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,
683 SCHED_FEAT_NORMALIZED_SLEEPER = 256,
686 const_debug unsigned int sysctl_sched_features =
687 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
688 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
689 SCHED_FEAT_START_DEBIT * 1 |
690 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
691 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
692 SCHED_FEAT_SYNC_WAKEUPS * 1 |
693 SCHED_FEAT_HRTICK * 1 |
694 SCHED_FEAT_DOUBLE_TICK * 0 |
695 SCHED_FEAT_NORMALIZED_SLEEPER * 1;
697 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
700 * Number of tasks to iterate in a single balance run.
701 * Limited because this is done with IRQs disabled.
703 const_debug unsigned int sysctl_sched_nr_migrate = 32;
706 * period over which we measure -rt task cpu usage in us.
709 unsigned int sysctl_sched_rt_period = 1000000;
711 static __read_mostly int scheduler_running;
714 * part of the period that we allow rt tasks to run in us.
717 int sysctl_sched_rt_runtime = 950000;
719 static inline u64 global_rt_period(void)
721 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
724 static inline u64 global_rt_runtime(void)
726 if (sysctl_sched_rt_period < 0)
729 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
732 static const unsigned long long time_sync_thresh = 100000;
734 static DEFINE_PER_CPU(unsigned long long, time_offset);
735 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
738 * Global lock which we take every now and then to synchronize
739 * the CPUs time. This method is not warp-safe, but it's good
740 * enough to synchronize slowly diverging time sources and thus
741 * it's good enough for tracing:
743 static DEFINE_SPINLOCK(time_sync_lock);
744 static unsigned long long prev_global_time;
746 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
750 spin_lock_irqsave(&time_sync_lock, flags);
752 if (time < prev_global_time) {
753 per_cpu(time_offset, cpu) += prev_global_time - time;
754 time = prev_global_time;
756 prev_global_time = time;
759 spin_unlock_irqrestore(&time_sync_lock, flags);
764 static unsigned long long __cpu_clock(int cpu)
766 unsigned long long now;
771 * Only call sched_clock() if the scheduler has already been
772 * initialized (some code might call cpu_clock() very early):
774 if (unlikely(!scheduler_running))
777 local_irq_save(flags);
781 local_irq_restore(flags);
787 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
788 * clock constructed from sched_clock():
790 unsigned long long cpu_clock(int cpu)
792 unsigned long long prev_cpu_time, time, delta_time;
794 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
795 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
796 delta_time = time-prev_cpu_time;
798 if (unlikely(delta_time > time_sync_thresh))
799 time = __sync_cpu_clock(time, cpu);
803 EXPORT_SYMBOL_GPL(cpu_clock);
805 #ifndef prepare_arch_switch
806 # define prepare_arch_switch(next) do { } while (0)
808 #ifndef finish_arch_switch
809 # define finish_arch_switch(prev) do { } while (0)
812 static inline int task_current(struct rq *rq, struct task_struct *p)
814 return rq->curr == p;
817 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
818 static inline int task_running(struct rq *rq, struct task_struct *p)
820 return task_current(rq, p);
823 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
827 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
829 #ifdef CONFIG_DEBUG_SPINLOCK
830 /* this is a valid case when another task releases the spinlock */
831 rq->lock.owner = current;
834 * If we are tracking spinlock dependencies then we have to
835 * fix up the runqueue lock - which gets 'carried over' from
838 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
840 spin_unlock_irq(&rq->lock);
843 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
844 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
853 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
857 * We can optimise this out completely for !SMP, because the
858 * SMP rebalancing from interrupt is the only thing that cares
863 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
864 spin_unlock_irq(&rq->lock);
866 spin_unlock(&rq->lock);
870 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
874 * After ->oncpu is cleared, the task can be moved to a different CPU.
875 * We must ensure this doesn't happen until the switch is completely
881 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
885 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
888 * __task_rq_lock - lock the runqueue a given task resides on.
889 * Must be called interrupts disabled.
891 static inline struct rq *__task_rq_lock(struct task_struct *p)
895 struct rq *rq = task_rq(p);
896 spin_lock(&rq->lock);
897 if (likely(rq == task_rq(p)))
899 spin_unlock(&rq->lock);
904 * task_rq_lock - lock the runqueue a given task resides on and disable
905 * interrupts. Note the ordering: we can safely lookup the task_rq without
906 * explicitly disabling preemption.
908 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
914 local_irq_save(*flags);
916 spin_lock(&rq->lock);
917 if (likely(rq == task_rq(p)))
919 spin_unlock_irqrestore(&rq->lock, *flags);
923 static void __task_rq_unlock(struct rq *rq)
926 spin_unlock(&rq->lock);
929 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
932 spin_unlock_irqrestore(&rq->lock, *flags);
936 * this_rq_lock - lock this runqueue and disable interrupts.
938 static struct rq *this_rq_lock(void)
945 spin_lock(&rq->lock);
951 * We are going deep-idle (irqs are disabled):
953 void sched_clock_idle_sleep_event(void)
955 struct rq *rq = cpu_rq(smp_processor_id());
957 spin_lock(&rq->lock);
958 __update_rq_clock(rq);
959 spin_unlock(&rq->lock);
960 rq->clock_deep_idle_events++;
962 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
965 * We just idled delta nanoseconds (called with irqs disabled):
967 void sched_clock_idle_wakeup_event(u64 delta_ns)
969 struct rq *rq = cpu_rq(smp_processor_id());
970 u64 now = sched_clock();
972 rq->idle_clock += delta_ns;
974 * Override the previous timestamp and ignore all
975 * sched_clock() deltas that occured while we idled,
976 * and use the PM-provided delta_ns to advance the
979 spin_lock(&rq->lock);
980 rq->prev_clock_raw = now;
981 rq->clock += delta_ns;
982 spin_unlock(&rq->lock);
983 touch_softlockup_watchdog();
985 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
987 static void __resched_task(struct task_struct *p, int tif_bit);
989 static inline void resched_task(struct task_struct *p)
991 __resched_task(p, TIF_NEED_RESCHED);
994 #ifdef CONFIG_SCHED_HRTICK
996 * Use HR-timers to deliver accurate preemption points.
998 * Its all a bit involved since we cannot program an hrt while holding the
999 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1002 * When we get rescheduled we reprogram the hrtick_timer outside of the
1005 static inline void resched_hrt(struct task_struct *p)
1007 __resched_task(p, TIF_HRTICK_RESCHED);
1010 static inline void resched_rq(struct rq *rq)
1012 unsigned long flags;
1014 spin_lock_irqsave(&rq->lock, flags);
1015 resched_task(rq->curr);
1016 spin_unlock_irqrestore(&rq->lock, flags);
1020 HRTICK_SET, /* re-programm hrtick_timer */
1021 HRTICK_RESET, /* not a new slice */
1026 * - enabled by features
1027 * - hrtimer is actually high res
1029 static inline int hrtick_enabled(struct rq *rq)
1031 if (!sched_feat(HRTICK))
1033 return hrtimer_is_hres_active(&rq->hrtick_timer);
1037 * Called to set the hrtick timer state.
1039 * called with rq->lock held and irqs disabled
1041 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1043 assert_spin_locked(&rq->lock);
1046 * preempt at: now + delay
1049 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1051 * indicate we need to program the timer
1053 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1055 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1058 * New slices are called from the schedule path and don't need a
1059 * forced reschedule.
1062 resched_hrt(rq->curr);
1065 static void hrtick_clear(struct rq *rq)
1067 if (hrtimer_active(&rq->hrtick_timer))
1068 hrtimer_cancel(&rq->hrtick_timer);
1072 * Update the timer from the possible pending state.
1074 static void hrtick_set(struct rq *rq)
1078 unsigned long flags;
1080 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1082 spin_lock_irqsave(&rq->lock, flags);
1083 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1084 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1085 time = rq->hrtick_expire;
1086 clear_thread_flag(TIF_HRTICK_RESCHED);
1087 spin_unlock_irqrestore(&rq->lock, flags);
1090 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1091 if (reset && !hrtimer_active(&rq->hrtick_timer))
1098 * High-resolution timer tick.
1099 * Runs from hardirq context with interrupts disabled.
1101 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1103 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1105 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1107 spin_lock(&rq->lock);
1108 __update_rq_clock(rq);
1109 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1110 spin_unlock(&rq->lock);
1112 return HRTIMER_NORESTART;
1115 static inline void init_rq_hrtick(struct rq *rq)
1117 rq->hrtick_flags = 0;
1118 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1119 rq->hrtick_timer.function = hrtick;
1120 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1123 void hrtick_resched(void)
1126 unsigned long flags;
1128 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1131 local_irq_save(flags);
1132 rq = cpu_rq(smp_processor_id());
1134 local_irq_restore(flags);
1137 static inline void hrtick_clear(struct rq *rq)
1141 static inline void hrtick_set(struct rq *rq)
1145 static inline void init_rq_hrtick(struct rq *rq)
1149 void hrtick_resched(void)
1155 * resched_task - mark a task 'to be rescheduled now'.
1157 * On UP this means the setting of the need_resched flag, on SMP it
1158 * might also involve a cross-CPU call to trigger the scheduler on
1163 #ifndef tsk_is_polling
1164 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1167 static void __resched_task(struct task_struct *p, int tif_bit)
1171 assert_spin_locked(&task_rq(p)->lock);
1173 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1176 set_tsk_thread_flag(p, tif_bit);
1179 if (cpu == smp_processor_id())
1182 /* NEED_RESCHED must be visible before we test polling */
1184 if (!tsk_is_polling(p))
1185 smp_send_reschedule(cpu);
1188 static void resched_cpu(int cpu)
1190 struct rq *rq = cpu_rq(cpu);
1191 unsigned long flags;
1193 if (!spin_trylock_irqsave(&rq->lock, flags))
1195 resched_task(cpu_curr(cpu));
1196 spin_unlock_irqrestore(&rq->lock, flags);
1201 * When add_timer_on() enqueues a timer into the timer wheel of an
1202 * idle CPU then this timer might expire before the next timer event
1203 * which is scheduled to wake up that CPU. In case of a completely
1204 * idle system the next event might even be infinite time into the
1205 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1206 * leaves the inner idle loop so the newly added timer is taken into
1207 * account when the CPU goes back to idle and evaluates the timer
1208 * wheel for the next timer event.
1210 void wake_up_idle_cpu(int cpu)
1212 struct rq *rq = cpu_rq(cpu);
1214 if (cpu == smp_processor_id())
1218 * This is safe, as this function is called with the timer
1219 * wheel base lock of (cpu) held. When the CPU is on the way
1220 * to idle and has not yet set rq->curr to idle then it will
1221 * be serialized on the timer wheel base lock and take the new
1222 * timer into account automatically.
1224 if (rq->curr != rq->idle)
1228 * We can set TIF_RESCHED on the idle task of the other CPU
1229 * lockless. The worst case is that the other CPU runs the
1230 * idle task through an additional NOOP schedule()
1232 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1234 /* NEED_RESCHED must be visible before we test polling */
1236 if (!tsk_is_polling(rq->idle))
1237 smp_send_reschedule(cpu);
1242 static void __resched_task(struct task_struct *p, int tif_bit)
1244 assert_spin_locked(&task_rq(p)->lock);
1245 set_tsk_thread_flag(p, tif_bit);
1249 #if BITS_PER_LONG == 32
1250 # define WMULT_CONST (~0UL)
1252 # define WMULT_CONST (1UL << 32)
1255 #define WMULT_SHIFT 32
1258 * Shift right and round:
1260 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1262 static unsigned long
1263 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1264 struct load_weight *lw)
1268 if (unlikely(!lw->inv_weight))
1269 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1271 tmp = (u64)delta_exec * weight;
1273 * Check whether we'd overflow the 64-bit multiplication:
1275 if (unlikely(tmp > WMULT_CONST))
1276 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1279 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1281 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1284 static inline unsigned long
1285 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1287 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1290 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1296 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1303 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1304 * of tasks with abnormal "nice" values across CPUs the contribution that
1305 * each task makes to its run queue's load is weighted according to its
1306 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1307 * scaled version of the new time slice allocation that they receive on time
1311 #define WEIGHT_IDLEPRIO 2
1312 #define WMULT_IDLEPRIO (1 << 31)
1315 * Nice levels are multiplicative, with a gentle 10% change for every
1316 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1317 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1318 * that remained on nice 0.
1320 * The "10% effect" is relative and cumulative: from _any_ nice level,
1321 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1322 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1323 * If a task goes up by ~10% and another task goes down by ~10% then
1324 * the relative distance between them is ~25%.)
1326 static const int prio_to_weight[40] = {
1327 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1328 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1329 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1330 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1331 /* 0 */ 1024, 820, 655, 526, 423,
1332 /* 5 */ 335, 272, 215, 172, 137,
1333 /* 10 */ 110, 87, 70, 56, 45,
1334 /* 15 */ 36, 29, 23, 18, 15,
1338 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1340 * In cases where the weight does not change often, we can use the
1341 * precalculated inverse to speed up arithmetics by turning divisions
1342 * into multiplications:
1344 static const u32 prio_to_wmult[40] = {
1345 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1346 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1347 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1348 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1349 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1350 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1351 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1352 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1355 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1358 * runqueue iterator, to support SMP load-balancing between different
1359 * scheduling classes, without having to expose their internal data
1360 * structures to the load-balancing proper:
1362 struct rq_iterator {
1364 struct task_struct *(*start)(void *);
1365 struct task_struct *(*next)(void *);
1369 static unsigned long
1370 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1371 unsigned long max_load_move, struct sched_domain *sd,
1372 enum cpu_idle_type idle, int *all_pinned,
1373 int *this_best_prio, struct rq_iterator *iterator);
1376 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1377 struct sched_domain *sd, enum cpu_idle_type idle,
1378 struct rq_iterator *iterator);
1381 #ifdef CONFIG_CGROUP_CPUACCT
1382 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1384 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1388 static unsigned long source_load(int cpu, int type);
1389 static unsigned long target_load(int cpu, int type);
1390 static unsigned long cpu_avg_load_per_task(int cpu);
1391 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1392 #endif /* CONFIG_SMP */
1394 #include "sched_stats.h"
1395 #include "sched_idletask.c"
1396 #include "sched_fair.c"
1397 #include "sched_rt.c"
1398 #ifdef CONFIG_SCHED_DEBUG
1399 # include "sched_debug.c"
1402 #define sched_class_highest (&rt_sched_class)
1404 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1406 update_load_add(&rq->load, p->se.load.weight);
1409 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1411 update_load_sub(&rq->load, p->se.load.weight);
1414 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1420 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1426 static void set_load_weight(struct task_struct *p)
1428 if (task_has_rt_policy(p)) {
1429 p->se.load.weight = prio_to_weight[0] * 2;
1430 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1435 * SCHED_IDLE tasks get minimal weight:
1437 if (p->policy == SCHED_IDLE) {
1438 p->se.load.weight = WEIGHT_IDLEPRIO;
1439 p->se.load.inv_weight = WMULT_IDLEPRIO;
1443 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1444 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1447 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1449 sched_info_queued(p);
1450 p->sched_class->enqueue_task(rq, p, wakeup);
1454 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1456 p->sched_class->dequeue_task(rq, p, sleep);
1461 * __normal_prio - return the priority that is based on the static prio
1463 static inline int __normal_prio(struct task_struct *p)
1465 return p->static_prio;
1469 * Calculate the expected normal priority: i.e. priority
1470 * without taking RT-inheritance into account. Might be
1471 * boosted by interactivity modifiers. Changes upon fork,
1472 * setprio syscalls, and whenever the interactivity
1473 * estimator recalculates.
1475 static inline int normal_prio(struct task_struct *p)
1479 if (task_has_rt_policy(p))
1480 prio = MAX_RT_PRIO-1 - p->rt_priority;
1482 prio = __normal_prio(p);
1487 * Calculate the current priority, i.e. the priority
1488 * taken into account by the scheduler. This value might
1489 * be boosted by RT tasks, or might be boosted by
1490 * interactivity modifiers. Will be RT if the task got
1491 * RT-boosted. If not then it returns p->normal_prio.
1493 static int effective_prio(struct task_struct *p)
1495 p->normal_prio = normal_prio(p);
1497 * If we are RT tasks or we were boosted to RT priority,
1498 * keep the priority unchanged. Otherwise, update priority
1499 * to the normal priority:
1501 if (!rt_prio(p->prio))
1502 return p->normal_prio;
1507 * activate_task - move a task to the runqueue.
1509 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1511 if (task_contributes_to_load(p))
1512 rq->nr_uninterruptible--;
1514 enqueue_task(rq, p, wakeup);
1515 inc_nr_running(p, rq);
1519 * deactivate_task - remove a task from the runqueue.
1521 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1523 if (task_contributes_to_load(p))
1524 rq->nr_uninterruptible++;
1526 dequeue_task(rq, p, sleep);
1527 dec_nr_running(p, rq);
1531 * task_curr - is this task currently executing on a CPU?
1532 * @p: the task in question.
1534 inline int task_curr(const struct task_struct *p)
1536 return cpu_curr(task_cpu(p)) == p;
1539 /* Used instead of source_load when we know the type == 0 */
1540 unsigned long weighted_cpuload(const int cpu)
1542 return cpu_rq(cpu)->load.weight;
1545 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1547 set_task_rq(p, cpu);
1550 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1551 * successfuly executed on another CPU. We must ensure that updates of
1552 * per-task data have been completed by this moment.
1555 task_thread_info(p)->cpu = cpu;
1559 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1560 const struct sched_class *prev_class,
1561 int oldprio, int running)
1563 if (prev_class != p->sched_class) {
1564 if (prev_class->switched_from)
1565 prev_class->switched_from(rq, p, running);
1566 p->sched_class->switched_to(rq, p, running);
1568 p->sched_class->prio_changed(rq, p, oldprio, running);
1574 * Is this task likely cache-hot:
1577 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1582 * Buddy candidates are cache hot:
1584 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1587 if (p->sched_class != &fair_sched_class)
1590 if (sysctl_sched_migration_cost == -1)
1592 if (sysctl_sched_migration_cost == 0)
1595 delta = now - p->se.exec_start;
1597 return delta < (s64)sysctl_sched_migration_cost;
1601 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1603 int old_cpu = task_cpu(p);
1604 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1605 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1606 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1609 clock_offset = old_rq->clock - new_rq->clock;
1611 #ifdef CONFIG_SCHEDSTATS
1612 if (p->se.wait_start)
1613 p->se.wait_start -= clock_offset;
1614 if (p->se.sleep_start)
1615 p->se.sleep_start -= clock_offset;
1616 if (p->se.block_start)
1617 p->se.block_start -= clock_offset;
1618 if (old_cpu != new_cpu) {
1619 schedstat_inc(p, se.nr_migrations);
1620 if (task_hot(p, old_rq->clock, NULL))
1621 schedstat_inc(p, se.nr_forced2_migrations);
1624 p->se.vruntime -= old_cfsrq->min_vruntime -
1625 new_cfsrq->min_vruntime;
1627 __set_task_cpu(p, new_cpu);
1630 struct migration_req {
1631 struct list_head list;
1633 struct task_struct *task;
1636 struct completion done;
1640 * The task's runqueue lock must be held.
1641 * Returns true if you have to wait for migration thread.
1644 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1646 struct rq *rq = task_rq(p);
1649 * If the task is not on a runqueue (and not running), then
1650 * it is sufficient to simply update the task's cpu field.
1652 if (!p->se.on_rq && !task_running(rq, p)) {
1653 set_task_cpu(p, dest_cpu);
1657 init_completion(&req->done);
1659 req->dest_cpu = dest_cpu;
1660 list_add(&req->list, &rq->migration_queue);
1666 * wait_task_inactive - wait for a thread to unschedule.
1668 * The caller must ensure that the task *will* unschedule sometime soon,
1669 * else this function might spin for a *long* time. This function can't
1670 * be called with interrupts off, or it may introduce deadlock with
1671 * smp_call_function() if an IPI is sent by the same process we are
1672 * waiting to become inactive.
1674 void wait_task_inactive(struct task_struct *p)
1676 unsigned long flags;
1682 * We do the initial early heuristics without holding
1683 * any task-queue locks at all. We'll only try to get
1684 * the runqueue lock when things look like they will
1690 * If the task is actively running on another CPU
1691 * still, just relax and busy-wait without holding
1694 * NOTE! Since we don't hold any locks, it's not
1695 * even sure that "rq" stays as the right runqueue!
1696 * But we don't care, since "task_running()" will
1697 * return false if the runqueue has changed and p
1698 * is actually now running somewhere else!
1700 while (task_running(rq, p))
1704 * Ok, time to look more closely! We need the rq
1705 * lock now, to be *sure*. If we're wrong, we'll
1706 * just go back and repeat.
1708 rq = task_rq_lock(p, &flags);
1709 running = task_running(rq, p);
1710 on_rq = p->se.on_rq;
1711 task_rq_unlock(rq, &flags);
1714 * Was it really running after all now that we
1715 * checked with the proper locks actually held?
1717 * Oops. Go back and try again..
1719 if (unlikely(running)) {
1725 * It's not enough that it's not actively running,
1726 * it must be off the runqueue _entirely_, and not
1729 * So if it wa still runnable (but just not actively
1730 * running right now), it's preempted, and we should
1731 * yield - it could be a while.
1733 if (unlikely(on_rq)) {
1734 schedule_timeout_uninterruptible(1);
1739 * Ahh, all good. It wasn't running, and it wasn't
1740 * runnable, which means that it will never become
1741 * running in the future either. We're all done!
1748 * kick_process - kick a running thread to enter/exit the kernel
1749 * @p: the to-be-kicked thread
1751 * Cause a process which is running on another CPU to enter
1752 * kernel-mode, without any delay. (to get signals handled.)
