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>
72 #include <asm/irq_regs.h>
75 * Scheduler clock - returns current time in nanosec units.
76 * This is default implementation.
77 * Architectures and sub-architectures can override this.
79 unsigned long long __attribute__((weak)) sched_clock(void)
81 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 #ifdef CONFIG_GROUP_SCHED
161 #include <linux/cgroup.h>
165 static LIST_HEAD(task_groups);
167 /* task group related information */
169 #ifdef CONFIG_CGROUP_SCHED
170 struct cgroup_subsys_state css;
173 #ifdef CONFIG_FAIR_GROUP_SCHED
174 /* schedulable entities of this group on each cpu */
175 struct sched_entity **se;
176 /* runqueue "owned" by this group on each cpu */
177 struct cfs_rq **cfs_rq;
178 unsigned long shares;
181 #ifdef CONFIG_RT_GROUP_SCHED
182 struct sched_rt_entity **rt_se;
183 struct rt_rq **rt_rq;
189 struct list_head list;
192 #ifdef CONFIG_FAIR_GROUP_SCHED
193 /* Default task group's sched entity on each cpu */
194 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
195 /* Default task group's cfs_rq on each cpu */
196 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
198 static struct sched_entity *init_sched_entity_p[NR_CPUS];
199 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
202 #ifdef CONFIG_RT_GROUP_SCHED
203 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
204 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
206 static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
207 static struct rt_rq *init_rt_rq_p[NR_CPUS];
210 /* task_group_lock serializes add/remove of task groups and also changes to
211 * a task group's cpu shares.
213 static DEFINE_SPINLOCK(task_group_lock);
215 /* doms_cur_mutex serializes access to doms_cur[] array */
216 static DEFINE_MUTEX(doms_cur_mutex);
218 #ifdef CONFIG_FAIR_GROUP_SCHED
219 #ifdef CONFIG_USER_SCHED
220 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
222 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
225 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
228 /* Default task group.
229 * Every task in system belong to this group at bootup.
231 struct task_group init_task_group = {
232 #ifdef CONFIG_FAIR_GROUP_SCHED
233 .se = init_sched_entity_p,
234 .cfs_rq = init_cfs_rq_p,
237 #ifdef CONFIG_RT_GROUP_SCHED
238 .rt_se = init_sched_rt_entity_p,
239 .rt_rq = init_rt_rq_p,
243 /* return group to which a task belongs */
244 static inline struct task_group *task_group(struct task_struct *p)
246 struct task_group *tg;
248 #ifdef CONFIG_USER_SCHED
250 #elif defined(CONFIG_CGROUP_SCHED)
251 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
252 struct task_group, css);
254 tg = &init_task_group;
259 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
260 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
262 #ifdef CONFIG_FAIR_GROUP_SCHED
263 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
264 p->se.parent = task_group(p)->se[cpu];
267 #ifdef CONFIG_RT_GROUP_SCHED
268 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
269 p->rt.parent = task_group(p)->rt_se[cpu];
273 static inline void lock_doms_cur(void)
275 mutex_lock(&doms_cur_mutex);
278 static inline void unlock_doms_cur(void)
280 mutex_unlock(&doms_cur_mutex);
285 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
286 static inline void lock_doms_cur(void) { }
287 static inline void unlock_doms_cur(void) { }
289 #endif /* CONFIG_GROUP_SCHED */
291 /* CFS-related fields in a runqueue */
293 struct load_weight load;
294 unsigned long nr_running;
299 struct rb_root tasks_timeline;
300 struct rb_node *rb_leftmost;
301 struct rb_node *rb_load_balance_curr;
302 /* 'curr' points to currently running entity on this cfs_rq.
303 * It is set to NULL otherwise (i.e when none are currently running).
305 struct sched_entity *curr, *next;
307 unsigned long nr_spread_over;
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
313 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
314 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
315 * (like users, containers etc.)
317 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
318 * list is used during load balance.
320 struct list_head leaf_cfs_rq_list;
321 struct task_group *tg; /* group that "owns" this runqueue */
325 /* Real-Time classes' related field in a runqueue: */
327 struct rt_prio_array active;
328 unsigned long rt_nr_running;
329 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
330 int highest_prio; /* highest queued rt task prio */
333 unsigned long rt_nr_migratory;
339 #ifdef CONFIG_RT_GROUP_SCHED
340 unsigned long rt_nr_boosted;
343 struct list_head leaf_rt_rq_list;
344 struct task_group *tg;
345 struct sched_rt_entity *rt_se;
352 * We add the notion of a root-domain which will be used to define per-domain
353 * variables. Each exclusive cpuset essentially defines an island domain by
354 * fully partitioning the member cpus from any other cpuset. Whenever a new
355 * exclusive cpuset is created, we also create and attach a new root-domain
365 * The "RT overload" flag: it gets set if a CPU has more than
366 * one runnable RT task.
373 * By default the system creates a single root-domain with all cpus as
374 * members (mimicking the global state we have today).
376 static struct root_domain def_root_domain;
381 * This is the main, per-CPU runqueue data structure.
383 * Locking rule: those places that want to lock multiple runqueues
384 * (such as the load balancing or the thread migration code), lock
385 * acquire operations must be ordered by ascending &runqueue.
392 * nr_running and cpu_load should be in the same cacheline because
393 * remote CPUs use both these fields when doing load calculation.
395 unsigned long nr_running;
396 #define CPU_LOAD_IDX_MAX 5
397 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
398 unsigned char idle_at_tick;
400 unsigned long last_tick_seen;
401 unsigned char in_nohz_recently;
403 /* capture load from *all* tasks on this cpu: */
404 struct load_weight load;
405 unsigned long nr_load_updates;
410 u64 rt_period_expire;
413 #ifdef CONFIG_FAIR_GROUP_SCHED
414 /* list of leaf cfs_rq on this cpu: */
415 struct list_head leaf_cfs_rq_list;
417 #ifdef CONFIG_RT_GROUP_SCHED
418 struct list_head leaf_rt_rq_list;
422 * This is part of a global counter where only the total sum
423 * over all CPUs matters. A task can increase this counter on
424 * one CPU and if it got migrated afterwards it may decrease
425 * it on another CPU. Always updated under the runqueue lock:
427 unsigned long nr_uninterruptible;
429 struct task_struct *curr, *idle;
430 unsigned long next_balance;
431 struct mm_struct *prev_mm;
433 u64 clock, prev_clock_raw;
436 unsigned int clock_warps, clock_overflows, clock_underflows;
438 unsigned int clock_deep_idle_events;
444 struct root_domain *rd;
445 struct sched_domain *sd;
447 /* For active balancing */
450 /* cpu of this runqueue: */
453 struct task_struct *migration_thread;
454 struct list_head migration_queue;
457 #ifdef CONFIG_SCHED_HRTICK
458 unsigned long hrtick_flags;
459 ktime_t hrtick_expire;
460 struct hrtimer hrtick_timer;
463 #ifdef CONFIG_SCHEDSTATS
465 struct sched_info rq_sched_info;
467 /* sys_sched_yield() stats */
468 unsigned int yld_exp_empty;
469 unsigned int yld_act_empty;
470 unsigned int yld_both_empty;
471 unsigned int yld_count;
473 /* schedule() stats */
474 unsigned int sched_switch;
475 unsigned int sched_count;
476 unsigned int sched_goidle;
478 /* try_to_wake_up() stats */
479 unsigned int ttwu_count;
480 unsigned int ttwu_local;
483 unsigned int bkl_count;
485 struct lock_class_key rq_lock_key;
488 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
490 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
492 rq->curr->sched_class->check_preempt_curr(rq, p);
495 static inline int cpu_of(struct rq *rq)
505 static inline bool nohz_on(int cpu)
507 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
510 static inline u64 max_skipped_ticks(struct rq *rq)
512 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
515 static inline void update_last_tick_seen(struct rq *rq)
517 rq->last_tick_seen = jiffies;
520 static inline u64 max_skipped_ticks(struct rq *rq)
525 static inline void update_last_tick_seen(struct rq *rq)
531 * Update the per-runqueue clock, as finegrained as the platform can give
532 * us, but without assuming monotonicity, etc.:
534 static void __update_rq_clock(struct rq *rq)
536 u64 prev_raw = rq->prev_clock_raw;
537 u64 now = sched_clock();
538 s64 delta = now - prev_raw;
539 u64 clock = rq->clock;
541 #ifdef CONFIG_SCHED_DEBUG
542 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
545 * Protect against sched_clock() occasionally going backwards:
547 if (unlikely(delta < 0)) {
552 * Catch too large forward jumps too:
554 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
555 u64 max_time = rq->tick_timestamp + max_jump;
557 if (unlikely(clock + delta > max_time)) {
558 if (clock < max_time)
562 rq->clock_overflows++;
564 if (unlikely(delta > rq->clock_max_delta))
565 rq->clock_max_delta = delta;
570 rq->prev_clock_raw = now;
574 static void update_rq_clock(struct rq *rq)
576 if (likely(smp_processor_id() == cpu_of(rq)))
577 __update_rq_clock(rq);
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
595 unsigned long rt_needs_cpu(int cpu)
597 struct rq *rq = cpu_rq(cpu);
600 if (!rq->rt_throttled)
603 if (rq->clock > rq->rt_period_expire)
606 delta = rq->rt_period_expire - rq->clock;
607 do_div(delta, NSEC_PER_SEC / HZ);
609 return (unsigned long)delta;
613 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
615 #ifdef CONFIG_SCHED_DEBUG
616 # define const_debug __read_mostly
618 # define const_debug static const
622 * Debugging: various feature bits
625 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
626 SCHED_FEAT_WAKEUP_PREEMPT = 2,
627 SCHED_FEAT_START_DEBIT = 4,
628 SCHED_FEAT_HRTICK = 8,
629 SCHED_FEAT_DOUBLE_TICK = 16,
630 SCHED_FEAT_SYNC_WAKEUPS = 32,
633 const_debug unsigned int sysctl_sched_features =
634 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
635 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
636 SCHED_FEAT_START_DEBIT * 1 |
637 SCHED_FEAT_HRTICK * 1 |
638 SCHED_FEAT_DOUBLE_TICK * 0 |
639 SCHED_FEAT_SYNC_WAKEUPS * 0;
641 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
644 * Number of tasks to iterate in a single balance run.
645 * Limited because this is done with IRQs disabled.
647 const_debug unsigned int sysctl_sched_nr_migrate = 32;
650 * period over which we measure -rt task cpu usage in us.
653 unsigned int sysctl_sched_rt_period = 1000000;
655 static __read_mostly int scheduler_running;
658 * part of the period that we allow rt tasks to run in us.
661 int sysctl_sched_rt_runtime = 950000;
664 * single value that denotes runtime == period, ie unlimited time.
666 #define RUNTIME_INF ((u64)~0ULL)
668 static const unsigned long long time_sync_thresh = 100000;
670 static DEFINE_PER_CPU(unsigned long long, time_offset);
671 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
674 * Global lock which we take every now and then to synchronize
675 * the CPUs time. This method is not warp-safe, but it's good
676 * enough to synchronize slowly diverging time sources and thus
677 * it's good enough for tracing:
679 static DEFINE_SPINLOCK(time_sync_lock);
680 static unsigned long long prev_global_time;
682 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
686 spin_lock_irqsave(&time_sync_lock, flags);
688 if (time < prev_global_time) {
689 per_cpu(time_offset, cpu) += prev_global_time - time;
690 time = prev_global_time;
692 prev_global_time = time;
695 spin_unlock_irqrestore(&time_sync_lock, flags);
700 static unsigned long long __cpu_clock(int cpu)
702 unsigned long long now;
707 * Only call sched_clock() if the scheduler has already been
708 * initialized (some code might call cpu_clock() very early):
710 if (unlikely(!scheduler_running))
713 local_irq_save(flags);
717 local_irq_restore(flags);
723 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
724 * clock constructed from sched_clock():
726 unsigned long long cpu_clock(int cpu)
728 unsigned long long prev_cpu_time, time, delta_time;
730 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
731 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
732 delta_time = time-prev_cpu_time;
734 if (unlikely(delta_time > time_sync_thresh))
735 time = __sync_cpu_clock(time, cpu);
739 EXPORT_SYMBOL_GPL(cpu_clock);
741 #ifndef prepare_arch_switch
742 # define prepare_arch_switch(next) do { } while (0)
744 #ifndef finish_arch_switch
745 # define finish_arch_switch(prev) do { } while (0)
748 static inline int task_current(struct rq *rq, struct task_struct *p)
750 return rq->curr == p;
753 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
754 static inline int task_running(struct rq *rq, struct task_struct *p)
756 return task_current(rq, p);
759 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
763 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
765 #ifdef CONFIG_DEBUG_SPINLOCK
766 /* this is a valid case when another task releases the spinlock */
767 rq->lock.owner = current;
770 * If we are tracking spinlock dependencies then we have to
771 * fix up the runqueue lock - which gets 'carried over' from
774 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
776 spin_unlock_irq(&rq->lock);
779 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
780 static inline int task_running(struct rq *rq, struct task_struct *p)
785 return task_current(rq, p);
789 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
793 * We can optimise this out completely for !SMP, because the
794 * SMP rebalancing from interrupt is the only thing that cares
799 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
800 spin_unlock_irq(&rq->lock);
802 spin_unlock(&rq->lock);
806 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
810 * After ->oncpu is cleared, the task can be moved to a different CPU.
811 * We must ensure this doesn't happen until the switch is completely
817 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
821 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
824 * __task_rq_lock - lock the runqueue a given task resides on.
825 * Must be called interrupts disabled.
827 static inline struct rq *__task_rq_lock(struct task_struct *p)
831 struct rq *rq = task_rq(p);
832 spin_lock(&rq->lock);
833 if (likely(rq == task_rq(p)))
835 spin_unlock(&rq->lock);
840 * task_rq_lock - lock the runqueue a given task resides on and disable
841 * interrupts. Note the ordering: we can safely lookup the task_rq without
842 * explicitly disabling preemption.
844 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
850 local_irq_save(*flags);
852 spin_lock(&rq->lock);
853 if (likely(rq == task_rq(p)))
855 spin_unlock_irqrestore(&rq->lock, *flags);
859 static void __task_rq_unlock(struct rq *rq)
862 spin_unlock(&rq->lock);
865 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
868 spin_unlock_irqrestore(&rq->lock, *flags);
872 * this_rq_lock - lock this runqueue and disable interrupts.
874 static struct rq *this_rq_lock(void)
881 spin_lock(&rq->lock);
887 * We are going deep-idle (irqs are disabled):
889 void sched_clock_idle_sleep_event(void)
891 struct rq *rq = cpu_rq(smp_processor_id());
893 spin_lock(&rq->lock);
894 __update_rq_clock(rq);
895 spin_unlock(&rq->lock);
896 rq->clock_deep_idle_events++;
898 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
901 * We just idled delta nanoseconds (called with irqs disabled):
903 void sched_clock_idle_wakeup_event(u64 delta_ns)
905 struct rq *rq = cpu_rq(smp_processor_id());
906 u64 now = sched_clock();
908 rq->idle_clock += delta_ns;
910 * Override the previous timestamp and ignore all
911 * sched_clock() deltas that occured while we idled,
912 * and use the PM-provided delta_ns to advance the
915 spin_lock(&rq->lock);
916 rq->prev_clock_raw = now;
917 rq->clock += delta_ns;
918 spin_unlock(&rq->lock);
919 touch_softlockup_watchdog();
921 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
923 static void __resched_task(struct task_struct *p, int tif_bit);
925 static inline void resched_task(struct task_struct *p)
927 __resched_task(p, TIF_NEED_RESCHED);
930 #ifdef CONFIG_SCHED_HRTICK
932 * Use HR-timers to deliver accurate preemption points.
934 * Its all a bit involved since we cannot program an hrt while holding the
935 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
938 * When we get rescheduled we reprogram the hrtick_timer outside of the
941 static inline void resched_hrt(struct task_struct *p)
943 __resched_task(p, TIF_HRTICK_RESCHED);
946 static inline void resched_rq(struct rq *rq)
950 spin_lock_irqsave(&rq->lock, flags);
951 resched_task(rq->curr);
952 spin_unlock_irqrestore(&rq->lock, flags);
956 HRTICK_SET, /* re-programm hrtick_timer */
957 HRTICK_RESET, /* not a new slice */
962 * - enabled by features
963 * - hrtimer is actually high res
965 static inline int hrtick_enabled(struct rq *rq)
967 if (!sched_feat(HRTICK))
969 return hrtimer_is_hres_active(&rq->hrtick_timer);
973 * Called to set the hrtick timer state.
975 * called with rq->lock held and irqs disabled
977 static void hrtick_start(struct rq *rq, u64 delay, int reset)
979 assert_spin_locked(&rq->lock);
982 * preempt at: now + delay
985 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
987 * indicate we need to program the timer
989 __set_bit(HRTICK_SET, &rq->hrtick_flags);
991 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
994 * New slices are called from the schedule path and don't need a
998 resched_hrt(rq->curr);
1001 static void hrtick_clear(struct rq *rq)
1003 if (hrtimer_active(&rq->hrtick_timer))
1004 hrtimer_cancel(&rq->hrtick_timer);
1008 * Update the timer from the possible pending state.
1010 static void hrtick_set(struct rq *rq)
1014 unsigned long flags;
1016 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1018 spin_lock_irqsave(&rq->lock, flags);
1019 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1020 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1021 time = rq->hrtick_expire;
1022 clear_thread_flag(TIF_HRTICK_RESCHED);
1023 spin_unlock_irqrestore(&rq->lock, flags);
1026 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1027 if (reset && !hrtimer_active(&rq->hrtick_timer))
1034 * High-resolution timer tick.
1035 * Runs from hardirq context with interrupts disabled.
1037 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1039 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1041 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1043 spin_lock(&rq->lock);
1044 __update_rq_clock(rq);
1045 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1046 spin_unlock(&rq->lock);
1048 return HRTIMER_NORESTART;
1051 static inline void init_rq_hrtick(struct rq *rq)
1053 rq->hrtick_flags = 0;
1054 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1055 rq->hrtick_timer.function = hrtick;
1056 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1059 void hrtick_resched(void)
1062 unsigned long flags;
1064 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1067 local_irq_save(flags);
1068 rq = cpu_rq(smp_processor_id());
1070 local_irq_restore(flags);
1073 static inline void hrtick_clear(struct rq *rq)
1077 static inline void hrtick_set(struct rq *rq)
1081 static inline void init_rq_hrtick(struct rq *rq)
1085 void hrtick_resched(void)
1091 * resched_task - mark a task 'to be rescheduled now'.
1093 * On UP this means the setting of the need_resched flag, on SMP it
1094 * might also involve a cross-CPU call to trigger the scheduler on
1099 #ifndef tsk_is_polling
1100 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1103 static void __resched_task(struct task_struct *p, int tif_bit)
1107 assert_spin_locked(&task_rq(p)->lock);
1109 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1112 set_tsk_thread_flag(p, tif_bit);
1115 if (cpu == smp_processor_id())
1118 /* NEED_RESCHED must be visible before we test polling */
1120 if (!tsk_is_polling(p))
1121 smp_send_reschedule(cpu);
1124 static void resched_cpu(int cpu)
1126 struct rq *rq = cpu_rq(cpu);
1127 unsigned long flags;
1129 if (!spin_trylock_irqsave(&rq->lock, flags))
1131 resched_task(cpu_curr(cpu));
1132 spin_unlock_irqrestore(&rq->lock, flags);
1137 * When add_timer_on() enqueues a timer into the timer wheel of an
1138 * idle CPU then this timer might expire before the next timer event
1139 * which is scheduled to wake up that CPU. In case of a completely
1140 * idle system the next event might even be infinite time into the
1141 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1142 * leaves the inner idle loop so the newly added timer is taken into
1143 * account when the CPU goes back to idle and evaluates the timer
1144 * wheel for the next timer event.
1146 void wake_up_idle_cpu(int cpu)
1148 struct rq *rq = cpu_rq(cpu);
1150 if (cpu == smp_processor_id())
1154 * This is safe, as this function is called with the timer
1155 * wheel base lock of (cpu) held. When the CPU is on the way
1156 * to idle and has not yet set rq->curr to idle then it will
1157 * be serialized on the timer wheel base lock and take the new
1158 * timer into account automatically.
1160 if (rq->curr != rq->idle)
1164 * We can set TIF_RESCHED on the idle task of the other CPU
1165 * lockless. The worst case is that the other CPU runs the
1166 * idle task through an additional NOOP schedule()
1168 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1170 /* NEED_RESCHED must be visible before we test polling */
1172 if (!tsk_is_polling(rq->idle))
1173 smp_send_reschedule(cpu);
1178 static void __resched_task(struct task_struct *p, int tif_bit)
1180 assert_spin_locked(&task_rq(p)->lock);
1181 set_tsk_thread_flag(p, tif_bit);
1185 #if BITS_PER_LONG == 32
1186 # define WMULT_CONST (~0UL)
1188 # define WMULT_CONST (1UL << 32)
1191 #define WMULT_SHIFT 32
1194 * Shift right and round:
1196 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1198 static unsigned long
1199 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1200 struct load_weight *lw)
1204 if (unlikely(!lw->inv_weight))
1205 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1207 tmp = (u64)delta_exec * weight;
1209 * Check whether we'd overflow the 64-bit multiplication:
1211 if (unlikely(tmp > WMULT_CONST))
1212 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1215 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1217 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1220 static inline unsigned long
1221 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1223 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1226 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1232 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1239 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1240 * of tasks with abnormal "nice" values across CPUs the contribution that
1241 * each task makes to its run queue's load is weighted according to its
1242 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1243 * scaled version of the new time slice allocation that they receive on time
1247 #define WEIGHT_IDLEPRIO 2
1248 #define WMULT_IDLEPRIO (1 << 31)
1251 * Nice levels are multiplicative, with a gentle 10% change for every
1252 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1253 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1254 * that remained on nice 0.
1256 * The "10% effect" is relative and cumulative: from _any_ nice level,
1257 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1258 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1259 * If a task goes up by ~10% and another task goes down by ~10% then
1260 * the relative distance between them is ~25%.)
