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/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak)) sched_clock(void)
77 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 static inline int rt_policy(int policy)
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
143 static inline int task_has_rt_policy(struct task_struct *p)
145 return rt_policy(p->policy);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
171 unsigned long shares;
175 /* Default task group's sched entity on each cpu */
176 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
177 /* Default task group's cfs_rq on each cpu */
178 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
180 static struct sched_entity *init_sched_entity_p[NR_CPUS];
181 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
183 /* task_group_mutex serializes add/remove of task groups and also changes to
184 * a task group's cpu shares.
186 static DEFINE_MUTEX(task_group_mutex);
188 /* doms_cur_mutex serializes access to doms_cur[] array */
189 static DEFINE_MUTEX(doms_cur_mutex);
191 /* Default task group.
192 * Every task in system belong to this group at bootup.
194 struct task_group init_task_group = {
195 .se = init_sched_entity_p,
196 .cfs_rq = init_cfs_rq_p,
199 #ifdef CONFIG_FAIR_USER_SCHED
200 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
202 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
205 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
207 /* return group to which a task belongs */
208 static inline struct task_group *task_group(struct task_struct *p)
210 struct task_group *tg;
212 #ifdef CONFIG_FAIR_USER_SCHED
214 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
215 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
216 struct task_group, css);
218 tg = &init_task_group;
223 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
224 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
226 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
227 p->se.parent = task_group(p)->se[cpu];
230 static inline void lock_task_group_list(void)
232 mutex_lock(&task_group_mutex);
235 static inline void unlock_task_group_list(void)
237 mutex_unlock(&task_group_mutex);
240 static inline void lock_doms_cur(void)
242 mutex_lock(&doms_cur_mutex);
245 static inline void unlock_doms_cur(void)
247 mutex_unlock(&doms_cur_mutex);
252 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
253 static inline void lock_task_group_list(void) { }
254 static inline void unlock_task_group_list(void) { }
255 static inline void lock_doms_cur(void) { }
256 static inline void unlock_doms_cur(void) { }
258 #endif /* CONFIG_FAIR_GROUP_SCHED */
260 /* CFS-related fields in a runqueue */
262 struct load_weight load;
263 unsigned long nr_running;
268 struct rb_root tasks_timeline;
269 struct rb_node *rb_leftmost;
270 struct rb_node *rb_load_balance_curr;
271 /* 'curr' points to currently running entity on this cfs_rq.
272 * It is set to NULL otherwise (i.e when none are currently running).
274 struct sched_entity *curr;
276 unsigned long nr_spread_over;
278 #ifdef CONFIG_FAIR_GROUP_SCHED
279 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
282 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
283 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
284 * (like users, containers etc.)
286 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
287 * list is used during load balance.
289 struct list_head leaf_cfs_rq_list;
290 struct task_group *tg; /* group that "owns" this runqueue */
294 /* Real-Time classes' related field in a runqueue: */
296 struct rt_prio_array active;
297 int rt_load_balance_idx;
298 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
302 * This is the main, per-CPU runqueue data structure.
304 * Locking rule: those places that want to lock multiple runqueues
305 * (such as the load balancing or the thread migration code), lock
306 * acquire operations must be ordered by ascending &runqueue.
313 * nr_running and cpu_load should be in the same cacheline because
314 * remote CPUs use both these fields when doing load calculation.
316 unsigned long nr_running;
317 #define CPU_LOAD_IDX_MAX 5
318 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
319 unsigned char idle_at_tick;
321 unsigned char in_nohz_recently;
323 /* capture load from *all* tasks on this cpu: */
324 struct load_weight load;
325 unsigned long nr_load_updates;
329 #ifdef CONFIG_FAIR_GROUP_SCHED
330 /* list of leaf cfs_rq on this cpu: */
331 struct list_head leaf_cfs_rq_list;
336 * This is part of a global counter where only the total sum
337 * over all CPUs matters. A task can increase this counter on
338 * one CPU and if it got migrated afterwards it may decrease
339 * it on another CPU. Always updated under the runqueue lock:
341 unsigned long nr_uninterruptible;
343 struct task_struct *curr, *idle;
344 unsigned long next_balance;
345 struct mm_struct *prev_mm;
347 u64 clock, prev_clock_raw;
350 unsigned int clock_warps, clock_overflows;
352 unsigned int clock_deep_idle_events;
358 struct sched_domain *sd;
360 /* For active balancing */
363 /* cpu of this runqueue: */
366 struct task_struct *migration_thread;
367 struct list_head migration_queue;
370 #ifdef CONFIG_SCHEDSTATS
372 struct sched_info rq_sched_info;
374 /* sys_sched_yield() stats */
375 unsigned int yld_exp_empty;
376 unsigned int yld_act_empty;
377 unsigned int yld_both_empty;
378 unsigned int yld_count;
380 /* schedule() stats */
381 unsigned int sched_switch;
382 unsigned int sched_count;
383 unsigned int sched_goidle;
385 /* try_to_wake_up() stats */
386 unsigned int ttwu_count;
387 unsigned int ttwu_local;
390 unsigned int bkl_count;
392 struct lock_class_key rq_lock_key;
395 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
396 static DEFINE_MUTEX(sched_hotcpu_mutex);
398 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
400 rq->curr->sched_class->check_preempt_curr(rq, p);
403 static inline int cpu_of(struct rq *rq)
413 * Update the per-runqueue clock, as finegrained as the platform can give
414 * us, but without assuming monotonicity, etc.:
416 static void __update_rq_clock(struct rq *rq)
418 u64 prev_raw = rq->prev_clock_raw;
419 u64 now = sched_clock();
420 s64 delta = now - prev_raw;
421 u64 clock = rq->clock;
423 #ifdef CONFIG_SCHED_DEBUG
424 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
427 * Protect against sched_clock() occasionally going backwards:
429 if (unlikely(delta < 0)) {
434 * Catch too large forward jumps too:
436 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
437 if (clock < rq->tick_timestamp + TICK_NSEC)
438 clock = rq->tick_timestamp + TICK_NSEC;
441 rq->clock_overflows++;
443 if (unlikely(delta > rq->clock_max_delta))
444 rq->clock_max_delta = delta;
449 rq->prev_clock_raw = now;
453 static void update_rq_clock(struct rq *rq)
455 if (likely(smp_processor_id() == cpu_of(rq)))
456 __update_rq_clock(rq);
460 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
461 * See detach_destroy_domains: synchronize_sched for details.
463 * The domain tree of any CPU may only be accessed from within
464 * preempt-disabled sections.
466 #define for_each_domain(cpu, __sd) \
467 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
469 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
470 #define this_rq() (&__get_cpu_var(runqueues))
471 #define task_rq(p) cpu_rq(task_cpu(p))
472 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
475 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
477 #ifdef CONFIG_SCHED_DEBUG
478 # define const_debug __read_mostly
480 # define const_debug static const
484 * Debugging: various feature bits
487 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
488 SCHED_FEAT_WAKEUP_PREEMPT = 2,
489 SCHED_FEAT_START_DEBIT = 4,
490 SCHED_FEAT_TREE_AVG = 8,
491 SCHED_FEAT_APPROX_AVG = 16,
494 const_debug unsigned int sysctl_sched_features =
495 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
496 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
497 SCHED_FEAT_START_DEBIT * 1 |
498 SCHED_FEAT_TREE_AVG * 0 |
499 SCHED_FEAT_APPROX_AVG * 0;
501 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
504 * Number of tasks to iterate in a single balance run.
505 * Limited because this is done with IRQs disabled.
507 const_debug unsigned int sysctl_sched_nr_migrate = 32;
510 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
511 * clock constructed from sched_clock():
513 unsigned long long cpu_clock(int cpu)
515 unsigned long long now;
519 local_irq_save(flags);
522 * Only call sched_clock() if the scheduler has already been
523 * initialized (some code might call cpu_clock() very early):
528 local_irq_restore(flags);
532 EXPORT_SYMBOL_GPL(cpu_clock);
534 #ifndef prepare_arch_switch
535 # define prepare_arch_switch(next) do { } while (0)
537 #ifndef finish_arch_switch
538 # define finish_arch_switch(prev) do { } while (0)
541 static inline int task_current(struct rq *rq, struct task_struct *p)
543 return rq->curr == p;
546 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
547 static inline int task_running(struct rq *rq, struct task_struct *p)
549 return task_current(rq, p);
552 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
556 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
558 #ifdef CONFIG_DEBUG_SPINLOCK
559 /* this is a valid case when another task releases the spinlock */
560 rq->lock.owner = current;
563 * If we are tracking spinlock dependencies then we have to
564 * fix up the runqueue lock - which gets 'carried over' from
567 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
569 spin_unlock_irq(&rq->lock);
572 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
573 static inline int task_running(struct rq *rq, struct task_struct *p)
578 return task_current(rq, p);
582 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
586 * We can optimise this out completely for !SMP, because the
587 * SMP rebalancing from interrupt is the only thing that cares
592 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
593 spin_unlock_irq(&rq->lock);
595 spin_unlock(&rq->lock);
599 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
603 * After ->oncpu is cleared, the task can be moved to a different CPU.
604 * We must ensure this doesn't happen until the switch is completely
610 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
614 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
617 * __task_rq_lock - lock the runqueue a given task resides on.
618 * Must be called interrupts disabled.
620 static inline struct rq *__task_rq_lock(struct task_struct *p)
624 struct rq *rq = task_rq(p);
625 spin_lock(&rq->lock);
626 if (likely(rq == task_rq(p)))
628 spin_unlock(&rq->lock);
633 * task_rq_lock - lock the runqueue a given task resides on and disable
634 * interrupts. Note the ordering: we can safely lookup the task_rq without
635 * explicitly disabling preemption.
637 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
643 local_irq_save(*flags);
645 spin_lock(&rq->lock);
646 if (likely(rq == task_rq(p)))
648 spin_unlock_irqrestore(&rq->lock, *flags);
652 static void __task_rq_unlock(struct rq *rq)
655 spin_unlock(&rq->lock);
658 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
661 spin_unlock_irqrestore(&rq->lock, *flags);
665 * this_rq_lock - lock this runqueue and disable interrupts.
667 static struct rq *this_rq_lock(void)
674 spin_lock(&rq->lock);
680 * We are going deep-idle (irqs are disabled):
682 void sched_clock_idle_sleep_event(void)
684 struct rq *rq = cpu_rq(smp_processor_id());
686 spin_lock(&rq->lock);
687 __update_rq_clock(rq);
688 spin_unlock(&rq->lock);
689 rq->clock_deep_idle_events++;
691 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
694 * We just idled delta nanoseconds (called with irqs disabled):
696 void sched_clock_idle_wakeup_event(u64 delta_ns)
698 struct rq *rq = cpu_rq(smp_processor_id());
699 u64 now = sched_clock();
701 touch_softlockup_watchdog();
702 rq->idle_clock += delta_ns;
704 * Override the previous timestamp and ignore all
705 * sched_clock() deltas that occured while we idled,
706 * and use the PM-provided delta_ns to advance the
709 spin_lock(&rq->lock);
710 rq->prev_clock_raw = now;
711 rq->clock += delta_ns;
712 spin_unlock(&rq->lock);
714 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
717 * resched_task - mark a task 'to be rescheduled now'.
719 * On UP this means the setting of the need_resched flag, on SMP it
720 * might also involve a cross-CPU call to trigger the scheduler on
725 #ifndef tsk_is_polling
726 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
729 static void resched_task(struct task_struct *p)
733 assert_spin_locked(&task_rq(p)->lock);
735 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
738 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
741 if (cpu == smp_processor_id())
744 /* NEED_RESCHED must be visible before we test polling */
746 if (!tsk_is_polling(p))
747 smp_send_reschedule(cpu);
750 static void resched_cpu(int cpu)
752 struct rq *rq = cpu_rq(cpu);
755 if (!spin_trylock_irqsave(&rq->lock, flags))
757 resched_task(cpu_curr(cpu));
758 spin_unlock_irqrestore(&rq->lock, flags);
761 static inline void resched_task(struct task_struct *p)
763 assert_spin_locked(&task_rq(p)->lock);
764 set_tsk_need_resched(p);
768 #if BITS_PER_LONG == 32
769 # define WMULT_CONST (~0UL)
771 # define WMULT_CONST (1UL << 32)
774 #define WMULT_SHIFT 32
777 * Shift right and round:
779 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
782 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
783 struct load_weight *lw)
787 if (unlikely(!lw->inv_weight))
788 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
790 tmp = (u64)delta_exec * weight;
792 * Check whether we'd overflow the 64-bit multiplication:
794 if (unlikely(tmp > WMULT_CONST))
795 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
798 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
800 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
803 static inline unsigned long
804 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
806 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
809 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
814 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
820 * To aid in avoiding the subversion of "niceness" due to uneven distribution
821 * of tasks with abnormal "nice" values across CPUs the contribution that
822 * each task makes to its run queue's load is weighted according to its
823 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
824 * scaled version of the new time slice allocation that they receive on time
828 #define WEIGHT_IDLEPRIO 2
829 #define WMULT_IDLEPRIO (1 << 31)
832 * Nice levels are multiplicative, with a gentle 10% change for every
833 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
834 * nice 1, it will get ~10% less CPU time than another CPU-bound task
835 * that remained on nice 0.
837 * The "10% effect" is relative and cumulative: from _any_ nice level,
838 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
839 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
840 * If a task goes up by ~10% and another task goes down by ~10% then
841 * the relative distance between them is ~25%.)
843 static const int prio_to_weight[40] = {
844 /* -20 */ 88761, 71755, 56483, 46273, 36291,
845 /* -15 */ 29154, 23254, 18705, 14949, 11916,
846 /* -10 */ 9548, 7620, 6100, 4904, 3906,
847 /* -5 */ 3121, 2501, 1991, 1586, 1277,
848 /* 0 */ 1024, 820, 655, 526, 423,
849 /* 5 */ 335, 272, 215, 172, 137,
850 /* 10 */ 110, 87, 70, 56, 45,
851 /* 15 */ 36, 29, 23, 18, 15,
855 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
857 * In cases where the weight does not change often, we can use the
858 * precalculated inverse to speed up arithmetics by turning divisions
859 * into multiplications:
861 static const u32 prio_to_wmult[40] = {
862 /* -20 */ 48388, 59856, 76040, 92818, 118348,
863 /* -15 */ 147320, 184698, 229616, 287308, 360437,
864 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
865 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
866 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
867 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
868 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
869 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
872 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
875 * runqueue iterator, to support SMP load-balancing between different
876 * scheduling classes, without having to expose their internal data
877 * structures to the load-balancing proper:
881 struct task_struct *(*start)(void *);
882 struct task_struct *(*next)(void *);
887 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
888 unsigned long max_load_move, struct sched_domain *sd,
889 enum cpu_idle_type idle, int *all_pinned,
890 int *this_best_prio, struct rq_iterator *iterator);
893 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
894 struct sched_domain *sd, enum cpu_idle_type idle,
895 struct rq_iterator *iterator);
898 #ifdef CONFIG_CGROUP_CPUACCT
899 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
901 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
904 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
906 update_load_add(&rq->load, load);
909 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
911 update_load_sub(&rq->load, load);
914 #include "sched_stats.h"
915 #include "sched_idletask.c"
916 #include "sched_fair.c"
917 #include "sched_rt.c"
918 #ifdef CONFIG_SCHED_DEBUG
919 # include "sched_debug.c"
922 #define sched_class_highest (&rt_sched_class)
924 static void inc_nr_running(struct task_struct *p, struct rq *rq)
929 static void dec_nr_running(struct task_struct *p, struct rq *rq)
934 static void set_load_weight(struct task_struct *p)
936 if (task_has_rt_policy(p)) {
937 p->se.load.weight = prio_to_weight[0] * 2;
938 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
943 * SCHED_IDLE tasks get minimal weight:
945 if (p->policy == SCHED_IDLE) {
946 p->se.load.weight = WEIGHT_IDLEPRIO;
947 p->se.load.inv_weight = WMULT_IDLEPRIO;
951 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
952 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
955 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
957 sched_info_queued(p);
958 p->sched_class->enqueue_task(rq, p, wakeup);
962 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
964 p->sched_class->dequeue_task(rq, p, sleep);
969 * __normal_prio - return the priority that is based on the static prio
971 static inline int __normal_prio(struct task_struct *p)
973 return p->static_prio;
977 * Calculate the expected normal priority: i.e. priority
978 * without taking RT-inheritance into account. Might be
979 * boosted by interactivity modifiers. Changes upon fork,
980 * setprio syscalls, and whenever the interactivity
981 * estimator recalculates.
983 static inline int normal_prio(struct task_struct *p)
987 if (task_has_rt_policy(p))
988 prio = MAX_RT_PRIO-1 - p->rt_priority;
990 prio = __normal_prio(p);
995 * Calculate the current priority, i.e. the priority
996 * taken into account by the scheduler. This value might
997 * be boosted by RT tasks, or might be boosted by
998 * interactivity modifiers. Will be RT if the task got
999 * RT-boosted. If not then it returns p->normal_prio.
1001 static int effective_prio(struct task_struct *p)
1003 p->normal_prio = normal_prio(p);
1005 * If we are RT tasks or we were boosted to RT priority,
1006 * keep the priority unchanged. Otherwise, update priority
1007 * to the normal priority:
1009 if (!rt_prio(p->prio))
1010 return p->normal_prio;
1015 * activate_task - move a task to the runqueue.
1017 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1019 if (p->state == TASK_UNINTERRUPTIBLE)
1020 rq->nr_uninterruptible--;
1022 enqueue_task(rq, p, wakeup);
1023 inc_nr_running(p, rq);
1027 * deactivate_task - remove a task from the runqueue.
1029 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1031 if (p->state == TASK_UNINTERRUPTIBLE)
1032 rq->nr_uninterruptible++;
1034 dequeue_task(rq, p, sleep);
1035 dec_nr_running(p, rq);
1039 * task_curr - is this task currently executing on a CPU?
1040 * @p: the task in question.
1042 inline int task_curr(const struct task_struct *p)
1044 return cpu_curr(task_cpu(p)) == p;
1047 /* Used instead of source_load when we know the type == 0 */
1048 unsigned long weighted_cpuload(const int cpu)
1050 return cpu_rq(cpu)->load.weight;
1053 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1055 set_task_cfs_rq(p, cpu);
1058 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1059 * successfuly executed on another CPU. We must ensure that updates of
1060 * per-task data have been completed by this moment.
1063 task_thread_info(p)->cpu = cpu;
1070 * Is this task likely cache-hot:
1073 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1077 if (p->sched_class != &fair_sched_class)
1080 if (sysctl_sched_migration_cost == -1)
1082 if (sysctl_sched_migration_cost == 0)
1085 delta = now - p->se.exec_start;
1087 return delta < (s64)sysctl_sched_migration_cost;
1091 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1093 int old_cpu = task_cpu(p);
1094 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1095 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1096 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1099 clock_offset = old_rq->clock - new_rq->clock;
1101 #ifdef CONFIG_SCHEDSTATS
1102 if (p->se.wait_start)
1103 p->se.wait_start -= clock_offset;
1104 if (p->se.sleep_start)
1105 p->se.sleep_start -= clock_offset;
1106 if (p->se.block_start)
1107 p->se.block_start -= clock_offset;
1108 if (old_cpu != new_cpu) {
1109 schedstat_inc(p, se.nr_migrations);
1110 if (task_hot(p, old_rq->clock, NULL))
1111 schedstat_inc(p, se.nr_forced2_migrations);
1114 p->se.vruntime -= old_cfsrq->min_vruntime -
1115 new_cfsrq->min_vruntime;
1117 __set_task_cpu(p, new_cpu);
1120 struct migration_req {
1121 struct list_head list;
1123 struct task_struct *task;
1126 struct completion done;
1130 * The task's runqueue lock must be held.
1131 * Returns true if you have to wait for migration thread.
1134 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1136 struct rq *rq = task_rq(p);
1139 * If the task is not on a runqueue (and not running), then
1140 * it is sufficient to simply update the task's cpu field.
1142 if (!p->se.on_rq && !task_running(rq, p)) {
1143 set_task_cpu(p, dest_cpu);
1147 init_completion(&req->done);
1149 req->dest_cpu = dest_cpu;
1150 list_add(&req->list, &rq->migration_queue);
1156 * wait_task_inactive - wait for a thread to unschedule.
