:
3408 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3409 p = fair_sched_class.pick_next_task(rq);
3414 class = sched_class_highest;
3416 p = class->pick_next_task(rq);
3420 * Will never be NULL as the idle class always
3421 * returns a non-NULL p:
3423 class = class->next;
3428 * schedule() is the main scheduler function.
3430 asmlinkage void __sched schedule(void)
3432 struct task_struct *prev, *next;
3439 cpu = smp_processor_id();
3443 switch_count = &prev->nivcsw;
3445 release_kernel_lock(prev);
3446 need_resched_nonpreemptible:
3448 schedule_debug(prev);
3451 * Do the rq-clock update outside the rq lock:
3453 local_irq_disable();
3454 __update_rq_clock(rq);
3455 spin_lock(&rq->lock);
3456 clear_tsk_need_resched(prev);
3458 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3459 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3460 unlikely(signal_pending(prev)))) {
3461 prev->state = TASK_RUNNING;
3463 deactivate_task(rq, prev, 1);
3465 switch_count = &prev->nvcsw;
3468 if (unlikely(!rq->nr_running))
3469 idle_balance(cpu, rq);
3471 prev->sched_class->put_prev_task(rq, prev);
3472 next = pick_next_task(rq, prev);
3474 sched_info_switch(prev, next);
3476 if (likely(prev != next)) {
3481 context_switch(rq, prev, next); /* unlocks the rq */
3483 spin_unlock_irq(&rq->lock);
3485 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3486 cpu = smp_processor_id();
3488 goto need_resched_nonpreemptible;
3490 preempt_enable_no_resched();
3491 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3494 EXPORT_SYMBOL(schedule);
3496 #ifdef CONFIG_PREEMPT
3498 * this is the entry point to schedule() from in-kernel preemption
3499 * off of preempt_enable. Kernel preemptions off return from interrupt
3500 * occur there and call schedule directly.
3502 asmlinkage void __sched preempt_schedule(void)
3504 struct thread_info *ti = current_thread_info();
3505 #ifdef CONFIG_PREEMPT_BKL
3506 struct task_struct *task = current;
3507 int saved_lock_depth;
3510 * If there is a non-zero preempt_count or interrupts are disabled,
3511 * we do not want to preempt the current task. Just return..
3513 if (likely(ti->preempt_count || irqs_disabled()))
3517 add_preempt_count(PREEMPT_ACTIVE);
3520 * We keep the big kernel semaphore locked, but we
3521 * clear ->lock_depth so that schedule() doesnt
3522 * auto-release the semaphore:
3524 #ifdef CONFIG_PREEMPT_BKL
3525 saved_lock_depth = task->lock_depth;
3526 task->lock_depth = -1;
3529 #ifdef CONFIG_PREEMPT_BKL
3530 task->lock_depth = saved_lock_depth;
3532 sub_preempt_count(PREEMPT_ACTIVE);
3535 * Check again in case we missed a preemption opportunity
3536 * between schedule and now.
3539 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3541 EXPORT_SYMBOL(preempt_schedule);
3544 * this is the entry point to schedule() from kernel preemption
3545 * off of irq context.
3546 * Note, that this is called and return with irqs disabled. This will
3547 * protect us against recursive calling from irq.
3549 asmlinkage void __sched preempt_schedule_irq(void)
3551 struct thread_info *ti = current_thread_info();
3552 #ifdef CONFIG_PREEMPT_BKL
3553 struct task_struct *task = current;
3554 int saved_lock_depth;
3556 /* Catch callers which need to be fixed */
3557 BUG_ON(ti->preempt_count || !irqs_disabled());
3560 add_preempt_count(PREEMPT_ACTIVE);
3563 * We keep the big kernel semaphore locked, but we
3564 * clear ->lock_depth so that schedule() doesnt
3565 * auto-release the semaphore:
3567 #ifdef CONFIG_PREEMPT_BKL
3568 saved_lock_depth = task->lock_depth;
3569 task->lock_depth = -1;
3573 local_irq_disable();
3574 #ifdef CONFIG_PREEMPT_BKL
3575 task->lock_depth = saved_lock_depth;
3577 sub_preempt_count(PREEMPT_ACTIVE);
3580 * Check again in case we missed a preemption opportunity
3581 * between schedule and now.
3584 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3587 #endif /* CONFIG_PREEMPT */
3589 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3592 return try_to_wake_up(curr->private, mode, sync);
3594 EXPORT_SYMBOL(default_wake_function);
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3605 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int sync, void *key)
3608 wait_queue_t *curr, *next;
3610 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3611 unsigned flags = curr->flags;
3613 if (curr->func(curr, mode, sync, key) &&
3614 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3620 * __wake_up - wake up threads blocked on a waitqueue.
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3626 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3627 int nr_exclusive, void *key)
3629 unsigned long flags;
3631 spin_lock_irqsave(&q->lock, flags);
3632 __wake_up_common(q, mode, nr_exclusive, 0, key);
3633 spin_unlock_irqrestore(&q->lock, flags);
3635 EXPORT_SYMBOL(__wake_up);
3638 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3640 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3642 __wake_up_common(q, mode, 1, 0, NULL);
3646 * __wake_up_sync - wake up threads blocked on a waitqueue.
3648 * @mode: which threads
3649 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3651 * The sync wakeup differs that the waker knows that it will schedule
3652 * away soon, so while the target thread will be woken up, it will not
3653 * be migrated to another CPU - ie. the two threads are 'synchronized'
3654 * with each other. This can prevent needless bouncing between CPUs.
3656 * On UP it can prevent extra preemption.
3659 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3661 unsigned long flags;
3667 if (unlikely(!nr_exclusive))
3670 spin_lock_irqsave(&q->lock, flags);
3671 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3672 spin_unlock_irqrestore(&q->lock, flags);
3674 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3676 void fastcall complete(struct completion *x)
3678 unsigned long flags;
3680 spin_lock_irqsave(&x->wait.lock, flags);
3682 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3684 spin_unlock_irqrestore(&x->wait.lock, flags);
3686 EXPORT_SYMBOL(complete);
3688 void fastcall complete_all(struct completion *x)
3690 unsigned long flags;
3692 spin_lock_irqsave(&x->wait.lock, flags);
3693 x->done += UINT_MAX/2;
3694 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3696 spin_unlock_irqrestore(&x->wait.lock, flags);
3698 EXPORT_SYMBOL(complete_all);
3700 static inline long __sched
3701 do_wait_for_common(struct completion *x, long timeout, int state)
3704 DECLARE_WAITQUEUE(wait, current);
3706 wait.flags |= WQ_FLAG_EXCLUSIVE;
3707 __add_wait_queue_tail(&x->wait, &wait);
3709 if (state == TASK_INTERRUPTIBLE &&
3710 signal_pending(current)) {
3711 __remove_wait_queue(&x->wait, &wait);
3712 return -ERESTARTSYS;
3714 __set_current_state(state);
3715 spin_unlock_irq(&x->wait.lock);
3716 timeout = schedule_timeout(timeout);
3717 spin_lock_irq(&x->wait.lock);
3719 __remove_wait_queue(&x->wait, &wait);
3723 __remove_wait_queue(&x->wait, &wait);
3730 wait_for_common(struct completion *x, long timeout, int state)
3734 spin_lock_irq(&x->wait.lock);
3735 timeout = do_wait_for_common(x, timeout, state);
3736 spin_unlock_irq(&x->wait.lock);
3740 void fastcall __sched wait_for_completion(struct completion *x)
3742 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3744 EXPORT_SYMBOL(wait_for_completion);
3746 unsigned long fastcall __sched
3747 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3749 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3751 EXPORT_SYMBOL(wait_for_completion_timeout);
3753 int __sched wait_for_completion_interruptible(struct completion *x)
3755 return wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3757 EXPORT_SYMBOL(wait_for_completion_interruptible);
3759 unsigned long fastcall __sched
3760 wait_for_completion_interruptible_timeout(struct completion *x,
3761 unsigned long timeout)
3763 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3765 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3768 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3770 unsigned long flags;
3773 init_waitqueue_entry(&wait, current);
3775 __set_current_state(state);
3777 spin_lock_irqsave(&q->lock, flags);
3778 __add_wait_queue(q, &wait);
3779 spin_unlock(&q->lock);
3780 timeout = schedule_timeout(timeout);
3781 spin_lock_irq(&q->lock);
3782 __remove_wait_queue(q, &wait);
3783 spin_unlock_irqrestore(&q->lock, flags);
3788 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3790 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3792 EXPORT_SYMBOL(interruptible_sleep_on);
3795 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3797 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3799 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3801 void __sched sleep_on(wait_queue_head_t *q)
3803 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3805 EXPORT_SYMBOL(sleep_on);
3807 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3809 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3811 EXPORT_SYMBOL(sleep_on_timeout);
3813 #ifdef CONFIG_RT_MUTEXES
3816 * rt_mutex_setprio - set the current priority of a task
3818 * @prio: prio value (kernel-internal form)
3820 * This function changes the 'effective' priority of a task. It does
3821 * not touch ->normal_prio like __setscheduler().
3823 * Used by the rt_mutex code to implement priority inheritance logic.