1262 static const int prio_to_weight[40] = {
1263 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1264 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1265 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1266 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1267 /* 0 */ 1024, 820, 655, 526, 423,
1268 /* 5 */ 335, 272, 215, 172, 137,
1269 /* 10 */ 110, 87, 70, 56, 45,
1270 /* 15 */ 36, 29, 23, 18, 15,
1274 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1276 * In cases where the weight does not change often, we can use the
1277 * precalculated inverse to speed up arithmetics by turning divisions
1278 * into multiplications:
1280 static const u32 prio_to_wmult[40] = {
1281 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1282 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1283 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1284 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1285 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1286 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1287 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1288 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1291 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1294 * runqueue iterator, to support SMP load-balancing between different
1295 * scheduling classes, without having to expose their internal data
1296 * structures to the load-balancing proper:
1298 struct rq_iterator {
1300 struct task_struct *(*start)(void *);
1301 struct task_struct *(*next)(void *);
1305 static unsigned long
1306 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1307 unsigned long max_load_move, struct sched_domain *sd,
1308 enum cpu_idle_type idle, int *all_pinned,
1309 int *this_best_prio, struct rq_iterator *iterator);
1312 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1313 struct sched_domain *sd, enum cpu_idle_type idle,
1314 struct rq_iterator *iterator);
1317 #ifdef CONFIG_CGROUP_CPUACCT
1318 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1320 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1324 static unsigned long source_load(int cpu, int type);
1325 static unsigned long target_load(int cpu, int type);
1326 static unsigned long cpu_avg_load_per_task(int cpu);
1327 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1328 #endif /* CONFIG_SMP */
1330 #include "sched_stats.h"
1331 #include "sched_idletask.c"
1332 #include "sched_fair.c"
1333 #include "sched_rt.c"
1334 #ifdef CONFIG_SCHED_DEBUG
1335 # include "sched_debug.c"
1338 #define sched_class_highest (&rt_sched_class)
1340 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1342 update_load_add(&rq->load, p->se.load.weight);
1345 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1347 update_load_sub(&rq->load, p->se.load.weight);
1350 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1356 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1362 static void set_load_weight(struct task_struct *p)
1364 if (task_has_rt_policy(p)) {
1365 p->se.load.weight = prio_to_weight[0] * 2;
1366 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1371 * SCHED_IDLE tasks get minimal weight:
1373 if (p->policy == SCHED_IDLE) {
1374 p->se.load.weight = WEIGHT_IDLEPRIO;
1375 p->se.load.inv_weight = WMULT_IDLEPRIO;
1379 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1380 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1383 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1385 sched_info_queued(p);
1386 p->sched_class->enqueue_task(rq, p, wakeup);
1390 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1392 p->sched_class->dequeue_task(rq, p, sleep);
1397 * __normal_prio - return the priority that is based on the static prio
1399 static inline int __normal_prio(struct task_struct *p)
1401 return p->static_prio;
1405 * Calculate the expected normal priority: i.e. priority
1406 * without taking RT-inheritance into account. Might be
1407 * boosted by interactivity modifiers. Changes upon fork,
1408 * setprio syscalls, and whenever the interactivity
1409 * estimator recalculates.
1411 static inline int normal_prio(struct task_struct *p)
1415 if (task_has_rt_policy(p))
1416 prio = MAX_RT_PRIO-1 - p->rt_priority;
1418 prio = __normal_prio(p);
1423 * Calculate the current priority, i.e. the priority
1424 * taken into account by the scheduler. This value might
1425 * be boosted by RT tasks, or might be boosted by
1426 * interactivity modifiers. Will be RT if the task got
1427 * RT-boosted. If not then it returns p->normal_prio.
1429 static int effective_prio(struct task_struct *p)
1431 p->normal_prio = normal_prio(p);
1433 * If we are RT tasks or we were boosted to RT priority,
1434 * keep the priority unchanged. Otherwise, update priority
1435 * to the normal priority:
1437 if (!rt_prio(p->prio))
1438 return p->normal_prio;
1443 * activate_task - move a task to the runqueue.
1445 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1447 if (task_contributes_to_load(p))
1448 rq->nr_uninterruptible--;
1450 enqueue_task(rq, p, wakeup);
1451 inc_nr_running(p, rq);
1455 * deactivate_task - remove a task from the runqueue.
1457 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1459 if (task_contributes_to_load(p))
1460 rq->nr_uninterruptible++;
1462 dequeue_task(rq, p, sleep);
1463 dec_nr_running(p, rq);
1467 * task_curr - is this task currently executing on a CPU?
1468 * @p: the task in question.
1470 inline int task_curr(const struct task_struct *p)
1472 return cpu_curr(task_cpu(p)) == p;
1475 /* Used instead of source_load when we know the type == 0 */
1476 unsigned long weighted_cpuload(const int cpu)
1478 return cpu_rq(cpu)->load.weight;
1481 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1483 set_task_rq(p, cpu);
1486 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1487 * successfuly executed on another CPU. We must ensure that updates of
1488 * per-task data have been completed by this moment.
1491 task_thread_info(p)->cpu = cpu;
1495 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1496 const struct sched_class *prev_class,
1497 int oldprio, int running)
1499 if (prev_class != p->sched_class) {
1500 if (prev_class->switched_from)
1501 prev_class->switched_from(rq, p, running);
1502 p->sched_class->switched_to(rq, p, running);
1504 p->sched_class->prio_changed(rq, p, oldprio, running);
1510 * Is this task likely cache-hot:
1513 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1518 * Buddy candidates are cache hot:
1520 if (&p->se == cfs_rq_of(&p->se)->next)
1523 if (p->sched_class != &fair_sched_class)
1526 if (sysctl_sched_migration_cost == -1)
1528 if (sysctl_sched_migration_cost == 0)
1531 delta = now - p->se.exec_start;
1533 return delta < (s64)sysctl_sched_migration_cost;
1537 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1539 int old_cpu = task_cpu(p);
1540 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1541 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1542 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1545 clock_offset = old_rq->clock - new_rq->clock;
1547 #ifdef CONFIG_SCHEDSTATS
1548 if (p->se.wait_start)
1549 p->se.wait_start -= clock_offset;
1550 if (p->se.sleep_start)
1551 p->se.sleep_start -= clock_offset;
1552 if (p->se.block_start)
1553 p->se.block_start -= clock_offset;
1554 if (old_cpu != new_cpu) {
1555 schedstat_inc(p, se.nr_migrations);
1556 if (task_hot(p, old_rq->clock, NULL))
1557 schedstat_inc(p, se.nr_forced2_migrations);
1560 p->se.vruntime -= old_cfsrq->min_vruntime -
1561 new_cfsrq->min_vruntime;
1563 __set_task_cpu(p, new_cpu);
1566 struct migration_req {
1567 struct list_head list;
1569 struct task_struct *task;
1572 struct completion done;
1576 * The task's runqueue lock must be held.
1577 * Returns true if you have to wait for migration thread.
1580 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1582 struct rq *rq = task_rq(p);
1585 * If the task is not on a runqueue (and not running), then
1586 * it is sufficient to simply update the task's cpu field.
1588 if (!p->se.on_rq && !task_running(rq, p)) {
1589 set_task_cpu(p, dest_cpu);
1593 init_completion(&req->done);
1595 req->dest_cpu = dest_cpu;
1596 list_add(&req->list, &rq->migration_queue);
1602 * wait_task_inactive - wait for a thread to unschedule.
1604 * The caller must ensure that the task *will* unschedule sometime soon,
1605 * else this function might spin for a *long* time. This function can't
1606 * be called with interrupts off, or it may introduce deadlock with
1607 * smp_call_function() if an IPI is sent by the same process we are
1608 * waiting to become inactive.
1610 void wait_task_inactive(struct task_struct *p)
1612 unsigned long flags;
1618 * We do the initial early heuristics without holding
1619 * any task-queue locks at all. We'll only try to get
1620 * the runqueue lock when things look like they will
1626 * If the task is actively running on another CPU
1627 * still, just relax and busy-wait without holding
1630 * NOTE! Since we don't hold any locks, it's not
1631 * even sure that "rq" stays as the right runqueue!
1632 * But we don't care, since "task_running()" will
1633 * return false if the runqueue has changed and p
1634 * is actually now running somewhere else!
1636 while (task_running(rq, p))
1640 * Ok, time to look more closely! We need the rq
1641 * lock now, to be *sure*. If we're wrong, we'll
1642 * just go back and repeat.
1644 rq = task_rq_lock(p, &flags);
1645 running = task_running(rq, p);
1646 on_rq = p->se.on_rq;
1647 task_rq_unlock(rq, &flags);
1650 * Was it really running after all now that we
1651 * checked with the proper locks actually held?
1653 * Oops. Go back and try again..
1655 if (unlikely(running)) {
1661 * It's not enough that it's not actively running,
1662 * it must be off the runqueue _entirely_, and not
1665 * So if it wa still runnable (but just not actively
1666 * running right now), it's preempted, and we should
1667 * yield - it could be a while.
1669 if (unlikely(on_rq)) {
1670 schedule_timeout_uninterruptible(1);
1675 * Ahh, all good. It wasn't running, and it wasn't
1676 * runnable, which means that it will never become
1677 * running in the future either. We're all done!
1684 * kick_process - kick a running thread to enter/exit the kernel
1685 * @p: the to-be-kicked thread
1687 * Cause a process which is running on another CPU to enter
1688 * kernel-mode, without any delay. (to get signals handled.)
1690 * NOTE: this function doesnt have to take the runqueue lock,
1691 * because all it wants to ensure is that the remote task enters
1692 * the kernel. If the IPI races and the task has been migrated
1693 * to another CPU then no harm is done and the purpose has been
1696 void kick_process(struct task_struct *p)
1702 if ((cpu != smp_processor_id()) && task_curr(p))
1703 smp_send_reschedule(cpu);
1708 * Return a low guess at the load of a migration-source cpu weighted
1709 * according to the scheduling class and "nice" value.
1711 * We want to under-estimate the load of migration sources, to
1712 * balance conservatively.
1714 static unsigned long source_load(int cpu, int type)
1716 struct rq *rq = cpu_rq(cpu);
1717 unsigned long total = weighted_cpuload(cpu);
1722 return min(rq->cpu_load[type-1], total);
1726 * Return a high guess at the load of a migration-target cpu weighted
1727 * according to the scheduling class and "nice" value.
1729 static unsigned long target_load(int cpu, int type)
1731 struct rq *rq = cpu_rq(cpu);
1732 unsigned long total = weighted_cpuload(cpu);
1737 return max(rq->cpu_load[type-1], total);
1741 * Return the average load per task on the cpu's run queue
1743 static unsigned long cpu_avg_load_per_task(int cpu)
1745 struct rq *rq = cpu_rq(cpu);
1746 unsigned long total = weighted_cpuload(cpu);
1747 unsigned long n = rq->nr_running;
1749 return n ? total / n : SCHED_LOAD_SCALE;
1753 * find_idlest_group finds and returns the least busy CPU group within the
1756 static struct sched_group *
1757 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1759 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1760 unsigned long min_load = ULONG_MAX, this_load = 0;
1761 int load_idx = sd->forkexec_idx;
1762 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1765 unsigned long load, avg_load;
1769 /* Skip over this group if it has no CPUs allowed */
1770 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1773 local_group = cpu_isset(this_cpu, group->cpumask);
1775 /* Tally up the load of all CPUs in the group */
1778 for_each_cpu_mask(i, group->cpumask) {
1779 /* Bias balancing toward cpus of our domain */
1781 load = source_load(i, load_idx);
1783 load = target_load(i, load_idx);
1788 /* Adjust by relative CPU power of the group */
1789 avg_load = sg_div_cpu_power(group,
1790 avg_load * SCHED_LOAD_SCALE);
1793 this_load = avg_load;
1795 } else if (avg_load < min_load) {
1796 min_load = avg_load;
1799 } while (group = group->next, group != sd->groups);
1801 if (!idlest || 100*this_load < imbalance*min_load)
1807 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1810 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1813 unsigned long load, min_load = ULONG_MAX;
1817 /* Traverse only the allowed CPUs */
1818 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1820 for_each_cpu_mask(i, tmp) {
1821 load = weighted_cpuload(i);
1823 if (load < min_load || (load == min_load && i == this_cpu)) {
1833 * sched_balance_self: balance the current task (running on cpu) in domains
1834 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1837 * Balance, ie. select the least loaded group.
1839 * Returns the target CPU number, or the same CPU if no balancing is needed.
1841 * preempt must be disabled.
1843 static int sched_balance_self(int cpu, int flag)
1845 struct task_struct *t = current;
1846 struct sched_domain *tmp, *sd = NULL;
1848 for_each_domain(cpu, tmp) {
1850 * If power savings logic is enabled for a domain, stop there.
1852 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1854 if (tmp->flags & flag)
1860 struct sched_group *group;
1861 int new_cpu, weight;
1863 if (!(sd->flags & flag)) {
1869 group = find_idlest_group(sd, t, cpu);
1875 new_cpu = find_idlest_cpu(group, t, cpu);
1876 if (new_cpu == -1 || new_cpu == cpu) {
1877 /* Now try balancing at a lower domain level of cpu */
1882 /* Now try balancing at a lower domain level of new_cpu */
1885 weight = cpus_weight(span);
1886 for_each_domain(cpu, tmp) {
1887 if (weight <= cpus_weight(tmp->span))
1889 if (tmp->flags & flag)
1892 /* while loop will break here if sd == NULL */
1898 #endif /* CONFIG_SMP */
1901 * try_to_wake_up - wake up a thread
1902 * @p: the to-be-woken-up thread
1903 * @state: the mask of task states that can be woken
1904 * @sync: do a synchronous wakeup?
1906 * Put it on the run-queue if it's not already there. The "current"
1907 * thread is always on the run-queue (except when the actual
1908 * re-schedule is in progress), and as such you're allowed to do
1909 * the simpler "current->state = TASK_RUNNING" to mark yourself
1910 * runnable without the overhead of this.
1912 * returns failure only if the task is already active.
1914 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1916 int cpu, orig_cpu, this_cpu, success = 0;
1917 unsigned long flags;
1921 if (!sched_feat(SYNC_WAKEUPS))
1925 rq = task_rq_lock(p, &flags);
1926 old_state = p->state;
1927 if (!(old_state & state))
1935 this_cpu = smp_processor_id();
1938 if (unlikely(task_running(rq, p)))
1941 cpu = p->sched_class->select_task_rq(p, sync);
1942 if (cpu != orig_cpu) {
1943 set_task_cpu(p, cpu);
1944 task_rq_unlock(rq, &flags);
1945 /* might preempt at this point */
1946 rq = task_rq_lock(p, &flags);
1947 old_state = p->state;
1948 if (!(old_state & state))
1953 this_cpu = smp_processor_id();
1957 #ifdef CONFIG_SCHEDSTATS
1958 schedstat_inc(rq, ttwu_count);
1959 if (cpu == this_cpu)
1960 schedstat_inc(rq, ttwu_local);
1962 struct sched_domain *sd;
1963 for_each_domain(this_cpu, sd) {
1964 if (cpu_isset(cpu, sd->span)) {
1965 schedstat_inc(sd, ttwu_wake_remote);
1973 #endif /* CONFIG_SMP */
1974 schedstat_inc(p, se.nr_wakeups);
1976 schedstat_inc(p, se.nr_wakeups_sync);
1977 if (orig_cpu != cpu)
1978 schedstat_inc(p, se.nr_wakeups_migrate);
1979 if (cpu == this_cpu)
1980 schedstat_inc(p, se.nr_wakeups_local);
1982 schedstat_inc(p, se.nr_wakeups_remote);
1983 update_rq_clock(rq);
1984 activate_task(rq, p, 1);
1988 check_preempt_curr(rq, p);
1990 p->state = TASK_RUNNING;
1992 if (p->sched_class->task_wake_up)
1993 p->sched_class->task_wake_up(rq, p);
1996 task_rq_unlock(rq, &flags);
2001 int wake_up_process(struct task_struct *p)
2003 return try_to_wake_up(p, TASK_ALL, 0);
2005 EXPORT_SYMBOL(wake_up_process);
2007 int wake_up_state(struct task_struct *p, unsigned int state)
2009 return try_to_wake_up(p, state, 0);
2013 * Perform scheduler related setup for a newly forked process p.
2014 * p is forked by current.
2016 * __sched_fork() is basic setup used by init_idle() too:
2018 static void __sched_fork(struct task_struct *p)
2020 p->se.exec_start = 0;
2021 p->se.sum_exec_runtime = 0;
2022 p->se.prev_sum_exec_runtime = 0;
2023 p->se.last_wakeup = 0;
2024 p->se.avg_overlap = 0;
2026 #ifdef CONFIG_SCHEDSTATS
2027 p->se.wait_start = 0;
2028 p->se.sum_sleep_runtime = 0;
2029 p->se.sleep_start = 0;
2030 p->se.block_start = 0;
2031 p->se.sleep_max = 0;
2032 p->se.block_max = 0;
2034 p->se.slice_max = 0;
2038 INIT_LIST_HEAD(&p->rt.run_list);
2041 #ifdef CONFIG_PREEMPT_NOTIFIERS
2042 INIT_HLIST_HEAD(&p->preempt_notifiers);
2046 * We mark the process as running here, but have not actually
2047 * inserted it onto the runqueue yet. This guarantees that
2048 * nobody will actually run it, and a signal or other external
2049 * event cannot wake it up and insert it on the runqueue either.
2051 p->state = TASK_RUNNING;
2055 * fork()/clone()-time setup:
2057 void sched_fork(struct task_struct *p, int clone_flags)
2059 int cpu = get_cpu();
2064 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2066 set_task_cpu(p, cpu);
2069 * Make sure we do not leak PI boosting priority to the child:
2071 p->prio = current->normal_prio;
2072 if (!rt_prio(p->prio))
2073 p->sched_class = &fair_sched_class;
2075 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2076 if (likely(sched_info_on()))
2077 memset(&p->sched_info, 0, sizeof(p->sched_info));
2079 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2082 #ifdef CONFIG_PREEMPT
2083 /* Want to start with kernel preemption disabled. */
2084 task_thread_info(p)->preempt_count = 1;
2090 * wake_up_new_task - wake up a newly created task for the first time.
2092 * This function will do some initial scheduler statistics housekeeping
2093 * that must be done for every newly created context, then puts the task
2094 * on the runqueue and wakes it.
2096 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2098 unsigned long flags;
2101 rq = task_rq_lock(p, &flags);
2102 BUG_ON(p->state != TASK_RUNNING);
2103 update_rq_clock(rq);
2105 p->prio = effective_prio(p);
2107 if (!p->sched_class->task_new || !current->se.on_rq) {
2108 activate_task(rq, p, 0);
2111 * Let the scheduling class do new task startup
2112 * management (if any):
2114 p->sched_class->task_new(rq, p);
2115 inc_nr_running(p, rq);
2117 check_preempt_curr(rq, p);
2119 if (p->sched_class->task_wake_up)
2120 p->sched_class->task_wake_up(rq, p);
2122 task_rq_unlock(rq, &flags);
2125 #ifdef CONFIG_PREEMPT_NOTIFIERS
2128 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2129 * @notifier: notifier struct to register
2131 void preempt_notifier_register(struct preempt_notifier *notifier)
2133 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2135 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2138 * preempt_notifier_unregister - no longer interested in preemption notifications
2139 * @notifier: notifier struct to unregister
2141 * This is safe to call from within a preemption notifier.
2143 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2145 hlist_del(¬ifier->link);
2147 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2149 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2151 struct preempt_notifier *notifier;
2152 struct hlist_node *node;
2154 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2155 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2159 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2160 struct task_struct *next)
2162 struct preempt_notifier *notifier;
2163 struct hlist_node *node;
2165 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2166 notifier->ops->sched_out(notifier, next);
2171 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2176 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2177 struct task_struct *next)
2184 * prepare_task_switch - prepare to switch tasks
2185 * @rq: the runqueue preparing to switch
2186 * @prev: the current task that is being switched out
2187 * @next: the task we are going to switch to.
2189 * This is called with the rq lock held and interrupts off. It must
2190 * be paired with a subsequent finish_task_switch after the context
2193 * prepare_task_switch sets up locking and calls architecture specific
2197 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2198 struct task_struct *next)
2200 fire_sched_out_preempt_notifiers(prev, next);
2201 prepare_lock_switch(rq, next);
2202 prepare_arch_switch(next);
2206 * finish_task_switch - clean up after a task-switch
2207 * @rq: runqueue associated with task-switch
2208 * @prev: the thread we just switched away from.
2210 * finish_task_switch must be called after the context switch, paired
2211 * with a prepare_task_switch call before the context switch.
2212 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2213 * and do any other architecture-specific cleanup actions.
2215 * Note that we may have delayed dropping an mm in context_switch(). If
2216 * so, we finish that here outside of the runqueue lock. (Doing it
2217 * with the lock held can cause deadlocks; see schedule() for
2220 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2221 __releases(rq->lock)
2223 struct mm_struct *mm = rq->prev_mm;
2229 * A task struct has one reference for the use as "current".
2230 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2231 * schedule one last time. The schedule call will never return, and
2232 * the scheduled task must drop that reference.
2233 * The test for TASK_DEAD must occur while the runqueue locks are
2234 * still held, otherwise prev could be scheduled on another cpu, die
2235 * there before we look at prev->state, and then the reference would
2237 * Manfred Spraul <manfred@colorfullife.com>
2239 prev_state = prev->state;
2240 finish_arch_switch(prev);
2241 finish_lock_switch(rq, prev);
2243 if (current->sched_class->post_schedule)
2244 current->sched_class->post_schedule(rq);
2247 fire_sched_in_preempt_notifiers(current);
2250 if (unlikely(prev_state == TASK_DEAD)) {
2252 * Remove function-return probe instances associated with this
2253 * task and put them back on the free list.
2255 kprobe_flush_task(prev);
2256 put_task_struct(prev);
2261 * schedule_tail - first thing a freshly forked thread must call.
2262 * @prev: the thread we just switched away from.
