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 * Is this task likely cache-hot:
2125 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2129 if (p->sched_class != &fair_sched_class)
2132 delta = now - p->se.exec_start;
2134 return delta < (s64)sysctl_sched_migration_cost;
2138 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2141 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2142 struct sched_domain *sd, enum cpu_idle_type idle,
2146 * We do not migrate tasks that are:
2147 * 1) running (obviously), or
2148 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2149 * 3) are cache-hot on their current CPU.
2151 if (!cpu_isset(this_cpu, p->cpus_allowed))
2155 if (task_running(rq, p))
2159 * Aggressive migration if:
2160 * 1) task is cache cold, or
2161 * 2) too many balance attempts have failed.
2164 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2165 #ifdef CONFIG_SCHEDSTATS
2166 if (task_hot(p, rq->clock, sd))
2167 schedstat_inc(sd, lb_hot_gained[idle]);
2172 if (task_hot(p, rq->clock, sd))
2177 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2178 unsigned long max_nr_move, unsigned long max_load_move,
2179 struct sched_domain *sd, enum cpu_idle_type idle,
2180 int *all_pinned, unsigned long *load_moved,
2181 int *this_best_prio, struct rq_iterator *iterator)
2183 int pulled = 0, pinned = 0, skip_for_load;
2184 struct task_struct *p;
2185 long rem_load_move = max_load_move;
2187 if (max_nr_move == 0 || max_load_move == 0)
2193 * Start the load-balancing iterator:
2195 p = iterator->start(iterator->arg);
2200 * To help distribute high priority tasks accross CPUs we don't
2201 * skip a task if it will be the highest priority task (i.e. smallest
2202 * prio value) on its new queue regardless of its load weight
2204 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2205 SCHED_LOAD_SCALE_FUZZ;
2206 if ((skip_for_load && p->prio >= *this_best_prio) ||
2207 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2208 p = iterator->next(iterator->arg);
2212 pull_task(busiest, p, this_rq, this_cpu);
2214 rem_load_move -= p->se.load.weight;
2217 * We only want to steal up to the prescribed number of tasks
2218 * and the prescribed amount of weighted load.
2220 if (pulled < max_nr_move && rem_load_move > 0) {
2221 if (p->prio < *this_best_prio)
2222 *this_best_prio = p->prio;
2223 p = iterator->next(iterator->arg);
2228 * Right now, this is the only place pull_task() is called,
2229 * so we can safely collect pull_task() stats here rather than
2230 * inside pull_task().
2232 schedstat_add(sd, lb_gained[idle], pulled);
2235 *all_pinned = pinned;
2236 *load_moved = max_load_move - rem_load_move;
2241 * move_tasks tries to move up to max_load_move weighted load from busiest to
2242 * this_rq, as part of a balancing operation within domain "sd".
2243 * Returns 1 if successful and 0 otherwise.
2245 * Called with both runqueues locked.
2247 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2248 unsigned long max_load_move,
2249 struct sched_domain *sd, enum cpu_idle_type idle,
2252 const struct sched_class *class = sched_class_highest;
2253 unsigned long total_load_moved = 0;
2254 int this_best_prio = this_rq->curr->prio;
2258 class->load_balance(this_rq, this_cpu, busiest,
2259 ULONG_MAX, max_load_move - total_load_moved,
2260 sd, idle, all_pinned, &this_best_prio);
2261 class = class->next;
2262 } while (class && max_load_move > total_load_moved);
2264 return total_load_moved > 0;
2268 * move_one_task tries to move exactly one task from busiest to this_rq, as
2269 * part of active balancing operations within "domain".
2270 * Returns 1 if successful and 0 otherwise.
2272 * Called with both runqueues locked.
2274 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2275 struct sched_domain *sd, enum cpu_idle_type idle)
2277 const struct sched_class *class;
2278 int this_best_prio = MAX_PRIO;
2280 for (class = sched_class_highest; class; class = class->next)
2281 if (class->load_balance(this_rq, this_cpu, busiest,
2282 1, ULONG_MAX, sd, idle, NULL,
2290 * find_busiest_group finds and returns the busiest CPU group within the
2291 * domain. It calculates and returns the amount of weighted load which
2292 * should be moved to restore balance via the imbalance parameter.
2294 static struct sched_group *
2295 find_busiest_group(struct sched_domain *sd, int this_cpu,
2296 unsigned long *imbalance, enum cpu_idle_type idle,
2297 int *sd_idle, cpumask_t *cpus, int *balance)
2299 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2300 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2301 unsigned long max_pull;
2302 unsigned long busiest_load_per_task, busiest_nr_running;
2303 unsigned long this_load_per_task, this_nr_running;
2305 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2306 int power_savings_balance = 1;
2307 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2308 unsigned long min_nr_running = ULONG_MAX;
2309 struct sched_group *group_min = NULL, *group_leader = NULL;
2312 max_load = this_load = total_load = total_pwr = 0;
2313 busiest_load_per_task = busiest_nr_running = 0;
2314 this_load_per_task = this_nr_running = 0;
2315 if (idle == CPU_NOT_IDLE)
2316 load_idx = sd->busy_idx;
2317 else if (idle == CPU_NEWLY_IDLE)
2318 load_idx = sd->newidle_idx;
2320 load_idx = sd->idle_idx;
2323 unsigned long load, group_capacity;
2326 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2327 unsigned long sum_nr_running, sum_weighted_load;
2329 local_group = cpu_isset(this_cpu, group->cpumask);
2332 balance_cpu = first_cpu(group->cpumask);
2334 /* Tally up the load of all CPUs in the group */
2335 sum_weighted_load = sum_nr_running = avg_load = 0;
2337 for_each_cpu_mask(i, group->cpumask) {
2340 if (!cpu_isset(i, *cpus))
2345 if (*sd_idle && rq->nr_running)
2348 /* Bias balancing toward cpus of our domain */
2350 if (idle_cpu(i) && !first_idle_cpu) {
2355 load = target_load(i, load_idx);
2357 load = source_load(i, load_idx);
2360 sum_nr_running += rq->nr_running;
2361 sum_weighted_load += weighted_cpuload(i);
2365 * First idle cpu or the first cpu(busiest) in this sched group
2366 * is eligible for doing load balancing at this and above
2367 * domains. In the newly idle case, we will allow all the cpu's
2368 * to do the newly idle load balance.
2370 if (idle != CPU_NEWLY_IDLE && local_group &&
2371 balance_cpu != this_cpu && balance) {
2376 total_load += avg_load;
2377 total_pwr += group->__cpu_power;
2379 /* Adjust by relative CPU power of the group */
2380 avg_load = sg_div_cpu_power(group,
2381 avg_load * SCHED_LOAD_SCALE);
2383 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2386 this_load = avg_load;
2388 this_nr_running = sum_nr_running;
2389 this_load_per_task = sum_weighted_load;
2390 } else if (avg_load > max_load &&
2391 sum_nr_running > group_capacity) {
2392 max_load = avg_load;
2394 busiest_nr_running = sum_nr_running;
2395 busiest_load_per_task = sum_weighted_load;
2398 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2400 * Busy processors will not participate in power savings
2403 if (idle == CPU_NOT_IDLE ||
2404 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2408 * If the local group is idle or completely loaded
2409 * no need to do power savings balance at this domain
2411 if (local_group && (this_nr_running >= group_capacity ||
2413 power_savings_balance = 0;
2416 * If a group is already running at full capacity or idle,
2417 * don't include that group in power savings calculations
2419 if (!power_savings_balance || sum_nr_running >= group_capacity
2424 * Calculate the group which has the least non-idle load.
2425 * This is the group from where we need to pick up the load
2428 if ((sum_nr_running < min_nr_running) ||
2429 (sum_nr_running == min_nr_running &&
2430 first_cpu(group->cpumask) <
2431 first_cpu(group_min->cpumask))) {
2433 min_nr_running = sum_nr_running;
2434 min_load_per_task = sum_weighted_load /
2439 * Calculate the group which is almost near its
2440 * capacity but still has some space to pick up some load
2441 * from other group and save more power
2443 if (sum_nr_running <= group_capacity - 1) {
2444 if (sum_nr_running > leader_nr_running ||
2445 (sum_nr_running == leader_nr_running &&
2446 first_cpu(group->cpumask) >
2447 first_cpu(group_leader->cpumask))) {
2448 group_leader = group;
2449 leader_nr_running = sum_nr_running;
2454 group = group->next;
2455 } while (group != sd->groups);
2457 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2460 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2462 if (this_load >= avg_load ||
2463 100*max_load <= sd->imbalance_pct*this_load)
2466 busiest_load_per_task /= busiest_nr_running;
2468 * We're trying to get all the cpus to the average_load, so we don't
2469 * want to push ourselves above the average load, nor do we wish to
2470 * reduce the max loaded cpu below the average load, as either of these
2471 * actions would just result in more rebalancing later, and ping-pong
2472 * tasks around. Thus we look for the minimum possible imbalance.
2473 * Negative imbalances (*we* are more loaded than anyone else) will
2474 * be counted as no imbalance for these purposes -- we can't fix that
2475 * by pulling tasks to us. Be careful of negative numbers as they'll
2476 * appear as very large values with unsigned longs.
2478 if (max_load <= busiest_load_per_task)
2482 * In the presence of smp nice balancing, certain scenarios can have
2483 * max load less than avg load(as we skip the groups at or below
2484 * its cpu_power, while calculating max_load..)
2486 if (max_load < avg_load) {
2488 goto small_imbalance;
2491 /* Don't want to pull so many tasks that a group would go idle */
2492 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2494 /* How much load to actually move to equalise the imbalance */
2495 *imbalance = min(max_pull * busiest->__cpu_power,
2496 (avg_load - this_load) * this->__cpu_power)
2500 * if *imbalance is less than the average load per runnable task
2501 * there is no gaurantee that any tasks will be moved so we'll have
2502 * a think about bumping its value to force at least one task to be
2505 if (*imbalance < busiest_load_per_task) {
2506 unsigned long tmp, pwr_now, pwr_move;
2510 pwr_move = pwr_now = 0;
2512 if (this_nr_running) {
2513 this_load_per_task /= this_nr_running;
2514 if (busiest_load_per_task > this_load_per_task)
2517 this_load_per_task = SCHED_LOAD_SCALE;
2519 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2520 busiest_load_per_task * imbn) {
2521 *imbalance = busiest_load_per_task;
2526 * OK, we don't have enough imbalance to justify moving tasks,
2527 * however we may be able to increase total CPU power used by
2531 pwr_now += busiest->__cpu_power *
2532 min(busiest_load_per_task, max_load);
2533 pwr_now += this->__cpu_power *
2534 min(this_load_per_task, this_load);
2535 pwr_now /= SCHED_LOAD_SCALE;
2537 /* Amount of load we'd subtract */
2538 tmp = sg_div_cpu_power(busiest,
2539 busiest_load_per_task * SCHED_LOAD_SCALE);
2541 pwr_move += busiest->__cpu_power *
2542 min(busiest_load_per_task, max_load - tmp);
2544 /* Amount of load we'd add */
2545 if (max_load * busiest->__cpu_power <
2546 busiest_load_per_task * SCHED_LOAD_SCALE)
2547 tmp = sg_div_cpu_power(this,
2548 max_load * busiest->__cpu_power);
2550 tmp = sg_div_cpu_power(this,
2551 busiest_load_per_task * SCHED_LOAD_SCALE);
2552 pwr_move += this->__cpu_power *
2553 min(this_load_per_task, this_load + tmp);
2554 pwr_move /= SCHED_LOAD_SCALE;
2556 /* Move if we gain throughput */
2557 if (pwr_move > pwr_now)
2558 *imbalance = busiest_load_per_task;
2564 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2565 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2568 if (this == group_leader && group_leader != group_min) {
2569 *imbalance = min_load_per_task;
2579 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2582 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2583 unsigned long imbalance, cpumask_t *cpus)
2585 struct rq *busiest = NULL, *rq;
2586 unsigned long max_load = 0;
2589 for_each_cpu_mask(i, group->cpumask) {
2592 if (!cpu_isset(i, *cpus))
2596 wl = weighted_cpuload(i);
2598 if (rq->nr_running == 1 && wl > imbalance)
2601 if (wl > max_load) {
2611 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2612 * so long as it is large enough.
2614 #define MAX_PINNED_INTERVAL 512
2617 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2618 * tasks if there is an imbalance.
2620 static int load_balance(int this_cpu, struct rq *this_rq,
2621 struct sched_domain *sd, enum cpu_idle_type idle,
2624 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2625 struct sched_group *group;
2626 unsigned long imbalance;
2628 cpumask_t cpus = CPU_MASK_ALL;
2629 unsigned long flags;
2632 * When power savings policy is enabled for the parent domain, idle
2633 * sibling can pick up load irrespective of busy siblings. In this case,
2634 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2635 * portraying it as CPU_NOT_IDLE.
2637 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2638 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2641 schedstat_inc(sd, lb_count[idle]);
2644 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2651 schedstat_inc(sd, lb_nobusyg[idle]);
2655 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2657 schedstat_inc(sd, lb_nobusyq[idle]);
2661 BUG_ON(busiest == this_rq);
2663 schedstat_add(sd, lb_imbalance[idle], imbalance);
2666 if (busiest->nr_running > 1) {
2668 * Attempt to move tasks. If find_busiest_group has found
2669 * an imbalance but busiest->nr_running <= 1, the group is
2670 * still unbalanced. ld_moved simply stays zero, so it is
2671 * correctly treated as an imbalance.