1158 * The caller must ensure that the task *will* unschedule sometime soon,
1159 * else this function might spin for a *long* time. This function can't
1160 * be called with interrupts off, or it may introduce deadlock with
1161 * smp_call_function() if an IPI is sent by the same process we are
1162 * waiting to become inactive.
1164 void wait_task_inactive(struct task_struct *p)
1166 unsigned long flags;
1172 * We do the initial early heuristics without holding
1173 * any task-queue locks at all. We'll only try to get
1174 * the runqueue lock when things look like they will
1180 * If the task is actively running on another CPU
1181 * still, just relax and busy-wait without holding
1184 * NOTE! Since we don't hold any locks, it's not
1185 * even sure that "rq" stays as the right runqueue!
1186 * But we don't care, since "task_running()" will
1187 * return false if the runqueue has changed and p
1188 * is actually now running somewhere else!
1190 while (task_running(rq, p))
1194 * Ok, time to look more closely! We need the rq
1195 * lock now, to be *sure*. If we're wrong, we'll
1196 * just go back and repeat.
1198 rq = task_rq_lock(p, &flags);
1199 running = task_running(rq, p);
1200 on_rq = p->se.on_rq;
1201 task_rq_unlock(rq, &flags);
1204 * Was it really running after all now that we
1205 * checked with the proper locks actually held?
1207 * Oops. Go back and try again..
1209 if (unlikely(running)) {
1215 * It's not enough that it's not actively running,
1216 * it must be off the runqueue _entirely_, and not
1219 * So if it wa still runnable (but just not actively
1220 * running right now), it's preempted, and we should
1221 * yield - it could be a while.
1223 if (unlikely(on_rq)) {
1224 schedule_timeout_uninterruptible(1);
1229 * Ahh, all good. It wasn't running, and it wasn't
1230 * runnable, which means that it will never become
1231 * running in the future either. We're all done!
1238 * kick_process - kick a running thread to enter/exit the kernel
1239 * @p: the to-be-kicked thread
1241 * Cause a process which is running on another CPU to enter
1242 * kernel-mode, without any delay. (to get signals handled.)
1244 * NOTE: this function doesnt have to take the runqueue lock,
1245 * because all it wants to ensure is that the remote task enters
1246 * the kernel. If the IPI races and the task has been migrated
1247 * to another CPU then no harm is done and the purpose has been
1250 void kick_process(struct task_struct *p)
1256 if ((cpu != smp_processor_id()) && task_curr(p))
1257 smp_send_reschedule(cpu);
1262 * Return a low guess at the load of a migration-source cpu weighted
1263 * according to the scheduling class and "nice" value.
1265 * We want to under-estimate the load of migration sources, to
1266 * balance conservatively.
1268 static unsigned long source_load(int cpu, int type)
1270 struct rq *rq = cpu_rq(cpu);
1271 unsigned long total = weighted_cpuload(cpu);
1276 return min(rq->cpu_load[type-1], total);
1280 * Return a high guess at the load of a migration-target cpu weighted
1281 * according to the scheduling class and "nice" value.
1283 static unsigned long target_load(int cpu, int type)
1285 struct rq *rq = cpu_rq(cpu);
1286 unsigned long total = weighted_cpuload(cpu);
1291 return max(rq->cpu_load[type-1], total);
1295 * Return the average load per task on the cpu's run queue
1297 static inline unsigned long cpu_avg_load_per_task(int cpu)
1299 struct rq *rq = cpu_rq(cpu);
1300 unsigned long total = weighted_cpuload(cpu);
1301 unsigned long n = rq->nr_running;
1303 return n ? total / n : SCHED_LOAD_SCALE;
1307 * find_idlest_group finds and returns the least busy CPU group within the
1310 static struct sched_group *
1311 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1313 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1314 unsigned long min_load = ULONG_MAX, this_load = 0;
1315 int load_idx = sd->forkexec_idx;
1316 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1319 unsigned long load, avg_load;
1323 /* Skip over this group if it has no CPUs allowed */
1324 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1327 local_group = cpu_isset(this_cpu, group->cpumask);
1329 /* Tally up the load of all CPUs in the group */
1332 for_each_cpu_mask(i, group->cpumask) {
1333 /* Bias balancing toward cpus of our domain */
1335 load = source_load(i, load_idx);
1337 load = target_load(i, load_idx);
1342 /* Adjust by relative CPU power of the group */
1343 avg_load = sg_div_cpu_power(group,
1344 avg_load * SCHED_LOAD_SCALE);
1347 this_load = avg_load;
1349 } else if (avg_load < min_load) {
1350 min_load = avg_load;
1353 } while (group = group->next, group != sd->groups);
1355 if (!idlest || 100*this_load < imbalance*min_load)
1361 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1364 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1367 unsigned long load, min_load = ULONG_MAX;
1371 /* Traverse only the allowed CPUs */
1372 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1374 for_each_cpu_mask(i, tmp) {
1375 load = weighted_cpuload(i);
1377 if (load < min_load || (load == min_load && i == this_cpu)) {
1387 * sched_balance_self: balance the current task (running on cpu) in domains
1388 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1391 * Balance, ie. select the least loaded group.
1393 * Returns the target CPU number, or the same CPU if no balancing is needed.
1395 * preempt must be disabled.
1397 static int sched_balance_self(int cpu, int flag)
1399 struct task_struct *t = current;
1400 struct sched_domain *tmp, *sd = NULL;
1402 for_each_domain(cpu, tmp) {
1404 * If power savings logic is enabled for a domain, stop there.
1406 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1408 if (tmp->flags & flag)
1414 struct sched_group *group;
1415 int new_cpu, weight;
1417 if (!(sd->flags & flag)) {
1423 group = find_idlest_group(sd, t, cpu);
1429 new_cpu = find_idlest_cpu(group, t, cpu);
1430 if (new_cpu == -1 || new_cpu == cpu) {
1431 /* Now try balancing at a lower domain level of cpu */
1436 /* Now try balancing at a lower domain level of new_cpu */
1439 weight = cpus_weight(span);
1440 for_each_domain(cpu, tmp) {
1441 if (weight <= cpus_weight(tmp->span))
1443 if (tmp->flags & flag)
1446 /* while loop will break here if sd == NULL */
1452 #endif /* CONFIG_SMP */
1455 * wake_idle() will wake a task on an idle cpu if task->cpu is
1456 * not idle and an idle cpu is available. The span of cpus to
1457 * search starts with cpus closest then further out as needed,
1458 * so we always favor a closer, idle cpu.
1460 * Returns the CPU we should wake onto.
1462 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1463 static int wake_idle(int cpu, struct task_struct *p)
1466 struct sched_domain *sd;
1470 * If it is idle, then it is the best cpu to run this task.
1472 * This cpu is also the best, if it has more than one task already.
1473 * Siblings must be also busy(in most cases) as they didn't already
1474 * pickup the extra load from this cpu and hence we need not check
1475 * sibling runqueue info. This will avoid the checks and cache miss
1476 * penalities associated with that.
1478 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1481 for_each_domain(cpu, sd) {
1482 if (sd->flags & SD_WAKE_IDLE) {
1483 cpus_and(tmp, sd->span, p->cpus_allowed);
1484 for_each_cpu_mask(i, tmp) {
1486 if (i != task_cpu(p)) {
1488 se.nr_wakeups_idle);
1500 static inline int wake_idle(int cpu, struct task_struct *p)
1507 * try_to_wake_up - wake up a thread
1508 * @p: the to-be-woken-up thread
1509 * @state: the mask of task states that can be woken
1510 * @sync: do a synchronous wakeup?
1512 * Put it on the run-queue if it's not already there. The "current"
1513 * thread is always on the run-queue (except when the actual
1514 * re-schedule is in progress), and as such you're allowed to do
1515 * the simpler "current->state = TASK_RUNNING" to mark yourself
1516 * runnable without the overhead of this.
1518 * returns failure only if the task is already active.
1520 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1522 int cpu, orig_cpu, this_cpu, success = 0;
1523 unsigned long flags;
1527 struct sched_domain *sd, *this_sd = NULL;
1528 unsigned long load, this_load;
1532 rq = task_rq_lock(p, &flags);
1533 old_state = p->state;
1534 if (!(old_state & state))
1542 this_cpu = smp_processor_id();
1545 if (unlikely(task_running(rq, p)))
1550 schedstat_inc(rq, ttwu_count);
1551 if (cpu == this_cpu) {
1552 schedstat_inc(rq, ttwu_local);
1556 for_each_domain(this_cpu, sd) {
1557 if (cpu_isset(cpu, sd->span)) {
1558 schedstat_inc(sd, ttwu_wake_remote);
1564 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1568 * Check for affine wakeup and passive balancing possibilities.
1571 int idx = this_sd->wake_idx;
1572 unsigned int imbalance;
1574 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1576 load = source_load(cpu, idx);
1577 this_load = target_load(this_cpu, idx);
1579 new_cpu = this_cpu; /* Wake to this CPU if we can */
1581 if (this_sd->flags & SD_WAKE_AFFINE) {
1582 unsigned long tl = this_load;
1583 unsigned long tl_per_task;
1586 * Attract cache-cold tasks on sync wakeups:
1588 if (sync && !task_hot(p, rq->clock, this_sd))
1591 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1592 tl_per_task = cpu_avg_load_per_task(this_cpu);
1595 * If sync wakeup then subtract the (maximum possible)
1596 * effect of the currently running task from the load
1597 * of the current CPU:
1600 tl -= current->se.load.weight;
1603 tl + target_load(cpu, idx) <= tl_per_task) ||
1604 100*(tl + p->se.load.weight) <= imbalance*load) {
1606 * This domain has SD_WAKE_AFFINE and
1607 * p is cache cold in this domain, and
1608 * there is no bad imbalance.
1610 schedstat_inc(this_sd, ttwu_move_affine);
1611 schedstat_inc(p, se.nr_wakeups_affine);
1617 * Start passive balancing when half the imbalance_pct
1620 if (this_sd->flags & SD_WAKE_BALANCE) {
1621 if (imbalance*this_load <= 100*load) {
1622 schedstat_inc(this_sd, ttwu_move_balance);
1623 schedstat_inc(p, se.nr_wakeups_passive);
1629 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1631 new_cpu = wake_idle(new_cpu, p);
1632 if (new_cpu != cpu) {
1633 set_task_cpu(p, new_cpu);
1634 task_rq_unlock(rq, &flags);
1635 /* might preempt at this point */
1636 rq = task_rq_lock(p, &flags);
1637 old_state = p->state;
1638 if (!(old_state & state))
1643 this_cpu = smp_processor_id();
1648 #endif /* CONFIG_SMP */
1649 schedstat_inc(p, se.nr_wakeups);
1651 schedstat_inc(p, se.nr_wakeups_sync);
1652 if (orig_cpu != cpu)
1653 schedstat_inc(p, se.nr_wakeups_migrate);
1654 if (cpu == this_cpu)
1655 schedstat_inc(p, se.nr_wakeups_local);
1657 schedstat_inc(p, se.nr_wakeups_remote);
1658 update_rq_clock(rq);
1659 activate_task(rq, p, 1);
1660 check_preempt_curr(rq, p);
1664 p->state = TASK_RUNNING;
1666 task_rq_unlock(rq, &flags);
1671 int fastcall wake_up_process(struct task_struct *p)
1673 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1674 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1676 EXPORT_SYMBOL(wake_up_process);
1678 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1680 return try_to_wake_up(p, state, 0);
1684 * Perform scheduler related setup for a newly forked process p.
1685 * p is forked by current.
1687 * __sched_fork() is basic setup used by init_idle() too:
1689 static void __sched_fork(struct task_struct *p)
1691 p->se.exec_start = 0;
1692 p->se.sum_exec_runtime = 0;
1693 p->se.prev_sum_exec_runtime = 0;
1695 #ifdef CONFIG_SCHEDSTATS
1696 p->se.wait_start = 0;
1697 p->se.sum_sleep_runtime = 0;
1698 p->se.sleep_start = 0;
1699 p->se.block_start = 0;
1700 p->se.sleep_max = 0;
1701 p->se.block_max = 0;
1703 p->se.slice_max = 0;
1707 INIT_LIST_HEAD(&p->run_list);
1710 #ifdef CONFIG_PREEMPT_NOTIFIERS
1711 INIT_HLIST_HEAD(&p->preempt_notifiers);
1715 * We mark the process as running here, but have not actually
1716 * inserted it onto the runqueue yet. This guarantees that
1717 * nobody will actually run it, and a signal or other external
1718 * event cannot wake it up and insert it on the runqueue either.
1720 p->state = TASK_RUNNING;
1724 * fork()/clone()-time setup:
1726 void sched_fork(struct task_struct *p, int clone_flags)
1728 int cpu = get_cpu();
1733 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1735 set_task_cpu(p, cpu);
1738 * Make sure we do not leak PI boosting priority to the child:
1740 p->prio = current->normal_prio;
1741 if (!rt_prio(p->prio))
1742 p->sched_class = &fair_sched_class;
1744 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1745 if (likely(sched_info_on()))
1746 memset(&p->sched_info, 0, sizeof(p->sched_info));
1748 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1751 #ifdef CONFIG_PREEMPT
1752 /* Want to start with kernel preemption disabled. */
1753 task_thread_info(p)->preempt_count = 1;
1759 * wake_up_new_task - wake up a newly created task for the first time.
1761 * This function will do some initial scheduler statistics housekeeping
1762 * that must be done for every newly created context, then puts the task
1763 * on the runqueue and wakes it.
1765 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1767 unsigned long flags;
1770 rq = task_rq_lock(p, &flags);
1771 BUG_ON(p->state != TASK_RUNNING);
1772 update_rq_clock(rq);
1774 p->prio = effective_prio(p);
1776 if (!p->sched_class->task_new || !current->se.on_rq) {
1777 activate_task(rq, p, 0);
1780 * Let the scheduling class do new task startup
1781 * management (if any):
1783 p->sched_class->task_new(rq, p);
1784 inc_nr_running(p, rq);
1786 check_preempt_curr(rq, p);
1787 task_rq_unlock(rq, &flags);
1790 #ifdef CONFIG_PREEMPT_NOTIFIERS
1793 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1794 * @notifier: notifier struct to register
1796 void preempt_notifier_register(struct preempt_notifier *notifier)
1798 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1800 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1803 * preempt_notifier_unregister - no longer interested in preemption notifications
1804 * @notifier: notifier struct to unregister
1806 * This is safe to call from within a preemption notifier.
1808 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1810 hlist_del(¬ifier->link);
1812 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1814 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1816 struct preempt_notifier *notifier;
1817 struct hlist_node *node;
1819 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1820 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1824 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1825 struct task_struct *next)
1827 struct preempt_notifier *notifier;
1828 struct hlist_node *node;
1830 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1831 notifier->ops->sched_out(notifier, next);
1836 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1841 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1842 struct task_struct *next)
1849 * prepare_task_switch - prepare to switch tasks
1850 * @rq: the runqueue preparing to switch
1851 * @prev: the current task that is being switched out
1852 * @next: the task we are going to switch to.
1854 * This is called with the rq lock held and interrupts off. It must
1855 * be paired with a subsequent finish_task_switch after the context
1858 * prepare_task_switch sets up locking and calls architecture specific
1862 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1863 struct task_struct *next)
1865 fire_sched_out_preempt_notifiers(prev, next);
1866 prepare_lock_switch(rq, next);
1867 prepare_arch_switch(next);
1871 * finish_task_switch - clean up after a task-switch
1872 * @rq: runqueue associated with task-switch
1873 * @prev: the thread we just switched away from.
1875 * finish_task_switch must be called after the context switch, paired
1876 * with a prepare_task_switch call before the context switch.
1877 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1878 * and do any other architecture-specific cleanup actions.
1880 * Note that we may have delayed dropping an mm in context_switch(). If
1881 * so, we finish that here outside of the runqueue lock. (Doing it
1882 * with the lock held can cause deadlocks; see schedule() for
1885 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1886 __releases(rq->lock)
1888 struct mm_struct *mm = rq->prev_mm;
1894 * A task struct has one reference for the use as "current".
1895 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1896 * schedule one last time. The schedule call will never return, and
1897 * the scheduled task must drop that reference.
1898 * The test for TASK_DEAD must occur while the runqueue locks are
1899 * still held, otherwise prev could be scheduled on another cpu, die
1900 * there before we look at prev->state, and then the reference would
1902 * Manfred Spraul <manfred@colorfullife.com>
1904 prev_state = prev->state;
1905 finish_arch_switch(prev);
1906 finish_lock_switch(rq, prev);
1907 fire_sched_in_preempt_notifiers(current);
1910 if (unlikely(prev_state == TASK_DEAD)) {
1912 * Remove function-return probe instances associated with this
1913 * task and put them back on the free list.
1915 kprobe_flush_task(prev);
1916 put_task_struct(prev);
1921 * schedule_tail - first thing a freshly forked thread must call.
1922 * @prev: the thread we just switched away from.
1924 asmlinkage void schedule_tail(struct task_struct *prev)
1925 __releases(rq->lock)
1927 struct rq *rq = this_rq();
1929 finish_task_switch(rq, prev);
1930 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1931 /* In this case, finish_task_switch does not reenable preemption */
1934 if (current->set_child_tid)
1935 put_user(task_pid_vnr(current), current->set_child_tid);
1939 * context_switch - switch to the new MM and the new
1940 * thread's register state.
1943 context_switch(struct rq *rq, struct task_struct *prev,
1944 struct task_struct *next)
1946 struct mm_struct *mm, *oldmm;
1948 prepare_task_switch(rq, prev, next);
1950 oldmm = prev->active_mm;
1952 * For paravirt, this is coupled with an exit in switch_to to
1953 * combine the page table reload and the switch backend into
1956 arch_enter_lazy_cpu_mode();
1958 if (unlikely(!mm)) {
1959 next->active_mm = oldmm;
1960 atomic_inc(&oldmm->mm_count);
1961 enter_lazy_tlb(oldmm, next);
1963 switch_mm(oldmm, mm, next);
1965 if (unlikely(!prev->mm)) {
1966 prev->active_mm = NULL;
1967 rq->prev_mm = oldmm;
1970 * Since the runqueue lock will be released by the next
1971 * task (which is an invalid locking op but in the case
1972 * of the scheduler it's an obvious special-case), so we
1973 * do an early lockdep release here:
1975 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1976 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1979 /* Here we just switch the register state and the stack. */
1980 switch_to(prev, next, prev);
1984 * this_rq must be evaluated again because prev may have moved
1985 * CPUs since it called schedule(), thus the 'rq' on its stack
1986 * frame will be invalid.
1988 finish_task_switch(this_rq(), prev);
1992 * nr_running, nr_uninterruptible and nr_context_switches:
1994 * externally visible scheduler statistics: current number of runnable
1995 * threads, current number of uninterruptible-sleeping threads, total
1996 * number of context switches performed since bootup.
1998 unsigned long nr_running(void)
2000 unsigned long i, sum = 0;
2002 for_each_online_cpu(i)
2003 sum += cpu_rq(i)->nr_running;
2008 unsigned long nr_uninterruptible(void)
2010 unsigned long i, sum = 0;
2012 for_each_possible_cpu(i)
2013 sum += cpu_rq(i)->nr_uninterruptible;
2016 * Since we read the counters lockless, it might be slightly
2017 * inaccurate. Do not allow it to go below zero though:
2019 if (unlikely((long)sum < 0))
2025 unsigned long long nr_context_switches(void)
2028 unsigned long long sum = 0;
2030 for_each_possible_cpu(i)
2031 sum += cpu_rq(i)->nr_switches;
2036 unsigned long nr_iowait(void)
2038 unsigned long i, sum = 0;
2040 for_each_possible_cpu(i)
2041 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2046 unsigned long nr_active(void)
2048 unsigned long i, running = 0, uninterruptible = 0;
2050 for_each_online_cpu(i) {
2051 running += cpu_rq(i)->nr_running;
2052 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2055 if (unlikely((long)uninterruptible < 0))
2056 uninterruptible = 0;
2058 return running + uninterruptible;
2062 * Update rq->cpu_load[] statistics. This function is usually called every
2063 * scheduler tick (TICK_NSEC).
