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)
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);
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, 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, 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,
3163 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3164 struct sched_group *group;
3165 unsigned long imbalance;
3167 cpumask_t cpus = CPU_MASK_ALL;
3168 unsigned long flags;
3171 * When power savings policy is enabled for the parent domain, idle
3172 * sibling can pick up load irrespective of busy siblings. In this case,
3173 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3174 * portraying it as CPU_NOT_IDLE.
3176 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3177 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3180 schedstat_inc(sd, lb_count[idle]);
3183 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3190 schedstat_inc(sd, lb_nobusyg[idle]);
3194 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3196 schedstat_inc(sd, lb_nobusyq[idle]);
3200 BUG_ON(busiest == this_rq);
3202 schedstat_add(sd, lb_imbalance[idle], imbalance);
3205 if (busiest->nr_running > 1) {
3207 * Attempt to move tasks. If find_busiest_group has found
3208 * an imbalance but busiest->nr_running <= 1, the group is
3209 * still unbalanced. ld_moved simply stays zero, so it is
3210 * correctly treated as an imbalance.
3212 local_irq_save(flags);
3213 double_rq_lock(this_rq, busiest);
3214 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3215 imbalance, sd, idle, &all_pinned);
3216 double_rq_unlock(this_rq, busiest);
3217 local_irq_restore(flags);
3220 * some other cpu did the load balance for us.
3222 if (ld_moved && this_cpu != smp_processor_id())
3223 resched_cpu(this_cpu);
3225 /* All tasks on this runqueue were pinned by CPU affinity */
3226 if (unlikely(all_pinned)) {
3227 cpu_clear(cpu_of(busiest), cpus);
3228 if (!cpus_empty(cpus))
3235 schedstat_inc(sd, lb_failed[idle]);
3236 sd->nr_balance_failed++;
3238 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3240 spin_lock_irqsave(&busiest->lock, flags);
3242 /* don't kick the migration_thread, if the curr
3243 * task on busiest cpu can't be moved to this_cpu
3245 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3246 spin_unlock_irqrestore(&busiest->lock, flags);
3248 goto out_one_pinned;
3251 if (!busiest->active_balance) {
3252 busiest->active_balance = 1;
3253 busiest->push_cpu = this_cpu;
3256 spin_unlock_irqrestore(&busiest->lock, flags);
3258 wake_up_process(busiest->migration_thread);
3261 * We've kicked active balancing, reset the failure
3264 sd->nr_balance_failed = sd->cache_nice_tries+1;
3267 sd->nr_balance_failed = 0;
3269 if (likely(!active_balance)) {
3270 /* We were unbalanced, so reset the balancing interval */
3271 sd->balance_interval = sd->min_interval;
3274 * If we've begun active balancing, start to back off. This
3275 * case may not be covered by the all_pinned logic if there
3276 * is only 1 task on the busy runqueue (because we don't call
3279 if (sd->balance_interval < sd->max_interval)
3280 sd->balance_interval *= 2;
3283 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3284 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3289 schedstat_inc(sd, lb_balanced[idle]);
3291 sd->nr_balance_failed = 0;
3294 /* tune up the balancing interval */
3295 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3296 (sd->balance_interval < sd->max_interval))
3297 sd->balance_interval *= 2;
3299 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3300 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3306 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3307 * tasks if there is an imbalance.
3309 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3310 * this_rq is locked.
3313 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3315 struct sched_group *group;
3316 struct rq *busiest = NULL;
3317 unsigned long imbalance;
3321 cpumask_t cpus = CPU_MASK_ALL;
3324 * When power savings policy is enabled for the parent domain, idle
3325 * sibling can pick up load irrespective of busy siblings. In this case,
3326 * let the state of idle sibling percolate up as IDLE, instead of
3327 * portraying it as CPU_NOT_IDLE.
3329 if (sd->flags & SD_SHARE_CPUPOWER &&
3330 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3333 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3335 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3336 &sd_idle, &cpus, NULL);
3338 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3342 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3345 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3349 BUG_ON(busiest == this_rq);
3351 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3354 if (busiest->nr_running > 1) {
3355 /* Attempt to move tasks */
3356 double_lock_balance(this_rq, busiest);
3357 /* this_rq->clock is already updated */
3358 update_rq_clock(busiest);
3359 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3360 imbalance, sd, CPU_NEWLY_IDLE,
3362 spin_unlock(&busiest->lock);
3364 if (unlikely(all_pinned)) {
3365 cpu_clear(cpu_of(busiest), cpus);
3366 if (!cpus_empty(cpus))
3372 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3373 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3374 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3377 sd->nr_balance_failed = 0;
3382 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3383 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3384 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3386 sd->nr_balance_failed = 0;
3392 * idle_balance is called by schedule() if this_cpu is about to become
3393 * idle. Attempts to pull tasks from other CPUs.
3395 static void idle_balance(int this_cpu, struct rq *this_rq)
3397 struct sched_domain *sd;
3398 int pulled_task = -1;
3399 unsigned long next_balance = jiffies + HZ;
3401 for_each_domain(this_cpu, sd) {
3402 unsigned long interval;
3404 if (!(sd->flags & SD_LOAD_BALANCE))
3407 if (sd->flags & SD_BALANCE_NEWIDLE)
3408 /* If we've pulled tasks over stop searching: */
3409 pulled_task = load_balance_newidle(this_cpu,
3412 interval = msecs_to_jiffies(sd->balance_interval);
3413 if (time_after(next_balance, sd->last_balance + interval))
3414 next_balance = sd->last_balance + interval;
3418 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3420 * We are going idle. next_balance may be set based on
3421 * a busy processor. So reset next_balance.
3423 this_rq->next_balance = next_balance;
3428 * active_load_balance is run by migration threads. It pushes running tasks
3429 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3430 * running on each physical CPU where possible, and avoids physical /
3431 * logical imbalances.
3433 * Called with busiest_rq locked.
3435 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3437 int target_cpu = busiest_rq->push_cpu;
3438 struct sched_domain *sd;
3439 struct rq *target_rq;
3441 /* Is there any task to move? */
3442 if (busiest_rq->nr_running <= 1)
3445 target_rq = cpu_rq(target_cpu);
3448 * This condition is "impossible", if it occurs
3449 * we need to fix it. Originally reported by
3450 * Bjorn Helgaas on a 128-cpu setup.
3452 BUG_ON(busiest_rq == target_rq);
3454 /* move a task from busiest_rq to target_rq */
3455 double_lock_balance(busiest_rq, target_rq);
3456 update_rq_clock(busiest_rq);
3457 update_rq_clock(target_rq);
3459 /* Search for an sd spanning us and the target CPU. */
3460 for_each_domain(target_cpu, sd) {
3461 if ((sd->flags & SD_LOAD_BALANCE) &&
3462 cpu_isset(busiest_cpu, sd->span))
3467 schedstat_inc(sd, alb_count);
3469 if (move_one_task(target_rq, target_cpu, busiest_rq,
3471 schedstat_inc(sd, alb_pushed);
3473 schedstat_inc(sd, alb_failed);
3475 spin_unlock(&target_rq->lock);
3480 atomic_t load_balancer;
3482 } nohz ____cacheline_aligned = {
3483 .load_balancer = ATOMIC_INIT(-1),
3484 .cpu_mask = CPU_MASK_NONE,
3488 * This routine will try to nominate the ilb (idle load balancing)
3489 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3490 * load balancing on behalf of all those cpus. If all the cpus in the system
3491 * go into this tickless mode, then there will be no ilb owner (as there is
3492 * no need for one) and all the cpus will sleep till the next wakeup event
3495 * For the ilb owner, tick is not stopped. And this tick will be used
3496 * for idle load balancing. ilb owner will still be part of
3499 * While stopping the tick, this cpu will become the ilb owner if there
3500 * is no other owner. And will be the owner till that cpu becomes busy
3501 * or if all cpus in the system stop their ticks at which point
3502 * there is no need for ilb owner.
3504 * When the ilb owner becomes busy, it nominates another owner, during the
3505 * next busy scheduler_tick()
3507 int select_nohz_load_balancer(int stop_tick)
3509 int cpu = smp_processor_id();
3512 cpu_set(cpu, nohz.cpu_mask);
3513 cpu_rq(cpu)->in_nohz_recently = 1;
3516 * If we are going offline and still the leader, give up!
3518 if (cpu_is_offline(cpu) &&
3519 atomic_read(&nohz.load_balancer) == cpu) {
3520 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3525 /* time for ilb owner also to sleep */
3526 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3527 if (atomic_read(&nohz.load_balancer) == cpu)
3528 atomic_set(&nohz.load_balancer, -1);
3532 if (atomic_read(&nohz.load_balancer) == -1) {
3533 /* make me the ilb owner */
3534 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3536 } else if (atomic_read(&nohz.load_balancer) == cpu)
3539 if (!cpu_isset(cpu, nohz.cpu_mask))
3542 cpu_clear(cpu, nohz.cpu_mask);
3544 if (atomic_read(&nohz.load_balancer) == cpu)
3545 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3552 static DEFINE_SPINLOCK(balancing);
3555 * It checks each scheduling domain to see if it is due to be balanced,
3556 * and initiates a balancing operation if so.
3558 * Balancing parameters are set up in arch_init_sched_domains.
3560 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3563 struct rq *rq = cpu_rq(cpu);
3564 unsigned long interval;
3565 struct sched_domain *sd;
3566 /* Earliest time when we have to do rebalance again */
3567 unsigned long next_balance = jiffies + 60*HZ;
3568 int update_next_balance = 0;
3570 for_each_domain(cpu, sd) {
3571 if (!(sd->flags & SD_LOAD_BALANCE))
3574 interval = sd->balance_interval;
3575 if (idle != CPU_IDLE)
3576 interval *= sd->busy_factor;
3578 /* scale ms to jiffies */
3579 interval = msecs_to_jiffies(interval);
3580 if (unlikely(!interval))
3582 if (interval > HZ*NR_CPUS/10)
3583 interval = HZ*NR_CPUS/10;
3586 if (sd->flags & SD_SERIALIZE) {
3587 if (!spin_trylock(&balancing))
3591 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3592 if (load_balance(cpu, rq, sd, idle, &balance)) {
3594 * We've pulled tasks over so either we're no
3595 * longer idle, or one of our SMT siblings is
3598 idle = CPU_NOT_IDLE;
3600 sd->last_balance = jiffies;
3602 if (sd->flags & SD_SERIALIZE)
3603 spin_unlock(&balancing);
3605 if (time_after(next_balance, sd->last_balance + interval)) {
3606 next_balance = sd->last_balance + interval;
3607 update_next_balance = 1;
3611 * Stop the load balance at this level. There is another
3612 * CPU in our sched group which is doing load balancing more
3620 * next_balance will be updated only when there is a need.
3621 * When the cpu is attached to null domain for ex, it will not be
3624 if (likely(update_next_balance))
3625 rq->next_balance = next_balance;
3629 * run_rebalance_domains is triggered when needed from the scheduler tick.
3630 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3631 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3633 static void run_rebalance_domains(struct softirq_action *h)
3635 int this_cpu = smp_processor_id();
3636 struct rq *this_rq = cpu_rq(this_cpu);
3637 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3638 CPU_IDLE : CPU_NOT_IDLE;
3640 rebalance_domains(this_cpu, idle);
3644 * If this cpu is the owner for idle load balancing, then do the
3645 * balancing on behalf of the other idle cpus whose ticks are
3648 if (this_rq->idle_at_tick &&
3649 atomic_read(&nohz.load_balancer) == this_cpu) {
3650 cpumask_t cpus = nohz.cpu_mask;
3654 cpu_clear(this_cpu, cpus);
3655 for_each_cpu_mask(balance_cpu, cpus) {
3657 * If this cpu gets work to do, stop the load balancing
3658 * work being done for other cpus. Next load
3659 * balancing owner will pick it up.
3664 rebalance_domains(balance_cpu, CPU_IDLE);
3666 rq = cpu_rq(balance_cpu);
3667 if (time_after(this_rq->next_balance, rq->next_balance))
3668 this_rq->next_balance = rq->next_balance;
3675 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3677 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3678 * idle load balancing owner or decide to stop the periodic load balancing,
3679 * if the whole system is idle.
3681 static inline void trigger_load_balance(struct rq *rq, int cpu)
3685 * If we were in the nohz mode recently and busy at the current
3686 * scheduler tick, then check if we need to nominate new idle
3689 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3690 rq->in_nohz_recently = 0;
3692 if (atomic_read(&nohz.load_balancer) == cpu) {
3693 cpu_clear(cpu, nohz.cpu_mask);
3694 atomic_set(&nohz.load_balancer, -1);
3697 if (atomic_read(&nohz.load_balancer) == -1) {
3699 * simple selection for now: Nominate the
3700 * first cpu in the nohz list to be the next
3703 * TBD: Traverse the sched domains and nominate
3704 * the nearest cpu in the nohz.cpu_mask.
3706 int ilb = first_cpu(nohz.cpu_mask);
3708 if (ilb < nr_cpu_ids)
3714 * If this cpu is idle and doing idle load balancing for all the
3715 * cpus with ticks stopped, is it time for that to stop?
3717 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3718 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3724 * If this cpu is idle and the idle load balancing is done by
3725 * someone else, then no need raise the SCHED_SOFTIRQ
3727 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3728 cpu_isset(cpu, nohz.cpu_mask))
3731 if (time_after_eq(jiffies, rq->next_balance))
3732 raise_softirq(SCHED_SOFTIRQ);
3735 #else /* CONFIG_SMP */
3738 * on UP we do not need to balance between CPUs:
3740 static inline void idle_balance(int cpu, struct rq *rq)
3746 DEFINE_PER_CPU(struct kernel_stat, kstat);
3748 EXPORT_PER_CPU_SYMBOL(kstat);
3751 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3752 * that have not yet been banked in case the task is currently running.
3754 unsigned long long task_sched_runtime(struct task_struct *p)
3756 unsigned long flags;
3760 rq = task_rq_lock(p, &flags);
3761 ns = p->se.sum_exec_runtime;
3762 if (task_current(rq, p)) {
3763 update_rq_clock(rq);
3764 delta_exec = rq->clock - p->se.exec_start;
3765 if ((s64)delta_exec > 0)
3768 task_rq_unlock(rq, &flags);
3774 * Account user cpu time to a process.
3775 * @p: the process that the cpu time gets accounted to
3776 * @cputime: the cpu time spent in user space since the last update
3778 void account_user_time(struct task_struct *p, cputime_t cputime)
3780 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3783 p->utime = cputime_add(p->utime, cputime);
3785 /* Add user time to cpustat. */
3786 tmp = cputime_to_cputime64(cputime);
3787 if (TASK_NICE(p) > 0)
3788 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3790 cpustat->user = cputime64_add(cpustat->user, tmp);
3794 * Account guest cpu time to a process.