1754 * NOTE: this function doesnt have to take the runqueue lock,
1755 * because all it wants to ensure is that the remote task enters
1756 * the kernel. If the IPI races and the task has been migrated
1757 * to another CPU then no harm is done and the purpose has been
1760 void kick_process(struct task_struct *p)
1766 if ((cpu != smp_processor_id()) && task_curr(p))
1767 smp_send_reschedule(cpu);
1772 * Return a low guess at the load of a migration-source cpu weighted
1773 * according to the scheduling class and "nice" value.
1775 * We want to under-estimate the load of migration sources, to
1776 * balance conservatively.
1778 static unsigned long source_load(int cpu, int type)
1780 struct rq *rq = cpu_rq(cpu);
1781 unsigned long total = weighted_cpuload(cpu);
1786 return min(rq->cpu_load[type-1], total);
1790 * Return a high guess at the load of a migration-target cpu weighted
1791 * according to the scheduling class and "nice" value.
1793 static unsigned long target_load(int cpu, int type)
1795 struct rq *rq = cpu_rq(cpu);
1796 unsigned long total = weighted_cpuload(cpu);
1801 return max(rq->cpu_load[type-1], total);
1805 * Return the average load per task on the cpu's run queue
1807 static unsigned long cpu_avg_load_per_task(int cpu)
1809 struct rq *rq = cpu_rq(cpu);
1810 unsigned long total = weighted_cpuload(cpu);
1811 unsigned long n = rq->nr_running;
1813 return n ? total / n : SCHED_LOAD_SCALE;
1817 * find_idlest_group finds and returns the least busy CPU group within the
1820 static struct sched_group *
1821 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1823 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1824 unsigned long min_load = ULONG_MAX, this_load = 0;
1825 int load_idx = sd->forkexec_idx;
1826 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1829 unsigned long load, avg_load;
1833 /* Skip over this group if it has no CPUs allowed */
1834 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1837 local_group = cpu_isset(this_cpu, group->cpumask);
1839 /* Tally up the load of all CPUs in the group */
1842 for_each_cpu_mask(i, group->cpumask) {
1843 /* Bias balancing toward cpus of our domain */
1845 load = source_load(i, load_idx);
1847 load = target_load(i, load_idx);
1852 /* Adjust by relative CPU power of the group */
1853 avg_load = sg_div_cpu_power(group,
1854 avg_load * SCHED_LOAD_SCALE);
1857 this_load = avg_load;
1859 } else if (avg_load < min_load) {
1860 min_load = avg_load;
1863 } while (group = group->next, group != sd->groups);
1865 if (!idlest || 100*this_load < imbalance*min_load)
1871 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1874 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1877 unsigned long load, min_load = ULONG_MAX;
1881 /* Traverse only the allowed CPUs */
1882 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1884 for_each_cpu_mask(i, *tmp) {
1885 load = weighted_cpuload(i);
1887 if (load < min_load || (load == min_load && i == this_cpu)) {
1897 * sched_balance_self: balance the current task (running on cpu) in domains
1898 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1901 * Balance, ie. select the least loaded group.
1903 * Returns the target CPU number, or the same CPU if no balancing is needed.
1905 * preempt must be disabled.
1907 static int sched_balance_self(int cpu, int flag)
1909 struct task_struct *t = current;
1910 struct sched_domain *tmp, *sd = NULL;
1912 for_each_domain(cpu, tmp) {
1914 * If power savings logic is enabled for a domain, stop there.
1916 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1918 if (tmp->flags & flag)
1923 cpumask_t span, tmpmask;
1924 struct sched_group *group;
1925 int new_cpu, weight;
1927 if (!(sd->flags & flag)) {
1933 group = find_idlest_group(sd, t, cpu);
1939 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1940 if (new_cpu == -1 || new_cpu == cpu) {
1941 /* Now try balancing at a lower domain level of cpu */
1946 /* Now try balancing at a lower domain level of new_cpu */
1949 weight = cpus_weight(span);
1950 for_each_domain(cpu, tmp) {
1951 if (weight <= cpus_weight(tmp->span))
1953 if (tmp->flags & flag)
1956 /* while loop will break here if sd == NULL */
1962 #endif /* CONFIG_SMP */
1965 * try_to_wake_up - wake up a thread
1966 * @p: the to-be-woken-up thread
1967 * @state: the mask of task states that can be woken
1968 * @sync: do a synchronous wakeup?
1970 * Put it on the run-queue if it's not already there. The "current"
1971 * thread is always on the run-queue (except when the actual
1972 * re-schedule is in progress), and as such you're allowed to do
1973 * the simpler "current->state = TASK_RUNNING" to mark yourself
1974 * runnable without the overhead of this.
1976 * returns failure only if the task is already active.
1978 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1980 int cpu, orig_cpu, this_cpu, success = 0;
1981 unsigned long flags;
1985 if (!sched_feat(SYNC_WAKEUPS))
1989 rq = task_rq_lock(p, &flags);
1990 old_state = p->state;
1991 if (!(old_state & state))
1999 this_cpu = smp_processor_id();
2002 if (unlikely(task_running(rq, p)))
2005 cpu = p->sched_class->select_task_rq(p, sync);
2006 if (cpu != orig_cpu) {
2007 set_task_cpu(p, cpu);
2008 task_rq_unlock(rq, &flags);
2009 /* might preempt at this point */
2010 rq = task_rq_lock(p, &flags);
2011 old_state = p->state;
2012 if (!(old_state & state))
2017 this_cpu = smp_processor_id();
2021 #ifdef CONFIG_SCHEDSTATS
2022 schedstat_inc(rq, ttwu_count);
2023 if (cpu == this_cpu)
2024 schedstat_inc(rq, ttwu_local);
2026 struct sched_domain *sd;
2027 for_each_domain(this_cpu, sd) {
2028 if (cpu_isset(cpu, sd->span)) {
2029 schedstat_inc(sd, ttwu_wake_remote);
2037 #endif /* CONFIG_SMP */
2038 schedstat_inc(p, se.nr_wakeups);
2040 schedstat_inc(p, se.nr_wakeups_sync);
2041 if (orig_cpu != cpu)
2042 schedstat_inc(p, se.nr_wakeups_migrate);
2043 if (cpu == this_cpu)
2044 schedstat_inc(p, se.nr_wakeups_local);
2046 schedstat_inc(p, se.nr_wakeups_remote);
2047 update_rq_clock(rq);
2048 activate_task(rq, p, 1);
2052 check_preempt_curr(rq, p);
2054 p->state = TASK_RUNNING;
2056 if (p->sched_class->task_wake_up)
2057 p->sched_class->task_wake_up(rq, p);
2060 task_rq_unlock(rq, &flags);
2065 int wake_up_process(struct task_struct *p)
2067 return try_to_wake_up(p, TASK_ALL, 0);
2069 EXPORT_SYMBOL(wake_up_process);
2071 int wake_up_state(struct task_struct *p, unsigned int state)
2073 return try_to_wake_up(p, state, 0);
2077 * Perform scheduler related setup for a newly forked process p.
2078 * p is forked by current.
2080 * __sched_fork() is basic setup used by init_idle() too:
2082 static void __sched_fork(struct task_struct *p)
2084 p->se.exec_start = 0;
2085 p->se.sum_exec_runtime = 0;
2086 p->se.prev_sum_exec_runtime = 0;
2087 p->se.last_wakeup = 0;
2088 p->se.avg_overlap = 0;
2090 #ifdef CONFIG_SCHEDSTATS
2091 p->se.wait_start = 0;
2092 p->se.sum_sleep_runtime = 0;
2093 p->se.sleep_start = 0;
2094 p->se.block_start = 0;
2095 p->se.sleep_max = 0;
2096 p->se.block_max = 0;
2098 p->se.slice_max = 0;
2102 INIT_LIST_HEAD(&p->rt.run_list);
2105 #ifdef CONFIG_PREEMPT_NOTIFIERS
2106 INIT_HLIST_HEAD(&p->preempt_notifiers);
2110 * We mark the process as running here, but have not actually
2111 * inserted it onto the runqueue yet. This guarantees that
2112 * nobody will actually run it, and a signal or other external
2113 * event cannot wake it up and insert it on the runqueue either.
2115 p->state = TASK_RUNNING;
2119 * fork()/clone()-time setup:
2121 void sched_fork(struct task_struct *p, int clone_flags)
2123 int cpu = get_cpu();
2128 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2130 set_task_cpu(p, cpu);
2133 * Make sure we do not leak PI boosting priority to the child:
2135 p->prio = current->normal_prio;
2136 if (!rt_prio(p->prio))
2137 p->sched_class = &fair_sched_class;
2139 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2140 if (likely(sched_info_on()))
2141 memset(&p->sched_info, 0, sizeof(p->sched_info));
2143 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2146 #ifdef CONFIG_PREEMPT
2147 /* Want to start with kernel preemption disabled. */
2148 task_thread_info(p)->preempt_count = 1;
2154 * wake_up_new_task - wake up a newly created task for the first time.
2156 * This function will do some initial scheduler statistics housekeeping
2157 * that must be done for every newly created context, then puts the task
2158 * on the runqueue and wakes it.
2160 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2162 unsigned long flags;
2165 rq = task_rq_lock(p, &flags);
2166 BUG_ON(p->state != TASK_RUNNING);
2167 update_rq_clock(rq);
2169 p->prio = effective_prio(p);
2171 if (!p->sched_class->task_new || !current->se.on_rq) {
2172 activate_task(rq, p, 0);
2175 * Let the scheduling class do new task startup
2176 * management (if any):
2178 p->sched_class->task_new(rq, p);
2179 inc_nr_running(p, rq);
2181 check_preempt_curr(rq, p);
2183 if (p->sched_class->task_wake_up)
2184 p->sched_class->task_wake_up(rq, p);
2186 task_rq_unlock(rq, &flags);
2189 #ifdef CONFIG_PREEMPT_NOTIFIERS
2192 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2193 * @notifier: notifier struct to register
2195 void preempt_notifier_register(struct preempt_notifier *notifier)
2197 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2199 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2202 * preempt_notifier_unregister - no longer interested in preemption notifications
2203 * @notifier: notifier struct to unregister
2205 * This is safe to call from within a preemption notifier.
2207 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2209 hlist_del(¬ifier->link);
2211 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2213 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2215 struct preempt_notifier *notifier;
2216 struct hlist_node *node;
2218 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2219 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2223 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2224 struct task_struct *next)
2226 struct preempt_notifier *notifier;
2227 struct hlist_node *node;
2229 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2230 notifier->ops->sched_out(notifier, next);
2235 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2240 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2241 struct task_struct *next)
2248 * prepare_task_switch - prepare to switch tasks
2249 * @rq: the runqueue preparing to switch
2250 * @prev: the current task that is being switched out
2251 * @next: the task we are going to switch to.
2253 * This is called with the rq lock held and interrupts off. It must
2254 * be paired with a subsequent finish_task_switch after the context
2257 * prepare_task_switch sets up locking and calls architecture specific
2261 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2262 struct task_struct *next)
2264 fire_sched_out_preempt_notifiers(prev, next);
2265 prepare_lock_switch(rq, next);
2266 prepare_arch_switch(next);
2270 * finish_task_switch - clean up after a task-switch
2271 * @rq: runqueue associated with task-switch
2272 * @prev: the thread we just switched away from.
2274 * finish_task_switch must be called after the context switch, paired
2275 * with a prepare_task_switch call before the context switch.
2276 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2277 * and do any other architecture-specific cleanup actions.
2279 * Note that we may have delayed dropping an mm in context_switch(). If
2280 * so, we finish that here outside of the runqueue lock. (Doing it
2281 * with the lock held can cause deadlocks; see schedule() for
2284 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2285 __releases(rq->lock)
2287 struct mm_struct *mm = rq->prev_mm;
2293 * A task struct has one reference for the use as "current".
2294 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2295 * schedule one last time. The schedule call will never return, and
2296 * the scheduled task must drop that reference.
2297 * The test for TASK_DEAD must occur while the runqueue locks are
2298 * still held, otherwise prev could be scheduled on another cpu, die
2299 * there before we look at prev->state, and then the reference would
2301 * Manfred Spraul <manfred@colorfullife.com>
2303 prev_state = prev->state;
2304 finish_arch_switch(prev);
2305 finish_lock_switch(rq, prev);
2307 if (current->sched_class->post_schedule)
2308 current->sched_class->post_schedule(rq);
2311 fire_sched_in_preempt_notifiers(current);
2314 if (unlikely(prev_state == TASK_DEAD)) {
2316 * Remove function-return probe instances associated with this
2317 * task and put them back on the free list.
2319 kprobe_flush_task(prev);
2320 put_task_struct(prev);
2325 * schedule_tail - first thing a freshly forked thread must call.
2326 * @prev: the thread we just switched away from.
2328 asmlinkage void schedule_tail(struct task_struct *prev)
2329 __releases(rq->lock)
2331 struct rq *rq = this_rq();
2333 finish_task_switch(rq, prev);
2334 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2335 /* In this case, finish_task_switch does not reenable preemption */
2338 if (current->set_child_tid)
2339 put_user(task_pid_vnr(current), current->set_child_tid);
2343 * context_switch - switch to the new MM and the new
2344 * thread's register state.
2347 context_switch(struct rq *rq, struct task_struct *prev,
2348 struct task_struct *next)
2350 struct mm_struct *mm, *oldmm;
2352 prepare_task_switch(rq, prev, next);
2354 oldmm = prev->active_mm;
2356 * For paravirt, this is coupled with an exit in switch_to to
2357 * combine the page table reload and the switch backend into
2360 arch_enter_lazy_cpu_mode();
2362 if (unlikely(!mm)) {
2363 next->active_mm = oldmm;
2364 atomic_inc(&oldmm->mm_count);
2365 enter_lazy_tlb(oldmm, next);
2367 switch_mm(oldmm, mm, next);
2369 if (unlikely(!prev->mm)) {
2370 prev->active_mm = NULL;
2371 rq->prev_mm = oldmm;
2374 * Since the runqueue lock will be released by the next
2375 * task (which is an invalid locking op but in the case
2376 * of the scheduler it's an obvious special-case), so we
2377 * do an early lockdep release here:
2379 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2380 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2383 /* Here we just switch the register state and the stack. */
2384 switch_to(prev, next, prev);
2388 * this_rq must be evaluated again because prev may have moved
2389 * CPUs since it called schedule(), thus the 'rq' on its stack
2390 * frame will be invalid.
2392 finish_task_switch(this_rq(), prev);
2396 * nr_running, nr_uninterruptible and nr_context_switches:
2398 * externally visible scheduler statistics: current number of runnable
2399 * threads, current number of uninterruptible-sleeping threads, total
2400 * number of context switches performed since bootup.
2402 unsigned long nr_running(void)
2404 unsigned long i, sum = 0;
2406 for_each_online_cpu(i)
2407 sum += cpu_rq(i)->nr_running;
2412 unsigned long nr_uninterruptible(void)
2414 unsigned long i, sum = 0;
2416 for_each_possible_cpu(i)
2417 sum += cpu_rq(i)->nr_uninterruptible;
2420 * Since we read the counters lockless, it might be slightly
2421 * inaccurate. Do not allow it to go below zero though:
2423 if (unlikely((long)sum < 0))
2429 unsigned long long nr_context_switches(void)
2432 unsigned long long sum = 0;
2434 for_each_possible_cpu(i)
2435 sum += cpu_rq(i)->nr_switches;
2440 unsigned long nr_iowait(void)
2442 unsigned long i, sum = 0;
2444 for_each_possible_cpu(i)
2445 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2450 unsigned long nr_active(void)
2452 unsigned long i, running = 0, uninterruptible = 0;
2454 for_each_online_cpu(i) {
2455 running += cpu_rq(i)->nr_running;
2456 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2459 if (unlikely((long)uninterruptible < 0))
2460 uninterruptible = 0;
2462 return running + uninterruptible;
2466 * Update rq->cpu_load[] statistics. This function is usually called every
2467 * scheduler tick (TICK_NSEC).
2469 static void update_cpu_load(struct rq *this_rq)
2471 unsigned long this_load = this_rq->load.weight;
2474 this_rq->nr_load_updates++;
2476 /* Update our load: */
2477 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2478 unsigned long old_load, new_load;
2480 /* scale is effectively 1 << i now, and >> i divides by scale */
2482 old_load = this_rq->cpu_load[i];
2483 new_load = this_load;
2485 * Round up the averaging division if load is increasing. This
2486 * prevents us from getting stuck on 9 if the load is 10, for
2489 if (new_load > old_load)
2490 new_load += scale-1;
2491 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2498 * double_rq_lock - safely lock two runqueues
2500 * Note this does not disable interrupts like task_rq_lock,
2501 * you need to do so manually before calling.
2503 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2504 __acquires(rq1->lock)
2505 __acquires(rq2->lock)
2507 BUG_ON(!irqs_disabled());
2509 spin_lock(&rq1->lock);
2510 __acquire(rq2->lock); /* Fake it out ;) */
2513 spin_lock(&rq1->lock);
2514 spin_lock(&rq2->lock);
2516 spin_lock(&rq2->lock);
2517 spin_lock(&rq1->lock);
2520 update_rq_clock(rq1);
2521 update_rq_clock(rq2);
2525 * double_rq_unlock - safely unlock two runqueues
2527 * Note this does not restore interrupts like task_rq_unlock,
2528 * you need to do so manually after calling.
2530 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2531 __releases(rq1->lock)
2532 __releases(rq2->lock)
2534 spin_unlock(&rq1->lock);
2536 spin_unlock(&rq2->lock);
2538 __release(rq2->lock);
2542 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2544 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2545 __releases(this_rq->lock)
2546 __acquires(busiest->lock)
2547 __acquires(this_rq->lock)
2551 if (unlikely(!irqs_disabled())) {
2552 /* printk() doesn't work good under rq->lock */
2553 spin_unlock(&this_rq->lock);
2556 if (unlikely(!spin_trylock(&busiest->lock))) {
2557 if (busiest < this_rq) {
2558 spin_unlock(&this_rq->lock);
2559 spin_lock(&busiest->lock);
2560 spin_lock(&this_rq->lock);
2563 spin_lock(&busiest->lock);
2569 * If dest_cpu is allowed for this process, migrate the task to it.
2570 * This is accomplished by forcing the cpu_allowed mask to only
2571 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2572 * the cpu_allowed mask is restored.
2574 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2576 struct migration_req req;
2577 unsigned long flags;
2580 rq = task_rq_lock(p, &flags);
2581 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2582 || unlikely(cpu_is_offline(dest_cpu)))
2585 /* force the process onto the specified CPU */
2586 if (migrate_task(p, dest_cpu, &req)) {
2587 /* Need to wait for migration thread (might exit: take ref). */
2588 struct task_struct *mt = rq->migration_thread;
2590 get_task_struct(mt);
2591 task_rq_unlock(rq, &flags);
2592 wake_up_process(mt);
2593 put_task_struct(mt);
2594 wait_for_completion(&req.done);
2599 task_rq_unlock(rq, &flags);
2603 * sched_exec - execve() is a valuable balancing opportunity, because at
2604 * this point the task has the smallest effective memory and cache footprint.
2606 void sched_exec(void)
2608 int new_cpu, this_cpu = get_cpu();
2609 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2611 if (new_cpu != this_cpu)
2612 sched_migrate_task(current, new_cpu);
2616 * pull_task - move a task from a remote runqueue to the local runqueue.
2617 * Both runqueues must be locked.
2619 static void pull_task(struct rq *src_rq, struct task_struct *p,
2620 struct rq *this_rq, int this_cpu)
2622 deactivate_task(src_rq, p, 0);
2623 set_task_cpu(p, this_cpu);
2624 activate_task(this_rq, p, 0);
2626 * Note that idle threads have a prio of MAX_PRIO, for this test
2627 * to be always true for them.
2629 check_preempt_curr(this_rq, p);
2633 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2636 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2637 struct sched_domain *sd, enum cpu_idle_type idle,
2641 * We do not migrate tasks that are:
2642 * 1) running (obviously), or
2643 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2644 * 3) are cache-hot on their current CPU.
2646 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2647 schedstat_inc(p, se.nr_failed_migrations_affine);
2652 if (task_running(rq, p)) {
2653 schedstat_inc(p, se.nr_failed_migrations_running);
2658 * Aggressive migration if:
2659 * 1) task is cache cold, or
2660 * 2) too many balance attempts have failed.
2663 if (!task_hot(p, rq->clock, sd) ||
2664 sd->nr_balance_failed > sd->cache_nice_tries) {
2665 #ifdef CONFIG_SCHEDSTATS
2666 if (task_hot(p, rq->clock, sd)) {
2667 schedstat_inc(sd, lb_hot_gained[idle]);
2668 schedstat_inc(p, se.nr_forced_migrations);
2674 if (task_hot(p, rq->clock, sd)) {
2675 schedstat_inc(p, se.nr_failed_migrations_hot);
2681 static unsigned long
2682 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2683 unsigned long max_load_move, struct sched_domain *sd,
2684 enum cpu_idle_type idle, int *all_pinned,
2685 int *this_best_prio, struct rq_iterator *iterator)
2687 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2688 struct task_struct *p;
2689 long rem_load_move = max_load_move;
2691 if (max_load_move == 0)
2697 * Start the load-balancing iterator:
2699 p = iterator->start(iterator->arg);
2701 if (!p || loops++ > sysctl_sched_nr_migrate)
2704 * To help distribute high priority tasks across CPUs we don't
2705 * skip a task if it will be the highest priority task (i.e. smallest
2706 * prio value) on its new queue regardless of its load weight
2708 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2709 SCHED_LOAD_SCALE_FUZZ;
2710 if ((skip_for_load && p->prio >= *this_best_prio) ||
2711 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2712 p = iterator->next(iterator->arg);
2716 pull_task(busiest, p, this_rq, this_cpu);
2718 rem_load_move -= p->se.load.weight;
2721 * We only want to steal up to the prescribed amount of weighted load.
