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
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak)) sched_clock(void)
80 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 #ifdef CONFIG_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups);
166 /* task group related information */
168 #ifdef CONFIG_CGROUP_SCHED
169 struct cgroup_subsys_state css;
172 #ifdef CONFIG_FAIR_GROUP_SCHED
173 /* schedulable entities of this group on each cpu */
174 struct sched_entity **se;
175 /* runqueue "owned" by this group on each cpu */
176 struct cfs_rq **cfs_rq;
177 unsigned long shares;
180 #ifdef CONFIG_RT_GROUP_SCHED
181 struct sched_rt_entity **rt_se;
182 struct rt_rq **rt_rq;
188 struct list_head list;
191 #ifdef CONFIG_FAIR_GROUP_SCHED
192 /* Default task group's sched entity on each cpu */
193 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
194 /* Default task group's cfs_rq on each cpu */
195 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
197 static struct sched_entity *init_sched_entity_p[NR_CPUS];
198 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
201 #ifdef CONFIG_RT_GROUP_SCHED
202 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
203 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
205 static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
206 static struct rt_rq *init_rt_rq_p[NR_CPUS];
209 /* task_group_lock serializes add/remove of task groups and also changes to
210 * a task group's cpu shares.
212 static DEFINE_SPINLOCK(task_group_lock);
214 /* doms_cur_mutex serializes access to doms_cur[] array */
215 static DEFINE_MUTEX(doms_cur_mutex);
217 #ifdef CONFIG_FAIR_GROUP_SCHED
218 #ifdef CONFIG_USER_SCHED
219 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
221 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
224 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
227 /* Default task group.
228 * Every task in system belong to this group at bootup.
230 struct task_group init_task_group = {
231 #ifdef CONFIG_FAIR_GROUP_SCHED
232 .se = init_sched_entity_p,
233 .cfs_rq = init_cfs_rq_p,
236 #ifdef CONFIG_RT_GROUP_SCHED
237 .rt_se = init_sched_rt_entity_p,
238 .rt_rq = init_rt_rq_p,
242 /* return group to which a task belongs */
243 static inline struct task_group *task_group(struct task_struct *p)
245 struct task_group *tg;
247 #ifdef CONFIG_USER_SCHED
249 #elif defined(CONFIG_CGROUP_SCHED)
250 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
251 struct task_group, css);
253 tg = &init_task_group;
258 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
259 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
261 #ifdef CONFIG_FAIR_GROUP_SCHED
262 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
263 p->se.parent = task_group(p)->se[cpu];
266 #ifdef CONFIG_RT_GROUP_SCHED
267 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
268 p->rt.parent = task_group(p)->rt_se[cpu];
272 static inline void lock_doms_cur(void)
274 mutex_lock(&doms_cur_mutex);
277 static inline void unlock_doms_cur(void)
279 mutex_unlock(&doms_cur_mutex);
284 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
285 static inline void lock_doms_cur(void) { }
286 static inline void unlock_doms_cur(void) { }
288 #endif /* CONFIG_GROUP_SCHED */
290 /* CFS-related fields in a runqueue */
292 struct load_weight load;
293 unsigned long nr_running;
298 struct rb_root tasks_timeline;
299 struct rb_node *rb_leftmost;
300 struct rb_node *rb_load_balance_curr;
301 /* 'curr' points to currently running entity on this cfs_rq.
302 * It is set to NULL otherwise (i.e when none are currently running).
304 struct sched_entity *curr, *next;
306 unsigned long nr_spread_over;
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
312 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
313 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
314 * (like users, containers etc.)
316 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
317 * list is used during load balance.
319 struct list_head leaf_cfs_rq_list;
320 struct task_group *tg; /* group that "owns" this runqueue */
324 /* Real-Time classes' related field in a runqueue: */
326 struct rt_prio_array active;
327 unsigned long rt_nr_running;
328 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
329 int highest_prio; /* highest queued rt task prio */
332 unsigned long rt_nr_migratory;
338 #ifdef CONFIG_RT_GROUP_SCHED
339 unsigned long rt_nr_boosted;
342 struct list_head leaf_rt_rq_list;
343 struct task_group *tg;
344 struct sched_rt_entity *rt_se;
351 * We add the notion of a root-domain which will be used to define per-domain
352 * variables. Each exclusive cpuset essentially defines an island domain by
353 * fully partitioning the member cpus from any other cpuset. Whenever a new
354 * exclusive cpuset is created, we also create and attach a new root-domain
364 * The "RT overload" flag: it gets set if a CPU has more than
365 * one runnable RT task.
372 * By default the system creates a single root-domain with all cpus as
373 * members (mimicking the global state we have today).
375 static struct root_domain def_root_domain;
380 * This is the main, per-CPU runqueue data structure.
382 * Locking rule: those places that want to lock multiple runqueues
383 * (such as the load balancing or the thread migration code), lock
384 * acquire operations must be ordered by ascending &runqueue.
391 * nr_running and cpu_load should be in the same cacheline because
392 * remote CPUs use both these fields when doing load calculation.
394 unsigned long nr_running;
395 #define CPU_LOAD_IDX_MAX 5
396 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
397 unsigned char idle_at_tick;
399 unsigned char in_nohz_recently;
401 /* capture load from *all* tasks on this cpu: */
402 struct load_weight load;
403 unsigned long nr_load_updates;
408 u64 rt_period_expire;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* list of leaf cfs_rq on this cpu: */
413 struct list_head leaf_cfs_rq_list;
415 #ifdef CONFIG_RT_GROUP_SCHED
416 struct list_head leaf_rt_rq_list;
420 * This is part of a global counter where only the total sum
421 * over all CPUs matters. A task can increase this counter on
422 * one CPU and if it got migrated afterwards it may decrease
423 * it on another CPU. Always updated under the runqueue lock:
425 unsigned long nr_uninterruptible;
427 struct task_struct *curr, *idle;
428 unsigned long next_balance;
429 struct mm_struct *prev_mm;
431 u64 clock, prev_clock_raw;
434 unsigned int clock_warps, clock_overflows, clock_underflows;
436 unsigned int clock_deep_idle_events;
442 struct root_domain *rd;
443 struct sched_domain *sd;
445 /* For active balancing */
448 /* cpu of this runqueue: */
451 struct task_struct *migration_thread;
452 struct list_head migration_queue;
455 #ifdef CONFIG_SCHED_HRTICK
456 unsigned long hrtick_flags;
457 ktime_t hrtick_expire;
458 struct hrtimer hrtick_timer;
461 #ifdef CONFIG_SCHEDSTATS
463 struct sched_info rq_sched_info;
465 /* sys_sched_yield() stats */
466 unsigned int yld_exp_empty;
467 unsigned int yld_act_empty;
468 unsigned int yld_both_empty;
469 unsigned int yld_count;
471 /* schedule() stats */
472 unsigned int sched_switch;
473 unsigned int sched_count;
474 unsigned int sched_goidle;
476 /* try_to_wake_up() stats */
477 unsigned int ttwu_count;
478 unsigned int ttwu_local;
481 unsigned int bkl_count;
483 struct lock_class_key rq_lock_key;
486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
488 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
490 rq->curr->sched_class->check_preempt_curr(rq, p);
493 static inline int cpu_of(struct rq *rq)
503 * Update the per-runqueue clock, as finegrained as the platform can give
504 * us, but without assuming monotonicity, etc.:
506 static void __update_rq_clock(struct rq *rq)
508 u64 prev_raw = rq->prev_clock_raw;
509 u64 now = sched_clock();
510 s64 delta = now - prev_raw;
511 u64 clock = rq->clock;
513 #ifdef CONFIG_SCHED_DEBUG
514 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
517 * Protect against sched_clock() occasionally going backwards:
519 if (unlikely(delta < 0)) {
524 * Catch too large forward jumps too:
526 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
527 if (clock < rq->tick_timestamp + TICK_NSEC)
528 clock = rq->tick_timestamp + TICK_NSEC;
531 rq->clock_overflows++;
533 if (unlikely(delta > rq->clock_max_delta))
534 rq->clock_max_delta = delta;
539 rq->prev_clock_raw = now;
543 static void update_rq_clock(struct rq *rq)
545 if (likely(smp_processor_id() == cpu_of(rq)))
546 __update_rq_clock(rq);
550 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
551 * See detach_destroy_domains: synchronize_sched for details.
553 * The domain tree of any CPU may only be accessed from within
554 * preempt-disabled sections.
556 #define for_each_domain(cpu, __sd) \
557 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
559 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
560 #define this_rq() (&__get_cpu_var(runqueues))
561 #define task_rq(p) cpu_rq(task_cpu(p))
562 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
564 unsigned long rt_needs_cpu(int cpu)
566 struct rq *rq = cpu_rq(cpu);
569 if (!rq->rt_throttled)
572 if (rq->clock > rq->rt_period_expire)
575 delta = rq->rt_period_expire - rq->clock;
576 do_div(delta, NSEC_PER_SEC / HZ);
578 return (unsigned long)delta;
582 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
584 #ifdef CONFIG_SCHED_DEBUG
585 # define const_debug __read_mostly
587 # define const_debug static const
591 * Debugging: various feature bits
594 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
595 SCHED_FEAT_WAKEUP_PREEMPT = 2,
596 SCHED_FEAT_START_DEBIT = 4,
597 SCHED_FEAT_HRTICK = 8,
598 SCHED_FEAT_DOUBLE_TICK = 16,
601 const_debug unsigned int sysctl_sched_features =
602 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
603 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
604 SCHED_FEAT_START_DEBIT * 1 |
605 SCHED_FEAT_HRTICK * 1 |
606 SCHED_FEAT_DOUBLE_TICK * 0;
608 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
611 * Number of tasks to iterate in a single balance run.
612 * Limited because this is done with IRQs disabled.
614 const_debug unsigned int sysctl_sched_nr_migrate = 32;
617 * period over which we measure -rt task cpu usage in us.
620 unsigned int sysctl_sched_rt_period = 1000000;
622 static __read_mostly int scheduler_running;
625 * part of the period that we allow rt tasks to run in us.
628 int sysctl_sched_rt_runtime = 950000;
631 * single value that denotes runtime == period, ie unlimited time.
633 #define RUNTIME_INF ((u64)~0ULL)
636 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
637 * clock constructed from sched_clock():
639 unsigned long long cpu_clock(int cpu)
641 unsigned long long now;
646 * Only call sched_clock() if the scheduler has already been
647 * initialized (some code might call cpu_clock() very early):
649 if (unlikely(!scheduler_running))
652 local_irq_save(flags);
656 local_irq_restore(flags);
660 EXPORT_SYMBOL_GPL(cpu_clock);
662 #ifndef prepare_arch_switch
663 # define prepare_arch_switch(next) do { } while (0)
665 #ifndef finish_arch_switch
666 # define finish_arch_switch(prev) do { } while (0)
669 static inline int task_current(struct rq *rq, struct task_struct *p)
671 return rq->curr == p;
674 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
675 static inline int task_running(struct rq *rq, struct task_struct *p)
677 return task_current(rq, p);
680 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
684 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
686 #ifdef CONFIG_DEBUG_SPINLOCK
687 /* this is a valid case when another task releases the spinlock */
688 rq->lock.owner = current;
691 * If we are tracking spinlock dependencies then we have to
692 * fix up the runqueue lock - which gets 'carried over' from
695 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
697 spin_unlock_irq(&rq->lock);
700 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
701 static inline int task_running(struct rq *rq, struct task_struct *p)
706 return task_current(rq, p);
710 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
714 * We can optimise this out completely for !SMP, because the
715 * SMP rebalancing from interrupt is the only thing that cares
720 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
721 spin_unlock_irq(&rq->lock);
723 spin_unlock(&rq->lock);
727 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
731 * After ->oncpu is cleared, the task can be moved to a different CPU.
732 * We must ensure this doesn't happen until the switch is completely
738 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
742 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
745 * __task_rq_lock - lock the runqueue a given task resides on.
746 * Must be called interrupts disabled.
748 static inline struct rq *__task_rq_lock(struct task_struct *p)
752 struct rq *rq = task_rq(p);
753 spin_lock(&rq->lock);
754 if (likely(rq == task_rq(p)))
756 spin_unlock(&rq->lock);
761 * task_rq_lock - lock the runqueue a given task resides on and disable
762 * interrupts. Note the ordering: we can safely lookup the task_rq without
763 * explicitly disabling preemption.
765 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
771 local_irq_save(*flags);
773 spin_lock(&rq->lock);
774 if (likely(rq == task_rq(p)))
776 spin_unlock_irqrestore(&rq->lock, *flags);
780 static void __task_rq_unlock(struct rq *rq)
783 spin_unlock(&rq->lock);
786 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
789 spin_unlock_irqrestore(&rq->lock, *flags);
793 * this_rq_lock - lock this runqueue and disable interrupts.
795 static struct rq *this_rq_lock(void)
802 spin_lock(&rq->lock);
808 * We are going deep-idle (irqs are disabled):
810 void sched_clock_idle_sleep_event(void)
812 struct rq *rq = cpu_rq(smp_processor_id());
814 spin_lock(&rq->lock);
815 __update_rq_clock(rq);
816 spin_unlock(&rq->lock);
817 rq->clock_deep_idle_events++;
819 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
822 * We just idled delta nanoseconds (called with irqs disabled):
824 void sched_clock_idle_wakeup_event(u64 delta_ns)
826 struct rq *rq = cpu_rq(smp_processor_id());
827 u64 now = sched_clock();
829 rq->idle_clock += delta_ns;
831 * Override the previous timestamp and ignore all
832 * sched_clock() deltas that occured while we idled,
833 * and use the PM-provided delta_ns to advance the
836 spin_lock(&rq->lock);
837 rq->prev_clock_raw = now;
838 rq->clock += delta_ns;
839 spin_unlock(&rq->lock);
840 touch_softlockup_watchdog();
842 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
844 static void __resched_task(struct task_struct *p, int tif_bit);
846 static inline void resched_task(struct task_struct *p)
848 __resched_task(p, TIF_NEED_RESCHED);
851 #ifdef CONFIG_SCHED_HRTICK
853 * Use HR-timers to deliver accurate preemption points.
855 * Its all a bit involved since we cannot program an hrt while holding the
856 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
859 * When we get rescheduled we reprogram the hrtick_timer outside of the
862 static inline void resched_hrt(struct task_struct *p)
864 __resched_task(p, TIF_HRTICK_RESCHED);
867 static inline void resched_rq(struct rq *rq)
871 spin_lock_irqsave(&rq->lock, flags);
872 resched_task(rq->curr);
873 spin_unlock_irqrestore(&rq->lock, flags);
877 HRTICK_SET, /* re-programm hrtick_timer */
878 HRTICK_RESET, /* not a new slice */
883 * - enabled by features
884 * - hrtimer is actually high res
886 static inline int hrtick_enabled(struct rq *rq)
888 if (!sched_feat(HRTICK))
890 return hrtimer_is_hres_active(&rq->hrtick_timer);
894 * Called to set the hrtick timer state.
896 * called with rq->lock held and irqs disabled
898 static void hrtick_start(struct rq *rq, u64 delay, int reset)
900 assert_spin_locked(&rq->lock);
903 * preempt at: now + delay
906 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
908 * indicate we need to program the timer
910 __set_bit(HRTICK_SET, &rq->hrtick_flags);
912 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
915 * New slices are called from the schedule path and don't need a
919 resched_hrt(rq->curr);
922 static void hrtick_clear(struct rq *rq)
924 if (hrtimer_active(&rq->hrtick_timer))
925 hrtimer_cancel(&rq->hrtick_timer);
929 * Update the timer from the possible pending state.
931 static void hrtick_set(struct rq *rq)
937 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
939 spin_lock_irqsave(&rq->lock, flags);
940 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
941 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
942 time = rq->hrtick_expire;
943 clear_thread_flag(TIF_HRTICK_RESCHED);
944 spin_unlock_irqrestore(&rq->lock, flags);
947 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
948 if (reset && !hrtimer_active(&rq->hrtick_timer))
955 * High-resolution timer tick.
956 * Runs from hardirq context with interrupts disabled.
958 static enum hrtimer_restart hrtick(struct hrtimer *timer)
960 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
962 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
964 spin_lock(&rq->lock);
965 __update_rq_clock(rq);
966 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
967 spin_unlock(&rq->lock);
969 return HRTIMER_NORESTART;
972 static inline void init_rq_hrtick(struct rq *rq)
974 rq->hrtick_flags = 0;
975 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
976 rq->hrtick_timer.function = hrtick;
977 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
980 void hrtick_resched(void)
985 if (!test_thread_flag(TIF_HRTICK_RESCHED))
988 local_irq_save(flags);
989 rq = cpu_rq(smp_processor_id());
991 local_irq_restore(flags);
994 static inline void hrtick_clear(struct rq *rq)
998 static inline void hrtick_set(struct rq *rq)
1002 static inline void init_rq_hrtick(struct rq *rq)
1006 void hrtick_resched(void)
1012 * resched_task - mark a task 'to be rescheduled now'.
1014 * On UP this means the setting of the need_resched flag, on SMP it
1015 * might also involve a cross-CPU call to trigger the scheduler on
1020 #ifndef tsk_is_polling
1021 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1024 static void __resched_task(struct task_struct *p, int tif_bit)
1028 assert_spin_locked(&task_rq(p)->lock);
1030 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1033 set_tsk_thread_flag(p, tif_bit);
1036 if (cpu == smp_processor_id())
1039 /* NEED_RESCHED must be visible before we test polling */
1041 if (!tsk_is_polling(p))
1042 smp_send_reschedule(cpu);
1045 static void resched_cpu(int cpu)
1047 struct rq *rq = cpu_rq(cpu);
1048 unsigned long flags;
1050 if (!spin_trylock_irqsave(&rq->lock, flags))
1052 resched_task(cpu_curr(cpu));
1053 spin_unlock_irqrestore(&rq->lock, flags);
1058 * When add_timer_on() enqueues a timer into the timer wheel of an
1059 * idle CPU then this timer might expire before the next timer event
1060 * which is scheduled to wake up that CPU. In case of a completely
1061 * idle system the next event might even be infinite time into the
1062 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1063 * leaves the inner idle loop so the newly added timer is taken into
1064 * account when the CPU goes back to idle and evaluates the timer
1065 * wheel for the next timer event.
1067 void wake_up_idle_cpu(int cpu)
1069 struct rq *rq = cpu_rq(cpu);
1071 if (cpu == smp_processor_id())
1075 * This is safe, as this function is called with the timer
1076 * wheel base lock of (cpu) held. When the CPU is on the way
1077 * to idle and has not yet set rq->curr to idle then it will
1078 * be serialized on the timer wheel base lock and take the new
1079 * timer into account automatically.
1081 if (rq->curr != rq->idle)
1085 * We can set TIF_RESCHED on the idle task of the other CPU
1086 * lockless. The worst case is that the other CPU runs the
1087 * idle task through an additional NOOP schedule()
1089 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1091 /* NEED_RESCHED must be visible before we test polling */
1093 if (!tsk_is_polling(rq->idle))
1094 smp_send_reschedule(cpu);
1099 static void __resched_task(struct task_struct *p, int tif_bit)
1101 assert_spin_locked(&task_rq(p)->lock);
1102 set_tsk_thread_flag(p, tif_bit);
1106 #if BITS_PER_LONG == 32
1107 # define WMULT_CONST (~0UL)
1109 # define WMULT_CONST (1UL << 32)
1112 #define WMULT_SHIFT 32
1115 * Shift right and round:
1117 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1119 static unsigned long
1120 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1121 struct load_weight *lw)
1125 if (unlikely(!lw->inv_weight))
1126 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1128 tmp = (u64)delta_exec * weight;
1130 * Check whether we'd overflow the 64-bit multiplication:
1132 if (unlikely(tmp > WMULT_CONST))
1133 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1136 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1138 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1141 static inline unsigned long
1142 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1144 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1147 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1153 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1160 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1161 * of tasks with abnormal "nice" values across CPUs the contribution that
1162 * each task makes to its run queue's load is weighted according to its
1163 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1164 * scaled version of the new time slice allocation that they receive on time
1168 #define WEIGHT_IDLEPRIO 2
1169 #define WMULT_IDLEPRIO (1 << 31)
1172 * Nice levels are multiplicative, with a gentle 10% change for every
1173 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1174 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1175 * that remained on nice 0.
1177 * The "10% effect" is relative and cumulative: from _any_ nice level,
1178 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1179 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1180 * If a task goes up by ~10% and another task goes down by ~10% then
1181 * the relative distance between them is ~25%.)