3795 * @p: the process that the cpu time gets accounted to
3796 * @cputime: the cpu time spent in virtual machine since the last update
3798 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3801 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3803 tmp = cputime_to_cputime64(cputime);
3805 p->utime = cputime_add(p->utime, cputime);
3806 p->gtime = cputime_add(p->gtime, cputime);
3808 cpustat->user = cputime64_add(cpustat->user, tmp);
3809 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3813 * Account scaled user cpu time to a process.
3814 * @p: the process that the cpu time gets accounted to
3815 * @cputime: the cpu time spent in user space since the last update
3817 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3819 p->utimescaled = cputime_add(p->utimescaled, cputime);
3823 * Account system cpu time to a process.
3824 * @p: the process that the cpu time gets accounted to
3825 * @hardirq_offset: the offset to subtract from hardirq_count()
3826 * @cputime: the cpu time spent in kernel space since the last update
3828 void account_system_time(struct task_struct *p, int hardirq_offset,
3831 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3832 struct rq *rq = this_rq();
3835 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3836 return account_guest_time(p, cputime);
3838 p->stime = cputime_add(p->stime, cputime);
3840 /* Add system time to cpustat. */
3841 tmp = cputime_to_cputime64(cputime);
3842 if (hardirq_count() - hardirq_offset)
3843 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3844 else if (softirq_count())
3845 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3846 else if (p != rq->idle)
3847 cpustat->system = cputime64_add(cpustat->system, tmp);
3848 else if (atomic_read(&rq->nr_iowait) > 0)
3849 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3851 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3852 /* Account for system time used */
3853 acct_update_integrals(p);
3857 * Account scaled system cpu time to a process.
3858 * @p: the process that the cpu time gets accounted to
3859 * @hardirq_offset: the offset to subtract from hardirq_count()
3860 * @cputime: the cpu time spent in kernel space since the last update
3862 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3864 p->stimescaled = cputime_add(p->stimescaled, cputime);
3868 * Account for involuntary wait time.
3869 * @p: the process from which the cpu time has been stolen
3870 * @steal: the cpu time spent in involuntary wait
3872 void account_steal_time(struct task_struct *p, cputime_t steal)
3874 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3875 cputime64_t tmp = cputime_to_cputime64(steal);
3876 struct rq *rq = this_rq();
3878 if (p == rq->idle) {
3879 p->stime = cputime_add(p->stime, steal);
3880 if (atomic_read(&rq->nr_iowait) > 0)
3881 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3883 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3885 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3889 * This function gets called by the timer code, with HZ frequency.
3890 * We call it with interrupts disabled.
3892 * It also gets called by the fork code, when changing the parent's
3895 void scheduler_tick(void)
3897 int cpu = smp_processor_id();
3898 struct rq *rq = cpu_rq(cpu);
3899 struct task_struct *curr = rq->curr;
3900 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3902 spin_lock(&rq->lock);
3903 __update_rq_clock(rq);
3905 * Let rq->clock advance by at least TICK_NSEC:
3907 if (unlikely(rq->clock < next_tick)) {
3908 rq->clock = next_tick;
3909 rq->clock_underflows++;
3911 rq->tick_timestamp = rq->clock;
3912 update_last_tick_seen(rq);
3913 update_cpu_load(rq);
3914 curr->sched_class->task_tick(rq, curr, 0);
3915 spin_unlock(&rq->lock);
3918 rq->idle_at_tick = idle_cpu(cpu);
3919 trigger_load_balance(rq, cpu);
3923 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3925 void __kprobes add_preempt_count(int val)
3930 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3932 preempt_count() += val;
3934 * Spinlock count overflowing soon?
3936 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3939 EXPORT_SYMBOL(add_preempt_count);
3941 void __kprobes sub_preempt_count(int val)
3946 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3949 * Is the spinlock portion underflowing?
3951 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3952 !(preempt_count() & PREEMPT_MASK)))
3955 preempt_count() -= val;
3957 EXPORT_SYMBOL(sub_preempt_count);
3962 * Print scheduling while atomic bug:
3964 static noinline void __schedule_bug(struct task_struct *prev)
3966 struct pt_regs *regs = get_irq_regs();
3968 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3969 prev->comm, prev->pid, preempt_count());
3971 debug_show_held_locks(prev);
3972 if (irqs_disabled())
3973 print_irqtrace_events(prev);
3982 * Various schedule()-time debugging checks and statistics:
3984 static inline void schedule_debug(struct task_struct *prev)
3987 * Test if we are atomic. Since do_exit() needs to call into
3988 * schedule() atomically, we ignore that path for now.
3989 * Otherwise, whine if we are scheduling when we should not be.
3991 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3992 __schedule_bug(prev);
3994 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3996 schedstat_inc(this_rq(), sched_count);
3997 #ifdef CONFIG_SCHEDSTATS
3998 if (unlikely(prev->lock_depth >= 0)) {
3999 schedstat_inc(this_rq(), bkl_count);
4000 schedstat_inc(prev, sched_info.bkl_count);
4006 * Pick up the highest-prio task:
4008 static inline struct task_struct *
4009 pick_next_task(struct rq *rq, struct task_struct *prev)
4011 const struct sched_class *class;
4012 struct task_struct *p;
4015 * Optimization: we know that if all tasks are in
4016 * the fair class we can call that function directly:
4018 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4019 p = fair_sched_class.pick_next_task(rq);
4024 class = sched_class_highest;
4026 p = class->pick_next_task(rq);
4030 * Will never be NULL as the idle class always
4031 * returns a non-NULL p:
4033 class = class->next;
4038 * schedule() is the main scheduler function.
4040 asmlinkage void __sched schedule(void)
4042 struct task_struct *prev, *next;
4043 unsigned long *switch_count;
4049 cpu = smp_processor_id();
4053 switch_count = &prev->nivcsw;
4055 release_kernel_lock(prev);
4056 need_resched_nonpreemptible:
4058 schedule_debug(prev);
4063 * Do the rq-clock update outside the rq lock:
4065 local_irq_disable();
4066 __update_rq_clock(rq);
4067 spin_lock(&rq->lock);
4068 clear_tsk_need_resched(prev);
4070 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4071 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4072 signal_pending(prev))) {
4073 prev->state = TASK_RUNNING;
4075 deactivate_task(rq, prev, 1);
4077 switch_count = &prev->nvcsw;
4081 if (prev->sched_class->pre_schedule)
4082 prev->sched_class->pre_schedule(rq, prev);
4085 if (unlikely(!rq->nr_running))
4086 idle_balance(cpu, rq);
4088 prev->sched_class->put_prev_task(rq, prev);
4089 next = pick_next_task(rq, prev);
4091 sched_info_switch(prev, next);
4093 if (likely(prev != next)) {
4098 context_switch(rq, prev, next); /* unlocks the rq */
4100 * the context switch might have flipped the stack from under
4101 * us, hence refresh the local variables.
4103 cpu = smp_processor_id();
4106 spin_unlock_irq(&rq->lock);
4110 if (unlikely(reacquire_kernel_lock(current) < 0))
4111 goto need_resched_nonpreemptible;
4113 preempt_enable_no_resched();
4114 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4117 EXPORT_SYMBOL(schedule);
4119 #ifdef CONFIG_PREEMPT
4121 * this is the entry point to schedule() from in-kernel preemption
4122 * off of preempt_enable. Kernel preemptions off return from interrupt
4123 * occur there and call schedule directly.
4125 asmlinkage void __sched preempt_schedule(void)
4127 struct thread_info *ti = current_thread_info();
4128 struct task_struct *task = current;
4129 int saved_lock_depth;
4132 * If there is a non-zero preempt_count or interrupts are disabled,
4133 * we do not want to preempt the current task. Just return..
4135 if (likely(ti->preempt_count || irqs_disabled()))
4139 add_preempt_count(PREEMPT_ACTIVE);
4142 * We keep the big kernel semaphore locked, but we
4143 * clear ->lock_depth so that schedule() doesnt
4144 * auto-release the semaphore:
4146 saved_lock_depth = task->lock_depth;
4147 task->lock_depth = -1;
4149 task->lock_depth = saved_lock_depth;
4150 sub_preempt_count(PREEMPT_ACTIVE);
4153 * Check again in case we missed a preemption opportunity
4154 * between schedule and now.
4157 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4159 EXPORT_SYMBOL(preempt_schedule);
4162 * this is the entry point to schedule() from kernel preemption
4163 * off of irq context.
4164 * Note, that this is called and return with irqs disabled. This will
4165 * protect us against recursive calling from irq.
4167 asmlinkage void __sched preempt_schedule_irq(void)
4169 struct thread_info *ti = current_thread_info();
4170 struct task_struct *task = current;
4171 int saved_lock_depth;
4173 /* Catch callers which need to be fixed */
4174 BUG_ON(ti->preempt_count || !irqs_disabled());
4177 add_preempt_count(PREEMPT_ACTIVE);
4180 * We keep the big kernel semaphore locked, but we
4181 * clear ->lock_depth so that schedule() doesnt
4182 * auto-release the semaphore:
4184 saved_lock_depth = task->lock_depth;
4185 task->lock_depth = -1;
4188 local_irq_disable();
4189 task->lock_depth = saved_lock_depth;
4190 sub_preempt_count(PREEMPT_ACTIVE);
4193 * Check again in case we missed a preemption opportunity
4194 * between schedule and now.
4197 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4200 #endif /* CONFIG_PREEMPT */
4202 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4205 return try_to_wake_up(curr->private, mode, sync);
4207 EXPORT_SYMBOL(default_wake_function);
4210 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4211 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4212 * number) then we wake all the non-exclusive tasks and one exclusive task.
4214 * There are circumstances in which we can try to wake a task which has already
4215 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4216 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4218 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4219 int nr_exclusive, int sync, void *key)
4221 wait_queue_t *curr, *next;
4223 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4224 unsigned flags = curr->flags;
4226 if (curr->func(curr, mode, sync, key) &&
4227 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4233 * __wake_up - wake up threads blocked on a waitqueue.
4235 * @mode: which threads
4236 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4237 * @key: is directly passed to the wakeup function
4239 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4240 int nr_exclusive, void *key)
4242 unsigned long flags;
4244 spin_lock_irqsave(&q->lock, flags);
4245 __wake_up_common(q, mode, nr_exclusive, 0, key);
4246 spin_unlock_irqrestore(&q->lock, flags);
4248 EXPORT_SYMBOL(__wake_up);
4251 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4253 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4255 __wake_up_common(q, mode, 1, 0, NULL);
4259 * __wake_up_sync - wake up threads blocked on a waitqueue.
4261 * @mode: which threads
4262 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4264 * The sync wakeup differs that the waker knows that it will schedule
4265 * away soon, so while the target thread will be woken up, it will not
4266 * be migrated to another CPU - ie. the two threads are 'synchronized'
4267 * with each other. This can prevent needless bouncing between CPUs.
4269 * On UP it can prevent extra preemption.
4272 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4274 unsigned long flags;
4280 if (unlikely(!nr_exclusive))
4283 spin_lock_irqsave(&q->lock, flags);
4284 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4285 spin_unlock_irqrestore(&q->lock, flags);
4287 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4289 void complete(struct completion *x)
4291 unsigned long flags;
4293 spin_lock_irqsave(&x->wait.lock, flags);
4295 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4296 spin_unlock_irqrestore(&x->wait.lock, flags);
4298 EXPORT_SYMBOL(complete);
4300 void complete_all(struct completion *x)
4302 unsigned long flags;
4304 spin_lock_irqsave(&x->wait.lock, flags);
4305 x->done += UINT_MAX/2;
4306 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4307 spin_unlock_irqrestore(&x->wait.lock, flags);
4309 EXPORT_SYMBOL(complete_all);
4311 static inline long __sched
4312 do_wait_for_common(struct completion *x, long timeout, int state)
4315 DECLARE_WAITQUEUE(wait, current);
4317 wait.flags |= WQ_FLAG_EXCLUSIVE;
4318 __add_wait_queue_tail(&x->wait, &wait);
4320 if ((state == TASK_INTERRUPTIBLE &&
4321 signal_pending(current)) ||
4322 (state == TASK_KILLABLE &&
4323 fatal_signal_pending(current))) {
4324 __remove_wait_queue(&x->wait, &wait);
4325 return -ERESTARTSYS;
4327 __set_current_state(state);
4328 spin_unlock_irq(&x->wait.lock);
4329 timeout = schedule_timeout(timeout);
4330 spin_lock_irq(&x->wait.lock);
4332 __remove_wait_queue(&x->wait, &wait);
4336 __remove_wait_queue(&x->wait, &wait);
4343 wait_for_common(struct completion *x, long timeout, int state)
4347 spin_lock_irq(&x->wait.lock);
4348 timeout = do_wait_for_common(x, timeout, state);
4349 spin_unlock_irq(&x->wait.lock);
4353 void __sched wait_for_completion(struct completion *x)
4355 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4357 EXPORT_SYMBOL(wait_for_completion);
4359 unsigned long __sched
4360 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4362 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4364 EXPORT_SYMBOL(wait_for_completion_timeout);
4366 int __sched wait_for_completion_interruptible(struct completion *x)
4368 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4369 if (t == -ERESTARTSYS)
4373 EXPORT_SYMBOL(wait_for_completion_interruptible);
4375 unsigned long __sched
4376 wait_for_completion_interruptible_timeout(struct completion *x,
4377 unsigned long timeout)
4379 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4381 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4383 int __sched wait_for_completion_killable(struct completion *x)
4385 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4386 if (t == -ERESTARTSYS)
4390 EXPORT_SYMBOL(wait_for_completion_killable);
4393 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4395 unsigned long flags;
4398 init_waitqueue_entry(&wait, current);
4400 __set_current_state(state);
4402 spin_lock_irqsave(&q->lock, flags);
4403 __add_wait_queue(q, &wait);
4404 spin_unlock(&q->lock);
4405 timeout = schedule_timeout(timeout);
4406 spin_lock_irq(&q->lock);
4407 __remove_wait_queue(q, &wait);
4408 spin_unlock_irqrestore(&q->lock, flags);
4413 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4415 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4417 EXPORT_SYMBOL(interruptible_sleep_on);
4420 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4422 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4424 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4426 void __sched sleep_on(wait_queue_head_t *q)
4428 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4430 EXPORT_SYMBOL(sleep_on);
4432 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4434 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4436 EXPORT_SYMBOL(sleep_on_timeout);
4438 #ifdef CONFIG_RT_MUTEXES
4441 * rt_mutex_setprio - set the current priority of a task
4443 * @prio: prio value (kernel-internal form)
4445 * This function changes the 'effective' priority of a task. It does
4446 * not touch ->normal_prio like __setscheduler().
