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>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
73 #include <asm/irq_regs.h>
76 * Scheduler clock - returns current time in nanosec units.
77 * This is default implementation.
78 * Architectures and sub-architectures can override this.
80 unsigned long long __attribute__((weak)) sched_clock(void)
82 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
126 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
127 * Since cpu_power is a 'constant', we can use a reciprocal divide.
129 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
131 return reciprocal_divide(load, sg->reciprocal_cpu_power);
135 * Each time a sched group cpu_power is changed,
136 * we must compute its reciprocal value
138 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
140 sg->__cpu_power += val;
141 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
145 static inline int rt_policy(int policy)
147 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
152 static inline int task_has_rt_policy(struct task_struct *p)
154 return rt_policy(p->policy);
158 * This is the priority-queue data structure of the RT scheduling class:
160 struct rt_prio_array {
161 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
162 struct list_head queue[MAX_RT_PRIO];
165 struct rt_bandwidth {
168 spinlock_t rt_runtime_lock;
169 struct hrtimer rt_period_timer;
172 static struct rt_bandwidth def_rt_bandwidth;
174 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
176 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
178 struct rt_bandwidth *rt_b =
179 container_of(timer, struct rt_bandwidth, rt_period_timer);
185 now = hrtimer_cb_get_time(timer);
186 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
191 idle = do_sched_rt_period_timer(rt_b, overrun);
194 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
198 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
200 rt_b->rt_period = ns_to_ktime(period);
201 rt_b->rt_runtime = runtime;
203 spin_lock_init(&rt_b->rt_runtime_lock);
205 hrtimer_init(&rt_b->rt_period_timer,
206 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
207 rt_b->rt_period_timer.function = sched_rt_period_timer;
208 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
211 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
215 if (rt_b->rt_runtime == RUNTIME_INF)
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 spin_lock(&rt_b->rt_runtime_lock);
223 if (hrtimer_active(&rt_b->rt_period_timer))
226 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
227 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
228 hrtimer_start(&rt_b->rt_period_timer,
229 rt_b->rt_period_timer.expires,
232 spin_unlock(&rt_b->rt_runtime_lock);
235 #ifdef CONFIG_RT_GROUP_SCHED
236 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
238 hrtimer_cancel(&rt_b->rt_period_timer);
242 #ifdef CONFIG_GROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 #ifdef CONFIG_CGROUP_SCHED
253 struct cgroup_subsys_state css;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
275 #ifdef CONFIG_FAIR_GROUP_SCHED
276 /* Default task group's sched entity on each cpu */
277 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
278 /* Default task group's cfs_rq on each cpu */
279 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
284 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
287 /* task_group_lock serializes add/remove of task groups and also changes to
288 * a task group's cpu shares.
290 static DEFINE_SPINLOCK(task_group_lock);
292 /* doms_cur_mutex serializes access to doms_cur[] array */
293 static DEFINE_MUTEX(doms_cur_mutex);
295 #ifdef CONFIG_FAIR_GROUP_SCHED
296 #ifdef CONFIG_USER_SCHED
297 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
299 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 /* return group to which a task belongs */
311 static inline struct task_group *task_group(struct task_struct *p)
313 struct task_group *tg;
315 #ifdef CONFIG_USER_SCHED
317 #elif defined(CONFIG_CGROUP_SCHED)
318 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
319 struct task_group, css);
321 tg = &init_task_group;
326 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
327 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
329 #ifdef CONFIG_FAIR_GROUP_SCHED
330 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
331 p->se.parent = task_group(p)->se[cpu];
334 #ifdef CONFIG_RT_GROUP_SCHED
335 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
336 p->rt.parent = task_group(p)->rt_se[cpu];
340 static inline void lock_doms_cur(void)
342 mutex_lock(&doms_cur_mutex);
345 static inline void unlock_doms_cur(void)
347 mutex_unlock(&doms_cur_mutex);
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
353 static inline void lock_doms_cur(void) { }
354 static inline void unlock_doms_cur(void) { }
356 #endif /* CONFIG_GROUP_SCHED */
358 /* CFS-related fields in a runqueue */
360 struct load_weight load;
361 unsigned long nr_running;
366 struct rb_root tasks_timeline;
367 struct rb_node *rb_leftmost;
368 struct rb_node *rb_load_balance_curr;
369 /* 'curr' points to currently running entity on this cfs_rq.
370 * It is set to NULL otherwise (i.e when none are currently running).
372 struct sched_entity *curr, *next;
374 unsigned long nr_spread_over;
376 #ifdef CONFIG_FAIR_GROUP_SCHED
377 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
380 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
381 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
382 * (like users, containers etc.)
384 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
385 * list is used during load balance.
387 struct list_head leaf_cfs_rq_list;
388 struct task_group *tg; /* group that "owns" this runqueue */
392 /* Real-Time classes' related field in a runqueue: */
394 struct rt_prio_array active;
395 unsigned long rt_nr_running;
396 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
397 int highest_prio; /* highest queued rt task prio */
400 unsigned long rt_nr_migratory;
406 spinlock_t rt_runtime_lock;
408 #ifdef CONFIG_RT_GROUP_SCHED
409 unsigned long rt_nr_boosted;
412 struct list_head leaf_rt_rq_list;
413 struct task_group *tg;
414 struct sched_rt_entity *rt_se;
421 * We add the notion of a root-domain which will be used to define per-domain
422 * variables. Each exclusive cpuset essentially defines an island domain by
423 * fully partitioning the member cpus from any other cpuset. Whenever a new
424 * exclusive cpuset is created, we also create and attach a new root-domain
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned char idle_at_tick;
469 unsigned long last_tick_seen;
470 unsigned char in_nohz_recently;
472 /* capture load from *all* tasks on this cpu: */
473 struct load_weight load;
474 unsigned long nr_load_updates;
480 #ifdef CONFIG_FAIR_GROUP_SCHED
481 /* list of leaf cfs_rq on this cpu: */
482 struct list_head leaf_cfs_rq_list;
484 #ifdef CONFIG_RT_GROUP_SCHED
485 struct list_head leaf_rt_rq_list;
489 * This is part of a global counter where only the total sum
490 * over all CPUs matters. A task can increase this counter on
491 * one CPU and if it got migrated afterwards it may decrease
492 * it on another CPU. Always updated under the runqueue lock:
494 unsigned long nr_uninterruptible;
496 struct task_struct *curr, *idle;
497 unsigned long next_balance;
498 struct mm_struct *prev_mm;
500 u64 clock, prev_clock_raw;
503 unsigned int clock_warps, clock_overflows, clock_underflows;
505 unsigned int clock_deep_idle_events;
511 struct root_domain *rd;
512 struct sched_domain *sd;
514 /* For active balancing */
517 /* cpu of this runqueue: */
520 struct task_struct *migration_thread;
521 struct list_head migration_queue;
524 #ifdef CONFIG_SCHED_HRTICK
525 unsigned long hrtick_flags;
526 ktime_t hrtick_expire;
527 struct hrtimer hrtick_timer;
530 #ifdef CONFIG_SCHEDSTATS
532 struct sched_info rq_sched_info;
534 /* sys_sched_yield() stats */
535 unsigned int yld_exp_empty;
536 unsigned int yld_act_empty;
537 unsigned int yld_both_empty;
538 unsigned int yld_count;
540 /* schedule() stats */
541 unsigned int sched_switch;
542 unsigned int sched_count;
543 unsigned int sched_goidle;
545 /* try_to_wake_up() stats */
546 unsigned int ttwu_count;
547 unsigned int ttwu_local;
550 unsigned int bkl_count;
552 struct lock_class_key rq_lock_key;
555 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
557 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
559 rq->curr->sched_class->check_preempt_curr(rq, p);
562 static inline int cpu_of(struct rq *rq)
572 static inline bool nohz_on(int cpu)
574 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
577 static inline u64 max_skipped_ticks(struct rq *rq)
579 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
582 static inline void update_last_tick_seen(struct rq *rq)
584 rq->last_tick_seen = jiffies;
587 static inline u64 max_skipped_ticks(struct rq *rq)
592 static inline void update_last_tick_seen(struct rq *rq)
598 * Update the per-runqueue clock, as finegrained as the platform can give
599 * us, but without assuming monotonicity, etc.:
601 static void __update_rq_clock(struct rq *rq)
603 u64 prev_raw = rq->prev_clock_raw;
604 u64 now = sched_clock();
605 s64 delta = now - prev_raw;
606 u64 clock = rq->clock;
608 #ifdef CONFIG_SCHED_DEBUG
609 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
612 * Protect against sched_clock() occasionally going backwards:
614 if (unlikely(delta < 0)) {
619 * Catch too large forward jumps too:
621 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
622 u64 max_time = rq->tick_timestamp + max_jump;
624 if (unlikely(clock + delta > max_time)) {
625 if (clock < max_time)
629 rq->clock_overflows++;
631 if (unlikely(delta > rq->clock_max_delta))
632 rq->clock_max_delta = delta;
637 rq->prev_clock_raw = now;
641 static void update_rq_clock(struct rq *rq)
643 if (likely(smp_processor_id() == cpu_of(rq)))
644 __update_rq_clock(rq);
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
665 #ifdef CONFIG_SCHED_DEBUG
666 # define const_debug __read_mostly
668 # define const_debug static const
672 * Debugging: various feature bits
675 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
676 SCHED_FEAT_WAKEUP_PREEMPT = 2,
677 SCHED_FEAT_START_DEBIT = 4,
678 SCHED_FEAT_AFFINE_WAKEUPS = 8,
679 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
680 SCHED_FEAT_SYNC_WAKEUPS = 32,
681 SCHED_FEAT_HRTICK = 64,
682 SCHED_FEAT_DOUBLE_TICK = 128,
685 const_debug unsigned int sysctl_sched_features =
686 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
687 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
688 SCHED_FEAT_START_DEBIT * 1 |
689 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
690 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
691 SCHED_FEAT_SYNC_WAKEUPS * 1 |
692 SCHED_FEAT_HRTICK * 1 |
693 SCHED_FEAT_DOUBLE_TICK * 0;
695 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
698 * Number of tasks to iterate in a single balance run.
699 * Limited because this is done with IRQs disabled.
701 const_debug unsigned int sysctl_sched_nr_migrate = 32;
704 * period over which we measure -rt task cpu usage in us.
707 unsigned int sysctl_sched_rt_period = 1000000;
709 static __read_mostly int scheduler_running;
712 * part of the period that we allow rt tasks to run in us.
715 int sysctl_sched_rt_runtime = 950000;
717 static inline u64 global_rt_period(void)
719 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
722 static inline u64 global_rt_runtime(void)
724 if (sysctl_sched_rt_period < 0)
727 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
730 static const unsigned long long time_sync_thresh = 100000;
732 static DEFINE_PER_CPU(unsigned long long, time_offset);
733 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
736 * Global lock which we take every now and then to synchronize
737 * the CPUs time. This method is not warp-safe, but it's good
738 * enough to synchronize slowly diverging time sources and thus
739 * it's good enough for tracing:
741 static DEFINE_SPINLOCK(time_sync_lock);
742 static unsigned long long prev_global_time;
744 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
748 spin_lock_irqsave(&time_sync_lock, flags);
750 if (time < prev_global_time) {
751 per_cpu(time_offset, cpu) += prev_global_time - time;
752 time = prev_global_time;
754 prev_global_time = time;
757 spin_unlock_irqrestore(&time_sync_lock, flags);
762 static unsigned long long __cpu_clock(int cpu)
764 unsigned long long now;
769 * Only call sched_clock() if the scheduler has already been
770 * initialized (some code might call cpu_clock() very early):
772 if (unlikely(!scheduler_running))
775 local_irq_save(flags);
779 local_irq_restore(flags);
785 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
786 * clock constructed from sched_clock():
788 unsigned long long cpu_clock(int cpu)
790 unsigned long long prev_cpu_time, time, delta_time;
792 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
793 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
794 delta_time = time-prev_cpu_time;
796 if (unlikely(delta_time > time_sync_thresh))
797 time = __sync_cpu_clock(time, cpu);
801 EXPORT_SYMBOL_GPL(cpu_clock);
803 #ifndef prepare_arch_switch
804 # define prepare_arch_switch(next) do { } while (0)
806 #ifndef finish_arch_switch
807 # define finish_arch_switch(prev) do { } while (0)
810 static inline int task_current(struct rq *rq, struct task_struct *p)
812 return rq->curr == p;
815 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
816 static inline int task_running(struct rq *rq, struct task_struct *p)
818 return task_current(rq, p);
821 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
825 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
827 #ifdef CONFIG_DEBUG_SPINLOCK
828 /* this is a valid case when another task releases the spinlock */
829 rq->lock.owner = current;
832 * If we are tracking spinlock dependencies then we have to
833 * fix up the runqueue lock - which gets 'carried over' from
836 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
838 spin_unlock_irq(&rq->lock);
841 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
842 static inline int task_running(struct rq *rq, struct task_struct *p)
847 return task_current(rq, p);
851 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
855 * We can optimise this out completely for !SMP, because the
856 * SMP rebalancing from interrupt is the only thing that cares
861 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
862 spin_unlock_irq(&rq->lock);
864 spin_unlock(&rq->lock);
868 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
872 * After ->oncpu is cleared, the task can be moved to a different CPU.
873 * We must ensure this doesn't happen until the switch is completely
879 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
883 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
886 * __task_rq_lock - lock the runqueue a given task resides on.
887 * Must be called interrupts disabled.
889 static inline struct rq *__task_rq_lock(struct task_struct *p)
893 struct rq *rq = task_rq(p);
894 spin_lock(&rq->lock);
895 if (likely(rq == task_rq(p)))
897 spin_unlock(&rq->lock);
902 * task_rq_lock - lock the runqueue a given task resides on and disable
903 * interrupts. Note the ordering: we can safely lookup the task_rq without
904 * explicitly disabling preemption.
906 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
912 local_irq_save(*flags);
914 spin_lock(&rq->lock);
915 if (likely(rq == task_rq(p)))
917 spin_unlock_irqrestore(&rq->lock, *flags);
921 static void __task_rq_unlock(struct rq *rq)
924 spin_unlock(&rq->lock);
927 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
930 spin_unlock_irqrestore(&rq->lock, *flags);
934 * this_rq_lock - lock this runqueue and disable interrupts.
936 static struct rq *this_rq_lock(void)
943 spin_lock(&rq->lock);
949 * We are going deep-idle (irqs are disabled):
951 void sched_clock_idle_sleep_event(void)
953 struct rq *rq = cpu_rq(smp_processor_id());
955 spin_lock(&rq->lock);
956 __update_rq_clock(rq);
957 spin_unlock(&rq->lock);
958 rq->clock_deep_idle_events++;
960 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
963 * We just idled delta nanoseconds (called with irqs disabled):
965 void sched_clock_idle_wakeup_event(u64 delta_ns)
967 struct rq *rq = cpu_rq(smp_processor_id());
968 u64 now = sched_clock();
970 rq->idle_clock += delta_ns;
972 * Override the previous timestamp and ignore all
973 * sched_clock() deltas that occured while we idled,
974 * and use the PM-provided delta_ns to advance the
977 spin_lock(&rq->lock);
978 rq->prev_clock_raw = now;
979 rq->clock += delta_ns;
980 spin_unlock(&rq->lock);
981 touch_softlockup_watchdog();
983 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
985 static void __resched_task(struct task_struct *p, int tif_bit);
987 static inline void resched_task(struct task_struct *p)
989 __resched_task(p, TIF_NEED_RESCHED);
992 #ifdef CONFIG_SCHED_HRTICK
994 * Use HR-timers to deliver accurate preemption points.
996 * Its all a bit involved since we cannot program an hrt while holding the
997 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1000 * When we get rescheduled we reprogram the hrtick_timer outside of the
1003 static inline void resched_hrt(struct task_struct *p)
1005 __resched_task(p, TIF_HRTICK_RESCHED);
1008 static inline void resched_rq(struct rq *rq)
1010 unsigned long flags;
1012 spin_lock_irqsave(&rq->lock, flags);
1013 resched_task(rq->curr);
1014 spin_unlock_irqrestore(&rq->lock, flags);
1018 HRTICK_SET, /* re-programm hrtick_timer */
1019 HRTICK_RESET, /* not a new slice */
1024 * - enabled by features
1025 * - hrtimer is actually high res
1027 static inline int hrtick_enabled(struct rq *rq)
1029 if (!sched_feat(HRTICK))
1031 return hrtimer_is_hres_active(&rq->hrtick_timer);
1035 * Called to set the hrtick timer state.
1037 * called with rq->lock held and irqs disabled
1039 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1041 assert_spin_locked(&rq->lock);
1044 * preempt at: now + delay
1047 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1049 * indicate we need to program the timer
1051 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1053 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1056 * New slices are called from the schedule path and don't need a
1057 * forced reschedule.
1060 resched_hrt(rq->curr);
1063 static void hrtick_clear(struct rq *rq)
1065 if (hrtimer_active(&rq->hrtick_timer))
1066 hrtimer_cancel(&rq->hrtick_timer);
1070 * Update the timer from the possible pending state.
1072 static void hrtick_set(struct rq *rq)
1076 unsigned long flags;
1078 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1080 spin_lock_irqsave(&rq->lock, flags);
1081 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1082 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1083 time = rq->hrtick_expire;
1084 clear_thread_flag(TIF_HRTICK_RESCHED);
1085 spin_unlock_irqrestore(&rq->lock, flags);
1088 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1089 if (reset && !hrtimer_active(&rq->hrtick_timer))
1096 * High-resolution timer tick.
1097 * Runs from hardirq context with interrupts disabled.
1099 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1101 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1103 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1105 spin_lock(&rq->lock);
1106 __update_rq_clock(rq);
1107 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1108 spin_unlock(&rq->lock);
1110 return HRTIMER_NORESTART;
1113 static inline void init_rq_hrtick(struct rq *rq)
1115 rq->hrtick_flags = 0;
1116 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1117 rq->hrtick_timer.function = hrtick;
1118 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1121 void hrtick_resched(void)
1124 unsigned long flags;
1126 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1129 local_irq_save(flags);
1130 rq = cpu_rq(smp_processor_id());
1132 local_irq_restore(flags);
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void hrtick_set(struct rq *rq)
1143 static inline void init_rq_hrtick(struct rq *rq)
1147 void hrtick_resched(void)
1153 * resched_task - mark a task 'to be rescheduled now'.
1155 * On UP this means the setting of the need_resched flag, on SMP it
1156 * might also involve a cross-CPU call to trigger the scheduler on
1161 #ifndef tsk_is_polling
1162 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1165 static void __resched_task(struct task_struct *p, int tif_bit)
1169 assert_spin_locked(&task_rq(p)->lock);
1171 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1174 set_tsk_thread_flag(p, tif_bit);
1177 if (cpu == smp_processor_id())
1180 /* NEED_RESCHED must be visible before we test polling */
1182 if (!tsk_is_polling(p))
1183 smp_send_reschedule(cpu);
1186 static void resched_cpu(int cpu)
1188 struct rq *rq = cpu_rq(cpu);
1189 unsigned long flags;
1191 if (!spin_trylock_irqsave(&rq->lock, flags))
1193 resched_task(cpu_curr(cpu));
1194 spin_unlock_irqrestore(&rq->lock, flags);
1199 * When add_timer_on() enqueues a timer into the timer wheel of an
1200 * idle CPU then this timer might expire before the next timer event
1201 * which is scheduled to wake up that CPU. In case of a completely
1202 * idle system the next event might even be infinite time into the
1203 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1204 * leaves the inner idle loop so the newly added timer is taken into
1205 * account when the CPU goes back to idle and evaluates the timer
1206 * wheel for the next timer event.
1208 void wake_up_idle_cpu(int cpu)
1210 struct rq *rq = cpu_rq(cpu);
1212 if (cpu == smp_processor_id())
1216 * This is safe, as this function is called with the timer
1217 * wheel base lock of (cpu) held. When the CPU is on the way
1218 * to idle and has not yet set rq->curr to idle then it will
1219 * be serialized on the timer wheel base lock and take the new
1220 * timer into account automatically.
1222 if (rq->curr != rq->idle)
1226 * We can set TIF_RESCHED on the idle task of the other CPU
1227 * lockless. The worst case is that the other CPU runs the
1228 * idle task through an additional NOOP schedule()
1230 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1232 /* NEED_RESCHED must be visible before we test polling */
1234 if (!tsk_is_polling(rq->idle))
1235 smp_send_reschedule(cpu);
1240 static void __resched_task(struct task_struct *p, int tif_bit)
1242 assert_spin_locked(&task_rq(p)->lock);
1243 set_tsk_thread_flag(p, tif_bit);
1247 #if BITS_PER_LONG == 32
1248 # define WMULT_CONST (~0UL)
1250 # define WMULT_CONST (1UL << 32)
1253 #define WMULT_SHIFT 32
1256 * Shift right and round:
1258 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1260 static unsigned long
1261 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1262 struct load_weight *lw)
1266 if (unlikely(!lw->inv_weight))
1267 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1269 tmp = (u64)delta_exec * weight;
1271 * Check whether we'd overflow the 64-bit multiplication:
1273 if (unlikely(tmp > WMULT_CONST))
1274 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1277 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1279 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1282 static inline unsigned long
1283 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1285 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1288 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1294 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1301 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1302 * of tasks with abnormal "nice" values across CPUs the contribution that
1303 * each task makes to its run queue's load is weighted according to its
1304 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1305 * scaled version of the new time slice allocation that they receive on time
1309 #define WEIGHT_IDLEPRIO 2
1310 #define WMULT_IDLEPRIO (1 << 31)
1313 * Nice levels are multiplicative, with a gentle 10% change for every
1314 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1315 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1316 * that remained on nice 0.
1318 * The "10% effect" is relative and cumulative: from _any_ nice level,
1319 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1320 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1321 * If a task goes up by ~10% and another task goes down by ~10% then
1322 * the relative distance between them is ~25%.)