3825 void rt_mutex_setprio(struct task_struct *p, int prio)
3827 unsigned long flags;
3828 int oldprio, on_rq, running;
3831 BUG_ON(prio < 0 || prio > MAX_PRIO);
3833 rq = task_rq_lock(p, &flags);
3834 update_rq_clock(rq);
3837 on_rq = p->se.on_rq;
3838 running = task_running(rq, p);
3840 dequeue_task(rq, p, 0);
3842 p->sched_class->put_prev_task(rq, p);
3846 p->sched_class = &rt_sched_class;
3848 p->sched_class = &fair_sched_class;
3854 p->sched_class->set_curr_task(rq);
3855 enqueue_task(rq, p, 0);
3857 * Reschedule if we are currently running on this runqueue and
3858 * our priority decreased, or if we are not currently running on
3859 * this runqueue and our priority is higher than the current's
3862 if (p->prio > oldprio)
3863 resched_task(rq->curr);
3865 check_preempt_curr(rq, p);
3868 task_rq_unlock(rq, &flags);
3873 void set_user_nice(struct task_struct *p, long nice)
3875 int old_prio, delta, on_rq;
3876 unsigned long flags;
3879 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3882 * We have to be careful, if called from sys_setpriority(),
3883 * the task might be in the middle of scheduling on another CPU.
3885 rq = task_rq_lock(p, &flags);
3886 update_rq_clock(rq);
3888 * The RT priorities are set via sched_setscheduler(), but we still
3889 * allow the 'normal' nice value to be set - but as expected
3890 * it wont have any effect on scheduling until the task is
3891 * SCHED_FIFO/SCHED_RR:
3893 if (task_has_rt_policy(p)) {
3894 p->static_prio = NICE_TO_PRIO(nice);
3897 on_rq = p->se.on_rq;
3899 dequeue_task(rq, p, 0);
3903 p->static_prio = NICE_TO_PRIO(nice);
3906 p->prio = effective_prio(p);
3907 delta = p->prio - old_prio;
3910 enqueue_task(rq, p, 0);
3913 * If the task increased its priority or is running and
3914 * lowered its priority, then reschedule its CPU:
3916 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3917 resched_task(rq->curr);
3920 task_rq_unlock(rq, &flags);
3922 EXPORT_SYMBOL(set_user_nice);
3925 * can_nice - check if a task can reduce its nice value
3929 int can_nice(const struct task_struct *p, const int nice)
3931 /* convert nice value [19,-20] to rlimit style value [1,40] */
3932 int nice_rlim = 20 - nice;
3934 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3935 capable(CAP_SYS_NICE));
3938 #ifdef __ARCH_WANT_SYS_NICE
3941 * sys_nice - change the priority of the current process.
3942 * @increment: priority increment
3944 * sys_setpriority is a more generic, but much slower function that
3945 * does similar things.
3947 asmlinkage long sys_nice(int increment)
3952 * Setpriority might change our priority at the same moment.
3953 * We don't have to worry. Conceptually one call occurs first
3954 * and we have a single winner.
3956 if (increment < -40)
3961 nice = PRIO_TO_NICE(current->static_prio) + increment;
3967 if (increment < 0 && !can_nice(current, nice))
3970 retval = security_task_setnice(current, nice);
3974 set_user_nice(current, nice);
3981 * task_prio - return the priority value of a given task.
3982 * @p: the task in question.
3984 * This is the priority value as seen by users in /proc.
3985 * RT tasks are offset by -200. Normal tasks are centered
3986 * around 0, value goes from -16 to +15.
3988 int task_prio(const struct task_struct *p)
3990 return p->prio - MAX_RT_PRIO;
3994 * task_nice - return the nice value of a given task.
3995 * @p: the task in question.
3997 int task_nice(const struct task_struct *p)
3999 return TASK_NICE(p);
4001 EXPORT_SYMBOL_GPL(task_nice);
4004 * idle_cpu - is a given cpu idle currently?
4005 * @cpu: the processor in question.
4007 int idle_cpu(int cpu)
4009 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4013 * idle_task - return the idle task for a given cpu.
4014 * @cpu: the processor in question.
4016 struct task_struct *idle_task(int cpu)
4018 return cpu_rq(cpu)->idle;
4022 * find_process_by_pid - find a process with a matching PID value.
4023 * @pid: the pid in question.
4025 static struct task_struct *find_process_by_pid(pid_t pid)
4027 return pid ? find_task_by_pid(pid) : current;
4030 /* Actually do priority change: must hold rq lock. */
4032 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4034 BUG_ON(p->se.on_rq);
4037 switch (p->policy) {
4041 p->sched_class = &fair_sched_class;
4045 p->sched_class = &rt_sched_class;
4049 p->rt_priority = prio;
4050 p->normal_prio = normal_prio(p);
4051 /* we are holding p->pi_lock already */
4052 p->prio = rt_mutex_getprio(p);
4057 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4058 * @p: the task in question.
4059 * @policy: new policy.
4060 * @param: structure containing the new RT priority.
4062 * NOTE that the task may be already dead.
4064 int sched_setscheduler(struct task_struct *p, int policy,
4065 struct sched_param *param)
4067 int retval, oldprio, oldpolicy = -1, on_rq, running;
4068 unsigned long flags;
4071 /* may grab non-irq protected spin_locks */
4072 BUG_ON(in_interrupt());
4074 /* double check policy once rq lock held */
4076 policy = oldpolicy = p->policy;
4077 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4078 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4079 policy != SCHED_IDLE)
4082 * Valid priorities for SCHED_FIFO and SCHED_RR are
4083 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4084 * SCHED_BATCH and SCHED_IDLE is 0.
4086 if (param->sched_priority < 0 ||
4087 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4088 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4090 if (rt_policy(policy) != (param->sched_priority != 0))
4094 * Allow unprivileged RT tasks to decrease priority:
4096 if (!capable(CAP_SYS_NICE)) {
4097 if (rt_policy(policy)) {
4098 unsigned long rlim_rtprio;
4100 if (!lock_task_sighand(p, &flags))
4102 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4103 unlock_task_sighand(p, &flags);
4105 /* can't set/change the rt policy */
4106 if (policy != p->policy && !rlim_rtprio)
4109 /* can't increase priority */
4110 if (param->sched_priority > p->rt_priority &&
4111 param->sched_priority > rlim_rtprio)
4115 * Like positive nice levels, dont allow tasks to
4116 * move out of SCHED_IDLE either:
4118 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4121 /* can't change other user's priorities */
4122 if ((current->euid != p->euid) &&
4123 (current->euid != p->uid))
4127 retval = security_task_setscheduler(p, policy, param);
4131 * make sure no PI-waiters arrive (or leave) while we are
4132 * changing the priority of the task:
4134 spin_lock_irqsave(&p->pi_lock, flags);
4136 * To be able to change p->policy safely, the apropriate
4137 * runqueue lock must be held.
4139 rq = __task_rq_lock(p);
4140 /* recheck policy now with rq lock held */
4141 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4142 policy = oldpolicy = -1;
4143 __task_rq_unlock(rq);
4144 spin_unlock_irqrestore(&p->pi_lock, flags);
4147 update_rq_clock(rq);
4148 on_rq = p->se.on_rq;
4149 running = task_running(rq, p);
4151 deactivate_task(rq, p, 0);
4153 p->sched_class->put_prev_task(rq, p);
4157 __setscheduler(rq, p, policy, param->sched_priority);
4161 p->sched_class->set_curr_task(rq);
4162 activate_task(rq, p, 0);
4164 * Reschedule if we are currently running on this runqueue and
4165 * our priority decreased, or if we are not currently running on
4166 * this runqueue and our priority is higher than the current's
4169 if (p->prio > oldprio)
4170 resched_task(rq->curr);
4172 check_preempt_curr(rq, p);
4175 __task_rq_unlock(rq);
4176 spin_unlock_irqrestore(&p->pi_lock, flags);
4178 rt_mutex_adjust_pi(p);
4182 EXPORT_SYMBOL_GPL(sched_setscheduler);
4185 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4187 struct sched_param lparam;
4188 struct task_struct *p;
4191 if (!param || pid < 0)
4193 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4198 p = find_process_by_pid(pid);
4200 retval = sched_setscheduler(p, policy, &lparam);
4207 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4208 * @pid: the pid in question.
4209 * @policy: new policy.
4210 * @param: structure containing the new RT priority.
4212 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4213 struct sched_param __user *param)
4215 /* negative values for policy are not valid */
4219 return do_sched_setscheduler(pid, policy, param);
4223 * sys_sched_setparam - set/change the RT priority of a thread
4224 * @pid: the pid in question.
4225 * @param: structure containing the new RT priority.
4227 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4229 return do_sched_setscheduler(pid, -1, param);
4233 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4234 * @pid: the pid in question.
4236 asmlinkage long sys_sched_getscheduler(pid_t pid)
4238 struct task_struct *p;
4245 read_lock(&tasklist_lock);
4246 p = find_process_by_pid(pid);
4248 retval = security_task_getscheduler(p);
4252 read_unlock(&tasklist_lock);
4257 * sys_sched_getscheduler - get the RT priority of a thread
4258 * @pid: the pid in question.
4259 * @param: structure containing the RT priority.
4261 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4263 struct sched_param lp;
4264 struct task_struct *p;
4267 if (!param || pid < 0)
4270 read_lock(&tasklist_lock);
4271 p = find_process_by_pid(pid);
4276 retval = security_task_getscheduler(p);
4280 lp.sched_priority = p->rt_priority;
4281 read_unlock(&tasklist_lock);
4284 * This one might sleep, we cannot do it with a spinlock held ...
4286 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4291 read_unlock(&tasklist_lock);
4295 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4297 cpumask_t cpus_allowed;
4298 struct task_struct *p;
4301 mutex_lock(&sched_hotcpu_mutex);
4302 read_lock(&tasklist_lock);
4304 p = find_process_by_pid(pid);
4306 read_unlock(&tasklist_lock);
4307 mutex_unlock(&sched_hotcpu_mutex);
4312 * It is not safe to call set_cpus_allowed with the
4313 * tasklist_lock held. We will bump the task_struct's
4314 * usage count and then drop tasklist_lock.