2264 asmlinkage void schedule_tail(struct task_struct *prev)
2265 __releases(rq->lock)
2267 struct rq *rq = this_rq();
2269 finish_task_switch(rq, prev);
2270 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2271 /* In this case, finish_task_switch does not reenable preemption */
2274 if (current->set_child_tid)
2275 put_user(task_pid_vnr(current), current->set_child_tid);
2279 * context_switch - switch to the new MM and the new
2280 * thread's register state.
2283 context_switch(struct rq *rq, struct task_struct *prev,
2284 struct task_struct *next)
2286 struct mm_struct *mm, *oldmm;
2288 prepare_task_switch(rq, prev, next);
2290 oldmm = prev->active_mm;
2292 * For paravirt, this is coupled with an exit in switch_to to
2293 * combine the page table reload and the switch backend into
2296 arch_enter_lazy_cpu_mode();
2298 if (unlikely(!mm)) {
2299 next->active_mm = oldmm;
2300 atomic_inc(&oldmm->mm_count);
2301 enter_lazy_tlb(oldmm, next);
2303 switch_mm(oldmm, mm, next);
2305 if (unlikely(!prev->mm)) {
2306 prev->active_mm = NULL;
2307 rq->prev_mm = oldmm;
2310 * Since the runqueue lock will be released by the next
2311 * task (which is an invalid locking op but in the case
2312 * of the scheduler it's an obvious special-case), so we
2313 * do an early lockdep release here:
2315 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2316 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2319 /* Here we just switch the register state and the stack. */
2320 switch_to(prev, next, prev);
2324 * this_rq must be evaluated again because prev may have moved
2325 * CPUs since it called schedule(), thus the 'rq' on its stack
2326 * frame will be invalid.
2328 finish_task_switch(this_rq(), prev);
2332 * nr_running, nr_uninterruptible and nr_context_switches:
2334 * externally visible scheduler statistics: current number of runnable
2335 * threads, current number of uninterruptible-sleeping threads, total
2336 * number of context switches performed since bootup.
2338 unsigned long nr_running(void)
2340 unsigned long i, sum = 0;
2342 for_each_online_cpu(i)
2343 sum += cpu_rq(i)->nr_running;
2348 unsigned long nr_uninterruptible(void)
2350 unsigned long i, sum = 0;
2352 for_each_possible_cpu(i)
2353 sum += cpu_rq(i)->nr_uninterruptible;
2356 * Since we read the counters lockless, it might be slightly
2357 * inaccurate. Do not allow it to go below zero though:
2359 if (unlikely((long)sum < 0))
2365 unsigned long long nr_context_switches(void)
2368 unsigned long long sum = 0;
2370 for_each_possible_cpu(i)
2371 sum += cpu_rq(i)->nr_switches;
2376 unsigned long nr_iowait(void)
2378 unsigned long i, sum = 0;
2380 for_each_possible_cpu(i)
2381 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2386 unsigned long nr_active(void)
2388 unsigned long i, running = 0, uninterruptible = 0;
2390 for_each_online_cpu(i) {
2391 running += cpu_rq(i)->nr_running;
2392 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2395 if (unlikely((long)uninterruptible < 0))
2396 uninterruptible = 0;
2398 return running + uninterruptible;
2402 * Update rq->cpu_load[] statistics. This function is usually called every
2403 * scheduler tick (TICK_NSEC).
2405 static void update_cpu_load(struct rq *this_rq)
2407 unsigned long this_load = this_rq->load.weight;
2410 this_rq->nr_load_updates++;
2412 /* Update our load: */
2413 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2414 unsigned long old_load, new_load;
2416 /* scale is effectively 1 << i now, and >> i divides by scale */
2418 old_load = this_rq->cpu_load[i];
2419 new_load = this_load;
2421 * Round up the averaging division if load is increasing. This
2422 * prevents us from getting stuck on 9 if the load is 10, for
2425 if (new_load > old_load)
2426 new_load += scale-1;
2427 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2434 * double_rq_lock - safely lock two runqueues
2436 * Note this does not disable interrupts like task_rq_lock,
2437 * you need to do so manually before calling.
2439 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2440 __acquires(rq1->lock)
2441 __acquires(rq2->lock)
2443 BUG_ON(!irqs_disabled());
2445 spin_lock(&rq1->lock);
2446 __acquire(rq2->lock); /* Fake it out ;) */
2449 spin_lock(&rq1->lock);
2450 spin_lock(&rq2->lock);
2452 spin_lock(&rq2->lock);
2453 spin_lock(&rq1->lock);
2456 update_rq_clock(rq1);
2457 update_rq_clock(rq2);
2461 * double_rq_unlock - safely unlock two runqueues
2463 * Note this does not restore interrupts like task_rq_unlock,
2464 * you need to do so manually after calling.
2466 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2467 __releases(rq1->lock)
2468 __releases(rq2->lock)
2470 spin_unlock(&rq1->lock);
2472 spin_unlock(&rq2->lock);
2474 __release(rq2->lock);
2478 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2480 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2481 __releases(this_rq->lock)
2482 __acquires(busiest->lock)
2483 __acquires(this_rq->lock)
2487 if (unlikely(!irqs_disabled())) {
2488 /* printk() doesn't work good under rq->lock */
2489 spin_unlock(&this_rq->lock);
2492 if (unlikely(!spin_trylock(&busiest->lock))) {
2493 if (busiest < this_rq) {
2494 spin_unlock(&this_rq->lock);
2495 spin_lock(&busiest->lock);
2496 spin_lock(&this_rq->lock);
2499 spin_lock(&busiest->lock);
2505 * If dest_cpu is allowed for this process, migrate the task to it.
2506 * This is accomplished by forcing the cpu_allowed mask to only
2507 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2508 * the cpu_allowed mask is restored.
2510 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2512 struct migration_req req;
2513 unsigned long flags;
2516 rq = task_rq_lock(p, &flags);
2517 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2518 || unlikely(cpu_is_offline(dest_cpu)))
2521 /* force the process onto the specified CPU */
2522 if (migrate_task(p, dest_cpu, &req)) {
2523 /* Need to wait for migration thread (might exit: take ref). */
2524 struct task_struct *mt = rq->migration_thread;
2526 get_task_struct(mt);
2527 task_rq_unlock(rq, &flags);
2528 wake_up_process(mt);
2529 put_task_struct(mt);
2530 wait_for_completion(&req.done);
2535 task_rq_unlock(rq, &flags);
2539 * sched_exec - execve() is a valuable balancing opportunity, because at
2540 * this point the task has the smallest effective memory and cache footprint.
2542 void sched_exec(void)
2544 int new_cpu, this_cpu = get_cpu();
2545 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2547 if (new_cpu != this_cpu)
2548 sched_migrate_task(current, new_cpu);
2552 * pull_task - move a task from a remote runqueue to the local runqueue.
2553 * Both runqueues must be locked.
2555 static void pull_task(struct rq *src_rq, struct task_struct *p,
2556 struct rq *this_rq, int this_cpu)
2558 deactivate_task(src_rq, p, 0);
2559 set_task_cpu(p, this_cpu);
2560 activate_task(this_rq, p, 0);
2562 * Note that idle threads have a prio of MAX_PRIO, for this test
2563 * to be always true for them.
2565 check_preempt_curr(this_rq, p);
2569 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2572 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2573 struct sched_domain *sd, enum cpu_idle_type idle,
2577 * We do not migrate tasks that are:
2578 * 1) running (obviously), or
2579 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2580 * 3) are cache-hot on their current CPU.
2582 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2583 schedstat_inc(p, se.nr_failed_migrations_affine);
2588 if (task_running(rq, p)) {
2589 schedstat_inc(p, se.nr_failed_migrations_running);
2594 * Aggressive migration if:
2595 * 1) task is cache cold, or
2596 * 2) too many balance attempts have failed.
2599 if (!task_hot(p, rq->clock, sd) ||
2600 sd->nr_balance_failed > sd->cache_nice_tries) {
2601 #ifdef CONFIG_SCHEDSTATS
2602 if (task_hot(p, rq->clock, sd)) {
2603 schedstat_inc(sd, lb_hot_gained[idle]);
2604 schedstat_inc(p, se.nr_forced_migrations);
2610 if (task_hot(p, rq->clock, sd)) {
2611 schedstat_inc(p, se.nr_failed_migrations_hot);
2617 static unsigned long
2618 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2619 unsigned long max_load_move, struct sched_domain *sd,
2620 enum cpu_idle_type idle, int *all_pinned,
2621 int *this_best_prio, struct rq_iterator *iterator)
2623 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2624 struct task_struct *p;
2625 long rem_load_move = max_load_move;
2627 if (max_load_move == 0)
2633 * Start the load-balancing iterator:
2635 p = iterator->start(iterator->arg);
2637 if (!p || loops++ > sysctl_sched_nr_migrate)
2640 * To help distribute high priority tasks across CPUs we don't
2641 * skip a task if it will be the highest priority task (i.e. smallest
2642 * prio value) on its new queue regardless of its load weight
2644 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2645 SCHED_LOAD_SCALE_FUZZ;
2646 if ((skip_for_load && p->prio >= *this_best_prio) ||
2647 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2648 p = iterator->next(iterator->arg);
2652 pull_task(busiest, p, this_rq, this_cpu);
2654 rem_load_move -= p->se.load.weight;
2657 * We only want to steal up to the prescribed amount of weighted load.
2659 if (rem_load_move > 0) {
2660 if (p->prio < *this_best_prio)
2661 *this_best_prio = p->prio;
2662 p = iterator->next(iterator->arg);
2667 * Right now, this is one of only two places pull_task() is called,
2668 * so we can safely collect pull_task() stats here rather than
2669 * inside pull_task().
2671 schedstat_add(sd, lb_gained[idle], pulled);
2674 *all_pinned = pinned;
2676 return max_load_move - rem_load_move;
2680 * move_tasks tries to move up to max_load_move weighted load from busiest to
2681 * this_rq, as part of a balancing operation within domain "sd".
2682 * Returns 1 if successful and 0 otherwise.
2684 * Called with both runqueues locked.
2686 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2687 unsigned long max_load_move,
2688 struct sched_domain *sd, enum cpu_idle_type idle,
2691 const struct sched_class *class = sched_class_highest;
2692 unsigned long total_load_moved = 0;
2693 int this_best_prio = this_rq->curr->prio;
2697 class->load_balance(this_rq, this_cpu, busiest,
2698 max_load_move - total_load_moved,
2699 sd, idle, all_pinned, &this_best_prio);
2700 class = class->next;
2701 } while (class && max_load_move > total_load_moved);
2703 return total_load_moved > 0;
2707 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2708 struct sched_domain *sd, enum cpu_idle_type idle,
2709 struct rq_iterator *iterator)
2711 struct task_struct *p = iterator->start(iterator->arg);
2715 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2716 pull_task(busiest, p, this_rq, this_cpu);
2718 * Right now, this is only the second place pull_task()
2719 * is called, so we can safely collect pull_task()
2720 * stats here rather than inside pull_task().
2722 schedstat_inc(sd, lb_gained[idle]);
2726 p = iterator->next(iterator->arg);
2733 * move_one_task tries to move exactly one task from busiest to this_rq, as
2734 * part of active balancing operations within "domain".
2735 * Returns 1 if successful and 0 otherwise.
2737 * Called with both runqueues locked.
2739 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2740 struct sched_domain *sd, enum cpu_idle_type idle)
2742 const struct sched_class *class;
2744 for (class = sched_class_highest; class; class = class->next)
2745 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2752 * find_busiest_group finds and returns the busiest CPU group within the
2753 * domain. It calculates and returns the amount of weighted load which
2754 * should be moved to restore balance via the imbalance parameter.
2756 static struct sched_group *
2757 find_busiest_group(struct sched_domain *sd, int this_cpu,
2758 unsigned long *imbalance, enum cpu_idle_type idle,
2759 int *sd_idle, cpumask_t *cpus, int *balance)
2761 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2762 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2763 unsigned long max_pull;
2764 unsigned long busiest_load_per_task, busiest_nr_running;
2765 unsigned long this_load_per_task, this_nr_running;
2766 int load_idx, group_imb = 0;
2767 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2768 int power_savings_balance = 1;
2769 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2770 unsigned long min_nr_running = ULONG_MAX;
2771 struct sched_group *group_min = NULL, *group_leader = NULL;
2774 max_load = this_load = total_load = total_pwr = 0;
2775 busiest_load_per_task = busiest_nr_running = 0;
2776 this_load_per_task = this_nr_running = 0;
2777 if (idle == CPU_NOT_IDLE)
2778 load_idx = sd->busy_idx;
2779 else if (idle == CPU_NEWLY_IDLE)
2780 load_idx = sd->newidle_idx;
2782 load_idx = sd->idle_idx;
2785 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2788 int __group_imb = 0;
2789 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2790 unsigned long sum_nr_running, sum_weighted_load;
2792 local_group = cpu_isset(this_cpu, group->cpumask);
2795 balance_cpu = first_cpu(group->cpumask);
2797 /* Tally up the load of all CPUs in the group */
2798 sum_weighted_load = sum_nr_running = avg_load = 0;
2800 min_cpu_load = ~0UL;
2802 for_each_cpu_mask(i, group->cpumask) {
2805 if (!cpu_isset(i, *cpus))
2810 if (*sd_idle && rq->nr_running)
2813 /* Bias balancing toward cpus of our domain */
2815 if (idle_cpu(i) && !first_idle_cpu) {
2820 load = target_load(i, load_idx);
2822 load = source_load(i, load_idx);
2823 if (load > max_cpu_load)
2824 max_cpu_load = load;
2825 if (min_cpu_load > load)
2826 min_cpu_load = load;
2830 sum_nr_running += rq->nr_running;
2831 sum_weighted_load += weighted_cpuload(i);
2835 * First idle cpu or the first cpu(busiest) in this sched group
2836 * is eligible for doing load balancing at this and above
2837 * domains. In the newly idle case, we will allow all the cpu's
2838 * to do the newly idle load balance.
2840 if (idle != CPU_NEWLY_IDLE && local_group &&
2841 balance_cpu != this_cpu && balance) {
2846 total_load += avg_load;
2847 total_pwr += group->__cpu_power;
2849 /* Adjust by relative CPU power of the group */
2850 avg_load = sg_div_cpu_power(group,
2851 avg_load * SCHED_LOAD_SCALE);
2853 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2856 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2859 this_load = avg_load;
2861 this_nr_running = sum_nr_running;
2862 this_load_per_task = sum_weighted_load;
2863 } else if (avg_load > max_load &&
2864 (sum_nr_running > group_capacity || __group_imb)) {
2865 max_load = avg_load;
2867 busiest_nr_running = sum_nr_running;
2868 busiest_load_per_task = sum_weighted_load;
2869 group_imb = __group_imb;
2872 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2874 * Busy processors will not participate in power savings
2877 if (idle == CPU_NOT_IDLE ||
2878 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2882 * If the local group is idle or completely loaded
2883 * no need to do power savings balance at this domain
2885 if (local_group && (this_nr_running >= group_capacity ||
2887 power_savings_balance = 0;
2890 * If a group is already running at full capacity or idle,
2891 * don't include that group in power savings calculations
2893 if (!power_savings_balance || sum_nr_running >= group_capacity
2898 * Calculate the group which has the least non-idle load.
2899 * This is the group from where we need to pick up the load
2902 if ((sum_nr_running < min_nr_running) ||
2903 (sum_nr_running == min_nr_running &&
2904 first_cpu(group->cpumask) <
2905 first_cpu(group_min->cpumask))) {
2907 min_nr_running = sum_nr_running;
2908 min_load_per_task = sum_weighted_load /
2913 * Calculate the group which is almost near its
2914 * capacity but still has some space to pick up some load
2915 * from other group and save more power
2917 if (sum_nr_running <= group_capacity - 1) {
2918 if (sum_nr_running > leader_nr_running ||
2919 (sum_nr_running == leader_nr_running &&
2920 first_cpu(group->cpumask) >
2921 first_cpu(group_leader->cpumask))) {
2922 group_leader = group;
2923 leader_nr_running = sum_nr_running;
2928 group = group->next;
2929 } while (group != sd->groups);
2931 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2934 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2936 if (this_load >= avg_load ||
2937 100*max_load <= sd->imbalance_pct*this_load)
2940 busiest_load_per_task /= busiest_nr_running;
2942 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2945 * We're trying to get all the cpus to the average_load, so we don't
2946 * want to push ourselves above the average load, nor do we wish to
2947 * reduce the max loaded cpu below the average load, as either of these
2948 * actions would just result in more rebalancing later, and ping-pong
2949 * tasks around. Thus we look for the minimum possible imbalance.
2950 * Negative imbalances (*we* are more loaded than anyone else) will
2951 * be counted as no imbalance for these purposes -- we can't fix that
2952 * by pulling tasks to us. Be careful of negative numbers as they'll
2953 * appear as very large values with unsigned longs.
2955 if (max_load <= busiest_load_per_task)
2959 * In the presence of smp nice balancing, certain scenarios can have
2960 * max load less than avg load(as we skip the groups at or below
2961 * its cpu_power, while calculating max_load..)
2963 if (max_load < avg_load) {
2965 goto small_imbalance;
2968 /* Don't want to pull so many tasks that a group would go idle */
2969 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2971 /* How much load to actually move to equalise the imbalance */
2972 *imbalance = min(max_pull * busiest->__cpu_power,
2973 (avg_load - this_load) * this->__cpu_power)
2977 * if *imbalance is less than the average load per runnable task
2978 * there is no gaurantee that any tasks will be moved so we'll have
2979 * a think about bumping its value to force at least one task to be
2982 if (*imbalance < busiest_load_per_task) {
2983 unsigned long tmp, pwr_now, pwr_move;
2987 pwr_move = pwr_now = 0;
2989 if (this_nr_running) {
2990 this_load_per_task /= this_nr_running;
2991 if (busiest_load_per_task > this_load_per_task)
2994 this_load_per_task = SCHED_LOAD_SCALE;
2996 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2997 busiest_load_per_task * imbn) {
2998 *imbalance = busiest_load_per_task;
3003 * OK, we don't have enough imbalance to justify moving tasks,
3004 * however we may be able to increase total CPU power used by
3008 pwr_now += busiest->__cpu_power *
3009 min(busiest_load_per_task, max_load);
3010 pwr_now += this->__cpu_power *
3011 min(this_load_per_task, this_load);
3012 pwr_now /= SCHED_LOAD_SCALE;
3014 /* Amount of load we'd subtract */
3015 tmp = sg_div_cpu_power(busiest,
3016 busiest_load_per_task * SCHED_LOAD_SCALE);
3018 pwr_move += busiest->__cpu_power *
3019 min(busiest_load_per_task, max_load - tmp);
3021 /* Amount of load we'd add */
3022 if (max_load * busiest->__cpu_power <
3023 busiest_load_per_task * SCHED_LOAD_SCALE)
3024 tmp = sg_div_cpu_power(this,
3025 max_load * busiest->__cpu_power);
3027 tmp = sg_div_cpu_power(this,
3028 busiest_load_per_task * SCHED_LOAD_SCALE);
3029 pwr_move += this->__cpu_power *
3030 min(this_load_per_task, this_load + tmp);
3031 pwr_move /= SCHED_LOAD_SCALE;
3033 /* Move if we gain throughput */
3034 if (pwr_move > pwr_now)
3035 *imbalance = busiest_load_per_task;
3041 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3042 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3045 if (this == group_leader && group_leader != group_min) {
3046 *imbalance = min_load_per_task;
3056 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3059 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3060 unsigned long imbalance, cpumask_t *cpus)
3062 struct rq *busiest = NULL, *rq;
3063 unsigned long max_load = 0;
3066 for_each_cpu_mask(i, group->cpumask) {
3069 if (!cpu_isset(i, *cpus))
3073 wl = weighted_cpuload(i);
3075 if (rq->nr_running == 1 && wl > imbalance)
3078 if (wl > max_load) {
3088 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3089 * so long as it is large enough.
3091 #define MAX_PINNED_INTERVAL 512
3094 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3095 * tasks if there is an imbalance.
3097 static int load_balance(int this_cpu, struct rq *this_rq,
3098 struct sched_domain *sd, enum cpu_idle_type idle,
3101 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3102 struct sched_group *group;
3103 unsigned long imbalance;
3105 cpumask_t cpus = CPU_MASK_ALL;
3106 unsigned long flags;
3109 * When power savings policy is enabled for the parent domain, idle
3110 * sibling can pick up load irrespective of busy siblings. In this case,
3111 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3112 * portraying it as CPU_NOT_IDLE.
3114 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3115 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3118 schedstat_inc(sd, lb_count[idle]);
3121 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3128 schedstat_inc(sd, lb_nobusyg[idle]);
3132 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3134 schedstat_inc(sd, lb_nobusyq[idle]);
3138 BUG_ON(busiest == this_rq);
3140 schedstat_add(sd, lb_imbalance[idle], imbalance);
3143 if (busiest->nr_running > 1) {
3145 * Attempt to move tasks. If find_busiest_group has found
3146 * an imbalance but busiest->nr_running <= 1, the group is
3147 * still unbalanced. ld_moved simply stays zero, so it is
3148 * correctly treated as an imbalance.
3150 local_irq_save(flags);
3151 double_rq_lock(this_rq, busiest);
3152 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3153 imbalance, sd, idle, &all_pinned);
3154 double_rq_unlock(this_rq, busiest);
3155 local_irq_restore(flags);
3158 * some other cpu did the load balance for us.