1183 static const int prio_to_weight[40] = {
1184 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1185 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1186 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1187 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1188 /* 0 */ 1024, 820, 655, 526, 423,
1189 /* 5 */ 335, 272, 215, 172, 137,
1190 /* 10 */ 110, 87, 70, 56, 45,
1191 /* 15 */ 36, 29, 23, 18, 15,
1195 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1197 * In cases where the weight does not change often, we can use the
1198 * precalculated inverse to speed up arithmetics by turning divisions
1199 * into multiplications:
1201 static const u32 prio_to_wmult[40] = {
1202 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1203 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1204 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1205 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1206 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1207 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1208 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1209 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1212 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1215 * runqueue iterator, to support SMP load-balancing between different
1216 * scheduling classes, without having to expose their internal data
1217 * structures to the load-balancing proper:
1219 struct rq_iterator {
1221 struct task_struct *(*start)(void *);
1222 struct task_struct *(*next)(void *);
1226 static unsigned long
1227 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1228 unsigned long max_load_move, struct sched_domain *sd,
1229 enum cpu_idle_type idle, int *all_pinned,
1230 int *this_best_prio, struct rq_iterator *iterator);
1233 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1234 struct sched_domain *sd, enum cpu_idle_type idle,
1235 struct rq_iterator *iterator);
1238 #ifdef CONFIG_CGROUP_CPUACCT
1239 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1241 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1245 static unsigned long source_load(int cpu, int type);
1246 static unsigned long target_load(int cpu, int type);
1247 static unsigned long cpu_avg_load_per_task(int cpu);
1248 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1249 #endif /* CONFIG_SMP */
1251 #include "sched_stats.h"
1252 #include "sched_idletask.c"
1253 #include "sched_fair.c"
1254 #include "sched_rt.c"
1255 #ifdef CONFIG_SCHED_DEBUG
1256 # include "sched_debug.c"
1259 #define sched_class_highest (&rt_sched_class)
1261 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1263 update_load_add(&rq->load, p->se.load.weight);
1266 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1268 update_load_sub(&rq->load, p->se.load.weight);
1271 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1277 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1283 static void set_load_weight(struct task_struct *p)
1285 if (task_has_rt_policy(p)) {
1286 p->se.load.weight = prio_to_weight[0] * 2;
1287 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1292 * SCHED_IDLE tasks get minimal weight:
1294 if (p->policy == SCHED_IDLE) {
1295 p->se.load.weight = WEIGHT_IDLEPRIO;
1296 p->se.load.inv_weight = WMULT_IDLEPRIO;
1300 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1301 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1304 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1306 sched_info_queued(p);
1307 p->sched_class->enqueue_task(rq, p, wakeup);
1311 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1313 p->sched_class->dequeue_task(rq, p, sleep);
1318 * __normal_prio - return the priority that is based on the static prio
1320 static inline int __normal_prio(struct task_struct *p)
1322 return p->static_prio;
1326 * Calculate the expected normal priority: i.e. priority
1327 * without taking RT-inheritance into account. Might be
1328 * boosted by interactivity modifiers. Changes upon fork,
1329 * setprio syscalls, and whenever the interactivity
1330 * estimator recalculates.
1332 static inline int normal_prio(struct task_struct *p)
1336 if (task_has_rt_policy(p))
1337 prio = MAX_RT_PRIO-1 - p->rt_priority;
1339 prio = __normal_prio(p);
1344 * Calculate the current priority, i.e. the priority
1345 * taken into account by the scheduler. This value might
1346 * be boosted by RT tasks, or might be boosted by
1347 * interactivity modifiers. Will be RT if the task got
1348 * RT-boosted. If not then it returns p->normal_prio.
1350 static int effective_prio(struct task_struct *p)
1352 p->normal_prio = normal_prio(p);
1354 * If we are RT tasks or we were boosted to RT priority,
1355 * keep the priority unchanged. Otherwise, update priority
1356 * to the normal priority:
1358 if (!rt_prio(p->prio))
1359 return p->normal_prio;
1364 * activate_task - move a task to the runqueue.
1366 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1368 if (task_contributes_to_load(p))
1369 rq->nr_uninterruptible--;
1371 enqueue_task(rq, p, wakeup);
1372 inc_nr_running(p, rq);
1376 * deactivate_task - remove a task from the runqueue.
1378 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1380 if (task_contributes_to_load(p))
1381 rq->nr_uninterruptible++;
1383 dequeue_task(rq, p, sleep);
1384 dec_nr_running(p, rq);
1388 * task_curr - is this task currently executing on a CPU?
1389 * @p: the task in question.
1391 inline int task_curr(const struct task_struct *p)
1393 return cpu_curr(task_cpu(p)) == p;
1396 /* Used instead of source_load when we know the type == 0 */
1397 unsigned long weighted_cpuload(const int cpu)
1399 return cpu_rq(cpu)->load.weight;
1402 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1404 set_task_rq(p, cpu);
1407 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1408 * successfuly executed on another CPU. We must ensure that updates of
1409 * per-task data have been completed by this moment.
1412 task_thread_info(p)->cpu = cpu;
1416 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1417 const struct sched_class *prev_class,
1418 int oldprio, int running)
1420 if (prev_class != p->sched_class) {
1421 if (prev_class->switched_from)
1422 prev_class->switched_from(rq, p, running);
1423 p->sched_class->switched_to(rq, p, running);
1425 p->sched_class->prio_changed(rq, p, oldprio, running);
1431 * Is this task likely cache-hot:
1434 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1439 * Buddy candidates are cache hot:
1441 if (&p->se == cfs_rq_of(&p->se)->next)
1444 if (p->sched_class != &fair_sched_class)
1447 if (sysctl_sched_migration_cost == -1)
1449 if (sysctl_sched_migration_cost == 0)
1452 delta = now - p->se.exec_start;
1454 return delta < (s64)sysctl_sched_migration_cost;
1458 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1460 int old_cpu = task_cpu(p);
1461 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1462 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1463 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1466 clock_offset = old_rq->clock - new_rq->clock;
1468 #ifdef CONFIG_SCHEDSTATS
1469 if (p->se.wait_start)
1470 p->se.wait_start -= clock_offset;
1471 if (p->se.sleep_start)
1472 p->se.sleep_start -= clock_offset;
1473 if (p->se.block_start)
1474 p->se.block_start -= clock_offset;
1475 if (old_cpu != new_cpu) {
1476 schedstat_inc(p, se.nr_migrations);
1477 if (task_hot(p, old_rq->clock, NULL))
1478 schedstat_inc(p, se.nr_forced2_migrations);
1481 p->se.vruntime -= old_cfsrq->min_vruntime -
1482 new_cfsrq->min_vruntime;
1484 __set_task_cpu(p, new_cpu);
1487 struct migration_req {
1488 struct list_head list;
1490 struct task_struct *task;
1493 struct completion done;
1497 * The task's runqueue lock must be held.
1498 * Returns true if you have to wait for migration thread.
1501 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1503 struct rq *rq = task_rq(p);
1506 * If the task is not on a runqueue (and not running), then
1507 * it is sufficient to simply update the task's cpu field.
1509 if (!p->se.on_rq && !task_running(rq, p)) {
1510 set_task_cpu(p, dest_cpu);
1514 init_completion(&req->done);
1516 req->dest_cpu = dest_cpu;
1517 list_add(&req->list, &rq->migration_queue);
1523 * wait_task_inactive - wait for a thread to unschedule.
1525 * The caller must ensure that the task *will* unschedule sometime soon,
1526 * else this function might spin for a *long* time. This function can't
1527 * be called with interrupts off, or it may introduce deadlock with
1528 * smp_call_function() if an IPI is sent by the same process we are
1529 * waiting to become inactive.
1531 void wait_task_inactive(struct task_struct *p)
1533 unsigned long flags;
1539 * We do the initial early heuristics without holding
1540 * any task-queue locks at all. We'll only try to get
1541 * the runqueue lock when things look like they will
1547 * If the task is actively running on another CPU
1548 * still, just relax and busy-wait without holding
1551 * NOTE! Since we don't hold any locks, it's not
1552 * even sure that "rq" stays as the right runqueue!
1553 * But we don't care, since "task_running()" will
1554 * return false if the runqueue has changed and p
1555 * is actually now running somewhere else!
1557 while (task_running(rq, p))
1561 * Ok, time to look more closely! We need the rq
1562 * lock now, to be *sure*. If we're wrong, we'll
1563 * just go back and repeat.
1565 rq = task_rq_lock(p, &flags);
1566 running = task_running(rq, p);
1567 on_rq = p->se.on_rq;
1568 task_rq_unlock(rq, &flags);
1571 * Was it really running after all now that we
1572 * checked with the proper locks actually held?
1574 * Oops. Go back and try again..
1576 if (unlikely(running)) {
1582 * It's not enough that it's not actively running,
1583 * it must be off the runqueue _entirely_, and not
1586 * So if it wa still runnable (but just not actively
1587 * running right now), it's preempted, and we should
1588 * yield - it could be a while.
1590 if (unlikely(on_rq)) {
1591 schedule_timeout_uninterruptible(1);
1596 * Ahh, all good. It wasn't running, and it wasn't
1597 * runnable, which means that it will never become
1598 * running in the future either. We're all done!
1605 * kick_process - kick a running thread to enter/exit the kernel
1606 * @p: the to-be-kicked thread
1608 * Cause a process which is running on another CPU to enter
1609 * kernel-mode, without any delay. (to get signals handled.)
1611 * NOTE: this function doesnt have to take the runqueue lock,
1612 * because all it wants to ensure is that the remote task enters
1613 * the kernel. If the IPI races and the task has been migrated
1614 * to another CPU then no harm is done and the purpose has been
1617 void kick_process(struct task_struct *p)
1623 if ((cpu != smp_processor_id()) && task_curr(p))
1624 smp_send_reschedule(cpu);
1629 * Return a low guess at the load of a migration-source cpu weighted
1630 * according to the scheduling class and "nice" value.
1632 * We want to under-estimate the load of migration sources, to
1633 * balance conservatively.
1635 static unsigned long source_load(int cpu, int type)
1637 struct rq *rq = cpu_rq(cpu);
1638 unsigned long total = weighted_cpuload(cpu);
1643 return min(rq->cpu_load[type-1], total);
1647 * Return a high guess at the load of a migration-target cpu weighted
1648 * according to the scheduling class and "nice" value.
1650 static unsigned long target_load(int cpu, int type)
1652 struct rq *rq = cpu_rq(cpu);
1653 unsigned long total = weighted_cpuload(cpu);
1658 return max(rq->cpu_load[type-1], total);
1662 * Return the average load per task on the cpu's run queue
1664 static unsigned long cpu_avg_load_per_task(int cpu)
1666 struct rq *rq = cpu_rq(cpu);
1667 unsigned long total = weighted_cpuload(cpu);
1668 unsigned long n = rq->nr_running;
1670 return n ? total / n : SCHED_LOAD_SCALE;
1674 * find_idlest_group finds and returns the least busy CPU group within the
1677 static struct sched_group *
1678 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1680 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1681 unsigned long min_load = ULONG_MAX, this_load = 0;
1682 int load_idx = sd->forkexec_idx;
1683 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1686 unsigned long load, avg_load;
1690 /* Skip over this group if it has no CPUs allowed */
1691 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1694 local_group = cpu_isset(this_cpu, group->cpumask);
1696 /* Tally up the load of all CPUs in the group */
1699 for_each_cpu_mask(i, group->cpumask) {
1700 /* Bias balancing toward cpus of our domain */
1702 load = source_load(i, load_idx);
1704 load = target_load(i, load_idx);
1709 /* Adjust by relative CPU power of the group */
1710 avg_load = sg_div_cpu_power(group,
1711 avg_load * SCHED_LOAD_SCALE);
1714 this_load = avg_load;
1716 } else if (avg_load < min_load) {
1717 min_load = avg_load;
1720 } while (group = group->next, group != sd->groups);
1722 if (!idlest || 100*this_load < imbalance*min_load)
1728 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1731 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1734 unsigned long load, min_load = ULONG_MAX;
1738 /* Traverse only the allowed CPUs */
1739 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1741 for_each_cpu_mask(i, tmp) {
1742 load = weighted_cpuload(i);
1744 if (load < min_load || (load == min_load && i == this_cpu)) {
1754 * sched_balance_self: balance the current task (running on cpu) in domains
1755 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1758 * Balance, ie. select the least loaded group.
1760 * Returns the target CPU number, or the same CPU if no balancing is needed.
1762 * preempt must be disabled.
1764 static int sched_balance_self(int cpu, int flag)
1766 struct task_struct *t = current;
1767 struct sched_domain *tmp, *sd = NULL;
1769 for_each_domain(cpu, tmp) {
1771 * If power savings logic is enabled for a domain, stop there.
1773 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1775 if (tmp->flags & flag)
1781 struct sched_group *group;
1782 int new_cpu, weight;
1784 if (!(sd->flags & flag)) {
1790 group = find_idlest_group(sd, t, cpu);
1796 new_cpu = find_idlest_cpu(group, t, cpu);
1797 if (new_cpu == -1 || new_cpu == cpu) {
1798 /* Now try balancing at a lower domain level of cpu */
1803 /* Now try balancing at a lower domain level of new_cpu */
1806 weight = cpus_weight(span);
1807 for_each_domain(cpu, tmp) {
1808 if (weight <= cpus_weight(tmp->span))
1810 if (tmp->flags & flag)
1813 /* while loop will break here if sd == NULL */
1819 #endif /* CONFIG_SMP */
1822 * try_to_wake_up - wake up a thread
1823 * @p: the to-be-woken-up thread
1824 * @state: the mask of task states that can be woken
1825 * @sync: do a synchronous wakeup?
1827 * Put it on the run-queue if it's not already there. The "current"
1828 * thread is always on the run-queue (except when the actual
1829 * re-schedule is in progress), and as such you're allowed to do
1830 * the simpler "current->state = TASK_RUNNING" to mark yourself
1831 * runnable without the overhead of this.
1833 * returns failure only if the task is already active.
1835 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1837 int cpu, orig_cpu, this_cpu, success = 0;
1838 unsigned long flags;
1843 rq = task_rq_lock(p, &flags);
1844 old_state = p->state;
1845 if (!(old_state & state))
1853 this_cpu = smp_processor_id();
1856 if (unlikely(task_running(rq, p)))
1859 cpu = p->sched_class->select_task_rq(p, sync);
1860 if (cpu != orig_cpu) {
1861 set_task_cpu(p, cpu);
1862 task_rq_unlock(rq, &flags);
1863 /* might preempt at this point */
1864 rq = task_rq_lock(p, &flags);
1865 old_state = p->state;
1866 if (!(old_state & state))
1871 this_cpu = smp_processor_id();
1875 #ifdef CONFIG_SCHEDSTATS
1876 schedstat_inc(rq, ttwu_count);
1877 if (cpu == this_cpu)
1878 schedstat_inc(rq, ttwu_local);
1880 struct sched_domain *sd;
1881 for_each_domain(this_cpu, sd) {
1882 if (cpu_isset(cpu, sd->span)) {
1883 schedstat_inc(sd, ttwu_wake_remote);
1891 #endif /* CONFIG_SMP */
1892 schedstat_inc(p, se.nr_wakeups);
1894 schedstat_inc(p, se.nr_wakeups_sync);
1895 if (orig_cpu != cpu)
1896 schedstat_inc(p, se.nr_wakeups_migrate);
1897 if (cpu == this_cpu)
1898 schedstat_inc(p, se.nr_wakeups_local);
1900 schedstat_inc(p, se.nr_wakeups_remote);
1901 update_rq_clock(rq);
1902 activate_task(rq, p, 1);
1906 check_preempt_curr(rq, p);
1908 p->state = TASK_RUNNING;
1910 if (p->sched_class->task_wake_up)
1911 p->sched_class->task_wake_up(rq, p);
1914 task_rq_unlock(rq, &flags);
1919 int wake_up_process(struct task_struct *p)
1921 return try_to_wake_up(p, TASK_ALL, 0);
1923 EXPORT_SYMBOL(wake_up_process);
1925 int wake_up_state(struct task_struct *p, unsigned int state)
1927 return try_to_wake_up(p, state, 0);
1931 * Perform scheduler related setup for a newly forked process p.
1932 * p is forked by current.
1934 * __sched_fork() is basic setup used by init_idle() too:
1936 static void __sched_fork(struct task_struct *p)
1938 p->se.exec_start = 0;
1939 p->se.sum_exec_runtime = 0;
1940 p->se.prev_sum_exec_runtime = 0;
1941 p->se.last_wakeup = 0;
1942 p->se.avg_overlap = 0;
1944 #ifdef CONFIG_SCHEDSTATS
1945 p->se.wait_start = 0;
1946 p->se.sum_sleep_runtime = 0;
1947 p->se.sleep_start = 0;
1948 p->se.block_start = 0;
1949 p->se.sleep_max = 0;
1950 p->se.block_max = 0;
1952 p->se.slice_max = 0;
1956 INIT_LIST_HEAD(&p->rt.run_list);
1959 #ifdef CONFIG_PREEMPT_NOTIFIERS
1960 INIT_HLIST_HEAD(&p->preempt_notifiers);
1964 * We mark the process as running here, but have not actually
1965 * inserted it onto the runqueue yet. This guarantees that
1966 * nobody will actually run it, and a signal or other external
1967 * event cannot wake it up and insert it on the runqueue either.
1969 p->state = TASK_RUNNING;
1973 * fork()/clone()-time setup:
1975 void sched_fork(struct task_struct *p, int clone_flags)
1977 int cpu = get_cpu();
1982 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1984 set_task_cpu(p, cpu);
1987 * Make sure we do not leak PI boosting priority to the child:
1989 p->prio = current->normal_prio;
1990 if (!rt_prio(p->prio))
1991 p->sched_class = &fair_sched_class;
1993 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1994 if (likely(sched_info_on()))
1995 memset(&p->sched_info, 0, sizeof(p->sched_info));
1997 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2000 #ifdef CONFIG_PREEMPT
2001 /* Want to start with kernel preemption disabled. */
2002 task_thread_info(p)->preempt_count = 1;
2008 * wake_up_new_task - wake up a newly created task for the first time.
2010 * This function will do some initial scheduler statistics housekeeping
2011 * that must be done for every newly created context, then puts the task
2012 * on the runqueue and wakes it.
2014 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2016 unsigned long flags;
2019 rq = task_rq_lock(p, &flags);
2020 BUG_ON(p->state != TASK_RUNNING);
2021 update_rq_clock(rq);
2023 p->prio = effective_prio(p);
2025 if (!p->sched_class->task_new || !current->se.on_rq) {
2026 activate_task(rq, p, 0);
2029 * Let the scheduling class do new task startup
2030 * management (if any):
2032 p->sched_class->task_new(rq, p);
2033 inc_nr_running(p, rq);
2035 check_preempt_curr(rq, p);
2037 if (p->sched_class->task_wake_up)
2038 p->sched_class->task_wake_up(rq, p);
2040 task_rq_unlock(rq, &flags);
2043 #ifdef CONFIG_PREEMPT_NOTIFIERS
2046 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2047 * @notifier: notifier struct to register
2049 void preempt_notifier_register(struct preempt_notifier *notifier)
2051 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2053 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2056 * preempt_notifier_unregister - no longer interested in preemption notifications
2057 * @notifier: notifier struct to unregister
2059 * This is safe to call from within a preemption notifier.
2061 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2063 hlist_del(¬ifier->link);
2065 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2067 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2069 struct preempt_notifier *notifier;
2070 struct hlist_node *node;
2072 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2073 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2077 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2078 struct task_struct *next)
2080 struct preempt_notifier *notifier;
2081 struct hlist_node *node;
2083 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2084 notifier->ops->sched_out(notifier, next);
2089 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2094 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2095 struct task_struct *next)
2102 * prepare_task_switch - prepare to switch tasks
2103 * @rq: the runqueue preparing to switch
2104 * @prev: the current task that is being switched out
2105 * @next: the task we are going to switch to.
2107 * This is called with the rq lock held and interrupts off. It must
2108 * be paired with a subsequent finish_task_switch after the context
2111 * prepare_task_switch sets up locking and calls architecture specific
2115 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2116 struct task_struct *next)
2118 fire_sched_out_preempt_notifiers(prev, next);
2119 prepare_lock_switch(rq, next);
2120 prepare_arch_switch(next);
2124 * finish_task_switch - clean up after a task-switch
2125 * @rq: runqueue associated with task-switch
2126 * @prev: the thread we just switched away from.
2128 * finish_task_switch must be called after the context switch, paired
2129 * with a prepare_task_switch call before the context switch.
2130 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2131 * and do any other architecture-specific cleanup actions.
2133 * Note that we may have delayed dropping an mm in context_switch(). If
2134 * so, we finish that here outside of the runqueue lock. (Doing it
2135 * with the lock held can cause deadlocks; see schedule() for
2138 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2139 __releases(rq->lock)
2141 struct mm_struct *mm = rq->prev_mm;
2147 * A task struct has one reference for the use as "current".
2148 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2149 * schedule one last time. The schedule call will never return, and
2150 * the scheduled task must drop that reference.
2151 * The test for TASK_DEAD must occur while the runqueue locks are
2152 * still held, otherwise prev could be scheduled on another cpu, die
2153 * there before we look at prev->state, and then the reference would
2155 * Manfred Spraul <manfred@colorfullife.com>
2157 prev_state = prev->state;
2158 finish_arch_switch(prev);
2159 finish_lock_switch(rq, prev);
2161 if (current->sched_class->post_schedule)
2162 current->sched_class->post_schedule(rq);
2165 fire_sched_in_preempt_notifiers(current);
2168 if (unlikely(prev_state == TASK_DEAD)) {
2170 * Remove function-return probe instances associated with this
2171 * task and put them back on the free list.
2173 kprobe_flush_task(prev);
2174 put_task_struct(prev);
2179 * schedule_tail - first thing a freshly forked thread must call.
2180 * @prev: the thread we just switched away from.
2182 asmlinkage void schedule_tail(struct task_struct *prev)
2183 __releases(rq->lock)
2185 struct rq *rq = this_rq();
2187 finish_task_switch(rq, prev);
2188 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2189 /* In this case, finish_task_switch does not reenable preemption */
2192 if (current->set_child_tid)
2193 put_user(task_pid_vnr(current), current->set_child_tid);
2197 * context_switch - switch to the new MM and the new
2198 * thread's register state.
2201 context_switch(struct rq *rq, struct task_struct *prev,
2202 struct task_struct *next)
2204 struct mm_struct *mm, *oldmm;
2206 prepare_task_switch(rq, prev, next);
2208 oldmm = prev->active_mm;
2210 * For paravirt, this is coupled with an exit in switch_to to
2211 * combine the page table reload and the switch backend into
2214 arch_enter_lazy_cpu_mode();
2216 if (unlikely(!mm)) {
2217 next->active_mm = oldmm;
2218 atomic_inc(&oldmm->mm_count);
2219 enter_lazy_tlb(oldmm, next);
2221 switch_mm(oldmm, mm, next);
2223 if (unlikely(!prev->mm)) {
2224 prev->active_mm = NULL;
2225 rq->prev_mm = oldmm;
2228 * Since the runqueue lock will be released by the next
2229 * task (which is an invalid locking op but in the case
2230 * of the scheduler it's an obvious special-case), so we
2231 * do an early lockdep release here:
2233 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2234 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2237 /* Here we just switch the register state and the stack. */
2238 switch_to(prev, next, prev);
2242 * this_rq must be evaluated again because prev may have moved
2243 * CPUs since it called schedule(), thus the 'rq' on its stack
2244 * frame will be invalid.