1324 static const int prio_to_weight[40] = {
1325 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1326 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1327 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1328 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1329 /* 0 */ 1024, 820, 655, 526, 423,
1330 /* 5 */ 335, 272, 215, 172, 137,
1331 /* 10 */ 110, 87, 70, 56, 45,
1332 /* 15 */ 36, 29, 23, 18, 15,
1336 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1338 * In cases where the weight does not change often, we can use the
1339 * precalculated inverse to speed up arithmetics by turning divisions
1340 * into multiplications:
1342 static const u32 prio_to_wmult[40] = {
1343 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1344 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1345 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1346 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1347 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1348 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1349 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1350 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1353 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1356 * runqueue iterator, to support SMP load-balancing between different
1357 * scheduling classes, without having to expose their internal data
1358 * structures to the load-balancing proper:
1360 struct rq_iterator {
1362 struct task_struct *(*start)(void *);
1363 struct task_struct *(*next)(void *);
1367 static unsigned long
1368 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1369 unsigned long max_load_move, struct sched_domain *sd,
1370 enum cpu_idle_type idle, int *all_pinned,
1371 int *this_best_prio, struct rq_iterator *iterator);
1374 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1375 struct sched_domain *sd, enum cpu_idle_type idle,
1376 struct rq_iterator *iterator);
1379 #ifdef CONFIG_CGROUP_CPUACCT
1380 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1382 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1386 static unsigned long source_load(int cpu, int type);
1387 static unsigned long target_load(int cpu, int type);
1388 static unsigned long cpu_avg_load_per_task(int cpu);
1389 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1390 #endif /* CONFIG_SMP */
1392 #include "sched_stats.h"
1393 #include "sched_idletask.c"
1394 #include "sched_fair.c"
1395 #include "sched_rt.c"
1396 #ifdef CONFIG_SCHED_DEBUG
1397 # include "sched_debug.c"
1400 #define sched_class_highest (&rt_sched_class)
1402 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1404 update_load_add(&rq->load, p->se.load.weight);
1407 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1409 update_load_sub(&rq->load, p->se.load.weight);
1412 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1418 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1424 static void set_load_weight(struct task_struct *p)
1426 if (task_has_rt_policy(p)) {
1427 p->se.load.weight = prio_to_weight[0] * 2;
1428 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1433 * SCHED_IDLE tasks get minimal weight:
1435 if (p->policy == SCHED_IDLE) {
1436 p->se.load.weight = WEIGHT_IDLEPRIO;
1437 p->se.load.inv_weight = WMULT_IDLEPRIO;
1441 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1442 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1445 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1447 sched_info_queued(p);
1448 p->sched_class->enqueue_task(rq, p, wakeup);
1452 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1454 p->sched_class->dequeue_task(rq, p, sleep);
1459 * __normal_prio - return the priority that is based on the static prio
1461 static inline int __normal_prio(struct task_struct *p)
1463 return p->static_prio;
1467 * Calculate the expected normal priority: i.e. priority
1468 * without taking RT-inheritance into account. Might be
1469 * boosted by interactivity modifiers. Changes upon fork,
1470 * setprio syscalls, and whenever the interactivity
1471 * estimator recalculates.
1473 static inline int normal_prio(struct task_struct *p)
1477 if (task_has_rt_policy(p))
1478 prio = MAX_RT_PRIO-1 - p->rt_priority;
1480 prio = __normal_prio(p);
1485 * Calculate the current priority, i.e. the priority
1486 * taken into account by the scheduler. This value might
1487 * be boosted by RT tasks, or might be boosted by
1488 * interactivity modifiers. Will be RT if the task got
1489 * RT-boosted. If not then it returns p->normal_prio.
1491 static int effective_prio(struct task_struct *p)
1493 p->normal_prio = normal_prio(p);
1495 * If we are RT tasks or we were boosted to RT priority,
1496 * keep the priority unchanged. Otherwise, update priority
1497 * to the normal priority:
1499 if (!rt_prio(p->prio))
1500 return p->normal_prio;
1505 * activate_task - move a task to the runqueue.
1507 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1509 if (task_contributes_to_load(p))
1510 rq->nr_uninterruptible--;
1512 enqueue_task(rq, p, wakeup);
1513 inc_nr_running(p, rq);
1517 * deactivate_task - remove a task from the runqueue.
1519 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1521 if (task_contributes_to_load(p))
1522 rq->nr_uninterruptible++;
1524 dequeue_task(rq, p, sleep);
1525 dec_nr_running(p, rq);
1529 * task_curr - is this task currently executing on a CPU?
1530 * @p: the task in question.
1532 inline int task_curr(const struct task_struct *p)
1534 return cpu_curr(task_cpu(p)) == p;
1537 /* Used instead of source_load when we know the type == 0 */
1538 unsigned long weighted_cpuload(const int cpu)
1540 return cpu_rq(cpu)->load.weight;
1543 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1545 set_task_rq(p, cpu);
1548 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1549 * successfuly executed on another CPU. We must ensure that updates of
1550 * per-task data have been completed by this moment.
1553 task_thread_info(p)->cpu = cpu;
1557 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1558 const struct sched_class *prev_class,
1559 int oldprio, int running)
1561 if (prev_class != p->sched_class) {
1562 if (prev_class->switched_from)
1563 prev_class->switched_from(rq, p, running);
1564 p->sched_class->switched_to(rq, p, running);
1566 p->sched_class->prio_changed(rq, p, oldprio, running);
1572 * Is this task likely cache-hot:
1575 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1580 * Buddy candidates are cache hot:
1582 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1585 if (p->sched_class != &fair_sched_class)
1588 if (sysctl_sched_migration_cost == -1)
1590 if (sysctl_sched_migration_cost == 0)
1593 delta = now - p->se.exec_start;
1595 return delta < (s64)sysctl_sched_migration_cost;
1599 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1601 int old_cpu = task_cpu(p);
1602 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1603 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1604 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1607 clock_offset = old_rq->clock - new_rq->clock;
1609 #ifdef CONFIG_SCHEDSTATS
1610 if (p->se.wait_start)
1611 p->se.wait_start -= clock_offset;
1612 if (p->se.sleep_start)
1613 p->se.sleep_start -= clock_offset;
1614 if (p->se.block_start)
1615 p->se.block_start -= clock_offset;
1616 if (old_cpu != new_cpu) {
1617 schedstat_inc(p, se.nr_migrations);
1618 if (task_hot(p, old_rq->clock, NULL))
1619 schedstat_inc(p, se.nr_forced2_migrations);
1622 p->se.vruntime -= old_cfsrq->min_vruntime -
1623 new_cfsrq->min_vruntime;
1625 __set_task_cpu(p, new_cpu);
1628 struct migration_req {
1629 struct list_head list;
1631 struct task_struct *task;
1634 struct completion done;
1638 * The task's runqueue lock must be held.
1639 * Returns true if you have to wait for migration thread.
1642 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1644 struct rq *rq = task_rq(p);
1647 * If the task is not on a runqueue (and not running), then
1648 * it is sufficient to simply update the task's cpu field.
1650 if (!p->se.on_rq && !task_running(rq, p)) {
1651 set_task_cpu(p, dest_cpu);
1655 init_completion(&req->done);
1657 req->dest_cpu = dest_cpu;
1658 list_add(&req->list, &rq->migration_queue);
1664 * wait_task_inactive - wait for a thread to unschedule.
1666 * The caller must ensure that the task *will* unschedule sometime soon,
1667 * else this function might spin for a *long* time. This function can't
1668 * be called with interrupts off, or it may introduce deadlock with
1669 * smp_call_function() if an IPI is sent by the same process we are
1670 * waiting to become inactive.
1672 void wait_task_inactive(struct task_struct *p)
1674 unsigned long flags;
1680 * We do the initial early heuristics without holding
1681 * any task-queue locks at all. We'll only try to get
1682 * the runqueue lock when things look like they will
1688 * If the task is actively running on another CPU
1689 * still, just relax and busy-wait without holding
1692 * NOTE! Since we don't hold any locks, it's not
1693 * even sure that "rq" stays as the right runqueue!
1694 * But we don't care, since "task_running()" will
1695 * return false if the runqueue has changed and p
1696 * is actually now running somewhere else!
1698 while (task_running(rq, p))
1702 * Ok, time to look more closely! We need the rq
1703 * lock now, to be *sure*. If we're wrong, we'll
1704 * just go back and repeat.
1706 rq = task_rq_lock(p, &flags);
1707 running = task_running(rq, p);
1708 on_rq = p->se.on_rq;
1709 task_rq_unlock(rq, &flags);
1712 * Was it really running after all now that we
1713 * checked with the proper locks actually held?
1715 * Oops. Go back and try again..
1717 if (unlikely(running)) {
1723 * It's not enough that it's not actively running,
1724 * it must be off the runqueue _entirely_, and not
1727 * So if it wa still runnable (but just not actively
1728 * running right now), it's preempted, and we should
1729 * yield - it could be a while.
1731 if (unlikely(on_rq)) {
1732 schedule_timeout_uninterruptible(1);
1737 * Ahh, all good. It wasn't running, and it wasn't
1738 * runnable, which means that it will never become
1739 * running in the future either. We're all done!
1746 * kick_process - kick a running thread to enter/exit the kernel
1747 * @p: the to-be-kicked thread
1749 * Cause a process which is running on another CPU to enter
1750 * kernel-mode, without any delay. (to get signals handled.)
1752 * NOTE: this function doesnt have to take the runqueue lock,
1753 * because all it wants to ensure is that the remote task enters
1754 * the kernel. If the IPI races and the task has been migrated
1755 * to another CPU then no harm is done and the purpose has been
1758 void kick_process(struct task_struct *p)
1764 if ((cpu != smp_processor_id()) && task_curr(p))
1765 smp_send_reschedule(cpu);
1770 * Return a low guess at the load of a migration-source cpu weighted
1771 * according to the scheduling class and "nice" value.
1773 * We want to under-estimate the load of migration sources, to
1774 * balance conservatively.
1776 static unsigned long source_load(int cpu, int type)
1778 struct rq *rq = cpu_rq(cpu);
1779 unsigned long total = weighted_cpuload(cpu);
1784 return min(rq->cpu_load[type-1], total);
1788 * Return a high guess at the load of a migration-target cpu weighted
1789 * according to the scheduling class and "nice" value.
1791 static unsigned long target_load(int cpu, int type)
1793 struct rq *rq = cpu_rq(cpu);
1794 unsigned long total = weighted_cpuload(cpu);
1799 return max(rq->cpu_load[type-1], total);
1803 * Return the average load per task on the cpu's run queue
1805 static unsigned long cpu_avg_load_per_task(int cpu)
1807 struct rq *rq = cpu_rq(cpu);
1808 unsigned long total = weighted_cpuload(cpu);
1809 unsigned long n = rq->nr_running;
1811 return n ? total / n : SCHED_LOAD_SCALE;
1815 * find_idlest_group finds and returns the least busy CPU group within the
1818 static struct sched_group *
1819 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1821 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1822 unsigned long min_load = ULONG_MAX, this_load = 0;
1823 int load_idx = sd->forkexec_idx;
1824 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1827 unsigned long load, avg_load;
1831 /* Skip over this group if it has no CPUs allowed */
1832 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1835 local_group = cpu_isset(this_cpu, group->cpumask);
1837 /* Tally up the load of all CPUs in the group */
1840 for_each_cpu_mask(i, group->cpumask) {
1841 /* Bias balancing toward cpus of our domain */
1843 load = source_load(i, load_idx);
1845 load = target_load(i, load_idx);
1850 /* Adjust by relative CPU power of the group */
1851 avg_load = sg_div_cpu_power(group,
1852 avg_load * SCHED_LOAD_SCALE);
1855 this_load = avg_load;
1857 } else if (avg_load < min_load) {
1858 min_load = avg_load;
1861 } while (group = group->next, group != sd->groups);
1863 if (!idlest || 100*this_load < imbalance*min_load)
1869 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1872 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1875 unsigned long load, min_load = ULONG_MAX;
1879 /* Traverse only the allowed CPUs */
1880 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1882 for_each_cpu_mask(i, *tmp) {
1883 load = weighted_cpuload(i);
1885 if (load < min_load || (load == min_load && i == this_cpu)) {
1895 * sched_balance_self: balance the current task (running on cpu) in domains
1896 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1899 * Balance, ie. select the least loaded group.
1901 * Returns the target CPU number, or the same CPU if no balancing is needed.
1903 * preempt must be disabled.
1905 static int sched_balance_self(int cpu, int flag)
1907 struct task_struct *t = current;
1908 struct sched_domain *tmp, *sd = NULL;
1910 for_each_domain(cpu, tmp) {
1912 * If power savings logic is enabled for a domain, stop there.
1914 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1916 if (tmp->flags & flag)
1921 cpumask_t span, tmpmask;
1922 struct sched_group *group;
1923 int new_cpu, weight;
1925 if (!(sd->flags & flag)) {
1931 group = find_idlest_group(sd, t, cpu);
1937 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1938 if (new_cpu == -1 || new_cpu == cpu) {
1939 /* Now try balancing at a lower domain level of cpu */
1944 /* Now try balancing at a lower domain level of new_cpu */
1947 weight = cpus_weight(span);
1948 for_each_domain(cpu, tmp) {
1949 if (weight <= cpus_weight(tmp->span))
1951 if (tmp->flags & flag)
1954 /* while loop will break here if sd == NULL */
1960 #endif /* CONFIG_SMP */
1963 * try_to_wake_up - wake up a thread
1964 * @p: the to-be-woken-up thread
1965 * @state: the mask of task states that can be woken
1966 * @sync: do a synchronous wakeup?
1968 * Put it on the run-queue if it's not already there. The "current"
1969 * thread is always on the run-queue (except when the actual
1970 * re-schedule is in progress), and as such you're allowed to do
1971 * the simpler "current->state = TASK_RUNNING" to mark yourself
1972 * runnable without the overhead of this.
1974 * returns failure only if the task is already active.
1976 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1978 int cpu, orig_cpu, this_cpu, success = 0;
1979 unsigned long flags;
1983 if (!sched_feat(SYNC_WAKEUPS))
1987 rq = task_rq_lock(p, &flags);
1988 old_state = p->state;
1989 if (!(old_state & state))
1997 this_cpu = smp_processor_id();
2000 if (unlikely(task_running(rq, p)))
2003 cpu = p->sched_class->select_task_rq(p, sync);
2004 if (cpu != orig_cpu) {
2005 set_task_cpu(p, cpu);
2006 task_rq_unlock(rq, &flags);
2007 /* might preempt at this point */
2008 rq = task_rq_lock(p, &flags);
2009 old_state = p->state;
2010 if (!(old_state & state))
2015 this_cpu = smp_processor_id();
2019 #ifdef CONFIG_SCHEDSTATS
2020 schedstat_inc(rq, ttwu_count);
2021 if (cpu == this_cpu)
2022 schedstat_inc(rq, ttwu_local);
2024 struct sched_domain *sd;
2025 for_each_domain(this_cpu, sd) {
2026 if (cpu_isset(cpu, sd->span)) {
2027 schedstat_inc(sd, ttwu_wake_remote);
2035 #endif /* CONFIG_SMP */
2036 schedstat_inc(p, se.nr_wakeups);
2038 schedstat_inc(p, se.nr_wakeups_sync);
2039 if (orig_cpu != cpu)
2040 schedstat_inc(p, se.nr_wakeups_migrate);
2041 if (cpu == this_cpu)
2042 schedstat_inc(p, se.nr_wakeups_local);
2044 schedstat_inc(p, se.nr_wakeups_remote);
2045 update_rq_clock(rq);
2046 activate_task(rq, p, 1);
2050 check_preempt_curr(rq, p);
2052 p->state = TASK_RUNNING;
2054 if (p->sched_class->task_wake_up)
2055 p->sched_class->task_wake_up(rq, p);
2058 task_rq_unlock(rq, &flags);
2063 int wake_up_process(struct task_struct *p)
2065 return try_to_wake_up(p, TASK_ALL, 0);
2067 EXPORT_SYMBOL(wake_up_process);
2069 int wake_up_state(struct task_struct *p, unsigned int state)
2071 return try_to_wake_up(p, state, 0);
2075 * Perform scheduler related setup for a newly forked process p.
2076 * p is forked by current.
2078 * __sched_fork() is basic setup used by init_idle() too:
2080 static void __sched_fork(struct task_struct *p)
2082 p->se.exec_start = 0;
2083 p->se.sum_exec_runtime = 0;
2084 p->se.prev_sum_exec_runtime = 0;
2085 p->se.last_wakeup = 0;
2086 p->se.avg_overlap = 0;
2088 #ifdef CONFIG_SCHEDSTATS
2089 p->se.wait_start = 0;
2090 p->se.sum_sleep_runtime = 0;
2091 p->se.sleep_start = 0;
2092 p->se.block_start = 0;
2093 p->se.sleep_max = 0;
2094 p->se.block_max = 0;
2096 p->se.slice_max = 0;
2100 INIT_LIST_HEAD(&p->rt.run_list);
2103 #ifdef CONFIG_PREEMPT_NOTIFIERS
2104 INIT_HLIST_HEAD(&p->preempt_notifiers);
2108 * We mark the process as running here, but have not actually
2109 * inserted it onto the runqueue yet. This guarantees that
2110 * nobody will actually run it, and a signal or other external
2111 * event cannot wake it up and insert it on the runqueue either.
2113 p->state = TASK_RUNNING;
2117 * fork()/clone()-time setup:
2119 void sched_fork(struct task_struct *p, int clone_flags)
2121 int cpu = get_cpu();
2126 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2128 set_task_cpu(p, cpu);
2131 * Make sure we do not leak PI boosting priority to the child:
2133 p->prio = current->normal_prio;
2134 if (!rt_prio(p->prio))
2135 p->sched_class = &fair_sched_class;
2137 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2138 if (likely(sched_info_on()))
2139 memset(&p->sched_info, 0, sizeof(p->sched_info));
2141 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2144 #ifdef CONFIG_PREEMPT
2145 /* Want to start with kernel preemption disabled. */
2146 task_thread_info(p)->preempt_count = 1;
2152 * wake_up_new_task - wake up a newly created task for the first time.
2154 * This function will do some initial scheduler statistics housekeeping
2155 * that must be done for every newly created context, then puts the task
2156 * on the runqueue and wakes it.
2158 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2160 unsigned long flags;
2163 rq = task_rq_lock(p, &flags);
2164 BUG_ON(p->state != TASK_RUNNING);
2165 update_rq_clock(rq);
2167 p->prio = effective_prio(p);
2169 if (!p->sched_class->task_new || !current->se.on_rq) {
2170 activate_task(rq, p, 0);
2173 * Let the scheduling class do new task startup
2174 * management (if any):
2176 p->sched_class->task_new(rq, p);
2177 inc_nr_running(p, rq);
2179 check_preempt_curr(rq, p);
2181 if (p->sched_class->task_wake_up)
2182 p->sched_class->task_wake_up(rq, p);
2184 task_rq_unlock(rq, &flags);
2187 #ifdef CONFIG_PREEMPT_NOTIFIERS
2190 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2191 * @notifier: notifier struct to register
2193 void preempt_notifier_register(struct preempt_notifier *notifier)
2195 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2197 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2200 * preempt_notifier_unregister - no longer interested in preemption notifications
2201 * @notifier: notifier struct to unregister
2203 * This is safe to call from within a preemption notifier.
2205 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2207 hlist_del(¬ifier->link);
2209 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2211 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2213 struct preempt_notifier *notifier;
2214 struct hlist_node *node;
2216 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2217 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2221 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2222 struct task_struct *next)
2224 struct preempt_notifier *notifier;
2225 struct hlist_node *node;
2227 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2228 notifier->ops->sched_out(notifier, next);
2233 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2238 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2239 struct task_struct *next)
2246 * prepare_task_switch - prepare to switch tasks
2247 * @rq: the runqueue preparing to switch
2248 * @prev: the current task that is being switched out
2249 * @next: the task we are going to switch to.
2251 * This is called with the rq lock held and interrupts off. It must
2252 * be paired with a subsequent finish_task_switch after the context
2255 * prepare_task_switch sets up locking and calls architecture specific
2259 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2260 struct task_struct *next)
2262 fire_sched_out_preempt_notifiers(prev, next);
2263 prepare_lock_switch(rq, next);
2264 prepare_arch_switch(next);
2268 * finish_task_switch - clean up after a task-switch
2269 * @rq: runqueue associated with task-switch
2270 * @prev: the thread we just switched away from.
2272 * finish_task_switch must be called after the context switch, paired
2273 * with a prepare_task_switch call before the context switch.
2274 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2275 * and do any other architecture-specific cleanup actions.
2277 * Note that we may have delayed dropping an mm in context_switch(). If
2278 * so, we finish that here outside of the runqueue lock. (Doing it
2279 * with the lock held can cause deadlocks; see schedule() for
2282 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2283 __releases(rq->lock)
2285 struct mm_struct *mm = rq->prev_mm;
2291 * A task struct has one reference for the use as "current".
2292 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2293 * schedule one last time. The schedule call will never return, and
2294 * the scheduled task must drop that reference.
2295 * The test for TASK_DEAD must occur while the runqueue locks are
2296 * still held, otherwise prev could be scheduled on another cpu, die
2297 * there before we look at prev->state, and then the reference would
2299 * Manfred Spraul <manfred@colorfullife.com>
2301 prev_state = prev->state;
2302 finish_arch_switch(prev);
2303 finish_lock_switch(rq, prev);
2305 if (current->sched_class->post_schedule)
2306 current->sched_class->post_schedule(rq);
2309 fire_sched_in_preempt_notifiers(current);
2312 if (unlikely(prev_state == TASK_DEAD)) {
2314 * Remove function-return probe instances associated with this
2315 * task and put them back on the free list.
2317 kprobe_flush_task(prev);
2318 put_task_struct(prev);
2323 * schedule_tail - first thing a freshly forked thread must call.
2324 * @prev: the thread we just switched away from.
2326 asmlinkage void schedule_tail(struct task_struct *prev)
2327 __releases(rq->lock)
2329 struct rq *rq = this_rq();
2331 finish_task_switch(rq, prev);
2332 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2333 /* In this case, finish_task_switch does not reenable preemption */
2336 if (current->set_child_tid)
2337 put_user(task_pid_vnr(current), current->set_child_tid);
2341 * context_switch - switch to the new MM and the new
2342 * thread's register state.
2345 context_switch(struct rq *rq, struct task_struct *prev,
2346 struct task_struct *next)
2348 struct mm_struct *mm, *oldmm;
2350 prepare_task_switch(rq, prev, next);
2352 oldmm = prev->active_mm;
2354 * For paravirt, this is coupled with an exit in switch_to to
2355 * combine the page table reload and the switch backend into
2358 arch_enter_lazy_cpu_mode();
2360 if (unlikely(!mm)) {
2361 next->active_mm = oldmm;
2362 atomic_inc(&oldmm->mm_count);
2363 enter_lazy_tlb(oldmm, next);
2365 switch_mm(oldmm, mm, next);
2367 if (unlikely(!prev->mm)) {
2368 prev->active_mm = NULL;
2369 rq->prev_mm = oldmm;
2372 * Since the runqueue lock will be released by the next
2373 * task (which is an invalid locking op but in the case
2374 * of the scheduler it's an obvious special-case), so we
2375 * do an early lockdep release here:
2377 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2378 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2381 /* Here we just switch the register state and the stack. */
2382 switch_to(prev, next, prev);
2386 * this_rq must be evaluated again because prev may have moved
2387 * CPUs since it called schedule(), thus the 'rq' on its stack
2388 * frame will be invalid.
2390 finish_task_switch(this_rq(), prev);
2394 * nr_running, nr_uninterruptible and nr_context_switches:
2396 * externally visible scheduler statistics: current number of runnable
2397 * threads, current number of uninterruptible-sleeping threads, total
2398 * number of context switches performed since bootup.