2673 local_irq_save(flags);
2674 double_rq_lock(this_rq, busiest);
2675 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2676 imbalance, sd, idle, &all_pinned);
2677 double_rq_unlock(this_rq, busiest);
2678 local_irq_restore(flags);
2681 * some other cpu did the load balance for us.
2683 if (ld_moved && this_cpu != smp_processor_id())
2684 resched_cpu(this_cpu);
2686 /* All tasks on this runqueue were pinned by CPU affinity */
2687 if (unlikely(all_pinned)) {
2688 cpu_clear(cpu_of(busiest), cpus);
2689 if (!cpus_empty(cpus))
2696 schedstat_inc(sd, lb_failed[idle]);
2697 sd->nr_balance_failed++;
2699 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2701 spin_lock_irqsave(&busiest->lock, flags);
2703 /* don't kick the migration_thread, if the curr
2704 * task on busiest cpu can't be moved to this_cpu
2706 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2707 spin_unlock_irqrestore(&busiest->lock, flags);
2709 goto out_one_pinned;
2712 if (!busiest->active_balance) {
2713 busiest->active_balance = 1;
2714 busiest->push_cpu = this_cpu;
2717 spin_unlock_irqrestore(&busiest->lock, flags);
2719 wake_up_process(busiest->migration_thread);
2722 * We've kicked active balancing, reset the failure
2725 sd->nr_balance_failed = sd->cache_nice_tries+1;
2728 sd->nr_balance_failed = 0;
2730 if (likely(!active_balance)) {
2731 /* We were unbalanced, so reset the balancing interval */
2732 sd->balance_interval = sd->min_interval;
2735 * If we've begun active balancing, start to back off. This
2736 * case may not be covered by the all_pinned logic if there
2737 * is only 1 task on the busy runqueue (because we don't call
2740 if (sd->balance_interval < sd->max_interval)
2741 sd->balance_interval *= 2;
2744 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2745 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2750 schedstat_inc(sd, lb_balanced[idle]);
2752 sd->nr_balance_failed = 0;
2755 /* tune up the balancing interval */
2756 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2757 (sd->balance_interval < sd->max_interval))
2758 sd->balance_interval *= 2;
2760 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2761 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2767 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2768 * tasks if there is an imbalance.
2770 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2771 * this_rq is locked.
2774 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2776 struct sched_group *group;
2777 struct rq *busiest = NULL;
2778 unsigned long imbalance;
2782 cpumask_t cpus = CPU_MASK_ALL;
2785 * When power savings policy is enabled for the parent domain, idle
2786 * sibling can pick up load irrespective of busy siblings. In this case,
2787 * let the state of idle sibling percolate up as IDLE, instead of
2788 * portraying it as CPU_NOT_IDLE.
2790 if (sd->flags & SD_SHARE_CPUPOWER &&
2791 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2794 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2796 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2797 &sd_idle, &cpus, NULL);
2799 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2803 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2806 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2810 BUG_ON(busiest == this_rq);
2812 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2815 if (busiest->nr_running > 1) {
2816 /* Attempt to move tasks */
2817 double_lock_balance(this_rq, busiest);
2818 /* this_rq->clock is already updated */
2819 update_rq_clock(busiest);
2820 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2821 imbalance, sd, CPU_NEWLY_IDLE,
2823 spin_unlock(&busiest->lock);
2825 if (unlikely(all_pinned)) {
2826 cpu_clear(cpu_of(busiest), cpus);
2827 if (!cpus_empty(cpus))
2833 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2834 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2835 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2838 sd->nr_balance_failed = 0;
2843 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2844 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2845 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2847 sd->nr_balance_failed = 0;
2853 * idle_balance is called by schedule() if this_cpu is about to become
2854 * idle. Attempts to pull tasks from other CPUs.
2856 static void idle_balance(int this_cpu, struct rq *this_rq)
2858 struct sched_domain *sd;
2859 int pulled_task = -1;
2860 unsigned long next_balance = jiffies + HZ;
2862 for_each_domain(this_cpu, sd) {
2863 unsigned long interval;
2865 if (!(sd->flags & SD_LOAD_BALANCE))
2868 if (sd->flags & SD_BALANCE_NEWIDLE)
2869 /* If we've pulled tasks over stop searching: */
2870 pulled_task = load_balance_newidle(this_cpu,
2873 interval = msecs_to_jiffies(sd->balance_interval);
2874 if (time_after(next_balance, sd->last_balance + interval))
2875 next_balance = sd->last_balance + interval;
2879 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2881 * We are going idle. next_balance may be set based on
2882 * a busy processor. So reset next_balance.
2884 this_rq->next_balance = next_balance;
2889 * active_load_balance is run by migration threads. It pushes running tasks
2890 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2891 * running on each physical CPU where possible, and avoids physical /
2892 * logical imbalances.
2894 * Called with busiest_rq locked.
2896 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2898 int target_cpu = busiest_rq->push_cpu;
2899 struct sched_domain *sd;
2900 struct rq *target_rq;
2902 /* Is there any task to move? */
2903 if (busiest_rq->nr_running <= 1)
2906 target_rq = cpu_rq(target_cpu);
2909 * This condition is "impossible", if it occurs
2910 * we need to fix it. Originally reported by
2911 * Bjorn Helgaas on a 128-cpu setup.
2913 BUG_ON(busiest_rq == target_rq);
2915 /* move a task from busiest_rq to target_rq */
2916 double_lock_balance(busiest_rq, target_rq);
2917 update_rq_clock(busiest_rq);
2918 update_rq_clock(target_rq);
2920 /* Search for an sd spanning us and the target CPU. */
2921 for_each_domain(target_cpu, sd) {
2922 if ((sd->flags & SD_LOAD_BALANCE) &&
2923 cpu_isset(busiest_cpu, sd->span))
2928 schedstat_inc(sd, alb_count);
2930 if (move_one_task(target_rq, target_cpu, busiest_rq,
2932 schedstat_inc(sd, alb_pushed);
2934 schedstat_inc(sd, alb_failed);
2936 spin_unlock(&target_rq->lock);
2941 atomic_t load_balancer;
2943 } nohz ____cacheline_aligned = {
2944 .load_balancer = ATOMIC_INIT(-1),
2945 .cpu_mask = CPU_MASK_NONE,
2949 * This routine will try to nominate the ilb (idle load balancing)
2950 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2951 * load balancing on behalf of all those cpus. If all the cpus in the system
2952 * go into this tickless mode, then there will be no ilb owner (as there is
2953 * no need for one) and all the cpus will sleep till the next wakeup event
2956 * For the ilb owner, tick is not stopped. And this tick will be used
2957 * for idle load balancing. ilb owner will still be part of
2960 * While stopping the tick, this cpu will become the ilb owner if there
2961 * is no other owner. And will be the owner till that cpu becomes busy
2962 * or if all cpus in the system stop their ticks at which point
2963 * there is no need for ilb owner.
2965 * When the ilb owner becomes busy, it nominates another owner, during the
2966 * next busy scheduler_tick()
2968 int select_nohz_load_balancer(int stop_tick)
2970 int cpu = smp_processor_id();
2973 cpu_set(cpu, nohz.cpu_mask);
2974 cpu_rq(cpu)->in_nohz_recently = 1;
2977 * If we are going offline and still the leader, give up!
2979 if (cpu_is_offline(cpu) &&
2980 atomic_read(&nohz.load_balancer) == cpu) {
2981 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2986 /* time for ilb owner also to sleep */
2987 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2988 if (atomic_read(&nohz.load_balancer) == cpu)
2989 atomic_set(&nohz.load_balancer, -1);
2993 if (atomic_read(&nohz.load_balancer) == -1) {
2994 /* make me the ilb owner */
2995 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2997 } else if (atomic_read(&nohz.load_balancer) == cpu)
3000 if (!cpu_isset(cpu, nohz.cpu_mask))
3003 cpu_clear(cpu, nohz.cpu_mask);
3005 if (atomic_read(&nohz.load_balancer) == cpu)
3006 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3013 static DEFINE_SPINLOCK(balancing);
3016 * It checks each scheduling domain to see if it is due to be balanced,
3017 * and initiates a balancing operation if so.
3019 * Balancing parameters are set up in arch_init_sched_domains.
3021 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3024 struct rq *rq = cpu_rq(cpu);
3025 unsigned long interval;
3026 struct sched_domain *sd;
3027 /* Earliest time when we have to do rebalance again */
3028 unsigned long next_balance = jiffies + 60*HZ;
3029 int update_next_balance = 0;
3031 for_each_domain(cpu, sd) {
3032 if (!(sd->flags & SD_LOAD_BALANCE))
3035 interval = sd->balance_interval;
3036 if (idle != CPU_IDLE)
3037 interval *= sd->busy_factor;
3039 /* scale ms to jiffies */
3040 interval = msecs_to_jiffies(interval);
3041 if (unlikely(!interval))
3043 if (interval > HZ*NR_CPUS/10)
3044 interval = HZ*NR_CPUS/10;
3047 if (sd->flags & SD_SERIALIZE) {
3048 if (!spin_trylock(&balancing))
3052 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3053 if (load_balance(cpu, rq, sd, idle, &balance)) {
3055 * We've pulled tasks over so either we're no
3056 * longer idle, or one of our SMT siblings is
3059 idle = CPU_NOT_IDLE;
3061 sd->last_balance = jiffies;
3063 if (sd->flags & SD_SERIALIZE)
3064 spin_unlock(&balancing);
3066 if (time_after(next_balance, sd->last_balance + interval)) {
3067 next_balance = sd->last_balance + interval;
3068 update_next_balance = 1;
3072 * Stop the load balance at this level. There is another
3073 * CPU in our sched group which is doing load balancing more
3081 * next_balance will be updated only when there is a need.
3082 * When the cpu is attached to null domain for ex, it will not be
3085 if (likely(update_next_balance))
3086 rq->next_balance = next_balance;
3090 * run_rebalance_domains is triggered when needed from the scheduler tick.
3091 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3092 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3094 static void run_rebalance_domains(struct softirq_action *h)
3096 int this_cpu = smp_processor_id();
3097 struct rq *this_rq = cpu_rq(this_cpu);
3098 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3099 CPU_IDLE : CPU_NOT_IDLE;
3101 rebalance_domains(this_cpu, idle);
3105 * If this cpu is the owner for idle load balancing, then do the
3106 * balancing on behalf of the other idle cpus whose ticks are
3109 if (this_rq->idle_at_tick &&
3110 atomic_read(&nohz.load_balancer) == this_cpu) {
3111 cpumask_t cpus = nohz.cpu_mask;
3115 cpu_clear(this_cpu, cpus);
3116 for_each_cpu_mask(balance_cpu, cpus) {
3118 * If this cpu gets work to do, stop the load balancing
3119 * work being done for other cpus. Next load
3120 * balancing owner will pick it up.
3125 rebalance_domains(balance_cpu, CPU_IDLE);
3127 rq = cpu_rq(balance_cpu);
3128 if (time_after(this_rq->next_balance, rq->next_balance))
3129 this_rq->next_balance = rq->next_balance;
3136 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3138 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3139 * idle load balancing owner or decide to stop the periodic load balancing,
3140 * if the whole system is idle.
3142 static inline void trigger_load_balance(struct rq *rq, int cpu)
3146 * If we were in the nohz mode recently and busy at the current
3147 * scheduler tick, then check if we need to nominate new idle
3150 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3151 rq->in_nohz_recently = 0;
3153 if (atomic_read(&nohz.load_balancer) == cpu) {
3154 cpu_clear(cpu, nohz.cpu_mask);
3155 atomic_set(&nohz.load_balancer, -1);
3158 if (atomic_read(&nohz.load_balancer) == -1) {
3160 * simple selection for now: Nominate the
3161 * first cpu in the nohz list to be the next
3164 * TBD: Traverse the sched domains and nominate
3165 * the nearest cpu in the nohz.cpu_mask.
3167 int ilb = first_cpu(nohz.cpu_mask);
3175 * If this cpu is idle and doing idle load balancing for all the
3176 * cpus with ticks stopped, is it time for that to stop?
3178 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3179 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3185 * If this cpu is idle and the idle load balancing is done by
3186 * someone else, then no need raise the SCHED_SOFTIRQ
3188 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3189 cpu_isset(cpu, nohz.cpu_mask))
3192 if (time_after_eq(jiffies, rq->next_balance))
3193 raise_softirq(SCHED_SOFTIRQ);
3196 #else /* CONFIG_SMP */
3199 * on UP we do not need to balance between CPUs:
3201 static inline void idle_balance(int cpu, struct rq *rq)
3205 /* Avoid "used but not defined" warning on UP */
3206 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3207 unsigned long max_nr_move, unsigned long max_load_move,
3208 struct sched_domain *sd, enum cpu_idle_type idle,
3209 int *all_pinned, unsigned long *load_moved,
3210 int *this_best_prio, struct rq_iterator *iterator)
3219 DEFINE_PER_CPU(struct kernel_stat, kstat);
3221 EXPORT_PER_CPU_SYMBOL(kstat);
3224 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3225 * that have not yet been banked in case the task is currently running.
3227 unsigned long long task_sched_runtime(struct task_struct *p)
3229 unsigned long flags;
3233 rq = task_rq_lock(p, &flags);
3234 ns = p->se.sum_exec_runtime;
3235 if (rq->curr == p) {
3236 update_rq_clock(rq);
3237 delta_exec = rq->clock - p->se.exec_start;
3238 if ((s64)delta_exec > 0)
3241 task_rq_unlock(rq, &flags);
3247 * Account user cpu time to a process.