2246 finish_task_switch(this_rq(), prev);
2250 * nr_running, nr_uninterruptible and nr_context_switches:
2252 * externally visible scheduler statistics: current number of runnable
2253 * threads, current number of uninterruptible-sleeping threads, total
2254 * number of context switches performed since bootup.
2256 unsigned long nr_running(void)
2258 unsigned long i, sum = 0;
2260 for_each_online_cpu(i)
2261 sum += cpu_rq(i)->nr_running;
2266 unsigned long nr_uninterruptible(void)
2268 unsigned long i, sum = 0;
2270 for_each_possible_cpu(i)
2271 sum += cpu_rq(i)->nr_uninterruptible;
2274 * Since we read the counters lockless, it might be slightly
2275 * inaccurate. Do not allow it to go below zero though:
2277 if (unlikely((long)sum < 0))
2283 unsigned long long nr_context_switches(void)
2286 unsigned long long sum = 0;
2288 for_each_possible_cpu(i)
2289 sum += cpu_rq(i)->nr_switches;
2294 unsigned long nr_iowait(void)
2296 unsigned long i, sum = 0;
2298 for_each_possible_cpu(i)
2299 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304 unsigned long nr_active(void)
2306 unsigned long i, running = 0, uninterruptible = 0;
2308 for_each_online_cpu(i) {
2309 running += cpu_rq(i)->nr_running;
2310 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2313 if (unlikely((long)uninterruptible < 0))
2314 uninterruptible = 0;
2316 return running + uninterruptible;
2320 * Update rq->cpu_load[] statistics. This function is usually called every
2321 * scheduler tick (TICK_NSEC).
2323 static void update_cpu_load(struct rq *this_rq)
2325 unsigned long this_load = this_rq->load.weight;
2328 this_rq->nr_load_updates++;
2330 /* Update our load: */
2331 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2332 unsigned long old_load, new_load;
2334 /* scale is effectively 1 << i now, and >> i divides by scale */
2336 old_load = this_rq->cpu_load[i];
2337 new_load = this_load;
2339 * Round up the averaging division if load is increasing. This
2340 * prevents us from getting stuck on 9 if the load is 10, for
2343 if (new_load > old_load)
2344 new_load += scale-1;
2345 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2352 * double_rq_lock - safely lock two runqueues
2354 * Note this does not disable interrupts like task_rq_lock,
2355 * you need to do so manually before calling.
2357 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2358 __acquires(rq1->lock)
2359 __acquires(rq2->lock)
2361 BUG_ON(!irqs_disabled());
2363 spin_lock(&rq1->lock);
2364 __acquire(rq2->lock); /* Fake it out ;) */
2367 spin_lock(&rq1->lock);
2368 spin_lock(&rq2->lock);
2370 spin_lock(&rq2->lock);
2371 spin_lock(&rq1->lock);
2374 update_rq_clock(rq1);
2375 update_rq_clock(rq2);
2379 * double_rq_unlock - safely unlock two runqueues
2381 * Note this does not restore interrupts like task_rq_unlock,
2382 * you need to do so manually after calling.
2384 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2385 __releases(rq1->lock)
2386 __releases(rq2->lock)
2388 spin_unlock(&rq1->lock);
2390 spin_unlock(&rq2->lock);
2392 __release(rq2->lock);
2396 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2398 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2399 __releases(this_rq->lock)
2400 __acquires(busiest->lock)
2401 __acquires(this_rq->lock)
2405 if (unlikely(!irqs_disabled())) {
2406 /* printk() doesn't work good under rq->lock */
2407 spin_unlock(&this_rq->lock);
2410 if (unlikely(!spin_trylock(&busiest->lock))) {
2411 if (busiest < this_rq) {
2412 spin_unlock(&this_rq->lock);
2413 spin_lock(&busiest->lock);
2414 spin_lock(&this_rq->lock);
2417 spin_lock(&busiest->lock);
2423 * If dest_cpu is allowed for this process, migrate the task to it.
2424 * This is accomplished by forcing the cpu_allowed mask to only
2425 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2426 * the cpu_allowed mask is restored.
2428 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2430 struct migration_req req;
2431 unsigned long flags;
2434 rq = task_rq_lock(p, &flags);
2435 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2436 || unlikely(cpu_is_offline(dest_cpu)))
2439 /* force the process onto the specified CPU */
2440 if (migrate_task(p, dest_cpu, &req)) {
2441 /* Need to wait for migration thread (might exit: take ref). */
2442 struct task_struct *mt = rq->migration_thread;
2444 get_task_struct(mt);
2445 task_rq_unlock(rq, &flags);
2446 wake_up_process(mt);
2447 put_task_struct(mt);
2448 wait_for_completion(&req.done);
2453 task_rq_unlock(rq, &flags);
2457 * sched_exec - execve() is a valuable balancing opportunity, because at
2458 * this point the task has the smallest effective memory and cache footprint.
2460 void sched_exec(void)
2462 int new_cpu, this_cpu = get_cpu();
2463 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2465 if (new_cpu != this_cpu)
2466 sched_migrate_task(current, new_cpu);
2470 * pull_task - move a task from a remote runqueue to the local runqueue.
2471 * Both runqueues must be locked.
2473 static void pull_task(struct rq *src_rq, struct task_struct *p,
2474 struct rq *this_rq, int this_cpu)
2476 deactivate_task(src_rq, p, 0);
2477 set_task_cpu(p, this_cpu);
2478 activate_task(this_rq, p, 0);
2480 * Note that idle threads have a prio of MAX_PRIO, for this test
2481 * to be always true for them.
2483 check_preempt_curr(this_rq, p);
2487 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2490 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2491 struct sched_domain *sd, enum cpu_idle_type idle,
2495 * We do not migrate tasks that are:
2496 * 1) running (obviously), or
2497 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2498 * 3) are cache-hot on their current CPU.
2500 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2501 schedstat_inc(p, se.nr_failed_migrations_affine);
2506 if (task_running(rq, p)) {
2507 schedstat_inc(p, se.nr_failed_migrations_running);
2512 * Aggressive migration if:
2513 * 1) task is cache cold, or
2514 * 2) too many balance attempts have failed.
2517 if (!task_hot(p, rq->clock, sd) ||
2518 sd->nr_balance_failed > sd->cache_nice_tries) {
2519 #ifdef CONFIG_SCHEDSTATS
2520 if (task_hot(p, rq->clock, sd)) {
2521 schedstat_inc(sd, lb_hot_gained[idle]);
2522 schedstat_inc(p, se.nr_forced_migrations);
2528 if (task_hot(p, rq->clock, sd)) {
2529 schedstat_inc(p, se.nr_failed_migrations_hot);
2535 static unsigned long
2536 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2537 unsigned long max_load_move, struct sched_domain *sd,
2538 enum cpu_idle_type idle, int *all_pinned,
2539 int *this_best_prio, struct rq_iterator *iterator)
2541 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2542 struct task_struct *p;
2543 long rem_load_move = max_load_move;
2545 if (max_load_move == 0)
2551 * Start the load-balancing iterator:
2553 p = iterator->start(iterator->arg);
2555 if (!p || loops++ > sysctl_sched_nr_migrate)
2558 * To help distribute high priority tasks across CPUs we don't
2559 * skip a task if it will be the highest priority task (i.e. smallest
2560 * prio value) on its new queue regardless of its load weight
2562 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2563 SCHED_LOAD_SCALE_FUZZ;
2564 if ((skip_for_load && p->prio >= *this_best_prio) ||
2565 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2566 p = iterator->next(iterator->arg);
2570 pull_task(busiest, p, this_rq, this_cpu);
2572 rem_load_move -= p->se.load.weight;
2575 * We only want to steal up to the prescribed amount of weighted load.
2577 if (rem_load_move > 0) {
2578 if (p->prio < *this_best_prio)
2579 *this_best_prio = p->prio;
2580 p = iterator->next(iterator->arg);
2585 * Right now, this is one of only two places pull_task() is called,
2586 * so we can safely collect pull_task() stats here rather than
2587 * inside pull_task().
2589 schedstat_add(sd, lb_gained[idle], pulled);
2592 *all_pinned = pinned;
2594 return max_load_move - rem_load_move;
2598 * move_tasks tries to move up to max_load_move weighted load from busiest to
2599 * this_rq, as part of a balancing operation within domain "sd".
2600 * Returns 1 if successful and 0 otherwise.
2602 * Called with both runqueues locked.
2604 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2605 unsigned long max_load_move,
2606 struct sched_domain *sd, enum cpu_idle_type idle,
2609 const struct sched_class *class = sched_class_highest;
2610 unsigned long total_load_moved = 0;
2611 int this_best_prio = this_rq->curr->prio;
2615 class->load_balance(this_rq, this_cpu, busiest,
2616 max_load_move - total_load_moved,
2617 sd, idle, all_pinned, &this_best_prio);
2618 class = class->next;
2619 } while (class && max_load_move > total_load_moved);
2621 return total_load_moved > 0;
2625 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2626 struct sched_domain *sd, enum cpu_idle_type idle,
2627 struct rq_iterator *iterator)
2629 struct task_struct *p = iterator->start(iterator->arg);
2633 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2634 pull_task(busiest, p, this_rq, this_cpu);
2636 * Right now, this is only the second place pull_task()
2637 * is called, so we can safely collect pull_task()
2638 * stats here rather than inside pull_task().
2640 schedstat_inc(sd, lb_gained[idle]);
2644 p = iterator->next(iterator->arg);
2651 * move_one_task tries to move exactly one task from busiest to this_rq, as
2652 * part of active balancing operations within "domain".
2653 * Returns 1 if successful and 0 otherwise.
2655 * Called with both runqueues locked.
2657 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2658 struct sched_domain *sd, enum cpu_idle_type idle)
2660 const struct sched_class *class;
2662 for (class = sched_class_highest; class; class = class->next)
2663 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2670 * find_busiest_group finds and returns the busiest CPU group within the
2671 * domain. It calculates and returns the amount of weighted load which
2672 * should be moved to restore balance via the imbalance parameter.
2674 static struct sched_group *
2675 find_busiest_group(struct sched_domain *sd, int this_cpu,
2676 unsigned long *imbalance, enum cpu_idle_type idle,
2677 int *sd_idle, cpumask_t *cpus, int *balance)
2679 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2680 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2681 unsigned long max_pull;
2682 unsigned long busiest_load_per_task, busiest_nr_running;
2683 unsigned long this_load_per_task, this_nr_running;
2684 int load_idx, group_imb = 0;
2685 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2686 int power_savings_balance = 1;
2687 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2688 unsigned long min_nr_running = ULONG_MAX;
2689 struct sched_group *group_min = NULL, *group_leader = NULL;
2692 max_load = this_load = total_load = total_pwr = 0;
2693 busiest_load_per_task = busiest_nr_running = 0;
2694 this_load_per_task = this_nr_running = 0;
2695 if (idle == CPU_NOT_IDLE)
2696 load_idx = sd->busy_idx;
2697 else if (idle == CPU_NEWLY_IDLE)
2698 load_idx = sd->newidle_idx;
2700 load_idx = sd->idle_idx;
2703 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2706 int __group_imb = 0;
2707 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2708 unsigned long sum_nr_running, sum_weighted_load;
2710 local_group = cpu_isset(this_cpu, group->cpumask);
2713 balance_cpu = first_cpu(group->cpumask);
2715 /* Tally up the load of all CPUs in the group */
2716 sum_weighted_load = sum_nr_running = avg_load = 0;
2718 min_cpu_load = ~0UL;
2720 for_each_cpu_mask(i, group->cpumask) {
2723 if (!cpu_isset(i, *cpus))
2728 if (*sd_idle && rq->nr_running)
2731 /* Bias balancing toward cpus of our domain */
2733 if (idle_cpu(i) && !first_idle_cpu) {
2738 load = target_load(i, load_idx);
2740 load = source_load(i, load_idx);
2741 if (load > max_cpu_load)
2742 max_cpu_load = load;
2743 if (min_cpu_load > load)
2744 min_cpu_load = load;
2748 sum_nr_running += rq->nr_running;
2749 sum_weighted_load += weighted_cpuload(i);
2753 * First idle cpu or the first cpu(busiest) in this sched group
2754 * is eligible for doing load balancing at this and above
2755 * domains. In the newly idle case, we will allow all the cpu's
2756 * to do the newly idle load balance.
2758 if (idle != CPU_NEWLY_IDLE && local_group &&
2759 balance_cpu != this_cpu && balance) {
2764 total_load += avg_load;
2765 total_pwr += group->__cpu_power;
2767 /* Adjust by relative CPU power of the group */
2768 avg_load = sg_div_cpu_power(group,
2769 avg_load * SCHED_LOAD_SCALE);
2771 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2774 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2777 this_load = avg_load;
2779 this_nr_running = sum_nr_running;
2780 this_load_per_task = sum_weighted_load;
2781 } else if (avg_load > max_load &&
2782 (sum_nr_running > group_capacity || __group_imb)) {
2783 max_load = avg_load;
2785 busiest_nr_running = sum_nr_running;
2786 busiest_load_per_task = sum_weighted_load;
2787 group_imb = __group_imb;
2790 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2792 * Busy processors will not participate in power savings
2795 if (idle == CPU_NOT_IDLE ||
2796 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2800 * If the local group is idle or completely loaded
2801 * no need to do power savings balance at this domain
2803 if (local_group && (this_nr_running >= group_capacity ||
2805 power_savings_balance = 0;
2808 * If a group is already running at full capacity or idle,
2809 * don't include that group in power savings calculations
2811 if (!power_savings_balance || sum_nr_running >= group_capacity
2816 * Calculate the group which has the least non-idle load.
2817 * This is the group from where we need to pick up the load
2820 if ((sum_nr_running < min_nr_running) ||
2821 (sum_nr_running == min_nr_running &&
2822 first_cpu(group->cpumask) <
2823 first_cpu(group_min->cpumask))) {
2825 min_nr_running = sum_nr_running;
2826 min_load_per_task = sum_weighted_load /
2831 * Calculate the group which is almost near its
2832 * capacity but still has some space to pick up some load
2833 * from other group and save more power
2835 if (sum_nr_running <= group_capacity - 1) {
2836 if (sum_nr_running > leader_nr_running ||
2837 (sum_nr_running == leader_nr_running &&
2838 first_cpu(group->cpumask) >
2839 first_cpu(group_leader->cpumask))) {
2840 group_leader = group;
2841 leader_nr_running = sum_nr_running;
2846 group = group->next;
2847 } while (group != sd->groups);
2849 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2852 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2854 if (this_load >= avg_load ||
2855 100*max_load <= sd->imbalance_pct*this_load)
2858 busiest_load_per_task /= busiest_nr_running;
2860 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2863 * We're trying to get all the cpus to the average_load, so we don't
2864 * want to push ourselves above the average load, nor do we wish to
2865 * reduce the max loaded cpu below the average load, as either of these
2866 * actions would just result in more rebalancing later, and ping-pong
2867 * tasks around. Thus we look for the minimum possible imbalance.
2868 * Negative imbalances (*we* are more loaded than anyone else) will
2869 * be counted as no imbalance for these purposes -- we can't fix that
2870 * by pulling tasks to us. Be careful of negative numbers as they'll
2871 * appear as very large values with unsigned longs.
2873 if (max_load <= busiest_load_per_task)
2877 * In the presence of smp nice balancing, certain scenarios can have
2878 * max load less than avg load(as we skip the groups at or below
2879 * its cpu_power, while calculating max_load..)
2881 if (max_load < avg_load) {
2883 goto small_imbalance;
2886 /* Don't want to pull so many tasks that a group would go idle */
2887 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2889 /* How much load to actually move to equalise the imbalance */
2890 *imbalance = min(max_pull * busiest->__cpu_power,
2891 (avg_load - this_load) * this->__cpu_power)
2895 * if *imbalance is less than the average load per runnable task
2896 * there is no gaurantee that any tasks will be moved so we'll have
2897 * a think about bumping its value to force at least one task to be
2900 if (*imbalance < busiest_load_per_task) {
2901 unsigned long tmp, pwr_now, pwr_move;
2905 pwr_move = pwr_now = 0;
2907 if (this_nr_running) {
2908 this_load_per_task /= this_nr_running;
2909 if (busiest_load_per_task > this_load_per_task)
2912 this_load_per_task = SCHED_LOAD_SCALE;
2914 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2915 busiest_load_per_task * imbn) {
2916 *imbalance = busiest_load_per_task;
2921 * OK, we don't have enough imbalance to justify moving tasks,
2922 * however we may be able to increase total CPU power used by
2926 pwr_now += busiest->__cpu_power *
2927 min(busiest_load_per_task, max_load);
2928 pwr_now += this->__cpu_power *
2929 min(this_load_per_task, this_load);
2930 pwr_now /= SCHED_LOAD_SCALE;
2932 /* Amount of load we'd subtract */
2933 tmp = sg_div_cpu_power(busiest,
2934 busiest_load_per_task * SCHED_LOAD_SCALE);
2936 pwr_move += busiest->__cpu_power *
2937 min(busiest_load_per_task, max_load - tmp);
2939 /* Amount of load we'd add */
2940 if (max_load * busiest->__cpu_power <
2941 busiest_load_per_task * SCHED_LOAD_SCALE)
2942 tmp = sg_div_cpu_power(this,
2943 max_load * busiest->__cpu_power);
2945 tmp = sg_div_cpu_power(this,
2946 busiest_load_per_task * SCHED_LOAD_SCALE);
2947 pwr_move += this->__cpu_power *
2948 min(this_load_per_task, this_load + tmp);
2949 pwr_move /= SCHED_LOAD_SCALE;
2951 /* Move if we gain throughput */
2952 if (pwr_move > pwr_now)
2953 *imbalance = busiest_load_per_task;
2959 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2960 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2963 if (this == group_leader && group_leader != group_min) {
2964 *imbalance = min_load_per_task;
2974 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2977 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2978 unsigned long imbalance, cpumask_t *cpus)
2980 struct rq *busiest = NULL, *rq;
2981 unsigned long max_load = 0;
2984 for_each_cpu_mask(i, group->cpumask) {
2987 if (!cpu_isset(i, *cpus))
2991 wl = weighted_cpuload(i);
2993 if (rq->nr_running == 1 && wl > imbalance)
2996 if (wl > max_load) {
3006 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3007 * so long as it is large enough.
3009 #define MAX_PINNED_INTERVAL 512
3012 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3013 * tasks if there is an imbalance.
3015 static int load_balance(int this_cpu, struct rq *this_rq,
3016 struct sched_domain *sd, enum cpu_idle_type idle,
3019 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3020 struct sched_group *group;
3021 unsigned long imbalance;
3023 cpumask_t cpus = CPU_MASK_ALL;
3024 unsigned long flags;
3027 * When power savings policy is enabled for the parent domain, idle
3028 * sibling can pick up load irrespective of busy siblings. In this case,
3029 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3030 * portraying it as CPU_NOT_IDLE.
3032 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3033 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3036 schedstat_inc(sd, lb_count[idle]);
3039 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3046 schedstat_inc(sd, lb_nobusyg[idle]);
3050 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3052 schedstat_inc(sd, lb_nobusyq[idle]);
3056 BUG_ON(busiest == this_rq);
3058 schedstat_add(sd, lb_imbalance[idle], imbalance);
3061 if (busiest->nr_running > 1) {
3063 * Attempt to move tasks. If find_busiest_group has found
3064 * an imbalance but busiest->nr_running <= 1, the group is
3065 * still unbalanced. ld_moved simply stays zero, so it is
3066 * correctly treated as an imbalance.
3068 local_irq_save(flags);
3069 double_rq_lock(this_rq, busiest);
3070 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3071 imbalance, sd, idle, &all_pinned);
3072 double_rq_unlock(this_rq, busiest);
3073 local_irq_restore(flags);
3076 * some other cpu did the load balance for us.
3078 if (ld_moved && this_cpu != smp_processor_id())
3079 resched_cpu(this_cpu);
3081 /* All tasks on this runqueue were pinned by CPU affinity */
3082 if (unlikely(all_pinned)) {
3083 cpu_clear(cpu_of(busiest), cpus);
3084 if (!cpus_empty(cpus))
3091 schedstat_inc(sd, lb_failed[idle]);
3092 sd->nr_balance_failed++;
3094 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3096 spin_lock_irqsave(&busiest->lock, flags);
3098 /* don't kick the migration_thread, if the curr
3099 * task on busiest cpu can't be moved to this_cpu
3101 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3102 spin_unlock_irqrestore(&busiest->lock, flags);
3104 goto out_one_pinned;
3107 if (!busiest->active_balance) {
3108 busiest->active_balance = 1;
3109 busiest->push_cpu = this_cpu;
3112 spin_unlock_irqrestore(&busiest->lock, flags);
3114 wake_up_process(busiest->migration_thread);
3117 * We've kicked active balancing, reset the failure
3120 sd->nr_balance_failed = sd->cache_nice_tries+1;
3123 sd->nr_balance_failed = 0;
3125 if (likely(!active_balance)) {
3126 /* We were unbalanced, so reset the balancing interval */
3127 sd->balance_interval = sd->min_interval;
3130 * If we've begun active balancing, start to back off. This
3131 * case may not be covered by the all_pinned logic if there
3132 * is only 1 task on the busy runqueue (because we don't call
3135 if (sd->balance_interval < sd->max_interval)
3136 sd->balance_interval *= 2;
3139 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3140 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3145 schedstat_inc(sd, lb_balanced[idle]);
3147 sd->nr_balance_failed = 0;
3150 /* tune up the balancing interval */
3151 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3152 (sd->balance_interval < sd->max_interval))
3153 sd->balance_interval *= 2;
3155 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3156 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3162 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3163 * tasks if there is an imbalance.