2400 unsigned long nr_running(void)
2402 unsigned long i, sum = 0;
2404 for_each_online_cpu(i)
2405 sum += cpu_rq(i)->nr_running;
2410 unsigned long nr_uninterruptible(void)
2412 unsigned long i, sum = 0;
2414 for_each_possible_cpu(i)
2415 sum += cpu_rq(i)->nr_uninterruptible;
2418 * Since we read the counters lockless, it might be slightly
2419 * inaccurate. Do not allow it to go below zero though:
2421 if (unlikely((long)sum < 0))
2427 unsigned long long nr_context_switches(void)
2430 unsigned long long sum = 0;
2432 for_each_possible_cpu(i)
2433 sum += cpu_rq(i)->nr_switches;
2438 unsigned long nr_iowait(void)
2440 unsigned long i, sum = 0;
2442 for_each_possible_cpu(i)
2443 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2448 unsigned long nr_active(void)
2450 unsigned long i, running = 0, uninterruptible = 0;
2452 for_each_online_cpu(i) {
2453 running += cpu_rq(i)->nr_running;
2454 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2457 if (unlikely((long)uninterruptible < 0))
2458 uninterruptible = 0;
2460 return running + uninterruptible;
2464 * Update rq->cpu_load[] statistics. This function is usually called every
2465 * scheduler tick (TICK_NSEC).
2467 static void update_cpu_load(struct rq *this_rq)
2469 unsigned long this_load = this_rq->load.weight;
2472 this_rq->nr_load_updates++;
2474 /* Update our load: */
2475 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2476 unsigned long old_load, new_load;
2478 /* scale is effectively 1 << i now, and >> i divides by scale */
2480 old_load = this_rq->cpu_load[i];
2481 new_load = this_load;
2483 * Round up the averaging division if load is increasing. This
2484 * prevents us from getting stuck on 9 if the load is 10, for
2487 if (new_load > old_load)
2488 new_load += scale-1;
2489 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2496 * double_rq_lock - safely lock two runqueues
2498 * Note this does not disable interrupts like task_rq_lock,
2499 * you need to do so manually before calling.
2501 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2502 __acquires(rq1->lock)
2503 __acquires(rq2->lock)
2505 BUG_ON(!irqs_disabled());
2507 spin_lock(&rq1->lock);
2508 __acquire(rq2->lock); /* Fake it out ;) */
2511 spin_lock(&rq1->lock);
2512 spin_lock(&rq2->lock);
2514 spin_lock(&rq2->lock);
2515 spin_lock(&rq1->lock);
2518 update_rq_clock(rq1);
2519 update_rq_clock(rq2);
2523 * double_rq_unlock - safely unlock two runqueues
2525 * Note this does not restore interrupts like task_rq_unlock,
2526 * you need to do so manually after calling.
2528 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2529 __releases(rq1->lock)
2530 __releases(rq2->lock)
2532 spin_unlock(&rq1->lock);
2534 spin_unlock(&rq2->lock);
2536 __release(rq2->lock);
2540 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2542 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2543 __releases(this_rq->lock)
2544 __acquires(busiest->lock)
2545 __acquires(this_rq->lock)
2549 if (unlikely(!irqs_disabled())) {
2550 /* printk() doesn't work good under rq->lock */
2551 spin_unlock(&this_rq->lock);
2554 if (unlikely(!spin_trylock(&busiest->lock))) {
2555 if (busiest < this_rq) {
2556 spin_unlock(&this_rq->lock);
2557 spin_lock(&busiest->lock);
2558 spin_lock(&this_rq->lock);
2561 spin_lock(&busiest->lock);
2567 * If dest_cpu is allowed for this process, migrate the task to it.
2568 * This is accomplished by forcing the cpu_allowed mask to only
2569 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2570 * the cpu_allowed mask is restored.
2572 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2574 struct migration_req req;
2575 unsigned long flags;
2578 rq = task_rq_lock(p, &flags);
2579 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2580 || unlikely(cpu_is_offline(dest_cpu)))
2583 /* force the process onto the specified CPU */
2584 if (migrate_task(p, dest_cpu, &req)) {
2585 /* Need to wait for migration thread (might exit: take ref). */
2586 struct task_struct *mt = rq->migration_thread;
2588 get_task_struct(mt);
2589 task_rq_unlock(rq, &flags);
2590 wake_up_process(mt);
2591 put_task_struct(mt);
2592 wait_for_completion(&req.done);
2597 task_rq_unlock(rq, &flags);
2601 * sched_exec - execve() is a valuable balancing opportunity, because at
2602 * this point the task has the smallest effective memory and cache footprint.
2604 void sched_exec(void)
2606 int new_cpu, this_cpu = get_cpu();
2607 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2609 if (new_cpu != this_cpu)
2610 sched_migrate_task(current, new_cpu);
2614 * pull_task - move a task from a remote runqueue to the local runqueue.
2615 * Both runqueues must be locked.
2617 static void pull_task(struct rq *src_rq, struct task_struct *p,
2618 struct rq *this_rq, int this_cpu)
2620 deactivate_task(src_rq, p, 0);
2621 set_task_cpu(p, this_cpu);
2622 activate_task(this_rq, p, 0);
2624 * Note that idle threads have a prio of MAX_PRIO, for this test
2625 * to be always true for them.
2627 check_preempt_curr(this_rq, p);
2631 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2634 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2635 struct sched_domain *sd, enum cpu_idle_type idle,
2639 * We do not migrate tasks that are:
2640 * 1) running (obviously), or
2641 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2642 * 3) are cache-hot on their current CPU.
2644 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2645 schedstat_inc(p, se.nr_failed_migrations_affine);
2650 if (task_running(rq, p)) {
2651 schedstat_inc(p, se.nr_failed_migrations_running);
2656 * Aggressive migration if:
2657 * 1) task is cache cold, or
2658 * 2) too many balance attempts have failed.
2661 if (!task_hot(p, rq->clock, sd) ||
2662 sd->nr_balance_failed > sd->cache_nice_tries) {
2663 #ifdef CONFIG_SCHEDSTATS
2664 if (task_hot(p, rq->clock, sd)) {
2665 schedstat_inc(sd, lb_hot_gained[idle]);
2666 schedstat_inc(p, se.nr_forced_migrations);
2672 if (task_hot(p, rq->clock, sd)) {
2673 schedstat_inc(p, se.nr_failed_migrations_hot);
2679 static unsigned long
2680 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2681 unsigned long max_load_move, struct sched_domain *sd,
2682 enum cpu_idle_type idle, int *all_pinned,
2683 int *this_best_prio, struct rq_iterator *iterator)
2685 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2686 struct task_struct *p;
2687 long rem_load_move = max_load_move;
2689 if (max_load_move == 0)
2695 * Start the load-balancing iterator:
2697 p = iterator->start(iterator->arg);
2699 if (!p || loops++ > sysctl_sched_nr_migrate)
2702 * To help distribute high priority tasks across CPUs we don't
2703 * skip a task if it will be the highest priority task (i.e. smallest
2704 * prio value) on its new queue regardless of its load weight
2706 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2707 SCHED_LOAD_SCALE_FUZZ;
2708 if ((skip_for_load && p->prio >= *this_best_prio) ||
2709 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2710 p = iterator->next(iterator->arg);
2714 pull_task(busiest, p, this_rq, this_cpu);
2716 rem_load_move -= p->se.load.weight;
2719 * We only want to steal up to the prescribed amount of weighted load.
2721 if (rem_load_move > 0) {
2722 if (p->prio < *this_best_prio)
2723 *this_best_prio = p->prio;
2724 p = iterator->next(iterator->arg);
2729 * Right now, this is one of only two places pull_task() is called,
2730 * so we can safely collect pull_task() stats here rather than
2731 * inside pull_task().
2733 schedstat_add(sd, lb_gained[idle], pulled);
2736 *all_pinned = pinned;
2738 return max_load_move - rem_load_move;
2742 * move_tasks tries to move up to max_load_move weighted load from busiest to
2743 * this_rq, as part of a balancing operation within domain "sd".
2744 * Returns 1 if successful and 0 otherwise.
2746 * Called with both runqueues locked.
2748 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2749 unsigned long max_load_move,
2750 struct sched_domain *sd, enum cpu_idle_type idle,
2753 const struct sched_class *class = sched_class_highest;
2754 unsigned long total_load_moved = 0;
2755 int this_best_prio = this_rq->curr->prio;
2759 class->load_balance(this_rq, this_cpu, busiest,
2760 max_load_move - total_load_moved,
2761 sd, idle, all_pinned, &this_best_prio);
2762 class = class->next;
2763 } while (class && max_load_move > total_load_moved);
2765 return total_load_moved > 0;
2769 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2770 struct sched_domain *sd, enum cpu_idle_type idle,
2771 struct rq_iterator *iterator)
2773 struct task_struct *p = iterator->start(iterator->arg);
2777 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2778 pull_task(busiest, p, this_rq, this_cpu);
2780 * Right now, this is only the second place pull_task()
2781 * is called, so we can safely collect pull_task()
2782 * stats here rather than inside pull_task().
2784 schedstat_inc(sd, lb_gained[idle]);
2788 p = iterator->next(iterator->arg);
2795 * move_one_task tries to move exactly one task from busiest to this_rq, as
2796 * part of active balancing operations within "domain".
2797 * Returns 1 if successful and 0 otherwise.
2799 * Called with both runqueues locked.
2801 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2802 struct sched_domain *sd, enum cpu_idle_type idle)
2804 const struct sched_class *class;
2806 for (class = sched_class_highest; class; class = class->next)
2807 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2814 * find_busiest_group finds and returns the busiest CPU group within the
2815 * domain. It calculates and returns the amount of weighted load which
2816 * should be moved to restore balance via the imbalance parameter.
2818 static struct sched_group *
2819 find_busiest_group(struct sched_domain *sd, int this_cpu,
2820 unsigned long *imbalance, enum cpu_idle_type idle,
2821 int *sd_idle, const cpumask_t *cpus, int *balance)
2823 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2824 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2825 unsigned long max_pull;
2826 unsigned long busiest_load_per_task, busiest_nr_running;
2827 unsigned long this_load_per_task, this_nr_running;
2828 int load_idx, group_imb = 0;
2829 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2830 int power_savings_balance = 1;
2831 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2832 unsigned long min_nr_running = ULONG_MAX;
2833 struct sched_group *group_min = NULL, *group_leader = NULL;
2836 max_load = this_load = total_load = total_pwr = 0;
2837 busiest_load_per_task = busiest_nr_running = 0;
2838 this_load_per_task = this_nr_running = 0;
2839 if (idle == CPU_NOT_IDLE)
2840 load_idx = sd->busy_idx;
2841 else if (idle == CPU_NEWLY_IDLE)
2842 load_idx = sd->newidle_idx;
2844 load_idx = sd->idle_idx;
2847 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2850 int __group_imb = 0;
2851 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2852 unsigned long sum_nr_running, sum_weighted_load;
2854 local_group = cpu_isset(this_cpu, group->cpumask);
2857 balance_cpu = first_cpu(group->cpumask);
2859 /* Tally up the load of all CPUs in the group */
2860 sum_weighted_load = sum_nr_running = avg_load = 0;
2862 min_cpu_load = ~0UL;
2864 for_each_cpu_mask(i, group->cpumask) {
2867 if (!cpu_isset(i, *cpus))
2872 if (*sd_idle && rq->nr_running)
2875 /* Bias balancing toward cpus of our domain */
2877 if (idle_cpu(i) && !first_idle_cpu) {
2882 load = target_load(i, load_idx);
2884 load = source_load(i, load_idx);
2885 if (load > max_cpu_load)
2886 max_cpu_load = load;
2887 if (min_cpu_load > load)
2888 min_cpu_load = load;
2892 sum_nr_running += rq->nr_running;
2893 sum_weighted_load += weighted_cpuload(i);
2897 * First idle cpu or the first cpu(busiest) in this sched group
2898 * is eligible for doing load balancing at this and above
2899 * domains. In the newly idle case, we will allow all the cpu's
2900 * to do the newly idle load balance.
2902 if (idle != CPU_NEWLY_IDLE && local_group &&
2903 balance_cpu != this_cpu && balance) {
2908 total_load += avg_load;
2909 total_pwr += group->__cpu_power;
2911 /* Adjust by relative CPU power of the group */
2912 avg_load = sg_div_cpu_power(group,
2913 avg_load * SCHED_LOAD_SCALE);
2915 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2918 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2921 this_load = avg_load;
2923 this_nr_running = sum_nr_running;
2924 this_load_per_task = sum_weighted_load;
2925 } else if (avg_load > max_load &&
2926 (sum_nr_running > group_capacity || __group_imb)) {
2927 max_load = avg_load;
2929 busiest_nr_running = sum_nr_running;
2930 busiest_load_per_task = sum_weighted_load;
2931 group_imb = __group_imb;
2934 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2936 * Busy processors will not participate in power savings
2939 if (idle == CPU_NOT_IDLE ||
2940 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2944 * If the local group is idle or completely loaded
2945 * no need to do power savings balance at this domain
2947 if (local_group && (this_nr_running >= group_capacity ||
2949 power_savings_balance = 0;
2952 * If a group is already running at full capacity or idle,
2953 * don't include that group in power savings calculations
2955 if (!power_savings_balance || sum_nr_running >= group_capacity
2960 * Calculate the group which has the least non-idle load.
2961 * This is the group from where we need to pick up the load
2964 if ((sum_nr_running < min_nr_running) ||
2965 (sum_nr_running == min_nr_running &&
2966 first_cpu(group->cpumask) <
2967 first_cpu(group_min->cpumask))) {
2969 min_nr_running = sum_nr_running;
2970 min_load_per_task = sum_weighted_load /
2975 * Calculate the group which is almost near its
2976 * capacity but still has some space to pick up some load
2977 * from other group and save more power
2979 if (sum_nr_running <= group_capacity - 1) {
2980 if (sum_nr_running > leader_nr_running ||
2981 (sum_nr_running == leader_nr_running &&
2982 first_cpu(group->cpumask) >
2983 first_cpu(group_leader->cpumask))) {
2984 group_leader = group;
2985 leader_nr_running = sum_nr_running;
2990 group = group->next;
2991 } while (group != sd->groups);
2993 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2996 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2998 if (this_load >= avg_load ||
2999 100*max_load <= sd->imbalance_pct*this_load)
3002 busiest_load_per_task /= busiest_nr_running;
3004 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3007 * We're trying to get all the cpus to the average_load, so we don't
3008 * want to push ourselves above the average load, nor do we wish to
3009 * reduce the max loaded cpu below the average load, as either of these
3010 * actions would just result in more rebalancing later, and ping-pong
3011 * tasks around. Thus we look for the minimum possible imbalance.
3012 * Negative imbalances (*we* are more loaded than anyone else) will
3013 * be counted as no imbalance for these purposes -- we can't fix that
3014 * by pulling tasks to us. Be careful of negative numbers as they'll
3015 * appear as very large values with unsigned longs.
3017 if (max_load <= busiest_load_per_task)
3021 * In the presence of smp nice balancing, certain scenarios can have
3022 * max load less than avg load(as we skip the groups at or below
3023 * its cpu_power, while calculating max_load..)
3025 if (max_load < avg_load) {
3027 goto small_imbalance;
3030 /* Don't want to pull so many tasks that a group would go idle */
3031 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3033 /* How much load to actually move to equalise the imbalance */
3034 *imbalance = min(max_pull * busiest->__cpu_power,
3035 (avg_load - this_load) * this->__cpu_power)
3039 * if *imbalance is less than the average load per runnable task
3040 * there is no gaurantee that any tasks will be moved so we'll have
3041 * a think about bumping its value to force at least one task to be
3044 if (*imbalance < busiest_load_per_task) {
3045 unsigned long tmp, pwr_now, pwr_move;
3049 pwr_move = pwr_now = 0;
3051 if (this_nr_running) {
3052 this_load_per_task /= this_nr_running;
3053 if (busiest_load_per_task > this_load_per_task)
3056 this_load_per_task = SCHED_LOAD_SCALE;
3058 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3059 busiest_load_per_task * imbn) {
3060 *imbalance = busiest_load_per_task;
3065 * OK, we don't have enough imbalance to justify moving tasks,
3066 * however we may be able to increase total CPU power used by
3070 pwr_now += busiest->__cpu_power *
3071 min(busiest_load_per_task, max_load);
3072 pwr_now += this->__cpu_power *
3073 min(this_load_per_task, this_load);
3074 pwr_now /= SCHED_LOAD_SCALE;
3076 /* Amount of load we'd subtract */
3077 tmp = sg_div_cpu_power(busiest,
3078 busiest_load_per_task * SCHED_LOAD_SCALE);
3080 pwr_move += busiest->__cpu_power *
3081 min(busiest_load_per_task, max_load - tmp);
3083 /* Amount of load we'd add */
3084 if (max_load * busiest->__cpu_power <
3085 busiest_load_per_task * SCHED_LOAD_SCALE)
3086 tmp = sg_div_cpu_power(this,
3087 max_load * busiest->__cpu_power);
3089 tmp = sg_div_cpu_power(this,
3090 busiest_load_per_task * SCHED_LOAD_SCALE);
3091 pwr_move += this->__cpu_power *
3092 min(this_load_per_task, this_load + tmp);
3093 pwr_move /= SCHED_LOAD_SCALE;
3095 /* Move if we gain throughput */
3096 if (pwr_move > pwr_now)
3097 *imbalance = busiest_load_per_task;
3103 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3104 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3107 if (this == group_leader && group_leader != group_min) {
3108 *imbalance = min_load_per_task;
3118 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3121 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3122 unsigned long imbalance, const cpumask_t *cpus)
3124 struct rq *busiest = NULL, *rq;
3125 unsigned long max_load = 0;
3128 for_each_cpu_mask(i, group->cpumask) {
3131 if (!cpu_isset(i, *cpus))
3135 wl = weighted_cpuload(i);
3137 if (rq->nr_running == 1 && wl > imbalance)
3140 if (wl > max_load) {
3150 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3151 * so long as it is large enough.
3153 #define MAX_PINNED_INTERVAL 512
3156 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3157 * tasks if there is an imbalance.
3159 static int load_balance(int this_cpu, struct rq *this_rq,
3160 struct sched_domain *sd, enum cpu_idle_type idle,
3161 int *balance, cpumask_t *cpus)
3163 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3164 struct sched_group *group;
3165 unsigned long imbalance;
3167 unsigned long flags;
3172 * When power savings policy is enabled for the parent domain, idle
3173 * sibling can pick up load irrespective of busy siblings. In this case,
3174 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3175 * portraying it as CPU_NOT_IDLE.
3177 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3178 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3181 schedstat_inc(sd, lb_count[idle]);
3184 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3191 schedstat_inc(sd, lb_nobusyg[idle]);
3195 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3197 schedstat_inc(sd, lb_nobusyq[idle]);
3201 BUG_ON(busiest == this_rq);
3203 schedstat_add(sd, lb_imbalance[idle], imbalance);
3206 if (busiest->nr_running > 1) {
3208 * Attempt to move tasks. If find_busiest_group has found
3209 * an imbalance but busiest->nr_running <= 1, the group is
3210 * still unbalanced. ld_moved simply stays zero, so it is
3211 * correctly treated as an imbalance.
3213 local_irq_save(flags);
3214 double_rq_lock(this_rq, busiest);
3215 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3216 imbalance, sd, idle, &all_pinned);
3217 double_rq_unlock(this_rq, busiest);
3218 local_irq_restore(flags);
3221 * some other cpu did the load balance for us.
3223 if (ld_moved && this_cpu != smp_processor_id())
3224 resched_cpu(this_cpu);
3226 /* All tasks on this runqueue were pinned by CPU affinity */
3227 if (unlikely(all_pinned)) {
3228 cpu_clear(cpu_of(busiest), *cpus);
3229 if (!cpus_empty(*cpus))
3236 schedstat_inc(sd, lb_failed[idle]);
3237 sd->nr_balance_failed++;
3239 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3241 spin_lock_irqsave(&busiest->lock, flags);
3243 /* don't kick the migration_thread, if the curr
3244 * task on busiest cpu can't be moved to this_cpu
3246 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3247 spin_unlock_irqrestore(&busiest->lock, flags);
3249 goto out_one_pinned;
3252 if (!busiest->active_balance) {
3253 busiest->active_balance = 1;
3254 busiest->push_cpu = this_cpu;
3257 spin_unlock_irqrestore(&busiest->lock, flags);
3259 wake_up_process(busiest->migration_thread);
3262 * We've kicked active balancing, reset the failure
3265 sd->nr_balance_failed = sd->cache_nice_tries+1;
3268 sd->nr_balance_failed = 0;
3270 if (likely(!active_balance)) {
3271 /* We were unbalanced, so reset the balancing interval */
3272 sd->balance_interval = sd->min_interval;
3275 * If we've begun active balancing, start to back off. This
3276 * case may not be covered by the all_pinned logic if there
3277 * is only 1 task on the busy runqueue (because we don't call
3280 if (sd->balance_interval < sd->max_interval)
3281 sd->balance_interval *= 2;
3284 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3285 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3290 schedstat_inc(sd, lb_balanced[idle]);
3292 sd->nr_balance_failed = 0;
3295 /* tune up the balancing interval */
3296 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3297 (sd->balance_interval < sd->max_interval))
3298 sd->balance_interval *= 2;
3300 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3301 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3307 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3308 * tasks if there is an imbalance.
3310 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3311 * this_rq is locked.
3314 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3317 struct sched_group *group;
3318 struct rq *busiest = NULL;
3319 unsigned long imbalance;
3327 * When power savings policy is enabled for the parent domain, idle
3328 * sibling can pick up load irrespective of busy siblings. In this case,
3329 * let the state of idle sibling percolate up as IDLE, instead of
3330 * portraying it as CPU_NOT_IDLE.
3332 if (sd->flags & SD_SHARE_CPUPOWER &&
3333 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3336 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3338 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3339 &sd_idle, cpus, NULL);
3341 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3345 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3347 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3351 BUG_ON(busiest == this_rq);
3353 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3356 if (busiest->nr_running > 1) {
3357 /* Attempt to move tasks */
3358 double_lock_balance(this_rq, busiest);
3359 /* this_rq->clock is already updated */
3360 update_rq_clock(busiest);
3361 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3362 imbalance, sd, CPU_NEWLY_IDLE,
3364 spin_unlock(&busiest->lock);
3366 if (unlikely(all_pinned)) {
3367 cpu_clear(cpu_of(busiest), *cpus);
3368 if (!cpus_empty(*cpus))
3374 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3375 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3376 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3379 sd->nr_balance_failed = 0;
3384 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3385 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3386 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3388 sd->nr_balance_failed = 0;
3394 * idle_balance is called by schedule() if this_cpu is about to become
3395 * idle. Attempts to pull tasks from other CPUs.