2065 static void update_cpu_load(struct rq *this_rq)
2067 unsigned long this_load = this_rq->load.weight;
2070 this_rq->nr_load_updates++;
2072 /* Update our load: */
2073 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2074 unsigned long old_load, new_load;
2076 /* scale is effectively 1 << i now, and >> i divides by scale */
2078 old_load = this_rq->cpu_load[i];
2079 new_load = this_load;
2081 * Round up the averaging division if load is increasing. This
2082 * prevents us from getting stuck on 9 if the load is 10, for
2085 if (new_load > old_load)
2086 new_load += scale-1;
2087 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2094 * double_rq_lock - safely lock two runqueues
2096 * Note this does not disable interrupts like task_rq_lock,
2097 * you need to do so manually before calling.
2099 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2100 __acquires(rq1->lock)
2101 __acquires(rq2->lock)
2103 BUG_ON(!irqs_disabled());
2105 spin_lock(&rq1->lock);
2106 __acquire(rq2->lock); /* Fake it out ;) */
2109 spin_lock(&rq1->lock);
2110 spin_lock(&rq2->lock);
2112 spin_lock(&rq2->lock);
2113 spin_lock(&rq1->lock);
2116 update_rq_clock(rq1);
2117 update_rq_clock(rq2);
2121 * double_rq_unlock - safely unlock two runqueues
2123 * Note this does not restore interrupts like task_rq_unlock,
2124 * you need to do so manually after calling.
2126 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2127 __releases(rq1->lock)
2128 __releases(rq2->lock)
2130 spin_unlock(&rq1->lock);
2132 spin_unlock(&rq2->lock);
2134 __release(rq2->lock);
2138 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2140 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2141 __releases(this_rq->lock)
2142 __acquires(busiest->lock)
2143 __acquires(this_rq->lock)
2145 if (unlikely(!irqs_disabled())) {
2146 /* printk() doesn't work good under rq->lock */
2147 spin_unlock(&this_rq->lock);
2150 if (unlikely(!spin_trylock(&busiest->lock))) {
2151 if (busiest < this_rq) {
2152 spin_unlock(&this_rq->lock);
2153 spin_lock(&busiest->lock);
2154 spin_lock(&this_rq->lock);
2156 spin_lock(&busiest->lock);
2161 * If dest_cpu is allowed for this process, migrate the task to it.
2162 * This is accomplished by forcing the cpu_allowed mask to only
2163 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2164 * the cpu_allowed mask is restored.
2166 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2168 struct migration_req req;
2169 unsigned long flags;
2172 rq = task_rq_lock(p, &flags);
2173 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2174 || unlikely(cpu_is_offline(dest_cpu)))
2177 /* force the process onto the specified CPU */
2178 if (migrate_task(p, dest_cpu, &req)) {
2179 /* Need to wait for migration thread (might exit: take ref). */
2180 struct task_struct *mt = rq->migration_thread;
2182 get_task_struct(mt);
2183 task_rq_unlock(rq, &flags);
2184 wake_up_process(mt);
2185 put_task_struct(mt);
2186 wait_for_completion(&req.done);
2191 task_rq_unlock(rq, &flags);
2195 * sched_exec - execve() is a valuable balancing opportunity, because at
2196 * this point the task has the smallest effective memory and cache footprint.
2198 void sched_exec(void)
2200 int new_cpu, this_cpu = get_cpu();
2201 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2203 if (new_cpu != this_cpu)
2204 sched_migrate_task(current, new_cpu);
2208 * pull_task - move a task from a remote runqueue to the local runqueue.
2209 * Both runqueues must be locked.
2211 static void pull_task(struct rq *src_rq, struct task_struct *p,
2212 struct rq *this_rq, int this_cpu)
2214 deactivate_task(src_rq, p, 0);
2215 set_task_cpu(p, this_cpu);
2216 activate_task(this_rq, p, 0);
2218 * Note that idle threads have a prio of MAX_PRIO, for this test
2219 * to be always true for them.
2221 check_preempt_curr(this_rq, p);
2225 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2228 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2229 struct sched_domain *sd, enum cpu_idle_type idle,
2233 * We do not migrate tasks that are:
2234 * 1) running (obviously), or
2235 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2236 * 3) are cache-hot on their current CPU.
2238 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2239 schedstat_inc(p, se.nr_failed_migrations_affine);
2244 if (task_running(rq, p)) {
2245 schedstat_inc(p, se.nr_failed_migrations_running);
2250 * Aggressive migration if:
2251 * 1) task is cache cold, or
2252 * 2) too many balance attempts have failed.
2255 if (!task_hot(p, rq->clock, sd) ||
2256 sd->nr_balance_failed > sd->cache_nice_tries) {
2257 #ifdef CONFIG_SCHEDSTATS
2258 if (task_hot(p, rq->clock, sd)) {
2259 schedstat_inc(sd, lb_hot_gained[idle]);
2260 schedstat_inc(p, se.nr_forced_migrations);
2266 if (task_hot(p, rq->clock, sd)) {
2267 schedstat_inc(p, se.nr_failed_migrations_hot);
2273 static unsigned long
2274 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2275 unsigned long max_load_move, struct sched_domain *sd,
2276 enum cpu_idle_type idle, int *all_pinned,
2277 int *this_best_prio, struct rq_iterator *iterator)
2279 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2280 struct task_struct *p;
2281 long rem_load_move = max_load_move;
2283 if (max_load_move == 0)
2289 * Start the load-balancing iterator:
2291 p = iterator->start(iterator->arg);
2293 if (!p || loops++ > sysctl_sched_nr_migrate)
2296 * To help distribute high priority tasks across CPUs we don't
2297 * skip a task if it will be the highest priority task (i.e. smallest
2298 * prio value) on its new queue regardless of its load weight
2300 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2301 SCHED_LOAD_SCALE_FUZZ;
2302 if ((skip_for_load && p->prio >= *this_best_prio) ||
2303 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2304 p = iterator->next(iterator->arg);
2308 pull_task(busiest, p, this_rq, this_cpu);
2310 rem_load_move -= p->se.load.weight;
2313 * We only want to steal up to the prescribed amount of weighted load.
2315 if (rem_load_move > 0) {
2316 if (p->prio < *this_best_prio)
2317 *this_best_prio = p->prio;
2318 p = iterator->next(iterator->arg);
2323 * Right now, this is one of only two places pull_task() is called,
2324 * so we can safely collect pull_task() stats here rather than
2325 * inside pull_task().
2327 schedstat_add(sd, lb_gained[idle], pulled);
2330 *all_pinned = pinned;
2332 return max_load_move - rem_load_move;
2336 * move_tasks tries to move up to max_load_move weighted load from busiest to
2337 * this_rq, as part of a balancing operation within domain "sd".
2338 * Returns 1 if successful and 0 otherwise.
2340 * Called with both runqueues locked.
2342 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2343 unsigned long max_load_move,
2344 struct sched_domain *sd, enum cpu_idle_type idle,
2347 const struct sched_class *class = sched_class_highest;
2348 unsigned long total_load_moved = 0;
2349 int this_best_prio = this_rq->curr->prio;
2353 class->load_balance(this_rq, this_cpu, busiest,
2354 max_load_move - total_load_moved,
2355 sd, idle, all_pinned, &this_best_prio);
2356 class = class->next;
2357 } while (class && max_load_move > total_load_moved);
2359 return total_load_moved > 0;
2363 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2364 struct sched_domain *sd, enum cpu_idle_type idle,
2365 struct rq_iterator *iterator)
2367 struct task_struct *p = iterator->start(iterator->arg);
2371 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2372 pull_task(busiest, p, this_rq, this_cpu);
2374 * Right now, this is only the second place pull_task()
2375 * is called, so we can safely collect pull_task()
2376 * stats here rather than inside pull_task().
2378 schedstat_inc(sd, lb_gained[idle]);
2382 p = iterator->next(iterator->arg);
2389 * move_one_task tries to move exactly one task from busiest to this_rq, as
2390 * part of active balancing operations within "domain".
2391 * Returns 1 if successful and 0 otherwise.
2393 * Called with both runqueues locked.
2395 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2396 struct sched_domain *sd, enum cpu_idle_type idle)
2398 const struct sched_class *class;
2400 for (class = sched_class_highest; class; class = class->next)
2401 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2408 * find_busiest_group finds and returns the busiest CPU group within the
2409 * domain. It calculates and returns the amount of weighted load which
2410 * should be moved to restore balance via the imbalance parameter.
2412 static struct sched_group *
2413 find_busiest_group(struct sched_domain *sd, int this_cpu,
2414 unsigned long *imbalance, enum cpu_idle_type idle,
2415 int *sd_idle, cpumask_t *cpus, int *balance)
2417 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2418 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2419 unsigned long max_pull;
2420 unsigned long busiest_load_per_task, busiest_nr_running;
2421 unsigned long this_load_per_task, this_nr_running;
2422 int load_idx, group_imb = 0;
2423 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2424 int power_savings_balance = 1;
2425 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2426 unsigned long min_nr_running = ULONG_MAX;
2427 struct sched_group *group_min = NULL, *group_leader = NULL;
2430 max_load = this_load = total_load = total_pwr = 0;
2431 busiest_load_per_task = busiest_nr_running = 0;
2432 this_load_per_task = this_nr_running = 0;
2433 if (idle == CPU_NOT_IDLE)
2434 load_idx = sd->busy_idx;
2435 else if (idle == CPU_NEWLY_IDLE)
2436 load_idx = sd->newidle_idx;
2438 load_idx = sd->idle_idx;
2441 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2444 int __group_imb = 0;
2445 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2446 unsigned long sum_nr_running, sum_weighted_load;
2448 local_group = cpu_isset(this_cpu, group->cpumask);
2451 balance_cpu = first_cpu(group->cpumask);
2453 /* Tally up the load of all CPUs in the group */
2454 sum_weighted_load = sum_nr_running = avg_load = 0;
2456 min_cpu_load = ~0UL;
2458 for_each_cpu_mask(i, group->cpumask) {
2461 if (!cpu_isset(i, *cpus))
2466 if (*sd_idle && rq->nr_running)
2469 /* Bias balancing toward cpus of our domain */
2471 if (idle_cpu(i) && !first_idle_cpu) {
2476 load = target_load(i, load_idx);
2478 load = source_load(i, load_idx);
2479 if (load > max_cpu_load)
2480 max_cpu_load = load;
2481 if (min_cpu_load > load)
2482 min_cpu_load = load;
2486 sum_nr_running += rq->nr_running;
2487 sum_weighted_load += weighted_cpuload(i);
2491 * First idle cpu or the first cpu(busiest) in this sched group
2492 * is eligible for doing load balancing at this and above
2493 * domains. In the newly idle case, we will allow all the cpu's
2494 * to do the newly idle load balance.
2496 if (idle != CPU_NEWLY_IDLE && local_group &&
2497 balance_cpu != this_cpu && balance) {
2502 total_load += avg_load;
2503 total_pwr += group->__cpu_power;
2505 /* Adjust by relative CPU power of the group */
2506 avg_load = sg_div_cpu_power(group,
2507 avg_load * SCHED_LOAD_SCALE);
2509 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2512 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2515 this_load = avg_load;
2517 this_nr_running = sum_nr_running;
2518 this_load_per_task = sum_weighted_load;
2519 } else if (avg_load > max_load &&
2520 (sum_nr_running > group_capacity || __group_imb)) {
2521 max_load = avg_load;
2523 busiest_nr_running = sum_nr_running;
2524 busiest_load_per_task = sum_weighted_load;
2525 group_imb = __group_imb;
2528 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2530 * Busy processors will not participate in power savings
2533 if (idle == CPU_NOT_IDLE ||
2534 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2538 * If the local group is idle or completely loaded
2539 * no need to do power savings balance at this domain
2541 if (local_group && (this_nr_running >= group_capacity ||
2543 power_savings_balance = 0;
2546 * If a group is already running at full capacity or idle,
2547 * don't include that group in power savings calculations
2549 if (!power_savings_balance || sum_nr_running >= group_capacity
2554 * Calculate the group which has the least non-idle load.
2555 * This is the group from where we need to pick up the load
2558 if ((sum_nr_running < min_nr_running) ||
2559 (sum_nr_running == min_nr_running &&
2560 first_cpu(group->cpumask) <
2561 first_cpu(group_min->cpumask))) {
2563 min_nr_running = sum_nr_running;
2564 min_load_per_task = sum_weighted_load /
2569 * Calculate the group which is almost near its
2570 * capacity but still has some space to pick up some load
2571 * from other group and save more power
2573 if (sum_nr_running <= group_capacity - 1) {
2574 if (sum_nr_running > leader_nr_running ||
2575 (sum_nr_running == leader_nr_running &&
2576 first_cpu(group->cpumask) >
2577 first_cpu(group_leader->cpumask))) {
2578 group_leader = group;
2579 leader_nr_running = sum_nr_running;
2584 group = group->next;
2585 } while (group != sd->groups);
2587 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2590 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2592 if (this_load >= avg_load ||
2593 100*max_load <= sd->imbalance_pct*this_load)
2596 busiest_load_per_task /= busiest_nr_running;
2598 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2601 * We're trying to get all the cpus to the average_load, so we don't
2602 * want to push ourselves above the average load, nor do we wish to
2603 * reduce the max loaded cpu below the average load, as either of these
2604 * actions would just result in more rebalancing later, and ping-pong
2605 * tasks around. Thus we look for the minimum possible imbalance.
2606 * Negative imbalances (*we* are more loaded than anyone else) will
2607 * be counted as no imbalance for these purposes -- we can't fix that
2608 * by pulling tasks to us. Be careful of negative numbers as they'll
2609 * appear as very large values with unsigned longs.
2611 if (max_load <= busiest_load_per_task)
2615 * In the presence of smp nice balancing, certain scenarios can have
2616 * max load less than avg load(as we skip the groups at or below
2617 * its cpu_power, while calculating max_load..)
2619 if (max_load < avg_load) {
2621 goto small_imbalance;
2624 /* Don't want to pull so many tasks that a group would go idle */
2625 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2627 /* How much load to actually move to equalise the imbalance */
2628 *imbalance = min(max_pull * busiest->__cpu_power,
2629 (avg_load - this_load) * this->__cpu_power)
2633 * if *imbalance is less than the average load per runnable task
2634 * there is no gaurantee that any tasks will be moved so we'll have
2635 * a think about bumping its value to force at least one task to be
2638 if (*imbalance < busiest_load_per_task) {
2639 unsigned long tmp, pwr_now, pwr_move;
2643 pwr_move = pwr_now = 0;
2645 if (this_nr_running) {
2646 this_load_per_task /= this_nr_running;
2647 if (busiest_load_per_task > this_load_per_task)
2650 this_load_per_task = SCHED_LOAD_SCALE;
2652 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2653 busiest_load_per_task * imbn) {
2654 *imbalance = busiest_load_per_task;
2659 * OK, we don't have enough imbalance to justify moving tasks,
2660 * however we may be able to increase total CPU power used by
2664 pwr_now += busiest->__cpu_power *
2665 min(busiest_load_per_task, max_load);
2666 pwr_now += this->__cpu_power *
2667 min(this_load_per_task, this_load);
2668 pwr_now /= SCHED_LOAD_SCALE;
2670 /* Amount of load we'd subtract */
2671 tmp = sg_div_cpu_power(busiest,
2672 busiest_load_per_task * SCHED_LOAD_SCALE);
2674 pwr_move += busiest->__cpu_power *
2675 min(busiest_load_per_task, max_load - tmp);
2677 /* Amount of load we'd add */
2678 if (max_load * busiest->__cpu_power <
2679 busiest_load_per_task * SCHED_LOAD_SCALE)
2680 tmp = sg_div_cpu_power(this,
2681 max_load * busiest->__cpu_power);
2683 tmp = sg_div_cpu_power(this,
2684 busiest_load_per_task * SCHED_LOAD_SCALE);
2685 pwr_move += this->__cpu_power *
2686 min(this_load_per_task, this_load + tmp);
2687 pwr_move /= SCHED_LOAD_SCALE;
2689 /* Move if we gain throughput */
2690 if (pwr_move > pwr_now)
2691 *imbalance = busiest_load_per_task;
2697 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2698 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2701 if (this == group_leader && group_leader != group_min) {
2702 *imbalance = min_load_per_task;
2712 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2715 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2716 unsigned long imbalance, cpumask_t *cpus)
2718 struct rq *busiest = NULL, *rq;
2719 unsigned long max_load = 0;
2722 for_each_cpu_mask(i, group->cpumask) {
2725 if (!cpu_isset(i, *cpus))
2729 wl = weighted_cpuload(i);
2731 if (rq->nr_running == 1 && wl > imbalance)
2734 if (wl > max_load) {
2744 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2745 * so long as it is large enough.
2747 #define MAX_PINNED_INTERVAL 512
2750 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2751 * tasks if there is an imbalance.
2753 static int load_balance(int this_cpu, struct rq *this_rq,
2754 struct sched_domain *sd, enum cpu_idle_type idle,
2757 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2758 struct sched_group *group;
2759 unsigned long imbalance;
2761 cpumask_t cpus = CPU_MASK_ALL;
2762 unsigned long flags;
2765 * When power savings policy is enabled for the parent domain, idle
2766 * sibling can pick up load irrespective of busy siblings. In this case,
2767 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2768 * portraying it as CPU_NOT_IDLE.
2770 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2771 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2774 schedstat_inc(sd, lb_count[idle]);
2777 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2784 schedstat_inc(sd, lb_nobusyg[idle]);
2788 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2790 schedstat_inc(sd, lb_nobusyq[idle]);
2794 BUG_ON(busiest == this_rq);
2796 schedstat_add(sd, lb_imbalance[idle], imbalance);
2799 if (busiest->nr_running > 1) {
2801 * Attempt to move tasks. If find_busiest_group has found
2802 * an imbalance but busiest->nr_running <= 1, the group is
2803 * still unbalanced. ld_moved simply stays zero, so it is
2804 * correctly treated as an imbalance.
2806 local_irq_save(flags);
2807 double_rq_lock(this_rq, busiest);
2808 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2809 imbalance, sd, idle, &all_pinned);
2810 double_rq_unlock(this_rq, busiest);
2811 local_irq_restore(flags);
2814 * some other cpu did the load balance for us.
2816 if (ld_moved && this_cpu != smp_processor_id())
2817 resched_cpu(this_cpu);
2819 /* All tasks on this runqueue were pinned by CPU affinity */
2820 if (unlikely(all_pinned)) {
2821 cpu_clear(cpu_of(busiest), cpus);
2822 if (!cpus_empty(cpus))
2829 schedstat_inc(sd, lb_failed[idle]);
2830 sd->nr_balance_failed++;
2832 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2834 spin_lock_irqsave(&busiest->lock, flags);
2836 /* don't kick the migration_thread, if the curr
2837 * task on busiest cpu can't be moved to this_cpu
2839 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2840 spin_unlock_irqrestore(&busiest->lock, flags);
2842 goto out_one_pinned;
2845 if (!busiest->active_balance) {
2846 busiest->active_balance = 1;
2847 busiest->push_cpu = this_cpu;
2850 spin_unlock_irqrestore(&busiest->lock, flags);
2852 wake_up_process(busiest->migration_thread);
2855 * We've kicked active balancing, reset the failure
2858 sd->nr_balance_failed = sd->cache_nice_tries+1;
2861 sd->nr_balance_failed = 0;
2863 if (likely(!active_balance)) {
2864 /* We were unbalanced, so reset the balancing interval */
2865 sd->balance_interval = sd->min_interval;
2868 * If we've begun active balancing, start to back off. This
2869 * case may not be covered by the all_pinned logic if there
2870 * is only 1 task on the busy runqueue (because we don't call
2873 if (sd->balance_interval < sd->max_interval)
2874 sd->balance_interval *= 2;
2877 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2878 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2883 schedstat_inc(sd, lb_balanced[idle]);
2885 sd->nr_balance_failed = 0;
2888 /* tune up the balancing interval */
2889 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2890 (sd->balance_interval < sd->max_interval))
2891 sd->balance_interval *= 2;
2893 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2894 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2900 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2901 * tasks if there is an imbalance.