3165 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3166 * this_rq is locked.
3169 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3171 struct sched_group *group;
3172 struct rq *busiest = NULL;
3173 unsigned long imbalance;
3177 cpumask_t cpus = CPU_MASK_ALL;
3180 * When power savings policy is enabled for the parent domain, idle
3181 * sibling can pick up load irrespective of busy siblings. In this case,
3182 * let the state of idle sibling percolate up as IDLE, instead of
3183 * portraying it as CPU_NOT_IDLE.
3185 if (sd->flags & SD_SHARE_CPUPOWER &&
3186 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3189 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3191 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3192 &sd_idle, &cpus, NULL);
3194 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3198 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3201 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3205 BUG_ON(busiest == this_rq);
3207 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3210 if (busiest->nr_running > 1) {
3211 /* Attempt to move tasks */
3212 double_lock_balance(this_rq, busiest);
3213 /* this_rq->clock is already updated */
3214 update_rq_clock(busiest);
3215 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3216 imbalance, sd, CPU_NEWLY_IDLE,
3218 spin_unlock(&busiest->lock);
3220 if (unlikely(all_pinned)) {
3221 cpu_clear(cpu_of(busiest), cpus);
3222 if (!cpus_empty(cpus))
3228 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3229 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3230 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3233 sd->nr_balance_failed = 0;
3238 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3239 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3240 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3242 sd->nr_balance_failed = 0;
3248 * idle_balance is called by schedule() if this_cpu is about to become
3249 * idle. Attempts to pull tasks from other CPUs.
3251 static void idle_balance(int this_cpu, struct rq *this_rq)
3253 struct sched_domain *sd;
3254 int pulled_task = -1;
3255 unsigned long next_balance = jiffies + HZ;
3257 for_each_domain(this_cpu, sd) {
3258 unsigned long interval;
3260 if (!(sd->flags & SD_LOAD_BALANCE))
3263 if (sd->flags & SD_BALANCE_NEWIDLE)
3264 /* If we've pulled tasks over stop searching: */
3265 pulled_task = load_balance_newidle(this_cpu,
3268 interval = msecs_to_jiffies(sd->balance_interval);
3269 if (time_after(next_balance, sd->last_balance + interval))
3270 next_balance = sd->last_balance + interval;
3274 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3276 * We are going idle. next_balance may be set based on
3277 * a busy processor. So reset next_balance.
3279 this_rq->next_balance = next_balance;
3284 * active_load_balance is run by migration threads. It pushes running tasks
3285 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3286 * running on each physical CPU where possible, and avoids physical /
3287 * logical imbalances.
3289 * Called with busiest_rq locked.
3291 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3293 int target_cpu = busiest_rq->push_cpu;
3294 struct sched_domain *sd;
3295 struct rq *target_rq;
3297 /* Is there any task to move? */
3298 if (busiest_rq->nr_running <= 1)
3301 target_rq = cpu_rq(target_cpu);
3304 * This condition is "impossible", if it occurs
3305 * we need to fix it. Originally reported by
3306 * Bjorn Helgaas on a 128-cpu setup.
3308 BUG_ON(busiest_rq == target_rq);
3310 /* move a task from busiest_rq to target_rq */
3311 double_lock_balance(busiest_rq, target_rq);
3312 update_rq_clock(busiest_rq);
3313 update_rq_clock(target_rq);
3315 /* Search for an sd spanning us and the target CPU. */
3316 for_each_domain(target_cpu, sd) {
3317 if ((sd->flags & SD_LOAD_BALANCE) &&
3318 cpu_isset(busiest_cpu, sd->span))
3323 schedstat_inc(sd, alb_count);
3325 if (move_one_task(target_rq, target_cpu, busiest_rq,
3327 schedstat_inc(sd, alb_pushed);
3329 schedstat_inc(sd, alb_failed);
3331 spin_unlock(&target_rq->lock);
3336 atomic_t load_balancer;
3338 } nohz ____cacheline_aligned = {
3339 .load_balancer = ATOMIC_INIT(-1),
3340 .cpu_mask = CPU_MASK_NONE,
3344 * This routine will try to nominate the ilb (idle load balancing)
3345 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3346 * load balancing on behalf of all those cpus. If all the cpus in the system
3347 * go into this tickless mode, then there will be no ilb owner (as there is
3348 * no need for one) and all the cpus will sleep till the next wakeup event
3351 * For the ilb owner, tick is not stopped. And this tick will be used
3352 * for idle load balancing. ilb owner will still be part of
3355 * While stopping the tick, this cpu will become the ilb owner if there
3356 * is no other owner. And will be the owner till that cpu becomes busy
3357 * or if all cpus in the system stop their ticks at which point
3358 * there is no need for ilb owner.
3360 * When the ilb owner becomes busy, it nominates another owner, during the
3361 * next busy scheduler_tick()
3363 int select_nohz_load_balancer(int stop_tick)
3365 int cpu = smp_processor_id();
3368 cpu_set(cpu, nohz.cpu_mask);
3369 cpu_rq(cpu)->in_nohz_recently = 1;
3372 * If we are going offline and still the leader, give up!
3374 if (cpu_is_offline(cpu) &&
3375 atomic_read(&nohz.load_balancer) == cpu) {
3376 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3381 /* time for ilb owner also to sleep */
3382 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3383 if (atomic_read(&nohz.load_balancer) == cpu)
3384 atomic_set(&nohz.load_balancer, -1);
3388 if (atomic_read(&nohz.load_balancer) == -1) {
3389 /* make me the ilb owner */
3390 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3392 } else if (atomic_read(&nohz.load_balancer) == cpu)
3395 if (!cpu_isset(cpu, nohz.cpu_mask))
3398 cpu_clear(cpu, nohz.cpu_mask);
3400 if (atomic_read(&nohz.load_balancer) == cpu)
3401 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3408 static DEFINE_SPINLOCK(balancing);
3411 * It checks each scheduling domain to see if it is due to be balanced,
3412 * and initiates a balancing operation if so.
3414 * Balancing parameters are set up in arch_init_sched_domains.
3416 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3419 struct rq *rq = cpu_rq(cpu);
3420 unsigned long interval;
3421 struct sched_domain *sd;
3422 /* Earliest time when we have to do rebalance again */
3423 unsigned long next_balance = jiffies + 60*HZ;
3424 int update_next_balance = 0;
3426 for_each_domain(cpu, sd) {
3427 if (!(sd->flags & SD_LOAD_BALANCE))
3430 interval = sd->balance_interval;
3431 if (idle != CPU_IDLE)
3432 interval *= sd->busy_factor;
3434 /* scale ms to jiffies */
3435 interval = msecs_to_jiffies(interval);
3436 if (unlikely(!interval))
3438 if (interval > HZ*NR_CPUS/10)
3439 interval = HZ*NR_CPUS/10;
3442 if (sd->flags & SD_SERIALIZE) {
3443 if (!spin_trylock(&balancing))
3447 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3448 if (load_balance(cpu, rq, sd, idle, &balance)) {
3450 * We've pulled tasks over so either we're no
3451 * longer idle, or one of our SMT siblings is
3454 idle = CPU_NOT_IDLE;
3456 sd->last_balance = jiffies;
3458 if (sd->flags & SD_SERIALIZE)
3459 spin_unlock(&balancing);
3461 if (time_after(next_balance, sd->last_balance + interval)) {
3462 next_balance = sd->last_balance + interval;
3463 update_next_balance = 1;
3467 * Stop the load balance at this level. There is another
3468 * CPU in our sched group which is doing load balancing more
3476 * next_balance will be updated only when there is a need.
3477 * When the cpu is attached to null domain for ex, it will not be
3480 if (likely(update_next_balance))
3481 rq->next_balance = next_balance;
3485 * run_rebalance_domains is triggered when needed from the scheduler tick.
3486 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3487 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3489 static void run_rebalance_domains(struct softirq_action *h)
3491 int this_cpu = smp_processor_id();
3492 struct rq *this_rq = cpu_rq(this_cpu);
3493 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3494 CPU_IDLE : CPU_NOT_IDLE;
3496 rebalance_domains(this_cpu, idle);
3500 * If this cpu is the owner for idle load balancing, then do the
3501 * balancing on behalf of the other idle cpus whose ticks are
3504 if (this_rq->idle_at_tick &&
3505 atomic_read(&nohz.load_balancer) == this_cpu) {
3506 cpumask_t cpus = nohz.cpu_mask;
3510 cpu_clear(this_cpu, cpus);
3511 for_each_cpu_mask(balance_cpu, cpus) {
3513 * If this cpu gets work to do, stop the load balancing
3514 * work being done for other cpus. Next load
3515 * balancing owner will pick it up.
3520 rebalance_domains(balance_cpu, CPU_IDLE);
3522 rq = cpu_rq(balance_cpu);
3523 if (time_after(this_rq->next_balance, rq->next_balance))
3524 this_rq->next_balance = rq->next_balance;
3531 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3533 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3534 * idle load balancing owner or decide to stop the periodic load balancing,
3535 * if the whole system is idle.
3537 static inline void trigger_load_balance(struct rq *rq, int cpu)
3541 * If we were in the nohz mode recently and busy at the current
3542 * scheduler tick, then check if we need to nominate new idle
3545 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3546 rq->in_nohz_recently = 0;
3548 if (atomic_read(&nohz.load_balancer) == cpu) {
3549 cpu_clear(cpu, nohz.cpu_mask);
3550 atomic_set(&nohz.load_balancer, -1);
3553 if (atomic_read(&nohz.load_balancer) == -1) {
3555 * simple selection for now: Nominate the
3556 * first cpu in the nohz list to be the next
3559 * TBD: Traverse the sched domains and nominate
3560 * the nearest cpu in the nohz.cpu_mask.
3562 int ilb = first_cpu(nohz.cpu_mask);
3570 * If this cpu is idle and doing idle load balancing for all the
3571 * cpus with ticks stopped, is it time for that to stop?
3573 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3574 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3580 * If this cpu is idle and the idle load balancing is done by
3581 * someone else, then no need raise the SCHED_SOFTIRQ
3583 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3584 cpu_isset(cpu, nohz.cpu_mask))
3587 if (time_after_eq(jiffies, rq->next_balance))
3588 raise_softirq(SCHED_SOFTIRQ);
3591 #else /* CONFIG_SMP */
3594 * on UP we do not need to balance between CPUs:
3596 static inline void idle_balance(int cpu, struct rq *rq)
3602 DEFINE_PER_CPU(struct kernel_stat, kstat);
3604 EXPORT_PER_CPU_SYMBOL(kstat);
3607 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3608 * that have not yet been banked in case the task is currently running.
3610 unsigned long long task_sched_runtime(struct task_struct *p)
3612 unsigned long flags;
3616 rq = task_rq_lock(p, &flags);
3617 ns = p->se.sum_exec_runtime;
3618 if (task_current(rq, p)) {
3619 update_rq_clock(rq);
3620 delta_exec = rq->clock - p->se.exec_start;
3621 if ((s64)delta_exec > 0)
3624 task_rq_unlock(rq, &flags);
3630 * Account user cpu time to a process.
3631 * @p: the process that the cpu time gets accounted to
3632 * @cputime: the cpu time spent in user space since the last update
3634 void account_user_time(struct task_struct *p, cputime_t cputime)
3636 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3639 p->utime = cputime_add(p->utime, cputime);
3641 /* Add user time to cpustat. */
3642 tmp = cputime_to_cputime64(cputime);
3643 if (TASK_NICE(p) > 0)
3644 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3646 cpustat->user = cputime64_add(cpustat->user, tmp);
3650 * Account guest cpu time to a process.
3651 * @p: the process that the cpu time gets accounted to
3652 * @cputime: the cpu time spent in virtual machine since the last update
3654 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3657 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3659 tmp = cputime_to_cputime64(cputime);
3661 p->utime = cputime_add(p->utime, cputime);
3662 p->gtime = cputime_add(p->gtime, cputime);
3664 cpustat->user = cputime64_add(cpustat->user, tmp);
3665 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3669 * Account scaled user cpu time to a process.
3670 * @p: the process that the cpu time gets accounted to
3671 * @cputime: the cpu time spent in user space since the last update
3673 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3675 p->utimescaled = cputime_add(p->utimescaled, cputime);
3679 * Account system cpu time to a process.
3680 * @p: the process that the cpu time gets accounted to
3681 * @hardirq_offset: the offset to subtract from hardirq_count()
3682 * @cputime: the cpu time spent in kernel space since the last update
3684 void account_system_time(struct task_struct *p, int hardirq_offset,
3687 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3688 struct rq *rq = this_rq();
3691 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3692 return account_guest_time(p, cputime);
3694 p->stime = cputime_add(p->stime, cputime);
3696 /* Add system time to cpustat. */
3697 tmp = cputime_to_cputime64(cputime);
3698 if (hardirq_count() - hardirq_offset)
3699 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3700 else if (softirq_count())
3701 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3702 else if (p != rq->idle)
3703 cpustat->system = cputime64_add(cpustat->system, tmp);
3704 else if (atomic_read(&rq->nr_iowait) > 0)
3705 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3707 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3708 /* Account for system time used */
3709 acct_update_integrals(p);
3713 * Account scaled system cpu time to a process.
3714 * @p: the process that the cpu time gets accounted to
3715 * @hardirq_offset: the offset to subtract from hardirq_count()
3716 * @cputime: the cpu time spent in kernel space since the last update
3718 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3720 p->stimescaled = cputime_add(p->stimescaled, cputime);
3724 * Account for involuntary wait time.
3725 * @p: the process from which the cpu time has been stolen
3726 * @steal: the cpu time spent in involuntary wait
3728 void account_steal_time(struct task_struct *p, cputime_t steal)
3730 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3731 cputime64_t tmp = cputime_to_cputime64(steal);
3732 struct rq *rq = this_rq();
3734 if (p == rq->idle) {
3735 p->stime = cputime_add(p->stime, steal);
3736 if (atomic_read(&rq->nr_iowait) > 0)
3737 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3739 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3741 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3745 * This function gets called by the timer code, with HZ frequency.
3746 * We call it with interrupts disabled.
3748 * It also gets called by the fork code, when changing the parent's
3751 void scheduler_tick(void)
3753 int cpu = smp_processor_id();
3754 struct rq *rq = cpu_rq(cpu);
3755 struct task_struct *curr = rq->curr;
3756 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3758 spin_lock(&rq->lock);
3759 __update_rq_clock(rq);
3761 * Let rq->clock advance by at least TICK_NSEC:
3763 if (unlikely(rq->clock < next_tick)) {
3764 rq->clock = next_tick;
3765 rq->clock_underflows++;
3767 rq->tick_timestamp = rq->clock;
3768 update_cpu_load(rq);
3769 curr->sched_class->task_tick(rq, curr, 0);
3770 update_sched_rt_period(rq);
3771 spin_unlock(&rq->lock);
3774 rq->idle_at_tick = idle_cpu(cpu);
3775 trigger_load_balance(rq, cpu);
3779 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3781 void __kprobes add_preempt_count(int val)
3786 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3788 preempt_count() += val;
3790 * Spinlock count overflowing soon?
3792 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3795 EXPORT_SYMBOL(add_preempt_count);
3797 void __kprobes sub_preempt_count(int val)
3802 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3805 * Is the spinlock portion underflowing?
3807 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3808 !(preempt_count() & PREEMPT_MASK)))
3811 preempt_count() -= val;
3813 EXPORT_SYMBOL(sub_preempt_count);
3818 * Print scheduling while atomic bug:
3820 static noinline void __schedule_bug(struct task_struct *prev)
3822 struct pt_regs *regs = get_irq_regs();
3824 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3825 prev->comm, prev->pid, preempt_count());
3827 debug_show_held_locks(prev);
3828 if (irqs_disabled())
3829 print_irqtrace_events(prev);
3838 * Various schedule()-time debugging checks and statistics:
3840 static inline void schedule_debug(struct task_struct *prev)
3843 * Test if we are atomic. Since do_exit() needs to call into
3844 * schedule() atomically, we ignore that path for now.
3845 * Otherwise, whine if we are scheduling when we should not be.
3847 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3848 __schedule_bug(prev);
3850 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3852 schedstat_inc(this_rq(), sched_count);
3853 #ifdef CONFIG_SCHEDSTATS
3854 if (unlikely(prev->lock_depth >= 0)) {
3855 schedstat_inc(this_rq(), bkl_count);
3856 schedstat_inc(prev, sched_info.bkl_count);
3862 * Pick up the highest-prio task:
3864 static inline struct task_struct *
3865 pick_next_task(struct rq *rq, struct task_struct *prev)
3867 const struct sched_class *class;
3868 struct task_struct *p;
3871 * Optimization: we know that if all tasks are in
3872 * the fair class we can call that function directly:
3874 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3875 p = fair_sched_class.pick_next_task(rq);
3880 class = sched_class_highest;
3882 p = class->pick_next_task(rq);
3886 * Will never be NULL as the idle class always
3887 * returns a non-NULL p:
3889 class = class->next;
3894 * schedule() is the main scheduler function.
3896 asmlinkage void __sched schedule(void)
3898 struct task_struct *prev, *next;
3899 unsigned long *switch_count;
3905 cpu = smp_processor_id();
3909 switch_count = &prev->nivcsw;
3911 release_kernel_lock(prev);
3912 need_resched_nonpreemptible:
3914 schedule_debug(prev);
3919 * Do the rq-clock update outside the rq lock:
3921 local_irq_disable();
3922 __update_rq_clock(rq);
3923 spin_lock(&rq->lock);
3924 clear_tsk_need_resched(prev);
3926 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3927 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3928 signal_pending(prev))) {
3929 prev->state = TASK_RUNNING;
3931 deactivate_task(rq, prev, 1);
3933 switch_count = &prev->nvcsw;
3937 if (prev->sched_class->pre_schedule)
3938 prev->sched_class->pre_schedule(rq, prev);
3941 if (unlikely(!rq->nr_running))
3942 idle_balance(cpu, rq);
3944 prev->sched_class->put_prev_task(rq, prev);
3945 next = pick_next_task(rq, prev);
3947 sched_info_switch(prev, next);
3949 if (likely(prev != next)) {
3954 context_switch(rq, prev, next); /* unlocks the rq */
3956 * the context switch might have flipped the stack from under
3957 * us, hence refresh the local variables.
3959 cpu = smp_processor_id();
3962 spin_unlock_irq(&rq->lock);
3966 if (unlikely(reacquire_kernel_lock(current) < 0))
3967 goto need_resched_nonpreemptible;
3969 preempt_enable_no_resched();
3970 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3973 EXPORT_SYMBOL(schedule);
3975 #ifdef CONFIG_PREEMPT
3977 * this is the entry point to schedule() from in-kernel preemption
3978 * off of preempt_enable. Kernel preemptions off return from interrupt
3979 * occur there and call schedule directly.
3981 asmlinkage void __sched preempt_schedule(void)
3983 struct thread_info *ti = current_thread_info();
3984 struct task_struct *task = current;
3985 int saved_lock_depth;
3988 * If there is a non-zero preempt_count or interrupts are disabled,
3989 * we do not want to preempt the current task. Just return..
3991 if (likely(ti->preempt_count || irqs_disabled()))
3995 add_preempt_count(PREEMPT_ACTIVE);
3998 * We keep the big kernel semaphore locked, but we
3999 * clear ->lock_depth so that schedule() doesnt
4000 * auto-release the semaphore:
4002 saved_lock_depth = task->lock_depth;
4003 task->lock_depth = -1;
4005 task->lock_depth = saved_lock_depth;
4006 sub_preempt_count(PREEMPT_ACTIVE);
4009 * Check again in case we missed a preemption opportunity
4010 * between schedule and now.
4013 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4015 EXPORT_SYMBOL(preempt_schedule);
4018 * this is the entry point to schedule() from kernel preemption
4019 * off of irq context.
4020 * Note, that this is called and return with irqs disabled. This will
4021 * protect us against recursive calling from irq.
4023 asmlinkage void __sched preempt_schedule_irq(void)
4025 struct thread_info *ti = current_thread_info();
4026 struct task_struct *task = current;
4027 int saved_lock_depth;
4029 /* Catch callers which need to be fixed */
4030 BUG_ON(ti->preempt_count || !irqs_disabled());
4033 add_preempt_count(PREEMPT_ACTIVE);
4036 * We keep the big kernel semaphore locked, but we
4037 * clear ->lock_depth so that schedule() doesnt
4038 * auto-release the semaphore:
4040 saved_lock_depth = task->lock_depth;
4041 task->lock_depth = -1;
4044 local_irq_disable();
4045 task->lock_depth = saved_lock_depth;
4046 sub_preempt_count(PREEMPT_ACTIVE);
4049 * Check again in case we missed a preemption opportunity
4050 * between schedule and now.
4053 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4056 #endif /* CONFIG_PREEMPT */
4058 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4061 return try_to_wake_up(curr->private, mode, sync);
4063 EXPORT_SYMBOL(default_wake_function);
4066 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4067 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4068 * number) then we wake all the non-exclusive tasks and one exclusive task.
4070 * There are circumstances in which we can try to wake a task which has already
4071 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4072 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4074 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4075 int nr_exclusive, int sync, void *key)
4077 wait_queue_t *curr, *next;
4079 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4080 unsigned flags = curr->flags;
4082 if (curr->func(curr, mode, sync, key) &&
4083 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4089 * __wake_up - wake up threads blocked on a waitqueue.
4091 * @mode: which threads
4092 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4093 * @key: is directly passed to the wakeup function
4095 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4096 int nr_exclusive, void *key)
4098 unsigned long flags;
4100 spin_lock_irqsave(&q->lock, flags);
4101 __wake_up_common(q, mode, nr_exclusive, 0, key);
4102 spin_unlock_irqrestore(&q->lock, flags);
4104 EXPORT_SYMBOL(__wake_up);
4107 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4109 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4111 __wake_up_common(q, mode, 1, 0, NULL);
4115 * __wake_up_sync - wake up threads blocked on a waitqueue.