3160 if (ld_moved && this_cpu != smp_processor_id())
3161 resched_cpu(this_cpu);
3163 /* All tasks on this runqueue were pinned by CPU affinity */
3164 if (unlikely(all_pinned)) {
3165 cpu_clear(cpu_of(busiest), cpus);
3166 if (!cpus_empty(cpus))
3173 schedstat_inc(sd, lb_failed[idle]);
3174 sd->nr_balance_failed++;
3176 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3178 spin_lock_irqsave(&busiest->lock, flags);
3180 /* don't kick the migration_thread, if the curr
3181 * task on busiest cpu can't be moved to this_cpu
3183 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3184 spin_unlock_irqrestore(&busiest->lock, flags);
3186 goto out_one_pinned;
3189 if (!busiest->active_balance) {
3190 busiest->active_balance = 1;
3191 busiest->push_cpu = this_cpu;
3194 spin_unlock_irqrestore(&busiest->lock, flags);
3196 wake_up_process(busiest->migration_thread);
3199 * We've kicked active balancing, reset the failure
3202 sd->nr_balance_failed = sd->cache_nice_tries+1;
3205 sd->nr_balance_failed = 0;
3207 if (likely(!active_balance)) {
3208 /* We were unbalanced, so reset the balancing interval */
3209 sd->balance_interval = sd->min_interval;
3212 * If we've begun active balancing, start to back off. This
3213 * case may not be covered by the all_pinned logic if there
3214 * is only 1 task on the busy runqueue (because we don't call
3217 if (sd->balance_interval < sd->max_interval)
3218 sd->balance_interval *= 2;
3221 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3222 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3227 schedstat_inc(sd, lb_balanced[idle]);
3229 sd->nr_balance_failed = 0;
3232 /* tune up the balancing interval */
3233 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3234 (sd->balance_interval < sd->max_interval))
3235 sd->balance_interval *= 2;
3237 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3238 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3244 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3245 * tasks if there is an imbalance.
3247 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3248 * this_rq is locked.
3251 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3253 struct sched_group *group;
3254 struct rq *busiest = NULL;
3255 unsigned long imbalance;
3259 cpumask_t cpus = CPU_MASK_ALL;
3262 * When power savings policy is enabled for the parent domain, idle
3263 * sibling can pick up load irrespective of busy siblings. In this case,
3264 * let the state of idle sibling percolate up as IDLE, instead of
3265 * portraying it as CPU_NOT_IDLE.
3267 if (sd->flags & SD_SHARE_CPUPOWER &&
3268 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3271 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3273 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3274 &sd_idle, &cpus, NULL);
3276 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3280 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3283 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3287 BUG_ON(busiest == this_rq);
3289 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3292 if (busiest->nr_running > 1) {
3293 /* Attempt to move tasks */
3294 double_lock_balance(this_rq, busiest);
3295 /* this_rq->clock is already updated */
3296 update_rq_clock(busiest);
3297 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3298 imbalance, sd, CPU_NEWLY_IDLE,
3300 spin_unlock(&busiest->lock);
3302 if (unlikely(all_pinned)) {
3303 cpu_clear(cpu_of(busiest), cpus);
3304 if (!cpus_empty(cpus))
3310 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3311 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3312 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3315 sd->nr_balance_failed = 0;
3320 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3321 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3322 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3324 sd->nr_balance_failed = 0;
3330 * idle_balance is called by schedule() if this_cpu is about to become
3331 * idle. Attempts to pull tasks from other CPUs.
3333 static void idle_balance(int this_cpu, struct rq *this_rq)
3335 struct sched_domain *sd;
3336 int pulled_task = -1;
3337 unsigned long next_balance = jiffies + HZ;
3339 for_each_domain(this_cpu, sd) {
3340 unsigned long interval;
3342 if (!(sd->flags & SD_LOAD_BALANCE))
3345 if (sd->flags & SD_BALANCE_NEWIDLE)
3346 /* If we've pulled tasks over stop searching: */
3347 pulled_task = load_balance_newidle(this_cpu,
3350 interval = msecs_to_jiffies(sd->balance_interval);
3351 if (time_after(next_balance, sd->last_balance + interval))
3352 next_balance = sd->last_balance + interval;
3356 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3358 * We are going idle. next_balance may be set based on
3359 * a busy processor. So reset next_balance.
3361 this_rq->next_balance = next_balance;
3366 * active_load_balance is run by migration threads. It pushes running tasks
3367 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3368 * running on each physical CPU where possible, and avoids physical /
3369 * logical imbalances.
3371 * Called with busiest_rq locked.
3373 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3375 int target_cpu = busiest_rq->push_cpu;
3376 struct sched_domain *sd;
3377 struct rq *target_rq;
3379 /* Is there any task to move? */
3380 if (busiest_rq->nr_running <= 1)
3383 target_rq = cpu_rq(target_cpu);
3386 * This condition is "impossible", if it occurs
3387 * we need to fix it. Originally reported by
3388 * Bjorn Helgaas on a 128-cpu setup.
3390 BUG_ON(busiest_rq == target_rq);
3392 /* move a task from busiest_rq to target_rq */
3393 double_lock_balance(busiest_rq, target_rq);
3394 update_rq_clock(busiest_rq);
3395 update_rq_clock(target_rq);
3397 /* Search for an sd spanning us and the target CPU. */
3398 for_each_domain(target_cpu, sd) {
3399 if ((sd->flags & SD_LOAD_BALANCE) &&
3400 cpu_isset(busiest_cpu, sd->span))
3405 schedstat_inc(sd, alb_count);
3407 if (move_one_task(target_rq, target_cpu, busiest_rq,
3409 schedstat_inc(sd, alb_pushed);
3411 schedstat_inc(sd, alb_failed);
3413 spin_unlock(&target_rq->lock);
3418 atomic_t load_balancer;
3420 } nohz ____cacheline_aligned = {
3421 .load_balancer = ATOMIC_INIT(-1),
3422 .cpu_mask = CPU_MASK_NONE,
3426 * This routine will try to nominate the ilb (idle load balancing)
3427 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3428 * load balancing on behalf of all those cpus. If all the cpus in the system
3429 * go into this tickless mode, then there will be no ilb owner (as there is
3430 * no need for one) and all the cpus will sleep till the next wakeup event
3433 * For the ilb owner, tick is not stopped. And this tick will be used
3434 * for idle load balancing. ilb owner will still be part of
3437 * While stopping the tick, this cpu will become the ilb owner if there
3438 * is no other owner. And will be the owner till that cpu becomes busy
3439 * or if all cpus in the system stop their ticks at which point
3440 * there is no need for ilb owner.
3442 * When the ilb owner becomes busy, it nominates another owner, during the
3443 * next busy scheduler_tick()
3445 int select_nohz_load_balancer(int stop_tick)
3447 int cpu = smp_processor_id();
3450 cpu_set(cpu, nohz.cpu_mask);
3451 cpu_rq(cpu)->in_nohz_recently = 1;
3454 * If we are going offline and still the leader, give up!
3456 if (cpu_is_offline(cpu) &&
3457 atomic_read(&nohz.load_balancer) == cpu) {
3458 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3463 /* time for ilb owner also to sleep */
3464 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3465 if (atomic_read(&nohz.load_balancer) == cpu)
3466 atomic_set(&nohz.load_balancer, -1);
3470 if (atomic_read(&nohz.load_balancer) == -1) {
3471 /* make me the ilb owner */
3472 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3474 } else if (atomic_read(&nohz.load_balancer) == cpu)
3477 if (!cpu_isset(cpu, nohz.cpu_mask))
3480 cpu_clear(cpu, nohz.cpu_mask);
3482 if (atomic_read(&nohz.load_balancer) == cpu)
3483 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3490 static DEFINE_SPINLOCK(balancing);
3493 * It checks each scheduling domain to see if it is due to be balanced,
3494 * and initiates a balancing operation if so.
3496 * Balancing parameters are set up in arch_init_sched_domains.
3498 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3501 struct rq *rq = cpu_rq(cpu);
3502 unsigned long interval;
3503 struct sched_domain *sd;
3504 /* Earliest time when we have to do rebalance again */
3505 unsigned long next_balance = jiffies + 60*HZ;
3506 int update_next_balance = 0;
3508 for_each_domain(cpu, sd) {
3509 if (!(sd->flags & SD_LOAD_BALANCE))
3512 interval = sd->balance_interval;
3513 if (idle != CPU_IDLE)
3514 interval *= sd->busy_factor;
3516 /* scale ms to jiffies */
3517 interval = msecs_to_jiffies(interval);
3518 if (unlikely(!interval))
3520 if (interval > HZ*NR_CPUS/10)
3521 interval = HZ*NR_CPUS/10;
3524 if (sd->flags & SD_SERIALIZE) {
3525 if (!spin_trylock(&balancing))
3529 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3530 if (load_balance(cpu, rq, sd, idle, &balance)) {
3532 * We've pulled tasks over so either we're no
3533 * longer idle, or one of our SMT siblings is
3536 idle = CPU_NOT_IDLE;
3538 sd->last_balance = jiffies;
3540 if (sd->flags & SD_SERIALIZE)
3541 spin_unlock(&balancing);
3543 if (time_after(next_balance, sd->last_balance + interval)) {
3544 next_balance = sd->last_balance + interval;
3545 update_next_balance = 1;
3549 * Stop the load balance at this level. There is another
3550 * CPU in our sched group which is doing load balancing more
3558 * next_balance will be updated only when there is a need.
3559 * When the cpu is attached to null domain for ex, it will not be
3562 if (likely(update_next_balance))
3563 rq->next_balance = next_balance;
3567 * run_rebalance_domains is triggered when needed from the scheduler tick.
3568 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3569 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3571 static void run_rebalance_domains(struct softirq_action *h)
3573 int this_cpu = smp_processor_id();
3574 struct rq *this_rq = cpu_rq(this_cpu);
3575 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3576 CPU_IDLE : CPU_NOT_IDLE;
3578 rebalance_domains(this_cpu, idle);
3582 * If this cpu is the owner for idle load balancing, then do the
3583 * balancing on behalf of the other idle cpus whose ticks are
3586 if (this_rq->idle_at_tick &&
3587 atomic_read(&nohz.load_balancer) == this_cpu) {
3588 cpumask_t cpus = nohz.cpu_mask;
3592 cpu_clear(this_cpu, cpus);
3593 for_each_cpu_mask(balance_cpu, cpus) {
3595 * If this cpu gets work to do, stop the load balancing
3596 * work being done for other cpus. Next load
3597 * balancing owner will pick it up.
3602 rebalance_domains(balance_cpu, CPU_IDLE);
3604 rq = cpu_rq(balance_cpu);
3605 if (time_after(this_rq->next_balance, rq->next_balance))
3606 this_rq->next_balance = rq->next_balance;
3613 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3615 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3616 * idle load balancing owner or decide to stop the periodic load balancing,
3617 * if the whole system is idle.
3619 static inline void trigger_load_balance(struct rq *rq, int cpu)
3623 * If we were in the nohz mode recently and busy at the current
3624 * scheduler tick, then check if we need to nominate new idle
3627 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3628 rq->in_nohz_recently = 0;
3630 if (atomic_read(&nohz.load_balancer) == cpu) {
3631 cpu_clear(cpu, nohz.cpu_mask);
3632 atomic_set(&nohz.load_balancer, -1);
3635 if (atomic_read(&nohz.load_balancer) == -1) {
3637 * simple selection for now: Nominate the
3638 * first cpu in the nohz list to be the next
3641 * TBD: Traverse the sched domains and nominate
3642 * the nearest cpu in the nohz.cpu_mask.
3644 int ilb = first_cpu(nohz.cpu_mask);
3652 * If this cpu is idle and doing idle load balancing for all the
3653 * cpus with ticks stopped, is it time for that to stop?
3655 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3656 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3662 * If this cpu is idle and the idle load balancing is done by
3663 * someone else, then no need raise the SCHED_SOFTIRQ
3665 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3666 cpu_isset(cpu, nohz.cpu_mask))
3669 if (time_after_eq(jiffies, rq->next_balance))
3670 raise_softirq(SCHED_SOFTIRQ);
3673 #else /* CONFIG_SMP */
3676 * on UP we do not need to balance between CPUs:
3678 static inline void idle_balance(int cpu, struct rq *rq)
3684 DEFINE_PER_CPU(struct kernel_stat, kstat);
3686 EXPORT_PER_CPU_SYMBOL(kstat);
3689 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3690 * that have not yet been banked in case the task is currently running.
3692 unsigned long long task_sched_runtime(struct task_struct *p)
3694 unsigned long flags;
3698 rq = task_rq_lock(p, &flags);
3699 ns = p->se.sum_exec_runtime;
3700 if (task_current(rq, p)) {
3701 update_rq_clock(rq);
3702 delta_exec = rq->clock - p->se.exec_start;
3703 if ((s64)delta_exec > 0)
3706 task_rq_unlock(rq, &flags);
3712 * Account user cpu time to a process.
3713 * @p: the process that the cpu time gets accounted to
3714 * @cputime: the cpu time spent in user space since the last update
3716 void account_user_time(struct task_struct *p, cputime_t cputime)
3718 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3721 p->utime = cputime_add(p->utime, cputime);
3723 /* Add user time to cpustat. */
3724 tmp = cputime_to_cputime64(cputime);
3725 if (TASK_NICE(p) > 0)
3726 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3728 cpustat->user = cputime64_add(cpustat->user, tmp);
3732 * Account guest cpu time to a process.
3733 * @p: the process that the cpu time gets accounted to
3734 * @cputime: the cpu time spent in virtual machine since the last update
3736 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3739 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3741 tmp = cputime_to_cputime64(cputime);
3743 p->utime = cputime_add(p->utime, cputime);
3744 p->gtime = cputime_add(p->gtime, cputime);
3746 cpustat->user = cputime64_add(cpustat->user, tmp);
3747 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3751 * Account scaled user cpu time to a process.
3752 * @p: the process that the cpu time gets accounted to
3753 * @cputime: the cpu time spent in user space since the last update
3755 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3757 p->utimescaled = cputime_add(p->utimescaled, cputime);
3761 * Account system cpu time to a process.
3762 * @p: the process that the cpu time gets accounted to
3763 * @hardirq_offset: the offset to subtract from hardirq_count()
3764 * @cputime: the cpu time spent in kernel space since the last update
3766 void account_system_time(struct task_struct *p, int hardirq_offset,
3769 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3770 struct rq *rq = this_rq();
3773 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3774 return account_guest_time(p, cputime);
3776 p->stime = cputime_add(p->stime, cputime);
3778 /* Add system time to cpustat. */
3779 tmp = cputime_to_cputime64(cputime);
3780 if (hardirq_count() - hardirq_offset)
3781 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3782 else if (softirq_count())
3783 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3784 else if (p != rq->idle)
3785 cpustat->system = cputime64_add(cpustat->system, tmp);
3786 else if (atomic_read(&rq->nr_iowait) > 0)
3787 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3789 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3790 /* Account for system time used */
3791 acct_update_integrals(p);
3795 * Account scaled system cpu time to a process.
3796 * @p: the process that the cpu time gets accounted to
3797 * @hardirq_offset: the offset to subtract from hardirq_count()
3798 * @cputime: the cpu time spent in kernel space since the last update
3800 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3802 p->stimescaled = cputime_add(p->stimescaled, cputime);
3806 * Account for involuntary wait time.
3807 * @p: the process from which the cpu time has been stolen
3808 * @steal: the cpu time spent in involuntary wait
3810 void account_steal_time(struct task_struct *p, cputime_t steal)
3812 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3813 cputime64_t tmp = cputime_to_cputime64(steal);
3814 struct rq *rq = this_rq();
3816 if (p == rq->idle) {
3817 p->stime = cputime_add(p->stime, steal);
3818 if (atomic_read(&rq->nr_iowait) > 0)
3819 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3821 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3823 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3827 * This function gets called by the timer code, with HZ frequency.
3828 * We call it with interrupts disabled.
3830 * It also gets called by the fork code, when changing the parent's
3833 void scheduler_tick(void)
3835 int cpu = smp_processor_id();
3836 struct rq *rq = cpu_rq(cpu);
3837 struct task_struct *curr = rq->curr;
3838 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3840 spin_lock(&rq->lock);
3841 __update_rq_clock(rq);
3843 * Let rq->clock advance by at least TICK_NSEC:
3845 if (unlikely(rq->clock < next_tick)) {
3846 rq->clock = next_tick;
3847 rq->clock_underflows++;
3849 rq->tick_timestamp = rq->clock;
3850 update_last_tick_seen(rq);
3851 update_cpu_load(rq);
3852 curr->sched_class->task_tick(rq, curr, 0);
3853 update_sched_rt_period(rq);
3854 spin_unlock(&rq->lock);
3857 rq->idle_at_tick = idle_cpu(cpu);
3858 trigger_load_balance(rq, cpu);
3862 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3864 void __kprobes add_preempt_count(int val)
3869 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3871 preempt_count() += val;
3873 * Spinlock count overflowing soon?
3875 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3878 EXPORT_SYMBOL(add_preempt_count);
3880 void __kprobes sub_preempt_count(int val)
3885 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3888 * Is the spinlock portion underflowing?
3890 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3891 !(preempt_count() & PREEMPT_MASK)))
3894 preempt_count() -= val;
3896 EXPORT_SYMBOL(sub_preempt_count);
3901 * Print scheduling while atomic bug:
3903 static noinline void __schedule_bug(struct task_struct *prev)
3905 struct pt_regs *regs = get_irq_regs();
3907 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3908 prev->comm, prev->pid, preempt_count());
3910 debug_show_held_locks(prev);
3911 if (irqs_disabled())
3912 print_irqtrace_events(prev);
3921 * Various schedule()-time debugging checks and statistics:
3923 static inline void schedule_debug(struct task_struct *prev)
3926 * Test if we are atomic. Since do_exit() needs to call into
3927 * schedule() atomically, we ignore that path for now.
3928 * Otherwise, whine if we are scheduling when we should not be.
3930 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3931 __schedule_bug(prev);
3933 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3935 schedstat_inc(this_rq(), sched_count);
3936 #ifdef CONFIG_SCHEDSTATS
3937 if (unlikely(prev->lock_depth >= 0)) {
3938 schedstat_inc(this_rq(), bkl_count);
3939 schedstat_inc(prev, sched_info.bkl_count);
3945 * Pick up the highest-prio task:
3947 static inline struct task_struct *
3948 pick_next_task(struct rq *rq, struct task_struct *prev)
3950 const struct sched_class *class;
3951 struct task_struct *p;
3954 * Optimization: we know that if all tasks are in
3955 * the fair class we can call that function directly:
3957 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3958 p = fair_sched_class.pick_next_task(rq);
3963 class = sched_class_highest;
3965 p = class->pick_next_task(rq);
3969 * Will never be NULL as the idle class always
3970 * returns a non-NULL p:
3972 class = class->next;
3977 * schedule() is the main scheduler function.
3979 asmlinkage void __sched schedule(void)
3981 struct task_struct *prev, *next;
3982 unsigned long *switch_count;
3988 cpu = smp_processor_id();
3992 switch_count = &prev->nivcsw;
3994 release_kernel_lock(prev);
3995 need_resched_nonpreemptible:
3997 schedule_debug(prev);
4002 * Do the rq-clock update outside the rq lock:
4004 local_irq_disable();
4005 __update_rq_clock(rq);
4006 spin_lock(&rq->lock);
4007 clear_tsk_need_resched(prev);
4009 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4010 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4011 signal_pending(prev))) {
4012 prev->state = TASK_RUNNING;
4014 deactivate_task(rq, prev, 1);
4016 switch_count = &prev->nvcsw;
4020 if (prev->sched_class->pre_schedule)
4021 prev->sched_class->pre_schedule(rq, prev);
4024 if (unlikely(!rq->nr_running))
4025 idle_balance(cpu, rq);
4027 prev->sched_class->put_prev_task(rq, prev);
4028 next = pick_next_task(rq, prev);
4030 sched_info_switch(prev, next);
4032 if (likely(prev != next)) {
4037 context_switch(rq, prev, next); /* unlocks the rq */
4039 * the context switch might have flipped the stack from under
4040 * us, hence refresh the local variables.
4042 cpu = smp_processor_id();
4045 spin_unlock_irq(&rq->lock);
4049 if (unlikely(reacquire_kernel_lock(current) < 0))
4050 goto need_resched_nonpreemptible;
4052 preempt_enable_no_resched();
4053 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4056 EXPORT_SYMBOL(schedule);
4058 #ifdef CONFIG_PREEMPT
4060 * this is the entry point to schedule() from in-kernel preemption
4061 * off of preempt_enable. Kernel preemptions off return from interrupt
4062 * occur there and call schedule directly.
4064 asmlinkage void __sched preempt_schedule(void)
4066 struct thread_info *ti = current_thread_info();
4067 struct task_struct *task = current;
4068 int saved_lock_depth;
4071 * If there is a non-zero preempt_count or interrupts are disabled,
4072 * we do not want to preempt the current task. Just return..
4074 if (likely(ti->preempt_count || irqs_disabled()))
4078 add_preempt_count(PREEMPT_ACTIVE);
4081 * We keep the big kernel semaphore locked, but we
4082 * clear ->lock_depth so that schedule() doesnt
4083 * auto-release the semaphore:
4085 saved_lock_depth = task->lock_depth;
4086 task->lock_depth = -1;
4088 task->lock_depth = saved_lock_depth;
4089 sub_preempt_count(PREEMPT_ACTIVE);
4092 * Check again in case we missed a preemption opportunity
4093 * between schedule and now.
4096 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4098 EXPORT_SYMBOL(preempt_schedule);
4101 * this is the entry point to schedule() from kernel preemption
4102 * off of irq context.
4103 * Note, that this is called and return with irqs disabled. This will
4104 * protect us against recursive calling from irq.
4106 asmlinkage void __sched preempt_schedule_irq(void)
4108 struct thread_info *ti = current_thread_info();
4109 struct task_struct *task = current;
4110 int saved_lock_depth;
4112 /* Catch callers which need to be fixed */
4113 BUG_ON(ti->preempt_count || !irqs_disabled());
4116 add_preempt_count(PREEMPT_ACTIVE);
4119 * We keep the big kernel semaphore locked, but we
4120 * clear ->lock_depth so that schedule() doesnt
4121 * auto-release the semaphore:
4123 saved_lock_depth = task->lock_depth;
4124 task->lock_depth = -1;
4127 local_irq_disable();
4128 task->lock_depth = saved_lock_depth;
4129 sub_preempt_count(PREEMPT_ACTIVE);
4132 * Check again in case we missed a preemption opportunity
4133 * between schedule and now.
4136 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4139 #endif /* CONFIG_PREEMPT */
4141 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4144 return try_to_wake_up(curr->private, mode, sync);
4146 EXPORT_SYMBOL(default_wake_function);
4149 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4150 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4151 * number) then we wake all the non-exclusive tasks and one exclusive task.