3248 * @p: the process that the cpu time gets accounted to
3249 * @hardirq_offset: the offset to subtract from hardirq_count()
3250 * @cputime: the cpu time spent in user space since the last update
3252 void account_user_time(struct task_struct *p, cputime_t cputime)
3254 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3257 p->utime = cputime_add(p->utime, cputime);
3259 /* Add user time to cpustat. */
3260 tmp = cputime_to_cputime64(cputime);
3261 if (TASK_NICE(p) > 0)
3262 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3264 cpustat->user = cputime64_add(cpustat->user, tmp);
3268 * Account system cpu time to a process.
3269 * @p: the process that the cpu time gets accounted to
3270 * @hardirq_offset: the offset to subtract from hardirq_count()
3271 * @cputime: the cpu time spent in kernel space since the last update
3273 void account_system_time(struct task_struct *p, int hardirq_offset,
3276 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3277 struct rq *rq = this_rq();
3280 p->stime = cputime_add(p->stime, cputime);
3282 /* Add system time to cpustat. */
3283 tmp = cputime_to_cputime64(cputime);
3284 if (hardirq_count() - hardirq_offset)
3285 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3286 else if (softirq_count())
3287 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3288 else if (p != rq->idle)
3289 cpustat->system = cputime64_add(cpustat->system, tmp);
3290 else if (atomic_read(&rq->nr_iowait) > 0)
3291 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3293 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3294 /* Account for system time used */
3295 acct_update_integrals(p);
3299 * Account for involuntary wait time.
3300 * @p: the process from which the cpu time has been stolen
3301 * @steal: the cpu time spent in involuntary wait
3303 void account_steal_time(struct task_struct *p, cputime_t steal)
3305 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3306 cputime64_t tmp = cputime_to_cputime64(steal);
3307 struct rq *rq = this_rq();
3309 if (p == rq->idle) {
3310 p->stime = cputime_add(p->stime, steal);
3311 if (atomic_read(&rq->nr_iowait) > 0)
3312 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3314 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3316 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3320 * This function gets called by the timer code, with HZ frequency.
3321 * We call it with interrupts disabled.
3323 * It also gets called by the fork code, when changing the parent's
3326 void scheduler_tick(void)
3328 int cpu = smp_processor_id();
3329 struct rq *rq = cpu_rq(cpu);
3330 struct task_struct *curr = rq->curr;
3331 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3333 spin_lock(&rq->lock);
3334 __update_rq_clock(rq);
3336 * Let rq->clock advance by at least TICK_NSEC:
3338 if (unlikely(rq->clock < next_tick))
3339 rq->clock = next_tick;
3340 rq->tick_timestamp = rq->clock;
3341 update_cpu_load(rq);
3342 if (curr != rq->idle) /* FIXME: needed? */
3343 curr->sched_class->task_tick(rq, curr);
3344 spin_unlock(&rq->lock);
3347 rq->idle_at_tick = idle_cpu(cpu);
3348 trigger_load_balance(rq, cpu);
3352 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3354 void fastcall add_preempt_count(int val)
3359 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3361 preempt_count() += val;
3363 * Spinlock count overflowing soon?
3365 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3368 EXPORT_SYMBOL(add_preempt_count);
3370 void fastcall sub_preempt_count(int val)
3375 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3378 * Is the spinlock portion underflowing?
3380 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3381 !(preempt_count() & PREEMPT_MASK)))
3384 preempt_count() -= val;
3386 EXPORT_SYMBOL(sub_preempt_count);
3391 * Print scheduling while atomic bug:
3393 static noinline void __schedule_bug(struct task_struct *prev)
3395 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3396 prev->comm, preempt_count(), prev->pid);
3397 debug_show_held_locks(prev);
3398 if (irqs_disabled())
3399 print_irqtrace_events(prev);
3404 * Various schedule()-time debugging checks and statistics:
3406 static inline void schedule_debug(struct task_struct *prev)
3409 * Test if we are atomic. Since do_exit() needs to call into
3410 * schedule() atomically, we ignore that path for now.
3411 * Otherwise, whine if we are scheduling when we should not be.
3413 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3414 __schedule_bug(prev);
3416 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3418 schedstat_inc(this_rq(), sched_count);
3419 #ifdef CONFIG_SCHEDSTATS
3420 if (unlikely(prev->lock_depth >= 0)) {
3421 schedstat_inc(this_rq(), bkl_count);
3422 schedstat_inc(prev, sched_info.bkl_count);
3428 * Pick up the highest-prio task:
3430 static inline struct task_struct *
3431 pick_next_task(struct rq *rq, struct task_struct *prev)
3433 const struct sched_class *class;
3434 struct task_struct *p;
3437 * Optimization: we know that if all tasks are in
3438 * the fair class we can call that function directly:
3440 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3441 p = fair_sched_class.pick_next_task(rq);
3446 class = sched_class_highest;
3448 p = class->pick_next_task(rq);
3452 * Will never be NULL as the idle class always
3453 * returns a non-NULL p:
3455 class = class->next;
3460 * schedule() is the main scheduler function.
3462 asmlinkage void __sched schedule(void)
3464 struct task_struct *prev, *next;
3471 cpu = smp_processor_id();
3475 switch_count = &prev->nivcsw;
3477 release_kernel_lock(prev);
3478 need_resched_nonpreemptible:
3480 schedule_debug(prev);
3483 * Do the rq-clock update outside the rq lock:
3485 local_irq_disable();
3486 __update_rq_clock(rq);
3487 spin_lock(&rq->lock);
3488 clear_tsk_need_resched(prev);
3490 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3491 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3492 unlikely(signal_pending(prev)))) {
3493 prev->state = TASK_RUNNING;
3495 deactivate_task(rq, prev, 1);
3497 switch_count = &prev->nvcsw;
3500 if (unlikely(!rq->nr_running))
3501 idle_balance(cpu, rq);
3503 prev->sched_class->put_prev_task(rq, prev);
3504 next = pick_next_task(rq, prev);
3506 sched_info_switch(prev, next);
3508 if (likely(prev != next)) {
3513 context_switch(rq, prev, next); /* unlocks the rq */
3515 spin_unlock_irq(&rq->lock);
3517 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3518 cpu = smp_processor_id();
3520 goto need_resched_nonpreemptible;
3522 preempt_enable_no_resched();
3523 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3526 EXPORT_SYMBOL(schedule);
3528 #ifdef CONFIG_PREEMPT
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3534 asmlinkage void __sched preempt_schedule(void)
3536 struct thread_info *ti = current_thread_info();
3537 #ifdef CONFIG_PREEMPT_BKL
3538 struct task_struct *task = current;
3539 int saved_lock_depth;
3542 * If there is a non-zero preempt_count or interrupts are disabled,
3543 * we do not want to preempt the current task. Just return..
3545 if (likely(ti->preempt_count || irqs_disabled()))
3549 add_preempt_count(PREEMPT_ACTIVE);
3552 * We keep the big kernel semaphore locked, but we
3553 * clear ->lock_depth so that schedule() doesnt
3554 * auto-release the semaphore:
3556 #ifdef CONFIG_PREEMPT_BKL
3557 saved_lock_depth = task->lock_depth;
3558 task->lock_depth = -1;
3561 #ifdef CONFIG_PREEMPT_BKL
3562 task->lock_depth = saved_lock_depth;
3564 sub_preempt_count(PREEMPT_ACTIVE);
3567 * Check again in case we missed a preemption opportunity
3568 * between schedule and now.
3571 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3573 EXPORT_SYMBOL(preempt_schedule);
3576 * this is the entry point to schedule() from kernel preemption
3577 * off of irq context.
3578 * Note, that this is called and return with irqs disabled. This will
3579 * protect us against recursive calling from irq.
3581 asmlinkage void __sched preempt_schedule_irq(void)
3583 struct thread_info *ti = current_thread_info();
3584 #ifdef CONFIG_PREEMPT_BKL
3585 struct task_struct *task = current;
3586 int saved_lock_depth;
3588 /* Catch callers which need to be fixed */
3589 BUG_ON(ti->preempt_count || !irqs_disabled());
3592 add_preempt_count(PREEMPT_ACTIVE);
3595 * We keep the big kernel semaphore locked, but we
3596 * clear ->lock_depth so that schedule() doesnt
3597 * auto-release the semaphore:
3599 #ifdef CONFIG_PREEMPT_BKL
3600 saved_lock_depth = task->lock_depth;
3601 task->lock_depth = -1;
3605 local_irq_disable();
3606 #ifdef CONFIG_PREEMPT_BKL
3607 task->lock_depth = saved_lock_depth;
3609 sub_preempt_count(PREEMPT_ACTIVE);
3612 * Check again in case we missed a preemption opportunity
3613 * between schedule and now.
3616 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3619 #endif /* CONFIG_PREEMPT */
3621 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3624 return try_to_wake_up(curr->private, mode, sync);
3626 EXPORT_SYMBOL(default_wake_function);
3629 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3630 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3631 * number) then we wake all the non-exclusive tasks and one exclusive task.
3633 * There are circumstances in which we can try to wake a task which has already
3634 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3635 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3637 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3638 int nr_exclusive, int sync, void *key)
3640 wait_queue_t *curr, *next;
3642 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3643 unsigned flags = curr->flags;
3645 if (curr->func(curr, mode, sync, key) &&
3646 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3652 * __wake_up - wake up threads blocked on a waitqueue.
3654 * @mode: which threads
3655 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3656 * @key: is directly passed to the wakeup function
3658 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3659 int nr_exclusive, void *key)
3661 unsigned long flags;
3663 spin_lock_irqsave(&q->lock, flags);
3664 __wake_up_common(q, mode, nr_exclusive, 0, key);
3665 spin_unlock_irqrestore(&q->lock, flags);
3667 EXPORT_SYMBOL(__wake_up);
3670 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3672 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3674 __wake_up_common(q, mode, 1, 0, NULL);
3678 * __wake_up_sync - wake up threads blocked on a waitqueue.
3680 * @mode: which threads
3681 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3683 * The sync wakeup differs that the waker knows that it will schedule
3684 * away soon, so while the target thread will be woken up, it will not
3685 * be migrated to another CPU - ie. the two threads are 'synchronized'
3686 * with each other. This can prevent needless bouncing between CPUs.
3688 * On UP it can prevent extra preemption.
3691 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3693 unsigned long flags;
3699 if (unlikely(!nr_exclusive))
3702 spin_lock_irqsave(&q->lock, flags);
3703 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3704 spin_unlock_irqrestore(&q->lock, flags);
3706 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3708 void fastcall complete(struct completion *x)
3710 unsigned long flags;
3712 spin_lock_irqsave(&x->wait.lock, flags);
3714 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3716 spin_unlock_irqrestore(&x->wait.lock, flags);
3718 EXPORT_SYMBOL(complete);
3720 void fastcall complete_all(struct completion *x)
3722 unsigned long flags;
3724 spin_lock_irqsave(&x->wait.lock, flags);
3725 x->done += UINT_MAX/2;
3726 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3728 spin_unlock_irqrestore(&x->wait.lock, flags);
3730 EXPORT_SYMBOL(complete_all);
3732 static inline long __sched
3733 do_wait_for_common(struct completion *x, long timeout, int state)
3736 DECLARE_WAITQUEUE(wait, current);
3738 wait.flags |= WQ_FLAG_EXCLUSIVE;
3739 __add_wait_queue_tail(&x->wait, &wait);
3741 if (state == TASK_INTERRUPTIBLE &&
3742 signal_pending(current)) {
3743 __remove_wait_queue(&x->wait, &wait);
3744 return -ERESTARTSYS;
3746 __set_current_state(state);
3747 spin_unlock_irq(&x->wait.lock);
3748 timeout = schedule_timeout(timeout);
3749 spin_lock_irq(&x->wait.lock);
3751 __remove_wait_queue(&x->wait, &wait);
3755 __remove_wait_queue(&x->wait, &wait);
3762 wait_for_common(struct completion *x, long timeout, int state)
3766 spin_lock_irq(&x->wait.lock);
3767 timeout = do_wait_for_common(x, timeout, state);
3768 spin_unlock_irq(&x->wait.lock);
3772 void fastcall __sched wait_for_completion(struct completion *x)
3774 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3776 EXPORT_SYMBOL(wait_for_completion);
3778 unsigned long fastcall __sched
3779 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3781 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3783 EXPORT_SYMBOL(wait_for_completion_timeout);
3785 int __sched wait_for_completion_interruptible(struct completion *x)
3787 return wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3789 EXPORT_SYMBOL(wait_for_completion_interruptible);
3791 unsigned long fastcall __sched
3792 wait_for_completion_interruptible_timeout(struct completion *x,
3793 unsigned long timeout)
3795 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3797 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3800 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3802 unsigned long flags;
3805 init_waitqueue_entry(&wait, current);
3807 __set_current_state(state);
3809 spin_lock_irqsave(&q->lock, flags);
3810 __add_wait_queue(q, &wait);
3811 spin_unlock(&q->lock);
3812 timeout = schedule_timeout(timeout);
3813 spin_lock_irq(&q->lock);
3814 __remove_wait_queue(q, &wait);
3815 spin_unlock_irqrestore(&q->lock, flags);
3820 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3822 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3824 EXPORT_SYMBOL(interruptible_sleep_on);
3827 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3829 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3831 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3833 void __sched sleep_on(wait_queue_head_t *q)
3835 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3837 EXPORT_SYMBOL(sleep_on);
3839 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3841 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3843 EXPORT_SYMBOL(sleep_on_timeout);
3845 #ifdef CONFIG_RT_MUTEXES
3848 * rt_mutex_setprio - set the current priority of a task
3850 * @prio: prio value (kernel-internal form)
3852 * This function changes the 'effective' priority of a task. It does
3853 * not touch ->normal_prio like __setscheduler().