4317 read_unlock(&tasklist_lock);
4320 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4321 !capable(CAP_SYS_NICE))
4324 retval = security_task_setscheduler(p, 0, NULL);
4328 cpus_allowed = cpuset_cpus_allowed(p);
4329 cpus_and(new_mask, new_mask, cpus_allowed);
4330 retval = set_cpus_allowed(p, new_mask);
4334 mutex_unlock(&sched_hotcpu_mutex);
4338 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4339 cpumask_t *new_mask)
4341 if (len < sizeof(cpumask_t)) {
4342 memset(new_mask, 0, sizeof(cpumask_t));
4343 } else if (len > sizeof(cpumask_t)) {
4344 len = sizeof(cpumask_t);
4346 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4350 * sys_sched_setaffinity - set the cpu affinity of a process
4351 * @pid: pid of the process
4352 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4353 * @user_mask_ptr: user-space pointer to the new cpu mask
4355 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4356 unsigned long __user *user_mask_ptr)
4361 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4365 return sched_setaffinity(pid, new_mask);
4369 * Represents all cpu's present in the system
4370 * In systems capable of hotplug, this map could dynamically grow
4371 * as new cpu's are detected in the system via any platform specific
4372 * method, such as ACPI for e.g.
4375 cpumask_t cpu_present_map __read_mostly;
4376 EXPORT_SYMBOL(cpu_present_map);
4379 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4380 EXPORT_SYMBOL(cpu_online_map);
4382 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4383 EXPORT_SYMBOL(cpu_possible_map);
4386 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4388 struct task_struct *p;
4391 mutex_lock(&sched_hotcpu_mutex);
4392 read_lock(&tasklist_lock);
4395 p = find_process_by_pid(pid);
4399 retval = security_task_getscheduler(p);
4403 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4406 read_unlock(&tasklist_lock);
4407 mutex_unlock(&sched_hotcpu_mutex);
4413 * sys_sched_getaffinity - get the cpu affinity of a process
4414 * @pid: pid of the process
4415 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4416 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4418 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4419 unsigned long __user *user_mask_ptr)
4424 if (len < sizeof(cpumask_t))
4427 ret = sched_getaffinity(pid, &mask);
4431 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4434 return sizeof(cpumask_t);
4438 * sys_sched_yield - yield the current processor to other threads.
4440 * This function yields the current CPU to other tasks. If there are no
4441 * other threads running on this CPU then this function will return.
4443 asmlinkage long sys_sched_yield(void)
4445 struct rq *rq = this_rq_lock();
4447 schedstat_inc(rq, yld_count);
4448 current->sched_class->yield_task(rq);
4451 * Since we are going to call schedule() anyway, there's
4452 * no need to preempt or enable interrupts:
4454 __release(rq->lock);
4455 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4456 _raw_spin_unlock(&rq->lock);
4457 preempt_enable_no_resched();
4464 static void __cond_resched(void)
4466 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4467 __might_sleep(__FILE__, __LINE__);
4470 * The BKS might be reacquired before we have dropped
4471 * PREEMPT_ACTIVE, which could trigger a second
4472 * cond_resched() call.
4475 add_preempt_count(PREEMPT_ACTIVE);
4477 sub_preempt_count(PREEMPT_ACTIVE);
4478 } while (need_resched());
4481 int __sched cond_resched(void)
4483 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4484 system_state == SYSTEM_RUNNING) {
4490 EXPORT_SYMBOL(cond_resched);
4493 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4494 * call schedule, and on return reacquire the lock.
4496 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4497 * operations here to prevent schedule() from being called twice (once via
4498 * spin_unlock(), once by hand).
4500 int cond_resched_lock(spinlock_t *lock)
4504 if (need_lockbreak(lock)) {
4510 if (need_resched() && system_state == SYSTEM_RUNNING) {
4511 spin_release(&lock->dep_map, 1, _THIS_IP_);
4512 _raw_spin_unlock(lock);
4513 preempt_enable_no_resched();
4520 EXPORT_SYMBOL(cond_resched_lock);
4522 int __sched cond_resched_softirq(void)
4524 BUG_ON(!in_softirq());
4526 if (need_resched() && system_state == SYSTEM_RUNNING) {
4534 EXPORT_SYMBOL(cond_resched_softirq);
4537 * yield - yield the current processor to other threads.
4539 * This is a shortcut for kernel-space yielding - it marks the
4540 * thread runnable and calls sys_sched_yield().
4542 void __sched yield(void)
4544 set_current_state(TASK_RUNNING);
4547 EXPORT_SYMBOL(yield);
4550 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4551 * that process accounting knows that this is a task in IO wait state.
4553 * But don't do that if it is a deliberate, throttling IO wait (this task
4554 * has set its backing_dev_info: the queue against which it should throttle)
4556 void __sched io_schedule(void)
4558 struct rq *rq = &__raw_get_cpu_var(runqueues);
4560 delayacct_blkio_start();
4561 atomic_inc(&rq->nr_iowait);
4563 atomic_dec(&rq->nr_iowait);
4564 delayacct_blkio_end();
4566 EXPORT_SYMBOL(io_schedule);
4568 long __sched io_schedule_timeout(long timeout)
4570 struct rq *rq = &__raw_get_cpu_var(runqueues);
4573 delayacct_blkio_start();
4574 atomic_inc(&rq->nr_iowait);
4575 ret = schedule_timeout(timeout);
4576 atomic_dec(&rq->nr_iowait);
4577 delayacct_blkio_end();
4582 * sys_sched_get_priority_max - return maximum RT priority.
4583 * @policy: scheduling class.
4585 * this syscall returns the maximum rt_priority that can be used
4586 * by a given scheduling class.
4588 asmlinkage long sys_sched_get_priority_max(int policy)
4595 ret = MAX_USER_RT_PRIO-1;
4607 * sys_sched_get_priority_min - return minimum RT priority.
4608 * @policy: scheduling class.
4610 * this syscall returns the minimum rt_priority that can be used
4611 * by a given scheduling class.
4613 asmlinkage long sys_sched_get_priority_min(int policy)
4631 * sys_sched_rr_get_interval - return the default timeslice of a process.
4632 * @pid: pid of the process.
4633 * @interval: userspace pointer to the timeslice value.
4635 * this syscall writes the default timeslice value of a given process
4636 * into the user-space timespec buffer. A value of '0' means infinity.
4639 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4641 struct task_struct *p;
4642 unsigned int time_slice;
4650 read_lock(&tasklist_lock);
4651 p = find_process_by_pid(pid);
4655 retval = security_task_getscheduler(p);
4659 if (p->policy == SCHED_FIFO)
4661 else if (p->policy == SCHED_RR)
4662 time_slice = DEF_TIMESLICE;
4664 struct sched_entity *se = &p->se;
4665 unsigned long flags;
4668 rq = task_rq_lock(p, &flags);
4669 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4670 task_rq_unlock(rq, &flags);
4672 read_unlock(&tasklist_lock);
4673 jiffies_to_timespec(time_slice, &t);
4674 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4678 read_unlock(&tasklist_lock);
4682 static const char stat_nam[] = "RSDTtZX";
4684 static void show_task(struct task_struct *p)
4686 unsigned long free = 0;
4689 state = p->state ? __ffs(p->state) + 1 : 0;
4690 printk("%-13.13s %c", p->comm,
4691 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4692 #if BITS_PER_LONG == 32
4693 if (state == TASK_RUNNING)
4694 printk(" running ");
4696 printk(" %08lx ", thread_saved_pc(p));
4698 if (state == TASK_RUNNING)
4699 printk(" running task ");
4701 printk(" %016lx ", thread_saved_pc(p));
4703 #ifdef CONFIG_DEBUG_STACK_USAGE
4705 unsigned long *n = end_of_stack(p);
4708 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4711 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4713 if (state != TASK_RUNNING)
4714 show_stack(p, NULL);
4717 void show_state_filter(unsigned long state_filter)
4719 struct task_struct *g, *p;
4721 #if BITS_PER_LONG == 32
4723 " task PC stack pid father\n");
4726 " task PC stack pid father\n");
4728 read_lock(&tasklist_lock);
4729 do_each_thread(g, p) {
4731 * reset the NMI-timeout, listing all files on a slow
4732 * console might take alot of time:
4734 touch_nmi_watchdog();
4735 if (!state_filter || (p->state & state_filter))
4737 } while_each_thread(g, p);
4739 touch_all_softlockup_watchdogs();
4741 #ifdef CONFIG_SCHED_DEBUG
4742 sysrq_sched_debug_show();
4744 read_unlock(&tasklist_lock);
4746 * Only show locks if all tasks are dumped:
4748 if (state_filter == -1)
4749 debug_show_all_locks();
4752 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4754 idle->sched_class = &idle_sched_class;
4758 * init_idle - set up an idle thread for a given CPU
4759 * @idle: task in question
4760 * @cpu: cpu the idle task belongs to
4762 * NOTE: this function does not set the idle thread's NEED_RESCHED
4763 * flag, to make booting more robust.
4765 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4767 struct rq *rq = cpu_rq(cpu);
4768 unsigned long flags;
4771 idle->se.exec_start = sched_clock();
4773 idle->prio = idle->normal_prio = MAX_PRIO;
4774 idle->cpus_allowed = cpumask_of_cpu(cpu);
4775 __set_task_cpu(idle, cpu);
4777 spin_lock_irqsave(&rq->lock, flags);
4778 rq->curr = rq->idle = idle;
4779 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4782 spin_unlock_irqrestore(&rq->lock, flags);
4784 /* Set the preempt count _outside_ the spinlocks! */
4785 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4786 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4788 task_thread_info(idle)->preempt_count = 0;
4791 * The idle tasks have their own, simple scheduling class:
4793 idle->sched_class = &idle_sched_class;
4797 * In a system that switches off the HZ timer nohz_cpu_mask
4798 * indicates which cpus entered this state. This is used
4799 * in the rcu update to wait only for active cpus. For system
4800 * which do not switch off the HZ timer nohz_cpu_mask should
4801 * always be CPU_MASK_NONE.