2723 if (rem_load_move > 0) {
2724 if (p->prio < *this_best_prio)
2725 *this_best_prio = p->prio;
2726 p = iterator->next(iterator->arg);
2731 * Right now, this is one of only two places pull_task() is called,
2732 * so we can safely collect pull_task() stats here rather than
2733 * inside pull_task().
2735 schedstat_add(sd, lb_gained[idle], pulled);
2738 *all_pinned = pinned;
2740 return max_load_move - rem_load_move;
2744 * move_tasks tries to move up to max_load_move weighted load from busiest to
2745 * this_rq, as part of a balancing operation within domain "sd".
2746 * Returns 1 if successful and 0 otherwise.
2748 * Called with both runqueues locked.
2750 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2751 unsigned long max_load_move,
2752 struct sched_domain *sd, enum cpu_idle_type idle,
2755 const struct sched_class *class = sched_class_highest;
2756 unsigned long total_load_moved = 0;
2757 int this_best_prio = this_rq->curr->prio;
2761 class->load_balance(this_rq, this_cpu, busiest,
2762 max_load_move - total_load_moved,
2763 sd, idle, all_pinned, &this_best_prio);
2764 class = class->next;
2765 } while (class && max_load_move > total_load_moved);
2767 return total_load_moved > 0;
2771 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2772 struct sched_domain *sd, enum cpu_idle_type idle,
2773 struct rq_iterator *iterator)
2775 struct task_struct *p = iterator->start(iterator->arg);
2779 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2780 pull_task(busiest, p, this_rq, this_cpu);
2782 * Right now, this is only the second place pull_task()
2783 * is called, so we can safely collect pull_task()
2784 * stats here rather than inside pull_task().
2786 schedstat_inc(sd, lb_gained[idle]);
2790 p = iterator->next(iterator->arg);
2797 * move_one_task tries to move exactly one task from busiest to this_rq, as
2798 * part of active balancing operations within "domain".
2799 * Returns 1 if successful and 0 otherwise.
2801 * Called with both runqueues locked.
2803 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2804 struct sched_domain *sd, enum cpu_idle_type idle)
2806 const struct sched_class *class;
2808 for (class = sched_class_highest; class; class = class->next)
2809 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2816 * find_busiest_group finds and returns the busiest CPU group within the
2817 * domain. It calculates and returns the amount of weighted load which
2818 * should be moved to restore balance via the imbalance parameter.
2820 static struct sched_group *
2821 find_busiest_group(struct sched_domain *sd, int this_cpu,
2822 unsigned long *imbalance, enum cpu_idle_type idle,
2823 int *sd_idle, const cpumask_t *cpus, int *balance)
2825 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2826 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2827 unsigned long max_pull;
2828 unsigned long busiest_load_per_task, busiest_nr_running;
2829 unsigned long this_load_per_task, this_nr_running;
2830 int load_idx, group_imb = 0;
2831 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2832 int power_savings_balance = 1;
2833 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2834 unsigned long min_nr_running = ULONG_MAX;
2835 struct sched_group *group_min = NULL, *group_leader = NULL;
2838 max_load = this_load = total_load = total_pwr = 0;
2839 busiest_load_per_task = busiest_nr_running = 0;
2840 this_load_per_task = this_nr_running = 0;
2841 if (idle == CPU_NOT_IDLE)
2842 load_idx = sd->busy_idx;
2843 else if (idle == CPU_NEWLY_IDLE)
2844 load_idx = sd->newidle_idx;
2846 load_idx = sd->idle_idx;
2849 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2852 int __group_imb = 0;
2853 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2854 unsigned long sum_nr_running, sum_weighted_load;
2856 local_group = cpu_isset(this_cpu, group->cpumask);
2859 balance_cpu = first_cpu(group->cpumask);
2861 /* Tally up the load of all CPUs in the group */
2862 sum_weighted_load = sum_nr_running = avg_load = 0;
2864 min_cpu_load = ~0UL;
2866 for_each_cpu_mask(i, group->cpumask) {
2869 if (!cpu_isset(i, *cpus))
2874 if (*sd_idle && rq->nr_running)
2877 /* Bias balancing toward cpus of our domain */
2879 if (idle_cpu(i) && !first_idle_cpu) {
2884 load = target_load(i, load_idx);
2886 load = source_load(i, load_idx);
2887 if (load > max_cpu_load)
2888 max_cpu_load = load;
2889 if (min_cpu_load > load)
2890 min_cpu_load = load;
2894 sum_nr_running += rq->nr_running;
2895 sum_weighted_load += weighted_cpuload(i);
2899 * First idle cpu or the first cpu(busiest) in this sched group
2900 * is eligible for doing load balancing at this and above
2901 * domains. In the newly idle case, we will allow all the cpu's
2902 * to do the newly idle load balance.
2904 if (idle != CPU_NEWLY_IDLE && local_group &&
2905 balance_cpu != this_cpu && balance) {
2910 total_load += avg_load;
2911 total_pwr += group->__cpu_power;
2913 /* Adjust by relative CPU power of the group */
2914 avg_load = sg_div_cpu_power(group,
2915 avg_load * SCHED_LOAD_SCALE);
2917 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2920 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2923 this_load = avg_load;
2925 this_nr_running = sum_nr_running;
2926 this_load_per_task = sum_weighted_load;
2927 } else if (avg_load > max_load &&
2928 (sum_nr_running > group_capacity || __group_imb)) {
2929 max_load = avg_load;
2931 busiest_nr_running = sum_nr_running;
2932 busiest_load_per_task = sum_weighted_load;
2933 group_imb = __group_imb;
2936 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2938 * Busy processors will not participate in power savings
2941 if (idle == CPU_NOT_IDLE ||
2942 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2946 * If the local group is idle or completely loaded
2947 * no need to do power savings balance at this domain
2949 if (local_group && (this_nr_running >= group_capacity ||
2951 power_savings_balance = 0;
2954 * If a group is already running at full capacity or idle,
2955 * don't include that group in power savings calculations
2957 if (!power_savings_balance || sum_nr_running >= group_capacity
2962 * Calculate the group which has the least non-idle load.
2963 * This is the group from where we need to pick up the load
2966 if ((sum_nr_running < min_nr_running) ||
2967 (sum_nr_running == min_nr_running &&
2968 first_cpu(group->cpumask) <
2969 first_cpu(group_min->cpumask))) {
2971 min_nr_running = sum_nr_running;
2972 min_load_per_task = sum_weighted_load /
2977 * Calculate the group which is almost near its
2978 * capacity but still has some space to pick up some load
2979 * from other group and save more power
2981 if (sum_nr_running <= group_capacity - 1) {
2982 if (sum_nr_running > leader_nr_running ||
2983 (sum_nr_running == leader_nr_running &&
2984 first_cpu(group->cpumask) >
2985 first_cpu(group_leader->cpumask))) {
2986 group_leader = group;
2987 leader_nr_running = sum_nr_running;
2992 group = group->next;
2993 } while (group != sd->groups);
2995 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2998 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3000 if (this_load >= avg_load ||
3001 100*max_load <= sd->imbalance_pct*this_load)
3004 busiest_load_per_task /= busiest_nr_running;
3006 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3009 * We're trying to get all the cpus to the average_load, so we don't
3010 * want to push ourselves above the average load, nor do we wish to
3011 * reduce the max loaded cpu below the average load, as either of these
3012 * actions would just result in more rebalancing later, and ping-pong
3013 * tasks around. Thus we look for the minimum possible imbalance.
3014 * Negative imbalances (*we* are more loaded than anyone else) will
3015 * be counted as no imbalance for these purposes -- we can't fix that
3016 * by pulling tasks to us. Be careful of negative numbers as they'll
3017 * appear as very large values with unsigned longs.
3019 if (max_load <= busiest_load_per_task)
3023 * In the presence of smp nice balancing, certain scenarios can have
3024 * max load less than avg load(as we skip the groups at or below
3025 * its cpu_power, while calculating max_load..)
3027 if (max_load < avg_load) {
3029 goto small_imbalance;
3032 /* Don't want to pull so many tasks that a group would go idle */
3033 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3035 /* How much load to actually move to equalise the imbalance */
3036 *imbalance = min(max_pull * busiest->__cpu_power,
3037 (avg_load - this_load) * this->__cpu_power)
3041 * if *imbalance is less than the average load per runnable task
3042 * there is no gaurantee that any tasks will be moved so we'll have
3043 * a think about bumping its value to force at least one task to be
3046 if (*imbalance < busiest_load_per_task) {
3047 unsigned long tmp, pwr_now, pwr_move;
3051 pwr_move = pwr_now = 0;
3053 if (this_nr_running) {
3054 this_load_per_task /= this_nr_running;
3055 if (busiest_load_per_task > this_load_per_task)
3058 this_load_per_task = SCHED_LOAD_SCALE;
3060 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3061 busiest_load_per_task * imbn) {
3062 *imbalance = busiest_load_per_task;
3067 * OK, we don't have enough imbalance to justify moving tasks,
3068 * however we may be able to increase total CPU power used by
3072 pwr_now += busiest->__cpu_power *
3073 min(busiest_load_per_task, max_load);
3074 pwr_now += this->__cpu_power *
3075 min(this_load_per_task, this_load);
3076 pwr_now /= SCHED_LOAD_SCALE;
3078 /* Amount of load we'd subtract */
3079 tmp = sg_div_cpu_power(busiest,
3080 busiest_load_per_task * SCHED_LOAD_SCALE);
3082 pwr_move += busiest->__cpu_power *
3083 min(busiest_load_per_task, max_load - tmp);
3085 /* Amount of load we'd add */
3086 if (max_load * busiest->__cpu_power <
3087 busiest_load_per_task * SCHED_LOAD_SCALE)
3088 tmp = sg_div_cpu_power(this,
3089 max_load * busiest->__cpu_power);
3091 tmp = sg_div_cpu_power(this,
3092 busiest_load_per_task * SCHED_LOAD_SCALE);
3093 pwr_move += this->__cpu_power *
3094 min(this_load_per_task, this_load + tmp);
3095 pwr_move /= SCHED_LOAD_SCALE;
3097 /* Move if we gain throughput */
3098 if (pwr_move > pwr_now)
3099 *imbalance = busiest_load_per_task;
3105 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3106 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3109 if (this == group_leader && group_leader != group_min) {
3110 *imbalance = min_load_per_task;
3120 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3123 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3124 unsigned long imbalance, const cpumask_t *cpus)
3126 struct rq *busiest = NULL, *rq;
3127 unsigned long max_load = 0;
3130 for_each_cpu_mask(i, group->cpumask) {
3133 if (!cpu_isset(i, *cpus))
3137 wl = weighted_cpuload(i);
3139 if (rq->nr_running == 1 && wl > imbalance)
3142 if (wl > max_load) {
3152 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3153 * so long as it is large enough.
3155 #define MAX_PINNED_INTERVAL 512
3158 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3159 * tasks if there is an imbalance.
3161 static int load_balance(int this_cpu, struct rq *this_rq,
3162 struct sched_domain *sd, enum cpu_idle_type idle,
3163 int *balance, cpumask_t *cpus)
3165 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3166 struct sched_group *group;
3167 unsigned long imbalance;
3169 unsigned long flags;
3174 * When power savings policy is enabled for the parent domain, idle
3175 * sibling can pick up load irrespective of busy siblings. In this case,
3176 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3177 * portraying it as CPU_NOT_IDLE.
3179 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3180 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3183 schedstat_inc(sd, lb_count[idle]);
3186 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3193 schedstat_inc(sd, lb_nobusyg[idle]);
3197 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3199 schedstat_inc(sd, lb_nobusyq[idle]);
3203 BUG_ON(busiest == this_rq);
3205 schedstat_add(sd, lb_imbalance[idle], imbalance);
3208 if (busiest->nr_running > 1) {
3210 * Attempt to move tasks. If find_busiest_group has found
3211 * an imbalance but busiest->nr_running <= 1, the group is
3212 * still unbalanced. ld_moved simply stays zero, so it is
3213 * correctly treated as an imbalance.
3215 local_irq_save(flags);
3216 double_rq_lock(this_rq, busiest);
3217 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3218 imbalance, sd, idle, &all_pinned);
3219 double_rq_unlock(this_rq, busiest);
3220 local_irq_restore(flags);
3223 * some other cpu did the load balance for us.
3225 if (ld_moved && this_cpu != smp_processor_id())
3226 resched_cpu(this_cpu);
3228 /* All tasks on this runqueue were pinned by CPU affinity */
3229 if (unlikely(all_pinned)) {
3230 cpu_clear(cpu_of(busiest), *cpus);
3231 if (!cpus_empty(*cpus))
3238 schedstat_inc(sd, lb_failed[idle]);
3239 sd->nr_balance_failed++;
3241 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3243 spin_lock_irqsave(&busiest->lock, flags);
3245 /* don't kick the migration_thread, if the curr
3246 * task on busiest cpu can't be moved to this_cpu
3248 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3249 spin_unlock_irqrestore(&busiest->lock, flags);
3251 goto out_one_pinned;
3254 if (!busiest->active_balance) {
3255 busiest->active_balance = 1;
3256 busiest->push_cpu = this_cpu;
3259 spin_unlock_irqrestore(&busiest->lock, flags);
3261 wake_up_process(busiest->migration_thread);
3264 * We've kicked active balancing, reset the failure
3267 sd->nr_balance_failed = sd->cache_nice_tries+1;
3270 sd->nr_balance_failed = 0;
3272 if (likely(!active_balance)) {
3273 /* We were unbalanced, so reset the balancing interval */
3274 sd->balance_interval = sd->min_interval;
3277 * If we've begun active balancing, start to back off. This
3278 * case may not be covered by the all_pinned logic if there
3279 * is only 1 task on the busy runqueue (because we don't call
3282 if (sd->balance_interval < sd->max_interval)
3283 sd->balance_interval *= 2;
3286 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3287 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3292 schedstat_inc(sd, lb_balanced[idle]);
3294 sd->nr_balance_failed = 0;
3297 /* tune up the balancing interval */
3298 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3299 (sd->balance_interval < sd->max_interval))
3300 sd->balance_interval *= 2;
3302 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3303 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3309 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3310 * tasks if there is an imbalance.
3312 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3313 * this_rq is locked.
3316 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3319 struct sched_group *group;
3320 struct rq *busiest = NULL;
3321 unsigned long imbalance;
3329 * When power savings policy is enabled for the parent domain, idle
3330 * sibling can pick up load irrespective of busy siblings. In this case,
3331 * let the state of idle sibling percolate up as IDLE, instead of
3332 * portraying it as CPU_NOT_IDLE.
3334 if (sd->flags & SD_SHARE_CPUPOWER &&
3335 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3338 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3340 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3341 &sd_idle, cpus, NULL);
3343 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3347 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3349 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3353 BUG_ON(busiest == this_rq);
3355 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3358 if (busiest->nr_running > 1) {
3359 /* Attempt to move tasks */
3360 double_lock_balance(this_rq, busiest);
3361 /* this_rq->clock is already updated */
3362 update_rq_clock(busiest);
3363 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3364 imbalance, sd, CPU_NEWLY_IDLE,
3366 spin_unlock(&busiest->lock);
3368 if (unlikely(all_pinned)) {
3369 cpu_clear(cpu_of(busiest), *cpus);
3370 if (!cpus_empty(*cpus))
3376 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3377 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3378 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3381 sd->nr_balance_failed = 0;
3386 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3387 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3388 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3390 sd->nr_balance_failed = 0;
3396 * idle_balance is called by schedule() if this_cpu is about to become
3397 * idle. Attempts to pull tasks from other CPUs.
3399 static void idle_balance(int this_cpu, struct rq *this_rq)
3401 struct sched_domain *sd;
3402 int pulled_task = -1;
3403 unsigned long next_balance = jiffies + HZ;
3406 for_each_domain(this_cpu, sd) {
3407 unsigned long interval;
3409 if (!(sd->flags & SD_LOAD_BALANCE))
3412 if (sd->flags & SD_BALANCE_NEWIDLE)
3413 /* If we've pulled tasks over stop searching: */
3414 pulled_task = load_balance_newidle(this_cpu, this_rq,
3417 interval = msecs_to_jiffies(sd->balance_interval);
3418 if (time_after(next_balance, sd->last_balance + interval))
3419 next_balance = sd->last_balance + interval;
3423 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3425 * We are going idle. next_balance may be set based on
3426 * a busy processor. So reset next_balance.
3428 this_rq->next_balance = next_balance;
3433 * active_load_balance is run by migration threads. It pushes running tasks
3434 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3435 * running on each physical CPU where possible, and avoids physical /
3436 * logical imbalances.
3438 * Called with busiest_rq locked.
3440 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3442 int target_cpu = busiest_rq->push_cpu;
3443 struct sched_domain *sd;
3444 struct rq *target_rq;
3446 /* Is there any task to move? */
3447 if (busiest_rq->nr_running <= 1)
3450 target_rq = cpu_rq(target_cpu);
3453 * This condition is "impossible", if it occurs
3454 * we need to fix it. Originally reported by
3455 * Bjorn Helgaas on a 128-cpu setup.
3457 BUG_ON(busiest_rq == target_rq);
3459 /* move a task from busiest_rq to target_rq */
3460 double_lock_balance(busiest_rq, target_rq);
3461 update_rq_clock(busiest_rq);
3462 update_rq_clock(target_rq);
3464 /* Search for an sd spanning us and the target CPU. */
3465 for_each_domain(target_cpu, sd) {
3466 if ((sd->flags & SD_LOAD_BALANCE) &&
3467 cpu_isset(busiest_cpu, sd->span))
3472 schedstat_inc(sd, alb_count);
3474 if (move_one_task(target_rq, target_cpu, busiest_rq,
3476 schedstat_inc(sd, alb_pushed);
3478 schedstat_inc(sd, alb_failed);
3480 spin_unlock(&target_rq->lock);
3485 atomic_t load_balancer;
3487 } nohz ____cacheline_aligned = {
3488 .load_balancer = ATOMIC_INIT(-1),
3489 .cpu_mask = CPU_MASK_NONE,
3493 * This routine will try to nominate the ilb (idle load balancing)
3494 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3495 * load balancing on behalf of all those cpus. If all the cpus in the system
3496 * go into this tickless mode, then there will be no ilb owner (as there is
3497 * no need for one) and all the cpus will sleep till the next wakeup event
3500 * For the ilb owner, tick is not stopped. And this tick will be used
3501 * for idle load balancing. ilb owner will still be part of
3504 * While stopping the tick, this cpu will become the ilb owner if there
3505 * is no other owner. And will be the owner till that cpu becomes busy
3506 * or if all cpus in the system stop their ticks at which point
3507 * there is no need for ilb owner.
3509 * When the ilb owner becomes busy, it nominates another owner, during the
3510 * next busy scheduler_tick()
3512 int select_nohz_load_balancer(int stop_tick)
3514 int cpu = smp_processor_id();
3517 cpu_set(cpu, nohz.cpu_mask);
3518 cpu_rq(cpu)->in_nohz_recently = 1;
3521 * If we are going offline and still the leader, give up!
3523 if (cpu_is_offline(cpu) &&
3524 atomic_read(&nohz.load_balancer) == cpu) {
3525 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3530 /* time for ilb owner also to sleep */
3531 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3532 if (atomic_read(&nohz.load_balancer) == cpu)
3533 atomic_set(&nohz.load_balancer, -1);
3537 if (atomic_read(&nohz.load_balancer) == -1) {
3538 /* make me the ilb owner */
3539 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3541 } else if (atomic_read(&nohz.load_balancer) == cpu)
3544 if (!cpu_isset(cpu, nohz.cpu_mask))
3547 cpu_clear(cpu, nohz.cpu_mask);
3549 if (atomic_read(&nohz.load_balancer) == cpu)
3550 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3557 static DEFINE_SPINLOCK(balancing);
3560 * It checks each scheduling domain to see if it is due to be balanced,
3561 * and initiates a balancing operation if so.
3563 * Balancing parameters are set up in arch_init_sched_domains.
3565 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3568 struct rq *rq = cpu_rq(cpu);
3569 unsigned long interval;
3570 struct sched_domain *sd;
3571 /* Earliest time when we have to do rebalance again */
3572 unsigned long next_balance = jiffies + 60*HZ;
3573 int update_next_balance = 0;
3576 for_each_domain(cpu, sd) {
3577 if (!(sd->flags & SD_LOAD_BALANCE))
3580 interval = sd->balance_interval;
3581 if (idle != CPU_IDLE)
3582 interval *= sd->busy_factor;
3584 /* scale ms to jiffies */
3585 interval = msecs_to_jiffies(interval);
3586 if (unlikely(!interval))
3588 if (interval > HZ*NR_CPUS/10)
3589 interval = HZ*NR_CPUS/10;
3592 if (sd->flags & SD_SERIALIZE) {
3593 if (!spin_trylock(&balancing))
3597 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3598 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3600 * We've pulled tasks over so either we're no
3601 * longer idle, or one of our SMT siblings is
3604 idle = CPU_NOT_IDLE;
3606 sd->last_balance = jiffies;
3608 if (sd->flags & SD_SERIALIZE)
3609 spin_unlock(&balancing);
3611 if (time_after(next_balance, sd->last_balance + interval)) {
3612 next_balance = sd->last_balance + interval;
3613 update_next_balance = 1;
3617 * Stop the load balance at this level. There is another
3618 * CPU in our sched group which is doing load balancing more
3626 * next_balance will be updated only when there is a need.