3855 * Used by the rt_mutex code to implement priority inheritance logic.
3857 void rt_mutex_setprio(struct task_struct *p, int prio)
3859 unsigned long flags;
3860 int oldprio, on_rq, running;
3863 BUG_ON(prio < 0 || prio > MAX_PRIO);
3865 rq = task_rq_lock(p, &flags);
3866 update_rq_clock(rq);
3869 on_rq = p->se.on_rq;
3870 running = task_running(rq, p);
3872 dequeue_task(rq, p, 0);
3874 p->sched_class->put_prev_task(rq, p);
3878 p->sched_class = &rt_sched_class;
3880 p->sched_class = &fair_sched_class;
3886 p->sched_class->set_curr_task(rq);
3887 enqueue_task(rq, p, 0);
3889 * Reschedule if we are currently running on this runqueue and
3890 * our priority decreased, or if we are not currently running on
3891 * this runqueue and our priority is higher than the current's
3894 if (p->prio > oldprio)
3895 resched_task(rq->curr);
3897 check_preempt_curr(rq, p);
3900 task_rq_unlock(rq, &flags);
3905 void set_user_nice(struct task_struct *p, long nice)
3907 int old_prio, delta, on_rq;
3908 unsigned long flags;
3911 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3914 * We have to be careful, if called from sys_setpriority(),
3915 * the task might be in the middle of scheduling on another CPU.
3917 rq = task_rq_lock(p, &flags);
3918 update_rq_clock(rq);
3920 * The RT priorities are set via sched_setscheduler(), but we still
3921 * allow the 'normal' nice value to be set - but as expected
3922 * it wont have any effect on scheduling until the task is
3923 * SCHED_FIFO/SCHED_RR:
3925 if (task_has_rt_policy(p)) {
3926 p->static_prio = NICE_TO_PRIO(nice);
3929 on_rq = p->se.on_rq;
3931 dequeue_task(rq, p, 0);
3935 p->static_prio = NICE_TO_PRIO(nice);
3938 p->prio = effective_prio(p);
3939 delta = p->prio - old_prio;
3942 enqueue_task(rq, p, 0);
3945 * If the task increased its priority or is running and
3946 * lowered its priority, then reschedule its CPU:
3948 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3949 resched_task(rq->curr);
3952 task_rq_unlock(rq, &flags);
3954 EXPORT_SYMBOL(set_user_nice);
3957 * can_nice - check if a task can reduce its nice value
3961 int can_nice(const struct task_struct *p, const int nice)
3963 /* convert nice value [19,-20] to rlimit style value [1,40] */
3964 int nice_rlim = 20 - nice;
3966 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3967 capable(CAP_SYS_NICE));
3970 #ifdef __ARCH_WANT_SYS_NICE
3973 * sys_nice - change the priority of the current process.
3974 * @increment: priority increment
3976 * sys_setpriority is a more generic, but much slower function that
3977 * does similar things.
3979 asmlinkage long sys_nice(int increment)
3984 * Setpriority might change our priority at the same moment.
3985 * We don't have to worry. Conceptually one call occurs first
3986 * and we have a single winner.
3988 if (increment < -40)
3993 nice = PRIO_TO_NICE(current->static_prio) + increment;
3999 if (increment < 0 && !can_nice(current, nice))
4002 retval = security_task_setnice(current, nice);
4006 set_user_nice(current, nice);
4013 * task_prio - return the priority value of a given task.
4014 * @p: the task in question.
4016 * This is the priority value as seen by users in /proc.
4017 * RT tasks are offset by -200. Normal tasks are centered
4018 * around 0, value goes from -16 to +15.
4020 int task_prio(const struct task_struct *p)
4022 return p->prio - MAX_RT_PRIO;
4026 * task_nice - return the nice value of a given task.
4027 * @p: the task in question.
4029 int task_nice(const struct task_struct *p)
4031 return TASK_NICE(p);
4033 EXPORT_SYMBOL_GPL(task_nice);
4036 * idle_cpu - is a given cpu idle currently?
4037 * @cpu: the processor in question.
4039 int idle_cpu(int cpu)
4041 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4045 * idle_task - return the idle task for a given cpu.
4046 * @cpu: the processor in question.
4048 struct task_struct *idle_task(int cpu)
4050 return cpu_rq(cpu)->idle;
4054 * find_process_by_pid - find a process with a matching PID value.
4055 * @pid: the pid in question.
4057 static struct task_struct *find_process_by_pid(pid_t pid)
4059 return pid ? find_task_by_pid(pid) : current;
4062 /* Actually do priority change: must hold rq lock. */
4064 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4066 BUG_ON(p->se.on_rq);
4069 switch (p->policy) {
4073 p->sched_class = &fair_sched_class;
4077 p->sched_class = &rt_sched_class;
4081 p->rt_priority = prio;
4082 p->normal_prio = normal_prio(p);
4083 /* we are holding p->pi_lock already */
4084 p->prio = rt_mutex_getprio(p);
4089 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4090 * @p: the task in question.
4091 * @policy: new policy.
4092 * @param: structure containing the new RT priority.
4094 * NOTE that the task may be already dead.
4096 int sched_setscheduler(struct task_struct *p, int policy,
4097 struct sched_param *param)
4099 int retval, oldprio, oldpolicy = -1, on_rq, running;
4100 unsigned long flags;
4103 /* may grab non-irq protected spin_locks */
4104 BUG_ON(in_interrupt());
4106 /* double check policy once rq lock held */
4108 policy = oldpolicy = p->policy;
4109 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4110 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4111 policy != SCHED_IDLE)
4114 * Valid priorities for SCHED_FIFO and SCHED_RR are
4115 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4116 * SCHED_BATCH and SCHED_IDLE is 0.
4118 if (param->sched_priority < 0 ||
4119 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4120 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4122 if (rt_policy(policy) != (param->sched_priority != 0))
4126 * Allow unprivileged RT tasks to decrease priority:
4128 if (!capable(CAP_SYS_NICE)) {
4129 if (rt_policy(policy)) {
4130 unsigned long rlim_rtprio;
4132 if (!lock_task_sighand(p, &flags))
4134 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4135 unlock_task_sighand(p, &flags);
4137 /* can't set/change the rt policy */
4138 if (policy != p->policy && !rlim_rtprio)
4141 /* can't increase priority */
4142 if (param->sched_priority > p->rt_priority &&
4143 param->sched_priority > rlim_rtprio)
4147 * Like positive nice levels, dont allow tasks to
4148 * move out of SCHED_IDLE either:
4150 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4153 /* can't change other user's priorities */
4154 if ((current->euid != p->euid) &&
4155 (current->euid != p->uid))
4159 retval = security_task_setscheduler(p, policy, param);
4163 * make sure no PI-waiters arrive (or leave) while we are
4164 * changing the priority of the task:
4166 spin_lock_irqsave(&p->pi_lock, flags);
4168 * To be able to change p->policy safely, the apropriate
4169 * runqueue lock must be held.
4171 rq = __task_rq_lock(p);
4172 /* recheck policy now with rq lock held */
4173 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4174 policy = oldpolicy = -1;
4175 __task_rq_unlock(rq);
4176 spin_unlock_irqrestore(&p->pi_lock, flags);
4179 update_rq_clock(rq);
4180 on_rq = p->se.on_rq;
4181 running = task_running(rq, p);
4183 deactivate_task(rq, p, 0);
4185 p->sched_class->put_prev_task(rq, p);
4189 __setscheduler(rq, p, policy, param->sched_priority);
4193 p->sched_class->set_curr_task(rq);
4194 activate_task(rq, p, 0);
4196 * Reschedule if we are currently running on this runqueue and
4197 * our priority decreased, or if we are not currently running on
4198 * this runqueue and our priority is higher than the current's
4201 if (p->prio > oldprio)
4202 resched_task(rq->curr);
4204 check_preempt_curr(rq, p);
4207 __task_rq_unlock(rq);
4208 spin_unlock_irqrestore(&p->pi_lock, flags);
4210 rt_mutex_adjust_pi(p);
4214 EXPORT_SYMBOL_GPL(sched_setscheduler);
4217 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4219 struct sched_param lparam;
4220 struct task_struct *p;
4223 if (!param || pid < 0)
4225 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4230 p = find_process_by_pid(pid);
4232 retval = sched_setscheduler(p, policy, &lparam);
4239 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4240 * @pid: the pid in question.
4241 * @policy: new policy.
4242 * @param: structure containing the new RT priority.
4244 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4245 struct sched_param __user *param)
4247 /* negative values for policy are not valid */
4251 return do_sched_setscheduler(pid, policy, param);
4255 * sys_sched_setparam - set/change the RT priority of a thread
4256 * @pid: the pid in question.
4257 * @param: structure containing the new RT priority.
4259 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4261 return do_sched_setscheduler(pid, -1, param);
4265 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4266 * @pid: the pid in question.
4268 asmlinkage long sys_sched_getscheduler(pid_t pid)
4270 struct task_struct *p;
4277 read_lock(&tasklist_lock);
4278 p = find_process_by_pid(pid);
4280 retval = security_task_getscheduler(p);
4284 read_unlock(&tasklist_lock);
4289 * sys_sched_getscheduler - get the RT priority of a thread
4290 * @pid: the pid in question.
4291 * @param: structure containing the RT priority.
4293 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4295 struct sched_param lp;
4296 struct task_struct *p;
4299 if (!param || pid < 0)
4302 read_lock(&tasklist_lock);
4303 p = find_process_by_pid(pid);
4308 retval = security_task_getscheduler(p);
4312 lp.sched_priority = p->rt_priority;
4313 read_unlock(&tasklist_lock);
4316 * This one might sleep, we cannot do it with a spinlock held ...
4318 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4323 read_unlock(&tasklist_lock);
4327 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4329 cpumask_t cpus_allowed;
4330 struct task_struct *p;
4333 mutex_lock(&sched_hotcpu_mutex);
4334 read_lock(&tasklist_lock);
4336 p = find_process_by_pid(pid);
4338 read_unlock(&tasklist_lock);
4339 mutex_unlock(&sched_hotcpu_mutex);
4344 * It is not safe to call set_cpus_allowed with the
4345 * tasklist_lock held. We will bump the task_struct's
4346 * usage count and then drop tasklist_lock.
4349 read_unlock(&tasklist_lock);
4352 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4353 !capable(CAP_SYS_NICE))
4356 retval = security_task_setscheduler(p, 0, NULL);
4360 cpus_allowed = cpuset_cpus_allowed(p);
4361 cpus_and(new_mask, new_mask, cpus_allowed);
4362 retval = set_cpus_allowed(p, new_mask);
4366 mutex_unlock(&sched_hotcpu_mutex);
4370 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4371 cpumask_t *new_mask)
4373 if (len < sizeof(cpumask_t)) {
4374 memset(new_mask, 0, sizeof(cpumask_t));
4375 } else if (len > sizeof(cpumask_t)) {
4376 len = sizeof(cpumask_t);
4378 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4382 * sys_sched_setaffinity - set the cpu affinity of a process
4383 * @pid: pid of the process
4384 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4385 * @user_mask_ptr: user-space pointer to the new cpu mask
4387 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4388 unsigned long __user *user_mask_ptr)
4393 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4397 return sched_setaffinity(pid, new_mask);
4401 * Represents all cpu's present in the system
4402 * In systems capable of hotplug, this map could dynamically grow
4403 * as new cpu's are detected in the system via any platform specific
4404 * method, such as ACPI for e.g.
4407 cpumask_t cpu_present_map __read_mostly;
4408 EXPORT_SYMBOL(cpu_present_map);
4411 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4412 EXPORT_SYMBOL(cpu_online_map);
4414 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4415 EXPORT_SYMBOL(cpu_possible_map);
4418 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4420 struct task_struct *p;
4423 mutex_lock(&sched_hotcpu_mutex);
4424 read_lock(&tasklist_lock);
4427 p = find_process_by_pid(pid);
4431 retval = security_task_getscheduler(p);
4435 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4438 read_unlock(&tasklist_lock);
4439 mutex_unlock(&sched_hotcpu_mutex);
4445 * sys_sched_getaffinity - get the cpu affinity of a process
4446 * @pid: pid of the process
4447 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4448 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4450 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4451 unsigned long __user *user_mask_ptr)
4456 if (len < sizeof(cpumask_t))
4459 ret = sched_getaffinity(pid, &mask);
4463 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4466 return sizeof(cpumask_t);
4470 * sys_sched_yield - yield the current processor to other threads.
4472 * This function yields the current CPU to other tasks. If there are no
4473 * other threads running on this CPU then this function will return.
4475 asmlinkage long sys_sched_yield(void)
4477 struct rq *rq = this_rq_lock();
4479 schedstat_inc(rq, yld_count);
4480 current->sched_class->yield_task(rq);
4483 * Since we are going to call schedule() anyway, there's
4484 * no need to preempt or enable interrupts:
4486 __release(rq->lock);
4487 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4488 _raw_spin_unlock(&rq->lock);
4489 preempt_enable_no_resched();
4496 static void __cond_resched(void)
4498 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4499 __might_sleep(__FILE__, __LINE__);
4502 * The BKS might be reacquired before we have dropped
4503 * PREEMPT_ACTIVE, which could trigger a second
4504 * cond_resched() call.
4507 add_preempt_count(PREEMPT_ACTIVE);
4509 sub_preempt_count(PREEMPT_ACTIVE);
4510 } while (need_resched());
4513 int __sched cond_resched(void)
4515 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4516 system_state == SYSTEM_RUNNING) {
4522 EXPORT_SYMBOL(cond_resched);
4525 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4526 * call schedule, and on return reacquire the lock.