3397 static void idle_balance(int this_cpu, struct rq *this_rq)
3399 struct sched_domain *sd;
3400 int pulled_task = -1;
3401 unsigned long next_balance = jiffies + HZ;
3404 for_each_domain(this_cpu, sd) {
3405 unsigned long interval;
3407 if (!(sd->flags & SD_LOAD_BALANCE))
3410 if (sd->flags & SD_BALANCE_NEWIDLE)
3411 /* If we've pulled tasks over stop searching: */
3412 pulled_task = load_balance_newidle(this_cpu, this_rq,
3415 interval = msecs_to_jiffies(sd->balance_interval);
3416 if (time_after(next_balance, sd->last_balance + interval))
3417 next_balance = sd->last_balance + interval;
3421 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3423 * We are going idle. next_balance may be set based on
3424 * a busy processor. So reset next_balance.
3426 this_rq->next_balance = next_balance;
3431 * active_load_balance is run by migration threads. It pushes running tasks
3432 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3433 * running on each physical CPU where possible, and avoids physical /
3434 * logical imbalances.
3436 * Called with busiest_rq locked.
3438 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3440 int target_cpu = busiest_rq->push_cpu;
3441 struct sched_domain *sd;
3442 struct rq *target_rq;
3444 /* Is there any task to move? */
3445 if (busiest_rq->nr_running <= 1)
3448 target_rq = cpu_rq(target_cpu);
3451 * This condition is "impossible", if it occurs
3452 * we need to fix it. Originally reported by
3453 * Bjorn Helgaas on a 128-cpu setup.
3455 BUG_ON(busiest_rq == target_rq);
3457 /* move a task from busiest_rq to target_rq */
3458 double_lock_balance(busiest_rq, target_rq);
3459 update_rq_clock(busiest_rq);
3460 update_rq_clock(target_rq);
3462 /* Search for an sd spanning us and the target CPU. */
3463 for_each_domain(target_cpu, sd) {
3464 if ((sd->flags & SD_LOAD_BALANCE) &&
3465 cpu_isset(busiest_cpu, sd->span))
3470 schedstat_inc(sd, alb_count);
3472 if (move_one_task(target_rq, target_cpu, busiest_rq,
3474 schedstat_inc(sd, alb_pushed);
3476 schedstat_inc(sd, alb_failed);
3478 spin_unlock(&target_rq->lock);
3483 atomic_t load_balancer;
3485 } nohz ____cacheline_aligned = {
3486 .load_balancer = ATOMIC_INIT(-1),
3487 .cpu_mask = CPU_MASK_NONE,
3491 * This routine will try to nominate the ilb (idle load balancing)
3492 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3493 * load balancing on behalf of all those cpus. If all the cpus in the system
3494 * go into this tickless mode, then there will be no ilb owner (as there is
3495 * no need for one) and all the cpus will sleep till the next wakeup event
3498 * For the ilb owner, tick is not stopped. And this tick will be used
3499 * for idle load balancing. ilb owner will still be part of
3502 * While stopping the tick, this cpu will become the ilb owner if there
3503 * is no other owner. And will be the owner till that cpu becomes busy
3504 * or if all cpus in the system stop their ticks at which point
3505 * there is no need for ilb owner.
3507 * When the ilb owner becomes busy, it nominates another owner, during the
3508 * next busy scheduler_tick()
3510 int select_nohz_load_balancer(int stop_tick)
3512 int cpu = smp_processor_id();
3515 cpu_set(cpu, nohz.cpu_mask);
3516 cpu_rq(cpu)->in_nohz_recently = 1;
3519 * If we are going offline and still the leader, give up!
3521 if (cpu_is_offline(cpu) &&
3522 atomic_read(&nohz.load_balancer) == cpu) {
3523 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3528 /* time for ilb owner also to sleep */
3529 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3530 if (atomic_read(&nohz.load_balancer) == cpu)
3531 atomic_set(&nohz.load_balancer, -1);
3535 if (atomic_read(&nohz.load_balancer) == -1) {
3536 /* make me the ilb owner */
3537 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3539 } else if (atomic_read(&nohz.load_balancer) == cpu)
3542 if (!cpu_isset(cpu, nohz.cpu_mask))
3545 cpu_clear(cpu, nohz.cpu_mask);
3547 if (atomic_read(&nohz.load_balancer) == cpu)
3548 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3555 static DEFINE_SPINLOCK(balancing);
3558 * It checks each scheduling domain to see if it is due to be balanced,
3559 * and initiates a balancing operation if so.
3561 * Balancing parameters are set up in arch_init_sched_domains.
3563 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3566 struct rq *rq = cpu_rq(cpu);
3567 unsigned long interval;
3568 struct sched_domain *sd;
3569 /* Earliest time when we have to do rebalance again */
3570 unsigned long next_balance = jiffies + 60*HZ;
3571 int update_next_balance = 0;
3574 for_each_domain(cpu, sd) {
3575 if (!(sd->flags & SD_LOAD_BALANCE))
3578 interval = sd->balance_interval;
3579 if (idle != CPU_IDLE)
3580 interval *= sd->busy_factor;
3582 /* scale ms to jiffies */
3583 interval = msecs_to_jiffies(interval);
3584 if (unlikely(!interval))
3586 if (interval > HZ*NR_CPUS/10)
3587 interval = HZ*NR_CPUS/10;
3590 if (sd->flags & SD_SERIALIZE) {
3591 if (!spin_trylock(&balancing))
3595 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3596 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3598 * We've pulled tasks over so either we're no
3599 * longer idle, or one of our SMT siblings is
3602 idle = CPU_NOT_IDLE;
3604 sd->last_balance = jiffies;
3606 if (sd->flags & SD_SERIALIZE)
3607 spin_unlock(&balancing);
3609 if (time_after(next_balance, sd->last_balance + interval)) {
3610 next_balance = sd->last_balance + interval;
3611 update_next_balance = 1;
3615 * Stop the load balance at this level. There is another
3616 * CPU in our sched group which is doing load balancing more
3624 * next_balance will be updated only when there is a need.
3625 * When the cpu is attached to null domain for ex, it will not be
3628 if (likely(update_next_balance))
3629 rq->next_balance = next_balance;
3633 * run_rebalance_domains is triggered when needed from the scheduler tick.
3634 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3635 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3637 static void run_rebalance_domains(struct softirq_action *h)
3639 int this_cpu = smp_processor_id();
3640 struct rq *this_rq = cpu_rq(this_cpu);
3641 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3642 CPU_IDLE : CPU_NOT_IDLE;
3644 rebalance_domains(this_cpu, idle);
3648 * If this cpu is the owner for idle load balancing, then do the
3649 * balancing on behalf of the other idle cpus whose ticks are
3652 if (this_rq->idle_at_tick &&
3653 atomic_read(&nohz.load_balancer) == this_cpu) {
3654 cpumask_t cpus = nohz.cpu_mask;
3658 cpu_clear(this_cpu, cpus);
3659 for_each_cpu_mask(balance_cpu, cpus) {
3661 * If this cpu gets work to do, stop the load balancing
3662 * work being done for other cpus. Next load
3663 * balancing owner will pick it up.
3668 rebalance_domains(balance_cpu, CPU_IDLE);
3670 rq = cpu_rq(balance_cpu);
3671 if (time_after(this_rq->next_balance, rq->next_balance))
3672 this_rq->next_balance = rq->next_balance;
3679 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3681 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3682 * idle load balancing owner or decide to stop the periodic load balancing,
3683 * if the whole system is idle.
3685 static inline void trigger_load_balance(struct rq *rq, int cpu)
3689 * If we were in the nohz mode recently and busy at the current
3690 * scheduler tick, then check if we need to nominate new idle
3693 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3694 rq->in_nohz_recently = 0;
3696 if (atomic_read(&nohz.load_balancer) == cpu) {
3697 cpu_clear(cpu, nohz.cpu_mask);
3698 atomic_set(&nohz.load_balancer, -1);
3701 if (atomic_read(&nohz.load_balancer) == -1) {
3703 * simple selection for now: Nominate the
3704 * first cpu in the nohz list to be the next
3707 * TBD: Traverse the sched domains and nominate
3708 * the nearest cpu in the nohz.cpu_mask.
3710 int ilb = first_cpu(nohz.cpu_mask);
3712 if (ilb < nr_cpu_ids)
3718 * If this cpu is idle and doing idle load balancing for all the
3719 * cpus with ticks stopped, is it time for that to stop?
3721 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3722 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3728 * If this cpu is idle and the idle load balancing is done by
3729 * someone else, then no need raise the SCHED_SOFTIRQ
3731 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3732 cpu_isset(cpu, nohz.cpu_mask))
3735 if (time_after_eq(jiffies, rq->next_balance))
3736 raise_softirq(SCHED_SOFTIRQ);
3739 #else /* CONFIG_SMP */
3742 * on UP we do not need to balance between CPUs:
3744 static inline void idle_balance(int cpu, struct rq *rq)
3750 DEFINE_PER_CPU(struct kernel_stat, kstat);
3752 EXPORT_PER_CPU_SYMBOL(kstat);
3755 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3756 * that have not yet been banked in case the task is currently running.
3758 unsigned long long task_sched_runtime(struct task_struct *p)
3760 unsigned long flags;
3764 rq = task_rq_lock(p, &flags);
3765 ns = p->se.sum_exec_runtime;
3766 if (task_current(rq, p)) {
3767 update_rq_clock(rq);
3768 delta_exec = rq->clock - p->se.exec_start;
3769 if ((s64)delta_exec > 0)
3772 task_rq_unlock(rq, &flags);
3778 * Account user cpu time to a process.
3779 * @p: the process that the cpu time gets accounted to
3780 * @cputime: the cpu time spent in user space since the last update
3782 void account_user_time(struct task_struct *p, cputime_t cputime)
3784 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3787 p->utime = cputime_add(p->utime, cputime);
3789 /* Add user time to cpustat. */
3790 tmp = cputime_to_cputime64(cputime);
3791 if (TASK_NICE(p) > 0)
3792 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3794 cpustat->user = cputime64_add(cpustat->user, tmp);
3798 * Account guest cpu time to a process.
3799 * @p: the process that the cpu time gets accounted to
3800 * @cputime: the cpu time spent in virtual machine since the last update
3802 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3805 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3807 tmp = cputime_to_cputime64(cputime);
3809 p->utime = cputime_add(p->utime, cputime);
3810 p->gtime = cputime_add(p->gtime, cputime);
3812 cpustat->user = cputime64_add(cpustat->user, tmp);
3813 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3817 * Account scaled user cpu time to a process.
3818 * @p: the process that the cpu time gets accounted to
3819 * @cputime: the cpu time spent in user space since the last update
3821 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3823 p->utimescaled = cputime_add(p->utimescaled, cputime);
3827 * Account system cpu time to a process.
3828 * @p: the process that the cpu time gets accounted to
3829 * @hardirq_offset: the offset to subtract from hardirq_count()
3830 * @cputime: the cpu time spent in kernel space since the last update
3832 void account_system_time(struct task_struct *p, int hardirq_offset,
3835 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3836 struct rq *rq = this_rq();
3839 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3840 return account_guest_time(p, cputime);
3842 p->stime = cputime_add(p->stime, cputime);
3844 /* Add system time to cpustat. */
3845 tmp = cputime_to_cputime64(cputime);
3846 if (hardirq_count() - hardirq_offset)
3847 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3848 else if (softirq_count())
3849 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3850 else if (p != rq->idle)
3851 cpustat->system = cputime64_add(cpustat->system, tmp);
3852 else if (atomic_read(&rq->nr_iowait) > 0)
3853 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3855 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3856 /* Account for system time used */
3857 acct_update_integrals(p);
3861 * Account scaled system cpu time to a process.
3862 * @p: the process that the cpu time gets accounted to
3863 * @hardirq_offset: the offset to subtract from hardirq_count()
3864 * @cputime: the cpu time spent in kernel space since the last update
3866 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3868 p->stimescaled = cputime_add(p->stimescaled, cputime);
3872 * Account for involuntary wait time.
3873 * @p: the process from which the cpu time has been stolen
3874 * @steal: the cpu time spent in involuntary wait
3876 void account_steal_time(struct task_struct *p, cputime_t steal)
3878 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3879 cputime64_t tmp = cputime_to_cputime64(steal);
3880 struct rq *rq = this_rq();
3882 if (p == rq->idle) {
3883 p->stime = cputime_add(p->stime, steal);
3884 if (atomic_read(&rq->nr_iowait) > 0)
3885 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3887 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3889 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3893 * This function gets called by the timer code, with HZ frequency.
3894 * We call it with interrupts disabled.
3896 * It also gets called by the fork code, when changing the parent's
3899 void scheduler_tick(void)
3901 int cpu = smp_processor_id();
3902 struct rq *rq = cpu_rq(cpu);
3903 struct task_struct *curr = rq->curr;
3904 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3906 spin_lock(&rq->lock);
3907 __update_rq_clock(rq);
3909 * Let rq->clock advance by at least TICK_NSEC:
3911 if (unlikely(rq->clock < next_tick)) {
3912 rq->clock = next_tick;
3913 rq->clock_underflows++;
3915 rq->tick_timestamp = rq->clock;
3916 update_last_tick_seen(rq);
3917 update_cpu_load(rq);
3918 curr->sched_class->task_tick(rq, curr, 0);
3919 spin_unlock(&rq->lock);
3922 rq->idle_at_tick = idle_cpu(cpu);
3923 trigger_load_balance(rq, cpu);
3927 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3929 void __kprobes add_preempt_count(int val)
3934 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3936 preempt_count() += val;
3938 * Spinlock count overflowing soon?
3940 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3943 EXPORT_SYMBOL(add_preempt_count);
3945 void __kprobes sub_preempt_count(int val)
3950 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3953 * Is the spinlock portion underflowing?
3955 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3956 !(preempt_count() & PREEMPT_MASK)))
3959 preempt_count() -= val;
3961 EXPORT_SYMBOL(sub_preempt_count);
3966 * Print scheduling while atomic bug:
3968 static noinline void __schedule_bug(struct task_struct *prev)
3970 struct pt_regs *regs = get_irq_regs();
3972 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3973 prev->comm, prev->pid, preempt_count());
3975 debug_show_held_locks(prev);
3976 if (irqs_disabled())
3977 print_irqtrace_events(prev);
3986 * Various schedule()-time debugging checks and statistics:
3988 static inline void schedule_debug(struct task_struct *prev)
3991 * Test if we are atomic. Since do_exit() needs to call into
3992 * schedule() atomically, we ignore that path for now.
3993 * Otherwise, whine if we are scheduling when we should not be.
3995 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3996 __schedule_bug(prev);
3998 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4000 schedstat_inc(this_rq(), sched_count);
4001 #ifdef CONFIG_SCHEDSTATS
4002 if (unlikely(prev->lock_depth >= 0)) {
4003 schedstat_inc(this_rq(), bkl_count);
4004 schedstat_inc(prev, sched_info.bkl_count);
4010 * Pick up the highest-prio task:
4012 static inline struct task_struct *
4013 pick_next_task(struct rq *rq, struct task_struct *prev)
4015 const struct sched_class *class;
4016 struct task_struct *p;
4019 * Optimization: we know that if all tasks are in
4020 * the fair class we can call that function directly:
4022 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4023 p = fair_sched_class.pick_next_task(rq);
4028 class = sched_class_highest;
4030 p = class->pick_next_task(rq);
4034 * Will never be NULL as the idle class always
4035 * returns a non-NULL p:
4037 class = class->next;
4042 * schedule() is the main scheduler function.
4044 asmlinkage void __sched schedule(void)
4046 struct task_struct *prev, *next;
4047 unsigned long *switch_count;
4053 cpu = smp_processor_id();
4057 switch_count = &prev->nivcsw;
4059 release_kernel_lock(prev);
4060 need_resched_nonpreemptible:
4062 schedule_debug(prev);
4067 * Do the rq-clock update outside the rq lock:
4069 local_irq_disable();
4070 __update_rq_clock(rq);
4071 spin_lock(&rq->lock);
4072 clear_tsk_need_resched(prev);
4074 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4075 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4076 signal_pending(prev))) {
4077 prev->state = TASK_RUNNING;
4079 deactivate_task(rq, prev, 1);
4081 switch_count = &prev->nvcsw;
4085 if (prev->sched_class->pre_schedule)
4086 prev->sched_class->pre_schedule(rq, prev);
4089 if (unlikely(!rq->nr_running))
4090 idle_balance(cpu, rq);
4092 prev->sched_class->put_prev_task(rq, prev);
4093 next = pick_next_task(rq, prev);
4095 sched_info_switch(prev, next);
4097 if (likely(prev != next)) {
4102 context_switch(rq, prev, next); /* unlocks the rq */
4104 * the context switch might have flipped the stack from under
4105 * us, hence refresh the local variables.
4107 cpu = smp_processor_id();
4110 spin_unlock_irq(&rq->lock);
4114 if (unlikely(reacquire_kernel_lock(current) < 0))
4115 goto need_resched_nonpreemptible;
4117 preempt_enable_no_resched();
4118 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4121 EXPORT_SYMBOL(schedule);
4123 #ifdef CONFIG_PREEMPT
4125 * this is the entry point to schedule() from in-kernel preemption
4126 * off of preempt_enable. Kernel preemptions off return from interrupt
4127 * occur there and call schedule directly.
4129 asmlinkage void __sched preempt_schedule(void)
4131 struct thread_info *ti = current_thread_info();
4132 struct task_struct *task = current;
4133 int saved_lock_depth;
4136 * If there is a non-zero preempt_count or interrupts are disabled,
4137 * we do not want to preempt the current task. Just return..
4139 if (likely(ti->preempt_count || irqs_disabled()))
4143 add_preempt_count(PREEMPT_ACTIVE);
4146 * We keep the big kernel semaphore locked, but we
4147 * clear ->lock_depth so that schedule() doesnt
4148 * auto-release the semaphore:
4150 saved_lock_depth = task->lock_depth;
4151 task->lock_depth = -1;
4153 task->lock_depth = saved_lock_depth;
4154 sub_preempt_count(PREEMPT_ACTIVE);
4157 * Check again in case we missed a preemption opportunity
4158 * between schedule and now.
4161 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4163 EXPORT_SYMBOL(preempt_schedule);
4166 * this is the entry point to schedule() from kernel preemption
4167 * off of irq context.
4168 * Note, that this is called and return with irqs disabled. This will
4169 * protect us against recursive calling from irq.
4171 asmlinkage void __sched preempt_schedule_irq(void)
4173 struct thread_info *ti = current_thread_info();
4174 struct task_struct *task = current;
4175 int saved_lock_depth;
4177 /* Catch callers which need to be fixed */
4178 BUG_ON(ti->preempt_count || !irqs_disabled());
4181 add_preempt_count(PREEMPT_ACTIVE);
4184 * We keep the big kernel semaphore locked, but we
4185 * clear ->lock_depth so that schedule() doesnt
4186 * auto-release the semaphore:
4188 saved_lock_depth = task->lock_depth;
4189 task->lock_depth = -1;
4192 local_irq_disable();
4193 task->lock_depth = saved_lock_depth;
4194 sub_preempt_count(PREEMPT_ACTIVE);
4197 * Check again in case we missed a preemption opportunity
4198 * between schedule and now.
4201 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4204 #endif /* CONFIG_PREEMPT */
4206 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4209 return try_to_wake_up(curr->private, mode, sync);
4211 EXPORT_SYMBOL(default_wake_function);
4214 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4215 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4216 * number) then we wake all the non-exclusive tasks and one exclusive task.
4218 * There are circumstances in which we can try to wake a task which has already
4219 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4220 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4222 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4223 int nr_exclusive, int sync, void *key)
4225 wait_queue_t *curr, *next;
4227 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4228 unsigned flags = curr->flags;
4230 if (curr->func(curr, mode, sync, key) &&
4231 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4237 * __wake_up - wake up threads blocked on a waitqueue.
4239 * @mode: which threads
4240 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4241 * @key: is directly passed to the wakeup function
4243 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4244 int nr_exclusive, void *key)
4246 unsigned long flags;
4248 spin_lock_irqsave(&q->lock, flags);
4249 __wake_up_common(q, mode, nr_exclusive, 0, key);
4250 spin_unlock_irqrestore(&q->lock, flags);
4252 EXPORT_SYMBOL(__wake_up);
4255 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4257 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4259 __wake_up_common(q, mode, 1, 0, NULL);
4263 * __wake_up_sync - wake up threads blocked on a waitqueue.
4265 * @mode: which threads
4266 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4268 * The sync wakeup differs that the waker knows that it will schedule
4269 * away soon, so while the target thread will be woken up, it will not
4270 * be migrated to another CPU - ie. the two threads are 'synchronized'
4271 * with each other. This can prevent needless bouncing between CPUs.
4273 * On UP it can prevent extra preemption.
4276 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4278 unsigned long flags;
4284 if (unlikely(!nr_exclusive))
4287 spin_lock_irqsave(&q->lock, flags);
4288 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4289 spin_unlock_irqrestore(&q->lock, flags);
4291 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4293 void complete(struct completion *x)
4295 unsigned long flags;
4297 spin_lock_irqsave(&x->wait.lock, flags);
4299 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4300 spin_unlock_irqrestore(&x->wait.lock, flags);
4302 EXPORT_SYMBOL(complete);
4304 void complete_all(struct completion *x)
4306 unsigned long flags;
4308 spin_lock_irqsave(&x->wait.lock, flags);
4309 x->done += UINT_MAX/2;
4310 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4311 spin_unlock_irqrestore(&x->wait.lock, flags);
4313 EXPORT_SYMBOL(complete_all);
4315 static inline long __sched
4316 do_wait_for_common(struct completion *x, long timeout, int state)
4319 DECLARE_WAITQUEUE(wait, current);
4321 wait.flags |= WQ_FLAG_EXCLUSIVE;
4322 __add_wait_queue_tail(&x->wait, &wait);
4324 if ((state == TASK_INTERRUPTIBLE &&
4325 signal_pending(current)) ||
4326 (state == TASK_KILLABLE &&
4327 fatal_signal_pending(current))) {
4328 __remove_wait_queue(&x->wait, &wait);
4329 return -ERESTARTSYS;
4331 __set_current_state(state);
4332 spin_unlock_irq(&x->wait.lock);
4333 timeout = schedule_timeout(timeout);
4334 spin_lock_irq(&x->wait.lock);
4336 __remove_wait_queue(&x->wait, &wait);
4340 __remove_wait_queue(&x->wait, &wait);
4347 wait_for_common(struct completion *x, long timeout, int state)
4351 spin_lock_irq(&x->wait.lock);
4352 timeout = do_wait_for_common(x, timeout, state);
4353 spin_unlock_irq(&x->wait.lock);
4357 void __sched wait_for_completion(struct completion *x)
4359 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4361 EXPORT_SYMBOL(wait_for_completion);
4363 unsigned long __sched
4364 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4366 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4368 EXPORT_SYMBOL(wait_for_completion_timeout);
4370 int __sched wait_for_completion_interruptible(struct completion *x)
4372 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4373 if (t == -ERESTARTSYS)
4377 EXPORT_SYMBOL(wait_for_completion_interruptible);
4379 unsigned long __sched
4380 wait_for_completion_interruptible_timeout(struct completion *x,
4381 unsigned long timeout)
4383 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4385 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4387 int __sched wait_for_completion_killable(struct completion *x)
4389 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4390 if (t == -ERESTARTSYS)
4394 EXPORT_SYMBOL(wait_for_completion_killable);
4397 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4399 unsigned long flags;
4402 init_waitqueue_entry(&wait, current);
4404 __set_current_state(state);
4406 spin_lock_irqsave(&q->lock, flags);
4407 __add_wait_queue(q, &wait);
4408 spin_unlock(&q->lock);
4409 timeout = schedule_timeout(timeout);
4410 spin_lock_irq(&q->lock);
4411 __remove_wait_queue(q, &wait);
4412 spin_unlock_irqrestore(&q->lock, flags);
4417 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4419 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4421 EXPORT_SYMBOL(interruptible_sleep_on);
4424 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4426 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4428 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4430 void __sched sleep_on(wait_queue_head_t *q)
4432 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4434 EXPORT_SYMBOL(sleep_on);
4436 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4438 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4440 EXPORT_SYMBOL(sleep_on_timeout);
4442 #ifdef CONFIG_RT_MUTEXES
4445 * rt_mutex_setprio - set the current priority of a task
4447 * @prio: prio value (kernel-internal form)
4449 * This function changes the 'effective' priority of a task. It does
4450 * not touch ->normal_prio like __setscheduler().