3627 * When the cpu is attached to null domain for ex, it will not be
3630 if (likely(update_next_balance))
3631 rq->next_balance = next_balance;
3635 * run_rebalance_domains is triggered when needed from the scheduler tick.
3636 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3637 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3639 static void run_rebalance_domains(struct softirq_action *h)
3641 int this_cpu = smp_processor_id();
3642 struct rq *this_rq = cpu_rq(this_cpu);
3643 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3644 CPU_IDLE : CPU_NOT_IDLE;
3646 rebalance_domains(this_cpu, idle);
3650 * If this cpu is the owner for idle load balancing, then do the
3651 * balancing on behalf of the other idle cpus whose ticks are
3654 if (this_rq->idle_at_tick &&
3655 atomic_read(&nohz.load_balancer) == this_cpu) {
3656 cpumask_t cpus = nohz.cpu_mask;
3660 cpu_clear(this_cpu, cpus);
3661 for_each_cpu_mask(balance_cpu, cpus) {
3663 * If this cpu gets work to do, stop the load balancing
3664 * work being done for other cpus. Next load
3665 * balancing owner will pick it up.
3670 rebalance_domains(balance_cpu, CPU_IDLE);
3672 rq = cpu_rq(balance_cpu);
3673 if (time_after(this_rq->next_balance, rq->next_balance))
3674 this_rq->next_balance = rq->next_balance;
3681 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3683 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3684 * idle load balancing owner or decide to stop the periodic load balancing,
3685 * if the whole system is idle.
3687 static inline void trigger_load_balance(struct rq *rq, int cpu)
3691 * If we were in the nohz mode recently and busy at the current
3692 * scheduler tick, then check if we need to nominate new idle
3695 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3696 rq->in_nohz_recently = 0;
3698 if (atomic_read(&nohz.load_balancer) == cpu) {
3699 cpu_clear(cpu, nohz.cpu_mask);
3700 atomic_set(&nohz.load_balancer, -1);
3703 if (atomic_read(&nohz.load_balancer) == -1) {
3705 * simple selection for now: Nominate the
3706 * first cpu in the nohz list to be the next
3709 * TBD: Traverse the sched domains and nominate
3710 * the nearest cpu in the nohz.cpu_mask.
3712 int ilb = first_cpu(nohz.cpu_mask);
3714 if (ilb < nr_cpu_ids)
3720 * If this cpu is idle and doing idle load balancing for all the
3721 * cpus with ticks stopped, is it time for that to stop?
3723 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3724 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3730 * If this cpu is idle and the idle load balancing is done by
3731 * someone else, then no need raise the SCHED_SOFTIRQ
3733 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3734 cpu_isset(cpu, nohz.cpu_mask))
3737 if (time_after_eq(jiffies, rq->next_balance))
3738 raise_softirq(SCHED_SOFTIRQ);
3741 #else /* CONFIG_SMP */
3744 * on UP we do not need to balance between CPUs:
3746 static inline void idle_balance(int cpu, struct rq *rq)
3752 DEFINE_PER_CPU(struct kernel_stat, kstat);
3754 EXPORT_PER_CPU_SYMBOL(kstat);
3757 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3758 * that have not yet been banked in case the task is currently running.
3760 unsigned long long task_sched_runtime(struct task_struct *p)
3762 unsigned long flags;
3766 rq = task_rq_lock(p, &flags);
3767 ns = p->se.sum_exec_runtime;
3768 if (task_current(rq, p)) {
3769 update_rq_clock(rq);
3770 delta_exec = rq->clock - p->se.exec_start;
3771 if ((s64)delta_exec > 0)
3774 task_rq_unlock(rq, &flags);
3780 * Account user cpu time to a process.
3781 * @p: the process that the cpu time gets accounted to
3782 * @cputime: the cpu time spent in user space since the last update
3784 void account_user_time(struct task_struct *p, cputime_t cputime)
3786 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3789 p->utime = cputime_add(p->utime, cputime);
3791 /* Add user time to cpustat. */
3792 tmp = cputime_to_cputime64(cputime);
3793 if (TASK_NICE(p) > 0)
3794 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3796 cpustat->user = cputime64_add(cpustat->user, tmp);
3800 * Account guest cpu time to a process.
3801 * @p: the process that the cpu time gets accounted to
3802 * @cputime: the cpu time spent in virtual machine since the last update
3804 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3807 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3809 tmp = cputime_to_cputime64(cputime);
3811 p->utime = cputime_add(p->utime, cputime);
3812 p->gtime = cputime_add(p->gtime, cputime);
3814 cpustat->user = cputime64_add(cpustat->user, tmp);
3815 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3819 * Account scaled user cpu time to a process.
3820 * @p: the process that the cpu time gets accounted to
3821 * @cputime: the cpu time spent in user space since the last update
3823 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3825 p->utimescaled = cputime_add(p->utimescaled, cputime);
3829 * Account system cpu time to a process.
3830 * @p: the process that the cpu time gets accounted to
3831 * @hardirq_offset: the offset to subtract from hardirq_count()
3832 * @cputime: the cpu time spent in kernel space since the last update
3834 void account_system_time(struct task_struct *p, int hardirq_offset,
3837 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3838 struct rq *rq = this_rq();
3841 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3842 return account_guest_time(p, cputime);
3844 p->stime = cputime_add(p->stime, cputime);
3846 /* Add system time to cpustat. */
3847 tmp = cputime_to_cputime64(cputime);
3848 if (hardirq_count() - hardirq_offset)
3849 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3850 else if (softirq_count())
3851 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3852 else if (p != rq->idle)
3853 cpustat->system = cputime64_add(cpustat->system, tmp);
3854 else if (atomic_read(&rq->nr_iowait) > 0)
3855 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3857 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3858 /* Account for system time used */
3859 acct_update_integrals(p);
3863 * Account scaled system cpu time to a process.
3864 * @p: the process that the cpu time gets accounted to
3865 * @hardirq_offset: the offset to subtract from hardirq_count()
3866 * @cputime: the cpu time spent in kernel space since the last update
3868 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3870 p->stimescaled = cputime_add(p->stimescaled, cputime);
3874 * Account for involuntary wait time.
3875 * @p: the process from which the cpu time has been stolen
3876 * @steal: the cpu time spent in involuntary wait
3878 void account_steal_time(struct task_struct *p, cputime_t steal)
3880 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3881 cputime64_t tmp = cputime_to_cputime64(steal);
3882 struct rq *rq = this_rq();
3884 if (p == rq->idle) {
3885 p->stime = cputime_add(p->stime, steal);
3886 if (atomic_read(&rq->nr_iowait) > 0)
3887 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3889 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3891 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3895 * This function gets called by the timer code, with HZ frequency.
3896 * We call it with interrupts disabled.
3898 * It also gets called by the fork code, when changing the parent's
3901 void scheduler_tick(void)
3903 int cpu = smp_processor_id();
3904 struct rq *rq = cpu_rq(cpu);
3905 struct task_struct *curr = rq->curr;
3906 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3908 spin_lock(&rq->lock);
3909 __update_rq_clock(rq);
3911 * Let rq->clock advance by at least TICK_NSEC:
3913 if (unlikely(rq->clock < next_tick)) {
3914 rq->clock = next_tick;
3915 rq->clock_underflows++;
3917 rq->tick_timestamp = rq->clock;
3918 update_last_tick_seen(rq);
3919 update_cpu_load(rq);
3920 curr->sched_class->task_tick(rq, curr, 0);
3921 spin_unlock(&rq->lock);
3924 rq->idle_at_tick = idle_cpu(cpu);
3925 trigger_load_balance(rq, cpu);
3929 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3931 void __kprobes add_preempt_count(int val)
3936 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3938 preempt_count() += val;
3940 * Spinlock count overflowing soon?
3942 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3945 EXPORT_SYMBOL(add_preempt_count);
3947 void __kprobes sub_preempt_count(int val)
3952 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3955 * Is the spinlock portion underflowing?
3957 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3958 !(preempt_count() & PREEMPT_MASK)))
3961 preempt_count() -= val;
3963 EXPORT_SYMBOL(sub_preempt_count);
3968 * Print scheduling while atomic bug:
3970 static noinline void __schedule_bug(struct task_struct *prev)
3972 struct pt_regs *regs = get_irq_regs();
3974 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3975 prev->comm, prev->pid, preempt_count());
3977 debug_show_held_locks(prev);
3978 if (irqs_disabled())
3979 print_irqtrace_events(prev);
3988 * Various schedule()-time debugging checks and statistics:
3990 static inline void schedule_debug(struct task_struct *prev)
3993 * Test if we are atomic. Since do_exit() needs to call into
3994 * schedule() atomically, we ignore that path for now.
3995 * Otherwise, whine if we are scheduling when we should not be.
3997 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3998 __schedule_bug(prev);
4000 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4002 schedstat_inc(this_rq(), sched_count);
4003 #ifdef CONFIG_SCHEDSTATS
4004 if (unlikely(prev->lock_depth >= 0)) {
4005 schedstat_inc(this_rq(), bkl_count);
4006 schedstat_inc(prev, sched_info.bkl_count);
4012 * Pick up the highest-prio task:
4014 static inline struct task_struct *
4015 pick_next_task(struct rq *rq, struct task_struct *prev)
4017 const struct sched_class *class;
4018 struct task_struct *p;
4021 * Optimization: we know that if all tasks are in
4022 * the fair class we can call that function directly:
4024 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4025 p = fair_sched_class.pick_next_task(rq);
4030 class = sched_class_highest;
4032 p = class->pick_next_task(rq);
4036 * Will never be NULL as the idle class always
4037 * returns a non-NULL p:
4039 class = class->next;
4044 * schedule() is the main scheduler function.
4046 asmlinkage void __sched schedule(void)
4048 struct task_struct *prev, *next;
4049 unsigned long *switch_count;
4055 cpu = smp_processor_id();
4059 switch_count = &prev->nivcsw;
4061 release_kernel_lock(prev);
4062 need_resched_nonpreemptible:
4064 schedule_debug(prev);
4069 * Do the rq-clock update outside the rq lock:
4071 local_irq_disable();
4072 __update_rq_clock(rq);
4073 spin_lock(&rq->lock);
4074 clear_tsk_need_resched(prev);
4076 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4077 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4078 signal_pending(prev))) {
4079 prev->state = TASK_RUNNING;
4081 deactivate_task(rq, prev, 1);
4083 switch_count = &prev->nvcsw;
4087 if (prev->sched_class->pre_schedule)
4088 prev->sched_class->pre_schedule(rq, prev);
4091 if (unlikely(!rq->nr_running))
4092 idle_balance(cpu, rq);
4094 prev->sched_class->put_prev_task(rq, prev);
4095 next = pick_next_task(rq, prev);
4097 sched_info_switch(prev, next);
4099 if (likely(prev != next)) {
4104 context_switch(rq, prev, next); /* unlocks the rq */
4106 * the context switch might have flipped the stack from under
4107 * us, hence refresh the local variables.
4109 cpu = smp_processor_id();
4112 spin_unlock_irq(&rq->lock);
4116 if (unlikely(reacquire_kernel_lock(current) < 0))
4117 goto need_resched_nonpreemptible;
4119 preempt_enable_no_resched();
4120 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4123 EXPORT_SYMBOL(schedule);
4125 #ifdef CONFIG_PREEMPT
4127 * this is the entry point to schedule() from in-kernel preemption
4128 * off of preempt_enable. Kernel preemptions off return from interrupt
4129 * occur there and call schedule directly.
4131 asmlinkage void __sched preempt_schedule(void)
4133 struct thread_info *ti = current_thread_info();
4134 struct task_struct *task = current;
4135 int saved_lock_depth;
4138 * If there is a non-zero preempt_count or interrupts are disabled,
4139 * we do not want to preempt the current task. Just return..
4141 if (likely(ti->preempt_count || irqs_disabled()))
4145 add_preempt_count(PREEMPT_ACTIVE);
4148 * We keep the big kernel semaphore locked, but we
4149 * clear ->lock_depth so that schedule() doesnt
4150 * auto-release the semaphore:
4152 saved_lock_depth = task->lock_depth;
4153 task->lock_depth = -1;
4155 task->lock_depth = saved_lock_depth;
4156 sub_preempt_count(PREEMPT_ACTIVE);
4159 * Check again in case we missed a preemption opportunity
4160 * between schedule and now.
4163 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4165 EXPORT_SYMBOL(preempt_schedule);
4168 * this is the entry point to schedule() from kernel preemption
4169 * off of irq context.
4170 * Note, that this is called and return with irqs disabled. This will
4171 * protect us against recursive calling from irq.
4173 asmlinkage void __sched preempt_schedule_irq(void)
4175 struct thread_info *ti = current_thread_info();
4176 struct task_struct *task = current;
4177 int saved_lock_depth;
4179 /* Catch callers which need to be fixed */
4180 BUG_ON(ti->preempt_count || !irqs_disabled());
4183 add_preempt_count(PREEMPT_ACTIVE);
4186 * We keep the big kernel semaphore locked, but we
4187 * clear ->lock_depth so that schedule() doesnt
4188 * auto-release the semaphore:
4190 saved_lock_depth = task->lock_depth;
4191 task->lock_depth = -1;
4194 local_irq_disable();
4195 task->lock_depth = saved_lock_depth;
4196 sub_preempt_count(PREEMPT_ACTIVE);
4199 * Check again in case we missed a preemption opportunity
4200 * between schedule and now.
4203 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4206 #endif /* CONFIG_PREEMPT */
4208 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4211 return try_to_wake_up(curr->private, mode, sync);
4213 EXPORT_SYMBOL(default_wake_function);
4216 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4217 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4218 * number) then we wake all the non-exclusive tasks and one exclusive task.
4220 * There are circumstances in which we can try to wake a task which has already
4221 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4222 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4224 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4225 int nr_exclusive, int sync, void *key)
4227 wait_queue_t *curr, *next;
4229 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4230 unsigned flags = curr->flags;
4232 if (curr->func(curr, mode, sync, key) &&
4233 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4239 * __wake_up - wake up threads blocked on a waitqueue.
4241 * @mode: which threads
4242 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4243 * @key: is directly passed to the wakeup function
4245 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4246 int nr_exclusive, void *key)
4248 unsigned long flags;
4250 spin_lock_irqsave(&q->lock, flags);
4251 __wake_up_common(q, mode, nr_exclusive, 0, key);
4252 spin_unlock_irqrestore(&q->lock, flags);
4254 EXPORT_SYMBOL(__wake_up);
4257 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4259 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4261 __wake_up_common(q, mode, 1, 0, NULL);
4265 * __wake_up_sync - wake up threads blocked on a waitqueue.
4267 * @mode: which threads
4268 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4270 * The sync wakeup differs that the waker knows that it will schedule
4271 * away soon, so while the target thread will be woken up, it will not
4272 * be migrated to another CPU - ie. the two threads are 'synchronized'
4273 * with each other. This can prevent needless bouncing between CPUs.
4275 * On UP it can prevent extra preemption.
4278 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4280 unsigned long flags;
4286 if (unlikely(!nr_exclusive))
4289 spin_lock_irqsave(&q->lock, flags);
4290 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4291 spin_unlock_irqrestore(&q->lock, flags);
4293 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4295 void complete(struct completion *x)
4297 unsigned long flags;
4299 spin_lock_irqsave(&x->wait.lock, flags);
4301 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4302 spin_unlock_irqrestore(&x->wait.lock, flags);
4304 EXPORT_SYMBOL(complete);
4306 void complete_all(struct completion *x)
4308 unsigned long flags;
4310 spin_lock_irqsave(&x->wait.lock, flags);
4311 x->done += UINT_MAX/2;
4312 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4313 spin_unlock_irqrestore(&x->wait.lock, flags);
4315 EXPORT_SYMBOL(complete_all);
4317 static inline long __sched
4318 do_wait_for_common(struct completion *x, long timeout, int state)
4321 DECLARE_WAITQUEUE(wait, current);
4323 wait.flags |= WQ_FLAG_EXCLUSIVE;
4324 __add_wait_queue_tail(&x->wait, &wait);
4326 if ((state == TASK_INTERRUPTIBLE &&
4327 signal_pending(current)) ||
4328 (state == TASK_KILLABLE &&
4329 fatal_signal_pending(current))) {
4330 __remove_wait_queue(&x->wait, &wait);
4331 return -ERESTARTSYS;
4333 __set_current_state(state);
4334 spin_unlock_irq(&x->wait.lock);
4335 timeout = schedule_timeout(timeout);
4336 spin_lock_irq(&x->wait.lock);
4338 __remove_wait_queue(&x->wait, &wait);
4342 __remove_wait_queue(&x->wait, &wait);
4349 wait_for_common(struct completion *x, long timeout, int state)
4353 spin_lock_irq(&x->wait.lock);
4354 timeout = do_wait_for_common(x, timeout, state);
4355 spin_unlock_irq(&x->wait.lock);
4359 void __sched wait_for_completion(struct completion *x)
4361 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4363 EXPORT_SYMBOL(wait_for_completion);
4365 unsigned long __sched
4366 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4368 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4370 EXPORT_SYMBOL(wait_for_completion_timeout);
4372 int __sched wait_for_completion_interruptible(struct completion *x)
4374 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4375 if (t == -ERESTARTSYS)
4379 EXPORT_SYMBOL(wait_for_completion_interruptible);
4381 unsigned long __sched
4382 wait_for_completion_interruptible_timeout(struct completion *x,
4383 unsigned long timeout)
4385 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4387 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4389 int __sched wait_for_completion_killable(struct completion *x)
4391 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4392 if (t == -ERESTARTSYS)
4396 EXPORT_SYMBOL(wait_for_completion_killable);
4399 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4401 unsigned long flags;
4404 init_waitqueue_entry(&wait, current);
4406 __set_current_state(state);
4408 spin_lock_irqsave(&q->lock, flags);
4409 __add_wait_queue(q, &wait);
4410 spin_unlock(&q->lock);
4411 timeout = schedule_timeout(timeout);
4412 spin_lock_irq(&q->lock);
4413 __remove_wait_queue(q, &wait);
4414 spin_unlock_irqrestore(&q->lock, flags);
4419 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4421 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4423 EXPORT_SYMBOL(interruptible_sleep_on);
4426 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4428 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4430 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4432 void __sched sleep_on(wait_queue_head_t *q)
4434 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4436 EXPORT_SYMBOL(sleep_on);
4438 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4440 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4442 EXPORT_SYMBOL(sleep_on_timeout);
4444 #ifdef CONFIG_RT_MUTEXES
4447 * rt_mutex_setprio - set the current priority of a task
4449 * @prio: prio value (kernel-internal form)
4451 * This function changes the 'effective' priority of a task. It does
4452 * not touch ->normal_prio like __setscheduler().
4454 * Used by the rt_mutex code to implement priority inheritance logic.
4456 void rt_mutex_setprio(struct task_struct *p, int prio)
4458 unsigned long flags;
4459 int oldprio, on_rq, running;
4461 const struct sched_class *prev_class = p->sched_class;
4463 BUG_ON(prio < 0 || prio > MAX_PRIO);
4465 rq = task_rq_lock(p, &flags);
4466 update_rq_clock(rq);
4469 on_rq = p->se.on_rq;
4470 running = task_current(rq, p);
4472 dequeue_task(rq, p, 0);
4474 p->sched_class->put_prev_task(rq, p);
4477 p->sched_class = &rt_sched_class;
4479 p->sched_class = &fair_sched_class;
4484 p->sched_class->set_curr_task(rq);
4486 enqueue_task(rq, p, 0);
4488 check_class_changed(rq, p, prev_class, oldprio, running);
4490 task_rq_unlock(rq, &flags);
4495 void set_user_nice(struct task_struct *p, long nice)
4497 int old_prio, delta, on_rq;
4498 unsigned long flags;
4501 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4504 * We have to be careful, if called from sys_setpriority(),
4505 * the task might be in the middle of scheduling on another CPU.
4507 rq = task_rq_lock(p, &flags);
4508 update_rq_clock(rq);
4510 * The RT priorities are set via sched_setscheduler(), but we still
4511 * allow the 'normal' nice value to be set - but as expected
4512 * it wont have any effect on scheduling until the task is
4513 * SCHED_FIFO/SCHED_RR:
4515 if (task_has_rt_policy(p)) {
4516 p->static_prio = NICE_TO_PRIO(nice);
4519 on_rq = p->se.on_rq;
4521 dequeue_task(rq, p, 0);
4525 p->static_prio = NICE_TO_PRIO(nice);
4528 p->prio = effective_prio(p);
4529 delta = p->prio - old_prio;
4532 enqueue_task(rq, p, 0);
4535 * If the task increased its priority or is running and
4536 * lowered its priority, then reschedule its CPU:
4538 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4539 resched_task(rq->curr);
4542 task_rq_unlock(rq, &flags);
4544 EXPORT_SYMBOL(set_user_nice);
4547 * can_nice - check if a task can reduce its nice value
4551 int can_nice(const struct task_struct *p, const int nice)
4553 /* convert nice value [19,-20] to rlimit style value [1,40] */
4554 int nice_rlim = 20 - nice;
4556 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4557 capable(CAP_SYS_NICE));
4560 #ifdef __ARCH_WANT_SYS_NICE
4563 * sys_nice - change the priority of the current process.
4564 * @increment: priority increment
4566 * sys_setpriority is a more generic, but much slower function that
4567 * does similar things.
4569 asmlinkage long sys_nice(int increment)
4574 * Setpriority might change our priority at the same moment.
4575 * We don't have to worry. Conceptually one call occurs first
4576 * and we have a single winner.