4153 * There are circumstances in which we can try to wake a task which has already
4154 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4155 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4157 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4158 int nr_exclusive, int sync, void *key)
4160 wait_queue_t *curr, *next;
4162 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4163 unsigned flags = curr->flags;
4165 if (curr->func(curr, mode, sync, key) &&
4166 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4172 * __wake_up - wake up threads blocked on a waitqueue.
4174 * @mode: which threads
4175 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4176 * @key: is directly passed to the wakeup function
4178 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4179 int nr_exclusive, void *key)
4181 unsigned long flags;
4183 spin_lock_irqsave(&q->lock, flags);
4184 __wake_up_common(q, mode, nr_exclusive, 0, key);
4185 spin_unlock_irqrestore(&q->lock, flags);
4187 EXPORT_SYMBOL(__wake_up);
4190 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4192 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4194 __wake_up_common(q, mode, 1, 0, NULL);
4198 * __wake_up_sync - wake up threads blocked on a waitqueue.
4200 * @mode: which threads
4201 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4203 * The sync wakeup differs that the waker knows that it will schedule
4204 * away soon, so while the target thread will be woken up, it will not
4205 * be migrated to another CPU - ie. the two threads are 'synchronized'
4206 * with each other. This can prevent needless bouncing between CPUs.
4208 * On UP it can prevent extra preemption.
4211 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4213 unsigned long flags;
4219 if (unlikely(!nr_exclusive))
4222 spin_lock_irqsave(&q->lock, flags);
4223 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4224 spin_unlock_irqrestore(&q->lock, flags);
4226 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4228 void complete(struct completion *x)
4230 unsigned long flags;
4232 spin_lock_irqsave(&x->wait.lock, flags);
4234 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4235 spin_unlock_irqrestore(&x->wait.lock, flags);
4237 EXPORT_SYMBOL(complete);
4239 void complete_all(struct completion *x)
4241 unsigned long flags;
4243 spin_lock_irqsave(&x->wait.lock, flags);
4244 x->done += UINT_MAX/2;
4245 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4246 spin_unlock_irqrestore(&x->wait.lock, flags);
4248 EXPORT_SYMBOL(complete_all);
4250 static inline long __sched
4251 do_wait_for_common(struct completion *x, long timeout, int state)
4254 DECLARE_WAITQUEUE(wait, current);
4256 wait.flags |= WQ_FLAG_EXCLUSIVE;
4257 __add_wait_queue_tail(&x->wait, &wait);
4259 if ((state == TASK_INTERRUPTIBLE &&
4260 signal_pending(current)) ||
4261 (state == TASK_KILLABLE &&
4262 fatal_signal_pending(current))) {
4263 __remove_wait_queue(&x->wait, &wait);
4264 return -ERESTARTSYS;
4266 __set_current_state(state);
4267 spin_unlock_irq(&x->wait.lock);
4268 timeout = schedule_timeout(timeout);
4269 spin_lock_irq(&x->wait.lock);
4271 __remove_wait_queue(&x->wait, &wait);
4275 __remove_wait_queue(&x->wait, &wait);
4282 wait_for_common(struct completion *x, long timeout, int state)
4286 spin_lock_irq(&x->wait.lock);
4287 timeout = do_wait_for_common(x, timeout, state);
4288 spin_unlock_irq(&x->wait.lock);
4292 void __sched wait_for_completion(struct completion *x)
4294 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4296 EXPORT_SYMBOL(wait_for_completion);
4298 unsigned long __sched
4299 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4301 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4303 EXPORT_SYMBOL(wait_for_completion_timeout);
4305 int __sched wait_for_completion_interruptible(struct completion *x)
4307 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4308 if (t == -ERESTARTSYS)
4312 EXPORT_SYMBOL(wait_for_completion_interruptible);
4314 unsigned long __sched
4315 wait_for_completion_interruptible_timeout(struct completion *x,
4316 unsigned long timeout)
4318 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4320 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4322 int __sched wait_for_completion_killable(struct completion *x)
4324 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4325 if (t == -ERESTARTSYS)
4329 EXPORT_SYMBOL(wait_for_completion_killable);
4332 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4334 unsigned long flags;
4337 init_waitqueue_entry(&wait, current);
4339 __set_current_state(state);
4341 spin_lock_irqsave(&q->lock, flags);
4342 __add_wait_queue(q, &wait);
4343 spin_unlock(&q->lock);
4344 timeout = schedule_timeout(timeout);
4345 spin_lock_irq(&q->lock);
4346 __remove_wait_queue(q, &wait);
4347 spin_unlock_irqrestore(&q->lock, flags);
4352 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4354 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4356 EXPORT_SYMBOL(interruptible_sleep_on);
4359 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4361 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4363 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4365 void __sched sleep_on(wait_queue_head_t *q)
4367 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4369 EXPORT_SYMBOL(sleep_on);
4371 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4373 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4375 EXPORT_SYMBOL(sleep_on_timeout);
4377 #ifdef CONFIG_RT_MUTEXES
4380 * rt_mutex_setprio - set the current priority of a task
4382 * @prio: prio value (kernel-internal form)
4384 * This function changes the 'effective' priority of a task. It does
4385 * not touch ->normal_prio like __setscheduler().
4387 * Used by the rt_mutex code to implement priority inheritance logic.
4389 void rt_mutex_setprio(struct task_struct *p, int prio)
4391 unsigned long flags;
4392 int oldprio, on_rq, running;
4394 const struct sched_class *prev_class = p->sched_class;
4396 BUG_ON(prio < 0 || prio > MAX_PRIO);
4398 rq = task_rq_lock(p, &flags);
4399 update_rq_clock(rq);
4402 on_rq = p->se.on_rq;
4403 running = task_current(rq, p);
4405 dequeue_task(rq, p, 0);
4407 p->sched_class->put_prev_task(rq, p);
4410 p->sched_class = &rt_sched_class;
4412 p->sched_class = &fair_sched_class;
4417 p->sched_class->set_curr_task(rq);
4419 enqueue_task(rq, p, 0);
4421 check_class_changed(rq, p, prev_class, oldprio, running);
4423 task_rq_unlock(rq, &flags);
4428 void set_user_nice(struct task_struct *p, long nice)
4430 int old_prio, delta, on_rq;
4431 unsigned long flags;
4434 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4437 * We have to be careful, if called from sys_setpriority(),
4438 * the task might be in the middle of scheduling on another CPU.
4440 rq = task_rq_lock(p, &flags);
4441 update_rq_clock(rq);
4443 * The RT priorities are set via sched_setscheduler(), but we still
4444 * allow the 'normal' nice value to be set - but as expected
4445 * it wont have any effect on scheduling until the task is
4446 * SCHED_FIFO/SCHED_RR:
4448 if (task_has_rt_policy(p)) {
4449 p->static_prio = NICE_TO_PRIO(nice);
4452 on_rq = p->se.on_rq;
4454 dequeue_task(rq, p, 0);
4458 p->static_prio = NICE_TO_PRIO(nice);
4461 p->prio = effective_prio(p);
4462 delta = p->prio - old_prio;
4465 enqueue_task(rq, p, 0);
4468 * If the task increased its priority or is running and
4469 * lowered its priority, then reschedule its CPU:
4471 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4472 resched_task(rq->curr);
4475 task_rq_unlock(rq, &flags);
4477 EXPORT_SYMBOL(set_user_nice);
4480 * can_nice - check if a task can reduce its nice value
4484 int can_nice(const struct task_struct *p, const int nice)
4486 /* convert nice value [19,-20] to rlimit style value [1,40] */
4487 int nice_rlim = 20 - nice;
4489 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4490 capable(CAP_SYS_NICE));
4493 #ifdef __ARCH_WANT_SYS_NICE
4496 * sys_nice - change the priority of the current process.
4497 * @increment: priority increment
4499 * sys_setpriority is a more generic, but much slower function that
4500 * does similar things.
4502 asmlinkage long sys_nice(int increment)
4507 * Setpriority might change our priority at the same moment.
4508 * We don't have to worry. Conceptually one call occurs first
4509 * and we have a single winner.
4511 if (increment < -40)
4516 nice = PRIO_TO_NICE(current->static_prio) + increment;
4522 if (increment < 0 && !can_nice(current, nice))
4525 retval = security_task_setnice(current, nice);
4529 set_user_nice(current, nice);
4536 * task_prio - return the priority value of a given task.
4537 * @p: the task in question.
4539 * This is the priority value as seen by users in /proc.
4540 * RT tasks are offset by -200. Normal tasks are centered
4541 * around 0, value goes from -16 to +15.
4543 int task_prio(const struct task_struct *p)
4545 return p->prio - MAX_RT_PRIO;
4549 * task_nice - return the nice value of a given task.
4550 * @p: the task in question.
4552 int task_nice(const struct task_struct *p)
4554 return TASK_NICE(p);
4556 EXPORT_SYMBOL(task_nice);
4559 * idle_cpu - is a given cpu idle currently?
4560 * @cpu: the processor in question.
4562 int idle_cpu(int cpu)
4564 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4568 * idle_task - return the idle task for a given cpu.
4569 * @cpu: the processor in question.
4571 struct task_struct *idle_task(int cpu)
4573 return cpu_rq(cpu)->idle;
4577 * find_process_by_pid - find a process with a matching PID value.
4578 * @pid: the pid in question.
4580 static struct task_struct *find_process_by_pid(pid_t pid)
4582 return pid ? find_task_by_vpid(pid) : current;
4585 /* Actually do priority change: must hold rq lock. */
4587 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4589 BUG_ON(p->se.on_rq);
4592 switch (p->policy) {
4596 p->sched_class = &fair_sched_class;
4600 p->sched_class = &rt_sched_class;
4604 p->rt_priority = prio;
4605 p->normal_prio = normal_prio(p);
4606 /* we are holding p->pi_lock already */
4607 p->prio = rt_mutex_getprio(p);
4612 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4613 * @p: the task in question.
4614 * @policy: new policy.
4615 * @param: structure containing the new RT priority.
4617 * NOTE that the task may be already dead.
4619 int sched_setscheduler(struct task_struct *p, int policy,
4620 struct sched_param *param)
4622 int retval, oldprio, oldpolicy = -1, on_rq, running;
4623 unsigned long flags;
4624 const struct sched_class *prev_class = p->sched_class;
4627 /* may grab non-irq protected spin_locks */
4628 BUG_ON(in_interrupt());
4630 /* double check policy once rq lock held */
4632 policy = oldpolicy = p->policy;
4633 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4634 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4635 policy != SCHED_IDLE)
4638 * Valid priorities for SCHED_FIFO and SCHED_RR are
4639 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4640 * SCHED_BATCH and SCHED_IDLE is 0.
4642 if (param->sched_priority < 0 ||
4643 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4644 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4646 if (rt_policy(policy) != (param->sched_priority != 0))
4650 * Allow unprivileged RT tasks to decrease priority:
4652 if (!capable(CAP_SYS_NICE)) {
4653 if (rt_policy(policy)) {
4654 unsigned long rlim_rtprio;
4656 if (!lock_task_sighand(p, &flags))
4658 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4659 unlock_task_sighand(p, &flags);
4661 /* can't set/change the rt policy */
4662 if (policy != p->policy && !rlim_rtprio)
4665 /* can't increase priority */
4666 if (param->sched_priority > p->rt_priority &&
4667 param->sched_priority > rlim_rtprio)
4671 * Like positive nice levels, dont allow tasks to
4672 * move out of SCHED_IDLE either:
4674 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4677 /* can't change other user's priorities */
4678 if ((current->euid != p->euid) &&
4679 (current->euid != p->uid))
4683 #ifdef CONFIG_RT_GROUP_SCHED
4685 * Do not allow realtime tasks into groups that have no runtime
4688 if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4692 retval = security_task_setscheduler(p, policy, param);
4696 * make sure no PI-waiters arrive (or leave) while we are
4697 * changing the priority of the task:
4699 spin_lock_irqsave(&p->pi_lock, flags);
4701 * To be able to change p->policy safely, the apropriate
4702 * runqueue lock must be held.
4704 rq = __task_rq_lock(p);
4705 /* recheck policy now with rq lock held */
4706 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4707 policy = oldpolicy = -1;
4708 __task_rq_unlock(rq);
4709 spin_unlock_irqrestore(&p->pi_lock, flags);
4712 update_rq_clock(rq);
4713 on_rq = p->se.on_rq;
4714 running = task_current(rq, p);
4716 deactivate_task(rq, p, 0);
4718 p->sched_class->put_prev_task(rq, p);
4721 __setscheduler(rq, p, policy, param->sched_priority);
4724 p->sched_class->set_curr_task(rq);
4726 activate_task(rq, p, 0);
4728 check_class_changed(rq, p, prev_class, oldprio, running);
4730 __task_rq_unlock(rq);
4731 spin_unlock_irqrestore(&p->pi_lock, flags);
4733 rt_mutex_adjust_pi(p);
4737 EXPORT_SYMBOL_GPL(sched_setscheduler);
4740 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4742 struct sched_param lparam;
4743 struct task_struct *p;
4746 if (!param || pid < 0)
4748 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4753 p = find_process_by_pid(pid);
4755 retval = sched_setscheduler(p, policy, &lparam);
4762 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4763 * @pid: the pid in question.
4764 * @policy: new policy.
4765 * @param: structure containing the new RT priority.
4768 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4770 /* negative values for policy are not valid */
4774 return do_sched_setscheduler(pid, policy, param);
4778 * sys_sched_setparam - set/change the RT priority of a thread
4779 * @pid: the pid in question.
4780 * @param: structure containing the new RT priority.
4782 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4784 return do_sched_setscheduler(pid, -1, param);
4788 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4789 * @pid: the pid in question.
4791 asmlinkage long sys_sched_getscheduler(pid_t pid)
4793 struct task_struct *p;
4800 read_lock(&tasklist_lock);
4801 p = find_process_by_pid(pid);
4803 retval = security_task_getscheduler(p);
4807 read_unlock(&tasklist_lock);
4812 * sys_sched_getscheduler - get the RT priority of a thread
4813 * @pid: the pid in question.
4814 * @param: structure containing the RT priority.
4816 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4818 struct sched_param lp;
4819 struct task_struct *p;
4822 if (!param || pid < 0)
4825 read_lock(&tasklist_lock);
4826 p = find_process_by_pid(pid);
4831 retval = security_task_getscheduler(p);
4835 lp.sched_priority = p->rt_priority;
4836 read_unlock(&tasklist_lock);
4839 * This one might sleep, we cannot do it with a spinlock held ...
4841 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4846 read_unlock(&tasklist_lock);
4850 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4852 cpumask_t cpus_allowed;
4853 struct task_struct *p;
4857 read_lock(&tasklist_lock);
4859 p = find_process_by_pid(pid);
4861 read_unlock(&tasklist_lock);
4867 * It is not safe to call set_cpus_allowed with the
4868 * tasklist_lock held. We will bump the task_struct's
4869 * usage count and then drop tasklist_lock.
4872 read_unlock(&tasklist_lock);
4875 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4876 !capable(CAP_SYS_NICE))
4879 retval = security_task_setscheduler(p, 0, NULL);
4883 cpus_allowed = cpuset_cpus_allowed(p);
4884 cpus_and(new_mask, new_mask, cpus_allowed);
4886 retval = set_cpus_allowed(p, new_mask);
4889 cpus_allowed = cpuset_cpus_allowed(p);
4890 if (!cpus_subset(new_mask, cpus_allowed)) {
4892 * We must have raced with a concurrent cpuset
4893 * update. Just reset the cpus_allowed to the
4894 * cpuset's cpus_allowed
4896 new_mask = cpus_allowed;
4906 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4907 cpumask_t *new_mask)
4909 if (len < sizeof(cpumask_t)) {
4910 memset(new_mask, 0, sizeof(cpumask_t));
4911 } else if (len > sizeof(cpumask_t)) {
4912 len = sizeof(cpumask_t);
4914 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4918 * sys_sched_setaffinity - set the cpu affinity of a process
4919 * @pid: pid of the process
4920 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4921 * @user_mask_ptr: user-space pointer to the new cpu mask
4923 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4924 unsigned long __user *user_mask_ptr)
4929 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4933 return sched_setaffinity(pid, new_mask);
4937 * Represents all cpu's present in the system
4938 * In systems capable of hotplug, this map could dynamically grow
4939 * as new cpu's are detected in the system via any platform specific
4940 * method, such as ACPI for e.g.
4943 cpumask_t cpu_present_map __read_mostly;
4944 EXPORT_SYMBOL(cpu_present_map);
4947 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4948 EXPORT_SYMBOL(cpu_online_map);
4950 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4951 EXPORT_SYMBOL(cpu_possible_map);
4954 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4956 struct task_struct *p;
4960 read_lock(&tasklist_lock);
4963 p = find_process_by_pid(pid);
4967 retval = security_task_getscheduler(p);
4971 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4974 read_unlock(&tasklist_lock);
4981 * sys_sched_getaffinity - get the cpu affinity of a process
4982 * @pid: pid of the process
4983 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4984 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4986 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4987 unsigned long __user *user_mask_ptr)
4992 if (len < sizeof(cpumask_t))
4995 ret = sched_getaffinity(pid, &mask);
4999 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5002 return sizeof(cpumask_t);
5006 * sys_sched_yield - yield the current processor to other threads.
5008 * This function yields the current CPU to other tasks. If there are no
5009 * other threads running on this CPU then this function will return.
5011 asmlinkage long sys_sched_yield(void)
5013 struct rq *rq = this_rq_lock();
5015 schedstat_inc(rq, yld_count);
5016 current->sched_class->yield_task(rq);
5019 * Since we are going to call schedule() anyway, there's
5020 * no need to preempt or enable interrupts:
5022 __release(rq->lock);
5023 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5024 _raw_spin_unlock(&rq->lock);
5025 preempt_enable_no_resched();
5032 static void __cond_resched(void)
5034 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5035 __might_sleep(__FILE__, __LINE__);
5038 * The BKS might be reacquired before we have dropped
5039 * PREEMPT_ACTIVE, which could trigger a second
5040 * cond_resched() call.
5043 add_preempt_count(PREEMPT_ACTIVE);
5045 sub_preempt_count(PREEMPT_ACTIVE);
5046 } while (need_resched());
5049 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5050 int __sched _cond_resched(void)
5052 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5053 system_state == SYSTEM_RUNNING) {
5059 EXPORT_SYMBOL(_cond_resched);
5063 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5064 * call schedule, and on return reacquire the lock.
5066 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5067 * operations here to prevent schedule() from being called twice (once via
5068 * spin_unlock(), once by hand).
5070 int cond_resched_lock(spinlock_t *lock)
5072 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5075 if (spin_needbreak(lock) || resched) {
5077 if (resched && need_resched())
5086 EXPORT_SYMBOL(cond_resched_lock);
5088 int __sched cond_resched_softirq(void)
5090 BUG_ON(!in_softirq());
5092 if (need_resched() && system_state == SYSTEM_RUNNING) {
5100 EXPORT_SYMBOL(cond_resched_softirq);
5103 * yield - yield the current processor to other threads.
5105 * This is a shortcut for kernel-space yielding - it marks the
5106 * thread runnable and calls sys_sched_yield().
5108 void __sched yield(void)
5110 set_current_state(TASK_RUNNING);
5113 EXPORT_SYMBOL(yield);
5116 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5117 * that process accounting knows that this is a task in IO wait state.
5119 * But don't do that if it is a deliberate, throttling IO wait (this task
5120 * has set its backing_dev_info: the queue against which it should throttle)
5122 void __sched io_schedule(void)
5124 struct rq *rq = &__raw_get_cpu_var(runqueues);
5126 delayacct_blkio_start();
5127 atomic_inc(&rq->nr_iowait);
5129 atomic_dec(&rq->nr_iowait);
5130 delayacct_blkio_end();
5132 EXPORT_SYMBOL(io_schedule);
5134 long __sched io_schedule_timeout(long timeout)
5136 struct rq *rq = &__raw_get_cpu_var(runqueues);
5139 delayacct_blkio_start();
5140 atomic_inc(&rq->nr_iowait);
5141 ret = schedule_timeout(timeout);
5142 atomic_dec(&rq->nr_iowait);
5143 delayacct_blkio_end();
5148 * sys_sched_get_priority_max - return maximum RT priority.
5149 * @policy: scheduling class.
5151 * this syscall returns the maximum rt_priority that can be used
5152 * by a given scheduling class.
5154 asmlinkage long sys_sched_get_priority_max(int policy)
5161 ret = MAX_USER_RT_PRIO-1;
5173 * sys_sched_get_priority_min - return minimum RT priority.
5174 * @policy: scheduling class.
5176 * this syscall returns the minimum rt_priority that can be used
5177 * by a given scheduling class.
5179 asmlinkage long sys_sched_get_priority_min(int policy)
5197 * sys_sched_rr_get_interval - return the default timeslice of a process.
5198 * @pid: pid of the process.
5199 * @interval: userspace pointer to the timeslice value.
5201 * this syscall writes the default timeslice value of a given process
5202 * into the user-space timespec buffer. A value of '0' means infinity.