4452 * Used by the rt_mutex code to implement priority inheritance logic.
4454 void rt_mutex_setprio(struct task_struct *p, int prio)
4456 unsigned long flags;
4457 int oldprio, on_rq, running;
4459 const struct sched_class *prev_class = p->sched_class;
4461 BUG_ON(prio < 0 || prio > MAX_PRIO);
4463 rq = task_rq_lock(p, &flags);
4464 update_rq_clock(rq);
4467 on_rq = p->se.on_rq;
4468 running = task_current(rq, p);
4470 dequeue_task(rq, p, 0);
4472 p->sched_class->put_prev_task(rq, p);
4475 p->sched_class = &rt_sched_class;
4477 p->sched_class = &fair_sched_class;
4482 p->sched_class->set_curr_task(rq);
4484 enqueue_task(rq, p, 0);
4486 check_class_changed(rq, p, prev_class, oldprio, running);
4488 task_rq_unlock(rq, &flags);
4493 void set_user_nice(struct task_struct *p, long nice)
4495 int old_prio, delta, on_rq;
4496 unsigned long flags;
4499 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4502 * We have to be careful, if called from sys_setpriority(),
4503 * the task might be in the middle of scheduling on another CPU.
4505 rq = task_rq_lock(p, &flags);
4506 update_rq_clock(rq);
4508 * The RT priorities are set via sched_setscheduler(), but we still
4509 * allow the 'normal' nice value to be set - but as expected
4510 * it wont have any effect on scheduling until the task is
4511 * SCHED_FIFO/SCHED_RR:
4513 if (task_has_rt_policy(p)) {
4514 p->static_prio = NICE_TO_PRIO(nice);
4517 on_rq = p->se.on_rq;
4519 dequeue_task(rq, p, 0);
4523 p->static_prio = NICE_TO_PRIO(nice);
4526 p->prio = effective_prio(p);
4527 delta = p->prio - old_prio;
4530 enqueue_task(rq, p, 0);
4533 * If the task increased its priority or is running and
4534 * lowered its priority, then reschedule its CPU:
4536 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4537 resched_task(rq->curr);
4540 task_rq_unlock(rq, &flags);
4542 EXPORT_SYMBOL(set_user_nice);
4545 * can_nice - check if a task can reduce its nice value
4549 int can_nice(const struct task_struct *p, const int nice)
4551 /* convert nice value [19,-20] to rlimit style value [1,40] */
4552 int nice_rlim = 20 - nice;
4554 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4555 capable(CAP_SYS_NICE));
4558 #ifdef __ARCH_WANT_SYS_NICE
4561 * sys_nice - change the priority of the current process.
4562 * @increment: priority increment
4564 * sys_setpriority is a more generic, but much slower function that
4565 * does similar things.
4567 asmlinkage long sys_nice(int increment)
4572 * Setpriority might change our priority at the same moment.
4573 * We don't have to worry. Conceptually one call occurs first
4574 * and we have a single winner.
4576 if (increment < -40)
4581 nice = PRIO_TO_NICE(current->static_prio) + increment;
4587 if (increment < 0 && !can_nice(current, nice))
4590 retval = security_task_setnice(current, nice);
4594 set_user_nice(current, nice);
4601 * task_prio - return the priority value of a given task.
4602 * @p: the task in question.
4604 * This is the priority value as seen by users in /proc.
4605 * RT tasks are offset by -200. Normal tasks are centered
4606 * around 0, value goes from -16 to +15.
4608 int task_prio(const struct task_struct *p)
4610 return p->prio - MAX_RT_PRIO;
4614 * task_nice - return the nice value of a given task.
4615 * @p: the task in question.
4617 int task_nice(const struct task_struct *p)
4619 return TASK_NICE(p);
4621 EXPORT_SYMBOL(task_nice);
4624 * idle_cpu - is a given cpu idle currently?
4625 * @cpu: the processor in question.
4627 int idle_cpu(int cpu)
4629 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4633 * idle_task - return the idle task for a given cpu.
4634 * @cpu: the processor in question.
4636 struct task_struct *idle_task(int cpu)
4638 return cpu_rq(cpu)->idle;
4642 * find_process_by_pid - find a process with a matching PID value.
4643 * @pid: the pid in question.
4645 static struct task_struct *find_process_by_pid(pid_t pid)
4647 return pid ? find_task_by_vpid(pid) : current;
4650 /* Actually do priority change: must hold rq lock. */
4652 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4654 BUG_ON(p->se.on_rq);
4657 switch (p->policy) {
4661 p->sched_class = &fair_sched_class;
4665 p->sched_class = &rt_sched_class;
4669 p->rt_priority = prio;
4670 p->normal_prio = normal_prio(p);
4671 /* we are holding p->pi_lock already */
4672 p->prio = rt_mutex_getprio(p);
4677 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4678 * @p: the task in question.
4679 * @policy: new policy.
4680 * @param: structure containing the new RT priority.
4682 * NOTE that the task may be already dead.
4684 int sched_setscheduler(struct task_struct *p, int policy,
4685 struct sched_param *param)
4687 int retval, oldprio, oldpolicy = -1, on_rq, running;
4688 unsigned long flags;
4689 const struct sched_class *prev_class = p->sched_class;
4692 /* may grab non-irq protected spin_locks */
4693 BUG_ON(in_interrupt());
4695 /* double check policy once rq lock held */
4697 policy = oldpolicy = p->policy;
4698 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4699 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4700 policy != SCHED_IDLE)
4703 * Valid priorities for SCHED_FIFO and SCHED_RR are
4704 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4705 * SCHED_BATCH and SCHED_IDLE is 0.
4707 if (param->sched_priority < 0 ||
4708 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4709 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4711 if (rt_policy(policy) != (param->sched_priority != 0))
4715 * Allow unprivileged RT tasks to decrease priority:
4717 if (!capable(CAP_SYS_NICE)) {
4718 if (rt_policy(policy)) {
4719 unsigned long rlim_rtprio;
4721 if (!lock_task_sighand(p, &flags))
4723 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4724 unlock_task_sighand(p, &flags);
4726 /* can't set/change the rt policy */
4727 if (policy != p->policy && !rlim_rtprio)
4730 /* can't increase priority */
4731 if (param->sched_priority > p->rt_priority &&
4732 param->sched_priority > rlim_rtprio)
4736 * Like positive nice levels, dont allow tasks to
4737 * move out of SCHED_IDLE either:
4739 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4742 /* can't change other user's priorities */
4743 if ((current->euid != p->euid) &&
4744 (current->euid != p->uid))
4748 #ifdef CONFIG_RT_GROUP_SCHED
4750 * Do not allow realtime tasks into groups that have no runtime
4753 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4757 retval = security_task_setscheduler(p, policy, param);
4761 * make sure no PI-waiters arrive (or leave) while we are
4762 * changing the priority of the task:
4764 spin_lock_irqsave(&p->pi_lock, flags);
4766 * To be able to change p->policy safely, the apropriate
4767 * runqueue lock must be held.
4769 rq = __task_rq_lock(p);
4770 /* recheck policy now with rq lock held */
4771 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4772 policy = oldpolicy = -1;
4773 __task_rq_unlock(rq);
4774 spin_unlock_irqrestore(&p->pi_lock, flags);
4777 update_rq_clock(rq);
4778 on_rq = p->se.on_rq;
4779 running = task_current(rq, p);
4781 deactivate_task(rq, p, 0);
4783 p->sched_class->put_prev_task(rq, p);
4786 __setscheduler(rq, p, policy, param->sched_priority);
4789 p->sched_class->set_curr_task(rq);
4791 activate_task(rq, p, 0);
4793 check_class_changed(rq, p, prev_class, oldprio, running);
4795 __task_rq_unlock(rq);
4796 spin_unlock_irqrestore(&p->pi_lock, flags);
4798 rt_mutex_adjust_pi(p);
4802 EXPORT_SYMBOL_GPL(sched_setscheduler);
4805 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4807 struct sched_param lparam;
4808 struct task_struct *p;
4811 if (!param || pid < 0)
4813 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4818 p = find_process_by_pid(pid);
4820 retval = sched_setscheduler(p, policy, &lparam);
4827 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4828 * @pid: the pid in question.
4829 * @policy: new policy.
4830 * @param: structure containing the new RT priority.
4833 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4835 /* negative values for policy are not valid */
4839 return do_sched_setscheduler(pid, policy, param);
4843 * sys_sched_setparam - set/change the RT priority of a thread
4844 * @pid: the pid in question.
4845 * @param: structure containing the new RT priority.
4847 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4849 return do_sched_setscheduler(pid, -1, param);
4853 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4854 * @pid: the pid in question.
4856 asmlinkage long sys_sched_getscheduler(pid_t pid)
4858 struct task_struct *p;
4865 read_lock(&tasklist_lock);
4866 p = find_process_by_pid(pid);
4868 retval = security_task_getscheduler(p);
4872 read_unlock(&tasklist_lock);
4877 * sys_sched_getscheduler - get the RT priority of a thread
4878 * @pid: the pid in question.
4879 * @param: structure containing the RT priority.
4881 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4883 struct sched_param lp;
4884 struct task_struct *p;
4887 if (!param || pid < 0)
4890 read_lock(&tasklist_lock);
4891 p = find_process_by_pid(pid);
4896 retval = security_task_getscheduler(p);
4900 lp.sched_priority = p->rt_priority;
4901 read_unlock(&tasklist_lock);
4904 * This one might sleep, we cannot do it with a spinlock held ...
4906 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4911 read_unlock(&tasklist_lock);
4915 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4917 cpumask_t cpus_allowed;
4918 cpumask_t new_mask = *in_mask;
4919 struct task_struct *p;
4923 read_lock(&tasklist_lock);
4925 p = find_process_by_pid(pid);
4927 read_unlock(&tasklist_lock);
4933 * It is not safe to call set_cpus_allowed with the
4934 * tasklist_lock held. We will bump the task_struct's
4935 * usage count and then drop tasklist_lock.
4938 read_unlock(&tasklist_lock);
4941 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4942 !capable(CAP_SYS_NICE))
4945 retval = security_task_setscheduler(p, 0, NULL);
4949 cpuset_cpus_allowed(p, &cpus_allowed);
4950 cpus_and(new_mask, new_mask, cpus_allowed);
4952 retval = set_cpus_allowed_ptr(p, &new_mask);
4955 cpuset_cpus_allowed(p, &cpus_allowed);
4956 if (!cpus_subset(new_mask, cpus_allowed)) {
4958 * We must have raced with a concurrent cpuset
4959 * update. Just reset the cpus_allowed to the
4960 * cpuset's cpus_allowed
4962 new_mask = cpus_allowed;
4972 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4973 cpumask_t *new_mask)
4975 if (len < sizeof(cpumask_t)) {
4976 memset(new_mask, 0, sizeof(cpumask_t));
4977 } else if (len > sizeof(cpumask_t)) {
4978 len = sizeof(cpumask_t);
4980 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4984 * sys_sched_setaffinity - set the cpu affinity of a process
4985 * @pid: pid of the process
4986 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4987 * @user_mask_ptr: user-space pointer to the new cpu mask
4989 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4990 unsigned long __user *user_mask_ptr)
4995 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4999 return sched_setaffinity(pid, &new_mask);
5003 * Represents all cpu's present in the system
5004 * In systems capable of hotplug, this map could dynamically grow
5005 * as new cpu's are detected in the system via any platform specific
5006 * method, such as ACPI for e.g.
5009 cpumask_t cpu_present_map __read_mostly;
5010 EXPORT_SYMBOL(cpu_present_map);
5013 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5014 EXPORT_SYMBOL(cpu_online_map);
5016 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5017 EXPORT_SYMBOL(cpu_possible_map);
5020 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5022 struct task_struct *p;
5026 read_lock(&tasklist_lock);
5029 p = find_process_by_pid(pid);
5033 retval = security_task_getscheduler(p);
5037 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5040 read_unlock(&tasklist_lock);
5047 * sys_sched_getaffinity - get the cpu affinity of a process
5048 * @pid: pid of the process
5049 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5050 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5052 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5053 unsigned long __user *user_mask_ptr)
5058 if (len < sizeof(cpumask_t))
5061 ret = sched_getaffinity(pid, &mask);
5065 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5068 return sizeof(cpumask_t);
5072 * sys_sched_yield - yield the current processor to other threads.
5074 * This function yields the current CPU to other tasks. If there are no
5075 * other threads running on this CPU then this function will return.
5077 asmlinkage long sys_sched_yield(void)
5079 struct rq *rq = this_rq_lock();
5081 schedstat_inc(rq, yld_count);
5082 current->sched_class->yield_task(rq);
5085 * Since we are going to call schedule() anyway, there's
5086 * no need to preempt or enable interrupts:
5088 __release(rq->lock);
5089 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5090 _raw_spin_unlock(&rq->lock);
5091 preempt_enable_no_resched();
5098 static void __cond_resched(void)
5100 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5101 __might_sleep(__FILE__, __LINE__);
5104 * The BKS might be reacquired before we have dropped
5105 * PREEMPT_ACTIVE, which could trigger a second
5106 * cond_resched() call.
5109 add_preempt_count(PREEMPT_ACTIVE);
5111 sub_preempt_count(PREEMPT_ACTIVE);
5112 } while (need_resched());
5115 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5116 int __sched _cond_resched(void)
5118 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5119 system_state == SYSTEM_RUNNING) {
5125 EXPORT_SYMBOL(_cond_resched);
5129 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5130 * call schedule, and on return reacquire the lock.
5132 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5133 * operations here to prevent schedule() from being called twice (once via
5134 * spin_unlock(), once by hand).
5136 int cond_resched_lock(spinlock_t *lock)
5138 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5141 if (spin_needbreak(lock) || resched) {
5143 if (resched && need_resched())
5152 EXPORT_SYMBOL(cond_resched_lock);
5154 int __sched cond_resched_softirq(void)
5156 BUG_ON(!in_softirq());
5158 if (need_resched() && system_state == SYSTEM_RUNNING) {
5166 EXPORT_SYMBOL(cond_resched_softirq);
5169 * yield - yield the current processor to other threads.
5171 * This is a shortcut for kernel-space yielding - it marks the
5172 * thread runnable and calls sys_sched_yield().
5174 void __sched yield(void)
5176 set_current_state(TASK_RUNNING);
5179 EXPORT_SYMBOL(yield);
5182 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5183 * that process accounting knows that this is a task in IO wait state.
5185 * But don't do that if it is a deliberate, throttling IO wait (this task
5186 * has set its backing_dev_info: the queue against which it should throttle)
5188 void __sched io_schedule(void)
5190 struct rq *rq = &__raw_get_cpu_var(runqueues);
5192 delayacct_blkio_start();
5193 atomic_inc(&rq->nr_iowait);
5195 atomic_dec(&rq->nr_iowait);
5196 delayacct_blkio_end();
5198 EXPORT_SYMBOL(io_schedule);
5200 long __sched io_schedule_timeout(long timeout)
5202 struct rq *rq = &__raw_get_cpu_var(runqueues);
5205 delayacct_blkio_start();
5206 atomic_inc(&rq->nr_iowait);
5207 ret = schedule_timeout(timeout);
5208 atomic_dec(&rq->nr_iowait);
5209 delayacct_blkio_end();
5214 * sys_sched_get_priority_max - return maximum RT priority.
5215 * @policy: scheduling class.
5217 * this syscall returns the maximum rt_priority that can be used
5218 * by a given scheduling class.
5220 asmlinkage long sys_sched_get_priority_max(int policy)
5227 ret = MAX_USER_RT_PRIO-1;
5239 * sys_sched_get_priority_min - return minimum RT priority.
5240 * @policy: scheduling class.
5242 * this syscall returns the minimum rt_priority that can be used
5243 * by a given scheduling class.
5245 asmlinkage long sys_sched_get_priority_min(int policy)
5263 * sys_sched_rr_get_interval - return the default timeslice of a process.
5264 * @pid: pid of the process.
5265 * @interval: userspace pointer to the timeslice value.
5267 * this syscall writes the default timeslice value of a given process
5268 * into the user-space timespec buffer. A value of '0' means infinity.
5271 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5273 struct task_struct *p;
5274 unsigned int time_slice;
5282 read_lock(&tasklist_lock);
5283 p = find_process_by_pid(pid);
5287 retval = security_task_getscheduler(p);
5292 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5293 * tasks that are on an otherwise idle runqueue:
5296 if (p->policy == SCHED_RR) {
5297 time_slice = DEF_TIMESLICE;
5298 } else if (p->policy != SCHED_FIFO) {
5299 struct sched_entity *se = &p->se;
5300 unsigned long flags;
5303 rq = task_rq_lock(p, &flags);
5304 if (rq->cfs.load.weight)
5305 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5306 task_rq_unlock(rq, &flags);
5308 read_unlock(&tasklist_lock);
5309 jiffies_to_timespec(time_slice, &t);
5310 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5314 read_unlock(&tasklist_lock);
5318 static const char stat_nam[] = "RSDTtZX";
5320 void sched_show_task(struct task_struct *p)
5322 unsigned long free = 0;
5325 state = p->state ? __ffs(p->state) + 1 : 0;
5326 printk(KERN_INFO "%-13.13s %c", p->comm,
5327 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5328 #if BITS_PER_LONG == 32
5329 if (state == TASK_RUNNING)
5330 printk(KERN_CONT " running ");
5332 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5334 if (state == TASK_RUNNING)
5335 printk(KERN_CONT " running task ");
5337 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5339 #ifdef CONFIG_DEBUG_STACK_USAGE
5341 unsigned long *n = end_of_stack(p);
5344 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5347 printk(KERN_CONT "%5lu %5d %6d\n", free,
5348 task_pid_nr(p), task_pid_nr(p->real_parent));
5350 show_stack(p, NULL);
5353 void show_state_filter(unsigned long state_filter)
5355 struct task_struct *g, *p;
5357 #if BITS_PER_LONG == 32
5359 " task PC stack pid father\n");
5362 " task PC stack pid father\n");
5364 read_lock(&tasklist_lock);
5365 do_each_thread(g, p) {
5367 * reset the NMI-timeout, listing all files on a slow
5368 * console might take alot of time:
5370 touch_nmi_watchdog();
5371 if (!state_filter || (p->state & state_filter))
5373 } while_each_thread(g, p);
5375 touch_all_softlockup_watchdogs();
5377 #ifdef CONFIG_SCHED_DEBUG
5378 sysrq_sched_debug_show();
5380 read_unlock(&tasklist_lock);
5382 * Only show locks if all tasks are dumped:
5384 if (state_filter == -1)
5385 debug_show_all_locks();
5388 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5390 idle->sched_class = &idle_sched_class;
5394 * init_idle - set up an idle thread for a given CPU
5395 * @idle: task in question
5396 * @cpu: cpu the idle task belongs to
5398 * NOTE: this function does not set the idle thread's NEED_RESCHED
5399 * flag, to make booting more robust.
5401 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5403 struct rq *rq = cpu_rq(cpu);
5404 unsigned long flags;
5407 idle->se.exec_start = sched_clock();
5409 idle->prio = idle->normal_prio = MAX_PRIO;
5410 idle->cpus_allowed = cpumask_of_cpu(cpu);
5411 __set_task_cpu(idle, cpu);
5413 spin_lock_irqsave(&rq->lock, flags);
5414 rq->curr = rq->idle = idle;
5415 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5418 spin_unlock_irqrestore(&rq->lock, flags);
5420 /* Set the preempt count _outside_ the spinlocks! */
5421 task_thread_info(idle)->preempt_count = 0;
5424 * The idle tasks have their own, simple scheduling class:
5426 idle->sched_class = &idle_sched_class;
5430 * In a system that switches off the HZ timer nohz_cpu_mask
5431 * indicates which cpus entered this state. This is used
5432 * in the rcu update to wait only for active cpus. For system
5433 * which do not switch off the HZ timer nohz_cpu_mask should
5434 * always be CPU_MASK_NONE.
5436 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5439 * Increase the granularity value when there are more CPUs,
5440 * because with more CPUs the 'effective latency' as visible
5441 * to users decreases. But the relationship is not linear,
5442 * so pick a second-best guess by going with the log2 of the
5445 * This idea comes from the SD scheduler of Con Kolivas:
5447 static inline void sched_init_granularity(void)
5449 unsigned int factor = 1 + ilog2(num_online_cpus());
5450 const unsigned long limit = 200000000;
5452 sysctl_sched_min_granularity *= factor;
5453 if (sysctl_sched_min_granularity > limit)
5454 sysctl_sched_min_granularity = limit;
5456 sysctl_sched_latency *= factor;
5457 if (sysctl_sched_latency > limit)
5458 sysctl_sched_latency = limit;
5460 sysctl_sched_wakeup_granularity *= factor;
5465 * This is how migration works:
5467 * 1) we queue a struct migration_req structure in the source CPU's
5468 * runqueue and wake up that CPU's migration thread.
5469 * 2) we down() the locked semaphore => thread blocks.
5470 * 3) migration thread wakes up (implicitly it forces the migrated
5471 * thread off the CPU)
5472 * 4) it gets the migration request and checks whether the migrated
5473 * task is still in the wrong runqueue.
5474 * 5) if it's in the wrong runqueue then the migration thread removes
5475 * it and puts it into the right queue.
5476 * 6) migration thread up()s the semaphore.
5477 * 7) we wake up and the migration is done.