4528 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4529 * operations here to prevent schedule() from being called twice (once via
4530 * spin_unlock(), once by hand).
4532 int cond_resched_lock(spinlock_t *lock)
4536 if (need_lockbreak(lock)) {
4542 if (need_resched() && system_state == SYSTEM_RUNNING) {
4543 spin_release(&lock->dep_map, 1, _THIS_IP_);
4544 _raw_spin_unlock(lock);
4545 preempt_enable_no_resched();
4552 EXPORT_SYMBOL(cond_resched_lock);
4554 int __sched cond_resched_softirq(void)
4556 BUG_ON(!in_softirq());
4558 if (need_resched() && system_state == SYSTEM_RUNNING) {
4566 EXPORT_SYMBOL(cond_resched_softirq);
4569 * yield - yield the current processor to other threads.
4571 * This is a shortcut for kernel-space yielding - it marks the
4572 * thread runnable and calls sys_sched_yield().
4574 void __sched yield(void)
4576 set_current_state(TASK_RUNNING);
4579 EXPORT_SYMBOL(yield);
4582 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4583 * that process accounting knows that this is a task in IO wait state.
4585 * But don't do that if it is a deliberate, throttling IO wait (this task
4586 * has set its backing_dev_info: the queue against which it should throttle)
4588 void __sched io_schedule(void)
4590 struct rq *rq = &__raw_get_cpu_var(runqueues);
4592 delayacct_blkio_start();
4593 atomic_inc(&rq->nr_iowait);
4595 atomic_dec(&rq->nr_iowait);
4596 delayacct_blkio_end();
4598 EXPORT_SYMBOL(io_schedule);
4600 long __sched io_schedule_timeout(long timeout)
4602 struct rq *rq = &__raw_get_cpu_var(runqueues);
4605 delayacct_blkio_start();
4606 atomic_inc(&rq->nr_iowait);
4607 ret = schedule_timeout(timeout);
4608 atomic_dec(&rq->nr_iowait);
4609 delayacct_blkio_end();
4614 * sys_sched_get_priority_max - return maximum RT priority.
4615 * @policy: scheduling class.
4617 * this syscall returns the maximum rt_priority that can be used
4618 * by a given scheduling class.
4620 asmlinkage long sys_sched_get_priority_max(int policy)
4627 ret = MAX_USER_RT_PRIO-1;
4639 * sys_sched_get_priority_min - return minimum RT priority.
4640 * @policy: scheduling class.
4642 * this syscall returns the minimum rt_priority that can be used
4643 * by a given scheduling class.
4645 asmlinkage long sys_sched_get_priority_min(int policy)
4663 * sys_sched_rr_get_interval - return the default timeslice of a process.
4664 * @pid: pid of the process.
4665 * @interval: userspace pointer to the timeslice value.
4667 * this syscall writes the default timeslice value of a given process
4668 * into the user-space timespec buffer. A value of '0' means infinity.
4671 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4673 struct task_struct *p;
4674 unsigned int time_slice;
4682 read_lock(&tasklist_lock);
4683 p = find_process_by_pid(pid);
4687 retval = security_task_getscheduler(p);
4691 if (p->policy == SCHED_FIFO)
4693 else if (p->policy == SCHED_RR)
4694 time_slice = DEF_TIMESLICE;
4696 struct sched_entity *se = &p->se;
4697 unsigned long flags;
4700 rq = task_rq_lock(p, &flags);
4701 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4702 task_rq_unlock(rq, &flags);
4704 read_unlock(&tasklist_lock);
4705 jiffies_to_timespec(time_slice, &t);
4706 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4710 read_unlock(&tasklist_lock);
4714 static const char stat_nam[] = "RSDTtZX";
4716 static void show_task(struct task_struct *p)
4718 unsigned long free = 0;
4721 state = p->state ? __ffs(p->state) + 1 : 0;
4722 printk("%-13.13s %c", p->comm,
4723 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4724 #if BITS_PER_LONG == 32
4725 if (state == TASK_RUNNING)
4726 printk(" running ");
4728 printk(" %08lx ", thread_saved_pc(p));
4730 if (state == TASK_RUNNING)
4731 printk(" running task ");
4733 printk(" %016lx ", thread_saved_pc(p));
4735 #ifdef CONFIG_DEBUG_STACK_USAGE
4737 unsigned long *n = end_of_stack(p);
4740 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4743 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4745 if (state != TASK_RUNNING)
4746 show_stack(p, NULL);
4749 void show_state_filter(unsigned long state_filter)
4751 struct task_struct *g, *p;
4753 #if BITS_PER_LONG == 32
4755 " task PC stack pid father\n");
4758 " task PC stack pid father\n");
4760 read_lock(&tasklist_lock);
4761 do_each_thread(g, p) {
4763 * reset the NMI-timeout, listing all files on a slow
4764 * console might take alot of time:
4766 touch_nmi_watchdog();
4767 if (!state_filter || (p->state & state_filter))
4769 } while_each_thread(g, p);
4771 touch_all_softlockup_watchdogs();
4773 #ifdef CONFIG_SCHED_DEBUG
4774 sysrq_sched_debug_show();
4776 read_unlock(&tasklist_lock);
4778 * Only show locks if all tasks are dumped:
4780 if (state_filter == -1)
4781 debug_show_all_locks();
4784 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4786 idle->sched_class = &idle_sched_class;
4790 * init_idle - set up an idle thread for a given CPU
4791 * @idle: task in question
4792 * @cpu: cpu the idle task belongs to
4794 * NOTE: this function does not set the idle thread's NEED_RESCHED
4795 * flag, to make booting more robust.
4797 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4799 struct rq *rq = cpu_rq(cpu);
4800 unsigned long flags;
4803 idle->se.exec_start = sched_clock();
4805 idle->prio = idle->normal_prio = MAX_PRIO;
4806 idle->cpus_allowed = cpumask_of_cpu(cpu);
4807 __set_task_cpu(idle, cpu);
4809 spin_lock_irqsave(&rq->lock, flags);
4810 rq->curr = rq->idle = idle;
4811 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4814 spin_unlock_irqrestore(&rq->lock, flags);
4816 /* Set the preempt count _outside_ the spinlocks! */
4817 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4818 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4820 task_thread_info(idle)->preempt_count = 0;
4823 * The idle tasks have their own, simple scheduling class:
4825 idle->sched_class = &idle_sched_class;
4829 * In a system that switches off the HZ timer nohz_cpu_mask
4830 * indicates which cpus entered this state. This is used
4831 * in the rcu update to wait only for active cpus. For system
4832 * which do not switch off the HZ timer nohz_cpu_mask should
4833 * always be CPU_MASK_NONE.
4835 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4839 * This is how migration works:
4841 * 1) we queue a struct migration_req structure in the source CPU's
4842 * runqueue and wake up that CPU's migration thread.
4843 * 2) we down() the locked semaphore => thread blocks.
4844 * 3) migration thread wakes up (implicitly it forces the migrated
4845 * thread off the CPU)
4846 * 4) it gets the migration request and checks whether the migrated
4847 * task is still in the wrong runqueue.
4848 * 5) if it's in the wrong runqueue then the migration thread removes
4849 * it and puts it into the right queue.
4850 * 6) migration thread up()s the semaphore.
4851 * 7) we wake up and the migration is done.
4855 * Change a given task's CPU affinity. Migrate the thread to a
4856 * proper CPU and schedule it away if the CPU it's executing on
4857 * is removed from the allowed bitmask.
4859 * NOTE: the caller must have a valid reference to the task, the
4860 * task must not exit() & deallocate itself prematurely. The
4861 * call is not atomic; no spinlocks may be held.
4863 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4865 struct migration_req req;
4866 unsigned long flags;
4870 rq = task_rq_lock(p, &flags);
4871 if (!cpus_intersects(new_mask, cpu_online_map)) {
4876 p->cpus_allowed = new_mask;
4877 /* Can the task run on the task's current CPU? If so, we're done */
4878 if (cpu_isset(task_cpu(p), new_mask))
4881 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4882 /* Need help from migration thread: drop lock and wait. */
4883 task_rq_unlock(rq, &flags);
4884 wake_up_process(rq->migration_thread);
4885 wait_for_completion(&req.done);
4886 tlb_migrate_finish(p->mm);
4890 task_rq_unlock(rq, &flags);
4894 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4897 * Move (not current) task off this cpu, onto dest cpu. We're doing
4898 * this because either it can't run here any more (set_cpus_allowed()
4899 * away from this CPU, or CPU going down), or because we're
4900 * attempting to rebalance this task on exec (sched_exec).
4902 * So we race with normal scheduler movements, but that's OK, as long
4903 * as the task is no longer on this CPU.
4905 * Returns non-zero if task was successfully migrated.
4907 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4909 struct rq *rq_dest, *rq_src;
4912 if (unlikely(cpu_is_offline(dest_cpu)))
4915 rq_src = cpu_rq(src_cpu);
4916 rq_dest = cpu_rq(dest_cpu);
4918 double_rq_lock(rq_src, rq_dest);
4919 /* Already moved. */
4920 if (task_cpu(p) != src_cpu)
4922 /* Affinity changed (again). */
4923 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4926 on_rq = p->se.on_rq;
4928 deactivate_task(rq_src, p, 0);
4930 set_task_cpu(p, dest_cpu);
4932 activate_task(rq_dest, p, 0);
4933 check_preempt_curr(rq_dest, p);
4937 double_rq_unlock(rq_src, rq_dest);
4942 * migration_thread - this is a highprio system thread that performs
4943 * thread migration by bumping thread off CPU then 'pushing' onto
4946 static int migration_thread(void *data)
4948 int cpu = (long)data;
4952 BUG_ON(rq->migration_thread != current);
4954 set_current_state(TASK_INTERRUPTIBLE);
4955 while (!kthread_should_stop()) {
4956 struct migration_req *req;
4957 struct list_head *head;
4959 spin_lock_irq(&rq->lock);
4961 if (cpu_is_offline(cpu)) {
4962 spin_unlock_irq(&rq->lock);
4966 if (rq->active_balance) {
4967 active_load_balance(rq, cpu);
4968 rq->active_balance = 0;
4971 head = &rq->migration_queue;
4973 if (list_empty(head)) {
4974 spin_unlock_irq(&rq->lock);
4976 set_current_state(TASK_INTERRUPTIBLE);
4979 req = list_entry(head->next, struct migration_req, list);
4980 list_del_init(head->next);
4982 spin_unlock(&rq->lock);
4983 __migrate_task(req->task, cpu, req->dest_cpu);
4986 complete(&req->done);
4988 __set_current_state(TASK_RUNNING);
4992 /* Wait for kthread_stop */
4993 set_current_state(TASK_INTERRUPTIBLE);
4994 while (!kthread_should_stop()) {
4996 set_current_state(TASK_INTERRUPTIBLE);
4998 __set_current_state(TASK_RUNNING);
5002 #ifdef CONFIG_HOTPLUG_CPU
5004 * Figure out where task on dead CPU should go, use force if neccessary.
5005 * NOTE: interrupts should be disabled by the caller
5007 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5009 unsigned long flags;
5016 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5017 cpus_and(mask, mask, p->cpus_allowed);
5018 dest_cpu = any_online_cpu(mask);
5020 /* On any allowed CPU? */
5021 if (dest_cpu == NR_CPUS)
5022 dest_cpu = any_online_cpu(p->cpus_allowed);
5024 /* No more Mr. Nice Guy. */
5025 if (dest_cpu == NR_CPUS) {
5026 rq = task_rq_lock(p, &flags);
5027 cpus_setall(p->cpus_allowed);
5028 dest_cpu = any_online_cpu(p->cpus_allowed);
5029 task_rq_unlock(rq, &flags);
5032 * Don't tell them about moving exiting tasks or
5033 * kernel threads (both mm NULL), since they never
5036 if (p->mm && printk_ratelimit())
5037 printk(KERN_INFO "process %d (%s) no "
5038 "longer affine to cpu%d\n",
5039 p->pid, p->comm, dead_cpu);
5041 } while (!__migrate_task(p, dead_cpu, dest_cpu));
5045 * While a dead CPU has no uninterruptible tasks queued at this point,
5046 * it might still have a nonzero ->nr_uninterruptible counter, because
5047 * for performance reasons the counter is not stricly tracking tasks to
5048 * their home CPUs. So we just add the counter to another CPU's counter,
5049 * to keep the global sum constant after CPU-down:
5051 static void migrate_nr_uninterruptible(struct rq *rq_src)
5053 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5054 unsigned long flags;
5056 local_irq_save(flags);
5057 double_rq_lock(rq_src, rq_dest);
5058 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5059 rq_src->nr_uninterruptible = 0;
5060 double_rq_unlock(rq_src, rq_dest);
5061 local_irq_restore(flags);
5064 /* Run through task list and migrate tasks from the dead cpu. */
5065 static void migrate_live_tasks(int src_cpu)
5067 struct task_struct *p, *t;
5069 write_lock_irq(&tasklist_lock);
5071 do_each_thread(t, p) {
5075 if (task_cpu(p) == src_cpu)
5076 move_task_off_dead_cpu(src_cpu, p);
5077 } while_each_thread(t, p);
5079 write_unlock_irq(&tasklist_lock);
5083 * activate_idle_task - move idle task to the _front_ of runqueue.
5085 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5087 update_rq_clock(rq);
5089 if (p->state == TASK_UNINTERRUPTIBLE)
5090 rq->nr_uninterruptible--;
5092 enqueue_task(rq, p, 0);
5093 inc_nr_running(p, rq);
5097 * Schedules idle task to be the next runnable task on current CPU.