4117 * @mode: which threads
4118 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4120 * The sync wakeup differs that the waker knows that it will schedule
4121 * away soon, so while the target thread will be woken up, it will not
4122 * be migrated to another CPU - ie. the two threads are 'synchronized'
4123 * with each other. This can prevent needless bouncing between CPUs.
4125 * On UP it can prevent extra preemption.
4128 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4130 unsigned long flags;
4136 if (unlikely(!nr_exclusive))
4139 spin_lock_irqsave(&q->lock, flags);
4140 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4141 spin_unlock_irqrestore(&q->lock, flags);
4143 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4145 void complete(struct completion *x)
4147 unsigned long flags;
4149 spin_lock_irqsave(&x->wait.lock, flags);
4151 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4152 spin_unlock_irqrestore(&x->wait.lock, flags);
4154 EXPORT_SYMBOL(complete);
4156 void complete_all(struct completion *x)
4158 unsigned long flags;
4160 spin_lock_irqsave(&x->wait.lock, flags);
4161 x->done += UINT_MAX/2;
4162 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4163 spin_unlock_irqrestore(&x->wait.lock, flags);
4165 EXPORT_SYMBOL(complete_all);
4167 static inline long __sched
4168 do_wait_for_common(struct completion *x, long timeout, int state)
4171 DECLARE_WAITQUEUE(wait, current);
4173 wait.flags |= WQ_FLAG_EXCLUSIVE;
4174 __add_wait_queue_tail(&x->wait, &wait);
4176 if ((state == TASK_INTERRUPTIBLE &&
4177 signal_pending(current)) ||
4178 (state == TASK_KILLABLE &&
4179 fatal_signal_pending(current))) {
4180 __remove_wait_queue(&x->wait, &wait);
4181 return -ERESTARTSYS;
4183 __set_current_state(state);
4184 spin_unlock_irq(&x->wait.lock);
4185 timeout = schedule_timeout(timeout);
4186 spin_lock_irq(&x->wait.lock);
4188 __remove_wait_queue(&x->wait, &wait);
4192 __remove_wait_queue(&x->wait, &wait);
4199 wait_for_common(struct completion *x, long timeout, int state)
4203 spin_lock_irq(&x->wait.lock);
4204 timeout = do_wait_for_common(x, timeout, state);
4205 spin_unlock_irq(&x->wait.lock);
4209 void __sched wait_for_completion(struct completion *x)
4211 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4213 EXPORT_SYMBOL(wait_for_completion);
4215 unsigned long __sched
4216 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4218 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4220 EXPORT_SYMBOL(wait_for_completion_timeout);
4222 int __sched wait_for_completion_interruptible(struct completion *x)
4224 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4225 if (t == -ERESTARTSYS)
4229 EXPORT_SYMBOL(wait_for_completion_interruptible);
4231 unsigned long __sched
4232 wait_for_completion_interruptible_timeout(struct completion *x,
4233 unsigned long timeout)
4235 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4237 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4239 int __sched wait_for_completion_killable(struct completion *x)
4241 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4242 if (t == -ERESTARTSYS)
4246 EXPORT_SYMBOL(wait_for_completion_killable);
4249 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4251 unsigned long flags;
4254 init_waitqueue_entry(&wait, current);
4256 __set_current_state(state);
4258 spin_lock_irqsave(&q->lock, flags);
4259 __add_wait_queue(q, &wait);
4260 spin_unlock(&q->lock);
4261 timeout = schedule_timeout(timeout);
4262 spin_lock_irq(&q->lock);
4263 __remove_wait_queue(q, &wait);
4264 spin_unlock_irqrestore(&q->lock, flags);
4269 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4271 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4273 EXPORT_SYMBOL(interruptible_sleep_on);
4276 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4278 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4280 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4282 void __sched sleep_on(wait_queue_head_t *q)
4284 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4286 EXPORT_SYMBOL(sleep_on);
4288 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4290 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4292 EXPORT_SYMBOL(sleep_on_timeout);
4294 #ifdef CONFIG_RT_MUTEXES
4297 * rt_mutex_setprio - set the current priority of a task
4299 * @prio: prio value (kernel-internal form)
4301 * This function changes the 'effective' priority of a task. It does
4302 * not touch ->normal_prio like __setscheduler().
4304 * Used by the rt_mutex code to implement priority inheritance logic.
4306 void rt_mutex_setprio(struct task_struct *p, int prio)
4308 unsigned long flags;
4309 int oldprio, on_rq, running;
4311 const struct sched_class *prev_class = p->sched_class;
4313 BUG_ON(prio < 0 || prio > MAX_PRIO);
4315 rq = task_rq_lock(p, &flags);
4316 update_rq_clock(rq);
4319 on_rq = p->se.on_rq;
4320 running = task_current(rq, p);
4322 dequeue_task(rq, p, 0);
4324 p->sched_class->put_prev_task(rq, p);
4327 p->sched_class = &rt_sched_class;
4329 p->sched_class = &fair_sched_class;
4334 p->sched_class->set_curr_task(rq);
4336 enqueue_task(rq, p, 0);
4338 check_class_changed(rq, p, prev_class, oldprio, running);
4340 task_rq_unlock(rq, &flags);
4345 void set_user_nice(struct task_struct *p, long nice)
4347 int old_prio, delta, on_rq;
4348 unsigned long flags;
4351 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4354 * We have to be careful, if called from sys_setpriority(),
4355 * the task might be in the middle of scheduling on another CPU.
4357 rq = task_rq_lock(p, &flags);
4358 update_rq_clock(rq);
4360 * The RT priorities are set via sched_setscheduler(), but we still
4361 * allow the 'normal' nice value to be set - but as expected
4362 * it wont have any effect on scheduling until the task is
4363 * SCHED_FIFO/SCHED_RR:
4365 if (task_has_rt_policy(p)) {
4366 p->static_prio = NICE_TO_PRIO(nice);
4369 on_rq = p->se.on_rq;
4371 dequeue_task(rq, p, 0);
4375 p->static_prio = NICE_TO_PRIO(nice);
4378 p->prio = effective_prio(p);
4379 delta = p->prio - old_prio;
4382 enqueue_task(rq, p, 0);
4385 * If the task increased its priority or is running and
4386 * lowered its priority, then reschedule its CPU:
4388 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4389 resched_task(rq->curr);
4392 task_rq_unlock(rq, &flags);
4394 EXPORT_SYMBOL(set_user_nice);
4397 * can_nice - check if a task can reduce its nice value
4401 int can_nice(const struct task_struct *p, const int nice)
4403 /* convert nice value [19,-20] to rlimit style value [1,40] */
4404 int nice_rlim = 20 - nice;
4406 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4407 capable(CAP_SYS_NICE));
4410 #ifdef __ARCH_WANT_SYS_NICE
4413 * sys_nice - change the priority of the current process.
4414 * @increment: priority increment
4416 * sys_setpriority is a more generic, but much slower function that
4417 * does similar things.
4419 asmlinkage long sys_nice(int increment)
4424 * Setpriority might change our priority at the same moment.
4425 * We don't have to worry. Conceptually one call occurs first
4426 * and we have a single winner.
4428 if (increment < -40)
4433 nice = PRIO_TO_NICE(current->static_prio) + increment;
4439 if (increment < 0 && !can_nice(current, nice))
4442 retval = security_task_setnice(current, nice);
4446 set_user_nice(current, nice);
4453 * task_prio - return the priority value of a given task.
4454 * @p: the task in question.
4456 * This is the priority value as seen by users in /proc.
4457 * RT tasks are offset by -200. Normal tasks are centered
4458 * around 0, value goes from -16 to +15.
4460 int task_prio(const struct task_struct *p)
4462 return p->prio - MAX_RT_PRIO;
4466 * task_nice - return the nice value of a given task.
4467 * @p: the task in question.
4469 int task_nice(const struct task_struct *p)
4471 return TASK_NICE(p);
4473 EXPORT_SYMBOL(task_nice);
4476 * idle_cpu - is a given cpu idle currently?
4477 * @cpu: the processor in question.
4479 int idle_cpu(int cpu)
4481 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4485 * idle_task - return the idle task for a given cpu.
4486 * @cpu: the processor in question.
4488 struct task_struct *idle_task(int cpu)
4490 return cpu_rq(cpu)->idle;
4494 * find_process_by_pid - find a process with a matching PID value.
4495 * @pid: the pid in question.
4497 static struct task_struct *find_process_by_pid(pid_t pid)
4499 return pid ? find_task_by_vpid(pid) : current;
4502 /* Actually do priority change: must hold rq lock. */
4504 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4506 BUG_ON(p->se.on_rq);
4509 switch (p->policy) {
4513 p->sched_class = &fair_sched_class;
4517 p->sched_class = &rt_sched_class;
4521 p->rt_priority = prio;
4522 p->normal_prio = normal_prio(p);
4523 /* we are holding p->pi_lock already */
4524 p->prio = rt_mutex_getprio(p);
4529 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4530 * @p: the task in question.
4531 * @policy: new policy.
4532 * @param: structure containing the new RT priority.
4534 * NOTE that the task may be already dead.
4536 int sched_setscheduler(struct task_struct *p, int policy,
4537 struct sched_param *param)
4539 int retval, oldprio, oldpolicy = -1, on_rq, running;
4540 unsigned long flags;
4541 const struct sched_class *prev_class = p->sched_class;
4544 /* may grab non-irq protected spin_locks */
4545 BUG_ON(in_interrupt());
4547 /* double check policy once rq lock held */
4549 policy = oldpolicy = p->policy;
4550 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4551 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4552 policy != SCHED_IDLE)
4555 * Valid priorities for SCHED_FIFO and SCHED_RR are
4556 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4557 * SCHED_BATCH and SCHED_IDLE is 0.
4559 if (param->sched_priority < 0 ||
4560 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4561 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4563 if (rt_policy(policy) != (param->sched_priority != 0))
4567 * Allow unprivileged RT tasks to decrease priority:
4569 if (!capable(CAP_SYS_NICE)) {
4570 if (rt_policy(policy)) {
4571 unsigned long rlim_rtprio;
4573 if (!lock_task_sighand(p, &flags))
4575 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4576 unlock_task_sighand(p, &flags);
4578 /* can't set/change the rt policy */
4579 if (policy != p->policy && !rlim_rtprio)
4582 /* can't increase priority */
4583 if (param->sched_priority > p->rt_priority &&
4584 param->sched_priority > rlim_rtprio)
4588 * Like positive nice levels, dont allow tasks to
4589 * move out of SCHED_IDLE either:
4591 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4594 /* can't change other user's priorities */
4595 if ((current->euid != p->euid) &&
4596 (current->euid != p->uid))
4600 #ifdef CONFIG_RT_GROUP_SCHED
4602 * Do not allow realtime tasks into groups that have no runtime
4605 if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4609 retval = security_task_setscheduler(p, policy, param);
4613 * make sure no PI-waiters arrive (or leave) while we are
4614 * changing the priority of the task:
4616 spin_lock_irqsave(&p->pi_lock, flags);
4618 * To be able to change p->policy safely, the apropriate
4619 * runqueue lock must be held.
4621 rq = __task_rq_lock(p);
4622 /* recheck policy now with rq lock held */
4623 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4624 policy = oldpolicy = -1;
4625 __task_rq_unlock(rq);
4626 spin_unlock_irqrestore(&p->pi_lock, flags);
4629 update_rq_clock(rq);
4630 on_rq = p->se.on_rq;
4631 running = task_current(rq, p);
4633 deactivate_task(rq, p, 0);
4635 p->sched_class->put_prev_task(rq, p);
4638 __setscheduler(rq, p, policy, param->sched_priority);
4641 p->sched_class->set_curr_task(rq);
4643 activate_task(rq, p, 0);
4645 check_class_changed(rq, p, prev_class, oldprio, running);
4647 __task_rq_unlock(rq);
4648 spin_unlock_irqrestore(&p->pi_lock, flags);
4650 rt_mutex_adjust_pi(p);
4654 EXPORT_SYMBOL_GPL(sched_setscheduler);
4657 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4659 struct sched_param lparam;
4660 struct task_struct *p;
4663 if (!param || pid < 0)
4665 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4670 p = find_process_by_pid(pid);
4672 retval = sched_setscheduler(p, policy, &lparam);
4679 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4680 * @pid: the pid in question.
4681 * @policy: new policy.
4682 * @param: structure containing the new RT priority.
4685 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4687 /* negative values for policy are not valid */
4691 return do_sched_setscheduler(pid, policy, param);
4695 * sys_sched_setparam - set/change the RT priority of a thread
4696 * @pid: the pid in question.
4697 * @param: structure containing the new RT priority.
4699 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4701 return do_sched_setscheduler(pid, -1, param);
4705 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4706 * @pid: the pid in question.
4708 asmlinkage long sys_sched_getscheduler(pid_t pid)
4710 struct task_struct *p;
4717 read_lock(&tasklist_lock);
4718 p = find_process_by_pid(pid);
4720 retval = security_task_getscheduler(p);
4724 read_unlock(&tasklist_lock);
4729 * sys_sched_getscheduler - get the RT priority of a thread
4730 * @pid: the pid in question.
4731 * @param: structure containing the RT priority.
4733 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4735 struct sched_param lp;
4736 struct task_struct *p;
4739 if (!param || pid < 0)
4742 read_lock(&tasklist_lock);
4743 p = find_process_by_pid(pid);
4748 retval = security_task_getscheduler(p);
4752 lp.sched_priority = p->rt_priority;
4753 read_unlock(&tasklist_lock);
4756 * This one might sleep, we cannot do it with a spinlock held ...
4758 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4763 read_unlock(&tasklist_lock);
4767 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4769 cpumask_t cpus_allowed;
4770 struct task_struct *p;
4774 read_lock(&tasklist_lock);
4776 p = find_process_by_pid(pid);
4778 read_unlock(&tasklist_lock);
4784 * It is not safe to call set_cpus_allowed with the
4785 * tasklist_lock held. We will bump the task_struct's
4786 * usage count and then drop tasklist_lock.
4789 read_unlock(&tasklist_lock);
4792 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4793 !capable(CAP_SYS_NICE))
4796 retval = security_task_setscheduler(p, 0, NULL);
4800 cpus_allowed = cpuset_cpus_allowed(p);
4801 cpus_and(new_mask, new_mask, cpus_allowed);
4803 retval = set_cpus_allowed(p, new_mask);
4806 cpus_allowed = cpuset_cpus_allowed(p);
4807 if (!cpus_subset(new_mask, cpus_allowed)) {
4809 * We must have raced with a concurrent cpuset
4810 * update. Just reset the cpus_allowed to the
4811 * cpuset's cpus_allowed
4813 new_mask = cpus_allowed;
4823 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4824 cpumask_t *new_mask)
4826 if (len < sizeof(cpumask_t)) {
4827 memset(new_mask, 0, sizeof(cpumask_t));
4828 } else if (len > sizeof(cpumask_t)) {
4829 len = sizeof(cpumask_t);
4831 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4835 * sys_sched_setaffinity - set the cpu affinity of a process
4836 * @pid: pid of the process
4837 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4838 * @user_mask_ptr: user-space pointer to the new cpu mask
4840 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4841 unsigned long __user *user_mask_ptr)
4846 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4850 return sched_setaffinity(pid, new_mask);
4854 * Represents all cpu's present in the system
4855 * In systems capable of hotplug, this map could dynamically grow
4856 * as new cpu's are detected in the system via any platform specific
4857 * method, such as ACPI for e.g.
4860 cpumask_t cpu_present_map __read_mostly;
4861 EXPORT_SYMBOL(cpu_present_map);
4864 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4865 EXPORT_SYMBOL(cpu_online_map);
4867 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4868 EXPORT_SYMBOL(cpu_possible_map);
4871 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4873 struct task_struct *p;
4877 read_lock(&tasklist_lock);
4880 p = find_process_by_pid(pid);
4884 retval = security_task_getscheduler(p);
4888 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4891 read_unlock(&tasklist_lock);
4898 * sys_sched_getaffinity - get the cpu affinity of a process
4899 * @pid: pid of the process
4900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4901 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4903 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4904 unsigned long __user *user_mask_ptr)
4909 if (len < sizeof(cpumask_t))
4912 ret = sched_getaffinity(pid, &mask);
4916 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4919 return sizeof(cpumask_t);
4923 * sys_sched_yield - yield the current processor to other threads.
4925 * This function yields the current CPU to other tasks. If there are no
4926 * other threads running on this CPU then this function will return.
4928 asmlinkage long sys_sched_yield(void)
4930 struct rq *rq = this_rq_lock();
4932 schedstat_inc(rq, yld_count);
4933 current->sched_class->yield_task(rq);
4936 * Since we are going to call schedule() anyway, there's
4937 * no need to preempt or enable interrupts:
4939 __release(rq->lock);
4940 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4941 _raw_spin_unlock(&rq->lock);
4942 preempt_enable_no_resched();
4949 static void __cond_resched(void)
4951 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4952 __might_sleep(__FILE__, __LINE__);
4955 * The BKS might be reacquired before we have dropped
4956 * PREEMPT_ACTIVE, which could trigger a second
4957 * cond_resched() call.
4960 add_preempt_count(PREEMPT_ACTIVE);
4962 sub_preempt_count(PREEMPT_ACTIVE);
4963 } while (need_resched());
4966 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4967 int __sched _cond_resched(void)
4969 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4970 system_state == SYSTEM_RUNNING) {
4976 EXPORT_SYMBOL(_cond_resched);
4980 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4981 * call schedule, and on return reacquire the lock.
4983 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4984 * operations here to prevent schedule() from being called twice (once via
4985 * spin_unlock(), once by hand).
4987 int cond_resched_lock(spinlock_t *lock)
4989 int resched = need_resched() && system_state == SYSTEM_RUNNING;
4992 if (spin_needbreak(lock) || resched) {
4994 if (resched && need_resched())
5003 EXPORT_SYMBOL(cond_resched_lock);
5005 int __sched cond_resched_softirq(void)
5007 BUG_ON(!in_softirq());
5009 if (need_resched() && system_state == SYSTEM_RUNNING) {
5017 EXPORT_SYMBOL(cond_resched_softirq);
5020 * yield - yield the current processor to other threads.
5022 * This is a shortcut for kernel-space yielding - it marks the
5023 * thread runnable and calls sys_sched_yield().
5025 void __sched yield(void)
5027 set_current_state(TASK_RUNNING);
5030 EXPORT_SYMBOL(yield);
5033 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5034 * that process accounting knows that this is a task in IO wait state.
5036 * But don't do that if it is a deliberate, throttling IO wait (this task
5037 * has set its backing_dev_info: the queue against which it should throttle)
5039 void __sched io_schedule(void)
5041 struct rq *rq = &__raw_get_cpu_var(runqueues);
5043 delayacct_blkio_start();
5044 atomic_inc(&rq->nr_iowait);
5046 atomic_dec(&rq->nr_iowait);
5047 delayacct_blkio_end();
5049 EXPORT_SYMBOL(io_schedule);
5051 long __sched io_schedule_timeout(long timeout)
5053 struct rq *rq = &__raw_get_cpu_var(runqueues);
5056 delayacct_blkio_start();
5057 atomic_inc(&rq->nr_iowait);
5058 ret = schedule_timeout(timeout);
5059 atomic_dec(&rq->nr_iowait);
5060 delayacct_blkio_end();
5065 * sys_sched_get_priority_max - return maximum RT priority.
5066 * @policy: scheduling class.
5068 * this syscall returns the maximum rt_priority that can be used
5069 * by a given scheduling class.
5071 asmlinkage long sys_sched_get_priority_max(int policy)
5078 ret = MAX_USER_RT_PRIO-1;
5090 * sys_sched_get_priority_min - return minimum RT priority.
5091 * @policy: scheduling class.
5093 * this syscall returns the minimum rt_priority that can be used
5094 * by a given scheduling class.
5096 asmlinkage long sys_sched_get_priority_min(int policy)
5114 * sys_sched_rr_get_interval - return the default timeslice of a process.
5115 * @pid: pid of the process.
5116 * @interval: userspace pointer to the timeslice value.
5118 * this syscall writes the default timeslice value of a given process
5119 * into the user-space timespec buffer. A value of '0' means infinity.
5122 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5124 struct task_struct *p;
5125 unsigned int time_slice;
5133 read_lock(&tasklist_lock);
5134 p = find_process_by_pid(pid);
5138 retval = security_task_getscheduler(p);
5143 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5144 * tasks that are on an otherwise idle runqueue:
5147 if (p->policy == SCHED_RR) {
5148 time_slice = DEF_TIMESLICE;
5149 } else if (p->policy != SCHED_FIFO) {
5150 struct sched_entity *se = &p->se;
5151 unsigned long flags;
5154 rq = task_rq_lock(p, &flags);
5155 if (rq->cfs.load.weight)
5156 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5157 task_rq_unlock(rq, &flags);
5159 read_unlock(&tasklist_lock);
5160 jiffies_to_timespec(time_slice, &t);
5161 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5165 read_unlock(&tasklist_lock);
5169 static const char stat_nam[] = "RSDTtZX";
5171 void sched_show_task(struct task_struct *p)
5173 unsigned long free = 0;
5176 state = p->state ? __ffs(p->state) + 1 : 0;
5177 printk(KERN_INFO "%-13.13s %c", p->comm,
5178 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5179 #if BITS_PER_LONG == 32
5180 if (state == TASK_RUNNING)
5181 printk(KERN_CONT " running ");
5183 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5185 if (state == TASK_RUNNING)
5186 printk(KERN_CONT " running task ");
5188 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5190 #ifdef CONFIG_DEBUG_STACK_USAGE
5192 unsigned long *n = end_of_stack(p);
5195 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5198 printk(KERN_CONT "%5lu %5d %6d\n", free,
5199 task_pid_nr(p), task_pid_nr(p->real_parent));
5201 show_stack(p, NULL);
5204 void show_state_filter(unsigned long state_filter)
5206 struct task_struct *g, *p;
5208 #if BITS_PER_LONG == 32
5210 " task PC stack pid father\n");
5213 " task PC stack pid father\n");
5215 read_lock(&tasklist_lock);
5216 do_each_thread(g, p) {
5218 * reset the NMI-timeout, listing all files on a slow
5219 * console might take alot of time:
5221 touch_nmi_watchdog();
5222 if (!state_filter || (p->state & state_filter))
5224 } while_each_thread(g, p);
5226 touch_all_softlockup_watchdogs();
5228 #ifdef CONFIG_SCHED_DEBUG
5229 sysrq_sched_debug_show();
5231 read_unlock(&tasklist_lock);
5233 * Only show locks if all tasks are dumped:
5235 if (state_filter == -1)
5236 debug_show_all_locks();
5239 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5241 idle->sched_class = &idle_sched_class;
5245 * init_idle - set up an idle thread for a given CPU
5246 * @idle: task in question
5247 * @cpu: cpu the idle task belongs to
5249 * NOTE: this function does not set the idle thread's NEED_RESCHED
5250 * flag, to make booting more robust.