5481 * Change a given task's CPU affinity. Migrate the thread to a
5482 * proper CPU and schedule it away if the CPU it's executing on
5483 * is removed from the allowed bitmask.
5485 * NOTE: the caller must have a valid reference to the task, the
5486 * task must not exit() & deallocate itself prematurely. The
5487 * call is not atomic; no spinlocks may be held.
5489 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5491 struct migration_req req;
5492 unsigned long flags;
5496 rq = task_rq_lock(p, &flags);
5497 if (!cpus_intersects(new_mask, cpu_online_map)) {
5502 if (p->sched_class->set_cpus_allowed)
5503 p->sched_class->set_cpus_allowed(p, &new_mask);
5505 p->cpus_allowed = new_mask;
5506 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5509 /* Can the task run on the task's current CPU? If so, we're done */
5510 if (cpu_isset(task_cpu(p), new_mask))
5513 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5514 /* Need help from migration thread: drop lock and wait. */
5515 task_rq_unlock(rq, &flags);
5516 wake_up_process(rq->migration_thread);
5517 wait_for_completion(&req.done);
5518 tlb_migrate_finish(p->mm);
5522 task_rq_unlock(rq, &flags);
5526 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5529 * Move (not current) task off this cpu, onto dest cpu. We're doing
5530 * this because either it can't run here any more (set_cpus_allowed()
5531 * away from this CPU, or CPU going down), or because we're
5532 * attempting to rebalance this task on exec (sched_exec).
5534 * So we race with normal scheduler movements, but that's OK, as long
5535 * as the task is no longer on this CPU.
5537 * Returns non-zero if task was successfully migrated.
5539 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5541 struct rq *rq_dest, *rq_src;
5544 if (unlikely(cpu_is_offline(dest_cpu)))
5547 rq_src = cpu_rq(src_cpu);
5548 rq_dest = cpu_rq(dest_cpu);
5550 double_rq_lock(rq_src, rq_dest);
5551 /* Already moved. */
5552 if (task_cpu(p) != src_cpu)
5554 /* Affinity changed (again). */
5555 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5558 on_rq = p->se.on_rq;
5560 deactivate_task(rq_src, p, 0);
5562 set_task_cpu(p, dest_cpu);
5564 activate_task(rq_dest, p, 0);
5565 check_preempt_curr(rq_dest, p);
5569 double_rq_unlock(rq_src, rq_dest);
5574 * migration_thread - this is a highprio system thread that performs
5575 * thread migration by bumping thread off CPU then 'pushing' onto
5578 static int migration_thread(void *data)
5580 int cpu = (long)data;
5584 BUG_ON(rq->migration_thread != current);
5586 set_current_state(TASK_INTERRUPTIBLE);
5587 while (!kthread_should_stop()) {
5588 struct migration_req *req;
5589 struct list_head *head;
5591 spin_lock_irq(&rq->lock);
5593 if (cpu_is_offline(cpu)) {
5594 spin_unlock_irq(&rq->lock);
5598 if (rq->active_balance) {
5599 active_load_balance(rq, cpu);
5600 rq->active_balance = 0;
5603 head = &rq->migration_queue;
5605 if (list_empty(head)) {
5606 spin_unlock_irq(&rq->lock);
5608 set_current_state(TASK_INTERRUPTIBLE);
5611 req = list_entry(head->next, struct migration_req, list);
5612 list_del_init(head->next);
5614 spin_unlock(&rq->lock);
5615 __migrate_task(req->task, cpu, req->dest_cpu);
5618 complete(&req->done);
5620 __set_current_state(TASK_RUNNING);
5624 /* Wait for kthread_stop */
5625 set_current_state(TASK_INTERRUPTIBLE);
5626 while (!kthread_should_stop()) {
5628 set_current_state(TASK_INTERRUPTIBLE);
5630 __set_current_state(TASK_RUNNING);
5634 #ifdef CONFIG_HOTPLUG_CPU
5636 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5640 local_irq_disable();
5641 ret = __migrate_task(p, src_cpu, dest_cpu);
5647 * Figure out where task on dead CPU should go, use force if necessary.
5648 * NOTE: interrupts should be disabled by the caller
5650 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5652 unsigned long flags;
5659 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5660 cpus_and(mask, mask, p->cpus_allowed);
5661 dest_cpu = any_online_cpu(mask);
5663 /* On any allowed CPU? */
5664 if (dest_cpu >= nr_cpu_ids)
5665 dest_cpu = any_online_cpu(p->cpus_allowed);
5667 /* No more Mr. Nice Guy. */
5668 if (dest_cpu >= nr_cpu_ids) {
5669 cpumask_t cpus_allowed;
5671 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5673 * Try to stay on the same cpuset, where the
5674 * current cpuset may be a subset of all cpus.
5675 * The cpuset_cpus_allowed_locked() variant of
5676 * cpuset_cpus_allowed() will not block. It must be
5677 * called within calls to cpuset_lock/cpuset_unlock.
5679 rq = task_rq_lock(p, &flags);
5680 p->cpus_allowed = cpus_allowed;
5681 dest_cpu = any_online_cpu(p->cpus_allowed);
5682 task_rq_unlock(rq, &flags);
5685 * Don't tell them about moving exiting tasks or
5686 * kernel threads (both mm NULL), since they never
5689 if (p->mm && printk_ratelimit()) {
5690 printk(KERN_INFO "process %d (%s) no "
5691 "longer affine to cpu%d\n",
5692 task_pid_nr(p), p->comm, dead_cpu);
5695 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5699 * While a dead CPU has no uninterruptible tasks queued at this point,
5700 * it might still have a nonzero ->nr_uninterruptible counter, because
5701 * for performance reasons the counter is not stricly tracking tasks to
5702 * their home CPUs. So we just add the counter to another CPU's counter,
5703 * to keep the global sum constant after CPU-down:
5705 static void migrate_nr_uninterruptible(struct rq *rq_src)
5707 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5708 unsigned long flags;
5710 local_irq_save(flags);
5711 double_rq_lock(rq_src, rq_dest);
5712 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5713 rq_src->nr_uninterruptible = 0;
5714 double_rq_unlock(rq_src, rq_dest);
5715 local_irq_restore(flags);
5718 /* Run through task list and migrate tasks from the dead cpu. */
5719 static void migrate_live_tasks(int src_cpu)
5721 struct task_struct *p, *t;
5723 read_lock(&tasklist_lock);
5725 do_each_thread(t, p) {
5729 if (task_cpu(p) == src_cpu)
5730 move_task_off_dead_cpu(src_cpu, p);
5731 } while_each_thread(t, p);
5733 read_unlock(&tasklist_lock);
5737 * Schedules idle task to be the next runnable task on current CPU.
5738 * It does so by boosting its priority to highest possible.
5739 * Used by CPU offline code.
5741 void sched_idle_next(void)
5743 int this_cpu = smp_processor_id();
5744 struct rq *rq = cpu_rq(this_cpu);
5745 struct task_struct *p = rq->idle;
5746 unsigned long flags;
5748 /* cpu has to be offline */
5749 BUG_ON(cpu_online(this_cpu));
5752 * Strictly not necessary since rest of the CPUs are stopped by now
5753 * and interrupts disabled on the current cpu.
5755 spin_lock_irqsave(&rq->lock, flags);
5757 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5759 update_rq_clock(rq);
5760 activate_task(rq, p, 0);
5762 spin_unlock_irqrestore(&rq->lock, flags);
5766 * Ensures that the idle task is using init_mm right before its cpu goes
5769 void idle_task_exit(void)
5771 struct mm_struct *mm = current->active_mm;
5773 BUG_ON(cpu_online(smp_processor_id()));
5776 switch_mm(mm, &init_mm, current);
5780 /* called under rq->lock with disabled interrupts */
5781 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5783 struct rq *rq = cpu_rq(dead_cpu);
5785 /* Must be exiting, otherwise would be on tasklist. */
5786 BUG_ON(!p->exit_state);
5788 /* Cannot have done final schedule yet: would have vanished. */
5789 BUG_ON(p->state == TASK_DEAD);
5794 * Drop lock around migration; if someone else moves it,
5795 * that's OK. No task can be added to this CPU, so iteration is
5798 spin_unlock_irq(&rq->lock);
5799 move_task_off_dead_cpu(dead_cpu, p);
5800 spin_lock_irq(&rq->lock);
5805 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5806 static void migrate_dead_tasks(unsigned int dead_cpu)
5808 struct rq *rq = cpu_rq(dead_cpu);
5809 struct task_struct *next;
5812 if (!rq->nr_running)
5814 update_rq_clock(rq);
5815 next = pick_next_task(rq, rq->curr);
5818 migrate_dead(dead_cpu, next);
5822 #endif /* CONFIG_HOTPLUG_CPU */
5824 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5826 static struct ctl_table sd_ctl_dir[] = {
5828 .procname = "sched_domain",
5834 static struct ctl_table sd_ctl_root[] = {
5836 .ctl_name = CTL_KERN,
5837 .procname = "kernel",
5839 .child = sd_ctl_dir,
5844 static struct ctl_table *sd_alloc_ctl_entry(int n)
5846 struct ctl_table *entry =
5847 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5852 static void sd_free_ctl_entry(struct ctl_table **tablep)
5854 struct ctl_table *entry;
5857 * In the intermediate directories, both the child directory and
5858 * procname are dynamically allocated and could fail but the mode
5859 * will always be set. In the lowest directory the names are
5860 * static strings and all have proc handlers.
5862 for (entry = *tablep; entry->mode; entry++) {
5864 sd_free_ctl_entry(&entry->child);
5865 if (entry->proc_handler == NULL)
5866 kfree(entry->procname);
5874 set_table_entry(struct ctl_table *entry,
5875 const char *procname, void *data, int maxlen,
5876 mode_t mode, proc_handler *proc_handler)
5878 entry->procname = procname;
5880 entry->maxlen = maxlen;
5882 entry->proc_handler = proc_handler;
5885 static struct ctl_table *
5886 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5888 struct ctl_table *table = sd_alloc_ctl_entry(12);
5893 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5894 sizeof(long), 0644, proc_doulongvec_minmax);
5895 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5896 sizeof(long), 0644, proc_doulongvec_minmax);
5897 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5898 sizeof(int), 0644, proc_dointvec_minmax);
5899 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5900 sizeof(int), 0644, proc_dointvec_minmax);
5901 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5902 sizeof(int), 0644, proc_dointvec_minmax);
5903 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5904 sizeof(int), 0644, proc_dointvec_minmax);
5905 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5906 sizeof(int), 0644, proc_dointvec_minmax);
5907 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5908 sizeof(int), 0644, proc_dointvec_minmax);
5909 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5910 sizeof(int), 0644, proc_dointvec_minmax);
5911 set_table_entry(&table[9], "cache_nice_tries",
5912 &sd->cache_nice_tries,
5913 sizeof(int), 0644, proc_dointvec_minmax);
5914 set_table_entry(&table[10], "flags", &sd->flags,
5915 sizeof(int), 0644, proc_dointvec_minmax);
5916 /* &table[11] is terminator */
5921 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5923 struct ctl_table *entry, *table;
5924 struct sched_domain *sd;
5925 int domain_num = 0, i;
5928 for_each_domain(cpu, sd)
5930 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5935 for_each_domain(cpu, sd) {
5936 snprintf(buf, 32, "domain%d", i);
5937 entry->procname = kstrdup(buf, GFP_KERNEL);
5939 entry->child = sd_alloc_ctl_domain_table(sd);
5946 static struct ctl_table_header *sd_sysctl_header;
5947 static void register_sched_domain_sysctl(void)
5949 int i, cpu_num = num_online_cpus();
5950 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5953 WARN_ON(sd_ctl_dir[0].child);
5954 sd_ctl_dir[0].child = entry;
5959 for_each_online_cpu(i) {
5960 snprintf(buf, 32, "cpu%d", i);
5961 entry->procname = kstrdup(buf, GFP_KERNEL);
5963 entry->child = sd_alloc_ctl_cpu_table(i);
5967 WARN_ON(sd_sysctl_header);
5968 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5971 /* may be called multiple times per register */
5972 static void unregister_sched_domain_sysctl(void)
5974 if (sd_sysctl_header)
5975 unregister_sysctl_table(sd_sysctl_header);
5976 sd_sysctl_header = NULL;
5977 if (sd_ctl_dir[0].child)
5978 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5981 static void register_sched_domain_sysctl(void)
5984 static void unregister_sched_domain_sysctl(void)
5990 * migration_call - callback that gets triggered when a CPU is added.
5991 * Here we can start up the necessary migration thread for the new CPU.
5993 static int __cpuinit
5994 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5996 struct task_struct *p;
5997 int cpu = (long)hcpu;
5998 unsigned long flags;
6003 case CPU_UP_PREPARE:
6004 case CPU_UP_PREPARE_FROZEN:
6005 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6008 kthread_bind(p, cpu);
6009 /* Must be high prio: stop_machine expects to yield to it. */
6010 rq = task_rq_lock(p, &flags);
6011 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6012 task_rq_unlock(rq, &flags);
6013 cpu_rq(cpu)->migration_thread = p;
6017 case CPU_ONLINE_FROZEN:
6018 /* Strictly unnecessary, as first user will wake it. */
6019 wake_up_process(cpu_rq(cpu)->migration_thread);
6021 /* Update our root-domain */
6023 spin_lock_irqsave(&rq->lock, flags);
6025 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6026 cpu_set(cpu, rq->rd->online);
6028 spin_unlock_irqrestore(&rq->lock, flags);
6031 #ifdef CONFIG_HOTPLUG_CPU
6032 case CPU_UP_CANCELED:
6033 case CPU_UP_CANCELED_FROZEN:
6034 if (!cpu_rq(cpu)->migration_thread)
6036 /* Unbind it from offline cpu so it can run. Fall thru. */
6037 kthread_bind(cpu_rq(cpu)->migration_thread,
6038 any_online_cpu(cpu_online_map));
6039 kthread_stop(cpu_rq(cpu)->migration_thread);
6040 cpu_rq(cpu)->migration_thread = NULL;
6044 case CPU_DEAD_FROZEN:
6045 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6046 migrate_live_tasks(cpu);
6048 kthread_stop(rq->migration_thread);
6049 rq->migration_thread = NULL;
6050 /* Idle task back to normal (off runqueue, low prio) */
6051 spin_lock_irq(&rq->lock);
6052 update_rq_clock(rq);
6053 deactivate_task(rq, rq->idle, 0);
6054 rq->idle->static_prio = MAX_PRIO;
6055 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6056 rq->idle->sched_class = &idle_sched_class;
6057 migrate_dead_tasks(cpu);
6058 spin_unlock_irq(&rq->lock);
6060 migrate_nr_uninterruptible(rq);
6061 BUG_ON(rq->nr_running != 0);
6064 * No need to migrate the tasks: it was best-effort if
6065 * they didn't take sched_hotcpu_mutex. Just wake up
6068 spin_lock_irq(&rq->lock);
6069 while (!list_empty(&rq->migration_queue)) {
6070 struct migration_req *req;
6072 req = list_entry(rq->migration_queue.next,
6073 struct migration_req, list);
6074 list_del_init(&req->list);
6075 complete(&req->done);
6077 spin_unlock_irq(&rq->lock);
6081 case CPU_DYING_FROZEN:
6082 /* Update our root-domain */
6084 spin_lock_irqsave(&rq->lock, flags);
6086 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6087 cpu_clear(cpu, rq->rd->online);
6089 spin_unlock_irqrestore(&rq->lock, flags);
6096 /* Register at highest priority so that task migration (migrate_all_tasks)
6097 * happens before everything else.
6099 static struct notifier_block __cpuinitdata migration_notifier = {
6100 .notifier_call = migration_call,
6104 void __init migration_init(void)
6106 void *cpu = (void *)(long)smp_processor_id();
6109 /* Start one for the boot CPU: */
6110 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6111 BUG_ON(err == NOTIFY_BAD);
6112 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6113 register_cpu_notifier(&migration_notifier);
6119 /* Number of possible processor ids */
6120 int nr_cpu_ids __read_mostly = NR_CPUS;
6121 EXPORT_SYMBOL(nr_cpu_ids);
6123 #ifdef CONFIG_SCHED_DEBUG
6125 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6126 cpumask_t *groupmask)
6128 struct sched_group *group = sd->groups;
6131 cpulist_scnprintf(str, sizeof(str), sd->span);
6132 cpus_clear(*groupmask);
6134 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6136 if (!(sd->flags & SD_LOAD_BALANCE)) {
6137 printk("does not load-balance\n");
6139 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6144 printk(KERN_CONT "span %s\n", str);
6146 if (!cpu_isset(cpu, sd->span)) {
6147 printk(KERN_ERR "ERROR: domain->span does not contain "
6150 if (!cpu_isset(cpu, group->cpumask)) {
6151 printk(KERN_ERR "ERROR: domain->groups does not contain"
6155 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6159 printk(KERN_ERR "ERROR: group is NULL\n");
6163 if (!group->__cpu_power) {
6164 printk(KERN_CONT "\n");
6165 printk(KERN_ERR "ERROR: domain->cpu_power not "
6170 if (!cpus_weight(group->cpumask)) {
6171 printk(KERN_CONT "\n");
6172 printk(KERN_ERR "ERROR: empty group\n");
6176 if (cpus_intersects(*groupmask, group->cpumask)) {
6177 printk(KERN_CONT "\n");
6178 printk(KERN_ERR "ERROR: repeated CPUs\n");
6182 cpus_or(*groupmask, *groupmask, group->cpumask);
6184 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6185 printk(KERN_CONT " %s", str);
6187 group = group->next;
6188 } while (group != sd->groups);
6189 printk(KERN_CONT "\n");
6191 if (!cpus_equal(sd->span, *groupmask))
6192 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6194 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6195 printk(KERN_ERR "ERROR: parent span is not a superset "
6196 "of domain->span\n");
6200 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6202 cpumask_t *groupmask;
6206 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6210 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6212 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6214 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6219 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6229 # define sched_domain_debug(sd, cpu) do { } while (0)
6232 static int sd_degenerate(struct sched_domain *sd)
6234 if (cpus_weight(sd->span) == 1)
6237 /* Following flags need at least 2 groups */
6238 if (sd->flags & (SD_LOAD_BALANCE |
6239 SD_BALANCE_NEWIDLE |
6243 SD_SHARE_PKG_RESOURCES)) {
6244 if (sd->groups != sd->groups->next)
6248 /* Following flags don't use groups */
6249 if (sd->flags & (SD_WAKE_IDLE |
6258 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6260 unsigned long cflags = sd->flags, pflags = parent->flags;
6262 if (sd_degenerate(parent))
6265 if (!cpus_equal(sd->span, parent->span))
6268 /* Does parent contain flags not in child? */
6269 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6270 if (cflags & SD_WAKE_AFFINE)
6271 pflags &= ~SD_WAKE_BALANCE;
6272 /* Flags needing groups don't count if only 1 group in parent */
6273 if (parent->groups == parent->groups->next) {
6274 pflags &= ~(SD_LOAD_BALANCE |
6275 SD_BALANCE_NEWIDLE |
6279 SD_SHARE_PKG_RESOURCES);
6281 if (~cflags & pflags)
6287 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6289 unsigned long flags;
6290 const struct sched_class *class;
6292 spin_lock_irqsave(&rq->lock, flags);
6295 struct root_domain *old_rd = rq->rd;
6297 for (class = sched_class_highest; class; class = class->next) {
6298 if (class->leave_domain)
6299 class->leave_domain(rq);
6302 cpu_clear(rq->cpu, old_rd->span);
6303 cpu_clear(rq->cpu, old_rd->online);
6305 if (atomic_dec_and_test(&old_rd->refcount))
6309 atomic_inc(&rd->refcount);
6312 cpu_set(rq->cpu, rd->span);
6313 if (cpu_isset(rq->cpu, cpu_online_map))
6314 cpu_set(rq->cpu, rd->online);
6316 for (class = sched_class_highest; class; class = class->next) {
6317 if (class->join_domain)
6318 class->join_domain(rq);
6321 spin_unlock_irqrestore(&rq->lock, flags);
6324 static void init_rootdomain(struct root_domain *rd)
6326 memset(rd, 0, sizeof(*rd));
6328 cpus_clear(rd->span);
6329 cpus_clear(rd->online);
6332 static void init_defrootdomain(void)
6334 init_rootdomain(&def_root_domain);
6335 atomic_set(&def_root_domain.refcount, 1);
6338 static struct root_domain *alloc_rootdomain(void)
6340 struct root_domain *rd;
6342 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6346 init_rootdomain(rd);
6352 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6353 * hold the hotplug lock.
6356 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6358 struct rq *rq = cpu_rq(cpu);
6359 struct sched_domain *tmp;
6361 /* Remove the sched domains which do not contribute to scheduling. */
6362 for (tmp = sd; tmp; tmp = tmp->parent) {
6363 struct sched_domain *parent = tmp->parent;
6366 if (sd_parent_degenerate(tmp, parent)) {
6367 tmp->parent = parent->parent;
6369 parent->parent->child = tmp;
6373 if (sd && sd_degenerate(sd)) {
6379 sched_domain_debug(sd, cpu);
6381 rq_attach_root(rq, rd);
6382 rcu_assign_pointer(rq->sd, sd);
6385 /* cpus with isolated domains */
6386 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6388 /* Setup the mask of cpus configured for isolated domains */
6389 static int __init isolated_cpu_setup(char *str)
6391 int ints[NR_CPUS], i;
6393 str = get_options(str, ARRAY_SIZE(ints), ints);
6394 cpus_clear(cpu_isolated_map);
6395 for (i = 1; i <= ints[0]; i++)
6396 if (ints[i] < NR_CPUS)
6397 cpu_set(ints[i], cpu_isolated_map);
6401 __setup("isolcpus=", isolated_cpu_setup);
6404 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6405 * to a function which identifies what group(along with sched group) a CPU
6406 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6407 * (due to the fact that we keep track of groups covered with a cpumask_t).
6409 * init_sched_build_groups will build a circular linked list of the groups
6410 * covered by the given span, and will set each group's ->cpumask correctly,
6411 * and ->cpu_power to 0.
6414 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6415 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6416 struct sched_group **sg,
6417 cpumask_t *tmpmask),
6418 cpumask_t *covered, cpumask_t *tmpmask)
6420 struct sched_group *first = NULL, *last = NULL;
6423 cpus_clear(*covered);
6425 for_each_cpu_mask(i, *span) {
6426 struct sched_group *sg;
6427 int group = group_fn(i, cpu_map, &sg, tmpmask);
6430 if (cpu_isset(i, *covered))
6433 cpus_clear(sg->cpumask);
6434 sg->__cpu_power = 0;
6436 for_each_cpu_mask(j, *span) {
6437 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6440 cpu_set(j, *covered);
6441 cpu_set(j, sg->cpumask);
6452 #define SD_NODES_PER_DOMAIN 16
6457 * find_next_best_node - find the next node to include in a sched_domain
6458 * @node: node whose sched_domain we're building
6459 * @used_nodes: nodes already in the sched_domain
6461 * Find the next node to include in a given scheduling domain. Simply
6462 * finds the closest node not already in the @used_nodes map.