4448 * Used by the rt_mutex code to implement priority inheritance logic.
4450 void rt_mutex_setprio(struct task_struct *p, int prio)
4452 unsigned long flags;
4453 int oldprio, on_rq, running;
4455 const struct sched_class *prev_class = p->sched_class;
4457 BUG_ON(prio < 0 || prio > MAX_PRIO);
4459 rq = task_rq_lock(p, &flags);
4460 update_rq_clock(rq);
4463 on_rq = p->se.on_rq;
4464 running = task_current(rq, p);
4466 dequeue_task(rq, p, 0);
4468 p->sched_class->put_prev_task(rq, p);
4471 p->sched_class = &rt_sched_class;
4473 p->sched_class = &fair_sched_class;
4478 p->sched_class->set_curr_task(rq);
4480 enqueue_task(rq, p, 0);
4482 check_class_changed(rq, p, prev_class, oldprio, running);
4484 task_rq_unlock(rq, &flags);
4489 void set_user_nice(struct task_struct *p, long nice)
4491 int old_prio, delta, on_rq;
4492 unsigned long flags;
4495 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4498 * We have to be careful, if called from sys_setpriority(),
4499 * the task might be in the middle of scheduling on another CPU.
4501 rq = task_rq_lock(p, &flags);
4502 update_rq_clock(rq);
4504 * The RT priorities are set via sched_setscheduler(), but we still
4505 * allow the 'normal' nice value to be set - but as expected
4506 * it wont have any effect on scheduling until the task is
4507 * SCHED_FIFO/SCHED_RR:
4509 if (task_has_rt_policy(p)) {
4510 p->static_prio = NICE_TO_PRIO(nice);
4513 on_rq = p->se.on_rq;
4515 dequeue_task(rq, p, 0);
4519 p->static_prio = NICE_TO_PRIO(nice);
4522 p->prio = effective_prio(p);
4523 delta = p->prio - old_prio;
4526 enqueue_task(rq, p, 0);
4529 * If the task increased its priority or is running and
4530 * lowered its priority, then reschedule its CPU:
4532 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4533 resched_task(rq->curr);
4536 task_rq_unlock(rq, &flags);
4538 EXPORT_SYMBOL(set_user_nice);
4541 * can_nice - check if a task can reduce its nice value
4545 int can_nice(const struct task_struct *p, const int nice)
4547 /* convert nice value [19,-20] to rlimit style value [1,40] */
4548 int nice_rlim = 20 - nice;
4550 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4551 capable(CAP_SYS_NICE));
4554 #ifdef __ARCH_WANT_SYS_NICE
4557 * sys_nice - change the priority of the current process.
4558 * @increment: priority increment
4560 * sys_setpriority is a more generic, but much slower function that
4561 * does similar things.
4563 asmlinkage long sys_nice(int increment)
4568 * Setpriority might change our priority at the same moment.
4569 * We don't have to worry. Conceptually one call occurs first
4570 * and we have a single winner.
4572 if (increment < -40)
4577 nice = PRIO_TO_NICE(current->static_prio) + increment;
4583 if (increment < 0 && !can_nice(current, nice))
4586 retval = security_task_setnice(current, nice);
4590 set_user_nice(current, nice);
4597 * task_prio - return the priority value of a given task.
4598 * @p: the task in question.
4600 * This is the priority value as seen by users in /proc.
4601 * RT tasks are offset by -200. Normal tasks are centered
4602 * around 0, value goes from -16 to +15.
4604 int task_prio(const struct task_struct *p)
4606 return p->prio - MAX_RT_PRIO;
4610 * task_nice - return the nice value of a given task.
4611 * @p: the task in question.
4613 int task_nice(const struct task_struct *p)
4615 return TASK_NICE(p);
4617 EXPORT_SYMBOL(task_nice);
4620 * idle_cpu - is a given cpu idle currently?
4621 * @cpu: the processor in question.
4623 int idle_cpu(int cpu)
4625 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4629 * idle_task - return the idle task for a given cpu.
4630 * @cpu: the processor in question.
4632 struct task_struct *idle_task(int cpu)
4634 return cpu_rq(cpu)->idle;
4638 * find_process_by_pid - find a process with a matching PID value.
4639 * @pid: the pid in question.
4641 static struct task_struct *find_process_by_pid(pid_t pid)
4643 return pid ? find_task_by_vpid(pid) : current;
4646 /* Actually do priority change: must hold rq lock. */
4648 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4650 BUG_ON(p->se.on_rq);
4653 switch (p->policy) {
4657 p->sched_class = &fair_sched_class;
4661 p->sched_class = &rt_sched_class;
4665 p->rt_priority = prio;
4666 p->normal_prio = normal_prio(p);
4667 /* we are holding p->pi_lock already */
4668 p->prio = rt_mutex_getprio(p);
4673 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4674 * @p: the task in question.
4675 * @policy: new policy.
4676 * @param: structure containing the new RT priority.
4678 * NOTE that the task may be already dead.
4680 int sched_setscheduler(struct task_struct *p, int policy,
4681 struct sched_param *param)
4683 int retval, oldprio, oldpolicy = -1, on_rq, running;
4684 unsigned long flags;
4685 const struct sched_class *prev_class = p->sched_class;
4688 /* may grab non-irq protected spin_locks */
4689 BUG_ON(in_interrupt());
4691 /* double check policy once rq lock held */
4693 policy = oldpolicy = p->policy;
4694 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4695 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4696 policy != SCHED_IDLE)
4699 * Valid priorities for SCHED_FIFO and SCHED_RR are
4700 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4701 * SCHED_BATCH and SCHED_IDLE is 0.
4703 if (param->sched_priority < 0 ||
4704 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4705 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4707 if (rt_policy(policy) != (param->sched_priority != 0))
4711 * Allow unprivileged RT tasks to decrease priority:
4713 if (!capable(CAP_SYS_NICE)) {
4714 if (rt_policy(policy)) {
4715 unsigned long rlim_rtprio;
4717 if (!lock_task_sighand(p, &flags))
4719 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4720 unlock_task_sighand(p, &flags);
4722 /* can't set/change the rt policy */
4723 if (policy != p->policy && !rlim_rtprio)
4726 /* can't increase priority */
4727 if (param->sched_priority > p->rt_priority &&
4728 param->sched_priority > rlim_rtprio)
4732 * Like positive nice levels, dont allow tasks to
4733 * move out of SCHED_IDLE either:
4735 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4738 /* can't change other user's priorities */
4739 if ((current->euid != p->euid) &&
4740 (current->euid != p->uid))
4744 #ifdef CONFIG_RT_GROUP_SCHED
4746 * Do not allow realtime tasks into groups that have no runtime
4749 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4753 retval = security_task_setscheduler(p, policy, param);
4757 * make sure no PI-waiters arrive (or leave) while we are
4758 * changing the priority of the task:
4760 spin_lock_irqsave(&p->pi_lock, flags);
4762 * To be able to change p->policy safely, the apropriate
4763 * runqueue lock must be held.
4765 rq = __task_rq_lock(p);
4766 /* recheck policy now with rq lock held */
4767 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4768 policy = oldpolicy = -1;
4769 __task_rq_unlock(rq);
4770 spin_unlock_irqrestore(&p->pi_lock, flags);
4773 update_rq_clock(rq);
4774 on_rq = p->se.on_rq;
4775 running = task_current(rq, p);
4777 deactivate_task(rq, p, 0);
4779 p->sched_class->put_prev_task(rq, p);
4782 __setscheduler(rq, p, policy, param->sched_priority);
4785 p->sched_class->set_curr_task(rq);
4787 activate_task(rq, p, 0);
4789 check_class_changed(rq, p, prev_class, oldprio, running);
4791 __task_rq_unlock(rq);
4792 spin_unlock_irqrestore(&p->pi_lock, flags);
4794 rt_mutex_adjust_pi(p);
4798 EXPORT_SYMBOL_GPL(sched_setscheduler);
4801 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4803 struct sched_param lparam;
4804 struct task_struct *p;
4807 if (!param || pid < 0)
4809 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4814 p = find_process_by_pid(pid);
4816 retval = sched_setscheduler(p, policy, &lparam);
4823 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4824 * @pid: the pid in question.
4825 * @policy: new policy.
4826 * @param: structure containing the new RT priority.
4829 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4831 /* negative values for policy are not valid */
4835 return do_sched_setscheduler(pid, policy, param);
4839 * sys_sched_setparam - set/change the RT priority of a thread
4840 * @pid: the pid in question.
4841 * @param: structure containing the new RT priority.
4843 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4845 return do_sched_setscheduler(pid, -1, param);
4849 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4850 * @pid: the pid in question.
4852 asmlinkage long sys_sched_getscheduler(pid_t pid)
4854 struct task_struct *p;
4861 read_lock(&tasklist_lock);
4862 p = find_process_by_pid(pid);
4864 retval = security_task_getscheduler(p);
4868 read_unlock(&tasklist_lock);
4873 * sys_sched_getscheduler - get the RT priority of a thread
4874 * @pid: the pid in question.
4875 * @param: structure containing the RT priority.
4877 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4879 struct sched_param lp;
4880 struct task_struct *p;
4883 if (!param || pid < 0)
4886 read_lock(&tasklist_lock);
4887 p = find_process_by_pid(pid);
4892 retval = security_task_getscheduler(p);
4896 lp.sched_priority = p->rt_priority;
4897 read_unlock(&tasklist_lock);
4900 * This one might sleep, we cannot do it with a spinlock held ...
4902 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4907 read_unlock(&tasklist_lock);
4911 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4913 cpumask_t cpus_allowed;
4914 struct task_struct *p;
4918 read_lock(&tasklist_lock);
4920 p = find_process_by_pid(pid);
4922 read_unlock(&tasklist_lock);
4928 * It is not safe to call set_cpus_allowed with the
4929 * tasklist_lock held. We will bump the task_struct's
4930 * usage count and then drop tasklist_lock.
4933 read_unlock(&tasklist_lock);
4936 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4937 !capable(CAP_SYS_NICE))
4940 retval = security_task_setscheduler(p, 0, NULL);
4944 cpuset_cpus_allowed(p, &cpus_allowed);
4945 cpus_and(new_mask, new_mask, cpus_allowed);
4947 retval = set_cpus_allowed(p, new_mask);
4950 cpuset_cpus_allowed(p, &cpus_allowed);
4951 if (!cpus_subset(new_mask, cpus_allowed)) {
4953 * We must have raced with a concurrent cpuset
4954 * update. Just reset the cpus_allowed to the
4955 * cpuset's cpus_allowed
4957 new_mask = cpus_allowed;
4967 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4968 cpumask_t *new_mask)
4970 if (len < sizeof(cpumask_t)) {
4971 memset(new_mask, 0, sizeof(cpumask_t));
4972 } else if (len > sizeof(cpumask_t)) {
4973 len = sizeof(cpumask_t);
4975 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4979 * sys_sched_setaffinity - set the cpu affinity of a process
4980 * @pid: pid of the process
4981 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4982 * @user_mask_ptr: user-space pointer to the new cpu mask
4984 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4985 unsigned long __user *user_mask_ptr)
4990 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4994 return sched_setaffinity(pid, new_mask);
4998 * Represents all cpu's present in the system
4999 * In systems capable of hotplug, this map could dynamically grow
5000 * as new cpu's are detected in the system via any platform specific
5001 * method, such as ACPI for e.g.
5004 cpumask_t cpu_present_map __read_mostly;
5005 EXPORT_SYMBOL(cpu_present_map);
5008 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5009 EXPORT_SYMBOL(cpu_online_map);
5011 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5012 EXPORT_SYMBOL(cpu_possible_map);
5015 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5017 struct task_struct *p;
5021 read_lock(&tasklist_lock);
5024 p = find_process_by_pid(pid);
5028 retval = security_task_getscheduler(p);
5032 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5035 read_unlock(&tasklist_lock);
5042 * sys_sched_getaffinity - get the cpu affinity of a process
5043 * @pid: pid of the process
5044 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5045 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5047 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5048 unsigned long __user *user_mask_ptr)
5053 if (len < sizeof(cpumask_t))
5056 ret = sched_getaffinity(pid, &mask);
5060 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5063 return sizeof(cpumask_t);
5067 * sys_sched_yield - yield the current processor to other threads.
5069 * This function yields the current CPU to other tasks. If there are no
5070 * other threads running on this CPU then this function will return.
5072 asmlinkage long sys_sched_yield(void)
5074 struct rq *rq = this_rq_lock();
5076 schedstat_inc(rq, yld_count);
5077 current->sched_class->yield_task(rq);
5080 * Since we are going to call schedule() anyway, there's
5081 * no need to preempt or enable interrupts:
5083 __release(rq->lock);
5084 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5085 _raw_spin_unlock(&rq->lock);
5086 preempt_enable_no_resched();
5093 static void __cond_resched(void)
5095 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5096 __might_sleep(__FILE__, __LINE__);
5099 * The BKS might be reacquired before we have dropped
5100 * PREEMPT_ACTIVE, which could trigger a second
5101 * cond_resched() call.
5104 add_preempt_count(PREEMPT_ACTIVE);
5106 sub_preempt_count(PREEMPT_ACTIVE);
5107 } while (need_resched());
5110 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5111 int __sched _cond_resched(void)
5113 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5114 system_state == SYSTEM_RUNNING) {
5120 EXPORT_SYMBOL(_cond_resched);
5124 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5125 * call schedule, and on return reacquire the lock.
5127 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5128 * operations here to prevent schedule() from being called twice (once via
5129 * spin_unlock(), once by hand).
5131 int cond_resched_lock(spinlock_t *lock)
5133 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5136 if (spin_needbreak(lock) || resched) {
5138 if (resched && need_resched())
5147 EXPORT_SYMBOL(cond_resched_lock);
5149 int __sched cond_resched_softirq(void)
5151 BUG_ON(!in_softirq());
5153 if (need_resched() && system_state == SYSTEM_RUNNING) {
5161 EXPORT_SYMBOL(cond_resched_softirq);
5164 * yield - yield the current processor to other threads.
5166 * This is a shortcut for kernel-space yielding - it marks the
5167 * thread runnable and calls sys_sched_yield().
5169 void __sched yield(void)
5171 set_current_state(TASK_RUNNING);
5174 EXPORT_SYMBOL(yield);
5177 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5178 * that process accounting knows that this is a task in IO wait state.