5252 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5254 struct rq *rq = cpu_rq(cpu);
5255 unsigned long flags;
5258 idle->se.exec_start = sched_clock();
5260 idle->prio = idle->normal_prio = MAX_PRIO;
5261 idle->cpus_allowed = cpumask_of_cpu(cpu);
5262 __set_task_cpu(idle, cpu);
5264 spin_lock_irqsave(&rq->lock, flags);
5265 rq->curr = rq->idle = idle;
5266 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5269 spin_unlock_irqrestore(&rq->lock, flags);
5271 /* Set the preempt count _outside_ the spinlocks! */
5272 task_thread_info(idle)->preempt_count = 0;
5275 * The idle tasks have their own, simple scheduling class:
5277 idle->sched_class = &idle_sched_class;
5281 * In a system that switches off the HZ timer nohz_cpu_mask
5282 * indicates which cpus entered this state. This is used
5283 * in the rcu update to wait only for active cpus. For system
5284 * which do not switch off the HZ timer nohz_cpu_mask should
5285 * always be CPU_MASK_NONE.
5287 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5290 * Increase the granularity value when there are more CPUs,
5291 * because with more CPUs the 'effective latency' as visible
5292 * to users decreases. But the relationship is not linear,
5293 * so pick a second-best guess by going with the log2 of the
5296 * This idea comes from the SD scheduler of Con Kolivas:
5298 static inline void sched_init_granularity(void)
5300 unsigned int factor = 1 + ilog2(num_online_cpus());
5301 const unsigned long limit = 200000000;
5303 sysctl_sched_min_granularity *= factor;
5304 if (sysctl_sched_min_granularity > limit)
5305 sysctl_sched_min_granularity = limit;
5307 sysctl_sched_latency *= factor;
5308 if (sysctl_sched_latency > limit)
5309 sysctl_sched_latency = limit;
5311 sysctl_sched_wakeup_granularity *= factor;
5312 sysctl_sched_batch_wakeup_granularity *= factor;
5317 * This is how migration works:
5319 * 1) we queue a struct migration_req structure in the source CPU's
5320 * runqueue and wake up that CPU's migration thread.
5321 * 2) we down() the locked semaphore => thread blocks.
5322 * 3) migration thread wakes up (implicitly it forces the migrated
5323 * thread off the CPU)
5324 * 4) it gets the migration request and checks whether the migrated
5325 * task is still in the wrong runqueue.
5326 * 5) if it's in the wrong runqueue then the migration thread removes
5327 * it and puts it into the right queue.
5328 * 6) migration thread up()s the semaphore.
5329 * 7) we wake up and the migration is done.
5333 * Change a given task's CPU affinity. Migrate the thread to a
5334 * proper CPU and schedule it away if the CPU it's executing on
5335 * is removed from the allowed bitmask.
5337 * NOTE: the caller must have a valid reference to the task, the
5338 * task must not exit() & deallocate itself prematurely. The
5339 * call is not atomic; no spinlocks may be held.
5341 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5343 struct migration_req req;
5344 unsigned long flags;
5348 rq = task_rq_lock(p, &flags);
5349 if (!cpus_intersects(new_mask, cpu_online_map)) {
5354 if (p->sched_class->set_cpus_allowed)
5355 p->sched_class->set_cpus_allowed(p, &new_mask);
5357 p->cpus_allowed = new_mask;
5358 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5361 /* Can the task run on the task's current CPU? If so, we're done */
5362 if (cpu_isset(task_cpu(p), new_mask))
5365 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5366 /* Need help from migration thread: drop lock and wait. */
5367 task_rq_unlock(rq, &flags);
5368 wake_up_process(rq->migration_thread);
5369 wait_for_completion(&req.done);
5370 tlb_migrate_finish(p->mm);
5374 task_rq_unlock(rq, &flags);
5378 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5381 * Move (not current) task off this cpu, onto dest cpu. We're doing
5382 * this because either it can't run here any more (set_cpus_allowed()
5383 * away from this CPU, or CPU going down), or because we're
5384 * attempting to rebalance this task on exec (sched_exec).
5386 * So we race with normal scheduler movements, but that's OK, as long
5387 * as the task is no longer on this CPU.
5389 * Returns non-zero if task was successfully migrated.
5391 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5393 struct rq *rq_dest, *rq_src;
5396 if (unlikely(cpu_is_offline(dest_cpu)))
5399 rq_src = cpu_rq(src_cpu);
5400 rq_dest = cpu_rq(dest_cpu);
5402 double_rq_lock(rq_src, rq_dest);
5403 /* Already moved. */
5404 if (task_cpu(p) != src_cpu)
5406 /* Affinity changed (again). */
5407 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5410 on_rq = p->se.on_rq;
5412 deactivate_task(rq_src, p, 0);
5414 set_task_cpu(p, dest_cpu);
5416 activate_task(rq_dest, p, 0);
5417 check_preempt_curr(rq_dest, p);
5421 double_rq_unlock(rq_src, rq_dest);
5426 * migration_thread - this is a highprio system thread that performs
5427 * thread migration by bumping thread off CPU then 'pushing' onto
5430 static int migration_thread(void *data)
5432 int cpu = (long)data;
5436 BUG_ON(rq->migration_thread != current);
5438 set_current_state(TASK_INTERRUPTIBLE);
5439 while (!kthread_should_stop()) {
5440 struct migration_req *req;
5441 struct list_head *head;
5443 spin_lock_irq(&rq->lock);
5445 if (cpu_is_offline(cpu)) {
5446 spin_unlock_irq(&rq->lock);
5450 if (rq->active_balance) {
5451 active_load_balance(rq, cpu);
5452 rq->active_balance = 0;
5455 head = &rq->migration_queue;
5457 if (list_empty(head)) {
5458 spin_unlock_irq(&rq->lock);
5460 set_current_state(TASK_INTERRUPTIBLE);
5463 req = list_entry(head->next, struct migration_req, list);
5464 list_del_init(head->next);
5466 spin_unlock(&rq->lock);
5467 __migrate_task(req->task, cpu, req->dest_cpu);
5470 complete(&req->done);
5472 __set_current_state(TASK_RUNNING);
5476 /* Wait for kthread_stop */
5477 set_current_state(TASK_INTERRUPTIBLE);
5478 while (!kthread_should_stop()) {
5480 set_current_state(TASK_INTERRUPTIBLE);
5482 __set_current_state(TASK_RUNNING);
5486 #ifdef CONFIG_HOTPLUG_CPU
5488 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5492 local_irq_disable();
5493 ret = __migrate_task(p, src_cpu, dest_cpu);
5499 * Figure out where task on dead CPU should go, use force if necessary.
5500 * NOTE: interrupts should be disabled by the caller
5502 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5504 unsigned long flags;
5511 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5512 cpus_and(mask, mask, p->cpus_allowed);
5513 dest_cpu = any_online_cpu(mask);
5515 /* On any allowed CPU? */
5516 if (dest_cpu == NR_CPUS)
5517 dest_cpu = any_online_cpu(p->cpus_allowed);
5519 /* No more Mr. Nice Guy. */
5520 if (dest_cpu == NR_CPUS) {
5521 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5523 * Try to stay on the same cpuset, where the
5524 * current cpuset may be a subset of all cpus.
5525 * The cpuset_cpus_allowed_locked() variant of
5526 * cpuset_cpus_allowed() will not block. It must be
5527 * called within calls to cpuset_lock/cpuset_unlock.
5529 rq = task_rq_lock(p, &flags);
5530 p->cpus_allowed = cpus_allowed;
5531 dest_cpu = any_online_cpu(p->cpus_allowed);
5532 task_rq_unlock(rq, &flags);
5535 * Don't tell them about moving exiting tasks or
5536 * kernel threads (both mm NULL), since they never
5539 if (p->mm && printk_ratelimit()) {
5540 printk(KERN_INFO "process %d (%s) no "
5541 "longer affine to cpu%d\n",
5542 task_pid_nr(p), p->comm, dead_cpu);
5545 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5549 * While a dead CPU has no uninterruptible tasks queued at this point,
5550 * it might still have a nonzero ->nr_uninterruptible counter, because
5551 * for performance reasons the counter is not stricly tracking tasks to
5552 * their home CPUs. So we just add the counter to another CPU's counter,
5553 * to keep the global sum constant after CPU-down:
5555 static void migrate_nr_uninterruptible(struct rq *rq_src)
5557 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5558 unsigned long flags;
5560 local_irq_save(flags);
5561 double_rq_lock(rq_src, rq_dest);
5562 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5563 rq_src->nr_uninterruptible = 0;
5564 double_rq_unlock(rq_src, rq_dest);
5565 local_irq_restore(flags);
5568 /* Run through task list and migrate tasks from the dead cpu. */
5569 static void migrate_live_tasks(int src_cpu)
5571 struct task_struct *p, *t;
5573 read_lock(&tasklist_lock);
5575 do_each_thread(t, p) {
5579 if (task_cpu(p) == src_cpu)
5580 move_task_off_dead_cpu(src_cpu, p);
5581 } while_each_thread(t, p);
5583 read_unlock(&tasklist_lock);
5587 * Schedules idle task to be the next runnable task on current CPU.
5588 * It does so by boosting its priority to highest possible.
5589 * Used by CPU offline code.
5591 void sched_idle_next(void)
5593 int this_cpu = smp_processor_id();
5594 struct rq *rq = cpu_rq(this_cpu);
5595 struct task_struct *p = rq->idle;
5596 unsigned long flags;
5598 /* cpu has to be offline */
5599 BUG_ON(cpu_online(this_cpu));
5602 * Strictly not necessary since rest of the CPUs are stopped by now
5603 * and interrupts disabled on the current cpu.
5605 spin_lock_irqsave(&rq->lock, flags);
5607 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5609 update_rq_clock(rq);
5610 activate_task(rq, p, 0);
5612 spin_unlock_irqrestore(&rq->lock, flags);
5616 * Ensures that the idle task is using init_mm right before its cpu goes
5619 void idle_task_exit(void)
5621 struct mm_struct *mm = current->active_mm;
5623 BUG_ON(cpu_online(smp_processor_id()));
5626 switch_mm(mm, &init_mm, current);
5630 /* called under rq->lock with disabled interrupts */
5631 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5633 struct rq *rq = cpu_rq(dead_cpu);
5635 /* Must be exiting, otherwise would be on tasklist. */
5636 BUG_ON(!p->exit_state);
5638 /* Cannot have done final schedule yet: would have vanished. */
5639 BUG_ON(p->state == TASK_DEAD);
5644 * Drop lock around migration; if someone else moves it,
5645 * that's OK. No task can be added to this CPU, so iteration is
5648 spin_unlock_irq(&rq->lock);
5649 move_task_off_dead_cpu(dead_cpu, p);
5650 spin_lock_irq(&rq->lock);
5655 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5656 static void migrate_dead_tasks(unsigned int dead_cpu)
5658 struct rq *rq = cpu_rq(dead_cpu);
5659 struct task_struct *next;
5662 if (!rq->nr_running)
5664 update_rq_clock(rq);
5665 next = pick_next_task(rq, rq->curr);
5668 migrate_dead(dead_cpu, next);
5672 #endif /* CONFIG_HOTPLUG_CPU */
5674 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5676 static struct ctl_table sd_ctl_dir[] = {
5678 .procname = "sched_domain",
5684 static struct ctl_table sd_ctl_root[] = {
5686 .ctl_name = CTL_KERN,
5687 .procname = "kernel",
5689 .child = sd_ctl_dir,
5694 static struct ctl_table *sd_alloc_ctl_entry(int n)
5696 struct ctl_table *entry =
5697 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5702 static void sd_free_ctl_entry(struct ctl_table **tablep)
5704 struct ctl_table *entry;
5707 * In the intermediate directories, both the child directory and
5708 * procname are dynamically allocated and could fail but the mode
5709 * will always be set. In the lowest directory the names are
5710 * static strings and all have proc handlers.
5712 for (entry = *tablep; entry->mode; entry++) {
5714 sd_free_ctl_entry(&entry->child);
5715 if (entry->proc_handler == NULL)
5716 kfree(entry->procname);
5724 set_table_entry(struct ctl_table *entry,
5725 const char *procname, void *data, int maxlen,
5726 mode_t mode, proc_handler *proc_handler)
5728 entry->procname = procname;
5730 entry->maxlen = maxlen;
5732 entry->proc_handler = proc_handler;
5735 static struct ctl_table *
5736 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5738 struct ctl_table *table = sd_alloc_ctl_entry(12);
5743 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5744 sizeof(long), 0644, proc_doulongvec_minmax);
5745 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5746 sizeof(long), 0644, proc_doulongvec_minmax);
5747 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5748 sizeof(int), 0644, proc_dointvec_minmax);
5749 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5750 sizeof(int), 0644, proc_dointvec_minmax);
5751 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5752 sizeof(int), 0644, proc_dointvec_minmax);
5753 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5754 sizeof(int), 0644, proc_dointvec_minmax);
5755 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5756 sizeof(int), 0644, proc_dointvec_minmax);
5757 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5758 sizeof(int), 0644, proc_dointvec_minmax);
5759 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5760 sizeof(int), 0644, proc_dointvec_minmax);
5761 set_table_entry(&table[9], "cache_nice_tries",
5762 &sd->cache_nice_tries,
5763 sizeof(int), 0644, proc_dointvec_minmax);
5764 set_table_entry(&table[10], "flags", &sd->flags,
5765 sizeof(int), 0644, proc_dointvec_minmax);
5766 /* &table[11] is terminator */
5771 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5773 struct ctl_table *entry, *table;
5774 struct sched_domain *sd;
5775 int domain_num = 0, i;
5778 for_each_domain(cpu, sd)
5780 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5785 for_each_domain(cpu, sd) {
5786 snprintf(buf, 32, "domain%d", i);
5787 entry->procname = kstrdup(buf, GFP_KERNEL);
5789 entry->child = sd_alloc_ctl_domain_table(sd);
5796 static struct ctl_table_header *sd_sysctl_header;
5797 static void register_sched_domain_sysctl(void)
5799 int i, cpu_num = num_online_cpus();
5800 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5803 WARN_ON(sd_ctl_dir[0].child);
5804 sd_ctl_dir[0].child = entry;
5809 for_each_online_cpu(i) {
5810 snprintf(buf, 32, "cpu%d", i);
5811 entry->procname = kstrdup(buf, GFP_KERNEL);
5813 entry->child = sd_alloc_ctl_cpu_table(i);
5817 WARN_ON(sd_sysctl_header);
5818 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5821 /* may be called multiple times per register */
5822 static void unregister_sched_domain_sysctl(void)
5824 if (sd_sysctl_header)
5825 unregister_sysctl_table(sd_sysctl_header);
5826 sd_sysctl_header = NULL;
5827 if (sd_ctl_dir[0].child)
5828 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5831 static void register_sched_domain_sysctl(void)
5834 static void unregister_sched_domain_sysctl(void)
5840 * migration_call - callback that gets triggered when a CPU is added.
5841 * Here we can start up the necessary migration thread for the new CPU.
5843 static int __cpuinit
5844 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5846 struct task_struct *p;
5847 int cpu = (long)hcpu;
5848 unsigned long flags;
5853 case CPU_UP_PREPARE:
5854 case CPU_UP_PREPARE_FROZEN:
5855 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5858 kthread_bind(p, cpu);
5859 /* Must be high prio: stop_machine expects to yield to it. */
5860 rq = task_rq_lock(p, &flags);
5861 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5862 task_rq_unlock(rq, &flags);
5863 cpu_rq(cpu)->migration_thread = p;
5867 case CPU_ONLINE_FROZEN:
5868 /* Strictly unnecessary, as first user will wake it. */
5869 wake_up_process(cpu_rq(cpu)->migration_thread);
5871 /* Update our root-domain */
5873 spin_lock_irqsave(&rq->lock, flags);
5875 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5876 cpu_set(cpu, rq->rd->online);
5878 spin_unlock_irqrestore(&rq->lock, flags);
5881 #ifdef CONFIG_HOTPLUG_CPU
5882 case CPU_UP_CANCELED:
5883 case CPU_UP_CANCELED_FROZEN:
5884 if (!cpu_rq(cpu)->migration_thread)
5886 /* Unbind it from offline cpu so it can run. Fall thru. */
5887 kthread_bind(cpu_rq(cpu)->migration_thread,
5888 any_online_cpu(cpu_online_map));
5889 kthread_stop(cpu_rq(cpu)->migration_thread);
5890 cpu_rq(cpu)->migration_thread = NULL;
5894 case CPU_DEAD_FROZEN:
5895 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5896 migrate_live_tasks(cpu);
5898 kthread_stop(rq->migration_thread);
5899 rq->migration_thread = NULL;
5900 /* Idle task back to normal (off runqueue, low prio) */
5901 spin_lock_irq(&rq->lock);
5902 update_rq_clock(rq);
5903 deactivate_task(rq, rq->idle, 0);
5904 rq->idle->static_prio = MAX_PRIO;
5905 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5906 rq->idle->sched_class = &idle_sched_class;
5907 migrate_dead_tasks(cpu);
5908 spin_unlock_irq(&rq->lock);
5910 migrate_nr_uninterruptible(rq);
5911 BUG_ON(rq->nr_running != 0);
5914 * No need to migrate the tasks: it was best-effort if
5915 * they didn't take sched_hotcpu_mutex. Just wake up
5918 spin_lock_irq(&rq->lock);
5919 while (!list_empty(&rq->migration_queue)) {
5920 struct migration_req *req;
5922 req = list_entry(rq->migration_queue.next,
5923 struct migration_req, list);
5924 list_del_init(&req->list);
5925 complete(&req->done);
5927 spin_unlock_irq(&rq->lock);
5931 case CPU_DYING_FROZEN:
5932 /* Update our root-domain */
5934 spin_lock_irqsave(&rq->lock, flags);
5936 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5937 cpu_clear(cpu, rq->rd->online);
5939 spin_unlock_irqrestore(&rq->lock, flags);
5946 /* Register at highest priority so that task migration (migrate_all_tasks)
5947 * happens before everything else.
5949 static struct notifier_block __cpuinitdata migration_notifier = {
5950 .notifier_call = migration_call,
5954 void __init migration_init(void)
5956 void *cpu = (void *)(long)smp_processor_id();
5959 /* Start one for the boot CPU: */
5960 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5961 BUG_ON(err == NOTIFY_BAD);
5962 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5963 register_cpu_notifier(&migration_notifier);
5969 /* Number of possible processor ids */
5970 int nr_cpu_ids __read_mostly = NR_CPUS;
5971 EXPORT_SYMBOL(nr_cpu_ids);
5973 #ifdef CONFIG_SCHED_DEBUG
5975 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5977 struct sched_group *group = sd->groups;
5978 cpumask_t groupmask;
5981 cpumask_scnprintf(str, NR_CPUS, sd->span);
5982 cpus_clear(groupmask);
5984 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5986 if (!(sd->flags & SD_LOAD_BALANCE)) {
5987 printk("does not load-balance\n");
5989 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5994 printk(KERN_CONT "span %s\n", str);
5996 if (!cpu_isset(cpu, sd->span)) {
5997 printk(KERN_ERR "ERROR: domain->span does not contain "
6000 if (!cpu_isset(cpu, group->cpumask)) {
6001 printk(KERN_ERR "ERROR: domain->groups does not contain"
6005 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6009 printk(KERN_ERR "ERROR: group is NULL\n");
6013 if (!group->__cpu_power) {
6014 printk(KERN_CONT "\n");
6015 printk(KERN_ERR "ERROR: domain->cpu_power not "
6020 if (!cpus_weight(group->cpumask)) {
6021 printk(KERN_CONT "\n");
6022 printk(KERN_ERR "ERROR: empty group\n");
6026 if (cpus_intersects(groupmask, group->cpumask)) {
6027 printk(KERN_CONT "\n");
6028 printk(KERN_ERR "ERROR: repeated CPUs\n");
6032 cpus_or(groupmask, groupmask, group->cpumask);
6034 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6035 printk(KERN_CONT " %s", str);
6037 group = group->next;
6038 } while (group != sd->groups);
6039 printk(KERN_CONT "\n");
6041 if (!cpus_equal(sd->span, groupmask))
6042 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6044 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6045 printk(KERN_ERR "ERROR: parent span is not a superset "
6046 "of domain->span\n");
6050 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6055 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6059 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6062 if (sched_domain_debug_one(sd, cpu, level))
6071 # define sched_domain_debug(sd, cpu) do { } while (0)
6074 static int sd_degenerate(struct sched_domain *sd)
6076 if (cpus_weight(sd->span) == 1)
6079 /* Following flags need at least 2 groups */
6080 if (sd->flags & (SD_LOAD_BALANCE |
6081 SD_BALANCE_NEWIDLE |
6085 SD_SHARE_PKG_RESOURCES)) {
6086 if (sd->groups != sd->groups->next)
6090 /* Following flags don't use groups */
6091 if (sd->flags & (SD_WAKE_IDLE |
6100 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6102 unsigned long cflags = sd->flags, pflags = parent->flags;
6104 if (sd_degenerate(parent))
6107 if (!cpus_equal(sd->span, parent->span))
6110 /* Does parent contain flags not in child? */
6111 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6112 if (cflags & SD_WAKE_AFFINE)
6113 pflags &= ~SD_WAKE_BALANCE;
6114 /* Flags needing groups don't count if only 1 group in parent */
6115 if (parent->groups == parent->groups->next) {
6116 pflags &= ~(SD_LOAD_BALANCE |
6117 SD_BALANCE_NEWIDLE |
6121 SD_SHARE_PKG_RESOURCES);
6123 if (~cflags & pflags)
6129 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6131 unsigned long flags;
6132 const struct sched_class *class;
6134 spin_lock_irqsave(&rq->lock, flags);
6137 struct root_domain *old_rd = rq->rd;
6139 for (class = sched_class_highest; class; class = class->next) {
6140 if (class->leave_domain)
6141 class->leave_domain(rq);
6144 cpu_clear(rq->cpu, old_rd->span);
6145 cpu_clear(rq->cpu, old_rd->online);
6147 if (atomic_dec_and_test(&old_rd->refcount))
6151 atomic_inc(&rd->refcount);
6154 cpu_set(rq->cpu, rd->span);
6155 if (cpu_isset(rq->cpu, cpu_online_map))
6156 cpu_set(rq->cpu, rd->online);
6158 for (class = sched_class_highest; class; class = class->next) {
6159 if (class->join_domain)
6160 class->join_domain(rq);
6163 spin_unlock_irqrestore(&rq->lock, flags);
6166 static void init_rootdomain(struct root_domain *rd)
6168 memset(rd, 0, sizeof(*rd));
6170 cpus_clear(rd->span);
6171 cpus_clear(rd->online);
6174 static void init_defrootdomain(void)
6176 init_rootdomain(&def_root_domain);
6177 atomic_set(&def_root_domain.refcount, 1);
6180 static struct root_domain *alloc_rootdomain(void)
6182 struct root_domain *rd;
6184 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6188 init_rootdomain(rd);
6194 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6195 * hold the hotplug lock.