4578 if (increment < -40)
4583 nice = PRIO_TO_NICE(current->static_prio) + increment;
4589 if (increment < 0 && !can_nice(current, nice))
4592 retval = security_task_setnice(current, nice);
4596 set_user_nice(current, nice);
4603 * task_prio - return the priority value of a given task.
4604 * @p: the task in question.
4606 * This is the priority value as seen by users in /proc.
4607 * RT tasks are offset by -200. Normal tasks are centered
4608 * around 0, value goes from -16 to +15.
4610 int task_prio(const struct task_struct *p)
4612 return p->prio - MAX_RT_PRIO;
4616 * task_nice - return the nice value of a given task.
4617 * @p: the task in question.
4619 int task_nice(const struct task_struct *p)
4621 return TASK_NICE(p);
4623 EXPORT_SYMBOL(task_nice);
4626 * idle_cpu - is a given cpu idle currently?
4627 * @cpu: the processor in question.
4629 int idle_cpu(int cpu)
4631 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4635 * idle_task - return the idle task for a given cpu.
4636 * @cpu: the processor in question.
4638 struct task_struct *idle_task(int cpu)
4640 return cpu_rq(cpu)->idle;
4644 * find_process_by_pid - find a process with a matching PID value.
4645 * @pid: the pid in question.
4647 static struct task_struct *find_process_by_pid(pid_t pid)
4649 return pid ? find_task_by_vpid(pid) : current;
4652 /* Actually do priority change: must hold rq lock. */
4654 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4656 BUG_ON(p->se.on_rq);
4659 switch (p->policy) {
4663 p->sched_class = &fair_sched_class;
4667 p->sched_class = &rt_sched_class;
4671 p->rt_priority = prio;
4672 p->normal_prio = normal_prio(p);
4673 /* we are holding p->pi_lock already */
4674 p->prio = rt_mutex_getprio(p);
4679 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4680 * @p: the task in question.
4681 * @policy: new policy.
4682 * @param: structure containing the new RT priority.
4684 * NOTE that the task may be already dead.
4686 int sched_setscheduler(struct task_struct *p, int policy,
4687 struct sched_param *param)
4689 int retval, oldprio, oldpolicy = -1, on_rq, running;
4690 unsigned long flags;
4691 const struct sched_class *prev_class = p->sched_class;
4694 /* may grab non-irq protected spin_locks */
4695 BUG_ON(in_interrupt());
4697 /* double check policy once rq lock held */
4699 policy = oldpolicy = p->policy;
4700 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4701 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4702 policy != SCHED_IDLE)
4705 * Valid priorities for SCHED_FIFO and SCHED_RR are
4706 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4707 * SCHED_BATCH and SCHED_IDLE is 0.
4709 if (param->sched_priority < 0 ||
4710 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4711 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4713 if (rt_policy(policy) != (param->sched_priority != 0))
4717 * Allow unprivileged RT tasks to decrease priority:
4719 if (!capable(CAP_SYS_NICE)) {
4720 if (rt_policy(policy)) {
4721 unsigned long rlim_rtprio;
4723 if (!lock_task_sighand(p, &flags))
4725 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4726 unlock_task_sighand(p, &flags);
4728 /* can't set/change the rt policy */
4729 if (policy != p->policy && !rlim_rtprio)
4732 /* can't increase priority */
4733 if (param->sched_priority > p->rt_priority &&
4734 param->sched_priority > rlim_rtprio)
4738 * Like positive nice levels, dont allow tasks to
4739 * move out of SCHED_IDLE either:
4741 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4744 /* can't change other user's priorities */
4745 if ((current->euid != p->euid) &&
4746 (current->euid != p->uid))
4750 #ifdef CONFIG_RT_GROUP_SCHED
4752 * Do not allow realtime tasks into groups that have no runtime
4755 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4759 retval = security_task_setscheduler(p, policy, param);
4763 * make sure no PI-waiters arrive (or leave) while we are
4764 * changing the priority of the task:
4766 spin_lock_irqsave(&p->pi_lock, flags);
4768 * To be able to change p->policy safely, the apropriate
4769 * runqueue lock must be held.
4771 rq = __task_rq_lock(p);
4772 /* recheck policy now with rq lock held */
4773 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4774 policy = oldpolicy = -1;
4775 __task_rq_unlock(rq);
4776 spin_unlock_irqrestore(&p->pi_lock, flags);
4779 update_rq_clock(rq);
4780 on_rq = p->se.on_rq;
4781 running = task_current(rq, p);
4783 deactivate_task(rq, p, 0);
4785 p->sched_class->put_prev_task(rq, p);
4788 __setscheduler(rq, p, policy, param->sched_priority);
4791 p->sched_class->set_curr_task(rq);
4793 activate_task(rq, p, 0);
4795 check_class_changed(rq, p, prev_class, oldprio, running);
4797 __task_rq_unlock(rq);
4798 spin_unlock_irqrestore(&p->pi_lock, flags);
4800 rt_mutex_adjust_pi(p);
4804 EXPORT_SYMBOL_GPL(sched_setscheduler);
4807 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4809 struct sched_param lparam;
4810 struct task_struct *p;
4813 if (!param || pid < 0)
4815 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4820 p = find_process_by_pid(pid);
4822 retval = sched_setscheduler(p, policy, &lparam);
4829 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4830 * @pid: the pid in question.
4831 * @policy: new policy.
4832 * @param: structure containing the new RT priority.
4835 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4837 /* negative values for policy are not valid */
4841 return do_sched_setscheduler(pid, policy, param);
4845 * sys_sched_setparam - set/change the RT priority of a thread
4846 * @pid: the pid in question.
4847 * @param: structure containing the new RT priority.
4849 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4851 return do_sched_setscheduler(pid, -1, param);
4855 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4856 * @pid: the pid in question.
4858 asmlinkage long sys_sched_getscheduler(pid_t pid)
4860 struct task_struct *p;
4867 read_lock(&tasklist_lock);
4868 p = find_process_by_pid(pid);
4870 retval = security_task_getscheduler(p);
4874 read_unlock(&tasklist_lock);
4879 * sys_sched_getscheduler - get the RT priority of a thread
4880 * @pid: the pid in question.
4881 * @param: structure containing the RT priority.
4883 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4885 struct sched_param lp;
4886 struct task_struct *p;
4889 if (!param || pid < 0)
4892 read_lock(&tasklist_lock);
4893 p = find_process_by_pid(pid);
4898 retval = security_task_getscheduler(p);
4902 lp.sched_priority = p->rt_priority;
4903 read_unlock(&tasklist_lock);
4906 * This one might sleep, we cannot do it with a spinlock held ...
4908 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4913 read_unlock(&tasklist_lock);
4917 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4919 cpumask_t cpus_allowed;
4920 cpumask_t new_mask = *in_mask;
4921 struct task_struct *p;
4925 read_lock(&tasklist_lock);
4927 p = find_process_by_pid(pid);
4929 read_unlock(&tasklist_lock);
4935 * It is not safe to call set_cpus_allowed with the
4936 * tasklist_lock held. We will bump the task_struct's
4937 * usage count and then drop tasklist_lock.
4940 read_unlock(&tasklist_lock);
4943 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4944 !capable(CAP_SYS_NICE))
4947 retval = security_task_setscheduler(p, 0, NULL);
4951 cpuset_cpus_allowed(p, &cpus_allowed);
4952 cpus_and(new_mask, new_mask, cpus_allowed);
4954 retval = set_cpus_allowed_ptr(p, &new_mask);
4957 cpuset_cpus_allowed(p, &cpus_allowed);
4958 if (!cpus_subset(new_mask, cpus_allowed)) {
4960 * We must have raced with a concurrent cpuset
4961 * update. Just reset the cpus_allowed to the
4962 * cpuset's cpus_allowed
4964 new_mask = cpus_allowed;
4974 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4975 cpumask_t *new_mask)
4977 if (len < sizeof(cpumask_t)) {
4978 memset(new_mask, 0, sizeof(cpumask_t));
4979 } else if (len > sizeof(cpumask_t)) {
4980 len = sizeof(cpumask_t);
4982 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4986 * sys_sched_setaffinity - set the cpu affinity of a process
4987 * @pid: pid of the process
4988 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4989 * @user_mask_ptr: user-space pointer to the new cpu mask
4991 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4992 unsigned long __user *user_mask_ptr)
4997 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5001 return sched_setaffinity(pid, &new_mask);
5005 * Represents all cpu's present in the system
5006 * In systems capable of hotplug, this map could dynamically grow
5007 * as new cpu's are detected in the system via any platform specific
5008 * method, such as ACPI for e.g.
5011 cpumask_t cpu_present_map __read_mostly;
5012 EXPORT_SYMBOL(cpu_present_map);
5015 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5016 EXPORT_SYMBOL(cpu_online_map);
5018 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5019 EXPORT_SYMBOL(cpu_possible_map);
5022 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5024 struct task_struct *p;
5028 read_lock(&tasklist_lock);
5031 p = find_process_by_pid(pid);
5035 retval = security_task_getscheduler(p);
5039 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5042 read_unlock(&tasklist_lock);
5049 * sys_sched_getaffinity - get the cpu affinity of a process
5050 * @pid: pid of the process
5051 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5052 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5054 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5055 unsigned long __user *user_mask_ptr)
5060 if (len < sizeof(cpumask_t))
5063 ret = sched_getaffinity(pid, &mask);
5067 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5070 return sizeof(cpumask_t);
5074 * sys_sched_yield - yield the current processor to other threads.
5076 * This function yields the current CPU to other tasks. If there are no
5077 * other threads running on this CPU then this function will return.
5079 asmlinkage long sys_sched_yield(void)
5081 struct rq *rq = this_rq_lock();
5083 schedstat_inc(rq, yld_count);
5084 current->sched_class->yield_task(rq);
5087 * Since we are going to call schedule() anyway, there's
5088 * no need to preempt or enable interrupts:
5090 __release(rq->lock);
5091 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5092 _raw_spin_unlock(&rq->lock);
5093 preempt_enable_no_resched();
5100 static void __cond_resched(void)
5102 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5103 __might_sleep(__FILE__, __LINE__);
5106 * The BKS might be reacquired before we have dropped
5107 * PREEMPT_ACTIVE, which could trigger a second
5108 * cond_resched() call.
5111 add_preempt_count(PREEMPT_ACTIVE);
5113 sub_preempt_count(PREEMPT_ACTIVE);
5114 } while (need_resched());
5117 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5118 int __sched _cond_resched(void)
5120 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5121 system_state == SYSTEM_RUNNING) {
5127 EXPORT_SYMBOL(_cond_resched);
5131 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5132 * call schedule, and on return reacquire the lock.
5134 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5135 * operations here to prevent schedule() from being called twice (once via
5136 * spin_unlock(), once by hand).
5138 int cond_resched_lock(spinlock_t *lock)
5140 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5143 if (spin_needbreak(lock) || resched) {
5145 if (resched && need_resched())
5154 EXPORT_SYMBOL(cond_resched_lock);
5156 int __sched cond_resched_softirq(void)
5158 BUG_ON(!in_softirq());
5160 if (need_resched() && system_state == SYSTEM_RUNNING) {
5168 EXPORT_SYMBOL(cond_resched_softirq);
5171 * yield - yield the current processor to other threads.
5173 * This is a shortcut for kernel-space yielding - it marks the
5174 * thread runnable and calls sys_sched_yield().
5176 void __sched yield(void)
5178 set_current_state(TASK_RUNNING);
5181 EXPORT_SYMBOL(yield);
5184 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5185 * that process accounting knows that this is a task in IO wait state.
5187 * But don't do that if it is a deliberate, throttling IO wait (this task
5188 * has set its backing_dev_info: the queue against which it should throttle)
5190 void __sched io_schedule(void)
5192 struct rq *rq = &__raw_get_cpu_var(runqueues);
5194 delayacct_blkio_start();
5195 atomic_inc(&rq->nr_iowait);
5197 atomic_dec(&rq->nr_iowait);
5198 delayacct_blkio_end();
5200 EXPORT_SYMBOL(io_schedule);
5202 long __sched io_schedule_timeout(long timeout)
5204 struct rq *rq = &__raw_get_cpu_var(runqueues);
5207 delayacct_blkio_start();
5208 atomic_inc(&rq->nr_iowait);
5209 ret = schedule_timeout(timeout);
5210 atomic_dec(&rq->nr_iowait);
5211 delayacct_blkio_end();
5216 * sys_sched_get_priority_max - return maximum RT priority.
5217 * @policy: scheduling class.
5219 * this syscall returns the maximum rt_priority that can be used
5220 * by a given scheduling class.
5222 asmlinkage long sys_sched_get_priority_max(int policy)
5229 ret = MAX_USER_RT_PRIO-1;
5241 * sys_sched_get_priority_min - return minimum RT priority.
5242 * @policy: scheduling class.
5244 * this syscall returns the minimum rt_priority that can be used
5245 * by a given scheduling class.
5247 asmlinkage long sys_sched_get_priority_min(int policy)
5265 * sys_sched_rr_get_interval - return the default timeslice of a process.
5266 * @pid: pid of the process.
5267 * @interval: userspace pointer to the timeslice value.
5269 * this syscall writes the default timeslice value of a given process
5270 * into the user-space timespec buffer. A value of '0' means infinity.
5273 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5275 struct task_struct *p;
5276 unsigned int time_slice;
5284 read_lock(&tasklist_lock);
5285 p = find_process_by_pid(pid);
5289 retval = security_task_getscheduler(p);
5294 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5295 * tasks that are on an otherwise idle runqueue:
5298 if (p->policy == SCHED_RR) {
5299 time_slice = DEF_TIMESLICE;
5300 } else if (p->policy != SCHED_FIFO) {
5301 struct sched_entity *se = &p->se;
5302 unsigned long flags;
5305 rq = task_rq_lock(p, &flags);
5306 if (rq->cfs.load.weight)
5307 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5308 task_rq_unlock(rq, &flags);
5310 read_unlock(&tasklist_lock);
5311 jiffies_to_timespec(time_slice, &t);
5312 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5316 read_unlock(&tasklist_lock);
5320 static const char stat_nam[] = "RSDTtZX";
5322 void sched_show_task(struct task_struct *p)
5324 unsigned long free = 0;
5327 state = p->state ? __ffs(p->state) + 1 : 0;
5328 printk(KERN_INFO "%-13.13s %c", p->comm,
5329 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5330 #if BITS_PER_LONG == 32
5331 if (state == TASK_RUNNING)
5332 printk(KERN_CONT " running ");
5334 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5336 if (state == TASK_RUNNING)
5337 printk(KERN_CONT " running task ");
5339 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5341 #ifdef CONFIG_DEBUG_STACK_USAGE
5343 unsigned long *n = end_of_stack(p);
5346 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5349 printk(KERN_CONT "%5lu %5d %6d\n", free,
5350 task_pid_nr(p), task_pid_nr(p->real_parent));
5352 show_stack(p, NULL);
5355 void show_state_filter(unsigned long state_filter)
5357 struct task_struct *g, *p;
5359 #if BITS_PER_LONG == 32
5361 " task PC stack pid father\n");
5364 " task PC stack pid father\n");
5366 read_lock(&tasklist_lock);
5367 do_each_thread(g, p) {
5369 * reset the NMI-timeout, listing all files on a slow
5370 * console might take alot of time:
5372 touch_nmi_watchdog();
5373 if (!state_filter || (p->state & state_filter))
5375 } while_each_thread(g, p);
5377 touch_all_softlockup_watchdogs();
5379 #ifdef CONFIG_SCHED_DEBUG
5380 sysrq_sched_debug_show();
5382 read_unlock(&tasklist_lock);
5384 * Only show locks if all tasks are dumped:
5386 if (state_filter == -1)
5387 debug_show_all_locks();
5390 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5392 idle->sched_class = &idle_sched_class;
5396 * init_idle - set up an idle thread for a given CPU
5397 * @idle: task in question
5398 * @cpu: cpu the idle task belongs to
5400 * NOTE: this function does not set the idle thread's NEED_RESCHED
5401 * flag, to make booting more robust.
5403 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5405 struct rq *rq = cpu_rq(cpu);
5406 unsigned long flags;
5409 idle->se.exec_start = sched_clock();
5411 idle->prio = idle->normal_prio = MAX_PRIO;
5412 idle->cpus_allowed = cpumask_of_cpu(cpu);
5413 __set_task_cpu(idle, cpu);
5415 spin_lock_irqsave(&rq->lock, flags);
5416 rq->curr = rq->idle = idle;
5417 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5420 spin_unlock_irqrestore(&rq->lock, flags);
5422 /* Set the preempt count _outside_ the spinlocks! */
5423 task_thread_info(idle)->preempt_count = 0;
5426 * The idle tasks have their own, simple scheduling class:
5428 idle->sched_class = &idle_sched_class;
5432 * In a system that switches off the HZ timer nohz_cpu_mask
5433 * indicates which cpus entered this state. This is used
5434 * in the rcu update to wait only for active cpus. For system
5435 * which do not switch off the HZ timer nohz_cpu_mask should
5436 * always be CPU_MASK_NONE.
5438 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5441 * Increase the granularity value when there are more CPUs,
5442 * because with more CPUs the 'effective latency' as visible
5443 * to users decreases. But the relationship is not linear,
5444 * so pick a second-best guess by going with the log2 of the
5447 * This idea comes from the SD scheduler of Con Kolivas:
5449 static inline void sched_init_granularity(void)
5451 unsigned int factor = 1 + ilog2(num_online_cpus());
5452 const unsigned long limit = 200000000;
5454 sysctl_sched_min_granularity *= factor;
5455 if (sysctl_sched_min_granularity > limit)
5456 sysctl_sched_min_granularity = limit;
5458 sysctl_sched_latency *= factor;
5459 if (sysctl_sched_latency > limit)
5460 sysctl_sched_latency = limit;
5462 sysctl_sched_wakeup_granularity *= factor;
5467 * This is how migration works:
5469 * 1) we queue a struct migration_req structure in the source CPU's
5470 * runqueue and wake up that CPU's migration thread.
5471 * 2) we down() the locked semaphore => thread blocks.
5472 * 3) migration thread wakes up (implicitly it forces the migrated
5473 * thread off the CPU)
5474 * 4) it gets the migration request and checks whether the migrated
5475 * task is still in the wrong runqueue.
5476 * 5) if it's in the wrong runqueue then the migration thread removes
5477 * it and puts it into the right queue.
5478 * 6) migration thread up()s the semaphore.
5479 * 7) we wake up and the migration is done.
5483 * Change a given task's CPU affinity. Migrate the thread to a
5484 * proper CPU and schedule it away if the CPU it's executing on
5485 * is removed from the allowed bitmask.
5487 * NOTE: the caller must have a valid reference to the task, the
5488 * task must not exit() & deallocate itself prematurely. The
5489 * call is not atomic; no spinlocks may be held.
5491 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5493 struct migration_req req;
5494 unsigned long flags;
5498 rq = task_rq_lock(p, &flags);
5499 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5504 if (p->sched_class->set_cpus_allowed)
5505 p->sched_class->set_cpus_allowed(p, new_mask);
5507 p->cpus_allowed = *new_mask;
5508 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5511 /* Can the task run on the task's current CPU? If so, we're done */
5512 if (cpu_isset(task_cpu(p), *new_mask))
5515 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5516 /* Need help from migration thread: drop lock and wait. */
5517 task_rq_unlock(rq, &flags);
5518 wake_up_process(rq->migration_thread);
5519 wait_for_completion(&req.done);
5520 tlb_migrate_finish(p->mm);
5524 task_rq_unlock(rq, &flags);
5528 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5531 * Move (not current) task off this cpu, onto dest cpu. We're doing
5532 * this because either it can't run here any more (set_cpus_allowed()
5533 * away from this CPU, or CPU going down), or because we're
5534 * attempting to rebalance this task on exec (sched_exec).
5536 * So we race with normal scheduler movements, but that's OK, as long
5537 * as the task is no longer on this CPU.
5539 * Returns non-zero if task was successfully migrated.
5541 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5543 struct rq *rq_dest, *rq_src;
5546 if (unlikely(cpu_is_offline(dest_cpu)))
5549 rq_src = cpu_rq(src_cpu);
5550 rq_dest = cpu_rq(dest_cpu);
5552 double_rq_lock(rq_src, rq_dest);
5553 /* Already moved. */
5554 if (task_cpu(p) != src_cpu)
5556 /* Affinity changed (again). */
5557 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5560 on_rq = p->se.on_rq;
5562 deactivate_task(rq_src, p, 0);
5564 set_task_cpu(p, dest_cpu);
5566 activate_task(rq_dest, p, 0);
5567 check_preempt_curr(rq_dest, p);
5571 double_rq_unlock(rq_src, rq_dest);
5576 * migration_thread - this is a highprio system thread that performs
5577 * thread migration by bumping thread off CPU then 'pushing' onto
5580 static int migration_thread(void *data)
5582 int cpu = (long)data;
5586 BUG_ON(rq->migration_thread != current);
5588 set_current_state(TASK_INTERRUPTIBLE);
5589 while (!kthread_should_stop()) {
5590 struct migration_req *req;
5591 struct list_head *head;
5593 spin_lock_irq(&rq->lock);
5595 if (cpu_is_offline(cpu)) {
5596 spin_unlock_irq(&rq->lock);
5600 if (rq->active_balance) {
5601 active_load_balance(rq, cpu);
5602 rq->active_balance = 0;
5605 head = &rq->migration_queue;
5607 if (list_empty(head)) {
5608 spin_unlock_irq(&rq->lock);
5610 set_current_state(TASK_INTERRUPTIBLE);
5613 req = list_entry(head->next, struct migration_req, list);
5614 list_del_init(head->next);
5616 spin_unlock(&rq->lock);
5617 __migrate_task(req->task, cpu, req->dest_cpu);
5620 complete(&req->done);
5622 __set_current_state(TASK_RUNNING);
5626 /* Wait for kthread_stop */
5627 set_current_state(TASK_INTERRUPTIBLE);
5628 while (!kthread_should_stop()) {
5630 set_current_state(TASK_INTERRUPTIBLE);
5632 __set_current_state(TASK_RUNNING);
5636 #ifdef CONFIG_HOTPLUG_CPU
5638 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5642 local_irq_disable();
5643 ret = __migrate_task(p, src_cpu, dest_cpu);
5649 * Figure out where task on dead CPU should go, use force if necessary.