2903 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2904 * this_rq is locked.
2907 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2909 struct sched_group *group;
2910 struct rq *busiest = NULL;
2911 unsigned long imbalance;
2915 cpumask_t cpus = CPU_MASK_ALL;
2918 * When power savings policy is enabled for the parent domain, idle
2919 * sibling can pick up load irrespective of busy siblings. In this case,
2920 * let the state of idle sibling percolate up as IDLE, instead of
2921 * portraying it as CPU_NOT_IDLE.
2923 if (sd->flags & SD_SHARE_CPUPOWER &&
2924 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2927 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2929 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2930 &sd_idle, &cpus, NULL);
2932 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2936 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2939 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2943 BUG_ON(busiest == this_rq);
2945 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2948 if (busiest->nr_running > 1) {
2949 /* Attempt to move tasks */
2950 double_lock_balance(this_rq, busiest);
2951 /* this_rq->clock is already updated */
2952 update_rq_clock(busiest);
2953 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2954 imbalance, sd, CPU_NEWLY_IDLE,
2956 spin_unlock(&busiest->lock);
2958 if (unlikely(all_pinned)) {
2959 cpu_clear(cpu_of(busiest), cpus);
2960 if (!cpus_empty(cpus))
2966 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2967 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2968 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2971 sd->nr_balance_failed = 0;
2976 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2977 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2978 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2980 sd->nr_balance_failed = 0;
2986 * idle_balance is called by schedule() if this_cpu is about to become
2987 * idle. Attempts to pull tasks from other CPUs.
2989 static void idle_balance(int this_cpu, struct rq *this_rq)
2991 struct sched_domain *sd;
2992 int pulled_task = -1;
2993 unsigned long next_balance = jiffies + HZ;
2995 for_each_domain(this_cpu, sd) {
2996 unsigned long interval;
2998 if (!(sd->flags & SD_LOAD_BALANCE))
3001 if (sd->flags & SD_BALANCE_NEWIDLE)
3002 /* If we've pulled tasks over stop searching: */
3003 pulled_task = load_balance_newidle(this_cpu,
3006 interval = msecs_to_jiffies(sd->balance_interval);
3007 if (time_after(next_balance, sd->last_balance + interval))
3008 next_balance = sd->last_balance + interval;
3012 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3014 * We are going idle. next_balance may be set based on
3015 * a busy processor. So reset next_balance.
3017 this_rq->next_balance = next_balance;
3022 * active_load_balance is run by migration threads. It pushes running tasks
3023 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3024 * running on each physical CPU where possible, and avoids physical /
3025 * logical imbalances.
3027 * Called with busiest_rq locked.
3029 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3031 int target_cpu = busiest_rq->push_cpu;
3032 struct sched_domain *sd;
3033 struct rq *target_rq;
3035 /* Is there any task to move? */
3036 if (busiest_rq->nr_running <= 1)
3039 target_rq = cpu_rq(target_cpu);
3042 * This condition is "impossible", if it occurs
3043 * we need to fix it. Originally reported by
3044 * Bjorn Helgaas on a 128-cpu setup.
3046 BUG_ON(busiest_rq == target_rq);
3048 /* move a task from busiest_rq to target_rq */
3049 double_lock_balance(busiest_rq, target_rq);
3050 update_rq_clock(busiest_rq);
3051 update_rq_clock(target_rq);
3053 /* Search for an sd spanning us and the target CPU. */
3054 for_each_domain(target_cpu, sd) {
3055 if ((sd->flags & SD_LOAD_BALANCE) &&
3056 cpu_isset(busiest_cpu, sd->span))
3061 schedstat_inc(sd, alb_count);
3063 if (move_one_task(target_rq, target_cpu, busiest_rq,
3065 schedstat_inc(sd, alb_pushed);
3067 schedstat_inc(sd, alb_failed);
3069 spin_unlock(&target_rq->lock);
3074 atomic_t load_balancer;
3076 } nohz ____cacheline_aligned = {
3077 .load_balancer = ATOMIC_INIT(-1),
3078 .cpu_mask = CPU_MASK_NONE,
3082 * This routine will try to nominate the ilb (idle load balancing)
3083 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3084 * load balancing on behalf of all those cpus. If all the cpus in the system
3085 * go into this tickless mode, then there will be no ilb owner (as there is
3086 * no need for one) and all the cpus will sleep till the next wakeup event
3089 * For the ilb owner, tick is not stopped. And this tick will be used
3090 * for idle load balancing. ilb owner will still be part of
3093 * While stopping the tick, this cpu will become the ilb owner if there
3094 * is no other owner. And will be the owner till that cpu becomes busy
3095 * or if all cpus in the system stop their ticks at which point
3096 * there is no need for ilb owner.
3098 * When the ilb owner becomes busy, it nominates another owner, during the
3099 * next busy scheduler_tick()
3101 int select_nohz_load_balancer(int stop_tick)
3103 int cpu = smp_processor_id();
3106 cpu_set(cpu, nohz.cpu_mask);
3107 cpu_rq(cpu)->in_nohz_recently = 1;
3110 * If we are going offline and still the leader, give up!
3112 if (cpu_is_offline(cpu) &&
3113 atomic_read(&nohz.load_balancer) == cpu) {
3114 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3119 /* time for ilb owner also to sleep */
3120 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3121 if (atomic_read(&nohz.load_balancer) == cpu)
3122 atomic_set(&nohz.load_balancer, -1);
3126 if (atomic_read(&nohz.load_balancer) == -1) {
3127 /* make me the ilb owner */
3128 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3130 } else if (atomic_read(&nohz.load_balancer) == cpu)
3133 if (!cpu_isset(cpu, nohz.cpu_mask))
3136 cpu_clear(cpu, nohz.cpu_mask);
3138 if (atomic_read(&nohz.load_balancer) == cpu)
3139 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3146 static DEFINE_SPINLOCK(balancing);
3149 * It checks each scheduling domain to see if it is due to be balanced,
3150 * and initiates a balancing operation if so.
3152 * Balancing parameters are set up in arch_init_sched_domains.
3154 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3157 struct rq *rq = cpu_rq(cpu);
3158 unsigned long interval;
3159 struct sched_domain *sd;
3160 /* Earliest time when we have to do rebalance again */
3161 unsigned long next_balance = jiffies + 60*HZ;
3162 int update_next_balance = 0;
3164 for_each_domain(cpu, sd) {
3165 if (!(sd->flags & SD_LOAD_BALANCE))
3168 interval = sd->balance_interval;
3169 if (idle != CPU_IDLE)
3170 interval *= sd->busy_factor;
3172 /* scale ms to jiffies */
3173 interval = msecs_to_jiffies(interval);
3174 if (unlikely(!interval))
3176 if (interval > HZ*NR_CPUS/10)
3177 interval = HZ*NR_CPUS/10;
3180 if (sd->flags & SD_SERIALIZE) {
3181 if (!spin_trylock(&balancing))
3185 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3186 if (load_balance(cpu, rq, sd, idle, &balance)) {
3188 * We've pulled tasks over so either we're no
3189 * longer idle, or one of our SMT siblings is
3192 idle = CPU_NOT_IDLE;
3194 sd->last_balance = jiffies;
3196 if (sd->flags & SD_SERIALIZE)
3197 spin_unlock(&balancing);
3199 if (time_after(next_balance, sd->last_balance + interval)) {
3200 next_balance = sd->last_balance + interval;
3201 update_next_balance = 1;
3205 * Stop the load balance at this level. There is another
3206 * CPU in our sched group which is doing load balancing more
3214 * next_balance will be updated only when there is a need.
3215 * When the cpu is attached to null domain for ex, it will not be
3218 if (likely(update_next_balance))
3219 rq->next_balance = next_balance;
3223 * run_rebalance_domains is triggered when needed from the scheduler tick.
3224 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3225 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3227 static void run_rebalance_domains(struct softirq_action *h)
3229 int this_cpu = smp_processor_id();
3230 struct rq *this_rq = cpu_rq(this_cpu);
3231 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3232 CPU_IDLE : CPU_NOT_IDLE;
3234 rebalance_domains(this_cpu, idle);
3238 * If this cpu is the owner for idle load balancing, then do the
3239 * balancing on behalf of the other idle cpus whose ticks are
3242 if (this_rq->idle_at_tick &&
3243 atomic_read(&nohz.load_balancer) == this_cpu) {
3244 cpumask_t cpus = nohz.cpu_mask;
3248 cpu_clear(this_cpu, cpus);
3249 for_each_cpu_mask(balance_cpu, cpus) {
3251 * If this cpu gets work to do, stop the load balancing
3252 * work being done for other cpus. Next load
3253 * balancing owner will pick it up.
3258 rebalance_domains(balance_cpu, CPU_IDLE);
3260 rq = cpu_rq(balance_cpu);
3261 if (time_after(this_rq->next_balance, rq->next_balance))
3262 this_rq->next_balance = rq->next_balance;
3269 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3271 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3272 * idle load balancing owner or decide to stop the periodic load balancing,
3273 * if the whole system is idle.
3275 static inline void trigger_load_balance(struct rq *rq, int cpu)
3279 * If we were in the nohz mode recently and busy at the current
3280 * scheduler tick, then check if we need to nominate new idle
3283 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3284 rq->in_nohz_recently = 0;
3286 if (atomic_read(&nohz.load_balancer) == cpu) {
3287 cpu_clear(cpu, nohz.cpu_mask);
3288 atomic_set(&nohz.load_balancer, -1);
3291 if (atomic_read(&nohz.load_balancer) == -1) {
3293 * simple selection for now: Nominate the
3294 * first cpu in the nohz list to be the next
3297 * TBD: Traverse the sched domains and nominate
3298 * the nearest cpu in the nohz.cpu_mask.
3300 int ilb = first_cpu(nohz.cpu_mask);
3308 * If this cpu is idle and doing idle load balancing for all the
3309 * cpus with ticks stopped, is it time for that to stop?
3311 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3312 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3318 * If this cpu is idle and the idle load balancing is done by
3319 * someone else, then no need raise the SCHED_SOFTIRQ
3321 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3322 cpu_isset(cpu, nohz.cpu_mask))
3325 if (time_after_eq(jiffies, rq->next_balance))
3326 raise_softirq(SCHED_SOFTIRQ);
3329 #else /* CONFIG_SMP */
3332 * on UP we do not need to balance between CPUs:
3334 static inline void idle_balance(int cpu, struct rq *rq)
3340 DEFINE_PER_CPU(struct kernel_stat, kstat);
3342 EXPORT_PER_CPU_SYMBOL(kstat);
3345 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3346 * that have not yet been banked in case the task is currently running.
3348 unsigned long long task_sched_runtime(struct task_struct *p)
3350 unsigned long flags;
3354 rq = task_rq_lock(p, &flags);
3355 ns = p->se.sum_exec_runtime;
3356 if (task_current(rq, p)) {
3357 update_rq_clock(rq);
3358 delta_exec = rq->clock - p->se.exec_start;
3359 if ((s64)delta_exec > 0)
3362 task_rq_unlock(rq, &flags);
3368 * Account user cpu time to a process.
3369 * @p: the process that the cpu time gets accounted to
3370 * @cputime: the cpu time spent in user space since the last update
3372 void account_user_time(struct task_struct *p, cputime_t cputime)
3374 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3377 p->utime = cputime_add(p->utime, cputime);
3379 /* Add user time to cpustat. */
3380 tmp = cputime_to_cputime64(cputime);
3381 if (TASK_NICE(p) > 0)
3382 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3384 cpustat->user = cputime64_add(cpustat->user, tmp);
3388 * Account guest cpu time to a process.
3389 * @p: the process that the cpu time gets accounted to
3390 * @cputime: the cpu time spent in virtual machine since the last update
3392 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3395 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3397 tmp = cputime_to_cputime64(cputime);
3399 p->utime = cputime_add(p->utime, cputime);
3400 p->gtime = cputime_add(p->gtime, cputime);
3402 cpustat->user = cputime64_add(cpustat->user, tmp);
3403 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3407 * Account scaled user cpu time to a process.
3408 * @p: the process that the cpu time gets accounted to
3409 * @cputime: the cpu time spent in user space since the last update
3411 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3413 p->utimescaled = cputime_add(p->utimescaled, cputime);
3417 * Account system cpu time to a process.
3418 * @p: the process that the cpu time gets accounted to
3419 * @hardirq_offset: the offset to subtract from hardirq_count()
3420 * @cputime: the cpu time spent in kernel space since the last update
3422 void account_system_time(struct task_struct *p, int hardirq_offset,
3425 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3426 struct rq *rq = this_rq();
3429 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3430 return account_guest_time(p, cputime);
3432 p->stime = cputime_add(p->stime, cputime);
3434 /* Add system time to cpustat. */
3435 tmp = cputime_to_cputime64(cputime);
3436 if (hardirq_count() - hardirq_offset)
3437 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3438 else if (softirq_count())
3439 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3440 else if (p != rq->idle)
3441 cpustat->system = cputime64_add(cpustat->system, tmp);
3442 else if (atomic_read(&rq->nr_iowait) > 0)
3443 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3445 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3446 /* Account for system time used */
3447 acct_update_integrals(p);
3451 * Account scaled system cpu time to a process.
3452 * @p: the process that the cpu time gets accounted to
3453 * @hardirq_offset: the offset to subtract from hardirq_count()
3454 * @cputime: the cpu time spent in kernel space since the last update
3456 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3458 p->stimescaled = cputime_add(p->stimescaled, cputime);
3462 * Account for involuntary wait time.
3463 * @p: the process from which the cpu time has been stolen
3464 * @steal: the cpu time spent in involuntary wait
3466 void account_steal_time(struct task_struct *p, cputime_t steal)
3468 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3469 cputime64_t tmp = cputime_to_cputime64(steal);
3470 struct rq *rq = this_rq();
3472 if (p == rq->idle) {
3473 p->stime = cputime_add(p->stime, steal);
3474 if (atomic_read(&rq->nr_iowait) > 0)
3475 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3477 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3479 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3483 * This function gets called by the timer code, with HZ frequency.
3484 * We call it with interrupts disabled.
3486 * It also gets called by the fork code, when changing the parent's
3489 void scheduler_tick(void)
3491 int cpu = smp_processor_id();
3492 struct rq *rq = cpu_rq(cpu);
3493 struct task_struct *curr = rq->curr;
3494 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3496 spin_lock(&rq->lock);
3497 __update_rq_clock(rq);
3499 * Let rq->clock advance by at least TICK_NSEC:
3501 if (unlikely(rq->clock < next_tick))
3502 rq->clock = next_tick;
3503 rq->tick_timestamp = rq->clock;
3504 update_cpu_load(rq);
3505 if (curr != rq->idle) /* FIXME: needed? */
3506 curr->sched_class->task_tick(rq, curr);
3507 spin_unlock(&rq->lock);
3510 rq->idle_at_tick = idle_cpu(cpu);
3511 trigger_load_balance(rq, cpu);
3515 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3517 void fastcall add_preempt_count(int val)
3522 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3524 preempt_count() += val;
3526 * Spinlock count overflowing soon?
3528 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3531 EXPORT_SYMBOL(add_preempt_count);
3533 void fastcall sub_preempt_count(int val)
3538 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3541 * Is the spinlock portion underflowing?
3543 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3544 !(preempt_count() & PREEMPT_MASK)))
3547 preempt_count() -= val;
3549 EXPORT_SYMBOL(sub_preempt_count);
3554 * Print scheduling while atomic bug:
3556 static noinline void __schedule_bug(struct task_struct *prev)
3558 struct pt_regs *regs = get_irq_regs();
3560 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3561 prev->comm, prev->pid, preempt_count());
3563 debug_show_held_locks(prev);
3564 if (irqs_disabled())
3565 print_irqtrace_events(prev);
3574 * Various schedule()-time debugging checks and statistics:
3576 static inline void schedule_debug(struct task_struct *prev)
3579 * Test if we are atomic. Since do_exit() needs to call into
3580 * schedule() atomically, we ignore that path for now.
3581 * Otherwise, whine if we are scheduling when we should not be.
3583 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3584 __schedule_bug(prev);
3586 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3588 schedstat_inc(this_rq(), sched_count);
3589 #ifdef CONFIG_SCHEDSTATS
3590 if (unlikely(prev->lock_depth >= 0)) {
3591 schedstat_inc(this_rq(), bkl_count);
3592 schedstat_inc(prev, sched_info.bkl_count);
3598 * Pick up the highest-prio task:
3600 static inline struct task_struct *
3601 pick_next_task(struct rq *rq, struct task_struct *prev)
3603 const struct sched_class *class;
3604 struct task_struct *p;
3607 * Optimization: we know that if all tasks are in
3608 * the fair class we can call that function directly:
3610 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3611 p = fair_sched_class.pick_next_task(rq);
3616 class = sched_class_highest;
3618 p = class->pick_next_task(rq);
3622 * Will never be NULL as the idle class always
3623 * returns a non-NULL p:
3625 class = class->next;
3630 * schedule() is the main scheduler function.
3632 asmlinkage void __sched schedule(void)
3634 struct task_struct *prev, *next;
3641 cpu = smp_processor_id();
3645 switch_count = &prev->nivcsw;
3647 release_kernel_lock(prev);
3648 need_resched_nonpreemptible:
3650 schedule_debug(prev);
3653 * Do the rq-clock update outside the rq lock:
3655 local_irq_disable();
3656 __update_rq_clock(rq);
3657 spin_lock(&rq->lock);
3658 clear_tsk_need_resched(prev);
3660 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3661 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3662 unlikely(signal_pending(prev)))) {
3663 prev->state = TASK_RUNNING;
3665 deactivate_task(rq, prev, 1);
3667 switch_count = &prev->nvcsw;
3670 if (unlikely(!rq->nr_running))
3671 idle_balance(cpu, rq);
3673 prev->sched_class->put_prev_task(rq, prev);
3674 next = pick_next_task(rq, prev);
3676 sched_info_switch(prev, next);
3678 if (likely(prev != next)) {
3683 context_switch(rq, prev, next); /* unlocks the rq */
3685 spin_unlock_irq(&rq->lock);
3687 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3688 cpu = smp_processor_id();
3690 goto need_resched_nonpreemptible;
3692 preempt_enable_no_resched();
3693 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3696 EXPORT_SYMBOL(schedule);
3698 #ifdef CONFIG_PREEMPT
3700 * this is the entry point to schedule() from in-kernel preemption
3701 * off of preempt_enable. Kernel preemptions off return from interrupt
3702 * occur there and call schedule directly.
3704 asmlinkage void __sched preempt_schedule(void)
3706 struct thread_info *ti = current_thread_info();
3707 #ifdef CONFIG_PREEMPT_BKL
3708 struct task_struct *task = current;
3709 int saved_lock_depth;
3712 * If there is a non-zero preempt_count or interrupts are disabled,
3713 * we do not want to preempt the current task. Just return..
3715 if (likely(ti->preempt_count || irqs_disabled()))
3719 add_preempt_count(PREEMPT_ACTIVE);
3722 * We keep the big kernel semaphore locked, but we
3723 * clear ->lock_depth so that schedule() doesnt
3724 * auto-release the semaphore:
3726 #ifdef CONFIG_PREEMPT_BKL
3727 saved_lock_depth = task->lock_depth;
3728 task->lock_depth = -1;
3731 #ifdef CONFIG_PREEMPT_BKL
3732 task->lock_depth = saved_lock_depth;
3734 sub_preempt_count(PREEMPT_ACTIVE);
3737 * Check again in case we missed a preemption opportunity
3738 * between schedule and now.
3741 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3743 EXPORT_SYMBOL(preempt_schedule);
3746 * this is the entry point to schedule() from kernel preemption
3747 * off of irq context.
3748 * Note, that this is called and return with irqs disabled. This will
3749 * protect us against recursive calling from irq.