5180 * But don't do that if it is a deliberate, throttling IO wait (this task
5181 * has set its backing_dev_info: the queue against which it should throttle)
5183 void __sched io_schedule(void)
5185 struct rq *rq = &__raw_get_cpu_var(runqueues);
5187 delayacct_blkio_start();
5188 atomic_inc(&rq->nr_iowait);
5190 atomic_dec(&rq->nr_iowait);
5191 delayacct_blkio_end();
5193 EXPORT_SYMBOL(io_schedule);
5195 long __sched io_schedule_timeout(long timeout)
5197 struct rq *rq = &__raw_get_cpu_var(runqueues);
5200 delayacct_blkio_start();
5201 atomic_inc(&rq->nr_iowait);
5202 ret = schedule_timeout(timeout);
5203 atomic_dec(&rq->nr_iowait);
5204 delayacct_blkio_end();
5209 * sys_sched_get_priority_max - return maximum RT priority.
5210 * @policy: scheduling class.
5212 * this syscall returns the maximum rt_priority that can be used
5213 * by a given scheduling class.
5215 asmlinkage long sys_sched_get_priority_max(int policy)
5222 ret = MAX_USER_RT_PRIO-1;
5234 * sys_sched_get_priority_min - return minimum RT priority.
5235 * @policy: scheduling class.
5237 * this syscall returns the minimum rt_priority that can be used
5238 * by a given scheduling class.
5240 asmlinkage long sys_sched_get_priority_min(int policy)
5258 * sys_sched_rr_get_interval - return the default timeslice of a process.
5259 * @pid: pid of the process.
5260 * @interval: userspace pointer to the timeslice value.
5262 * this syscall writes the default timeslice value of a given process
5263 * into the user-space timespec buffer. A value of '0' means infinity.
5266 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5268 struct task_struct *p;
5269 unsigned int time_slice;
5277 read_lock(&tasklist_lock);
5278 p = find_process_by_pid(pid);
5282 retval = security_task_getscheduler(p);
5287 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5288 * tasks that are on an otherwise idle runqueue:
5291 if (p->policy == SCHED_RR) {
5292 time_slice = DEF_TIMESLICE;
5293 } else if (p->policy != SCHED_FIFO) {
5294 struct sched_entity *se = &p->se;
5295 unsigned long flags;
5298 rq = task_rq_lock(p, &flags);
5299 if (rq->cfs.load.weight)
5300 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5301 task_rq_unlock(rq, &flags);
5303 read_unlock(&tasklist_lock);
5304 jiffies_to_timespec(time_slice, &t);
5305 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5309 read_unlock(&tasklist_lock);
5313 static const char stat_nam[] = "RSDTtZX";
5315 void sched_show_task(struct task_struct *p)
5317 unsigned long free = 0;
5320 state = p->state ? __ffs(p->state) + 1 : 0;
5321 printk(KERN_INFO "%-13.13s %c", p->comm,
5322 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5323 #if BITS_PER_LONG == 32
5324 if (state == TASK_RUNNING)
5325 printk(KERN_CONT " running ");
5327 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5329 if (state == TASK_RUNNING)
5330 printk(KERN_CONT " running task ");
5332 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5334 #ifdef CONFIG_DEBUG_STACK_USAGE
5336 unsigned long *n = end_of_stack(p);
5339 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5342 printk(KERN_CONT "%5lu %5d %6d\n", free,
5343 task_pid_nr(p), task_pid_nr(p->real_parent));
5345 show_stack(p, NULL);
5348 void show_state_filter(unsigned long state_filter)
5350 struct task_struct *g, *p;
5352 #if BITS_PER_LONG == 32
5354 " task PC stack pid father\n");
5357 " task PC stack pid father\n");
5359 read_lock(&tasklist_lock);
5360 do_each_thread(g, p) {
5362 * reset the NMI-timeout, listing all files on a slow
5363 * console might take alot of time:
5365 touch_nmi_watchdog();
5366 if (!state_filter || (p->state & state_filter))
5368 } while_each_thread(g, p);
5370 touch_all_softlockup_watchdogs();
5372 #ifdef CONFIG_SCHED_DEBUG
5373 sysrq_sched_debug_show();
5375 read_unlock(&tasklist_lock);
5377 * Only show locks if all tasks are dumped:
5379 if (state_filter == -1)
5380 debug_show_all_locks();
5383 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5385 idle->sched_class = &idle_sched_class;
5389 * init_idle - set up an idle thread for a given CPU
5390 * @idle: task in question
5391 * @cpu: cpu the idle task belongs to
5393 * NOTE: this function does not set the idle thread's NEED_RESCHED
5394 * flag, to make booting more robust.
5396 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5398 struct rq *rq = cpu_rq(cpu);
5399 unsigned long flags;
5402 idle->se.exec_start = sched_clock();
5404 idle->prio = idle->normal_prio = MAX_PRIO;
5405 idle->cpus_allowed = cpumask_of_cpu(cpu);
5406 __set_task_cpu(idle, cpu);
5408 spin_lock_irqsave(&rq->lock, flags);
5409 rq->curr = rq->idle = idle;
5410 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5413 spin_unlock_irqrestore(&rq->lock, flags);
5415 /* Set the preempt count _outside_ the spinlocks! */
5416 task_thread_info(idle)->preempt_count = 0;
5419 * The idle tasks have their own, simple scheduling class:
5421 idle->sched_class = &idle_sched_class;
5425 * In a system that switches off the HZ timer nohz_cpu_mask
5426 * indicates which cpus entered this state. This is used
5427 * in the rcu update to wait only for active cpus. For system
5428 * which do not switch off the HZ timer nohz_cpu_mask should
5429 * always be CPU_MASK_NONE.
5431 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5434 * Increase the granularity value when there are more CPUs,
5435 * because with more CPUs the 'effective latency' as visible
5436 * to users decreases. But the relationship is not linear,
5437 * so pick a second-best guess by going with the log2 of the
5440 * This idea comes from the SD scheduler of Con Kolivas:
5442 static inline void sched_init_granularity(void)
5444 unsigned int factor = 1 + ilog2(num_online_cpus());
5445 const unsigned long limit = 200000000;
5447 sysctl_sched_min_granularity *= factor;
5448 if (sysctl_sched_min_granularity > limit)
5449 sysctl_sched_min_granularity = limit;
5451 sysctl_sched_latency *= factor;
5452 if (sysctl_sched_latency > limit)
5453 sysctl_sched_latency = limit;
5455 sysctl_sched_wakeup_granularity *= factor;
5460 * This is how migration works:
5462 * 1) we queue a struct migration_req structure in the source CPU's
5463 * runqueue and wake up that CPU's migration thread.
5464 * 2) we down() the locked semaphore => thread blocks.
5465 * 3) migration thread wakes up (implicitly it forces the migrated
5466 * thread off the CPU)
5467 * 4) it gets the migration request and checks whether the migrated
5468 * task is still in the wrong runqueue.
5469 * 5) if it's in the wrong runqueue then the migration thread removes
5470 * it and puts it into the right queue.
5471 * 6) migration thread up()s the semaphore.
5472 * 7) we wake up and the migration is done.
5476 * Change a given task's CPU affinity. Migrate the thread to a
5477 * proper CPU and schedule it away if the CPU it's executing on
5478 * is removed from the allowed bitmask.
5480 * NOTE: the caller must have a valid reference to the task, the
5481 * task must not exit() & deallocate itself prematurely. The
5482 * call is not atomic; no spinlocks may be held.
5484 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5486 struct migration_req req;
5487 unsigned long flags;
5491 rq = task_rq_lock(p, &flags);
5492 if (!cpus_intersects(new_mask, cpu_online_map)) {
5497 if (p->sched_class->set_cpus_allowed)
5498 p->sched_class->set_cpus_allowed(p, &new_mask);
5500 p->cpus_allowed = new_mask;
5501 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5504 /* Can the task run on the task's current CPU? If so, we're done */
5505 if (cpu_isset(task_cpu(p), new_mask))
5508 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5509 /* Need help from migration thread: drop lock and wait. */
5510 task_rq_unlock(rq, &flags);
5511 wake_up_process(rq->migration_thread);
5512 wait_for_completion(&req.done);
5513 tlb_migrate_finish(p->mm);
5517 task_rq_unlock(rq, &flags);
5521 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5524 * Move (not current) task off this cpu, onto dest cpu. We're doing
5525 * this because either it can't run here any more (set_cpus_allowed()
5526 * away from this CPU, or CPU going down), or because we're
5527 * attempting to rebalance this task on exec (sched_exec).
5529 * So we race with normal scheduler movements, but that's OK, as long
5530 * as the task is no longer on this CPU.
5532 * Returns non-zero if task was successfully migrated.
5534 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5536 struct rq *rq_dest, *rq_src;
5539 if (unlikely(cpu_is_offline(dest_cpu)))
5542 rq_src = cpu_rq(src_cpu);
5543 rq_dest = cpu_rq(dest_cpu);
5545 double_rq_lock(rq_src, rq_dest);
5546 /* Already moved. */
5547 if (task_cpu(p) != src_cpu)
5549 /* Affinity changed (again). */
5550 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5553 on_rq = p->se.on_rq;
5555 deactivate_task(rq_src, p, 0);
5557 set_task_cpu(p, dest_cpu);
5559 activate_task(rq_dest, p, 0);
5560 check_preempt_curr(rq_dest, p);
5564 double_rq_unlock(rq_src, rq_dest);
5569 * migration_thread - this is a highprio system thread that performs
5570 * thread migration by bumping thread off CPU then 'pushing' onto
5573 static int migration_thread(void *data)
5575 int cpu = (long)data;
5579 BUG_ON(rq->migration_thread != current);
5581 set_current_state(TASK_INTERRUPTIBLE);
5582 while (!kthread_should_stop()) {
5583 struct migration_req *req;
5584 struct list_head *head;
5586 spin_lock_irq(&rq->lock);
5588 if (cpu_is_offline(cpu)) {
5589 spin_unlock_irq(&rq->lock);
5593 if (rq->active_balance) {
5594 active_load_balance(rq, cpu);
5595 rq->active_balance = 0;
5598 head = &rq->migration_queue;
5600 if (list_empty(head)) {
5601 spin_unlock_irq(&rq->lock);
5603 set_current_state(TASK_INTERRUPTIBLE);
5606 req = list_entry(head->next, struct migration_req, list);
5607 list_del_init(head->next);
5609 spin_unlock(&rq->lock);
5610 __migrate_task(req->task, cpu, req->dest_cpu);
5613 complete(&req->done);
5615 __set_current_state(TASK_RUNNING);
5619 /* Wait for kthread_stop */
5620 set_current_state(TASK_INTERRUPTIBLE);
5621 while (!kthread_should_stop()) {
5623 set_current_state(TASK_INTERRUPTIBLE);
5625 __set_current_state(TASK_RUNNING);
5629 #ifdef CONFIG_HOTPLUG_CPU
5631 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5635 local_irq_disable();
5636 ret = __migrate_task(p, src_cpu, dest_cpu);
5642 * Figure out where task on dead CPU should go, use force if necessary.
5643 * NOTE: interrupts should be disabled by the caller
5645 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5647 unsigned long flags;
5654 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5655 cpus_and(mask, mask, p->cpus_allowed);
5656 dest_cpu = any_online_cpu(mask);
5658 /* On any allowed CPU? */
5659 if (dest_cpu >= nr_cpu_ids)
5660 dest_cpu = any_online_cpu(p->cpus_allowed);
5662 /* No more Mr. Nice Guy. */
5663 if (dest_cpu >= nr_cpu_ids) {
5664 cpumask_t cpus_allowed;
5666 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5668 * Try to stay on the same cpuset, where the
5669 * current cpuset may be a subset of all cpus.
5670 * The cpuset_cpus_allowed_locked() variant of
5671 * cpuset_cpus_allowed() will not block. It must be
5672 * called within calls to cpuset_lock/cpuset_unlock.
5674 rq = task_rq_lock(p, &flags);
5675 p->cpus_allowed = cpus_allowed;
5676 dest_cpu = any_online_cpu(p->cpus_allowed);
5677 task_rq_unlock(rq, &flags);
5680 * Don't tell them about moving exiting tasks or
5681 * kernel threads (both mm NULL), since they never
5684 if (p->mm && printk_ratelimit()) {
5685 printk(KERN_INFO "process %d (%s) no "
5686 "longer affine to cpu%d\n",
5687 task_pid_nr(p), p->comm, dead_cpu);
5690 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5694 * While a dead CPU has no uninterruptible tasks queued at this point,
5695 * it might still have a nonzero ->nr_uninterruptible counter, because
5696 * for performance reasons the counter is not stricly tracking tasks to
5697 * their home CPUs. So we just add the counter to another CPU's counter,
5698 * to keep the global sum constant after CPU-down:
5700 static void migrate_nr_uninterruptible(struct rq *rq_src)
5702 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5703 unsigned long flags;
5705 local_irq_save(flags);
5706 double_rq_lock(rq_src, rq_dest);
5707 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5708 rq_src->nr_uninterruptible = 0;
5709 double_rq_unlock(rq_src, rq_dest);
5710 local_irq_restore(flags);
5713 /* Run through task list and migrate tasks from the dead cpu. */
5714 static void migrate_live_tasks(int src_cpu)
5716 struct task_struct *p, *t;
5718 read_lock(&tasklist_lock);
5720 do_each_thread(t, p) {
5724 if (task_cpu(p) == src_cpu)
5725 move_task_off_dead_cpu(src_cpu, p);
5726 } while_each_thread(t, p);
5728 read_unlock(&tasklist_lock);
5732 * Schedules idle task to be the next runnable task on current CPU.
5733 * It does so by boosting its priority to highest possible.
5734 * Used by CPU offline code.
5736 void sched_idle_next(void)
5738 int this_cpu = smp_processor_id();
5739 struct rq *rq = cpu_rq(this_cpu);
5740 struct task_struct *p = rq->idle;
5741 unsigned long flags;
5743 /* cpu has to be offline */
5744 BUG_ON(cpu_online(this_cpu));
5747 * Strictly not necessary since rest of the CPUs are stopped by now
5748 * and interrupts disabled on the current cpu.