5650 * NOTE: interrupts should be disabled by the caller
5652 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5654 unsigned long flags;
5661 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5662 cpus_and(mask, mask, p->cpus_allowed);
5663 dest_cpu = any_online_cpu(mask);
5665 /* On any allowed CPU? */
5666 if (dest_cpu >= nr_cpu_ids)
5667 dest_cpu = any_online_cpu(p->cpus_allowed);
5669 /* No more Mr. Nice Guy. */
5670 if (dest_cpu >= nr_cpu_ids) {
5671 cpumask_t cpus_allowed;
5673 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5675 * Try to stay on the same cpuset, where the
5676 * current cpuset may be a subset of all cpus.
5677 * The cpuset_cpus_allowed_locked() variant of
5678 * cpuset_cpus_allowed() will not block. It must be
5679 * called within calls to cpuset_lock/cpuset_unlock.
5681 rq = task_rq_lock(p, &flags);
5682 p->cpus_allowed = cpus_allowed;
5683 dest_cpu = any_online_cpu(p->cpus_allowed);
5684 task_rq_unlock(rq, &flags);
5687 * Don't tell them about moving exiting tasks or
5688 * kernel threads (both mm NULL), since they never
5691 if (p->mm && printk_ratelimit()) {
5692 printk(KERN_INFO "process %d (%s) no "
5693 "longer affine to cpu%d\n",
5694 task_pid_nr(p), p->comm, dead_cpu);
5697 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5701 * While a dead CPU has no uninterruptible tasks queued at this point,
5702 * it might still have a nonzero ->nr_uninterruptible counter, because
5703 * for performance reasons the counter is not stricly tracking tasks to
5704 * their home CPUs. So we just add the counter to another CPU's counter,
5705 * to keep the global sum constant after CPU-down:
5707 static void migrate_nr_uninterruptible(struct rq *rq_src)
5709 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5710 unsigned long flags;
5712 local_irq_save(flags);
5713 double_rq_lock(rq_src, rq_dest);
5714 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5715 rq_src->nr_uninterruptible = 0;
5716 double_rq_unlock(rq_src, rq_dest);
5717 local_irq_restore(flags);
5720 /* Run through task list and migrate tasks from the dead cpu. */
5721 static void migrate_live_tasks(int src_cpu)
5723 struct task_struct *p, *t;
5725 read_lock(&tasklist_lock);
5727 do_each_thread(t, p) {
5731 if (task_cpu(p) == src_cpu)
5732 move_task_off_dead_cpu(src_cpu, p);
5733 } while_each_thread(t, p);
5735 read_unlock(&tasklist_lock);
5739 * Schedules idle task to be the next runnable task on current CPU.
5740 * It does so by boosting its priority to highest possible.
5741 * Used by CPU offline code.
5743 void sched_idle_next(void)
5745 int this_cpu = smp_processor_id();
5746 struct rq *rq = cpu_rq(this_cpu);
5747 struct task_struct *p = rq->idle;
5748 unsigned long flags;
5750 /* cpu has to be offline */
5751 BUG_ON(cpu_online(this_cpu));
5754 * Strictly not necessary since rest of the CPUs are stopped by now
5755 * and interrupts disabled on the current cpu.
5757 spin_lock_irqsave(&rq->lock, flags);
5759 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5761 update_rq_clock(rq);
5762 activate_task(rq, p, 0);
5764 spin_unlock_irqrestore(&rq->lock, flags);
5768 * Ensures that the idle task is using init_mm right before its cpu goes
5771 void idle_task_exit(void)
5773 struct mm_struct *mm = current->active_mm;
5775 BUG_ON(cpu_online(smp_processor_id()));
5778 switch_mm(mm, &init_mm, current);
5782 /* called under rq->lock with disabled interrupts */
5783 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5785 struct rq *rq = cpu_rq(dead_cpu);
5787 /* Must be exiting, otherwise would be on tasklist. */
5788 BUG_ON(!p->exit_state);
5790 /* Cannot have done final schedule yet: would have vanished. */
5791 BUG_ON(p->state == TASK_DEAD);
5796 * Drop lock around migration; if someone else moves it,
5797 * that's OK. No task can be added to this CPU, so iteration is
5800 spin_unlock_irq(&rq->lock);
5801 move_task_off_dead_cpu(dead_cpu, p);
5802 spin_lock_irq(&rq->lock);
5807 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5808 static void migrate_dead_tasks(unsigned int dead_cpu)
5810 struct rq *rq = cpu_rq(dead_cpu);
5811 struct task_struct *next;
5814 if (!rq->nr_running)
5816 update_rq_clock(rq);
5817 next = pick_next_task(rq, rq->curr);
5820 migrate_dead(dead_cpu, next);
5824 #endif /* CONFIG_HOTPLUG_CPU */
5826 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5828 static struct ctl_table sd_ctl_dir[] = {
5830 .procname = "sched_domain",
5836 static struct ctl_table sd_ctl_root[] = {
5838 .ctl_name = CTL_KERN,
5839 .procname = "kernel",
5841 .child = sd_ctl_dir,
5846 static struct ctl_table *sd_alloc_ctl_entry(int n)
5848 struct ctl_table *entry =
5849 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5854 static void sd_free_ctl_entry(struct ctl_table **tablep)
5856 struct ctl_table *entry;
5859 * In the intermediate directories, both the child directory and
5860 * procname are dynamically allocated and could fail but the mode
5861 * will always be set. In the lowest directory the names are
5862 * static strings and all have proc handlers.
5864 for (entry = *tablep; entry->mode; entry++) {
5866 sd_free_ctl_entry(&entry->child);
5867 if (entry->proc_handler == NULL)
5868 kfree(entry->procname);
5876 set_table_entry(struct ctl_table *entry,
5877 const char *procname, void *data, int maxlen,
5878 mode_t mode, proc_handler *proc_handler)
5880 entry->procname = procname;
5882 entry->maxlen = maxlen;
5884 entry->proc_handler = proc_handler;
5887 static struct ctl_table *
5888 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5890 struct ctl_table *table = sd_alloc_ctl_entry(12);
5895 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5896 sizeof(long), 0644, proc_doulongvec_minmax);
5897 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5898 sizeof(long), 0644, proc_doulongvec_minmax);
5899 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5900 sizeof(int), 0644, proc_dointvec_minmax);
5901 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5902 sizeof(int), 0644, proc_dointvec_minmax);
5903 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5904 sizeof(int), 0644, proc_dointvec_minmax);
5905 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5906 sizeof(int), 0644, proc_dointvec_minmax);
5907 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5908 sizeof(int), 0644, proc_dointvec_minmax);
5909 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5910 sizeof(int), 0644, proc_dointvec_minmax);
5911 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5912 sizeof(int), 0644, proc_dointvec_minmax);
5913 set_table_entry(&table[9], "cache_nice_tries",
5914 &sd->cache_nice_tries,
5915 sizeof(int), 0644, proc_dointvec_minmax);
5916 set_table_entry(&table[10], "flags", &sd->flags,
5917 sizeof(int), 0644, proc_dointvec_minmax);
5918 /* &table[11] is terminator */
5923 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5925 struct ctl_table *entry, *table;
5926 struct sched_domain *sd;
5927 int domain_num = 0, i;
5930 for_each_domain(cpu, sd)
5932 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5937 for_each_domain(cpu, sd) {
5938 snprintf(buf, 32, "domain%d", i);
5939 entry->procname = kstrdup(buf, GFP_KERNEL);
5941 entry->child = sd_alloc_ctl_domain_table(sd);
5948 static struct ctl_table_header *sd_sysctl_header;
5949 static void register_sched_domain_sysctl(void)
5951 int i, cpu_num = num_online_cpus();
5952 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5955 WARN_ON(sd_ctl_dir[0].child);
5956 sd_ctl_dir[0].child = entry;
5961 for_each_online_cpu(i) {
5962 snprintf(buf, 32, "cpu%d", i);
5963 entry->procname = kstrdup(buf, GFP_KERNEL);
5965 entry->child = sd_alloc_ctl_cpu_table(i);
5969 WARN_ON(sd_sysctl_header);
5970 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5973 /* may be called multiple times per register */
5974 static void unregister_sched_domain_sysctl(void)
5976 if (sd_sysctl_header)
5977 unregister_sysctl_table(sd_sysctl_header);
5978 sd_sysctl_header = NULL;
5979 if (sd_ctl_dir[0].child)
5980 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5983 static void register_sched_domain_sysctl(void)
5986 static void unregister_sched_domain_sysctl(void)
5992 * migration_call - callback that gets triggered when a CPU is added.
5993 * Here we can start up the necessary migration thread for the new CPU.
5995 static int __cpuinit
5996 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5998 struct task_struct *p;
5999 int cpu = (long)hcpu;
6000 unsigned long flags;
6005 case CPU_UP_PREPARE:
6006 case CPU_UP_PREPARE_FROZEN:
6007 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6010 kthread_bind(p, cpu);
6011 /* Must be high prio: stop_machine expects to yield to it. */
6012 rq = task_rq_lock(p, &flags);
6013 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6014 task_rq_unlock(rq, &flags);
6015 cpu_rq(cpu)->migration_thread = p;
6019 case CPU_ONLINE_FROZEN:
6020 /* Strictly unnecessary, as first user will wake it. */
6021 wake_up_process(cpu_rq(cpu)->migration_thread);
6023 /* Update our root-domain */
6025 spin_lock_irqsave(&rq->lock, flags);
6027 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6028 cpu_set(cpu, rq->rd->online);
6030 spin_unlock_irqrestore(&rq->lock, flags);
6033 #ifdef CONFIG_HOTPLUG_CPU
6034 case CPU_UP_CANCELED:
6035 case CPU_UP_CANCELED_FROZEN:
6036 if (!cpu_rq(cpu)->migration_thread)
6038 /* Unbind it from offline cpu so it can run. Fall thru. */
6039 kthread_bind(cpu_rq(cpu)->migration_thread,
6040 any_online_cpu(cpu_online_map));
6041 kthread_stop(cpu_rq(cpu)->migration_thread);
6042 cpu_rq(cpu)->migration_thread = NULL;
6046 case CPU_DEAD_FROZEN:
6047 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6048 migrate_live_tasks(cpu);
6050 kthread_stop(rq->migration_thread);
6051 rq->migration_thread = NULL;
6052 /* Idle task back to normal (off runqueue, low prio) */
6053 spin_lock_irq(&rq->lock);
6054 update_rq_clock(rq);
6055 deactivate_task(rq, rq->idle, 0);
6056 rq->idle->static_prio = MAX_PRIO;
6057 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6058 rq->idle->sched_class = &idle_sched_class;
6059 migrate_dead_tasks(cpu);
6060 spin_unlock_irq(&rq->lock);
6062 migrate_nr_uninterruptible(rq);
6063 BUG_ON(rq->nr_running != 0);
6066 * No need to migrate the tasks: it was best-effort if
6067 * they didn't take sched_hotcpu_mutex. Just wake up
6070 spin_lock_irq(&rq->lock);
6071 while (!list_empty(&rq->migration_queue)) {
6072 struct migration_req *req;
6074 req = list_entry(rq->migration_queue.next,
6075 struct migration_req, list);
6076 list_del_init(&req->list);
6077 complete(&req->done);
6079 spin_unlock_irq(&rq->lock);
6083 case CPU_DYING_FROZEN:
6084 /* Update our root-domain */
6086 spin_lock_irqsave(&rq->lock, flags);
6088 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6089 cpu_clear(cpu, rq->rd->online);
6091 spin_unlock_irqrestore(&rq->lock, flags);
6098 /* Register at highest priority so that task migration (migrate_all_tasks)
6099 * happens before everything else.
6101 static struct notifier_block __cpuinitdata migration_notifier = {
6102 .notifier_call = migration_call,
6106 void __init migration_init(void)
6108 void *cpu = (void *)(long)smp_processor_id();
6111 /* Start one for the boot CPU: */
6112 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6113 BUG_ON(err == NOTIFY_BAD);
6114 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6115 register_cpu_notifier(&migration_notifier);
6121 #ifdef CONFIG_SCHED_DEBUG
6123 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6124 cpumask_t *groupmask)
6126 struct sched_group *group = sd->groups;
6129 cpulist_scnprintf(str, sizeof(str), sd->span);
6130 cpus_clear(*groupmask);
6132 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6134 if (!(sd->flags & SD_LOAD_BALANCE)) {
6135 printk("does not load-balance\n");
6137 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6142 printk(KERN_CONT "span %s\n", str);
6144 if (!cpu_isset(cpu, sd->span)) {
6145 printk(KERN_ERR "ERROR: domain->span does not contain "
6148 if (!cpu_isset(cpu, group->cpumask)) {
6149 printk(KERN_ERR "ERROR: domain->groups does not contain"
6153 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6157 printk(KERN_ERR "ERROR: group is NULL\n");
6161 if (!group->__cpu_power) {
6162 printk(KERN_CONT "\n");
6163 printk(KERN_ERR "ERROR: domain->cpu_power not "
6168 if (!cpus_weight(group->cpumask)) {
6169 printk(KERN_CONT "\n");
6170 printk(KERN_ERR "ERROR: empty group\n");
6174 if (cpus_intersects(*groupmask, group->cpumask)) {
6175 printk(KERN_CONT "\n");
6176 printk(KERN_ERR "ERROR: repeated CPUs\n");
6180 cpus_or(*groupmask, *groupmask, group->cpumask);
6182 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6183 printk(KERN_CONT " %s", str);
6185 group = group->next;
6186 } while (group != sd->groups);
6187 printk(KERN_CONT "\n");
6189 if (!cpus_equal(sd->span, *groupmask))
6190 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6192 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6193 printk(KERN_ERR "ERROR: parent span is not a superset "
6194 "of domain->span\n");
6198 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6200 cpumask_t *groupmask;
6204 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6208 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6210 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6212 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6217 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6227 # define sched_domain_debug(sd, cpu) do { } while (0)
6230 static int sd_degenerate(struct sched_domain *sd)
6232 if (cpus_weight(sd->span) == 1)
6235 /* Following flags need at least 2 groups */
6236 if (sd->flags & (SD_LOAD_BALANCE |
6237 SD_BALANCE_NEWIDLE |
6241 SD_SHARE_PKG_RESOURCES)) {
6242 if (sd->groups != sd->groups->next)
6246 /* Following flags don't use groups */
6247 if (sd->flags & (SD_WAKE_IDLE |
6256 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6258 unsigned long cflags = sd->flags, pflags = parent->flags;
6260 if (sd_degenerate(parent))
6263 if (!cpus_equal(sd->span, parent->span))
6266 /* Does parent contain flags not in child? */
6267 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6268 if (cflags & SD_WAKE_AFFINE)
6269 pflags &= ~SD_WAKE_BALANCE;
6270 /* Flags needing groups don't count if only 1 group in parent */
6271 if (parent->groups == parent->groups->next) {
6272 pflags &= ~(SD_LOAD_BALANCE |
6273 SD_BALANCE_NEWIDLE |
6277 SD_SHARE_PKG_RESOURCES);
6279 if (~cflags & pflags)
6285 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6287 unsigned long flags;
6288 const struct sched_class *class;
6290 spin_lock_irqsave(&rq->lock, flags);
6293 struct root_domain *old_rd = rq->rd;
6295 for (class = sched_class_highest; class; class = class->next) {
6296 if (class->leave_domain)
6297 class->leave_domain(rq);
6300 cpu_clear(rq->cpu, old_rd->span);
6301 cpu_clear(rq->cpu, old_rd->online);
6303 if (atomic_dec_and_test(&old_rd->refcount))
6307 atomic_inc(&rd->refcount);
6310 cpu_set(rq->cpu, rd->span);
6311 if (cpu_isset(rq->cpu, cpu_online_map))
6312 cpu_set(rq->cpu, rd->online);
6314 for (class = sched_class_highest; class; class = class->next) {
6315 if (class->join_domain)
6316 class->join_domain(rq);
6319 spin_unlock_irqrestore(&rq->lock, flags);
6322 static void init_rootdomain(struct root_domain *rd)
6324 memset(rd, 0, sizeof(*rd));
6326 cpus_clear(rd->span);
6327 cpus_clear(rd->online);
6330 static void init_defrootdomain(void)
6332 init_rootdomain(&def_root_domain);
6333 atomic_set(&def_root_domain.refcount, 1);
6336 static struct root_domain *alloc_rootdomain(void)
6338 struct root_domain *rd;
6340 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6344 init_rootdomain(rd);
6350 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6351 * hold the hotplug lock.
6354 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6356 struct rq *rq = cpu_rq(cpu);
6357 struct sched_domain *tmp;
6359 /* Remove the sched domains which do not contribute to scheduling. */
6360 for (tmp = sd; tmp; tmp = tmp->parent) {
6361 struct sched_domain *parent = tmp->parent;
6364 if (sd_parent_degenerate(tmp, parent)) {
6365 tmp->parent = parent->parent;
6367 parent->parent->child = tmp;
6371 if (sd && sd_degenerate(sd)) {
6377 sched_domain_debug(sd, cpu);
6379 rq_attach_root(rq, rd);
6380 rcu_assign_pointer(rq->sd, sd);
6383 /* cpus with isolated domains */
6384 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6386 /* Setup the mask of cpus configured for isolated domains */
6387 static int __init isolated_cpu_setup(char *str)
6389 int ints[NR_CPUS], i;
6391 str = get_options(str, ARRAY_SIZE(ints), ints);
6392 cpus_clear(cpu_isolated_map);
6393 for (i = 1; i <= ints[0]; i++)
6394 if (ints[i] < NR_CPUS)
6395 cpu_set(ints[i], cpu_isolated_map);
6399 __setup("isolcpus=", isolated_cpu_setup);
6402 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6403 * to a function which identifies what group(along with sched group) a CPU
6404 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6405 * (due to the fact that we keep track of groups covered with a cpumask_t).
6407 * init_sched_build_groups will build a circular linked list of the groups
6408 * covered by the given span, and will set each group's ->cpumask correctly,
6409 * and ->cpu_power to 0.
6412 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6413 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6414 struct sched_group **sg,
6415 cpumask_t *tmpmask),
6416 cpumask_t *covered, cpumask_t *tmpmask)
6418 struct sched_group *first = NULL, *last = NULL;
6421 cpus_clear(*covered);
6423 for_each_cpu_mask(i, *span) {
6424 struct sched_group *sg;
6425 int group = group_fn(i, cpu_map, &sg, tmpmask);
6428 if (cpu_isset(i, *covered))
6431 cpus_clear(sg->cpumask);
6432 sg->__cpu_power = 0;
6434 for_each_cpu_mask(j, *span) {
6435 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6438 cpu_set(j, *covered);
6439 cpu_set(j, sg->cpumask);
6450 #define SD_NODES_PER_DOMAIN 16
6455 * find_next_best_node - find the next node to include in a sched_domain
6456 * @node: node whose sched_domain we're building
6457 * @used_nodes: nodes already in the sched_domain
6459 * Find the next node to include in a given scheduling domain. Simply
6460 * finds the closest node not already in the @used_nodes map.
6462 * Should use nodemask_t.
6464 static int find_next_best_node(int node, nodemask_t *used_nodes)
6466 int i, n, val, min_val, best_node = 0;
6470 for (i = 0; i < MAX_NUMNODES; i++) {
6471 /* Start at @node */
6472 n = (node + i) % MAX_NUMNODES;
6474 if (!nr_cpus_node(n))
6477 /* Skip already used nodes */
6478 if (node_isset(n, *used_nodes))
6481 /* Simple min distance search */
6482 val = node_distance(node, n);
6484 if (val < min_val) {
6490 node_set(best_node, *used_nodes);
6495 * sched_domain_node_span - get a cpumask for a node's sched_domain
6496 * @node: node whose cpumask we're constructing
6498 * Given a node, construct a good cpumask for its sched_domain to span. It
6499 * should be one that prevents unnecessary balancing, but also spreads tasks
6502 static void sched_domain_node_span(int node, cpumask_t *span)
6504 nodemask_t used_nodes;
6505 node_to_cpumask_ptr(nodemask, node);
6509 nodes_clear(used_nodes);
6511 cpus_or(*span, *span, *nodemask);
6512 node_set(node, used_nodes);
6514 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6515 int next_node = find_next_best_node(node, &used_nodes);
6517 node_to_cpumask_ptr_next(nodemask, next_node);
6518 cpus_or(*span, *span, *nodemask);
6523 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6526 * SMT sched-domains:
6528 #ifdef CONFIG_SCHED_SMT
6529 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6530 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6533 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6537 *sg = &per_cpu(sched_group_cpus, cpu);
6543 * multi-core sched-domains:
6545 #ifdef CONFIG_SCHED_MC
6546 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6547 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6550 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6552 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6557 *mask = per_cpu(cpu_sibling_map, cpu);
6558 cpus_and(*mask, *mask, *cpu_map);
6559 group = first_cpu(*mask);
6561 *sg = &per_cpu(sched_group_core, group);
6564 #elif defined(CONFIG_SCHED_MC)
6566 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6570 *sg = &per_cpu(sched_group_core, cpu);
6575 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6576 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6579 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6583 #ifdef CONFIG_SCHED_MC
6584 *mask = cpu_coregroup_map(cpu);
6585 cpus_and(*mask, *mask, *cpu_map);
6586 group = first_cpu(*mask);
6587 #elif defined(CONFIG_SCHED_SMT)
6588 *mask = per_cpu(cpu_sibling_map, cpu);
6589 cpus_and(*mask, *mask, *cpu_map);
6590 group = first_cpu(*mask);
6595 *sg = &per_cpu(sched_group_phys, group);
6601 * The init_sched_build_groups can't handle what we want to do with node
6602 * groups, so roll our own. Now each node has its own list of groups which
6603 * gets dynamically allocated.