6464 * Should use nodemask_t.
6466 static int find_next_best_node(int node, nodemask_t *used_nodes)
6468 int i, n, val, min_val, best_node = 0;
6472 for (i = 0; i < MAX_NUMNODES; i++) {
6473 /* Start at @node */
6474 n = (node + i) % MAX_NUMNODES;
6476 if (!nr_cpus_node(n))
6479 /* Skip already used nodes */
6480 if (node_isset(n, *used_nodes))
6483 /* Simple min distance search */
6484 val = node_distance(node, n);
6486 if (val < min_val) {
6492 node_set(best_node, *used_nodes);
6497 * sched_domain_node_span - get a cpumask for a node's sched_domain
6498 * @node: node whose cpumask we're constructing
6500 * Given a node, construct a good cpumask for its sched_domain to span. It
6501 * should be one that prevents unnecessary balancing, but also spreads tasks
6504 static void sched_domain_node_span(int node, cpumask_t *span)
6506 nodemask_t used_nodes;
6507 node_to_cpumask_ptr(nodemask, node);
6511 nodes_clear(used_nodes);
6513 cpus_or(*span, *span, *nodemask);
6514 node_set(node, used_nodes);
6516 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6517 int next_node = find_next_best_node(node, &used_nodes);
6519 node_to_cpumask_ptr_next(nodemask, next_node);
6520 cpus_or(*span, *span, *nodemask);
6525 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6528 * SMT sched-domains:
6530 #ifdef CONFIG_SCHED_SMT
6531 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6532 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6535 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6539 *sg = &per_cpu(sched_group_cpus, cpu);
6545 * multi-core sched-domains:
6547 #ifdef CONFIG_SCHED_MC
6548 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6549 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6552 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6554 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6559 *mask = per_cpu(cpu_sibling_map, cpu);
6560 cpus_and(*mask, *mask, *cpu_map);
6561 group = first_cpu(*mask);
6563 *sg = &per_cpu(sched_group_core, group);
6566 #elif defined(CONFIG_SCHED_MC)
6568 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6572 *sg = &per_cpu(sched_group_core, cpu);
6577 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6578 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6581 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6585 #ifdef CONFIG_SCHED_MC
6586 *mask = cpu_coregroup_map(cpu);
6587 cpus_and(*mask, *mask, *cpu_map);
6588 group = first_cpu(*mask);
6589 #elif defined(CONFIG_SCHED_SMT)
6590 *mask = per_cpu(cpu_sibling_map, cpu);
6591 cpus_and(*mask, *mask, *cpu_map);
6592 group = first_cpu(*mask);
6597 *sg = &per_cpu(sched_group_phys, group);
6603 * The init_sched_build_groups can't handle what we want to do with node
6604 * groups, so roll our own. Now each node has its own list of groups which
6605 * gets dynamically allocated.
6607 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6608 static struct sched_group ***sched_group_nodes_bycpu;
6610 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6611 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6613 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6614 struct sched_group **sg, cpumask_t *nodemask)
6618 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6619 cpus_and(*nodemask, *nodemask, *cpu_map);
6620 group = first_cpu(*nodemask);
6623 *sg = &per_cpu(sched_group_allnodes, group);
6627 static void init_numa_sched_groups_power(struct sched_group *group_head)
6629 struct sched_group *sg = group_head;
6635 for_each_cpu_mask(j, sg->cpumask) {
6636 struct sched_domain *sd;
6638 sd = &per_cpu(phys_domains, j);
6639 if (j != first_cpu(sd->groups->cpumask)) {
6641 * Only add "power" once for each
6647 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6650 } while (sg != group_head);
6655 /* Free memory allocated for various sched_group structures */
6656 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6660 for_each_cpu_mask(cpu, *cpu_map) {
6661 struct sched_group **sched_group_nodes
6662 = sched_group_nodes_bycpu[cpu];
6664 if (!sched_group_nodes)
6667 for (i = 0; i < MAX_NUMNODES; i++) {
6668 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6670 *nodemask = node_to_cpumask(i);
6671 cpus_and(*nodemask, *nodemask, *cpu_map);
6672 if (cpus_empty(*nodemask))
6682 if (oldsg != sched_group_nodes[i])
6685 kfree(sched_group_nodes);
6686 sched_group_nodes_bycpu[cpu] = NULL;
6690 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6696 * Initialize sched groups cpu_power.
6698 * cpu_power indicates the capacity of sched group, which is used while
6699 * distributing the load between different sched groups in a sched domain.
6700 * Typically cpu_power for all the groups in a sched domain will be same unless
6701 * there are asymmetries in the topology. If there are asymmetries, group
6702 * having more cpu_power will pickup more load compared to the group having
6705 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6706 * the maximum number of tasks a group can handle in the presence of other idle
6707 * or lightly loaded groups in the same sched domain.
6709 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6711 struct sched_domain *child;
6712 struct sched_group *group;
6714 WARN_ON(!sd || !sd->groups);
6716 if (cpu != first_cpu(sd->groups->cpumask))
6721 sd->groups->__cpu_power = 0;
6724 * For perf policy, if the groups in child domain share resources
6725 * (for example cores sharing some portions of the cache hierarchy
6726 * or SMT), then set this domain groups cpu_power such that each group
6727 * can handle only one task, when there are other idle groups in the
6728 * same sched domain.
6730 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6732 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6733 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6738 * add cpu_power of each child group to this groups cpu_power
6740 group = child->groups;
6742 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6743 group = group->next;
6744 } while (group != child->groups);
6748 * Initializers for schedule domains
6749 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6752 #define SD_INIT(sd, type) sd_init_##type(sd)
6753 #define SD_INIT_FUNC(type) \
6754 static noinline void sd_init_##type(struct sched_domain *sd) \
6756 memset(sd, 0, sizeof(*sd)); \
6757 *sd = SD_##type##_INIT; \
6762 SD_INIT_FUNC(ALLNODES)
6765 #ifdef CONFIG_SCHED_SMT
6766 SD_INIT_FUNC(SIBLING)
6768 #ifdef CONFIG_SCHED_MC
6773 * To minimize stack usage kmalloc room for cpumasks and share the
6774 * space as the usage in build_sched_domains() dictates. Used only
6775 * if the amount of space is significant.
6778 cpumask_t tmpmask; /* make this one first */
6781 cpumask_t this_sibling_map;
6782 cpumask_t this_core_map;
6784 cpumask_t send_covered;
6787 cpumask_t domainspan;
6789 cpumask_t notcovered;
6794 #define SCHED_CPUMASK_ALLOC 1
6795 #define SCHED_CPUMASK_FREE(v) kfree(v)
6796 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6798 #define SCHED_CPUMASK_ALLOC 0
6799 #define SCHED_CPUMASK_FREE(v)
6800 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6803 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6804 ((unsigned long)(a) + offsetof(struct allmasks, v))
6807 * Build sched domains for a given set of cpus and attach the sched domains
6808 * to the individual cpus
6810 static int build_sched_domains(const cpumask_t *cpu_map)
6813 struct root_domain *rd;
6814 SCHED_CPUMASK_DECLARE(allmasks);
6817 struct sched_group **sched_group_nodes = NULL;
6818 int sd_allnodes = 0;
6821 * Allocate the per-node list of sched groups
6823 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6825 if (!sched_group_nodes) {
6826 printk(KERN_WARNING "Can not alloc sched group node list\n");
6831 rd = alloc_rootdomain();
6833 printk(KERN_WARNING "Cannot alloc root domain\n");
6835 kfree(sched_group_nodes);
6840 #if SCHED_CPUMASK_ALLOC
6841 /* get space for all scratch cpumask variables */
6842 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6844 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6847 kfree(sched_group_nodes);
6852 tmpmask = (cpumask_t *)allmasks;
6856 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6860 * Set up domains for cpus specified by the cpu_map.
6862 for_each_cpu_mask(i, *cpu_map) {
6863 struct sched_domain *sd = NULL, *p;
6864 SCHED_CPUMASK_VAR(nodemask, allmasks);
6866 *nodemask = node_to_cpumask(cpu_to_node(i));
6867 cpus_and(*nodemask, *nodemask, *cpu_map);
6870 if (cpus_weight(*cpu_map) >
6871 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6872 sd = &per_cpu(allnodes_domains, i);
6873 SD_INIT(sd, ALLNODES);
6874 sd->span = *cpu_map;
6875 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6881 sd = &per_cpu(node_domains, i);
6883 sched_domain_node_span(cpu_to_node(i), &sd->span);
6887 cpus_and(sd->span, sd->span, *cpu_map);
6891 sd = &per_cpu(phys_domains, i);
6893 sd->span = *nodemask;
6897 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6899 #ifdef CONFIG_SCHED_MC
6901 sd = &per_cpu(core_domains, i);
6903 sd->span = cpu_coregroup_map(i);
6904 cpus_and(sd->span, sd->span, *cpu_map);
6907 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
6910 #ifdef CONFIG_SCHED_SMT
6912 sd = &per_cpu(cpu_domains, i);
6913 SD_INIT(sd, SIBLING);
6914 sd->span = per_cpu(cpu_sibling_map, i);
6915 cpus_and(sd->span, sd->span, *cpu_map);
6918 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
6922 #ifdef CONFIG_SCHED_SMT
6923 /* Set up CPU (sibling) groups */
6924 for_each_cpu_mask(i, *cpu_map) {
6925 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
6926 SCHED_CPUMASK_VAR(send_covered, allmasks);
6928 *this_sibling_map = per_cpu(cpu_sibling_map, i);
6929 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
6930 if (i != first_cpu(*this_sibling_map))
6933 init_sched_build_groups(this_sibling_map, cpu_map,
6935 send_covered, tmpmask);
6939 #ifdef CONFIG_SCHED_MC
6940 /* Set up multi-core groups */
6941 for_each_cpu_mask(i, *cpu_map) {
6942 SCHED_CPUMASK_VAR(this_core_map, allmasks);
6943 SCHED_CPUMASK_VAR(send_covered, allmasks);
6945 *this_core_map = cpu_coregroup_map(i);
6946 cpus_and(*this_core_map, *this_core_map, *cpu_map);
6947 if (i != first_cpu(*this_core_map))
6950 init_sched_build_groups(this_core_map, cpu_map,
6952 send_covered, tmpmask);
6956 /* Set up physical groups */
6957 for (i = 0; i < MAX_NUMNODES; i++) {
6958 SCHED_CPUMASK_VAR(nodemask, allmasks);
6959 SCHED_CPUMASK_VAR(send_covered, allmasks);
6961 *nodemask = node_to_cpumask(i);
6962 cpus_and(*nodemask, *nodemask, *cpu_map);
6963 if (cpus_empty(*nodemask))
6966 init_sched_build_groups(nodemask, cpu_map,
6968 send_covered, tmpmask);
6972 /* Set up node groups */
6974 SCHED_CPUMASK_VAR(send_covered, allmasks);
6976 init_sched_build_groups(cpu_map, cpu_map,
6977 &cpu_to_allnodes_group,
6978 send_covered, tmpmask);
6981 for (i = 0; i < MAX_NUMNODES; i++) {
6982 /* Set up node groups */
6983 struct sched_group *sg, *prev;
6984 SCHED_CPUMASK_VAR(nodemask, allmasks);
6985 SCHED_CPUMASK_VAR(domainspan, allmasks);
6986 SCHED_CPUMASK_VAR(covered, allmasks);
6989 *nodemask = node_to_cpumask(i);
6990 cpus_clear(*covered);
6992 cpus_and(*nodemask, *nodemask, *cpu_map);
6993 if (cpus_empty(*nodemask)) {
6994 sched_group_nodes[i] = NULL;
6998 sched_domain_node_span(i, domainspan);
6999 cpus_and(*domainspan, *domainspan, *cpu_map);
7001 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7003 printk(KERN_WARNING "Can not alloc domain group for "
7007 sched_group_nodes[i] = sg;
7008 for_each_cpu_mask(j, *nodemask) {
7009 struct sched_domain *sd;
7011 sd = &per_cpu(node_domains, j);
7014 sg->__cpu_power = 0;
7015 sg->cpumask = *nodemask;
7017 cpus_or(*covered, *covered, *nodemask);
7020 for (j = 0; j < MAX_NUMNODES; j++) {
7021 SCHED_CPUMASK_VAR(notcovered, allmasks);
7022 int n = (i + j) % MAX_NUMNODES;
7023 node_to_cpumask_ptr(pnodemask, n);
7025 cpus_complement(*notcovered, *covered);
7026 cpus_and(*tmpmask, *notcovered, *cpu_map);
7027 cpus_and(*tmpmask, *tmpmask, *domainspan);
7028 if (cpus_empty(*tmpmask))
7031 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7032 if (cpus_empty(*tmpmask))
7035 sg = kmalloc_node(sizeof(struct sched_group),
7039 "Can not alloc domain group for node %d\n", j);
7042 sg->__cpu_power = 0;
7043 sg->cpumask = *tmpmask;
7044 sg->next = prev->next;
7045 cpus_or(*covered, *covered, *tmpmask);
7052 /* Calculate CPU power for physical packages and nodes */
7053 #ifdef CONFIG_SCHED_SMT
7054 for_each_cpu_mask(i, *cpu_map) {
7055 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7057 init_sched_groups_power(i, sd);
7060 #ifdef CONFIG_SCHED_MC
7061 for_each_cpu_mask(i, *cpu_map) {
7062 struct sched_domain *sd = &per_cpu(core_domains, i);
7064 init_sched_groups_power(i, sd);
7068 for_each_cpu_mask(i, *cpu_map) {
7069 struct sched_domain *sd = &per_cpu(phys_domains, i);
7071 init_sched_groups_power(i, sd);
7075 for (i = 0; i < MAX_NUMNODES; i++)
7076 init_numa_sched_groups_power(sched_group_nodes[i]);
7079 struct sched_group *sg;
7081 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7083 init_numa_sched_groups_power(sg);
7087 /* Attach the domains */
7088 for_each_cpu_mask(i, *cpu_map) {
7089 struct sched_domain *sd;
7090 #ifdef CONFIG_SCHED_SMT
7091 sd = &per_cpu(cpu_domains, i);
7092 #elif defined(CONFIG_SCHED_MC)
7093 sd = &per_cpu(core_domains, i);
7095 sd = &per_cpu(phys_domains, i);
7097 cpu_attach_domain(sd, rd, i);
7100 SCHED_CPUMASK_FREE((void *)allmasks);
7105 free_sched_groups(cpu_map, tmpmask);
7106 SCHED_CPUMASK_FREE((void *)allmasks);
7111 static cpumask_t *doms_cur; /* current sched domains */
7112 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7115 * Special case: If a kmalloc of a doms_cur partition (array of
7116 * cpumask_t) fails, then fallback to a single sched domain,
7117 * as determined by the single cpumask_t fallback_doms.
7119 static cpumask_t fallback_doms;
7121 void __attribute__((weak)) arch_update_cpu_topology(void)
7126 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7127 * For now this just excludes isolated cpus, but could be used to
7128 * exclude other special cases in the future.
7130 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7134 arch_update_cpu_topology();
7136 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7138 doms_cur = &fallback_doms;
7139 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7140 err = build_sched_domains(doms_cur);
7141 register_sched_domain_sysctl();
7146 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7149 free_sched_groups(cpu_map, tmpmask);
7153 * Detach sched domains from a group of cpus specified in cpu_map
7154 * These cpus will now be attached to the NULL domain
7156 static void detach_destroy_domains(const cpumask_t *cpu_map)
7161 unregister_sched_domain_sysctl();
7163 for_each_cpu_mask(i, *cpu_map)
7164 cpu_attach_domain(NULL, &def_root_domain, i);
7165 synchronize_sched();
7166 arch_destroy_sched_domains(cpu_map, &tmpmask);
7170 * Partition sched domains as specified by the 'ndoms_new'
7171 * cpumasks in the array doms_new[] of cpumasks. This compares
7172 * doms_new[] to the current sched domain partitioning, doms_cur[].
7173 * It destroys each deleted domain and builds each new domain.
7175 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7176 * The masks don't intersect (don't overlap.) We should setup one
7177 * sched domain for each mask. CPUs not in any of the cpumasks will
7178 * not be load balanced. If the same cpumask appears both in the
7179 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7182 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7183 * ownership of it and will kfree it when done with it. If the caller
7184 * failed the kmalloc call, then it can pass in doms_new == NULL,
7185 * and partition_sched_domains() will fallback to the single partition
7188 * Call with hotplug lock held
7190 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7196 /* always unregister in case we don't destroy any domains */
7197 unregister_sched_domain_sysctl();
7199 if (doms_new == NULL) {
7201 doms_new = &fallback_doms;
7202 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7205 /* Destroy deleted domains */
7206 for (i = 0; i < ndoms_cur; i++) {
7207 for (j = 0; j < ndoms_new; j++) {
7208 if (cpus_equal(doms_cur[i], doms_new[j]))
7211 /* no match - a current sched domain not in new doms_new[] */
7212 detach_destroy_domains(doms_cur + i);
7217 /* Build new domains */
7218 for (i = 0; i < ndoms_new; i++) {
7219 for (j = 0; j < ndoms_cur; j++) {
7220 if (cpus_equal(doms_new[i], doms_cur[j]))
7223 /* no match - add a new doms_new */
7224 build_sched_domains(doms_new + i);
7229 /* Remember the new sched domains */
7230 if (doms_cur != &fallback_doms)
7232 doms_cur = doms_new;
7233 ndoms_cur = ndoms_new;
7235 register_sched_domain_sysctl();
7240 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7241 int arch_reinit_sched_domains(void)
7246 detach_destroy_domains(&cpu_online_map);
7247 err = arch_init_sched_domains(&cpu_online_map);
7253 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7257 if (buf[0] != '0' && buf[0] != '1')
7261 sched_smt_power_savings = (buf[0] == '1');
7263 sched_mc_power_savings = (buf[0] == '1');
7265 ret = arch_reinit_sched_domains();
7267 return ret ? ret : count;
7270 #ifdef CONFIG_SCHED_MC
7271 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7273 return sprintf(page, "%u\n", sched_mc_power_savings);
7275 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7276 const char *buf, size_t count)
7278 return sched_power_savings_store(buf, count, 0);
7280 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7281 sched_mc_power_savings_store);
7284 #ifdef CONFIG_SCHED_SMT
7285 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7287 return sprintf(page, "%u\n", sched_smt_power_savings);
7289 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7290 const char *buf, size_t count)
7292 return sched_power_savings_store(buf, count, 1);
7294 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7295 sched_smt_power_savings_store);
7298 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7302 #ifdef CONFIG_SCHED_SMT
7304 err = sysfs_create_file(&cls->kset.kobj,
7305 &attr_sched_smt_power_savings.attr);
7307 #ifdef CONFIG_SCHED_MC
7308 if (!err && mc_capable())
7309 err = sysfs_create_file(&cls->kset.kobj,
7310 &attr_sched_mc_power_savings.attr);
7317 * Force a reinitialization of the sched domains hierarchy. The domains
7318 * and groups cannot be updated in place without racing with the balancing
7319 * code, so we temporarily attach all running cpus to the NULL domain
7320 * which will prevent rebalancing while the sched domains are recalculated.
7322 static int update_sched_domains(struct notifier_block *nfb,
7323 unsigned long action, void *hcpu)
7326 case CPU_UP_PREPARE:
7327 case CPU_UP_PREPARE_FROZEN:
7328 case CPU_DOWN_PREPARE:
7329 case CPU_DOWN_PREPARE_FROZEN:
7330 detach_destroy_domains(&cpu_online_map);
7333 case CPU_UP_CANCELED:
7334 case CPU_UP_CANCELED_FROZEN:
7335 case CPU_DOWN_FAILED:
7336 case CPU_DOWN_FAILED_FROZEN:
7338 case CPU_ONLINE_FROZEN:
7340 case CPU_DEAD_FROZEN:
7342 * Fall through and re-initialise the domains.
7349 /* The hotplug lock is already held by cpu_up/cpu_down */
7350 arch_init_sched_domains(&cpu_online_map);
7355 void __init sched_init_smp(void)
7357 cpumask_t non_isolated_cpus;
7359 #if defined(CONFIG_NUMA)
7360 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7362 BUG_ON(sched_group_nodes_bycpu == NULL);
7365 arch_init_sched_domains(&cpu_online_map);
7366 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7367 if (cpus_empty(non_isolated_cpus))
7368 cpu_set(smp_processor_id(), non_isolated_cpus);
7370 /* XXX: Theoretical race here - CPU may be hotplugged now */
7371 hotcpu_notifier(update_sched_domains, 0);
7373 /* Move init over to a non-isolated CPU */
7374 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7376 sched_init_granularity();
7379 void __init sched_init_smp(void)
7381 #if defined(CONFIG_NUMA)
7382 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7384 BUG_ON(sched_group_nodes_bycpu == NULL);
7386 sched_init_granularity();
7388 #endif /* CONFIG_SMP */
7390 int in_sched_functions(unsigned long addr)
7392 return in_lock_functions(addr) ||
7393 (addr >= (unsigned long)__sched_text_start
7394 && addr < (unsigned long)__sched_text_end);
7397 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7399 cfs_rq->tasks_timeline = RB_ROOT;
7400 #ifdef CONFIG_FAIR_GROUP_SCHED
7403 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7406 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7408 struct rt_prio_array *array;
7411 array = &rt_rq->active;
7412 for (i = 0; i < MAX_RT_PRIO; i++) {
7413 INIT_LIST_HEAD(array->queue + i);
7414 __clear_bit(i, array->bitmap);
7416 /* delimiter for bitsearch: */
7417 __set_bit(MAX_RT_PRIO, array->bitmap);
7419 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7420 rt_rq->highest_prio = MAX_RT_PRIO;
7423 rt_rq->rt_nr_migratory = 0;
7424 rt_rq->overloaded = 0;
7428 rt_rq->rt_throttled = 0;
7429 rt_rq->rt_runtime = 0;
7430 spin_lock_init(&rt_rq->rt_runtime_lock);
7432 #ifdef CONFIG_RT_GROUP_SCHED
7433 rt_rq->rt_nr_boosted = 0;
7438 #ifdef CONFIG_FAIR_GROUP_SCHED
7439 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7440 struct cfs_rq *cfs_rq, struct sched_entity *se,
7443 tg->cfs_rq[cpu] = cfs_rq;
7444 init_cfs_rq(cfs_rq, rq);
7447 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7450 se->cfs_rq = &rq->cfs;
7452 se->load.weight = tg->shares;
7453 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7458 #ifdef CONFIG_RT_GROUP_SCHED
7459 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7460 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7463 tg->rt_rq[cpu] = rt_rq;
7464 init_rt_rq(rt_rq, rq);
7466 rt_rq->rt_se = rt_se;
7467 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7469 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7471 tg->rt_se[cpu] = rt_se;
7472 rt_se->rt_rq = &rq->rt;
7473 rt_se->my_q = rt_rq;
7474 rt_se->parent = NULL;
7475 INIT_LIST_HEAD(&rt_se->run_list);
7479 void __init sched_init(void)
7481 int highest_cpu = 0;
7483 unsigned long alloc_size = 0, ptr;
7485 #ifdef CONFIG_FAIR_GROUP_SCHED
7486 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7488 #ifdef CONFIG_RT_GROUP_SCHED
7489 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7492 * As sched_init() is called before page_alloc is setup,
7493 * we use alloc_bootmem().