4803 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4807 * This is how migration works:
4809 * 1) we queue a struct migration_req structure in the source CPU's
4810 * runqueue and wake up that CPU's migration thread.
4811 * 2) we down() the locked semaphore => thread blocks.
4812 * 3) migration thread wakes up (implicitly it forces the migrated
4813 * thread off the CPU)
4814 * 4) it gets the migration request and checks whether the migrated
4815 * task is still in the wrong runqueue.
4816 * 5) if it's in the wrong runqueue then the migration thread removes
4817 * it and puts it into the right queue.
4818 * 6) migration thread up()s the semaphore.
4819 * 7) we wake up and the migration is done.
4823 * Change a given task's CPU affinity. Migrate the thread to a
4824 * proper CPU and schedule it away if the CPU it's executing on
4825 * is removed from the allowed bitmask.
4827 * NOTE: the caller must have a valid reference to the task, the
4828 * task must not exit() & deallocate itself prematurely. The
4829 * call is not atomic; no spinlocks may be held.
4831 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4833 struct migration_req req;
4834 unsigned long flags;
4838 rq = task_rq_lock(p, &flags);
4839 if (!cpus_intersects(new_mask, cpu_online_map)) {
4844 p->cpus_allowed = new_mask;
4845 /* Can the task run on the task's current CPU? If so, we're done */
4846 if (cpu_isset(task_cpu(p), new_mask))
4849 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4850 /* Need help from migration thread: drop lock and wait. */
4851 task_rq_unlock(rq, &flags);
4852 wake_up_process(rq->migration_thread);
4853 wait_for_completion(&req.done);
4854 tlb_migrate_finish(p->mm);
4858 task_rq_unlock(rq, &flags);
4862 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4865 * Move (not current) task off this cpu, onto dest cpu. We're doing
4866 * this because either it can't run here any more (set_cpus_allowed()
4867 * away from this CPU, or CPU going down), or because we're
4868 * attempting to rebalance this task on exec (sched_exec).
4870 * So we race with normal scheduler movements, but that's OK, as long
4871 * as the task is no longer on this CPU.
4873 * Returns non-zero if task was successfully migrated.
4875 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4877 struct rq *rq_dest, *rq_src;
4880 if (unlikely(cpu_is_offline(dest_cpu)))
4883 rq_src = cpu_rq(src_cpu);
4884 rq_dest = cpu_rq(dest_cpu);
4886 double_rq_lock(rq_src, rq_dest);
4887 /* Already moved. */
4888 if (task_cpu(p) != src_cpu)
4890 /* Affinity changed (again). */
4891 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4894 on_rq = p->se.on_rq;
4896 deactivate_task(rq_src, p, 0);
4898 set_task_cpu(p, dest_cpu);
4900 activate_task(rq_dest, p, 0);
4901 check_preempt_curr(rq_dest, p);
4905 double_rq_unlock(rq_src, rq_dest);
4910 * migration_thread - this is a highprio system thread that performs
4911 * thread migration by bumping thread off CPU then 'pushing' onto
4914 static int migration_thread(void *data)
4916 int cpu = (long)data;
4920 BUG_ON(rq->migration_thread != current);
4922 set_current_state(TASK_INTERRUPTIBLE);
4923 while (!kthread_should_stop()) {
4924 struct migration_req *req;
4925 struct list_head *head;
4927 spin_lock_irq(&rq->lock);
4929 if (cpu_is_offline(cpu)) {
4930 spin_unlock_irq(&rq->lock);
4934 if (rq->active_balance) {
4935 active_load_balance(rq, cpu);
4936 rq->active_balance = 0;
4939 head = &rq->migration_queue;
4941 if (list_empty(head)) {
4942 spin_unlock_irq(&rq->lock);
4944 set_current_state(TASK_INTERRUPTIBLE);
4947 req = list_entry(head->next, struct migration_req, list);
4948 list_del_init(head->next);
4950 spin_unlock(&rq->lock);
4951 __migrate_task(req->task, cpu, req->dest_cpu);
4954 complete(&req->done);
4956 __set_current_state(TASK_RUNNING);
4960 /* Wait for kthread_stop */
4961 set_current_state(TASK_INTERRUPTIBLE);
4962 while (!kthread_should_stop()) {
4964 set_current_state(TASK_INTERRUPTIBLE);
4966 __set_current_state(TASK_RUNNING);
4970 #ifdef CONFIG_HOTPLUG_CPU
4972 * Figure out where task on dead CPU should go, use force if neccessary.
4973 * NOTE: interrupts should be disabled by the caller
4975 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
4977 unsigned long flags;
4984 mask = node_to_cpumask(cpu_to_node(dead_cpu));
4985 cpus_and(mask, mask, p->cpus_allowed);
4986 dest_cpu = any_online_cpu(mask);
4988 /* On any allowed CPU? */
4989 if (dest_cpu == NR_CPUS)
4990 dest_cpu = any_online_cpu(p->cpus_allowed);
4992 /* No more Mr. Nice Guy. */
4993 if (dest_cpu == NR_CPUS) {
4994 rq = task_rq_lock(p, &flags);
4995 cpus_setall(p->cpus_allowed);
4996 dest_cpu = any_online_cpu(p->cpus_allowed);
4997 task_rq_unlock(rq, &flags);
5000 * Don't tell them about moving exiting tasks or
5001 * kernel threads (both mm NULL), since they never
5004 if (p->mm && printk_ratelimit())
5005 printk(KERN_INFO "process %d (%s) no "
5006 "longer affine to cpu%d\n",
5007 p->pid, p->comm, dead_cpu);
5009 } while (!__migrate_task(p, dead_cpu, dest_cpu));
5013 * While a dead CPU has no uninterruptible tasks queued at this point,
5014 * it might still have a nonzero ->nr_uninterruptible counter, because
5015 * for performance reasons the counter is not stricly tracking tasks to
5016 * their home CPUs. So we just add the counter to another CPU's counter,
5017 * to keep the global sum constant after CPU-down:
5019 static void migrate_nr_uninterruptible(struct rq *rq_src)
5021 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5022 unsigned long flags;
5024 local_irq_save(flags);
5025 double_rq_lock(rq_src, rq_dest);
5026 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5027 rq_src->nr_uninterruptible = 0;
5028 double_rq_unlock(rq_src, rq_dest);
5029 local_irq_restore(flags);
5032 /* Run through task list and migrate tasks from the dead cpu. */
5033 static void migrate_live_tasks(int src_cpu)
5035 struct task_struct *p, *t;
5037 write_lock_irq(&tasklist_lock);
5039 do_each_thread(t, p) {
5043 if (task_cpu(p) == src_cpu)
5044 move_task_off_dead_cpu(src_cpu, p);
5045 } while_each_thread(t, p);
5047 write_unlock_irq(&tasklist_lock);
5051 * activate_idle_task - move idle task to the _front_ of runqueue.
5053 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5055 update_rq_clock(rq);
5057 if (p->state == TASK_UNINTERRUPTIBLE)
5058 rq->nr_uninterruptible--;
5060 enqueue_task(rq, p, 0);
5061 inc_nr_running(p, rq);
5065 * Schedules idle task to be the next runnable task on current CPU.
5066 * It does so by boosting its priority to highest possible and adding it to
5067 * the _front_ of the runqueue. Used by CPU offline code.
5069 void sched_idle_next(void)
5071 int this_cpu = smp_processor_id();
5072 struct rq *rq = cpu_rq(this_cpu);
5073 struct task_struct *p = rq->idle;
5074 unsigned long flags;
5076 /* cpu has to be offline */
5077 BUG_ON(cpu_online(this_cpu));
5080 * Strictly not necessary since rest of the CPUs are stopped by now
5081 * and interrupts disabled on the current cpu.