5205 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5207 struct task_struct *p;
5208 unsigned int time_slice;
5216 read_lock(&tasklist_lock);
5217 p = find_process_by_pid(pid);
5221 retval = security_task_getscheduler(p);
5226 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5227 * tasks that are on an otherwise idle runqueue:
5230 if (p->policy == SCHED_RR) {
5231 time_slice = DEF_TIMESLICE;
5232 } else if (p->policy != SCHED_FIFO) {
5233 struct sched_entity *se = &p->se;
5234 unsigned long flags;
5237 rq = task_rq_lock(p, &flags);
5238 if (rq->cfs.load.weight)
5239 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5240 task_rq_unlock(rq, &flags);
5242 read_unlock(&tasklist_lock);
5243 jiffies_to_timespec(time_slice, &t);
5244 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5248 read_unlock(&tasklist_lock);
5252 static const char stat_nam[] = "RSDTtZX";
5254 void sched_show_task(struct task_struct *p)
5256 unsigned long free = 0;
5259 state = p->state ? __ffs(p->state) + 1 : 0;
5260 printk(KERN_INFO "%-13.13s %c", p->comm,
5261 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5262 #if BITS_PER_LONG == 32
5263 if (state == TASK_RUNNING)
5264 printk(KERN_CONT " running ");
5266 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5268 if (state == TASK_RUNNING)
5269 printk(KERN_CONT " running task ");
5271 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5273 #ifdef CONFIG_DEBUG_STACK_USAGE
5275 unsigned long *n = end_of_stack(p);
5278 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5281 printk(KERN_CONT "%5lu %5d %6d\n", free,
5282 task_pid_nr(p), task_pid_nr(p->real_parent));
5284 show_stack(p, NULL);
5287 void show_state_filter(unsigned long state_filter)
5289 struct task_struct *g, *p;
5291 #if BITS_PER_LONG == 32
5293 " task PC stack pid father\n");
5296 " task PC stack pid father\n");
5298 read_lock(&tasklist_lock);
5299 do_each_thread(g, p) {
5301 * reset the NMI-timeout, listing all files on a slow
5302 * console might take alot of time:
5304 touch_nmi_watchdog();
5305 if (!state_filter || (p->state & state_filter))
5307 } while_each_thread(g, p);
5309 touch_all_softlockup_watchdogs();
5311 #ifdef CONFIG_SCHED_DEBUG
5312 sysrq_sched_debug_show();
5314 read_unlock(&tasklist_lock);
5316 * Only show locks if all tasks are dumped:
5318 if (state_filter == -1)
5319 debug_show_all_locks();
5322 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5324 idle->sched_class = &idle_sched_class;
5328 * init_idle - set up an idle thread for a given CPU
5329 * @idle: task in question
5330 * @cpu: cpu the idle task belongs to
5332 * NOTE: this function does not set the idle thread's NEED_RESCHED
5333 * flag, to make booting more robust.
5335 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5337 struct rq *rq = cpu_rq(cpu);
5338 unsigned long flags;
5341 idle->se.exec_start = sched_clock();
5343 idle->prio = idle->normal_prio = MAX_PRIO;
5344 idle->cpus_allowed = cpumask_of_cpu(cpu);
5345 __set_task_cpu(idle, cpu);
5347 spin_lock_irqsave(&rq->lock, flags);
5348 rq->curr = rq->idle = idle;
5349 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5352 spin_unlock_irqrestore(&rq->lock, flags);
5354 /* Set the preempt count _outside_ the spinlocks! */
5355 task_thread_info(idle)->preempt_count = 0;
5358 * The idle tasks have their own, simple scheduling class:
5360 idle->sched_class = &idle_sched_class;
5364 * In a system that switches off the HZ timer nohz_cpu_mask
5365 * indicates which cpus entered this state. This is used
5366 * in the rcu update to wait only for active cpus. For system
5367 * which do not switch off the HZ timer nohz_cpu_mask should
5368 * always be CPU_MASK_NONE.
5370 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5373 * Increase the granularity value when there are more CPUs,
5374 * because with more CPUs the 'effective latency' as visible
5375 * to users decreases. But the relationship is not linear,
5376 * so pick a second-best guess by going with the log2 of the
5379 * This idea comes from the SD scheduler of Con Kolivas:
5381 static inline void sched_init_granularity(void)
5383 unsigned int factor = 1 + ilog2(num_online_cpus());
5384 const unsigned long limit = 200000000;
5386 sysctl_sched_min_granularity *= factor;
5387 if (sysctl_sched_min_granularity > limit)
5388 sysctl_sched_min_granularity = limit;
5390 sysctl_sched_latency *= factor;
5391 if (sysctl_sched_latency > limit)
5392 sysctl_sched_latency = limit;
5394 sysctl_sched_wakeup_granularity *= factor;
5395 sysctl_sched_batch_wakeup_granularity *= factor;
5400 * This is how migration works:
5402 * 1) we queue a struct migration_req structure in the source CPU's
5403 * runqueue and wake up that CPU's migration thread.
5404 * 2) we down() the locked semaphore => thread blocks.
5405 * 3) migration thread wakes up (implicitly it forces the migrated
5406 * thread off the CPU)
5407 * 4) it gets the migration request and checks whether the migrated
5408 * task is still in the wrong runqueue.
5409 * 5) if it's in the wrong runqueue then the migration thread removes
5410 * it and puts it into the right queue.
5411 * 6) migration thread up()s the semaphore.
5412 * 7) we wake up and the migration is done.
5416 * Change a given task's CPU affinity. Migrate the thread to a
5417 * proper CPU and schedule it away if the CPU it's executing on
5418 * is removed from the allowed bitmask.
5420 * NOTE: the caller must have a valid reference to the task, the
5421 * task must not exit() & deallocate itself prematurely. The
5422 * call is not atomic; no spinlocks may be held.
5424 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5426 struct migration_req req;
5427 unsigned long flags;
5431 rq = task_rq_lock(p, &flags);
5432 if (!cpus_intersects(new_mask, cpu_online_map)) {
5437 if (p->sched_class->set_cpus_allowed)
5438 p->sched_class->set_cpus_allowed(p, &new_mask);
5440 p->cpus_allowed = new_mask;
5441 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5444 /* Can the task run on the task's current CPU? If so, we're done */
5445 if (cpu_isset(task_cpu(p), new_mask))
5448 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5449 /* Need help from migration thread: drop lock and wait. */
5450 task_rq_unlock(rq, &flags);
5451 wake_up_process(rq->migration_thread);
5452 wait_for_completion(&req.done);
5453 tlb_migrate_finish(p->mm);
5457 task_rq_unlock(rq, &flags);
5461 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5464 * Move (not current) task off this cpu, onto dest cpu. We're doing
5465 * this because either it can't run here any more (set_cpus_allowed()
5466 * away from this CPU, or CPU going down), or because we're
5467 * attempting to rebalance this task on exec (sched_exec).
5469 * So we race with normal scheduler movements, but that's OK, as long
5470 * as the task is no longer on this CPU.
5472 * Returns non-zero if task was successfully migrated.
5474 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5476 struct rq *rq_dest, *rq_src;
5479 if (unlikely(cpu_is_offline(dest_cpu)))
5482 rq_src = cpu_rq(src_cpu);
5483 rq_dest = cpu_rq(dest_cpu);
5485 double_rq_lock(rq_src, rq_dest);
5486 /* Already moved. */
5487 if (task_cpu(p) != src_cpu)
5489 /* Affinity changed (again). */
5490 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5493 on_rq = p->se.on_rq;
5495 deactivate_task(rq_src, p, 0);
5497 set_task_cpu(p, dest_cpu);
5499 activate_task(rq_dest, p, 0);
5500 check_preempt_curr(rq_dest, p);
5504 double_rq_unlock(rq_src, rq_dest);
5509 * migration_thread - this is a highprio system thread that performs
5510 * thread migration by bumping thread off CPU then 'pushing' onto
5513 static int migration_thread(void *data)
5515 int cpu = (long)data;
5519 BUG_ON(rq->migration_thread != current);
5521 set_current_state(TASK_INTERRUPTIBLE);
5522 while (!kthread_should_stop()) {
5523 struct migration_req *req;
5524 struct list_head *head;
5526 spin_lock_irq(&rq->lock);
5528 if (cpu_is_offline(cpu)) {
5529 spin_unlock_irq(&rq->lock);
5533 if (rq->active_balance) {
5534 active_load_balance(rq, cpu);
5535 rq->active_balance = 0;
5538 head = &rq->migration_queue;
5540 if (list_empty(head)) {
5541 spin_unlock_irq(&rq->lock);
5543 set_current_state(TASK_INTERRUPTIBLE);
5546 req = list_entry(head->next, struct migration_req, list);
5547 list_del_init(head->next);
5549 spin_unlock(&rq->lock);
5550 __migrate_task(req->task, cpu, req->dest_cpu);
5553 complete(&req->done);
5555 __set_current_state(TASK_RUNNING);
5559 /* Wait for kthread_stop */
5560 set_current_state(TASK_INTERRUPTIBLE);
5561 while (!kthread_should_stop()) {
5563 set_current_state(TASK_INTERRUPTIBLE);
5565 __set_current_state(TASK_RUNNING);
5569 #ifdef CONFIG_HOTPLUG_CPU
5571 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5575 local_irq_disable();
5576 ret = __migrate_task(p, src_cpu, dest_cpu);
5582 * Figure out where task on dead CPU should go, use force if necessary.
5583 * NOTE: interrupts should be disabled by the caller
5585 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5587 unsigned long flags;
5594 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5595 cpus_and(mask, mask, p->cpus_allowed);
5596 dest_cpu = any_online_cpu(mask);
5598 /* On any allowed CPU? */
5599 if (dest_cpu == NR_CPUS)
5600 dest_cpu = any_online_cpu(p->cpus_allowed);
5602 /* No more Mr. Nice Guy. */
5603 if (dest_cpu == NR_CPUS) {
5604 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5606 * Try to stay on the same cpuset, where the
5607 * current cpuset may be a subset of all cpus.
5608 * The cpuset_cpus_allowed_locked() variant of
5609 * cpuset_cpus_allowed() will not block. It must be
5610 * called within calls to cpuset_lock/cpuset_unlock.
5612 rq = task_rq_lock(p, &flags);
5613 p->cpus_allowed = cpus_allowed;
5614 dest_cpu = any_online_cpu(p->cpus_allowed);
5615 task_rq_unlock(rq, &flags);
5618 * Don't tell them about moving exiting tasks or
5619 * kernel threads (both mm NULL), since they never
5622 if (p->mm && printk_ratelimit()) {
5623 printk(KERN_INFO "process %d (%s) no "
5624 "longer affine to cpu%d\n",
5625 task_pid_nr(p), p->comm, dead_cpu);
5628 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5632 * While a dead CPU has no uninterruptible tasks queued at this point,
5633 * it might still have a nonzero ->nr_uninterruptible counter, because
5634 * for performance reasons the counter is not stricly tracking tasks to
5635 * their home CPUs. So we just add the counter to another CPU's counter,
5636 * to keep the global sum constant after CPU-down:
5638 static void migrate_nr_uninterruptible(struct rq *rq_src)
5640 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5641 unsigned long flags;
5643 local_irq_save(flags);
5644 double_rq_lock(rq_src, rq_dest);
5645 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5646 rq_src->nr_uninterruptible = 0;
5647 double_rq_unlock(rq_src, rq_dest);
5648 local_irq_restore(flags);
5651 /* Run through task list and migrate tasks from the dead cpu. */
5652 static void migrate_live_tasks(int src_cpu)
5654 struct task_struct *p, *t;
5656 read_lock(&tasklist_lock);
5658 do_each_thread(t, p) {
5662 if (task_cpu(p) == src_cpu)
5663 move_task_off_dead_cpu(src_cpu, p);
5664 } while_each_thread(t, p);
5666 read_unlock(&tasklist_lock);
5670 * Schedules idle task to be the next runnable task on current CPU.
5671 * It does so by boosting its priority to highest possible.
5672 * Used by CPU offline code.
5674 void sched_idle_next(void)
5676 int this_cpu = smp_processor_id();
5677 struct rq *rq = cpu_rq(this_cpu);
5678 struct task_struct *p = rq->idle;
5679 unsigned long flags;
5681 /* cpu has to be offline */
5682 BUG_ON(cpu_online(this_cpu));
5685 * Strictly not necessary since rest of the CPUs are stopped by now
5686 * and interrupts disabled on the current cpu.
5688 spin_lock_irqsave(&rq->lock, flags);
5690 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5692 update_rq_clock(rq);
5693 activate_task(rq, p, 0);
5695 spin_unlock_irqrestore(&rq->lock, flags);
5699 * Ensures that the idle task is using init_mm right before its cpu goes
5702 void idle_task_exit(void)
5704 struct mm_struct *mm = current->active_mm;
5706 BUG_ON(cpu_online(smp_processor_id()));
5709 switch_mm(mm, &init_mm, current);
5713 /* called under rq->lock with disabled interrupts */
5714 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5716 struct rq *rq = cpu_rq(dead_cpu);
5718 /* Must be exiting, otherwise would be on tasklist. */
5719 BUG_ON(!p->exit_state);
5721 /* Cannot have done final schedule yet: would have vanished. */
5722 BUG_ON(p->state == TASK_DEAD);
5727 * Drop lock around migration; if someone else moves it,
5728 * that's OK. No task can be added to this CPU, so iteration is
5731 spin_unlock_irq(&rq->lock);
5732 move_task_off_dead_cpu(dead_cpu, p);
5733 spin_lock_irq(&rq->lock);
5738 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5739 static void migrate_dead_tasks(unsigned int dead_cpu)
5741 struct rq *rq = cpu_rq(dead_cpu);
5742 struct task_struct *next;
5745 if (!rq->nr_running)
5747 update_rq_clock(rq);
5748 next = pick_next_task(rq, rq->curr);
5751 migrate_dead(dead_cpu, next);
5755 #endif /* CONFIG_HOTPLUG_CPU */
5757 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5759 static struct ctl_table sd_ctl_dir[] = {
5761 .procname = "sched_domain",
5767 static struct ctl_table sd_ctl_root[] = {
5769 .ctl_name = CTL_KERN,
5770 .procname = "kernel",
5772 .child = sd_ctl_dir,
5777 static struct ctl_table *sd_alloc_ctl_entry(int n)
5779 struct ctl_table *entry =
5780 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5785 static void sd_free_ctl_entry(struct ctl_table **tablep)
5787 struct ctl_table *entry;
5790 * In the intermediate directories, both the child directory and
5791 * procname are dynamically allocated and could fail but the mode
5792 * will always be set. In the lowest directory the names are
5793 * static strings and all have proc handlers.
5795 for (entry = *tablep; entry->mode; entry++) {
5797 sd_free_ctl_entry(&entry->child);
5798 if (entry->proc_handler == NULL)
5799 kfree(entry->procname);
5807 set_table_entry(struct ctl_table *entry,
5808 const char *procname, void *data, int maxlen,
5809 mode_t mode, proc_handler *proc_handler)
5811 entry->procname = procname;
5813 entry->maxlen = maxlen;
5815 entry->proc_handler = proc_handler;
5818 static struct ctl_table *
5819 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5821 struct ctl_table *table = sd_alloc_ctl_entry(12);
5826 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5827 sizeof(long), 0644, proc_doulongvec_minmax);
5828 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5829 sizeof(long), 0644, proc_doulongvec_minmax);
5830 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5831 sizeof(int), 0644, proc_dointvec_minmax);
5832 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5833 sizeof(int), 0644, proc_dointvec_minmax);
5834 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5835 sizeof(int), 0644, proc_dointvec_minmax);
5836 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5837 sizeof(int), 0644, proc_dointvec_minmax);
5838 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5839 sizeof(int), 0644, proc_dointvec_minmax);
5840 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5841 sizeof(int), 0644, proc_dointvec_minmax);
5842 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5843 sizeof(int), 0644, proc_dointvec_minmax);
5844 set_table_entry(&table[9], "cache_nice_tries",
5845 &sd->cache_nice_tries,
5846 sizeof(int), 0644, proc_dointvec_minmax);
5847 set_table_entry(&table[10], "flags", &sd->flags,
5848 sizeof(int), 0644, proc_dointvec_minmax);
5849 /* &table[11] is terminator */
5854 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5856 struct ctl_table *entry, *table;
5857 struct sched_domain *sd;
5858 int domain_num = 0, i;
5861 for_each_domain(cpu, sd)
5863 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5868 for_each_domain(cpu, sd) {
5869 snprintf(buf, 32, "domain%d", i);
5870 entry->procname = kstrdup(buf, GFP_KERNEL);
5872 entry->child = sd_alloc_ctl_domain_table(sd);
5879 static struct ctl_table_header *sd_sysctl_header;
5880 static void register_sched_domain_sysctl(void)
5882 int i, cpu_num = num_online_cpus();
5883 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5886 WARN_ON(sd_ctl_dir[0].child);
5887 sd_ctl_dir[0].child = entry;
5892 for_each_online_cpu(i) {
5893 snprintf(buf, 32, "cpu%d", i);
5894 entry->procname = kstrdup(buf, GFP_KERNEL);
5896 entry->child = sd_alloc_ctl_cpu_table(i);
5900 WARN_ON(sd_sysctl_header);
5901 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5904 /* may be called multiple times per register */
5905 static void unregister_sched_domain_sysctl(void)
5907 if (sd_sysctl_header)
5908 unregister_sysctl_table(sd_sysctl_header);
5909 sd_sysctl_header = NULL;
5910 if (sd_ctl_dir[0].child)
5911 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5914 static void register_sched_domain_sysctl(void)
5917 static void unregister_sched_domain_sysctl(void)
5923 * migration_call - callback that gets triggered when a CPU is added.
5924 * Here we can start up the necessary migration thread for the new CPU.
5926 static int __cpuinit
5927 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5929 struct task_struct *p;
5930 int cpu = (long)hcpu;
5931 unsigned long flags;
5936 case CPU_UP_PREPARE:
5937 case CPU_UP_PREPARE_FROZEN:
5938 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5941 kthread_bind(p, cpu);
5942 /* Must be high prio: stop_machine expects to yield to it. */
5943 rq = task_rq_lock(p, &flags);
5944 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5945 task_rq_unlock(rq, &flags);
5946 cpu_rq(cpu)->migration_thread = p;
5950 case CPU_ONLINE_FROZEN:
5951 /* Strictly unnecessary, as first user will wake it. */
5952 wake_up_process(cpu_rq(cpu)->migration_thread);
5954 /* Update our root-domain */
5956 spin_lock_irqsave(&rq->lock, flags);
5958 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5959 cpu_set(cpu, rq->rd->online);
5961 spin_unlock_irqrestore(&rq->lock, flags);
5964 #ifdef CONFIG_HOTPLUG_CPU
5965 case CPU_UP_CANCELED:
5966 case CPU_UP_CANCELED_FROZEN:
5967 if (!cpu_rq(cpu)->migration_thread)
5969 /* Unbind it from offline cpu so it can run. Fall thru. */
5970 kthread_bind(cpu_rq(cpu)->migration_thread,
5971 any_online_cpu(cpu_online_map));
5972 kthread_stop(cpu_rq(cpu)->migration_thread);
5973 cpu_rq(cpu)->migration_thread = NULL;
5977 case CPU_DEAD_FROZEN:
5978 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5979 migrate_live_tasks(cpu);
5981 kthread_stop(rq->migration_thread);
5982 rq->migration_thread = NULL;
5983 /* Idle task back to normal (off runqueue, low prio) */
5984 spin_lock_irq(&rq->lock);
5985 update_rq_clock(rq);
5986 deactivate_task(rq, rq->idle, 0);
5987 rq->idle->static_prio = MAX_PRIO;
5988 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5989 rq->idle->sched_class = &idle_sched_class;
5990 migrate_dead_tasks(cpu);
5991 spin_unlock_irq(&rq->lock);
5993 migrate_nr_uninterruptible(rq);
5994 BUG_ON(rq->nr_running != 0);
5997 * No need to migrate the tasks: it was best-effort if
5998 * they didn't take sched_hotcpu_mutex. Just wake up
6001 spin_lock_irq(&rq->lock);
6002 while (!list_empty(&rq->migration_queue)) {
6003 struct migration_req *req;
6005 req = list_entry(rq->migration_queue.next,
6006 struct migration_req, list);
6007 list_del_init(&req->list);
6008 complete(&req->done);
6010 spin_unlock_irq(&rq->lock);
6014 case CPU_DYING_FROZEN:
6015 /* Update our root-domain */
6017 spin_lock_irqsave(&rq->lock, flags);
6019 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6020 cpu_clear(cpu, rq->rd->online);
6022 spin_unlock_irqrestore(&rq->lock, flags);
6029 /* Register at highest priority so that task migration (migrate_all_tasks)
6030 * happens before everything else.