5750 spin_lock_irqsave(&rq->lock, flags);
5752 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5754 update_rq_clock(rq);
5755 activate_task(rq, p, 0);
5757 spin_unlock_irqrestore(&rq->lock, flags);
5761 * Ensures that the idle task is using init_mm right before its cpu goes
5764 void idle_task_exit(void)
5766 struct mm_struct *mm = current->active_mm;
5768 BUG_ON(cpu_online(smp_processor_id()));
5771 switch_mm(mm, &init_mm, current);
5775 /* called under rq->lock with disabled interrupts */
5776 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5778 struct rq *rq = cpu_rq(dead_cpu);
5780 /* Must be exiting, otherwise would be on tasklist. */
5781 BUG_ON(!p->exit_state);
5783 /* Cannot have done final schedule yet: would have vanished. */
5784 BUG_ON(p->state == TASK_DEAD);
5789 * Drop lock around migration; if someone else moves it,
5790 * that's OK. No task can be added to this CPU, so iteration is
5793 spin_unlock_irq(&rq->lock);
5794 move_task_off_dead_cpu(dead_cpu, p);
5795 spin_lock_irq(&rq->lock);
5800 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5801 static void migrate_dead_tasks(unsigned int dead_cpu)
5803 struct rq *rq = cpu_rq(dead_cpu);
5804 struct task_struct *next;
5807 if (!rq->nr_running)
5809 update_rq_clock(rq);
5810 next = pick_next_task(rq, rq->curr);
5813 migrate_dead(dead_cpu, next);
5817 #endif /* CONFIG_HOTPLUG_CPU */
5819 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5821 static struct ctl_table sd_ctl_dir[] = {
5823 .procname = "sched_domain",
5829 static struct ctl_table sd_ctl_root[] = {
5831 .ctl_name = CTL_KERN,
5832 .procname = "kernel",
5834 .child = sd_ctl_dir,
5839 static struct ctl_table *sd_alloc_ctl_entry(int n)
5841 struct ctl_table *entry =
5842 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5847 static void sd_free_ctl_entry(struct ctl_table **tablep)
5849 struct ctl_table *entry;
5852 * In the intermediate directories, both the child directory and
5853 * procname are dynamically allocated and could fail but the mode
5854 * will always be set. In the lowest directory the names are
5855 * static strings and all have proc handlers.
5857 for (entry = *tablep; entry->mode; entry++) {
5859 sd_free_ctl_entry(&entry->child);
5860 if (entry->proc_handler == NULL)
5861 kfree(entry->procname);
5869 set_table_entry(struct ctl_table *entry,
5870 const char *procname, void *data, int maxlen,
5871 mode_t mode, proc_handler *proc_handler)
5873 entry->procname = procname;
5875 entry->maxlen = maxlen;
5877 entry->proc_handler = proc_handler;
5880 static struct ctl_table *
5881 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5883 struct ctl_table *table = sd_alloc_ctl_entry(12);
5888 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5889 sizeof(long), 0644, proc_doulongvec_minmax);
5890 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5891 sizeof(long), 0644, proc_doulongvec_minmax);
5892 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5893 sizeof(int), 0644, proc_dointvec_minmax);
5894 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5895 sizeof(int), 0644, proc_dointvec_minmax);
5896 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5897 sizeof(int), 0644, proc_dointvec_minmax);
5898 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5899 sizeof(int), 0644, proc_dointvec_minmax);
5900 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5901 sizeof(int), 0644, proc_dointvec_minmax);
5902 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5903 sizeof(int), 0644, proc_dointvec_minmax);
5904 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5905 sizeof(int), 0644, proc_dointvec_minmax);
5906 set_table_entry(&table[9], "cache_nice_tries",
5907 &sd->cache_nice_tries,
5908 sizeof(int), 0644, proc_dointvec_minmax);
5909 set_table_entry(&table[10], "flags", &sd->flags,
5910 sizeof(int), 0644, proc_dointvec_minmax);
5911 /* &table[11] is terminator */
5916 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5918 struct ctl_table *entry, *table;
5919 struct sched_domain *sd;
5920 int domain_num = 0, i;
5923 for_each_domain(cpu, sd)
5925 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5930 for_each_domain(cpu, sd) {
5931 snprintf(buf, 32, "domain%d", i);
5932 entry->procname = kstrdup(buf, GFP_KERNEL);
5934 entry->child = sd_alloc_ctl_domain_table(sd);
5941 static struct ctl_table_header *sd_sysctl_header;
5942 static void register_sched_domain_sysctl(void)
5944 int i, cpu_num = num_online_cpus();
5945 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5948 WARN_ON(sd_ctl_dir[0].child);
5949 sd_ctl_dir[0].child = entry;
5954 for_each_online_cpu(i) {
5955 snprintf(buf, 32, "cpu%d", i);
5956 entry->procname = kstrdup(buf, GFP_KERNEL);
5958 entry->child = sd_alloc_ctl_cpu_table(i);
5962 WARN_ON(sd_sysctl_header);
5963 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5966 /* may be called multiple times per register */
5967 static void unregister_sched_domain_sysctl(void)
5969 if (sd_sysctl_header)
5970 unregister_sysctl_table(sd_sysctl_header);
5971 sd_sysctl_header = NULL;
5972 if (sd_ctl_dir[0].child)
5973 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5976 static void register_sched_domain_sysctl(void)
5979 static void unregister_sched_domain_sysctl(void)
5985 * migration_call - callback that gets triggered when a CPU is added.
5986 * Here we can start up the necessary migration thread for the new CPU.
5988 static int __cpuinit
5989 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5991 struct task_struct *p;
5992 int cpu = (long)hcpu;
5993 unsigned long flags;
5998 case CPU_UP_PREPARE:
5999 case CPU_UP_PREPARE_FROZEN:
6000 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6003 kthread_bind(p, cpu);
6004 /* Must be high prio: stop_machine expects to yield to it. */
6005 rq = task_rq_lock(p, &flags);
6006 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6007 task_rq_unlock(rq, &flags);
6008 cpu_rq(cpu)->migration_thread = p;
6012 case CPU_ONLINE_FROZEN:
6013 /* Strictly unnecessary, as first user will wake it. */
6014 wake_up_process(cpu_rq(cpu)->migration_thread);
6016 /* Update our root-domain */
6018 spin_lock_irqsave(&rq->lock, flags);
6020 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6021 cpu_set(cpu, rq->rd->online);
6023 spin_unlock_irqrestore(&rq->lock, flags);
6026 #ifdef CONFIG_HOTPLUG_CPU
6027 case CPU_UP_CANCELED:
6028 case CPU_UP_CANCELED_FROZEN:
6029 if (!cpu_rq(cpu)->migration_thread)
6031 /* Unbind it from offline cpu so it can run. Fall thru. */
6032 kthread_bind(cpu_rq(cpu)->migration_thread,
6033 any_online_cpu(cpu_online_map));
6034 kthread_stop(cpu_rq(cpu)->migration_thread);
6035 cpu_rq(cpu)->migration_thread = NULL;
6039 case CPU_DEAD_FROZEN:
6040 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6041 migrate_live_tasks(cpu);
6043 kthread_stop(rq->migration_thread);
6044 rq->migration_thread = NULL;
6045 /* Idle task back to normal (off runqueue, low prio) */
6046 spin_lock_irq(&rq->lock);
6047 update_rq_clock(rq);
6048 deactivate_task(rq, rq->idle, 0);
6049 rq->idle->static_prio = MAX_PRIO;
6050 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6051 rq->idle->sched_class = &idle_sched_class;
6052 migrate_dead_tasks(cpu);
6053 spin_unlock_irq(&rq->lock);
6055 migrate_nr_uninterruptible(rq);
6056 BUG_ON(rq->nr_running != 0);
6059 * No need to migrate the tasks: it was best-effort if
6060 * they didn't take sched_hotcpu_mutex. Just wake up
6063 spin_lock_irq(&rq->lock);
6064 while (!list_empty(&rq->migration_queue)) {
6065 struct migration_req *req;
6067 req = list_entry(rq->migration_queue.next,
6068 struct migration_req, list);
6069 list_del_init(&req->list);
6070 complete(&req->done);
6072 spin_unlock_irq(&rq->lock);
6076 case CPU_DYING_FROZEN:
6077 /* Update our root-domain */
6079 spin_lock_irqsave(&rq->lock, flags);
6081 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6082 cpu_clear(cpu, rq->rd->online);
6084 spin_unlock_irqrestore(&rq->lock, flags);
6091 /* Register at highest priority so that task migration (migrate_all_tasks)
6092 * happens before everything else.
6094 static struct notifier_block __cpuinitdata migration_notifier = {
6095 .notifier_call = migration_call,
6099 void __init migration_init(void)
6101 void *cpu = (void *)(long)smp_processor_id();
6104 /* Start one for the boot CPU: */
6105 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6106 BUG_ON(err == NOTIFY_BAD);
6107 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6108 register_cpu_notifier(&migration_notifier);
6114 /* Number of possible processor ids */
6115 int nr_cpu_ids __read_mostly = NR_CPUS;
6116 EXPORT_SYMBOL(nr_cpu_ids);
6118 #ifdef CONFIG_SCHED_DEBUG
6120 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
6122 struct sched_group *group = sd->groups;
6123 cpumask_t groupmask;
6126 cpulist_scnprintf(str, sizeof(str), sd->span);
6127 cpus_clear(groupmask);
6129 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6131 if (!(sd->flags & SD_LOAD_BALANCE)) {
6132 printk("does not load-balance\n");
6134 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6139 printk(KERN_CONT "span %s\n", str);
6141 if (!cpu_isset(cpu, sd->span)) {
6142 printk(KERN_ERR "ERROR: domain->span does not contain "
6145 if (!cpu_isset(cpu, group->cpumask)) {
6146 printk(KERN_ERR "ERROR: domain->groups does not contain"
6150 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6154 printk(KERN_ERR "ERROR: group is NULL\n");
6158 if (!group->__cpu_power) {
6159 printk(KERN_CONT "\n");
6160 printk(KERN_ERR "ERROR: domain->cpu_power not "
6165 if (!cpus_weight(group->cpumask)) {
6166 printk(KERN_CONT "\n");
6167 printk(KERN_ERR "ERROR: empty group\n");
6171 if (cpus_intersects(groupmask, group->cpumask)) {
6172 printk(KERN_CONT "\n");
6173 printk(KERN_ERR "ERROR: repeated CPUs\n");
6177 cpus_or(groupmask, groupmask, group->cpumask);
6179 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6180 printk(KERN_CONT " %s", str);
6182 group = group->next;
6183 } while (group != sd->groups);
6184 printk(KERN_CONT "\n");
6186 if (!cpus_equal(sd->span, groupmask))
6187 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6189 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6190 printk(KERN_ERR "ERROR: parent span is not a superset "
6191 "of domain->span\n");
6195 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6200 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6204 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6207 if (sched_domain_debug_one(sd, cpu, level))
6216 # define sched_domain_debug(sd, cpu) do { } while (0)
6219 static int sd_degenerate(struct sched_domain *sd)
6221 if (cpus_weight(sd->span) == 1)
6224 /* Following flags need at least 2 groups */
6225 if (sd->flags & (SD_LOAD_BALANCE |
6226 SD_BALANCE_NEWIDLE |
6230 SD_SHARE_PKG_RESOURCES)) {
6231 if (sd->groups != sd->groups->next)
6235 /* Following flags don't use groups */
6236 if (sd->flags & (SD_WAKE_IDLE |
6245 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6247 unsigned long cflags = sd->flags, pflags = parent->flags;
6249 if (sd_degenerate(parent))
6252 if (!cpus_equal(sd->span, parent->span))
6255 /* Does parent contain flags not in child? */
6256 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6257 if (cflags & SD_WAKE_AFFINE)
6258 pflags &= ~SD_WAKE_BALANCE;
6259 /* Flags needing groups don't count if only 1 group in parent */
6260 if (parent->groups == parent->groups->next) {
6261 pflags &= ~(SD_LOAD_BALANCE |
6262 SD_BALANCE_NEWIDLE |
6266 SD_SHARE_PKG_RESOURCES);
6268 if (~cflags & pflags)
6274 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6276 unsigned long flags;
6277 const struct sched_class *class;
6279 spin_lock_irqsave(&rq->lock, flags);
6282 struct root_domain *old_rd = rq->rd;
6284 for (class = sched_class_highest; class; class = class->next) {
6285 if (class->leave_domain)
6286 class->leave_domain(rq);
6289 cpu_clear(rq->cpu, old_rd->span);
6290 cpu_clear(rq->cpu, old_rd->online);
6292 if (atomic_dec_and_test(&old_rd->refcount))
6296 atomic_inc(&rd->refcount);
6299 cpu_set(rq->cpu, rd->span);
6300 if (cpu_isset(rq->cpu, cpu_online_map))
6301 cpu_set(rq->cpu, rd->online);
6303 for (class = sched_class_highest; class; class = class->next) {
6304 if (class->join_domain)
6305 class->join_domain(rq);
6308 spin_unlock_irqrestore(&rq->lock, flags);
6311 static void init_rootdomain(struct root_domain *rd)
6313 memset(rd, 0, sizeof(*rd));
6315 cpus_clear(rd->span);
6316 cpus_clear(rd->online);
6319 static void init_defrootdomain(void)
6321 init_rootdomain(&def_root_domain);
6322 atomic_set(&def_root_domain.refcount, 1);
6325 static struct root_domain *alloc_rootdomain(void)
6327 struct root_domain *rd;
6329 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6333 init_rootdomain(rd);
6339 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6340 * hold the hotplug lock.
6343 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6345 struct rq *rq = cpu_rq(cpu);
6346 struct sched_domain *tmp;
6348 /* Remove the sched domains which do not contribute to scheduling. */
6349 for (tmp = sd; tmp; tmp = tmp->parent) {
6350 struct sched_domain *parent = tmp->parent;
6353 if (sd_parent_degenerate(tmp, parent)) {
6354 tmp->parent = parent->parent;
6356 parent->parent->child = tmp;
6360 if (sd && sd_degenerate(sd)) {
6366 sched_domain_debug(sd, cpu);
6368 rq_attach_root(rq, rd);
6369 rcu_assign_pointer(rq->sd, sd);
6372 /* cpus with isolated domains */
6373 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6375 /* Setup the mask of cpus configured for isolated domains */
6376 static int __init isolated_cpu_setup(char *str)
6378 int ints[NR_CPUS], i;
6380 str = get_options(str, ARRAY_SIZE(ints), ints);
6381 cpus_clear(cpu_isolated_map);
6382 for (i = 1; i <= ints[0]; i++)
6383 if (ints[i] < NR_CPUS)
6384 cpu_set(ints[i], cpu_isolated_map);
6388 __setup("isolcpus=", isolated_cpu_setup);
6391 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6392 * to a function which identifies what group(along with sched group) a CPU
6393 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6394 * (due to the fact that we keep track of groups covered with a cpumask_t).
6396 * init_sched_build_groups will build a circular linked list of the groups
6397 * covered by the given span, and will set each group's ->cpumask correctly,
6398 * and ->cpu_power to 0.