5083 spin_lock_irqsave(&rq->lock, flags);
5085 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5087 /* Add idle task to the _front_ of its priority queue: */
5088 activate_idle_task(p, rq);
5090 spin_unlock_irqrestore(&rq->lock, flags);
5094 * Ensures that the idle task is using init_mm right before its cpu goes
5097 void idle_task_exit(void)
5099 struct mm_struct *mm = current->active_mm;
5101 BUG_ON(cpu_online(smp_processor_id()));
5104 switch_mm(mm, &init_mm, current);
5108 /* called under rq->lock with disabled interrupts */
5109 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5111 struct rq *rq = cpu_rq(dead_cpu);
5113 /* Must be exiting, otherwise would be on tasklist. */
5114 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5116 /* Cannot have done final schedule yet: would have vanished. */
5117 BUG_ON(p->state == TASK_DEAD);
5122 * Drop lock around migration; if someone else moves it,
5123 * that's OK. No task can be added to this CPU, so iteration is
5125 * NOTE: interrupts should be left disabled --dev@
5127 spin_unlock(&rq->lock);
5128 move_task_off_dead_cpu(dead_cpu, p);
5129 spin_lock(&rq->lock);
5134 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5135 static void migrate_dead_tasks(unsigned int dead_cpu)
5137 struct rq *rq = cpu_rq(dead_cpu);
5138 struct task_struct *next;
5141 if (!rq->nr_running)
5143 update_rq_clock(rq);
5144 next = pick_next_task(rq, rq->curr);
5147 migrate_dead(dead_cpu, next);
5151 #endif /* CONFIG_HOTPLUG_CPU */
5153 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5155 static struct ctl_table sd_ctl_dir[] = {
5157 .procname = "sched_domain",
5163 static struct ctl_table sd_ctl_root[] = {
5165 .ctl_name = CTL_KERN,
5166 .procname = "kernel",
5168 .child = sd_ctl_dir,
5173 static struct ctl_table *sd_alloc_ctl_entry(int n)
5175 struct ctl_table *entry =
5176 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5179 memset(entry, 0, n * sizeof(struct ctl_table));
5185 set_table_entry(struct ctl_table *entry,
5186 const char *procname, void *data, int maxlen,
5187 mode_t mode, proc_handler *proc_handler)
5189 entry->procname = procname;
5191 entry->maxlen = maxlen;
5193 entry->proc_handler = proc_handler;
5196 static struct ctl_table *
5197 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5199 struct ctl_table *table = sd_alloc_ctl_entry(12);
5201 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5202 sizeof(long), 0644, proc_doulongvec_minmax);
5203 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5204 sizeof(long), 0644, proc_doulongvec_minmax);
5205 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5206 sizeof(int), 0644, proc_dointvec_minmax);
5207 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5208 sizeof(int), 0644, proc_dointvec_minmax);
5209 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5210 sizeof(int), 0644, proc_dointvec_minmax);
5211 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5212 sizeof(int), 0644, proc_dointvec_minmax);
5213 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5214 sizeof(int), 0644, proc_dointvec_minmax);
5215 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5216 sizeof(int), 0644, proc_dointvec_minmax);
5217 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5218 sizeof(int), 0644, proc_dointvec_minmax);
5219 set_table_entry(&table[9], "cache_nice_tries",
5220 &sd->cache_nice_tries,
5221 sizeof(int), 0644, proc_dointvec_minmax);
5222 set_table_entry(&table[10], "flags", &sd->flags,
5223 sizeof(int), 0644, proc_dointvec_minmax);
5228 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5230 struct ctl_table *entry, *table;
5231 struct sched_domain *sd;
5232 int domain_num = 0, i;
5235 for_each_domain(cpu, sd)
5237 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5240 for_each_domain(cpu, sd) {
5241 snprintf(buf, 32, "domain%d", i);
5242 entry->procname = kstrdup(buf, GFP_KERNEL);
5244 entry->child = sd_alloc_ctl_domain_table(sd);
5251 static struct ctl_table_header *sd_sysctl_header;
5252 static void init_sched_domain_sysctl(void)
5254 int i, cpu_num = num_online_cpus();
5255 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5258 sd_ctl_dir[0].child = entry;
5260 for (i = 0; i < cpu_num; i++, entry++) {
5261 snprintf(buf, 32, "cpu%d", i);
5262 entry->procname = kstrdup(buf, GFP_KERNEL);
5264 entry->child = sd_alloc_ctl_cpu_table(i);
5266 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5269 static void init_sched_domain_sysctl(void)
5275 * migration_call - callback that gets triggered when a CPU is added.
5276 * Here we can start up the necessary migration thread for the new CPU.
5278 static int __cpuinit
5279 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5281 struct task_struct *p;
5282 int cpu = (long)hcpu;
5283 unsigned long flags;
5287 case CPU_LOCK_ACQUIRE:
5288 mutex_lock(&sched_hotcpu_mutex);
5291 case CPU_UP_PREPARE:
5292 case CPU_UP_PREPARE_FROZEN:
5293 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5296 kthread_bind(p, cpu);
5297 /* Must be high prio: stop_machine expects to yield to it. */
5298 rq = task_rq_lock(p, &flags);
5299 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5300 task_rq_unlock(rq, &flags);
5301 cpu_rq(cpu)->migration_thread = p;
5305 case CPU_ONLINE_FROZEN:
5306 /* Strictly unneccessary, as first user will wake it. */
5307 wake_up_process(cpu_rq(cpu)->migration_thread);
5310 #ifdef CONFIG_HOTPLUG_CPU
5311 case CPU_UP_CANCELED:
5312 case CPU_UP_CANCELED_FROZEN:
5313 if (!cpu_rq(cpu)->migration_thread)
5315 /* Unbind it from offline cpu so it can run. Fall thru. */
5316 kthread_bind(cpu_rq(cpu)->migration_thread,
5317 any_online_cpu(cpu_online_map));
5318 kthread_stop(cpu_rq(cpu)->migration_thread);
5319 cpu_rq(cpu)->migration_thread = NULL;
5323 case CPU_DEAD_FROZEN:
5324 migrate_live_tasks(cpu);
5326 kthread_stop(rq->migration_thread);
5327 rq->migration_thread = NULL;
5328 /* Idle task back to normal (off runqueue, low prio) */
5329 rq = task_rq_lock(rq->idle, &flags);
5330 update_rq_clock(rq);
5331 deactivate_task(rq, rq->idle, 0);
5332 rq->idle->static_prio = MAX_PRIO;
5333 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5334 rq->idle->sched_class = &idle_sched_class;
5335 migrate_dead_tasks(cpu);
5336 task_rq_unlock(rq, &flags);
5337 migrate_nr_uninterruptible(rq);
5338 BUG_ON(rq->nr_running != 0);
5340 /* No need to migrate the tasks: it was best-effort if
5341 * they didn't take sched_hotcpu_mutex. Just wake up
5342 * the requestors. */
5343 spin_lock_irq(&rq->lock);
5344 while (!list_empty(&rq->migration_queue)) {
5345 struct migration_req *req;
5347 req = list_entry(rq->migration_queue.next,
5348 struct migration_req, list);
5349 list_del_init(&req->list);
5350 complete(&req->done);
5352 spin_unlock_irq(&rq->lock);
5355 case CPU_LOCK_RELEASE:
5356 mutex_unlock(&sched_hotcpu_mutex);
5362 /* Register at highest priority so that task migration (migrate_all_tasks)
5363 * happens before everything else.
5365 static struct notifier_block __cpuinitdata migration_notifier = {
5366 .notifier_call = migration_call,
5370 int __init migration_init(void)
5372 void *cpu = (void *)(long)smp_processor_id();
5375 /* Start one for the boot CPU: */
5376 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5377 BUG_ON(err == NOTIFY_BAD);
5378 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5379 register_cpu_notifier(&migration_notifier);
5387 /* Number of possible processor ids */
5388 int nr_cpu_ids __read_mostly = NR_CPUS;
5389 EXPORT_SYMBOL(nr_cpu_ids);
5391 #ifdef CONFIG_SCHED_DEBUG
5392 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5397 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5401 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5406 struct sched_group *group = sd->groups;
5407 cpumask_t groupmask;
5409 cpumask_scnprintf(str, NR_CPUS, sd->span);
5410 cpus_clear(groupmask);
5413 for (i = 0; i < level + 1; i++)
5415 printk("domain %d: ", level);
5417 if (!(sd->flags & SD_LOAD_BALANCE)) {
5418 printk("does not load-balance\n");
5420 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5425 printk("span %s\n", str);
5427 if (!cpu_isset(cpu, sd->span))
5428 printk(KERN_ERR "ERROR: domain->span does not contain "
5430 if (!cpu_isset(cpu, group->cpumask))
5431 printk(KERN_ERR "ERROR: domain->groups does not contain"
5435 for (i = 0; i < level + 2; i++)
5441 printk(KERN_ERR "ERROR: group is NULL\n");
5445 if (!group->__cpu_power) {
5447 printk(KERN_ERR "ERROR: domain->cpu_power not "
5452 if (!cpus_weight(group->cpumask)) {
5454 printk(KERN_ERR "ERROR: empty group\n");
5458 if (cpus_intersects(groupmask, group->cpumask)) {
5460 printk(KERN_ERR "ERROR: repeated CPUs\n");
5464 cpus_or(groupmask, groupmask, group->cpumask);
5466 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5469 group = group->next;
5470 } while (group != sd->groups);
5473 if (!cpus_equal(sd->span, groupmask))
5474 printk(KERN_ERR "ERROR: groups don't span "
5482 if (!cpus_subset(groupmask, sd->span))
5483 printk(KERN_ERR "ERROR: parent span is not a superset "
5484 "of domain->span\n");
5489 # define sched_domain_debug(sd, cpu) do { } while (0)
5492 static int sd_degenerate(struct sched_domain *sd)
5494 if (cpus_weight(sd->span) == 1)
5497 /* Following flags need at least 2 groups */
5498 if (sd->flags & (SD_LOAD_BALANCE |
5499 SD_BALANCE_NEWIDLE |
5503 SD_SHARE_PKG_RESOURCES)) {
5504 if (sd->groups != sd->groups->next)
5508 /* Following flags don't use groups */
5509 if (sd->flags & (SD_WAKE_IDLE |
5518 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5520 unsigned long cflags = sd->flags, pflags = parent->flags;
5522 if (sd_degenerate(parent))
5525 if (!cpus_equal(sd->span, parent->span))
5528 /* Does parent contain flags not in child? */
5529 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5530 if (cflags & SD_WAKE_AFFINE)
5531 pflags &= ~SD_WAKE_BALANCE;
5532 /* Flags needing groups don't count if only 1 group in parent */
5533 if (parent->groups == parent->groups->next) {
5534 pflags &= ~(SD_LOAD_BALANCE |
5535 SD_BALANCE_NEWIDLE |
5539 SD_SHARE_PKG_RESOURCES);
5541 if (~cflags & pflags)
5548 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5549 * hold the hotplug lock.