5098 * It does so by boosting its priority to highest possible and adding it to
5099 * the _front_ of the runqueue. Used by CPU offline code.
5101 void sched_idle_next(void)
5103 int this_cpu = smp_processor_id();
5104 struct rq *rq = cpu_rq(this_cpu);
5105 struct task_struct *p = rq->idle;
5106 unsigned long flags;
5108 /* cpu has to be offline */
5109 BUG_ON(cpu_online(this_cpu));
5112 * Strictly not necessary since rest of the CPUs are stopped by now
5113 * and interrupts disabled on the current cpu.
5115 spin_lock_irqsave(&rq->lock, flags);
5117 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5119 /* Add idle task to the _front_ of its priority queue: */
5120 activate_idle_task(p, rq);
5122 spin_unlock_irqrestore(&rq->lock, flags);
5126 * Ensures that the idle task is using init_mm right before its cpu goes
5129 void idle_task_exit(void)
5131 struct mm_struct *mm = current->active_mm;
5133 BUG_ON(cpu_online(smp_processor_id()));
5136 switch_mm(mm, &init_mm, current);
5140 /* called under rq->lock with disabled interrupts */
5141 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5143 struct rq *rq = cpu_rq(dead_cpu);
5145 /* Must be exiting, otherwise would be on tasklist. */
5146 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5148 /* Cannot have done final schedule yet: would have vanished. */
5149 BUG_ON(p->state == TASK_DEAD);
5154 * Drop lock around migration; if someone else moves it,
5155 * that's OK. No task can be added to this CPU, so iteration is
5157 * NOTE: interrupts should be left disabled --dev@
5159 spin_unlock(&rq->lock);
5160 move_task_off_dead_cpu(dead_cpu, p);
5161 spin_lock(&rq->lock);
5166 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5167 static void migrate_dead_tasks(unsigned int dead_cpu)
5169 struct rq *rq = cpu_rq(dead_cpu);
5170 struct task_struct *next;
5173 if (!rq->nr_running)
5175 update_rq_clock(rq);
5176 next = pick_next_task(rq, rq->curr);
5179 migrate_dead(dead_cpu, next);
5183 #endif /* CONFIG_HOTPLUG_CPU */
5185 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5187 static struct ctl_table sd_ctl_dir[] = {
5189 .procname = "sched_domain",
5195 static struct ctl_table sd_ctl_root[] = {
5197 .ctl_name = CTL_KERN,
5198 .procname = "kernel",
5200 .child = sd_ctl_dir,
5205 static struct ctl_table *sd_alloc_ctl_entry(int n)
5207 struct ctl_table *entry =
5208 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5211 memset(entry, 0, n * sizeof(struct ctl_table));
5217 set_table_entry(struct ctl_table *entry,
5218 const char *procname, void *data, int maxlen,
5219 mode_t mode, proc_handler *proc_handler)
5221 entry->procname = procname;
5223 entry->maxlen = maxlen;
5225 entry->proc_handler = proc_handler;
5228 static struct ctl_table *
5229 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5231 struct ctl_table *table = sd_alloc_ctl_entry(12);
5233 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5234 sizeof(long), 0644, proc_doulongvec_minmax);
5235 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5236 sizeof(long), 0644, proc_doulongvec_minmax);
5237 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5238 sizeof(int), 0644, proc_dointvec_minmax);
5239 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5240 sizeof(int), 0644, proc_dointvec_minmax);
5241 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5242 sizeof(int), 0644, proc_dointvec_minmax);
5243 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5244 sizeof(int), 0644, proc_dointvec_minmax);
5245 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5246 sizeof(int), 0644, proc_dointvec_minmax);
5247 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5248 sizeof(int), 0644, proc_dointvec_minmax);
5249 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5250 sizeof(int), 0644, proc_dointvec_minmax);
5251 set_table_entry(&table[9], "cache_nice_tries",
5252 &sd->cache_nice_tries,
5253 sizeof(int), 0644, proc_dointvec_minmax);
5254 set_table_entry(&table[10], "flags", &sd->flags,
5255 sizeof(int), 0644, proc_dointvec_minmax);
5260 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5262 struct ctl_table *entry, *table;
5263 struct sched_domain *sd;
5264 int domain_num = 0, i;
5267 for_each_domain(cpu, sd)
5269 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5272 for_each_domain(cpu, sd) {
5273 snprintf(buf, 32, "domain%d", i);
5274 entry->procname = kstrdup(buf, GFP_KERNEL);
5276 entry->child = sd_alloc_ctl_domain_table(sd);
5283 static struct ctl_table_header *sd_sysctl_header;
5284 static void init_sched_domain_sysctl(void)
5286 int i, cpu_num = num_online_cpus();
5287 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5290 sd_ctl_dir[0].child = entry;
5292 for (i = 0; i < cpu_num; i++, entry++) {
5293 snprintf(buf, 32, "cpu%d", i);
5294 entry->procname = kstrdup(buf, GFP_KERNEL);
5296 entry->child = sd_alloc_ctl_cpu_table(i);
5298 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5301 static void init_sched_domain_sysctl(void)
5307 * migration_call - callback that gets triggered when a CPU is added.
5308 * Here we can start up the necessary migration thread for the new CPU.
5310 static int __cpuinit
5311 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5313 struct task_struct *p;
5314 int cpu = (long)hcpu;
5315 unsigned long flags;
5319 case CPU_LOCK_ACQUIRE:
5320 mutex_lock(&sched_hotcpu_mutex);
5323 case CPU_UP_PREPARE:
5324 case CPU_UP_PREPARE_FROZEN:
5325 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5328 kthread_bind(p, cpu);
5329 /* Must be high prio: stop_machine expects to yield to it. */
5330 rq = task_rq_lock(p, &flags);
5331 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5332 task_rq_unlock(rq, &flags);
5333 cpu_rq(cpu)->migration_thread = p;
5337 case CPU_ONLINE_FROZEN:
5338 /* Strictly unneccessary, as first user will wake it. */
5339 wake_up_process(cpu_rq(cpu)->migration_thread);
5342 #ifdef CONFIG_HOTPLUG_CPU
5343 case CPU_UP_CANCELED:
5344 case CPU_UP_CANCELED_FROZEN:
5345 if (!cpu_rq(cpu)->migration_thread)
5347 /* Unbind it from offline cpu so it can run. Fall thru. */
5348 kthread_bind(cpu_rq(cpu)->migration_thread,
5349 any_online_cpu(cpu_online_map));
5350 kthread_stop(cpu_rq(cpu)->migration_thread);
5351 cpu_rq(cpu)->migration_thread = NULL;
5355 case CPU_DEAD_FROZEN:
5356 migrate_live_tasks(cpu);
5358 kthread_stop(rq->migration_thread);
5359 rq->migration_thread = NULL;
5360 /* Idle task back to normal (off runqueue, low prio) */
5361 rq = task_rq_lock(rq->idle, &flags);
5362 update_rq_clock(rq);
5363 deactivate_task(rq, rq->idle, 0);
5364 rq->idle->static_prio = MAX_PRIO;
5365 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5366 rq->idle->sched_class = &idle_sched_class;
5367 migrate_dead_tasks(cpu);
5368 task_rq_unlock(rq, &flags);
5369 migrate_nr_uninterruptible(rq);
5370 BUG_ON(rq->nr_running != 0);
5372 /* No need to migrate the tasks: it was best-effort if
5373 * they didn't take sched_hotcpu_mutex. Just wake up
5374 * the requestors. */
5375 spin_lock_irq(&rq->lock);
5376 while (!list_empty(&rq->migration_queue)) {
5377 struct migration_req *req;
5379 req = list_entry(rq->migration_queue.next,
5380 struct migration_req, list);
5381 list_del_init(&req->list);
5382 complete(&req->done);
5384 spin_unlock_irq(&rq->lock);
5387 case CPU_LOCK_RELEASE:
5388 mutex_unlock(&sched_hotcpu_mutex);
5394 /* Register at highest priority so that task migration (migrate_all_tasks)
5395 * happens before everything else.
5397 static struct notifier_block __cpuinitdata migration_notifier = {
5398 .notifier_call = migration_call,
5402 int __init migration_init(void)
5404 void *cpu = (void *)(long)smp_processor_id();
5407 /* Start one for the boot CPU: */
5408 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5409 BUG_ON(err == NOTIFY_BAD);
5410 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5411 register_cpu_notifier(&migration_notifier);
5419 /* Number of possible processor ids */
5420 int nr_cpu_ids __read_mostly = NR_CPUS;
5421 EXPORT_SYMBOL(nr_cpu_ids);
5423 #ifdef CONFIG_SCHED_DEBUG
5424 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5429 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5433 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5438 struct sched_group *group = sd->groups;
5439 cpumask_t groupmask;
5441 cpumask_scnprintf(str, NR_CPUS, sd->span);
5442 cpus_clear(groupmask);
5445 for (i = 0; i < level + 1; i++)
5447 printk("domain %d: ", level);
5449 if (!(sd->flags & SD_LOAD_BALANCE)) {
5450 printk("does not load-balance\n");
5452 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5457 printk("span %s\n", str);
5459 if (!cpu_isset(cpu, sd->span))
5460 printk(KERN_ERR "ERROR: domain->span does not contain "
5462 if (!cpu_isset(cpu, group->cpumask))
5463 printk(KERN_ERR "ERROR: domain->groups does not contain"
5467 for (i = 0; i < level + 2; i++)
5473 printk(KERN_ERR "ERROR: group is NULL\n");
5477 if (!group->__cpu_power) {
5479 printk(KERN_ERR "ERROR: domain->cpu_power not "
5484 if (!cpus_weight(group->cpumask)) {
5486 printk(KERN_ERR "ERROR: empty group\n");
5490 if (cpus_intersects(groupmask, group->cpumask)) {
5492 printk(KERN_ERR "ERROR: repeated CPUs\n");
5496 cpus_or(groupmask, groupmask, group->cpumask);
5498 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5501 group = group->next;
5502 } while (group != sd->groups);
5505 if (!cpus_equal(sd->span, groupmask))
5506 printk(KERN_ERR "ERROR: groups don't span "
5514 if (!cpus_subset(groupmask, sd->span))
5515 printk(KERN_ERR "ERROR: parent span is not a superset "
5516 "of domain->span\n");
5521 # define sched_domain_debug(sd, cpu) do { } while (0)
5524 static int sd_degenerate(struct sched_domain *sd)
5526 if (cpus_weight(sd->span) == 1)
5529 /* Following flags need at least 2 groups */
5530 if (sd->flags & (SD_LOAD_BALANCE |
5531 SD_BALANCE_NEWIDLE |
5535 SD_SHARE_PKG_RESOURCES)) {
5536 if (sd->groups != sd->groups->next)
5540 /* Following flags don't use groups */
5541 if (sd->flags & (SD_WAKE_IDLE |
5550 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5552 unsigned long cflags = sd->flags, pflags = parent->flags;
5554 if (sd_degenerate(parent))
5557 if (!cpus_equal(sd->span, parent->span))
5560 /* Does parent contain flags not in child? */
5561 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5562 if (cflags & SD_WAKE_AFFINE)
5563 pflags &= ~SD_WAKE_BALANCE;
5564 /* Flags needing groups don't count if only 1 group in parent */
5565 if (parent->groups == parent->groups->next) {
5566 pflags &= ~(SD_LOAD_BALANCE |
5567 SD_BALANCE_NEWIDLE |
5571 SD_SHARE_PKG_RESOURCES);
5573 if (~cflags & pflags)
5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5581 * hold the hotplug lock.
5583 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5585 struct rq *rq = cpu_rq(cpu);
5586 struct sched_domain *tmp;
5588 /* Remove the sched domains which do not contribute to scheduling. */
5589 for (tmp = sd; tmp; tmp = tmp->parent) {
5590 struct sched_domain *parent = tmp->parent;
5593 if (sd_parent_degenerate(tmp, parent)) {
5594 tmp->parent = parent->parent;
5596 parent->parent->child = tmp;
5600 if (sd && sd_degenerate(sd)) {
5606 sched_domain_debug(sd, cpu);
5608 rcu_assign_pointer(rq->sd, sd);
5611 /* cpus with isolated domains */
5612 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5614 /* Setup the mask of cpus configured for isolated domains */
5615 static int __init isolated_cpu_setup(char *str)
5617 int ints[NR_CPUS], i;
5619 str = get_options(str, ARRAY_SIZE(ints), ints);
5620 cpus_clear(cpu_isolated_map);
5621 for (i = 1; i <= ints[0]; i++)
5622 if (ints[i] < NR_CPUS)
5623 cpu_set(ints[i], cpu_isolated_map);
5627 __setup("isolcpus=", isolated_cpu_setup);
5630 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5631 * to a function which identifies what group(along with sched group) a CPU
5632 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5633 * (due to the fact that we keep track of groups covered with a cpumask_t).
5635 * init_sched_build_groups will build a circular linked list of the groups
5636 * covered by the given span, and will set each group's ->cpumask correctly,
5637 * and ->cpu_power to 0.
5640 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5641 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5642 struct sched_group **sg))
5644 struct sched_group *first = NULL, *last = NULL;
5645 cpumask_t covered = CPU_MASK_NONE;
5648 for_each_cpu_mask(i, span) {
5649 struct sched_group *sg;
5650 int group = group_fn(i, cpu_map, &sg);
5653 if (cpu_isset(i, covered))
5656 sg->cpumask = CPU_MASK_NONE;
5657 sg->__cpu_power = 0;
5659 for_each_cpu_mask(j, span) {
5660 if (group_fn(j, cpu_map, NULL) != group)
5663 cpu_set(j, covered);
5664 cpu_set(j, sg->cpumask);
5675 #define SD_NODES_PER_DOMAIN 16
5680 * find_next_best_node - find the next node to include in a sched_domain
5681 * @node: node whose sched_domain we're building
5682 * @used_nodes: nodes already in the sched_domain
5684 * Find the next node to include in a given scheduling domain. Simply
5685 * finds the closest node not already in the @used_nodes map.