6198 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6200 struct rq *rq = cpu_rq(cpu);
6201 struct sched_domain *tmp;
6203 /* Remove the sched domains which do not contribute to scheduling. */
6204 for (tmp = sd; tmp; tmp = tmp->parent) {
6205 struct sched_domain *parent = tmp->parent;
6208 if (sd_parent_degenerate(tmp, parent)) {
6209 tmp->parent = parent->parent;
6211 parent->parent->child = tmp;
6215 if (sd && sd_degenerate(sd)) {
6221 sched_domain_debug(sd, cpu);
6223 rq_attach_root(rq, rd);
6224 rcu_assign_pointer(rq->sd, sd);
6227 /* cpus with isolated domains */
6228 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6230 /* Setup the mask of cpus configured for isolated domains */
6231 static int __init isolated_cpu_setup(char *str)
6233 int ints[NR_CPUS], i;
6235 str = get_options(str, ARRAY_SIZE(ints), ints);
6236 cpus_clear(cpu_isolated_map);
6237 for (i = 1; i <= ints[0]; i++)
6238 if (ints[i] < NR_CPUS)
6239 cpu_set(ints[i], cpu_isolated_map);
6243 __setup("isolcpus=", isolated_cpu_setup);
6246 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6247 * to a function which identifies what group(along with sched group) a CPU
6248 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6249 * (due to the fact that we keep track of groups covered with a cpumask_t).
6251 * init_sched_build_groups will build a circular linked list of the groups
6252 * covered by the given span, and will set each group's ->cpumask correctly,
6253 * and ->cpu_power to 0.
6256 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6257 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6258 struct sched_group **sg))
6260 struct sched_group *first = NULL, *last = NULL;
6261 cpumask_t covered = CPU_MASK_NONE;
6264 for_each_cpu_mask(i, span) {
6265 struct sched_group *sg;
6266 int group = group_fn(i, cpu_map, &sg);
6269 if (cpu_isset(i, covered))
6272 sg->cpumask = CPU_MASK_NONE;
6273 sg->__cpu_power = 0;
6275 for_each_cpu_mask(j, span) {
6276 if (group_fn(j, cpu_map, NULL) != group)
6279 cpu_set(j, covered);
6280 cpu_set(j, sg->cpumask);
6291 #define SD_NODES_PER_DOMAIN 16
6296 * find_next_best_node - find the next node to include in a sched_domain
6297 * @node: node whose sched_domain we're building
6298 * @used_nodes: nodes already in the sched_domain
6300 * Find the next node to include in a given scheduling domain. Simply
6301 * finds the closest node not already in the @used_nodes map.
6303 * Should use nodemask_t.
6305 static int find_next_best_node(int node, unsigned long *used_nodes)
6307 int i, n, val, min_val, best_node = 0;
6311 for (i = 0; i < MAX_NUMNODES; i++) {
6312 /* Start at @node */
6313 n = (node + i) % MAX_NUMNODES;
6315 if (!nr_cpus_node(n))
6318 /* Skip already used nodes */
6319 if (test_bit(n, used_nodes))
6322 /* Simple min distance search */
6323 val = node_distance(node, n);
6325 if (val < min_val) {
6331 set_bit(best_node, used_nodes);
6336 * sched_domain_node_span - get a cpumask for a node's sched_domain
6337 * @node: node whose cpumask we're constructing
6338 * @size: number of nodes to include in this span
6340 * Given a node, construct a good cpumask for its sched_domain to span. It
6341 * should be one that prevents unnecessary balancing, but also spreads tasks
6344 static cpumask_t sched_domain_node_span(int node)
6346 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6347 cpumask_t span, nodemask;
6351 bitmap_zero(used_nodes, MAX_NUMNODES);
6353 nodemask = node_to_cpumask(node);
6354 cpus_or(span, span, nodemask);
6355 set_bit(node, used_nodes);
6357 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6358 int next_node = find_next_best_node(node, used_nodes);
6360 nodemask = node_to_cpumask(next_node);
6361 cpus_or(span, span, nodemask);
6368 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6371 * SMT sched-domains:
6373 #ifdef CONFIG_SCHED_SMT
6374 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6375 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6378 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6381 *sg = &per_cpu(sched_group_cpus, cpu);
6387 * multi-core sched-domains:
6389 #ifdef CONFIG_SCHED_MC
6390 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6391 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6394 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6396 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6399 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6400 cpus_and(mask, mask, *cpu_map);
6401 group = first_cpu(mask);
6403 *sg = &per_cpu(sched_group_core, group);
6406 #elif defined(CONFIG_SCHED_MC)
6408 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6411 *sg = &per_cpu(sched_group_core, cpu);
6416 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6417 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6420 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6423 #ifdef CONFIG_SCHED_MC
6424 cpumask_t mask = cpu_coregroup_map(cpu);
6425 cpus_and(mask, mask, *cpu_map);
6426 group = first_cpu(mask);
6427 #elif defined(CONFIG_SCHED_SMT)
6428 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6429 cpus_and(mask, mask, *cpu_map);
6430 group = first_cpu(mask);
6435 *sg = &per_cpu(sched_group_phys, group);
6441 * The init_sched_build_groups can't handle what we want to do with node
6442 * groups, so roll our own. Now each node has its own list of groups which
6443 * gets dynamically allocated.
6445 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6446 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6448 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6449 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6451 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6452 struct sched_group **sg)
6454 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6457 cpus_and(nodemask, nodemask, *cpu_map);
6458 group = first_cpu(nodemask);
6461 *sg = &per_cpu(sched_group_allnodes, group);
6465 static void init_numa_sched_groups_power(struct sched_group *group_head)
6467 struct sched_group *sg = group_head;
6473 for_each_cpu_mask(j, sg->cpumask) {
6474 struct sched_domain *sd;
6476 sd = &per_cpu(phys_domains, j);
6477 if (j != first_cpu(sd->groups->cpumask)) {
6479 * Only add "power" once for each
6485 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6488 } while (sg != group_head);
6493 /* Free memory allocated for various sched_group structures */
6494 static void free_sched_groups(const cpumask_t *cpu_map)
6498 for_each_cpu_mask(cpu, *cpu_map) {
6499 struct sched_group **sched_group_nodes
6500 = sched_group_nodes_bycpu[cpu];
6502 if (!sched_group_nodes)
6505 for (i = 0; i < MAX_NUMNODES; i++) {
6506 cpumask_t nodemask = node_to_cpumask(i);
6507 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6509 cpus_and(nodemask, nodemask, *cpu_map);
6510 if (cpus_empty(nodemask))
6520 if (oldsg != sched_group_nodes[i])
6523 kfree(sched_group_nodes);
6524 sched_group_nodes_bycpu[cpu] = NULL;
6528 static void free_sched_groups(const cpumask_t *cpu_map)
6534 * Initialize sched groups cpu_power.
6536 * cpu_power indicates the capacity of sched group, which is used while
6537 * distributing the load between different sched groups in a sched domain.
6538 * Typically cpu_power for all the groups in a sched domain will be same unless
6539 * there are asymmetries in the topology. If there are asymmetries, group
6540 * having more cpu_power will pickup more load compared to the group having
6543 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6544 * the maximum number of tasks a group can handle in the presence of other idle
6545 * or lightly loaded groups in the same sched domain.
6547 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6549 struct sched_domain *child;
6550 struct sched_group *group;
6552 WARN_ON(!sd || !sd->groups);
6554 if (cpu != first_cpu(sd->groups->cpumask))
6559 sd->groups->__cpu_power = 0;
6562 * For perf policy, if the groups in child domain share resources
6563 * (for example cores sharing some portions of the cache hierarchy
6564 * or SMT), then set this domain groups cpu_power such that each group
6565 * can handle only one task, when there are other idle groups in the
6566 * same sched domain.
6568 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6570 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6571 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6576 * add cpu_power of each child group to this groups cpu_power
6578 group = child->groups;
6580 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6581 group = group->next;
6582 } while (group != child->groups);
6586 * Build sched domains for a given set of cpus and attach the sched domains
6587 * to the individual cpus
6589 static int build_sched_domains(const cpumask_t *cpu_map)
6592 struct root_domain *rd;
6594 struct sched_group **sched_group_nodes = NULL;
6595 int sd_allnodes = 0;
6598 * Allocate the per-node list of sched groups
6600 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6602 if (!sched_group_nodes) {
6603 printk(KERN_WARNING "Can not alloc sched group node list\n");
6606 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6609 rd = alloc_rootdomain();
6611 printk(KERN_WARNING "Cannot alloc root domain\n");
6616 * Set up domains for cpus specified by the cpu_map.
6618 for_each_cpu_mask(i, *cpu_map) {
6619 struct sched_domain *sd = NULL, *p;
6620 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6622 cpus_and(nodemask, nodemask, *cpu_map);
6625 if (cpus_weight(*cpu_map) >
6626 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6627 sd = &per_cpu(allnodes_domains, i);
6628 *sd = SD_ALLNODES_INIT;
6629 sd->span = *cpu_map;
6630 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6636 sd = &per_cpu(node_domains, i);
6638 sd->span = sched_domain_node_span(cpu_to_node(i));
6642 cpus_and(sd->span, sd->span, *cpu_map);
6646 sd = &per_cpu(phys_domains, i);
6648 sd->span = nodemask;
6652 cpu_to_phys_group(i, cpu_map, &sd->groups);
6654 #ifdef CONFIG_SCHED_MC
6656 sd = &per_cpu(core_domains, i);
6658 sd->span = cpu_coregroup_map(i);
6659 cpus_and(sd->span, sd->span, *cpu_map);
6662 cpu_to_core_group(i, cpu_map, &sd->groups);
6665 #ifdef CONFIG_SCHED_SMT
6667 sd = &per_cpu(cpu_domains, i);
6668 *sd = SD_SIBLING_INIT;
6669 sd->span = per_cpu(cpu_sibling_map, i);
6670 cpus_and(sd->span, sd->span, *cpu_map);
6673 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6677 #ifdef CONFIG_SCHED_SMT
6678 /* Set up CPU (sibling) groups */
6679 for_each_cpu_mask(i, *cpu_map) {
6680 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6681 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6682 if (i != first_cpu(this_sibling_map))
6685 init_sched_build_groups(this_sibling_map, cpu_map,
6690 #ifdef CONFIG_SCHED_MC
6691 /* Set up multi-core groups */
6692 for_each_cpu_mask(i, *cpu_map) {
6693 cpumask_t this_core_map = cpu_coregroup_map(i);
6694 cpus_and(this_core_map, this_core_map, *cpu_map);
6695 if (i != first_cpu(this_core_map))
6697 init_sched_build_groups(this_core_map, cpu_map,
6698 &cpu_to_core_group);
6702 /* Set up physical groups */
6703 for (i = 0; i < MAX_NUMNODES; i++) {
6704 cpumask_t nodemask = node_to_cpumask(i);
6706 cpus_and(nodemask, nodemask, *cpu_map);
6707 if (cpus_empty(nodemask))
6710 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6714 /* Set up node groups */
6716 init_sched_build_groups(*cpu_map, cpu_map,
6717 &cpu_to_allnodes_group);
6719 for (i = 0; i < MAX_NUMNODES; i++) {
6720 /* Set up node groups */
6721 struct sched_group *sg, *prev;
6722 cpumask_t nodemask = node_to_cpumask(i);
6723 cpumask_t domainspan;
6724 cpumask_t covered = CPU_MASK_NONE;
6727 cpus_and(nodemask, nodemask, *cpu_map);
6728 if (cpus_empty(nodemask)) {
6729 sched_group_nodes[i] = NULL;
6733 domainspan = sched_domain_node_span(i);
6734 cpus_and(domainspan, domainspan, *cpu_map);
6736 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6738 printk(KERN_WARNING "Can not alloc domain group for "
6742 sched_group_nodes[i] = sg;
6743 for_each_cpu_mask(j, nodemask) {
6744 struct sched_domain *sd;
6746 sd = &per_cpu(node_domains, j);
6749 sg->__cpu_power = 0;
6750 sg->cpumask = nodemask;
6752 cpus_or(covered, covered, nodemask);
6755 for (j = 0; j < MAX_NUMNODES; j++) {
6756 cpumask_t tmp, notcovered;
6757 int n = (i + j) % MAX_NUMNODES;
6759 cpus_complement(notcovered, covered);
6760 cpus_and(tmp, notcovered, *cpu_map);
6761 cpus_and(tmp, tmp, domainspan);
6762 if (cpus_empty(tmp))
6765 nodemask = node_to_cpumask(n);
6766 cpus_and(tmp, tmp, nodemask);
6767 if (cpus_empty(tmp))
6770 sg = kmalloc_node(sizeof(struct sched_group),
6774 "Can not alloc domain group for node %d\n", j);
6777 sg->__cpu_power = 0;
6779 sg->next = prev->next;
6780 cpus_or(covered, covered, tmp);
6787 /* Calculate CPU power for physical packages and nodes */
6788 #ifdef CONFIG_SCHED_SMT
6789 for_each_cpu_mask(i, *cpu_map) {
6790 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6792 init_sched_groups_power(i, sd);
6795 #ifdef CONFIG_SCHED_MC
6796 for_each_cpu_mask(i, *cpu_map) {
6797 struct sched_domain *sd = &per_cpu(core_domains, i);
6799 init_sched_groups_power(i, sd);
6803 for_each_cpu_mask(i, *cpu_map) {
6804 struct sched_domain *sd = &per_cpu(phys_domains, i);
6806 init_sched_groups_power(i, sd);
6810 for (i = 0; i < MAX_NUMNODES; i++)
6811 init_numa_sched_groups_power(sched_group_nodes[i]);
6814 struct sched_group *sg;
6816 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6817 init_numa_sched_groups_power(sg);
6821 /* Attach the domains */
6822 for_each_cpu_mask(i, *cpu_map) {
6823 struct sched_domain *sd;
6824 #ifdef CONFIG_SCHED_SMT
6825 sd = &per_cpu(cpu_domains, i);
6826 #elif defined(CONFIG_SCHED_MC)
6827 sd = &per_cpu(core_domains, i);
6829 sd = &per_cpu(phys_domains, i);
6831 cpu_attach_domain(sd, rd, i);
6838 free_sched_groups(cpu_map);
6843 static cpumask_t *doms_cur; /* current sched domains */
6844 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6847 * Special case: If a kmalloc of a doms_cur partition (array of
6848 * cpumask_t) fails, then fallback to a single sched domain,
6849 * as determined by the single cpumask_t fallback_doms.
6851 static cpumask_t fallback_doms;
6853 void __attribute__((weak)) arch_update_cpu_topology(void)
6858 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6859 * For now this just excludes isolated cpus, but could be used to
6860 * exclude other special cases in the future.
6862 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6866 arch_update_cpu_topology();
6868 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6870 doms_cur = &fallback_doms;
6871 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6872 err = build_sched_domains(doms_cur);
6873 register_sched_domain_sysctl();
6878 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6880 free_sched_groups(cpu_map);
6884 * Detach sched domains from a group of cpus specified in cpu_map
6885 * These cpus will now be attached to the NULL domain
6887 static void detach_destroy_domains(const cpumask_t *cpu_map)
6891 unregister_sched_domain_sysctl();
6893 for_each_cpu_mask(i, *cpu_map)
6894 cpu_attach_domain(NULL, &def_root_domain, i);
6895 synchronize_sched();
6896 arch_destroy_sched_domains(cpu_map);
6900 * Partition sched domains as specified by the 'ndoms_new'
6901 * cpumasks in the array doms_new[] of cpumasks. This compares
6902 * doms_new[] to the current sched domain partitioning, doms_cur[].
6903 * It destroys each deleted domain and builds each new domain.
6905 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6906 * The masks don't intersect (don't overlap.) We should setup one
6907 * sched domain for each mask. CPUs not in any of the cpumasks will
6908 * not be load balanced. If the same cpumask appears both in the
6909 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6912 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6913 * ownership of it and will kfree it when done with it. If the caller
6914 * failed the kmalloc call, then it can pass in doms_new == NULL,
6915 * and partition_sched_domains() will fallback to the single partition
6918 * Call with hotplug lock held
6920 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6926 /* always unregister in case we don't destroy any domains */
6927 unregister_sched_domain_sysctl();
6929 if (doms_new == NULL) {
6931 doms_new = &fallback_doms;
6932 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6935 /* Destroy deleted domains */
6936 for (i = 0; i < ndoms_cur; i++) {
6937 for (j = 0; j < ndoms_new; j++) {
6938 if (cpus_equal(doms_cur[i], doms_new[j]))
6941 /* no match - a current sched domain not in new doms_new[] */
6942 detach_destroy_domains(doms_cur + i);
6947 /* Build new domains */
6948 for (i = 0; i < ndoms_new; i++) {
6949 for (j = 0; j < ndoms_cur; j++) {
6950 if (cpus_equal(doms_new[i], doms_cur[j]))
6953 /* no match - add a new doms_new */
6954 build_sched_domains(doms_new + i);
6959 /* Remember the new sched domains */
6960 if (doms_cur != &fallback_doms)
6962 doms_cur = doms_new;
6963 ndoms_cur = ndoms_new;
6965 register_sched_domain_sysctl();
6970 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6971 int arch_reinit_sched_domains(void)
6976 detach_destroy_domains(&cpu_online_map);
6977 err = arch_init_sched_domains(&cpu_online_map);
6983 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6987 if (buf[0] != '0' && buf[0] != '1')
6991 sched_smt_power_savings = (buf[0] == '1');
6993 sched_mc_power_savings = (buf[0] == '1');
6995 ret = arch_reinit_sched_domains();
6997 return ret ? ret : count;
7000 #ifdef CONFIG_SCHED_MC
7001 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7003 return sprintf(page, "%u\n", sched_mc_power_savings);
7005 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7006 const char *buf, size_t count)
7008 return sched_power_savings_store(buf, count, 0);
7010 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7011 sched_mc_power_savings_store);
7014 #ifdef CONFIG_SCHED_SMT
7015 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7017 return sprintf(page, "%u\n", sched_smt_power_savings);
7019 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7020 const char *buf, size_t count)
7022 return sched_power_savings_store(buf, count, 1);
7024 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7025 sched_smt_power_savings_store);
7028 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7032 #ifdef CONFIG_SCHED_SMT
7034 err = sysfs_create_file(&cls->kset.kobj,
7035 &attr_sched_smt_power_savings.attr);
7037 #ifdef CONFIG_SCHED_MC
7038 if (!err && mc_capable())
7039 err = sysfs_create_file(&cls->kset.kobj,
7040 &attr_sched_mc_power_savings.attr);
7047 * Force a reinitialization of the sched domains hierarchy. The domains
7048 * and groups cannot be updated in place without racing with the balancing
7049 * code, so we temporarily attach all running cpus to the NULL domain
7050 * which will prevent rebalancing while the sched domains are recalculated.
7052 static int update_sched_domains(struct notifier_block *nfb,
7053 unsigned long action, void *hcpu)
7056 case CPU_UP_PREPARE:
7057 case CPU_UP_PREPARE_FROZEN:
7058 case CPU_DOWN_PREPARE:
7059 case CPU_DOWN_PREPARE_FROZEN:
7060 detach_destroy_domains(&cpu_online_map);
7063 case CPU_UP_CANCELED:
7064 case CPU_UP_CANCELED_FROZEN:
7065 case CPU_DOWN_FAILED:
7066 case CPU_DOWN_FAILED_FROZEN:
7068 case CPU_ONLINE_FROZEN:
7070 case CPU_DEAD_FROZEN:
7072 * Fall through and re-initialise the domains.