6605 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6606 static struct sched_group ***sched_group_nodes_bycpu;
6608 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6609 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6611 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6612 struct sched_group **sg, cpumask_t *nodemask)
6616 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6617 cpus_and(*nodemask, *nodemask, *cpu_map);
6618 group = first_cpu(*nodemask);
6621 *sg = &per_cpu(sched_group_allnodes, group);
6625 static void init_numa_sched_groups_power(struct sched_group *group_head)
6627 struct sched_group *sg = group_head;
6633 for_each_cpu_mask(j, sg->cpumask) {
6634 struct sched_domain *sd;
6636 sd = &per_cpu(phys_domains, j);
6637 if (j != first_cpu(sd->groups->cpumask)) {
6639 * Only add "power" once for each
6645 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6648 } while (sg != group_head);
6653 /* Free memory allocated for various sched_group structures */
6654 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6658 for_each_cpu_mask(cpu, *cpu_map) {
6659 struct sched_group **sched_group_nodes
6660 = sched_group_nodes_bycpu[cpu];
6662 if (!sched_group_nodes)
6665 for (i = 0; i < MAX_NUMNODES; i++) {
6666 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6668 *nodemask = node_to_cpumask(i);
6669 cpus_and(*nodemask, *nodemask, *cpu_map);
6670 if (cpus_empty(*nodemask))
6680 if (oldsg != sched_group_nodes[i])
6683 kfree(sched_group_nodes);
6684 sched_group_nodes_bycpu[cpu] = NULL;
6688 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6694 * Initialize sched groups cpu_power.
6696 * cpu_power indicates the capacity of sched group, which is used while
6697 * distributing the load between different sched groups in a sched domain.
6698 * Typically cpu_power for all the groups in a sched domain will be same unless
6699 * there are asymmetries in the topology. If there are asymmetries, group
6700 * having more cpu_power will pickup more load compared to the group having
6703 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6704 * the maximum number of tasks a group can handle in the presence of other idle
6705 * or lightly loaded groups in the same sched domain.
6707 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6709 struct sched_domain *child;
6710 struct sched_group *group;
6712 WARN_ON(!sd || !sd->groups);
6714 if (cpu != first_cpu(sd->groups->cpumask))
6719 sd->groups->__cpu_power = 0;
6722 * For perf policy, if the groups in child domain share resources
6723 * (for example cores sharing some portions of the cache hierarchy
6724 * or SMT), then set this domain groups cpu_power such that each group
6725 * can handle only one task, when there are other idle groups in the
6726 * same sched domain.
6728 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6730 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6731 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6736 * add cpu_power of each child group to this groups cpu_power
6738 group = child->groups;
6740 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6741 group = group->next;
6742 } while (group != child->groups);
6746 * Initializers for schedule domains
6747 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6750 #define SD_INIT(sd, type) sd_init_##type(sd)
6751 #define SD_INIT_FUNC(type) \
6752 static noinline void sd_init_##type(struct sched_domain *sd) \
6754 memset(sd, 0, sizeof(*sd)); \
6755 *sd = SD_##type##_INIT; \
6760 SD_INIT_FUNC(ALLNODES)
6763 #ifdef CONFIG_SCHED_SMT
6764 SD_INIT_FUNC(SIBLING)
6766 #ifdef CONFIG_SCHED_MC
6771 * To minimize stack usage kmalloc room for cpumasks and share the
6772 * space as the usage in build_sched_domains() dictates. Used only
6773 * if the amount of space is significant.
6776 cpumask_t tmpmask; /* make this one first */
6779 cpumask_t this_sibling_map;
6780 cpumask_t this_core_map;
6782 cpumask_t send_covered;
6785 cpumask_t domainspan;
6787 cpumask_t notcovered;
6792 #define SCHED_CPUMASK_ALLOC 1
6793 #define SCHED_CPUMASK_FREE(v) kfree(v)
6794 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6796 #define SCHED_CPUMASK_ALLOC 0
6797 #define SCHED_CPUMASK_FREE(v)
6798 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6801 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6802 ((unsigned long)(a) + offsetof(struct allmasks, v))
6805 * Build sched domains for a given set of cpus and attach the sched domains
6806 * to the individual cpus
6808 static int build_sched_domains(const cpumask_t *cpu_map)
6811 struct root_domain *rd;
6812 SCHED_CPUMASK_DECLARE(allmasks);
6815 struct sched_group **sched_group_nodes = NULL;
6816 int sd_allnodes = 0;
6819 * Allocate the per-node list of sched groups
6821 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6823 if (!sched_group_nodes) {
6824 printk(KERN_WARNING "Can not alloc sched group node list\n");
6829 rd = alloc_rootdomain();
6831 printk(KERN_WARNING "Cannot alloc root domain\n");
6833 kfree(sched_group_nodes);
6838 #if SCHED_CPUMASK_ALLOC
6839 /* get space for all scratch cpumask variables */
6840 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6842 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6845 kfree(sched_group_nodes);
6850 tmpmask = (cpumask_t *)allmasks;
6854 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6858 * Set up domains for cpus specified by the cpu_map.
6860 for_each_cpu_mask(i, *cpu_map) {
6861 struct sched_domain *sd = NULL, *p;
6862 SCHED_CPUMASK_VAR(nodemask, allmasks);
6864 *nodemask = node_to_cpumask(cpu_to_node(i));
6865 cpus_and(*nodemask, *nodemask, *cpu_map);
6868 if (cpus_weight(*cpu_map) >
6869 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6870 sd = &per_cpu(allnodes_domains, i);
6871 SD_INIT(sd, ALLNODES);
6872 sd->span = *cpu_map;
6873 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6879 sd = &per_cpu(node_domains, i);
6881 sched_domain_node_span(cpu_to_node(i), &sd->span);
6885 cpus_and(sd->span, sd->span, *cpu_map);
6889 sd = &per_cpu(phys_domains, i);
6891 sd->span = *nodemask;
6895 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6897 #ifdef CONFIG_SCHED_MC
6899 sd = &per_cpu(core_domains, i);
6901 sd->span = cpu_coregroup_map(i);
6902 cpus_and(sd->span, sd->span, *cpu_map);
6905 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
6908 #ifdef CONFIG_SCHED_SMT
6910 sd = &per_cpu(cpu_domains, i);
6911 SD_INIT(sd, SIBLING);
6912 sd->span = per_cpu(cpu_sibling_map, i);
6913 cpus_and(sd->span, sd->span, *cpu_map);
6916 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
6920 #ifdef CONFIG_SCHED_SMT
6921 /* Set up CPU (sibling) groups */
6922 for_each_cpu_mask(i, *cpu_map) {
6923 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
6924 SCHED_CPUMASK_VAR(send_covered, allmasks);
6926 *this_sibling_map = per_cpu(cpu_sibling_map, i);
6927 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
6928 if (i != first_cpu(*this_sibling_map))
6931 init_sched_build_groups(this_sibling_map, cpu_map,
6933 send_covered, tmpmask);
6937 #ifdef CONFIG_SCHED_MC
6938 /* Set up multi-core groups */
6939 for_each_cpu_mask(i, *cpu_map) {
6940 SCHED_CPUMASK_VAR(this_core_map, allmasks);
6941 SCHED_CPUMASK_VAR(send_covered, allmasks);
6943 *this_core_map = cpu_coregroup_map(i);
6944 cpus_and(*this_core_map, *this_core_map, *cpu_map);
6945 if (i != first_cpu(*this_core_map))
6948 init_sched_build_groups(this_core_map, cpu_map,
6950 send_covered, tmpmask);
6954 /* Set up physical groups */
6955 for (i = 0; i < MAX_NUMNODES; i++) {
6956 SCHED_CPUMASK_VAR(nodemask, allmasks);
6957 SCHED_CPUMASK_VAR(send_covered, allmasks);
6959 *nodemask = node_to_cpumask(i);
6960 cpus_and(*nodemask, *nodemask, *cpu_map);
6961 if (cpus_empty(*nodemask))
6964 init_sched_build_groups(nodemask, cpu_map,
6966 send_covered, tmpmask);
6970 /* Set up node groups */
6972 SCHED_CPUMASK_VAR(send_covered, allmasks);
6974 init_sched_build_groups(cpu_map, cpu_map,
6975 &cpu_to_allnodes_group,
6976 send_covered, tmpmask);
6979 for (i = 0; i < MAX_NUMNODES; i++) {
6980 /* Set up node groups */
6981 struct sched_group *sg, *prev;
6982 SCHED_CPUMASK_VAR(nodemask, allmasks);
6983 SCHED_CPUMASK_VAR(domainspan, allmasks);
6984 SCHED_CPUMASK_VAR(covered, allmasks);
6987 *nodemask = node_to_cpumask(i);
6988 cpus_clear(*covered);
6990 cpus_and(*nodemask, *nodemask, *cpu_map);
6991 if (cpus_empty(*nodemask)) {
6992 sched_group_nodes[i] = NULL;
6996 sched_domain_node_span(i, domainspan);
6997 cpus_and(*domainspan, *domainspan, *cpu_map);
6999 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7001 printk(KERN_WARNING "Can not alloc domain group for "
7005 sched_group_nodes[i] = sg;
7006 for_each_cpu_mask(j, *nodemask) {
7007 struct sched_domain *sd;
7009 sd = &per_cpu(node_domains, j);
7012 sg->__cpu_power = 0;
7013 sg->cpumask = *nodemask;
7015 cpus_or(*covered, *covered, *nodemask);
7018 for (j = 0; j < MAX_NUMNODES; j++) {
7019 SCHED_CPUMASK_VAR(notcovered, allmasks);
7020 int n = (i + j) % MAX_NUMNODES;
7021 node_to_cpumask_ptr(pnodemask, n);
7023 cpus_complement(*notcovered, *covered);
7024 cpus_and(*tmpmask, *notcovered, *cpu_map);
7025 cpus_and(*tmpmask, *tmpmask, *domainspan);
7026 if (cpus_empty(*tmpmask))
7029 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7030 if (cpus_empty(*tmpmask))
7033 sg = kmalloc_node(sizeof(struct sched_group),
7037 "Can not alloc domain group for node %d\n", j);
7040 sg->__cpu_power = 0;
7041 sg->cpumask = *tmpmask;
7042 sg->next = prev->next;
7043 cpus_or(*covered, *covered, *tmpmask);
7050 /* Calculate CPU power for physical packages and nodes */
7051 #ifdef CONFIG_SCHED_SMT
7052 for_each_cpu_mask(i, *cpu_map) {
7053 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7055 init_sched_groups_power(i, sd);
7058 #ifdef CONFIG_SCHED_MC
7059 for_each_cpu_mask(i, *cpu_map) {
7060 struct sched_domain *sd = &per_cpu(core_domains, i);
7062 init_sched_groups_power(i, sd);
7066 for_each_cpu_mask(i, *cpu_map) {
7067 struct sched_domain *sd = &per_cpu(phys_domains, i);
7069 init_sched_groups_power(i, sd);
7073 for (i = 0; i < MAX_NUMNODES; i++)
7074 init_numa_sched_groups_power(sched_group_nodes[i]);
7077 struct sched_group *sg;
7079 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7081 init_numa_sched_groups_power(sg);
7085 /* Attach the domains */
7086 for_each_cpu_mask(i, *cpu_map) {
7087 struct sched_domain *sd;
7088 #ifdef CONFIG_SCHED_SMT
7089 sd = &per_cpu(cpu_domains, i);
7090 #elif defined(CONFIG_SCHED_MC)
7091 sd = &per_cpu(core_domains, i);
7093 sd = &per_cpu(phys_domains, i);
7095 cpu_attach_domain(sd, rd, i);
7098 SCHED_CPUMASK_FREE((void *)allmasks);
7103 free_sched_groups(cpu_map, tmpmask);
7104 SCHED_CPUMASK_FREE((void *)allmasks);
7109 static cpumask_t *doms_cur; /* current sched domains */
7110 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7113 * Special case: If a kmalloc of a doms_cur partition (array of
7114 * cpumask_t) fails, then fallback to a single sched domain,
7115 * as determined by the single cpumask_t fallback_doms.
7117 static cpumask_t fallback_doms;
7119 void __attribute__((weak)) arch_update_cpu_topology(void)
7124 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7125 * For now this just excludes isolated cpus, but could be used to
7126 * exclude other special cases in the future.
7128 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7132 arch_update_cpu_topology();
7134 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7136 doms_cur = &fallback_doms;
7137 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7138 err = build_sched_domains(doms_cur);
7139 register_sched_domain_sysctl();
7144 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7147 free_sched_groups(cpu_map, tmpmask);
7151 * Detach sched domains from a group of cpus specified in cpu_map
7152 * These cpus will now be attached to the NULL domain
7154 static void detach_destroy_domains(const cpumask_t *cpu_map)
7159 unregister_sched_domain_sysctl();
7161 for_each_cpu_mask(i, *cpu_map)
7162 cpu_attach_domain(NULL, &def_root_domain, i);
7163 synchronize_sched();
7164 arch_destroy_sched_domains(cpu_map, &tmpmask);
7168 * Partition sched domains as specified by the 'ndoms_new'
7169 * cpumasks in the array doms_new[] of cpumasks. This compares
7170 * doms_new[] to the current sched domain partitioning, doms_cur[].
7171 * It destroys each deleted domain and builds each new domain.
7173 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7174 * The masks don't intersect (don't overlap.) We should setup one
7175 * sched domain for each mask. CPUs not in any of the cpumasks will
7176 * not be load balanced. If the same cpumask appears both in the
7177 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7180 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7181 * ownership of it and will kfree it when done with it. If the caller
7182 * failed the kmalloc call, then it can pass in doms_new == NULL,
7183 * and partition_sched_domains() will fallback to the single partition
7186 * Call with hotplug lock held
7188 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7194 /* always unregister in case we don't destroy any domains */
7195 unregister_sched_domain_sysctl();
7197 if (doms_new == NULL) {
7199 doms_new = &fallback_doms;
7200 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7203 /* Destroy deleted domains */
7204 for (i = 0; i < ndoms_cur; i++) {
7205 for (j = 0; j < ndoms_new; j++) {
7206 if (cpus_equal(doms_cur[i], doms_new[j]))
7209 /* no match - a current sched domain not in new doms_new[] */
7210 detach_destroy_domains(doms_cur + i);
7215 /* Build new domains */
7216 for (i = 0; i < ndoms_new; i++) {
7217 for (j = 0; j < ndoms_cur; j++) {
7218 if (cpus_equal(doms_new[i], doms_cur[j]))
7221 /* no match - add a new doms_new */
7222 build_sched_domains(doms_new + i);
7227 /* Remember the new sched domains */
7228 if (doms_cur != &fallback_doms)
7230 doms_cur = doms_new;
7231 ndoms_cur = ndoms_new;
7233 register_sched_domain_sysctl();
7238 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7239 int arch_reinit_sched_domains(void)
7244 detach_destroy_domains(&cpu_online_map);
7245 err = arch_init_sched_domains(&cpu_online_map);
7251 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7255 if (buf[0] != '0' && buf[0] != '1')
7259 sched_smt_power_savings = (buf[0] == '1');
7261 sched_mc_power_savings = (buf[0] == '1');
7263 ret = arch_reinit_sched_domains();
7265 return ret ? ret : count;
7268 #ifdef CONFIG_SCHED_MC
7269 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7271 return sprintf(page, "%u\n", sched_mc_power_savings);
7273 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7274 const char *buf, size_t count)
7276 return sched_power_savings_store(buf, count, 0);
7278 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7279 sched_mc_power_savings_store);
7282 #ifdef CONFIG_SCHED_SMT
7283 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7285 return sprintf(page, "%u\n", sched_smt_power_savings);
7287 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7288 const char *buf, size_t count)
7290 return sched_power_savings_store(buf, count, 1);
7292 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7293 sched_smt_power_savings_store);
7296 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7300 #ifdef CONFIG_SCHED_SMT
7302 err = sysfs_create_file(&cls->kset.kobj,
7303 &attr_sched_smt_power_savings.attr);
7305 #ifdef CONFIG_SCHED_MC
7306 if (!err && mc_capable())
7307 err = sysfs_create_file(&cls->kset.kobj,
7308 &attr_sched_mc_power_savings.attr);
7315 * Force a reinitialization of the sched domains hierarchy. The domains
7316 * and groups cannot be updated in place without racing with the balancing
7317 * code, so we temporarily attach all running cpus to the NULL domain
7318 * which will prevent rebalancing while the sched domains are recalculated.
7320 static int update_sched_domains(struct notifier_block *nfb,
7321 unsigned long action, void *hcpu)
7324 case CPU_UP_PREPARE:
7325 case CPU_UP_PREPARE_FROZEN:
7326 case CPU_DOWN_PREPARE:
7327 case CPU_DOWN_PREPARE_FROZEN:
7328 detach_destroy_domains(&cpu_online_map);
7331 case CPU_UP_CANCELED:
7332 case CPU_UP_CANCELED_FROZEN:
7333 case CPU_DOWN_FAILED:
7334 case CPU_DOWN_FAILED_FROZEN:
7336 case CPU_ONLINE_FROZEN:
7338 case CPU_DEAD_FROZEN:
7340 * Fall through and re-initialise the domains.
7347 /* The hotplug lock is already held by cpu_up/cpu_down */
7348 arch_init_sched_domains(&cpu_online_map);
7353 void __init sched_init_smp(void)
7355 cpumask_t non_isolated_cpus;
7357 #if defined(CONFIG_NUMA)
7358 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7360 BUG_ON(sched_group_nodes_bycpu == NULL);
7363 arch_init_sched_domains(&cpu_online_map);
7364 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7365 if (cpus_empty(non_isolated_cpus))
7366 cpu_set(smp_processor_id(), non_isolated_cpus);
7368 /* XXX: Theoretical race here - CPU may be hotplugged now */
7369 hotcpu_notifier(update_sched_domains, 0);
7371 /* Move init over to a non-isolated CPU */
7372 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7374 sched_init_granularity();
7377 void __init sched_init_smp(void)
7379 #if defined(CONFIG_NUMA)
7380 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7382 BUG_ON(sched_group_nodes_bycpu == NULL);
7384 sched_init_granularity();
7386 #endif /* CONFIG_SMP */
7388 int in_sched_functions(unsigned long addr)
7390 return in_lock_functions(addr) ||
7391 (addr >= (unsigned long)__sched_text_start
7392 && addr < (unsigned long)__sched_text_end);
7395 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7397 cfs_rq->tasks_timeline = RB_ROOT;
7398 #ifdef CONFIG_FAIR_GROUP_SCHED
7401 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7404 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7406 struct rt_prio_array *array;
7409 array = &rt_rq->active;
7410 for (i = 0; i < MAX_RT_PRIO; i++) {
7411 INIT_LIST_HEAD(array->queue + i);
7412 __clear_bit(i, array->bitmap);
7414 /* delimiter for bitsearch: */
7415 __set_bit(MAX_RT_PRIO, array->bitmap);
7417 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7418 rt_rq->highest_prio = MAX_RT_PRIO;
7421 rt_rq->rt_nr_migratory = 0;
7422 rt_rq->overloaded = 0;
7426 rt_rq->rt_throttled = 0;
7427 rt_rq->rt_runtime = 0;
7428 spin_lock_init(&rt_rq->rt_runtime_lock);
7430 #ifdef CONFIG_RT_GROUP_SCHED
7431 rt_rq->rt_nr_boosted = 0;
7436 #ifdef CONFIG_FAIR_GROUP_SCHED
7437 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7438 struct cfs_rq *cfs_rq, struct sched_entity *se,
7441 tg->cfs_rq[cpu] = cfs_rq;
7442 init_cfs_rq(cfs_rq, rq);
7445 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7448 se->cfs_rq = &rq->cfs;
7450 se->load.weight = tg->shares;
7451 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7456 #ifdef CONFIG_RT_GROUP_SCHED
7457 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7458 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7461 tg->rt_rq[cpu] = rt_rq;
7462 init_rt_rq(rt_rq, rq);
7464 rt_rq->rt_se = rt_se;
7465 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7467 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7469 tg->rt_se[cpu] = rt_se;
7470 rt_se->rt_rq = &rq->rt;
7471 rt_se->my_q = rt_rq;
7472 rt_se->parent = NULL;
7473 INIT_LIST_HEAD(&rt_se->run_list);
7477 void __init sched_init(void)
7480 unsigned long alloc_size = 0, ptr;
7482 #ifdef CONFIG_FAIR_GROUP_SCHED
7483 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7485 #ifdef CONFIG_RT_GROUP_SCHED
7486 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7489 * As sched_init() is called before page_alloc is setup,
7490 * we use alloc_bootmem().