5687 * Should use nodemask_t.
5689 static int find_next_best_node(int node, unsigned long *used_nodes)
5691 int i, n, val, min_val, best_node = 0;
5695 for (i = 0; i < MAX_NUMNODES; i++) {
5696 /* Start at @node */
5697 n = (node + i) % MAX_NUMNODES;
5699 if (!nr_cpus_node(n))
5702 /* Skip already used nodes */
5703 if (test_bit(n, used_nodes))
5706 /* Simple min distance search */
5707 val = node_distance(node, n);
5709 if (val < min_val) {
5715 set_bit(best_node, used_nodes);
5720 * sched_domain_node_span - get a cpumask for a node's sched_domain
5721 * @node: node whose cpumask we're constructing
5722 * @size: number of nodes to include in this span
5724 * Given a node, construct a good cpumask for its sched_domain to span. It
5725 * should be one that prevents unnecessary balancing, but also spreads tasks
5728 static cpumask_t sched_domain_node_span(int node)
5730 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5731 cpumask_t span, nodemask;
5735 bitmap_zero(used_nodes, MAX_NUMNODES);
5737 nodemask = node_to_cpumask(node);
5738 cpus_or(span, span, nodemask);
5739 set_bit(node, used_nodes);
5741 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5742 int next_node = find_next_best_node(node, used_nodes);
5744 nodemask = node_to_cpumask(next_node);
5745 cpus_or(span, span, nodemask);
5752 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5755 * SMT sched-domains:
5757 #ifdef CONFIG_SCHED_SMT
5758 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5759 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5761 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5762 struct sched_group **sg)
5765 *sg = &per_cpu(sched_group_cpus, cpu);
5771 * multi-core sched-domains:
5773 #ifdef CONFIG_SCHED_MC
5774 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5775 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5778 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5779 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5780 struct sched_group **sg)
5783 cpumask_t mask = cpu_sibling_map[cpu];
5784 cpus_and(mask, mask, *cpu_map);
5785 group = first_cpu(mask);
5787 *sg = &per_cpu(sched_group_core, group);
5790 #elif defined(CONFIG_SCHED_MC)
5791 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5792 struct sched_group **sg)
5795 *sg = &per_cpu(sched_group_core, cpu);
5800 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5801 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5803 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5804 struct sched_group **sg)
5807 #ifdef CONFIG_SCHED_MC
5808 cpumask_t mask = cpu_coregroup_map(cpu);
5809 cpus_and(mask, mask, *cpu_map);
5810 group = first_cpu(mask);
5811 #elif defined(CONFIG_SCHED_SMT)
5812 cpumask_t mask = cpu_sibling_map[cpu];
5813 cpus_and(mask, mask, *cpu_map);
5814 group = first_cpu(mask);
5819 *sg = &per_cpu(sched_group_phys, group);
5825 * The init_sched_build_groups can't handle what we want to do with node
5826 * groups, so roll our own. Now each node has its own list of groups which
5827 * gets dynamically allocated.
5829 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5830 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5832 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5833 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5835 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5836 struct sched_group **sg)
5838 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5841 cpus_and(nodemask, nodemask, *cpu_map);
5842 group = first_cpu(nodemask);
5845 *sg = &per_cpu(sched_group_allnodes, group);
5849 static void init_numa_sched_groups_power(struct sched_group *group_head)
5851 struct sched_group *sg = group_head;
5857 for_each_cpu_mask(j, sg->cpumask) {
5858 struct sched_domain *sd;
5860 sd = &per_cpu(phys_domains, j);
5861 if (j != first_cpu(sd->groups->cpumask)) {
5863 * Only add "power" once for each
5869 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5872 } while (sg != group_head);
5877 /* Free memory allocated for various sched_group structures */
5878 static void free_sched_groups(const cpumask_t *cpu_map)
5882 for_each_cpu_mask(cpu, *cpu_map) {
5883 struct sched_group **sched_group_nodes
5884 = sched_group_nodes_bycpu[cpu];
5886 if (!sched_group_nodes)
5889 for (i = 0; i < MAX_NUMNODES; i++) {
5890 cpumask_t nodemask = node_to_cpumask(i);
5891 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5893 cpus_and(nodemask, nodemask, *cpu_map);
5894 if (cpus_empty(nodemask))
5904 if (oldsg != sched_group_nodes[i])
5907 kfree(sched_group_nodes);
5908 sched_group_nodes_bycpu[cpu] = NULL;
5912 static void free_sched_groups(const cpumask_t *cpu_map)
5918 * Initialize sched groups cpu_power.
5920 * cpu_power indicates the capacity of sched group, which is used while
5921 * distributing the load between different sched groups in a sched domain.
5922 * Typically cpu_power for all the groups in a sched domain will be same unless
5923 * there are asymmetries in the topology. If there are asymmetries, group
5924 * having more cpu_power will pickup more load compared to the group having
5927 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5928 * the maximum number of tasks a group can handle in the presence of other idle
5929 * or lightly loaded groups in the same sched domain.
5931 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5933 struct sched_domain *child;
5934 struct sched_group *group;
5936 WARN_ON(!sd || !sd->groups);
5938 if (cpu != first_cpu(sd->groups->cpumask))
5943 sd->groups->__cpu_power = 0;
5946 * For perf policy, if the groups in child domain share resources
5947 * (for example cores sharing some portions of the cache hierarchy
5948 * or SMT), then set this domain groups cpu_power such that each group
5949 * can handle only one task, when there are other idle groups in the
5950 * same sched domain.
5952 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5954 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5955 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5960 * add cpu_power of each child group to this groups cpu_power
5962 group = child->groups;
5964 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5965 group = group->next;
5966 } while (group != child->groups);
5970 * Build sched domains for a given set of cpus and attach the sched domains
5971 * to the individual cpus
5973 static int build_sched_domains(const cpumask_t *cpu_map)
5977 struct sched_group **sched_group_nodes = NULL;
5978 int sd_allnodes = 0;
5981 * Allocate the per-node list of sched groups
5983 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
5985 if (!sched_group_nodes) {
5986 printk(KERN_WARNING "Can not alloc sched group node list\n");
5989 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
5993 * Set up domains for cpus specified by the cpu_map.
5995 for_each_cpu_mask(i, *cpu_map) {
5996 struct sched_domain *sd = NULL, *p;
5997 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
5999 cpus_and(nodemask, nodemask, *cpu_map);
6002 if (cpus_weight(*cpu_map) >
6003 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6004 sd = &per_cpu(allnodes_domains, i);
6005 *sd = SD_ALLNODES_INIT;
6006 sd->span = *cpu_map;
6007 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6013 sd = &per_cpu(node_domains, i);
6015 sd->span = sched_domain_node_span(cpu_to_node(i));
6019 cpus_and(sd->span, sd->span, *cpu_map);
6023 sd = &per_cpu(phys_domains, i);
6025 sd->span = nodemask;
6029 cpu_to_phys_group(i, cpu_map, &sd->groups);
6031 #ifdef CONFIG_SCHED_MC
6033 sd = &per_cpu(core_domains, i);
6035 sd->span = cpu_coregroup_map(i);
6036 cpus_and(sd->span, sd->span, *cpu_map);
6039 cpu_to_core_group(i, cpu_map, &sd->groups);
6042 #ifdef CONFIG_SCHED_SMT
6044 sd = &per_cpu(cpu_domains, i);
6045 *sd = SD_SIBLING_INIT;
6046 sd->span = cpu_sibling_map[i];
6047 cpus_and(sd->span, sd->span, *cpu_map);
6050 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6054 #ifdef CONFIG_SCHED_SMT
6055 /* Set up CPU (sibling) groups */
6056 for_each_cpu_mask(i, *cpu_map) {
6057 cpumask_t this_sibling_map = cpu_sibling_map[i];
6058 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6059 if (i != first_cpu(this_sibling_map))
6062 init_sched_build_groups(this_sibling_map, cpu_map,
6067 #ifdef CONFIG_SCHED_MC
6068 /* Set up multi-core groups */
6069 for_each_cpu_mask(i, *cpu_map) {
6070 cpumask_t this_core_map = cpu_coregroup_map(i);
6071 cpus_and(this_core_map, this_core_map, *cpu_map);
6072 if (i != first_cpu(this_core_map))
6074 init_sched_build_groups(this_core_map, cpu_map,
6075 &cpu_to_core_group);
6079 /* Set up physical groups */
6080 for (i = 0; i < MAX_NUMNODES; i++) {
6081 cpumask_t nodemask = node_to_cpumask(i);
6083 cpus_and(nodemask, nodemask, *cpu_map);
6084 if (cpus_empty(nodemask))
6087 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6091 /* Set up node groups */
6093 init_sched_build_groups(*cpu_map, cpu_map,
6094 &cpu_to_allnodes_group);
6096 for (i = 0; i < MAX_NUMNODES; i++) {
6097 /* Set up node groups */
6098 struct sched_group *sg, *prev;
6099 cpumask_t nodemask = node_to_cpumask(i);
6100 cpumask_t domainspan;
6101 cpumask_t covered = CPU_MASK_NONE;
6104 cpus_and(nodemask, nodemask, *cpu_map);
6105 if (cpus_empty(nodemask)) {
6106 sched_group_nodes[i] = NULL;
6110 domainspan = sched_domain_node_span(i);
6111 cpus_and(domainspan, domainspan, *cpu_map);
6113 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6115 printk(KERN_WARNING "Can not alloc domain group for "
6119 sched_group_nodes[i] = sg;
6120 for_each_cpu_mask(j, nodemask) {
6121 struct sched_domain *sd;
6123 sd = &per_cpu(node_domains, j);
6126 sg->__cpu_power = 0;
6127 sg->cpumask = nodemask;
6129 cpus_or(covered, covered, nodemask);
6132 for (j = 0; j < MAX_NUMNODES; j++) {
6133 cpumask_t tmp, notcovered;
6134 int n = (i + j) % MAX_NUMNODES;
6136 cpus_complement(notcovered, covered);
6137 cpus_and(tmp, notcovered, *cpu_map);
6138 cpus_and(tmp, tmp, domainspan);
6139 if (cpus_empty(tmp))
6142 nodemask = node_to_cpumask(n);
6143 cpus_and(tmp, tmp, nodemask);
6144 if (cpus_empty(tmp))
6147 sg = kmalloc_node(sizeof(struct sched_group),
6151 "Can not alloc domain group for node %d\n", j);
6154 sg->__cpu_power = 0;
6156 sg->next = prev->next;
6157 cpus_or(covered, covered, tmp);
6164 /* Calculate CPU power for physical packages and nodes */
6165 #ifdef CONFIG_SCHED_SMT
6166 for_each_cpu_mask(i, *cpu_map) {
6167 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6169 init_sched_groups_power(i, sd);
6172 #ifdef CONFIG_SCHED_MC
6173 for_each_cpu_mask(i, *cpu_map) {
6174 struct sched_domain *sd = &per_cpu(core_domains, i);
6176 init_sched_groups_power(i, sd);
6180 for_each_cpu_mask(i, *cpu_map) {
6181 struct sched_domain *sd = &per_cpu(phys_domains, i);
6183 init_sched_groups_power(i, sd);
6187 for (i = 0; i < MAX_NUMNODES; i++)
6188 init_numa_sched_groups_power(sched_group_nodes[i]);
6191 struct sched_group *sg;
6193 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6194 init_numa_sched_groups_power(sg);
6198 /* Attach the domains */
6199 for_each_cpu_mask(i, *cpu_map) {
6200 struct sched_domain *sd;
6201 #ifdef CONFIG_SCHED_SMT
6202 sd = &per_cpu(cpu_domains, i);
6203 #elif defined(CONFIG_SCHED_MC)
6204 sd = &per_cpu(core_domains, i);
6206 sd = &per_cpu(phys_domains, i);
6208 cpu_attach_domain(sd, i);
6215 free_sched_groups(cpu_map);
6220 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6222 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6224 cpumask_t cpu_default_map;
6228 * Setup mask for cpus without special case scheduling requirements.
6229 * For now this just excludes isolated cpus, but could be used to
6230 * exclude other special cases in the future.
6232 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6234 err = build_sched_domains(&cpu_default_map);
6239 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6241 free_sched_groups(cpu_map);
6245 * Detach sched domains from a group of cpus specified in cpu_map
6246 * These cpus will now be attached to the NULL domain
6248 static void detach_destroy_domains(const cpumask_t *cpu_map)
6252 for_each_cpu_mask(i, *cpu_map)
6253 cpu_attach_domain(NULL, i);
6254 synchronize_sched();
6255 arch_destroy_sched_domains(cpu_map);
6259 * Partition sched domains as specified by the cpumasks below.