6032 static struct notifier_block __cpuinitdata migration_notifier = {
6033 .notifier_call = migration_call,
6037 void __init migration_init(void)
6039 void *cpu = (void *)(long)smp_processor_id();
6042 /* Start one for the boot CPU: */
6043 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6044 BUG_ON(err == NOTIFY_BAD);
6045 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6046 register_cpu_notifier(&migration_notifier);
6052 /* Number of possible processor ids */
6053 int nr_cpu_ids __read_mostly = NR_CPUS;
6054 EXPORT_SYMBOL(nr_cpu_ids);
6056 #ifdef CONFIG_SCHED_DEBUG
6058 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
6060 struct sched_group *group = sd->groups;
6061 cpumask_t groupmask;
6064 cpumask_scnprintf(str, NR_CPUS, sd->span);
6065 cpus_clear(groupmask);
6067 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6069 if (!(sd->flags & SD_LOAD_BALANCE)) {
6070 printk("does not load-balance\n");
6072 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6077 printk(KERN_CONT "span %s\n", str);
6079 if (!cpu_isset(cpu, sd->span)) {
6080 printk(KERN_ERR "ERROR: domain->span does not contain "
6083 if (!cpu_isset(cpu, group->cpumask)) {
6084 printk(KERN_ERR "ERROR: domain->groups does not contain"
6088 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6092 printk(KERN_ERR "ERROR: group is NULL\n");
6096 if (!group->__cpu_power) {
6097 printk(KERN_CONT "\n");
6098 printk(KERN_ERR "ERROR: domain->cpu_power not "
6103 if (!cpus_weight(group->cpumask)) {
6104 printk(KERN_CONT "\n");
6105 printk(KERN_ERR "ERROR: empty group\n");
6109 if (cpus_intersects(groupmask, group->cpumask)) {
6110 printk(KERN_CONT "\n");
6111 printk(KERN_ERR "ERROR: repeated CPUs\n");
6115 cpus_or(groupmask, groupmask, group->cpumask);
6117 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6118 printk(KERN_CONT " %s", str);
6120 group = group->next;
6121 } while (group != sd->groups);
6122 printk(KERN_CONT "\n");
6124 if (!cpus_equal(sd->span, groupmask))
6125 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6127 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6128 printk(KERN_ERR "ERROR: parent span is not a superset "
6129 "of domain->span\n");
6133 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6138 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6142 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6145 if (sched_domain_debug_one(sd, cpu, level))
6154 # define sched_domain_debug(sd, cpu) do { } while (0)
6157 static int sd_degenerate(struct sched_domain *sd)
6159 if (cpus_weight(sd->span) == 1)
6162 /* Following flags need at least 2 groups */
6163 if (sd->flags & (SD_LOAD_BALANCE |
6164 SD_BALANCE_NEWIDLE |
6168 SD_SHARE_PKG_RESOURCES)) {
6169 if (sd->groups != sd->groups->next)
6173 /* Following flags don't use groups */
6174 if (sd->flags & (SD_WAKE_IDLE |
6183 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6185 unsigned long cflags = sd->flags, pflags = parent->flags;
6187 if (sd_degenerate(parent))
6190 if (!cpus_equal(sd->span, parent->span))
6193 /* Does parent contain flags not in child? */
6194 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6195 if (cflags & SD_WAKE_AFFINE)
6196 pflags &= ~SD_WAKE_BALANCE;
6197 /* Flags needing groups don't count if only 1 group in parent */
6198 if (parent->groups == parent->groups->next) {
6199 pflags &= ~(SD_LOAD_BALANCE |
6200 SD_BALANCE_NEWIDLE |
6204 SD_SHARE_PKG_RESOURCES);
6206 if (~cflags & pflags)
6212 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6214 unsigned long flags;
6215 const struct sched_class *class;
6217 spin_lock_irqsave(&rq->lock, flags);
6220 struct root_domain *old_rd = rq->rd;
6222 for (class = sched_class_highest; class; class = class->next) {
6223 if (class->leave_domain)
6224 class->leave_domain(rq);
6227 cpu_clear(rq->cpu, old_rd->span);
6228 cpu_clear(rq->cpu, old_rd->online);
6230 if (atomic_dec_and_test(&old_rd->refcount))
6234 atomic_inc(&rd->refcount);
6237 cpu_set(rq->cpu, rd->span);
6238 if (cpu_isset(rq->cpu, cpu_online_map))
6239 cpu_set(rq->cpu, rd->online);
6241 for (class = sched_class_highest; class; class = class->next) {
6242 if (class->join_domain)
6243 class->join_domain(rq);
6246 spin_unlock_irqrestore(&rq->lock, flags);
6249 static void init_rootdomain(struct root_domain *rd)
6251 memset(rd, 0, sizeof(*rd));
6253 cpus_clear(rd->span);
6254 cpus_clear(rd->online);
6257 static void init_defrootdomain(void)
6259 init_rootdomain(&def_root_domain);
6260 atomic_set(&def_root_domain.refcount, 1);
6263 static struct root_domain *alloc_rootdomain(void)
6265 struct root_domain *rd;
6267 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6271 init_rootdomain(rd);
6277 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6278 * hold the hotplug lock.
6281 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6283 struct rq *rq = cpu_rq(cpu);
6284 struct sched_domain *tmp;
6286 /* Remove the sched domains which do not contribute to scheduling. */
6287 for (tmp = sd; tmp; tmp = tmp->parent) {
6288 struct sched_domain *parent = tmp->parent;
6291 if (sd_parent_degenerate(tmp, parent)) {
6292 tmp->parent = parent->parent;
6294 parent->parent->child = tmp;
6298 if (sd && sd_degenerate(sd)) {
6304 sched_domain_debug(sd, cpu);
6306 rq_attach_root(rq, rd);
6307 rcu_assign_pointer(rq->sd, sd);
6310 /* cpus with isolated domains */
6311 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6313 /* Setup the mask of cpus configured for isolated domains */
6314 static int __init isolated_cpu_setup(char *str)
6316 int ints[NR_CPUS], i;
6318 str = get_options(str, ARRAY_SIZE(ints), ints);
6319 cpus_clear(cpu_isolated_map);
6320 for (i = 1; i <= ints[0]; i++)
6321 if (ints[i] < NR_CPUS)
6322 cpu_set(ints[i], cpu_isolated_map);
6326 __setup("isolcpus=", isolated_cpu_setup);
6329 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6330 * to a function which identifies what group(along with sched group) a CPU
6331 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6332 * (due to the fact that we keep track of groups covered with a cpumask_t).
6334 * init_sched_build_groups will build a circular linked list of the groups
6335 * covered by the given span, and will set each group's ->cpumask correctly,
6336 * and ->cpu_power to 0.
6339 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6340 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6341 struct sched_group **sg))
6343 struct sched_group *first = NULL, *last = NULL;
6344 cpumask_t covered = CPU_MASK_NONE;
6347 for_each_cpu_mask(i, span) {
6348 struct sched_group *sg;
6349 int group = group_fn(i, cpu_map, &sg);
6352 if (cpu_isset(i, covered))
6355 sg->cpumask = CPU_MASK_NONE;
6356 sg->__cpu_power = 0;
6358 for_each_cpu_mask(j, span) {
6359 if (group_fn(j, cpu_map, NULL) != group)
6362 cpu_set(j, covered);
6363 cpu_set(j, sg->cpumask);
6374 #define SD_NODES_PER_DOMAIN 16
6379 * find_next_best_node - find the next node to include in a sched_domain
6380 * @node: node whose sched_domain we're building
6381 * @used_nodes: nodes already in the sched_domain
6383 * Find the next node to include in a given scheduling domain. Simply
6384 * finds the closest node not already in the @used_nodes map.
6386 * Should use nodemask_t.
6388 static int find_next_best_node(int node, unsigned long *used_nodes)
6390 int i, n, val, min_val, best_node = 0;
6394 for (i = 0; i < MAX_NUMNODES; i++) {
6395 /* Start at @node */
6396 n = (node + i) % MAX_NUMNODES;
6398 if (!nr_cpus_node(n))
6401 /* Skip already used nodes */
6402 if (test_bit(n, used_nodes))
6405 /* Simple min distance search */
6406 val = node_distance(node, n);
6408 if (val < min_val) {
6414 set_bit(best_node, used_nodes);
6419 * sched_domain_node_span - get a cpumask for a node's sched_domain
6420 * @node: node whose cpumask we're constructing
6421 * @size: number of nodes to include in this span
6423 * Given a node, construct a good cpumask for its sched_domain to span. It
6424 * should be one that prevents unnecessary balancing, but also spreads tasks
6427 static cpumask_t sched_domain_node_span(int node)
6429 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6430 cpumask_t span, nodemask;
6434 bitmap_zero(used_nodes, MAX_NUMNODES);
6436 nodemask = node_to_cpumask(node);
6437 cpus_or(span, span, nodemask);
6438 set_bit(node, used_nodes);
6440 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6441 int next_node = find_next_best_node(node, used_nodes);
6443 nodemask = node_to_cpumask(next_node);
6444 cpus_or(span, span, nodemask);
6451 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6454 * SMT sched-domains:
6456 #ifdef CONFIG_SCHED_SMT
6457 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6458 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6461 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6464 *sg = &per_cpu(sched_group_cpus, cpu);
6470 * multi-core sched-domains:
6472 #ifdef CONFIG_SCHED_MC
6473 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6474 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6477 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6479 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6482 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6483 cpus_and(mask, mask, *cpu_map);
6484 group = first_cpu(mask);
6486 *sg = &per_cpu(sched_group_core, group);
6489 #elif defined(CONFIG_SCHED_MC)
6491 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6494 *sg = &per_cpu(sched_group_core, cpu);
6499 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6500 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6503 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6506 #ifdef CONFIG_SCHED_MC
6507 cpumask_t mask = cpu_coregroup_map(cpu);
6508 cpus_and(mask, mask, *cpu_map);
6509 group = first_cpu(mask);
6510 #elif defined(CONFIG_SCHED_SMT)
6511 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6512 cpus_and(mask, mask, *cpu_map);
6513 group = first_cpu(mask);
6518 *sg = &per_cpu(sched_group_phys, group);
6524 * The init_sched_build_groups can't handle what we want to do with node
6525 * groups, so roll our own. Now each node has its own list of groups which
6526 * gets dynamically allocated.
6528 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6529 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6531 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6532 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6534 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6535 struct sched_group **sg)
6537 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6540 cpus_and(nodemask, nodemask, *cpu_map);
6541 group = first_cpu(nodemask);
6544 *sg = &per_cpu(sched_group_allnodes, group);
6548 static void init_numa_sched_groups_power(struct sched_group *group_head)
6550 struct sched_group *sg = group_head;
6556 for_each_cpu_mask(j, sg->cpumask) {
6557 struct sched_domain *sd;
6559 sd = &per_cpu(phys_domains, j);
6560 if (j != first_cpu(sd->groups->cpumask)) {
6562 * Only add "power" once for each
6568 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6571 } while (sg != group_head);
6576 /* Free memory allocated for various sched_group structures */
6577 static void free_sched_groups(const cpumask_t *cpu_map)
6581 for_each_cpu_mask(cpu, *cpu_map) {
6582 struct sched_group **sched_group_nodes
6583 = sched_group_nodes_bycpu[cpu];
6585 if (!sched_group_nodes)
6588 for (i = 0; i < MAX_NUMNODES; i++) {
6589 cpumask_t nodemask = node_to_cpumask(i);
6590 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6592 cpus_and(nodemask, nodemask, *cpu_map);
6593 if (cpus_empty(nodemask))
6603 if (oldsg != sched_group_nodes[i])
6606 kfree(sched_group_nodes);
6607 sched_group_nodes_bycpu[cpu] = NULL;
6611 static void free_sched_groups(const cpumask_t *cpu_map)
6617 * Initialize sched groups cpu_power.
6619 * cpu_power indicates the capacity of sched group, which is used while
6620 * distributing the load between different sched groups in a sched domain.
6621 * Typically cpu_power for all the groups in a sched domain will be same unless
6622 * there are asymmetries in the topology. If there are asymmetries, group
6623 * having more cpu_power will pickup more load compared to the group having
6626 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6627 * the maximum number of tasks a group can handle in the presence of other idle
6628 * or lightly loaded groups in the same sched domain.
6630 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6632 struct sched_domain *child;
6633 struct sched_group *group;
6635 WARN_ON(!sd || !sd->groups);
6637 if (cpu != first_cpu(sd->groups->cpumask))
6642 sd->groups->__cpu_power = 0;
6645 * For perf policy, if the groups in child domain share resources
6646 * (for example cores sharing some portions of the cache hierarchy
6647 * or SMT), then set this domain groups cpu_power such that each group
6648 * can handle only one task, when there are other idle groups in the
6649 * same sched domain.
6651 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6653 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6654 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6659 * add cpu_power of each child group to this groups cpu_power
6661 group = child->groups;
6663 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6664 group = group->next;
6665 } while (group != child->groups);
6669 * Build sched domains for a given set of cpus and attach the sched domains
6670 * to the individual cpus
6672 static int build_sched_domains(const cpumask_t *cpu_map)
6675 struct root_domain *rd;
6677 struct sched_group **sched_group_nodes = NULL;
6678 int sd_allnodes = 0;
6681 * Allocate the per-node list of sched groups
6683 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6685 if (!sched_group_nodes) {
6686 printk(KERN_WARNING "Can not alloc sched group node list\n");
6689 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6692 rd = alloc_rootdomain();
6694 printk(KERN_WARNING "Cannot alloc root domain\n");
6699 * Set up domains for cpus specified by the cpu_map.
6701 for_each_cpu_mask(i, *cpu_map) {
6702 struct sched_domain *sd = NULL, *p;
6703 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6705 cpus_and(nodemask, nodemask, *cpu_map);
6708 if (cpus_weight(*cpu_map) >
6709 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6710 sd = &per_cpu(allnodes_domains, i);
6711 *sd = SD_ALLNODES_INIT;
6712 sd->span = *cpu_map;
6713 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6719 sd = &per_cpu(node_domains, i);
6721 sd->span = sched_domain_node_span(cpu_to_node(i));
6725 cpus_and(sd->span, sd->span, *cpu_map);
6729 sd = &per_cpu(phys_domains, i);
6731 sd->span = nodemask;
6735 cpu_to_phys_group(i, cpu_map, &sd->groups);
6737 #ifdef CONFIG_SCHED_MC
6739 sd = &per_cpu(core_domains, i);
6741 sd->span = cpu_coregroup_map(i);
6742 cpus_and(sd->span, sd->span, *cpu_map);
6745 cpu_to_core_group(i, cpu_map, &sd->groups);
6748 #ifdef CONFIG_SCHED_SMT
6750 sd = &per_cpu(cpu_domains, i);
6751 *sd = SD_SIBLING_INIT;
6752 sd->span = per_cpu(cpu_sibling_map, i);
6753 cpus_and(sd->span, sd->span, *cpu_map);
6756 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6760 #ifdef CONFIG_SCHED_SMT
6761 /* Set up CPU (sibling) groups */
6762 for_each_cpu_mask(i, *cpu_map) {
6763 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6764 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6765 if (i != first_cpu(this_sibling_map))
6768 init_sched_build_groups(this_sibling_map, cpu_map,
6773 #ifdef CONFIG_SCHED_MC
6774 /* Set up multi-core groups */
6775 for_each_cpu_mask(i, *cpu_map) {
6776 cpumask_t this_core_map = cpu_coregroup_map(i);
6777 cpus_and(this_core_map, this_core_map, *cpu_map);
6778 if (i != first_cpu(this_core_map))
6780 init_sched_build_groups(this_core_map, cpu_map,
6781 &cpu_to_core_group);
6785 /* Set up physical groups */
6786 for (i = 0; i < MAX_NUMNODES; i++) {
6787 cpumask_t nodemask = node_to_cpumask(i);
6789 cpus_and(nodemask, nodemask, *cpu_map);
6790 if (cpus_empty(nodemask))
6793 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6797 /* Set up node groups */
6799 init_sched_build_groups(*cpu_map, cpu_map,
6800 &cpu_to_allnodes_group);
6802 for (i = 0; i < MAX_NUMNODES; i++) {
6803 /* Set up node groups */
6804 struct sched_group *sg, *prev;
6805 cpumask_t nodemask = node_to_cpumask(i);
6806 cpumask_t domainspan;
6807 cpumask_t covered = CPU_MASK_NONE;
6810 cpus_and(nodemask, nodemask, *cpu_map);
6811 if (cpus_empty(nodemask)) {
6812 sched_group_nodes[i] = NULL;
6816 domainspan = sched_domain_node_span(i);
6817 cpus_and(domainspan, domainspan, *cpu_map);
6819 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6821 printk(KERN_WARNING "Can not alloc domain group for "
6825 sched_group_nodes[i] = sg;
6826 for_each_cpu_mask(j, nodemask) {
6827 struct sched_domain *sd;
6829 sd = &per_cpu(node_domains, j);
6832 sg->__cpu_power = 0;
6833 sg->cpumask = nodemask;
6835 cpus_or(covered, covered, nodemask);
6838 for (j = 0; j < MAX_NUMNODES; j++) {
6839 cpumask_t tmp, notcovered;
6840 int n = (i + j) % MAX_NUMNODES;
6842 cpus_complement(notcovered, covered);
6843 cpus_and(tmp, notcovered, *cpu_map);
6844 cpus_and(tmp, tmp, domainspan);
6845 if (cpus_empty(tmp))
6848 nodemask = node_to_cpumask(n);
6849 cpus_and(tmp, tmp, nodemask);
6850 if (cpus_empty(tmp))
6853 sg = kmalloc_node(sizeof(struct sched_group),
6857 "Can not alloc domain group for node %d\n", j);
6860 sg->__cpu_power = 0;
6862 sg->next = prev->next;
6863 cpus_or(covered, covered, tmp);
6870 /* Calculate CPU power for physical packages and nodes */
6871 #ifdef CONFIG_SCHED_SMT
6872 for_each_cpu_mask(i, *cpu_map) {
6873 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6875 init_sched_groups_power(i, sd);
6878 #ifdef CONFIG_SCHED_MC
6879 for_each_cpu_mask(i, *cpu_map) {
6880 struct sched_domain *sd = &per_cpu(core_domains, i);
6882 init_sched_groups_power(i, sd);
6886 for_each_cpu_mask(i, *cpu_map) {
6887 struct sched_domain *sd = &per_cpu(phys_domains, i);
6889 init_sched_groups_power(i, sd);
6893 for (i = 0; i < MAX_NUMNODES; i++)
6894 init_numa_sched_groups_power(sched_group_nodes[i]);
6897 struct sched_group *sg;
6899 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6900 init_numa_sched_groups_power(sg);
6904 /* Attach the domains */
6905 for_each_cpu_mask(i, *cpu_map) {
6906 struct sched_domain *sd;
6907 #ifdef CONFIG_SCHED_SMT
6908 sd = &per_cpu(cpu_domains, i);
6909 #elif defined(CONFIG_SCHED_MC)
6910 sd = &per_cpu(core_domains, i);
6912 sd = &per_cpu(phys_domains, i);
6914 cpu_attach_domain(sd, rd, i);
6921 free_sched_groups(cpu_map);
6926 static cpumask_t *doms_cur; /* current sched domains */
6927 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6930 * Special case: If a kmalloc of a doms_cur partition (array of
6931 * cpumask_t) fails, then fallback to a single sched domain,
6932 * as determined by the single cpumask_t fallback_doms.
6934 static cpumask_t fallback_doms;
6936 void __attribute__((weak)) arch_update_cpu_topology(void)
6941 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6942 * For now this just excludes isolated cpus, but could be used to
6943 * exclude other special cases in the future.
6945 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6949 arch_update_cpu_topology();
6951 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6953 doms_cur = &fallback_doms;
6954 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6955 err = build_sched_domains(doms_cur);
6956 register_sched_domain_sysctl();
6961 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6963 free_sched_groups(cpu_map);
6967 * Detach sched domains from a group of cpus specified in cpu_map
6968 * These cpus will now be attached to the NULL domain
6970 static void detach_destroy_domains(const cpumask_t *cpu_map)
6974 unregister_sched_domain_sysctl();
6976 for_each_cpu_mask(i, *cpu_map)
6977 cpu_attach_domain(NULL, &def_root_domain, i);
6978 synchronize_sched();
6979 arch_destroy_sched_domains(cpu_map);
6983 * Partition sched domains as specified by the 'ndoms_new'
6984 * cpumasks in the array doms_new[] of cpumasks. This compares
6985 * doms_new[] to the current sched domain partitioning, doms_cur[].
6986 * It destroys each deleted domain and builds each new domain.
6988 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6989 * The masks don't intersect (don't overlap.) We should setup one
6990 * sched domain for each mask. CPUs not in any of the cpumasks will
6991 * not be load balanced. If the same cpumask appears both in the
6992 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6995 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6996 * ownership of it and will kfree it when done with it. If the caller
6997 * failed the kmalloc call, then it can pass in doms_new == NULL,
6998 * and partition_sched_domains() will fallback to the single partition
7001 * Call with hotplug lock held
7003 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7009 /* always unregister in case we don't destroy any domains */
7010 unregister_sched_domain_sysctl();
7012 if (doms_new == NULL) {
7014 doms_new = &fallback_doms;
7015 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7018 /* Destroy deleted domains */
7019 for (i = 0; i < ndoms_cur; i++) {
7020 for (j = 0; j < ndoms_new; j++) {
7021 if (cpus_equal(doms_cur[i], doms_new[j]))
7024 /* no match - a current sched domain not in new doms_new[] */
7025 detach_destroy_domains(doms_cur + i);
7030 /* Build new domains */
7031 for (i = 0; i < ndoms_new; i++) {
7032 for (j = 0; j < ndoms_cur; j++) {
7033 if (cpus_equal(doms_new[i], doms_cur[j]))
7036 /* no match - add a new doms_new */
7037 build_sched_domains(doms_new + i);
7042 /* Remember the new sched domains */
7043 if (doms_cur != &fallback_doms)
7045 doms_cur = doms_new;
7046 ndoms_cur = ndoms_new;
7048 register_sched_domain_sysctl();
7053 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7054 int arch_reinit_sched_domains(void)
7059 detach_destroy_domains(&cpu_online_map);
7060 err = arch_init_sched_domains(&cpu_online_map);
7066 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7070 if (buf[0] != '0' && buf[0] != '1')
7074 sched_smt_power_savings = (buf[0] == '1');
7076 sched_mc_power_savings = (buf[0] == '1');
7078 ret = arch_reinit_sched_domains();
7080 return ret ? ret : count;
7083 #ifdef CONFIG_SCHED_MC
7084 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7086 return sprintf(page, "%u\n", sched_mc_power_savings);
7088 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7089 const char *buf, size_t count)
7091 return sched_power_savings_store(buf, count, 0);
7093 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7094 sched_mc_power_savings_store);
7097 #ifdef CONFIG_SCHED_SMT
7098 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7100 return sprintf(page, "%u\n", sched_smt_power_savings);
7102 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7103 const char *buf, size_t count)
7105 return sched_power_savings_store(buf, count, 1);
7107 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7108 sched_smt_power_savings_store);
7111 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7115 #ifdef CONFIG_SCHED_SMT
7117 err = sysfs_create_file(&cls->kset.kobj,
7118 &attr_sched_smt_power_savings.attr);
7120 #ifdef CONFIG_SCHED_MC
7121 if (!err && mc_capable())
7122 err = sysfs_create_file(&cls->kset.kobj,
7123 &attr_sched_mc_power_savings.attr);
7130 * Force a reinitialization of the sched domains hierarchy. The domains
7131 * and groups cannot be updated in place without racing with the balancing
7132 * code, so we temporarily attach all running cpus to the NULL domain
7133 * which will prevent rebalancing while the sched domains are recalculated.
7135 static int update_sched_domains(struct notifier_block *nfb,
7136 unsigned long action, void *hcpu)
7139 case CPU_UP_PREPARE:
7140 case CPU_UP_PREPARE_FROZEN:
7141 case CPU_DOWN_PREPARE:
7142 case CPU_DOWN_PREPARE_FROZEN:
7143 detach_destroy_domains(&cpu_online_map);
7146 case CPU_UP_CANCELED:
7147 case CPU_UP_CANCELED_FROZEN:
7148 case CPU_DOWN_FAILED:
7149 case CPU_DOWN_FAILED_FROZEN:
7151 case CPU_ONLINE_FROZEN:
7153 case CPU_DEAD_FROZEN:
7155 * Fall through and re-initialise the domains.