3751 asmlinkage void __sched preempt_schedule_irq(void)
3753 struct thread_info *ti = current_thread_info();
3754 #ifdef CONFIG_PREEMPT_BKL
3755 struct task_struct *task = current;
3756 int saved_lock_depth;
3758 /* Catch callers which need to be fixed */
3759 BUG_ON(ti->preempt_count || !irqs_disabled());
3762 add_preempt_count(PREEMPT_ACTIVE);
3765 * We keep the big kernel semaphore locked, but we
3766 * clear ->lock_depth so that schedule() doesnt
3767 * auto-release the semaphore:
3769 #ifdef CONFIG_PREEMPT_BKL
3770 saved_lock_depth = task->lock_depth;
3771 task->lock_depth = -1;
3775 local_irq_disable();
3776 #ifdef CONFIG_PREEMPT_BKL
3777 task->lock_depth = saved_lock_depth;
3779 sub_preempt_count(PREEMPT_ACTIVE);
3782 * Check again in case we missed a preemption opportunity
3783 * between schedule and now.
3786 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3789 #endif /* CONFIG_PREEMPT */
3791 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3794 return try_to_wake_up(curr->private, mode, sync);
3796 EXPORT_SYMBOL(default_wake_function);
3799 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3800 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3801 * number) then we wake all the non-exclusive tasks and one exclusive task.
3803 * There are circumstances in which we can try to wake a task which has already
3804 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3805 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3807 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3808 int nr_exclusive, int sync, void *key)
3810 wait_queue_t *curr, *next;
3812 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3813 unsigned flags = curr->flags;
3815 if (curr->func(curr, mode, sync, key) &&
3816 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3822 * __wake_up - wake up threads blocked on a waitqueue.
3824 * @mode: which threads
3825 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3826 * @key: is directly passed to the wakeup function
3828 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3829 int nr_exclusive, void *key)
3831 unsigned long flags;
3833 spin_lock_irqsave(&q->lock, flags);
3834 __wake_up_common(q, mode, nr_exclusive, 0, key);
3835 spin_unlock_irqrestore(&q->lock, flags);
3837 EXPORT_SYMBOL(__wake_up);
3840 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3842 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3844 __wake_up_common(q, mode, 1, 0, NULL);
3848 * __wake_up_sync - wake up threads blocked on a waitqueue.
3850 * @mode: which threads
3851 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3853 * The sync wakeup differs that the waker knows that it will schedule
3854 * away soon, so while the target thread will be woken up, it will not
3855 * be migrated to another CPU - ie. the two threads are 'synchronized'
3856 * with each other. This can prevent needless bouncing between CPUs.
3858 * On UP it can prevent extra preemption.
3861 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3863 unsigned long flags;
3869 if (unlikely(!nr_exclusive))
3872 spin_lock_irqsave(&q->lock, flags);
3873 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3874 spin_unlock_irqrestore(&q->lock, flags);
3876 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3878 void complete(struct completion *x)
3880 unsigned long flags;
3882 spin_lock_irqsave(&x->wait.lock, flags);
3884 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3886 spin_unlock_irqrestore(&x->wait.lock, flags);
3888 EXPORT_SYMBOL(complete);
3890 void complete_all(struct completion *x)
3892 unsigned long flags;
3894 spin_lock_irqsave(&x->wait.lock, flags);
3895 x->done += UINT_MAX/2;
3896 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3898 spin_unlock_irqrestore(&x->wait.lock, flags);
3900 EXPORT_SYMBOL(complete_all);
3902 static inline long __sched
3903 do_wait_for_common(struct completion *x, long timeout, int state)
3906 DECLARE_WAITQUEUE(wait, current);
3908 wait.flags |= WQ_FLAG_EXCLUSIVE;
3909 __add_wait_queue_tail(&x->wait, &wait);
3911 if (state == TASK_INTERRUPTIBLE &&
3912 signal_pending(current)) {
3913 __remove_wait_queue(&x->wait, &wait);
3914 return -ERESTARTSYS;
3916 __set_current_state(state);
3917 spin_unlock_irq(&x->wait.lock);
3918 timeout = schedule_timeout(timeout);
3919 spin_lock_irq(&x->wait.lock);
3921 __remove_wait_queue(&x->wait, &wait);
3925 __remove_wait_queue(&x->wait, &wait);
3932 wait_for_common(struct completion *x, long timeout, int state)
3936 spin_lock_irq(&x->wait.lock);
3937 timeout = do_wait_for_common(x, timeout, state);
3938 spin_unlock_irq(&x->wait.lock);
3942 void __sched wait_for_completion(struct completion *x)
3944 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3946 EXPORT_SYMBOL(wait_for_completion);
3948 unsigned long __sched
3949 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3951 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3953 EXPORT_SYMBOL(wait_for_completion_timeout);
3955 int __sched wait_for_completion_interruptible(struct completion *x)
3957 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3958 if (t == -ERESTARTSYS)
3962 EXPORT_SYMBOL(wait_for_completion_interruptible);
3964 unsigned long __sched
3965 wait_for_completion_interruptible_timeout(struct completion *x,
3966 unsigned long timeout)
3968 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3970 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3973 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3975 unsigned long flags;
3978 init_waitqueue_entry(&wait, current);
3980 __set_current_state(state);
3982 spin_lock_irqsave(&q->lock, flags);
3983 __add_wait_queue(q, &wait);
3984 spin_unlock(&q->lock);
3985 timeout = schedule_timeout(timeout);
3986 spin_lock_irq(&q->lock);
3987 __remove_wait_queue(q, &wait);
3988 spin_unlock_irqrestore(&q->lock, flags);
3993 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3995 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3997 EXPORT_SYMBOL(interruptible_sleep_on);
4000 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4002 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4004 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4006 void __sched sleep_on(wait_queue_head_t *q)
4008 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4010 EXPORT_SYMBOL(sleep_on);
4012 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4014 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4016 EXPORT_SYMBOL(sleep_on_timeout);
4018 #ifdef CONFIG_RT_MUTEXES
4021 * rt_mutex_setprio - set the current priority of a task
4023 * @prio: prio value (kernel-internal form)
4025 * This function changes the 'effective' priority of a task. It does
4026 * not touch ->normal_prio like __setscheduler().
4028 * Used by the rt_mutex code to implement priority inheritance logic.
4030 void rt_mutex_setprio(struct task_struct *p, int prio)
4032 unsigned long flags;
4033 int oldprio, on_rq, running;
4036 BUG_ON(prio < 0 || prio > MAX_PRIO);
4038 rq = task_rq_lock(p, &flags);
4039 update_rq_clock(rq);
4042 on_rq = p->se.on_rq;
4043 running = task_current(rq, p);
4045 dequeue_task(rq, p, 0);
4047 p->sched_class->put_prev_task(rq, p);
4051 p->sched_class = &rt_sched_class;
4053 p->sched_class = &fair_sched_class;
4059 p->sched_class->set_curr_task(rq);
4060 enqueue_task(rq, p, 0);
4062 * Reschedule if we are currently running on this runqueue and
4063 * our priority decreased, or if we are not currently running on
4064 * this runqueue and our priority is higher than the current's
4067 if (p->prio > oldprio)
4068 resched_task(rq->curr);
4070 check_preempt_curr(rq, p);
4073 task_rq_unlock(rq, &flags);
4078 void set_user_nice(struct task_struct *p, long nice)
4080 int old_prio, delta, on_rq;
4081 unsigned long flags;
4084 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4087 * We have to be careful, if called from sys_setpriority(),
4088 * the task might be in the middle of scheduling on another CPU.
4090 rq = task_rq_lock(p, &flags);
4091 update_rq_clock(rq);
4093 * The RT priorities are set via sched_setscheduler(), but we still
4094 * allow the 'normal' nice value to be set - but as expected
4095 * it wont have any effect on scheduling until the task is
4096 * SCHED_FIFO/SCHED_RR:
4098 if (task_has_rt_policy(p)) {
4099 p->static_prio = NICE_TO_PRIO(nice);
4102 on_rq = p->se.on_rq;
4104 dequeue_task(rq, p, 0);
4106 p->static_prio = NICE_TO_PRIO(nice);
4109 p->prio = effective_prio(p);
4110 delta = p->prio - old_prio;
4113 enqueue_task(rq, p, 0);
4115 * If the task increased its priority or is running and
4116 * lowered its priority, then reschedule its CPU:
4118 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4119 resched_task(rq->curr);
4122 task_rq_unlock(rq, &flags);
4124 EXPORT_SYMBOL(set_user_nice);
4127 * can_nice - check if a task can reduce its nice value
4131 int can_nice(const struct task_struct *p, const int nice)
4133 /* convert nice value [19,-20] to rlimit style value [1,40] */
4134 int nice_rlim = 20 - nice;
4136 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4137 capable(CAP_SYS_NICE));
4140 #ifdef __ARCH_WANT_SYS_NICE
4143 * sys_nice - change the priority of the current process.
4144 * @increment: priority increment
4146 * sys_setpriority is a more generic, but much slower function that
4147 * does similar things.
4149 asmlinkage long sys_nice(int increment)
4154 * Setpriority might change our priority at the same moment.
4155 * We don't have to worry. Conceptually one call occurs first
4156 * and we have a single winner.
4158 if (increment < -40)
4163 nice = PRIO_TO_NICE(current->static_prio) + increment;
4169 if (increment < 0 && !can_nice(current, nice))
4172 retval = security_task_setnice(current, nice);
4176 set_user_nice(current, nice);
4183 * task_prio - return the priority value of a given task.
4184 * @p: the task in question.
4186 * This is the priority value as seen by users in /proc.
4187 * RT tasks are offset by -200. Normal tasks are centered
4188 * around 0, value goes from -16 to +15.
4190 int task_prio(const struct task_struct *p)
4192 return p->prio - MAX_RT_PRIO;
4196 * task_nice - return the nice value of a given task.
4197 * @p: the task in question.
4199 int task_nice(const struct task_struct *p)
4201 return TASK_NICE(p);
4203 EXPORT_SYMBOL_GPL(task_nice);
4206 * idle_cpu - is a given cpu idle currently?
4207 * @cpu: the processor in question.
4209 int idle_cpu(int cpu)
4211 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4215 * idle_task - return the idle task for a given cpu.
4216 * @cpu: the processor in question.
4218 struct task_struct *idle_task(int cpu)
4220 return cpu_rq(cpu)->idle;
4224 * find_process_by_pid - find a process with a matching PID value.
4225 * @pid: the pid in question.
4227 static struct task_struct *find_process_by_pid(pid_t pid)
4229 return pid ? find_task_by_vpid(pid) : current;
4232 /* Actually do priority change: must hold rq lock. */
4234 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4236 BUG_ON(p->se.on_rq);
4239 switch (p->policy) {
4243 p->sched_class = &fair_sched_class;
4247 p->sched_class = &rt_sched_class;
4251 p->rt_priority = prio;
4252 p->normal_prio = normal_prio(p);
4253 /* we are holding p->pi_lock already */
4254 p->prio = rt_mutex_getprio(p);
4259 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4260 * @p: the task in question.
4261 * @policy: new policy.
4262 * @param: structure containing the new RT priority.
4264 * NOTE that the task may be already dead.
4266 int sched_setscheduler(struct task_struct *p, int policy,
4267 struct sched_param *param)
4269 int retval, oldprio, oldpolicy = -1, on_rq, running;
4270 unsigned long flags;
4273 /* may grab non-irq protected spin_locks */
4274 BUG_ON(in_interrupt());
4276 /* double check policy once rq lock held */
4278 policy = oldpolicy = p->policy;
4279 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4280 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4281 policy != SCHED_IDLE)
4284 * Valid priorities for SCHED_FIFO and SCHED_RR are
4285 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4286 * SCHED_BATCH and SCHED_IDLE is 0.
4288 if (param->sched_priority < 0 ||
4289 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4290 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4292 if (rt_policy(policy) != (param->sched_priority != 0))
4296 * Allow unprivileged RT tasks to decrease priority:
4298 if (!capable(CAP_SYS_NICE)) {
4299 if (rt_policy(policy)) {
4300 unsigned long rlim_rtprio;
4302 if (!lock_task_sighand(p, &flags))
4304 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4305 unlock_task_sighand(p, &flags);
4307 /* can't set/change the rt policy */
4308 if (policy != p->policy && !rlim_rtprio)
4311 /* can't increase priority */
4312 if (param->sched_priority > p->rt_priority &&
4313 param->sched_priority > rlim_rtprio)
4317 * Like positive nice levels, dont allow tasks to
4318 * move out of SCHED_IDLE either:
4320 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4323 /* can't change other user's priorities */
4324 if ((current->euid != p->euid) &&
4325 (current->euid != p->uid))
4329 retval = security_task_setscheduler(p, policy, param);
4333 * make sure no PI-waiters arrive (or leave) while we are
4334 * changing the priority of the task:
4336 spin_lock_irqsave(&p->pi_lock, flags);
4338 * To be able to change p->policy safely, the apropriate
4339 * runqueue lock must be held.
4341 rq = __task_rq_lock(p);
4342 /* recheck policy now with rq lock held */
4343 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4344 policy = oldpolicy = -1;
4345 __task_rq_unlock(rq);
4346 spin_unlock_irqrestore(&p->pi_lock, flags);
4349 update_rq_clock(rq);
4350 on_rq = p->se.on_rq;
4351 running = task_current(rq, p);
4353 deactivate_task(rq, p, 0);
4355 p->sched_class->put_prev_task(rq, p);
4359 __setscheduler(rq, p, policy, param->sched_priority);
4363 p->sched_class->set_curr_task(rq);
4364 activate_task(rq, p, 0);
4366 * Reschedule if we are currently running on this runqueue and
4367 * our priority decreased, or if we are not currently running on
4368 * this runqueue and our priority is higher than the current's
4371 if (p->prio > oldprio)
4372 resched_task(rq->curr);
4374 check_preempt_curr(rq, p);
4377 __task_rq_unlock(rq);
4378 spin_unlock_irqrestore(&p->pi_lock, flags);
4380 rt_mutex_adjust_pi(p);
4384 EXPORT_SYMBOL_GPL(sched_setscheduler);
4387 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4389 struct sched_param lparam;
4390 struct task_struct *p;
4393 if (!param || pid < 0)
4395 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4400 p = find_process_by_pid(pid);
4402 retval = sched_setscheduler(p, policy, &lparam);
4409 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4410 * @pid: the pid in question.
4411 * @policy: new policy.
4412 * @param: structure containing the new RT priority.
4415 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4417 /* negative values for policy are not valid */
4421 return do_sched_setscheduler(pid, policy, param);
4425 * sys_sched_setparam - set/change the RT priority of a thread
4426 * @pid: the pid in question.
4427 * @param: structure containing the new RT priority.
4429 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4431 return do_sched_setscheduler(pid, -1, param);
4435 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4436 * @pid: the pid in question.
4438 asmlinkage long sys_sched_getscheduler(pid_t pid)
4440 struct task_struct *p;
4447 read_lock(&tasklist_lock);
4448 p = find_process_by_pid(pid);
4450 retval = security_task_getscheduler(p);
4454 read_unlock(&tasklist_lock);
4459 * sys_sched_getscheduler - get the RT priority of a thread
4460 * @pid: the pid in question.
4461 * @param: structure containing the RT priority.
4463 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4465 struct sched_param lp;
4466 struct task_struct *p;
4469 if (!param || pid < 0)
4472 read_lock(&tasklist_lock);
4473 p = find_process_by_pid(pid);
4478 retval = security_task_getscheduler(p);
4482 lp.sched_priority = p->rt_priority;
4483 read_unlock(&tasklist_lock);
4486 * This one might sleep, we cannot do it with a spinlock held ...
4488 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4493 read_unlock(&tasklist_lock);
4497 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4499 cpumask_t cpus_allowed;
4500 struct task_struct *p;
4503 mutex_lock(&sched_hotcpu_mutex);
4504 read_lock(&tasklist_lock);
4506 p = find_process_by_pid(pid);
4508 read_unlock(&tasklist_lock);
4509 mutex_unlock(&sched_hotcpu_mutex);
4514 * It is not safe to call set_cpus_allowed with the
4515 * tasklist_lock held. We will bump the task_struct's
4516 * usage count and then drop tasklist_lock.
4519 read_unlock(&tasklist_lock);
4522 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4523 !capable(CAP_SYS_NICE))
4526 retval = security_task_setscheduler(p, 0, NULL);
4530 cpus_allowed = cpuset_cpus_allowed(p);
4531 cpus_and(new_mask, new_mask, cpus_allowed);
4533 retval = set_cpus_allowed(p, new_mask);
4536 cpus_allowed = cpuset_cpus_allowed(p);
4537 if (!cpus_subset(new_mask, cpus_allowed)) {
4539 * We must have raced with a concurrent cpuset
4540 * update. Just reset the cpus_allowed to the
4541 * cpuset's cpus_allowed
4543 new_mask = cpus_allowed;
4549 mutex_unlock(&sched_hotcpu_mutex);
4553 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4554 cpumask_t *new_mask)
4556 if (len < sizeof(cpumask_t)) {
4557 memset(new_mask, 0, sizeof(cpumask_t));
4558 } else if (len > sizeof(cpumask_t)) {
4559 len = sizeof(cpumask_t);
4561 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4565 * sys_sched_setaffinity - set the cpu affinity of a process
4566 * @pid: pid of the process
4567 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4568 * @user_mask_ptr: user-space pointer to the new cpu mask
4570 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4571 unsigned long __user *user_mask_ptr)
4576 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4580 return sched_setaffinity(pid, new_mask);
4584 * Represents all cpu's present in the system
4585 * In systems capable of hotplug, this map could dynamically grow
4586 * as new cpu's are detected in the system via any platform specific
4587 * method, such as ACPI for e.g.
4590 cpumask_t cpu_present_map __read_mostly;
4591 EXPORT_SYMBOL(cpu_present_map);
4594 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4595 EXPORT_SYMBOL(cpu_online_map);
4597 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4598 EXPORT_SYMBOL(cpu_possible_map);
4601 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4603 struct task_struct *p;
4606 mutex_lock(&sched_hotcpu_mutex);
4607 read_lock(&tasklist_lock);
4610 p = find_process_by_pid(pid);
4614 retval = security_task_getscheduler(p);
4618 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4621 read_unlock(&tasklist_lock);
4622 mutex_unlock(&sched_hotcpu_mutex);
4628 * sys_sched_getaffinity - get the cpu affinity of a process
4629 * @pid: pid of the process
4630 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4631 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4633 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4634 unsigned long __user *user_mask_ptr)
4639 if (len < sizeof(cpumask_t))
4642 ret = sched_getaffinity(pid, &mask);
4646 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4649 return sizeof(cpumask_t);
4653 * sys_sched_yield - yield the current processor to other threads.
4655 * This function yields the current CPU to other tasks. If there are no
4656 * other threads running on this CPU then this function will return.
4658 asmlinkage long sys_sched_yield(void)
4660 struct rq *rq = this_rq_lock();
4662 schedstat_inc(rq, yld_count);
4663 current->sched_class->yield_task(rq);
4666 * Since we are going to call schedule() anyway, there's
4667 * no need to preempt or enable interrupts:
4669 __release(rq->lock);
4670 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4671 _raw_spin_unlock(&rq->lock);
4672 preempt_enable_no_resched();
4679 static void __cond_resched(void)
4681 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4682 __might_sleep(__FILE__, __LINE__);
4685 * The BKS might be reacquired before we have dropped
4686 * PREEMPT_ACTIVE, which could trigger a second
4687 * cond_resched() call.
4690 add_preempt_count(PREEMPT_ACTIVE);
4692 sub_preempt_count(PREEMPT_ACTIVE);
4693 } while (need_resched());
4696 int __sched cond_resched(void)
4698 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4699 system_state == SYSTEM_RUNNING) {
4705 EXPORT_SYMBOL(cond_resched);
4708 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4709 * call schedule, and on return reacquire the lock.
4711 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4712 * operations here to prevent schedule() from being called twice (once via
4713 * spin_unlock(), once by hand).
4715 int cond_resched_lock(spinlock_t *lock)
4719 if (need_lockbreak(lock)) {
4725 if (need_resched() && system_state == SYSTEM_RUNNING) {
4726 spin_release(&lock->dep_map, 1, _THIS_IP_);
4727 _raw_spin_unlock(lock);
4728 preempt_enable_no_resched();
4735 EXPORT_SYMBOL(cond_resched_lock);
4737 int __sched cond_resched_softirq(void)
4739 BUG_ON(!in_softirq());
4741 if (need_resched() && system_state == SYSTEM_RUNNING) {
4749 EXPORT_SYMBOL(cond_resched_softirq);
4752 * yield - yield the current processor to other threads.