5551 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5553 struct rq *rq = cpu_rq(cpu);
5554 struct sched_domain *tmp;
5556 /* Remove the sched domains which do not contribute to scheduling. */
5557 for (tmp = sd; tmp; tmp = tmp->parent) {
5558 struct sched_domain *parent = tmp->parent;
5561 if (sd_parent_degenerate(tmp, parent)) {
5562 tmp->parent = parent->parent;
5564 parent->parent->child = tmp;
5568 if (sd && sd_degenerate(sd)) {
5574 sched_domain_debug(sd, cpu);
5576 rcu_assign_pointer(rq->sd, sd);
5579 /* cpus with isolated domains */
5580 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5582 /* Setup the mask of cpus configured for isolated domains */
5583 static int __init isolated_cpu_setup(char *str)
5585 int ints[NR_CPUS], i;
5587 str = get_options(str, ARRAY_SIZE(ints), ints);
5588 cpus_clear(cpu_isolated_map);
5589 for (i = 1; i <= ints[0]; i++)
5590 if (ints[i] < NR_CPUS)
5591 cpu_set(ints[i], cpu_isolated_map);
5595 __setup("isolcpus=", isolated_cpu_setup);
5598 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5599 * to a function which identifies what group(along with sched group) a CPU
5600 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5601 * (due to the fact that we keep track of groups covered with a cpumask_t).
5603 * init_sched_build_groups will build a circular linked list of the groups
5604 * covered by the given span, and will set each group's ->cpumask correctly,
5605 * and ->cpu_power to 0.
5608 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5609 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5610 struct sched_group **sg))
5612 struct sched_group *first = NULL, *last = NULL;
5613 cpumask_t covered = CPU_MASK_NONE;
5616 for_each_cpu_mask(i, span) {
5617 struct sched_group *sg;
5618 int group = group_fn(i, cpu_map, &sg);
5621 if (cpu_isset(i, covered))
5624 sg->cpumask = CPU_MASK_NONE;
5625 sg->__cpu_power = 0;
5627 for_each_cpu_mask(j, span) {
5628 if (group_fn(j, cpu_map, NULL) != group)
5631 cpu_set(j, covered);
5632 cpu_set(j, sg->cpumask);
5643 #define SD_NODES_PER_DOMAIN 16
5648 * find_next_best_node - find the next node to include in a sched_domain
5649 * @node: node whose sched_domain we're building
5650 * @used_nodes: nodes already in the sched_domain
5652 * Find the next node to include in a given scheduling domain. Simply
5653 * finds the closest node not already in the @used_nodes map.
5655 * Should use nodemask_t.
5657 static int find_next_best_node(int node, unsigned long *used_nodes)
5659 int i, n, val, min_val, best_node = 0;
5663 for (i = 0; i < MAX_NUMNODES; i++) {
5664 /* Start at @node */
5665 n = (node + i) % MAX_NUMNODES;
5667 if (!nr_cpus_node(n))
5670 /* Skip already used nodes */
5671 if (test_bit(n, used_nodes))
5674 /* Simple min distance search */
5675 val = node_distance(node, n);
5677 if (val < min_val) {
5683 set_bit(best_node, used_nodes);
5688 * sched_domain_node_span - get a cpumask for a node's sched_domain
5689 * @node: node whose cpumask we're constructing
5690 * @size: number of nodes to include in this span
5692 * Given a node, construct a good cpumask for its sched_domain to span. It
5693 * should be one that prevents unnecessary balancing, but also spreads tasks
5696 static cpumask_t sched_domain_node_span(int node)
5698 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5699 cpumask_t span, nodemask;
5703 bitmap_zero(used_nodes, MAX_NUMNODES);
5705 nodemask = node_to_cpumask(node);
5706 cpus_or(span, span, nodemask);
5707 set_bit(node, used_nodes);
5709 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5710 int next_node = find_next_best_node(node, used_nodes);
5712 nodemask = node_to_cpumask(next_node);
5713 cpus_or(span, span, nodemask);
5720 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5723 * SMT sched-domains:
5725 #ifdef CONFIG_SCHED_SMT
5726 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5727 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5729 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5730 struct sched_group **sg)
5733 *sg = &per_cpu(sched_group_cpus, cpu);
5739 * multi-core sched-domains:
5741 #ifdef CONFIG_SCHED_MC
5742 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5743 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5746 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5747 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5748 struct sched_group **sg)
5751 cpumask_t mask = cpu_sibling_map[cpu];
5752 cpus_and(mask, mask, *cpu_map);
5753 group = first_cpu(mask);
5755 *sg = &per_cpu(sched_group_core, group);
5758 #elif defined(CONFIG_SCHED_MC)
5759 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5760 struct sched_group **sg)
5763 *sg = &per_cpu(sched_group_core, cpu);
5768 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5769 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5771 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5772 struct sched_group **sg)
5775 #ifdef CONFIG_SCHED_MC
5776 cpumask_t mask = cpu_coregroup_map(cpu);
5777 cpus_and(mask, mask, *cpu_map);
5778 group = first_cpu(mask);
5779 #elif defined(CONFIG_SCHED_SMT)
5780 cpumask_t mask = cpu_sibling_map[cpu];
5781 cpus_and(mask, mask, *cpu_map);
5782 group = first_cpu(mask);
5787 *sg = &per_cpu(sched_group_phys, group);
5793 * The init_sched_build_groups can't handle what we want to do with node
5794 * groups, so roll our own. Now each node has its own list of groups which
5795 * gets dynamically allocated.
5797 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5798 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5800 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5801 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5803 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5804 struct sched_group **sg)
5806 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5809 cpus_and(nodemask, nodemask, *cpu_map);
5810 group = first_cpu(nodemask);
5813 *sg = &per_cpu(sched_group_allnodes, group);
5817 static void init_numa_sched_groups_power(struct sched_group *group_head)
5819 struct sched_group *sg = group_head;
5825 for_each_cpu_mask(j, sg->cpumask) {
5826 struct sched_domain *sd;
5828 sd = &per_cpu(phys_domains, j);
5829 if (j != first_cpu(sd->groups->cpumask)) {
5831 * Only add "power" once for each
5837 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5840 } while (sg != group_head);
5845 /* Free memory allocated for various sched_group structures */
5846 static void free_sched_groups(const cpumask_t *cpu_map)
5850 for_each_cpu_mask(cpu, *cpu_map) {
5851 struct sched_group **sched_group_nodes
5852 = sched_group_nodes_bycpu[cpu];
5854 if (!sched_group_nodes)
5857 for (i = 0; i < MAX_NUMNODES; i++) {
5858 cpumask_t nodemask = node_to_cpumask(i);
5859 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5861 cpus_and(nodemask, nodemask, *cpu_map);
5862 if (cpus_empty(nodemask))
5872 if (oldsg != sched_group_nodes[i])
5875 kfree(sched_group_nodes);
5876 sched_group_nodes_bycpu[cpu] = NULL;
5880 static void free_sched_groups(const cpumask_t *cpu_map)
5886 * Initialize sched groups cpu_power.
5888 * cpu_power indicates the capacity of sched group, which is used while
5889 * distributing the load between different sched groups in a sched domain.
5890 * Typically cpu_power for all the groups in a sched domain will be same unless
5891 * there are asymmetries in the topology. If there are asymmetries, group
5892 * having more cpu_power will pickup more load compared to the group having
5895 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5896 * the maximum number of tasks a group can handle in the presence of other idle
5897 * or lightly loaded groups in the same sched domain.
5899 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5901 struct sched_domain *child;
5902 struct sched_group *group;
5904 WARN_ON(!sd || !sd->groups);
5906 if (cpu != first_cpu(sd->groups->cpumask))
5911 sd->groups->__cpu_power = 0;
5914 * For perf policy, if the groups in child domain share resources
5915 * (for example cores sharing some portions of the cache hierarchy
5916 * or SMT), then set this domain groups cpu_power such that each group
5917 * can handle only one task, when there are other idle groups in the
5918 * same sched domain.
5920 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5922 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5923 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5928 * add cpu_power of each child group to this groups cpu_power
5930 group = child->groups;
5932 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5933 group = group->next;
5934 } while (group != child->groups);
5938 * Build sched domains for a given set of cpus and attach the sched domains
5939 * to the individual cpus
5941 static int build_sched_domains(const cpumask_t *cpu_map)
5945 struct sched_group **sched_group_nodes = NULL;
5946 int sd_allnodes = 0;
5949 * Allocate the per-node list of sched groups
5951 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
5953 if (!sched_group_nodes) {
5954 printk(KERN_WARNING "Can not alloc sched group node list\n");
5957 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5961 * Set up domains for cpus specified by the cpu_map.