6401 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6402 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6403 struct sched_group **sg))
6405 struct sched_group *first = NULL, *last = NULL;
6406 cpumask_t covered = CPU_MASK_NONE;
6409 for_each_cpu_mask(i, span) {
6410 struct sched_group *sg;
6411 int group = group_fn(i, cpu_map, &sg);
6414 if (cpu_isset(i, covered))
6417 sg->cpumask = CPU_MASK_NONE;
6418 sg->__cpu_power = 0;
6420 for_each_cpu_mask(j, span) {
6421 if (group_fn(j, cpu_map, NULL) != group)
6424 cpu_set(j, covered);
6425 cpu_set(j, sg->cpumask);
6436 #define SD_NODES_PER_DOMAIN 16
6441 * find_next_best_node - find the next node to include in a sched_domain
6442 * @node: node whose sched_domain we're building
6443 * @used_nodes: nodes already in the sched_domain
6445 * Find the next node to include in a given scheduling domain. Simply
6446 * finds the closest node not already in the @used_nodes map.
6448 * Should use nodemask_t.
6450 static int find_next_best_node(int node, unsigned long *used_nodes)
6452 int i, n, val, min_val, best_node = 0;
6456 for (i = 0; i < MAX_NUMNODES; i++) {
6457 /* Start at @node */
6458 n = (node + i) % MAX_NUMNODES;
6460 if (!nr_cpus_node(n))
6463 /* Skip already used nodes */
6464 if (test_bit(n, used_nodes))
6467 /* Simple min distance search */
6468 val = node_distance(node, n);
6470 if (val < min_val) {
6476 set_bit(best_node, used_nodes);
6481 * sched_domain_node_span - get a cpumask for a node's sched_domain
6482 * @node: node whose cpumask we're constructing
6483 * @size: number of nodes to include in this span
6485 * Given a node, construct a good cpumask for its sched_domain to span. It
6486 * should be one that prevents unnecessary balancing, but also spreads tasks
6489 static cpumask_t sched_domain_node_span(int node)
6491 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6492 cpumask_t span, nodemask;
6496 bitmap_zero(used_nodes, MAX_NUMNODES);
6498 nodemask = node_to_cpumask(node);
6499 cpus_or(span, span, nodemask);
6500 set_bit(node, used_nodes);
6502 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6503 int next_node = find_next_best_node(node, used_nodes);
6505 nodemask = node_to_cpumask(next_node);
6506 cpus_or(span, span, nodemask);
6513 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6516 * SMT sched-domains:
6518 #ifdef CONFIG_SCHED_SMT
6519 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6520 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6523 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6526 *sg = &per_cpu(sched_group_cpus, cpu);
6532 * multi-core sched-domains:
6534 #ifdef CONFIG_SCHED_MC
6535 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6536 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6539 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6541 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6544 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6545 cpus_and(mask, mask, *cpu_map);
6546 group = first_cpu(mask);
6548 *sg = &per_cpu(sched_group_core, group);
6551 #elif defined(CONFIG_SCHED_MC)
6553 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6556 *sg = &per_cpu(sched_group_core, cpu);
6561 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6562 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6565 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6568 #ifdef CONFIG_SCHED_MC
6569 cpumask_t mask = cpu_coregroup_map(cpu);
6570 cpus_and(mask, mask, *cpu_map);
6571 group = first_cpu(mask);
6572 #elif defined(CONFIG_SCHED_SMT)
6573 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6574 cpus_and(mask, mask, *cpu_map);
6575 group = first_cpu(mask);
6580 *sg = &per_cpu(sched_group_phys, group);
6586 * The init_sched_build_groups can't handle what we want to do with node
6587 * groups, so roll our own. Now each node has its own list of groups which
6588 * gets dynamically allocated.
6590 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6591 static struct sched_group ***sched_group_nodes_bycpu;
6593 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6594 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6596 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6597 struct sched_group **sg)
6599 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6602 cpus_and(nodemask, nodemask, *cpu_map);
6603 group = first_cpu(nodemask);
6606 *sg = &per_cpu(sched_group_allnodes, group);
6610 static void init_numa_sched_groups_power(struct sched_group *group_head)
6612 struct sched_group *sg = group_head;
6618 for_each_cpu_mask(j, sg->cpumask) {
6619 struct sched_domain *sd;
6621 sd = &per_cpu(phys_domains, j);
6622 if (j != first_cpu(sd->groups->cpumask)) {
6624 * Only add "power" once for each
6630 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6633 } while (sg != group_head);
6638 /* Free memory allocated for various sched_group structures */
6639 static void free_sched_groups(const cpumask_t *cpu_map)
6643 for_each_cpu_mask(cpu, *cpu_map) {
6644 struct sched_group **sched_group_nodes
6645 = sched_group_nodes_bycpu[cpu];
6647 if (!sched_group_nodes)
6650 for (i = 0; i < MAX_NUMNODES; i++) {
6651 cpumask_t nodemask = node_to_cpumask(i);
6652 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6654 cpus_and(nodemask, nodemask, *cpu_map);
6655 if (cpus_empty(nodemask))
6665 if (oldsg != sched_group_nodes[i])
6668 kfree(sched_group_nodes);
6669 sched_group_nodes_bycpu[cpu] = NULL;
6673 static void free_sched_groups(const cpumask_t *cpu_map)
6679 * Initialize sched groups cpu_power.
6681 * cpu_power indicates the capacity of sched group, which is used while
6682 * distributing the load between different sched groups in a sched domain.
6683 * Typically cpu_power for all the groups in a sched domain will be same unless
6684 * there are asymmetries in the topology. If there are asymmetries, group
6685 * having more cpu_power will pickup more load compared to the group having
6688 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6689 * the maximum number of tasks a group can handle in the presence of other idle
6690 * or lightly loaded groups in the same sched domain.
6692 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6694 struct sched_domain *child;
6695 struct sched_group *group;
6697 WARN_ON(!sd || !sd->groups);
6699 if (cpu != first_cpu(sd->groups->cpumask))
6704 sd->groups->__cpu_power = 0;
6707 * For perf policy, if the groups in child domain share resources
6708 * (for example cores sharing some portions of the cache hierarchy
6709 * or SMT), then set this domain groups cpu_power such that each group
6710 * can handle only one task, when there are other idle groups in the
6711 * same sched domain.
6713 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6715 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6716 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6721 * add cpu_power of each child group to this groups cpu_power
6723 group = child->groups;
6725 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6726 group = group->next;
6727 } while (group != child->groups);
6731 * Build sched domains for a given set of cpus and attach the sched domains
6732 * to the individual cpus
6734 static int build_sched_domains(const cpumask_t *cpu_map)
6737 struct root_domain *rd;
6739 struct sched_group **sched_group_nodes = NULL;
6740 int sd_allnodes = 0;
6743 * Allocate the per-node list of sched groups
6745 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6747 if (!sched_group_nodes) {
6748 printk(KERN_WARNING "Can not alloc sched group node list\n");
6751 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6754 rd = alloc_rootdomain();
6756 printk(KERN_WARNING "Cannot alloc root domain\n");
6761 * Set up domains for cpus specified by the cpu_map.
6763 for_each_cpu_mask(i, *cpu_map) {
6764 struct sched_domain *sd = NULL, *p;
6765 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6767 cpus_and(nodemask, nodemask, *cpu_map);
6770 if (cpus_weight(*cpu_map) >
6771 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6772 sd = &per_cpu(allnodes_domains, i);
6773 *sd = SD_ALLNODES_INIT;
6774 sd->span = *cpu_map;
6775 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6781 sd = &per_cpu(node_domains, i);
6783 sd->span = sched_domain_node_span(cpu_to_node(i));
6787 cpus_and(sd->span, sd->span, *cpu_map);
6791 sd = &per_cpu(phys_domains, i);
6793 sd->span = nodemask;
6797 cpu_to_phys_group(i, cpu_map, &sd->groups);
6799 #ifdef CONFIG_SCHED_MC
6801 sd = &per_cpu(core_domains, i);
6803 sd->span = cpu_coregroup_map(i);
6804 cpus_and(sd->span, sd->span, *cpu_map);
6807 cpu_to_core_group(i, cpu_map, &sd->groups);
6810 #ifdef CONFIG_SCHED_SMT
6812 sd = &per_cpu(cpu_domains, i);
6813 *sd = SD_SIBLING_INIT;
6814 sd->span = per_cpu(cpu_sibling_map, i);
6815 cpus_and(sd->span, sd->span, *cpu_map);
6818 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6822 #ifdef CONFIG_SCHED_SMT
6823 /* Set up CPU (sibling) groups */
6824 for_each_cpu_mask(i, *cpu_map) {
6825 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6826 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6827 if (i != first_cpu(this_sibling_map))
6830 init_sched_build_groups(this_sibling_map, cpu_map,
6835 #ifdef CONFIG_SCHED_MC
6836 /* Set up multi-core groups */
6837 for_each_cpu_mask(i, *cpu_map) {
6838 cpumask_t this_core_map = cpu_coregroup_map(i);
6839 cpus_and(this_core_map, this_core_map, *cpu_map);
6840 if (i != first_cpu(this_core_map))
6842 init_sched_build_groups(this_core_map, cpu_map,
6843 &cpu_to_core_group);
6847 /* Set up physical groups */
6848 for (i = 0; i < MAX_NUMNODES; i++) {
6849 cpumask_t nodemask = node_to_cpumask(i);
6851 cpus_and(nodemask, nodemask, *cpu_map);
6852 if (cpus_empty(nodemask))
6855 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6859 /* Set up node groups */
6861 init_sched_build_groups(*cpu_map, cpu_map,
6862 &cpu_to_allnodes_group);
6864 for (i = 0; i < MAX_NUMNODES; i++) {
6865 /* Set up node groups */
6866 struct sched_group *sg, *prev;
6867 cpumask_t nodemask = node_to_cpumask(i);
6868 cpumask_t domainspan;
6869 cpumask_t covered = CPU_MASK_NONE;
6872 cpus_and(nodemask, nodemask, *cpu_map);
6873 if (cpus_empty(nodemask)) {
6874 sched_group_nodes[i] = NULL;
6878 domainspan = sched_domain_node_span(i);
6879 cpus_and(domainspan, domainspan, *cpu_map);
6881 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6883 printk(KERN_WARNING "Can not alloc domain group for "
6887 sched_group_nodes[i] = sg;
6888 for_each_cpu_mask(j, nodemask) {
6889 struct sched_domain *sd;
6891 sd = &per_cpu(node_domains, j);
6894 sg->__cpu_power = 0;
6895 sg->cpumask = nodemask;
6897 cpus_or(covered, covered, nodemask);
6900 for (j = 0; j < MAX_NUMNODES; j++) {
6901 cpumask_t tmp, notcovered;
6902 int n = (i + j) % MAX_NUMNODES;
6904 cpus_complement(notcovered, covered);
6905 cpus_and(tmp, notcovered, *cpu_map);
6906 cpus_and(tmp, tmp, domainspan);
6907 if (cpus_empty(tmp))
6910 nodemask = node_to_cpumask(n);
6911 cpus_and(tmp, tmp, nodemask);
6912 if (cpus_empty(tmp))
6915 sg = kmalloc_node(sizeof(struct sched_group),
6919 "Can not alloc domain group for node %d\n", j);
6922 sg->__cpu_power = 0;
6924 sg->next = prev->next;
6925 cpus_or(covered, covered, tmp);
6932 /* Calculate CPU power for physical packages and nodes */
6933 #ifdef CONFIG_SCHED_SMT
6934 for_each_cpu_mask(i, *cpu_map) {
6935 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6937 init_sched_groups_power(i, sd);
6940 #ifdef CONFIG_SCHED_MC
6941 for_each_cpu_mask(i, *cpu_map) {
6942 struct sched_domain *sd = &per_cpu(core_domains, i);
6944 init_sched_groups_power(i, sd);
6948 for_each_cpu_mask(i, *cpu_map) {
6949 struct sched_domain *sd = &per_cpu(phys_domains, i);
6951 init_sched_groups_power(i, sd);
6955 for (i = 0; i < MAX_NUMNODES; i++)
6956 init_numa_sched_groups_power(sched_group_nodes[i]);
6959 struct sched_group *sg;
6961 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6962 init_numa_sched_groups_power(sg);
6966 /* Attach the domains */
6967 for_each_cpu_mask(i, *cpu_map) {
6968 struct sched_domain *sd;
6969 #ifdef CONFIG_SCHED_SMT
6970 sd = &per_cpu(cpu_domains, i);
6971 #elif defined(CONFIG_SCHED_MC)
6972 sd = &per_cpu(core_domains, i);
6974 sd = &per_cpu(phys_domains, i);
6976 cpu_attach_domain(sd, rd, i);
6983 free_sched_groups(cpu_map);
6988 static cpumask_t *doms_cur; /* current sched domains */
6989 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6992 * Special case: If a kmalloc of a doms_cur partition (array of
6993 * cpumask_t) fails, then fallback to a single sched domain,
6994 * as determined by the single cpumask_t fallback_doms.
6996 static cpumask_t fallback_doms;
6998 void __attribute__((weak)) arch_update_cpu_topology(void)
7003 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7004 * For now this just excludes isolated cpus, but could be used to
7005 * exclude other special cases in the future.
7007 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7011 arch_update_cpu_topology();
7013 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7015 doms_cur = &fallback_doms;
7016 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7017 err = build_sched_domains(doms_cur);
7018 register_sched_domain_sysctl();
7023 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
7025 free_sched_groups(cpu_map);
7029 * Detach sched domains from a group of cpus specified in cpu_map
7030 * These cpus will now be attached to the NULL domain
7032 static void detach_destroy_domains(const cpumask_t *cpu_map)
7036 unregister_sched_domain_sysctl();
7038 for_each_cpu_mask(i, *cpu_map)
7039 cpu_attach_domain(NULL, &def_root_domain, i);
7040 synchronize_sched();
7041 arch_destroy_sched_domains(cpu_map);
7045 * Partition sched domains as specified by the 'ndoms_new'
7046 * cpumasks in the array doms_new[] of cpumasks. This compares
7047 * doms_new[] to the current sched domain partitioning, doms_cur[].
7048 * It destroys each deleted domain and builds each new domain.