4754 * This is a shortcut for kernel-space yielding - it marks the
4755 * thread runnable and calls sys_sched_yield().
4757 void __sched yield(void)
4759 set_current_state(TASK_RUNNING);
4762 EXPORT_SYMBOL(yield);
4765 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4766 * that process accounting knows that this is a task in IO wait state.
4768 * But don't do that if it is a deliberate, throttling IO wait (this task
4769 * has set its backing_dev_info: the queue against which it should throttle)
4771 void __sched io_schedule(void)
4773 struct rq *rq = &__raw_get_cpu_var(runqueues);
4775 delayacct_blkio_start();
4776 atomic_inc(&rq->nr_iowait);
4778 atomic_dec(&rq->nr_iowait);
4779 delayacct_blkio_end();
4781 EXPORT_SYMBOL(io_schedule);
4783 long __sched io_schedule_timeout(long timeout)
4785 struct rq *rq = &__raw_get_cpu_var(runqueues);
4788 delayacct_blkio_start();
4789 atomic_inc(&rq->nr_iowait);
4790 ret = schedule_timeout(timeout);
4791 atomic_dec(&rq->nr_iowait);
4792 delayacct_blkio_end();
4797 * sys_sched_get_priority_max - return maximum RT priority.
4798 * @policy: scheduling class.
4800 * this syscall returns the maximum rt_priority that can be used
4801 * by a given scheduling class.
4803 asmlinkage long sys_sched_get_priority_max(int policy)
4810 ret = MAX_USER_RT_PRIO-1;
4822 * sys_sched_get_priority_min - return minimum RT priority.
4823 * @policy: scheduling class.
4825 * this syscall returns the minimum rt_priority that can be used
4826 * by a given scheduling class.
4828 asmlinkage long sys_sched_get_priority_min(int policy)
4846 * sys_sched_rr_get_interval - return the default timeslice of a process.
4847 * @pid: pid of the process.
4848 * @interval: userspace pointer to the timeslice value.
4850 * this syscall writes the default timeslice value of a given process
4851 * into the user-space timespec buffer. A value of '0' means infinity.
4854 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4856 struct task_struct *p;
4857 unsigned int time_slice;
4865 read_lock(&tasklist_lock);
4866 p = find_process_by_pid(pid);
4870 retval = security_task_getscheduler(p);
4875 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4876 * tasks that are on an otherwise idle runqueue:
4879 if (p->policy == SCHED_RR) {
4880 time_slice = DEF_TIMESLICE;
4882 struct sched_entity *se = &p->se;
4883 unsigned long flags;
4886 rq = task_rq_lock(p, &flags);
4887 if (rq->cfs.load.weight)
4888 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4889 task_rq_unlock(rq, &flags);
4891 read_unlock(&tasklist_lock);
4892 jiffies_to_timespec(time_slice, &t);
4893 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4897 read_unlock(&tasklist_lock);
4901 static const char stat_nam[] = "RSDTtZX";
4903 static void show_task(struct task_struct *p)
4905 unsigned long free = 0;
4908 state = p->state ? __ffs(p->state) + 1 : 0;
4909 printk(KERN_INFO "%-13.13s %c", p->comm,
4910 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4911 #if BITS_PER_LONG == 32
4912 if (state == TASK_RUNNING)
4913 printk(KERN_CONT " running ");
4915 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4917 if (state == TASK_RUNNING)
4918 printk(KERN_CONT " running task ");
4920 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4922 #ifdef CONFIG_DEBUG_STACK_USAGE
4924 unsigned long *n = end_of_stack(p);
4927 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4930 printk(KERN_CONT "%5lu %5d %6d\n", free,
4931 task_pid_nr(p), task_pid_nr(p->real_parent));
4933 if (state != TASK_RUNNING)
4934 show_stack(p, NULL);
4937 void show_state_filter(unsigned long state_filter)
4939 struct task_struct *g, *p;
4941 #if BITS_PER_LONG == 32
4943 " task PC stack pid father\n");
4946 " task PC stack pid father\n");
4948 read_lock(&tasklist_lock);
4949 do_each_thread(g, p) {
4951 * reset the NMI-timeout, listing all files on a slow
4952 * console might take alot of time:
4954 touch_nmi_watchdog();
4955 if (!state_filter || (p->state & state_filter))
4957 } while_each_thread(g, p);
4959 touch_all_softlockup_watchdogs();
4961 #ifdef CONFIG_SCHED_DEBUG
4962 sysrq_sched_debug_show();
4964 read_unlock(&tasklist_lock);
4966 * Only show locks if all tasks are dumped:
4968 if (state_filter == -1)
4969 debug_show_all_locks();
4972 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4974 idle->sched_class = &idle_sched_class;
4978 * init_idle - set up an idle thread for a given CPU
4979 * @idle: task in question
4980 * @cpu: cpu the idle task belongs to
4982 * NOTE: this function does not set the idle thread's NEED_RESCHED
4983 * flag, to make booting more robust.
4985 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4987 struct rq *rq = cpu_rq(cpu);
4988 unsigned long flags;
4991 idle->se.exec_start = sched_clock();
4993 idle->prio = idle->normal_prio = MAX_PRIO;
4994 idle->cpus_allowed = cpumask_of_cpu(cpu);
4995 __set_task_cpu(idle, cpu);
4997 spin_lock_irqsave(&rq->lock, flags);
4998 rq->curr = rq->idle = idle;
4999 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5002 spin_unlock_irqrestore(&rq->lock, flags);
5004 /* Set the preempt count _outside_ the spinlocks! */
5005 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5006 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5008 task_thread_info(idle)->preempt_count = 0;
5011 * The idle tasks have their own, simple scheduling class:
5013 idle->sched_class = &idle_sched_class;
5017 * In a system that switches off the HZ timer nohz_cpu_mask
5018 * indicates which cpus entered this state. This is used
5019 * in the rcu update to wait only for active cpus. For system
5020 * which do not switch off the HZ timer nohz_cpu_mask should
5021 * always be CPU_MASK_NONE.
5023 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5026 * Increase the granularity value when there are more CPUs,
5027 * because with more CPUs the 'effective latency' as visible
5028 * to users decreases. But the relationship is not linear,
5029 * so pick a second-best guess by going with the log2 of the
5032 * This idea comes from the SD scheduler of Con Kolivas:
5034 static inline void sched_init_granularity(void)
5036 unsigned int factor = 1 + ilog2(num_online_cpus());
5037 const unsigned long limit = 200000000;
5039 sysctl_sched_min_granularity *= factor;
5040 if (sysctl_sched_min_granularity > limit)
5041 sysctl_sched_min_granularity = limit;
5043 sysctl_sched_latency *= factor;
5044 if (sysctl_sched_latency > limit)
5045 sysctl_sched_latency = limit;
5047 sysctl_sched_wakeup_granularity *= factor;
5048 sysctl_sched_batch_wakeup_granularity *= factor;
5053 * This is how migration works:
5055 * 1) we queue a struct migration_req structure in the source CPU's
5056 * runqueue and wake up that CPU's migration thread.
5057 * 2) we down() the locked semaphore => thread blocks.
5058 * 3) migration thread wakes up (implicitly it forces the migrated
5059 * thread off the CPU)
5060 * 4) it gets the migration request and checks whether the migrated
5061 * task is still in the wrong runqueue.
5062 * 5) if it's in the wrong runqueue then the migration thread removes
5063 * it and puts it into the right queue.
5064 * 6) migration thread up()s the semaphore.
5065 * 7) we wake up and the migration is done.
5069 * Change a given task's CPU affinity. Migrate the thread to a
5070 * proper CPU and schedule it away if the CPU it's executing on
5071 * is removed from the allowed bitmask.
5073 * NOTE: the caller must have a valid reference to the task, the
5074 * task must not exit() & deallocate itself prematurely. The
5075 * call is not atomic; no spinlocks may be held.
5077 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5079 struct migration_req req;
5080 unsigned long flags;
5084 rq = task_rq_lock(p, &flags);
5085 if (!cpus_intersects(new_mask, cpu_online_map)) {
5090 p->cpus_allowed = new_mask;
5091 /* Can the task run on the task's current CPU? If so, we're done */
5092 if (cpu_isset(task_cpu(p), new_mask))
5095 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5096 /* Need help from migration thread: drop lock and wait. */
5097 task_rq_unlock(rq, &flags);
5098 wake_up_process(rq->migration_thread);
5099 wait_for_completion(&req.done);
5100 tlb_migrate_finish(p->mm);
5104 task_rq_unlock(rq, &flags);
5108 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5111 * Move (not current) task off this cpu, onto dest cpu. We're doing
5112 * this because either it can't run here any more (set_cpus_allowed()
5113 * away from this CPU, or CPU going down), or because we're
5114 * attempting to rebalance this task on exec (sched_exec).
5116 * So we race with normal scheduler movements, but that's OK, as long
5117 * as the task is no longer on this CPU.
5119 * Returns non-zero if task was successfully migrated.
5121 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5123 struct rq *rq_dest, *rq_src;
5126 if (unlikely(cpu_is_offline(dest_cpu)))
5129 rq_src = cpu_rq(src_cpu);
5130 rq_dest = cpu_rq(dest_cpu);
5132 double_rq_lock(rq_src, rq_dest);
5133 /* Already moved. */
5134 if (task_cpu(p) != src_cpu)
5136 /* Affinity changed (again). */
5137 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5140 on_rq = p->se.on_rq;
5142 deactivate_task(rq_src, p, 0);
5144 set_task_cpu(p, dest_cpu);
5146 activate_task(rq_dest, p, 0);
5147 check_preempt_curr(rq_dest, p);
5151 double_rq_unlock(rq_src, rq_dest);
5156 * migration_thread - this is a highprio system thread that performs
5157 * thread migration by bumping thread off CPU then 'pushing' onto
5160 static int migration_thread(void *data)
5162 int cpu = (long)data;
5166 BUG_ON(rq->migration_thread != current);
5168 set_current_state(TASK_INTERRUPTIBLE);
5169 while (!kthread_should_stop()) {
5170 struct migration_req *req;
5171 struct list_head *head;
5173 spin_lock_irq(&rq->lock);
5175 if (cpu_is_offline(cpu)) {
5176 spin_unlock_irq(&rq->lock);
5180 if (rq->active_balance) {
5181 active_load_balance(rq, cpu);
5182 rq->active_balance = 0;
5185 head = &rq->migration_queue;
5187 if (list_empty(head)) {
5188 spin_unlock_irq(&rq->lock);
5190 set_current_state(TASK_INTERRUPTIBLE);
5193 req = list_entry(head->next, struct migration_req, list);
5194 list_del_init(head->next);
5196 spin_unlock(&rq->lock);
5197 __migrate_task(req->task, cpu, req->dest_cpu);
5200 complete(&req->done);
5202 __set_current_state(TASK_RUNNING);
5206 /* Wait for kthread_stop */
5207 set_current_state(TASK_INTERRUPTIBLE);
5208 while (!kthread_should_stop()) {
5210 set_current_state(TASK_INTERRUPTIBLE);
5212 __set_current_state(TASK_RUNNING);
5216 #ifdef CONFIG_HOTPLUG_CPU
5218 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5222 local_irq_disable();
5223 ret = __migrate_task(p, src_cpu, dest_cpu);
5229 * Figure out where task on dead CPU should go, use force if necessary.
5230 * NOTE: interrupts should be disabled by the caller
5232 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5234 unsigned long flags;
5241 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5242 cpus_and(mask, mask, p->cpus_allowed);
5243 dest_cpu = any_online_cpu(mask);
5245 /* On any allowed CPU? */
5246 if (dest_cpu == NR_CPUS)
5247 dest_cpu = any_online_cpu(p->cpus_allowed);
5249 /* No more Mr. Nice Guy. */
5250 if (dest_cpu == NR_CPUS) {
5251 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5253 * Try to stay on the same cpuset, where the
5254 * current cpuset may be a subset of all cpus.
5255 * The cpuset_cpus_allowed_locked() variant of
5256 * cpuset_cpus_allowed() will not block. It must be
5257 * called within calls to cpuset_lock/cpuset_unlock.
5259 rq = task_rq_lock(p, &flags);
5260 p->cpus_allowed = cpus_allowed;
5261 dest_cpu = any_online_cpu(p->cpus_allowed);
5262 task_rq_unlock(rq, &flags);
5265 * Don't tell them about moving exiting tasks or
5266 * kernel threads (both mm NULL), since they never
5269 if (p->mm && printk_ratelimit()) {
5270 printk(KERN_INFO "process %d (%s) no "
5271 "longer affine to cpu%d\n",
5272 task_pid_nr(p), p->comm, dead_cpu);
5275 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5279 * While a dead CPU has no uninterruptible tasks queued at this point,
5280 * it might still have a nonzero ->nr_uninterruptible counter, because
5281 * for performance reasons the counter is not stricly tracking tasks to
5282 * their home CPUs. So we just add the counter to another CPU's counter,
5283 * to keep the global sum constant after CPU-down:
5285 static void migrate_nr_uninterruptible(struct rq *rq_src)
5287 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5288 unsigned long flags;
5290 local_irq_save(flags);
5291 double_rq_lock(rq_src, rq_dest);
5292 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5293 rq_src->nr_uninterruptible = 0;
5294 double_rq_unlock(rq_src, rq_dest);
5295 local_irq_restore(flags);
5298 /* Run through task list and migrate tasks from the dead cpu. */
5299 static void migrate_live_tasks(int src_cpu)
5301 struct task_struct *p, *t;
5303 read_lock(&tasklist_lock);
5305 do_each_thread(t, p) {
5309 if (task_cpu(p) == src_cpu)
5310 move_task_off_dead_cpu(src_cpu, p);
5311 } while_each_thread(t, p);
5313 read_unlock(&tasklist_lock);
5317 * Schedules idle task to be the next runnable task on current CPU.
5318 * It does so by boosting its priority to highest possible.
5319 * Used by CPU offline code.
5321 void sched_idle_next(void)
5323 int this_cpu = smp_processor_id();
5324 struct rq *rq = cpu_rq(this_cpu);
5325 struct task_struct *p = rq->idle;
5326 unsigned long flags;
5328 /* cpu has to be offline */
5329 BUG_ON(cpu_online(this_cpu));
5332 * Strictly not necessary since rest of the CPUs are stopped by now
5333 * and interrupts disabled on the current cpu.
5335 spin_lock_irqsave(&rq->lock, flags);
5337 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5339 update_rq_clock(rq);
5340 activate_task(rq, p, 0);
5342 spin_unlock_irqrestore(&rq->lock, flags);
5346 * Ensures that the idle task is using init_mm right before its cpu goes
5349 void idle_task_exit(void)
5351 struct mm_struct *mm = current->active_mm;
5353 BUG_ON(cpu_online(smp_processor_id()));
5356 switch_mm(mm, &init_mm, current);
5360 /* called under rq->lock with disabled interrupts */
5361 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5363 struct rq *rq = cpu_rq(dead_cpu);
5365 /* Must be exiting, otherwise would be on tasklist. */
5366 BUG_ON(!p->exit_state);
5368 /* Cannot have done final schedule yet: would have vanished. */
5369 BUG_ON(p->state == TASK_DEAD);
5374 * Drop lock around migration; if someone else moves it,
5375 * that's OK. No task can be added to this CPU, so iteration is
5378 spin_unlock_irq(&rq->lock);
5379 move_task_off_dead_cpu(dead_cpu, p);
5380 spin_lock_irq(&rq->lock);
5385 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5386 static void migrate_dead_tasks(unsigned int dead_cpu)
5388 struct rq *rq = cpu_rq(dead_cpu);
5389 struct task_struct *next;
5392 if (!rq->nr_running)
5394 update_rq_clock(rq);
5395 next = pick_next_task(rq, rq->curr);
5398 migrate_dead(dead_cpu, next);
5402 #endif /* CONFIG_HOTPLUG_CPU */
5404 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5406 static struct ctl_table sd_ctl_dir[] = {
5408 .procname = "sched_domain",
5414 static struct ctl_table sd_ctl_root[] = {
5416 .ctl_name = CTL_KERN,
5417 .procname = "kernel",
5419 .child = sd_ctl_dir,
5424 static struct ctl_table *sd_alloc_ctl_entry(int n)
5426 struct ctl_table *entry =
5427 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5432 static void sd_free_ctl_entry(struct ctl_table **tablep)
5434 struct ctl_table *entry;
5437 * In the intermediate directories, both the child directory and
5438 * procname are dynamically allocated and could fail but the mode
5439 * will always be set. In the lowest directory the names are
5440 * static strings and all have proc handlers.
5442 for (entry = *tablep; entry->mode; entry++) {
5444 sd_free_ctl_entry(&entry->child);
5445 if (entry->proc_handler == NULL)
5446 kfree(entry->procname);
5454 set_table_entry(struct ctl_table *entry,
5455 const char *procname, void *data, int maxlen,
5456 mode_t mode, proc_handler *proc_handler)
5458 entry->procname = procname;
5460 entry->maxlen = maxlen;
5462 entry->proc_handler = proc_handler;
5465 static struct ctl_table *
5466 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5468 struct ctl_table *table = sd_alloc_ctl_entry(12);
5473 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5474 sizeof(long), 0644, proc_doulongvec_minmax);
5475 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5476 sizeof(long), 0644, proc_doulongvec_minmax);
5477 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5478 sizeof(int), 0644, proc_dointvec_minmax);
5479 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5480 sizeof(int), 0644, proc_dointvec_minmax);
5481 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5482 sizeof(int), 0644, proc_dointvec_minmax);
5483 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5484 sizeof(int), 0644, proc_dointvec_minmax);
5485 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5486 sizeof(int), 0644, proc_dointvec_minmax);
5487 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5488 sizeof(int), 0644, proc_dointvec_minmax);
5489 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5490 sizeof(int), 0644, proc_dointvec_minmax);
5491 set_table_entry(&table[9], "cache_nice_tries",
5492 &sd->cache_nice_tries,
5493 sizeof(int), 0644, proc_dointvec_minmax);
5494 set_table_entry(&table[10], "flags", &sd->flags,
5495 sizeof(int), 0644, proc_dointvec_minmax);
5496 /* &table[11] is terminator */
5501 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5503 struct ctl_table *entry, *table;
5504 struct sched_domain *sd;
5505 int domain_num = 0, i;
5508 for_each_domain(cpu, sd)
5510 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5515 for_each_domain(cpu, sd) {
5516 snprintf(buf, 32, "domain%d", i);
5517 entry->procname = kstrdup(buf, GFP_KERNEL);
5519 entry->child = sd_alloc_ctl_domain_table(sd);
5526 static struct ctl_table_header *sd_sysctl_header;
5527 static void register_sched_domain_sysctl(void)
5529 int i, cpu_num = num_online_cpus();
5530 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5533 WARN_ON(sd_ctl_dir[0].child);
5534 sd_ctl_dir[0].child = entry;
5539 for_each_online_cpu(i) {
5540 snprintf(buf, 32, "cpu%d", i);
5541 entry->procname = kstrdup(buf, GFP_KERNEL);
5543 entry->child = sd_alloc_ctl_cpu_table(i);
5547 WARN_ON(sd_sysctl_header);
5548 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5551 /* may be called multiple times per register */
5552 static void unregister_sched_domain_sysctl(void)
5554 if (sd_sysctl_header)
5555 unregister_sysctl_table(sd_sysctl_header);
5556 sd_sysctl_header = NULL;
5557 if (sd_ctl_dir[0].child)
5558 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5561 static void register_sched_domain_sysctl(void)
5564 static void unregister_sched_domain_sysctl(void)
5570 * migration_call - callback that gets triggered when a CPU is added.
5571 * Here we can start up the necessary migration thread for the new CPU.