5963 for_each_cpu_mask(i, *cpu_map) {
5964 struct sched_domain *sd = NULL, *p;
5965 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5967 cpus_and(nodemask, nodemask, *cpu_map);
5970 if (cpus_weight(*cpu_map) >
5971 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
5972 sd = &per_cpu(allnodes_domains, i);
5973 *sd = SD_ALLNODES_INIT;
5974 sd->span = *cpu_map;
5975 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
5981 sd = &per_cpu(node_domains, i);
5983 sd->span = sched_domain_node_span(cpu_to_node(i));
5987 cpus_and(sd->span, sd->span, *cpu_map);
5991 sd = &per_cpu(phys_domains, i);
5993 sd->span = nodemask;
5997 cpu_to_phys_group(i, cpu_map, &sd->groups);
5999 #ifdef CONFIG_SCHED_MC
6001 sd = &per_cpu(core_domains, i);
6003 sd->span = cpu_coregroup_map(i);
6004 cpus_and(sd->span, sd->span, *cpu_map);
6007 cpu_to_core_group(i, cpu_map, &sd->groups);
6010 #ifdef CONFIG_SCHED_SMT
6012 sd = &per_cpu(cpu_domains, i);
6013 *sd = SD_SIBLING_INIT;
6014 sd->span = cpu_sibling_map[i];
6015 cpus_and(sd->span, sd->span, *cpu_map);
6018 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6022 #ifdef CONFIG_SCHED_SMT
6023 /* Set up CPU (sibling) groups */
6024 for_each_cpu_mask(i, *cpu_map) {
6025 cpumask_t this_sibling_map = cpu_sibling_map[i];
6026 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6027 if (i != first_cpu(this_sibling_map))
6030 init_sched_build_groups(this_sibling_map, cpu_map,
6035 #ifdef CONFIG_SCHED_MC
6036 /* Set up multi-core groups */
6037 for_each_cpu_mask(i, *cpu_map) {
6038 cpumask_t this_core_map = cpu_coregroup_map(i);
6039 cpus_and(this_core_map, this_core_map, *cpu_map);
6040 if (i != first_cpu(this_core_map))
6042 init_sched_build_groups(this_core_map, cpu_map,
6043 &cpu_to_core_group);
6047 /* Set up physical groups */
6048 for (i = 0; i < MAX_NUMNODES; i++) {
6049 cpumask_t nodemask = node_to_cpumask(i);
6051 cpus_and(nodemask, nodemask, *cpu_map);
6052 if (cpus_empty(nodemask))
6055 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6059 /* Set up node groups */
6061 init_sched_build_groups(*cpu_map, cpu_map,
6062 &cpu_to_allnodes_group);
6064 for (i = 0; i < MAX_NUMNODES; i++) {
6065 /* Set up node groups */
6066 struct sched_group *sg, *prev;
6067 cpumask_t nodemask = node_to_cpumask(i);
6068 cpumask_t domainspan;
6069 cpumask_t covered = CPU_MASK_NONE;
6072 cpus_and(nodemask, nodemask, *cpu_map);
6073 if (cpus_empty(nodemask)) {
6074 sched_group_nodes[i] = NULL;
6078 domainspan = sched_domain_node_span(i);
6079 cpus_and(domainspan, domainspan, *cpu_map);
6081 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6083 printk(KERN_WARNING "Can not alloc domain group for "
6087 sched_group_nodes[i] = sg;
6088 for_each_cpu_mask(j, nodemask) {
6089 struct sched_domain *sd;
6091 sd = &per_cpu(node_domains, j);
6094 sg->__cpu_power = 0;
6095 sg->cpumask = nodemask;
6097 cpus_or(covered, covered, nodemask);
6100 for (j = 0; j < MAX_NUMNODES; j++) {
6101 cpumask_t tmp, notcovered;
6102 int n = (i + j) % MAX_NUMNODES;
6104 cpus_complement(notcovered, covered);
6105 cpus_and(tmp, notcovered, *cpu_map);
6106 cpus_and(tmp, tmp, domainspan);
6107 if (cpus_empty(tmp))
6110 nodemask = node_to_cpumask(n);
6111 cpus_and(tmp, tmp, nodemask);
6112 if (cpus_empty(tmp))
6115 sg = kmalloc_node(sizeof(struct sched_group),
6119 "Can not alloc domain group for node %d\n", j);
6122 sg->__cpu_power = 0;
6124 sg->next = prev->next;
6125 cpus_or(covered, covered, tmp);
6132 /* Calculate CPU power for physical packages and nodes */
6133 #ifdef CONFIG_SCHED_SMT
6134 for_each_cpu_mask(i, *cpu_map) {
6135 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6137 init_sched_groups_power(i, sd);
6140 #ifdef CONFIG_SCHED_MC
6141 for_each_cpu_mask(i, *cpu_map) {
6142 struct sched_domain *sd = &per_cpu(core_domains, i);
6144 init_sched_groups_power(i, sd);
6148 for_each_cpu_mask(i, *cpu_map) {
6149 struct sched_domain *sd = &per_cpu(phys_domains, i);
6151 init_sched_groups_power(i, sd);
6155 for (i = 0; i < MAX_NUMNODES; i++)
6156 init_numa_sched_groups_power(sched_group_nodes[i]);
6159 struct sched_group *sg;
6161 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6162 init_numa_sched_groups_power(sg);
6166 /* Attach the domains */
6167 for_each_cpu_mask(i, *cpu_map) {
6168 struct sched_domain *sd;
6169 #ifdef CONFIG_SCHED_SMT
6170 sd = &per_cpu(cpu_domains, i);
6171 #elif defined(CONFIG_SCHED_MC)
6172 sd = &per_cpu(core_domains, i);
6174 sd = &per_cpu(phys_domains, i);
6176 cpu_attach_domain(sd, i);
6183 free_sched_groups(cpu_map);
6188 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6190 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6192 cpumask_t cpu_default_map;
6196 * Setup mask for cpus without special case scheduling requirements.
6197 * For now this just excludes isolated cpus, but could be used to
6198 * exclude other special cases in the future.
6200 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6202 err = build_sched_domains(&cpu_default_map);
6207 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6209 free_sched_groups(cpu_map);
6213 * Detach sched domains from a group of cpus specified in cpu_map
6214 * These cpus will now be attached to the NULL domain
6216 static void detach_destroy_domains(const cpumask_t *cpu_map)
6220 for_each_cpu_mask(i, *cpu_map)
6221 cpu_attach_domain(NULL, i);
6222 synchronize_sched();
6223 arch_destroy_sched_domains(cpu_map);
6227 * Partition sched domains as specified by the cpumasks below.
6228 * This attaches all cpus from the cpumasks to the NULL domain,
6229 * waits for a RCU quiescent period, recalculates sched
6230 * domain information and then attaches them back to the
6231 * correct sched domains
6232 * Call with hotplug lock held
6234 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6236 cpumask_t change_map;
6239 cpus_and(*partition1, *partition1, cpu_online_map);
6240 cpus_and(*partition2, *partition2, cpu_online_map);
6241 cpus_or(change_map, *partition1, *partition2);
6243 /* Detach sched domains from all of the affected cpus */
6244 detach_destroy_domains(&change_map);
6245 if (!cpus_empty(*partition1))
6246 err = build_sched_domains(partition1);
6247 if (!err && !cpus_empty(*partition2))
6248 err = build_sched_domains(partition2);
6253 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6254 static int arch_reinit_sched_domains(void)
6258 mutex_lock(&sched_hotcpu_mutex);
6259 detach_destroy_domains(&cpu_online_map);
6260 err = arch_init_sched_domains(&cpu_online_map);
6261 mutex_unlock(&sched_hotcpu_mutex);
6266 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6270 if (buf[0] != '0' && buf[0] != '1')
6274 sched_smt_power_savings = (buf[0] == '1');
6276 sched_mc_power_savings = (buf[0] == '1');
6278 ret = arch_reinit_sched_domains();
6280 return ret ? ret : count;
6283 #ifdef CONFIG_SCHED_MC
6284 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6286 return sprintf(page, "%u\n", sched_mc_power_savings);
6288 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6289 const char *buf, size_t count)
6291 return sched_power_savings_store(buf, count, 0);
6293 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6294 sched_mc_power_savings_store);
6297 #ifdef CONFIG_SCHED_SMT
6298 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6300 return sprintf(page, "%u\n", sched_smt_power_savings);
6302 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6303 const char *buf, size_t count)
6305 return sched_power_savings_store(buf, count, 1);
6307 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6308 sched_smt_power_savings_store);
6311 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6315 #ifdef CONFIG_SCHED_SMT
6317 err = sysfs_create_file(&cls->kset.kobj,
6318 &attr_sched_smt_power_savings.attr);
6320 #ifdef CONFIG_SCHED_MC
6321 if (!err && mc_capable())
6322 err = sysfs_create_file(&cls->kset.kobj,
6323 &attr_sched_mc_power_savings.attr);
6330 * Force a reinitialization of the sched domains hierarchy. The domains
6331 * and groups cannot be updated in place without racing with the balancing
6332 * code, so we temporarily attach all running cpus to the NULL domain
6333 * which will prevent rebalancing while the sched domains are recalculated.
6335 static int update_sched_domains(struct notifier_block *nfb,
6336 unsigned long action, void *hcpu)
6339 case CPU_UP_PREPARE:
6340 case CPU_UP_PREPARE_FROZEN:
6341 case CPU_DOWN_PREPARE:
6342 case CPU_DOWN_PREPARE_FROZEN:
6343 detach_destroy_domains(&cpu_online_map);
6346 case CPU_UP_CANCELED:
6347 case CPU_UP_CANCELED_FROZEN:
6348 case CPU_DOWN_FAILED:
6349 case CPU_DOWN_FAILED_FROZEN:
6351 case CPU_ONLINE_FROZEN:
6353 case CPU_DEAD_FROZEN:
6355 * Fall through and re-initialise the domains.