7050 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7051 * The masks don't intersect (don't overlap.) We should setup one
7052 * sched domain for each mask. CPUs not in any of the cpumasks will
7053 * not be load balanced. If the same cpumask appears both in the
7054 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7057 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7058 * ownership of it and will kfree it when done with it. If the caller
7059 * failed the kmalloc call, then it can pass in doms_new == NULL,
7060 * and partition_sched_domains() will fallback to the single partition
7063 * Call with hotplug lock held
7065 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7071 /* always unregister in case we don't destroy any domains */
7072 unregister_sched_domain_sysctl();
7074 if (doms_new == NULL) {
7076 doms_new = &fallback_doms;
7077 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7080 /* Destroy deleted domains */
7081 for (i = 0; i < ndoms_cur; i++) {
7082 for (j = 0; j < ndoms_new; j++) {
7083 if (cpus_equal(doms_cur[i], doms_new[j]))
7086 /* no match - a current sched domain not in new doms_new[] */
7087 detach_destroy_domains(doms_cur + i);
7092 /* Build new domains */
7093 for (i = 0; i < ndoms_new; i++) {
7094 for (j = 0; j < ndoms_cur; j++) {
7095 if (cpus_equal(doms_new[i], doms_cur[j]))
7098 /* no match - add a new doms_new */
7099 build_sched_domains(doms_new + i);
7104 /* Remember the new sched domains */
7105 if (doms_cur != &fallback_doms)
7107 doms_cur = doms_new;
7108 ndoms_cur = ndoms_new;
7110 register_sched_domain_sysctl();
7115 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7116 int arch_reinit_sched_domains(void)
7121 detach_destroy_domains(&cpu_online_map);
7122 err = arch_init_sched_domains(&cpu_online_map);
7128 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7132 if (buf[0] != '0' && buf[0] != '1')
7136 sched_smt_power_savings = (buf[0] == '1');
7138 sched_mc_power_savings = (buf[0] == '1');
7140 ret = arch_reinit_sched_domains();
7142 return ret ? ret : count;
7145 #ifdef CONFIG_SCHED_MC
7146 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7148 return sprintf(page, "%u\n", sched_mc_power_savings);
7150 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7151 const char *buf, size_t count)
7153 return sched_power_savings_store(buf, count, 0);
7155 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7156 sched_mc_power_savings_store);
7159 #ifdef CONFIG_SCHED_SMT
7160 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7162 return sprintf(page, "%u\n", sched_smt_power_savings);
7164 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7165 const char *buf, size_t count)
7167 return sched_power_savings_store(buf, count, 1);
7169 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7170 sched_smt_power_savings_store);
7173 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7177 #ifdef CONFIG_SCHED_SMT
7179 err = sysfs_create_file(&cls->kset.kobj,
7180 &attr_sched_smt_power_savings.attr);
7182 #ifdef CONFIG_SCHED_MC
7183 if (!err && mc_capable())
7184 err = sysfs_create_file(&cls->kset.kobj,
7185 &attr_sched_mc_power_savings.attr);
7192 * Force a reinitialization of the sched domains hierarchy. The domains
7193 * and groups cannot be updated in place without racing with the balancing
7194 * code, so we temporarily attach all running cpus to the NULL domain
7195 * which will prevent rebalancing while the sched domains are recalculated.
7197 static int update_sched_domains(struct notifier_block *nfb,
7198 unsigned long action, void *hcpu)
7201 case CPU_UP_PREPARE:
7202 case CPU_UP_PREPARE_FROZEN:
7203 case CPU_DOWN_PREPARE:
7204 case CPU_DOWN_PREPARE_FROZEN:
7205 detach_destroy_domains(&cpu_online_map);
7208 case CPU_UP_CANCELED:
7209 case CPU_UP_CANCELED_FROZEN:
7210 case CPU_DOWN_FAILED:
7211 case CPU_DOWN_FAILED_FROZEN:
7213 case CPU_ONLINE_FROZEN:
7215 case CPU_DEAD_FROZEN:
7217 * Fall through and re-initialise the domains.
7224 /* The hotplug lock is already held by cpu_up/cpu_down */
7225 arch_init_sched_domains(&cpu_online_map);
7230 void __init sched_init_smp(void)
7232 cpumask_t non_isolated_cpus;
7234 #if defined(CONFIG_NUMA)
7235 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7237 BUG_ON(sched_group_nodes_bycpu == NULL);
7240 arch_init_sched_domains(&cpu_online_map);
7241 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7242 if (cpus_empty(non_isolated_cpus))
7243 cpu_set(smp_processor_id(), non_isolated_cpus);
7245 /* XXX: Theoretical race here - CPU may be hotplugged now */
7246 hotcpu_notifier(update_sched_domains, 0);
7248 /* Move init over to a non-isolated CPU */
7249 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7251 sched_init_granularity();
7254 void __init sched_init_smp(void)
7256 #if defined(CONFIG_NUMA)
7257 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7259 BUG_ON(sched_group_nodes_bycpu == NULL);
7261 sched_init_granularity();
7263 #endif /* CONFIG_SMP */
7265 int in_sched_functions(unsigned long addr)
7267 return in_lock_functions(addr) ||
7268 (addr >= (unsigned long)__sched_text_start
7269 && addr < (unsigned long)__sched_text_end);
7272 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7274 cfs_rq->tasks_timeline = RB_ROOT;
7275 #ifdef CONFIG_FAIR_GROUP_SCHED
7278 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7281 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7283 struct rt_prio_array *array;
7286 array = &rt_rq->active;
7287 for (i = 0; i < MAX_RT_PRIO; i++) {
7288 INIT_LIST_HEAD(array->queue + i);
7289 __clear_bit(i, array->bitmap);
7291 /* delimiter for bitsearch: */
7292 __set_bit(MAX_RT_PRIO, array->bitmap);
7294 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7295 rt_rq->highest_prio = MAX_RT_PRIO;
7298 rt_rq->rt_nr_migratory = 0;
7299 rt_rq->overloaded = 0;
7303 rt_rq->rt_throttled = 0;
7304 rt_rq->rt_runtime = 0;
7305 spin_lock_init(&rt_rq->rt_runtime_lock);
7307 #ifdef CONFIG_RT_GROUP_SCHED
7308 rt_rq->rt_nr_boosted = 0;
7313 #ifdef CONFIG_FAIR_GROUP_SCHED
7314 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7315 struct cfs_rq *cfs_rq, struct sched_entity *se,
7318 tg->cfs_rq[cpu] = cfs_rq;
7319 init_cfs_rq(cfs_rq, rq);
7322 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7325 se->cfs_rq = &rq->cfs;
7327 se->load.weight = tg->shares;
7328 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7333 #ifdef CONFIG_RT_GROUP_SCHED
7334 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7335 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7338 tg->rt_rq[cpu] = rt_rq;
7339 init_rt_rq(rt_rq, rq);
7341 rt_rq->rt_se = rt_se;
7342 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7344 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7346 tg->rt_se[cpu] = rt_se;
7347 rt_se->rt_rq = &rq->rt;
7348 rt_se->my_q = rt_rq;
7349 rt_se->parent = NULL;
7350 INIT_LIST_HEAD(&rt_se->run_list);
7354 void __init sched_init(void)
7356 int highest_cpu = 0;
7358 unsigned long alloc_size = 0, ptr;
7360 #ifdef CONFIG_FAIR_GROUP_SCHED
7361 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7363 #ifdef CONFIG_RT_GROUP_SCHED
7364 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7367 * As sched_init() is called before page_alloc is setup,
7368 * we use alloc_bootmem().
7371 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7373 #ifdef CONFIG_FAIR_GROUP_SCHED
7374 init_task_group.se = (struct sched_entity **)ptr;
7375 ptr += nr_cpu_ids * sizeof(void **);
7377 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7378 ptr += nr_cpu_ids * sizeof(void **);
7380 #ifdef CONFIG_RT_GROUP_SCHED
7381 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7382 ptr += nr_cpu_ids * sizeof(void **);
7384 init_task_group.rt_rq = (struct rt_rq **)ptr;
7389 init_defrootdomain();
7392 init_rt_bandwidth(&def_rt_bandwidth,
7393 global_rt_period(), global_rt_runtime());
7395 #ifdef CONFIG_RT_GROUP_SCHED
7396 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7397 global_rt_period(), global_rt_runtime());
7400 #ifdef CONFIG_GROUP_SCHED
7401 list_add(&init_task_group.list, &task_groups);
7404 for_each_possible_cpu(i) {
7408 spin_lock_init(&rq->lock);
7409 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7412 update_last_tick_seen(rq);
7413 init_cfs_rq(&rq->cfs, rq);
7414 init_rt_rq(&rq->rt, rq);
7415 #ifdef CONFIG_FAIR_GROUP_SCHED
7416 init_task_group.shares = init_task_group_load;
7417 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7418 init_tg_cfs_entry(rq, &init_task_group,
7419 &per_cpu(init_cfs_rq, i),
7420 &per_cpu(init_sched_entity, i), i, 1);
7423 #ifdef CONFIG_RT_GROUP_SCHED
7424 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7425 init_tg_rt_entry(rq, &init_task_group,
7426 &per_cpu(init_rt_rq, i),
7427 &per_cpu(init_sched_rt_entity, i), i, 1);
7429 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7432 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7433 rq->cpu_load[j] = 0;
7437 rq->active_balance = 0;
7438 rq->next_balance = jiffies;
7441 rq->migration_thread = NULL;
7442 INIT_LIST_HEAD(&rq->migration_queue);
7443 rq_attach_root(rq, &def_root_domain);
7446 atomic_set(&rq->nr_iowait, 0);
7450 set_load_weight(&init_task);
7452 #ifdef CONFIG_PREEMPT_NOTIFIERS
7453 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7457 nr_cpu_ids = highest_cpu + 1;
7458 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7461 #ifdef CONFIG_RT_MUTEXES
7462 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7466 * The boot idle thread does lazy MMU switching as well:
7468 atomic_inc(&init_mm.mm_count);
7469 enter_lazy_tlb(&init_mm, current);
7472 * Make us the idle thread. Technically, schedule() should not be
7473 * called from this thread, however somewhere below it might be,
7474 * but because we are the idle thread, we just pick up running again
7475 * when this runqueue becomes "idle".
7477 init_idle(current, smp_processor_id());
7479 * During early bootup we pretend to be a normal task:
7481 current->sched_class = &fair_sched_class;
7483 scheduler_running = 1;
7486 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7487 void __might_sleep(char *file, int line)
7490 static unsigned long prev_jiffy; /* ratelimiting */
7492 if ((in_atomic() || irqs_disabled()) &&
7493 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7494 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7496 prev_jiffy = jiffies;
7497 printk(KERN_ERR "BUG: sleeping function called from invalid"
7498 " context at %s:%d\n", file, line);
7499 printk("in_atomic():%d, irqs_disabled():%d\n",
7500 in_atomic(), irqs_disabled());
7501 debug_show_held_locks(current);
7502 if (irqs_disabled())
7503 print_irqtrace_events(current);
7508 EXPORT_SYMBOL(__might_sleep);
7511 #ifdef CONFIG_MAGIC_SYSRQ
7512 static void normalize_task(struct rq *rq, struct task_struct *p)
7515 update_rq_clock(rq);
7516 on_rq = p->se.on_rq;
7518 deactivate_task(rq, p, 0);
7519 __setscheduler(rq, p, SCHED_NORMAL, 0);
7521 activate_task(rq, p, 0);
7522 resched_task(rq->curr);
7526 void normalize_rt_tasks(void)
7528 struct task_struct *g, *p;
7529 unsigned long flags;
7532 read_lock_irqsave(&tasklist_lock, flags);
7533 do_each_thread(g, p) {
7535 * Only normalize user tasks:
7540 p->se.exec_start = 0;
7541 #ifdef CONFIG_SCHEDSTATS
7542 p->se.wait_start = 0;
7543 p->se.sleep_start = 0;
7544 p->se.block_start = 0;
7546 task_rq(p)->clock = 0;
7550 * Renice negative nice level userspace
7553 if (TASK_NICE(p) < 0 && p->mm)
7554 set_user_nice(p, 0);
7558 spin_lock(&p->pi_lock);
7559 rq = __task_rq_lock(p);
7561 normalize_task(rq, p);
7563 __task_rq_unlock(rq);
7564 spin_unlock(&p->pi_lock);
7565 } while_each_thread(g, p);
7567 read_unlock_irqrestore(&tasklist_lock, flags);
7570 #endif /* CONFIG_MAGIC_SYSRQ */
7574 * These functions are only useful for the IA64 MCA handling.
7576 * They can only be called when the whole system has been
7577 * stopped - every CPU needs to be quiescent, and no scheduling
7578 * activity can take place. Using them for anything else would
7579 * be a serious bug, and as a result, they aren't even visible
7580 * under any other configuration.
7584 * curr_task - return the current task for a given cpu.
7585 * @cpu: the processor in question.
7587 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7589 struct task_struct *curr_task(int cpu)
7591 return cpu_curr(cpu);
7595 * set_curr_task - set the current task for a given cpu.
7596 * @cpu: the processor in question.
7597 * @p: the task pointer to set.
7599 * Description: This function must only be used when non-maskable interrupts
7600 * are serviced on a separate stack. It allows the architecture to switch the
7601 * notion of the current task on a cpu in a non-blocking manner. This function
7602 * must be called with all CPU's synchronized, and interrupts disabled, the
7603 * and caller must save the original value of the current task (see
7604 * curr_task() above) and restore that value before reenabling interrupts and
7605 * re-starting the system.