7496 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7498 #ifdef CONFIG_FAIR_GROUP_SCHED
7499 init_task_group.se = (struct sched_entity **)ptr;
7500 ptr += nr_cpu_ids * sizeof(void **);
7502 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7503 ptr += nr_cpu_ids * sizeof(void **);
7505 #ifdef CONFIG_RT_GROUP_SCHED
7506 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7507 ptr += nr_cpu_ids * sizeof(void **);
7509 init_task_group.rt_rq = (struct rt_rq **)ptr;
7514 init_defrootdomain();
7517 init_rt_bandwidth(&def_rt_bandwidth,
7518 global_rt_period(), global_rt_runtime());
7520 #ifdef CONFIG_RT_GROUP_SCHED
7521 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7522 global_rt_period(), global_rt_runtime());
7525 #ifdef CONFIG_GROUP_SCHED
7526 list_add(&init_task_group.list, &task_groups);
7529 for_each_possible_cpu(i) {
7533 spin_lock_init(&rq->lock);
7534 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7537 update_last_tick_seen(rq);
7538 init_cfs_rq(&rq->cfs, rq);
7539 init_rt_rq(&rq->rt, rq);
7540 #ifdef CONFIG_FAIR_GROUP_SCHED
7541 init_task_group.shares = init_task_group_load;
7542 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7543 init_tg_cfs_entry(rq, &init_task_group,
7544 &per_cpu(init_cfs_rq, i),
7545 &per_cpu(init_sched_entity, i), i, 1);
7548 #ifdef CONFIG_RT_GROUP_SCHED
7549 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7550 init_tg_rt_entry(rq, &init_task_group,
7551 &per_cpu(init_rt_rq, i),
7552 &per_cpu(init_sched_rt_entity, i), i, 1);
7554 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7557 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7558 rq->cpu_load[j] = 0;
7562 rq->active_balance = 0;
7563 rq->next_balance = jiffies;
7566 rq->migration_thread = NULL;
7567 INIT_LIST_HEAD(&rq->migration_queue);
7568 rq_attach_root(rq, &def_root_domain);
7571 atomic_set(&rq->nr_iowait, 0);
7575 set_load_weight(&init_task);
7577 #ifdef CONFIG_PREEMPT_NOTIFIERS
7578 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7582 nr_cpu_ids = highest_cpu + 1;
7583 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7586 #ifdef CONFIG_RT_MUTEXES
7587 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7591 * The boot idle thread does lazy MMU switching as well:
7593 atomic_inc(&init_mm.mm_count);
7594 enter_lazy_tlb(&init_mm, current);
7597 * Make us the idle thread. Technically, schedule() should not be
7598 * called from this thread, however somewhere below it might be,
7599 * but because we are the idle thread, we just pick up running again
7600 * when this runqueue becomes "idle".
7602 init_idle(current, smp_processor_id());
7604 * During early bootup we pretend to be a normal task:
7606 current->sched_class = &fair_sched_class;
7608 scheduler_running = 1;
7611 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7612 void __might_sleep(char *file, int line)
7615 static unsigned long prev_jiffy; /* ratelimiting */
7617 if ((in_atomic() || irqs_disabled()) &&
7618 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7619 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7621 prev_jiffy = jiffies;
7622 printk(KERN_ERR "BUG: sleeping function called from invalid"
7623 " context at %s:%d\n", file, line);
7624 printk("in_atomic():%d, irqs_disabled():%d\n",
7625 in_atomic(), irqs_disabled());
7626 debug_show_held_locks(current);
7627 if (irqs_disabled())
7628 print_irqtrace_events(current);
7633 EXPORT_SYMBOL(__might_sleep);
7636 #ifdef CONFIG_MAGIC_SYSRQ
7637 static void normalize_task(struct rq *rq, struct task_struct *p)
7640 update_rq_clock(rq);
7641 on_rq = p->se.on_rq;
7643 deactivate_task(rq, p, 0);
7644 __setscheduler(rq, p, SCHED_NORMAL, 0);
7646 activate_task(rq, p, 0);
7647 resched_task(rq->curr);
7651 void normalize_rt_tasks(void)
7653 struct task_struct *g, *p;
7654 unsigned long flags;
7657 read_lock_irqsave(&tasklist_lock, flags);
7658 do_each_thread(g, p) {
7660 * Only normalize user tasks:
7665 p->se.exec_start = 0;
7666 #ifdef CONFIG_SCHEDSTATS
7667 p->se.wait_start = 0;
7668 p->se.sleep_start = 0;
7669 p->se.block_start = 0;
7671 task_rq(p)->clock = 0;
7675 * Renice negative nice level userspace
7678 if (TASK_NICE(p) < 0 && p->mm)
7679 set_user_nice(p, 0);
7683 spin_lock(&p->pi_lock);
7684 rq = __task_rq_lock(p);
7686 normalize_task(rq, p);
7688 __task_rq_unlock(rq);
7689 spin_unlock(&p->pi_lock);
7690 } while_each_thread(g, p);
7692 read_unlock_irqrestore(&tasklist_lock, flags);
7695 #endif /* CONFIG_MAGIC_SYSRQ */
7699 * These functions are only useful for the IA64 MCA handling.
7701 * They can only be called when the whole system has been
7702 * stopped - every CPU needs to be quiescent, and no scheduling
7703 * activity can take place. Using them for anything else would
7704 * be a serious bug, and as a result, they aren't even visible
7705 * under any other configuration.
7709 * curr_task - return the current task for a given cpu.
7710 * @cpu: the processor in question.
7712 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7714 struct task_struct *curr_task(int cpu)
7716 return cpu_curr(cpu);
7720 * set_curr_task - set the current task for a given cpu.
7721 * @cpu: the processor in question.
7722 * @p: the task pointer to set.
7724 * Description: This function must only be used when non-maskable interrupts
7725 * are serviced on a separate stack. It allows the architecture to switch the
7726 * notion of the current task on a cpu in a non-blocking manner. This function
7727 * must be called with all CPU's synchronized, and interrupts disabled, the
7728 * and caller must save the original value of the current task (see
7729 * curr_task() above) and restore that value before reenabling interrupts and
7730 * re-starting the system.
7732 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7734 void set_curr_task(int cpu, struct task_struct *p)
7741 #ifdef CONFIG_FAIR_GROUP_SCHED
7742 static void free_fair_sched_group(struct task_group *tg)
7746 for_each_possible_cpu(i) {
7748 kfree(tg->cfs_rq[i]);
7757 static int alloc_fair_sched_group(struct task_group *tg)
7759 struct cfs_rq *cfs_rq;
7760 struct sched_entity *se;
7764 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7767 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7771 tg->shares = NICE_0_LOAD;
7773 for_each_possible_cpu(i) {
7776 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7777 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7781 se = kmalloc_node(sizeof(struct sched_entity),
7782 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7786 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7795 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7797 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7798 &cpu_rq(cpu)->leaf_cfs_rq_list);
7801 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7803 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7806 static inline void free_fair_sched_group(struct task_group *tg)
7810 static inline int alloc_fair_sched_group(struct task_group *tg)
7815 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7819 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7824 #ifdef CONFIG_RT_GROUP_SCHED
7825 static void free_rt_sched_group(struct task_group *tg)
7829 destroy_rt_bandwidth(&tg->rt_bandwidth);
7831 for_each_possible_cpu(i) {
7833 kfree(tg->rt_rq[i]);
7835 kfree(tg->rt_se[i]);
7842 static int alloc_rt_sched_group(struct task_group *tg)
7844 struct rt_rq *rt_rq;
7845 struct sched_rt_entity *rt_se;
7849 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7852 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7856 init_rt_bandwidth(&tg->rt_bandwidth,
7857 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7859 for_each_possible_cpu(i) {
7862 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7863 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7867 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7868 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7872 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7881 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7883 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7884 &cpu_rq(cpu)->leaf_rt_rq_list);
7887 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7889 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7892 static inline void free_rt_sched_group(struct task_group *tg)
7896 static inline int alloc_rt_sched_group(struct task_group *tg)
7901 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7905 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7910 #ifdef CONFIG_GROUP_SCHED
7911 static void free_sched_group(struct task_group *tg)
7913 free_fair_sched_group(tg);
7914 free_rt_sched_group(tg);
7918 /* allocate runqueue etc for a new task group */
7919 struct task_group *sched_create_group(void)
7921 struct task_group *tg;
7922 unsigned long flags;
7925 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7927 return ERR_PTR(-ENOMEM);
7929 if (!alloc_fair_sched_group(tg))
7932 if (!alloc_rt_sched_group(tg))
7935 spin_lock_irqsave(&task_group_lock, flags);
7936 for_each_possible_cpu(i) {
7937 register_fair_sched_group(tg, i);
7938 register_rt_sched_group(tg, i);
7940 list_add_rcu(&tg->list, &task_groups);
7941 spin_unlock_irqrestore(&task_group_lock, flags);
7946 free_sched_group(tg);
7947 return ERR_PTR(-ENOMEM);
7950 /* rcu callback to free various structures associated with a task group */
7951 static void free_sched_group_rcu(struct rcu_head *rhp)
7953 /* now it should be safe to free those cfs_rqs */
7954 free_sched_group(container_of(rhp, struct task_group, rcu));
7957 /* Destroy runqueue etc associated with a task group */
7958 void sched_destroy_group(struct task_group *tg)
7960 unsigned long flags;
7963 spin_lock_irqsave(&task_group_lock, flags);
7964 for_each_possible_cpu(i) {
7965 unregister_fair_sched_group(tg, i);
7966 unregister_rt_sched_group(tg, i);
7968 list_del_rcu(&tg->list);
7969 spin_unlock_irqrestore(&task_group_lock, flags);
7971 /* wait for possible concurrent references to cfs_rqs complete */
7972 call_rcu(&tg->rcu, free_sched_group_rcu);
7975 /* change task's runqueue when it moves between groups.
7976 * The caller of this function should have put the task in its new group
7977 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7978 * reflect its new group.
7980 void sched_move_task(struct task_struct *tsk)
7983 unsigned long flags;
7986 rq = task_rq_lock(tsk, &flags);
7988 update_rq_clock(rq);
7990 running = task_current(rq, tsk);
7991 on_rq = tsk->se.on_rq;
7994 dequeue_task(rq, tsk, 0);
7995 if (unlikely(running))
7996 tsk->sched_class->put_prev_task(rq, tsk);
7998 set_task_rq(tsk, task_cpu(tsk));
8000 #ifdef CONFIG_FAIR_GROUP_SCHED
8001 if (tsk->sched_class->moved_group)
8002 tsk->sched_class->moved_group(tsk);
8005 if (unlikely(running))
8006 tsk->sched_class->set_curr_task(rq);
8008 enqueue_task(rq, tsk, 0);
8010 task_rq_unlock(rq, &flags);
8014 #ifdef CONFIG_FAIR_GROUP_SCHED
8015 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8017 struct cfs_rq *cfs_rq = se->cfs_rq;
8018 struct rq *rq = cfs_rq->rq;
8021 spin_lock_irq(&rq->lock);
8025 dequeue_entity(cfs_rq, se, 0);
8027 se->load.weight = shares;
8028 se->load.inv_weight = div64_64((1ULL<<32), shares);
8031 enqueue_entity(cfs_rq, se, 0);
8033 spin_unlock_irq(&rq->lock);
8036 static DEFINE_MUTEX(shares_mutex);
8038 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8041 unsigned long flags;
8044 * A weight of 0 or 1 can cause arithmetics problems.
8045 * (The default weight is 1024 - so there's no practical
8046 * limitation from this.)
8051 mutex_lock(&shares_mutex);
8052 if (tg->shares == shares)
8055 spin_lock_irqsave(&task_group_lock, flags);
8056 for_each_possible_cpu(i)
8057 unregister_fair_sched_group(tg, i);
8058 spin_unlock_irqrestore(&task_group_lock, flags);
8060 /* wait for any ongoing reference to this group to finish */
8061 synchronize_sched();
8064 * Now we are free to modify the group's share on each cpu
8065 * w/o tripping rebalance_share or load_balance_fair.
8067 tg->shares = shares;
8068 for_each_possible_cpu(i)
8069 set_se_shares(tg->se[i], shares);
8072 * Enable load balance activity on this group, by inserting it back on
8073 * each cpu's rq->leaf_cfs_rq_list.
8075 spin_lock_irqsave(&task_group_lock, flags);
8076 for_each_possible_cpu(i)
8077 register_fair_sched_group(tg, i);
8078 spin_unlock_irqrestore(&task_group_lock, flags);
8080 mutex_unlock(&shares_mutex);
8084 unsigned long sched_group_shares(struct task_group *tg)
8090 #ifdef CONFIG_RT_GROUP_SCHED
8092 * Ensure that the real time constraints are schedulable.
8094 static DEFINE_MUTEX(rt_constraints_mutex);
8096 static unsigned long to_ratio(u64 period, u64 runtime)
8098 if (runtime == RUNTIME_INF)
8101 return div64_64(runtime << 16, period);
8104 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8106 struct task_group *tgi;
8107 unsigned long total = 0;
8108 unsigned long global_ratio =
8109 to_ratio(global_rt_period(), global_rt_runtime());
8112 list_for_each_entry_rcu(tgi, &task_groups, list) {
8116 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8117 tgi->rt_bandwidth.rt_runtime);
8121 return total + to_ratio(period, runtime) < global_ratio;
8124 /* Must be called with tasklist_lock held */
8125 static inline int tg_has_rt_tasks(struct task_group *tg)
8127 struct task_struct *g, *p;
8128 do_each_thread(g, p) {
8129 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8131 } while_each_thread(g, p);
8135 static int tg_set_bandwidth(struct task_group *tg,
8136 u64 rt_period, u64 rt_runtime)
8140 mutex_lock(&rt_constraints_mutex);
8141 read_lock(&tasklist_lock);
8142 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8146 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8151 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8152 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8153 tg->rt_bandwidth.rt_runtime = rt_runtime;
8155 for_each_possible_cpu(i) {
8156 struct rt_rq *rt_rq = tg->rt_rq[i];
8158 spin_lock(&rt_rq->rt_runtime_lock);
8159 rt_rq->rt_runtime = rt_runtime;
8160 spin_unlock(&rt_rq->rt_runtime_lock);
8162 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8164 read_unlock(&tasklist_lock);
8165 mutex_unlock(&rt_constraints_mutex);
8170 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8172 u64 rt_runtime, rt_period;
8174 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8175 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8176 if (rt_runtime_us < 0)
8177 rt_runtime = RUNTIME_INF;
8179 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8182 long sched_group_rt_runtime(struct task_group *tg)
8186 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8189 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8190 do_div(rt_runtime_us, NSEC_PER_USEC);
8191 return rt_runtime_us;
8194 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8196 u64 rt_runtime, rt_period;
8198 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8199 rt_runtime = tg->rt_bandwidth.rt_runtime;
8201 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8204 long sched_group_rt_period(struct task_group *tg)
8208 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8209 do_div(rt_period_us, NSEC_PER_USEC);
8210 return rt_period_us;
8213 static int sched_rt_global_constraints(void)
8217 mutex_lock(&rt_constraints_mutex);
8218 if (!__rt_schedulable(NULL, 1, 0))
8220 mutex_unlock(&rt_constraints_mutex);
8225 static int sched_rt_global_constraints(void)
8227 unsigned long flags;
8230 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8231 for_each_possible_cpu(i) {
8232 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8234 spin_lock(&rt_rq->rt_runtime_lock);
8235 rt_rq->rt_runtime = global_rt_runtime();
8236 spin_unlock(&rt_rq->rt_runtime_lock);
8238 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8244 int sched_rt_handler(struct ctl_table *table, int write,
8245 struct file *filp, void __user *buffer, size_t *lenp,
8249 int old_period, old_runtime;
8250 static DEFINE_MUTEX(mutex);
8253 old_period = sysctl_sched_rt_period;
8254 old_runtime = sysctl_sched_rt_runtime;
8256 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8258 if (!ret && write) {
8259 ret = sched_rt_global_constraints();
8261 sysctl_sched_rt_period = old_period;
8262 sysctl_sched_rt_runtime = old_runtime;
8264 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8265 def_rt_bandwidth.rt_period =
8266 ns_to_ktime(global_rt_period());
8269 mutex_unlock(&mutex);
8274 #ifdef CONFIG_CGROUP_SCHED
8276 /* return corresponding task_group object of a cgroup */
8277 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8279 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8280 struct task_group, css);
8283 static struct cgroup_subsys_state *
8284 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8286 struct task_group *tg;
8288 if (!cgrp->parent) {
8289 /* This is early initialization for the top cgroup */
8290 init_task_group.css.cgroup = cgrp;
8291 return &init_task_group.css;
8294 /* we support only 1-level deep hierarchical scheduler atm */
8295 if (cgrp->parent->parent)
8296 return ERR_PTR(-EINVAL);
8298 tg = sched_create_group();
8300 return ERR_PTR(-ENOMEM);
8302 /* Bind the cgroup to task_group object we just created */
8303 tg->css.cgroup = cgrp;
8309 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8311 struct task_group *tg = cgroup_tg(cgrp);
8313 sched_destroy_group(tg);
8317 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8318 struct task_struct *tsk)
8320 #ifdef CONFIG_RT_GROUP_SCHED
8321 /* Don't accept realtime tasks when there is no way for them to run */
8322 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8325 /* We don't support RT-tasks being in separate groups */
8326 if (tsk->sched_class != &fair_sched_class)
8334 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8335 struct cgroup *old_cont, struct task_struct *tsk)
8337 sched_move_task(tsk);
8340 #ifdef CONFIG_FAIR_GROUP_SCHED
8341 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8344 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8347 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8349 struct task_group *tg = cgroup_tg(cgrp);
8351 return (u64) tg->shares;
8355 #ifdef CONFIG_RT_GROUP_SCHED
8356 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8358 const char __user *userbuf,
8359 size_t nbytes, loff_t *unused_ppos)
8368 if (nbytes >= sizeof(buffer))
8370 if (copy_from_user(buffer, userbuf, nbytes))
8373 buffer[nbytes] = 0; /* nul-terminate */
8375 /* strip newline if necessary */
8376 if (nbytes && (buffer[nbytes-1] == '\n'))
8377 buffer[nbytes-1] = 0;
8378 val = simple_strtoll(buffer, &end, 0);
8382 /* Pass to subsystem */
8383 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8389 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8391 char __user *buf, size_t nbytes,
8395 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8396 int len = sprintf(tmp, "%ld\n", val);
8398 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8401 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8404 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8407 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8409 return sched_group_rt_period(cgroup_tg(cgrp));
8413 static struct cftype cpu_files[] = {
8414 #ifdef CONFIG_FAIR_GROUP_SCHED
8417 .read_uint = cpu_shares_read_uint,
8418 .write_uint = cpu_shares_write_uint,
8421 #ifdef CONFIG_RT_GROUP_SCHED
8423 .name = "rt_runtime_us",
8424 .read = cpu_rt_runtime_read,
8425 .write = cpu_rt_runtime_write,
8428 .name = "rt_period_us",
8429 .read_uint = cpu_rt_period_read_uint,
8430 .write_uint = cpu_rt_period_write_uint,
8435 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8437 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8440 struct cgroup_subsys cpu_cgroup_subsys = {
8442 .create = cpu_cgroup_create,
8443 .destroy = cpu_cgroup_destroy,
8444 .can_attach = cpu_cgroup_can_attach,
8445 .attach = cpu_cgroup_attach,
8446 .populate = cpu_cgroup_populate,
8447 .subsys_id = cpu_cgroup_subsys_id,
8451 #endif /* CONFIG_CGROUP_SCHED */
8453 #ifdef CONFIG_CGROUP_CPUACCT
8456 * CPU accounting code for task groups.
8458 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8459 * (balbir@in.ibm.com).
8462 /* track cpu usage of a group of tasks */
8464 struct cgroup_subsys_state css;
8465 /* cpuusage holds pointer to a u64-type object on every cpu */
8469 struct cgroup_subsys cpuacct_subsys;
8471 /* return cpu accounting group corresponding to this container */
8472 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8474 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8475 struct cpuacct, css);
8478 /* return cpu accounting group to which this task belongs */
8479 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8481 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8482 struct cpuacct, css);
8485 /* create a new cpu accounting group */
8486 static struct cgroup_subsys_state *cpuacct_create(
8487 struct cgroup_subsys *ss, struct cgroup *cgrp)
8489 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8492 return ERR_PTR(-ENOMEM);
8494 ca->cpuusage = alloc_percpu(u64);
8495 if (!ca->cpuusage) {
8497 return ERR_PTR(-ENOMEM);
8503 /* destroy an existing cpu accounting group */
8505 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8507 struct cpuacct *ca = cgroup_ca(cgrp);
8509 free_percpu(ca->cpuusage);
8513 /* return total cpu usage (in nanoseconds) of a group */
8514 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8516 struct cpuacct *ca = cgroup_ca(cgrp);
8517 u64 totalcpuusage = 0;
8520 for_each_possible_cpu(i) {
8521 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8524 * Take rq->lock to make 64-bit addition safe on 32-bit
8527 spin_lock_irq(&cpu_rq(i)->lock);
8528 totalcpuusage += *cpuusage;
8529 spin_unlock_irq(&cpu_rq(i)->lock);
8532 return totalcpuusage;
8535 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8538 struct cpuacct *ca = cgroup_ca(cgrp);
8547 for_each_possible_cpu(i) {
8548 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8550 spin_lock_irq(&cpu_rq(i)->lock);
8552 spin_unlock_irq(&cpu_rq(i)->lock);
8558 static struct cftype files[] = {
8561 .read_uint = cpuusage_read,
8562 .write_uint = cpuusage_write,
8566 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8568 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8572 * charge this task's execution time to its accounting group.
8574 * called with rq->lock held.
8576 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8580 if (!cpuacct_subsys.active)
8585 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8587 *cpuusage += cputime;
8591 struct cgroup_subsys cpuacct_subsys = {
8593 .create = cpuacct_create,
8594 .destroy = cpuacct_destroy,
8595 .populate = cpuacct_populate,
8596 .subsys_id = cpuacct_subsys_id,
8598 #endif /* CONFIG_CGROUP_CPUACCT */