5573 static int __cpuinit
5574 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5576 struct task_struct *p;
5577 int cpu = (long)hcpu;
5578 unsigned long flags;
5582 case CPU_LOCK_ACQUIRE:
5583 mutex_lock(&sched_hotcpu_mutex);
5586 case CPU_UP_PREPARE:
5587 case CPU_UP_PREPARE_FROZEN:
5588 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5591 kthread_bind(p, cpu);
5592 /* Must be high prio: stop_machine expects to yield to it. */
5593 rq = task_rq_lock(p, &flags);
5594 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5595 task_rq_unlock(rq, &flags);
5596 cpu_rq(cpu)->migration_thread = p;
5600 case CPU_ONLINE_FROZEN:
5601 /* Strictly unnecessary, as first user will wake it. */
5602 wake_up_process(cpu_rq(cpu)->migration_thread);
5605 #ifdef CONFIG_HOTPLUG_CPU
5606 case CPU_UP_CANCELED:
5607 case CPU_UP_CANCELED_FROZEN:
5608 if (!cpu_rq(cpu)->migration_thread)
5610 /* Unbind it from offline cpu so it can run. Fall thru. */
5611 kthread_bind(cpu_rq(cpu)->migration_thread,
5612 any_online_cpu(cpu_online_map));
5613 kthread_stop(cpu_rq(cpu)->migration_thread);
5614 cpu_rq(cpu)->migration_thread = NULL;
5618 case CPU_DEAD_FROZEN:
5619 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5620 migrate_live_tasks(cpu);
5622 kthread_stop(rq->migration_thread);
5623 rq->migration_thread = NULL;
5624 /* Idle task back to normal (off runqueue, low prio) */
5625 spin_lock_irq(&rq->lock);
5626 update_rq_clock(rq);
5627 deactivate_task(rq, rq->idle, 0);
5628 rq->idle->static_prio = MAX_PRIO;
5629 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5630 rq->idle->sched_class = &idle_sched_class;
5631 migrate_dead_tasks(cpu);
5632 spin_unlock_irq(&rq->lock);
5634 migrate_nr_uninterruptible(rq);
5635 BUG_ON(rq->nr_running != 0);
5638 * No need to migrate the tasks: it was best-effort if
5639 * they didn't take sched_hotcpu_mutex. Just wake up
5642 spin_lock_irq(&rq->lock);
5643 while (!list_empty(&rq->migration_queue)) {
5644 struct migration_req *req;
5646 req = list_entry(rq->migration_queue.next,
5647 struct migration_req, list);
5648 list_del_init(&req->list);
5649 complete(&req->done);
5651 spin_unlock_irq(&rq->lock);
5654 case CPU_LOCK_RELEASE:
5655 mutex_unlock(&sched_hotcpu_mutex);
5661 /* Register at highest priority so that task migration (migrate_all_tasks)
5662 * happens before everything else.
5664 static struct notifier_block __cpuinitdata migration_notifier = {
5665 .notifier_call = migration_call,
5669 void __init migration_init(void)
5671 void *cpu = (void *)(long)smp_processor_id();
5674 /* Start one for the boot CPU: */
5675 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5676 BUG_ON(err == NOTIFY_BAD);
5677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5678 register_cpu_notifier(&migration_notifier);
5684 /* Number of possible processor ids */
5685 int nr_cpu_ids __read_mostly = NR_CPUS;
5686 EXPORT_SYMBOL(nr_cpu_ids);
5688 #ifdef CONFIG_SCHED_DEBUG
5690 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5692 struct sched_group *group = sd->groups;
5693 cpumask_t groupmask;
5696 cpumask_scnprintf(str, NR_CPUS, sd->span);
5697 cpus_clear(groupmask);
5699 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5701 if (!(sd->flags & SD_LOAD_BALANCE)) {
5702 printk("does not load-balance\n");
5704 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5709 printk(KERN_CONT "span %s\n", str);
5711 if (!cpu_isset(cpu, sd->span)) {
5712 printk(KERN_ERR "ERROR: domain->span does not contain "
5715 if (!cpu_isset(cpu, group->cpumask)) {
5716 printk(KERN_ERR "ERROR: domain->groups does not contain"
5720 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5724 printk(KERN_ERR "ERROR: group is NULL\n");
5728 if (!group->__cpu_power) {
5729 printk(KERN_CONT "\n");
5730 printk(KERN_ERR "ERROR: domain->cpu_power not "
5735 if (!cpus_weight(group->cpumask)) {
5736 printk(KERN_CONT "\n");
5737 printk(KERN_ERR "ERROR: empty group\n");
5741 if (cpus_intersects(groupmask, group->cpumask)) {
5742 printk(KERN_CONT "\n");
5743 printk(KERN_ERR "ERROR: repeated CPUs\n");
5747 cpus_or(groupmask, groupmask, group->cpumask);
5749 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5750 printk(KERN_CONT " %s", str);
5752 group = group->next;
5753 } while (group != sd->groups);
5754 printk(KERN_CONT "\n");
5756 if (!cpus_equal(sd->span, groupmask))
5757 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5759 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5760 printk(KERN_ERR "ERROR: parent span is not a superset "
5761 "of domain->span\n");
5765 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5770 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5774 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5777 if (sched_domain_debug_one(sd, cpu, level))
5786 # define sched_domain_debug(sd, cpu) do { } while (0)
5789 static int sd_degenerate(struct sched_domain *sd)
5791 if (cpus_weight(sd->span) == 1)
5794 /* Following flags need at least 2 groups */
5795 if (sd->flags & (SD_LOAD_BALANCE |
5796 SD_BALANCE_NEWIDLE |
5800 SD_SHARE_PKG_RESOURCES)) {
5801 if (sd->groups != sd->groups->next)
5805 /* Following flags don't use groups */
5806 if (sd->flags & (SD_WAKE_IDLE |
5815 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5817 unsigned long cflags = sd->flags, pflags = parent->flags;
5819 if (sd_degenerate(parent))
5822 if (!cpus_equal(sd->span, parent->span))
5825 /* Does parent contain flags not in child? */
5826 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5827 if (cflags & SD_WAKE_AFFINE)
5828 pflags &= ~SD_WAKE_BALANCE;
5829 /* Flags needing groups don't count if only 1 group in parent */
5830 if (parent->groups == parent->groups->next) {
5831 pflags &= ~(SD_LOAD_BALANCE |
5832 SD_BALANCE_NEWIDLE |
5836 SD_SHARE_PKG_RESOURCES);
5838 if (~cflags & pflags)
5845 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5846 * hold the hotplug lock.
5848 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5850 struct rq *rq = cpu_rq(cpu);
5851 struct sched_domain *tmp;
5853 /* Remove the sched domains which do not contribute to scheduling. */
5854 for (tmp = sd; tmp; tmp = tmp->parent) {
5855 struct sched_domain *parent = tmp->parent;
5858 if (sd_parent_degenerate(tmp, parent)) {
5859 tmp->parent = parent->parent;
5861 parent->parent->child = tmp;
5865 if (sd && sd_degenerate(sd)) {
5871 sched_domain_debug(sd, cpu);
5873 rcu_assign_pointer(rq->sd, sd);
5876 /* cpus with isolated domains */
5877 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5879 /* Setup the mask of cpus configured for isolated domains */
5880 static int __init isolated_cpu_setup(char *str)
5882 int ints[NR_CPUS], i;
5884 str = get_options(str, ARRAY_SIZE(ints), ints);
5885 cpus_clear(cpu_isolated_map);
5886 for (i = 1; i <= ints[0]; i++)
5887 if (ints[i] < NR_CPUS)
5888 cpu_set(ints[i], cpu_isolated_map);
5892 __setup("isolcpus=", isolated_cpu_setup);
5895 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5896 * to a function which identifies what group(along with sched group) a CPU
5897 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5898 * (due to the fact that we keep track of groups covered with a cpumask_t).
5900 * init_sched_build_groups will build a circular linked list of the groups
5901 * covered by the given span, and will set each group's ->cpumask correctly,
5902 * and ->cpu_power to 0.
5905 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5906 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5907 struct sched_group **sg))
5909 struct sched_group *first = NULL, *last = NULL;
5910 cpumask_t covered = CPU_MASK_NONE;
5913 for_each_cpu_mask(i, span) {
5914 struct sched_group *sg;
5915 int group = group_fn(i, cpu_map, &sg);
5918 if (cpu_isset(i, covered))
5921 sg->cpumask = CPU_MASK_NONE;
5922 sg->__cpu_power = 0;
5924 for_each_cpu_mask(j, span) {
5925 if (group_fn(j, cpu_map, NULL) != group)
5928 cpu_set(j, covered);
5929 cpu_set(j, sg->cpumask);
5940 #define SD_NODES_PER_DOMAIN 16
5945 * find_next_best_node - find the next node to include in a sched_domain
5946 * @node: node whose sched_domain we're building
5947 * @used_nodes: nodes already in the sched_domain
5949 * Find the next node to include in a given scheduling domain. Simply
5950 * finds the closest node not already in the @used_nodes map.
5952 * Should use nodemask_t.
5954 static int find_next_best_node(int node, unsigned long *used_nodes)
5956 int i, n, val, min_val, best_node = 0;
5960 for (i = 0; i < MAX_NUMNODES; i++) {
5961 /* Start at @node */
5962 n = (node + i) % MAX_NUMNODES;
5964 if (!nr_cpus_node(n))
5967 /* Skip already used nodes */
5968 if (test_bit(n, used_nodes))
5971 /* Simple min distance search */
5972 val = node_distance(node, n);
5974 if (val < min_val) {
5980 set_bit(best_node, used_nodes);
5985 * sched_domain_node_span - get a cpumask for a node's sched_domain
5986 * @node: node whose cpumask we're constructing
5987 * @size: number of nodes to include in this span
5989 * Given a node, construct a good cpumask for its sched_domain to span. It
5990 * should be one that prevents unnecessary balancing, but also spreads tasks
5993 static cpumask_t sched_domain_node_span(int node)
5995 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5996 cpumask_t span, nodemask;
6000 bitmap_zero(used_nodes, MAX_NUMNODES);
6002 nodemask = node_to_cpumask(node);
6003 cpus_or(span, span, nodemask);
6004 set_bit(node, used_nodes);
6006 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6007 int next_node = find_next_best_node(node, used_nodes);
6009 nodemask = node_to_cpumask(next_node);
6010 cpus_or(span, span, nodemask);
6017 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6020 * SMT sched-domains:
6022 #ifdef CONFIG_SCHED_SMT
6023 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6024 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6027 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6030 *sg = &per_cpu(sched_group_cpus, cpu);
6036 * multi-core sched-domains:
6038 #ifdef CONFIG_SCHED_MC
6039 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6040 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6043 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6045 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6048 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6049 cpus_and(mask, mask, *cpu_map);
6050 group = first_cpu(mask);
6052 *sg = &per_cpu(sched_group_core, group);
6055 #elif defined(CONFIG_SCHED_MC)
6057 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6060 *sg = &per_cpu(sched_group_core, cpu);
6065 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6066 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6069 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6072 #ifdef CONFIG_SCHED_MC
6073 cpumask_t mask = cpu_coregroup_map(cpu);
6074 cpus_and(mask, mask, *cpu_map);
6075 group = first_cpu(mask);
6076 #elif defined(CONFIG_SCHED_SMT)
6077 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6078 cpus_and(mask, mask, *cpu_map);
6079 group = first_cpu(mask);
6084 *sg = &per_cpu(sched_group_phys, group);
6090 * The init_sched_build_groups can't handle what we want to do with node
6091 * groups, so roll our own. Now each node has its own list of groups which
6092 * gets dynamically allocated.
6094 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6095 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6097 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6098 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6100 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6101 struct sched_group **sg)
6103 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6106 cpus_and(nodemask, nodemask, *cpu_map);
6107 group = first_cpu(nodemask);
6110 *sg = &per_cpu(sched_group_allnodes, group);
6114 static void init_numa_sched_groups_power(struct sched_group *group_head)
6116 struct sched_group *sg = group_head;
6122 for_each_cpu_mask(j, sg->cpumask) {
6123 struct sched_domain *sd;
6125 sd = &per_cpu(phys_domains, j);
6126 if (j != first_cpu(sd->groups->cpumask)) {
6128 * Only add "power" once for each
6134 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6137 } while (sg != group_head);
6142 /* Free memory allocated for various sched_group structures */
6143 static void free_sched_groups(const cpumask_t *cpu_map)
6147 for_each_cpu_mask(cpu, *cpu_map) {
6148 struct sched_group **sched_group_nodes
6149 = sched_group_nodes_bycpu[cpu];
6151 if (!sched_group_nodes)
6154 for (i = 0; i < MAX_NUMNODES; i++) {
6155 cpumask_t nodemask = node_to_cpumask(i);
6156 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6158 cpus_and(nodemask, nodemask, *cpu_map);
6159 if (cpus_empty(nodemask))
6169 if (oldsg != sched_group_nodes[i])
6172 kfree(sched_group_nodes);
6173 sched_group_nodes_bycpu[cpu] = NULL;
6177 static void free_sched_groups(const cpumask_t *cpu_map)
6183 * Initialize sched groups cpu_power.
6185 * cpu_power indicates the capacity of sched group, which is used while
6186 * distributing the load between different sched groups in a sched domain.
6187 * Typically cpu_power for all the groups in a sched domain will be same unless
6188 * there are asymmetries in the topology. If there are asymmetries, group
6189 * having more cpu_power will pickup more load compared to the group having
6192 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6193 * the maximum number of tasks a group can handle in the presence of other idle
6194 * or lightly loaded groups in the same sched domain.
6196 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6198 struct sched_domain *child;
6199 struct sched_group *group;
6201 WARN_ON(!sd || !sd->groups);
6203 if (cpu != first_cpu(sd->groups->cpumask))
6208 sd->groups->__cpu_power = 0;
6211 * For perf policy, if the groups in child domain share resources
6212 * (for example cores sharing some portions of the cache hierarchy
6213 * or SMT), then set this domain groups cpu_power such that each group
6214 * can handle only one task, when there are other idle groups in the
6215 * same sched domain.
6217 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6219 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6220 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6225 * add cpu_power of each child group to this groups cpu_power
6227 group = child->groups;
6229 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6230 group = group->next;
6231 } while (group != child->groups);
6235 * Build sched domains for a given set of cpus and attach the sched domains
6236 * to the individual cpus
6238 static int build_sched_domains(const cpumask_t *cpu_map)
6242 struct sched_group **sched_group_nodes = NULL;
6243 int sd_allnodes = 0;
6246 * Allocate the per-node list of sched groups
6248 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6250 if (!sched_group_nodes) {
6251 printk(KERN_WARNING "Can not alloc sched group node list\n");
6254 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6258 * Set up domains for cpus specified by the cpu_map.
6260 for_each_cpu_mask(i, *cpu_map) {
6261 struct sched_domain *sd = NULL, *p;
6262 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6264 cpus_and(nodemask, nodemask, *cpu_map);
6267 if (cpus_weight(*cpu_map) >
6268 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6269 sd = &per_cpu(allnodes_domains, i);
6270 *sd = SD_ALLNODES_INIT;
6271 sd->span = *cpu_map;
6272 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6278 sd = &per_cpu(node_domains, i);
6280 sd->span = sched_domain_node_span(cpu_to_node(i));
6284 cpus_and(sd->span, sd->span, *cpu_map);
6288 sd = &per_cpu(phys_domains, i);
6290 sd->span = nodemask;
6294 cpu_to_phys_group(i, cpu_map, &sd->groups);
6296 #ifdef CONFIG_SCHED_MC
6298 sd = &per_cpu(core_domains, i);
6300 sd->span = cpu_coregroup_map(i);
6301 cpus_and(sd->span, sd->span, *cpu_map);
6304 cpu_to_core_group(i, cpu_map, &sd->groups);
6307 #ifdef CONFIG_SCHED_SMT
6309 sd = &per_cpu(cpu_domains, i);
6310 *sd = SD_SIBLING_INIT;
6311 sd->span = per_cpu(cpu_sibling_map, i);
6312 cpus_and(sd->span, sd->span, *cpu_map);
6315 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6319 #ifdef CONFIG_SCHED_SMT
6320 /* Set up CPU (sibling) groups */
6321 for_each_cpu_mask(i, *cpu_map) {
6322 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6323 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6324 if (i != first_cpu(this_sibling_map))
6327 init_sched_build_groups(this_sibling_map, cpu_map,
6332 #ifdef CONFIG_SCHED_MC
6333 /* Set up multi-core groups */
6334 for_each_cpu_mask(i, *cpu_map) {
6335 cpumask_t this_core_map = cpu_coregroup_map(i);
6336 cpus_and(this_core_map, this_core_map, *cpu_map);
6337 if (i != first_cpu(this_core_map))
6339 init_sched_build_groups(this_core_map, cpu_map,
6340 &cpu_to_core_group);
6344 /* Set up physical groups */
6345 for (i = 0; i < MAX_NUMNODES; i++) {
6346 cpumask_t nodemask = node_to_cpumask(i);
6348 cpus_and(nodemask, nodemask, *cpu_map);
6349 if (cpus_empty(nodemask))
6352 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6356 /* Set up node groups */
6358 init_sched_build_groups(*cpu_map, cpu_map,
6359 &cpu_to_allnodes_group);
6361 for (i = 0; i < MAX_NUMNODES; i++) {
6362 /* Set up node groups */
6363 struct sched_group *sg, *prev;
6364 cpumask_t nodemask = node_to_cpumask(i);
6365 cpumask_t domainspan;
6366 cpumask_t covered = CPU_MASK_NONE;
6369 cpus_and(nodemask, nodemask, *cpu_map);
6370 if (cpus_empty(nodemask)) {
6371 sched_group_nodes[i] = NULL;
6375 domainspan = sched_domain_node_span(i);
6376 cpus_and(domainspan, domainspan, *cpu_map);
6378 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6380 printk(KERN_WARNING "Can not alloc domain group for "
6384 sched_group_nodes[i] = sg;
6385 for_each_cpu_mask(j, nodemask) {
6386 struct sched_domain *sd;
6388 sd = &per_cpu(node_domains, j);
6391 sg->__cpu_power = 0;
6392 sg->cpumask = nodemask;
6394 cpus_or(covered, covered, nodemask);
6397 for (j = 0; j < MAX_NUMNODES; j++) {
6398 cpumask_t tmp, notcovered;
6399 int n = (i + j) % MAX_NUMNODES;
6401 cpus_complement(notcovered, covered);
6402 cpus_and(tmp, notcovered, *cpu_map);
6403 cpus_and(tmp, tmp, domainspan);
6404 if (cpus_empty(tmp))
6407 nodemask = node_to_cpumask(n);
6408 cpus_and(tmp, tmp, nodemask);
6409 if (cpus_empty(tmp))
6412 sg = kmalloc_node(sizeof(struct sched_group),
6416 "Can not alloc domain group for node %d\n", j);
6419 sg->__cpu_power = 0;
6421 sg->next = prev->next;
6422 cpus_or(covered, covered, tmp);
6429 /* Calculate CPU power for physical packages and nodes */
6430 #ifdef CONFIG_SCHED_SMT
6431 for_each_cpu_mask(i, *cpu_map) {
6432 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6434 init_sched_groups_power(i, sd);
6437 #ifdef CONFIG_SCHED_MC
6438 for_each_cpu_mask(i, *cpu_map) {
6439 struct sched_domain *sd = &per_cpu(core_domains, i);
6441 init_sched_groups_power(i, sd);
6445 for_each_cpu_mask(i, *cpu_map) {
6446 struct sched_domain *sd = &per_cpu(phys_domains, i);
6448 init_sched_groups_power(i, sd);
6452 for (i = 0; i < MAX_NUMNODES; i++)
6453 init_numa_sched_groups_power(sched_group_nodes[i]);
6456 struct sched_group *sg;
6458 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6459 init_numa_sched_groups_power(sg);
6463 /* Attach the domains */
6464 for_each_cpu_mask(i, *cpu_map) {
6465 struct sched_domain *sd;
6466 #ifdef CONFIG_SCHED_SMT
6467 sd = &per_cpu(cpu_domains, i);
6468 #elif defined(CONFIG_SCHED_MC)
6469 sd = &per_cpu(core_domains, i);
6471 sd = &per_cpu(phys_domains, i);
6473 cpu_attach_domain(sd, i);
6480 free_sched_groups(cpu_map);
6485 static cpumask_t *doms_cur; /* current sched domains */
6486 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6489 * Special case: If a kmalloc of a doms_cur partition (array of
6490 * cpumask_t) fails, then fallback to a single sched domain,
6491 * as determined by the single cpumask_t fallback_doms.