7607 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7609 void set_curr_task(int cpu, struct task_struct *p)
7616 #ifdef CONFIG_FAIR_GROUP_SCHED
7617 static void free_fair_sched_group(struct task_group *tg)
7621 for_each_possible_cpu(i) {
7623 kfree(tg->cfs_rq[i]);
7632 static int alloc_fair_sched_group(struct task_group *tg)
7634 struct cfs_rq *cfs_rq;
7635 struct sched_entity *se;
7639 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7642 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7646 tg->shares = NICE_0_LOAD;
7648 for_each_possible_cpu(i) {
7651 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7652 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7656 se = kmalloc_node(sizeof(struct sched_entity),
7657 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7661 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7670 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7672 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7673 &cpu_rq(cpu)->leaf_cfs_rq_list);
7676 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7678 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7681 static inline void free_fair_sched_group(struct task_group *tg)
7685 static inline int alloc_fair_sched_group(struct task_group *tg)
7690 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7694 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7699 #ifdef CONFIG_RT_GROUP_SCHED
7700 static void free_rt_sched_group(struct task_group *tg)
7704 destroy_rt_bandwidth(&tg->rt_bandwidth);
7706 for_each_possible_cpu(i) {
7708 kfree(tg->rt_rq[i]);
7710 kfree(tg->rt_se[i]);
7717 static int alloc_rt_sched_group(struct task_group *tg)
7719 struct rt_rq *rt_rq;
7720 struct sched_rt_entity *rt_se;
7724 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7727 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7731 init_rt_bandwidth(&tg->rt_bandwidth,
7732 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7734 for_each_possible_cpu(i) {
7737 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7738 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7742 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7743 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7747 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7756 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7758 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7759 &cpu_rq(cpu)->leaf_rt_rq_list);
7762 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7764 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7767 static inline void free_rt_sched_group(struct task_group *tg)
7771 static inline int alloc_rt_sched_group(struct task_group *tg)
7776 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7780 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7785 #ifdef CONFIG_GROUP_SCHED
7786 static void free_sched_group(struct task_group *tg)
7788 free_fair_sched_group(tg);
7789 free_rt_sched_group(tg);
7793 /* allocate runqueue etc for a new task group */
7794 struct task_group *sched_create_group(void)
7796 struct task_group *tg;
7797 unsigned long flags;
7800 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7802 return ERR_PTR(-ENOMEM);
7804 if (!alloc_fair_sched_group(tg))
7807 if (!alloc_rt_sched_group(tg))
7810 spin_lock_irqsave(&task_group_lock, flags);
7811 for_each_possible_cpu(i) {
7812 register_fair_sched_group(tg, i);
7813 register_rt_sched_group(tg, i);
7815 list_add_rcu(&tg->list, &task_groups);
7816 spin_unlock_irqrestore(&task_group_lock, flags);
7821 free_sched_group(tg);
7822 return ERR_PTR(-ENOMEM);
7825 /* rcu callback to free various structures associated with a task group */
7826 static void free_sched_group_rcu(struct rcu_head *rhp)
7828 /* now it should be safe to free those cfs_rqs */
7829 free_sched_group(container_of(rhp, struct task_group, rcu));
7832 /* Destroy runqueue etc associated with a task group */
7833 void sched_destroy_group(struct task_group *tg)
7835 unsigned long flags;
7838 spin_lock_irqsave(&task_group_lock, flags);
7839 for_each_possible_cpu(i) {
7840 unregister_fair_sched_group(tg, i);
7841 unregister_rt_sched_group(tg, i);
7843 list_del_rcu(&tg->list);
7844 spin_unlock_irqrestore(&task_group_lock, flags);
7846 /* wait for possible concurrent references to cfs_rqs complete */
7847 call_rcu(&tg->rcu, free_sched_group_rcu);
7850 /* change task's runqueue when it moves between groups.
7851 * The caller of this function should have put the task in its new group
7852 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7853 * reflect its new group.
7855 void sched_move_task(struct task_struct *tsk)
7858 unsigned long flags;
7861 rq = task_rq_lock(tsk, &flags);
7863 update_rq_clock(rq);
7865 running = task_current(rq, tsk);
7866 on_rq = tsk->se.on_rq;
7869 dequeue_task(rq, tsk, 0);
7870 if (unlikely(running))
7871 tsk->sched_class->put_prev_task(rq, tsk);
7873 set_task_rq(tsk, task_cpu(tsk));
7875 #ifdef CONFIG_FAIR_GROUP_SCHED
7876 if (tsk->sched_class->moved_group)
7877 tsk->sched_class->moved_group(tsk);
7880 if (unlikely(running))
7881 tsk->sched_class->set_curr_task(rq);
7883 enqueue_task(rq, tsk, 0);
7885 task_rq_unlock(rq, &flags);
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7890 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7892 struct cfs_rq *cfs_rq = se->cfs_rq;
7893 struct rq *rq = cfs_rq->rq;
7896 spin_lock_irq(&rq->lock);
7900 dequeue_entity(cfs_rq, se, 0);
7902 se->load.weight = shares;
7903 se->load.inv_weight = div64_64((1ULL<<32), shares);
7906 enqueue_entity(cfs_rq, se, 0);
7908 spin_unlock_irq(&rq->lock);
7911 static DEFINE_MUTEX(shares_mutex);
7913 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7916 unsigned long flags;
7919 * A weight of 0 or 1 can cause arithmetics problems.
7920 * (The default weight is 1024 - so there's no practical
7921 * limitation from this.)
7926 mutex_lock(&shares_mutex);
7927 if (tg->shares == shares)
7930 spin_lock_irqsave(&task_group_lock, flags);
7931 for_each_possible_cpu(i)
7932 unregister_fair_sched_group(tg, i);
7933 spin_unlock_irqrestore(&task_group_lock, flags);
7935 /* wait for any ongoing reference to this group to finish */
7936 synchronize_sched();
7939 * Now we are free to modify the group's share on each cpu
7940 * w/o tripping rebalance_share or load_balance_fair.
7942 tg->shares = shares;
7943 for_each_possible_cpu(i)
7944 set_se_shares(tg->se[i], shares);
7947 * Enable load balance activity on this group, by inserting it back on
7948 * each cpu's rq->leaf_cfs_rq_list.
7950 spin_lock_irqsave(&task_group_lock, flags);
7951 for_each_possible_cpu(i)
7952 register_fair_sched_group(tg, i);
7953 spin_unlock_irqrestore(&task_group_lock, flags);
7955 mutex_unlock(&shares_mutex);
7959 unsigned long sched_group_shares(struct task_group *tg)
7965 #ifdef CONFIG_RT_GROUP_SCHED
7967 * Ensure that the real time constraints are schedulable.
7969 static DEFINE_MUTEX(rt_constraints_mutex);
7971 static unsigned long to_ratio(u64 period, u64 runtime)
7973 if (runtime == RUNTIME_INF)
7976 return div64_64(runtime << 16, period);
7979 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7981 struct task_group *tgi;
7982 unsigned long total = 0;
7983 unsigned long global_ratio =
7984 to_ratio(global_rt_period(), global_rt_runtime());
7987 list_for_each_entry_rcu(tgi, &task_groups, list) {
7991 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
7992 tgi->rt_bandwidth.rt_runtime);
7996 return total + to_ratio(period, runtime) < global_ratio;
7999 /* Must be called with tasklist_lock held */
8000 static inline int tg_has_rt_tasks(struct task_group *tg)
8002 struct task_struct *g, *p;
8003 do_each_thread(g, p) {
8004 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8006 } while_each_thread(g, p);
8010 static int tg_set_bandwidth(struct task_group *tg,
8011 u64 rt_period, u64 rt_runtime)
8015 mutex_lock(&rt_constraints_mutex);
8016 read_lock(&tasklist_lock);
8017 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8021 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8026 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8027 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8028 tg->rt_bandwidth.rt_runtime = rt_runtime;
8030 for_each_possible_cpu(i) {
8031 struct rt_rq *rt_rq = tg->rt_rq[i];
8033 spin_lock(&rt_rq->rt_runtime_lock);
8034 rt_rq->rt_runtime = rt_runtime;
8035 spin_unlock(&rt_rq->rt_runtime_lock);
8037 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8039 read_unlock(&tasklist_lock);
8040 mutex_unlock(&rt_constraints_mutex);
8045 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8047 u64 rt_runtime, rt_period;
8049 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8050 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8051 if (rt_runtime_us < 0)
8052 rt_runtime = RUNTIME_INF;
8054 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8057 long sched_group_rt_runtime(struct task_group *tg)
8061 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8064 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8065 do_div(rt_runtime_us, NSEC_PER_USEC);
8066 return rt_runtime_us;
8069 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8071 u64 rt_runtime, rt_period;
8073 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8074 rt_runtime = tg->rt_bandwidth.rt_runtime;
8076 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8079 long sched_group_rt_period(struct task_group *tg)
8083 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8084 do_div(rt_period_us, NSEC_PER_USEC);
8085 return rt_period_us;
8088 static int sched_rt_global_constraints(void)
8092 mutex_lock(&rt_constraints_mutex);
8093 if (!__rt_schedulable(NULL, 1, 0))
8095 mutex_unlock(&rt_constraints_mutex);
8100 static int sched_rt_global_constraints(void)
8102 unsigned long flags;
8105 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8106 for_each_possible_cpu(i) {
8107 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8109 spin_lock(&rt_rq->rt_runtime_lock);
8110 rt_rq->rt_runtime = global_rt_runtime();
8111 spin_unlock(&rt_rq->rt_runtime_lock);
8113 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8119 int sched_rt_handler(struct ctl_table *table, int write,
8120 struct file *filp, void __user *buffer, size_t *lenp,
8124 int old_period, old_runtime;
8125 static DEFINE_MUTEX(mutex);
8128 old_period = sysctl_sched_rt_period;
8129 old_runtime = sysctl_sched_rt_runtime;
8131 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8133 if (!ret && write) {
8134 ret = sched_rt_global_constraints();
8136 sysctl_sched_rt_period = old_period;
8137 sysctl_sched_rt_runtime = old_runtime;
8139 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8140 def_rt_bandwidth.rt_period =
8141 ns_to_ktime(global_rt_period());
8144 mutex_unlock(&mutex);
8149 #ifdef CONFIG_CGROUP_SCHED
8151 /* return corresponding task_group object of a cgroup */
8152 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8154 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8155 struct task_group, css);
8158 static struct cgroup_subsys_state *
8159 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8161 struct task_group *tg;
8163 if (!cgrp->parent) {
8164 /* This is early initialization for the top cgroup */
8165 init_task_group.css.cgroup = cgrp;
8166 return &init_task_group.css;
8169 /* we support only 1-level deep hierarchical scheduler atm */
8170 if (cgrp->parent->parent)
8171 return ERR_PTR(-EINVAL);
8173 tg = sched_create_group();
8175 return ERR_PTR(-ENOMEM);
8177 /* Bind the cgroup to task_group object we just created */
8178 tg->css.cgroup = cgrp;
8184 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8186 struct task_group *tg = cgroup_tg(cgrp);
8188 sched_destroy_group(tg);
8192 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8193 struct task_struct *tsk)
8195 #ifdef CONFIG_RT_GROUP_SCHED
8196 /* Don't accept realtime tasks when there is no way for them to run */
8197 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8200 /* We don't support RT-tasks being in separate groups */
8201 if (tsk->sched_class != &fair_sched_class)
8209 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8210 struct cgroup *old_cont, struct task_struct *tsk)
8212 sched_move_task(tsk);
8215 #ifdef CONFIG_FAIR_GROUP_SCHED
8216 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8219 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8222 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8224 struct task_group *tg = cgroup_tg(cgrp);
8226 return (u64) tg->shares;
8230 #ifdef CONFIG_RT_GROUP_SCHED
8231 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8233 const char __user *userbuf,
8234 size_t nbytes, loff_t *unused_ppos)
8243 if (nbytes >= sizeof(buffer))
8245 if (copy_from_user(buffer, userbuf, nbytes))
8248 buffer[nbytes] = 0; /* nul-terminate */
8250 /* strip newline if necessary */
8251 if (nbytes && (buffer[nbytes-1] == '\n'))
8252 buffer[nbytes-1] = 0;
8253 val = simple_strtoll(buffer, &end, 0);
8257 /* Pass to subsystem */
8258 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8264 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8266 char __user *buf, size_t nbytes,
8270 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8271 int len = sprintf(tmp, "%ld\n", val);
8273 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8276 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8279 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8282 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8284 return sched_group_rt_period(cgroup_tg(cgrp));
8288 static struct cftype cpu_files[] = {
8289 #ifdef CONFIG_FAIR_GROUP_SCHED
8292 .read_uint = cpu_shares_read_uint,
8293 .write_uint = cpu_shares_write_uint,
8296 #ifdef CONFIG_RT_GROUP_SCHED
8298 .name = "rt_runtime_us",
8299 .read = cpu_rt_runtime_read,
8300 .write = cpu_rt_runtime_write,
8303 .name = "rt_period_us",
8304 .read_uint = cpu_rt_period_read_uint,
8305 .write_uint = cpu_rt_period_write_uint,
8310 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8312 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8315 struct cgroup_subsys cpu_cgroup_subsys = {
8317 .create = cpu_cgroup_create,
8318 .destroy = cpu_cgroup_destroy,
8319 .can_attach = cpu_cgroup_can_attach,
8320 .attach = cpu_cgroup_attach,
8321 .populate = cpu_cgroup_populate,
8322 .subsys_id = cpu_cgroup_subsys_id,
8326 #endif /* CONFIG_CGROUP_SCHED */
8328 #ifdef CONFIG_CGROUP_CPUACCT
8331 * CPU accounting code for task groups.
8333 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8334 * (balbir@in.ibm.com).
8337 /* track cpu usage of a group of tasks */
8339 struct cgroup_subsys_state css;
8340 /* cpuusage holds pointer to a u64-type object on every cpu */
8344 struct cgroup_subsys cpuacct_subsys;
8346 /* return cpu accounting group corresponding to this container */
8347 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8349 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8350 struct cpuacct, css);
8353 /* return cpu accounting group to which this task belongs */
8354 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8356 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8357 struct cpuacct, css);
8360 /* create a new cpu accounting group */
8361 static struct cgroup_subsys_state *cpuacct_create(
8362 struct cgroup_subsys *ss, struct cgroup *cgrp)
8364 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8367 return ERR_PTR(-ENOMEM);
8369 ca->cpuusage = alloc_percpu(u64);
8370 if (!ca->cpuusage) {
8372 return ERR_PTR(-ENOMEM);
8378 /* destroy an existing cpu accounting group */
8380 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8382 struct cpuacct *ca = cgroup_ca(cgrp);
8384 free_percpu(ca->cpuusage);
8388 /* return total cpu usage (in nanoseconds) of a group */
8389 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8391 struct cpuacct *ca = cgroup_ca(cgrp);
8392 u64 totalcpuusage = 0;
8395 for_each_possible_cpu(i) {
8396 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8399 * Take rq->lock to make 64-bit addition safe on 32-bit
8402 spin_lock_irq(&cpu_rq(i)->lock);
8403 totalcpuusage += *cpuusage;
8404 spin_unlock_irq(&cpu_rq(i)->lock);
8407 return totalcpuusage;
8410 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8413 struct cpuacct *ca = cgroup_ca(cgrp);
8422 for_each_possible_cpu(i) {
8423 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8425 spin_lock_irq(&cpu_rq(i)->lock);
8427 spin_unlock_irq(&cpu_rq(i)->lock);
8433 static struct cftype files[] = {
8436 .read_uint = cpuusage_read,
8437 .write_uint = cpuusage_write,
8441 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8443 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8447 * charge this task's execution time to its accounting group.
8449 * called with rq->lock held.
8451 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8455 if (!cpuacct_subsys.active)
8460 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8462 *cpuusage += cputime;
8466 struct cgroup_subsys cpuacct_subsys = {
8468 .create = cpuacct_create,
8469 .destroy = cpuacct_destroy,
8470 .populate = cpuacct_populate,
8471 .subsys_id = cpuacct_subsys_id,
8473 #endif /* CONFIG_CGROUP_CPUACCT */