6260 * This attaches all cpus from the cpumasks to the NULL domain,
6261 * waits for a RCU quiescent period, recalculates sched
6262 * domain information and then attaches them back to the
6263 * correct sched domains
6264 * Call with hotplug lock held
6266 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6268 cpumask_t change_map;
6271 cpus_and(*partition1, *partition1, cpu_online_map);
6272 cpus_and(*partition2, *partition2, cpu_online_map);
6273 cpus_or(change_map, *partition1, *partition2);
6275 /* Detach sched domains from all of the affected cpus */
6276 detach_destroy_domains(&change_map);
6277 if (!cpus_empty(*partition1))
6278 err = build_sched_domains(partition1);
6279 if (!err && !cpus_empty(*partition2))
6280 err = build_sched_domains(partition2);
6285 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6286 static int arch_reinit_sched_domains(void)
6290 mutex_lock(&sched_hotcpu_mutex);
6291 detach_destroy_domains(&cpu_online_map);
6292 err = arch_init_sched_domains(&cpu_online_map);
6293 mutex_unlock(&sched_hotcpu_mutex);
6298 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6302 if (buf[0] != '0' && buf[0] != '1')
6306 sched_smt_power_savings = (buf[0] == '1');
6308 sched_mc_power_savings = (buf[0] == '1');
6310 ret = arch_reinit_sched_domains();
6312 return ret ? ret : count;
6315 #ifdef CONFIG_SCHED_MC
6316 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6318 return sprintf(page, "%u\n", sched_mc_power_savings);
6320 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6321 const char *buf, size_t count)
6323 return sched_power_savings_store(buf, count, 0);
6325 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6326 sched_mc_power_savings_store);
6329 #ifdef CONFIG_SCHED_SMT
6330 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6332 return sprintf(page, "%u\n", sched_smt_power_savings);
6334 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6335 const char *buf, size_t count)
6337 return sched_power_savings_store(buf, count, 1);
6339 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6340 sched_smt_power_savings_store);
6343 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6347 #ifdef CONFIG_SCHED_SMT
6349 err = sysfs_create_file(&cls->kset.kobj,
6350 &attr_sched_smt_power_savings.attr);
6352 #ifdef CONFIG_SCHED_MC
6353 if (!err && mc_capable())
6354 err = sysfs_create_file(&cls->kset.kobj,
6355 &attr_sched_mc_power_savings.attr);
6362 * Force a reinitialization of the sched domains hierarchy. The domains
6363 * and groups cannot be updated in place without racing with the balancing
6364 * code, so we temporarily attach all running cpus to the NULL domain
6365 * which will prevent rebalancing while the sched domains are recalculated.
6367 static int update_sched_domains(struct notifier_block *nfb,
6368 unsigned long action, void *hcpu)
6371 case CPU_UP_PREPARE:
6372 case CPU_UP_PREPARE_FROZEN:
6373 case CPU_DOWN_PREPARE:
6374 case CPU_DOWN_PREPARE_FROZEN:
6375 detach_destroy_domains(&cpu_online_map);
6378 case CPU_UP_CANCELED:
6379 case CPU_UP_CANCELED_FROZEN:
6380 case CPU_DOWN_FAILED:
6381 case CPU_DOWN_FAILED_FROZEN:
6383 case CPU_ONLINE_FROZEN:
6385 case CPU_DEAD_FROZEN:
6387 * Fall through and re-initialise the domains.
6394 /* The hotplug lock is already held by cpu_up/cpu_down */
6395 arch_init_sched_domains(&cpu_online_map);
6400 void __init sched_init_smp(void)
6402 cpumask_t non_isolated_cpus;
6404 mutex_lock(&sched_hotcpu_mutex);
6405 arch_init_sched_domains(&cpu_online_map);
6406 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6407 if (cpus_empty(non_isolated_cpus))
6408 cpu_set(smp_processor_id(), non_isolated_cpus);
6409 mutex_unlock(&sched_hotcpu_mutex);
6410 /* XXX: Theoretical race here - CPU may be hotplugged now */
6411 hotcpu_notifier(update_sched_domains, 0);
6413 init_sched_domain_sysctl();
6415 /* Move init over to a non-isolated CPU */
6416 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6420 void __init sched_init_smp(void)
6423 #endif /* CONFIG_SMP */
6425 int in_sched_functions(unsigned long addr)
6427 /* Linker adds these: start and end of __sched functions */
6428 extern char __sched_text_start[], __sched_text_end[];
6430 return in_lock_functions(addr) ||
6431 (addr >= (unsigned long)__sched_text_start
6432 && addr < (unsigned long)__sched_text_end);
6435 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6437 cfs_rq->tasks_timeline = RB_ROOT;
6438 #ifdef CONFIG_FAIR_GROUP_SCHED
6441 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6444 void __init sched_init(void)
6446 int highest_cpu = 0;
6449 for_each_possible_cpu(i) {
6450 struct rt_prio_array *array;
6454 spin_lock_init(&rq->lock);
6455 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6458 init_cfs_rq(&rq->cfs, rq);
6459 #ifdef CONFIG_FAIR_GROUP_SCHED
6460 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6462 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6463 struct sched_entity *se =
6464 &per_cpu(init_sched_entity, i);
6466 init_cfs_rq_p[i] = cfs_rq;
6467 init_cfs_rq(cfs_rq, rq);
6468 cfs_rq->tg = &init_task_group;
6469 list_add(&cfs_rq->leaf_cfs_rq_list,
6470 &rq->leaf_cfs_rq_list);
6472 init_sched_entity_p[i] = se;
6473 se->cfs_rq = &rq->cfs;
6475 se->load.weight = init_task_group_load;
6476 se->load.inv_weight =
6477 div64_64(1ULL<<32, init_task_group_load);
6480 init_task_group.shares = init_task_group_load;
6481 spin_lock_init(&init_task_group.lock);
6484 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6485 rq->cpu_load[j] = 0;
6488 rq->active_balance = 0;
6489 rq->next_balance = jiffies;
6492 rq->migration_thread = NULL;
6493 INIT_LIST_HEAD(&rq->migration_queue);
6495 atomic_set(&rq->nr_iowait, 0);
6497 array = &rq->rt.active;
6498 for (j = 0; j < MAX_RT_PRIO; j++) {
6499 INIT_LIST_HEAD(array->queue + j);
6500 __clear_bit(j, array->bitmap);
6503 /* delimiter for bitsearch: */
6504 __set_bit(MAX_RT_PRIO, array->bitmap);
6507 set_load_weight(&init_task);
6509 #ifdef CONFIG_PREEMPT_NOTIFIERS
6510 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6514 nr_cpu_ids = highest_cpu + 1;
6515 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6518 #ifdef CONFIG_RT_MUTEXES
6519 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6523 * The boot idle thread does lazy MMU switching as well:
6525 atomic_inc(&init_mm.mm_count);
6526 enter_lazy_tlb(&init_mm, current);
6529 * Make us the idle thread. Technically, schedule() should not be
6530 * called from this thread, however somewhere below it might be,
6531 * but because we are the idle thread, we just pick up running again
6532 * when this runqueue becomes "idle".
6534 init_idle(current, smp_processor_id());
6536 * During early bootup we pretend to be a normal task:
6538 current->sched_class = &fair_sched_class;
6541 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6542 void __might_sleep(char *file, int line)
6545 static unsigned long prev_jiffy; /* ratelimiting */
6547 if ((in_atomic() || irqs_disabled()) &&
6548 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6549 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6551 prev_jiffy = jiffies;
6552 printk(KERN_ERR "BUG: sleeping function called from invalid"
6553 " context at %s:%d\n", file, line);
6554 printk("in_atomic():%d, irqs_disabled():%d\n",
6555 in_atomic(), irqs_disabled());
6556 debug_show_held_locks(current);
6557 if (irqs_disabled())
6558 print_irqtrace_events(current);
6563 EXPORT_SYMBOL(__might_sleep);
6566 #ifdef CONFIG_MAGIC_SYSRQ
6567 static void normalize_task(struct rq *rq, struct task_struct *p)
6570 update_rq_clock(rq);
6571 on_rq = p->se.on_rq;
6573 deactivate_task(rq, p, 0);
6574 __setscheduler(rq, p, SCHED_NORMAL, 0);
6576 activate_task(rq, p, 0);
6577 resched_task(rq->curr);
6581 void normalize_rt_tasks(void)
6583 struct task_struct *g, *p;
6584 unsigned long flags;
6587 read_lock_irq(&tasklist_lock);
6588 do_each_thread(g, p) {
6590 * Only normalize user tasks:
6595 p->se.exec_start = 0;
6596 #ifdef CONFIG_SCHEDSTATS
6597 p->se.wait_start = 0;
6598 p->se.sleep_start = 0;
6599 p->se.block_start = 0;
6601 task_rq(p)->clock = 0;
6605 * Renice negative nice level userspace
6608 if (TASK_NICE(p) < 0 && p->mm)
6609 set_user_nice(p, 0);
6613 spin_lock_irqsave(&p->pi_lock, flags);
6614 rq = __task_rq_lock(p);
6616 normalize_task(rq, p);
6618 __task_rq_unlock(rq);
6619 spin_unlock_irqrestore(&p->pi_lock, flags);
6620 } while_each_thread(g, p);
6622 read_unlock_irq(&tasklist_lock);
6625 #endif /* CONFIG_MAGIC_SYSRQ */
6629 * These functions are only useful for the IA64 MCA handling.
6631 * They can only be called when the whole system has been
6632 * stopped - every CPU needs to be quiescent, and no scheduling
6633 * activity can take place. Using them for anything else would
6634 * be a serious bug, and as a result, they aren't even visible
6635 * under any other configuration.
6639 * curr_task - return the current task for a given cpu.
6640 * @cpu: the processor in question.
6642 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6644 struct task_struct *curr_task(int cpu)
6646 return cpu_curr(cpu);
6650 * set_curr_task - set the current task for a given cpu.
6651 * @cpu: the processor in question.
6652 * @p: the task pointer to set.
6654 * Description: This function must only be used when non-maskable interrupts
6655 * are serviced on a separate stack. It allows the architecture to switch the
6656 * notion of the current task on a cpu in a non-blocking manner. This function
6657 * must be called with all CPU's synchronized, and interrupts disabled, the
6658 * and caller must save the original value of the current task (see
6659 * curr_task() above) and restore that value before reenabling interrupts and
6660 * re-starting the system.
6662 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6664 void set_curr_task(int cpu, struct task_struct *p)
6671 #ifdef CONFIG_FAIR_GROUP_SCHED
6673 /* allocate runqueue etc for a new task group */
6674 struct task_group *sched_create_group(void)
6676 struct task_group *tg;
6677 struct cfs_rq *cfs_rq;
6678 struct sched_entity *se;
6682 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6684 return ERR_PTR(-ENOMEM);
6686 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6689 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6693 for_each_possible_cpu(i) {
6696 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6701 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6706 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6707 memset(se, 0, sizeof(struct sched_entity));
6709 tg->cfs_rq[i] = cfs_rq;
6710 init_cfs_rq(cfs_rq, rq);
6714 se->cfs_rq = &rq->cfs;
6716 se->load.weight = NICE_0_LOAD;
6717 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6721 for_each_possible_cpu(i) {
6723 cfs_rq = tg->cfs_rq[i];
6724 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6727 tg->shares = NICE_0_LOAD;
6728 spin_lock_init(&tg->lock);
6733 for_each_possible_cpu(i) {
6735 kfree(tg->cfs_rq[i]);
6743 return ERR_PTR(-ENOMEM);
6746 /* rcu callback to free various structures associated with a task group */
6747 static void free_sched_group(struct rcu_head *rhp)
6749 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6750 struct task_group *tg = cfs_rq->tg;
6751 struct sched_entity *se;
6754 /* now it should be safe to free those cfs_rqs */
6755 for_each_possible_cpu(i) {
6756 cfs_rq = tg->cfs_rq[i];
6768 /* Destroy runqueue etc associated with a task group */
6769 void sched_destroy_group(struct task_group *tg)
6771 struct cfs_rq *cfs_rq;
6774 for_each_possible_cpu(i) {
6775 cfs_rq = tg->cfs_rq[i];
6776 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6779 cfs_rq = tg->cfs_rq[0];
6781 /* wait for possible concurrent references to cfs_rqs complete */
6782 call_rcu(&cfs_rq->rcu, free_sched_group);
6785 /* change task's runqueue when it moves between groups.
6786 * The caller of this function should have put the task in its new group
6787 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6788 * reflect its new group.
6790 void sched_move_task(struct task_struct *tsk)
6793 unsigned long flags;
6796 rq = task_rq_lock(tsk, &flags);
6798 if (tsk->sched_class != &fair_sched_class)
6801 update_rq_clock(rq);
6803 running = task_running(rq, tsk);
6804 on_rq = tsk->se.on_rq;
6807 dequeue_task(rq, tsk, 0);
6808 if (unlikely(running))
6809 tsk->sched_class->put_prev_task(rq, tsk);
6812 set_task_cfs_rq(tsk);
6815 if (unlikely(running))
6816 tsk->sched_class->set_curr_task(rq);
6817 enqueue_task(rq, tsk, 0);
6821 task_rq_unlock(rq, &flags);
6824 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6826 struct cfs_rq *cfs_rq = se->cfs_rq;
6827 struct rq *rq = cfs_rq->rq;
6830 spin_lock_irq(&rq->lock);
6834 dequeue_entity(cfs_rq, se, 0);
6836 se->load.weight = shares;
6837 se->load.inv_weight = div64_64((1ULL<<32), shares);
6840 enqueue_entity(cfs_rq, se, 0);
6842 spin_unlock_irq(&rq->lock);
6845 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6849 spin_lock(&tg->lock);
6850 if (tg->shares == shares)
6853 tg->shares = shares;
6854 for_each_possible_cpu(i)
6855 set_se_shares(tg->se[i], shares);
6858 spin_unlock(&tg->lock);
6862 unsigned long sched_group_shares(struct task_group *tg)
6867 #endif /* CONFIG_FAIR_GROUP_SCHED */