6493 static cpumask_t fallback_doms;
6496 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6497 * For now this just excludes isolated cpus, but could be used to
6498 * exclude other special cases in the future.
6500 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6505 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6507 doms_cur = &fallback_doms;
6508 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6509 err = build_sched_domains(doms_cur);
6510 register_sched_domain_sysctl();
6515 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6517 free_sched_groups(cpu_map);
6521 * Detach sched domains from a group of cpus specified in cpu_map
6522 * These cpus will now be attached to the NULL domain
6524 static void detach_destroy_domains(const cpumask_t *cpu_map)
6528 unregister_sched_domain_sysctl();
6530 for_each_cpu_mask(i, *cpu_map)
6531 cpu_attach_domain(NULL, i);
6532 synchronize_sched();
6533 arch_destroy_sched_domains(cpu_map);
6537 * Partition sched domains as specified by the 'ndoms_new'
6538 * cpumasks in the array doms_new[] of cpumasks. This compares
6539 * doms_new[] to the current sched domain partitioning, doms_cur[].
6540 * It destroys each deleted domain and builds each new domain.
6542 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6543 * The masks don't intersect (don't overlap.) We should setup one
6544 * sched domain for each mask. CPUs not in any of the cpumasks will
6545 * not be load balanced. If the same cpumask appears both in the
6546 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6549 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6550 * ownership of it and will kfree it when done with it. If the caller
6551 * failed the kmalloc call, then it can pass in doms_new == NULL,
6552 * and partition_sched_domains() will fallback to the single partition
6555 * Call with hotplug lock held
6557 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6563 /* always unregister in case we don't destroy any domains */
6564 unregister_sched_domain_sysctl();
6566 if (doms_new == NULL) {
6568 doms_new = &fallback_doms;
6569 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6572 /* Destroy deleted domains */
6573 for (i = 0; i < ndoms_cur; i++) {
6574 for (j = 0; j < ndoms_new; j++) {
6575 if (cpus_equal(doms_cur[i], doms_new[j]))
6578 /* no match - a current sched domain not in new doms_new[] */
6579 detach_destroy_domains(doms_cur + i);
6584 /* Build new domains */
6585 for (i = 0; i < ndoms_new; i++) {
6586 for (j = 0; j < ndoms_cur; j++) {
6587 if (cpus_equal(doms_new[i], doms_cur[j]))
6590 /* no match - add a new doms_new */
6591 build_sched_domains(doms_new + i);
6596 /* Remember the new sched domains */
6597 if (doms_cur != &fallback_doms)
6599 doms_cur = doms_new;
6600 ndoms_cur = ndoms_new;
6602 register_sched_domain_sysctl();
6607 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6608 static int arch_reinit_sched_domains(void)
6612 mutex_lock(&sched_hotcpu_mutex);
6613 detach_destroy_domains(&cpu_online_map);
6614 err = arch_init_sched_domains(&cpu_online_map);
6615 mutex_unlock(&sched_hotcpu_mutex);
6620 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6624 if (buf[0] != '0' && buf[0] != '1')
6628 sched_smt_power_savings = (buf[0] == '1');
6630 sched_mc_power_savings = (buf[0] == '1');
6632 ret = arch_reinit_sched_domains();
6634 return ret ? ret : count;
6637 #ifdef CONFIG_SCHED_MC
6638 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6640 return sprintf(page, "%u\n", sched_mc_power_savings);
6642 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6643 const char *buf, size_t count)
6645 return sched_power_savings_store(buf, count, 0);
6647 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6648 sched_mc_power_savings_store);
6651 #ifdef CONFIG_SCHED_SMT
6652 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6654 return sprintf(page, "%u\n", sched_smt_power_savings);
6656 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6657 const char *buf, size_t count)
6659 return sched_power_savings_store(buf, count, 1);
6661 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6662 sched_smt_power_savings_store);
6665 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6669 #ifdef CONFIG_SCHED_SMT
6671 err = sysfs_create_file(&cls->kset.kobj,
6672 &attr_sched_smt_power_savings.attr);
6674 #ifdef CONFIG_SCHED_MC
6675 if (!err && mc_capable())
6676 err = sysfs_create_file(&cls->kset.kobj,
6677 &attr_sched_mc_power_savings.attr);
6684 * Force a reinitialization of the sched domains hierarchy. The domains
6685 * and groups cannot be updated in place without racing with the balancing
6686 * code, so we temporarily attach all running cpus to the NULL domain
6687 * which will prevent rebalancing while the sched domains are recalculated.
6689 static int update_sched_domains(struct notifier_block *nfb,
6690 unsigned long action, void *hcpu)
6693 case CPU_UP_PREPARE:
6694 case CPU_UP_PREPARE_FROZEN:
6695 case CPU_DOWN_PREPARE:
6696 case CPU_DOWN_PREPARE_FROZEN:
6697 detach_destroy_domains(&cpu_online_map);
6700 case CPU_UP_CANCELED:
6701 case CPU_UP_CANCELED_FROZEN:
6702 case CPU_DOWN_FAILED:
6703 case CPU_DOWN_FAILED_FROZEN:
6705 case CPU_ONLINE_FROZEN:
6707 case CPU_DEAD_FROZEN:
6709 * Fall through and re-initialise the domains.
6716 /* The hotplug lock is already held by cpu_up/cpu_down */
6717 arch_init_sched_domains(&cpu_online_map);
6722 void __init sched_init_smp(void)
6724 cpumask_t non_isolated_cpus;
6726 mutex_lock(&sched_hotcpu_mutex);
6727 arch_init_sched_domains(&cpu_online_map);
6728 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6729 if (cpus_empty(non_isolated_cpus))
6730 cpu_set(smp_processor_id(), non_isolated_cpus);
6731 mutex_unlock(&sched_hotcpu_mutex);
6732 /* XXX: Theoretical race here - CPU may be hotplugged now */
6733 hotcpu_notifier(update_sched_domains, 0);
6735 /* Move init over to a non-isolated CPU */
6736 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6738 sched_init_granularity();
6741 void __init sched_init_smp(void)
6743 sched_init_granularity();
6745 #endif /* CONFIG_SMP */
6747 int in_sched_functions(unsigned long addr)
6749 return in_lock_functions(addr) ||
6750 (addr >= (unsigned long)__sched_text_start
6751 && addr < (unsigned long)__sched_text_end);
6754 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6756 cfs_rq->tasks_timeline = RB_ROOT;
6757 #ifdef CONFIG_FAIR_GROUP_SCHED
6760 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6763 void __init sched_init(void)
6765 int highest_cpu = 0;
6768 for_each_possible_cpu(i) {
6769 struct rt_prio_array *array;
6773 spin_lock_init(&rq->lock);
6774 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6777 init_cfs_rq(&rq->cfs, rq);
6778 #ifdef CONFIG_FAIR_GROUP_SCHED
6779 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6781 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6782 struct sched_entity *se =
6783 &per_cpu(init_sched_entity, i);
6785 init_cfs_rq_p[i] = cfs_rq;
6786 init_cfs_rq(cfs_rq, rq);
6787 cfs_rq->tg = &init_task_group;
6788 list_add(&cfs_rq->leaf_cfs_rq_list,
6789 &rq->leaf_cfs_rq_list);
6791 init_sched_entity_p[i] = se;
6792 se->cfs_rq = &rq->cfs;
6794 se->load.weight = init_task_group_load;
6795 se->load.inv_weight =
6796 div64_64(1ULL<<32, init_task_group_load);
6799 init_task_group.shares = init_task_group_load;
6802 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6803 rq->cpu_load[j] = 0;
6806 rq->active_balance = 0;
6807 rq->next_balance = jiffies;
6810 rq->migration_thread = NULL;
6811 INIT_LIST_HEAD(&rq->migration_queue);
6813 atomic_set(&rq->nr_iowait, 0);
6815 array = &rq->rt.active;
6816 for (j = 0; j < MAX_RT_PRIO; j++) {
6817 INIT_LIST_HEAD(array->queue + j);
6818 __clear_bit(j, array->bitmap);
6821 /* delimiter for bitsearch: */
6822 __set_bit(MAX_RT_PRIO, array->bitmap);
6825 set_load_weight(&init_task);
6827 #ifdef CONFIG_PREEMPT_NOTIFIERS
6828 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6832 nr_cpu_ids = highest_cpu + 1;
6833 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6836 #ifdef CONFIG_RT_MUTEXES
6837 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6841 * The boot idle thread does lazy MMU switching as well:
6843 atomic_inc(&init_mm.mm_count);
6844 enter_lazy_tlb(&init_mm, current);
6847 * Make us the idle thread. Technically, schedule() should not be
6848 * called from this thread, however somewhere below it might be,
6849 * but because we are the idle thread, we just pick up running again
6850 * when this runqueue becomes "idle".
6852 init_idle(current, smp_processor_id());
6854 * During early bootup we pretend to be a normal task:
6856 current->sched_class = &fair_sched_class;
6859 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6860 void __might_sleep(char *file, int line)
6863 static unsigned long prev_jiffy; /* ratelimiting */
6865 if ((in_atomic() || irqs_disabled()) &&
6866 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6867 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6869 prev_jiffy = jiffies;
6870 printk(KERN_ERR "BUG: sleeping function called from invalid"
6871 " context at %s:%d\n", file, line);
6872 printk("in_atomic():%d, irqs_disabled():%d\n",
6873 in_atomic(), irqs_disabled());
6874 debug_show_held_locks(current);
6875 if (irqs_disabled())
6876 print_irqtrace_events(current);
6881 EXPORT_SYMBOL(__might_sleep);
6884 #ifdef CONFIG_MAGIC_SYSRQ
6885 static void normalize_task(struct rq *rq, struct task_struct *p)
6888 update_rq_clock(rq);
6889 on_rq = p->se.on_rq;
6891 deactivate_task(rq, p, 0);
6892 __setscheduler(rq, p, SCHED_NORMAL, 0);
6894 activate_task(rq, p, 0);
6895 resched_task(rq->curr);
6899 void normalize_rt_tasks(void)
6901 struct task_struct *g, *p;
6902 unsigned long flags;
6905 read_lock_irq(&tasklist_lock);
6906 do_each_thread(g, p) {
6908 * Only normalize user tasks:
6913 p->se.exec_start = 0;
6914 #ifdef CONFIG_SCHEDSTATS
6915 p->se.wait_start = 0;
6916 p->se.sleep_start = 0;
6917 p->se.block_start = 0;
6919 task_rq(p)->clock = 0;
6923 * Renice negative nice level userspace
6926 if (TASK_NICE(p) < 0 && p->mm)
6927 set_user_nice(p, 0);
6931 spin_lock_irqsave(&p->pi_lock, flags);
6932 rq = __task_rq_lock(p);
6934 normalize_task(rq, p);
6936 __task_rq_unlock(rq);
6937 spin_unlock_irqrestore(&p->pi_lock, flags);
6938 } while_each_thread(g, p);
6940 read_unlock_irq(&tasklist_lock);
6943 #endif /* CONFIG_MAGIC_SYSRQ */
6947 * These functions are only useful for the IA64 MCA handling.
6949 * They can only be called when the whole system has been
6950 * stopped - every CPU needs to be quiescent, and no scheduling
6951 * activity can take place. Using them for anything else would
6952 * be a serious bug, and as a result, they aren't even visible
6953 * under any other configuration.
6957 * curr_task - return the current task for a given cpu.
6958 * @cpu: the processor in question.
6960 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6962 struct task_struct *curr_task(int cpu)
6964 return cpu_curr(cpu);
6968 * set_curr_task - set the current task for a given cpu.
6969 * @cpu: the processor in question.
6970 * @p: the task pointer to set.
6972 * Description: This function must only be used when non-maskable interrupts
6973 * are serviced on a separate stack. It allows the architecture to switch the
6974 * notion of the current task on a cpu in a non-blocking manner. This function
6975 * must be called with all CPU's synchronized, and interrupts disabled, the
6976 * and caller must save the original value of the current task (see
6977 * curr_task() above) and restore that value before reenabling interrupts and
6978 * re-starting the system.
6980 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6982 void set_curr_task(int cpu, struct task_struct *p)
6989 #ifdef CONFIG_FAIR_GROUP_SCHED
6991 /* allocate runqueue etc for a new task group */
6992 struct task_group *sched_create_group(void)
6994 struct task_group *tg;
6995 struct cfs_rq *cfs_rq;
6996 struct sched_entity *se;
7000 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7002 return ERR_PTR(-ENOMEM);
7004 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7007 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7011 for_each_possible_cpu(i) {
7014 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7019 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7024 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7025 memset(se, 0, sizeof(struct sched_entity));
7027 tg->cfs_rq[i] = cfs_rq;
7028 init_cfs_rq(cfs_rq, rq);
7032 se->cfs_rq = &rq->cfs;
7034 se->load.weight = NICE_0_LOAD;
7035 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7039 tg->shares = NICE_0_LOAD;
7041 lock_task_group_list();
7042 for_each_possible_cpu(i) {
7044 cfs_rq = tg->cfs_rq[i];
7045 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7047 unlock_task_group_list();
7052 for_each_possible_cpu(i) {
7054 kfree(tg->cfs_rq[i]);
7062 return ERR_PTR(-ENOMEM);
7065 /* rcu callback to free various structures associated with a task group */
7066 static void free_sched_group(struct rcu_head *rhp)
7068 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7069 struct cfs_rq *cfs_rq;
7070 struct sched_entity *se;
7073 /* now it should be safe to free those cfs_rqs */
7074 for_each_possible_cpu(i) {
7075 cfs_rq = tg->cfs_rq[i];
7087 /* Destroy runqueue etc associated with a task group */
7088 void sched_destroy_group(struct task_group *tg)
7090 struct cfs_rq *cfs_rq = NULL;
7093 lock_task_group_list();
7094 for_each_possible_cpu(i) {
7095 cfs_rq = tg->cfs_rq[i];
7096 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7098 unlock_task_group_list();
7102 /* wait for possible concurrent references to cfs_rqs complete */
7103 call_rcu(&tg->rcu, free_sched_group);
7106 /* change task's runqueue when it moves between groups.
7107 * The caller of this function should have put the task in its new group
7108 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7109 * reflect its new group.
7111 void sched_move_task(struct task_struct *tsk)
7114 unsigned long flags;
7117 rq = task_rq_lock(tsk, &flags);
7119 if (tsk->sched_class != &fair_sched_class) {
7120 set_task_cfs_rq(tsk, task_cpu(tsk));
7124 update_rq_clock(rq);
7126 running = task_current(rq, tsk);
7127 on_rq = tsk->se.on_rq;
7130 dequeue_task(rq, tsk, 0);
7131 if (unlikely(running))
7132 tsk->sched_class->put_prev_task(rq, tsk);
7135 set_task_cfs_rq(tsk, task_cpu(tsk));
7138 if (unlikely(running))
7139 tsk->sched_class->set_curr_task(rq);
7140 enqueue_task(rq, tsk, 0);
7144 task_rq_unlock(rq, &flags);
7147 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7149 struct cfs_rq *cfs_rq = se->cfs_rq;
7150 struct rq *rq = cfs_rq->rq;
7153 spin_lock_irq(&rq->lock);
7157 dequeue_entity(cfs_rq, se, 0);
7159 se->load.weight = shares;
7160 se->load.inv_weight = div64_64((1ULL<<32), shares);
7163 enqueue_entity(cfs_rq, se, 0);
7165 spin_unlock_irq(&rq->lock);
7168 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7173 * A weight of 0 or 1 can cause arithmetics problems.
7174 * (The default weight is 1024 - so there's no practical
7175 * limitation from this.)
7180 lock_task_group_list();
7181 if (tg->shares == shares)
7184 tg->shares = shares;
7185 for_each_possible_cpu(i)
7186 set_se_shares(tg->se[i], shares);
7189 unlock_task_group_list();
7193 unsigned long sched_group_shares(struct task_group *tg)
7198 #endif /* CONFIG_FAIR_GROUP_SCHED */
7200 #ifdef CONFIG_FAIR_CGROUP_SCHED
7202 /* return corresponding task_group object of a cgroup */
7203 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7205 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7206 struct task_group, css);
7209 static struct cgroup_subsys_state *
7210 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7212 struct task_group *tg;
7214 if (!cgrp->parent) {
7215 /* This is early initialization for the top cgroup */
7216 init_task_group.css.cgroup = cgrp;
7217 return &init_task_group.css;
7220 /* we support only 1-level deep hierarchical scheduler atm */
7221 if (cgrp->parent->parent)
7222 return ERR_PTR(-EINVAL);
7224 tg = sched_create_group();
7226 return ERR_PTR(-ENOMEM);
7228 /* Bind the cgroup to task_group object we just created */
7229 tg->css.cgroup = cgrp;
7235 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7237 struct task_group *tg = cgroup_tg(cgrp);
7239 sched_destroy_group(tg);
7243 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7244 struct task_struct *tsk)
7246 /* We don't support RT-tasks being in separate groups */
7247 if (tsk->sched_class != &fair_sched_class)
7254 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7255 struct cgroup *old_cont, struct task_struct *tsk)
7257 sched_move_task(tsk);
7260 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7263 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7266 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7268 struct task_group *tg = cgroup_tg(cgrp);
7270 return (u64) tg->shares;
7273 static struct cftype cpu_files[] = {
7276 .read_uint = cpu_shares_read_uint,
7277 .write_uint = cpu_shares_write_uint,
7281 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7283 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7286 struct cgroup_subsys cpu_cgroup_subsys = {
7288 .create = cpu_cgroup_create,
7289 .destroy = cpu_cgroup_destroy,
7290 .can_attach = cpu_cgroup_can_attach,
7291 .attach = cpu_cgroup_attach,
7292 .populate = cpu_cgroup_populate,
7293 .subsys_id = cpu_cgroup_subsys_id,
7297 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7299 #ifdef CONFIG_CGROUP_CPUACCT
7302 * CPU accounting code for task groups.
7304 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7305 * (balbir@in.ibm.com).
7308 /* track cpu usage of a group of tasks */
7310 struct cgroup_subsys_state css;
7311 /* cpuusage holds pointer to a u64-type object on every cpu */
7315 struct cgroup_subsys cpuacct_subsys;
7317 /* return cpu accounting group corresponding to this container */
7318 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7320 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7321 struct cpuacct, css);
7324 /* return cpu accounting group to which this task belongs */
7325 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7327 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7328 struct cpuacct, css);
7331 /* create a new cpu accounting group */
7332 static struct cgroup_subsys_state *cpuacct_create(
7333 struct cgroup_subsys *ss, struct cgroup *cont)
7335 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7338 return ERR_PTR(-ENOMEM);
7340 ca->cpuusage = alloc_percpu(u64);
7341 if (!ca->cpuusage) {
7343 return ERR_PTR(-ENOMEM);
7349 /* destroy an existing cpu accounting group */
7351 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7353 struct cpuacct *ca = cgroup_ca(cont);
7355 free_percpu(ca->cpuusage);
7359 /* return total cpu usage (in nanoseconds) of a group */
7360 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7362 struct cpuacct *ca = cgroup_ca(cont);
7363 u64 totalcpuusage = 0;
7366 for_each_possible_cpu(i) {
7367 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7370 * Take rq->lock to make 64-bit addition safe on 32-bit
7373 spin_lock_irq(&cpu_rq(i)->lock);
7374 totalcpuusage += *cpuusage;
7375 spin_unlock_irq(&cpu_rq(i)->lock);
7378 return totalcpuusage;
7381 static struct cftype files[] = {
7384 .read_uint = cpuusage_read,
7388 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7390 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7394 * charge this task's execution time to its accounting group.
7396 * called with rq->lock held.
7398 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7402 if (!cpuacct_subsys.active)
7407 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7409 *cpuusage += cputime;
7413 struct cgroup_subsys cpuacct_subsys = {
7415 .create = cpuacct_create,
7416 .destroy = cpuacct_destroy,
7417 .populate = cpuacct_populate,
7418 .subsys_id = cpuacct_subsys_id,
7420 #endif /* CONFIG_CGROUP_CPUACCT */