6362 /* The hotplug lock is already held by cpu_up/cpu_down */
6363 arch_init_sched_domains(&cpu_online_map);
6368 void __init sched_init_smp(void)
6370 cpumask_t non_isolated_cpus;
6372 mutex_lock(&sched_hotcpu_mutex);
6373 arch_init_sched_domains(&cpu_online_map);
6374 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6375 if (cpus_empty(non_isolated_cpus))
6376 cpu_set(smp_processor_id(), non_isolated_cpus);
6377 mutex_unlock(&sched_hotcpu_mutex);
6378 /* XXX: Theoretical race here - CPU may be hotplugged now */
6379 hotcpu_notifier(update_sched_domains, 0);
6381 init_sched_domain_sysctl();
6383 /* Move init over to a non-isolated CPU */
6384 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6388 void __init sched_init_smp(void)
6391 #endif /* CONFIG_SMP */
6393 int in_sched_functions(unsigned long addr)
6395 /* Linker adds these: start and end of __sched functions */
6396 extern char __sched_text_start[], __sched_text_end[];
6398 return in_lock_functions(addr) ||
6399 (addr >= (unsigned long)__sched_text_start
6400 && addr < (unsigned long)__sched_text_end);
6403 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6405 cfs_rq->tasks_timeline = RB_ROOT;
6406 #ifdef CONFIG_FAIR_GROUP_SCHED
6409 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6412 void __init sched_init(void)
6414 int highest_cpu = 0;
6417 for_each_possible_cpu(i) {
6418 struct rt_prio_array *array;
6422 spin_lock_init(&rq->lock);
6423 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6426 init_cfs_rq(&rq->cfs, rq);
6427 #ifdef CONFIG_FAIR_GROUP_SCHED
6428 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6430 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6431 struct sched_entity *se =
6432 &per_cpu(init_sched_entity, i);
6434 init_cfs_rq_p[i] = cfs_rq;
6435 init_cfs_rq(cfs_rq, rq);
6436 cfs_rq->tg = &init_task_group;
6437 list_add(&cfs_rq->leaf_cfs_rq_list,
6438 &rq->leaf_cfs_rq_list);
6440 init_sched_entity_p[i] = se;
6441 se->cfs_rq = &rq->cfs;
6443 se->load.weight = init_task_group_load;
6444 se->load.inv_weight =
6445 div64_64(1ULL<<32, init_task_group_load);
6448 init_task_group.shares = init_task_group_load;
6449 spin_lock_init(&init_task_group.lock);
6452 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6453 rq->cpu_load[j] = 0;
6456 rq->active_balance = 0;
6457 rq->next_balance = jiffies;
6460 rq->migration_thread = NULL;
6461 INIT_LIST_HEAD(&rq->migration_queue);
6463 atomic_set(&rq->nr_iowait, 0);
6465 array = &rq->rt.active;
6466 for (j = 0; j < MAX_RT_PRIO; j++) {
6467 INIT_LIST_HEAD(array->queue + j);
6468 __clear_bit(j, array->bitmap);
6471 /* delimiter for bitsearch: */
6472 __set_bit(MAX_RT_PRIO, array->bitmap);
6475 set_load_weight(&init_task);
6477 #ifdef CONFIG_PREEMPT_NOTIFIERS
6478 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6482 nr_cpu_ids = highest_cpu + 1;
6483 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6486 #ifdef CONFIG_RT_MUTEXES
6487 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6491 * The boot idle thread does lazy MMU switching as well:
6493 atomic_inc(&init_mm.mm_count);
6494 enter_lazy_tlb(&init_mm, current);
6497 * Make us the idle thread. Technically, schedule() should not be
6498 * called from this thread, however somewhere below it might be,
6499 * but because we are the idle thread, we just pick up running again
6500 * when this runqueue becomes "idle".
6502 init_idle(current, smp_processor_id());
6504 * During early bootup we pretend to be a normal task:
6506 current->sched_class = &fair_sched_class;
6509 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6510 void __might_sleep(char *file, int line)
6513 static unsigned long prev_jiffy; /* ratelimiting */
6515 if ((in_atomic() || irqs_disabled()) &&
6516 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6517 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6519 prev_jiffy = jiffies;
6520 printk(KERN_ERR "BUG: sleeping function called from invalid"
6521 " context at %s:%d\n", file, line);
6522 printk("in_atomic():%d, irqs_disabled():%d\n",
6523 in_atomic(), irqs_disabled());
6524 debug_show_held_locks(current);
6525 if (irqs_disabled())
6526 print_irqtrace_events(current);
6531 EXPORT_SYMBOL(__might_sleep);
6534 #ifdef CONFIG_MAGIC_SYSRQ
6535 void normalize_rt_tasks(void)
6537 struct task_struct *g, *p;
6538 unsigned long flags;
6542 read_lock_irq(&tasklist_lock);
6543 do_each_thread(g, p) {
6544 p->se.exec_start = 0;
6545 #ifdef CONFIG_SCHEDSTATS
6546 p->se.wait_start = 0;
6547 p->se.sleep_start = 0;
6548 p->se.block_start = 0;
6550 task_rq(p)->clock = 0;
6554 * Renice negative nice level userspace
6557 if (TASK_NICE(p) < 0 && p->mm)
6558 set_user_nice(p, 0);
6562 spin_lock_irqsave(&p->pi_lock, flags);
6563 rq = __task_rq_lock(p);
6566 * Do not touch the migration thread:
6568 if (p == rq->migration_thread)
6572 update_rq_clock(rq);
6573 on_rq = p->se.on_rq;
6575 deactivate_task(rq, p, 0);
6576 __setscheduler(rq, p, SCHED_NORMAL, 0);
6578 activate_task(rq, p, 0);
6579 resched_task(rq->curr);
6584 __task_rq_unlock(rq);
6585 spin_unlock_irqrestore(&p->pi_lock, flags);
6586 } while_each_thread(g, p);
6588 read_unlock_irq(&tasklist_lock);
6591 #endif /* CONFIG_MAGIC_SYSRQ */
6595 * These functions are only useful for the IA64 MCA handling.
6597 * They can only be called when the whole system has been
6598 * stopped - every CPU needs to be quiescent, and no scheduling
6599 * activity can take place. Using them for anything else would
6600 * be a serious bug, and as a result, they aren't even visible
6601 * under any other configuration.
6605 * curr_task - return the current task for a given cpu.
6606 * @cpu: the processor in question.
6608 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6610 struct task_struct *curr_task(int cpu)
6612 return cpu_curr(cpu);
6616 * set_curr_task - set the current task for a given cpu.
6617 * @cpu: the processor in question.
6618 * @p: the task pointer to set.
6620 * Description: This function must only be used when non-maskable interrupts
6621 * are serviced on a separate stack. It allows the architecture to switch the
6622 * notion of the current task on a cpu in a non-blocking manner. This function
6623 * must be called with all CPU's synchronized, and interrupts disabled, the
6624 * and caller must save the original value of the current task (see
6625 * curr_task() above) and restore that value before reenabling interrupts and
6626 * re-starting the system.
6628 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6630 void set_curr_task(int cpu, struct task_struct *p)
6637 #ifdef CONFIG_FAIR_GROUP_SCHED
6639 /* allocate runqueue etc for a new task group */
6640 struct task_group *sched_create_group(void)
6642 struct task_group *tg;
6643 struct cfs_rq *cfs_rq;
6644 struct sched_entity *se;
6648 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6650 return ERR_PTR(-ENOMEM);
6652 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6655 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6659 for_each_possible_cpu(i) {
6662 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6667 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6672 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6673 memset(se, 0, sizeof(struct sched_entity));
6675 tg->cfs_rq[i] = cfs_rq;
6676 init_cfs_rq(cfs_rq, rq);
6680 se->cfs_rq = &rq->cfs;
6682 se->load.weight = NICE_0_LOAD;
6683 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6687 for_each_possible_cpu(i) {
6689 cfs_rq = tg->cfs_rq[i];
6690 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6693 tg->shares = NICE_0_LOAD;
6694 spin_lock_init(&tg->lock);
6699 for_each_possible_cpu(i) {
6701 kfree(tg->cfs_rq[i]);
6709 return ERR_PTR(-ENOMEM);
6712 /* rcu callback to free various structures associated with a task group */
6713 static void free_sched_group(struct rcu_head *rhp)
6715 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6716 struct task_group *tg = cfs_rq->tg;
6717 struct sched_entity *se;
6720 /* now it should be safe to free those cfs_rqs */
6721 for_each_possible_cpu(i) {
6722 cfs_rq = tg->cfs_rq[i];
6734 /* Destroy runqueue etc associated with a task group */
6735 void sched_destroy_group(struct task_group *tg)
6737 struct cfs_rq *cfs_rq;
6740 for_each_possible_cpu(i) {
6741 cfs_rq = tg->cfs_rq[i];
6742 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6745 cfs_rq = tg->cfs_rq[0];
6747 /* wait for possible concurrent references to cfs_rqs complete */
6748 call_rcu(&cfs_rq->rcu, free_sched_group);
6751 /* change task's runqueue when it moves between groups.
6752 * The caller of this function should have put the task in its new group
6753 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6754 * reflect its new group.
6756 void sched_move_task(struct task_struct *tsk)
6759 unsigned long flags;
6762 rq = task_rq_lock(tsk, &flags);
6764 if (tsk->sched_class != &fair_sched_class)
6767 update_rq_clock(rq);
6769 running = task_running(rq, tsk);
6770 on_rq = tsk->se.on_rq;
6773 dequeue_task(rq, tsk, 0);
6774 if (unlikely(running))
6775 tsk->sched_class->put_prev_task(rq, tsk);
6778 set_task_cfs_rq(tsk);
6781 if (unlikely(running))
6782 tsk->sched_class->set_curr_task(rq);
6783 enqueue_task(rq, tsk, 0);
6787 task_rq_unlock(rq, &flags);
6790 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6792 struct cfs_rq *cfs_rq = se->cfs_rq;
6793 struct rq *rq = cfs_rq->rq;
6796 spin_lock_irq(&rq->lock);
6800 dequeue_entity(cfs_rq, se, 0);
6802 se->load.weight = shares;
6803 se->load.inv_weight = div64_64((1ULL<<32), shares);
6806 enqueue_entity(cfs_rq, se, 0);
6808 spin_unlock_irq(&rq->lock);
6811 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6815 spin_lock(&tg->lock);
6816 if (tg->shares == shares)
6819 /* return -EINVAL if the new value is not sane */
6821 tg->shares = shares;
6822 for_each_possible_cpu(i)
6823 set_se_shares(tg->se[i], shares);
6826 spin_unlock(&tg->lock);
6830 unsigned long sched_group_shares(struct task_group *tg)
6835 #endif /* CONFIG_FAIR_GROUP_SCHED */