7079 /* The hotplug lock is already held by cpu_up/cpu_down */
7080 arch_init_sched_domains(&cpu_online_map);
7085 void __init sched_init_smp(void)
7087 cpumask_t non_isolated_cpus;
7090 arch_init_sched_domains(&cpu_online_map);
7091 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7092 if (cpus_empty(non_isolated_cpus))
7093 cpu_set(smp_processor_id(), non_isolated_cpus);
7095 /* XXX: Theoretical race here - CPU may be hotplugged now */
7096 hotcpu_notifier(update_sched_domains, 0);
7098 /* Move init over to a non-isolated CPU */
7099 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7101 sched_init_granularity();
7104 void __init sched_init_smp(void)
7106 sched_init_granularity();
7108 #endif /* CONFIG_SMP */
7110 int in_sched_functions(unsigned long addr)
7112 return in_lock_functions(addr) ||
7113 (addr >= (unsigned long)__sched_text_start
7114 && addr < (unsigned long)__sched_text_end);
7117 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7119 cfs_rq->tasks_timeline = RB_ROOT;
7120 #ifdef CONFIG_FAIR_GROUP_SCHED
7123 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7126 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7128 struct rt_prio_array *array;
7131 array = &rt_rq->active;
7132 for (i = 0; i < MAX_RT_PRIO; i++) {
7133 INIT_LIST_HEAD(array->queue + i);
7134 __clear_bit(i, array->bitmap);
7136 /* delimiter for bitsearch: */
7137 __set_bit(MAX_RT_PRIO, array->bitmap);
7139 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7140 rt_rq->highest_prio = MAX_RT_PRIO;
7143 rt_rq->rt_nr_migratory = 0;
7144 rt_rq->overloaded = 0;
7148 rt_rq->rt_throttled = 0;
7150 #ifdef CONFIG_RT_GROUP_SCHED
7151 rt_rq->rt_nr_boosted = 0;
7156 #ifdef CONFIG_FAIR_GROUP_SCHED
7157 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7158 struct cfs_rq *cfs_rq, struct sched_entity *se,
7161 tg->cfs_rq[cpu] = cfs_rq;
7162 init_cfs_rq(cfs_rq, rq);
7165 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7168 se->cfs_rq = &rq->cfs;
7170 se->load.weight = tg->shares;
7171 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7176 #ifdef CONFIG_RT_GROUP_SCHED
7177 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7178 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7181 tg->rt_rq[cpu] = rt_rq;
7182 init_rt_rq(rt_rq, rq);
7184 rt_rq->rt_se = rt_se;
7186 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7188 tg->rt_se[cpu] = rt_se;
7189 rt_se->rt_rq = &rq->rt;
7190 rt_se->my_q = rt_rq;
7191 rt_se->parent = NULL;
7192 INIT_LIST_HEAD(&rt_se->run_list);
7196 void __init sched_init(void)
7198 int highest_cpu = 0;
7202 init_defrootdomain();
7205 #ifdef CONFIG_GROUP_SCHED
7206 list_add(&init_task_group.list, &task_groups);
7209 for_each_possible_cpu(i) {
7213 spin_lock_init(&rq->lock);
7214 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7217 init_cfs_rq(&rq->cfs, rq);
7218 init_rt_rq(&rq->rt, rq);
7219 #ifdef CONFIG_FAIR_GROUP_SCHED
7220 init_task_group.shares = init_task_group_load;
7221 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7222 init_tg_cfs_entry(rq, &init_task_group,
7223 &per_cpu(init_cfs_rq, i),
7224 &per_cpu(init_sched_entity, i), i, 1);
7227 #ifdef CONFIG_RT_GROUP_SCHED
7228 init_task_group.rt_runtime =
7229 sysctl_sched_rt_runtime * NSEC_PER_USEC;
7230 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7231 init_tg_rt_entry(rq, &init_task_group,
7232 &per_cpu(init_rt_rq, i),
7233 &per_cpu(init_sched_rt_entity, i), i, 1);
7235 rq->rt_period_expire = 0;
7236 rq->rt_throttled = 0;
7238 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7239 rq->cpu_load[j] = 0;
7243 rq->active_balance = 0;
7244 rq->next_balance = jiffies;
7247 rq->migration_thread = NULL;
7248 INIT_LIST_HEAD(&rq->migration_queue);
7249 rq_attach_root(rq, &def_root_domain);
7252 atomic_set(&rq->nr_iowait, 0);
7256 set_load_weight(&init_task);
7258 #ifdef CONFIG_PREEMPT_NOTIFIERS
7259 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7263 nr_cpu_ids = highest_cpu + 1;
7264 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7267 #ifdef CONFIG_RT_MUTEXES
7268 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7272 * The boot idle thread does lazy MMU switching as well:
7274 atomic_inc(&init_mm.mm_count);
7275 enter_lazy_tlb(&init_mm, current);
7278 * Make us the idle thread. Technically, schedule() should not be
7279 * called from this thread, however somewhere below it might be,
7280 * but because we are the idle thread, we just pick up running again
7281 * when this runqueue becomes "idle".
7283 init_idle(current, smp_processor_id());
7285 * During early bootup we pretend to be a normal task:
7287 current->sched_class = &fair_sched_class;
7289 scheduler_running = 1;
7292 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7293 void __might_sleep(char *file, int line)
7296 static unsigned long prev_jiffy; /* ratelimiting */
7298 if ((in_atomic() || irqs_disabled()) &&
7299 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7300 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7302 prev_jiffy = jiffies;
7303 printk(KERN_ERR "BUG: sleeping function called from invalid"
7304 " context at %s:%d\n", file, line);
7305 printk("in_atomic():%d, irqs_disabled():%d\n",
7306 in_atomic(), irqs_disabled());
7307 debug_show_held_locks(current);
7308 if (irqs_disabled())
7309 print_irqtrace_events(current);
7314 EXPORT_SYMBOL(__might_sleep);
7317 #ifdef CONFIG_MAGIC_SYSRQ
7318 static void normalize_task(struct rq *rq, struct task_struct *p)
7321 update_rq_clock(rq);
7322 on_rq = p->se.on_rq;
7324 deactivate_task(rq, p, 0);
7325 __setscheduler(rq, p, SCHED_NORMAL, 0);
7327 activate_task(rq, p, 0);
7328 resched_task(rq->curr);
7332 void normalize_rt_tasks(void)
7334 struct task_struct *g, *p;
7335 unsigned long flags;
7338 read_lock_irqsave(&tasklist_lock, flags);
7339 do_each_thread(g, p) {
7341 * Only normalize user tasks:
7346 p->se.exec_start = 0;
7347 #ifdef CONFIG_SCHEDSTATS
7348 p->se.wait_start = 0;
7349 p->se.sleep_start = 0;
7350 p->se.block_start = 0;
7352 task_rq(p)->clock = 0;
7356 * Renice negative nice level userspace
7359 if (TASK_NICE(p) < 0 && p->mm)
7360 set_user_nice(p, 0);
7364 spin_lock(&p->pi_lock);
7365 rq = __task_rq_lock(p);
7367 normalize_task(rq, p);
7369 __task_rq_unlock(rq);
7370 spin_unlock(&p->pi_lock);
7371 } while_each_thread(g, p);
7373 read_unlock_irqrestore(&tasklist_lock, flags);
7376 #endif /* CONFIG_MAGIC_SYSRQ */
7380 * These functions are only useful for the IA64 MCA handling.
7382 * They can only be called when the whole system has been
7383 * stopped - every CPU needs to be quiescent, and no scheduling
7384 * activity can take place. Using them for anything else would
7385 * be a serious bug, and as a result, they aren't even visible
7386 * under any other configuration.
7390 * curr_task - return the current task for a given cpu.
7391 * @cpu: the processor in question.
7393 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7395 struct task_struct *curr_task(int cpu)
7397 return cpu_curr(cpu);
7401 * set_curr_task - set the current task for a given cpu.
7402 * @cpu: the processor in question.
7403 * @p: the task pointer to set.
7405 * Description: This function must only be used when non-maskable interrupts
7406 * are serviced on a separate stack. It allows the architecture to switch the
7407 * notion of the current task on a cpu in a non-blocking manner. This function
7408 * must be called with all CPU's synchronized, and interrupts disabled, the
7409 * and caller must save the original value of the current task (see
7410 * curr_task() above) and restore that value before reenabling interrupts and
7411 * re-starting the system.
7413 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7415 void set_curr_task(int cpu, struct task_struct *p)
7422 #ifdef CONFIG_GROUP_SCHED
7424 #ifdef CONFIG_FAIR_GROUP_SCHED
7425 static void free_fair_sched_group(struct task_group *tg)
7429 for_each_possible_cpu(i) {
7431 kfree(tg->cfs_rq[i]);
7440 static int alloc_fair_sched_group(struct task_group *tg)
7442 struct cfs_rq *cfs_rq;
7443 struct sched_entity *se;
7447 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7450 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7454 tg->shares = NICE_0_LOAD;
7456 for_each_possible_cpu(i) {
7459 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7460 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7464 se = kmalloc_node(sizeof(struct sched_entity),
7465 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7469 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7478 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7480 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7481 &cpu_rq(cpu)->leaf_cfs_rq_list);
7484 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7486 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7489 static inline void free_fair_sched_group(struct task_group *tg)
7493 static inline int alloc_fair_sched_group(struct task_group *tg)
7498 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7502 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7507 #ifdef CONFIG_RT_GROUP_SCHED
7508 static void free_rt_sched_group(struct task_group *tg)
7512 for_each_possible_cpu(i) {
7514 kfree(tg->rt_rq[i]);
7516 kfree(tg->rt_se[i]);
7523 static int alloc_rt_sched_group(struct task_group *tg)
7525 struct rt_rq *rt_rq;
7526 struct sched_rt_entity *rt_se;
7530 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7533 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7539 for_each_possible_cpu(i) {
7542 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7543 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7547 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7548 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7552 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7561 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7563 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7564 &cpu_rq(cpu)->leaf_rt_rq_list);
7567 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7569 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7572 static inline void free_rt_sched_group(struct task_group *tg)
7576 static inline int alloc_rt_sched_group(struct task_group *tg)
7581 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7585 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7590 static void free_sched_group(struct task_group *tg)
7592 free_fair_sched_group(tg);
7593 free_rt_sched_group(tg);
7597 /* allocate runqueue etc for a new task group */
7598 struct task_group *sched_create_group(void)
7600 struct task_group *tg;
7601 unsigned long flags;
7604 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7606 return ERR_PTR(-ENOMEM);
7608 if (!alloc_fair_sched_group(tg))
7611 if (!alloc_rt_sched_group(tg))
7614 spin_lock_irqsave(&task_group_lock, flags);
7615 for_each_possible_cpu(i) {
7616 register_fair_sched_group(tg, i);
7617 register_rt_sched_group(tg, i);
7619 list_add_rcu(&tg->list, &task_groups);
7620 spin_unlock_irqrestore(&task_group_lock, flags);
7625 free_sched_group(tg);
7626 return ERR_PTR(-ENOMEM);
7629 /* rcu callback to free various structures associated with a task group */
7630 static void free_sched_group_rcu(struct rcu_head *rhp)
7632 /* now it should be safe to free those cfs_rqs */
7633 free_sched_group(container_of(rhp, struct task_group, rcu));
7636 /* Destroy runqueue etc associated with a task group */
7637 void sched_destroy_group(struct task_group *tg)
7639 unsigned long flags;
7642 spin_lock_irqsave(&task_group_lock, flags);
7643 for_each_possible_cpu(i) {
7644 unregister_fair_sched_group(tg, i);
7645 unregister_rt_sched_group(tg, i);
7647 list_del_rcu(&tg->list);
7648 spin_unlock_irqrestore(&task_group_lock, flags);
7650 /* wait for possible concurrent references to cfs_rqs complete */
7651 call_rcu(&tg->rcu, free_sched_group_rcu);
7654 /* change task's runqueue when it moves between groups.
7655 * The caller of this function should have put the task in its new group
7656 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7657 * reflect its new group.
7659 void sched_move_task(struct task_struct *tsk)
7662 unsigned long flags;
7665 rq = task_rq_lock(tsk, &flags);
7667 update_rq_clock(rq);
7669 running = task_current(rq, tsk);
7670 on_rq = tsk->se.on_rq;
7673 dequeue_task(rq, tsk, 0);
7674 if (unlikely(running))
7675 tsk->sched_class->put_prev_task(rq, tsk);
7677 set_task_rq(tsk, task_cpu(tsk));
7679 #ifdef CONFIG_FAIR_GROUP_SCHED
7680 if (tsk->sched_class->moved_group)
7681 tsk->sched_class->moved_group(tsk);
7684 if (unlikely(running))
7685 tsk->sched_class->set_curr_task(rq);
7687 enqueue_task(rq, tsk, 0);
7689 task_rq_unlock(rq, &flags);
7692 #ifdef CONFIG_FAIR_GROUP_SCHED
7693 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7695 struct cfs_rq *cfs_rq = se->cfs_rq;
7696 struct rq *rq = cfs_rq->rq;
7699 spin_lock_irq(&rq->lock);
7703 dequeue_entity(cfs_rq, se, 0);
7705 se->load.weight = shares;
7706 se->load.inv_weight = div64_64((1ULL<<32), shares);
7709 enqueue_entity(cfs_rq, se, 0);
7711 spin_unlock_irq(&rq->lock);
7714 static DEFINE_MUTEX(shares_mutex);
7716 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7719 unsigned long flags;
7722 * A weight of 0 or 1 can cause arithmetics problems.
7723 * (The default weight is 1024 - so there's no practical
7724 * limitation from this.)
7729 mutex_lock(&shares_mutex);
7730 if (tg->shares == shares)
7733 spin_lock_irqsave(&task_group_lock, flags);
7734 for_each_possible_cpu(i)
7735 unregister_fair_sched_group(tg, i);
7736 spin_unlock_irqrestore(&task_group_lock, flags);
7738 /* wait for any ongoing reference to this group to finish */
7739 synchronize_sched();
7742 * Now we are free to modify the group's share on each cpu
7743 * w/o tripping rebalance_share or load_balance_fair.
7745 tg->shares = shares;
7746 for_each_possible_cpu(i)
7747 set_se_shares(tg->se[i], shares);
7750 * Enable load balance activity on this group, by inserting it back on
7751 * each cpu's rq->leaf_cfs_rq_list.
7753 spin_lock_irqsave(&task_group_lock, flags);
7754 for_each_possible_cpu(i)
7755 register_fair_sched_group(tg, i);
7756 spin_unlock_irqrestore(&task_group_lock, flags);
7758 mutex_unlock(&shares_mutex);
7762 unsigned long sched_group_shares(struct task_group *tg)
7768 #ifdef CONFIG_RT_GROUP_SCHED
7770 * Ensure that the real time constraints are schedulable.
7772 static DEFINE_MUTEX(rt_constraints_mutex);
7774 static unsigned long to_ratio(u64 period, u64 runtime)
7776 if (runtime == RUNTIME_INF)
7779 return div64_64(runtime << 16, period);
7782 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7784 struct task_group *tgi;
7785 unsigned long total = 0;
7786 unsigned long global_ratio =
7787 to_ratio(sysctl_sched_rt_period,
7788 sysctl_sched_rt_runtime < 0 ?
7789 RUNTIME_INF : sysctl_sched_rt_runtime);
7792 list_for_each_entry_rcu(tgi, &task_groups, list) {
7796 total += to_ratio(period, tgi->rt_runtime);
7800 return total + to_ratio(period, runtime) < global_ratio;
7803 /* Must be called with tasklist_lock held */
7804 static inline int tg_has_rt_tasks(struct task_group *tg)
7806 struct task_struct *g, *p;
7807 do_each_thread(g, p) {
7808 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
7810 } while_each_thread(g, p);
7814 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7816 u64 rt_runtime, rt_period;
7819 rt_period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
7820 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7821 if (rt_runtime_us == -1)
7822 rt_runtime = RUNTIME_INF;
7824 mutex_lock(&rt_constraints_mutex);
7825 read_lock(&tasklist_lock);
7826 if (rt_runtime_us == 0 && tg_has_rt_tasks(tg)) {
7830 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7834 tg->rt_runtime = rt_runtime;
7836 read_unlock(&tasklist_lock);
7837 mutex_unlock(&rt_constraints_mutex);
7842 long sched_group_rt_runtime(struct task_group *tg)
7846 if (tg->rt_runtime == RUNTIME_INF)
7849 rt_runtime_us = tg->rt_runtime;
7850 do_div(rt_runtime_us, NSEC_PER_USEC);
7851 return rt_runtime_us;
7854 #endif /* CONFIG_GROUP_SCHED */
7856 #ifdef CONFIG_CGROUP_SCHED
7858 /* return corresponding task_group object of a cgroup */
7859 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7861 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7862 struct task_group, css);
7865 static struct cgroup_subsys_state *
7866 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7868 struct task_group *tg;
7870 if (!cgrp->parent) {
7871 /* This is early initialization for the top cgroup */
7872 init_task_group.css.cgroup = cgrp;
7873 return &init_task_group.css;
7876 /* we support only 1-level deep hierarchical scheduler atm */
7877 if (cgrp->parent->parent)
7878 return ERR_PTR(-EINVAL);
7880 tg = sched_create_group();
7882 return ERR_PTR(-ENOMEM);
7884 /* Bind the cgroup to task_group object we just created */
7885 tg->css.cgroup = cgrp;
7891 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7893 struct task_group *tg = cgroup_tg(cgrp);
7895 sched_destroy_group(tg);
7899 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7900 struct task_struct *tsk)
7902 #ifdef CONFIG_RT_GROUP_SCHED
7903 /* Don't accept realtime tasks when there is no way for them to run */
7904 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7907 /* We don't support RT-tasks being in separate groups */
7908 if (tsk->sched_class != &fair_sched_class)
7916 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7917 struct cgroup *old_cont, struct task_struct *tsk)
7919 sched_move_task(tsk);
7922 #ifdef CONFIG_FAIR_GROUP_SCHED
7923 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7926 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7929 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7931 struct task_group *tg = cgroup_tg(cgrp);
7933 return (u64) tg->shares;
7937 #ifdef CONFIG_RT_GROUP_SCHED
7938 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7940 const char __user *userbuf,
7941 size_t nbytes, loff_t *unused_ppos)
7950 if (nbytes >= sizeof(buffer))
7952 if (copy_from_user(buffer, userbuf, nbytes))
7955 buffer[nbytes] = 0; /* nul-terminate */
7957 /* strip newline if necessary */
7958 if (nbytes && (buffer[nbytes-1] == '\n'))
7959 buffer[nbytes-1] = 0;
7960 val = simple_strtoll(buffer, &end, 0);
7964 /* Pass to subsystem */
7965 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7971 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
7973 char __user *buf, size_t nbytes,
7977 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
7978 int len = sprintf(tmp, "%ld\n", val);
7980 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
7984 static struct cftype cpu_files[] = {
7985 #ifdef CONFIG_FAIR_GROUP_SCHED
7988 .read_uint = cpu_shares_read_uint,
7989 .write_uint = cpu_shares_write_uint,
7992 #ifdef CONFIG_RT_GROUP_SCHED
7994 .name = "rt_runtime_us",
7995 .read = cpu_rt_runtime_read,
7996 .write = cpu_rt_runtime_write,
8001 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8003 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8006 struct cgroup_subsys cpu_cgroup_subsys = {
8008 .create = cpu_cgroup_create,
8009 .destroy = cpu_cgroup_destroy,
8010 .can_attach = cpu_cgroup_can_attach,
8011 .attach = cpu_cgroup_attach,
8012 .populate = cpu_cgroup_populate,
8013 .subsys_id = cpu_cgroup_subsys_id,
8017 #endif /* CONFIG_CGROUP_SCHED */
8019 #ifdef CONFIG_CGROUP_CPUACCT
8022 * CPU accounting code for task groups.
8024 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8025 * (balbir@in.ibm.com).
8028 /* track cpu usage of a group of tasks */
8030 struct cgroup_subsys_state css;
8031 /* cpuusage holds pointer to a u64-type object on every cpu */
8035 struct cgroup_subsys cpuacct_subsys;
8037 /* return cpu accounting group corresponding to this container */
8038 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
8040 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
8041 struct cpuacct, css);
8044 /* return cpu accounting group to which this task belongs */
8045 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8047 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8048 struct cpuacct, css);
8051 /* create a new cpu accounting group */
8052 static struct cgroup_subsys_state *cpuacct_create(
8053 struct cgroup_subsys *ss, struct cgroup *cont)
8055 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8058 return ERR_PTR(-ENOMEM);
8060 ca->cpuusage = alloc_percpu(u64);
8061 if (!ca->cpuusage) {
8063 return ERR_PTR(-ENOMEM);
8069 /* destroy an existing cpu accounting group */
8071 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8073 struct cpuacct *ca = cgroup_ca(cont);
8075 free_percpu(ca->cpuusage);
8079 /* return total cpu usage (in nanoseconds) of a group */
8080 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8082 struct cpuacct *ca = cgroup_ca(cont);
8083 u64 totalcpuusage = 0;
8086 for_each_possible_cpu(i) {
8087 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8090 * Take rq->lock to make 64-bit addition safe on 32-bit
8093 spin_lock_irq(&cpu_rq(i)->lock);
8094 totalcpuusage += *cpuusage;
8095 spin_unlock_irq(&cpu_rq(i)->lock);
8098 return totalcpuusage;
8101 static struct cftype files[] = {
8104 .read_uint = cpuusage_read,
8108 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8110 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8114 * charge this task's execution time to its accounting group.
8116 * called with rq->lock held.
8118 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8122 if (!cpuacct_subsys.active)
8127 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8129 *cpuusage += cputime;
8133 struct cgroup_subsys cpuacct_subsys = {
8135 .create = cpuacct_create,
8136 .destroy = cpuacct_destroy,
8137 .populate = cpuacct_populate,
8138 .subsys_id = cpuacct_subsys_id,
8140 #endif /* CONFIG_CGROUP_CPUACCT */