7493 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7495 #ifdef CONFIG_FAIR_GROUP_SCHED
7496 init_task_group.se = (struct sched_entity **)ptr;
7497 ptr += nr_cpu_ids * sizeof(void **);
7499 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7500 ptr += nr_cpu_ids * sizeof(void **);
7502 #ifdef CONFIG_RT_GROUP_SCHED
7503 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7504 ptr += nr_cpu_ids * sizeof(void **);
7506 init_task_group.rt_rq = (struct rt_rq **)ptr;
7511 init_defrootdomain();
7514 init_rt_bandwidth(&def_rt_bandwidth,
7515 global_rt_period(), global_rt_runtime());
7517 #ifdef CONFIG_RT_GROUP_SCHED
7518 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7519 global_rt_period(), global_rt_runtime());
7522 #ifdef CONFIG_GROUP_SCHED
7523 list_add(&init_task_group.list, &task_groups);
7526 for_each_possible_cpu(i) {
7530 spin_lock_init(&rq->lock);
7531 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7534 update_last_tick_seen(rq);
7535 init_cfs_rq(&rq->cfs, rq);
7536 init_rt_rq(&rq->rt, rq);
7537 #ifdef CONFIG_FAIR_GROUP_SCHED
7538 init_task_group.shares = init_task_group_load;
7539 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7540 init_tg_cfs_entry(rq, &init_task_group,
7541 &per_cpu(init_cfs_rq, i),
7542 &per_cpu(init_sched_entity, i), i, 1);
7545 #ifdef CONFIG_RT_GROUP_SCHED
7546 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7547 init_tg_rt_entry(rq, &init_task_group,
7548 &per_cpu(init_rt_rq, i),
7549 &per_cpu(init_sched_rt_entity, i), i, 1);
7551 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7554 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7555 rq->cpu_load[j] = 0;
7559 rq->active_balance = 0;
7560 rq->next_balance = jiffies;
7563 rq->migration_thread = NULL;
7564 INIT_LIST_HEAD(&rq->migration_queue);
7565 rq_attach_root(rq, &def_root_domain);
7568 atomic_set(&rq->nr_iowait, 0);
7571 set_load_weight(&init_task);
7573 #ifdef CONFIG_PREEMPT_NOTIFIERS
7574 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7578 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7581 #ifdef CONFIG_RT_MUTEXES
7582 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7586 * The boot idle thread does lazy MMU switching as well:
7588 atomic_inc(&init_mm.mm_count);
7589 enter_lazy_tlb(&init_mm, current);
7592 * Make us the idle thread. Technically, schedule() should not be
7593 * called from this thread, however somewhere below it might be,
7594 * but because we are the idle thread, we just pick up running again
7595 * when this runqueue becomes "idle".
7597 init_idle(current, smp_processor_id());
7599 * During early bootup we pretend to be a normal task:
7601 current->sched_class = &fair_sched_class;
7603 scheduler_running = 1;
7606 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7607 void __might_sleep(char *file, int line)
7610 static unsigned long prev_jiffy; /* ratelimiting */
7612 if ((in_atomic() || irqs_disabled()) &&
7613 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7614 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7616 prev_jiffy = jiffies;
7617 printk(KERN_ERR "BUG: sleeping function called from invalid"
7618 " context at %s:%d\n", file, line);
7619 printk("in_atomic():%d, irqs_disabled():%d\n",
7620 in_atomic(), irqs_disabled());
7621 debug_show_held_locks(current);
7622 if (irqs_disabled())
7623 print_irqtrace_events(current);
7628 EXPORT_SYMBOL(__might_sleep);
7631 #ifdef CONFIG_MAGIC_SYSRQ
7632 static void normalize_task(struct rq *rq, struct task_struct *p)
7635 update_rq_clock(rq);
7636 on_rq = p->se.on_rq;
7638 deactivate_task(rq, p, 0);
7639 __setscheduler(rq, p, SCHED_NORMAL, 0);
7641 activate_task(rq, p, 0);
7642 resched_task(rq->curr);
7646 void normalize_rt_tasks(void)
7648 struct task_struct *g, *p;
7649 unsigned long flags;
7652 read_lock_irqsave(&tasklist_lock, flags);
7653 do_each_thread(g, p) {
7655 * Only normalize user tasks:
7660 p->se.exec_start = 0;
7661 #ifdef CONFIG_SCHEDSTATS
7662 p->se.wait_start = 0;
7663 p->se.sleep_start = 0;
7664 p->se.block_start = 0;
7666 task_rq(p)->clock = 0;
7670 * Renice negative nice level userspace
7673 if (TASK_NICE(p) < 0 && p->mm)
7674 set_user_nice(p, 0);
7678 spin_lock(&p->pi_lock);
7679 rq = __task_rq_lock(p);
7681 normalize_task(rq, p);
7683 __task_rq_unlock(rq);
7684 spin_unlock(&p->pi_lock);
7685 } while_each_thread(g, p);
7687 read_unlock_irqrestore(&tasklist_lock, flags);
7690 #endif /* CONFIG_MAGIC_SYSRQ */
7694 * These functions are only useful for the IA64 MCA handling.
7696 * They can only be called when the whole system has been
7697 * stopped - every CPU needs to be quiescent, and no scheduling
7698 * activity can take place. Using them for anything else would
7699 * be a serious bug, and as a result, they aren't even visible
7700 * under any other configuration.
7704 * curr_task - return the current task for a given cpu.
7705 * @cpu: the processor in question.
7707 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7709 struct task_struct *curr_task(int cpu)
7711 return cpu_curr(cpu);
7715 * set_curr_task - set the current task for a given cpu.
7716 * @cpu: the processor in question.
7717 * @p: the task pointer to set.
7719 * Description: This function must only be used when non-maskable interrupts
7720 * are serviced on a separate stack. It allows the architecture to switch the
7721 * notion of the current task on a cpu in a non-blocking manner. This function
7722 * must be called with all CPU's synchronized, and interrupts disabled, the
7723 * and caller must save the original value of the current task (see
7724 * curr_task() above) and restore that value before reenabling interrupts and
7725 * re-starting the system.
7727 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7729 void set_curr_task(int cpu, struct task_struct *p)
7736 #ifdef CONFIG_FAIR_GROUP_SCHED
7737 static void free_fair_sched_group(struct task_group *tg)
7741 for_each_possible_cpu(i) {
7743 kfree(tg->cfs_rq[i]);
7752 static int alloc_fair_sched_group(struct task_group *tg)
7754 struct cfs_rq *cfs_rq;
7755 struct sched_entity *se;
7759 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7762 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7766 tg->shares = NICE_0_LOAD;
7768 for_each_possible_cpu(i) {
7771 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7772 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7776 se = kmalloc_node(sizeof(struct sched_entity),
7777 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7781 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7790 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7792 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7793 &cpu_rq(cpu)->leaf_cfs_rq_list);
7796 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7798 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7801 static inline void free_fair_sched_group(struct task_group *tg)
7805 static inline int alloc_fair_sched_group(struct task_group *tg)
7810 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7814 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7819 #ifdef CONFIG_RT_GROUP_SCHED
7820 static void free_rt_sched_group(struct task_group *tg)
7824 destroy_rt_bandwidth(&tg->rt_bandwidth);
7826 for_each_possible_cpu(i) {
7828 kfree(tg->rt_rq[i]);
7830 kfree(tg->rt_se[i]);
7837 static int alloc_rt_sched_group(struct task_group *tg)
7839 struct rt_rq *rt_rq;
7840 struct sched_rt_entity *rt_se;
7844 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7847 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7851 init_rt_bandwidth(&tg->rt_bandwidth,
7852 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7854 for_each_possible_cpu(i) {
7857 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7858 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7862 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7863 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7867 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7876 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7878 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7879 &cpu_rq(cpu)->leaf_rt_rq_list);
7882 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7884 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7887 static inline void free_rt_sched_group(struct task_group *tg)
7891 static inline int alloc_rt_sched_group(struct task_group *tg)
7896 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7900 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7905 #ifdef CONFIG_GROUP_SCHED
7906 static void free_sched_group(struct task_group *tg)
7908 free_fair_sched_group(tg);
7909 free_rt_sched_group(tg);
7913 /* allocate runqueue etc for a new task group */
7914 struct task_group *sched_create_group(void)
7916 struct task_group *tg;
7917 unsigned long flags;
7920 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7922 return ERR_PTR(-ENOMEM);
7924 if (!alloc_fair_sched_group(tg))
7927 if (!alloc_rt_sched_group(tg))
7930 spin_lock_irqsave(&task_group_lock, flags);
7931 for_each_possible_cpu(i) {
7932 register_fair_sched_group(tg, i);
7933 register_rt_sched_group(tg, i);
7935 list_add_rcu(&tg->list, &task_groups);
7936 spin_unlock_irqrestore(&task_group_lock, flags);
7941 free_sched_group(tg);
7942 return ERR_PTR(-ENOMEM);
7945 /* rcu callback to free various structures associated with a task group */
7946 static void free_sched_group_rcu(struct rcu_head *rhp)
7948 /* now it should be safe to free those cfs_rqs */
7949 free_sched_group(container_of(rhp, struct task_group, rcu));
7952 /* Destroy runqueue etc associated with a task group */
7953 void sched_destroy_group(struct task_group *tg)
7955 unsigned long flags;
7958 spin_lock_irqsave(&task_group_lock, flags);
7959 for_each_possible_cpu(i) {
7960 unregister_fair_sched_group(tg, i);
7961 unregister_rt_sched_group(tg, i);
7963 list_del_rcu(&tg->list);
7964 spin_unlock_irqrestore(&task_group_lock, flags);
7966 /* wait for possible concurrent references to cfs_rqs complete */
7967 call_rcu(&tg->rcu, free_sched_group_rcu);
7970 /* change task's runqueue when it moves between groups.
7971 * The caller of this function should have put the task in its new group
7972 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7973 * reflect its new group.
7975 void sched_move_task(struct task_struct *tsk)
7978 unsigned long flags;
7981 rq = task_rq_lock(tsk, &flags);
7983 update_rq_clock(rq);
7985 running = task_current(rq, tsk);
7986 on_rq = tsk->se.on_rq;
7989 dequeue_task(rq, tsk, 0);
7990 if (unlikely(running))
7991 tsk->sched_class->put_prev_task(rq, tsk);
7993 set_task_rq(tsk, task_cpu(tsk));
7995 #ifdef CONFIG_FAIR_GROUP_SCHED
7996 if (tsk->sched_class->moved_group)
7997 tsk->sched_class->moved_group(tsk);
8000 if (unlikely(running))
8001 tsk->sched_class->set_curr_task(rq);
8003 enqueue_task(rq, tsk, 0);
8005 task_rq_unlock(rq, &flags);
8009 #ifdef CONFIG_FAIR_GROUP_SCHED
8010 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8012 struct cfs_rq *cfs_rq = se->cfs_rq;
8013 struct rq *rq = cfs_rq->rq;
8016 spin_lock_irq(&rq->lock);
8020 dequeue_entity(cfs_rq, se, 0);
8022 se->load.weight = shares;
8023 se->load.inv_weight = div64_64((1ULL<<32), shares);
8026 enqueue_entity(cfs_rq, se, 0);
8028 spin_unlock_irq(&rq->lock);
8031 static DEFINE_MUTEX(shares_mutex);
8033 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8036 unsigned long flags;
8039 * A weight of 0 or 1 can cause arithmetics problems.
8040 * (The default weight is 1024 - so there's no practical
8041 * limitation from this.)
8046 mutex_lock(&shares_mutex);
8047 if (tg->shares == shares)
8050 spin_lock_irqsave(&task_group_lock, flags);
8051 for_each_possible_cpu(i)
8052 unregister_fair_sched_group(tg, i);
8053 spin_unlock_irqrestore(&task_group_lock, flags);
8055 /* wait for any ongoing reference to this group to finish */
8056 synchronize_sched();
8059 * Now we are free to modify the group's share on each cpu
8060 * w/o tripping rebalance_share or load_balance_fair.
8062 tg->shares = shares;
8063 for_each_possible_cpu(i)
8064 set_se_shares(tg->se[i], shares);
8067 * Enable load balance activity on this group, by inserting it back on
8068 * each cpu's rq->leaf_cfs_rq_list.
8070 spin_lock_irqsave(&task_group_lock, flags);
8071 for_each_possible_cpu(i)
8072 register_fair_sched_group(tg, i);
8073 spin_unlock_irqrestore(&task_group_lock, flags);
8075 mutex_unlock(&shares_mutex);
8079 unsigned long sched_group_shares(struct task_group *tg)
8085 #ifdef CONFIG_RT_GROUP_SCHED
8087 * Ensure that the real time constraints are schedulable.
8089 static DEFINE_MUTEX(rt_constraints_mutex);
8091 static unsigned long to_ratio(u64 period, u64 runtime)
8093 if (runtime == RUNTIME_INF)
8096 return div64_64(runtime << 16, period);
8099 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8101 struct task_group *tgi;
8102 unsigned long total = 0;
8103 unsigned long global_ratio =
8104 to_ratio(global_rt_period(), global_rt_runtime());
8107 list_for_each_entry_rcu(tgi, &task_groups, list) {
8111 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8112 tgi->rt_bandwidth.rt_runtime);
8116 return total + to_ratio(period, runtime) < global_ratio;
8119 /* Must be called with tasklist_lock held */
8120 static inline int tg_has_rt_tasks(struct task_group *tg)
8122 struct task_struct *g, *p;
8123 do_each_thread(g, p) {
8124 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8126 } while_each_thread(g, p);
8130 static int tg_set_bandwidth(struct task_group *tg,
8131 u64 rt_period, u64 rt_runtime)
8135 mutex_lock(&rt_constraints_mutex);
8136 read_lock(&tasklist_lock);
8137 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8141 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8146 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8147 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8148 tg->rt_bandwidth.rt_runtime = rt_runtime;
8150 for_each_possible_cpu(i) {
8151 struct rt_rq *rt_rq = tg->rt_rq[i];
8153 spin_lock(&rt_rq->rt_runtime_lock);
8154 rt_rq->rt_runtime = rt_runtime;
8155 spin_unlock(&rt_rq->rt_runtime_lock);
8157 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8159 read_unlock(&tasklist_lock);
8160 mutex_unlock(&rt_constraints_mutex);
8165 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8167 u64 rt_runtime, rt_period;
8169 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8170 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8171 if (rt_runtime_us < 0)
8172 rt_runtime = RUNTIME_INF;
8174 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8177 long sched_group_rt_runtime(struct task_group *tg)
8181 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8184 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8185 do_div(rt_runtime_us, NSEC_PER_USEC);
8186 return rt_runtime_us;
8189 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8191 u64 rt_runtime, rt_period;
8193 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8194 rt_runtime = tg->rt_bandwidth.rt_runtime;
8196 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8199 long sched_group_rt_period(struct task_group *tg)
8203 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8204 do_div(rt_period_us, NSEC_PER_USEC);
8205 return rt_period_us;
8208 static int sched_rt_global_constraints(void)
8212 mutex_lock(&rt_constraints_mutex);
8213 if (!__rt_schedulable(NULL, 1, 0))
8215 mutex_unlock(&rt_constraints_mutex);
8220 static int sched_rt_global_constraints(void)
8222 unsigned long flags;
8225 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8226 for_each_possible_cpu(i) {
8227 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8229 spin_lock(&rt_rq->rt_runtime_lock);
8230 rt_rq->rt_runtime = global_rt_runtime();
8231 spin_unlock(&rt_rq->rt_runtime_lock);
8233 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8239 int sched_rt_handler(struct ctl_table *table, int write,
8240 struct file *filp, void __user *buffer, size_t *lenp,
8244 int old_period, old_runtime;
8245 static DEFINE_MUTEX(mutex);
8248 old_period = sysctl_sched_rt_period;
8249 old_runtime = sysctl_sched_rt_runtime;
8251 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8253 if (!ret && write) {
8254 ret = sched_rt_global_constraints();
8256 sysctl_sched_rt_period = old_period;
8257 sysctl_sched_rt_runtime = old_runtime;
8259 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8260 def_rt_bandwidth.rt_period =
8261 ns_to_ktime(global_rt_period());
8264 mutex_unlock(&mutex);
8269 #ifdef CONFIG_CGROUP_SCHED
8271 /* return corresponding task_group object of a cgroup */
8272 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8274 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8275 struct task_group, css);
8278 static struct cgroup_subsys_state *
8279 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8281 struct task_group *tg;
8283 if (!cgrp->parent) {
8284 /* This is early initialization for the top cgroup */
8285 init_task_group.css.cgroup = cgrp;
8286 return &init_task_group.css;
8289 /* we support only 1-level deep hierarchical scheduler atm */
8290 if (cgrp->parent->parent)
8291 return ERR_PTR(-EINVAL);
8293 tg = sched_create_group();
8295 return ERR_PTR(-ENOMEM);
8297 /* Bind the cgroup to task_group object we just created */
8298 tg->css.cgroup = cgrp;
8304 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8306 struct task_group *tg = cgroup_tg(cgrp);
8308 sched_destroy_group(tg);
8312 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8313 struct task_struct *tsk)
8315 #ifdef CONFIG_RT_GROUP_SCHED
8316 /* Don't accept realtime tasks when there is no way for them to run */
8317 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8320 /* We don't support RT-tasks being in separate groups */
8321 if (tsk->sched_class != &fair_sched_class)
8329 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8330 struct cgroup *old_cont, struct task_struct *tsk)
8332 sched_move_task(tsk);
8335 #ifdef CONFIG_FAIR_GROUP_SCHED
8336 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8339 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8342 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8344 struct task_group *tg = cgroup_tg(cgrp);
8346 return (u64) tg->shares;
8350 #ifdef CONFIG_RT_GROUP_SCHED
8351 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8353 const char __user *userbuf,
8354 size_t nbytes, loff_t *unused_ppos)
8363 if (nbytes >= sizeof(buffer))
8365 if (copy_from_user(buffer, userbuf, nbytes))
8368 buffer[nbytes] = 0; /* nul-terminate */
8370 /* strip newline if necessary */
8371 if (nbytes && (buffer[nbytes-1] == '\n'))
8372 buffer[nbytes-1] = 0;
8373 val = simple_strtoll(buffer, &end, 0);
8377 /* Pass to subsystem */
8378 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8384 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8386 char __user *buf, size_t nbytes,
8390 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8391 int len = sprintf(tmp, "%ld\n", val);
8393 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8396 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8399 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8402 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8404 return sched_group_rt_period(cgroup_tg(cgrp));
8408 static struct cftype cpu_files[] = {
8409 #ifdef CONFIG_FAIR_GROUP_SCHED
8412 .read_uint = cpu_shares_read_uint,
8413 .write_uint = cpu_shares_write_uint,
8416 #ifdef CONFIG_RT_GROUP_SCHED
8418 .name = "rt_runtime_us",
8419 .read = cpu_rt_runtime_read,
8420 .write = cpu_rt_runtime_write,
8423 .name = "rt_period_us",
8424 .read_uint = cpu_rt_period_read_uint,
8425 .write_uint = cpu_rt_period_write_uint,
8430 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8432 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8435 struct cgroup_subsys cpu_cgroup_subsys = {
8437 .create = cpu_cgroup_create,
8438 .destroy = cpu_cgroup_destroy,
8439 .can_attach = cpu_cgroup_can_attach,
8440 .attach = cpu_cgroup_attach,
8441 .populate = cpu_cgroup_populate,
8442 .subsys_id = cpu_cgroup_subsys_id,
8446 #endif /* CONFIG_CGROUP_SCHED */
8448 #ifdef CONFIG_CGROUP_CPUACCT
8451 * CPU accounting code for task groups.
8453 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8454 * (balbir@in.ibm.com).
8457 /* track cpu usage of a group of tasks */
8459 struct cgroup_subsys_state css;
8460 /* cpuusage holds pointer to a u64-type object on every cpu */
8464 struct cgroup_subsys cpuacct_subsys;
8466 /* return cpu accounting group corresponding to this container */
8467 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8469 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8470 struct cpuacct, css);
8473 /* return cpu accounting group to which this task belongs */
8474 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8476 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8477 struct cpuacct, css);
8480 /* create a new cpu accounting group */
8481 static struct cgroup_subsys_state *cpuacct_create(
8482 struct cgroup_subsys *ss, struct cgroup *cgrp)
8484 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8487 return ERR_PTR(-ENOMEM);
8489 ca->cpuusage = alloc_percpu(u64);
8490 if (!ca->cpuusage) {
8492 return ERR_PTR(-ENOMEM);
8498 /* destroy an existing cpu accounting group */
8500 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8502 struct cpuacct *ca = cgroup_ca(cgrp);
8504 free_percpu(ca->cpuusage);
8508 /* return total cpu usage (in nanoseconds) of a group */
8509 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8511 struct cpuacct *ca = cgroup_ca(cgrp);
8512 u64 totalcpuusage = 0;
8515 for_each_possible_cpu(i) {
8516 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8519 * Take rq->lock to make 64-bit addition safe on 32-bit
8522 spin_lock_irq(&cpu_rq(i)->lock);
8523 totalcpuusage += *cpuusage;
8524 spin_unlock_irq(&cpu_rq(i)->lock);
8527 return totalcpuusage;
8530 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8533 struct cpuacct *ca = cgroup_ca(cgrp);
8542 for_each_possible_cpu(i) {
8543 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8545 spin_lock_irq(&cpu_rq(i)->lock);
8547 spin_unlock_irq(&cpu_rq(i)->lock);
8553 static struct cftype files[] = {
8556 .read_uint = cpuusage_read,
8557 .write_uint = cpuusage_write,
8561 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8563 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8567 * charge this task's execution time to its accounting group.
8569 * called with rq->lock held.
8571 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8575 if (!cpuacct_subsys.active)
8580 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8582 *cpuusage += cputime;
8586 struct cgroup_subsys cpuacct_subsys = {
8588 .create = cpuacct_create,
8589 .destroy = cpuacct_destroy,
8590 .populate = cpuacct_populate,
8591 .subsys_id = cpuacct_subsys_id,
8593 #endif /* CONFIG_CGROUP_CPUACCT */