7162 /* The hotplug lock is already held by cpu_up/cpu_down */
7163 arch_init_sched_domains(&cpu_online_map);
7168 void __init sched_init_smp(void)
7170 cpumask_t non_isolated_cpus;
7173 arch_init_sched_domains(&cpu_online_map);
7174 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7175 if (cpus_empty(non_isolated_cpus))
7176 cpu_set(smp_processor_id(), non_isolated_cpus);
7178 /* XXX: Theoretical race here - CPU may be hotplugged now */
7179 hotcpu_notifier(update_sched_domains, 0);
7181 /* Move init over to a non-isolated CPU */
7182 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7184 sched_init_granularity();
7187 void __init sched_init_smp(void)
7189 sched_init_granularity();
7191 #endif /* CONFIG_SMP */
7193 int in_sched_functions(unsigned long addr)
7195 return in_lock_functions(addr) ||
7196 (addr >= (unsigned long)__sched_text_start
7197 && addr < (unsigned long)__sched_text_end);
7200 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7202 cfs_rq->tasks_timeline = RB_ROOT;
7203 #ifdef CONFIG_FAIR_GROUP_SCHED
7206 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7209 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7211 struct rt_prio_array *array;
7214 array = &rt_rq->active;
7215 for (i = 0; i < MAX_RT_PRIO; i++) {
7216 INIT_LIST_HEAD(array->queue + i);
7217 __clear_bit(i, array->bitmap);
7219 /* delimiter for bitsearch: */
7220 __set_bit(MAX_RT_PRIO, array->bitmap);
7222 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7223 rt_rq->highest_prio = MAX_RT_PRIO;
7226 rt_rq->rt_nr_migratory = 0;
7227 rt_rq->overloaded = 0;
7231 rt_rq->rt_throttled = 0;
7233 #ifdef CONFIG_RT_GROUP_SCHED
7234 rt_rq->rt_nr_boosted = 0;
7239 #ifdef CONFIG_FAIR_GROUP_SCHED
7240 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7241 struct cfs_rq *cfs_rq, struct sched_entity *se,
7244 tg->cfs_rq[cpu] = cfs_rq;
7245 init_cfs_rq(cfs_rq, rq);
7248 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7251 se->cfs_rq = &rq->cfs;
7253 se->load.weight = tg->shares;
7254 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7259 #ifdef CONFIG_RT_GROUP_SCHED
7260 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7261 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7264 tg->rt_rq[cpu] = rt_rq;
7265 init_rt_rq(rt_rq, rq);
7267 rt_rq->rt_se = rt_se;
7269 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7271 tg->rt_se[cpu] = rt_se;
7272 rt_se->rt_rq = &rq->rt;
7273 rt_se->my_q = rt_rq;
7274 rt_se->parent = NULL;
7275 INIT_LIST_HEAD(&rt_se->run_list);
7279 void __init sched_init(void)
7281 int highest_cpu = 0;
7285 init_defrootdomain();
7288 #ifdef CONFIG_GROUP_SCHED
7289 list_add(&init_task_group.list, &task_groups);
7292 for_each_possible_cpu(i) {
7296 spin_lock_init(&rq->lock);
7297 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7300 update_last_tick_seen(rq);
7301 init_cfs_rq(&rq->cfs, rq);
7302 init_rt_rq(&rq->rt, rq);
7303 #ifdef CONFIG_FAIR_GROUP_SCHED
7304 init_task_group.shares = init_task_group_load;
7305 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7306 init_tg_cfs_entry(rq, &init_task_group,
7307 &per_cpu(init_cfs_rq, i),
7308 &per_cpu(init_sched_entity, i), i, 1);
7311 #ifdef CONFIG_RT_GROUP_SCHED
7312 init_task_group.rt_runtime =
7313 sysctl_sched_rt_runtime * NSEC_PER_USEC;
7314 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7315 init_tg_rt_entry(rq, &init_task_group,
7316 &per_cpu(init_rt_rq, i),
7317 &per_cpu(init_sched_rt_entity, i), i, 1);
7319 rq->rt_period_expire = 0;
7320 rq->rt_throttled = 0;
7322 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7323 rq->cpu_load[j] = 0;
7327 rq->active_balance = 0;
7328 rq->next_balance = jiffies;
7331 rq->migration_thread = NULL;
7332 INIT_LIST_HEAD(&rq->migration_queue);
7333 rq_attach_root(rq, &def_root_domain);
7336 atomic_set(&rq->nr_iowait, 0);
7340 set_load_weight(&init_task);
7342 #ifdef CONFIG_PREEMPT_NOTIFIERS
7343 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7347 nr_cpu_ids = highest_cpu + 1;
7348 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7351 #ifdef CONFIG_RT_MUTEXES
7352 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7356 * The boot idle thread does lazy MMU switching as well:
7358 atomic_inc(&init_mm.mm_count);
7359 enter_lazy_tlb(&init_mm, current);
7362 * Make us the idle thread. Technically, schedule() should not be
7363 * called from this thread, however somewhere below it might be,
7364 * but because we are the idle thread, we just pick up running again
7365 * when this runqueue becomes "idle".
7367 init_idle(current, smp_processor_id());
7369 * During early bootup we pretend to be a normal task:
7371 current->sched_class = &fair_sched_class;
7373 scheduler_running = 1;
7376 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7377 void __might_sleep(char *file, int line)
7380 static unsigned long prev_jiffy; /* ratelimiting */
7382 if ((in_atomic() || irqs_disabled()) &&
7383 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7384 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7386 prev_jiffy = jiffies;
7387 printk(KERN_ERR "BUG: sleeping function called from invalid"
7388 " context at %s:%d\n", file, line);
7389 printk("in_atomic():%d, irqs_disabled():%d\n",
7390 in_atomic(), irqs_disabled());
7391 debug_show_held_locks(current);
7392 if (irqs_disabled())
7393 print_irqtrace_events(current);
7398 EXPORT_SYMBOL(__might_sleep);
7401 #ifdef CONFIG_MAGIC_SYSRQ
7402 static void normalize_task(struct rq *rq, struct task_struct *p)
7405 update_rq_clock(rq);
7406 on_rq = p->se.on_rq;
7408 deactivate_task(rq, p, 0);
7409 __setscheduler(rq, p, SCHED_NORMAL, 0);
7411 activate_task(rq, p, 0);
7412 resched_task(rq->curr);
7416 void normalize_rt_tasks(void)
7418 struct task_struct *g, *p;
7419 unsigned long flags;
7422 read_lock_irqsave(&tasklist_lock, flags);
7423 do_each_thread(g, p) {
7425 * Only normalize user tasks:
7430 p->se.exec_start = 0;
7431 #ifdef CONFIG_SCHEDSTATS
7432 p->se.wait_start = 0;
7433 p->se.sleep_start = 0;
7434 p->se.block_start = 0;
7436 task_rq(p)->clock = 0;
7440 * Renice negative nice level userspace
7443 if (TASK_NICE(p) < 0 && p->mm)
7444 set_user_nice(p, 0);
7448 spin_lock(&p->pi_lock);
7449 rq = __task_rq_lock(p);
7451 normalize_task(rq, p);
7453 __task_rq_unlock(rq);
7454 spin_unlock(&p->pi_lock);
7455 } while_each_thread(g, p);
7457 read_unlock_irqrestore(&tasklist_lock, flags);
7460 #endif /* CONFIG_MAGIC_SYSRQ */
7464 * These functions are only useful for the IA64 MCA handling.
7466 * They can only be called when the whole system has been
7467 * stopped - every CPU needs to be quiescent, and no scheduling
7468 * activity can take place. Using them for anything else would
7469 * be a serious bug, and as a result, they aren't even visible
7470 * under any other configuration.
7474 * curr_task - return the current task for a given cpu.
7475 * @cpu: the processor in question.
7477 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7479 struct task_struct *curr_task(int cpu)
7481 return cpu_curr(cpu);
7485 * set_curr_task - set the current task for a given cpu.
7486 * @cpu: the processor in question.
7487 * @p: the task pointer to set.
7489 * Description: This function must only be used when non-maskable interrupts
7490 * are serviced on a separate stack. It allows the architecture to switch the
7491 * notion of the current task on a cpu in a non-blocking manner. This function
7492 * must be called with all CPU's synchronized, and interrupts disabled, the
7493 * and caller must save the original value of the current task (see
7494 * curr_task() above) and restore that value before reenabling interrupts and
7495 * re-starting the system.
7497 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7499 void set_curr_task(int cpu, struct task_struct *p)
7506 #ifdef CONFIG_GROUP_SCHED
7508 #ifdef CONFIG_FAIR_GROUP_SCHED
7509 static void free_fair_sched_group(struct task_group *tg)
7513 for_each_possible_cpu(i) {
7515 kfree(tg->cfs_rq[i]);
7524 static int alloc_fair_sched_group(struct task_group *tg)
7526 struct cfs_rq *cfs_rq;
7527 struct sched_entity *se;
7531 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7534 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7538 tg->shares = NICE_0_LOAD;
7540 for_each_possible_cpu(i) {
7543 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7544 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7548 se = kmalloc_node(sizeof(struct sched_entity),
7549 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7553 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7562 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7564 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7565 &cpu_rq(cpu)->leaf_cfs_rq_list);
7568 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7570 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7573 static inline void free_fair_sched_group(struct task_group *tg)
7577 static inline int alloc_fair_sched_group(struct task_group *tg)
7582 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7586 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7591 #ifdef CONFIG_RT_GROUP_SCHED
7592 static void free_rt_sched_group(struct task_group *tg)
7596 for_each_possible_cpu(i) {
7598 kfree(tg->rt_rq[i]);
7600 kfree(tg->rt_se[i]);
7607 static int alloc_rt_sched_group(struct task_group *tg)
7609 struct rt_rq *rt_rq;
7610 struct sched_rt_entity *rt_se;
7614 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7617 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7623 for_each_possible_cpu(i) {
7626 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7627 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7631 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7632 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7636 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7645 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7647 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7648 &cpu_rq(cpu)->leaf_rt_rq_list);
7651 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7653 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7656 static inline void free_rt_sched_group(struct task_group *tg)
7660 static inline int alloc_rt_sched_group(struct task_group *tg)
7665 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7669 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7674 static void free_sched_group(struct task_group *tg)
7676 free_fair_sched_group(tg);
7677 free_rt_sched_group(tg);
7681 /* allocate runqueue etc for a new task group */
7682 struct task_group *sched_create_group(void)
7684 struct task_group *tg;
7685 unsigned long flags;
7688 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7690 return ERR_PTR(-ENOMEM);
7692 if (!alloc_fair_sched_group(tg))
7695 if (!alloc_rt_sched_group(tg))
7698 spin_lock_irqsave(&task_group_lock, flags);
7699 for_each_possible_cpu(i) {
7700 register_fair_sched_group(tg, i);
7701 register_rt_sched_group(tg, i);
7703 list_add_rcu(&tg->list, &task_groups);
7704 spin_unlock_irqrestore(&task_group_lock, flags);
7709 free_sched_group(tg);
7710 return ERR_PTR(-ENOMEM);
7713 /* rcu callback to free various structures associated with a task group */
7714 static void free_sched_group_rcu(struct rcu_head *rhp)
7716 /* now it should be safe to free those cfs_rqs */
7717 free_sched_group(container_of(rhp, struct task_group, rcu));
7720 /* Destroy runqueue etc associated with a task group */
7721 void sched_destroy_group(struct task_group *tg)
7723 unsigned long flags;
7726 spin_lock_irqsave(&task_group_lock, flags);
7727 for_each_possible_cpu(i) {
7728 unregister_fair_sched_group(tg, i);
7729 unregister_rt_sched_group(tg, i);
7731 list_del_rcu(&tg->list);
7732 spin_unlock_irqrestore(&task_group_lock, flags);
7734 /* wait for possible concurrent references to cfs_rqs complete */
7735 call_rcu(&tg->rcu, free_sched_group_rcu);
7738 /* change task's runqueue when it moves between groups.
7739 * The caller of this function should have put the task in its new group
7740 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7741 * reflect its new group.
7743 void sched_move_task(struct task_struct *tsk)
7746 unsigned long flags;
7749 rq = task_rq_lock(tsk, &flags);
7751 update_rq_clock(rq);
7753 running = task_current(rq, tsk);
7754 on_rq = tsk->se.on_rq;
7757 dequeue_task(rq, tsk, 0);
7758 if (unlikely(running))
7759 tsk->sched_class->put_prev_task(rq, tsk);
7761 set_task_rq(tsk, task_cpu(tsk));
7763 #ifdef CONFIG_FAIR_GROUP_SCHED
7764 if (tsk->sched_class->moved_group)
7765 tsk->sched_class->moved_group(tsk);
7768 if (unlikely(running))
7769 tsk->sched_class->set_curr_task(rq);
7771 enqueue_task(rq, tsk, 0);
7773 task_rq_unlock(rq, &flags);
7776 #ifdef CONFIG_FAIR_GROUP_SCHED
7777 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7779 struct cfs_rq *cfs_rq = se->cfs_rq;
7780 struct rq *rq = cfs_rq->rq;
7783 spin_lock_irq(&rq->lock);
7787 dequeue_entity(cfs_rq, se, 0);
7789 se->load.weight = shares;
7790 se->load.inv_weight = div64_64((1ULL<<32), shares);
7793 enqueue_entity(cfs_rq, se, 0);
7795 spin_unlock_irq(&rq->lock);
7798 static DEFINE_MUTEX(shares_mutex);
7800 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7803 unsigned long flags;
7806 * A weight of 0 or 1 can cause arithmetics problems.
7807 * (The default weight is 1024 - so there's no practical
7808 * limitation from this.)
7813 mutex_lock(&shares_mutex);
7814 if (tg->shares == shares)
7817 spin_lock_irqsave(&task_group_lock, flags);
7818 for_each_possible_cpu(i)
7819 unregister_fair_sched_group(tg, i);
7820 spin_unlock_irqrestore(&task_group_lock, flags);
7822 /* wait for any ongoing reference to this group to finish */
7823 synchronize_sched();
7826 * Now we are free to modify the group's share on each cpu
7827 * w/o tripping rebalance_share or load_balance_fair.
7829 tg->shares = shares;
7830 for_each_possible_cpu(i)
7831 set_se_shares(tg->se[i], shares);
7834 * Enable load balance activity on this group, by inserting it back on
7835 * each cpu's rq->leaf_cfs_rq_list.
7837 spin_lock_irqsave(&task_group_lock, flags);
7838 for_each_possible_cpu(i)
7839 register_fair_sched_group(tg, i);
7840 spin_unlock_irqrestore(&task_group_lock, flags);
7842 mutex_unlock(&shares_mutex);
7846 unsigned long sched_group_shares(struct task_group *tg)
7852 #ifdef CONFIG_RT_GROUP_SCHED
7854 * Ensure that the real time constraints are schedulable.
7856 static DEFINE_MUTEX(rt_constraints_mutex);
7858 static unsigned long to_ratio(u64 period, u64 runtime)
7860 if (runtime == RUNTIME_INF)
7863 return div64_64(runtime << 16, period);
7866 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7868 struct task_group *tgi;
7869 unsigned long total = 0;
7870 unsigned long global_ratio =
7871 to_ratio(sysctl_sched_rt_period,
7872 sysctl_sched_rt_runtime < 0 ?
7873 RUNTIME_INF : sysctl_sched_rt_runtime);
7876 list_for_each_entry_rcu(tgi, &task_groups, list) {
7880 total += to_ratio(period, tgi->rt_runtime);
7884 return total + to_ratio(period, runtime) < global_ratio;
7887 /* Must be called with tasklist_lock held */
7888 static inline int tg_has_rt_tasks(struct task_group *tg)
7890 struct task_struct *g, *p;
7891 do_each_thread(g, p) {
7892 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
7894 } while_each_thread(g, p);
7898 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7900 u64 rt_runtime, rt_period;
7903 rt_period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
7904 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7905 if (rt_runtime_us == -1)
7906 rt_runtime = RUNTIME_INF;
7908 mutex_lock(&rt_constraints_mutex);
7909 read_lock(&tasklist_lock);
7910 if (rt_runtime_us == 0 && tg_has_rt_tasks(tg)) {
7914 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7918 tg->rt_runtime = rt_runtime;
7920 read_unlock(&tasklist_lock);
7921 mutex_unlock(&rt_constraints_mutex);
7926 long sched_group_rt_runtime(struct task_group *tg)
7930 if (tg->rt_runtime == RUNTIME_INF)
7933 rt_runtime_us = tg->rt_runtime;
7934 do_div(rt_runtime_us, NSEC_PER_USEC);
7935 return rt_runtime_us;
7938 #endif /* CONFIG_GROUP_SCHED */
7940 #ifdef CONFIG_CGROUP_SCHED
7942 /* return corresponding task_group object of a cgroup */
7943 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7945 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7946 struct task_group, css);
7949 static struct cgroup_subsys_state *
7950 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7952 struct task_group *tg;
7954 if (!cgrp->parent) {
7955 /* This is early initialization for the top cgroup */
7956 init_task_group.css.cgroup = cgrp;
7957 return &init_task_group.css;
7960 /* we support only 1-level deep hierarchical scheduler atm */
7961 if (cgrp->parent->parent)
7962 return ERR_PTR(-EINVAL);
7964 tg = sched_create_group();
7966 return ERR_PTR(-ENOMEM);
7968 /* Bind the cgroup to task_group object we just created */
7969 tg->css.cgroup = cgrp;
7975 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7977 struct task_group *tg = cgroup_tg(cgrp);
7979 sched_destroy_group(tg);
7983 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7984 struct task_struct *tsk)
7986 #ifdef CONFIG_RT_GROUP_SCHED
7987 /* Don't accept realtime tasks when there is no way for them to run */
7988 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7991 /* We don't support RT-tasks being in separate groups */
7992 if (tsk->sched_class != &fair_sched_class)
8000 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8001 struct cgroup *old_cont, struct task_struct *tsk)
8003 sched_move_task(tsk);
8006 #ifdef CONFIG_FAIR_GROUP_SCHED
8007 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8010 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8013 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8015 struct task_group *tg = cgroup_tg(cgrp);
8017 return (u64) tg->shares;
8021 #ifdef CONFIG_RT_GROUP_SCHED
8022 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8024 const char __user *userbuf,
8025 size_t nbytes, loff_t *unused_ppos)
8034 if (nbytes >= sizeof(buffer))
8036 if (copy_from_user(buffer, userbuf, nbytes))
8039 buffer[nbytes] = 0; /* nul-terminate */
8041 /* strip newline if necessary */
8042 if (nbytes && (buffer[nbytes-1] == '\n'))
8043 buffer[nbytes-1] = 0;
8044 val = simple_strtoll(buffer, &end, 0);
8048 /* Pass to subsystem */
8049 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8055 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8057 char __user *buf, size_t nbytes,
8061 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8062 int len = sprintf(tmp, "%ld\n", val);
8064 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8068 static struct cftype cpu_files[] = {
8069 #ifdef CONFIG_FAIR_GROUP_SCHED
8072 .read_uint = cpu_shares_read_uint,
8073 .write_uint = cpu_shares_write_uint,
8076 #ifdef CONFIG_RT_GROUP_SCHED
8078 .name = "rt_runtime_us",
8079 .read = cpu_rt_runtime_read,
8080 .write = cpu_rt_runtime_write,
8085 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8087 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8090 struct cgroup_subsys cpu_cgroup_subsys = {
8092 .create = cpu_cgroup_create,
8093 .destroy = cpu_cgroup_destroy,
8094 .can_attach = cpu_cgroup_can_attach,
8095 .attach = cpu_cgroup_attach,
8096 .populate = cpu_cgroup_populate,
8097 .subsys_id = cpu_cgroup_subsys_id,
8101 #endif /* CONFIG_CGROUP_SCHED */
8103 #ifdef CONFIG_CGROUP_CPUACCT
8106 * CPU accounting code for task groups.
8108 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8109 * (balbir@in.ibm.com).
8112 /* track cpu usage of a group of tasks */
8114 struct cgroup_subsys_state css;
8115 /* cpuusage holds pointer to a u64-type object on every cpu */
8119 struct cgroup_subsys cpuacct_subsys;
8121 /* return cpu accounting group corresponding to this container */
8122 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
8124 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
8125 struct cpuacct, css);
8128 /* return cpu accounting group to which this task belongs */
8129 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8131 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8132 struct cpuacct, css);
8135 /* create a new cpu accounting group */
8136 static struct cgroup_subsys_state *cpuacct_create(
8137 struct cgroup_subsys *ss, struct cgroup *cont)
8139 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8142 return ERR_PTR(-ENOMEM);
8144 ca->cpuusage = alloc_percpu(u64);
8145 if (!ca->cpuusage) {
8147 return ERR_PTR(-ENOMEM);
8153 /* destroy an existing cpu accounting group */
8155 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8157 struct cpuacct *ca = cgroup_ca(cont);
8159 free_percpu(ca->cpuusage);
8163 /* return total cpu usage (in nanoseconds) of a group */
8164 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8166 struct cpuacct *ca = cgroup_ca(cont);
8167 u64 totalcpuusage = 0;
8170 for_each_possible_cpu(i) {
8171 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8174 * Take rq->lock to make 64-bit addition safe on 32-bit
8177 spin_lock_irq(&cpu_rq(i)->lock);
8178 totalcpuusage += *cpuusage;
8179 spin_unlock_irq(&cpu_rq(i)->lock);
8182 return totalcpuusage;
8185 static struct cftype files[] = {
8188 .read_uint = cpuusage_read,
8192 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8194 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8198 * charge this task's execution time to its accounting group.
8200 * called with rq->lock held.
8202 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8206 if (!cpuacct_subsys.active)
8211 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8213 *cpuusage += cputime;
8217 struct cgroup_subsys cpuacct_subsys = {
8219 .create = cpuacct_create,
8220 .destroy = cpuacct_destroy,
8221 .populate = cpuacct_populate,
8222 .subsys_id = cpuacct_subsys_id,
8224 #endif /* CONFIG_CGROUP_CPUACCT */